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_lock.h> 160 #include <net/netdev_rx_queue.h> 161 #include <net/page_pool/types.h> 162 #include <net/page_pool/helpers.h> 163 #include <net/page_pool/memory_provider.h> 164 #include <net/rps.h> 165 #include <linux/phy_link_topology.h> 166 167 #include "dev.h" 168 #include "devmem.h" 169 #include "net-sysfs.h" 170 171 static DEFINE_SPINLOCK(ptype_lock); 172 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly; 173 174 static int netif_rx_internal(struct sk_buff *skb); 175 static int call_netdevice_notifiers_extack(unsigned long val, 176 struct net_device *dev, 177 struct netlink_ext_ack *extack); 178 179 static DEFINE_MUTEX(ifalias_mutex); 180 181 /* protects napi_hash addition/deletion and napi_gen_id */ 182 static DEFINE_SPINLOCK(napi_hash_lock); 183 184 static unsigned int napi_gen_id = NR_CPUS; 185 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8); 186 187 static inline void dev_base_seq_inc(struct net *net) 188 { 189 unsigned int val = net->dev_base_seq + 1; 190 191 WRITE_ONCE(net->dev_base_seq, val ?: 1); 192 } 193 194 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name) 195 { 196 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ)); 197 198 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)]; 199 } 200 201 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex) 202 { 203 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)]; 204 } 205 206 #ifndef CONFIG_PREEMPT_RT 207 208 static DEFINE_STATIC_KEY_FALSE(use_backlog_threads_key); 209 210 static int __init setup_backlog_napi_threads(char *arg) 211 { 212 static_branch_enable(&use_backlog_threads_key); 213 return 0; 214 } 215 early_param("thread_backlog_napi", setup_backlog_napi_threads); 216 217 static bool use_backlog_threads(void) 218 { 219 return static_branch_unlikely(&use_backlog_threads_key); 220 } 221 222 #else 223 224 static bool use_backlog_threads(void) 225 { 226 return true; 227 } 228 229 #endif 230 231 static inline void backlog_lock_irq_save(struct softnet_data *sd, 232 unsigned long *flags) 233 { 234 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 235 spin_lock_irqsave(&sd->input_pkt_queue.lock, *flags); 236 else 237 local_irq_save(*flags); 238 } 239 240 static inline void backlog_lock_irq_disable(struct softnet_data *sd) 241 { 242 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 243 spin_lock_irq(&sd->input_pkt_queue.lock); 244 else 245 local_irq_disable(); 246 } 247 248 static inline void backlog_unlock_irq_restore(struct softnet_data *sd, 249 unsigned long *flags) 250 { 251 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 252 spin_unlock_irqrestore(&sd->input_pkt_queue.lock, *flags); 253 else 254 local_irq_restore(*flags); 255 } 256 257 static inline void backlog_unlock_irq_enable(struct softnet_data *sd) 258 { 259 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 260 spin_unlock_irq(&sd->input_pkt_queue.lock); 261 else 262 local_irq_enable(); 263 } 264 265 static struct netdev_name_node *netdev_name_node_alloc(struct net_device *dev, 266 const char *name) 267 { 268 struct netdev_name_node *name_node; 269 270 name_node = kmalloc(sizeof(*name_node), GFP_KERNEL); 271 if (!name_node) 272 return NULL; 273 INIT_HLIST_NODE(&name_node->hlist); 274 name_node->dev = dev; 275 name_node->name = name; 276 return name_node; 277 } 278 279 static struct netdev_name_node * 280 netdev_name_node_head_alloc(struct net_device *dev) 281 { 282 struct netdev_name_node *name_node; 283 284 name_node = netdev_name_node_alloc(dev, dev->name); 285 if (!name_node) 286 return NULL; 287 INIT_LIST_HEAD(&name_node->list); 288 return name_node; 289 } 290 291 static void netdev_name_node_free(struct netdev_name_node *name_node) 292 { 293 kfree(name_node); 294 } 295 296 static void netdev_name_node_add(struct net *net, 297 struct netdev_name_node *name_node) 298 { 299 hlist_add_head_rcu(&name_node->hlist, 300 dev_name_hash(net, name_node->name)); 301 } 302 303 static void netdev_name_node_del(struct netdev_name_node *name_node) 304 { 305 hlist_del_rcu(&name_node->hlist); 306 } 307 308 static struct netdev_name_node *netdev_name_node_lookup(struct net *net, 309 const char *name) 310 { 311 struct hlist_head *head = dev_name_hash(net, name); 312 struct netdev_name_node *name_node; 313 314 hlist_for_each_entry(name_node, head, hlist) 315 if (!strcmp(name_node->name, name)) 316 return name_node; 317 return NULL; 318 } 319 320 static struct netdev_name_node *netdev_name_node_lookup_rcu(struct net *net, 321 const char *name) 322 { 323 struct hlist_head *head = dev_name_hash(net, name); 324 struct netdev_name_node *name_node; 325 326 hlist_for_each_entry_rcu(name_node, head, hlist) 327 if (!strcmp(name_node->name, name)) 328 return name_node; 329 return NULL; 330 } 331 332 bool netdev_name_in_use(struct net *net, const char *name) 333 { 334 return netdev_name_node_lookup(net, name); 335 } 336 EXPORT_SYMBOL(netdev_name_in_use); 337 338 int netdev_name_node_alt_create(struct net_device *dev, const char *name) 339 { 340 struct netdev_name_node *name_node; 341 struct net *net = dev_net(dev); 342 343 name_node = netdev_name_node_lookup(net, name); 344 if (name_node) 345 return -EEXIST; 