1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * NET3 Protocol independent device support routines. 4 * 5 * Derived from the non IP parts of dev.c 1.0.19 6 * Authors: Ross Biro 7 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 8 * Mark Evans, <evansmp@uhura.aston.ac.uk> 9 * 10 * Additional Authors: 11 * Florian la Roche <rzsfl@rz.uni-sb.de> 12 * Alan Cox <gw4pts@gw4pts.ampr.org> 13 * David Hinds <dahinds@users.sourceforge.net> 14 * Alexey Kuznetsov <kuznet@ms2.inr.ac.ru> 15 * Adam Sulmicki <adam@cfar.umd.edu> 16 * Pekka Riikonen <priikone@poesidon.pspt.fi> 17 * 18 * Changes: 19 * D.J. Barrow : Fixed bug where dev->refcnt gets set 20 * to 2 if register_netdev gets called 21 * before net_dev_init & also removed a 22 * few lines of code in the process. 23 * Alan Cox : device private ioctl copies fields back. 24 * Alan Cox : Transmit queue code does relevant 25 * stunts to keep the queue safe. 26 * Alan Cox : Fixed double lock. 27 * Alan Cox : Fixed promisc NULL pointer trap 28 * ???????? : Support the full private ioctl range 29 * Alan Cox : Moved ioctl permission check into 30 * drivers 31 * Tim Kordas : SIOCADDMULTI/SIOCDELMULTI 32 * Alan Cox : 100 backlog just doesn't cut it when 33 * you start doing multicast video 8) 34 * Alan Cox : Rewrote net_bh and list manager. 35 * Alan Cox : Fix ETH_P_ALL echoback lengths. 36 * Alan Cox : Took out transmit every packet pass 37 * Saved a few bytes in the ioctl handler 38 * Alan Cox : Network driver sets packet type before 39 * calling netif_rx. Saves a function 40 * call a packet. 41 * Alan Cox : Hashed net_bh() 42 * Richard Kooijman: Timestamp fixes. 43 * Alan Cox : Wrong field in SIOCGIFDSTADDR 44 * Alan Cox : Device lock protection. 45 * Alan Cox : Fixed nasty side effect of device close 46 * changes. 47 * Rudi Cilibrasi : Pass the right thing to 48 * set_mac_address() 49 * Dave Miller : 32bit quantity for the device lock to 50 * make it work out on a Sparc. 51 * Bjorn Ekwall : Added KERNELD hack. 52 * Alan Cox : Cleaned up the backlog initialise. 53 * Craig Metz : SIOCGIFCONF fix if space for under 54 * 1 device. 55 * Thomas Bogendoerfer : Return ENODEV for dev_open, if there 56 * is no device open function. 57 * Andi Kleen : Fix error reporting for SIOCGIFCONF 58 * Michael Chastain : Fix signed/unsigned for SIOCGIFCONF 59 * Cyrus Durgin : Cleaned for KMOD 60 * Adam Sulmicki : Bug Fix : Network Device Unload 61 * A network device unload needs to purge 62 * the backlog queue. 63 * Paul Rusty Russell : SIOCSIFNAME 64 * Pekka Riikonen : Netdev boot-time settings code 65 * Andrew Morton : Make unregister_netdevice wait 66 * indefinitely on dev->refcnt 67 * J Hadi Salim : - Backlog queue sampling 68 * - netif_rx() feedback 69 */ 70 71 #include <linux/uaccess.h> 72 #include <linux/bitmap.h> 73 #include <linux/capability.h> 74 #include <linux/cpu.h> 75 #include <linux/types.h> 76 #include <linux/kernel.h> 77 #include <linux/hash.h> 78 #include <linux/slab.h> 79 #include <linux/sched.h> 80 #include <linux/sched/isolation.h> 81 #include <linux/sched/mm.h> 82 #include <linux/smpboot.h> 83 #include <linux/mutex.h> 84 #include <linux/rwsem.h> 85 #include <linux/string.h> 86 #include <linux/mm.h> 87 #include <linux/socket.h> 88 #include <linux/sockios.h> 89 #include <linux/errno.h> 90 #include <linux/interrupt.h> 91 #include <linux/if_ether.h> 92 #include <linux/netdevice.h> 93 #include <linux/etherdevice.h> 94 #include <linux/ethtool.h> 95 #include <linux/ethtool_netlink.h> 96 #include <linux/skbuff.h> 97 #include <linux/kthread.h> 98 #include <linux/bpf.h> 99 #include <linux/bpf_trace.h> 100 #include <net/net_namespace.h> 101 #include <net/sock.h> 102 #include <net/busy_poll.h> 103 #include <linux/rtnetlink.h> 104 #include <linux/stat.h> 105 #include <net/dsa.h> 106 #include <net/dst.h> 107 #include <net/dst_metadata.h> 108 #include <net/gro.h> 109 #include <net/netdev_queues.h> 110 #include <net/pkt_sched.h> 111 #include <net/pkt_cls.h> 112 #include <net/checksum.h> 113 #include <net/xfrm.h> 114 #include <net/tcx.h> 115 #include <linux/highmem.h> 116 #include <linux/init.h> 117 #include <linux/module.h> 118 #include <linux/netpoll.h> 119 #include <linux/rcupdate.h> 120 #include <linux/delay.h> 121 #include <net/iw_handler.h> 122 #include <asm/current.h> 123 #include <linux/audit.h> 124 #include <linux/dmaengine.h> 125 #include <linux/err.h> 126 #include <linux/ctype.h> 127 #include <linux/if_arp.h> 128 #include <linux/if_vlan.h> 129 #include <linux/ip.h> 130 #include <net/ip.h> 131 #include <net/mpls.h> 132 #include <linux/ipv6.h> 133 #include <linux/in.h> 134 #include <linux/jhash.h> 135 #include <linux/random.h> 136 #include <trace/events/napi.h> 137 #include <trace/events/net.h> 138 #include <trace/events/skb.h> 139 #include <trace/events/qdisc.h> 140 #include <trace/events/xdp.h> 141 #include <linux/inetdevice.h> 142 #include <linux/cpu_rmap.h> 143 #include <linux/static_key.h> 144 #include <linux/hashtable.h> 145 #include <linux/vmalloc.h> 146 #include <linux/if_macvlan.h> 147 #include <linux/errqueue.h> 148 #include <linux/hrtimer.h> 149 #include <linux/netfilter_netdev.h> 150 #include <linux/crash_dump.h> 151 #include <linux/sctp.h> 152 #include <net/udp_tunnel.h> 153 #include <linux/net_namespace.h> 154 #include <linux/indirect_call_wrapper.h> 155 #include <net/devlink.h> 156 #include <linux/pm_runtime.h> 157 #include <linux/prandom.h> 158 #include <linux/once_lite.h> 159 #include <net/netdev_rx_queue.h> 160 #include <net/page_pool/types.h> 161 #include <net/page_pool/helpers.h> 162 #include <net/rps.h> 163 #include <linux/phy_link_topology.h> 164 165 #include "dev.h" 166 #include "devmem.h" 167 #include "net-sysfs.h" 168 169 static DEFINE_SPINLOCK(ptype_lock); 170 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly; 171 172 static int netif_rx_internal(struct sk_buff *skb); 173 static int call_netdevice_notifiers_extack(unsigned long val, 174 struct net_device *dev, 175 struct netlink_ext_ack *extack); 176 177 static DEFINE_MUTEX(ifalias_mutex); 178 179 /* protects napi_hash addition/deletion and napi_gen_id */ 180 static DEFINE_SPINLOCK(napi_hash_lock); 181 182 static unsigned int napi_gen_id = NR_CPUS; 183 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8); 184 185 static inline void dev_base_seq_inc(struct net *net) 186 { 187 unsigned int val = net->dev_base_seq + 1; 188 189 WRITE_ONCE(net->dev_base_seq, val ?: 1); 190 } 191 192 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name) 193 { 194 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ)); 195 196 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)]; 197 } 198 199 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex) 200 { 201 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)]; 202 } 203 204 #ifndef CONFIG_PREEMPT_RT 205 206 static DEFINE_STATIC_KEY_FALSE(use_backlog_threads_key); 207 208 static int __init setup_backlog_napi_threads(char *arg) 209 { 210 static_branch_enable(&use_backlog_threads_key); 211 return 0; 212 } 213 early_param("thread_backlog_napi", setup_backlog_napi_threads); 214 215 static bool use_backlog_threads(void) 216 { 217 return static_branch_unlikely(&use_backlog_threads_key); 218 } 219 220 #else 221 222 static bool use_backlog_threads(void) 223 { 224 return true; 225 } 226 227 #endif 228 229 static inline void backlog_lock_irq_save(struct softnet_data *sd, 230 unsigned long *flags) 231 { 232 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 233 spin_lock_irqsave(&sd->input_pkt_queue.lock, *flags); 234 else 235 local_irq_save(*flags); 236 } 237 238 static inline void backlog_lock_irq_disable(struct softnet_data *sd) 239 { 240 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 241 spin_lock_irq(&sd->input_pkt_queue.lock); 242 else 243 local_irq_disable(); 244 } 245 246 static inline void backlog_unlock_irq_restore(struct softnet_data *sd, 247 unsigned long *flags) 248 { 249 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 250 spin_unlock_irqrestore(&sd->input_pkt_queue.lock, *flags); 251 else 252 local_irq_restore(*flags); 253 } 254 255 static inline void backlog_unlock_irq_enable(struct softnet_data *sd) 256 { 257 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 258 spin_unlock_irq(&sd->input_pkt_queue.lock); 259 else 260 local_irq_enable(); 261 } 262 263 static struct netdev_name_node *netdev_name_node_alloc(struct net_device *dev, 264 const char *name) 265 { 266 struct netdev_name_node *name_node; 267 268 name_node = kmalloc(sizeof(*name_node), GFP_KERNEL); 269 if (!name_node) 270 return NULL; 271 INIT_HLIST_NODE(&name_node->hlist); 272 name_node->dev = dev; 273 name_node->name = name; 274 return name_node; 275 } 276 277 static struct netdev_name_node * 278 netdev_name_node_head_alloc(struct net_device *dev) 279 { 280 struct netdev_name_node *name_node; 281 282 name_node = netdev_name_node_alloc(dev, dev->name); 283 if (!name_node) 284 return NULL; 285 INIT_LIST_HEAD(&name_node->list); 286 return name_node; 287 } 288 289 static void netdev_name_node_free(struct netdev_name_node *name_node) 290 { 291 kfree(name_node); 292 } 293 294 static void netdev_name_node_add(struct net *net, 295 struct netdev_name_node *name_node) 296 { 297 hlist_add_head_rcu(&name_node->hlist, 298 dev_name_hash(net, name_node->name)); 299 } 300 301 static void netdev_name_node_del(struct netdev_name_node *name_node) 302 { 303 hlist_del_rcu(&name_node->hlist); 304 } 305 306 static struct netdev_name_node *netdev_name_node_lookup(struct net *net, 307 const char *name) 308 { 309 struct hlist_head *head = dev_name_hash(net, name); 310 struct netdev_name_node *name_node; 311 312 hlist_for_each_entry(name_node, head, hlist) 313 if (!strcmp(name_node->name, name)) 314 return name_node; 315 return NULL; 316 } 317 318 static struct netdev_name_node *netdev_name_node_lookup_rcu(struct net *net, 319 const char *name) 320 { 321 struct hlist_head *head = dev_name_hash(net, name); 322 struct netdev_name_node *name_node; 323 324 hlist_for_each_entry_rcu(name_node, head, hlist) 325 if (!strcmp(name_node->name, name)) 326 return name_node; 327 return NULL; 328 } 329 330 bool netdev_name_in_use(struct net *net, const char *name) 331 { 332 return netdev_name_node_lookup(net, name); 333 } 334 EXPORT_SYMBOL(netdev_name_in_use); 335 336 int netdev_name_node_alt_create(struct net_device *dev, const char *name) 337 { 338 struct netdev_name_node *name_node; 339 struct net *net = dev_net(dev); 340 341 name_node = netdev_name_node_lookup(net, name); 342 if (name_node) 343 return -EEXIST; 344 name_node = netdev_name_node_alloc(dev, name); 345 if (!name_node) 346 return -ENOMEM; 347 netdev_name_node_add(net, name_node); 348 /* The node that holds dev->name acts as a head of per-device list. */ 349 list_add_tail_rcu(&name_node->list, &dev->name_node->list); 350 351 return 0; 352 } 353 354 static void netdev_name_node_alt_free(struct rcu_head *head) 355 { 356 struct netdev_name_node *name_node = 357 container_of(head, struct netdev_name_node, rcu); 358 359 kfree(name_node->name); 360 netdev_name_node_free(name_node); 361 } 362 363 static void __netdev_name_node_alt_destroy(struct netdev_name_node *name_node) 364 { 365 netdev_name_node_del(name_node); 366 list_del(&name_node->list); 367 call_rcu(&name_node->rcu, netdev_name_node_alt_free); 368 } 369 370 int netdev_name_node_alt_destroy(struct net_device *dev, const char *name) 371 { 372 struct netdev_name_node *name_node; 373 struct net *net = dev_net(dev); 374 375 name_node = netdev_name_node_lookup(net, name); 376 if (!name_node) 377 return -ENOENT; 378 /* lookup might have found our primary name or a name belonging 379 * to another device. 380 */ 381 if (name_node == dev->name_node || name_node->dev != dev) 382 return -EINVAL; 383 384 __netdev_name_node_alt_destroy(name_node); 385 return 0; 386 } 387 388 static void netdev_name_node_alt_flush(struct net_device *dev) 389 { 390 struct netdev_name_node *name_node, *tmp; 391 392 list_for_each_entry_safe(name_node, tmp, &dev->name_node->list, list) { 393 list_del(&name_node->list); 394 netdev_name_node_alt_free(&name_node->rcu); 395 } 396 } 397 398 /* Device list insertion */ 399 static void list_netdevice(struct net_device *dev) 400 { 401 struct netdev_name_node *name_node; 402 struct net *net = dev_net(dev); 403 404 ASSERT_RTNL(); 405 406 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head); 407 netdev_name_node_add(net, dev->name_node); 408 hlist_add_head_rcu(&dev->index_hlist, 409 dev_index_hash(net, dev->ifindex)); 410 411 netdev_for_each_altname(dev, name_node) 412 netdev_name_node_add(net, name_node); 413 414 /* We reserved the ifindex, this can't fail */ 415 WARN_ON(xa_store(&net->dev_by_index, dev->ifindex, dev, GFP_KERNEL)); 416 417 dev_base_seq_inc(net); 418 } 419 420 /* Device list removal 421 * caller must respect a RCU grace period before freeing/reusing dev 422 */ 423 static void unlist_netdevice(struct net_device *dev) 424 { 425 struct netdev_name_node *name_node; 426 struct net *net = dev_net(dev); 427 428 ASSERT_RTNL(); 429 430 xa_erase(&net->dev_by_index, dev->ifindex); 431 432 netdev_for_each_altname(dev, name_node) 433 netdev_name_node_del(name_node); 434 435 /* Unlink dev from the device chain */ 436 list_del_rcu(&dev->dev_list); 437 netdev_name_node_del(dev->name_node); 438 hlist_del_rcu(&dev->index_hlist); 439 440 dev_base_seq_inc(dev_net(dev)); 441 } 442 443 /* 444 * Our notifier list 445 */ 446 447 static RAW_NOTIFIER_HEAD(netdev_chain); 448 449 /* 450 * Device drivers call our routines to queue packets here. We empty the 451 * queue in the local softnet handler. 452 */ 453 454 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data) = { 455 .process_queue_bh_lock = INIT_LOCAL_LOCK(process_queue_bh_lock), 456 }; 457 EXPORT_PER_CPU_SYMBOL(softnet_data); 458 459 /* Page_pool has a lockless array/stack to alloc/recycle pages. 460 * PP consumers must pay attention to run APIs in the appropriate context 461 * (e.g. NAPI context). 462 */ 463 DEFINE_PER_CPU(struct page_pool *, system_page_pool); 464 465 #ifdef CONFIG_LOCKDEP 466 /* 467 * register_netdevice() inits txq->_xmit_lock and sets lockdep class 468 * according to dev->type 469 */ 470 static const unsigned short netdev_lock_type[] = { 471 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25, 472 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET, 473 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM, 474 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP, 475 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD, 476 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25, 477 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP, 478 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD, 479 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI, 480 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE, 481 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET, 482 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL, 483 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM, 484 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE, 485 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE}; 486 487 static const char *const netdev_lock_name[] = { 488 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25", 489 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET", 490 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM", 491 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP", 492 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD", 493 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25", 494 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP", 495 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD", 496 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI", 497 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE", 498 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET", 499 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL", 500 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM", 501 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE", 502 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"}; 503 504 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)]; 505 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)]; 506 507 static inline unsigned short netdev_lock_pos(unsigned short dev_type) 508 { 509 int i; 510 511 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++) 512 if (netdev_lock_type[i] == dev_type) 513 return i; 514 /* the last key is used by default */ 515 return ARRAY_SIZE(netdev_lock_type) - 1; 516 } 517 518 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 519 unsigned short dev_type) 520 { 521 int i; 522 523 i = netdev_lock_pos(dev_type); 524 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i], 525 netdev_lock_name[i]); 526 } 527 528 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 529 { 530 int i; 531 532 i = netdev_lock_pos(dev->type); 533 lockdep_set_class_and_name(&dev->addr_list_lock, 534 &netdev_addr_lock_key[i], 535 netdev_lock_name[i]); 536 } 537 #else 538 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 539 unsigned short dev_type) 540 { 541 } 542 543 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 544 { 545 } 546 #endif 547 548 /******************************************************************************* 549 * 550 * Protocol management and registration routines 551 * 552 *******************************************************************************/ 553 554 555 /* 556 * Add a protocol ID to the list. Now that the input handler is 557 * smarter we can dispense with all the messy stuff that used to be 558 * here. 559 * 560 * BEWARE!!! Protocol handlers, mangling input packets, 561 * MUST BE last in hash buckets and checking protocol handlers 562 * MUST start from promiscuous ptype_all chain in net_bh. 563 * It is true now, do not change it. 564 * Explanation follows: if protocol handler, mangling packet, will 565 * be the first on list, it is not able to sense, that packet 566 * is cloned and should be copied-on-write, so that it will 567 * change it and subsequent readers will get broken packet. 568 * --ANK (980803) 569 */ 570 571 static inline struct list_head *ptype_head(const struct packet_type *pt) 572 { 573 if (pt->type == htons(ETH_P_ALL)) 574 return pt->dev ? &pt->dev->ptype_all : &net_hotdata.ptype_all; 575 else 576 return pt->dev ? &pt->dev->ptype_specific : 577 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK]; 578 } 579 580 /** 581 * dev_add_pack - add packet handler 582 * @pt: packet type declaration 583 * 584 * Add a protocol handler to the networking stack. The passed &packet_type 585 * is linked into kernel lists and may not be freed until it has been 586 * removed from the kernel lists. 587 * 588 * This call does not sleep therefore it can not 589 * guarantee all CPU's that are in middle of receiving packets 590 * will see the new packet type (until the next received packet). 591 */ 592 593 void dev_add_pack(struct packet_type *pt) 594 { 595 struct list_head *head = ptype_head(pt); 596 597 spin_lock(&ptype_lock); 598 list_add_rcu(&pt->list, head); 599 spin_unlock(&ptype_lock); 600 } 601 EXPORT_SYMBOL(dev_add_pack); 602 603 /** 604 * __dev_remove_pack - remove packet handler 605 * @pt: packet type declaration 606 * 607 * Remove a protocol handler that was previously added to the kernel 608 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 609 * from the kernel lists and can be freed or reused once this function 610 * returns. 611 * 612 * The packet type might still be in use by receivers 613 * and must not be freed until after all the CPU's have gone 614 * through a quiescent state. 615 */ 616 void __dev_remove_pack(struct packet_type *pt) 617 { 618 struct list_head *head = ptype_head(pt); 619 struct packet_type *pt1; 620 621 spin_lock(&ptype_lock); 622 623 list_for_each_entry(pt1, head, list) { 624 if (pt == pt1) { 625 list_del_rcu(&pt->list); 626 goto out; 627 } 628 } 629 630 pr_warn("dev_remove_pack: %p not found\n", pt); 631 out: 632 spin_unlock(&ptype_lock); 633 } 634 EXPORT_SYMBOL(__dev_remove_pack); 635 636 /** 637 * dev_remove_pack - remove packet handler 638 * @pt: packet type declaration 639 * 640 * Remove a protocol handler that was previously added to the kernel 641 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 642 * from the kernel lists and can be freed or reused once this function 643 * returns. 644 * 645 * This call sleeps to guarantee that no CPU is looking at the packet 646 * type after return. 647 */ 648 void dev_remove_pack(struct packet_type *pt) 649 { 650 __dev_remove_pack(pt); 651 652 synchronize_net(); 653 } 654 EXPORT_SYMBOL(dev_remove_pack); 655 656 657 /******************************************************************************* 658 * 659 * Device Interface Subroutines 660 * 661 *******************************************************************************/ 662 663 /** 664 * dev_get_iflink - get 'iflink' value of a interface 665 * @dev: targeted interface 666 * 667 * Indicates the ifindex the interface is linked to. 668 * Physical interfaces have the same 'ifindex' and 'iflink' values. 669 */ 670 671 int dev_get_iflink(const struct net_device *dev) 672 { 673 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink) 674 return dev->netdev_ops->ndo_get_iflink(dev); 675 676 return READ_ONCE(dev->ifindex); 677 } 678 EXPORT_SYMBOL(dev_get_iflink); 679 680 /** 681 * dev_fill_metadata_dst - Retrieve tunnel egress information. 682 * @dev: targeted interface 683 * @skb: The packet. 684 * 685 * For better visibility of tunnel traffic OVS needs to retrieve 686 * egress tunnel information for a packet. Following API allows 687 * user to get this info. 688 */ 689 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb) 690 { 691 struct ip_tunnel_info *info; 692 693 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst) 694 return -EINVAL; 695 696 info = skb_tunnel_info_unclone(skb); 697 if (!info) 698 return -ENOMEM; 699 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX))) 700 return -EINVAL; 701 702 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb); 703 } 704 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst); 705 706 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack) 707 { 708 int k = stack->num_paths++; 709 710 if (WARN_ON_ONCE(k >= NET_DEVICE_PATH_STACK_MAX)) 711 return NULL; 712 713 return &stack->path[k]; 714 } 715 716 int dev_fill_forward_path(const struct net_device *dev, const u8 *daddr, 717 struct net_device_path_stack *stack) 718 { 719 const struct net_device *last_dev; 720 struct net_device_path_ctx ctx = { 721 .dev = dev, 722 }; 723 struct net_device_path *path; 724 int ret = 0; 725 726 memcpy(ctx.daddr, daddr, sizeof(ctx.daddr)); 727 stack->num_paths = 0; 728 while (ctx.dev && ctx.dev->netdev_ops->ndo_fill_forward_path) { 729 last_dev = ctx.dev; 730 path = dev_fwd_path(stack); 731 if (!path) 732 return -1; 733 734 memset(path, 0, sizeof(struct net_device_path)); 735 ret = ctx.dev->netdev_ops->ndo_fill_forward_path(&ctx, path); 736 if (ret < 0) 737 return -1; 738 739 if (WARN_ON_ONCE(last_dev == ctx.dev)) 740 return -1; 741 } 742 743 if (!ctx.dev) 744 return ret; 745 746 path = dev_fwd_path(stack); 747 if (!path) 748 return -1; 749 path->type = DEV_PATH_ETHERNET; 750 path->dev = ctx.dev; 751 752 return ret; 753 } 754 EXPORT_SYMBOL_GPL(dev_fill_forward_path); 755 756 /* must be called under rcu_read_lock(), as we dont take a reference */ 757 static struct napi_struct *napi_by_id(unsigned int napi_id) 758 { 759 unsigned int hash = napi_id % HASH_SIZE(napi_hash); 760 struct napi_struct *napi; 761 762 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node) 763 if (napi->napi_id == napi_id) 764 return napi; 765 766 return NULL; 767 } 768 769 /* must be called under rcu_read_lock(), as we dont take a reference */ 770 static struct napi_struct * 771 netdev_napi_by_id(struct net *net, unsigned int napi_id) 772 { 773 struct napi_struct *napi; 774 775 napi = napi_by_id(napi_id); 776 if (!napi) 777 return NULL; 778 779 if (WARN_ON_ONCE(!napi->dev)) 780 return NULL; 781 if (!net_eq(net, dev_net(napi->dev))) 782 return NULL; 783 784 return napi; 785 } 786 787 /** 788 * netdev_napi_by_id_lock() - find a device by NAPI ID and lock it 789 * @net: the applicable net namespace 790 * @napi_id: ID of a NAPI of a target device 791 * 792 * Find a NAPI instance with @napi_id. Lock its device. 793 * The device must be in %NETREG_REGISTERED state for lookup to succeed. 794 * netdev_unlock() must be called to release it. 795 * 796 * Return: pointer to NAPI, its device with lock held, NULL if not found. 797 */ 798 struct napi_struct * 799 netdev_napi_by_id_lock(struct net *net, unsigned int napi_id) 800 { 801 struct napi_struct *napi; 802 struct net_device *dev; 803 804 rcu_read_lock(); 805 napi = netdev_napi_by_id(net, napi_id); 806 if (!napi || READ_ONCE(napi->dev->reg_state) != NETREG_REGISTERED) { 807 rcu_read_unlock(); 808 return NULL; 809 } 810 811 dev = napi->dev; 812 dev_hold(dev); 813 rcu_read_unlock(); 814 815 dev = __netdev_put_lock(dev); 816 if (!dev) 817 return NULL; 818 819 rcu_read_lock(); 820 napi = netdev_napi_by_id(net, napi_id); 821 if (napi && napi->dev != dev) 822 napi = NULL; 823 rcu_read_unlock(); 824 825 if (!napi) 826 netdev_unlock(dev); 827 return napi; 828 } 829 830 /** 831 * __dev_get_by_name - find a device by its name 832 * @net: the applicable net namespace 833 * @name: name to find 834 * 835 * Find an interface by name. Must be called under RTNL semaphore. 836 * If the name is found a pointer to the device is returned. 837 * If the name is not found then %NULL is returned. The 838 * reference counters are not incremented so the caller must be 839 * careful with locks. 840 */ 841 842 struct net_device *__dev_get_by_name(struct net *net, const char *name) 843 { 844 struct netdev_name_node *node_name; 845 846 node_name = netdev_name_node_lookup(net, name); 847 return node_name ? node_name->dev : NULL; 848 } 849 EXPORT_SYMBOL(__dev_get_by_name); 850 851 /** 852 * dev_get_by_name_rcu - find a device by its name 853 * @net: the applicable net namespace 854 * @name: name to find 855 * 856 * Find an interface by name. 857 * If the name is found a pointer to the device is returned. 858 * If the name is not found then %NULL is returned. 859 * The reference counters are not incremented so the caller must be 860 * careful with locks. The caller must hold RCU lock. 861 */ 862 863 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name) 864 { 865 struct netdev_name_node *node_name; 866 867 node_name = netdev_name_node_lookup_rcu(net, name); 868 return node_name ? node_name->dev : NULL; 869 } 870 EXPORT_SYMBOL(dev_get_by_name_rcu); 871 872 /* Deprecated for new users, call netdev_get_by_name() instead */ 873 struct net_device *dev_get_by_name(struct net *net, const char *name) 874 { 875 struct net_device *dev; 876 877 rcu_read_lock(); 878 dev = dev_get_by_name_rcu(net, name); 879 dev_hold(dev); 880 rcu_read_unlock(); 881 return dev; 882 } 883 EXPORT_SYMBOL(dev_get_by_name); 884 885 /** 886 * netdev_get_by_name() - find a device by its name 887 * @net: the applicable net namespace 888 * @name: name to find 889 * @tracker: tracking object for the acquired reference 890 * @gfp: allocation flags for the tracker 891 * 892 * Find an interface by name. This can be called from any 893 * context and does its own locking. The returned handle has 894 * the usage count incremented and the caller must use netdev_put() to 895 * release it when it is no longer needed. %NULL is returned if no 896 * matching device is found. 897 */ 898 struct net_device *netdev_get_by_name(struct net *net, const char *name, 899 netdevice_tracker *tracker, gfp_t gfp) 900 { 901 struct net_device *dev; 902 903 dev = dev_get_by_name(net, name); 904 if (dev) 905 netdev_tracker_alloc(dev, tracker, gfp); 906 return dev; 907 } 908 EXPORT_SYMBOL(netdev_get_by_name); 909 910 /** 911 * __dev_get_by_index - find a device by its ifindex 912 * @net: the applicable net namespace 913 * @ifindex: index of device 914 * 915 * Search for an interface by index. Returns %NULL if the device 916 * is not found or a pointer to the device. The device has not 917 * had its reference counter increased so the caller must be careful 918 * about locking. The caller must hold the RTNL semaphore. 919 */ 920 921 struct net_device *__dev_get_by_index(struct net *net, int ifindex) 922 { 923 struct net_device *dev; 924 struct hlist_head *head = dev_index_hash(net, ifindex); 925 926 hlist_for_each_entry(dev, head, index_hlist) 927 if (dev->ifindex == ifindex) 928 return dev; 929 930 return NULL; 931 } 932 EXPORT_SYMBOL(__dev_get_by_index); 933 934 /** 935 * dev_get_by_index_rcu - find a device by its ifindex 936 * @net: the applicable net namespace 937 * @ifindex: index of device 938 * 939 * Search for an interface by index. Returns %NULL if the device 940 * is not found or a pointer to the device. The device has not 941 * had its reference counter increased so the caller must be careful 942 * about locking. The caller must hold RCU lock. 943 */ 944 945 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex) 946 { 947 struct net_device *dev; 948 struct hlist_head *head = dev_index_hash(net, ifindex); 949 950 hlist_for_each_entry_rcu(dev, head, index_hlist) 951 if (dev->ifindex == ifindex) 952 return dev; 953 954 return NULL; 955 } 956 EXPORT_SYMBOL(dev_get_by_index_rcu); 957 958 /* Deprecated for new users, call netdev_get_by_index() instead */ 959 struct net_device *dev_get_by_index(struct net *net, int ifindex) 960 { 961 struct net_device *dev; 962 963 rcu_read_lock(); 964 dev = dev_get_by_index_rcu(net, ifindex); 965 dev_hold(dev); 966 rcu_read_unlock(); 967 return dev; 968 } 969 EXPORT_SYMBOL(dev_get_by_index); 970 971 /** 972 * netdev_get_by_index() - find a device by its ifindex 973 * @net: the applicable net namespace 974 * @ifindex: index of device 975 * @tracker: tracking object for the acquired reference 976 * @gfp: allocation flags for the tracker 977 * 978 * Search for an interface by index. Returns NULL if the device 979 * is not found or a pointer to the device. The device returned has 980 * had a reference added and the pointer is safe until the user calls 981 * netdev_put() to indicate they have finished with it. 982 */ 983 struct net_device *netdev_get_by_index(struct net *net, int ifindex, 984 netdevice_tracker *tracker, gfp_t gfp) 985 { 986 struct net_device *dev; 987 988 dev = dev_get_by_index(net, ifindex); 989 if (dev) 990 netdev_tracker_alloc(dev, tracker, gfp); 991 return dev; 992 } 993 EXPORT_SYMBOL(netdev_get_by_index); 994 995 /** 996 * dev_get_by_napi_id - find a device by napi_id 997 * @napi_id: ID of the NAPI struct 998 * 999 * Search for an interface by NAPI ID. Returns %NULL if the device 1000 * is not found or a pointer to the device. The device has not had 1001 * its reference counter increased so the caller must be careful 1002 * about locking. The caller must hold RCU lock. 1003 */ 1004 struct net_device *dev_get_by_napi_id(unsigned int napi_id) 1005 { 1006 struct napi_struct *napi; 1007 1008 WARN_ON_ONCE(!rcu_read_lock_held()); 1009 1010 if (napi_id < MIN_NAPI_ID) 1011 return NULL; 1012 1013 napi = napi_by_id(napi_id); 1014 1015 return napi ? napi->dev : NULL; 1016 } 1017 1018 /* Release the held reference on the net_device, and if the net_device 1019 * is still registered try to lock the instance lock. If device is being 1020 * unregistered NULL will be returned (but the reference has been released, 1021 * either way!) 1022 * 1023 * This helper is intended for locking net_device after it has been looked up 1024 * using a lockless lookup helper. Lock prevents the instance from going away. 1025 */ 1026 struct net_device *__netdev_put_lock(struct net_device *dev) 1027 { 1028 netdev_lock(dev); 1029 if (dev->reg_state > NETREG_REGISTERED) { 1030 netdev_unlock(dev); 1031 dev_put(dev); 1032 return NULL; 1033 } 1034 dev_put(dev); 1035 return dev; 1036 } 1037 1038 /** 1039 * netdev_get_by_index_lock() - find a device by its ifindex 1040 * @net: the applicable net namespace 1041 * @ifindex: index of device 1042 * 1043 * Search for an interface by index. If a valid device 1044 * with @ifindex is found it will be returned with netdev->lock held. 1045 * netdev_unlock() must be called to release it. 1046 * 1047 * Return: pointer to a device with lock held, NULL if not found. 1048 */ 1049 struct net_device *netdev_get_by_index_lock(struct net *net, int ifindex) 1050 { 1051 struct net_device *dev; 1052 1053 dev = dev_get_by_index(net, ifindex); 1054 if (!dev) 1055 return NULL; 1056 1057 return __netdev_put_lock(dev); 1058 } 1059 1060 struct net_device * 1061 netdev_xa_find_lock(struct net *net, struct net_device *dev, 1062 unsigned long *index) 1063 { 1064 if (dev) 1065 netdev_unlock(dev); 1066 1067 do { 1068 rcu_read_lock(); 1069 dev = xa_find(&net->dev_by_index, index, ULONG_MAX, XA_PRESENT); 1070 if (!dev) { 1071 rcu_read_unlock(); 1072 return NULL; 1073 } 1074 dev_hold(dev); 1075 rcu_read_unlock(); 1076 1077 dev = __netdev_put_lock(dev); 1078 if (dev) 1079 return dev; 1080 1081 (*index)++; 1082 } while (true); 1083 } 1084 1085 static DEFINE_SEQLOCK(netdev_rename_lock); 1086 1087 void netdev_copy_name(struct net_device *dev, char *name) 1088 { 1089 unsigned int seq; 1090 1091 do { 1092 seq = read_seqbegin(&netdev_rename_lock); 1093 strscpy(name, dev->name, IFNAMSIZ); 1094 } while (read_seqretry(&netdev_rename_lock, seq)); 1095 } 1096 1097 /** 1098 * netdev_get_name - get a netdevice name, knowing its ifindex. 1099 * @net: network namespace 1100 * @name: a pointer to the buffer where the name will be stored. 1101 * @ifindex: the ifindex of the interface to get the name from. 1102 */ 1103 int netdev_get_name(struct net *net, char *name, int ifindex) 1104 { 1105 struct net_device *dev; 1106 int ret; 1107 1108 rcu_read_lock(); 1109 1110 dev = dev_get_by_index_rcu(net, ifindex); 1111 if (!dev) { 1112 ret = -ENODEV; 1113 goto out; 1114 } 1115 1116 netdev_copy_name(dev, name); 1117 1118 ret = 0; 1119 out: 1120 rcu_read_unlock(); 1121 return ret; 1122 } 1123 1124 static bool dev_addr_cmp(struct net_device *dev, unsigned short type, 1125 const char *ha) 1126 { 1127 return dev->type == type && !memcmp(dev->dev_addr, ha, dev->addr_len); 1128 } 1129 1130 /** 1131 * dev_getbyhwaddr_rcu - find a device by its hardware address 1132 * @net: the applicable net namespace 1133 * @type: media type of device 1134 * @ha: hardware address 1135 * 1136 * Search for an interface by MAC address. Returns NULL if the device 1137 * is not found or a pointer to the device. 1138 * The caller must hold RCU. 1139 * The returned device has not had its ref count increased 1140 * and the caller must therefore be careful about locking 1141 * 1142 */ 1143 1144 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type, 1145 const char *ha) 1146 { 1147 struct net_device *dev; 1148 1149 for_each_netdev_rcu(net, dev) 1150 if (dev_addr_cmp(dev, type, ha)) 1151 return dev; 1152 1153 return NULL; 1154 } 1155 EXPORT_SYMBOL(dev_getbyhwaddr_rcu); 1156 1157 /** 1158 * dev_getbyhwaddr() - find a device by its hardware address 1159 * @net: the applicable net namespace 1160 * @type: media type of device 1161 * @ha: hardware address 1162 * 1163 * Similar to dev_getbyhwaddr_rcu(), but the owner needs to hold 1164 * rtnl_lock. 1165 * 1166 * Context: rtnl_lock() must be held. 1167 * Return: pointer to the net_device, or NULL if not found 1168 */ 1169 struct net_device *dev_getbyhwaddr(struct net *net, unsigned short type, 1170 const char *ha) 1171 { 1172 struct net_device *dev; 1173 1174 ASSERT_RTNL(); 1175 for_each_netdev(net, dev) 1176 if (dev_addr_cmp(dev, type, ha)) 1177 return dev; 1178 1179 return NULL; 1180 } 1181 EXPORT_SYMBOL(dev_getbyhwaddr); 1182 1183 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type) 1184 { 1185 struct net_device *dev, *ret = NULL; 1186 1187 rcu_read_lock(); 1188 for_each_netdev_rcu(net, dev) 1189 if (dev->type == type) { 1190 dev_hold(dev); 1191 ret = dev; 1192 break; 1193 } 1194 rcu_read_unlock(); 1195 return ret; 1196 } 1197 EXPORT_SYMBOL(dev_getfirstbyhwtype); 1198 1199 /** 1200 * __dev_get_by_flags - find any device with given flags 1201 * @net: the applicable net namespace 1202 * @if_flags: IFF_* values 1203 * @mask: bitmask of bits in if_flags to check 1204 * 1205 * Search for any interface with the given flags. Returns NULL if a device 1206 * is not found or a pointer to the device. Must be called inside 1207 * rtnl_lock(), and result refcount is unchanged. 1208 */ 1209 1210 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags, 1211 unsigned short mask) 1212 { 1213 struct net_device *dev, *ret; 1214 1215 ASSERT_RTNL(); 1216 1217 ret = NULL; 1218 for_each_netdev(net, dev) { 1219 if (((dev->flags ^ if_flags) & mask) == 0) { 1220 ret = dev; 1221 break; 1222 } 1223 } 1224 return ret; 1225 } 1226 EXPORT_SYMBOL(__dev_get_by_flags); 1227 1228 /** 1229 * dev_valid_name - check if name is okay for network device 1230 * @name: name string 1231 * 1232 * Network device names need to be valid file names to 1233 * allow sysfs to work. We also disallow any kind of 1234 * whitespace. 1235 */ 1236 bool dev_valid_name(const char *name) 1237 { 1238 if (*name == '\0') 1239 return false; 1240 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ) 1241 return false; 1242 if (!strcmp(name, ".") || !strcmp(name, "..")) 1243 return false; 1244 1245 while (*name) { 1246 if (*name == '/' || *name == ':' || isspace(*name)) 1247 return false; 1248 name++; 1249 } 1250 return true; 1251 } 1252 EXPORT_SYMBOL(dev_valid_name); 1253 1254 /** 1255 * __dev_alloc_name - allocate a name for a device 1256 * @net: network namespace to allocate the device name in 1257 * @name: name format string 1258 * @res: result name string 1259 * 1260 * Passed a format string - eg "lt%d" it will try and find a suitable 1261 * id. It scans list of devices to build up a free map, then chooses 1262 * the first empty slot. The caller must hold the dev_base or rtnl lock 1263 * while allocating the name and adding the device in order to avoid 1264 * duplicates. 1265 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1266 * Returns the number of the unit assigned or a negative errno code. 1267 */ 1268 1269 static int __dev_alloc_name(struct net *net, const char *name, char *res) 1270 { 1271 int i = 0; 1272 const char *p; 1273 const int max_netdevices = 8*PAGE_SIZE; 1274 unsigned long *inuse; 1275 struct net_device *d; 1276 char buf[IFNAMSIZ]; 1277 1278 /* Verify the string as this thing may have come from the user. 1279 * There must be one "%d" and no other "%" characters. 1280 */ 1281 p = strchr(name, '%'); 1282 if (!p || p[1] != 'd' || strchr(p + 2, '%')) 1283 return -EINVAL; 1284 1285 /* Use one page as a bit array of possible slots */ 1286 inuse = bitmap_zalloc(max_netdevices, GFP_ATOMIC); 1287 if (!inuse) 1288 return -ENOMEM; 1289 1290 for_each_netdev(net, d) { 1291 struct netdev_name_node *name_node; 1292 1293 netdev_for_each_altname(d, name_node) { 1294 if (!sscanf(name_node->name, name, &i)) 1295 continue; 1296 if (i < 0 || i >= max_netdevices) 1297 continue; 1298 1299 /* avoid cases where sscanf is not exact inverse of printf */ 1300 snprintf(buf, IFNAMSIZ, name, i); 1301 if (!strncmp(buf, name_node->name, IFNAMSIZ)) 1302 __set_bit(i, inuse); 1303 } 1304 if (!sscanf(d->name, name, &i)) 1305 continue; 1306 if (i < 0 || i >= max_netdevices) 1307 continue; 1308 1309 /* avoid cases where sscanf is not exact inverse of printf */ 1310 snprintf(buf, IFNAMSIZ, name, i); 1311 if (!strncmp(buf, d->name, IFNAMSIZ)) 1312 __set_bit(i, inuse); 1313 } 1314 1315 i = find_first_zero_bit(inuse, max_netdevices); 1316 bitmap_free(inuse); 1317 if (i == max_netdevices) 1318 return -ENFILE; 1319 1320 /* 'res' and 'name' could overlap, use 'buf' as an intermediate buffer */ 1321 strscpy(buf, name, IFNAMSIZ); 1322 snprintf(res, IFNAMSIZ, buf, i); 1323 return i; 1324 } 1325 1326 /* Returns negative errno or allocated unit id (see __dev_alloc_name()) */ 1327 static int dev_prep_valid_name(struct net *net, struct net_device *dev, 1328 const char *want_name, char *out_name, 1329 int dup_errno) 1330 { 1331 if (!dev_valid_name(want_name)) 1332 return -EINVAL; 1333 1334 if (strchr(want_name, '%')) 1335 return __dev_alloc_name(net, want_name, out_name); 1336 1337 if (netdev_name_in_use(net, want_name)) 1338 return -dup_errno; 1339 if (out_name != want_name) 1340 strscpy(out_name, want_name, IFNAMSIZ); 1341 return 0; 1342 } 1343 1344 /** 1345 * dev_alloc_name - allocate a name for a device 1346 * @dev: device 1347 * @name: name format string 1348 * 1349 * Passed a format string - eg "lt%d" it will try and find a suitable 1350 * id. It scans list of devices to build up a free map, then chooses 1351 * the first empty slot. The caller must hold the dev_base or rtnl lock 1352 * while allocating the name and adding the device in order to avoid 1353 * duplicates. 1354 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1355 * Returns the number of the unit assigned or a negative errno code. 1356 */ 1357 1358 int dev_alloc_name(struct net_device *dev, const char *name) 1359 { 1360 return dev_prep_valid_name(dev_net(dev), dev, name, dev->name, ENFILE); 1361 } 1362 EXPORT_SYMBOL(dev_alloc_name); 1363 1364 static int dev_get_valid_name(struct net *net, struct net_device *dev, 1365 const char *name) 1366 { 1367 int ret; 1368 1369 ret = dev_prep_valid_name(net, dev, name, dev->name, EEXIST); 1370 return ret < 0 ? ret : 0; 1371 } 1372 1373 /** 1374 * dev_change_name - change name of a device 1375 * @dev: device 1376 * @newname: name (or format string) must be at least IFNAMSIZ 1377 * 1378 * Change name of a device, can pass format strings "eth%d". 1379 * for wildcarding. 1380 */ 1381 int dev_change_name(struct net_device *dev, const char *newname) 1382 { 1383 struct net *net = dev_net(dev); 1384 unsigned char old_assign_type; 1385 char oldname[IFNAMSIZ]; 1386 int err = 0; 1387 int ret; 1388 1389 ASSERT_RTNL_NET(net); 1390 1391 if (!strncmp(newname, dev->name, IFNAMSIZ)) 1392 return 0; 1393 1394 memcpy(oldname, dev->name, IFNAMSIZ); 1395 1396 write_seqlock_bh(&netdev_rename_lock); 1397 err = dev_get_valid_name(net, dev, newname); 1398 write_sequnlock_bh(&netdev_rename_lock); 1399 1400 if (err < 0) 1401 return err; 1402 1403 if (oldname[0] && !strchr(oldname, '%')) 1404 netdev_info(dev, "renamed from %s%s\n", oldname, 1405 dev->flags & IFF_UP ? " (while UP)" : ""); 1406 1407 old_assign_type = dev->name_assign_type; 1408 WRITE_ONCE(dev->name_assign_type, NET_NAME_RENAMED); 1409 1410 rollback: 1411 ret = device_rename(&dev->dev, dev->name); 1412 if (ret) { 1413 write_seqlock_bh(&netdev_rename_lock); 1414 memcpy(dev->name, oldname, IFNAMSIZ); 1415 write_sequnlock_bh(&netdev_rename_lock); 1416 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1417 return ret; 1418 } 1419 1420 netdev_adjacent_rename_links(dev, oldname); 1421 1422 netdev_name_node_del(dev->name_node); 1423 1424 synchronize_net(); 1425 1426 netdev_name_node_add(net, dev->name_node); 1427 1428 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev); 1429 ret = notifier_to_errno(ret); 1430 1431 if (ret) { 1432 /* err >= 0 after dev_alloc_name() or stores the first errno */ 1433 if (err >= 0) { 1434 err = ret; 1435 write_seqlock_bh(&netdev_rename_lock); 1436 memcpy(dev->name, oldname, IFNAMSIZ); 1437 write_sequnlock_bh(&netdev_rename_lock); 1438 memcpy(oldname, newname, IFNAMSIZ); 1439 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1440 old_assign_type = NET_NAME_RENAMED; 1441 goto rollback; 1442 } else { 1443 netdev_err(dev, "name change rollback failed: %d\n", 1444 ret); 1445 } 1446 } 1447 1448 return err; 1449 } 1450 1451 /** 1452 * dev_set_alias - change ifalias of a device 1453 * @dev: device 1454 * @alias: name up to IFALIASZ 1455 * @len: limit of bytes to copy from info 1456 * 1457 * Set ifalias for a device, 1458 */ 1459 int dev_set_alias(struct net_device *dev, const char *alias, size_t len) 1460 { 1461 struct dev_ifalias *new_alias = NULL; 1462 1463 if (len >= IFALIASZ) 1464 return -EINVAL; 1465 1466 if (len) { 1467 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL); 1468 if (!new_alias) 1469 return -ENOMEM; 1470 1471 memcpy(new_alias->ifalias, alias, len); 1472 new_alias->ifalias[len] = 0; 1473 } 1474 1475 mutex_lock(&ifalias_mutex); 1476 new_alias = rcu_replace_pointer(dev->ifalias, new_alias, 1477 mutex_is_locked(&ifalias_mutex)); 1478 mutex_unlock(&ifalias_mutex); 1479 1480 if (new_alias) 1481 kfree_rcu(new_alias, rcuhead); 1482 1483 return len; 1484 } 1485 EXPORT_SYMBOL(dev_set_alias); 1486 1487 /** 1488 * dev_get_alias - get ifalias of a device 1489 * @dev: device 1490 * @name: buffer to store name of ifalias 1491 * @len: size of buffer 1492 * 1493 * get ifalias for a device. Caller must make sure dev cannot go 1494 * away, e.g. rcu read lock or own a reference count to device. 1495 */ 1496 int dev_get_alias(const struct net_device *dev, char *name, size_t len) 1497 { 1498 const struct dev_ifalias *alias; 1499 int ret = 0; 1500 1501 rcu_read_lock(); 1502 alias = rcu_dereference(dev->ifalias); 1503 if (alias) 1504 ret = snprintf(name, len, "%s", alias->ifalias); 1505 rcu_read_unlock(); 1506 1507 return ret; 1508 } 1509 1510 /** 1511 * netdev_features_change - device changes features 1512 * @dev: device to cause notification 1513 * 1514 * Called to indicate a device has changed features. 1515 */ 1516 void netdev_features_change(struct net_device *dev) 1517 { 1518 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev); 1519 } 1520 EXPORT_SYMBOL(netdev_features_change); 1521 1522 /** 1523 * netdev_state_change - device changes state 1524 * @dev: device to cause notification 1525 * 1526 * Called to indicate a device has changed state. This function calls 1527 * the notifier chains for netdev_chain and sends a NEWLINK message 1528 * to the routing socket. 1529 */ 1530 void netdev_state_change(struct net_device *dev) 1531 { 1532 if (dev->flags & IFF_UP) { 1533 struct netdev_notifier_change_info change_info = { 1534 .info.dev = dev, 1535 }; 1536 1537 call_netdevice_notifiers_info(NETDEV_CHANGE, 1538 &change_info.info); 1539 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL, 0, NULL); 1540 } 1541 } 1542 EXPORT_SYMBOL(netdev_state_change); 1543 1544 /** 1545 * __netdev_notify_peers - notify network peers about existence of @dev, 1546 * to be called when rtnl lock is already held. 1547 * @dev: network device 1548 * 1549 * Generate traffic such that interested network peers are aware of 1550 * @dev, such as by generating a gratuitous ARP. This may be used when 1551 * a device wants to inform the rest of the network about some sort of 1552 * reconfiguration such as a failover event or virtual machine 1553 * migration. 1554 */ 1555 void __netdev_notify_peers(struct net_device *dev) 1556 { 1557 ASSERT_RTNL(); 1558 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev); 1559 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev); 1560 } 1561 EXPORT_SYMBOL(__netdev_notify_peers); 1562 1563 /** 1564 * netdev_notify_peers - notify network peers about existence of @dev 1565 * @dev: network device 1566 * 1567 * Generate traffic such that interested network peers are aware of 1568 * @dev, such as by generating a gratuitous ARP. This may be used when 1569 * a device wants to inform the rest of the network about some sort of 1570 * reconfiguration such as a failover event or virtual machine 1571 * migration. 1572 */ 1573 void netdev_notify_peers(struct net_device *dev) 1574 { 1575 rtnl_lock(); 1576 __netdev_notify_peers(dev); 1577 rtnl_unlock(); 1578 } 1579 EXPORT_SYMBOL(netdev_notify_peers); 1580 1581 static int napi_threaded_poll(void *data); 1582 1583 static int napi_kthread_create(struct napi_struct *n) 1584 { 1585 int err = 0; 1586 1587 /* Create and wake up the kthread once to put it in 1588 * TASK_INTERRUPTIBLE mode to avoid the blocked task 1589 * warning and work with loadavg. 1590 */ 1591 n->thread = kthread_run(napi_threaded_poll, n, "napi/%s-%d", 1592 n->dev->name, n->napi_id); 1593 if (IS_ERR(n->thread)) { 1594 err = PTR_ERR(n->thread); 1595 pr_err("kthread_run failed with err %d\n", err); 1596 n->thread = NULL; 1597 } 1598 1599 return err; 1600 } 1601 1602 static int __dev_open(struct net_device *dev, struct netlink_ext_ack *extack) 1603 { 1604 const struct net_device_ops *ops = dev->netdev_ops; 1605 int ret; 1606 1607 ASSERT_RTNL(); 1608 dev_addr_check(dev); 1609 1610 if (!netif_device_present(dev)) { 1611 /* may be detached because parent is runtime-suspended */ 1612 if (dev->dev.parent) 1613 pm_runtime_resume(dev->dev.parent); 1614 if (!netif_device_present(dev)) 1615 return -ENODEV; 1616 } 1617 1618 /* Block netpoll from trying to do any rx path servicing. 1619 * If we don't do this there is a chance ndo_poll_controller 1620 * or ndo_poll may be running while we open the device 1621 */ 1622 netpoll_poll_disable(dev); 1623 1624 ret = call_netdevice_notifiers_extack(NETDEV_PRE_UP, dev, extack); 1625 ret = notifier_to_errno(ret); 1626 if (ret) 1627 return ret; 1628 1629 set_bit(__LINK_STATE_START, &dev->state); 1630 1631 if (ops->ndo_validate_addr) 1632 ret = ops->ndo_validate_addr(dev); 1633 1634 if (!ret && ops->ndo_open) 1635 ret = ops->ndo_open(dev); 1636 1637 netpoll_poll_enable(dev); 1638 1639 if (ret) 1640 clear_bit(__LINK_STATE_START, &dev->state); 1641 else { 1642 netif_set_up(dev, true); 1643 dev_set_rx_mode(dev); 1644 dev_activate(dev); 1645 add_device_randomness(dev->dev_addr, dev->addr_len); 1646 } 1647 1648 return ret; 1649 } 1650 1651 /** 1652 * dev_open - prepare an interface for use. 1653 * @dev: device to open 1654 * @extack: netlink extended ack 1655 * 1656 * Takes a device from down to up state. The device's private open 1657 * function is invoked and then the multicast lists are loaded. Finally 1658 * the device is moved into the up state and a %NETDEV_UP message is 1659 * sent to the netdev notifier chain. 1660 * 1661 * Calling this function on an active interface is a nop. On a failure 1662 * a negative errno code is returned. 1663 */ 1664 int dev_open(struct net_device *dev, struct netlink_ext_ack *extack) 1665 { 1666 int ret; 1667 1668 if (dev->flags & IFF_UP) 1669 return 0; 1670 1671 ret = __dev_open(dev, extack); 1672 if (ret < 0) 1673 return ret; 1674 1675 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1676 call_netdevice_notifiers(NETDEV_UP, dev); 1677 1678 return ret; 1679 } 1680 EXPORT_SYMBOL(dev_open); 1681 1682 static void __dev_close_many(struct list_head *head) 1683 { 1684 struct net_device *dev; 1685 1686 ASSERT_RTNL(); 1687 might_sleep(); 1688 1689 list_for_each_entry(dev, head, close_list) { 1690 /* Temporarily disable netpoll until the interface is down */ 1691 netpoll_poll_disable(dev); 1692 1693 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev); 1694 1695 clear_bit(__LINK_STATE_START, &dev->state); 1696 1697 /* Synchronize to scheduled poll. We cannot touch poll list, it 1698 * can be even on different cpu. So just clear netif_running(). 1699 * 1700 * dev->stop() will invoke napi_disable() on all of it's 1701 * napi_struct instances on this device. 1702 */ 1703 smp_mb__after_atomic(); /* Commit netif_running(). */ 1704 } 1705 1706 dev_deactivate_many(head); 1707 1708 list_for_each_entry(dev, head, close_list) { 1709 const struct net_device_ops *ops = dev->netdev_ops; 1710 1711 /* 1712 * Call the device specific close. This cannot fail. 1713 * Only if device is UP 1714 * 1715 * We allow it to be called even after a DETACH hot-plug 1716 * event. 1717 */ 1718 if (ops->ndo_stop) 1719 ops->ndo_stop(dev); 1720 1721 netif_set_up(dev, false); 1722 netpoll_poll_enable(dev); 1723 } 1724 } 1725 1726 static void __dev_close(struct net_device *dev) 1727 { 1728 LIST_HEAD(single); 1729 1730 list_add(&dev->close_list, &single); 1731 __dev_close_many(&single); 1732 list_del(&single); 1733 } 1734 1735 void dev_close_many(struct list_head *head, bool unlink) 1736 { 1737 struct net_device *dev, *tmp; 1738 1739 /* Remove the devices that don't need to be closed */ 1740 list_for_each_entry_safe(dev, tmp, head, close_list) 1741 if (!(dev->flags & IFF_UP)) 1742 list_del_init(&dev->close_list); 1743 1744 __dev_close_many(head); 1745 1746 list_for_each_entry_safe(dev, tmp, head, close_list) { 1747 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1748 call_netdevice_notifiers(NETDEV_DOWN, dev); 1749 if (unlink) 1750 list_del_init(&dev->close_list); 1751 } 1752 } 1753 EXPORT_SYMBOL(dev_close_many); 1754 1755 /** 1756 * dev_close - shutdown an interface. 1757 * @dev: device to shutdown 1758 * 1759 * This function moves an active device into down state. A 1760 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device 1761 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier 1762 * chain. 1763 */ 1764 void dev_close(struct net_device *dev) 1765 { 1766 if (dev->flags & IFF_UP) { 1767 LIST_HEAD(single); 1768 1769 list_add(&dev->close_list, &single); 1770 dev_close_many(&single, true); 1771 list_del(&single); 1772 } 1773 } 1774 EXPORT_SYMBOL(dev_close); 1775 1776 1777 /** 1778 * dev_disable_lro - disable Large Receive Offload on a device 1779 * @dev: device 1780 * 1781 * Disable Large Receive Offload (LRO) on a net device. Must be 1782 * called under RTNL. This is needed if received packets may be 1783 * forwarded to another interface. 1784 */ 1785 void dev_disable_lro(struct net_device *dev) 1786 { 1787 struct net_device *lower_dev; 1788 struct list_head *iter; 1789 1790 dev->wanted_features &= ~NETIF_F_LRO; 1791 netdev_update_features(dev); 1792 1793 if (unlikely(dev->features & NETIF_F_LRO)) 1794 netdev_WARN(dev, "failed to disable LRO!\n"); 1795 1796 netdev_for_each_lower_dev(dev, lower_dev, iter) 1797 dev_disable_lro(lower_dev); 1798 } 1799 EXPORT_SYMBOL(dev_disable_lro); 1800 1801 /** 1802 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device 1803 * @dev: device 1804 * 1805 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be 1806 * called under RTNL. This is needed if Generic XDP is installed on 1807 * the device. 1808 */ 1809 static void dev_disable_gro_hw(struct net_device *dev) 1810 { 1811 dev->wanted_features &= ~NETIF_F_GRO_HW; 1812 netdev_update_features(dev); 1813 1814 if (unlikely(dev->features & NETIF_F_GRO_HW)) 1815 netdev_WARN(dev, "failed to disable GRO_HW!\n"); 1816 } 1817 1818 const char *netdev_cmd_to_name(enum netdev_cmd cmd) 1819 { 1820 #define N(val) \ 1821 case NETDEV_##val: \ 1822 return "NETDEV_" __stringify(val); 1823 switch (cmd) { 1824 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER) 1825 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE) 1826 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE) 1827 N(POST_INIT) N(PRE_UNINIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) 1828 N(CHANGEUPPER) N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) 1829 N(BONDING_INFO) N(PRECHANGEUPPER) N(CHANGELOWERSTATE) 1830 N(UDP_TUNNEL_PUSH_INFO) N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN) 1831 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO) 1832 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO) 1833 N(PRE_CHANGEADDR) N(OFFLOAD_XSTATS_ENABLE) N(OFFLOAD_XSTATS_DISABLE) 1834 N(OFFLOAD_XSTATS_REPORT_USED) N(OFFLOAD_XSTATS_REPORT_DELTA) 1835 N(XDP_FEAT_CHANGE) 1836 } 1837 #undef N 1838 return "UNKNOWN_NETDEV_EVENT"; 1839 } 1840 EXPORT_SYMBOL_GPL(netdev_cmd_to_name); 1841 1842 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val, 1843 struct net_device *dev) 1844 { 1845 struct netdev_notifier_info info = { 1846 .dev = dev, 1847 }; 1848 1849 return nb->notifier_call(nb, val, &info); 1850 } 1851 1852 static int call_netdevice_register_notifiers(struct notifier_block *nb, 1853 struct net_device *dev) 1854 { 1855 int err; 1856 1857 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev); 1858 err = notifier_to_errno(err); 1859 if (err) 1860 return err; 1861 1862 if (!(dev->flags & IFF_UP)) 1863 return 0; 1864 1865 call_netdevice_notifier(nb, NETDEV_UP, dev); 1866 return 0; 1867 } 1868 1869 static void call_netdevice_unregister_notifiers(struct notifier_block *nb, 1870 struct net_device *dev) 1871 { 1872 if (dev->flags & IFF_UP) { 1873 call_netdevice_notifier(nb, NETDEV_GOING_DOWN, 1874 dev); 1875 call_netdevice_notifier(nb, NETDEV_DOWN, dev); 1876 } 1877 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev); 1878 } 1879 1880 static int call_netdevice_register_net_notifiers(struct notifier_block *nb, 1881 struct net *net) 1882 { 1883 struct net_device *dev; 1884 int err; 1885 1886 for_each_netdev(net, dev) { 1887 err = call_netdevice_register_notifiers(nb, dev); 1888 if (err) 1889 goto rollback; 1890 } 1891 return 0; 1892 1893 rollback: 1894 for_each_netdev_continue_reverse(net, dev) 1895 call_netdevice_unregister_notifiers(nb, dev); 1896 return err; 1897 } 1898 1899 static void call_netdevice_unregister_net_notifiers(struct notifier_block *nb, 1900 struct net *net) 1901 { 1902 struct net_device *dev; 1903 1904 for_each_netdev(net, dev) 1905 call_netdevice_unregister_notifiers(nb, dev); 1906 } 1907 1908 static int dev_boot_phase = 1; 1909 1910 /** 1911 * register_netdevice_notifier - register a network notifier block 1912 * @nb: notifier 1913 * 1914 * Register a notifier to be called when network device events occur. 1915 * The notifier passed is linked into the kernel structures and must 1916 * not be reused until it has been unregistered. A negative errno code 1917 * is returned on a failure. 1918 * 1919 * When registered all registration and up events are replayed 1920 * to the new notifier to allow device to have a race free 1921 * view of the network device list. 1922 */ 1923 1924 int register_netdevice_notifier(struct notifier_block *nb) 1925 { 1926 struct net *net; 1927 int err; 1928 1929 /* Close race with setup_net() and cleanup_net() */ 1930 down_write(&pernet_ops_rwsem); 1931 1932 /* When RTNL is removed, we need protection for netdev_chain. */ 1933 rtnl_lock(); 1934 1935 err = raw_notifier_chain_register(&netdev_chain, nb); 1936 if (err) 1937 goto unlock; 1938 if (dev_boot_phase) 1939 goto unlock; 1940 for_each_net(net) { 1941 __rtnl_net_lock(net); 1942 err = call_netdevice_register_net_notifiers(nb, net); 1943 __rtnl_net_unlock(net); 1944 if (err) 1945 goto rollback; 1946 } 1947 1948 unlock: 1949 rtnl_unlock(); 1950 up_write(&pernet_ops_rwsem); 1951 return err; 1952 1953 rollback: 1954 for_each_net_continue_reverse(net) { 1955 __rtnl_net_lock(net); 1956 call_netdevice_unregister_net_notifiers(nb, net); 1957 __rtnl_net_unlock(net); 1958 } 1959 1960 raw_notifier_chain_unregister(&netdev_chain, nb); 1961 goto unlock; 1962 } 1963 EXPORT_SYMBOL(register_netdevice_notifier); 1964 1965 /** 1966 * unregister_netdevice_notifier - unregister a network notifier block 1967 * @nb: notifier 1968 * 1969 * Unregister a notifier previously registered by 1970 * register_netdevice_notifier(). The notifier is unlinked into the 1971 * kernel structures and may then be reused. A negative errno code 1972 * is returned on a failure. 1973 * 1974 * After unregistering unregister and down device events are synthesized 1975 * for all devices on the device list to the removed notifier to remove 1976 * the need for special case cleanup code. 1977 */ 1978 1979 int unregister_netdevice_notifier(struct notifier_block *nb) 1980 { 1981 struct net *net; 1982 int err; 1983 1984 /* Close race with setup_net() and cleanup_net() */ 1985 down_write(&pernet_ops_rwsem); 1986 rtnl_lock(); 1987 err = raw_notifier_chain_unregister(&netdev_chain, nb); 1988 if (err) 1989 goto unlock; 1990 1991 for_each_net(net) { 1992 __rtnl_net_lock(net); 1993 call_netdevice_unregister_net_notifiers(nb, net); 1994 __rtnl_net_unlock(net); 1995 } 1996 1997 unlock: 1998 rtnl_unlock(); 1999 up_write(&pernet_ops_rwsem); 2000 return err; 2001 } 2002 EXPORT_SYMBOL(unregister_netdevice_notifier); 2003 2004 static int __register_netdevice_notifier_net(struct net *net, 2005 struct notifier_block *nb, 2006 bool ignore_call_fail) 2007 { 2008 int err; 2009 2010 err = raw_notifier_chain_register(&net->netdev_chain, nb); 2011 if (err) 2012 return err; 2013 if (dev_boot_phase) 2014 return 0; 2015 2016 err = call_netdevice_register_net_notifiers(nb, net); 2017 if (err && !ignore_call_fail) 2018 goto chain_unregister; 2019 2020 return 0; 2021 2022 chain_unregister: 2023 raw_notifier_chain_unregister(&net->netdev_chain, nb); 2024 return err; 2025 } 2026 2027 static int __unregister_netdevice_notifier_net(struct net *net, 2028 struct notifier_block *nb) 2029 { 2030 int err; 2031 2032 err = raw_notifier_chain_unregister(&net->netdev_chain, nb); 2033 if (err) 2034 return err; 2035 2036 call_netdevice_unregister_net_notifiers(nb, net); 2037 return 0; 2038 } 2039 2040 /** 2041 * register_netdevice_notifier_net - register a per-netns network notifier block 2042 * @net: network namespace 2043 * @nb: notifier 2044 * 2045 * Register a notifier to be called when network device events occur. 2046 * The notifier passed is linked into the kernel structures and must 2047 * not be reused until it has been unregistered. A negative errno code 2048 * is returned on a failure. 2049 * 2050 * When registered all registration and up events are replayed 2051 * to the new notifier to allow device to have a race free 2052 * view of the network device list. 2053 */ 2054 2055 int register_netdevice_notifier_net(struct net *net, struct notifier_block *nb) 2056 { 2057 int err; 2058 2059 rtnl_net_lock(net); 2060 err = __register_netdevice_notifier_net(net, nb, false); 2061 rtnl_net_unlock(net); 2062 2063 return err; 2064 } 2065 EXPORT_SYMBOL(register_netdevice_notifier_net); 2066 2067 /** 2068 * unregister_netdevice_notifier_net - unregister a per-netns 2069 * network notifier block 2070 * @net: network namespace 2071 * @nb: notifier 2072 * 2073 * Unregister a notifier previously registered by 2074 * register_netdevice_notifier_net(). The notifier is unlinked from the 2075 * kernel structures and may then be reused. A negative errno code 2076 * is returned on a failure. 2077 * 2078 * After unregistering unregister and down device events are synthesized 2079 * for all devices on the device list to the removed notifier to remove 2080 * the need for special case cleanup code. 2081 */ 2082 2083 int unregister_netdevice_notifier_net(struct net *net, 2084 struct notifier_block *nb) 2085 { 2086 int err; 2087 2088 rtnl_net_lock(net); 2089 err = __unregister_netdevice_notifier_net(net, nb); 2090 rtnl_net_unlock(net); 2091 2092 return err; 2093 } 2094 EXPORT_SYMBOL(unregister_netdevice_notifier_net); 2095 2096 static void __move_netdevice_notifier_net(struct net *src_net, 2097 struct net *dst_net, 2098 struct notifier_block *nb) 2099 { 2100 __unregister_netdevice_notifier_net(src_net, nb); 2101 __register_netdevice_notifier_net(dst_net, nb, true); 2102 } 2103 2104 static void rtnl_net_dev_lock(struct net_device *dev) 2105 { 2106 bool again; 2107 2108 do { 2109 struct net *net; 2110 2111 again = false; 2112 2113 /* netns might be being dismantled. */ 2114 rcu_read_lock(); 2115 net = dev_net_rcu(dev); 2116 net_passive_inc(net); 2117 rcu_read_unlock(); 2118 2119 rtnl_net_lock(net); 2120 2121 #ifdef CONFIG_NET_NS 2122 /* dev might have been moved to another netns. */ 2123 if (!net_eq(net, rcu_access_pointer(dev->nd_net.net))) { 2124 rtnl_net_unlock(net); 2125 net_passive_dec(net); 2126 again = true; 2127 } 2128 #endif 2129 } while (again); 2130 } 2131 2132 static void rtnl_net_dev_unlock(struct net_device *dev) 2133 { 2134 struct net *net = dev_net(dev); 2135 2136 rtnl_net_unlock(net); 2137 net_passive_dec(net); 2138 } 2139 2140 int register_netdevice_notifier_dev_net(struct net_device *dev, 2141 struct notifier_block *nb, 2142 struct netdev_net_notifier *nn) 2143 { 2144 struct net *net = dev_net(dev); 2145 int err; 2146 2147 /* rtnl_net_lock() assumes dev is not yet published by 2148 * register_netdevice(). 2149 */ 2150 DEBUG_NET_WARN_ON_ONCE(!list_empty(&dev->dev_list)); 2151 2152 rtnl_net_lock(net); 2153 err = __register_netdevice_notifier_net(net, nb, false); 2154 if (!err) { 2155 nn->nb = nb; 2156 list_add(&nn->list, &dev->net_notifier_list); 2157 } 2158 rtnl_net_unlock(net); 2159 2160 return err; 2161 } 2162 EXPORT_SYMBOL(register_netdevice_notifier_dev_net); 2163 2164 int unregister_netdevice_notifier_dev_net(struct net_device *dev, 2165 struct notifier_block *nb, 2166 struct netdev_net_notifier *nn) 2167 { 2168 int err; 2169 2170 rtnl_net_dev_lock(dev); 2171 list_del(&nn->list); 2172 err = __unregister_netdevice_notifier_net(dev_net(dev), nb); 2173 rtnl_net_dev_unlock(dev); 2174 2175 return err; 2176 } 2177 EXPORT_SYMBOL(unregister_netdevice_notifier_dev_net); 2178 2179 static void move_netdevice_notifiers_dev_net(struct net_device *dev, 2180 struct net *net) 2181 { 2182 struct netdev_net_notifier *nn; 2183 2184 list_for_each_entry(nn, &dev->net_notifier_list, list) 2185 __move_netdevice_notifier_net(dev_net(dev), net, nn->nb); 2186 } 2187 2188 /** 2189 * call_netdevice_notifiers_info - call all network notifier blocks 2190 * @val: value passed unmodified to notifier function 2191 * @info: notifier information data 2192 * 2193 * Call all network notifier blocks. Parameters and return value 2194 * are as for raw_notifier_call_chain(). 2195 */ 2196 2197 int call_netdevice_notifiers_info(unsigned long val, 2198 struct netdev_notifier_info *info) 2199 { 2200 struct net *net = dev_net(info->dev); 2201 int ret; 2202 2203 ASSERT_RTNL(); 2204 2205 /* Run per-netns notifier block chain first, then run the global one. 2206 * Hopefully, one day, the global one is going to be removed after 2207 * all notifier block registrators get converted to be per-netns. 2208 */ 2209 ret = raw_notifier_call_chain(&net->netdev_chain, val, info); 2210 if (ret & NOTIFY_STOP_MASK) 2211 return ret; 2212 return raw_notifier_call_chain(&netdev_chain, val, info); 2213 } 2214 2215 /** 2216 * call_netdevice_notifiers_info_robust - call per-netns notifier blocks 2217 * for and rollback on error 2218 * @val_up: value passed unmodified to notifier function 2219 * @val_down: value passed unmodified to the notifier function when 2220 * recovering from an error on @val_up 2221 * @info: notifier information data 2222 * 2223 * Call all per-netns network notifier blocks, but not notifier blocks on 2224 * the global notifier chain. Parameters and return value are as for 2225 * raw_notifier_call_chain_robust(). 2226 */ 2227 2228 static int 2229 call_netdevice_notifiers_info_robust(unsigned long val_up, 2230 unsigned long val_down, 2231 struct netdev_notifier_info *info) 2232 { 2233 struct net *net = dev_net(info->dev); 2234 2235 ASSERT_RTNL(); 2236 2237 return raw_notifier_call_chain_robust(&net->netdev_chain, 2238 val_up, val_down, info); 2239 } 2240 2241 static int call_netdevice_notifiers_extack(unsigned long val, 2242 struct net_device *dev, 2243 struct netlink_ext_ack *extack) 2244 { 2245 struct netdev_notifier_info info = { 2246 .dev = dev, 2247 .extack = extack, 2248 }; 2249 2250 return call_netdevice_notifiers_info(val, &info); 2251 } 2252 2253 /** 2254 * call_netdevice_notifiers - call all network notifier blocks 2255 * @val: value passed unmodified to notifier function 2256 * @dev: net_device pointer passed unmodified to notifier function 2257 * 2258 * Call all network notifier blocks. Parameters and return value 2259 * are as for raw_notifier_call_chain(). 2260 */ 2261 2262 int call_netdevice_notifiers(unsigned long val, struct net_device *dev) 2263 { 2264 return call_netdevice_notifiers_extack(val, dev, NULL); 2265 } 2266 EXPORT_SYMBOL(call_netdevice_notifiers); 2267 2268 /** 2269 * call_netdevice_notifiers_mtu - call all network notifier blocks 2270 * @val: value passed unmodified to notifier function 2271 * @dev: net_device pointer passed unmodified to notifier function 2272 * @arg: additional u32 argument passed to the notifier function 2273 * 2274 * Call all network notifier blocks. Parameters and return value 2275 * are as for raw_notifier_call_chain(). 2276 */ 2277 static int call_netdevice_notifiers_mtu(unsigned long val, 2278 struct net_device *dev, u32 arg) 2279 { 2280 struct netdev_notifier_info_ext info = { 2281 .info.dev = dev, 2282 .ext.mtu = arg, 2283 }; 2284 2285 BUILD_BUG_ON(offsetof(struct netdev_notifier_info_ext, info) != 0); 2286 2287 return call_netdevice_notifiers_info(val, &info.info); 2288 } 2289 2290 #ifdef CONFIG_NET_INGRESS 2291 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key); 2292 2293 void net_inc_ingress_queue(void) 2294 { 2295 static_branch_inc(&ingress_needed_key); 2296 } 2297 EXPORT_SYMBOL_GPL(net_inc_ingress_queue); 2298 2299 void net_dec_ingress_queue(void) 2300 { 2301 static_branch_dec(&ingress_needed_key); 2302 } 2303 EXPORT_SYMBOL_GPL(net_dec_ingress_queue); 2304 #endif 2305 2306 #ifdef CONFIG_NET_EGRESS 2307 static DEFINE_STATIC_KEY_FALSE(egress_needed_key); 2308 2309 void net_inc_egress_queue(void) 2310 { 2311 static_branch_inc(&egress_needed_key); 2312 } 2313 EXPORT_SYMBOL_GPL(net_inc_egress_queue); 2314 2315 void net_dec_egress_queue(void) 2316 { 2317 static_branch_dec(&egress_needed_key); 2318 } 2319 EXPORT_SYMBOL_GPL(net_dec_egress_queue); 2320 #endif 2321 2322 #ifdef CONFIG_NET_CLS_ACT 2323 DEFINE_STATIC_KEY_FALSE(tcf_sw_enabled_key); 2324 EXPORT_SYMBOL(tcf_sw_enabled_key); 2325 #endif 2326 2327 DEFINE_STATIC_KEY_FALSE(netstamp_needed_key); 2328 EXPORT_SYMBOL(netstamp_needed_key); 2329 #ifdef CONFIG_JUMP_LABEL 2330 static atomic_t netstamp_needed_deferred; 2331 static atomic_t netstamp_wanted; 2332 static void netstamp_clear(struct work_struct *work) 2333 { 2334 int deferred = atomic_xchg(&netstamp_needed_deferred, 0); 2335 int wanted; 2336 2337 wanted = atomic_add_return(deferred, &netstamp_wanted); 2338 if (wanted > 0) 2339 static_branch_enable(&netstamp_needed_key); 2340 else 2341 static_branch_disable(&netstamp_needed_key); 2342 } 2343 static DECLARE_WORK(netstamp_work, netstamp_clear); 2344 #endif 2345 2346 void net_enable_timestamp(void) 2347 { 2348 #ifdef CONFIG_JUMP_LABEL 2349 int wanted = atomic_read(&netstamp_wanted); 2350 2351 while (wanted > 0) { 2352 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted + 1)) 2353 return; 2354 } 2355 atomic_inc(&netstamp_needed_deferred); 2356 schedule_work(&netstamp_work); 2357 #else 2358 static_branch_inc(&netstamp_needed_key); 2359 #endif 2360 } 2361 EXPORT_SYMBOL(net_enable_timestamp); 2362 2363 void net_disable_timestamp(void) 2364 { 2365 #ifdef CONFIG_JUMP_LABEL 2366 int wanted = atomic_read(&netstamp_wanted); 2367 2368 while (wanted > 1) { 2369 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted - 1)) 2370 return; 2371 } 2372 atomic_dec(&netstamp_needed_deferred); 2373 schedule_work(&netstamp_work); 2374 #else 2375 static_branch_dec(&netstamp_needed_key); 2376 #endif 2377 } 2378 EXPORT_SYMBOL(net_disable_timestamp); 2379 2380 static inline void net_timestamp_set(struct sk_buff *skb) 2381 { 2382 skb->tstamp = 0; 2383 skb->tstamp_type = SKB_CLOCK_REALTIME; 2384 if (static_branch_unlikely(&netstamp_needed_key)) 2385 skb->tstamp = ktime_get_real(); 2386 } 2387 2388 #define net_timestamp_check(COND, SKB) \ 2389 if (static_branch_unlikely(&netstamp_needed_key)) { \ 2390 if ((COND) && !(SKB)->tstamp) \ 2391 (SKB)->tstamp = ktime_get_real(); \ 2392 } \ 2393 2394 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb) 2395 { 2396 return __is_skb_forwardable(dev, skb, true); 2397 } 2398 EXPORT_SYMBOL_GPL(is_skb_forwardable); 2399 2400 static int __dev_forward_skb2(struct net_device *dev, struct sk_buff *skb, 2401 bool check_mtu) 2402 { 2403 int ret = ____dev_forward_skb(dev, skb, check_mtu); 2404 2405 if (likely(!ret)) { 2406 skb->protocol = eth_type_trans(skb, dev); 2407 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN); 2408 } 2409 2410 return ret; 2411 } 2412 2413 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2414 { 2415 return __dev_forward_skb2(dev, skb, true); 2416 } 2417 EXPORT_SYMBOL_GPL(__dev_forward_skb); 2418 2419 /** 2420 * dev_forward_skb - loopback an skb to another netif 2421 * 2422 * @dev: destination network device 2423 * @skb: buffer to forward 2424 * 2425 * return values: 2426 * NET_RX_SUCCESS (no congestion) 2427 * NET_RX_DROP (packet was dropped, but freed) 2428 * 2429 * dev_forward_skb can be used for injecting an skb from the 2430 * start_xmit function of one device into the receive queue 2431 * of another device. 2432 * 2433 * The receiving device may be in another namespace, so 2434 * we have to clear all information in the skb that could 2435 * impact namespace isolation. 2436 */ 2437 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2438 { 2439 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb); 2440 } 2441 EXPORT_SYMBOL_GPL(dev_forward_skb); 2442 2443 int dev_forward_skb_nomtu(struct net_device *dev, struct sk_buff *skb) 2444 { 2445 return __dev_forward_skb2(dev, skb, false) ?: netif_rx_internal(skb); 2446 } 2447 2448 static inline int deliver_skb(struct sk_buff *skb, 2449 struct packet_type *pt_prev, 2450 struct net_device *orig_dev) 2451 { 2452 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 2453 return -ENOMEM; 2454 refcount_inc(&skb->users); 2455 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 2456 } 2457 2458 static inline void deliver_ptype_list_skb(struct sk_buff *skb, 2459 struct packet_type **pt, 2460 struct net_device *orig_dev, 2461 __be16 type, 2462 struct list_head *ptype_list) 2463 { 2464 struct packet_type *ptype, *pt_prev = *pt; 2465 2466 list_for_each_entry_rcu(ptype, ptype_list, list) { 2467 if (ptype->type != type) 2468 continue; 2469 if (pt_prev) 2470 deliver_skb(skb, pt_prev, orig_dev); 2471 pt_prev = ptype; 2472 } 2473 *pt = pt_prev; 2474 } 2475 2476 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb) 2477 { 2478 if (!ptype->af_packet_priv || !skb->sk) 2479 return false; 2480 2481 if (ptype->id_match) 2482 return ptype->id_match(ptype, skb->sk); 2483 else if ((struct sock *)ptype->af_packet_priv == skb->sk) 2484 return true; 2485 2486 return false; 2487 } 2488 2489 /** 2490 * dev_nit_active - return true if any network interface taps are in use 2491 * 2492 * @dev: network device to check for the presence of taps 2493 */ 2494 bool dev_nit_active(struct net_device *dev) 2495 { 2496 return !list_empty(&net_hotdata.ptype_all) || 2497 !list_empty(&dev->ptype_all); 2498 } 2499 EXPORT_SYMBOL_GPL(dev_nit_active); 2500 2501 /* 2502 * Support routine. Sends outgoing frames to any network 2503 * taps currently in use. 2504 */ 2505 2506 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev) 2507 { 2508 struct list_head *ptype_list = &net_hotdata.ptype_all; 2509 struct packet_type *ptype, *pt_prev = NULL; 2510 struct sk_buff *skb2 = NULL; 2511 2512 rcu_read_lock(); 2513 again: 2514 list_for_each_entry_rcu(ptype, ptype_list, list) { 2515 if (READ_ONCE(ptype->ignore_outgoing)) 2516 continue; 2517 2518 /* Never send packets back to the socket 2519 * they originated from - MvS (miquels@drinkel.ow.org) 2520 */ 2521 if (skb_loop_sk(ptype, skb)) 2522 continue; 2523 2524 if (pt_prev) { 2525 deliver_skb(skb2, pt_prev, skb->dev); 2526 pt_prev = ptype; 2527 continue; 2528 } 2529 2530 /* need to clone skb, done only once */ 2531 skb2 = skb_clone(skb, GFP_ATOMIC); 2532 if (!skb2) 2533 goto out_unlock; 2534 2535 net_timestamp_set(skb2); 2536 2537 /* skb->nh should be correctly 2538 * set by sender, so that the second statement is 2539 * just protection against buggy protocols. 2540 */ 2541 skb_reset_mac_header(skb2); 2542 2543 if (skb_network_header(skb2) < skb2->data || 2544 skb_network_header(skb2) > skb_tail_pointer(skb2)) { 2545 net_crit_ratelimited("protocol %04x is buggy, dev %s\n", 2546 ntohs(skb2->protocol), 2547 dev->name); 2548 skb_reset_network_header(skb2); 2549 } 2550 2551 skb2->transport_header = skb2->network_header; 2552 skb2->pkt_type = PACKET_OUTGOING; 2553 pt_prev = ptype; 2554 } 2555 2556 if (ptype_list == &net_hotdata.ptype_all) { 2557 ptype_list = &dev->ptype_all; 2558 goto again; 2559 } 2560 out_unlock: 2561 if (pt_prev) { 2562 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC)) 2563 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev); 2564 else 2565 kfree_skb(skb2); 2566 } 2567 rcu_read_unlock(); 2568 } 2569 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit); 2570 2571 /** 2572 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change 2573 * @dev: Network device 2574 * @txq: number of queues available 2575 * 2576 * If real_num_tx_queues is changed the tc mappings may no longer be 2577 * valid. To resolve this verify the tc mapping remains valid and if 2578 * not NULL the mapping. With no priorities mapping to this 2579 * offset/count pair it will no longer be used. In the worst case TC0 2580 * is invalid nothing can be done so disable priority mappings. If is 2581 * expected that drivers will fix this mapping if they can before 2582 * calling netif_set_real_num_tx_queues. 2583 */ 2584 static void netif_setup_tc(struct net_device *dev, unsigned int txq) 2585 { 2586 int i; 2587 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2588 2589 /* If TC0 is invalidated disable TC mapping */ 2590 if (tc->offset + tc->count > txq) { 2591 netdev_warn(dev, "Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n"); 2592 dev->num_tc = 0; 2593 return; 2594 } 2595 2596 /* Invalidated prio to tc mappings set to TC0 */ 2597 for (i = 1; i < TC_BITMASK + 1; i++) { 2598 int q = netdev_get_prio_tc_map(dev, i); 2599 2600 tc = &dev->tc_to_txq[q]; 2601 if (tc->offset + tc->count > txq) { 2602 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", 2603 i, q); 2604 netdev_set_prio_tc_map(dev, i, 0); 2605 } 2606 } 2607 } 2608 2609 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq) 2610 { 2611 if (dev->num_tc) { 2612 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2613 int i; 2614 2615 /* walk through the TCs and see if it falls into any of them */ 2616 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) { 2617 if ((txq - tc->offset) < tc->count) 2618 return i; 2619 } 2620 2621 /* didn't find it, just return -1 to indicate no match */ 2622 return -1; 2623 } 2624 2625 return 0; 2626 } 2627 EXPORT_SYMBOL(netdev_txq_to_tc); 2628 2629 #ifdef CONFIG_XPS 2630 static struct static_key xps_needed __read_mostly; 2631 static struct static_key xps_rxqs_needed __read_mostly; 2632 static DEFINE_MUTEX(xps_map_mutex); 2633 #define xmap_dereference(P) \ 2634 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex)) 2635 2636 static bool remove_xps_queue(struct xps_dev_maps *dev_maps, 2637 struct xps_dev_maps *old_maps, int tci, u16 index) 2638 { 2639 struct xps_map *map = NULL; 2640 int pos; 2641 2642 map = xmap_dereference(dev_maps->attr_map[tci]); 2643 if (!map) 2644 return false; 2645 2646 for (pos = map->len; pos--;) { 2647 if (map->queues[pos] != index) 2648 continue; 2649 2650 if (map->len > 1) { 2651 map->queues[pos] = map->queues[--map->len]; 2652 break; 2653 } 2654 2655 if (old_maps) 2656 RCU_INIT_POINTER(old_maps->attr_map[tci], NULL); 2657 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2658 kfree_rcu(map, rcu); 2659 return false; 2660 } 2661 2662 return true; 2663 } 2664 2665 static bool remove_xps_queue_cpu(struct net_device *dev, 2666 struct xps_dev_maps *dev_maps, 2667 int cpu, u16 offset, u16 count) 2668 { 2669 int num_tc = dev_maps->num_tc; 2670 bool active = false; 2671 int tci; 2672 2673 for (tci = cpu * num_tc; num_tc--; tci++) { 2674 int i, j; 2675 2676 for (i = count, j = offset; i--; j++) { 2677 if (!remove_xps_queue(dev_maps, NULL, tci, j)) 2678 break; 2679 } 2680 2681 active |= i < 0; 2682 } 2683 2684 return active; 2685 } 2686 2687 static void reset_xps_maps(struct net_device *dev, 2688 struct xps_dev_maps *dev_maps, 2689 enum xps_map_type type) 2690 { 2691 static_key_slow_dec_cpuslocked(&xps_needed); 2692 if (type == XPS_RXQS) 2693 static_key_slow_dec_cpuslocked(&xps_rxqs_needed); 2694 2695 RCU_INIT_POINTER(dev->xps_maps[type], NULL); 2696 2697 kfree_rcu(dev_maps, rcu); 2698 } 2699 2700 static void clean_xps_maps(struct net_device *dev, enum xps_map_type type, 2701 u16 offset, u16 count) 2702 { 2703 struct xps_dev_maps *dev_maps; 2704 bool active = false; 2705 int i, j; 2706 2707 dev_maps = xmap_dereference(dev->xps_maps[type]); 2708 if (!dev_maps) 2709 return; 2710 2711 for (j = 0; j < dev_maps->nr_ids; j++) 2712 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, count); 2713 if (!active) 2714 reset_xps_maps(dev, dev_maps, type); 2715 2716 if (type == XPS_CPUS) { 2717 for (i = offset + (count - 1); count--; i--) 2718 netdev_queue_numa_node_write( 2719 netdev_get_tx_queue(dev, i), NUMA_NO_NODE); 2720 } 2721 } 2722 2723 static void netif_reset_xps_queues(struct net_device *dev, u16 offset, 2724 u16 count) 2725 { 2726 if (!static_key_false(&xps_needed)) 2727 return; 2728 2729 cpus_read_lock(); 2730 mutex_lock(&xps_map_mutex); 2731 2732 if (static_key_false(&xps_rxqs_needed)) 2733 clean_xps_maps(dev, XPS_RXQS, offset, count); 2734 2735 clean_xps_maps(dev, XPS_CPUS, offset, count); 2736 2737 mutex_unlock(&xps_map_mutex); 2738 cpus_read_unlock(); 2739 } 2740 2741 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index) 2742 { 2743 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index); 2744 } 2745 2746 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index, 2747 u16 index, bool is_rxqs_map) 2748 { 2749 struct xps_map *new_map; 2750 int alloc_len = XPS_MIN_MAP_ALLOC; 2751 int i, pos; 2752 2753 for (pos = 0; map && pos < map->len; pos++) { 2754 if (map->queues[pos] != index) 2755 continue; 2756 return map; 2757 } 2758 2759 /* Need to add tx-queue to this CPU's/rx-queue's existing map */ 2760 if (map) { 2761 if (pos < map->alloc_len) 2762 return map; 2763 2764 alloc_len = map->alloc_len * 2; 2765 } 2766 2767 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's 2768 * map 2769 */ 2770 if (is_rxqs_map) 2771 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL); 2772 else 2773 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL, 2774 cpu_to_node(attr_index)); 2775 if (!new_map) 2776 return NULL; 2777 2778 for (i = 0; i < pos; i++) 2779 new_map->queues[i] = map->queues[i]; 2780 new_map->alloc_len = alloc_len; 2781 new_map->len = pos; 2782 2783 return new_map; 2784 } 2785 2786 /* Copy xps maps at a given index */ 2787 static void xps_copy_dev_maps(struct xps_dev_maps *dev_maps, 2788 struct xps_dev_maps *new_dev_maps, int index, 2789 int tc, bool skip_tc) 2790 { 2791 int i, tci = index * dev_maps->num_tc; 2792 struct xps_map *map; 2793 2794 /* copy maps belonging to foreign traffic classes */ 2795 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2796 if (i == tc && skip_tc) 2797 continue; 2798 2799 /* fill in the new device map from the old device map */ 2800 map = xmap_dereference(dev_maps->attr_map[tci]); 2801 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2802 } 2803 } 2804 2805 /* Must be called under cpus_read_lock */ 2806 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask, 2807 u16 index, enum xps_map_type type) 2808 { 2809 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL, *old_dev_maps = NULL; 2810 const unsigned long *online_mask = NULL; 2811 bool active = false, copy = false; 2812 int i, j, tci, numa_node_id = -2; 2813 int maps_sz, num_tc = 1, tc = 0; 2814 struct xps_map *map, *new_map; 2815 unsigned int nr_ids; 2816 2817 WARN_ON_ONCE(index >= dev->num_tx_queues); 2818 2819 if (dev->num_tc) { 2820 /* Do not allow XPS on subordinate device directly */ 2821 num_tc = dev->num_tc; 2822 if (num_tc < 0) 2823 return -EINVAL; 2824 2825 /* If queue belongs to subordinate dev use its map */ 2826 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev; 2827 2828 tc = netdev_txq_to_tc(dev, index); 2829 if (tc < 0) 2830 return -EINVAL; 2831 } 2832 2833 mutex_lock(&xps_map_mutex); 2834 2835 dev_maps = xmap_dereference(dev->xps_maps[type]); 2836 if (type == XPS_RXQS) { 2837 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues); 2838 nr_ids = dev->num_rx_queues; 2839 } else { 2840 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc); 2841 if (num_possible_cpus() > 1) 2842 online_mask = cpumask_bits(cpu_online_mask); 2843 nr_ids = nr_cpu_ids; 2844 } 2845 2846 if (maps_sz < L1_CACHE_BYTES) 2847 maps_sz = L1_CACHE_BYTES; 2848 2849 /* The old dev_maps could be larger or smaller than the one we're 2850 * setting up now, as dev->num_tc or nr_ids could have been updated in 2851 * between. We could try to be smart, but let's be safe instead and only 2852 * copy foreign traffic classes if the two map sizes match. 2853 */ 2854 if (dev_maps && 2855 dev_maps->num_tc == num_tc && dev_maps->nr_ids == nr_ids) 2856 copy = true; 2857 2858 /* allocate memory for queue storage */ 2859 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids), 2860 j < nr_ids;) { 2861 if (!new_dev_maps) { 2862 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL); 2863 if (!new_dev_maps) { 2864 mutex_unlock(&xps_map_mutex); 2865 return -ENOMEM; 2866 } 2867 2868 new_dev_maps->nr_ids = nr_ids; 2869 new_dev_maps->num_tc = num_tc; 2870 } 2871 2872 tci = j * num_tc + tc; 2873 map = copy ? xmap_dereference(dev_maps->attr_map[tci]) : NULL; 2874 2875 map = expand_xps_map(map, j, index, type == XPS_RXQS); 2876 if (!map) 2877 goto error; 2878 2879 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2880 } 2881 2882 if (!new_dev_maps) 2883 goto out_no_new_maps; 2884 2885 if (!dev_maps) { 2886 /* Increment static keys at most once per type */ 2887 static_key_slow_inc_cpuslocked(&xps_needed); 2888 if (type == XPS_RXQS) 2889 static_key_slow_inc_cpuslocked(&xps_rxqs_needed); 2890 } 2891 2892 for (j = 0; j < nr_ids; j++) { 2893 bool skip_tc = false; 2894 2895 tci = j * num_tc + tc; 2896 if (netif_attr_test_mask(j, mask, nr_ids) && 2897 netif_attr_test_online(j, online_mask, nr_ids)) { 2898 /* add tx-queue to CPU/rx-queue maps */ 2899 int pos = 0; 2900 2901 skip_tc = true; 2902 2903 map = xmap_dereference(new_dev_maps->attr_map[tci]); 2904 while ((pos < map->len) && (map->queues[pos] != index)) 2905 pos++; 2906 2907 if (pos == map->len) 2908 map->queues[map->len++] = index; 2909 #ifdef CONFIG_NUMA 2910 if (type == XPS_CPUS) { 2911 if (numa_node_id == -2) 2912 numa_node_id = cpu_to_node(j); 2913 else if (numa_node_id != cpu_to_node(j)) 2914 numa_node_id = -1; 2915 } 2916 #endif 2917 } 2918 2919 if (copy) 2920 xps_copy_dev_maps(dev_maps, new_dev_maps, j, tc, 2921 skip_tc); 2922 } 2923 2924 rcu_assign_pointer(dev->xps_maps[type], new_dev_maps); 2925 2926 /* Cleanup old maps */ 2927 if (!dev_maps) 2928 goto out_no_old_maps; 2929 2930 for (j = 0; j < dev_maps->nr_ids; j++) { 2931 for (i = num_tc, tci = j * dev_maps->num_tc; i--; tci++) { 2932 map = xmap_dereference(dev_maps->attr_map[tci]); 2933 if (!map) 2934 continue; 2935 2936 if (copy) { 2937 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2938 if (map == new_map) 2939 continue; 2940 } 2941 2942 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2943 kfree_rcu(map, rcu); 2944 } 2945 } 2946 2947 old_dev_maps = dev_maps; 2948 2949 out_no_old_maps: 2950 dev_maps = new_dev_maps; 2951 active = true; 2952 2953 out_no_new_maps: 2954 if (type == XPS_CPUS) 2955 /* update Tx queue numa node */ 2956 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index), 2957 (numa_node_id >= 0) ? 2958 numa_node_id : NUMA_NO_NODE); 2959 2960 if (!dev_maps) 2961 goto out_no_maps; 2962 2963 /* removes tx-queue from unused CPUs/rx-queues */ 2964 for (j = 0; j < dev_maps->nr_ids; j++) { 2965 tci = j * dev_maps->num_tc; 2966 2967 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2968 if (i == tc && 2969 netif_attr_test_mask(j, mask, dev_maps->nr_ids) && 2970 netif_attr_test_online(j, online_mask, dev_maps->nr_ids)) 2971 continue; 2972 2973 active |= remove_xps_queue(dev_maps, 2974 copy ? old_dev_maps : NULL, 2975 tci, index); 2976 } 2977 } 2978 2979 if (old_dev_maps) 2980 kfree_rcu(old_dev_maps, rcu); 2981 2982 /* free map if not active */ 2983 if (!active) 2984 reset_xps_maps(dev, dev_maps, type); 2985 2986 out_no_maps: 2987 mutex_unlock(&xps_map_mutex); 2988 2989 return 0; 2990 error: 2991 /* remove any maps that we added */ 2992 for (j = 0; j < nr_ids; j++) { 2993 for (i = num_tc, tci = j * num_tc; i--; tci++) { 2994 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2995 map = copy ? 2996 xmap_dereference(dev_maps->attr_map[tci]) : 2997 NULL; 2998 if (new_map && new_map != map) 2999 kfree(new_map); 3000 } 3001 } 3002 3003 mutex_unlock(&xps_map_mutex); 3004 3005 kfree(new_dev_maps); 3006 return -ENOMEM; 3007 } 3008 EXPORT_SYMBOL_GPL(__netif_set_xps_queue); 3009 3010 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, 3011 u16 index) 3012 { 3013 int ret; 3014 3015 cpus_read_lock(); 3016 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, XPS_CPUS); 3017 cpus_read_unlock(); 3018 3019 return ret; 3020 } 3021 EXPORT_SYMBOL(netif_set_xps_queue); 3022 3023 #endif 3024 static void netdev_unbind_all_sb_channels(struct net_device *dev) 3025 { 3026 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 3027 3028 /* Unbind any subordinate channels */ 3029 while (txq-- != &dev->_tx[0]) { 3030 if (txq->sb_dev) 3031 netdev_unbind_sb_channel(dev, txq->sb_dev); 3032 } 3033 } 3034 3035 void netdev_reset_tc(struct net_device *dev) 3036 { 3037 #ifdef CONFIG_XPS 3038 netif_reset_xps_queues_gt(dev, 0); 3039 #endif 3040 netdev_unbind_all_sb_channels(dev); 3041 3042 /* Reset TC configuration of device */ 3043 dev->num_tc = 0; 3044 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq)); 3045 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map)); 3046 } 3047 EXPORT_SYMBOL(netdev_reset_tc); 3048 3049 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset) 3050 { 3051 if (tc >= dev->num_tc) 3052 return -EINVAL; 3053 3054 #ifdef CONFIG_XPS 3055 netif_reset_xps_queues(dev, offset, count); 3056 #endif 3057 dev->tc_to_txq[tc].count = count; 3058 dev->tc_to_txq[tc].offset = offset; 3059 return 0; 3060 } 3061 EXPORT_SYMBOL(netdev_set_tc_queue); 3062 3063 int netdev_set_num_tc(struct net_device *dev, u8 num_tc) 3064 { 3065 if (num_tc > TC_MAX_QUEUE) 3066 return -EINVAL; 3067 3068 #ifdef CONFIG_XPS 3069 netif_reset_xps_queues_gt(dev, 0); 3070 #endif 3071 netdev_unbind_all_sb_channels(dev); 3072 3073 dev->num_tc = num_tc; 3074 return 0; 3075 } 3076 EXPORT_SYMBOL(netdev_set_num_tc); 3077 3078 void netdev_unbind_sb_channel(struct net_device *dev, 3079 struct net_device *sb_dev) 3080 { 3081 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 3082 3083 #ifdef CONFIG_XPS 3084 netif_reset_xps_queues_gt(sb_dev, 0); 3085 #endif 3086 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq)); 3087 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map)); 3088 3089 while (txq-- != &dev->_tx[0]) { 3090 if (txq->sb_dev == sb_dev) 3091 txq->sb_dev = NULL; 3092 } 3093 } 3094 EXPORT_SYMBOL(netdev_unbind_sb_channel); 3095 3096 int netdev_bind_sb_channel_queue(struct net_device *dev, 3097 struct net_device *sb_dev, 3098 u8 tc, u16 count, u16 offset) 3099 { 3100 /* Make certain the sb_dev and dev are already configured */ 3101 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc) 3102 return -EINVAL; 3103 3104 /* We cannot hand out queues we don't have */ 3105 if ((offset + count) > dev->real_num_tx_queues) 3106 return -EINVAL; 3107 3108 /* Record the mapping */ 3109 sb_dev->tc_to_txq[tc].count = count; 3110 sb_dev->tc_to_txq[tc].offset = offset; 3111 3112 /* Provide a way for Tx queue to find the tc_to_txq map or 3113 * XPS map for itself. 3114 */ 3115 while (count--) 3116 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev; 3117 3118 return 0; 3119 } 3120 EXPORT_SYMBOL(netdev_bind_sb_channel_queue); 3121 3122 int netdev_set_sb_channel(struct net_device *dev, u16 channel) 3123 { 3124 /* Do not use a multiqueue device to represent a subordinate channel */ 3125 if (netif_is_multiqueue(dev)) 3126 return -ENODEV; 3127 3128 /* We allow channels 1 - 32767 to be used for subordinate channels. 3129 * Channel 0 is meant to be "native" mode and used only to represent 3130 * the main root device. We allow writing 0 to reset the device back 3131 * to normal mode after being used as a subordinate channel. 