346 name_node = netdev_name_node_alloc(dev, name); 347 if (!name_node) 348 return -ENOMEM; 349 netdev_name_node_add(net, name_node); 350 /* The node that holds dev->name acts as a head of per-device list. */ 351 list_add_tail_rcu(&name_node->list, &dev->name_node->list); 352 353 return 0; 354 } 355 356 static void netdev_name_node_alt_free(struct rcu_head *head) 357 { 358 struct netdev_name_node *name_node = 359 container_of(head, struct netdev_name_node, rcu); 360 361 kfree(name_node->name); 362 netdev_name_node_free(name_node); 363 } 364 365 static void __netdev_name_node_alt_destroy(struct netdev_name_node *name_node) 366 { 367 netdev_name_node_del(name_node); 368 list_del(&name_node->list); 369 call_rcu(&name_node->rcu, netdev_name_node_alt_free); 370 } 371 372 int netdev_name_node_alt_destroy(struct net_device *dev, const char *name) 373 { 374 struct netdev_name_node *name_node; 375 struct net *net = dev_net(dev); 376 377 name_node = netdev_name_node_lookup(net, name); 378 if (!name_node) 379 return -ENOENT; 380 /* lookup might have found our primary name or a name belonging 381 * to another device. 382 */ 383 if (name_node == dev->name_node || name_node->dev != dev) 384 return -EINVAL; 385 386 __netdev_name_node_alt_destroy(name_node); 387 return 0; 388 } 389 390 static void netdev_name_node_alt_flush(struct net_device *dev) 391 { 392 struct netdev_name_node *name_node, *tmp; 393 394 list_for_each_entry_safe(name_node, tmp, &dev->name_node->list, list) { 395 list_del(&name_node->list); 396 netdev_name_node_alt_free(&name_node->rcu); 397 } 398 } 399 400 /* Device list insertion */ 401 static void list_netdevice(struct net_device *dev) 402 { 403 struct netdev_name_node *name_node; 404 struct net *net = dev_net(dev); 405 406 ASSERT_RTNL(); 407 408 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head); 409 netdev_name_node_add(net, dev->name_node); 410 hlist_add_head_rcu(&dev->index_hlist, 411 dev_index_hash(net, dev->ifindex)); 412 413 netdev_for_each_altname(dev, name_node) 414 netdev_name_node_add(net, name_node); 415 416 /* We reserved the ifindex, this can't fail */ 417 WARN_ON(xa_store(&net->dev_by_index, dev->ifindex, dev, GFP_KERNEL)); 418 419 dev_base_seq_inc(net); 420 } 421 422 /* Device list removal 423 * caller must respect a RCU grace period before freeing/reusing dev 424 */ 425 static void unlist_netdevice(struct net_device *dev) 426 { 427 struct netdev_name_node *name_node; 428 struct net *net = dev_net(dev); 429 430 ASSERT_RTNL(); 431 432 xa_erase(&net->dev_by_index, dev->ifindex); 433 434 netdev_for_each_altname(dev, name_node) 435 netdev_name_node_del(name_node); 436 437 /* Unlink dev from the device chain */ 438 list_del_rcu(&dev->dev_list); 439 netdev_name_node_del(dev->name_node); 440 hlist_del_rcu(&dev->index_hlist); 441 442 dev_base_seq_inc(dev_net(dev)); 443 } 444 445 /* 446 * Our notifier list 447 */ 448 449 static RAW_NOTIFIER_HEAD(netdev_chain); 450 451 /* 452 * Device drivers call our routines to queue packets here. We empty the 453 * queue in the local softnet handler. 454 */ 455 456 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data) = { 457 .process_queue_bh_lock = INIT_LOCAL_LOCK(process_queue_bh_lock), 458 }; 459 EXPORT_PER_CPU_SYMBOL(softnet_data); 460 461 /* Page_pool has a lockless array/stack to alloc/recycle pages. 462 * PP consumers must pay attention to run APIs in the appropriate context 463 * (e.g. NAPI context). 464 */ 465 DEFINE_PER_CPU(struct page_pool *, system_page_pool); 466 467 #ifdef CONFIG_LOCKDEP 468 /* 469 * register_netdevice() inits txq->_xmit_lock and sets lockdep class 470 * according to dev->type 471 */ 472 static const unsigned short netdev_lock_type[] = { 473 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25, 474 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET, 475 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM, 476 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP, 477 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD, 478 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25, 479 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP, 480 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD, 481 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI, 482 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE, 483 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET, 484 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL, 485 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM, 486 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE, 487 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE}; 488 489 static const char *const netdev_lock_name[] = { 490 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25", 491 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET", 492 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM", 493 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP", 494 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD", 495 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25", 496 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP", 497 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD", 498 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI", 