3132 */ 3133 if (channel > S16_MAX) 3134 return -EINVAL; 3135 3136 dev->num_tc = -channel; 3137 3138 return 0; 3139 } 3140 EXPORT_SYMBOL(netdev_set_sb_channel); 3141 3142 /* 3143 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues 3144 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed. 3145 */ 3146 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq) 3147 { 3148 bool disabling; 3149 int rc; 3150 3151 disabling = txq < dev->real_num_tx_queues; 3152 3153 if (txq < 1 || txq > dev->num_tx_queues) 3154 return -EINVAL; 3155 3156 if (dev->reg_state == NETREG_REGISTERED || 3157 dev->reg_state == NETREG_UNREGISTERING) { 3158 ASSERT_RTNL(); 3159 3160 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues, 3161 txq); 3162 if (rc) 3163 return rc; 3164 3165 if (dev->num_tc) 3166 netif_setup_tc(dev, txq); 3167 3168 net_shaper_set_real_num_tx_queues(dev, txq); 3169 3170 dev_qdisc_change_real_num_tx(dev, txq); 3171 3172 dev->real_num_tx_queues = txq; 3173 3174 if (disabling) { 3175 synchronize_net(); 3176 qdisc_reset_all_tx_gt(dev, txq); 3177 #ifdef CONFIG_XPS 3178 netif_reset_xps_queues_gt(dev, txq); 3179 #endif 3180 } 3181 } else { 3182 dev->real_num_tx_queues = txq; 3183 } 3184 3185 return 0; 3186 } 3187 EXPORT_SYMBOL(netif_set_real_num_tx_queues); 3188 3189 #ifdef CONFIG_SYSFS 3190 /** 3191 * netif_set_real_num_rx_queues - set actual number of RX queues used 3192 * @dev: Network device 3193 * @rxq: Actual number of RX queues 3194 * 3195 * This must be called either with the rtnl_lock held or before 3196 * registration of the net device. Returns 0 on success, or a 3197 * negative error code. If called before registration, it always 3198 * succeeds. 3199 */ 3200 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq) 3201 { 3202 int rc; 3203 3204 if (rxq < 1 || rxq > dev->num_rx_queues) 3205 return -EINVAL; 3206 3207 if (dev->reg_state == NETREG_REGISTERED) { 3208 ASSERT_RTNL(); 3209 3210 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues, 3211 rxq); 3212 if (rc) 3213 return rc; 3214 } 3215 3216 dev->real_num_rx_queues = rxq; 3217 return 0; 3218 } 3219 EXPORT_SYMBOL(netif_set_real_num_rx_queues); 3220 #endif 3221 3222 /** 3223 * netif_set_real_num_queues - set actual number of RX and TX queues used 3224 * @dev: Network device 3225 * @txq: Actual number of TX queues 3226 * @rxq: Actual number of RX queues 3227 * 3228 * Set the real number of both TX and RX queues. 3229 * Does nothing if the number of queues is already correct. 3230 */ 3231 int netif_set_real_num_queues(struct net_device *dev, 3232 unsigned int txq, unsigned int rxq) 3233 { 3234 unsigned int old_rxq = dev->real_num_rx_queues; 3235 int err; 3236 3237 if (txq < 1 || txq > dev->num_tx_queues || 3238 rxq < 1 || rxq > dev->num_rx_queues) 3239 return -EINVAL; 3240 3241 /* Start from increases, so the error path only does decreases - 3242 * decreases can't fail. 3243 */ 3244 if (rxq > dev->real_num_rx_queues) { 3245 err = netif_set_real_num_rx_queues(dev, rxq); 3246 if (err) 3247 return err; 3248 } 3249 if (txq > dev->real_num_tx_queues) { 3250 err = netif_set_real_num_tx_queues(dev, txq); 3251 if (err) 3252 goto undo_rx; 3253 } 3254 if (rxq < dev->real_num_rx_queues) 3255 WARN_ON(netif_set_real_num_rx_queues(dev, rxq)); 3256 if (txq < dev->real_num_tx_queues) 3257 WARN_ON(netif_set_real_num_tx_queues(dev, txq)); 3258 3259 return 0; 3260 undo_rx: 3261 WARN_ON(netif_set_real_num_rx_queues(dev, old_rxq)); 3262 return err; 3263 } 3264 EXPORT_SYMBOL(netif_set_real_num_queues); 3265 3266 /** 3267 * netif_set_tso_max_size() - set the max size of TSO frames supported 3268 * @dev: netdev to update 3269 * @size: max skb->len of a TSO frame 3270 * 3271 * Set the limit on the size of TSO super-frames the device can handle. 3272 * Unless explicitly set the stack will assume the value of 3273 * %GSO_LEGACY_MAX_SIZE. 3274 */ 3275 void netif_set_tso_max_size(struct net_device *dev, unsigned int size) 3276 { 3277 dev->tso_max_size = min(GSO_MAX_SIZE, size); 3278 if (size < READ_ONCE(dev->gso_max_size)) 3279 netif_set_gso_max_size(dev, size); 3280 if (size < READ_ONCE(dev->gso_ipv4_max_size)) 3281 netif_set_gso_ipv4_max_size(dev, size); 3282 } 3283 EXPORT_SYMBOL(netif_set_tso_max_size); 3284 3285 /** 3286 * netif_set_tso_max_segs() - set the max number of segs supported for TSO 3287 * @dev: netdev to update 3288 * @segs: max number of TCP segments 3289 * 3290 * Set the limit on the number of TCP segments the device can generate from 3291 * a single TSO super-frame. 3292 * Unless explicitly set the stack will assume the value of %GSO_MAX_SEGS. 3293 */ 3294 void netif_set_tso_max_segs(struct net_device *dev, unsigned int segs) 3295 { 3296 dev->tso_max_segs = segs; 3297 if (segs < READ_ONCE(dev->gso_max_segs)) 3298 netif_set_gso_max_segs(dev, segs); 3299 } 3300 EXPORT_SYMBOL(netif_set_tso_max_segs); 3301 3302 /** 3303 * netif_inherit_tso_max() - copy all TSO limits from a lower device to an upper 3304 * @to: netdev to update 3305 * @from: netdev from which to copy the limits 3306 */ 3307 void netif_inherit_tso_max(struct net_device *to, const struct net_device *from) 3308 { 3309 netif_set_tso_max_size(to, from->tso_max_size); 3310 netif_set_tso_max_segs(to, from->tso_max_segs); 3311 } 3312 EXPORT_SYMBOL(netif_inherit_tso_max); 3313 3314 /** 3315 * netif_get_num_default_rss_queues - default number of RSS queues 3316 * 3317 * Default value is the number of physical cores if there are only 1 or 2, or 3318 * divided by 2 if there are more. 3319 */ 3320 int netif_get_num_default_rss_queues(void) 3321 { 3322 cpumask_var_t cpus; 3323 int cpu, count = 0; 3324 3325 if (unlikely(is_kdump_kernel() || !zalloc_cpumask_var(&cpus, GFP_KERNEL))) 3326 return 1; 3327 3328 cpumask_copy(cpus, cpu_online_mask); 3329 for_each_cpu(cpu, cpus) { 3330 ++count; 3331 cpumask_andnot(cpus, cpus, topology_sibling_cpumask(cpu)); 3332 } 3333 free_cpumask_var(cpus); 3334 3335 return count > 2 ? DIV_ROUND_UP(count, 2) : count; 3336 } 3337 EXPORT_SYMBOL(netif_get_num_default_rss_queues); 3338 3339 static void __netif_reschedule(struct Qdisc *q) 3340 { 3341 struct softnet_data *sd; 3342 unsigned long flags; 3343 3344 local_irq_save(flags); 3345 sd = this_cpu_ptr(&softnet_data); 3346 q->next_sched = NULL; 3347 *sd->output_queue_tailp = q; 3348 sd->output_queue_tailp = &q->next_sched; 3349 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3350 local_irq_restore(flags); 3351 } 3352 3353 void __netif_schedule(struct Qdisc *q) 3354 { 3355 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state)) 3356 __netif_reschedule(q); 3357 } 3358 EXPORT_SYMBOL(__netif_schedule); 3359 3360 struct dev_kfree_skb_cb { 3361 enum skb_drop_reason reason; 3362 }; 3363 3364 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb) 3365 { 3366 return (struct dev_kfree_skb_cb *)skb->cb; 3367 } 3368 3369 void netif_schedule_queue(struct netdev_queue *txq) 3370 { 3371 rcu_read_lock(); 3372 if (!netif_xmit_stopped(txq)) { 3373 struct Qdisc *q = rcu_dereference(txq->qdisc); 3374 3375 __netif_schedule(q); 3376 } 3377 rcu_read_unlock(); 3378 } 3379 EXPORT_SYMBOL(netif_schedule_queue); 3380 3381 void netif_tx_wake_queue(struct netdev_queue *dev_queue) 3382 { 3383 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) { 3384 struct Qdisc *q; 3385 3386 rcu_read_lock(); 3387 q = rcu_dereference(dev_queue->qdisc); 3388 __netif_schedule(q); 3389 rcu_read_unlock(); 3390 } 3391 } 3392 EXPORT_SYMBOL(netif_tx_wake_queue); 3393 3394 void dev_kfree_skb_irq_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3395 { 3396 unsigned long flags; 3397 3398 if (unlikely(!skb)) 3399 return; 3400 3401 if (likely(refcount_read(&skb->users) == 1)) { 3402 smp_rmb(); 3403 refcount_set(&skb->users, 0); 3404 } else if (likely(!refcount_dec_and_test(&skb->users))) { 3405 return; 3406 } 3407 get_kfree_skb_cb(skb)->reason = reason; 3408 local_irq_save(flags); 3409 skb->next = __this_cpu_read(softnet_data.completion_queue); 3410 __this_cpu_write(softnet_data.completion_queue, skb); 3411 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3412 local_irq_restore(flags); 3413 } 3414 EXPORT_SYMBOL(dev_kfree_skb_irq_reason); 3415 3416 void dev_kfree_skb_any_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3417 { 3418 if (in_hardirq() || irqs_disabled()) 3419 dev_kfree_skb_irq_reason(skb, reason); 3420 else 3421 kfree_skb_reason(skb, reason); 3422 } 3423 EXPORT_SYMBOL(dev_kfree_skb_any_reason); 3424 3425 3426 /** 3427 * netif_device_detach - mark device as removed 3428 * @dev: network device 3429 * 3430 * Mark device as removed from system and therefore no longer available. 3431 */ 3432 void netif_device_detach(struct net_device *dev) 3433 { 3434 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) && 3435 netif_running(dev)) { 3436 netif_tx_stop_all_queues(dev); 3437 } 3438 } 3439 EXPORT_SYMBOL(netif_device_detach); 3440 3441 /** 3442 * netif_device_attach - mark device as attached 3443 * @dev: network device 3444 * 3445 * Mark device as attached from system and restart if needed. 3446 */ 3447 void netif_device_attach(struct net_device *dev) 3448 { 3449 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) && 3450 netif_running(dev)) { 3451 netif_tx_wake_all_queues(dev); 3452 netdev_watchdog_up(dev); 3453 } 3454 } 3455 EXPORT_SYMBOL(netif_device_attach); 3456 3457 /* 3458 * Returns a Tx hash based on the given packet descriptor a Tx queues' number 3459 * to be used as a distribution range. 3460 */ 3461 static u16 skb_tx_hash(const struct net_device *dev, 3462 const struct net_device *sb_dev, 3463 struct sk_buff *skb) 3464 { 3465 u32 hash; 3466 u16 qoffset = 0; 3467 u16 qcount = dev->real_num_tx_queues; 3468 3469 if (dev->num_tc) { 3470 u8 tc = netdev_get_prio_tc_map(dev, skb->priority); 3471 3472 qoffset = sb_dev->tc_to_txq[tc].offset; 3473 qcount = sb_dev->tc_to_txq[tc].count; 3474 if (unlikely(!qcount)) { 3475 net_warn_ratelimited("%s: invalid qcount, qoffset %u for tc %u\n", 3476 sb_dev->name, qoffset, tc); 3477 qoffset = 0; 3478 qcount = dev->real_num_tx_queues; 3479 } 3480 } 3481 3482 if (skb_rx_queue_recorded(skb)) { 3483 DEBUG_NET_WARN_ON_ONCE(qcount == 0); 3484 hash = skb_get_rx_queue(skb); 3485 if (hash >= qoffset) 3486 hash -= qoffset; 3487 while (unlikely(hash >= qcount)) 3488 hash -= qcount; 3489 return hash + qoffset; 3490 } 3491 3492 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset; 3493 } 3494 3495 void skb_warn_bad_offload(const struct sk_buff *skb) 3496 { 3497 static const netdev_features_t null_features; 3498 struct net_device *dev = skb->dev; 3499 const char *name = ""; 3500 3501 if (!net_ratelimit()) 3502 return; 3503 3504 if (dev) { 3505 if (dev->dev.parent) 3506 name = dev_driver_string(dev->dev.parent); 3507 else 3508 name = netdev_name(dev); 3509 } 3510 skb_dump(KERN_WARNING, skb, false); 3511 WARN(1, "%s: caps=(%pNF, %pNF)\n", 3512 name, dev ? &dev->features : &null_features, 3513 skb->sk ? &skb->sk->sk_route_caps : &null_features); 3514 } 3515 3516 /* 3517 * Invalidate hardware checksum when packet is to be mangled, and 3518 * complete checksum manually on outgoing path. 3519 */ 3520 int skb_checksum_help(struct sk_buff *skb) 3521 { 3522 __wsum csum; 3523 int ret = 0, offset; 3524 3525 if (skb->ip_summed == CHECKSUM_COMPLETE) 3526 goto out_set_summed; 3527 3528 if (unlikely(skb_is_gso(skb))) { 3529 skb_warn_bad_offload(skb); 3530 return -EINVAL; 3531 } 3532 3533 if (!skb_frags_readable(skb)) { 3534 return -EFAULT; 3535 } 3536 3537 /* Before computing a checksum, we should make sure no frag could 3538 * be modified by an external entity : checksum could be wrong. 3539 */ 3540 if (skb_has_shared_frag(skb)) { 3541 ret = __skb_linearize(skb); 3542 if (ret) 3543 goto out; 3544 } 3545 3546 offset = skb_checksum_start_offset(skb); 3547 ret = -EINVAL; 3548 if (unlikely(offset >= skb_headlen(skb))) { 3549 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3550 WARN_ONCE(true, "offset (%d) >= skb_headlen() (%u)\n", 3551 offset, skb_headlen(skb)); 3552 goto out; 3553 } 3554 csum = skb_checksum(skb, offset, skb->len - offset, 0); 3555 3556 offset += skb->csum_offset; 3557 if (unlikely(offset + sizeof(__sum16) > skb_headlen(skb))) { 3558 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3559 WARN_ONCE(true, "offset+2 (%zu) > skb_headlen() (%u)\n", 3560 offset + sizeof(__sum16), skb_headlen(skb)); 3561 goto out; 3562 } 3563 ret = skb_ensure_writable(skb, offset + sizeof(__sum16)); 3564 if (ret) 3565 goto out; 3566 3567 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0; 3568 out_set_summed: 3569 skb->ip_summed = CHECKSUM_NONE; 3570 out: 3571 return ret; 3572 } 3573 EXPORT_SYMBOL(skb_checksum_help); 3574 3575 int skb_crc32c_csum_help(struct sk_buff *skb) 3576 { 3577 __le32 crc32c_csum; 3578 int ret = 0, offset, start; 3579 3580 if (skb->ip_summed != CHECKSUM_PARTIAL) 3581 goto out; 3582 3583 if (unlikely(skb_is_gso(skb))) 3584 goto out; 3585 3586 /* Before computing a checksum, we should make sure no frag could 3587 * be modified by an external entity : checksum could be wrong. 3588 */ 3589 if (unlikely(skb_has_shared_frag(skb))) { 3590 ret = __skb_linearize(skb); 3591 if (ret) 3592 goto out; 3593 } 3594 start = skb_checksum_start_offset(skb); 3595 offset = start + offsetof(struct sctphdr, checksum); 3596 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) { 3597 ret = -EINVAL; 3598 goto out; 3599 } 3600 3601 ret = skb_ensure_writable(skb, offset + sizeof(__le32)); 3602 if (ret) 3603 goto out; 3604 3605 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start, 3606 skb->len - start, ~(__u32)0, 3607 crc32c_csum_stub)); 3608 *(__le32 *)(skb->data + offset) = crc32c_csum; 3609 skb_reset_csum_not_inet(skb); 3610 out: 3611 return ret; 3612 } 3613 EXPORT_SYMBOL(skb_crc32c_csum_help); 3614 3615 __be16 skb_network_protocol(struct sk_buff *skb, int *depth) 3616 { 3617 __be16 type = skb->protocol; 3618 3619 /* Tunnel gso handlers can set protocol to ethernet. */ 3620 if (type == htons(ETH_P_TEB)) { 3621 struct ethhdr *eth; 3622 3623 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr)))) 3624 return 0; 3625 3626 eth = (struct ethhdr *)skb->data; 3627 type = eth->h_proto; 3628 } 3629 3630 return vlan_get_protocol_and_depth(skb, type, depth); 3631 } 3632 3633 3634 /* Take action when hardware reception checksum errors are detected. */ 3635 #ifdef CONFIG_BUG 3636 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3637 { 3638 netdev_err(dev, "hw csum failure\n"); 3639 skb_dump(KERN_ERR, skb, true); 3640 dump_stack(); 3641 } 3642 3643 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3644 { 3645 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb); 3646 } 3647 EXPORT_SYMBOL(netdev_rx_csum_fault); 3648 #endif 3649 3650 /* XXX: check that highmem exists at all on the given machine. */ 3651 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb) 3652 { 3653 #ifdef CONFIG_HIGHMEM 3654 int i; 3655 3656 if (!(dev->features & NETIF_F_HIGHDMA)) { 3657 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 3658 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 3659 struct page *page = skb_frag_page(frag); 3660 3661 if (page && PageHighMem(page)) 3662 return 1; 3663 } 3664 } 3665 #endif 3666 return 0; 3667 } 3668 3669 /* If MPLS offload request, verify we are testing hardware MPLS features 3670 * instead of standard features for the netdev. 3671 */ 3672 #if IS_ENABLED(CONFIG_NET_MPLS_GSO) 3673 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3674 netdev_features_t features, 3675 __be16 type) 3676 { 3677 if (eth_p_mpls(type)) 3678 features &= skb->dev->mpls_features; 3679 3680 return features; 3681 } 3682 #else 3683 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3684 netdev_features_t features, 3685 __be16 type) 3686 { 3687 return features; 3688 } 3689 #endif 3690 3691 static netdev_features_t harmonize_features(struct sk_buff *skb, 3692 netdev_features_t features) 3693 { 3694 __be16 type; 3695 3696 type = skb_network_protocol(skb, NULL); 3697 features = net_mpls_features(skb, features, type); 3698 3699 if (skb->ip_summed != CHECKSUM_NONE && 3700 !can_checksum_protocol(features, type)) { 3701 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3702 } 3703 if (illegal_highdma(skb->dev, skb)) 3704 features &= ~NETIF_F_SG; 3705 3706 return features; 3707 } 3708 3709 netdev_features_t passthru_features_check(struct sk_buff *skb, 3710 struct net_device *dev, 3711 netdev_features_t features) 3712 { 3713 return features; 3714 } 3715 EXPORT_SYMBOL(passthru_features_check); 3716 3717 static netdev_features_t dflt_features_check(struct sk_buff *skb, 3718 struct net_device *dev, 3719 netdev_features_t features) 3720 { 3721 return vlan_features_check(skb, features); 3722 } 3723 3724 static netdev_features_t gso_features_check(const struct sk_buff *skb, 3725 struct net_device *dev, 3726 netdev_features_t features) 3727 { 3728 u16 gso_segs = skb_shinfo(skb)->gso_segs; 3729 3730 if (gso_segs > READ_ONCE(dev->gso_max_segs)) 3731 return features & ~NETIF_F_GSO_MASK; 3732 3733 if (unlikely(skb->len >= netif_get_gso_max_size(dev, skb))) 3734 return features & ~NETIF_F_GSO_MASK; 3735 3736 if (!skb_shinfo(skb)->gso_type) { 3737 skb_warn_bad_offload(skb); 3738 return features & ~NETIF_F_GSO_MASK; 3739 } 3740 3741 /* Support for GSO partial features requires software 3742 * intervention before we can actually process the packets 3743 * so we need to strip support for any partial features now 3744 * and we can pull them back in after we have partially 3745 * segmented the frame. 3746 */ 3747 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)) 3748 features &= ~dev->gso_partial_features; 3749 3750 /* Make sure to clear the IPv4 ID mangling feature if the 3751 * IPv4 header has the potential to be fragmented. 3752 */ 3753 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) { 3754 struct iphdr *iph = skb->encapsulation ? 3755 inner_ip_hdr(skb) : ip_hdr(skb); 3756 3757 if (!(iph->frag_off & htons(IP_DF))) 3758 features &= ~NETIF_F_TSO_MANGLEID; 3759 } 3760 3761 return features; 3762 } 3763 3764 netdev_features_t netif_skb_features(struct sk_buff *skb) 3765 { 3766 struct net_device *dev = skb->dev; 3767 netdev_features_t features = dev->features; 3768 3769 if (skb_is_gso(skb)) 3770 features = gso_features_check(skb, dev, features); 3771 3772 /* If encapsulation offload request, verify we are testing 3773 * hardware encapsulation features instead of standard 3774 * features for the netdev 3775 */ 3776 if (skb->encapsulation) 3777 features &= dev->hw_enc_features; 3778 3779 if (skb_vlan_tagged(skb)) 3780 features = netdev_intersect_features(features, 3781 dev->vlan_features | 3782 NETIF_F_HW_VLAN_CTAG_TX | 3783 NETIF_F_HW_VLAN_STAG_TX); 3784 3785 if (dev->netdev_ops->ndo_features_check) 3786 features &= dev->netdev_ops->ndo_features_check(skb, dev, 3787 features); 3788 else 3789 features &= dflt_features_check(skb, dev, features); 3790 3791 return harmonize_features(skb, features); 3792 } 3793 EXPORT_SYMBOL(netif_skb_features); 3794 3795 static int xmit_one(struct sk_buff *skb, struct net_device *dev, 3796 struct netdev_queue *txq, bool more) 3797 { 3798 unsigned int len; 3799 int rc; 3800 3801 if (dev_nit_active(dev)) 3802 dev_queue_xmit_nit(skb, dev); 3803 3804 len = skb->len; 3805 trace_net_dev_start_xmit(skb, dev); 3806 rc = netdev_start_xmit(skb, dev, txq, more); 3807 trace_net_dev_xmit(skb, rc, dev, len); 3808 3809 return rc; 3810 } 3811 3812 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev, 3813 struct netdev_queue *txq, int *ret) 3814 { 3815 struct sk_buff *skb = first; 3816 int rc = NETDEV_TX_OK; 3817 3818 while (skb) { 3819 struct sk_buff *next = skb->next; 3820 3821 skb_mark_not_on_list(skb); 3822 rc = xmit_one(skb, dev, txq, next != NULL); 3823 if (unlikely(!dev_xmit_complete(rc))) { 3824 skb->next = next; 3825 goto out; 3826 } 3827 3828 skb = next; 3829 if (netif_tx_queue_stopped(txq) && skb) { 3830 rc = NETDEV_TX_BUSY; 3831 break; 3832 } 3833 } 3834 3835 out: 3836 *ret = rc; 3837 return skb; 3838 } 3839 3840 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb, 3841 netdev_features_t features) 3842 { 3843 if (skb_vlan_tag_present(skb) && 3844 !vlan_hw_offload_capable(features, skb->vlan_proto)) 3845 skb = __vlan_hwaccel_push_inside(skb); 3846 return skb; 3847 } 3848 3849 int skb_csum_hwoffload_help(struct sk_buff *skb, 3850 const netdev_features_t features) 3851 { 3852 if (unlikely(skb_csum_is_sctp(skb))) 3853 return !!(features & NETIF_F_SCTP_CRC) ? 0 : 3854 skb_crc32c_csum_help(skb); 3855 3856 if (features & NETIF_F_HW_CSUM) 3857 return 0; 3858 3859 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) { 3860 if (vlan_get_protocol(skb) == htons(ETH_P_IPV6) && 3861 skb_network_header_len(skb) != sizeof(struct ipv6hdr) && 3862 !ipv6_has_hopopt_jumbo(skb)) 3863 goto sw_checksum; 3864 3865 switch (skb->csum_offset) { 3866 case offsetof(struct tcphdr, check): 3867 case offsetof(struct udphdr, check): 3868 return 0; 3869 } 3870 } 3871 3872 sw_checksum: 3873 return skb_checksum_help(skb); 3874 } 3875 EXPORT_SYMBOL(skb_csum_hwoffload_help); 3876 3877 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again) 3878 { 3879 netdev_features_t features; 3880 3881 features = netif_skb_features(skb); 3882 skb = validate_xmit_vlan(skb, features); 3883 if (unlikely(!skb)) 3884 goto out_null; 3885 3886 skb = sk_validate_xmit_skb(skb, dev); 3887 if (unlikely(!skb)) 3888 goto out_null; 3889 3890 if (netif_needs_gso(skb, features)) { 3891 struct sk_buff *segs; 3892 3893 segs = skb_gso_segment(skb, features); 3894 if (IS_ERR(segs)) { 3895 goto out_kfree_skb; 3896 } else if (segs) { 3897 consume_skb(skb); 3898 skb = segs; 3899 } 3900 } else { 3901 if (skb_needs_linearize(skb, features) && 3902 __skb_linearize(skb)) 3903 goto out_kfree_skb; 3904 3905 /* If packet is not checksummed and device does not 3906 * support checksumming for this protocol, complete 3907 * checksumming here. 3908 */ 3909 if (skb->ip_summed == CHECKSUM_PARTIAL) { 3910 if (skb->encapsulation) 3911 skb_set_inner_transport_header(skb, 3912 skb_checksum_start_offset(skb)); 3913 else 3914 skb_set_transport_header(skb, 3915 skb_checksum_start_offset(skb)); 3916 if (skb_csum_hwoffload_help(skb, features)) 3917 goto out_kfree_skb; 3918 } 3919 } 3920 3921 skb = validate_xmit_xfrm(skb, features, again); 3922 3923 return skb; 3924 3925 out_kfree_skb: 3926 kfree_skb(skb); 3927 out_null: 3928 dev_core_stats_tx_dropped_inc(dev); 3929 return NULL; 3930 } 3931 3932 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again) 3933 { 3934 struct sk_buff *next, *head = NULL, *tail; 3935 3936 for (; skb != NULL; skb = next) { 3937 next = skb->next; 3938 skb_mark_not_on_list(skb); 3939 3940 /* in case skb won't be segmented, point to itself */ 3941 skb->prev = skb; 3942 3943 skb = validate_xmit_skb(skb, dev, again); 3944 if (!skb) 3945 continue; 3946 3947 if (!head) 3948 head = skb; 3949 else 3950 tail->next = skb; 3951 /* If skb was segmented, skb->prev points to 3952 * the last segment. If not, it still contains skb. 3953 */ 3954 tail = skb->prev; 3955 } 3956 return head; 3957 } 3958 EXPORT_SYMBOL_GPL(validate_xmit_skb_list); 3959 3960 static void qdisc_pkt_len_init(struct sk_buff *skb) 3961 { 3962 const struct skb_shared_info *shinfo = skb_shinfo(skb); 3963 3964 qdisc_skb_cb(skb)->pkt_len = skb->len; 3965 3966 /* To get more precise estimation of bytes sent on wire, 3967 * we add to pkt_len the headers size of all segments 3968 */ 3969 if (shinfo->gso_size && skb_transport_header_was_set(skb)) { 3970 u16 gso_segs = shinfo->gso_segs; 3971 unsigned int hdr_len; 3972 3973 /* mac layer + network layer */ 3974 hdr_len = skb_transport_offset(skb); 3975 3976 /* + transport layer */ 3977 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) { 3978 const struct tcphdr *th; 3979 struct tcphdr _tcphdr; 3980 3981 th = skb_header_pointer(skb, hdr_len, 3982 sizeof(_tcphdr), &_tcphdr); 3983 if (likely(th)) 3984 hdr_len += __tcp_hdrlen(th); 3985 } else if (shinfo->gso_type & SKB_GSO_UDP_L4) { 3986 struct udphdr _udphdr; 3987 3988 if (skb_header_pointer(skb, hdr_len, 3989 sizeof(_udphdr), &_udphdr)) 3990 hdr_len += sizeof(struct udphdr); 3991 } 3992 3993 if (unlikely(shinfo->gso_type & SKB_GSO_DODGY)) { 3994 int payload = skb->len - hdr_len; 3995 3996 /* Malicious packet. */ 3997 if (payload <= 0) 3998 return; 3999 gso_segs = DIV_ROUND_UP(payload, shinfo->gso_size); 4000 } 4001 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len; 4002 } 4003 } 4004 4005 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q, 4006 struct sk_buff **to_free, 4007 struct netdev_queue *txq) 4008 { 4009 int rc; 4010 4011 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK; 4012 if (rc == NET_XMIT_SUCCESS) 4013 trace_qdisc_enqueue(q, txq, skb); 4014 return rc; 4015 } 4016 4017 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q, 4018 struct net_device *dev, 4019 struct netdev_queue *txq) 4020 { 4021 spinlock_t *root_lock = qdisc_lock(q); 4022 struct sk_buff *to_free = NULL; 4023 bool contended; 4024 int rc; 4025 4026 qdisc_calculate_pkt_len(skb, q); 4027 4028 tcf_set_drop_reason(skb, SKB_DROP_REASON_QDISC_DROP); 4029 4030 if (q->flags & TCQ_F_NOLOCK) { 4031 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) && 4032 qdisc_run_begin(q)) { 4033 /* Retest nolock_qdisc_is_empty() within the protection 4034 * of q->seqlock to protect from racing with requeuing. 4035 */ 4036 if (unlikely(!nolock_qdisc_is_empty(q))) { 4037 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4038 __qdisc_run(q); 4039 qdisc_run_end(q); 4040 4041 goto no_lock_out; 4042 } 4043 4044 qdisc_bstats_cpu_update(q, skb); 4045 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) && 4046 !nolock_qdisc_is_empty(q)) 4047 __qdisc_run(q); 4048 4049 qdisc_run_end(q); 4050 return NET_XMIT_SUCCESS; 4051 } 4052 4053 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4054 qdisc_run(q); 4055 4056 no_lock_out: 4057 if (unlikely(to_free)) 4058 kfree_skb_list_reason(to_free, 4059 tcf_get_drop_reason(to_free)); 4060 return rc; 4061 } 4062 4063 if (unlikely(READ_ONCE(q->owner) == smp_processor_id())) { 4064 kfree_skb_reason(skb, SKB_DROP_REASON_TC_RECLASSIFY_LOOP); 4065 return NET_XMIT_DROP; 4066 } 4067 /* 4068 * Heuristic to force contended enqueues to serialize on a 4069 * separate lock before trying to get qdisc main lock. 4070 * This permits qdisc->running owner to get the lock more 4071 * often and dequeue packets faster. 4072 * On PREEMPT_RT it is possible to preempt the qdisc owner during xmit 4073 * and then other tasks will only enqueue packets. The packets will be 4074 * sent after the qdisc owner is scheduled again. To prevent this 4075 * scenario the task always serialize on the lock. 4076 */ 4077 contended = qdisc_is_running(q) || IS_ENABLED(CONFIG_PREEMPT_RT); 4078 if (unlikely(contended)) 4079 spin_lock(&q->busylock); 4080 4081 spin_lock(root_lock); 4082 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) { 4083 __qdisc_drop(skb, &to_free); 4084 rc = NET_XMIT_DROP; 4085 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && 4086 qdisc_run_begin(q)) { 4087 /* 4088 * This is a work-conserving queue; there are no old skbs 4089 * waiting to be sent out; and the qdisc is not running - 4090 * xmit the skb directly. 4091 */ 4092 4093 qdisc_bstats_update(q, skb); 4094 4095 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) { 4096 if (unlikely(contended)) { 4097 spin_unlock(&q->busylock); 4098 contended = false; 4099 } 4100 __qdisc_run(q); 4101 } 4102 4103 qdisc_run_end(q); 4104 rc = NET_XMIT_SUCCESS; 4105 } else { 4106 WRITE_ONCE(q->owner, smp_processor_id()); 4107 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4108 WRITE_ONCE(q->owner, -1); 4109 if (qdisc_run_begin(q)) { 4110 if (unlikely(contended)) { 4111 spin_unlock(&q->busylock); 4112 contended = false; 4113 } 4114 __qdisc_run(q); 4115 qdisc_run_end(q); 4116 } 4117 } 4118 spin_unlock(root_lock); 4119 if (unlikely(to_free)) 4120 kfree_skb_list_reason(to_free, 4121 tcf_get_drop_reason(to_free)); 4122 if (unlikely(contended)) 4123 spin_unlock(&q->busylock); 4124 return rc; 4125 } 4126 4127 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO) 4128 static void skb_update_prio(struct sk_buff *skb) 4129 { 4130 const struct netprio_map *map; 4131 const struct sock *sk; 4132 unsigned int prioidx; 4133 4134 if (skb->priority) 4135 return; 4136 map = rcu_dereference_bh(skb->dev->priomap); 4137 if (!map) 4138 return; 4139 sk = skb_to_full_sk(skb); 4140 if (!sk) 4141 return; 4142 4143 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data); 4144 4145 if (prioidx < map->priomap_len) 4146 skb->priority = map->priomap[prioidx]; 4147 } 4148 #else 4149 #define skb_update_prio(skb) 4150 #endif 4151 4152 /** 4153 * dev_loopback_xmit - loop back @skb 4154 * @net: network namespace this loopback is happening in 4155 * @sk: sk needed to be a netfilter okfn 4156 * @skb: buffer to transmit 4157 */ 4158 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb) 4159 { 4160 skb_reset_mac_header(skb); 4161 __skb_pull(skb, skb_network_offset(skb)); 4162 skb->pkt_type = PACKET_LOOPBACK; 4163 if (skb->ip_summed == CHECKSUM_NONE) 4164 skb->ip_summed = CHECKSUM_UNNECESSARY; 4165 DEBUG_NET_WARN_ON_ONCE(!skb_dst(skb)); 4166 skb_dst_force(skb); 4167 netif_rx(skb); 4168 return 0; 4169 } 4170 EXPORT_SYMBOL(dev_loopback_xmit); 4171 4172 #ifdef CONFIG_NET_EGRESS 4173 static struct netdev_queue * 4174 netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb) 4175 { 4176 int qm = skb_get_queue_mapping(skb); 4177 4178 return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm)); 4179 } 4180 4181 #ifndef CONFIG_PREEMPT_RT 4182 static bool netdev_xmit_txqueue_skipped(void) 4183 { 4184 return __this_cpu_read(softnet_data.xmit.skip_txqueue); 4185 } 4186 4187 void netdev_xmit_skip_txqueue(bool skip) 4188 { 4189 __this_cpu_write(softnet_data.xmit.skip_txqueue, skip); 4190 } 4191 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4192 4193 #else 4194 static bool netdev_xmit_txqueue_skipped(void) 4195 { 4196 return current->net_xmit.skip_txqueue; 4197 } 4198 4199 void netdev_xmit_skip_txqueue(bool skip) 4200 { 4201 current->net_xmit.skip_txqueue = skip; 4202 } 4203 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4204 #endif 4205 #endif /* CONFIG_NET_EGRESS */ 4206 4207 #ifdef CONFIG_NET_XGRESS 4208 static int tc_run(struct tcx_entry *entry, struct sk_buff *skb, 4209 enum skb_drop_reason *drop_reason) 4210 { 4211 int ret = TC_ACT_UNSPEC; 4212 #ifdef CONFIG_NET_CLS_ACT 4213 struct mini_Qdisc *miniq = rcu_dereference_bh(entry->miniq); 4214 struct tcf_result res; 4215 4216 if (!miniq) 4217 return ret; 4218 4219 /* Global bypass */ 4220 if (!static_branch_likely(&tcf_sw_enabled_key)) 4221 return ret; 4222 4223 /* Block-wise bypass */ 4224 if (tcf_block_bypass_sw(miniq->block)) 4225 return ret; 4226 4227 tc_skb_cb(skb)->mru = 0; 4228 tc_skb_cb(skb)->post_ct = false; 4229 tcf_set_drop_reason(skb, *drop_reason); 4230 4231 mini_qdisc_bstats_cpu_update(miniq, skb); 4232 ret = tcf_classify(skb, miniq->block, miniq->filter_list, &res, false); 4233 /* Only tcf related quirks below. */ 4234 switch (ret) { 4235 case TC_ACT_SHOT: 4236 *drop_reason = tcf_get_drop_reason(skb); 4237 mini_qdisc_qstats_cpu_drop(miniq); 4238 break; 4239 case TC_ACT_OK: 4240 case TC_ACT_RECLASSIFY: 4241 skb->tc_index = TC_H_MIN(res.classid); 4242 break; 4243 } 4244 #endif /* CONFIG_NET_CLS_ACT */ 4245 return ret; 4246 } 4247 4248 static DEFINE_STATIC_KEY_FALSE(tcx_needed_key); 4249 4250 void tcx_inc(void) 4251 { 4252 static_branch_inc(&tcx_needed_key); 4253 } 4254 4255 void tcx_dec(void) 4256 { 4257 static_branch_dec(&tcx_needed_key); 4258 } 4259 4260 static __always_inline enum tcx_action_base 4261 tcx_run(const struct bpf_mprog_entry *entry, struct sk_buff *skb, 4262 const bool needs_mac) 4263 { 4264 const struct bpf_mprog_fp *fp; 4265 const struct bpf_prog *prog; 4266 int ret = TCX_NEXT; 4267 4268 if (needs_mac) 4269 __skb_push(skb, skb->mac_len); 4270 bpf_mprog_foreach_prog(entry, fp, prog) { 4271 bpf_compute_data_pointers(skb); 4272 ret = bpf_prog_run(prog, skb); 4273 if (ret != TCX_NEXT) 4274 break; 4275 } 4276 if (needs_mac) 4277 __skb_pull(skb, skb->mac_len); 4278 return tcx_action_code(skb, ret); 4279 } 4280 4281 static __always_inline struct sk_buff * 4282 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4283 struct net_device *orig_dev, bool *another) 4284 { 4285 struct bpf_mprog_entry *entry = rcu_dereference_bh(skb->dev->tcx_ingress); 4286 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_INGRESS; 4287 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4288 int sch_ret; 4289 4290 if (!entry) 4291 return skb; 4292 4293 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4294 if (*pt_prev) { 4295 *ret = deliver_skb(skb, *pt_prev, orig_dev); 4296 *pt_prev = NULL; 4297 } 4298 4299 qdisc_skb_cb(skb)->pkt_len = skb->len; 4300 tcx_set_ingress(skb, true); 4301 4302 if (static_branch_unlikely(&tcx_needed_key)) { 4303 sch_ret = tcx_run(entry, skb, true); 4304 if (sch_ret != TC_ACT_UNSPEC) 4305 goto ingress_verdict; 4306 } 4307 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4308 ingress_verdict: 4309 switch (sch_ret) { 4310 case TC_ACT_REDIRECT: 4311 /* skb_mac_header check was done by BPF, so we can safely 4312 * push the L2 header back before redirecting to another 4313 * netdev. 4314 */ 4315 __skb_push(skb, skb->mac_len); 4316 if (skb_do_redirect(skb) == -EAGAIN) { 4317 __skb_pull(skb, skb->mac_len); 4318 *another = true; 4319 break; 4320 } 4321 *ret = NET_RX_SUCCESS; 4322 bpf_net_ctx_clear(bpf_net_ctx); 4323 return NULL; 4324 case TC_ACT_SHOT: 4325 kfree_skb_reason(skb, drop_reason); 4326 *ret = NET_RX_DROP; 4327 bpf_net_ctx_clear(bpf_net_ctx); 4328 return NULL; 4329 /* used by tc_run */ 4330 case TC_ACT_STOLEN: 4331 case TC_ACT_QUEUED: 4332 case TC_ACT_TRAP: 4333 consume_skb(skb); 4334 fallthrough; 4335 case TC_ACT_CONSUMED: 4336 *ret = NET_RX_SUCCESS; 4337 bpf_net_ctx_clear(bpf_net_ctx); 4338 return NULL; 4339 } 4340 bpf_net_ctx_clear(bpf_net_ctx); 4341 4342 return skb; 4343 } 4344 4345 static __always_inline struct sk_buff * 4346 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4347 { 4348 struct bpf_mprog_entry *entry = rcu_dereference_bh(dev->tcx_egress); 4349 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_EGRESS; 4350 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4351 int sch_ret; 4352 4353 if (!entry) 4354 return skb; 4355 4356 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4357 4358 /* qdisc_skb_cb(skb)->pkt_len & tcx_set_ingress() was 4359 * already set by the caller. 4360 */ 4361 if (static_branch_unlikely(&tcx_needed_key)) { 4362 sch_ret = tcx_run(entry, skb, false); 4363 if (sch_ret != TC_ACT_UNSPEC) 4364 goto egress_verdict; 4365 } 4366 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4367 egress_verdict: 4368 switch (sch_ret) { 4369 case TC_ACT_REDIRECT: 4370 /* No need to push/pop skb's mac_header here on egress! */ 4371 skb_do_redirect(skb); 4372 *ret = NET_XMIT_SUCCESS; 4373 bpf_net_ctx_clear(bpf_net_ctx); 4374 return NULL; 4375 case TC_ACT_SHOT: 4376 kfree_skb_reason(skb, drop_reason); 4377 *ret = NET_XMIT_DROP; 4378 bpf_net_ctx_clear(bpf_net_ctx); 4379 return NULL; 4380 /* used by tc_run */ 4381 case TC_ACT_STOLEN: 4382 case TC_ACT_QUEUED: 4383 case TC_ACT_TRAP: 4384 consume_skb(skb); 4385 fallthrough; 4386 case TC_ACT_CONSUMED: 4387 *ret = NET_XMIT_SUCCESS; 4388 bpf_net_ctx_clear(bpf_net_ctx); 4389 return NULL; 4390 } 4391 bpf_net_ctx_clear(bpf_net_ctx); 4392 4393 return skb; 4394 } 4395 #else 4396 static __always_inline struct sk_buff * 4397 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4398 struct net_device *orig_dev, bool *another) 4399 { 4400 return skb; 4401 } 4402 4403 static __always_inline struct sk_buff * 4404 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4405 { 4406 return skb; 4407 } 4408 #endif /* CONFIG_NET_XGRESS */ 4409 4410 #ifdef CONFIG_XPS 4411 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb, 4412 struct xps_dev_maps *dev_maps, unsigned int tci) 4413 { 4414 int tc = netdev_get_prio_tc_map(dev, skb->priority); 4415 struct xps_map *map; 4416 int queue_index = -1; 4417 4418 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids) 4419 return queue_index; 4420 4421 tci *= dev_maps->num_tc; 4422 tci += tc; 4423 4424 map = rcu_dereference(dev_maps->attr_map[tci]); 4425 if (map) { 4426 if (map->len == 1) 4427 queue_index = map->queues[0]; 4428 else 4429 queue_index = map->queues[reciprocal_scale( 4430 skb_get_hash(skb), map->len)]; 4431 if (unlikely(queue_index >= dev->real_num_tx_queues)) 4432 queue_index = -1; 4433 } 4434 return queue_index; 4435 } 4436 #endif 4437 4438 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev, 4439 struct sk_buff *skb) 4440 { 4441 #ifdef CONFIG_XPS 4442 struct xps_dev_maps *dev_maps; 4443 struct sock *sk = skb->sk; 4444 int queue_index = -1; 4445 4446 if (!static_key_false(&xps_needed)) 4447 return -1; 4448 4449 rcu_read_lock(); 4450 if (!static_key_false(&xps_rxqs_needed)) 4451 goto get_cpus_map; 4452 4453 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]); 4454 if (dev_maps) { 4455 int tci = sk_rx_queue_get(sk); 4456 4457 if (tci >= 0) 4458 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4459 tci); 4460 } 4461 4462 get_cpus_map: 4463 if (queue_index < 0) { 4464 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]); 4465 if (dev_maps) { 4466 unsigned int tci = skb->sender_cpu - 1; 4467 4468 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4469 tci); 4470 } 4471 } 4472 rcu_read_unlock(); 4473 4474 return queue_index; 4475 #else 4476 return -1; 4477 #endif 4478 } 4479 4480 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb, 4481 struct net_device *sb_dev) 4482 { 4483 return 0; 4484 } 4485 EXPORT_SYMBOL(dev_pick_tx_zero); 4486 4487 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb, 4488 struct net_device *sb_dev) 4489 { 4490 struct sock *sk = skb->sk; 4491 int queue_index = sk_tx_queue_get(sk); 4492 4493 sb_dev = sb_dev ? : dev; 4494 4495 if (queue_index < 0 || skb->ooo_okay || 4496 queue_index >= dev->real_num_tx_queues) { 4497 int new_index = get_xps_queue(dev, sb_dev, skb); 4498 4499 if (new_index < 0) 4500 new_index = skb_tx_hash(dev, sb_dev, skb); 4501 4502 if (queue_index != new_index && sk && 4503 sk_fullsock(sk) && 4504 rcu_access_pointer(sk->sk_dst_cache)) 4505 sk_tx_queue_set(sk, new_index); 4506 4507 queue_index = new_index; 4508 } 4509 4510 return queue_index; 4511 } 4512 EXPORT_SYMBOL(netdev_pick_tx); 4513 4514 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev, 4515 struct sk_buff *skb, 4516 struct net_device *sb_dev) 4517 { 4518 int queue_index = 0; 4519 4520 #ifdef CONFIG_XPS 4521 u32 sender_cpu = skb->sender_cpu - 1; 4522 4523 if (sender_cpu >= (u32)NR_CPUS) 4524 skb->sender_cpu = raw_smp_processor_id() + 1; 4525 #endif 4526 4527 if (dev->real_num_tx_queues != 1) { 4528 const struct net_device_ops *ops = dev->netdev_ops; 4529 4530 if (ops->ndo_select_queue) 4531 queue_index = ops->ndo_select_queue(dev, skb, sb_dev); 4532 else 4533 queue_index = netdev_pick_tx(dev, skb, sb_dev); 4534 4535 queue_index = netdev_cap_txqueue(dev, queue_index); 4536 } 4537 4538 skb_set_queue_mapping(skb, queue_index); 4539 return netdev_get_tx_queue(dev, queue_index); 4540 } 4541 4542 /** 4543 * __dev_queue_xmit() - transmit a buffer 4544 * @skb: buffer to transmit 4545 * @sb_dev: suboordinate device used for L2 forwarding offload 4546 * 4547 * Queue a buffer for transmission to a network device. The caller must 4548 * have set the device and priority and built the buffer before calling 4549 * this function. The function can be called from an interrupt. 4550 * 4551 * When calling this method, interrupts MUST be enabled. This is because 4552 * the BH enable code must have IRQs enabled so that it will not deadlock. 4553 * 4554 * Regardless of the return value, the skb is consumed, so it is currently 4555 * difficult to retry a send to this method. (You can bump the ref count 4556 * before sending to hold a reference for retry if you are careful.) 4557 * 4558 * Return: 4559 * * 0 - buffer successfully transmitted 4560 * * positive qdisc return code - NET_XMIT_DROP etc. 4561 * * negative errno - other errors 4562 */ 4563 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev) 4564 { 4565 struct net_device *dev = skb->dev; 4566 struct netdev_queue *txq = NULL; 4567 struct Qdisc *q; 4568 int rc = -ENOMEM; 4569 bool again = false; 4570 4571 skb_reset_mac_header(skb); 4572 skb_assert_len(skb); 4573 4574 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP)) 4575 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED); 4576 4577 /* Disable soft irqs for various locks below. Also 4578 * stops preemption for RCU. 4579 */ 4580 rcu_read_lock_bh(); 4581 4582 skb_update_prio(skb); 4583 4584 qdisc_pkt_len_init(skb); 4585 tcx_set_ingress(skb, false); 4586 #ifdef CONFIG_NET_EGRESS 4587 if (static_branch_unlikely(&egress_needed_key)) { 4588 if (nf_hook_egress_active()) { 4589 skb = nf_hook_egress(skb, &rc, dev); 4590 if (!skb) 4591 goto out; 4592 } 4593 4594 netdev_xmit_skip_txqueue(false); 4595 4596 nf_skip_egress(skb, true); 4597 skb = sch_handle_egress(skb, &rc, dev); 4598 if (!skb) 4599 goto out; 4600 nf_skip_egress(skb, false); 4601 4602 if (netdev_xmit_txqueue_skipped()) 4603 txq = netdev_tx_queue_mapping(dev, skb); 4604 } 4605 #endif 4606 /* If device/qdisc don't need skb->dst, release it right now while 4607 * its hot in this cpu cache. 4608 */ 4609 if (dev->priv_flags & IFF_XMIT_DST_RELEASE) 4610 skb_dst_drop(skb); 4611 else 4612 skb_dst_force(skb); 4613 4614 if (!txq) 4615 txq = netdev_core_pick_tx(dev, skb, sb_dev); 4616 4617 q = rcu_dereference_bh(txq->qdisc); 4618 4619 trace_net_dev_queue(skb); 4620 if (q->enqueue) { 4621 rc = __dev_xmit_skb(skb, q, dev, txq); 4622 goto out; 4623 } 4624 4625 /* The device has no queue. Common case for software devices: 4626 * loopback, all the sorts of tunnels... 4627 4628 * Really, it is unlikely that netif_tx_lock protection is necessary 4629 * here. (f.e. loopback and IP tunnels are clean ignoring statistics 4630 * counters.) 4631 * However, it is possible, that they rely on protection 4632 * made by us here. 4633 4634 * Check this and shot the lock. It is not prone from deadlocks. 4635 *Either shot noqueue qdisc, it is even simpler 8) 4636 */ 4637 if (dev->flags & IFF_UP) { 4638 int cpu = smp_processor_id(); /* ok because BHs are off */ 4639 4640 /* Other cpus might concurrently change txq->xmit_lock_owner 4641 * to -1 or to their cpu id, but not to our id. 4642 */ 4643 if (READ_ONCE(txq->xmit_lock_owner) != cpu) { 4644 if (dev_xmit_recursion()) 4645 goto recursion_alert; 4646 4647 skb = validate_xmit_skb(skb, dev, &again); 4648 if (!skb) 4649 goto out; 4650 4651 HARD_TX_LOCK(dev, txq, cpu); 4652 4653 if (!netif_xmit_stopped(txq)) { 4654 dev_xmit_recursion_inc(); 4655 skb = dev_hard_start_xmit(skb, dev, txq, &rc); 4656 dev_xmit_recursion_dec(); 4657 if (dev_xmit_complete(rc)) { 4658 HARD_TX_UNLOCK(dev, txq); 4659 goto out; 4660 } 4661 } 4662 HARD_TX_UNLOCK(dev, txq); 4663 net_crit_ratelimited("Virtual device %s asks to queue packet!\n", 4664 dev->name); 4665 } else { 4666 /* Recursion is detected! It is possible, 4667 * unfortunately 4668 */ 4669 recursion_alert: 4670 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n", 4671 dev->name); 4672 } 4673 } 4674 4675 rc = -ENETDOWN; 4676 rcu_read_unlock_bh(); 4677 4678 dev_core_stats_tx_dropped_inc(dev); 4679 kfree_skb_list(skb); 4680 return rc; 4681 out: 4682 rcu_read_unlock_bh(); 4683 return rc; 4684 } 4685 EXPORT_SYMBOL(__dev_queue_xmit); 4686 4687 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id) 4688 { 4689 struct net_device *dev = skb->dev; 4690 struct sk_buff *orig_skb = skb; 4691 struct netdev_queue *txq; 4692 int ret = NETDEV_TX_BUSY; 4693 bool again = false; 4694 4695 if (unlikely(!netif_running(dev) || 4696 !netif_carrier_ok(dev))) 4697 goto drop; 4698 4699 skb = validate_xmit_skb_list(skb, dev, &again); 4700 if (skb != orig_skb) 4701 goto drop; 4702 4703 skb_set_queue_mapping(skb, queue_id); 4704 txq = skb_get_tx_queue(dev, skb); 4705 4706 local_bh_disable(); 4707 4708 dev_xmit_recursion_inc(); 4709 HARD_TX_LOCK(dev, txq, smp_processor_id()); 4710 if (!netif_xmit_frozen_or_drv_stopped(txq)) 4711 ret = netdev_start_xmit(skb, dev, txq, false); 4712 HARD_TX_UNLOCK(dev, txq); 4713 dev_xmit_recursion_dec(); 4714 4715 local_bh_enable(); 4716 return ret; 4717 drop: 4718 dev_core_stats_tx_dropped_inc(dev); 4719 kfree_skb_list(skb); 4720 return NET_XMIT_DROP; 4721 } 4722 EXPORT_SYMBOL(__dev_direct_xmit); 4723 4724 /************************************************************************* 4725 * Receiver routines 4726 *************************************************************************/ 4727 static DEFINE_PER_CPU(struct task_struct *, backlog_napi); 4728 4729 int weight_p __read_mostly = 64; /* old backlog weight */ 4730 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */ 4731 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */ 4732 4733 /* Called with irq disabled */ 4734 static inline void ____napi_schedule(struct softnet_data *sd, 4735 struct napi_struct *napi) 4736 { 4737 struct task_struct *thread; 4738 4739 lockdep_assert_irqs_disabled(); 4740 4741 if (test_bit(NAPI_STATE_THREADED, &napi->state)) { 4742 /* Paired with smp_mb__before_atomic() in 4743 * napi_enable()/dev_set_threaded(). 4744 * Use READ_ONCE() to guarantee a complete 4745 * read on napi->thread. Only call 4746 * wake_up_process() when it's not NULL. 4747 */ 4748 thread = READ_ONCE(napi->thread); 4749 if (thread) { 4750 if (use_backlog_threads() && thread == raw_cpu_read(backlog_napi)) 4751 goto use_local_napi; 4752 4753 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 4754 wake_up_process(thread); 4755 return; 4756 } 4757 } 4758 4759 use_local_napi: 4760 list_add_tail(&napi->poll_list, &sd->poll_list); 4761 WRITE_ONCE(napi->list_owner, smp_processor_id()); 4762 /* If not called from net_rx_action() 4763 * we have to raise NET_RX_SOFTIRQ. 4764 */ 4765 if (!sd->in_net_rx_action) 4766 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4767 } 4768 4769 #ifdef CONFIG_RPS 4770 4771 struct static_key_false rps_needed __read_mostly; 4772 EXPORT_SYMBOL(rps_needed); 4773 struct static_key_false rfs_needed __read_mostly; 4774 EXPORT_SYMBOL(rfs_needed); 4775 4776 static struct rps_dev_flow * 4777 set_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4778 struct rps_dev_flow *rflow, u16 next_cpu) 4779 { 4780 if (next_cpu < nr_cpu_ids) { 4781 u32 head; 4782 #ifdef CONFIG_RFS_ACCEL 4783 struct netdev_rx_queue *rxqueue; 4784 struct rps_dev_flow_table *flow_table; 4785 struct rps_dev_flow *old_rflow; 4786 u16 rxq_index; 4787 u32 flow_id; 4788 int rc; 4789 4790 /* Should we steer this flow to a different hardware queue? */ 4791 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap || 4792 !