499 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE", 500 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET", 501 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL", 502 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM", 503 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE", 504 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"}; 505 506 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)]; 507 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)]; 508 509 static inline unsigned short netdev_lock_pos(unsigned short dev_type) 510 { 511 int i; 512 513 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++) 514 if (netdev_lock_type[i] == dev_type) 515 return i; 516 /* the last key is used by default */ 517 return ARRAY_SIZE(netdev_lock_type) - 1; 518 } 519 520 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 521 unsigned short dev_type) 522 { 523 int i; 524 525 i = netdev_lock_pos(dev_type); 526 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i], 527 netdev_lock_name[i]); 528 } 529 530 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 531 { 532 int i; 533 534 i = netdev_lock_pos(dev->type); 535 lockdep_set_class_and_name(&dev->addr_list_lock, 536 &netdev_addr_lock_key[i], 537 netdev_lock_name[i]); 538 } 539 #else 540 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 541 unsigned short dev_type) 542 { 543 } 544 545 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 546 { 547 } 548 #endif 549 550 /******************************************************************************* 551 * 552 * Protocol management and registration routines 553 * 554 *******************************************************************************/ 555 556 557 /* 558 * Add a protocol ID to the list. Now that the input handler is 559 * smarter we can dispense with all the messy stuff that used to be 560 * here. 561 * 562 * BEWARE!!! Protocol handlers, mangling input packets, 563 * MUST BE last in hash buckets and checking protocol handlers 564 * MUST start from promiscuous ptype_all chain in net_bh. 565 * It is true now, do not change it. 566 * Explanation follows: if protocol handler, mangling packet, will 567 * be the first on list, it is not able to sense, that packet 568 * is cloned and should be copied-on-write, so that it will 569 * change it and subsequent readers will get broken packet. 570 * --ANK (980803) 571 */ 572 573 static inline struct list_head *ptype_head(const struct packet_type *pt) 574 { 575 if (pt->type == htons(ETH_P_ALL)) { 576 if (!pt->af_packet_net && !pt->dev) 577 return NULL; 578 579 return pt->dev ? &pt->dev->ptype_all : 580 &pt->af_packet_net->ptype_all; 581 } 582 583 if (pt->dev) 584 return &pt->dev->ptype_specific; 585 586 return pt->af_packet_net ? &pt->af_packet_net->ptype_specific : 587 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK]; 588 } 589 590 /** 591 * dev_add_pack - add packet handler 592 * @pt: packet type declaration 593 * 594 * Add a protocol handler to the networking stack. The passed &packet_type 595 * is linked into kernel lists and may not be freed until it has been 596 * removed from the kernel lists. 597 * 598 * This call does not sleep therefore it can not 599 * guarantee all CPU's that are in middle of receiving packets 600 * will see the new packet type (until the next received packet). 601 */ 602 603 void dev_add_pack(struct packet_type *pt) 604 { 605 struct list_head *head = ptype_head(pt); 606 607 if (WARN_ON_ONCE(!head)) 608 return; 609 610 spin_lock(&ptype_lock); 611 list_add_rcu(&pt->list, head); 612 spin_unlock(&ptype_lock); 613 } 614 EXPORT_SYMBOL(dev_add_pack); 615 616 /** 617 * __dev_remove_pack - remove packet handler 618 * @pt: packet type declaration 619 * 620 * Remove a protocol handler that was previously added to the kernel 621 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 622 * from the kernel lists and can be freed or reused once this function 623 * returns. 624 * 625 * The packet type might still be in use by receivers 626 * and must not be freed until after all the CPU's have gone 627 * through a quiescent state. 628 */ 629 void __dev_remove_pack(struct packet_type *pt) 630 { 631 struct list_head *head = ptype_head(pt); 632 struct packet_type *pt1; 633 634 if (!head) 635 return; 636 637 spin_lock(&ptype_lock); 638 639 list_for_each_entry(pt1, head, list) { 640 if (pt == pt1) { 641 list_del_rcu(&pt->list); 642 goto out; 643 } 644 } 645 646 pr_warn("dev_remove_pack: %p not found\n", pt); 647 out: 648 spin_unlock(&ptype_lock); 649 } 650 EXPORT_SYMBOL(__dev_remove_pack); 651 652 /** 653 * dev_remove_pack - remove packet handler 654 * @pt: packet type declaration 655 * 656 * Remove a protocol handler that was previously added to the kernel 657 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 658 * from the kernel lists and can be freed or reused once this function 659 * returns. 660 * 661 * This call sleeps to guarantee that no CPU is looking at the packet 662 * type after return. 