(dev->features & NETIF_F_NTUPLE)) 4793 goto out; 4794 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu); 4795 if (rxq_index == skb_get_rx_queue(skb)) 4796 goto out; 4797 4798 rxqueue = dev->_rx + rxq_index; 4799 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4800 if (!flow_table) 4801 goto out; 4802 flow_id = skb_get_hash(skb) & flow_table->mask; 4803 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb, 4804 rxq_index, flow_id); 4805 if (rc < 0) 4806 goto out; 4807 old_rflow = rflow; 4808 rflow = &flow_table->flows[flow_id]; 4809 WRITE_ONCE(rflow->filter, rc); 4810 if (old_rflow->filter == rc) 4811 WRITE_ONCE(old_rflow->filter, RPS_NO_FILTER); 4812 out: 4813 #endif 4814 head = READ_ONCE(per_cpu(softnet_data, next_cpu).input_queue_head); 4815 rps_input_queue_tail_save(&rflow->last_qtail, head); 4816 } 4817 4818 WRITE_ONCE(rflow->cpu, next_cpu); 4819 return rflow; 4820 } 4821 4822 /* 4823 * get_rps_cpu is called from netif_receive_skb and returns the target 4824 * CPU from the RPS map of the receiving queue for a given skb. 4825 * rcu_read_lock must be held on entry. 4826 */ 4827 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4828 struct rps_dev_flow **rflowp) 4829 { 4830 const struct rps_sock_flow_table *sock_flow_table; 4831 struct netdev_rx_queue *rxqueue = dev->_rx; 4832 struct rps_dev_flow_table *flow_table; 4833 struct rps_map *map; 4834 int cpu = -1; 4835 u32 tcpu; 4836 u32 hash; 4837 4838 if (skb_rx_queue_recorded(skb)) { 4839 u16 index = skb_get_rx_queue(skb); 4840 4841 if (unlikely(index >= dev->real_num_rx_queues)) { 4842 WARN_ONCE(dev->real_num_rx_queues > 1, 4843 "%s received packet on queue %u, but number " 4844 "of RX queues is %u\n", 4845 dev->name, index, dev->real_num_rx_queues); 4846 goto done; 4847 } 4848 rxqueue += index; 4849 } 4850 4851 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */ 4852 4853 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4854 map = rcu_dereference(rxqueue->rps_map); 4855 if (!flow_table && !map) 4856 goto done; 4857 4858 skb_reset_network_header(skb); 4859 hash = skb_get_hash(skb); 4860 if (!hash) 4861 goto done; 4862 4863 sock_flow_table = rcu_dereference(net_hotdata.rps_sock_flow_table); 4864 if (flow_table && sock_flow_table) { 4865 struct rps_dev_flow *rflow; 4866 u32 next_cpu; 4867 u32 ident; 4868 4869 /* First check into global flow table if there is a match. 4870 * This READ_ONCE() pairs with WRITE_ONCE() from rps_record_sock_flow(). 4871 */ 4872 ident = READ_ONCE(sock_flow_table->ents[hash & sock_flow_table->mask]); 4873 if ((ident ^ hash) & ~net_hotdata.rps_cpu_mask) 4874 goto try_rps; 4875 4876 next_cpu = ident & net_hotdata.rps_cpu_mask; 4877 4878 /* OK, now we know there is a match, 4879 * we can look at the local (per receive queue) flow table 4880 */ 4881 rflow = &flow_table->flows[hash & flow_table->mask]; 4882 tcpu = rflow->cpu; 4883 4884 /* 4885 * If the desired CPU (where last recvmsg was done) is 4886 * different from current CPU (one in the rx-queue flow 4887 * table entry), switch if one of the following holds: 4888 * - Current CPU is unset (>= nr_cpu_ids). 4889 * - Current CPU is offline. 4890 * - The current CPU's queue tail has advanced beyond the 4891 * last packet that was enqueued using this table entry. 4892 * This guarantees that all previous packets for the flow 4893 * have been dequeued, thus preserving in order delivery. 4894 */ 4895 if (unlikely(tcpu != next_cpu) && 4896 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) || 4897 ((int)(READ_ONCE(per_cpu(softnet_data, tcpu).input_queue_head) - 4898 rflow->last_qtail)) >= 0)) { 4899 tcpu = next_cpu; 4900 rflow = set_rps_cpu(dev, skb, rflow, next_cpu); 4901 } 4902 4903 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) { 4904 *rflowp = rflow; 4905 cpu = tcpu; 4906 goto done; 4907 } 4908 } 4909 4910 try_rps: 4911 4912 if (map) { 4913 tcpu = map->cpus[reciprocal_scale(hash, map->len)]; 4914 if (cpu_online(tcpu)) { 4915 cpu = tcpu; 4916 goto done; 4917 } 4918 } 4919 4920 done: 4921 return cpu; 4922 } 4923 4924 #ifdef CONFIG_RFS_ACCEL 4925 4926 /** 4927 * rps_may_expire_flow - check whether an RFS hardware filter may be removed 4928 * @dev: Device on which the filter was set 4929 * @rxq_index: RX queue index 4930 * @flow_id: Flow ID passed to ndo_rx_flow_steer() 4931 * @filter_id: Filter ID returned by ndo_rx_flow_steer() 4932 * 4933 * Drivers that implement ndo_rx_flow_steer() should periodically call 4934 * this function for each installed filter and remove the filters for 4935 * which it returns %true. 4936 */ 4937 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index, 4938 u32 flow_id, u16 filter_id) 4939 { 4940 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index; 4941 struct rps_dev_flow_table *flow_table; 4942 struct rps_dev_flow *rflow; 4943 bool expire = true; 4944 unsigned int cpu; 4945 4946 rcu_read_lock(); 4947 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4948 if (flow_table && flow_id <= flow_table->mask) { 4949 rflow = &flow_table->flows[flow_id]; 4950 cpu = READ_ONCE(rflow->cpu); 4951 if (READ_ONCE(rflow->filter) == filter_id && cpu < nr_cpu_ids && 4952 ((int)(READ_ONCE(per_cpu(softnet_data, cpu).input_queue_head) - 4953 READ_ONCE(rflow->last_qtail)) < 4954 (int)(10 * flow_table->mask))) 4955 expire = false; 4956 } 4957 rcu_read_unlock(); 4958 return expire; 4959 } 4960 EXPORT_SYMBOL(rps_may_expire_flow); 4961 4962 #endif /* CONFIG_RFS_ACCEL */ 4963 4964 /* Called from hardirq (IPI) context */ 4965 static void rps_trigger_softirq(void *data) 4966 { 4967 struct softnet_data *sd = data; 4968 4969 ____napi_schedule(sd, &sd->backlog); 4970 sd->received_rps++; 4971 } 4972 4973 #endif /* CONFIG_RPS */ 4974 4975 /* Called from hardirq (IPI) context */ 4976 static void trigger_rx_softirq(void *data) 4977 { 4978 struct softnet_data *sd = data; 4979 4980 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4981 smp_store_release(&sd->defer_ipi_scheduled, 0); 4982 } 4983 4984 /* 4985 * After we queued a packet into sd->input_pkt_queue, 4986 * we need to make sure this queue is serviced soon. 4987 * 4988 * - If this is another cpu queue, link it to our rps_ipi_list, 4989 * and make sure we will process rps_ipi_list from net_rx_action(). 4990 * 4991 * - If this is our own queue, NAPI schedule our backlog. 4992 * Note that this also raises NET_RX_SOFTIRQ. 4993 */ 4994 static void napi_schedule_rps(struct softnet_data *sd) 4995 { 4996 struct softnet_data *mysd = this_cpu_ptr(&softnet_data); 4997 4998 #ifdef CONFIG_RPS 4999 if (sd != mysd) { 5000 if (use_backlog_threads()) { 5001 __napi_schedule_irqoff(&sd->backlog); 5002 return; 5003 } 5004 5005 sd->rps_ipi_next = mysd->rps_ipi_list; 5006 mysd->rps_ipi_list = sd; 5007 5008 /* If not called from net_rx_action() or napi_threaded_poll() 5009 * we have to raise NET_RX_SOFTIRQ. 5010 */ 5011 if (!mysd->in_net_rx_action && !mysd->in_napi_threaded_poll) 5012 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 5013 return; 5014 } 5015 #endif /* CONFIG_RPS */ 5016 __napi_schedule_irqoff(&mysd->backlog); 5017 } 5018 5019 void kick_defer_list_purge(struct softnet_data *sd, unsigned int cpu) 5020 { 5021 unsigned long flags; 5022 5023 if (use_backlog_threads()) { 5024 backlog_lock_irq_save(sd, &flags); 5025 5026 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) 5027 __napi_schedule_irqoff(&sd->backlog); 5028 5029 backlog_unlock_irq_restore(sd, &flags); 5030 5031 } else if (!cmpxchg(&sd->defer_ipi_scheduled, 0, 1)) { 5032 smp_call_function_single_async(cpu, &sd->defer_csd); 5033 } 5034 } 5035 5036 #ifdef CONFIG_NET_FLOW_LIMIT 5037 int netdev_flow_limit_table_len __read_mostly = (1 << 12); 5038 #endif 5039 5040 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen) 5041 { 5042 #ifdef CONFIG_NET_FLOW_LIMIT 5043 struct sd_flow_limit *fl; 5044 struct softnet_data *sd; 5045 unsigned int old_flow, new_flow; 5046 5047 if (qlen < (READ_ONCE(net_hotdata.max_backlog) >> 1)) 5048 return false; 5049 5050 sd = this_cpu_ptr(&softnet_data); 5051 5052 rcu_read_lock(); 5053 fl = rcu_dereference(sd->flow_limit); 5054 if (fl) { 5055 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1); 5056 old_flow = fl->history[fl->history_head]; 5057 fl->history[fl->history_head] = new_flow; 5058 5059 fl->history_head++; 5060 fl->history_head &= FLOW_LIMIT_HISTORY - 1; 5061 5062 if (likely(fl->buckets[old_flow])) 5063 fl->buckets[old_flow]--; 5064 5065 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) { 5066 fl->count++; 5067 rcu_read_unlock(); 5068 return true; 5069 } 5070 } 5071 rcu_read_unlock(); 5072 #endif 5073 return false; 5074 } 5075 5076 /* 5077 * enqueue_to_backlog is called to queue an skb to a per CPU backlog 5078 * queue (may be a remote CPU queue). 5079 */ 5080 static int enqueue_to_backlog(struct sk_buff *skb, int cpu, 5081 unsigned int *qtail) 5082 { 5083 enum skb_drop_reason reason; 5084 struct softnet_data *sd; 5085 unsigned long flags; 5086 unsigned int qlen; 5087 int max_backlog; 5088 u32 tail; 5089 5090 reason = SKB_DROP_REASON_DEV_READY; 5091 if (!netif_running(skb->dev)) 5092 goto bad_dev; 5093 5094 reason = SKB_DROP_REASON_CPU_BACKLOG; 5095 sd = &per_cpu(softnet_data, cpu); 5096 5097 qlen = skb_queue_len_lockless(&sd->input_pkt_queue); 5098 max_backlog = READ_ONCE(net_hotdata.max_backlog); 5099 if (unlikely(qlen > max_backlog)) 5100 goto cpu_backlog_drop; 5101 backlog_lock_irq_save(sd, &flags); 5102 qlen = skb_queue_len(&sd->input_pkt_queue); 5103 if (qlen <= max_backlog && !skb_flow_limit(skb, qlen)) { 5104 if (!qlen) { 5105 /* Schedule NAPI for backlog device. We can use 5106 * non atomic operation as we own the queue lock. 5107 */ 5108 if (!__test_and_set_bit(NAPI_STATE_SCHED, 5109 &sd->backlog.state)) 5110 napi_schedule_rps(sd); 5111 } 5112 __skb_queue_tail(&sd->input_pkt_queue, skb); 5113 tail = rps_input_queue_tail_incr(sd); 5114 backlog_unlock_irq_restore(sd, &flags); 5115 5116 /* save the tail outside of the critical section */ 5117 rps_input_queue_tail_save(qtail, tail); 5118 return NET_RX_SUCCESS; 5119 } 5120 5121 backlog_unlock_irq_restore(sd, &flags); 5122 5123 cpu_backlog_drop: 5124 atomic_inc(&sd->dropped); 5125 bad_dev: 5126 dev_core_stats_rx_dropped_inc(skb->dev); 5127 kfree_skb_reason(skb, reason); 5128 return NET_RX_DROP; 5129 } 5130 5131 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb) 5132 { 5133 struct net_device *dev = skb->dev; 5134 struct netdev_rx_queue *rxqueue; 5135 5136 rxqueue = dev->_rx; 5137 5138 if (skb_rx_queue_recorded(skb)) { 5139 u16 index = skb_get_rx_queue(skb); 5140 5141 if (unlikely(index >= dev->real_num_rx_queues)) { 5142 WARN_ONCE(dev->real_num_rx_queues > 1, 5143 "%s received packet on queue %u, but number " 5144 "of RX queues is %u\n", 5145 dev->name, index, dev->real_num_rx_queues); 5146 5147 return rxqueue; /* Return first rxqueue */ 5148 } 5149 rxqueue += index; 5150 } 5151 return rxqueue; 5152 } 5153 5154 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, 5155 const struct bpf_prog *xdp_prog) 5156 { 5157 void *orig_data, *orig_data_end, *hard_start; 5158 struct netdev_rx_queue *rxqueue; 5159 bool orig_bcast, orig_host; 5160 u32 mac_len, frame_sz; 5161 __be16 orig_eth_type; 5162 struct ethhdr *eth; 5163 u32 metalen, act; 5164 int off; 5165 5166 /* The XDP program wants to see the packet starting at the MAC 5167 * header. 5168 */ 5169 mac_len = skb->data - skb_mac_header(skb); 5170 hard_start = skb->data - skb_headroom(skb); 5171 5172 /* SKB "head" area always have tailroom for skb_shared_info */ 5173 frame_sz = (void *)skb_end_pointer(skb) - hard_start; 5174 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 5175 5176 rxqueue = netif_get_rxqueue(skb); 5177 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq); 5178 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len, 5179 skb_headlen(skb) + mac_len, true); 5180 if (skb_is_nonlinear(skb)) { 5181 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 5182 xdp_buff_set_frags_flag(xdp); 5183 } else { 5184 xdp_buff_clear_frags_flag(xdp); 5185 } 5186 5187 orig_data_end = xdp->data_end; 5188 orig_data = xdp->data; 5189 eth = (struct ethhdr *)xdp->data; 5190 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr); 5191 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest); 5192 orig_eth_type = eth->h_proto; 5193 5194 act = bpf_prog_run_xdp(xdp_prog, xdp); 5195 5196 /* check if bpf_xdp_adjust_head was used */ 5197 off = xdp->data - orig_data; 5198 if (off) { 5199 if (off > 0) 5200 __skb_pull(skb, off); 5201 else if (off < 0) 5202 __skb_push(skb, -off); 5203 5204 skb->mac_header += off; 5205 skb_reset_network_header(skb); 5206 } 5207 5208 /* check if bpf_xdp_adjust_tail was used */ 5209 off = xdp->data_end - orig_data_end; 5210 if (off != 0) { 5211 skb_set_tail_pointer(skb, xdp->data_end - xdp->data); 5212 skb->len += off; /* positive on grow, negative on shrink */ 5213 } 5214 5215 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 5216 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here. 5217 */ 5218 if (xdp_buff_has_frags(xdp)) 5219 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 5220 else 5221 skb->data_len = 0; 5222 5223 /* check if XDP changed eth hdr such SKB needs update */ 5224 eth = (struct ethhdr *)xdp->data; 5225 if ((orig_eth_type != eth->h_proto) || 5226 (orig_host != ether_addr_equal_64bits(eth->h_dest, 5227 skb->dev->dev_addr)) || 5228 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) { 5229 __skb_push(skb, ETH_HLEN); 5230 skb->pkt_type = PACKET_HOST; 5231 skb->protocol = eth_type_trans(skb, skb->dev); 5232 } 5233 5234 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull 5235 * before calling us again on redirect path. We do not call do_redirect 5236 * as we leave that up to the caller. 5237 * 5238 * Caller is responsible for managing lifetime of skb (i.e. calling 5239 * kfree_skb in response to actions it cannot handle/XDP_DROP). 5240 */ 5241 switch (act) { 5242 case XDP_REDIRECT: 5243 case XDP_TX: 5244 __skb_push(skb, mac_len); 5245 break; 5246 case XDP_PASS: 5247 metalen = xdp->data - xdp->data_meta; 5248 if (metalen) 5249 skb_metadata_set(skb, metalen); 5250 break; 5251 } 5252 5253 return act; 5254 } 5255 5256 static int 5257 netif_skb_check_for_xdp(struct sk_buff **pskb, const struct bpf_prog *prog) 5258 { 5259 struct sk_buff *skb = *pskb; 5260 int err, hroom, troom; 5261 5262 if (!skb_cow_data_for_xdp(this_cpu_read(system_page_pool), pskb, prog)) 5263 return 0; 5264 5265 /* In case we have to go down the path and also linearize, 5266 * then lets do the pskb_expand_head() work just once here. 5267 */ 5268 hroom = XDP_PACKET_HEADROOM - skb_headroom(skb); 5269 troom = skb->tail + skb->data_len - skb->end; 5270 err = pskb_expand_head(skb, 5271 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0, 5272 troom > 0 ? troom + 128 : 0, GFP_ATOMIC); 5273 if (err) 5274 return err; 5275 5276 return skb_linearize(skb); 5277 } 5278 5279 static u32 netif_receive_generic_xdp(struct sk_buff **pskb, 5280 struct xdp_buff *xdp, 5281 const struct bpf_prog *xdp_prog) 5282 { 5283 struct sk_buff *skb = *pskb; 5284 u32 mac_len, act = XDP_DROP; 5285 5286 /* Reinjected packets coming from act_mirred or similar should 5287 * not get XDP generic processing. 5288 */ 5289 if (skb_is_redirected(skb)) 5290 return XDP_PASS; 5291 5292 /* XDP packets must have sufficient headroom of XDP_PACKET_HEADROOM 5293 * bytes. This is the guarantee that also native XDP provides, 5294 * thus we need to do it here as well. 5295 */ 5296 mac_len = skb->data - skb_mac_header(skb); 5297 __skb_push(skb, mac_len); 5298 5299 if (skb_cloned(skb) || skb_is_nonlinear(skb) || 5300 skb_headroom(skb) < XDP_PACKET_HEADROOM) { 5301 if (netif_skb_check_for_xdp(pskb, xdp_prog)) 5302 goto do_drop; 5303 } 5304 5305 __skb_pull(*pskb, mac_len); 5306 5307 act = bpf_prog_run_generic_xdp(*pskb, xdp, xdp_prog); 5308 switch (act) { 5309 case XDP_REDIRECT: 5310 case XDP_TX: 5311 case XDP_PASS: 5312 break; 5313 default: 5314 bpf_warn_invalid_xdp_action((*pskb)->dev, xdp_prog, act); 5315 fallthrough; 5316 case XDP_ABORTED: 5317 trace_xdp_exception((*pskb)->dev, xdp_prog, act); 5318 fallthrough; 5319 case XDP_DROP: 5320 do_drop: 5321 kfree_skb(*pskb); 5322 break; 5323 } 5324 5325 return act; 5326 } 5327 5328 /* When doing generic XDP we have to bypass the qdisc layer and the 5329 * network taps in order to match in-driver-XDP behavior. This also means 5330 * that XDP packets are able to starve other packets going through a qdisc, 5331 * and DDOS attacks will be more effective. In-driver-XDP use dedicated TX 5332 * queues, so they do not have this starvation issue. 5333 */ 5334 void generic_xdp_tx(struct sk_buff *skb, const struct bpf_prog *xdp_prog) 5335 { 5336 struct net_device *dev = skb->dev; 5337 struct netdev_queue *txq; 5338 bool free_skb = true; 5339 int cpu, rc; 5340 5341 txq = netdev_core_pick_tx(dev, skb, NULL); 5342 cpu = smp_processor_id(); 5343 HARD_TX_LOCK(dev, txq, cpu); 5344 if (!netif_xmit_frozen_or_drv_stopped(txq)) { 5345 rc = netdev_start_xmit(skb, dev, txq, 0); 5346 if (dev_xmit_complete(rc)) 5347 free_skb = false; 5348 } 5349 HARD_TX_UNLOCK(dev, txq); 5350 if (free_skb) { 5351 trace_xdp_exception(dev, xdp_prog, XDP_TX); 5352 dev_core_stats_tx_dropped_inc(dev); 5353 kfree_skb(skb); 5354 } 5355 } 5356 5357 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key); 5358 5359 int do_xdp_generic(const struct bpf_prog *xdp_prog, struct sk_buff **pskb) 5360 { 5361 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 5362 5363 if (xdp_prog) { 5364 struct xdp_buff xdp; 5365 u32 act; 5366 int err; 5367 5368 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 5369 act = netif_receive_generic_xdp(pskb, &xdp, xdp_prog); 5370 if (act != XDP_PASS) { 5371 switch (act) { 5372 case XDP_REDIRECT: 5373 err = xdp_do_generic_redirect((*pskb)->dev, *pskb, 5374 &xdp, xdp_prog); 5375 if (err) 5376 goto out_redir; 5377 break; 5378 case XDP_TX: 5379 generic_xdp_tx(*pskb, xdp_prog); 5380 break; 5381 } 5382 bpf_net_ctx_clear(bpf_net_ctx); 5383 return XDP_DROP; 5384 } 5385 bpf_net_ctx_clear(bpf_net_ctx); 5386 } 5387 return XDP_PASS; 5388 out_redir: 5389 bpf_net_ctx_clear(bpf_net_ctx); 5390 kfree_skb_reason(*pskb, SKB_DROP_REASON_XDP); 5391 return XDP_DROP; 5392 } 5393 EXPORT_SYMBOL_GPL(do_xdp_generic); 5394 5395 static int netif_rx_internal(struct sk_buff *skb) 5396 { 5397 int ret; 5398 5399 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5400 5401 trace_netif_rx(skb); 5402 5403 #ifdef CONFIG_RPS 5404 if (static_branch_unlikely(&rps_needed)) { 5405 struct rps_dev_flow voidflow, *rflow = &voidflow; 5406 int cpu; 5407 5408 rcu_read_lock(); 5409 5410 cpu = get_rps_cpu(skb->dev, skb, &rflow); 5411 if (cpu < 0) 5412 cpu = smp_processor_id(); 5413 5414 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5415 5416 rcu_read_unlock(); 5417 } else 5418 #endif 5419 { 5420 unsigned int qtail; 5421 5422 ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail); 5423 } 5424 return ret; 5425 } 5426 5427 /** 5428 * __netif_rx - Slightly optimized version of netif_rx 5429 * @skb: buffer to post 5430 * 5431 * This behaves as netif_rx except that it does not disable bottom halves. 5432 * As a result this function may only be invoked from the interrupt context 5433 * (either hard or soft interrupt). 5434 */ 5435 int __netif_rx(struct sk_buff *skb) 5436 { 5437 int ret; 5438 5439 lockdep_assert_once(hardirq_count() | softirq_count()); 5440 5441 trace_netif_rx_entry(skb); 5442 ret = netif_rx_internal(skb); 5443 trace_netif_rx_exit(ret); 5444 return ret; 5445 } 5446 EXPORT_SYMBOL(__netif_rx); 5447 5448 /** 5449 * netif_rx - post buffer to the network code 5450 * @skb: buffer to post 5451 * 5452 * This function receives a packet from a device driver and queues it for 5453 * the upper (protocol) levels to process via the backlog NAPI device. It 5454 * always succeeds. The buffer may be dropped during processing for 5455 * congestion control or by the protocol layers. 5456 * The network buffer is passed via the backlog NAPI device. Modern NIC 5457 * driver should use NAPI and GRO. 5458 * This function can used from interrupt and from process context. The 5459 * caller from process context must not disable interrupts before invoking 5460 * this function. 5461 * 5462 * return values: 5463 * NET_RX_SUCCESS (no congestion) 5464 * NET_RX_DROP (packet was dropped) 5465 * 5466 */ 5467 int netif_rx(struct sk_buff *skb) 5468 { 5469 bool need_bh_off = !(hardirq_count() | softirq_count()); 5470 int ret; 5471 5472 if (need_bh_off) 5473 local_bh_disable(); 5474 trace_netif_rx_entry(skb); 5475 ret = netif_rx_internal(skb); 5476 trace_netif_rx_exit(ret); 5477 if (need_bh_off) 5478 local_bh_enable(); 5479 return ret; 5480 } 5481 EXPORT_SYMBOL(netif_rx); 5482 5483 static __latent_entropy void net_tx_action(void) 5484 { 5485 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 5486 5487 if (sd->completion_queue) { 5488 struct sk_buff *clist; 5489 5490 local_irq_disable(); 5491 clist = sd->completion_queue; 5492 sd->completion_queue = NULL; 5493 local_irq_enable(); 5494 5495 while (clist) { 5496 struct sk_buff *skb = clist; 5497 5498 clist = clist->next; 5499 5500 WARN_ON(refcount_read(&skb->users)); 5501 if (likely(get_kfree_skb_cb(skb)->reason == SKB_CONSUMED)) 5502 trace_consume_skb(skb, net_tx_action); 5503 else 5504 trace_kfree_skb(skb, net_tx_action, 5505 get_kfree_skb_cb(skb)->reason, NULL); 5506 5507 if (skb->fclone != SKB_FCLONE_UNAVAILABLE) 5508 __kfree_skb(skb); 5509 else 5510 __napi_kfree_skb(skb, 5511 get_kfree_skb_cb(skb)->reason); 5512 } 5513 } 5514 5515 if (sd->output_queue) { 5516 struct Qdisc *head; 5517 5518 local_irq_disable(); 5519 head = sd->output_queue; 5520 sd->output_queue = NULL; 5521 sd->output_queue_tailp = &sd->output_queue; 5522 local_irq_enable(); 5523 5524 rcu_read_lock(); 5525 5526 while (head) { 5527 struct Qdisc *q = head; 5528 spinlock_t *root_lock = NULL; 5529 5530 head = head->next_sched; 5531 5532 /* We need to make sure head->next_sched is read 5533 * before clearing __QDISC_STATE_SCHED 5534 */ 5535 smp_mb__before_atomic(); 5536 5537 if (!(q->flags & TCQ_F_NOLOCK)) { 5538 root_lock = qdisc_lock(q); 5539 spin_lock(root_lock); 5540 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, 5541 &q->state))) { 5542 /* There is a synchronize_net() between 5543 * STATE_DEACTIVATED flag being set and 5544 * qdisc_reset()/some_qdisc_is_busy() in 5545 * dev_deactivate(), so we can safely bail out 5546 * early here to avoid data race between 5547 * qdisc_deactivate() and some_qdisc_is_busy() 5548 * for lockless qdisc. 5549 */ 5550 clear_bit(__QDISC_STATE_SCHED, &q->state); 5551 continue; 5552 } 5553 5554 clear_bit(__QDISC_STATE_SCHED, &q->state); 5555 qdisc_run(q); 5556 if (root_lock) 5557 spin_unlock(root_lock); 5558 } 5559 5560 rcu_read_unlock(); 5561 } 5562 5563 xfrm_dev_backlog(sd); 5564 } 5565 5566 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE) 5567 /* This hook is defined here for ATM LANE */ 5568 int (*br_fdb_test_addr_hook)(struct net_device *dev, 5569 unsigned char *addr) __read_mostly; 5570 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook); 5571 #endif 5572 5573 /** 5574 * netdev_is_rx_handler_busy - check if receive handler is registered 5575 * @dev: device to check 5576 * 5577 * Check if a receive handler is already registered for a given device. 5578 * Return true if there one. 5579 * 5580 * The caller must hold the rtnl_mutex. 5581 */ 5582 bool netdev_is_rx_handler_busy(struct net_device *dev) 5583 { 5584 ASSERT_RTNL(); 5585 return dev && rtnl_dereference(dev->rx_handler); 5586 } 5587 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy); 5588 5589 /** 5590 * netdev_rx_handler_register - register receive handler 5591 * @dev: device to register a handler for 5592 * @rx_handler: receive handler to register 5593 * @rx_handler_data: data pointer that is used by rx handler 5594 * 5595 * Register a receive handler for a device. This handler will then be 5596 * called from __netif_receive_skb. A negative errno code is returned 5597 * on a failure. 5598 * 5599 * The caller must hold the rtnl_mutex. 5600 * 5601 * For a general description of rx_handler, see enum rx_handler_result. 5602 */ 5603 int netdev_rx_handler_register(struct net_device *dev, 5604 rx_handler_func_t *rx_handler, 5605 void *rx_handler_data) 5606 { 5607 if (netdev_is_rx_handler_busy(dev)) 5608 return -EBUSY; 5609 5610 if (dev->priv_flags & IFF_NO_RX_HANDLER) 5611 return -EINVAL; 5612 5613 /* Note: rx_handler_data must be set before rx_handler */ 5614 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data); 5615 rcu_assign_pointer(dev->rx_handler, rx_handler); 5616 5617 return 0; 5618 } 5619 EXPORT_SYMBOL_GPL(netdev_rx_handler_register); 5620 5621 /** 5622 * netdev_rx_handler_unregister - unregister receive handler 5623 * @dev: device to unregister a handler from 5624 * 5625 * Unregister a receive handler from a device. 5626 * 5627 * The caller must hold the rtnl_mutex. 5628 */ 5629 void netdev_rx_handler_unregister(struct net_device *dev) 5630 { 5631 5632 ASSERT_RTNL(); 5633 RCU_INIT_POINTER(dev->rx_handler, NULL); 5634 /* a reader seeing a non NULL rx_handler in a rcu_read_lock() 5635 * section has a guarantee to see a non NULL rx_handler_data 5636 * as well. 5637 */ 5638 synchronize_net(); 5639 RCU_INIT_POINTER(dev->rx_handler_data, NULL); 5640 } 5641 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister); 5642 5643 /* 5644 * Limit the use of PFMEMALLOC reserves to those protocols that implement 5645 * the special handling of PFMEMALLOC skbs. 5646 */ 5647 static bool skb_pfmemalloc_protocol(struct sk_buff *skb) 5648 { 5649 switch (skb->protocol) { 5650 case htons(ETH_P_ARP): 5651 case htons(ETH_P_IP): 5652 case htons(ETH_P_IPV6): 5653 case htons(ETH_P_8021Q): 5654 case htons(ETH_P_8021AD): 5655 return true; 5656 default: 5657 return false; 5658 } 5659 } 5660 5661 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev, 5662 int *ret, struct net_device *orig_dev) 5663 { 5664 if (nf_hook_ingress_active(skb)) { 5665 int ingress_retval; 5666 5667 if (*pt_prev) { 5668 *ret = deliver_skb(skb, *pt_prev, orig_dev); 5669 *pt_prev = NULL; 5670 } 5671 5672 rcu_read_lock(); 5673 ingress_retval = nf_hook_ingress(skb); 5674 rcu_read_unlock(); 5675 return ingress_retval; 5676 } 5677 return 0; 5678 } 5679 5680 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc, 5681 struct packet_type **ppt_prev) 5682 { 5683 struct packet_type *ptype, *pt_prev; 5684 rx_handler_func_t *rx_handler; 5685 struct sk_buff *skb = *pskb; 5686 struct net_device *orig_dev; 5687 bool deliver_exact = false; 5688 int ret = NET_RX_DROP; 5689 __be16 type; 5690 5691 net_timestamp_check(!READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5692 5693 trace_netif_receive_skb(skb); 5694 5695 orig_dev = skb->dev; 5696 5697 skb_reset_network_header(skb); 5698 #if !defined(CONFIG_DEBUG_NET) 5699 /* We plan to no longer reset the transport header here. 5700 * Give some time to fuzzers and dev build to catch bugs 5701 * in network stacks. 5702 */ 5703 if (!skb_transport_header_was_set(skb)) 5704 skb_reset_transport_header(skb); 5705 #endif 5706 skb_reset_mac_len(skb); 5707 5708 pt_prev = NULL; 5709 5710 another_round: 5711 skb->skb_iif = skb->dev->ifindex; 5712 5713 __this_cpu_inc(softnet_data.processed); 5714 5715 if (static_branch_unlikely(&generic_xdp_needed_key)) { 5716 int ret2; 5717 5718 migrate_disable(); 5719 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), 5720 &skb); 5721 migrate_enable(); 5722 5723 if (ret2 != XDP_PASS) { 5724 ret = NET_RX_DROP; 5725 goto out; 5726 } 5727 } 5728 5729 if (eth_type_vlan(skb->protocol)) { 5730 skb = skb_vlan_untag(skb); 5731 if (unlikely(!skb)) 5732 goto out; 5733 } 5734 5735 if (skb_skip_tc_classify(skb)) 5736 goto skip_classify; 5737 5738 if (pfmemalloc) 5739 goto skip_taps; 5740 5741 list_for_each_entry_rcu(ptype, &net_hotdata.ptype_all, list) { 5742 if (pt_prev) 5743 ret = deliver_skb(skb, pt_prev, orig_dev); 5744 pt_prev = ptype; 5745 } 5746 5747 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) { 5748 if (pt_prev) 5749 ret = deliver_skb(skb, pt_prev, orig_dev); 5750 pt_prev = ptype; 5751 } 5752 5753 skip_taps: 5754 #ifdef CONFIG_NET_INGRESS 5755 if (static_branch_unlikely(&ingress_needed_key)) { 5756 bool another = false; 5757 5758 nf_skip_egress(skb, true); 5759 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev, 5760 &another); 5761 if (another) 5762 goto another_round; 5763 if (!skb) 5764 goto out; 5765 5766 nf_skip_egress(skb, false); 5767 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0) 5768 goto out; 5769 } 5770 #endif 5771 skb_reset_redirect(skb); 5772 skip_classify: 5773 if (pfmemalloc && !skb_pfmemalloc_protocol(skb)) 5774 goto drop; 5775 5776 if (skb_vlan_tag_present(skb)) { 5777 if (pt_prev) { 5778 ret = deliver_skb(skb, pt_prev, orig_dev); 5779 pt_prev = NULL; 5780 } 5781 if (vlan_do_receive(&skb)) 5782 goto another_round; 5783 else if (unlikely(!skb)) 5784 goto out; 5785 } 5786 5787 rx_handler = rcu_dereference(skb->dev->rx_handler); 5788 if (rx_handler) { 5789 if (pt_prev) { 5790 ret = deliver_skb(skb, pt_prev, orig_dev); 5791 pt_prev = NULL; 5792 } 5793 switch (rx_handler(&skb)) { 5794 case RX_HANDLER_CONSUMED: 5795 ret = NET_RX_SUCCESS; 5796 goto out; 5797 case RX_HANDLER_ANOTHER: 5798 goto another_round; 5799 case RX_HANDLER_EXACT: 5800 deliver_exact = true; 5801 break; 5802 case RX_HANDLER_PASS: 5803 break; 5804 default: 5805 BUG(); 5806 } 5807 } 5808 5809 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) { 5810 check_vlan_id: 5811 if (skb_vlan_tag_get_id(skb)) { 5812 /* Vlan id is non 0 and vlan_do_receive() above couldn't 5813 * find vlan device. 5814 */ 5815 skb->pkt_type = PACKET_OTHERHOST; 5816 } else if (eth_type_vlan(skb->protocol)) { 5817 /* Outer header is 802.1P with vlan 0, inner header is 5818 * 802.1Q or 802.1AD and vlan_do_receive() above could 5819 * not find vlan dev for vlan id 0. 5820 */ 5821 __vlan_hwaccel_clear_tag(skb); 5822 skb = skb_vlan_untag(skb); 5823 if (unlikely(!skb)) 5824 goto out; 5825 if (vlan_do_receive(&skb)) 5826 /* After stripping off 802.1P header with vlan 0 5827 * vlan dev is found for inner header. 5828 */ 5829 goto another_round; 5830 else if (unlikely(!skb)) 5831 goto out; 5832 else 5833 /* We have stripped outer 802.1P vlan 0 header. 5834 * But could not find vlan dev. 5835 * check again for vlan id to set OTHERHOST. 5836 */ 5837 goto check_vlan_id; 5838 } 5839 /* Note: we might in the future use prio bits 5840 * and set skb->priority like in vlan_do_receive() 5841 * For the time being, just ignore Priority Code Point 5842 */ 5843 __vlan_hwaccel_clear_tag(skb); 5844 } 5845 5846 type = skb->protocol; 5847 5848 /* deliver only exact match when indicated */ 5849 if (likely(!deliver_exact)) { 5850 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5851 &ptype_base[ntohs(type) & 5852 PTYPE_HASH_MASK]); 5853 } 5854 5855 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5856 &orig_dev->ptype_specific); 5857 5858 if (unlikely(skb->dev != orig_dev)) { 5859 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5860 &skb->dev->ptype_specific); 5861 } 5862 5863 if (pt_prev) { 5864 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 5865 goto drop; 5866 *ppt_prev = pt_prev; 5867 } else { 5868 drop: 5869 if (!deliver_exact) 5870 dev_core_stats_rx_dropped_inc(skb->dev); 5871 else 5872 dev_core_stats_rx_nohandler_inc(skb->dev); 5873 kfree_skb_reason(skb, SKB_DROP_REASON_UNHANDLED_PROTO); 5874 /* Jamal, now you will not able to escape explaining 5875 * me how you were going to use this. :-) 5876 */ 5877 ret = NET_RX_DROP; 5878 } 5879 5880 out: 5881 /* The invariant here is that if *ppt_prev is not NULL 5882 * then skb should also be non-NULL. 5883 * 5884 * Apparently *ppt_prev assignment above holds this invariant due to 5885 * skb dereferencing near it. 5886 */ 5887 *pskb = skb; 5888 return ret; 5889 } 5890 5891 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc) 5892 { 5893 struct net_device *orig_dev = skb->dev; 5894 struct packet_type *pt_prev = NULL; 5895 int ret; 5896 5897 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5898 if (pt_prev) 5899 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb, 5900 skb->dev, pt_prev, orig_dev); 5901 return ret; 5902 } 5903 5904 /** 5905 * netif_receive_skb_core - special purpose version of netif_receive_skb 5906 * @skb: buffer to process 5907 * 5908 * More direct receive version of netif_receive_skb(). It should 5909 * only be used by callers that have a need to skip RPS and Generic XDP. 5910 * Caller must also take care of handling if ``(page_is_)pfmemalloc``. 5911 * 5912 * This function may only be called from softirq context and interrupts 5913 * should be enabled. 5914 * 5915 * Return values (usually ignored): 5916 * NET_RX_SUCCESS: no congestion 5917 * NET_RX_DROP: packet was dropped 5918 */ 5919 int netif_receive_skb_core(struct sk_buff *skb) 5920 { 5921 int ret; 5922 5923 rcu_read_lock(); 5924 ret = __netif_receive_skb_one_core(skb, false); 5925 rcu_read_unlock(); 5926 5927 return ret; 5928 } 5929 EXPORT_SYMBOL(netif_receive_skb_core); 5930 5931 static inline void __netif_receive_skb_list_ptype(struct list_head *head, 5932 struct packet_type *pt_prev, 5933 struct net_device *orig_dev) 5934 { 5935 struct sk_buff *skb, *next; 5936 5937 if (!pt_prev) 5938 return; 5939 if (list_empty(head)) 5940 return; 5941 if (pt_prev->list_func != NULL) 5942 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv, 5943 ip_list_rcv, head, pt_prev, orig_dev); 5944 else 5945 list_for_each_entry_safe(skb, next, head, list) { 5946 skb_list_del_init(skb); 5947 pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 5948 } 5949 } 5950 5951 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc) 5952 { 5953 /* Fast-path assumptions: 5954 * - There is no RX handler. 5955 * - Only one packet_type matches. 5956 * If either of these fails, we will end up doing some per-packet 5957 * processing in-line, then handling the 'last ptype' for the whole 5958 * sublist. This can't cause out-of-order delivery to any single ptype, 5959 * because the 'last ptype' must be constant across the sublist, and all 5960 * other ptypes are handled per-packet. 5961 */ 5962 /* Current (common) ptype of sublist */ 5963 struct packet_type *pt_curr = NULL; 5964 /* Current (common) orig_dev of sublist */ 5965 struct net_device *od_curr = NULL; 5966 struct sk_buff *skb, *next; 5967 LIST_HEAD(sublist); 5968 5969 list_for_each_entry_safe(skb, next, head, list) { 5970 struct net_device *orig_dev = skb->dev; 5971 struct packet_type *pt_prev = NULL; 5972 5973 skb_list_del_init(skb); 5974 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5975 if (!pt_prev) 5976 continue; 5977 if (pt_curr != pt_prev || od_curr != orig_dev) { 5978 /* dispatch old sublist */ 5979 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5980 /* start new sublist */ 5981 INIT_LIST_HEAD(&sublist); 5982 pt_curr = pt_prev; 5983 od_curr = orig_dev; 5984 } 5985 list_add_tail(&skb->list, &sublist); 5986 } 5987 5988 /* dispatch final sublist */ 5989 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5990 } 5991 5992 static int __netif_receive_skb(struct sk_buff *skb) 5993 { 5994 int ret; 5995 5996 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) { 5997 unsigned int noreclaim_flag; 5998 5999 /* 6000 * PFMEMALLOC skbs are special, they should 6001 * - be delivered to SOCK_MEMALLOC sockets only 6002 * - stay away from userspace 6003 * - have bounded memory usage 6004 * 6005 * Use PF_MEMALLOC as this saves us from propagating the allocation 6006 * context down to all allocation sites. 6007 */ 6008 noreclaim_flag = memalloc_noreclaim_save(); 6009 ret = __netif_receive_skb_one_core(skb, true); 6010 memalloc_noreclaim_restore(noreclaim_flag); 6011 } else 6012 ret = __netif_receive_skb_one_core(skb, false); 6013 6014 return ret; 6015 } 6016 6017 static void __netif_receive_skb_list(struct list_head *head) 6018 { 6019 unsigned long noreclaim_flag = 0; 6020 struct sk_buff *skb, *next; 6021 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */ 6022 6023 list_for_each_entry_safe(skb, next, head, list) { 6024 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) { 6025 struct list_head sublist; 6026 6027 /* Handle the previous sublist */ 6028 list_cut_before(&sublist, head, &skb->list); 6029 if (!list_empty(&sublist)) 6030 __netif_receive_skb_list_core(&sublist, pfmemalloc); 6031 pfmemalloc = !pfmemalloc; 6032 /* See comments in __netif_receive_skb */ 6033 if (pfmemalloc) 6034 noreclaim_flag = memalloc_noreclaim_save(); 6035 else 6036 memalloc_noreclaim_restore(noreclaim_flag); 6037 } 6038 } 6039 /* Handle the remaining sublist */ 6040 if (!list_empty(head)) 6041 __netif_receive_skb_list_core(head, pfmemalloc); 6042 /* Restore pflags */ 6043 if (pfmemalloc) 6044 memalloc_noreclaim_restore(noreclaim_flag); 6045 } 6046 6047 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp) 6048 { 6049 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog); 6050 struct bpf_prog *new = xdp->prog; 6051 int ret = 0; 6052 6053 switch (xdp->command) { 6054 case XDP_SETUP_PROG: 6055 rcu_assign_pointer(dev->xdp_prog, new); 6056 if (old) 6057 bpf_prog_put(old); 6058 6059 if (old && !new) { 6060 static_branch_dec(&generic_xdp_needed_key); 6061 } else if (new && !old) { 6062 static_branch_inc(&generic_xdp_needed_key); 6063 dev_disable_lro(dev); 6064 dev_disable_gro_hw(dev); 6065 } 6066 break; 6067 6068 default: 6069 ret = -EINVAL; 6070 break; 6071 } 6072 6073 return ret; 6074 } 6075 6076 static int netif_receive_skb_internal(struct sk_buff *skb) 6077 { 6078 int ret; 6079 6080 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 6081 6082 if (skb_defer_rx_timestamp(skb)) 6083 return NET_RX_SUCCESS; 6084 6085 rcu_read_lock(); 6086 #ifdef CONFIG_RPS 6087 if (static_branch_unlikely(&rps_needed)) { 6088 struct rps_dev_flow voidflow, *rflow = &voidflow; 6089 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6090 6091 if (cpu >= 0) { 6092 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6093 rcu_read_unlock(); 6094 return ret; 6095 } 6096 } 6097 #endif 6098 ret = __netif_receive_skb(skb); 6099 rcu_read_unlock(); 6100 return ret; 6101 } 6102 6103 void netif_receive_skb_list_internal(struct list_head *head) 6104 { 6105 struct sk_buff *skb, *next; 6106 LIST_HEAD(sublist); 6107 6108 list_for_each_entry_safe(skb, next, head, list) { 6109 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), 6110 skb); 6111 skb_list_del_init(skb); 6112 if (!skb_defer_rx_timestamp(skb)) 6113 list_add_tail(&skb->list, &sublist); 6114 } 6115 list_splice_init(&sublist, head); 6116 6117 rcu_read_lock(); 6118 #ifdef CONFIG_RPS 6119 if (static_branch_unlikely(&rps_needed)) { 6120 list_for_each_entry_safe(skb, next, head, list) { 6121 struct rps_dev_flow voidflow, *rflow = &voidflow; 6122 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6123 6124 if (cpu >= 0) { 6125 /* Will be handled, remove from list */ 6126 skb_list_del_init(skb); 6127 enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6128 } 6129 } 6130 } 6131 #endif 6132 __netif_receive_skb_list(head); 6133 rcu_read_unlock(); 6134 } 6135 6136 /** 6137 * netif_receive_skb - process receive buffer from network 6138 * @skb: buffer to process 6139 * 6140 * netif_receive_skb() is the main receive data processing function. 6141 * It always succeeds. The buffer may be dropped during processing 6142 * for congestion control or by the protocol layers. 6143 * 6144 * This function may only be called from softirq context and interrupts 6145 * should be enabled. 6146 * 6147 * Return values (usually ignored): 6148 * NET_RX_SUCCESS: no congestion 6149 * NET_RX_DROP: packet was dropped 6150 */ 6151 int netif_receive_skb(struct sk_buff *skb) 6152 { 6153 int ret; 6154 6155 trace_netif_receive_skb_entry(skb); 6156 6157 ret = netif_receive_skb_internal(skb); 6158 trace_netif_receive_skb_exit(ret); 6159 6160 return ret; 6161 } 6162 EXPORT_SYMBOL(netif_receive_skb); 6163 6164 /** 6165 * netif_receive_skb_list - process many receive buffers from network 6166 * @head: list of skbs to process. 6167 * 6168 * Since return value of netif_receive_skb() is normally ignored, and 6169 * wouldn't be meaningful for a list, this function returns void. 6170 * 6171 * This function may only be called from softirq context and interrupts 6172 * should be enabled. 6173 */ 6174 void netif_receive_skb_list(struct list_head *head) 6175 { 6176 struct sk_buff *skb; 6177 6178 if (list_empty(head)) 6179 return; 6180 if (trace_netif_receive_skb_list_entry_enabled()) { 6181 list_for_each_entry(skb, head, list) 6182 trace_netif_receive_skb_list_entry(skb); 6183 } 6184 netif_receive_skb_list_internal(head); 6185 trace_netif_receive_skb_list_exit(0); 6186 } 6187 EXPORT_SYMBOL(netif_receive_skb_list); 6188 6189 /* Network device is going away, flush any packets still pending */ 6190 static void flush_backlog(struct work_struct *work) 6191 { 6192 struct sk_buff *skb, *tmp; 6193 struct softnet_data *sd; 6194 6195 local_bh_disable(); 6196 sd = this_cpu_ptr(&softnet_data); 6197 6198 backlog_lock_irq_disable(sd); 6199 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) { 6200 if (skb->dev->reg_state == NETREG_UNREGISTERING) { 6201 __skb_unlink(skb, &sd->input_pkt_queue); 6202 dev_kfree_skb_irq(skb); 6203 rps_input_queue_head_incr(sd); 6204 } 6205 } 6206 backlog_unlock_irq_enable(sd); 6207 6208 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6209 skb_queue_walk_safe(&sd->process_queue, skb, tmp) { 6210 if (skb->dev->reg_state == NETREG_UNREGISTERING) { 6211 __skb_unlink(skb, &sd->process_queue); 6212 kfree_skb(skb); 6213 rps_input_queue_head_incr(sd); 6214 } 6215 } 6216 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6217 local_bh_enable(); 6218 } 6219 6220 static bool flush_required(int cpu) 6221 { 6222 #if IS_ENABLED(CONFIG_RPS) 6223 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 6224 bool do_flush; 6225 6226 backlog_lock_irq_disable(sd); 6227 6228 /* as insertion into process_queue happens with the rps lock held, 6229 * process_queue access may race only with dequeue 6230 */ 6231 do_flush = !skb_queue_empty(&sd->input_pkt_queue) || 6232 !skb_queue_empty_lockless(&sd->process_queue); 6233 backlog_unlock_irq_enable(sd); 6234 6235 return do_flush; 6236 #endif 6237 /* without RPS we can't safely check input_pkt_queue: during a 6238 * concurrent remote skb_queue_splice() we can detect as empty both 6239 * input_pkt_queue and process_queue even if the latter could end-up 6240 * containing a lot of packets. 6241 */ 6242 return true; 6243 } 6244 6245 struct flush_backlogs { 6246 cpumask_t flush_cpus; 6247 struct work_struct w[]; 6248 }; 6249 6250 static struct flush_backlogs *flush_backlogs_alloc(void) 6251 { 6252 return kmalloc(struct_size_t(struct flush_backlogs, w, nr_cpu_ids), 6253 GFP_KERNEL); 6254 } 6255 6256 static struct flush_backlogs *flush_backlogs_fallback; 6257 static DEFINE_MUTEX(flush_backlogs_mutex); 6258 6259 static void flush_all_backlogs(void) 6260 { 6261 struct flush_backlogs *ptr = flush_backlogs_alloc(); 6262 unsigned int cpu; 6263 6264 if (!ptr) { 6265 mutex_lock(&flush_backlogs_mutex); 6266 ptr = flush_backlogs_fallback; 6267 } 6268 cpumask_clear(&ptr->flush_cpus); 6269 6270 cpus_read_lock(); 6271 6272 for_each_online_cpu(cpu) { 6273 if (flush_required(cpu)) { 6274 INIT_WORK(&ptr->w[cpu], flush_backlog); 6275 queue_work_on(cpu, system_highpri_wq, &ptr->w[cpu]); 6276 __cpumask_set_cpu(cpu, &ptr->flush_cpus); 6277 } 6278 } 6279 6280 /* we can have in flight packet[s] on the cpus we are not flushing, 6281 * synchronize_net() in unregister_netdevice_many() will take care of 6282 * them. 6283 */ 6284 for_each_cpu(cpu, &ptr->flush_cpus) 6285 flush_work(&ptr->w[cpu]); 6286 6287 cpus_read_unlock(); 6288 6289 if (ptr != flush_backlogs_fallback) 6290 kfree(ptr); 6291 else 6292 mutex_unlock(&flush_backlogs_mutex); 6293 } 6294 6295 static void net_rps_send_ipi(struct softnet_data *remsd) 6296 { 6297 #ifdef CONFIG_RPS 6298 while (remsd) { 6299 struct softnet_data *next = remsd->rps_ipi_next; 6300 6301 if (cpu_online(remsd->cpu)) 6302 smp_call_function_single_async(remsd->cpu, &remsd->csd); 6303 remsd = next; 6304 } 6305 #endif 6306 } 6307 6308 /* 6309 * net_rps_action_and_irq_enable sends any pending IPI's for rps. 6310 * Note: called with local irq disabled, but exits with local irq enabled. 6311 */ 6312 static void net_rps_action_and_irq_enable(struct softnet_data *sd) 6313 { 6314 #ifdef CONFIG_RPS 6315 struct softnet_data *remsd = sd->rps_ipi_list; 6316 6317 if (!use_backlog_threads() && remsd) { 6318 sd->rps_ipi_list = NULL; 6319 6320 local_irq_enable(); 6321 6322 /* Send pending IPI's to kick RPS processing on remote cpus. */ 6323 net_rps_send_ipi(remsd); 6324 } else 6325 #endif 6326 local_irq_enable(); 6327 } 6328 6329 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd) 6330 { 6331 #ifdef CONFIG_RPS 6332 return !use_backlog_threads() && sd->rps_ipi_list; 6333 #else 6334 return false; 6335 #endif 6336 } 6337 6338 static int process_backlog(struct napi_struct *napi, int quota) 6339 { 6340 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog); 6341 bool again = true; 6342 int work = 0; 6343 6344 /* Check if we have pending ipi, its better to send them now, 6345 * not waiting net_rx_action() end. 6346 */ 6347 if (sd_has_rps_ipi_waiting(sd)) { 6348 local_irq_disable(); 6349 net_rps_action_and_irq_enable(sd); 6350 } 6351 6352 napi->weight = READ_ONCE(net_hotdata.dev_rx_weight); 6353 while (again) { 6354 struct sk_buff *skb; 6355 6356 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6357 while ((skb = __skb_dequeue(&sd->process_queue))) { 6358 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6359 rcu_read_lock(); 6360 __netif_receive_skb(skb); 6361 rcu_read_unlock(); 6362 if (++work >= quota) { 6363 rps_input_queue_head_add(sd, work); 6364 return work; 6365 } 6366 6367 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6368 } 6369 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6370 6371 backlog_lock_irq_disable(sd); 6372 if (skb_queue_empty(&sd->input_pkt_queue)) { 6373 /* 6374 * Inline a custom version of __napi_complete(). 