663 */ 664 void dev_remove_pack(struct packet_type *pt) 665 { 666 __dev_remove_pack(pt); 667 668 synchronize_net(); 669 } 670 EXPORT_SYMBOL(dev_remove_pack); 671 672 673 /******************************************************************************* 674 * 675 * Device Interface Subroutines 676 * 677 *******************************************************************************/ 678 679 /** 680 * dev_get_iflink - get 'iflink' value of a interface 681 * @dev: targeted interface 682 * 683 * Indicates the ifindex the interface is linked to. 684 * Physical interfaces have the same 'ifindex' and 'iflink' values. 685 */ 686 687 int dev_get_iflink(const struct net_device *dev) 688 { 689 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink) 690 return dev->netdev_ops->ndo_get_iflink(dev); 691 692 return READ_ONCE(dev->ifindex); 693 } 694 EXPORT_SYMBOL(dev_get_iflink); 695 696 /** 697 * dev_fill_metadata_dst - Retrieve tunnel egress information. 698 * @dev: targeted interface 699 * @skb: The packet. 700 * 701 * For better visibility of tunnel traffic OVS needs to retrieve 702 * egress tunnel information for a packet. Following API allows 703 * user to get this info. 704 */ 705 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb) 706 { 707 struct ip_tunnel_info *info; 708 709 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst) 710 return -EINVAL; 711 712 info = skb_tunnel_info_unclone(skb); 713 if (!info) 714 return -ENOMEM; 715 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX))) 716 return -EINVAL; 717 718 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb); 719 } 720 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst); 721 722 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack) 723 { 724 int k = stack->num_paths++; 725 726 if (WARN_ON_ONCE(k >= NET_DEVICE_PATH_STACK_MAX)) 727 return NULL; 728 729 return &stack->path[k]; 730 } 731 732 int dev_fill_forward_path(const struct net_device *dev, const u8 *daddr, 733 struct net_device_path_stack *stack) 734 { 735 const struct net_device *last_dev; 736 struct net_device_path_ctx ctx = { 737 .dev = dev, 738 }; 739 struct net_device_path *path; 740 int ret = 0; 741 742 memcpy(ctx.daddr, daddr, sizeof(ctx.daddr)); 743 stack->num_paths = 0; 744 while (ctx.dev && ctx.dev->netdev_ops->ndo_fill_forward_path) { 745 last_dev = ctx.dev; 746 path = dev_fwd_path(stack); 747 if (!path) 748 return -1; 749 750 memset(path, 0, sizeof(struct net_device_path)); 751 ret = ctx.dev->netdev_ops->ndo_fill_forward_path(&ctx, path); 752 if (ret < 0) 753 return -1; 754 755 if (WARN_ON_ONCE(last_dev == ctx.dev)) 756 return -1; 757 } 758 759 if (!ctx.dev) 760 return ret; 761 762 path = dev_fwd_path(stack); 763 if (!path) 764 return -1; 765 path->type = DEV_PATH_ETHERNET; 766 path->dev = ctx.dev; 767 768 return ret; 769 } 770 EXPORT_SYMBOL_GPL(dev_fill_forward_path); 771 772 /* must be called under rcu_read_lock(), as we dont take a reference */ 773 static struct napi_struct *napi_by_id(unsigned int napi_id) 774 { 775 unsigned int hash = napi_id % HASH_SIZE(napi_hash); 776 struct napi_struct *napi; 777 778 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node) 779 if (napi->napi_id == napi_id) 780 return napi; 781 782 return NULL; 783 } 784 785 /* must be called under rcu_read_lock(), as we dont take a reference */ 786 static struct napi_struct * 787 netdev_napi_by_id(struct net *net, unsigned int napi_id) 788 { 789 struct napi_struct *napi; 790 791 napi = napi_by_id(napi_id); 792 if (!napi) 793 return NULL; 794 795 if (WARN_ON_ONCE(!napi->dev)) 796 return NULL; 797 if (!net_eq(net, dev_net(napi->dev))) 798 return NULL; 799 800 return napi; 801 } 802 803 /** 804 * netdev_napi_by_id_lock() - find a device by NAPI ID and lock it 805 * @net: the applicable net namespace 806 * @napi_id: ID of a NAPI of a target device 807 * 808 * Find a NAPI instance with @napi_id. Lock its device. 809 * The device must be in %NETREG_REGISTERED state for lookup to succeed. 810 * netdev_unlock() must be called to release it. 811 * 812 * Return: pointer to NAPI, its device with lock held, NULL if not found. 813 */ 814 struct napi_struct * 815 netdev_napi_by_id_lock(struct net *net, unsigned int napi_id) 816 { 817 struct napi_struct *napi; 818 struct net_device *dev; 819 820 rcu_read_lock(); 821 napi = netdev_napi_by_id(net, napi_id); 822 if (!napi || READ_ONCE(napi->dev->reg_state) != NETREG_REGISTERED) { 823 rcu_read_unlock(); 824 return NULL; 825 } 826 827 dev = napi->dev; 828 dev_hold(dev); 829 rcu_read_unlock(); 830 831 dev = __netdev_put_lock(dev, net); 832 if (!dev) 833 return NULL; 834 835 rcu_read_lock(); 836 napi = netdev_napi_by_id(net, napi_id); 837 if (napi && napi->dev != dev) 838 napi = NULL; 839 rcu_read_unlock(); 840 841 if (!napi) 842 netdev_unlock(dev); 843 return napi; 844 } 845 846 /** 847 * __dev_get_by_name - find a device by its name 848 * @net: the applicable net namespace 849 * @name: name to find 850 * 851 * Find an interface by name. Must be called under RTNL semaphore. 852 * If the name is found a pointer to the device is returned. 853 * If the name is not found then %NULL is returned. The 854 * reference counters are not incremented so the caller must be 855 * careful with locks. 856 */ 857 858 struct net_device *__dev_get_by_name(struct net *net, const char *name) 859 { 860 struct netdev_name_node *node_name; 861 862 node_name = netdev_name_node_lookup(net, name); 863 return node_name ? node_name->dev : NULL; 864 } 865 EXPORT_SYMBOL(__dev_get_by_name); 866 867 /** 868 * dev_get_by_name_rcu - find a device by its name 869 * @net: the applicable net namespace 870 * @name: name to find 871 * 872 * Find an interface by name. 873 * If the name is found a pointer to the device is returned. 874 * If the name is not found then %NULL is returned. 875 * The reference counters are not incremented so the caller must be 876 * careful with locks. The caller must hold RCU lock. 877 */ 878 879 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name) 880 { 881 struct netdev_name_node *node_name; 882 883 node_name = netdev_name_node_lookup_rcu(net, name); 884 return node_name ? node_name->dev : NULL; 885 } 886 EXPORT_SYMBOL(dev_get_by_name_rcu); 887 888 /* Deprecated for new users, call netdev_get_by_name() instead */ 889 struct net_device *dev_get_by_name(struct net *net, const char *name) 890 { 891 struct net_device *dev; 892 893 rcu_read_lock(); 894 dev = dev_get_by_name_rcu(net, name); 895 dev_hold(dev); 896 rcu_read_unlock(); 897 return dev; 898 } 899 EXPORT_SYMBOL(dev_get_by_name); 900 901 /** 902 * netdev_get_by_name() - find a device by its name 903 * @net: the applicable net namespace 904 * @name: name to find 905 * @tracker: tracking object for the acquired reference 906 * @gfp: allocation flags for the tracker 907 * 908 * Find an interface by name. This can be called from any 909 * context and does its own locking. The returned handle has 910 * the usage count incremented and the caller must use netdev_put() to 911 * release it when it is no longer needed. %NULL is returned if no 912 * matching device is found. 913 */ 914 struct net_device *netdev_get_by_name(struct net *net, const char *name, 915 netdevice_tracker *tracker, gfp_t gfp) 916 { 917 struct net_device *dev; 918 919 dev = dev_get_by_name(net, name); 920 if (dev) 921 netdev_tracker_alloc(dev, tracker, gfp); 922 return dev; 923 } 924 EXPORT_SYMBOL(netdev_get_by_name); 925 926 /** 927 * __dev_get_by_index - find a device by its ifindex 928 * @net: the applicable net namespace 929 * @ifindex: index of device 930 * 931 * Search for an interface by index. Returns %NULL if the device 932 * is not found or a pointer to the device. The device has not 933 * had its reference counter increased so the caller must be careful 934 * about locking. The caller must hold the RTNL semaphore. 935 */ 936 937 struct net_device *__dev_get_by_index(struct net *net, int ifindex) 938 { 939 struct net_device *dev; 940 struct hlist_head *head = dev_index_hash(net, ifindex); 941 942 hlist_for_each_entry(dev, head, index_hlist) 943 if (dev->ifindex == ifindex) 944 return dev; 945 946 return NULL; 947 } 948 EXPORT_SYMBOL(__dev_get_by_index); 949 950 /** 951 * dev_get_by_index_rcu - find a device by its ifindex 952 * @net: the applicable net namespace 953 * @ifindex: index of device 954 * 955 * Search for an interface by index. Returns %NULL if the device 956 * is not found or a pointer to the device. The device has not 957 * had its reference counter increased so the caller must be careful 958 * about locking. The caller must hold RCU lock. 959 */ 960 961 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex) 962 { 963 struct net_device *dev; 964 struct hlist_head *head = dev_index_hash(net, ifindex); 965 966 hlist_for_each_entry_rcu(dev, head, index_hlist) 967 if (dev->ifindex == ifindex) 968 return dev; 969 970 return NULL; 971 } 972 EXPORT_SYMBOL(dev_get_by_index_rcu); 973 974 /* Deprecated for new users, call netdev_get_by_index() instead */ 975 struct net_device *dev_get_by_index(struct net *net, int ifindex) 976 { 977 struct net_device *dev; 978 979 rcu_read_lock(); 980 dev = dev_get_by_index_rcu(net, ifindex); 981 dev_hold(dev); 982 rcu_read_unlock(); 983 return dev; 984 } 985 EXPORT_SYMBOL(dev_get_by_index); 986 987 /** 988 * netdev_get_by_index() - find a device by its ifindex 989 * @net: the applicable net namespace 990 * @ifindex: index of device 991 * @tracker: tracking object for the acquired reference 992 * @gfp: allocation flags for the tracker 993 * 994 * Search for an interface by index. Returns NULL if the device 995 * is not found or a pointer to the device. The device returned has 996 * had a reference added and the pointer is safe until the user calls 997 * netdev_put() to indicate they have finished with it. 998 */ 999 struct net_device *netdev_get_by_index(struct net *net, int ifindex, 1000 netdevice_tracker *tracker, gfp_t gfp) 1001 { 1002 struct net_device *dev; 1003 1004 dev = dev_get_by_index(net, ifindex); 1005 if (dev) 1006 netdev_tracker_alloc(dev, tracker, gfp); 1007 return dev; 1008 } 1009 EXPORT_SYMBOL(netdev_get_by_index); 1010 1011 /** 1012 * dev_get_by_napi_id - find a device by napi_id 1013 * @napi_id: ID of the NAPI struct 1014 * 1015 * Search for an interface by NAPI ID. Returns %NULL if the device 1016 * is not found or a pointer to the device. The device has not had 1017 * its reference counter increased so the caller must be careful 1018 * about locking. The caller must hold RCU lock. 1019 */ 1020 struct net_device *dev_get_by_napi_id(unsigned int napi_id) 1021 { 1022 struct napi_struct *napi; 1023 1024 WARN_ON_ONCE(!rcu_read_lock_held()); 1025 1026 if (!napi_id_valid(napi_id)) 1027 return NULL; 1028 1029 napi = napi_by_id(napi_id); 1030 1031 return napi ? napi->dev : NULL; 1032 } 1033 1034 /* Release the held reference on the net_device, and if the net_device 1035 * is still registered try to lock the instance lock. If device is being 1036 * unregistered NULL will be returned (but the reference has been released, 1037 * either way!) 1038 * 1039 * This helper is intended for locking net_device after it has been looked up 1040 * using a lockless lookup helper. Lock prevents the instance from going away. 1041 */ 1042 struct net_device *__netdev_put_lock(struct net_device *dev, struct net *net) 1043 { 1044 netdev_lock(dev); 1045 if (dev->reg_state > NETREG_REGISTERED || 1046 dev->moving_ns || !net_eq(dev_net(dev), net)) { 1047 netdev_unlock(dev); 1048 dev_put(dev); 1049 return NULL; 1050 } 1051 dev_put(dev); 1052 return dev; 1053 } 1054 1055 static struct net_device * 1056 __netdev_put_lock_ops_compat(struct net_device *dev, struct net *net) 1057 { 1058 netdev_lock_ops_compat(dev); 1059 if (dev->reg_state > NETREG_REGISTERED || 1060 dev->moving_ns || !net_eq(dev_net(dev), net)) { 1061 netdev_unlock_ops_compat(dev); 1062 dev_put(dev); 1063 return NULL; 1064 } 1065 dev_put(dev); 1066 return dev; 1067 } 1068 1069 /** 1070 * netdev_get_by_index_lock() - find a device by its ifindex 1071 * @net: the applicable net namespace 1072 * @ifindex: index of device 1073 * 1074 * Search for an interface by index. If a valid device 1075 * with @ifindex is found it will be returned with netdev->lock held. 1076 * netdev_unlock() must be called to release it. 1077 * 1078 * Return: pointer to a device with lock held, NULL if not found. 1079 */ 1080 struct net_device *netdev_get_by_index_lock(struct net *net, int ifindex) 1081 { 1082 struct net_device *dev; 1083 1084 dev = dev_get_by_index(net, ifindex); 1085 if (!dev) 1086 return NULL; 1087 1088 return __netdev_put_lock(dev, net); 1089 } 1090 1091 struct net_device * 1092 netdev_get_by_index_lock_ops_compat(struct net *net, int ifindex) 1093 { 1094 struct net_device *dev; 1095 1096 dev = dev_get_by_index(net, ifindex); 1097 if (!dev) 1098 return NULL; 1099 1100 return __netdev_put_lock_ops_compat(dev, net); 1101 } 1102 1103 struct net_device * 1104 netdev_xa_find_lock(struct net *net, struct net_device *dev, 1105 unsigned long *index) 1106 { 1107 if (dev) 1108 netdev_unlock(dev); 1109 1110 do { 1111 rcu_read_lock(); 1112 dev = xa_find(&net->dev_by_index, index, ULONG_MAX, XA_PRESENT); 1113 if (!dev) { 1114 rcu_read_unlock(); 1115 return NULL; 1116 } 1117 dev_hold(dev); 1118 rcu_read_unlock(); 1119 1120 dev = __netdev_put_lock(dev, net); 1121 if (dev) 1122 return dev; 1123 1124 (*index)++; 1125 } while (true); 1126 } 1127 1128 struct net_device * 1129 netdev_xa_find_lock_ops_compat(struct net *net, struct net_device *dev, 1130 unsigned long *index) 1131 { 1132 if (dev) 1133 netdev_unlock_ops_compat(dev); 1134 1135 do { 1136 rcu_read_lock(); 1137 dev = xa_find(&net->dev_by_index, index, ULONG_MAX, XA_PRESENT); 1138 if (!dev) { 1139 rcu_read_unlock(); 1140 return NULL; 1141 } 1142 dev_hold(dev); 1143 rcu_read_unlock(); 1144 1145 dev = __netdev_put_lock_ops_compat(dev, net); 1146 if (dev) 1147 return dev; 1148 1149 (*index)++; 1150 } while (true); 1151 } 1152 1153 static DEFINE_SEQLOCK(netdev_rename_lock); 1154 1155 void netdev_copy_name(struct net_device *dev, char *name) 1156 { 1157 unsigned int seq; 1158 1159 do { 1160 seq = read_seqbegin(&netdev_rename_lock); 1161 strscpy(name, dev->name, IFNAMSIZ); 1162 } while (read_seqretry(&netdev_rename_lock, seq)); 1163 } 1164 1165 /** 1166 * netdev_get_name - get a netdevice name, knowing its ifindex. 1167 * @net: network namespace 1168 * @name: a pointer to the buffer where the name will be stored. 1169 * @ifindex: the ifindex of the interface to get the name from. 1170 */ 1171 int netdev_get_name(struct net *net, char *name, int ifindex) 1172 { 1173 struct net_device *dev; 1174 int ret; 1175 1176 rcu_read_lock(); 1177 1178 dev = dev_get_by_index_rcu(net, ifindex); 1179 if (!dev) { 1180 ret = -ENODEV; 1181 goto out; 1182 } 1183 1184 netdev_copy_name(dev, name); 1185 1186 ret = 0; 1187 out: 1188 rcu_read_unlock(); 1189 return ret; 1190 } 1191 1192 static bool dev_addr_cmp(struct net_device *dev, unsigned short type, 1193 const char *ha) 1194 { 1195 return dev->type == type && !memcmp(dev->dev_addr, ha, dev->addr_len); 1196 } 1197 1198 /** 1199 * dev_getbyhwaddr_rcu - find a device by its hardware address 1200 * @net: the applicable net namespace 1201 * @type: media type of device 1202 * @ha: hardware address 1203 * 1204 * Search for an interface by MAC address. Returns NULL if the device 1205 * is not found or a pointer to the device. 