6375 * only current cpu owns and manipulates this napi, 6376 * and NAPI_STATE_SCHED is the only possible flag set 6377 * on backlog. 6378 * We can use a plain write instead of clear_bit(), 6379 * and we dont need an smp_mb() memory barrier. 6380 */ 6381 napi->state &= NAPIF_STATE_THREADED; 6382 again = false; 6383 } else { 6384 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6385 skb_queue_splice_tail_init(&sd->input_pkt_queue, 6386 &sd->process_queue); 6387 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6388 } 6389 backlog_unlock_irq_enable(sd); 6390 } 6391 6392 if (work) 6393 rps_input_queue_head_add(sd, work); 6394 return work; 6395 } 6396 6397 /** 6398 * __napi_schedule - schedule for receive 6399 * @n: entry to schedule 6400 * 6401 * The entry's receive function will be scheduled to run. 6402 * Consider using __napi_schedule_irqoff() if hard irqs are masked. 6403 */ 6404 void __napi_schedule(struct napi_struct *n) 6405 { 6406 unsigned long flags; 6407 6408 local_irq_save(flags); 6409 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6410 local_irq_restore(flags); 6411 } 6412 EXPORT_SYMBOL(__napi_schedule); 6413 6414 /** 6415 * napi_schedule_prep - check if napi can be scheduled 6416 * @n: napi context 6417 * 6418 * Test if NAPI routine is already running, and if not mark 6419 * it as running. This is used as a condition variable to 6420 * insure only one NAPI poll instance runs. We also make 6421 * sure there is no pending NAPI disable. 6422 */ 6423 bool napi_schedule_prep(struct napi_struct *n) 6424 { 6425 unsigned long new, val = READ_ONCE(n->state); 6426 6427 do { 6428 if (unlikely(val & NAPIF_STATE_DISABLE)) 6429 return false; 6430 new = val | NAPIF_STATE_SCHED; 6431 6432 /* Sets STATE_MISSED bit if STATE_SCHED was already set 6433 * This was suggested by Alexander Duyck, as compiler 6434 * emits better code than : 6435 * if (val & NAPIF_STATE_SCHED) 6436 * new |= NAPIF_STATE_MISSED; 6437 */ 6438 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED * 6439 NAPIF_STATE_MISSED; 6440 } while (!try_cmpxchg(&n->state, &val, new)); 6441 6442 return !(val & NAPIF_STATE_SCHED); 6443 } 6444 EXPORT_SYMBOL(napi_schedule_prep); 6445 6446 /** 6447 * __napi_schedule_irqoff - schedule for receive 6448 * @n: entry to schedule 6449 * 6450 * Variant of __napi_schedule() assuming hard irqs are masked. 6451 * 6452 * On PREEMPT_RT enabled kernels this maps to __napi_schedule() 6453 * because the interrupt disabled assumption might not be true 6454 * due to force-threaded interrupts and spinlock substitution. 6455 */ 6456 void __napi_schedule_irqoff(struct napi_struct *n) 6457 { 6458 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6459 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6460 else 6461 __napi_schedule(n); 6462 } 6463 EXPORT_SYMBOL(__napi_schedule_irqoff); 6464 6465 bool napi_complete_done(struct napi_struct *n, int work_done) 6466 { 6467 unsigned long flags, val, new, timeout = 0; 6468 bool ret = true; 6469 6470 /* 6471 * 1) Don't let napi dequeue from the cpu poll list 6472 * just in case its running on a different cpu. 6473 * 2) If we are busy polling, do nothing here, we have 6474 * the guarantee we will be called later. 6475 */ 6476 if (unlikely(n->state & (NAPIF_STATE_NPSVC | 6477 NAPIF_STATE_IN_BUSY_POLL))) 6478 return false; 6479 6480 if (work_done) { 6481 if (n->gro_bitmask) 6482 timeout = napi_get_gro_flush_timeout(n); 6483 n->defer_hard_irqs_count = napi_get_defer_hard_irqs(n); 6484 } 6485 if (n->defer_hard_irqs_count > 0) { 6486 n->defer_hard_irqs_count--; 6487 timeout = napi_get_gro_flush_timeout(n); 6488 if (timeout) 6489 ret = false; 6490 } 6491 if (n->gro_bitmask) { 6492 /* When the NAPI instance uses a timeout and keeps postponing 6493 * it, we need to bound somehow the time packets are kept in 6494 * the GRO layer 6495 */ 6496 napi_gro_flush(n, !!timeout); 6497 } 6498 6499 gro_normal_list(n); 6500 6501 if (unlikely(!list_empty(&n->poll_list))) { 6502 /* If n->poll_list is not empty, we need to mask irqs */ 6503 local_irq_save(flags); 6504 list_del_init(&n->poll_list); 6505 local_irq_restore(flags); 6506 } 6507 WRITE_ONCE(n->list_owner, -1); 6508 6509 val = READ_ONCE(n->state); 6510 do { 6511 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED)); 6512 6513 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED | 6514 NAPIF_STATE_SCHED_THREADED | 6515 NAPIF_STATE_PREFER_BUSY_POLL); 6516 6517 /* If STATE_MISSED was set, leave STATE_SCHED set, 6518 * because we will call napi->poll() one more time. 6519 * This C code was suggested by Alexander Duyck to help gcc. 6520 */ 6521 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED * 6522 NAPIF_STATE_SCHED; 6523 } while (!try_cmpxchg(&n->state, &val, new)); 6524 6525 if (unlikely(val & NAPIF_STATE_MISSED)) { 6526 __napi_schedule(n); 6527 return false; 6528 } 6529 6530 if (timeout) 6531 hrtimer_start(&n->timer, ns_to_ktime(timeout), 6532 HRTIMER_MODE_REL_PINNED); 6533 return ret; 6534 } 6535 EXPORT_SYMBOL(napi_complete_done); 6536 6537 static void skb_defer_free_flush(struct softnet_data *sd) 6538 { 6539 struct sk_buff *skb, *next; 6540 6541 /* Paired with WRITE_ONCE() in skb_attempt_defer_free() */ 6542 if (!READ_ONCE(sd->defer_list)) 6543 return; 6544 6545 spin_lock(&sd->defer_lock); 6546 skb = sd->defer_list; 6547 sd->defer_list = NULL; 6548 sd->defer_count = 0; 6549 spin_unlock(&sd->defer_lock); 6550 6551 while (skb != NULL) { 6552 next = skb->next; 6553 napi_consume_skb(skb, 1); 6554 skb = next; 6555 } 6556 } 6557 6558 #if defined(CONFIG_NET_RX_BUSY_POLL) 6559 6560 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule) 6561 { 6562 if (!skip_schedule) { 6563 gro_normal_list(napi); 6564 __napi_schedule(napi); 6565 return; 6566 } 6567 6568 if (napi->gro_bitmask) { 6569 /* flush too old packets 6570 * If HZ < 1000, flush all packets. 6571 */ 6572 napi_gro_flush(napi, HZ >= 1000); 6573 } 6574 6575 gro_normal_list(napi); 6576 clear_bit(NAPI_STATE_SCHED, &napi->state); 6577 } 6578 6579 enum { 6580 NAPI_F_PREFER_BUSY_POLL = 1, 6581 NAPI_F_END_ON_RESCHED = 2, 6582 }; 6583 6584 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, 6585 unsigned flags, u16 budget) 6586 { 6587 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6588 bool skip_schedule = false; 6589 unsigned long timeout; 6590 int rc; 6591 6592 /* Busy polling means there is a high chance device driver hard irq 6593 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was 6594 * set in napi_schedule_prep(). 6595 * Since we are about to call napi->poll() once more, we can safely 6596 * clear NAPI_STATE_MISSED. 6597 * 6598 * Note: x86 could use a single "lock and ..." instruction 6599 * to perform these two clear_bit() 6600 */ 6601 clear_bit(NAPI_STATE_MISSED, &napi->state); 6602 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state); 6603 6604 local_bh_disable(); 6605 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6606 6607 if (flags & NAPI_F_PREFER_BUSY_POLL) { 6608 napi->defer_hard_irqs_count = napi_get_defer_hard_irqs(napi); 6609 timeout = napi_get_gro_flush_timeout(napi); 6610 if (napi->defer_hard_irqs_count && timeout) { 6611 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED); 6612 skip_schedule = true; 6613 } 6614 } 6615 6616 /* All we really want here is to re-enable device interrupts. 6617 * Ideally, a new ndo_busy_poll_stop() could avoid another round. 6618 */ 6619 rc = napi->poll(napi, budget); 6620 /* We can't gro_normal_list() here, because napi->poll() might have 6621 * rearmed the napi (napi_complete_done()) in which case it could 6622 * already be running on another CPU. 6623 */ 6624 trace_napi_poll(napi, rc, budget); 6625 netpoll_poll_unlock(have_poll_lock); 6626 if (rc == budget) 6627 __busy_poll_stop(napi, skip_schedule); 6628 bpf_net_ctx_clear(bpf_net_ctx); 6629 local_bh_enable(); 6630 } 6631 6632 static void __napi_busy_loop(unsigned int napi_id, 6633 bool (*loop_end)(void *, unsigned long), 6634 void *loop_end_arg, unsigned flags, u16 budget) 6635 { 6636 unsigned long start_time = loop_end ? busy_loop_current_time() : 0; 6637 int (*napi_poll)(struct napi_struct *napi, int budget); 6638 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6639 void *have_poll_lock = NULL; 6640 struct napi_struct *napi; 6641 6642 WARN_ON_ONCE(!rcu_read_lock_held()); 6643 6644 restart: 6645 napi_poll = NULL; 6646 6647 napi = napi_by_id(napi_id); 6648 if (!napi) 6649 return; 6650 6651 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6652 preempt_disable(); 6653 for (;;) { 6654 int work = 0; 6655 6656 local_bh_disable(); 6657 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6658 if (!napi_poll) { 6659 unsigned long val = READ_ONCE(napi->state); 6660 6661 /* If multiple threads are competing for this napi, 6662 * we avoid dirtying napi->state as much as we can. 6663 */ 6664 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED | 6665 NAPIF_STATE_IN_BUSY_POLL)) { 6666 if (flags & NAPI_F_PREFER_BUSY_POLL) 6667 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6668 goto count; 6669 } 6670 if (cmpxchg(&napi->state, val, 6671 val | NAPIF_STATE_IN_BUSY_POLL | 6672 NAPIF_STATE_SCHED) != val) { 6673 if (flags & NAPI_F_PREFER_BUSY_POLL) 6674 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6675 goto count; 6676 } 6677 have_poll_lock = netpoll_poll_lock(napi); 6678 napi_poll = napi->poll; 6679 } 6680 work = napi_poll(napi, budget); 6681 trace_napi_poll(napi, work, budget); 6682 gro_normal_list(napi); 6683 count: 6684 if (work > 0) 6685 __NET_ADD_STATS(dev_net(napi->dev), 6686 LINUX_MIB_BUSYPOLLRXPACKETS, work); 6687 skb_defer_free_flush(this_cpu_ptr(&softnet_data)); 6688 bpf_net_ctx_clear(bpf_net_ctx); 6689 local_bh_enable(); 6690 6691 if (!loop_end || loop_end(loop_end_arg, start_time)) 6692 break; 6693 6694 if (unlikely(need_resched())) { 6695 if (flags & NAPI_F_END_ON_RESCHED) 6696 break; 6697 if (napi_poll) 6698 busy_poll_stop(napi, have_poll_lock, flags, budget); 6699 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6700 preempt_enable(); 6701 rcu_read_unlock(); 6702 cond_resched(); 6703 rcu_read_lock(); 6704 if (loop_end(loop_end_arg, start_time)) 6705 return; 6706 goto restart; 6707 } 6708 cpu_relax(); 6709 } 6710 if (napi_poll) 6711 busy_poll_stop(napi, have_poll_lock, flags, budget); 6712 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6713 preempt_enable(); 6714 } 6715 6716 void napi_busy_loop_rcu(unsigned int napi_id, 6717 bool (*loop_end)(void *, unsigned long), 6718 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6719 { 6720 unsigned flags = NAPI_F_END_ON_RESCHED; 6721 6722 if (prefer_busy_poll) 6723 flags |= NAPI_F_PREFER_BUSY_POLL; 6724 6725 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6726 } 6727 6728 void napi_busy_loop(unsigned int napi_id, 6729 bool (*loop_end)(void *, unsigned long), 6730 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6731 { 6732 unsigned flags = prefer_busy_poll ? NAPI_F_PREFER_BUSY_POLL : 0; 6733 6734 rcu_read_lock(); 6735 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6736 rcu_read_unlock(); 6737 } 6738 EXPORT_SYMBOL(napi_busy_loop); 6739 6740 void napi_suspend_irqs(unsigned int napi_id) 6741 { 6742 struct napi_struct *napi; 6743 6744 rcu_read_lock(); 6745 napi = napi_by_id(napi_id); 6746 if (napi) { 6747 unsigned long timeout = napi_get_irq_suspend_timeout(napi); 6748 6749 if (timeout) 6750 hrtimer_start(&napi->timer, ns_to_ktime(timeout), 6751 HRTIMER_MODE_REL_PINNED); 6752 } 6753 rcu_read_unlock(); 6754 } 6755 6756 void napi_resume_irqs(unsigned int napi_id) 6757 { 6758 struct napi_struct *napi; 6759 6760 rcu_read_lock(); 6761 napi = napi_by_id(napi_id); 6762 if (napi) { 6763 /* If irq_suspend_timeout is set to 0 between the call to 6764 * napi_suspend_irqs and now, the original value still 6765 * determines the safety timeout as intended and napi_watchdog 6766 * will resume irq processing. 6767 */ 6768 if (napi_get_irq_suspend_timeout(napi)) { 6769 local_bh_disable(); 6770 napi_schedule(napi); 6771 local_bh_enable(); 6772 } 6773 } 6774 rcu_read_unlock(); 6775 } 6776 6777 #endif /* CONFIG_NET_RX_BUSY_POLL */ 6778 6779 static void __napi_hash_add_with_id(struct napi_struct *napi, 6780 unsigned int napi_id) 6781 { 6782 WRITE_ONCE(napi->napi_id, napi_id); 6783 hlist_add_head_rcu(&napi->napi_hash_node, 6784 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]); 6785 } 6786 6787 static void napi_hash_add_with_id(struct napi_struct *napi, 6788 unsigned int napi_id) 6789 { 6790 unsigned long flags; 6791 6792 spin_lock_irqsave(&napi_hash_lock, flags); 6793 WARN_ON_ONCE(napi_by_id(napi_id)); 6794 __napi_hash_add_with_id(napi, napi_id); 6795 spin_unlock_irqrestore(&napi_hash_lock, flags); 6796 } 6797 6798 static void napi_hash_add(struct napi_struct *napi) 6799 { 6800 unsigned long flags; 6801 6802 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state)) 6803 return; 6804 6805 spin_lock_irqsave(&napi_hash_lock, flags); 6806 6807 /* 0..NR_CPUS range is reserved for sender_cpu use */ 6808 do { 6809 if (unlikely(++napi_gen_id < MIN_NAPI_ID)) 6810 napi_gen_id = MIN_NAPI_ID; 6811 } while (napi_by_id(napi_gen_id)); 6812 6813 __napi_hash_add_with_id(napi, napi_gen_id); 6814 6815 spin_unlock_irqrestore(&napi_hash_lock, flags); 6816 } 6817 6818 /* Warning : caller is responsible to make sure rcu grace period 6819 * is respected before freeing memory containing @napi 6820 */ 6821 static void napi_hash_del(struct napi_struct *napi) 6822 { 6823 unsigned long flags; 6824 6825 spin_lock_irqsave(&napi_hash_lock, flags); 6826 6827 hlist_del_init_rcu(&napi->napi_hash_node); 6828 6829 spin_unlock_irqrestore(&napi_hash_lock, flags); 6830 } 6831 6832 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer) 6833 { 6834 struct napi_struct *napi; 6835 6836 napi = container_of(timer, struct napi_struct, timer); 6837 6838 /* Note : we use a relaxed variant of napi_schedule_prep() not setting 6839 * NAPI_STATE_MISSED, since we do not react to a device IRQ. 6840 */ 6841 if (!napi_disable_pending(napi) && 6842 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) { 6843 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6844 __napi_schedule_irqoff(napi); 6845 } 6846 6847 return HRTIMER_NORESTART; 6848 } 6849 6850 static void init_gro_hash(struct napi_struct *napi) 6851 { 6852 int i; 6853 6854 for (i = 0; i < GRO_HASH_BUCKETS; i++) { 6855 INIT_LIST_HEAD(&napi->gro_hash[i].list); 6856 napi->gro_hash[i].count = 0; 6857 } 6858 napi->gro_bitmask = 0; 6859 } 6860 6861 int dev_set_threaded(struct net_device *dev, bool threaded) 6862 { 6863 struct napi_struct *napi; 6864 int err = 0; 6865 6866 netdev_assert_locked_or_invisible(dev); 6867 6868 if (dev->threaded == threaded) 6869 return 0; 6870 6871 if (threaded) { 6872 list_for_each_entry(napi, &dev->napi_list, dev_list) { 6873 if (!napi->thread) { 6874 err = napi_kthread_create(napi); 6875 if (err) { 6876 threaded = false; 6877 break; 6878 } 6879 } 6880 } 6881 } 6882 6883 WRITE_ONCE(dev->threaded, threaded); 6884 6885 /* Make sure kthread is created before THREADED bit 6886 * is set. 6887 */ 6888 smp_mb__before_atomic(); 6889 6890 /* Setting/unsetting threaded mode on a napi might not immediately 6891 * take effect, if the current napi instance is actively being 6892 * polled. In this case, the switch between threaded mode and 6893 * softirq mode will happen in the next round of napi_schedule(). 6894 * This should not cause hiccups/stalls to the live traffic. 6895 */ 6896 list_for_each_entry(napi, &dev->napi_list, dev_list) 6897 assign_bit(NAPI_STATE_THREADED, &napi->state, threaded); 6898 6899 return err; 6900 } 6901 EXPORT_SYMBOL(dev_set_threaded); 6902 6903 /** 6904 * netif_queue_set_napi - Associate queue with the napi 6905 * @dev: device to which NAPI and queue belong 6906 * @queue_index: Index of queue 6907 * @type: queue type as RX or TX 6908 * @napi: NAPI context, pass NULL to clear previously set NAPI 6909 * 6910 * Set queue with its corresponding napi context. This should be done after 6911 * registering the NAPI handler for the queue-vector and the queues have been 6912 * mapped to the corresponding interrupt vector. 6913 */ 6914 void netif_queue_set_napi(struct net_device *dev, unsigned int queue_index, 6915 enum netdev_queue_type type, struct napi_struct *napi) 6916 { 6917 struct netdev_rx_queue *rxq; 6918 struct netdev_queue *txq; 6919 6920 if (WARN_ON_ONCE(napi && !napi->dev)) 6921 return; 6922 if (dev->reg_state >= NETREG_REGISTERED) 6923 ASSERT_RTNL(); 6924 6925 switch (type) { 6926 case NETDEV_QUEUE_TYPE_RX: 6927 rxq = __netif_get_rx_queue(dev, queue_index); 6928 rxq->napi = napi; 6929 return; 6930 case NETDEV_QUEUE_TYPE_TX: 6931 txq = netdev_get_tx_queue(dev, queue_index); 6932 txq->napi = napi; 6933 return; 6934 default: 6935 return; 6936 } 6937 } 6938 EXPORT_SYMBOL(netif_queue_set_napi); 6939 6940 static void napi_restore_config(struct napi_struct *n) 6941 { 6942 n->defer_hard_irqs = n->config->defer_hard_irqs; 6943 n->gro_flush_timeout = n->config->gro_flush_timeout; 6944 n->irq_suspend_timeout = n->config->irq_suspend_timeout; 6945 /* a NAPI ID might be stored in the config, if so use it. if not, use 6946 * napi_hash_add to generate one for us. 6947 */ 6948 if (n->config->napi_id) { 6949 napi_hash_add_with_id(n, n->config->napi_id); 6950 } else { 6951 napi_hash_add(n); 6952 n->config->napi_id = n->napi_id; 6953 } 6954 } 6955 6956 static void napi_save_config(struct napi_struct *n) 6957 { 6958 n->config->defer_hard_irqs = n->defer_hard_irqs; 6959 n->config->gro_flush_timeout = n->gro_flush_timeout; 6960 n->config->irq_suspend_timeout = n->irq_suspend_timeout; 6961 napi_hash_del(n); 6962 } 6963 6964 /* Netlink wants the NAPI list to be sorted by ID, if adding a NAPI which will 6965 * inherit an existing ID try to insert it at the right position. 6966 */ 6967 static void 6968 netif_napi_dev_list_add(struct net_device *dev, struct napi_struct *napi) 6969 { 6970 unsigned int new_id, pos_id; 6971 struct list_head *higher; 6972 struct napi_struct *pos; 6973 6974 new_id = UINT_MAX; 6975 if (napi->config && napi->config->napi_id) 6976 new_id = napi->config->napi_id; 6977 6978 higher = &dev->napi_list; 6979 list_for_each_entry(pos, &dev->napi_list, dev_list) { 6980 if (pos->napi_id >= MIN_NAPI_ID) 6981 pos_id = pos->napi_id; 6982 else if (pos->config) 6983 pos_id = pos->config->napi_id; 6984 else 6985 pos_id = UINT_MAX; 6986 6987 if (pos_id <= new_id) 6988 break; 6989 higher = &pos->dev_list; 6990 } 6991 list_add_rcu(&napi->dev_list, higher); /* adds after higher */ 6992 } 6993 6994 /* Double check that napi_get_frags() allocates skbs with 6995 * skb->head being backed by slab, not a page fragment. 6996 * This is to make sure bug fixed in 3226b158e67c 6997 * ("net: avoid 32 x truesize under-estimation for tiny skbs") 6998 * does not accidentally come back. 6999 */ 7000 static void napi_get_frags_check(struct napi_struct *napi) 7001 { 7002 struct sk_buff *skb; 7003 7004 local_bh_disable(); 7005 skb = napi_get_frags(napi); 7006 WARN_ON_ONCE(skb && skb->head_frag); 7007 napi_free_frags(napi); 7008 local_bh_enable(); 7009 } 7010 7011 void netif_napi_add_weight_locked(struct net_device *dev, 7012 struct napi_struct *napi, 7013 int (*poll)(struct napi_struct *, int), 7014 int weight) 7015 { 7016 netdev_assert_locked(dev); 7017 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state))) 7018 return; 7019 7020 INIT_LIST_HEAD(&napi->poll_list); 7021 INIT_HLIST_NODE(&napi->napi_hash_node); 7022 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); 7023 napi->timer.function = napi_watchdog; 7024 init_gro_hash(napi); 7025 napi->skb = NULL; 7026 INIT_LIST_HEAD(&napi->rx_list); 7027 napi->rx_count = 0; 7028 napi->poll = poll; 7029 if (weight > NAPI_POLL_WEIGHT) 7030 netdev_err_once(dev, "%s() called with weight %d\n", __func__, 7031 weight); 7032 napi->weight = weight; 7033 napi->dev = dev; 7034 #ifdef CONFIG_NETPOLL 7035 napi->poll_owner = -1; 7036 #endif 7037 napi->list_owner = -1; 7038 set_bit(NAPI_STATE_SCHED, &napi->state); 7039 set_bit(NAPI_STATE_NPSVC, &napi->state); 7040 netif_napi_dev_list_add(dev, napi); 7041 7042 /* default settings from sysfs are applied to all NAPIs. any per-NAPI 7043 * configuration will be loaded in napi_enable 7044 */ 7045 napi_set_defer_hard_irqs(napi, READ_ONCE(dev->napi_defer_hard_irqs)); 7046 napi_set_gro_flush_timeout(napi, READ_ONCE(dev->gro_flush_timeout)); 7047 7048 napi_get_frags_check(napi); 7049 /* Create kthread for this napi if dev->threaded is set. 7050 * Clear dev->threaded if kthread creation failed so that 7051 * threaded mode will not be enabled in napi_enable(). 7052 */ 7053 if (dev->threaded && napi_kthread_create(napi)) 7054 dev->threaded = false; 7055 netif_napi_set_irq_locked(napi, -1); 7056 } 7057 EXPORT_SYMBOL(netif_napi_add_weight_locked); 7058 7059 void napi_disable_locked(struct napi_struct *n) 7060 { 7061 unsigned long val, new; 7062 7063 might_sleep(); 7064 netdev_assert_locked(n->dev); 7065 7066 set_bit(NAPI_STATE_DISABLE, &n->state); 7067 7068 val = READ_ONCE(n->state); 7069 do { 7070 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) { 7071 usleep_range(20, 200); 7072 val = READ_ONCE(n->state); 7073 } 7074 7075 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC; 7076 new &= ~(NAPIF_STATE_THREADED | NAPIF_STATE_PREFER_BUSY_POLL); 7077 } while (!try_cmpxchg(&n->state, &val, new)); 7078 7079 hrtimer_cancel(&n->timer); 7080 7081 if (n->config) 7082 napi_save_config(n); 7083 else 7084 napi_hash_del(n); 7085 7086 clear_bit(NAPI_STATE_DISABLE, &n->state); 7087 } 7088 EXPORT_SYMBOL(napi_disable_locked); 7089 7090 /** 7091 * napi_disable() - prevent NAPI from scheduling 7092 * @n: NAPI context 7093 * 7094 * Stop NAPI from being scheduled on this context. 7095 * Waits till any outstanding processing completes. 7096 * Takes netdev_lock() for associated net_device. 7097 */ 7098 void napi_disable(struct napi_struct *n) 7099 { 7100 netdev_lock(n->dev); 7101 napi_disable_locked(n); 7102 netdev_unlock(n->dev); 7103 } 7104 EXPORT_SYMBOL(napi_disable); 7105 7106 void napi_enable_locked(struct napi_struct *n) 7107 { 7108 unsigned long new, val = READ_ONCE(n->state); 7109 7110 if (n->config) 7111 napi_restore_config(n); 7112 else 7113 napi_hash_add(n); 7114 7115 do { 7116 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val)); 7117 7118 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC); 7119 if (n->dev->threaded && n->thread) 7120 new |= NAPIF_STATE_THREADED; 7121 } while (!try_cmpxchg(&n->state, &val, new)); 7122 } 7123 EXPORT_SYMBOL(napi_enable_locked); 7124 7125 /** 7126 * napi_enable() - enable NAPI scheduling 7127 * @n: NAPI context 7128 * 7129 * Enable scheduling of a NAPI instance. 7130 * Must be paired with napi_disable(). 7131 * Takes netdev_lock() for associated net_device. 7132 */ 7133 void napi_enable(struct napi_struct *n) 7134 { 7135 netdev_lock(n->dev); 7136 napi_enable_locked(n); 7137 netdev_unlock(n->dev); 7138 } 7139 EXPORT_SYMBOL(napi_enable); 7140 7141 static void flush_gro_hash(struct napi_struct *napi) 7142 { 7143 int i; 7144 7145 for (i = 0; i < GRO_HASH_BUCKETS; i++) { 7146 struct sk_buff *skb, *n; 7147 7148 list_for_each_entry_safe(skb, n, &napi->gro_hash[i].list, list) 7149 kfree_skb(skb); 7150 napi->gro_hash[i].count = 0; 7151 } 7152 } 7153 7154 /* Must be called in process context */ 7155 void __netif_napi_del_locked(struct napi_struct *napi) 7156 { 7157 netdev_assert_locked(napi->dev); 7158 7159 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state)) 7160 return; 7161 7162 if (napi->config) { 7163 napi->index = -1; 7164 napi->config = NULL; 7165 } 7166 7167 list_del_rcu(&napi->dev_list); 7168 napi_free_frags(napi); 7169 7170 flush_gro_hash(napi); 7171 napi->gro_bitmask = 0; 7172 7173 if (napi->thread) { 7174 kthread_stop(napi->thread); 7175 napi->thread = NULL; 7176 } 7177 } 7178 EXPORT_SYMBOL(__netif_napi_del_locked); 7179 7180 static int __napi_poll(struct napi_struct *n, bool *repoll) 7181 { 7182 int work, weight; 7183 7184 weight = n->weight; 7185 7186 /* This NAPI_STATE_SCHED test is for avoiding a race 7187 * with netpoll's poll_napi(). Only the entity which 7188 * obtains the lock and sees NAPI_STATE_SCHED set will 7189 * actually make the ->poll() call. Therefore we avoid 7190 * accidentally calling ->poll() when NAPI is not scheduled. 7191 */ 7192 work = 0; 7193 if (napi_is_scheduled(n)) { 7194 work = n->poll(n, weight); 7195 trace_napi_poll(n, work, weight); 7196 7197 xdp_do_check_flushed(n); 7198 } 7199 7200 if (unlikely(work > weight)) 7201 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n", 7202 n->poll, work, weight); 7203 7204 if (likely(work < weight)) 7205 return work; 7206 7207 /* Drivers must not modify the NAPI state if they 7208 * consume the entire weight. In such cases this code 7209 * still "owns" the NAPI instance and therefore can 7210 * move the instance around on the list at-will. 7211 */ 7212 if (unlikely(napi_disable_pending(n))) { 7213 napi_complete(n); 7214 return work; 7215 } 7216 7217 /* The NAPI context has more processing work, but busy-polling 7218 * is preferred. Exit early. 7219 */ 7220 if (napi_prefer_busy_poll(n)) { 7221 if (napi_complete_done(n, work)) { 7222 /* If timeout is not set, we need to make sure 7223 * that the NAPI is re-scheduled. 7224 */ 7225 napi_schedule(n); 7226 } 7227 return work; 7228 } 7229 7230 if (n->gro_bitmask) { 7231 /* flush too old packets 7232 * If HZ < 1000, flush all packets. 7233 */ 7234 napi_gro_flush(n, HZ >= 1000); 7235 } 7236 7237 gro_normal_list(n); 7238 7239 /* Some drivers may have called napi_schedule 7240 * prior to exhausting their budget. 7241 */ 7242 if (unlikely(!list_empty(&n->poll_list))) { 7243 pr_warn_once("%s: Budget exhausted after napi rescheduled\n", 7244 n->dev ? n->dev->name : "backlog"); 7245 return work; 7246 } 7247 7248 *repoll = true; 7249 7250 return work; 7251 } 7252 7253 static int napi_poll(struct napi_struct *n, struct list_head *repoll) 7254 { 7255 bool do_repoll = false; 7256 void *have; 7257 int work; 7258 7259 list_del_init(&n->poll_list); 7260 7261 have = netpoll_poll_lock(n); 7262 7263 work = __napi_poll(n, &do_repoll); 7264 7265 if (do_repoll) 7266 list_add_tail(&n->poll_list, repoll); 7267 7268 netpoll_poll_unlock(have); 7269 7270 return work; 7271 } 7272 7273 static int napi_thread_wait(struct napi_struct *napi) 7274 { 7275 set_current_state(TASK_INTERRUPTIBLE); 7276 7277 while (!kthread_should_stop()) { 7278 /* Testing SCHED_THREADED bit here to make sure the current 7279 * kthread owns this napi and could poll on this napi. 7280 * Testing SCHED bit is not enough because SCHED bit might be 7281 * set by some other busy poll thread or by napi_disable(). 7282 */ 7283 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) { 7284 WARN_ON(!list_empty(&napi->poll_list)); 7285 __set_current_state(TASK_RUNNING); 7286 return 0; 7287 } 7288 7289 schedule(); 7290 set_current_state(TASK_INTERRUPTIBLE); 7291 } 7292 __set_current_state(TASK_RUNNING); 7293 7294 return -1; 7295 } 7296 7297 static void napi_threaded_poll_loop(struct napi_struct *napi) 7298 { 7299 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7300 struct softnet_data *sd; 7301 unsigned long last_qs = jiffies; 7302 7303 for (;;) { 7304 bool repoll = false; 7305 void *have; 7306 7307 local_bh_disable(); 7308 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7309 7310 sd = this_cpu_ptr(&softnet_data); 7311 sd->in_napi_threaded_poll = true; 7312 7313 have = netpoll_poll_lock(napi); 7314 __napi_poll(napi, &repoll); 7315 netpoll_poll_unlock(have); 7316 7317 sd->in_napi_threaded_poll = false; 7318 barrier(); 7319 7320 if (sd_has_rps_ipi_waiting(sd)) { 7321 local_irq_disable(); 7322 net_rps_action_and_irq_enable(sd); 7323 } 7324 skb_defer_free_flush(sd); 7325 bpf_net_ctx_clear(bpf_net_ctx); 7326 local_bh_enable(); 7327 7328 if (!repoll) 7329 break; 7330 7331 rcu_softirq_qs_periodic(last_qs); 7332 cond_resched(); 7333 } 7334 } 7335 7336 static int napi_threaded_poll(void *data) 7337 { 7338 struct napi_struct *napi = data; 7339 7340 while (!napi_thread_wait(napi)) 7341 napi_threaded_poll_loop(napi); 7342 7343 return 0; 7344 } 7345 7346 static __latent_entropy void net_rx_action(void) 7347 { 7348 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 7349 unsigned long time_limit = jiffies + 7350 usecs_to_jiffies(READ_ONCE(net_hotdata.netdev_budget_usecs)); 7351 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7352 int budget = READ_ONCE(net_hotdata.netdev_budget); 7353 LIST_HEAD(list); 7354 LIST_HEAD(repoll); 7355 7356 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7357 start: 7358 sd->in_net_rx_action = true; 7359 local_irq_disable(); 7360 list_splice_init(&sd->poll_list, &list); 7361 local_irq_enable(); 7362 7363 for (;;) { 7364 struct napi_struct *n; 7365 7366 skb_defer_free_flush(sd); 7367 7368 if (list_empty(&list)) { 7369 if (list_empty(&repoll)) { 7370 sd->in_net_rx_action = false; 7371 barrier(); 7372 /* We need to check if ____napi_schedule() 7373 * had refilled poll_list while 7374 * sd->in_net_rx_action was true. 7375 */ 7376 if (!list_empty(&sd->poll_list)) 7377 goto start; 7378 if (!sd_has_rps_ipi_waiting(sd)) 7379 goto end; 7380 } 7381 break; 7382 } 7383 7384 n = list_first_entry(&list, struct napi_struct, poll_list); 7385 budget -= napi_poll(n, &repoll); 7386 7387 /* If softirq window is exhausted then punt. 7388 * Allow this to run for 2 jiffies since which will allow 7389 * an average latency of 1.5/HZ. 7390 */ 7391 if (unlikely(budget <= 0 || 7392 time_after_eq(jiffies, time_limit))) { 7393 sd->time_squeeze++; 7394 break; 7395 } 7396 } 7397 7398 local_irq_disable(); 7399 7400 list_splice_tail_init(&sd->poll_list, &list); 7401 list_splice_tail(&repoll, &list); 7402 list_splice(&list, &sd->poll_list); 7403 if (!list_empty(&sd->poll_list)) 7404 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 7405 else 7406 sd->in_net_rx_action = false; 7407 7408 net_rps_action_and_irq_enable(sd); 7409 end: 7410 bpf_net_ctx_clear(bpf_net_ctx); 7411 } 7412 7413 struct netdev_adjacent { 7414 struct net_device *dev; 7415 netdevice_tracker dev_tracker; 7416 7417 /* upper master flag, there can only be one master device per list */ 7418 bool master; 7419 7420 /* lookup ignore flag */ 7421 bool ignore; 7422 7423 /* counter for the number of times this device was added to us */ 7424 u16 ref_nr; 7425 7426 /* private field for the users */ 7427 void *private; 7428 7429 struct list_head list; 7430 struct rcu_head rcu; 7431 }; 7432 7433 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev, 7434 struct list_head *adj_list) 7435 { 7436 struct netdev_adjacent *adj; 7437 7438 list_for_each_entry(adj, adj_list, list) { 7439 if (adj->dev == adj_dev) 7440 return adj; 7441 } 7442 return NULL; 7443 } 7444 7445 static int ____netdev_has_upper_dev(struct net_device *upper_dev, 7446 struct netdev_nested_priv *priv) 7447 { 7448 struct net_device *dev = (struct net_device *)priv->data; 7449 7450 return upper_dev == dev; 7451 } 7452 7453 /** 7454 * netdev_has_upper_dev - Check if device is linked to an upper device 7455 * @dev: device 7456 * @upper_dev: upper device to check 7457 * 7458 * Find out if a device is linked to specified upper device and return true 7459 * in case it is. Note that this checks only immediate upper device, 7460 * not through a complete stack of devices. The caller must hold the RTNL lock. 7461 */ 7462 bool netdev_has_upper_dev(struct net_device *dev, 7463 struct net_device *upper_dev) 7464 { 7465 struct netdev_nested_priv priv = { 7466 .data = (void *)upper_dev, 7467 }; 7468 7469 ASSERT_RTNL(); 7470 7471 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7472 &priv); 7473 } 7474 EXPORT_SYMBOL(netdev_has_upper_dev); 7475 7476 /** 7477 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device 7478 * @dev: device 7479 * @upper_dev: upper device to check 7480 * 7481 * Find out if a device is linked to specified upper device and return true 7482 * in case it is. Note that this checks the entire upper device chain. 7483 * The caller must hold rcu lock. 7484 */ 7485 7486 bool netdev_has_upper_dev_all_rcu(struct net_device *dev, 7487 struct net_device *upper_dev) 7488 { 7489 struct netdev_nested_priv priv = { 7490 .data = (void *)upper_dev, 7491 }; 7492 7493 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7494 &priv); 7495 } 7496 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu); 7497 7498 /** 7499 * netdev_has_any_upper_dev - Check if device is linked to some device 7500 * @dev: device 7501 * 7502 * Find out if a device is linked to an upper device and return true in case 7503 * it is. The caller must hold the RTNL lock. 7504 */ 7505 bool netdev_has_any_upper_dev(struct net_device *dev) 7506 { 7507 ASSERT_RTNL(); 7508 7509 return !list_empty(&dev->adj_list.upper); 7510 } 7511 EXPORT_SYMBOL(netdev_has_any_upper_dev); 7512 7513 /** 7514 * netdev_master_upper_dev_get - Get master upper device 7515 * @dev: device 7516 * 7517 * Find a master upper device and return pointer to it or NULL in case 7518 * it's not there. The caller must hold the RTNL lock. 7519 */ 7520 struct net_device *netdev_master_upper_dev_get(struct net_device *dev) 7521 { 7522 struct netdev_adjacent *upper; 7523 7524 ASSERT_RTNL(); 7525 7526 if (list_empty(&dev->adj_list.upper)) 7527 return NULL; 7528 7529 upper = list_first_entry(&dev->adj_list.upper, 7530 struct netdev_adjacent, list); 7531 if (likely(upper->master)) 7532 return upper->dev; 7533 return NULL; 7534 } 7535 EXPORT_SYMBOL(netdev_master_upper_dev_get); 7536 7537 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev) 7538 { 7539 struct netdev_adjacent *upper; 7540 7541 ASSERT_RTNL(); 7542 7543 if (list_empty(&dev->adj_list.upper)) 7544 return NULL; 7545 7546 upper = list_first_entry(&dev->adj_list.upper, 7547 struct netdev_adjacent, list); 7548 if (likely(upper->master) && !upper->ignore) 7549 return upper->dev; 7550 return NULL; 7551 } 7552 7553 /** 7554 * netdev_has_any_lower_dev - Check if device is linked to some device 7555 * @dev: device 7556 * 7557 * Find out if a device is linked to a lower device and return true in case 7558 * it is. The caller must hold the RTNL lock. 7559 */ 7560 static bool netdev_has_any_lower_dev(struct net_device *dev) 7561 { 7562 ASSERT_RTNL(); 7563 7564 return !list_empty(&dev->adj_list.lower); 7565 } 7566 7567 void *netdev_adjacent_get_private(struct list_head *adj_list) 7568 { 7569 struct netdev_adjacent *adj; 7570 7571 adj = list_entry(adj_list, struct netdev_adjacent, list); 7572 7573 return adj->private; 7574 } 7575 EXPORT_SYMBOL(netdev_adjacent_get_private); 7576 7577 /** 7578 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list 7579 * @dev: device 7580 * @iter: list_head ** of the current position 7581 * 7582 * Gets the next device from the dev's upper list, starting from iter 7583 * position. The caller must hold RCU read lock. 7584 */ 7585 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev, 7586 struct list_head **iter) 7587 { 7588 struct netdev_adjacent *upper; 7589 7590 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7591 7592 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7593 7594 if (&upper->list == &dev->adj_list.upper) 7595 return NULL; 7596 7597 *iter = &upper->list; 7598 7599 return upper->dev; 7600 } 7601 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu); 7602 7603 static struct net_device *__netdev_next_upper_dev(struct net_device *dev, 7604 struct list_head **iter, 7605 bool *ignore) 7606 { 7607 struct netdev_adjacent *upper; 7608 7609 upper = list_entry((*iter)->next, struct netdev_adjacent, list); 7610 7611 if (&upper->list == &dev->adj_list.upper) 7612 return NULL; 7613 7614 *iter = &upper->list; 7615 *ignore = upper->ignore; 7616 7617 return upper->dev; 7618 } 7619 7620 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev, 7621 struct list_head **iter) 7622 { 7623 struct netdev_adjacent *upper; 7624 7625 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7626 7627 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7628 7629 if (&upper->list == &dev->adj_list.upper) 7630 return NULL; 7631 7632 *iter = &upper->list; 7633 7634 return upper->dev; 7635 } 7636 7637 static int __netdev_walk_all_upper_dev(struct net_device *dev, 7638 int (*fn)(struct net_device *dev, 7639 struct netdev_nested_priv *priv), 7640 struct netdev_nested_priv *priv) 7641 { 7642 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7643 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7644 int ret, cur = 0; 7645 bool ignore; 7646 7647 now = dev; 7648 iter = &dev->adj_list.upper; 7649 7650 while (1) { 7651 if (now != dev) { 7652 ret = fn(now, priv); 7653 if (ret) 7654 return ret; 7655 } 7656 7657 next = NULL; 7658 while (1) { 7659 udev = __netdev_next_upper_dev(now, &iter, &ignore); 7660 if (!udev) 7661 break; 7662 if (ignore) 7663 continue; 7664 7665 next = udev; 7666 niter = &udev->adj_list.upper; 7667 dev_stack[cur] = now; 7668 iter_stack[cur++] = iter; 7669 break; 7670 } 7671 7672 if (!next) { 7673 if (!cur) 7674 return 0; 7675 next = dev_stack[--cur]; 7676 niter = iter_stack[cur]; 7677 } 7678 7679 now = next; 7680 iter = niter; 7681 } 7682 7683 return 0; 7684 } 7685 7686 int netdev_walk_all_upper_dev_rcu(struct net_device *dev, 7687 int (*fn)(struct net_device *dev, 7688 struct netdev_nested_priv *priv), 7689 struct netdev_nested_priv *priv) 7690 { 7691 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7692 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7693 int ret, cur = 0; 7694 7695 now = dev; 7696 iter = &dev->adj_list.upper; 7697 7698 while (1) { 7699 if (now != dev) { 7700 ret = fn(now, priv); 7701 if (ret) 7702 return ret; 7703 } 7704 7705 next = NULL; 7706 while (1) { 7707 udev = netdev_next_upper_dev_rcu(now, &iter); 7708 if (!udev) 7709 break; 7710 7711 next = udev; 7712 niter = &udev->adj_list.upper; 7713 dev_stack[cur] = now; 7714 iter_stack[cur++] = iter; 7715 break; 7716 } 7717 7718 if (!next) { 7719 if (!cur) 7720 return 0; 7721 next = dev_stack[--cur]; 7722 niter = iter_stack[cur]; 7723 } 7724 7725 now = next; 7726 iter = niter; 7727 } 7728 7729 return 0; 7730 } 7731 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu); 7732 7733 static bool __netdev_has_upper_dev(struct net_device *dev, 7734 struct net_device *upper_dev) 7735 { 7736 struct netdev_nested_priv priv = { 7737 .flags = 0, 7738 .data = (void *)upper_dev, 7739 }; 7740 7741 ASSERT_RTNL(); 7742 7743 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev, 7744 &priv); 7745 } 7746 7747 /** 7748 * netdev_lower_get_next_private - Get the next ->private from the 7749 * lower neighbour list 7750 * @dev: device 7751 * @iter: list_head ** of the current position 7752 * 7753 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7754 * list, starting from iter position. The caller must hold either hold the 7755 * RTNL lock or its own locking that guarantees that the neighbour lower 7756 * list will remain unchanged. 7757 */ 7758 void *netdev_lower_get_next_private(struct net_device *dev, 7759 struct list_head **iter) 7760 { 7761 struct netdev_adjacent *lower; 7762 7763 lower = list_entry(*iter, struct netdev_adjacent, list); 7764 7765 if (&lower->list == &dev->adj_list.lower) 7766 return NULL; 7767 7768 *iter = lower->list.next; 7769 7770 return lower->private; 7771 } 7772 EXPORT_SYMBOL(netdev_lower_get_next_private); 7773 7774 /** 7775 * netdev_lower_get_next_private_rcu - Get the next ->private from the 7776 * lower neighbour list, RCU 7777 * variant 7778 * @dev: device 7779 * @iter: list_head ** of the current position 7780 * 7781 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7782 * list, starting from iter position. The caller must hold RCU read lock. 7783 */ 7784 void *netdev_lower_get_next_private_rcu(struct net_device *dev, 7785 struct list_head **iter) 7786 { 7787 struct netdev_adjacent *lower; 7788 7789 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 7790 7791 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7792 7793 if (&lower->list == &dev->adj_list.lower) 7794 return NULL; 7795 7796 *iter = &lower->list; 7797 7798 return lower->private; 7799 } 7800 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu); 7801 7802 /** 7803 * netdev_lower_get_next - Get the next device from the lower neighbour 7804 * list 7805 * @dev: device 7806 * @iter: list_head ** of the current position 7807 * 7808 * Gets the next netdev_adjacent from the dev's lower neighbour 7809 * list, starting from iter position. The caller must hold RTNL lock or 7810 * its own locking that guarantees that the neighbour lower 7811 * list will remain unchanged. 