1206 * The caller must hold RCU. 1207 * The returned device has not had its ref count increased 1208 * and the caller must therefore be careful about locking 1209 * 1210 */ 1211 1212 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type, 1213 const char *ha) 1214 { 1215 struct net_device *dev; 1216 1217 for_each_netdev_rcu(net, dev) 1218 if (dev_addr_cmp(dev, type, ha)) 1219 return dev; 1220 1221 return NULL; 1222 } 1223 EXPORT_SYMBOL(dev_getbyhwaddr_rcu); 1224 1225 /** 1226 * dev_getbyhwaddr() - find a device by its hardware address 1227 * @net: the applicable net namespace 1228 * @type: media type of device 1229 * @ha: hardware address 1230 * 1231 * Similar to dev_getbyhwaddr_rcu(), but the owner needs to hold 1232 * rtnl_lock. 1233 * 1234 * Context: rtnl_lock() must be held. 1235 * Return: pointer to the net_device, or NULL if not found 1236 */ 1237 struct net_device *dev_getbyhwaddr(struct net *net, unsigned short type, 1238 const char *ha) 1239 { 1240 struct net_device *dev; 1241 1242 ASSERT_RTNL(); 1243 for_each_netdev(net, dev) 1244 if (dev_addr_cmp(dev, type, ha)) 1245 return dev; 1246 1247 return NULL; 1248 } 1249 EXPORT_SYMBOL(dev_getbyhwaddr); 1250 1251 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type) 1252 { 1253 struct net_device *dev, *ret = NULL; 1254 1255 rcu_read_lock(); 1256 for_each_netdev_rcu(net, dev) 1257 if (dev->type == type) { 1258 dev_hold(dev); 1259 ret = dev; 1260 break; 1261 } 1262 rcu_read_unlock(); 1263 return ret; 1264 } 1265 EXPORT_SYMBOL(dev_getfirstbyhwtype); 1266 1267 /** 1268 * __dev_get_by_flags - find any device with given flags 1269 * @net: the applicable net namespace 1270 * @if_flags: IFF_* values 1271 * @mask: bitmask of bits in if_flags to check 1272 * 1273 * Search for any interface with the given flags. Returns NULL if a device 1274 * is not found or a pointer to the device. Must be called inside 1275 * rtnl_lock(), and result refcount is unchanged. 1276 */ 1277 1278 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags, 1279 unsigned short mask) 1280 { 1281 struct net_device *dev, *ret; 1282 1283 ASSERT_RTNL(); 1284 1285 ret = NULL; 1286 for_each_netdev(net, dev) { 1287 if (((dev->flags ^ if_flags) & mask) == 0) { 1288 ret = dev; 1289 break; 1290 } 1291 } 1292 return ret; 1293 } 1294 EXPORT_SYMBOL(__dev_get_by_flags); 1295 1296 /** 1297 * dev_valid_name - check if name is okay for network device 1298 * @name: name string 1299 * 1300 * Network device names need to be valid file names to 1301 * allow sysfs to work. We also disallow any kind of 1302 * whitespace. 1303 */ 1304 bool dev_valid_name(const char *name) 1305 { 1306 if (*name == '\0') 1307 return false; 1308 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ) 1309 return false; 1310 if (!strcmp(name, ".") || !strcmp(name, "..")) 1311 return false; 1312 1313 while (*name) { 1314 if (*name == '/' || *name == ':' || isspace(*name)) 1315 return false; 1316 name++; 1317 } 1318 return true; 1319 } 1320 EXPORT_SYMBOL(dev_valid_name); 1321 1322 /** 1323 * __dev_alloc_name - allocate a name for a device 1324 * @net: network namespace to allocate the device name in 1325 * @name: name format string 1326 * @res: result name string 1327 * 1328 * Passed a format string - eg "lt%d" it will try and find a suitable 1329 * id. It scans list of devices to build up a free map, then chooses 1330 * the first empty slot. The caller must hold the dev_base or rtnl lock 1331 * while allocating the name and adding the device in order to avoid 1332 * duplicates. 1333 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1334 * Returns the number of the unit assigned or a negative errno code. 1335 */ 1336 1337 static int __dev_alloc_name(struct net *net, const char *name, char *res) 1338 { 1339 int i = 0; 1340 const char *p; 1341 const int max_netdevices = 8*PAGE_SIZE; 1342 unsigned long *inuse; 1343 struct net_device *d; 1344 char buf[IFNAMSIZ]; 1345 1346 /* Verify the string as this thing may have come from the user. 1347 * There must be one "%d" and no other "%" characters. 1348 */ 1349 p = strchr(name, '%'); 1350 if (!p || p[1] != 'd' || strchr(p + 2, '%')) 1351 return -EINVAL; 1352 1353 /* Use one page as a bit array of possible slots */ 1354 inuse = bitmap_zalloc(max_netdevices, GFP_ATOMIC); 1355 if (!inuse) 1356 return -ENOMEM; 1357 1358 for_each_netdev(net, d) { 1359 struct netdev_name_node *name_node; 1360 1361 netdev_for_each_altname(d, name_node) { 1362 if (!sscanf(name_node->name, name, &i)) 1363 continue; 1364 if (i < 0 || i >= max_netdevices) 1365 continue; 1366 1367 /* avoid cases where sscanf is not exact inverse of printf */ 1368 snprintf(buf, IFNAMSIZ, name, i); 1369 if (!strncmp(buf, name_node->name, IFNAMSIZ)) 1370 __set_bit(i, inuse); 1371 } 1372 if (!sscanf(d->name, name, &i)) 1373 continue; 1374 if (i < 0 || i >= max_netdevices) 1375 continue; 1376 1377 /* avoid cases where sscanf is not exact inverse of printf */ 1378 snprintf(buf, IFNAMSIZ, name, i); 1379 if (!strncmp(buf, d->name, IFNAMSIZ)) 1380 __set_bit(i, inuse); 1381 } 1382 1383 i = find_first_zero_bit(inuse, max_netdevices); 1384 bitmap_free(inuse); 1385 if (i == max_netdevices) 1386 return -ENFILE; 1387 1388 /* 'res' and 'name' could overlap, use 'buf' as an intermediate buffer */ 1389 strscpy(buf, name, IFNAMSIZ); 1390 snprintf(res, IFNAMSIZ, buf, i); 1391 return i; 1392 } 1393 1394 /* Returns negative errno or allocated unit id (see __dev_alloc_name()) */ 1395 static int dev_prep_valid_name(struct net *net, struct net_device *dev, 1396 const char *want_name, char *out_name, 1397 int dup_errno) 1398 { 1399 if (!dev_valid_name(want_name)) 1400 return -EINVAL; 1401 1402 if (strchr(want_name, '%')) 1403 return __dev_alloc_name(net, want_name, out_name); 1404 1405 if (netdev_name_in_use(net, want_name)) 1406 return -dup_errno; 1407 if (out_name != want_name) 1408 strscpy(out_name, want_name, IFNAMSIZ); 1409 return 0; 1410 } 1411 1412 /** 1413 * dev_alloc_name - allocate a name for a device 1414 * @dev: device 1415 * @name: name format string 1416 * 1417 * Passed a format string - eg "lt%d" it will try and find a suitable 1418 * id. It scans list of devices to build up a free map, then chooses 1419 * the first empty slot. The caller must hold the dev_base or rtnl lock 1420 * while allocating the name and adding the device in order to avoid 1421 * duplicates. 