7812 */ 7813 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter) 7814 { 7815 struct netdev_adjacent *lower; 7816 7817 lower = list_entry(*iter, struct netdev_adjacent, list); 7818 7819 if (&lower->list == &dev->adj_list.lower) 7820 return NULL; 7821 7822 *iter = lower->list.next; 7823 7824 return lower->dev; 7825 } 7826 EXPORT_SYMBOL(netdev_lower_get_next); 7827 7828 static struct net_device *netdev_next_lower_dev(struct net_device *dev, 7829 struct list_head **iter) 7830 { 7831 struct netdev_adjacent *lower; 7832 7833 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7834 7835 if (&lower->list == &dev->adj_list.lower) 7836 return NULL; 7837 7838 *iter = &lower->list; 7839 7840 return lower->dev; 7841 } 7842 7843 static struct net_device *__netdev_next_lower_dev(struct net_device *dev, 7844 struct list_head **iter, 7845 bool *ignore) 7846 { 7847 struct netdev_adjacent *lower; 7848 7849 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7850 7851 if (&lower->list == &dev->adj_list.lower) 7852 return NULL; 7853 7854 *iter = &lower->list; 7855 *ignore = lower->ignore; 7856 7857 return lower->dev; 7858 } 7859 7860 int netdev_walk_all_lower_dev(struct net_device *dev, 7861 int (*fn)(struct net_device *dev, 7862 struct netdev_nested_priv *priv), 7863 struct netdev_nested_priv *priv) 7864 { 7865 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7866 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7867 int ret, cur = 0; 7868 7869 now = dev; 7870 iter = &dev->adj_list.lower; 7871 7872 while (1) { 7873 if (now != dev) { 7874 ret = fn(now, priv); 7875 if (ret) 7876 return ret; 7877 } 7878 7879 next = NULL; 7880 while (1) { 7881 ldev = netdev_next_lower_dev(now, &iter); 7882 if (!ldev) 7883 break; 7884 7885 next = ldev; 7886 niter = &ldev->adj_list.lower; 7887 dev_stack[cur] = now; 7888 iter_stack[cur++] = iter; 7889 break; 7890 } 7891 7892 if (!next) { 7893 if (!cur) 7894 return 0; 7895 next = dev_stack[--cur]; 7896 niter = iter_stack[cur]; 7897 } 7898 7899 now = next; 7900 iter = niter; 7901 } 7902 7903 return 0; 7904 } 7905 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev); 7906 7907 static int __netdev_walk_all_lower_dev(struct net_device *dev, 7908 int (*fn)(struct net_device *dev, 7909 struct netdev_nested_priv *priv), 7910 struct netdev_nested_priv *priv) 7911 { 7912 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7913 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7914 int ret, cur = 0; 7915 bool ignore; 7916 7917 now = dev; 7918 iter = &dev->adj_list.lower; 7919 7920 while (1) { 7921 if (now != dev) { 7922 ret = fn(now, priv); 7923 if (ret) 7924 return ret; 7925 } 7926 7927 next = NULL; 7928 while (1) { 7929 ldev = __netdev_next_lower_dev(now, &iter, &ignore); 7930 if (!ldev) 7931 break; 7932 if (ignore) 7933 continue; 7934 7935 next = ldev; 7936 niter = &ldev->adj_list.lower; 7937 dev_stack[cur] = now; 7938 iter_stack[cur++] = iter; 7939 break; 7940 } 7941 7942 if (!next) { 7943 if (!cur) 7944 return 0; 7945 next = dev_stack[--cur]; 7946 niter = iter_stack[cur]; 7947 } 7948 7949 now = next; 7950 iter = niter; 7951 } 7952 7953 return 0; 7954 } 7955 7956 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev, 7957 struct list_head **iter) 7958 { 7959 struct netdev_adjacent *lower; 7960 7961 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7962 if (&lower->list == &dev->adj_list.lower) 7963 return NULL; 7964 7965 *iter = &lower->list; 7966 7967 return lower->dev; 7968 } 7969 EXPORT_SYMBOL(netdev_next_lower_dev_rcu); 7970 7971 static u8 __netdev_upper_depth(struct net_device *dev) 7972 { 7973 struct net_device *udev; 7974 struct list_head *iter; 7975 u8 max_depth = 0; 7976 bool ignore; 7977 7978 for (iter = &dev->adj_list.upper, 7979 udev = __netdev_next_upper_dev(dev, &iter, &ignore); 7980 udev; 7981 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) { 7982 if (ignore) 7983 continue; 7984 if (max_depth < udev->upper_level) 7985 max_depth = udev->upper_level; 7986 } 7987 7988 return max_depth; 7989 } 7990 7991 static u8 __netdev_lower_depth(struct net_device *dev) 7992 { 7993 struct net_device *ldev; 7994 struct list_head *iter; 7995 u8 max_depth = 0; 7996 bool ignore; 7997 7998 for (iter = &dev->adj_list.lower, 7999 ldev = __netdev_next_lower_dev(dev, &iter, &ignore); 8000 ldev; 8001 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) { 8002 if (ignore) 8003 continue; 8004 if (max_depth < ldev->lower_level) 8005 max_depth = ldev->lower_level; 8006 } 8007 8008 return max_depth; 8009 } 8010 8011 static int __netdev_update_upper_level(struct net_device *dev, 8012 struct netdev_nested_priv *__unused) 8013 { 8014 dev->upper_level = __netdev_upper_depth(dev) + 1; 8015 return 0; 8016 } 8017 8018 #ifdef CONFIG_LOCKDEP 8019 static LIST_HEAD(net_unlink_list); 8020 8021 static void net_unlink_todo(struct net_device *dev) 8022 { 8023 if (list_empty(&dev->unlink_list)) 8024 list_add_tail(&dev->unlink_list, &net_unlink_list); 8025 } 8026 #endif 8027 8028 static int __netdev_update_lower_level(struct net_device *dev, 8029 struct netdev_nested_priv *priv) 8030 { 8031 dev->lower_level = __netdev_lower_depth(dev) + 1; 8032 8033 #ifdef CONFIG_LOCKDEP 8034 if (!priv) 8035 return 0; 8036 8037 if (priv->flags & NESTED_SYNC_IMM) 8038 dev->nested_level = dev->lower_level - 1; 8039 if (priv->flags & NESTED_SYNC_TODO) 8040 net_unlink_todo(dev); 8041 #endif 8042 return 0; 8043 } 8044 8045 int netdev_walk_all_lower_dev_rcu(struct net_device *dev, 8046 int (*fn)(struct net_device *dev, 8047 struct netdev_nested_priv *priv), 8048 struct netdev_nested_priv *priv) 8049 { 8050 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8051 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8052 int ret, cur = 0; 8053 8054 now = dev; 8055 iter = &dev->adj_list.lower; 8056 8057 while (1) { 8058 if (now != dev) { 8059 ret = fn(now, priv); 8060 if (ret) 8061 return ret; 8062 } 8063 8064 next = NULL; 8065 while (1) { 8066 ldev = netdev_next_lower_dev_rcu(now, &iter); 8067 if (!ldev) 8068 break; 8069 8070 next = ldev; 8071 niter = &ldev->adj_list.lower; 8072 dev_stack[cur] = now; 8073 iter_stack[cur++] = iter; 8074 break; 8075 } 8076 8077 if (!next) { 8078 if (!cur) 8079 return 0; 8080 next = dev_stack[--cur]; 8081 niter = iter_stack[cur]; 8082 } 8083 8084 now = next; 8085 iter = niter; 8086 } 8087 8088 return 0; 8089 } 8090 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu); 8091 8092 /** 8093 * netdev_lower_get_first_private_rcu - Get the first ->private from the 8094 * lower neighbour list, RCU 8095 * variant 8096 * @dev: device 8097 * 8098 * Gets the first netdev_adjacent->private from the dev's lower neighbour 8099 * list. The caller must hold RCU read lock. 8100 */ 8101 void *netdev_lower_get_first_private_rcu(struct net_device *dev) 8102 { 8103 struct netdev_adjacent *lower; 8104 8105 lower = list_first_or_null_rcu(&dev->adj_list.lower, 8106 struct netdev_adjacent, list); 8107 if (lower) 8108 return lower->private; 8109 return NULL; 8110 } 8111 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu); 8112 8113 /** 8114 * netdev_master_upper_dev_get_rcu - Get master upper device 8115 * @dev: device 8116 * 8117 * Find a master upper device and return pointer to it or NULL in case 8118 * it's not there. The caller must hold the RCU read lock. 8119 */ 8120 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev) 8121 { 8122 struct netdev_adjacent *upper; 8123 8124 upper = list_first_or_null_rcu(&dev->adj_list.upper, 8125 struct netdev_adjacent, list); 8126 if (upper && likely(upper->master)) 8127 return upper->dev; 8128 return NULL; 8129 } 8130 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu); 8131 8132 static int netdev_adjacent_sysfs_add(struct net_device *dev, 8133 struct net_device *adj_dev, 8134 struct list_head *dev_list) 8135 { 8136 char linkname[IFNAMSIZ+7]; 8137 8138 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8139 "upper_%s" : "lower_%s", adj_dev->name); 8140 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj), 8141 linkname); 8142 } 8143 static void netdev_adjacent_sysfs_del(struct net_device *dev, 8144 char *name, 8145 struct list_head *dev_list) 8146 { 8147 char linkname[IFNAMSIZ+7]; 8148 8149 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8150 "upper_%s" : "lower_%s", name); 8151 sysfs_remove_link(&(dev->dev.kobj), linkname); 8152 } 8153 8154 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev, 8155 struct net_device *adj_dev, 8156 struct list_head *dev_list) 8157 { 8158 return (dev_list == &dev->adj_list.upper || 8159 dev_list == &dev->adj_list.lower) && 8160 net_eq(dev_net(dev), dev_net(adj_dev)); 8161 } 8162 8163 static int __netdev_adjacent_dev_insert(struct net_device *dev, 8164 struct net_device *adj_dev, 8165 struct list_head *dev_list, 8166 void *private, bool master) 8167 { 8168 struct netdev_adjacent *adj; 8169 int ret; 8170 8171 adj = __netdev_find_adj(adj_dev, dev_list); 8172 8173 if (adj) { 8174 adj->ref_nr += 1; 8175 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n", 8176 dev->name, adj_dev->name, adj->ref_nr); 8177 8178 return 0; 8179 } 8180 8181 adj = kmalloc(sizeof(*adj), GFP_KERNEL); 8182 if (!adj) 8183 return -ENOMEM; 8184 8185 adj->dev = adj_dev; 8186 adj->master = master; 8187 adj->ref_nr = 1; 8188 adj->private = private; 8189 adj->ignore = false; 8190 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL); 8191 8192 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n", 8193 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name); 8194 8195 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) { 8196 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list); 8197 if (ret) 8198 goto free_adj; 8199 } 8200 8201 /* Ensure that master link is always the first item in list. */ 8202 if (master) { 8203 ret = sysfs_create_link(&(dev->dev.kobj), 8204 &(adj_dev->dev.kobj), "master"); 8205 if (ret) 8206 goto remove_symlinks; 8207 8208 list_add_rcu(&adj->list, dev_list); 8209 } else { 8210 list_add_tail_rcu(&adj->list, dev_list); 8211 } 8212 8213 return 0; 8214 8215 remove_symlinks: 8216 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8217 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8218 free_adj: 8219 netdev_put(adj_dev, &adj->dev_tracker); 8220 kfree(adj); 8221 8222 return ret; 8223 } 8224 8225 static void __netdev_adjacent_dev_remove(struct net_device *dev, 8226 struct net_device *adj_dev, 8227 u16 ref_nr, 8228 struct list_head *dev_list) 8229 { 8230 struct netdev_adjacent *adj; 8231 8232 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n", 8233 dev->name, adj_dev->name, ref_nr); 8234 8235 adj = __netdev_find_adj(adj_dev, dev_list); 8236 8237 if (!adj) { 8238 pr_err("Adjacency does not exist for device %s from %s\n", 8239 dev->name, adj_dev->name); 8240 WARN_ON(1); 8241 return; 8242 } 8243 8244 if (adj->ref_nr > ref_nr) { 8245 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n", 8246 dev->name, adj_dev->name, ref_nr, 8247 adj->ref_nr - ref_nr); 8248 adj->ref_nr -= ref_nr; 8249 return; 8250 } 8251 8252 if (adj->master) 8253 sysfs_remove_link(&(dev->dev.kobj), "master"); 8254 8255 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8256 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8257 8258 list_del_rcu(&adj->list); 8259 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n", 8260 adj_dev->name, dev->name, adj_dev->name); 8261 netdev_put(adj_dev, &adj->dev_tracker); 8262 kfree_rcu(adj, rcu); 8263 } 8264 8265 static int __netdev_adjacent_dev_link_lists(struct net_device *dev, 8266 struct net_device *upper_dev, 8267 struct list_head *up_list, 8268 struct list_head *down_list, 8269 void *private, bool master) 8270 { 8271 int ret; 8272 8273 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list, 8274 private, master); 8275 if (ret) 8276 return ret; 8277 8278 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list, 8279 private, false); 8280 if (ret) { 8281 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list); 8282 return ret; 8283 } 8284 8285 return 0; 8286 } 8287 8288 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev, 8289 struct net_device *upper_dev, 8290 u16 ref_nr, 8291 struct list_head *up_list, 8292 struct list_head *down_list) 8293 { 8294 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list); 8295 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list); 8296 } 8297 8298 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev, 8299 struct net_device *upper_dev, 8300 void *private, bool master) 8301 { 8302 return __netdev_adjacent_dev_link_lists(dev, upper_dev, 8303 &dev->adj_list.upper, 8304 &upper_dev->adj_list.lower, 8305 private, master); 8306 } 8307 8308 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev, 8309 struct net_device *upper_dev) 8310 { 8311 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1, 8312 &dev->adj_list.upper, 8313 &upper_dev->adj_list.lower); 8314 } 8315 8316 static int __netdev_upper_dev_link(struct net_device *dev, 8317 struct net_device *upper_dev, bool master, 8318 void *upper_priv, void *upper_info, 8319 struct netdev_nested_priv *priv, 8320 struct netlink_ext_ack *extack) 8321 { 8322 struct netdev_notifier_changeupper_info changeupper_info = { 8323 .info = { 8324 .dev = dev, 8325 .extack = extack, 8326 }, 8327 .upper_dev = upper_dev, 8328 .master = master, 8329 .linking = true, 8330 .upper_info = upper_info, 8331 }; 8332 struct net_device *master_dev; 8333 int ret = 0; 8334 8335 ASSERT_RTNL(); 8336 8337 if (dev == upper_dev) 8338 return -EBUSY; 8339 8340 /* To prevent loops, check if dev is not upper device to upper_dev. */ 8341 if (__netdev_has_upper_dev(upper_dev, dev)) 8342 return -EBUSY; 8343 8344 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV) 8345 return -EMLINK; 8346 8347 if (!master) { 8348 if (__netdev_has_upper_dev(dev, upper_dev)) 8349 return -EEXIST; 8350 } else { 8351 master_dev = __netdev_master_upper_dev_get(dev); 8352 if (master_dev) 8353 return master_dev == upper_dev ? -EEXIST : -EBUSY; 8354 } 8355 8356 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8357 &changeupper_info.info); 8358 ret = notifier_to_errno(ret); 8359 if (ret) 8360 return ret; 8361 8362 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv, 8363 master); 8364 if (ret) 8365 return ret; 8366 8367 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8368 &changeupper_info.info); 8369 ret = notifier_to_errno(ret); 8370 if (ret) 8371 goto rollback; 8372 8373 __netdev_update_upper_level(dev, NULL); 8374 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8375 8376 __netdev_update_lower_level(upper_dev, priv); 8377 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8378 priv); 8379 8380 return 0; 8381 8382 rollback: 8383 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8384 8385 return ret; 8386 } 8387 8388 /** 8389 * netdev_upper_dev_link - Add a link to the upper device 8390 * @dev: device 8391 * @upper_dev: new upper device 8392 * @extack: netlink extended ack 8393 * 8394 * Adds a link to device which is upper to this one. The caller must hold 8395 * the RTNL lock. On a failure a negative errno code is returned. 8396 * On success the reference counts are adjusted and the function 8397 * returns zero. 8398 */ 8399 int netdev_upper_dev_link(struct net_device *dev, 8400 struct net_device *upper_dev, 8401 struct netlink_ext_ack *extack) 8402 { 8403 struct netdev_nested_priv priv = { 8404 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8405 .data = NULL, 8406 }; 8407 8408 return __netdev_upper_dev_link(dev, upper_dev, false, 8409 NULL, NULL, &priv, extack); 8410 } 8411 EXPORT_SYMBOL(netdev_upper_dev_link); 8412 8413 /** 8414 * netdev_master_upper_dev_link - Add a master link to the upper device 8415 * @dev: device 8416 * @upper_dev: new upper device 8417 * @upper_priv: upper device private 8418 * @upper_info: upper info to be passed down via notifier 8419 * @extack: netlink extended ack 8420 * 8421 * Adds a link to device which is upper to this one. In this case, only 8422 * one master upper device can be linked, although other non-master devices 8423 * might be linked as well. The caller must hold the RTNL lock. 8424 * On a failure a negative errno code is returned. On success the reference 8425 * counts are adjusted and the function returns zero. 8426 */ 8427 int netdev_master_upper_dev_link(struct net_device *dev, 8428 struct net_device *upper_dev, 8429 void *upper_priv, void *upper_info, 8430 struct netlink_ext_ack *extack) 8431 { 8432 struct netdev_nested_priv priv = { 8433 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8434 .data = NULL, 8435 }; 8436 8437 return __netdev_upper_dev_link(dev, upper_dev, true, 8438 upper_priv, upper_info, &priv, extack); 8439 } 8440 EXPORT_SYMBOL(netdev_master_upper_dev_link); 8441 8442 static void __netdev_upper_dev_unlink(struct net_device *dev, 8443 struct net_device *upper_dev, 8444 struct netdev_nested_priv *priv) 8445 { 8446 struct netdev_notifier_changeupper_info changeupper_info = { 8447 .info = { 8448 .dev = dev, 8449 }, 8450 .upper_dev = upper_dev, 8451 .linking = false, 8452 }; 8453 8454 ASSERT_RTNL(); 8455 8456 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev; 8457 8458 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8459 &changeupper_info.info); 8460 8461 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8462 8463 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8464 &changeupper_info.info); 8465 8466 __netdev_update_upper_level(dev, NULL); 8467 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8468 8469 __netdev_update_lower_level(upper_dev, priv); 8470 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8471 priv); 8472 } 8473 8474 /** 8475 * netdev_upper_dev_unlink - Removes a link to upper device 8476 * @dev: device 8477 * @upper_dev: new upper device 8478 * 8479 * Removes a link to device which is upper to this one. The caller must hold 8480 * the RTNL lock. 8481 */ 8482 void netdev_upper_dev_unlink(struct net_device *dev, 8483 struct net_device *upper_dev) 8484 { 8485 struct netdev_nested_priv priv = { 8486 .flags = NESTED_SYNC_TODO, 8487 .data = NULL, 8488 }; 8489 8490 __netdev_upper_dev_unlink(dev, upper_dev, &priv); 8491 } 8492 EXPORT_SYMBOL(netdev_upper_dev_unlink); 8493 8494 static void __netdev_adjacent_dev_set(struct net_device *upper_dev, 8495 struct net_device *lower_dev, 8496 bool val) 8497 { 8498 struct netdev_adjacent *adj; 8499 8500 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower); 8501 if (adj) 8502 adj->ignore = val; 8503 8504 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper); 8505 if (adj) 8506 adj->ignore = val; 8507 } 8508 8509 static void netdev_adjacent_dev_disable(struct net_device *upper_dev, 8510 struct net_device *lower_dev) 8511 { 8512 __netdev_adjacent_dev_set(upper_dev, lower_dev, true); 8513 } 8514 8515 static void netdev_adjacent_dev_enable(struct net_device *upper_dev, 8516 struct net_device *lower_dev) 8517 { 8518 __netdev_adjacent_dev_set(upper_dev, lower_dev, false); 8519 } 8520 8521 int netdev_adjacent_change_prepare(struct net_device *old_dev, 8522 struct net_device *new_dev, 8523 struct net_device *dev, 8524 struct netlink_ext_ack *extack) 8525 { 8526 struct netdev_nested_priv priv = { 8527 .flags = 0, 8528 .data = NULL, 8529 }; 8530 int err; 8531 8532 if (!new_dev) 8533 return 0; 8534 8535 if (old_dev && new_dev != old_dev) 8536 netdev_adjacent_dev_disable(dev, old_dev); 8537 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv, 8538 extack); 8539 if (err) { 8540 if (old_dev && new_dev != old_dev) 8541 netdev_adjacent_dev_enable(dev, old_dev); 8542 return err; 8543 } 8544 8545 return 0; 8546 } 8547 EXPORT_SYMBOL(netdev_adjacent_change_prepare); 8548 8549 void netdev_adjacent_change_commit(struct net_device *old_dev, 8550 struct net_device *new_dev, 8551 struct net_device *dev) 8552 { 8553 struct netdev_nested_priv priv = { 8554 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8555 .data = NULL, 8556 }; 8557 8558 if (!new_dev || !old_dev) 8559 return; 8560 8561 if (new_dev == old_dev) 8562 return; 8563 8564 netdev_adjacent_dev_enable(dev, old_dev); 8565 __netdev_upper_dev_unlink(old_dev, dev, &priv); 8566 } 8567 EXPORT_SYMBOL(netdev_adjacent_change_commit); 8568 8569 void netdev_adjacent_change_abort(struct net_device *old_dev, 8570 struct net_device *new_dev, 8571 struct net_device *dev) 8572 { 8573 struct netdev_nested_priv priv = { 8574 .flags = 0, 8575 .data = NULL, 8576 }; 8577 8578 if (!new_dev) 8579 return; 8580 8581 if (old_dev && new_dev != old_dev) 8582 netdev_adjacent_dev_enable(dev, old_dev); 8583 8584 __netdev_upper_dev_unlink(new_dev, dev, &priv); 8585 } 8586 EXPORT_SYMBOL(netdev_adjacent_change_abort); 8587 8588 /** 8589 * netdev_bonding_info_change - Dispatch event about slave change 8590 * @dev: device 8591 * @bonding_info: info to dispatch 8592 * 8593 * Send NETDEV_BONDING_INFO to netdev notifiers with info. 8594 * The caller must hold the RTNL lock. 8595 */ 8596 void netdev_bonding_info_change(struct net_device *dev, 8597 struct netdev_bonding_info *bonding_info) 8598 { 8599 struct netdev_notifier_bonding_info info = { 8600 .info.dev = dev, 8601 }; 8602 8603 memcpy(&info.bonding_info, bonding_info, 8604 sizeof(struct netdev_bonding_info)); 8605 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, 8606 &info.info); 8607 } 8608 EXPORT_SYMBOL(netdev_bonding_info_change); 8609 8610 static int netdev_offload_xstats_enable_l3(struct net_device *dev, 8611 struct netlink_ext_ack *extack) 8612 { 8613 struct netdev_notifier_offload_xstats_info info = { 8614 .info.dev = dev, 8615 .info.extack = extack, 8616 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8617 }; 8618 int err; 8619 int rc; 8620 8621 dev->offload_xstats_l3 = kzalloc(sizeof(*dev->offload_xstats_l3), 8622 GFP_KERNEL); 8623 if (!dev->offload_xstats_l3) 8624 return -ENOMEM; 8625 8626 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE, 8627 NETDEV_OFFLOAD_XSTATS_DISABLE, 8628 &info.info); 8629 err = notifier_to_errno(rc); 8630 if (err) 8631 goto free_stats; 8632 8633 return 0; 8634 8635 free_stats: 8636 kfree(dev->offload_xstats_l3); 8637 dev->offload_xstats_l3 = NULL; 8638 return err; 8639 } 8640 8641 int netdev_offload_xstats_enable(struct net_device *dev, 8642 enum netdev_offload_xstats_type type, 8643 struct netlink_ext_ack *extack) 8644 { 8645 ASSERT_RTNL(); 8646 8647 if (netdev_offload_xstats_enabled(dev, type)) 8648 return -EALREADY; 8649 8650 switch (type) { 8651 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8652 return netdev_offload_xstats_enable_l3(dev, extack); 8653 } 8654 8655 WARN_ON(1); 8656 return -EINVAL; 8657 } 8658 EXPORT_SYMBOL(netdev_offload_xstats_enable); 8659 8660 static void netdev_offload_xstats_disable_l3(struct net_device *dev) 8661 { 8662 struct netdev_notifier_offload_xstats_info info = { 8663 .info.dev = dev, 8664 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8665 }; 8666 8667 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE, 8668 &info.info); 8669 kfree(dev->offload_xstats_l3); 8670 dev->offload_xstats_l3 = NULL; 8671 } 8672 8673 int netdev_offload_xstats_disable(struct net_device *dev, 8674 enum netdev_offload_xstats_type type) 8675 { 8676 ASSERT_RTNL(); 8677 8678 if (!netdev_offload_xstats_enabled(dev, type)) 8679 return -EALREADY; 8680 8681 switch (type) { 8682 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8683 netdev_offload_xstats_disable_l3(dev); 8684 return 0; 8685 } 8686 8687 WARN_ON(1); 8688 return -EINVAL; 8689 } 8690 EXPORT_SYMBOL(netdev_offload_xstats_disable); 8691 8692 static void netdev_offload_xstats_disable_all(struct net_device *dev) 8693 { 8694 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3); 8695 } 8696 8697 static struct rtnl_hw_stats64 * 8698 netdev_offload_xstats_get_ptr(const struct net_device *dev, 8699 enum netdev_offload_xstats_type type) 8700 { 8701 switch (type) { 8702 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8703 return dev->offload_xstats_l3; 8704 } 8705 8706 WARN_ON(1); 8707 return NULL; 8708 } 8709 8710 bool netdev_offload_xstats_enabled(const struct net_device *dev, 8711 enum netdev_offload_xstats_type type) 8712 { 8713 ASSERT_RTNL(); 8714 8715 return netdev_offload_xstats_get_ptr(dev, type); 8716 } 8717 EXPORT_SYMBOL(netdev_offload_xstats_enabled); 8718 8719 struct netdev_notifier_offload_xstats_ru { 8720 bool used; 8721 }; 8722 8723 struct netdev_notifier_offload_xstats_rd { 8724 struct rtnl_hw_stats64 stats; 8725 bool used; 8726 }; 8727 8728 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest, 8729 const struct rtnl_hw_stats64 *src) 8730 { 8731 dest->rx_packets += src->rx_packets; 8732 dest->tx_packets += src->tx_packets; 8733 dest->rx_bytes += src->rx_bytes; 8734 dest->tx_bytes += src->tx_bytes; 8735 dest->rx_errors += src->rx_errors; 8736 dest->tx_errors += src->tx_errors; 8737 dest->rx_dropped += src->rx_dropped; 8738 dest->tx_dropped += src->tx_dropped; 8739 dest->multicast += src->multicast; 8740 } 8741 8742 static int netdev_offload_xstats_get_used(struct net_device *dev, 8743 enum netdev_offload_xstats_type type, 8744 bool *p_used, 8745 struct netlink_ext_ack *extack) 8746 { 8747 struct netdev_notifier_offload_xstats_ru report_used = {}; 8748 struct netdev_notifier_offload_xstats_info info = { 8749 .info.dev = dev, 8750 .info.extack = extack, 8751 .type = type, 8752 .report_used = &report_used, 8753 }; 8754 int rc; 8755 8756 WARN_ON(!netdev_offload_xstats_enabled(dev, type)); 8757 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED, 8758 &info.info); 8759 *p_used = report_used.used; 8760 return notifier_to_errno(rc); 8761 } 8762 8763 static int netdev_offload_xstats_get_stats(struct net_device *dev, 8764 enum netdev_offload_xstats_type type, 8765 struct rtnl_hw_stats64 *p_stats, 8766 bool *p_used, 8767 struct netlink_ext_ack *extack) 8768 { 8769 struct netdev_notifier_offload_xstats_rd report_delta = {}; 8770 struct netdev_notifier_offload_xstats_info info = { 8771 .info.dev = dev, 8772 .info.extack = extack, 8773 .type = type, 8774 .report_delta = &report_delta, 8775 }; 8776 struct rtnl_hw_stats64 *stats; 8777 int rc; 8778 8779 stats = netdev_offload_xstats_get_ptr(dev, type); 8780 if (WARN_ON(!stats)) 8781 return -EINVAL; 8782 8783 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA, 8784 &info.info); 8785 8786 /* Cache whatever we got, even if there was an error, otherwise the 8787 * successful stats retrievals would get lost. 8788 */ 8789 netdev_hw_stats64_add(stats, &report_delta.stats); 8790 8791 if (p_stats) 8792 *p_stats = *stats; 8793 *p_used = report_delta.used; 8794 8795 return notifier_to_errno(rc); 8796 } 8797 8798 int netdev_offload_xstats_get(struct net_device *dev, 8799 enum netdev_offload_xstats_type type, 8800 struct rtnl_hw_stats64 *p_stats, bool *p_used, 8801 struct netlink_ext_ack *extack) 8802 { 8803 ASSERT_RTNL(); 8804 8805 if (p_stats) 8806 return netdev_offload_xstats_get_stats(dev, type, p_stats, 8807 p_used, extack); 8808 else 8809 return netdev_offload_xstats_get_used(dev, type, p_used, 8810 extack); 8811 } 8812 EXPORT_SYMBOL(netdev_offload_xstats_get); 8813 8814 void 8815 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta, 8816 const struct rtnl_hw_stats64 *stats) 8817 { 8818 report_delta->used = true; 8819 netdev_hw_stats64_add(&report_delta->stats, stats); 8820 } 8821 EXPORT_SYMBOL(netdev_offload_xstats_report_delta); 8822 8823 void 8824 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used) 8825 { 8826 report_used->used = true; 8827 } 8828 EXPORT_SYMBOL(netdev_offload_xstats_report_used); 8829 8830 void netdev_offload_xstats_push_delta(struct net_device *dev, 8831 enum netdev_offload_xstats_type type, 8832 const struct rtnl_hw_stats64 *p_stats) 8833 { 8834 struct rtnl_hw_stats64 *stats; 8835 8836 ASSERT_RTNL(); 8837 8838 stats = netdev_offload_xstats_get_ptr(dev, type); 8839 if (WARN_ON(!stats)) 8840 return; 8841 8842 netdev_hw_stats64_add(stats, p_stats); 8843 } 8844 EXPORT_SYMBOL(netdev_offload_xstats_push_delta); 8845 8846 /** 8847 * netdev_get_xmit_slave - Get the xmit slave of master device 8848 * @dev: device 8849 * @skb: The packet 8850 * @all_slaves: assume all the slaves are active 8851 * 8852 * The reference counters are not incremented so the caller must be 8853 * careful with locks. The caller must hold RCU lock. 8854 * %NULL is returned if no slave is found. 8855 */ 8856 8857 struct net_device *netdev_get_xmit_slave(struct net_device *dev, 8858 struct sk_buff *skb, 8859 bool all_slaves) 8860 { 8861 const struct net_device_ops *ops = dev->netdev_ops; 8862 8863 if (!ops->ndo_get_xmit_slave) 8864 return NULL; 8865 return ops->ndo_get_xmit_slave(dev, skb, all_slaves); 8866 } 8867 EXPORT_SYMBOL(netdev_get_xmit_slave); 8868 8869 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev, 8870 struct sock *sk) 8871 { 8872 const struct net_device_ops *ops = dev->netdev_ops; 8873 8874 if (!ops->ndo_sk_get_lower_dev) 8875 return NULL; 8876 return ops->ndo_sk_get_lower_dev(dev, sk); 8877 } 8878 8879 /** 8880 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket 8881 * @dev: device 8882 * @sk: the socket 8883 * 8884 * %NULL is returned if no lower device is found. 8885 */ 8886 8887 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev, 8888 struct sock *sk) 8889 { 8890 struct net_device *lower; 8891 8892 lower = netdev_sk_get_lower_dev(dev, sk); 8893 while (lower) { 8894 dev = lower; 8895 lower = netdev_sk_get_lower_dev(dev, sk); 8896 } 8897 8898 return dev; 8899 } 8900 EXPORT_SYMBOL(netdev_sk_get_lowest_dev); 8901 8902 static void netdev_adjacent_add_links(struct net_device *dev) 8903 { 8904 struct netdev_adjacent *iter; 8905 8906 struct net *net = dev_net(dev); 8907 8908 list_for_each_entry(iter, &dev->adj_list.upper, list) { 8909 if (!net_eq(net, dev_net(iter->dev))) 8910 continue; 8911 netdev_adjacent_sysfs_add(iter->dev, dev, 8912 &iter->dev->adj_list.lower); 8913 netdev_adjacent_sysfs_add(dev, iter->dev, 8914 &dev->adj_list.upper); 8915 } 8916 8917 list_for_each_entry(iter, &dev->adj_list.lower, list) { 8918 if (!net_eq(net, dev_net(iter->dev))) 8919 continue; 8920 netdev_adjacent_sysfs_add(iter->dev, dev, 8921 &iter->dev->adj_list.upper); 8922 netdev_adjacent_sysfs_add(dev, iter->dev, 8923 &dev->adj_list.lower); 8924 } 8925 } 8926 8927 static void netdev_adjacent_del_links(struct net_device *dev) 8928 { 8929 struct netdev_adjacent *iter; 8930 8931 struct net *net = dev_net(dev); 8932 8933 list_for_each_entry(iter, &dev->adj_list.upper, list) { 8934 if (!net_eq(net, dev_net(iter->dev))) 8935 continue; 8936 netdev_adjacent_sysfs_del(iter->dev, dev->name, 8937 &iter->dev->adj_list.lower); 8938 netdev_adjacent_sysfs_del(dev, iter->dev->name, 8939 &dev->adj_list.upper); 8940 } 8941 8942 list_for_each_entry(iter, &dev->adj_list.lower, list) { 8943 if (!net_eq(net, dev_net(iter->dev))) 8944 continue; 8945 netdev_adjacent_sysfs_del(iter->dev, dev->name, 8946 &iter->dev->adj_list.upper); 8947 netdev_adjacent_sysfs_del(dev, iter->dev->name, 8948 &dev->adj_list.lower); 8949 } 8950 } 8951 8952 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname) 8953 { 8954 struct netdev_adjacent *iter; 8955 8956 struct net *net = dev_net(dev); 8957 8958 list_for_each_entry(iter, &dev->adj_list.upper, list) { 8959 if (!net_eq(net, dev_net(iter->dev))) 8960 continue; 8961 netdev_adjacent_sysfs_del(iter->dev, oldname, 8962 &iter->dev->adj_list.lower); 8963 netdev_adjacent_sysfs_add(iter->dev, dev, 8964 &iter->dev->adj_list.lower); 8965 } 8966 8967 list_for_each_entry(iter, &dev->adj_list.lower, list) { 8968 if (!net_eq(net, dev_net(iter->dev))) 8969 continue; 8970 netdev_adjacent_sysfs_del(iter->dev, oldname, 8971 &iter->dev->adj_list.upper); 8972 netdev_adjacent_sysfs_add(iter->dev, dev, 8973 &iter->dev->adj_list.upper); 8974 } 8975 } 8976 8977 void *netdev_lower_dev_get_private(struct net_device *dev, 8978 struct net_device *lower_dev) 8979 { 8980 struct netdev_adjacent *lower; 8981 8982 if (!lower_dev) 8983 return NULL; 8984 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower); 8985 if (!lower) 8986 return NULL; 8987 8988 return lower->private; 8989 } 8990 EXPORT_SYMBOL(netdev_lower_dev_get_private); 8991 8992 8993 /** 8994 * netdev_lower_state_changed - Dispatch event about lower device state change 8995 * @lower_dev: device 8996 * @lower_state_info: state to dispatch 8997 * 8998 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info. 8999 * The caller must hold the RTNL lock. 9000 */ 9001 void netdev_lower_state_changed(struct net_device *lower_dev, 9002 void *lower_state_info) 9003 { 9004 struct netdev_notifier_changelowerstate_info changelowerstate_info = { 9005 .info.dev = lower_dev, 9006 }; 9007 9008 ASSERT_RTNL(); 9009 changelowerstate_info.lower_state_info = lower_state_info; 9010 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, 9011 &changelowerstate_info.info); 9012 } 9013 EXPORT_SYMBOL(netdev_lower_state_changed); 9014 9015 static void dev_change_rx_flags(struct net_device *dev, int flags) 9016 { 9017 const struct net_device_ops *ops = dev->netdev_ops; 9018 9019 if (ops->ndo_change_rx_flags) 9020 ops->ndo_change_rx_flags(dev, flags); 9021 } 9022 9023 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify) 9024 { 9025 unsigned int old_flags = dev->flags; 9026 unsigned int promiscuity, flags; 9027 kuid_t uid; 9028 kgid_t gid; 9029 9030 ASSERT_RTNL(); 9031 9032 promiscuity = dev->promiscuity + inc; 9033 if (promiscuity == 0) { 9034 /* 9035 * Avoid overflow. 9036 * If inc causes overflow, untouch promisc and return error. 9037 */ 9038 if (unlikely(inc > 0)) { 9039 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n"); 9040 return -EOVERFLOW; 9041 } 9042 flags = old_flags & ~IFF_PROMISC; 9043 } else { 9044 flags = old_flags | IFF_PROMISC; 9045 } 9046 WRITE_ONCE(dev->promiscuity, promiscuity); 9047 if (flags != old_flags) { 9048 WRITE_ONCE(dev->flags, flags); 9049 netdev_info(dev, "%s promiscuous mode\n", 9050 dev->flags & IFF_PROMISC ? "entered" : "left"); 9051 if (audit_enabled) { 9052 current_uid_gid(&uid, &gid); 9053 audit_log(audit_context(), GFP_ATOMIC, 9054 AUDIT_ANOM_PROMISCUOUS, 9055 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u", 9056 dev->name, (dev->flags & IFF_PROMISC), 9057 (old_flags & IFF_PROMISC), 9058 from_kuid(&init_user_ns, audit_get_loginuid(current)), 9059 from_kuid(&init_user_ns, uid), 9060 from_kgid(&init_user_ns, gid), 9061 audit_get_sessionid(current)); 9062 } 9063 9064 dev_change_rx_flags(dev, IFF_PROMISC); 9065 } 9066 if (notify) 9067 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL); 9068 return 0; 9069 } 9070 9071 /** 9072 * dev_set_promiscuity - update promiscuity count on a device 9073 * @dev: device 9074 * @inc: modifier 9075 * 9076 * Add or remove promiscuity from a device. While the count in the device 9077 * remains above zero the interface remains promiscuous. Once it hits zero 9078 * the device reverts back to normal filtering operation. A negative inc 9079 * value is used to drop promiscuity on the device. 9080 * Return 0 if successful or a negative errno code on error. 9081 */ 9082 int dev_set_promiscuity(struct net_device *dev, int inc) 9083 { 9084 unsigned int old_flags = dev->flags; 9085 int err; 9086 9087 err = __dev_set_promiscuity(dev, inc, true); 9088 if (err < 0) 9089 return err; 9090 if (dev->flags != old_flags) 9091 dev_set_rx_mode(dev); 9092 return err; 9093 } 9094 EXPORT_SYMBOL(dev_set_promiscuity); 9095 9096 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify) 9097 { 9098 unsigned int old_flags = dev->flags, old_gflags = dev->gflags; 9099 unsigned int allmulti, flags; 9100 9101 ASSERT_RTNL(); 9102 9103 allmulti = dev->allmulti + inc; 9104 if (allmulti == 0) { 9105 /* 9106 * Avoid overflow. 9107 * If inc causes overflow, untouch allmulti and return error. 9108 */ 9109 if (unlikely(inc > 0)) { 9110 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n"); 9111 return -EOVERFLOW; 9112 } 9113 flags = old_flags & ~IFF_ALLMULTI; 9114 } else { 9115 flags = old_flags | IFF_ALLMULTI; 9116 } 9117 WRITE_ONCE(dev->allmulti, allmulti); 9118 if (flags != old_flags) { 9119 WRITE_ONCE(dev->flags, flags); 9120 netdev_info(dev, "%s allmulticast mode\n", 9121 dev->flags & IFF_ALLMULTI ? "entered" : "left"); 9122 dev_change_rx_flags(dev, IFF_ALLMULTI); 9123 dev_set_rx_mode(dev); 9124 if (notify) 9125 __dev_notify_flags(dev, old_flags, 9126 dev->gflags ^ old_gflags, 0, NULL); 9127 } 9128 return 0; 9129 } 9130 9131 /** 9132 * dev_set_allmulti - update allmulti count on a device 9133 * @dev: device 9134 * @inc: modifier 9135 * 9136 * Add or remove reception of all multicast frames to a device. While the 9137 * count in the device remains above zero the interface remains listening 9138 * to all interfaces. Once it hits zero the device reverts back to normal 9139 * filtering operation. A negative @inc value is used to drop the counter 9140 * when releasing a resource needing all multicasts. 9141 * Return 0 if successful or a negative errno code on error. 9142 */ 9143 9144 int dev_set_allmulti(struct net_device *dev, int inc) 9145 { 9146 return __dev_set_allmulti(dev, inc, true); 9147 } 9148 EXPORT_SYMBOL(dev_set_allmulti); 9149 9150 /* 9151 * Upload unicast and multicast address lists to device and 9152 * configure RX filtering. When the device doesn't support unicast 9153 * filtering it is put in promiscuous mode while unicast addresses 9154 * are present. 9155 */ 9156 void __dev_set_rx_mode(struct net_device *dev) 9157 { 9158 const struct net_device_ops *ops = dev->netdev_ops; 9159 9160 /* dev_open will call this function so the list will stay sane. */ 9161 if (!(dev->flags&IFF_UP)) 9162 return; 9163 9164 if (!netif_device_present(dev)) 9165 return; 9166 9167 if (!(dev->priv_flags & IFF_UNICAST_FLT)) { 9168 /* Unicast addresses changes may only happen under the rtnl, 9169 * therefore calling __dev_set_promiscuity here is safe. 9170 */ 9171 if (!netdev_uc_empty(dev) && !dev->uc_promisc) { 9172 __dev_set_promiscuity(dev, 1, false); 9173 dev->uc_promisc = true; 9174 } else if (netdev_uc_empty(dev) && dev->uc_promisc) { 9175 __dev_set_promiscuity(dev, -1, false); 9176 dev->uc_promisc = false; 9177 } 9178 } 9179 9180 if (ops->ndo_set_rx_mode) 9181 ops->ndo_set_rx_mode(dev); 9182 } 9183 9184 void dev_set_rx_mode(struct net_device *dev) 9185 { 9186 netif_addr_lock_bh(dev); 9187 __dev_set_rx_mode(dev); 9188 netif_addr_unlock_bh(dev); 9189 } 9190 9191 /** 9192 * dev_get_flags - get flags reported to userspace 9193 * @dev: device 9194 * 9195 * Get the combination of flag bits exported through APIs to userspace. 9196 */ 9197 unsigned int dev_get_flags(const struct net_device *dev) 9198 { 9199 unsigned int flags; 9200 9201 flags = (READ_ONCE(dev->flags) & ~(IFF_PROMISC | 9202 IFF_ALLMULTI | 9203 IFF_RUNNING | 9204 IFF_LOWER_UP | 9205 IFF_DORMANT)) | 9206 (READ_ONCE(dev->gflags) & (IFF_PROMISC | 9207 IFF_ALLMULTI)); 9208 9209 if (netif_running(dev)) { 9210 if (netif_oper_up(dev)) 9211 flags |= IFF_RUNNING; 9212 if (netif_carrier_ok(dev)) 9213 flags |= IFF_LOWER_UP; 9214 if (netif_dormant(dev)) 9215 flags |= IFF_DORMANT; 9216 } 9217 9218 return flags; 9219 } 9220 EXPORT_SYMBOL(dev_get_flags); 9221 9222 int __dev_change_flags(struct net_device *dev, unsigned int flags, 9223 struct netlink_ext_ack *extack) 9224 { 9225 unsigned int old_flags = dev->flags; 9226 int ret; 9227 9228 ASSERT_RTNL(); 9229 9230 /* 9231 * Set the flags on our device. 9232 */ 9233 9234 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP | 9235 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL | 9236 IFF_AUTOMEDIA)) | 9237 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC | 9238 IFF_ALLMULTI)); 9239 9240 /* 9241 * Load in the correct multicast list now the flags have changed. 9242 */ 9243 9244 if ((old_flags ^ flags) & IFF_MULTICAST) 9245 dev_change_rx_flags(dev, IFF_MULTICAST); 9246 9247 dev_set_rx_mode(dev); 9248 9249 /* 9250 * Have we downed the interface. We handle IFF_UP ourselves 9251 * according to user attempts to set it, rather than blindly 9252 * setting it. 9253 */ 9254 9255 ret = 0; 9256 if ((old_flags ^ flags) & IFF_UP) { 9257 if (old_flags & IFF_UP) 9258 __dev_close(dev); 9259 else 9260 ret = __dev_open(dev, extack); 9261 } 9262 9263 if ((flags ^ dev->gflags) & IFF_PROMISC) { 9264 int inc = (flags & IFF_PROMISC) ? 1 : -1; 9265 unsigned int old_flags = dev->flags; 9266 9267 dev->gflags ^= IFF_PROMISC; 9268 9269 if (__dev_set_promiscuity(dev, inc, false) >= 0) 9270 if (dev->flags != old_flags) 9271 dev_set_rx_mode(dev); 9272 } 9273 9274 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI 9275 * is important. Some (broken) drivers set IFF_PROMISC, when 9276 * IFF_ALLMULTI is requested not asking us and not reporting. 9277 */ 9278 if ((flags ^ dev->gflags) & IFF_ALLMULTI) { 9279 int inc = (flags & IFF_ALLMULTI) ? 1 : -1; 9280 9281 dev->gflags ^= IFF_ALLMULTI; 9282 __dev_set_allmulti(dev, inc, false); 9283 } 9284 9285 return ret; 9286 } 9287 9288 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags, 9289 unsigned int gchanges, u32 portid, 9290 const struct nlmsghdr *nlh) 9291 { 9292 unsigned int changes = dev->flags ^ old_flags; 9293 9294 if (gchanges) 9295 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh); 9296 9297 if (changes & IFF_UP) { 9298 if (dev->flags & IFF_UP) 9299 call_netdevice_notifiers(NETDEV_UP, dev); 9300 else 9301 call_netdevice_notifiers(NETDEV_DOWN, dev); 9302 } 9303 9304 if (dev->flags & IFF_UP && 9305 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) { 9306 struct netdev_notifier_change_info change_info = { 9307 .info = { 9308 .dev = dev, 9309 }, 9310 .flags_changed = changes, 9311 }; 9312 9313 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info); 9314 } 9315 } 9316 9317 /** 9318 * dev_change_flags - change device settings 9319 * @dev: device 9320 * @flags: device state flags 9321 * @extack: netlink extended ack 9322 * 9323 * Change settings on device based state flags. The flags are 9324 * in the userspace exported format. 9325 */ 9326 int dev_change_flags(struct net_device *dev, unsigned int flags, 9327 struct netlink_ext_ack *extack) 9328 { 9329 int ret; 9330 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags; 9331 9332 ret = __dev_change_flags(dev, flags, extack); 9333 if (ret < 0) 9334 return ret; 9335 9336 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags); 9337 __dev_notify_flags(dev, old_flags, changes, 0, NULL); 9338 return ret; 9339 } 9340 EXPORT_SYMBOL(dev_change_flags); 9341 9342 int __dev_set_mtu(struct net_device *dev, int new_mtu) 9343 { 9344 const struct net_device_ops *ops = dev->netdev_ops; 9345 9346 if (ops->ndo_change_mtu) 9347 return ops->ndo_change_mtu(dev, new_mtu); 9348 9349 /* Pairs with all the lockless reads of dev->mtu in the stack */ 9350 WRITE_ONCE(dev->mtu, new_mtu); 9351 return 0; 9352 } 9353 EXPORT_SYMBOL(__dev_set_mtu); 9354 9355 int dev_validate_mtu(struct net_device *dev, int new_mtu, 9356 struct netlink_ext_ack *extack) 9357 { 9358 /* MTU must be positive, and in range */ 9359 if (new_mtu < 0 || new_mtu < dev->min_mtu) { 9360 NL_SET_ERR_MSG(extack, "mtu less than device minimum"); 9361 return -EINVAL; 9362 } 9363 9364 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) { 9365 NL_SET_ERR_MSG(extack, "mtu greater than device maximum"); 9366 return -EINVAL; 9367 } 9368 return 0; 9369 } 9370 9371 /** 9372 * dev_set_mtu_ext - Change maximum transfer unit 9373 * @dev: device 9374 * @new_mtu: new transfer unit 9375 * @extack: netlink extended ack 9376 * 9377 * Change the maximum transfer size of the network device. 9378 */ 9379 int dev_set_mtu_ext(struct net_device *dev, int new_mtu, 9380 struct netlink_ext_ack *extack) 9381 { 9382 int err, orig_mtu; 9383 9384 if (new_mtu == dev->mtu) 9385 return 0; 9386 9387 err = dev_validate_mtu(dev, new_mtu, extack); 9388 if (err) 9389 return err; 9390 9391 if (!netif_device_present(dev)) 9392 return -ENODEV; 9393 9394 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev); 9395 err = notifier_to_errno(err); 9396 if (err) 9397 return err; 9398 9399 orig_mtu = dev->mtu; 9400 err = __dev_set_mtu(dev, new_mtu); 9401 9402 if (!err) { 9403 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9404 orig_mtu); 9405 err = notifier_to_errno(err); 9406 if (err) { 9407 /* setting mtu back and notifying everyone again, 9408 * so that they have a chance to revert changes. 9409 */ 9410 __dev_set_mtu(dev, orig_mtu); 9411 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9412 new_mtu); 9413 } 9414 } 9415 return err; 9416 } 9417 9418 int dev_set_mtu(struct net_device *dev, int new_mtu) 9419 { 9420 struct netlink_ext_ack extack; 9421 int err; 9422 9423 memset(&extack, 0, sizeof(extack)); 9424 err = dev_set_mtu_ext(dev, new_mtu, &extack); 9425 if (err && extack._msg) 9426 net_err_ratelimited("%s: %s\n", dev->name, extack._msg); 9427 return err; 9428 } 9429 EXPORT_SYMBOL(dev_set_mtu); 9430 9431 /** 9432 * dev_change_tx_queue_len - Change TX queue length of a netdevice 9433 * @dev: device 9434 * @new_len: new tx queue length 9435 */ 9436 int dev_change_tx_queue_len(struct net_device *dev, unsigned long new_len) 9437 { 9438 unsigned int orig_len = dev->tx_queue_len; 9439 int res; 9440 9441 if (new_len != (unsigned int)new_len) 9442 return -ERANGE; 9443 9444 if (new_len != orig_len) { 9445 WRITE_ONCE(dev->tx_queue_len, new_len); 9446 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev); 9447 res = notifier_to_errno(res); 9448 if (res) 9449 goto err_rollback; 9450 res = dev_qdisc_change_tx_queue_len(dev); 9451 if (res) 9452 goto err_rollback; 9453 } 9454 9455 return 0; 9456 9457 err_rollback: 9458 netdev_err(dev, "refused to change device tx_queue_len\n"); 9459 WRITE_ONCE(dev->tx_queue_len, orig_len); 9460 return res; 9461 } 9462 9463 /** 9464 * dev_set_group - Change group this device belongs to 9465 * @dev: device 9466 * @new_group: group this device should belong to 9467 */ 9468 void dev_set_group(struct net_device *dev, int new_group) 9469 { 9470 dev->group = new_group; 9471 } 9472 9473 /** 9474 * dev_pre_changeaddr_notify - Call NETDEV_PRE_CHANGEADDR. 9475 * @dev: device 9476 * @addr: new address 9477 * @extack: netlink extended ack 9478 */ 9479 int dev_pre_changeaddr_notify(struct net_device *dev, const char *addr, 9480 struct netlink_ext_ack *extack) 9481 { 9482 struct netdev_notifier_pre_changeaddr_info info = { 9483 .info.dev = dev, 9484 .info.extack = extack, 9485 .dev_addr = addr, 9486 }; 9487 int rc; 9488 9489 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info); 9490 return notifier_to_errno(rc); 9491 } 9492 EXPORT_SYMBOL(dev_pre_changeaddr_notify); 9493 9494 /** 9495 * dev_set_mac_address - Change Media Access Control Address 9496 * @dev: device 9497 * @sa: new address 9498 * @extack: netlink extended ack 9499 * 9500 * Change the hardware (MAC) address of the device 9501 */ 9502 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa, 9503 struct netlink_ext_ack *extack) 9504 { 9505 const struct net_device_ops *ops = dev->netdev_ops; 9506 int err; 9507 9508 if (!ops->ndo_set_mac_address) 9509 return -EOPNOTSUPP; 9510 if (sa->sa_family != dev->type) 9511 return -EINVAL; 9512 if (!netif_device_present(dev)) 9513 return -ENODEV; 9514 err = dev_pre_changeaddr_notify(dev, sa->sa_data, extack); 9515 if (err) 9516 return err; 9517 if (memcmp(dev->dev_addr, sa->sa_data, dev->addr_len)) { 9518 err = ops->ndo_set_mac_address(dev, sa); 9519 if (err) 9520 return err; 9521 } 9522 dev->addr_assign_type = NET_ADDR_SET; 9523 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev); 9524 add_device_randomness(dev->dev_addr, dev->addr_len); 9525 return 0; 9526 } 9527 EXPORT_SYMBOL(dev_set_mac_address); 9528 9529 DECLARE_RWSEM(dev_addr_sem); 9530 9531 int dev_set_mac_address_user(struct net_device *dev, struct sockaddr *sa, 9532 struct netlink_ext_ack *extack) 9533 { 9534 int ret; 9535 9536 down_write(&dev_addr_sem); 9537 ret = dev_set_mac_address(dev, sa, extack); 9538 up_write(&dev_addr_sem); 9539 return ret; 9540 } 9541 EXPORT_SYMBOL(dev_set_mac_address_user); 9542 9543 int dev_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name) 9544 { 9545 size_t size = sizeof(sa->sa_data_min); 9546 struct net_device *dev; 9547 int ret = 0; 9548 9549 down_read(&dev_addr_sem); 9550 rcu_read_lock(); 9551 9552 dev = dev_get_by_name_rcu(net, dev_name); 9553 if (!dev) { 9554 ret = -ENODEV; 9555 goto unlock; 9556 } 9557 if (!dev->addr_len) 9558 memset(sa->sa_data, 0, size); 9559 else 9560 memcpy(sa->sa_data, dev->dev_addr, 9561 min_t(size_t, size, dev->addr_len)); 9562 sa->sa_family = dev->type; 9563 9564 unlock: 9565 rcu_read_unlock(); 9566 up_read(&dev_addr_sem); 9567 return ret; 9568 } 9569 EXPORT_SYMBOL(dev_get_mac_address); 9570 9571 /** 9572 * dev_change_carrier - Change device carrier 9573 * @dev: device 9574 * @new_carrier: new value 9575 * 9576 * Change device carrier 9577 */ 9578 int dev_change_carrier(struct net_device *dev, bool new_carrier) 9579 { 9580 const struct net_device_ops *ops = dev->netdev_ops; 9581 9582 if (!ops->ndo_change_carrier) 9583 return -EOPNOTSUPP; 9584 if (!netif_device_present(dev)) 9585 return -ENODEV; 9586 return ops->ndo_change_carrier(dev, new_carrier); 9587 } 9588 9589 /** 9590 * dev_get_phys_port_id - Get device physical port ID 9591 * @dev: device 9592 * @ppid: port ID 9593 * 9594 * Get device physical port ID 9595 */ 9596 int dev_get_phys_port_id(struct net_device *dev, 9597 struct netdev_phys_item_id *ppid) 9598 { 9599 const struct net_device_ops *ops = dev->netdev_ops; 9600 9601 if (!ops->ndo_get_phys_port_id) 9602 return -EOPNOTSUPP; 9603 return ops->ndo_get_phys_port_id(dev, ppid); 9604 } 9605 9606 /** 9607 * dev_get_phys_port_name - Get device physical port name 9608 * @dev: device 9609 * @name: port name 9610 * @len: limit of bytes to copy to name 9611 * 9612 * Get device physical port name 9613 */ 9614 int dev_get_phys_port_name(struct net_device *dev, 9615 char *name, size_t len) 9616 { 9617 const struct net_device_ops *ops = dev->netdev_ops; 9618 int err; 9619 9620 if (ops->ndo_get_phys_port_name) { 9621 err = ops->ndo_get_phys_port_name(dev, name, len); 9622 if (err != -EOPNOTSUPP) 9623 return err; 9624 } 9625 return devlink_compat_phys_port_name_get(dev, name, len); 9626 } 9627 9628 /** 9629 * dev_get_port_parent_id - Get the device's port parent identifier 9630 * @dev: network device 9631 * @ppid: pointer to a storage for the port's parent identifier 9632 * @recurse: allow/disallow recursion to lower devices 9633 * 9634 * Get the devices's port parent identifier 9635 */ 9636 int dev_get_port_parent_id(struct net_device *dev, 9637 struct netdev_phys_item_id *ppid, 9638 bool recurse) 9639 { 9640 const struct net_device_ops *ops = dev->netdev_ops; 9641 struct netdev_phys_item_id first = { }; 9642 struct net_device *lower_dev; 9643 struct list_head *iter; 9644 int err; 9645 9646 if (ops->ndo_get_port_parent_id) { 9647 err = ops->ndo_get_port_parent_id(dev, ppid); 9648 if (err != -EOPNOTSUPP) 9649 return err; 9650 } 9651 9652 err = devlink_compat_switch_id_get(dev, ppid); 9653 if (!recurse || err != -EOPNOTSUPP) 9654 return err; 9655 9656 netdev_for_each_lower_dev(dev, lower_dev, iter) { 9657 err = dev_get_port_parent_id(lower_dev, ppid, true); 9658 if (err) 9659 break; 9660 if (!first.id_len) 9661 first = *ppid; 9662 else if (memcmp(&first, ppid, sizeof(*ppid))) 9663 return -EOPNOTSUPP; 9664 } 9665 9666 return err; 9667 } 9668 EXPORT_SYMBOL(dev_get_port_parent_id); 9669 9670 /** 9671 * netdev_port_same_parent_id - Indicate if two network devices have 9672 * the same port parent identifier 9673 * @a: first network device 9674 * @b: second network device 9675 */ 9676 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b) 9677 { 9678 struct netdev_phys_item_id a_id = { }; 9679 struct netdev_phys_item_id b_id = { }; 9680 9681 if (dev_get_port_parent_id(a, &a_id, true) || 9682 dev_get_port_parent_id(b, &b_id, true)) 9683 return false; 9684 9685 return netdev_phys_item_id_same(&a_id, &b_id); 9686 } 9687 EXPORT_SYMBOL(netdev_port_same_parent_id); 9688 9689 /** 9690 * dev_change_proto_down - set carrier according to proto_down. 9691 * 9692 * @dev: device 9693 * @proto_down: new value 9694 */ 9695 int dev_change_proto_down(struct net_device *dev, bool proto_down) 9696 { 9697 if (!dev->change_proto_down) 9698 return -EOPNOTSUPP; 9699 if (!netif_device_present(dev)) 9700 return -ENODEV; 9701 if (proto_down) 9702 netif_carrier_off(dev); 9703 else 9704 netif_carrier_on(dev); 9705 WRITE_ONCE(dev->proto_down, proto_down); 9706 return 0; 9707 } 9708 9709 /** 9710 * dev_change_proto_down_reason - proto down reason 9711 * 9712 * @dev: device 9713 * @mask: proto down mask 9714 * @value: proto down value 9715 */ 9716 void dev_change_proto_down_reason(struct net_device *dev, unsigned long mask, 9717 u32 value) 9718 { 9719 u32 proto_down_reason; 9720 int b; 9721 9722 if (!mask) { 9723 proto_down_reason = value; 9724 } else { 9725 proto_down_reason = dev->proto_down_reason; 9726 for_each_set_bit(b, &mask, 32) { 9727 if (value & (1 << b)) 9728 proto_down_reason |= BIT(b); 9729 else 9730 proto_down_reason &= ~BIT(b); 9731 } 9732 } 9733 WRITE_ONCE(dev->proto_down_reason, proto_down_reason); 9734 } 9735 9736 struct bpf_xdp_link { 9737 struct bpf_link link; 9738 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */ 9739 int flags; 9740 }; 9741 9742 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags) 9743 { 9744 if (flags & XDP_FLAGS_HW_MODE) 9745 return XDP_MODE_HW; 9746 if (flags & XDP_FLAGS_DRV_MODE) 9747 return XDP_MODE_DRV; 9748 if (flags & XDP_FLAGS_SKB_MODE) 9749 return XDP_MODE_SKB; 9750 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB; 9751 } 9752 9753 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode) 9754 { 9755 switch (mode) { 9756 case XDP_MODE_SKB: 9757 return generic_xdp_install; 9758 case XDP_MODE_DRV: 9759 case XDP_MODE_HW: 9760 return dev->netdev_ops->ndo_bpf; 9761 default: 9762 return NULL; 9763 } 9764 } 9765 9766 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev, 9767 enum bpf_xdp_mode mode) 9768 { 9769 return dev->xdp_state[mode].link; 9770 } 9771 9772 static struct bpf_prog *dev_xdp_prog(struct net_device *dev, 9773 enum bpf_xdp_mode mode) 9774 { 9775 struct bpf_xdp_link *link = dev_xdp_link(dev, mode); 9776 9777 if (link) 9778 return link->link.prog; 9779 return dev->xdp_state[mode].prog; 9780 } 9781 9782 u8 dev_xdp_prog_count(struct net_device *dev) 9783 { 9784 u8 count = 0; 9785 int i; 9786 9787 for (i = 0; i < __MAX_XDP_MODE; i++) 9788 if (dev->xdp_state[i].prog || dev->xdp_state[i].link) 9789 count++; 9790 return count; 9791 } 9792 EXPORT_SYMBOL_GPL(dev_xdp_prog_count); 9793 9794 u8 dev_xdp_sb_prog_count(struct net_device *dev) 9795 { 9796 u8 count = 0; 9797 int i; 9798 9799 for (i = 0; i < __MAX_XDP_MODE; i++) 9800 if (dev->xdp_state[i].prog && 9801 !dev->xdp_state[i].prog->aux->xdp_has_frags) 9802 count++; 9803 return count; 9804 } 9805 9806 int dev_xdp_propagate(struct net_device *dev, struct netdev_bpf *bpf) 9807 { 9808 if (!dev->netdev_ops->ndo_bpf) 9809 return -EOPNOTSUPP; 9810 9811 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 9812 bpf->command == XDP_SETUP_PROG && 9813 bpf->prog && !bpf->prog->aux->xdp_has_frags) { 9814 NL_SET_ERR_MSG(bpf->extack, 9815 "unable to propagate XDP to device using tcp-data-split"); 9816 return -EBUSY; 9817 } 9818 9819 if (dev_get_min_mp_channel_count(dev)) { 9820 NL_SET_ERR_MSG(bpf->extack, "unable to propagate XDP to device using memory provider"); 9821 return -EBUSY; 9822 } 9823 9824 return dev->netdev_ops->ndo_bpf(dev, bpf); 9825 } 9826 EXPORT_SYMBOL_GPL(dev_xdp_propagate); 9827 9828 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode) 9829 { 9830 struct bpf_prog *prog = dev_xdp_prog(dev, mode); 9831 9832 return prog ? prog->aux->id : 0; 9833 } 9834 9835 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode, 9836 struct bpf_xdp_link *link) 9837 { 9838 dev->xdp_state[mode].link = link; 9839 dev->xdp_state[mode].prog = NULL; 9840 } 9841 9842 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode, 9843 struct bpf_prog *prog) 9844 { 9845 dev->xdp_state[mode].link = NULL; 9846 dev->xdp_state[mode].prog = prog; 9847 } 9848 9849 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode, 9850 bpf_op_t bpf_op, struct netlink_ext_ack *extack, 9851 u32 flags, struct bpf_prog *prog) 9852 { 9853 struct netdev_bpf xdp; 9854 int err; 9855 9856 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 9857 prog && !prog->aux->xdp_has_frags) { 9858 NL_SET_ERR_MSG(extack, "unable to install XDP to device using tcp-data-split"); 9859 return -EBUSY; 9860 } 9861 9862 if (dev_get_min_mp_channel_count(dev)) { 9863 NL_SET_ERR_MSG(extack, "unable to install XDP to device using memory provider"); 9864 return -EBUSY; 9865 } 9866 9867 memset(&xdp, 0, sizeof(xdp)); 9868 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG; 9869 xdp.extack = extack; 9870 xdp.flags = flags; 9871 xdp.prog = prog; 9872 9873 /* Drivers assume refcnt is already incremented (i.e, prog pointer is 9874 * "moved" into driver), so they don't increment it on their own, but 9875 * they do decrement refcnt when program is detached or replaced. 9876 * Given net_device also owns link/prog, we need to bump refcnt here 9877 * to prevent drivers from underflowing it. 9878 */ 9879 if (prog) 9880 bpf_prog_inc(prog); 9881 err = bpf_op(dev, &xdp); 9882 if (err) { 9883 if (prog) 9884 bpf_prog_put(prog); 9885 return err; 9886 } 9887 9888 if (mode != XDP_MODE_HW) 9889 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog); 9890 9891 return 0; 9892 } 9893 9894 static void dev_xdp_uninstall(struct net_device *dev) 9895 { 9896 struct bpf_xdp_link *link; 9897 struct bpf_prog *prog; 9898 enum bpf_xdp_mode mode; 9899 bpf_op_t bpf_op; 9900 9901 ASSERT_RTNL(); 9902 9903 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) { 9904 prog = dev_xdp_prog(dev, mode); 9905 if (!prog) 9906 continue; 9907 9908 bpf_op = dev_xdp_bpf_op(dev, mode); 9909 if (!bpf_op) 9910 continue; 9911 9912 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 9913 9914 /* auto-detach link from net device */ 9915 link = dev_xdp_link(dev, mode); 9916 if (link) 9917 link->dev = NULL; 9918 else 9919 bpf_prog_put(prog); 9920 9921 dev_xdp_set_link(dev, mode, NULL); 9922 } 9923 } 9924 9925 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack, 9926 struct bpf_xdp_link *link, struct bpf_prog *new_prog, 9927 struct bpf_prog *old_prog, u32 flags) 9928 { 9929 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES); 9930 struct bpf_prog *cur_prog; 9931 struct net_device *upper; 9932 struct list_head *iter; 9933 enum bpf_xdp_mode mode; 9934 bpf_op_t bpf_op; 9935 int err; 9936 9937 ASSERT_RTNL(); 9938 9939 /* either link or prog attachment, never both */ 9940 if (link && (new_prog || old_prog)) 9941 return -EINVAL; 9942 /* link supports only XDP mode flags */ 9943 if (link && (flags & ~XDP_FLAGS_MODES)) { 9944 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment"); 9945 return -EINVAL; 9946 } 9947 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */ 9948 if (num_modes > 1) { 9949 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set"); 9950 return -EINVAL; 9951 } 9952 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */ 9953 if (!num_modes && dev_xdp_prog_count(dev) > 1) { 9954 NL_SET_ERR_MSG(extack, 9955 "More than one program loaded, unset mode is ambiguous"); 9956 return -EINVAL; 9957 } 9958 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */ 9959 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) { 9960 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified"); 9961 return -EINVAL; 9962 } 9963 9964 mode = dev_xdp_mode(dev, flags); 9965 /* can't replace attached link */ 9966 if (dev_xdp_link(dev, mode)) { 9967 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link"); 9968 return -EBUSY; 9969 } 9970 9971 /* don't allow if an upper device already has a program */ 9972 netdev_for_each_upper_dev_rcu(dev, upper, iter) { 9973 if (dev_xdp_prog_count(upper) > 0) { 9974 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program"); 9975 return -EEXIST; 9976 } 9977 } 9978 9979 cur_prog = dev_xdp_prog(dev, mode); 9980 /* can't replace attached prog with link */ 9981 if (link && cur_prog) { 9982 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link"); 9983 return -EBUSY; 9984 } 9985 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) { 9986 NL_SET_ERR_MSG(extack, "Active program does not match expected"); 9987 return -EEXIST; 9988 } 9989 9990 /* put effective new program into new_prog */ 9991 if (link) 9992 new_prog = link->link.prog; 9993 9994 if (new_prog) { 9995 bool offload = mode == XDP_MODE_HW; 9996 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB 9997 ? XDP_MODE_DRV : XDP_MODE_SKB; 9998 9999 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) { 10000 NL_SET_ERR_MSG(extack, "XDP program already attached"); 10001 return -EBUSY; 10002 } 10003 if (!offload && dev_xdp_prog(dev, other_mode)) { 10004 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time"); 10005 return -EEXIST; 10006 } 10007 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) { 10008 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported"); 10009 return -EINVAL; 10010 } 10011 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) { 10012 NL_SET_ERR_MSG(extack, "Program bound to different device"); 10013 return -EINVAL; 10014 } 10015 if (bpf_prog_is_dev_bound(new_prog->aux) && mode == XDP_MODE_SKB) { 10016 NL_SET_ERR_MSG(extack, "Can't attach device-bound programs in generic mode"); 10017 return -EINVAL; 10018 } 10019 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) { 10020 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device"); 10021 return -EINVAL; 10022 } 10023 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) { 10024 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device"); 10025 return -EINVAL; 10026 } 10027 } 10028 10029 /* don't call drivers if the effective program didn't change */ 10030 if (new_prog != cur_prog) { 10031 bpf_op = dev_xdp_bpf_op(dev, mode); 10032 if (!bpf_op) { 10033 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode"); 10034 return -EOPNOTSUPP; 10035 } 10036 10037 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog); 10038 if (err) 10039 return err; 10040 } 10041 10042 if (link) 10043 dev_xdp_set_link(dev, mode, link); 10044 else 10045 dev_xdp_set_prog(dev, mode, new_prog); 10046 if (cur_prog) 10047 bpf_prog_put(cur_prog); 10048 10049 return 0; 10050 } 10051 10052 static int dev_xdp_attach_link(struct net_device *dev, 10053 struct netlink_ext_ack *extack, 10054 struct bpf_xdp_link *link) 10055 { 10056 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags); 10057 } 10058 10059 static int dev_xdp_detach_link(struct net_device *dev, 10060 struct netlink_ext_ack *extack, 10061 struct bpf_xdp_link *link) 10062 { 10063 enum bpf_xdp_mode mode; 10064 bpf_op_t bpf_op; 10065 10066 ASSERT_RTNL(); 10067 10068 mode = dev_xdp_mode(dev, link->flags); 10069 if (dev_xdp_link(dev, mode) != link) 10070 return -EINVAL; 10071 10072 bpf_op = dev_xdp_bpf_op(dev, mode); 10073 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 10074 dev_xdp_set_link(dev, mode, NULL); 10075 return 0; 10076 } 10077 10078 static void bpf_xdp_link_release(struct bpf_link *link) 10079 { 10080 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10081 10082 rtnl_lock(); 10083 10084 /* if racing with net_device's tear down, xdp_link->dev might be 10085 * already NULL, in which case link was already auto-detached 10086 */ 10087 if (xdp_link->dev) { 10088 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link)); 10089 xdp_link->dev = NULL; 10090 } 10091 10092 rtnl_unlock(); 10093 } 10094 10095 static int bpf_xdp_link_detach(struct bpf_link *link) 10096 { 10097 bpf_xdp_link_release(link); 10098 return 0; 10099 } 10100 10101 static void bpf_xdp_link_dealloc(struct bpf_link *link) 10102 { 10103 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10104 10105 kfree(xdp_link); 10106 } 10107 10108 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link, 10109 struct seq_file *seq) 10110 { 10111 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10112 u32 ifindex = 0; 10113 10114 rtnl_lock(); 10115 if (xdp_link->dev) 10116 ifindex = xdp_link->dev->ifindex; 10117 rtnl_unlock(); 10118 10119 seq_printf(seq, "ifindex:\t%u\n", ifindex); 10120 } 10121 10122 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link, 10123 struct bpf_link_info *info) 10124 { 10125 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10126 u32 ifindex = 0; 10127 10128 rtnl_lock(); 10129 if (xdp_link->dev) 10130 ifindex = xdp_link->dev->ifindex; 10131 rtnl_unlock(); 10132 10133 info->xdp.ifindex = ifindex; 10134 return 0; 10135 } 10136 10137 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog, 10138 struct bpf_prog *old_prog) 10139 { 10140 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10141 enum bpf_xdp_mode mode; 10142 bpf_op_t bpf_op; 10143 int err = 0; 10144 10145 rtnl_lock(); 10146 10147 /* link might have been auto-released already, so fail */ 10148 if (!xdp_link->dev) { 10149 err = -ENOLINK; 10150 goto out_unlock; 10151 } 10152 10153 if (old_prog && link->prog != old_prog) { 10154 err = -EPERM; 10155 goto out_unlock; 10156 } 10157 old_prog = link->prog; 10158 if (old_prog->type != new_prog->type || 10159 old_prog->expected_attach_type != new_prog->expected_attach_type) { 10160 err = -EINVAL; 10161 goto out_unlock; 10162 } 10163 10164 if (old_prog == new_prog) { 10165 /* no-op, don't disturb drivers */ 10166 bpf_prog_put(new_prog); 10167 goto out_unlock; 10168 } 10169 10170 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags); 10171 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode); 10172 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL, 10173 xdp_link->flags, new_prog); 10174 if (err) 10175 goto out_unlock; 10176 10177 old_prog = xchg(&link->prog, new_prog); 10178 bpf_prog_put(old_prog); 10179 10180 out_unlock: 10181 rtnl_unlock(); 10182 return err; 10183 } 10184 10185 static const struct bpf_link_ops bpf_xdp_link_lops = { 10186 .release = bpf_xdp_link_release, 10187 .dealloc = bpf_xdp_link_dealloc, 10188 .detach = bpf_xdp_link_detach, 10189 .show_fdinfo = bpf_xdp_link_show_fdinfo, 10190 .fill_link_info = bpf_xdp_link_fill_link_info, 10191 .update_prog = bpf_xdp_link_update, 10192 }; 10193 10194 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog) 10195 { 10196 struct net *net = current->nsproxy->net_ns; 10197 struct bpf_link_primer link_primer; 10198 struct netlink_ext_ack extack = {}; 10199 struct bpf_xdp_link *link; 10200 struct net_device *dev; 10201 int err, fd; 10202 10203 rtnl_lock(); 10204 dev = dev_get_by_index(net, attr->link_create.target_ifindex); 10205 if (!dev) { 10206 rtnl_unlock(); 10207 return -EINVAL; 10208 } 10209 10210 link = kzalloc(sizeof(*link), GFP_USER); 10211 if (!link) { 10212 err = -ENOMEM; 10213 goto unlock; 10214 } 10215 10216 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog); 10217 link->dev = dev; 10218 link->flags = attr->link_create.flags; 10219 10220 err = bpf_link_prime(&link->link, &link_primer); 10221 if (err) { 10222 kfree(link); 10223 goto unlock; 10224 } 10225 10226 err = dev_xdp_attach_link(dev, &extack, link); 10227 rtnl_unlock(); 10228 10229 if (err) { 10230 link->dev = NULL; 10231 bpf_link_cleanup(&link_primer); 10232 trace_bpf_xdp_link_attach_failed(extack._msg); 10233 goto out_put_dev; 10234 } 10235 10236 fd = bpf_link_settle(&link_primer); 10237 /* link itself doesn't hold dev's refcnt to not complicate shutdown */ 10238 dev_put(dev); 10239 return fd; 10240 10241 unlock: 10242 rtnl_unlock(); 10243 10244 out_put_dev: 10245 dev_put(dev); 10246 return err; 10247 } 10248 10249 /** 10250 * dev_change_xdp_fd - set or clear a bpf program for a device rx path 10251 * @dev: device 10252 * @extack: netlink extended ack 10253 * @fd: new program fd or negative value to clear 10254 * @expected_fd: old program fd that userspace expects to replace or clear 10255 * @flags: xdp-related flags 10256 * 10257 * Set or clear a bpf program for a device 10258 */ 10259 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, 10260 int fd, int expected_fd, u32 flags) 10261 { 10262 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags); 10263 struct bpf_prog *new_prog = NULL, *old_prog = NULL; 10264 int err; 10265 10266 ASSERT_RTNL(); 10267 10268 if (fd >= 0) { 10269 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP, 10270 mode != XDP_MODE_SKB); 10271 if (IS_ERR(new_prog)) 10272 return PTR_ERR(new_prog); 10273 } 10274 10275 if (expected_fd >= 0) { 10276 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP, 10277 mode != XDP_MODE_SKB); 10278 if (IS_ERR(old_prog)) { 10279 err = PTR_ERR(old_prog); 10280 old_prog = NULL; 10281 goto err_out; 10282 } 10283 } 10284 10285 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags); 10286 10287 err_out: 10288 if (err && new_prog) 10289 bpf_prog_put(new_prog); 10290 if (old_prog) 10291 bpf_prog_put(old_prog); 10292 return err; 10293 } 10294 10295 u32 dev_get_min_mp_channel_count(const struct net_device *dev) 10296 { 10297 int i; 10298 10299 ASSERT_RTNL(); 10300 10301 for (i = dev->real_num_rx_queues - 1; i >= 0; i--) 10302 if (dev->_rx[i].mp_params.mp_priv) 10303 /* The channel count is the idx plus 1. */ 10304 return i + 1; 10305 10306 return 0; 10307 } 10308 10309 /** 10310 * dev_index_reserve() - allocate an ifindex in a namespace 10311 * @net: the applicable net namespace 10312 * @ifindex: requested ifindex, pass %0 to get one allocated 10313 * 10314 * Allocate a ifindex for a new device. Caller must either use the ifindex 10315 * to store the device (via list_netdevice()) or call dev_index_release() 10316 * to give the index up. 10317 * 10318 * Return: a suitable unique value for a new device interface number or -errno. 10319 */ 10320 static int dev_index_reserve(struct net *net, u32 ifindex) 10321 { 10322 int err; 10323 10324 if (ifindex > INT_MAX) { 10325 DEBUG_NET_WARN_ON_ONCE(1); 10326 return -EINVAL; 10327 } 10328 10329 if (!ifindex) 10330 err = xa_alloc_cyclic(&net->dev_by_index, &ifindex, NULL, 10331 xa_limit_31b, &net->ifindex, GFP_KERNEL); 10332 else 10333 err = xa_insert(&net->dev_by_index, ifindex, NULL, GFP_KERNEL); 10334 if (err < 0) 10335 return err; 10336 10337 return ifindex; 10338 } 10339 10340 static void dev_index_release(struct net *net, int ifindex) 10341 { 10342 /* Expect only unused indexes, unlist_netdevice() removes the used */ 10343 WARN_ON(xa_erase(&net->dev_by_index, ifindex)); 10344 } 10345 10346 static bool from_cleanup_net(void) 10347 { 10348 #ifdef CONFIG_NET_NS 10349 return current == cleanup_net_task; 10350 #else 10351 return false; 10352 #endif 10353 } 10354 10355 /* Delayed registration/unregisteration */ 10356 LIST_HEAD(net_todo_list); 10357 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq); 10358 atomic_t dev_unreg_count = ATOMIC_INIT(0); 10359 10360 static void net_set_todo(struct net_device *dev) 10361 { 10362 list_add_tail(&dev->todo_list, &net_todo_list); 10363 } 10364 10365 static netdev_features_t netdev_sync_upper_features(struct net_device *lower, 10366 struct net_device *upper, netdev_features_t features) 10367 { 10368 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10369 netdev_features_t feature; 10370 int feature_bit; 10371 10372 for_each_netdev_feature(upper_disables, feature_bit) { 10373 feature = __NETIF_F_BIT(feature_bit); 10374 if (!(upper->wanted_features & feature) 10375 && (features & feature)) { 10376 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n", 10377 &feature, upper->name); 10378 features &= ~feature; 10379 } 10380 } 10381 10382 return features; 10383 } 10384 10385 static void netdev_sync_lower_features(struct net_device *upper, 10386 struct net_device *lower, netdev_features_t features) 10387 { 10388 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10389 netdev_features_t feature; 10390 int feature_bit; 10391 10392 for_each_netdev_feature(upper_disables, feature_bit) { 10393 feature = __NETIF_F_BIT(feature_bit); 10394 if (!(features & feature) && (lower->features & feature)) { 10395 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n", 10396 &feature, lower->name); 10397 lower->wanted_features &= ~feature; 10398 __netdev_update_features(lower); 10399 10400 if (unlikely(lower->features & feature)) 10401 netdev_WARN(upper, "failed to disable %pNF on %s!\n", 10402 &feature, lower->name); 10403 else 10404 netdev_features_change(lower); 10405 } 10406 } 10407 } 10408 10409 static bool netdev_has_ip_or_hw_csum(netdev_features_t features) 10410 { 10411 netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 10412 bool ip_csum = (features & ip_csum_mask) == ip_csum_mask; 10413 bool hw_csum = features & NETIF_F_HW_CSUM; 10414 10415 return ip_csum || hw_csum; 10416 } 10417 10418 static netdev_features_t netdev_fix_features(struct net_device *dev, 10419 netdev_features_t features) 10420 { 10421 /* Fix illegal checksum combinations */ 10422 if ((features & NETIF_F_HW_CSUM) && 10423 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) { 10424 netdev_warn(dev, "mixed HW and IP checksum settings.\n"); 10425 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM); 10426 } 10427 10428 /* TSO requires that SG is present as well. */ 10429 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) { 10430 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n"); 10431 features &= ~NETIF_F_ALL_TSO; 10432 } 10433 10434 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) && 10435 !(features & NETIF_F_IP_CSUM)) { 10436 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n"); 10437 features &= ~NETIF_F_TSO; 10438 features &= ~NETIF_F_TSO_ECN; 10439 } 10440 10441 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) && 10442 !(features & NETIF_F_IPV6_CSUM)) { 10443 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n"); 10444 features &= ~NETIF_F_TSO6; 10445 } 10446 10447 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */ 10448 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO)) 10449 features &= ~NETIF_F_TSO_MANGLEID; 10450 10451 /* TSO ECN requires that TSO is present as well. */ 10452 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN) 10453 features &= ~NETIF_F_TSO_ECN; 10454 10455 /* Software GSO depends on SG. */ 10456 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) { 10457 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n"); 10458 features &= ~NETIF_F_GSO; 10459 } 10460 10461 /* GSO partial features require GSO partial be set */ 10462 if ((features & dev->gso_partial_features) && 10463 !(features & NETIF_F_GSO_PARTIAL)) { 10464 netdev_dbg(dev, 10465 "Dropping partially supported GSO features since no GSO partial.\n"); 10466 features &= ~dev->gso_partial_features; 10467 } 10468 10469 if (!(features & NETIF_F_RXCSUM)) { 10470 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet 10471 * successfully merged by hardware must also have the 10472 * checksum verified by hardware. If the user does not 10473 * want to enable RXCSUM, logically, we should disable GRO_HW. 10474 */ 10475 if (features & NETIF_F_GRO_HW) { 10476 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n"); 10477 features &= ~NETIF_F_GRO_HW; 10478 } 10479 } 10480 10481 /* LRO/HW-GRO features cannot be combined with RX-FCS */ 10482 if (features & NETIF_F_RXFCS) { 10483 if (features & NETIF_F_LRO) { 10484 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n"); 10485 features &= ~NETIF_F_LRO; 10486 } 10487 10488 if (features & NETIF_F_GRO_HW) { 10489 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n"); 10490 features &= ~NETIF_F_GRO_HW; 10491 } 10492 } 10493 10494 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) { 10495 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n"); 10496 features &= ~NETIF_F_LRO; 10497 } 10498 10499 if ((features & NETIF_F_HW_TLS_TX) && !netdev_has_ip_or_hw_csum(features)) { 10500 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n"); 10501 features &= ~NETIF_F_HW_TLS_TX; 10502 } 10503 10504 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) { 10505 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n"); 10506 features &= ~NETIF_F_HW_TLS_RX; 10507 } 10508 10509 if ((features & NETIF_F_GSO_UDP_L4) && !netdev_has_ip_or_hw_csum(features)) { 10510 netdev_dbg(dev, "Dropping USO feature since no CSUM feature.\n"); 10511 features &= ~NETIF_F_GSO_UDP_L4; 10512 } 10513 10514 return features; 10515 } 10516 10517 int __netdev_update_features(struct net_device *dev) 10518 { 10519 struct net_device *upper, *lower; 10520 netdev_features_t features; 10521 struct list_head *iter; 10522 int err = -1; 10523 10524 ASSERT_RTNL(); 10525 10526 features = netdev_get_wanted_features(dev); 10527 10528 if (dev->netdev_ops->ndo_fix_features) 10529 features = dev->netdev_ops->ndo_fix_features(dev, features); 10530 10531 /* driver might be less strict about feature dependencies */ 10532 features = netdev_fix_features(dev, features); 10533 10534 /* some features can't be enabled if they're off on an upper device */ 10535 netdev_for_each_upper_dev_rcu(dev, upper, iter) 10536 features = netdev_sync_upper_features(dev, upper, features); 10537 10538 if (dev->features == features) 10539 goto sync_lower; 10540 10541 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n", 10542 &dev->features, &features); 10543 10544 if (dev->netdev_ops->ndo_set_features) 10545 err = dev->netdev_ops->ndo_set_features(dev, features); 10546 else 10547 err = 0; 10548 10549 if (unlikely(err < 0)) { 10550 netdev_err(dev, 10551 "set_features() failed (%d); wanted %pNF, left %pNF\n", 10552 err, &features, &dev->features); 10553 /* return non-0 since some features might have changed and 10554 * it's better to fire a spurious notification than miss it 10555 */ 10556 return -1; 10557 } 10558 10559 sync_lower: 10560 /* some features must be disabled on lower devices when disabled 10561 * on an upper device (think: bonding master or bridge) 10562 */ 10563 netdev_for_each_lower_dev(dev, lower, iter) 10564 netdev_sync_lower_features(dev, lower, features); 10565 10566 if (!err) { 10567 netdev_features_t diff = features ^ dev->features; 10568 10569 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) { 10570 /* udp_tunnel_{get,drop}_rx_info both need 10571 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the 10572 * device, or they won't do anything. 10573 * Thus we need to update dev->features 10574 * *before* calling udp_tunnel_get_rx_info, 10575 * but *after* calling udp_tunnel_drop_rx_info. 10576 */ 10577 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) { 10578 dev->features = features; 10579 udp_tunnel_get_rx_info(dev); 10580 } else { 10581 udp_tunnel_drop_rx_info(dev); 10582 } 10583 } 10584 10585 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) { 10586 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) { 10587 dev->features = features; 10588 err |= vlan_get_rx_ctag_filter_info(dev); 10589 } else { 10590 vlan_drop_rx_ctag_filter_info(dev); 10591 } 10592 } 10593 10594 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) { 10595 if (features & NETIF_F_HW_VLAN_STAG_FILTER) { 10596 dev->features = features; 10597 err |= vlan_get_rx_stag_filter_info(dev); 10598 } else { 10599 vlan_drop_rx_stag_filter_info(dev); 10600 } 10601 } 10602 10603 dev->features = features; 10604 } 10605 10606 return err < 0 ? 0 : 1; 10607 } 10608 10609 /** 10610 * netdev_update_features - recalculate device features 10611 * @dev: the device to check 10612 * 10613 * Recalculate dev->features set and send notifications if it 10614 * has changed. Should be called after driver or hardware dependent 10615 * conditions might have changed that influence the features. 10616 */ 10617 void netdev_update_features(struct net_device *dev) 10618 { 10619 if (__netdev_update_features(dev)) 10620 netdev_features_change(dev); 10621 } 10622 EXPORT_SYMBOL(netdev_update_features); 10623 10624 /** 10625 * netdev_change_features - recalculate device features 10626 * @dev: the device to check 10627 * 10628 * Recalculate dev->features set and send notifications even 10629 * if they have not changed. Should be called instead of 10630 * netdev_update_features() if also dev->vlan_features might 10631 * have changed to allow the changes to be propagated to stacked 10632 * VLAN devices. 10633 */ 10634 void netdev_change_features(struct net_device *dev) 10635 { 10636 __netdev_update_features(dev); 10637 netdev_features_change(dev); 10638 } 10639 EXPORT_SYMBOL(netdev_change_features); 10640 10641 /** 10642 * netif_stacked_transfer_operstate - transfer operstate 10643 * @rootdev: the root or lower level device to transfer state from 10644 * @dev: the device to transfer operstate to 10645 * 10646 * Transfer operational state from root to device. This is normally 10647 * called when a stacking relationship exists between the root 10648 * device and the device(a leaf device). 10649 */ 10650 void netif_stacked_transfer_operstate(const struct net_device *rootdev, 10651 struct net_device *dev) 10652 { 10653 if (rootdev->operstate == IF_OPER_DORMANT) 10654 netif_dormant_on(dev); 10655 else 10656 netif_dormant_off(dev); 10657 10658 if (rootdev->operstate == IF_OPER_TESTING) 10659 netif_testing_on(dev); 10660 else 10661 netif_testing_off(dev); 10662 10663 if (netif_carrier_ok(rootdev)) 10664 netif_carrier_on(dev); 10665 else 10666 netif_carrier_off(dev); 10667 } 10668 EXPORT_SYMBOL(netif_stacked_transfer_operstate); 10669 10670 static int netif_alloc_rx_queues(struct net_device *dev) 10671 { 10672 unsigned int i, count = dev->num_rx_queues; 10673 struct netdev_rx_queue *rx; 10674 size_t sz = count * sizeof(*rx); 10675 int err = 0; 10676 10677 BUG_ON(count < 1); 10678 10679 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10680 if (!rx) 10681 return -ENOMEM; 10682 10683 dev->_rx = rx; 10684 10685 for (i = 0; i < count; i++) { 10686 rx[i].dev = dev; 10687 10688 /* XDP RX-queue setup */ 10689 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0); 10690 if (err < 0) 10691 goto err_rxq_info; 10692 } 10693 return 0; 10694 10695 err_rxq_info: 10696 /* Rollback successful reg's and free other resources */ 10697 while (i--) 10698 xdp_rxq_info_unreg(&rx[i].xdp_rxq); 10699 kvfree(dev->_rx); 10700 dev->_rx = NULL; 10701 return err; 10702 } 10703 10704 static void netif_free_rx_queues(struct net_device *dev) 10705 { 10706 unsigned int i, count = dev->num_rx_queues; 10707 10708 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */ 10709 if (!dev->_rx) 10710 return; 10711 10712 for (i = 0; i < count; i++) 10713 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq); 10714 10715 kvfree(dev->_rx); 10716 } 10717 10718 static void netdev_init_one_queue(struct net_device *dev, 10719 struct netdev_queue *queue, void *_unused) 10720 { 10721 /* Initialize queue lock */ 10722 spin_lock_init(&queue->_xmit_lock); 10723 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type); 10724 queue->xmit_lock_owner = -1; 10725 netdev_queue_numa_node_write(queue, NUMA_NO_NODE); 10726 queue->dev = dev; 10727 #ifdef CONFIG_BQL 10728 dql_init(&queue->dql, HZ); 10729 #endif 10730 } 10731 10732 static void netif_free_tx_queues(struct net_device *dev) 10733 { 10734 kvfree(dev->_tx); 10735 } 10736 10737 static int netif_alloc_netdev_queues(struct net_device *dev) 10738 { 10739 unsigned int count = dev->num_tx_queues; 10740 struct netdev_queue *tx; 10741 size_t sz = count * sizeof(*tx); 10742 10743 if (count < 1 || count > 0xffff) 10744 return -EINVAL; 10745 10746 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10747 if (!tx) 10748 return -ENOMEM; 10749 10750 dev->_tx = tx; 10751 10752 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL); 10753 spin_lock_init(&dev->tx_global_lock); 10754 10755 return 0; 10756 } 10757 10758 void netif_tx_stop_all_queues(struct net_device *dev) 10759 { 10760 unsigned int i; 10761 10762 for (i = 0; i < dev->num_tx_queues; i++) { 10763 struct netdev_queue *txq = netdev_get_tx_queue(dev, i); 10764 10765 netif_tx_stop_queue(txq); 10766 } 10767 } 10768 EXPORT_SYMBOL(netif_tx_stop_all_queues); 10769 10770 static int netdev_do_alloc_pcpu_stats(struct net_device *dev) 10771 { 10772 void __percpu *v; 10773 10774 /* Drivers implementing ndo_get_peer_dev must support tstat 10775 * accounting, so that skb_do_redirect() can bump the dev's 10776 * RX stats upon network namespace switch. 10777 */ 10778 if (dev->netdev_ops->ndo_get_peer_dev && 10779 dev->pcpu_stat_type != NETDEV_PCPU_STAT_TSTATS) 10780 return -EOPNOTSUPP; 10781 10782 switch (dev->pcpu_stat_type) { 10783 case NETDEV_PCPU_STAT_NONE: 10784 return 0; 10785 case NETDEV_PCPU_STAT_LSTATS: 10786 v = dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats); 10787 break; 10788 case NETDEV_PCPU_STAT_TSTATS: 10789 v = dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats); 10790 break; 10791 case NETDEV_PCPU_STAT_DSTATS: 10792 v = dev->dstats = netdev_alloc_pcpu_stats(struct pcpu_dstats); 10793 break; 10794 default: 10795 return -EINVAL; 10796 } 10797 10798 return v ? 0 : -ENOMEM; 10799 } 10800 10801 static void netdev_do_free_pcpu_stats(struct net_device *dev) 10802 { 10803 switch (dev->pcpu_stat_type) { 10804 case NETDEV_PCPU_STAT_NONE: 10805 return; 10806 case NETDEV_PCPU_STAT_LSTATS: 10807 free_percpu(dev->lstats); 10808 break; 10809 case NETDEV_PCPU_STAT_TSTATS: 10810 free_percpu(dev->tstats); 10811 break; 10812 case NETDEV_PCPU_STAT_DSTATS: 10813 free_percpu(dev->dstats); 10814 break; 10815 } 10816 } 10817 10818 static void netdev_free_phy_link_topology(struct net_device *dev) 10819 { 10820 struct phy_link_topology *topo = dev->link_topo; 10821 10822 if (IS_ENABLED(CONFIG_PHYLIB) && topo) { 10823 xa_destroy(&topo->phys); 10824 kfree(topo); 10825 dev->link_topo = NULL; 10826 } 10827 } 10828 10829 /** 10830 * register_netdevice() - register a network device 10831 * @dev: device to register 10832 * 10833 * Take a prepared network device structure and make it externally accessible. 10834 * A %NETDEV_REGISTER message is sent to the netdev notifier chain. 10835 * Callers must hold the rtnl lock - you may want register_netdev() 10836 * instead of this. 10837 */ 10838 int register_netdevice(struct net_device *dev) 10839 { 10840 int ret; 10841 struct net *net = dev_net(dev); 10842 10843 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE < 10844 NETDEV_FEATURE_COUNT); 10845 BUG_ON(dev_boot_phase); 10846 ASSERT_RTNL(); 10847 10848 might_sleep(); 10849 10850 /* When net_device's are persistent, this will be fatal. */ 10851 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED); 10852 BUG_ON(!net); 10853 10854 ret = ethtool_check_ops(dev->ethtool_ops); 10855 if (ret) 10856 return ret; 10857 10858 /* rss ctx ID 0 is reserved for the default context, start from 1 */ 10859 xa_init_flags(&dev->ethtool->rss_ctx, XA_FLAGS_ALLOC1); 10860 mutex_init(&dev->ethtool->rss_lock); 10861 10862 spin_lock_init(&dev->addr_list_lock); 10863 netdev_set_addr_lockdep_class(dev); 10864 10865 ret = dev_get_valid_name(net, dev, dev->name); 10866 if (ret < 0) 10867 goto out; 10868 10869 ret = -ENOMEM; 10870 dev->name_node = netdev_name_node_head_alloc(dev); 10871 if (!dev->name_node) 10872 goto out; 10873 10874 /* Init, if this function is available */ 10875 if (dev->netdev_ops->ndo_init) { 10876 ret = dev->netdev_ops->ndo_init(dev); 10877 if (ret) { 10878 if (ret > 0) 10879 ret = -EIO; 10880 goto err_free_name; 10881 } 10882 } 10883 10884 if (((dev->hw_features | dev->features) & 10885 NETIF_F_HW_VLAN_CTAG_FILTER) && 10886 (!dev->netdev_ops->ndo_vlan_rx_add_vid || 10887 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) { 10888 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n"); 10889 ret = -EINVAL; 10890 goto err_uninit; 10891 } 10892 10893 ret = netdev_do_alloc_pcpu_stats(dev); 10894 if (ret) 10895 goto err_uninit; 10896 10897 ret = dev_index_reserve(net, dev->ifindex); 10898 if (ret < 0) 10899 goto err_free_pcpu; 10900 dev->ifindex = ret; 10901 10902 /* Transfer changeable features to wanted_features and enable 10903 * software offloads (GSO and GRO). 10904 */ 10905 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF); 10906 dev->features |= NETIF_F_SOFT_FEATURES; 10907 10908 if (dev->udp_tunnel_nic_info) { 10909 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10910 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10911 } 10912 10913 dev->wanted_features = dev->features & dev->hw_features; 10914 10915 if (!(dev->flags & IFF_LOOPBACK)) 10916 dev->hw_features |= NETIF_F_NOCACHE_COPY; 10917 10918 /* If IPv4 TCP segmentation offload is supported we should also 10919 * allow the device to enable segmenting the frame with the option 10920 * of ignoring a static IP ID value. This doesn't enable the 10921 * feature itself but allows the user to enable it later. 10922 */ 10923 if (dev->hw_features & NETIF_F_TSO) 10924 dev->hw_features |= NETIF_F_TSO_MANGLEID; 10925 if (dev->vlan_features & NETIF_F_TSO) 10926 dev->vlan_features |= NETIF_F_TSO_MANGLEID; 10927 if (dev->mpls_features & NETIF_F_TSO) 10928 dev->mpls_features |= NETIF_F_TSO_MANGLEID; 10929 if (dev->hw_enc_features & NETIF_F_TSO) 10930 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID; 10931 10932 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices. 10933 */ 10934 dev->vlan_features |= NETIF_F_HIGHDMA; 10935 10936 /* Make NETIF_F_SG inheritable to tunnel devices. 10937 */ 10938 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL; 10939 10940 /* Make NETIF_F_SG inheritable to MPLS. 10941 */ 10942 dev->mpls_features |= NETIF_F_SG; 10943 10944 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev); 10945 ret = notifier_to_errno(ret); 10946 if (ret) 10947 goto err_ifindex_release; 10948 10949 ret = netdev_register_kobject(dev); 10950 10951 netdev_lock(dev); 10952 WRITE_ONCE(dev->reg_state, ret ? NETREG_UNREGISTERED : NETREG_REGISTERED); 10953 netdev_unlock(dev); 10954 10955 if (ret) 10956 goto err_uninit_notify; 10957 10958 __netdev_update_features(dev); 10959 10960 /* 10961 * Default initial state at registry is that the 10962 * device is present. 10963 */ 10964 10965 set_bit(__LINK_STATE_PRESENT, &dev->state); 10966 10967 linkwatch_init_dev(dev); 10968 10969 dev_init_scheduler(dev); 10970 10971 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL); 10972 list_netdevice(dev); 10973 10974 add_device_randomness(dev->dev_addr, dev->addr_len); 10975 10976 /* If the device has permanent device address, driver should 10977 * set dev_addr and also addr_assign_type should be set to 10978 * NET_ADDR_PERM (default value). 10979 */ 10980 if (dev->addr_assign_type == NET_ADDR_PERM) 10981 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len); 10982 10983 /* Notify protocols, that a new device appeared. */ 10984 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev); 10985 ret = notifier_to_errno(ret); 10986 if (ret) { 10987 /* Expect explicit free_netdev() on failure */ 10988 dev->needs_free_netdev = false; 10989 unregister_netdevice_queue(dev, NULL); 10990 goto out; 10991 } 10992 /* 10993 * Prevent userspace races by waiting until the network 10994 * device is fully setup before sending notifications. 10995 */ 10996 if (!dev->rtnl_link_ops || 10997 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 10998 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 10999 11000 out: 11001 return ret; 11002 11003 err_uninit_notify: 11004 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11005 err_ifindex_release: 11006 dev_index_release(net, dev->ifindex); 11007 err_free_pcpu: 11008 netdev_do_free_pcpu_stats(dev); 11009 err_uninit: 11010 if (dev->netdev_ops->ndo_uninit) 11011 dev->netdev_ops->ndo_uninit(dev); 11012 if (dev->priv_destructor) 11013 dev->priv_destructor(dev); 11014 err_free_name: 11015 netdev_name_node_free(dev->name_node); 11016 goto out; 11017 } 11018 EXPORT_SYMBOL(register_netdevice); 11019 11020 /* Initialize the core of a dummy net device. 11021 * The setup steps dummy netdevs need which normal netdevs get by going 11022 * through register_netdevice(). 11023 */ 11024 static void init_dummy_netdev(struct net_device *dev) 11025 { 11026 /* make sure we BUG if trying to hit standard 11027 * register/unregister code path 11028 */ 11029 dev->reg_state = NETREG_DUMMY; 11030 11031 /* a dummy interface is started by default */ 11032 set_bit(__LINK_STATE_PRESENT, &dev->state); 11033 set_bit(__LINK_STATE_START, &dev->state); 11034 11035 /* Note : We dont allocate pcpu_refcnt for dummy devices, 11036 * because users of this 'device' dont need to change 11037 * its refcount. 11038 */ 11039 } 11040 11041 /** 11042 * register_netdev - register a network device 11043 * @dev: device to register 11044 * 11045 * Take a completed network device structure and add it to the kernel 11046 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier 11047 * chain. 0 is returned on success. A negative errno code is returned 11048 * on a failure to set up the device, or if the name is a duplicate. 11049 * 11050 * This is a wrapper around register_netdevice that takes the rtnl semaphore 11051 * and expands the device name if you passed a format string to 11052 * alloc_netdev. 11053 */ 11054 int register_netdev(struct net_device *dev) 11055 { 11056 struct net *net = dev_net(dev); 11057 int err; 11058 11059 if (rtnl_net_lock_killable(net)) 11060 return -EINTR; 11061 11062 err = register_netdevice(dev); 11063 11064 rtnl_net_unlock(net); 11065 11066 return err; 11067 } 11068 EXPORT_SYMBOL(register_netdev); 11069 11070 int netdev_refcnt_read(const struct net_device *dev) 11071 { 11072 #ifdef CONFIG_PCPU_DEV_REFCNT 11073 int i, refcnt = 0; 11074 11075 for_each_possible_cpu(i) 11076 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i); 11077 return refcnt; 11078 #else 11079 return refcount_read(&dev->dev_refcnt); 11080 #endif 11081 } 11082 EXPORT_SYMBOL(netdev_refcnt_read); 11083 11084 int netdev_unregister_timeout_secs __read_mostly = 10; 11085 11086 #define WAIT_REFS_MIN_MSECS 1 11087 #define WAIT_REFS_MAX_MSECS 250 11088 /** 11089 * netdev_wait_allrefs_any - wait until all references are gone. 11090 * @list: list of net_devices to wait on 11091 * 11092 * This is called when unregistering network devices. 11093 * 11094 * Any protocol or device that holds a reference should register 11095 * for netdevice notification, and cleanup and put back the 11096 * reference if they receive an UNREGISTER event. 11097 * We can get stuck here if buggy protocols don't correctly 11098 * call dev_put. 11099 */ 11100 static struct net_device *netdev_wait_allrefs_any(struct list_head *list) 11101 { 11102 unsigned long rebroadcast_time, warning_time; 11103 struct net_device *dev; 11104 int wait = 0; 11105 11106 rebroadcast_time = warning_time = jiffies; 11107 11108 list_for_each_entry(dev, list, todo_list) 11109 if (netdev_refcnt_read(dev) == 1) 11110 return dev; 11111 11112 while (true) { 11113 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) { 11114 rtnl_lock(); 11115 11116 /* Rebroadcast unregister notification */ 11117 list_for_each_entry(dev, list, todo_list) 11118 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11119 11120 __rtnl_unlock(); 11121 rcu_barrier(); 11122 rtnl_lock(); 11123 11124 list_for_each_entry(dev, list, todo_list) 11125 if (test_bit(__LINK_STATE_LINKWATCH_PENDING, 11126 &dev->state)) { 11127 /* We must not have linkwatch events 11128 * pending on unregister. If this 11129 * happens, we simply run the queue 11130 * unscheduled, resulting in a noop 11131 * for this device. 11132 */ 11133 linkwatch_run_queue(); 11134 break; 11135 } 11136 11137 __rtnl_unlock(); 11138 11139 rebroadcast_time = jiffies; 11140 } 11141 11142 rcu_barrier(); 11143 11144 if (!wait) { 11145 wait = WAIT_REFS_MIN_MSECS; 11146 } else { 11147 msleep(wait); 11148 wait = min(wait << 1, WAIT_REFS_MAX_MSECS); 11149 } 11150 11151 list_for_each_entry(dev, list, todo_list) 11152 if (netdev_refcnt_read(dev) == 1) 11153 return dev; 11154 11155 if (time_after(jiffies, warning_time + 11156 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) { 11157 list_for_each_entry(dev, list, todo_list) { 11158 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n", 11159 dev->name, netdev_refcnt_read(dev)); 11160 ref_tracker_dir_print(&dev->refcnt_tracker, 10); 11161 } 11162 11163 warning_time = jiffies; 11164 } 11165 } 11166 } 11167 11168 /* The sequence is: 11169 * 11170 * rtnl_lock(); 11171 * ... 11172 * register_netdevice(x1); 11173 * register_netdevice(x2); 11174 * ... 11175 * unregister_netdevice(y1); 11176 * unregister_netdevice(y2); 11177 * ... 11178 * rtnl_unlock(); 11179 * free_netdev(y1); 11180 * free_netdev(y2); 11181 * 11182 * We are invoked by rtnl_unlock(). 11183 * This allows us to deal with problems: 11184 * 1) We can delete sysfs objects which invoke hotplug 11185 * without deadlocking with linkwatch via keventd. 11186 * 2) Since we run with the RTNL semaphore not held, we can sleep 11187 * safely in order to wait for the netdev refcnt to drop to zero. 11188 * 11189 * We must not return until all unregister events added during 11190 * the interval the lock was held have been completed. 11191 */ 11192 void netdev_run_todo(void) 11193 { 11194 struct net_device *dev, *tmp; 11195 struct list_head list; 11196 int cnt; 11197 #ifdef CONFIG_LOCKDEP 11198 struct list_head unlink_list; 11199 11200 list_replace_init(&net_unlink_list, &unlink_list); 11201 11202 while (!list_empty(&unlink_list)) { 11203 struct net_device *dev = list_first_entry(&unlink_list, 11204 struct net_device, 11205 unlink_list); 11206 list_del_init(&dev->unlink_list); 11207 dev->nested_level = dev->lower_level - 1; 11208 } 11209 #endif 11210 11211 /* Snapshot list, allow later requests */ 11212 list_replace_init(&net_todo_list, &list); 11213 11214 __rtnl_unlock(); 11215 11216 /* Wait for rcu callbacks to finish before next phase */ 11217 if (!list_empty(&list)) 11218 rcu_barrier(); 11219 11220 list_for_each_entry_safe(dev, tmp, &list, todo_list) { 11221 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) { 11222 netdev_WARN(dev, "run_todo but not unregistering\n"); 11223 list_del(&dev->todo_list); 11224 continue; 11225 } 11226 11227 netdev_lock(dev); 11228 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERED); 11229 netdev_unlock(dev); 11230 linkwatch_sync_dev(dev); 11231 } 11232 11233 cnt = 0; 11234 while (!list_empty(&list)) { 11235 dev = netdev_wait_allrefs_any(&list); 11236 list_del(&dev->todo_list); 11237 11238 /* paranoia */ 11239 BUG_ON(netdev_refcnt_read(dev) != 1); 11240 BUG_ON(!list_empty(&dev->ptype_all)); 11241 BUG_ON(!list_empty(&dev->ptype_specific)); 11242 WARN_ON(rcu_access_pointer(dev->ip_ptr)); 11243 WARN_ON(rcu_access_pointer(dev->ip6_ptr)); 11244 11245 netdev_do_free_pcpu_stats(dev); 11246 if (dev->priv_destructor) 11247 dev->priv_destructor(dev); 11248 if (dev->needs_free_netdev) 11249 free_netdev(dev); 11250 11251 cnt++; 11252 11253 /* Free network device */ 11254 kobject_put(&dev->dev.kobj); 11255 } 11256 if (cnt && atomic_sub_and_test(cnt, &dev_unreg_count)) 11257 wake_up(&netdev_unregistering_wq); 11258 } 11259 11260 /* Collate per-cpu network dstats statistics 11261 * 11262 * Read per-cpu network statistics from dev->dstats and populate the related 11263 * fields in @s. 11264 */ 11265 static void dev_fetch_dstats(struct rtnl_link_stats64 *s, 11266 const struct pcpu_dstats __percpu *dstats) 11267 { 11268 int cpu; 11269 11270 for_each_possible_cpu(cpu) { 11271 u64 rx_packets, rx_bytes, rx_drops; 11272 u64 tx_packets, tx_bytes, tx_drops; 11273 const struct pcpu_dstats *stats; 11274 unsigned int start; 11275 11276 stats = per_cpu_ptr(dstats, cpu); 11277 do { 11278 start = u64_stats_fetch_begin(&stats->syncp); 11279 rx_packets = u64_stats_read(&stats->rx_packets); 11280 rx_bytes = u64_stats_read(&stats->rx_bytes); 11281 rx_drops = u64_stats_read(&stats->rx_drops); 11282 tx_packets = u64_stats_read(&stats->tx_packets); 11283 tx_bytes = u64_stats_read(&stats->tx_bytes); 11284 tx_drops = u64_stats_read(&stats->tx_drops); 11285 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11286 11287 s->rx_packets += rx_packets; 11288 s->rx_bytes += rx_bytes; 11289 s->rx_dropped += rx_drops; 11290 s->tx_packets += tx_packets; 11291 s->tx_bytes += tx_bytes; 11292 s->tx_dropped += tx_drops; 11293 } 11294 } 11295 11296 /* ndo_get_stats64 implementation for dtstats-based accounting. 11297 * 11298 * Populate @s from dev->stats and dev->dstats. This is used internally by the 11299 * core for NETDEV_PCPU_STAT_DSTAT-type stats collection. 11300 */ 11301 static void dev_get_dstats64(const struct net_device *dev, 11302 struct rtnl_link_stats64 *s) 11303 { 11304 netdev_stats_to_stats64(s, &dev->stats); 11305 dev_fetch_dstats(s, dev->dstats); 11306 } 11307 11308 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has 11309 * all the same fields in the same order as net_device_stats, with only 11310 * the type differing, but rtnl_link_stats64 may have additional fields 11311 * at the end for newer counters. 11312 */ 11313 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64, 11314 const struct net_device_stats *netdev_stats) 11315 { 11316 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t); 11317 const atomic_long_t *src = (atomic_long_t *)netdev_stats; 11318 u64 *dst = (u64 *)stats64; 11319 11320 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64)); 11321 for (i = 0; i < n; i++) 11322 dst[i] = (unsigned long)atomic_long_read(&src[i]); 11323 /* zero out counters that only exist in rtnl_link_stats64 */ 11324 memset((char *)stats64 + n * sizeof(u64), 0, 11325 sizeof(*stats64) - n * sizeof(u64)); 11326 } 11327 EXPORT_SYMBOL(netdev_stats_to_stats64); 11328 11329 static __cold struct net_device_core_stats __percpu *netdev_core_stats_alloc( 11330 struct net_device *dev) 11331 { 11332 struct net_device_core_stats __percpu *p; 11333 11334 p = alloc_percpu_gfp(struct net_device_core_stats, 11335 GFP_ATOMIC | __GFP_NOWARN); 11336 11337 if (p && cmpxchg(&dev->core_stats, NULL, p)) 11338 free_percpu(p); 11339 11340 /* This READ_ONCE() pairs with the cmpxchg() above */ 11341 return READ_ONCE(dev->core_stats); 11342 } 11343 11344 noinline void netdev_core_stats_inc(struct net_device *dev, u32 offset) 11345 { 11346 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11347 struct net_device_core_stats __percpu *p = READ_ONCE(dev->core_stats); 11348 unsigned long __percpu *field; 11349 11350 if (unlikely(!p)) { 11351 p = netdev_core_stats_alloc(dev); 11352 if (!p) 11353 return; 11354 } 11355 11356 field = (unsigned long __percpu *)((void __percpu *)p + offset); 11357 this_cpu_inc(*field); 11358 } 11359 EXPORT_SYMBOL_GPL(netdev_core_stats_inc); 11360 11361 /** 11362 * dev_get_stats - get network device statistics 11363 * @dev: device to get statistics from 11364 * @storage: place to store stats 11365 * 11366 * Get network statistics from device. Return @storage. 11367 * The device driver may provide its own method by setting 11368 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats; 11369 * otherwise the internal statistics structure is used. 11370 */ 11371 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev, 11372 struct rtnl_link_stats64 *storage) 11373 { 11374 const struct net_device_ops *ops = dev->netdev_ops; 11375 const struct net_device_core_stats __percpu *p; 11376 11377 /* 11378 * IPv{4,6} and udp tunnels share common stat helpers and use 11379 * different stat type (NETDEV_PCPU_STAT_TSTATS vs 11380 * NETDEV_PCPU_STAT_DSTATS). Ensure the accounting is consistent. 11381 */ 11382 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_bytes) != 11383 offsetof(struct pcpu_dstats, rx_bytes)); 11384 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_packets) != 11385 offsetof(struct pcpu_dstats, rx_packets)); 11386 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_bytes) != 11387 offsetof(struct pcpu_dstats, tx_bytes)); 11388 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_packets) != 11389 offsetof(struct pcpu_dstats, tx_packets)); 11390 11391 if (ops->ndo_get_stats64) { 11392 memset(storage, 0, sizeof(*storage)); 11393 ops->ndo_get_stats64(dev, storage); 11394 } else if (ops->ndo_get_stats) { 11395 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev)); 11396 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_TSTATS) { 11397 dev_get_tstats64(dev, storage); 11398 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_DSTATS) { 11399 dev_get_dstats64(dev, storage); 11400 } else { 11401 netdev_stats_to_stats64(storage, &dev->stats); 11402 } 11403 11404 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11405 p = READ_ONCE(dev->core_stats); 11406 if (p) { 11407 const struct net_device_core_stats *core_stats; 11408 int i; 11409 11410 for_each_possible_cpu(i) { 11411 core_stats = per_cpu_ptr(p, i); 11412 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped); 11413 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped); 11414 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler); 11415 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped); 11416 } 11417 } 11418 return storage; 11419 } 11420 EXPORT_SYMBOL(dev_get_stats); 11421 11422 /** 11423 * dev_fetch_sw_netstats - get per-cpu network device statistics 11424 * @s: place to store stats 11425 * @netstats: per-cpu network stats to read from 11426 * 11427 * Read per-cpu network statistics and populate the related fields in @s. 11428 */ 11429 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s, 11430 const struct pcpu_sw_netstats __percpu *netstats) 11431 { 11432 int cpu; 11433 11434 for_each_possible_cpu(cpu) { 11435 u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 11436 const struct pcpu_sw_netstats *stats; 11437 unsigned int start; 11438 11439 stats = per_cpu_ptr(netstats, cpu); 11440 do { 11441 start = u64_stats_fetch_begin(&stats->syncp); 11442 rx_packets = u64_stats_read(&stats->rx_packets); 11443 rx_bytes = u64_stats_read(&stats->rx_bytes); 11444 tx_packets = u64_stats_read(&stats->tx_packets); 11445 tx_bytes = u64_stats_read(&stats->tx_bytes); 11446 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11447 11448 s->rx_packets += rx_packets; 11449 s->rx_bytes += rx_bytes; 11450 s->tx_packets += tx_packets; 11451 s->tx_bytes += tx_bytes; 11452 } 11453 } 11454 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats); 11455 11456 /** 11457 * dev_get_tstats64 - ndo_get_stats64 implementation 11458 * @dev: device to get statistics from 11459 * @s: place to store stats 11460 * 11461 * Populate @s from dev->stats and dev->tstats. Can be used as 11462 * ndo_get_stats64() callback. 11463 */ 11464 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s) 11465 { 11466 netdev_stats_to_stats64(s, &dev->stats); 11467 dev_fetch_sw_netstats(s, dev->tstats); 11468 } 11469 EXPORT_SYMBOL_GPL(dev_get_tstats64); 11470 11471 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev) 11472 { 11473 struct netdev_queue *queue = dev_ingress_queue(dev); 11474 11475 #ifdef CONFIG_NET_CLS_ACT 11476 if (queue) 11477 return queue; 11478 queue = kzalloc(sizeof(*queue), GFP_KERNEL); 11479 if (!queue) 11480 return NULL; 11481 netdev_init_one_queue(dev, queue, NULL); 11482 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc); 11483 RCU_INIT_POINTER(queue->qdisc_sleeping, &noop_qdisc); 11484 rcu_assign_pointer(dev->ingress_queue, queue); 11485 #endif 11486 return queue; 11487 } 11488 11489 static const struct ethtool_ops default_ethtool_ops; 11490 11491 void netdev_set_default_ethtool_ops(struct net_device *dev, 11492 const struct ethtool_ops *ops) 11493 { 11494 if (dev->ethtool_ops == &default_ethtool_ops) 11495 dev->ethtool_ops = ops; 11496 } 11497 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops); 11498 11499 /** 11500 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default 11501 * @dev: netdev to enable the IRQ coalescing on 11502 * 11503 * Sets a conservative default for SW IRQ coalescing. Users can use 11504 * sysfs attributes to override the default values. 11505 */ 11506 void netdev_sw_irq_coalesce_default_on(struct net_device *dev) 11507 { 11508 WARN_ON(dev->reg_state == NETREG_REGISTERED); 11509 11510 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) { 11511 netdev_set_gro_flush_timeout(dev, 20000); 11512 netdev_set_defer_hard_irqs(dev, 1); 11513 } 11514 } 11515 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on); 11516 11517 /** 11518 * alloc_netdev_mqs - allocate network device 11519 * @sizeof_priv: size of private data to allocate space for 11520 * @name: device name format string 11521 * @name_assign_type: origin of device name 11522 * @setup: callback to initialize device 11523 * @txqs: the number of TX subqueues to allocate 11524 * @rxqs: the number of RX subqueues to allocate 11525 * 11526 * Allocates a struct net_device with private data area for driver use 11527 * and performs basic initialization. Also allocates subqueue structs 11528 * for each queue on the device. 11529 */ 11530 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name, 11531 unsigned char name_assign_type, 11532 void (*setup)(struct net_device *), 11533 unsigned int txqs, unsigned int rxqs) 11534 { 11535 struct net_device *dev; 11536 size_t napi_config_sz; 11537 unsigned int maxqs; 11538 11539 BUG_ON(strlen(name) >= sizeof(dev->name)); 11540 11541 if (txqs < 1) { 11542 pr_err("alloc_netdev: Unable to allocate device with zero queues\n"); 11543 return NULL; 11544 } 11545 11546 if (rxqs < 1) { 11547 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n"); 11548 return NULL; 11549 } 11550 11551 maxqs = max(txqs, rxqs); 11552 11553 dev = kvzalloc(struct_size(dev, priv, sizeof_priv), 11554 GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11555 if (!dev) 11556 return NULL; 11557 11558 dev->priv_len = sizeof_priv; 11559 11560 ref_tracker_dir_init(&dev->refcnt_tracker, 128, name); 11561 #ifdef CONFIG_PCPU_DEV_REFCNT 11562 dev->pcpu_refcnt = alloc_percpu(int); 11563 if (!dev->pcpu_refcnt) 11564 goto free_dev; 11565 __dev_hold(dev); 11566 #else 11567 refcount_set(&dev->dev_refcnt, 1); 11568 #endif 11569 11570 if (dev_addr_init(dev)) 11571 goto free_pcpu; 11572 11573 dev_mc_init(dev); 11574 dev_uc_init(dev); 11575 11576 dev_net_set(dev, &init_net); 11577 11578 dev->gso_max_size = GSO_LEGACY_MAX_SIZE; 11579 dev->xdp_zc_max_segs = 1; 11580 dev->gso_max_segs = GSO_MAX_SEGS; 11581 dev->gro_max_size = GRO_LEGACY_MAX_SIZE; 11582 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE; 11583 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE; 11584 dev->tso_max_size = TSO_LEGACY_MAX_SIZE; 11585 dev->tso_max_segs = TSO_MAX_SEGS; 11586 dev->upper_level = 1; 11587 dev->lower_level = 1; 11588 #ifdef CONFIG_LOCKDEP 11589 dev->nested_level = 0; 11590 INIT_LIST_HEAD(&dev->unlink_list); 11591 #endif 11592 11593 INIT_LIST_HEAD(&dev->napi_list); 11594 INIT_LIST_HEAD(&dev->unreg_list); 11595 INIT_LIST_HEAD(&dev->close_list); 11596 INIT_LIST_HEAD(&dev->link_watch_list); 11597 INIT_LIST_HEAD(&dev->adj_list.upper); 11598 INIT_LIST_HEAD(&dev->adj_list.lower); 11599 INIT_LIST_HEAD(&dev->ptype_all); 11600 INIT_LIST_HEAD(&dev->ptype_specific); 11601 INIT_LIST_HEAD(&dev->net_notifier_list); 11602 #ifdef CONFIG_NET_SCHED 11603 hash_init(dev->qdisc_hash); 11604 #endif 11605 11606 mutex_init(&dev->lock); 11607 11608 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 11609 setup(dev); 11610 11611 if (!dev->tx_queue_len) { 11612 dev->priv_flags |= IFF_NO_QUEUE; 11613 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN; 11614 } 11615 11616 dev->num_tx_queues = txqs; 11617 dev->real_num_tx_queues = txqs; 11618 if (netif_alloc_netdev_queues(dev)) 11619 goto free_all; 11620 11621 dev->num_rx_queues = rxqs; 11622 dev->real_num_rx_queues = rxqs; 11623 if (netif_alloc_rx_queues(dev)) 11624 goto free_all; 11625 dev->ethtool = kzalloc(sizeof(*dev->ethtool), GFP_KERNEL_ACCOUNT); 11626 if (!dev->ethtool) 11627 goto free_all; 11628 11629 dev->cfg = kzalloc(sizeof(*dev->cfg), GFP_KERNEL_ACCOUNT); 11630 if (!dev->cfg) 11631 goto free_all; 11632 dev->cfg_pending = dev->cfg; 11633 11634 napi_config_sz = array_size(maxqs, sizeof(*dev->napi_config)); 11635 dev->napi_config = kvzalloc(napi_config_sz, GFP_KERNEL_ACCOUNT); 11636 if (!dev->napi_config) 11637 goto free_all; 11638 11639 strscpy(dev->name, name); 11640 dev->name_assign_type = name_assign_type; 11641 dev->group = INIT_NETDEV_GROUP; 11642 if (!dev->ethtool_ops) 11643 dev->ethtool_ops = &default_ethtool_ops; 11644 11645 nf_hook_netdev_init(dev); 11646 11647 return dev; 11648 11649 free_all: 11650 free_netdev(dev); 11651 return NULL; 11652 11653 free_pcpu: 11654 #ifdef CONFIG_PCPU_DEV_REFCNT 11655 free_percpu(dev->pcpu_refcnt); 11656 free_dev: 11657 #endif 11658 kvfree(dev); 11659 return NULL; 11660 } 11661 EXPORT_SYMBOL(alloc_netdev_mqs); 11662 11663 static void netdev_napi_exit(struct net_device *dev) 11664 { 11665 if (!list_empty(&dev->napi_list)) { 11666 struct napi_struct *p, *n; 11667 11668 netdev_lock(dev); 11669 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list) 11670 __netif_napi_del_locked(p); 11671 netdev_unlock(dev); 11672 11673 synchronize_net(); 11674 } 11675 11676 kvfree(dev->napi_config); 11677 } 11678 11679 /** 11680 * free_netdev - free network device 11681 * @dev: device 11682 * 11683 * This function does the last stage of destroying an allocated device 11684 * interface. The reference to the device object is released. If this 11685 * is the last reference then it will be freed.Must be called in process 11686 * context. 11687 */ 11688 void free_netdev(struct net_device *dev) 11689 { 11690 might_sleep(); 11691 11692 /* When called immediately after register_netdevice() failed the unwind 11693 * handling may still be dismantling the device. Handle that case by 11694 * deferring the free. 11695 */ 11696 if (dev->reg_state == NETREG_UNREGISTERING) { 11697 ASSERT_RTNL(); 11698 dev->needs_free_netdev = true; 11699 return; 11700 } 11701 11702 WARN_ON(dev->cfg != dev->cfg_pending); 11703 kfree(dev->cfg); 11704 kfree(dev->ethtool); 11705 netif_free_tx_queues(dev); 11706 netif_free_rx_queues(dev); 11707 11708 kfree(rcu_dereference_protected(dev->ingress_queue, 1)); 11709 11710 /* Flush device addresses */ 11711 dev_addr_flush(dev); 11712 11713 netdev_napi_exit(dev); 11714 11715 ref_tracker_dir_exit(&dev->refcnt_tracker); 11716 #ifdef CONFIG_PCPU_DEV_REFCNT 11717 free_percpu(dev->pcpu_refcnt); 11718 dev->pcpu_refcnt = NULL; 11719 #endif 11720 free_percpu(dev->core_stats); 11721 dev->core_stats = NULL; 11722 free_percpu(dev->xdp_bulkq); 11723 dev->xdp_bulkq = NULL; 11724 11725 netdev_free_phy_link_topology(dev); 11726 11727 mutex_destroy(&dev->lock); 11728 11729 /* Compatibility with error handling in drivers */ 11730 if (dev->reg_state == NETREG_UNINITIALIZED || 11731 dev->reg_state == NETREG_DUMMY) { 11732 kvfree(dev); 11733 return; 11734 } 11735 11736 BUG_ON(dev->reg_state != NETREG_UNREGISTERED); 11737 WRITE_ONCE(dev->reg_state, NETREG_RELEASED); 11738 11739 /* will free via device release */ 11740 put_device(&dev->dev); 11741 } 11742 EXPORT_SYMBOL(free_netdev); 11743 11744 /** 11745 * alloc_netdev_dummy - Allocate and initialize a dummy net device. 11746 * @sizeof_priv: size of private data to allocate space for 11747 * 11748 * Return: the allocated net_device on success, NULL otherwise 11749 */ 11750 struct net_device *alloc_netdev_dummy(int sizeof_priv) 11751 { 11752 return alloc_netdev(sizeof_priv, "dummy#", NET_NAME_UNKNOWN, 11753 init_dummy_netdev); 11754 } 11755 EXPORT_SYMBOL_GPL(alloc_netdev_dummy); 11756 11757 /** 11758 * synchronize_net - Synchronize with packet receive processing 11759 * 11760 * Wait for packets currently being received to be done. 11761 * Does not block later packets from starting. 11762 */ 11763 void synchronize_net(void) 11764 { 11765 might_sleep(); 11766 if (from_cleanup_net() || rtnl_is_locked()) 11767 synchronize_rcu_expedited(); 11768 else 11769 synchronize_rcu(); 11770 } 11771 EXPORT_SYMBOL(synchronize_net); 11772 11773 static void netdev_rss_contexts_free(struct net_device *dev) 11774 { 11775 struct ethtool_rxfh_context *ctx; 11776 unsigned long context; 11777 11778 mutex_lock(&dev->ethtool->rss_lock); 11779 xa_for_each(&dev->ethtool->rss_ctx, context, ctx) { 11780 struct ethtool_rxfh_param rxfh; 11781 11782 rxfh.indir = ethtool_rxfh_context_indir(ctx); 11783 rxfh.key = ethtool_rxfh_context_key(ctx); 11784 rxfh.hfunc = ctx->hfunc; 11785 rxfh.input_xfrm = ctx->input_xfrm; 11786 rxfh.rss_context = context; 11787 rxfh.rss_delete = true; 11788 11789 xa_erase(&dev->ethtool->rss_ctx, context); 11790 if (dev->ethtool_ops->create_rxfh_context) 11791 dev->ethtool_ops->remove_rxfh_context(dev, ctx, 11792 context, NULL); 11793 else 11794 dev->ethtool_ops->set_rxfh(dev, &rxfh, NULL); 11795 kfree(ctx); 11796 } 11797 xa_destroy(&dev->ethtool->rss_ctx); 11798 mutex_unlock(&dev->ethtool->rss_lock); 11799 } 11800 11801 /** 11802 * unregister_netdevice_queue - remove device from the kernel 11803 * @dev: device 11804 * @head: list 11805 * 11806 * This function shuts down a device interface and removes it 11807 * from the kernel tables. 11808 * If head not NULL, device is queued to be unregistered later. 11809 * 11810 * Callers must hold the rtnl semaphore. You may want 11811 * unregister_netdev() instead of this. 11812 */ 11813 11814 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head) 11815 { 11816 ASSERT_RTNL(); 11817 11818 if (head) { 11819 list_move_tail(&dev->unreg_list, head); 11820 } else { 11821 LIST_HEAD(single); 11822 11823 list_add(&dev->unreg_list, &single); 11824 unregister_netdevice_many(&single); 11825 } 11826 } 11827 EXPORT_SYMBOL(unregister_netdevice_queue); 11828 11829 void unregister_netdevice_many_notify(struct list_head *head, 11830 u32 portid, const struct nlmsghdr *nlh) 11831 { 11832 struct net_device *dev, *tmp; 11833 LIST_HEAD(close_head); 11834 int cnt = 0; 11835 11836 BUG_ON(dev_boot_phase); 11837 ASSERT_RTNL(); 11838 11839 if (list_empty(head)) 11840 return; 11841 11842 list_for_each_entry_safe(dev, tmp, head, unreg_list) { 11843 /* Some devices call without registering 11844 * for initialization unwind. Remove those 11845 * devices and proceed with the remaining. 11846 */ 11847 if (dev->reg_state == NETREG_UNINITIALIZED) { 11848 pr_debug("unregister_netdevice: device %s/%p never was registered\n", 11849 dev->name, dev); 11850 11851 WARN_ON(1); 11852 list_del(&dev->unreg_list); 11853 continue; 11854 } 11855 dev->dismantle = true; 11856 BUG_ON(dev->reg_state != NETREG_REGISTERED); 11857 } 11858 11859 /* If device is running, close it first. */ 11860 list_for_each_entry(dev, head, unreg_list) 11861 list_add_tail(&dev->close_list, &close_head); 11862 dev_close_many(&close_head, true); 11863 11864 list_for_each_entry(dev, head, unreg_list) { 11865 /* And unlink it from device chain. */ 11866 unlist_netdevice(dev); 11867 netdev_lock(dev); 11868 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERING); 11869 netdev_unlock(dev); 11870 } 11871 flush_all_backlogs(); 11872 11873 synchronize_net(); 11874 11875 list_for_each_entry(dev, head, unreg_list) { 11876 struct sk_buff *skb = NULL; 11877 11878 /* Shutdown queueing discipline. */ 11879 dev_shutdown(dev); 11880 dev_tcx_uninstall(dev); 11881 dev_xdp_uninstall(dev); 11882 bpf_dev_bound_netdev_unregister(dev); 11883 dev_dmabuf_uninstall(dev); 11884 11885 netdev_offload_xstats_disable_all(dev); 11886 11887 /* Notify protocols, that we are about to destroy 11888 * this device. They should clean all the things. 11889 */ 11890 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11891 11892 if (!dev->rtnl_link_ops || 11893 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 11894 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0, 11895 GFP_KERNEL, NULL, 0, 11896 portid, nlh); 11897 11898 /* 11899 * Flush the unicast and multicast chains 11900 */ 11901 dev_uc_flush(dev); 11902 dev_mc_flush(dev); 11903 11904 netdev_name_node_alt_flush(dev); 11905 netdev_name_node_free(dev->name_node); 11906 11907 netdev_rss_contexts_free(dev); 11908 11909 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11910 11911 if (dev->netdev_ops->ndo_uninit) 11912 dev->netdev_ops->ndo_uninit(dev); 11913 11914 mutex_destroy(&dev->ethtool->rss_lock); 11915 11916 net_shaper_flush_netdev(dev); 11917 11918 if (skb) 11919 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh); 11920 11921 /* Notifier chain MUST detach us all upper devices. */ 11922 WARN_ON(netdev_has_any_upper_dev(dev)); 11923 WARN_ON(netdev_has_any_lower_dev(dev)); 11924 11925 /* Remove entries from kobject tree */ 11926 netdev_unregister_kobject(dev); 11927 #ifdef CONFIG_XPS 11928 /* Remove XPS queueing entries */ 11929 netif_reset_xps_queues_gt(dev, 0); 11930 #endif 11931 } 11932 11933 synchronize_net(); 11934 11935 list_for_each_entry(dev, head, unreg_list) { 11936 netdev_put(dev, &dev->dev_registered_tracker); 11937 net_set_todo(dev); 11938 cnt++; 11939 } 11940 atomic_add(cnt, &dev_unreg_count); 11941 11942 list_del(head); 11943 } 11944 11945 /** 11946 * unregister_netdevice_many - unregister many devices 11947 * @head: list of devices 11948 * 11949 * Note: As most callers use a stack allocated list_head, 11950 * we force a list_del() to make sure stack won't be corrupted later. 11951 */ 11952 void unregister_netdevice_many(struct list_head *head) 11953 { 11954 unregister_netdevice_many_notify(head, 0, NULL); 11955 } 11956 EXPORT_SYMBOL(unregister_netdevice_many); 11957 11958 /** 11959 * unregister_netdev - remove device from the kernel 11960 * @dev: device 11961 * 11962 * This function shuts down a device interface and removes it 11963 * from the kernel tables. 11964 * 11965 * This is just a wrapper for unregister_netdevice that takes 11966 * the rtnl semaphore. In general you want to use this and not 11967 * unregister_netdevice. 11968 */ 11969 void unregister_netdev(struct net_device *dev) 11970 { 11971 rtnl_net_dev_lock(dev); 11972 unregister_netdevice(dev); 11973 rtnl_net_dev_unlock(dev); 11974 } 11975 EXPORT_SYMBOL(unregister_netdev); 11976 11977 /** 11978 * __dev_change_net_namespace - move device to different nethost namespace 11979 * @dev: device 11980 * @net: network namespace 11981 * @pat: If not NULL name pattern to try if the current device name 11982 * is already taken in the destination network namespace. 11983 * @new_ifindex: If not zero, specifies device index in the target 11984 * namespace. 11985 * 11986 * This function shuts down a device interface and moves it 11987 * to a new network namespace. On success 0 is returned, on 11988 * a failure a netagive errno code is returned. 11989 * 11990 * Callers must hold the rtnl semaphore. 11991 */ 11992 11993 int __dev_change_net_namespace(struct net_device *dev, struct net *net, 11994 const char *pat, int new_ifindex) 11995 { 11996 struct netdev_name_node *name_node; 11997 struct net *net_old = dev_net(dev); 11998 char new_name[IFNAMSIZ] = {}; 11999 int err, new_nsid; 12000 12001 ASSERT_RTNL(); 12002 12003 /* Don't allow namespace local devices to be moved. */ 12004 err = -EINVAL; 12005 if (dev->netns_local) 12006 goto out; 12007 12008 /* Ensure the device has been registered */ 12009 if (dev->reg_state != NETREG_REGISTERED) 12010 goto out; 12011 12012 /* Get out if there is nothing todo */ 12013 err = 0; 12014 if (net_eq(net_old, net)) 12015 goto out; 12016 12017 /* Pick the destination device name, and ensure 12018 * we can use it in the destination network namespace. 12019 */ 12020 err = -EEXIST; 12021 if (netdev_name_in_use(net, dev->name)) { 12022 /* We get here if we can't use the current device name */ 12023 if (!pat) 12024 goto out; 12025 err = dev_prep_valid_name(net, dev, pat, new_name, EEXIST); 12026 if (err < 0) 12027 goto out; 12028 } 12029 /* Check that none of the altnames conflicts. */ 12030 err = -EEXIST; 12031 netdev_for_each_altname(dev, name_node) 12032 if (netdev_name_in_use(net, name_node->name)) 12033 goto out; 12034 12035 /* Check that new_ifindex isn't used yet. */ 12036 if (new_ifindex) { 12037 err = dev_index_reserve(net, new_ifindex); 12038 if (err < 0) 12039 goto out; 12040 } else { 12041 /* If there is an ifindex conflict assign a new one */ 12042 err = dev_index_reserve(net, dev->ifindex); 12043 if (err == -EBUSY) 12044 err = dev_index_reserve(net, 0); 12045 if (err < 0) 12046 goto out; 12047 new_ifindex = err; 12048 } 12049 12050 /* 12051 * And now a mini version of register_netdevice unregister_netdevice. 12052 */ 12053 12054 /* If device is running close it first. */ 12055 dev_close(dev); 12056 12057 /* And unlink it from device chain */ 12058 unlist_netdevice(dev); 12059 12060 synchronize_net(); 12061 12062 /* Shutdown queueing discipline. */ 12063 dev_shutdown(dev); 12064 12065 /* Notify protocols, that we are about to destroy 12066 * this device. They should clean all the things. 12067 * 12068 * Note that dev->reg_state stays at NETREG_REGISTERED. 12069 * This is wanted because this way 8021q and macvlan know 12070 * the device is just moving and can keep their slaves up. 12071 */ 12072 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 12073 rcu_barrier(); 12074 12075 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL); 12076 12077 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid, 12078 new_ifindex); 12079 12080 /* 12081 * Flush the unicast and multicast chains 12082 */ 12083 dev_uc_flush(dev); 12084 dev_mc_flush(dev); 12085 12086 /* Send a netdev-removed uevent to the old namespace */ 12087 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE); 12088 netdev_adjacent_del_links(dev); 12089 12090 /* Move per-net netdevice notifiers that are following the netdevice */ 12091 move_netdevice_notifiers_dev_net(dev, net); 12092 12093 /* Actually switch the network namespace */ 12094 dev_net_set(dev, net); 12095 dev->ifindex = new_ifindex; 12096 12097 if (new_name[0]) { 12098 /* Rename the netdev to prepared name */ 12099 write_seqlock_bh(&netdev_rename_lock); 12100 strscpy(dev->name, new_name, IFNAMSIZ); 12101 write_sequnlock_bh(&netdev_rename_lock); 12102 } 12103 12104 /* Fixup kobjects */ 12105 dev_set_uevent_suppress(&dev->dev, 1); 12106 err = device_rename(&dev->dev, dev->name); 12107 dev_set_uevent_suppress(&dev->dev, 0); 12108 WARN_ON(err); 12109 12110 /* Send a netdev-add uevent to the new namespace */ 12111 kobject_uevent(&dev->dev.kobj, KOBJ_ADD); 12112 netdev_adjacent_add_links(dev); 12113 12114 /* Adapt owner in case owning user namespace of target network 12115 * namespace is different from the original one. 12116 */ 12117 err = netdev_change_owner(dev, net_old, net); 12118 WARN_ON(err); 12119 12120 /* Add the device back in the hashes */ 12121 list_netdevice(dev); 12122 12123 /* Notify protocols, that a new device appeared. */ 12124 call_netdevice_notifiers(NETDEV_REGISTER, dev); 12125 12126 /* 12127 * Prevent userspace races by waiting until the network 12128 * device is fully setup before sending notifications. 12129 */ 12130 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 12131 12132 synchronize_net(); 12133 err = 0; 12134 out: 12135 return err; 12136 } 12137 EXPORT_SYMBOL_GPL(__dev_change_net_namespace); 12138 12139 static int dev_cpu_dead(unsigned int oldcpu) 12140 { 12141 struct sk_buff **list_skb; 12142 struct sk_buff *skb; 12143 unsigned int cpu; 12144 struct softnet_data *sd, *oldsd, *remsd = NULL; 12145 12146 local_irq_disable(); 12147 cpu = smp_processor_id(); 12148 sd = &per_cpu(softnet_data, cpu); 12149 oldsd = &per_cpu(softnet_data, oldcpu); 12150 12151 /* Find end of our completion_queue. */ 12152 list_skb = &sd->completion_queue; 12153 while (*list_skb) 12154 list_skb = &(*list_skb)->next; 12155 /* Append completion queue from offline CPU. */ 12156 *list_skb = oldsd->completion_queue; 12157 oldsd->completion_queue = NULL; 12158 12159 /* Append output queue from offline CPU. */ 12160 if (oldsd->output_queue) { 12161 *sd->output_queue_tailp = oldsd->output_queue; 12162 sd->output_queue_tailp = oldsd->output_queue_tailp; 12163 oldsd->output_queue = NULL; 12164 oldsd->output_queue_tailp = &oldsd->output_queue; 12165 } 12166 /* Append NAPI poll list from offline CPU, with one exception : 12167 * process_backlog() must be called by cpu owning percpu backlog. 12168 * We properly handle process_queue & input_pkt_queue later. 12169 */ 12170 while (!list_empty(&oldsd->poll_list)) { 12171 struct napi_struct *napi = list_first_entry(&oldsd->poll_list, 12172 struct napi_struct, 12173 poll_list); 12174 12175 list_del_init(&napi->poll_list); 12176 if (napi->poll == process_backlog) 12177 napi->state &= NAPIF_STATE_THREADED; 12178 else 12179 ____napi_schedule(sd, napi); 12180 } 12181 12182 raise_softirq_irqoff(NET_TX_SOFTIRQ); 12183 local_irq_enable(); 12184 12185 if (!use_backlog_threads()) { 12186 #ifdef CONFIG_RPS 12187 remsd = oldsd->rps_ipi_list; 12188 oldsd->rps_ipi_list = NULL; 12189 #endif 12190 /* send out pending IPI's on offline CPU */ 12191 net_rps_send_ipi(remsd); 12192 } 12193 12194 /* Process offline CPU's input_pkt_queue */ 12195 while ((skb = __skb_dequeue(&oldsd->process_queue))) { 12196 netif_rx(skb); 12197 rps_input_queue_head_incr(oldsd); 12198 } 12199 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) { 12200 netif_rx(skb); 12201 rps_input_queue_head_incr(oldsd); 12202 } 12203 12204 return 0; 12205 } 12206 12207 /** 12208 * netdev_increment_features - increment feature set by one 12209 * @all: current feature set 12210 * @one: new feature set 12211 * @mask: mask feature set 12212 * 12213 * Computes a new feature set after adding a device with feature set 12214 * @one to the master device with current feature set @all. Will not 12215 * enable anything that is off in @mask. Returns the new feature set. 12216 */ 12217 netdev_features_t netdev_increment_features(netdev_features_t all, 12218 netdev_features_t one, netdev_features_t mask) 12219 { 12220 if (mask & NETIF_F_HW_CSUM) 12221 mask |= NETIF_F_CSUM_MASK; 12222 mask |= NETIF_F_VLAN_CHALLENGED; 12223 12224 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask; 12225 all &= one | ~NETIF_F_ALL_FOR_ALL; 12226 12227 /* If one device supports hw checksumming, set for all. */ 12228 if (all & NETIF_F_HW_CSUM) 12229 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM); 12230 12231 return all; 12232 } 12233 EXPORT_SYMBOL(netdev_increment_features); 12234 12235 static struct hlist_head * __net_init netdev_create_hash(void) 12236 { 12237 int i; 12238 struct hlist_head *hash; 12239 12240 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL); 12241 if (hash != NULL) 12242 for (i = 0; i < NETDEV_HASHENTRIES; i++) 12243 INIT_HLIST_HEAD(&hash[i]); 12244 12245 return hash; 12246 } 12247 12248 /* Initialize per network namespace state */ 12249 static int __net_init netdev_init(struct net *net) 12250 { 12251 BUILD_BUG_ON(GRO_HASH_BUCKETS > 12252 8 * sizeof_field(struct napi_struct, gro_bitmask)); 12253 12254 INIT_LIST_HEAD(&net->dev_base_head); 12255 12256 net->dev_name_head = netdev_create_hash(); 12257 if (net->dev_name_head == NULL) 12258 goto err_name; 12259 12260 net->dev_index_head = netdev_create_hash(); 12261 if (net->dev_index_head == NULL) 12262 goto err_idx; 12263 12264 xa_init_flags(&net->dev_by_index, XA_FLAGS_ALLOC1); 12265 12266 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain); 12267 12268 return 0; 12269 12270 err_idx: 12271 kfree(net->dev_name_head); 12272 err_name: 12273 return -ENOMEM; 12274 } 12275 12276 /** 12277 * netdev_drivername - network driver for the device 12278 * @dev: network device 12279 * 12280 * Determine network driver for device. 12281 */ 12282 const char *netdev_drivername(const struct net_device *dev) 12283 { 12284 const struct device_driver *driver; 12285 const struct device *parent; 12286 const char *empty = ""; 12287 12288 parent = dev->dev.parent; 12289 if (!parent) 12290 return empty; 12291 12292 driver = parent->driver; 12293 if (driver && driver->name) 12294 return driver->name; 12295 return empty; 12296 } 12297 12298 static void __netdev_printk(const char *level, const struct net_device *dev, 12299 struct va_format *vaf) 12300 { 12301 if (dev && dev->dev.parent) { 12302 dev_printk_emit(level[1] - '0', 12303 dev->dev.parent, 12304 "%s %s %s%s: %pV", 12305 dev_driver_string(dev->dev.parent), 12306 dev_name(dev->dev.parent), 12307 netdev_name(dev), netdev_reg_state(dev), 12308 vaf); 12309 } else if (dev) { 12310 printk("%s%s%s: %pV", 12311 level, netdev_name(dev), netdev_reg_state(dev), vaf); 12312 } else { 12313 printk("%s(NULL net_device): %pV", level, vaf); 12314 } 12315 } 12316 12317 void netdev_printk(const char *level, const struct net_device *dev, 12318 const char *format, ...) 12319 { 12320 struct va_format vaf; 12321 va_list args; 12322 12323 va_start(args, format); 12324 12325 vaf.fmt = format; 12326 vaf.va = &args; 12327 12328 __netdev_printk(level, dev, &vaf); 12329 12330 va_end(args); 12331 } 12332 EXPORT_SYMBOL(netdev_printk); 12333 12334 #define define_netdev_printk_level(func, level) \ 12335 void func(const struct net_device *dev, const char *fmt, ...) \ 12336 { \ 12337 struct va_format vaf; \ 12338 va_list args; \ 12339 \ 12340 va_start(args, fmt); \ 12341 \ 12342 vaf.fmt = fmt; \ 12343 vaf.va = &args; \ 12344 \ 12345 __netdev_printk(level, dev, &vaf); \ 12346 \ 12347 va_end(args); \ 12348 } \ 12349 EXPORT_SYMBOL(func); 12350 12351 define_netdev_printk_level(netdev_emerg, KERN_EMERG); 12352 define_netdev_printk_level(netdev_alert, KERN_ALERT); 12353 define_netdev_printk_level(netdev_crit, KERN_CRIT); 12354 define_netdev_printk_level(netdev_err, KERN_ERR); 12355 define_netdev_printk_level(netdev_warn, KERN_WARNING); 12356 define_netdev_printk_level(netdev_notice, KERN_NOTICE); 12357 define_netdev_printk_level(netdev_info, KERN_INFO); 12358 12359 static void __net_exit netdev_exit(struct net *net) 12360 { 12361 kfree(net->dev_name_head); 12362 kfree(net->dev_index_head); 12363 xa_destroy(&net->dev_by_index); 12364 if (net != &init_net) 12365 WARN_ON_ONCE(!list_empty(&net->dev_base_head)); 12366 } 12367 12368 static struct pernet_operations __net_initdata netdev_net_ops = { 12369 .init = netdev_init, 12370 .exit = netdev_exit, 12371 }; 12372 12373 static void __net_exit default_device_exit_net(struct net *net) 12374 { 12375 struct netdev_name_node *name_node, *tmp; 12376 struct net_device *dev, *aux; 12377 /* 12378 * Push all migratable network devices back to the 12379 * initial network namespace 12380 */ 12381 ASSERT_RTNL(); 12382 for_each_netdev_safe(net, dev, aux) { 12383 int err; 12384 char fb_name[IFNAMSIZ]; 12385 12386 /* Ignore unmoveable devices (i.e. loopback) */ 12387 if (dev->netns_local) 12388 continue; 12389 12390 /* Leave virtual devices for the generic cleanup */ 12391 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund) 12392 continue; 12393 12394 /* Push remaining network devices to init_net */ 12395 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex); 12396 if (netdev_name_in_use(&init_net, fb_name)) 12397 snprintf(fb_name, IFNAMSIZ, "dev%%d"); 12398 12399 netdev_for_each_altname_safe(dev, name_node, tmp) 12400 if (netdev_name_in_use(&init_net, name_node->name)) 12401 __netdev_name_node_alt_destroy(name_node); 12402 12403 err = dev_change_net_namespace(dev, &init_net, fb_name); 12404 if (err) { 12405 pr_emerg("%s: failed to move %s to init_net: %d\n", 12406 __func__, dev->name, err); 12407 BUG(); 12408 } 12409 } 12410 } 12411 12412 static void __net_exit default_device_exit_batch(struct list_head *net_list) 12413 { 12414 /* At exit all network devices most be removed from a network 12415 * namespace. Do this in the reverse order of registration. 12416 * Do this across as many network namespaces as possible to 12417 * improve batching efficiency. 12418 */ 12419 struct net_device *dev; 12420 struct net *net; 12421 LIST_HEAD(dev_kill_list); 12422 12423 rtnl_lock(); 12424 list_for_each_entry(net, net_list, exit_list) { 12425 default_device_exit_net(net); 12426 cond_resched(); 12427 } 12428 12429 list_for_each_entry(net, net_list, exit_list) { 12430 for_each_netdev_reverse(net, dev) { 12431 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink) 12432 dev->rtnl_link_ops->dellink(dev, &dev_kill_list); 12433 else 12434 unregister_netdevice_queue(dev, &dev_kill_list); 12435 } 12436 } 12437 unregister_netdevice_many(&dev_kill_list); 12438 rtnl_unlock(); 12439 } 12440 12441 static struct pernet_operations __net_initdata default_device_ops = { 12442 .exit_batch = default_device_exit_batch, 12443 }; 12444 12445 static void __init net_dev_struct_check(void) 12446 { 12447 /* TX read-mostly hotpath */ 12448 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, priv_flags_fast); 12449 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, netdev_ops); 12450 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, header_ops); 12451 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, _tx); 12452 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, real_num_tx_queues); 12453 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_size); 12454 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_ipv4_max_size); 12455 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_segs); 12456 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_partial_features); 12457 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, num_tc); 12458 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, mtu); 12459 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, needed_headroom); 12460 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tc_to_txq); 12461 #ifdef CONFIG_XPS 12462 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, xps_maps); 12463 #endif 12464 #ifdef CONFIG_NETFILTER_EGRESS 12465 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, nf_hooks_egress); 12466 #endif 12467 #ifdef CONFIG_NET_XGRESS 12468 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tcx_egress); 12469 #endif 12470 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_tx, 160); 12471 12472 /* TXRX read-mostly hotpath */ 12473 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, lstats); 12474 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, state); 12475 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, flags); 12476 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, hard_header_len); 12477 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, features); 12478 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, ip6_ptr); 12479 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_txrx, 46); 12480 12481 /* RX read-mostly hotpath */ 12482 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ptype_specific); 12483 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ifindex); 12484 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, real_num_rx_queues); 12485 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, _rx); 12486 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_max_size); 12487 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_ipv4_max_size); 12488 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler); 12489 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler_data); 12490 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, nd_net); 12491 #ifdef CONFIG_NETPOLL 12492 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, npinfo); 12493 #endif 12494 #ifdef CONFIG_NET_XGRESS 12495 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, tcx_ingress); 12496 #endif 12497 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_rx, 92); 12498 } 12499 12500 /* 12501 * Initialize the DEV module. At boot time this walks the device list and 12502 * unhooks any devices that fail to initialise (normally hardware not 12503 * present) and leaves us with a valid list of present and active devices. 12504 * 12505 */ 12506 12507 /* We allocate 256 pages for each CPU if PAGE_SHIFT is 12 */ 12508 #define SYSTEM_PERCPU_PAGE_POOL_SIZE ((1 << 20) / PAGE_SIZE) 12509 12510 static int net_page_pool_create(int cpuid) 12511 { 12512 #if IS_ENABLED(CONFIG_PAGE_POOL) 12513 struct page_pool_params page_pool_params = { 12514 .pool_size = SYSTEM_PERCPU_PAGE_POOL_SIZE, 12515 .flags = PP_FLAG_SYSTEM_POOL, 12516 .nid = cpu_to_mem(cpuid), 12517 }; 12518 struct page_pool *pp_ptr; 12519 int err; 12520 12521 pp_ptr = page_pool_create_percpu(&page_pool_params, cpuid); 12522 if (IS_ERR(pp_ptr)) 12523 return -ENOMEM; 12524 12525 err = xdp_reg_page_pool(pp_ptr); 12526 if (err) { 12527 page_pool_destroy(pp_ptr); 12528 return err; 12529 } 12530 12531 per_cpu(system_page_pool, cpuid) = pp_ptr; 12532 #endif 12533 return 0; 12534 } 12535 12536 static int backlog_napi_should_run(unsigned int cpu) 12537 { 12538 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12539 struct napi_struct *napi = &sd->backlog; 12540 12541 return test_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 12542 } 12543 12544 static void run_backlog_napi(unsigned int cpu) 12545 { 12546 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12547 12548 napi_threaded_poll_loop(&sd->backlog); 12549 } 12550 12551 static void backlog_napi_setup(unsigned int cpu) 12552 { 12553 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12554 struct napi_struct *napi = &sd->backlog; 12555 12556 napi->thread = this_cpu_read(backlog_napi); 12557 set_bit(NAPI_STATE_THREADED, &napi->state); 12558 } 12559 12560 static struct smp_hotplug_thread backlog_threads = { 12561 .store = &backlog_napi, 12562 .thread_should_run = backlog_napi_should_run, 12563 .thread_fn = run_backlog_napi, 12564 .thread_comm = "backlog_napi/%u", 12565 .setup = backlog_napi_setup, 12566 }; 12567 12568 /* 12569 * This is called single threaded during boot, so no need 12570 * to take the rtnl semaphore. 12571 */ 12572 static int __init net_dev_init(void) 12573 { 12574 int i, rc = -ENOMEM; 12575 12576 BUG_ON(!dev_boot_phase); 12577 12578 net_dev_struct_check(); 12579 12580 if (dev_proc_init()) 12581 goto out; 12582 12583 if (netdev_kobject_init()) 12584 goto out; 12585 12586 for (i = 0; i < PTYPE_HASH_SIZE; i++) 12587 INIT_LIST_HEAD(&ptype_base[i]); 12588 12589 if (register_pernet_subsys(&netdev_net_ops)) 12590 goto out; 12591 12592 /* 12593 * Initialise the packet receive queues. 12594 */ 12595 12596 flush_backlogs_fallback = flush_backlogs_alloc(); 12597 if (!flush_backlogs_fallback) 12598 goto out; 12599 12600 for_each_possible_cpu(i) { 12601 struct softnet_data *sd = &per_cpu(softnet_data, i); 12602 12603 skb_queue_head_init(&sd->input_pkt_queue); 12604 skb_queue_head_init(&sd->process_queue); 12605 #ifdef CONFIG_XFRM_OFFLOAD 12606 skb_queue_head_init(&sd->xfrm_backlog); 12607 #endif 12608 INIT_LIST_HEAD(&sd->poll_list); 12609 sd->output_queue_tailp = &sd->output_queue; 12610 #ifdef CONFIG_RPS 12611 INIT_CSD(&sd->csd, rps_trigger_softirq, sd); 12612 sd->cpu = i; 12613 #endif 12614 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd); 12615 spin_lock_init(&sd->defer_lock); 12616 12617 init_gro_hash(&sd->backlog); 12618 sd->backlog.poll = process_backlog; 12619 sd->backlog.weight = weight_p; 12620 INIT_LIST_HEAD(&sd->backlog.poll_list); 12621 12622 if (net_page_pool_create(i)) 12623 goto out; 12624 } 12625 if (use_backlog_threads()) 12626 smpboot_register_percpu_thread(&backlog_threads); 12627 12628 dev_boot_phase = 0; 12629 12630 /* The loopback device is special if any other network devices 12631 * is present in a network namespace the loopback device must 12632 * be present. Since we now dynamically allocate and free the 12633 * loopback device ensure this invariant is maintained by 12634 * keeping the loopback device as the first device on the 12635 * list of network devices. Ensuring the loopback devices 12636 * is the first device that appears and the last network device 12637 * that disappears. 12638 */ 12639 if (register_pernet_device(&loopback_net_ops)) 12640 goto out; 12641 12642 if (register_pernet_device(&default_device_ops)) 12643 goto out; 12644 12645 open_softirq(NET_TX_SOFTIRQ, net_tx_action); 12646 open_softirq(NET_RX_SOFTIRQ, net_rx_action); 12647 12648 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead", 12649 NULL, dev_cpu_dead); 12650 WARN_ON(rc < 0); 12651 rc = 0; 12652 12653 /* avoid static key IPIs to isolated CPUs */ 12654 if (housekeeping_enabled(HK_TYPE_MISC)) 12655 net_enable_timestamp(); 12656 out: 12657 if (rc < 0) { 12658 for_each_possible_cpu(i) { 12659 struct page_pool *pp_ptr; 12660 12661 pp_ptr = per_cpu(system_page_pool, i); 12662 if (!pp_ptr) 12663 continue; 12664 12665 xdp_unreg_page_pool(pp_ptr); 12666 page_pool_destroy(pp_ptr); 12667 per_cpu(system_page_pool, i) = NULL; 12668 } 12669 } 12670 12671 return rc; 12672 } 12673 12674 subsys_initcall(net_dev_init); 12675