1422 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1423 * Returns the number of the unit assigned or a negative errno code. 1424 */ 1425 1426 int dev_alloc_name(struct net_device *dev, const char *name) 1427 { 1428 return dev_prep_valid_name(dev_net(dev), dev, name, dev->name, ENFILE); 1429 } 1430 EXPORT_SYMBOL(dev_alloc_name); 1431 1432 static int dev_get_valid_name(struct net *net, struct net_device *dev, 1433 const char *name) 1434 { 1435 int ret; 1436 1437 ret = dev_prep_valid_name(net, dev, name, dev->name, EEXIST); 1438 return ret < 0 ? ret : 0; 1439 } 1440 1441 int netif_change_name(struct net_device *dev, const char *newname) 1442 { 1443 struct net *net = dev_net(dev); 1444 unsigned char old_assign_type; 1445 char oldname[IFNAMSIZ]; 1446 int err = 0; 1447 int ret; 1448 1449 ASSERT_RTNL_NET(net); 1450 1451 if (!strncmp(newname, dev->name, IFNAMSIZ)) 1452 return 0; 1453 1454 memcpy(oldname, dev->name, IFNAMSIZ); 1455 1456 write_seqlock_bh(&netdev_rename_lock); 1457 err = dev_get_valid_name(net, dev, newname); 1458 write_sequnlock_bh(&netdev_rename_lock); 1459 1460 if (err < 0) 1461 return err; 1462 1463 if (oldname[0] && !strchr(oldname, '%')) 1464 netdev_info(dev, "renamed from %s%s\n", oldname, 1465 dev->flags & IFF_UP ? " (while UP)" : ""); 1466 1467 old_assign_type = dev->name_assign_type; 1468 WRITE_ONCE(dev->name_assign_type, NET_NAME_RENAMED); 1469 1470 rollback: 1471 ret = device_rename(&dev->dev, dev->name); 1472 if (ret) { 1473 write_seqlock_bh(&netdev_rename_lock); 1474 memcpy(dev->name, oldname, IFNAMSIZ); 1475 write_sequnlock_bh(&netdev_rename_lock); 1476 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1477 return ret; 1478 } 1479 1480 netdev_adjacent_rename_links(dev, oldname); 1481 1482 netdev_name_node_del(dev->name_node); 1483 1484 synchronize_net(); 1485 1486 netdev_name_node_add(net, dev->name_node); 1487 1488 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev); 1489 ret = notifier_to_errno(ret); 1490 1491 if (ret) { 1492 /* err >= 0 after dev_alloc_name() or stores the first errno */ 1493 if (err >= 0) { 1494 err = ret; 1495 write_seqlock_bh(&netdev_rename_lock); 1496 memcpy(dev->name, oldname, IFNAMSIZ); 1497 write_sequnlock_bh(&netdev_rename_lock); 1498 memcpy(oldname, newname, IFNAMSIZ); 1499 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1500 old_assign_type = NET_NAME_RENAMED; 1501 goto rollback; 1502 } else { 1503 netdev_err(dev, "name change rollback failed: %d\n", 1504 ret); 1505 } 1506 } 1507 1508 return err; 1509 } 1510 1511 int netif_set_alias(struct net_device *dev, const char *alias, size_t len) 1512 { 1513 struct dev_ifalias *new_alias = NULL; 1514 1515 if (len >= IFALIASZ) 1516 return -EINVAL; 1517 1518 if (len) { 1519 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL); 1520 if (!new_alias) 1521 return -ENOMEM; 1522 1523 memcpy(new_alias->ifalias, alias, len); 1524 new_alias->ifalias[len] = 0; 1525 } 1526 1527 mutex_lock(&ifalias_mutex); 1528 new_alias = rcu_replace_pointer(dev->ifalias, new_alias, 1529 mutex_is_locked(&ifalias_mutex)); 1530 mutex_unlock(&ifalias_mutex); 1531 1532 if (new_alias) 1533 kfree_rcu(new_alias, rcuhead); 1534 1535 return len; 1536 } 1537 1538 /** 1539 * dev_get_alias - get ifalias of a device 1540 * @dev: device 1541 * @name: buffer to store name of ifalias 1542 * @len: size of buffer 1543 * 1544 * get ifalias for a device. Caller must make sure dev cannot go 1545 * away, e.g. rcu read lock or own a reference count to device. 1546 */ 1547 int dev_get_alias(const struct net_device *dev, char *name, size_t len) 1548 { 1549 const struct dev_ifalias *alias; 1550 int ret = 0; 1551 1552 rcu_read_lock(); 1553 alias = rcu_dereference(dev->ifalias); 1554 if (alias) 1555 ret = snprintf(name, len, "%s", alias->ifalias); 1556 rcu_read_unlock(); 1557 1558 return ret; 1559 } 1560 1561 /** 1562 * netdev_features_change - device changes features 1563 * @dev: device to cause notification 1564 * 1565 * Called to indicate a device has changed features. 1566 */ 1567 void netdev_features_change(struct net_device *dev) 1568 { 1569 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev); 1570 } 1571 EXPORT_SYMBOL(netdev_features_change); 1572 1573 void netif_state_change(struct net_device *dev) 1574 { 1575 netdev_ops_assert_locked_or_invisible(dev); 1576 1577 if (dev->flags & IFF_UP) { 1578 struct netdev_notifier_change_info change_info = { 1579 .info.dev = dev, 1580 }; 1581 1582 call_netdevice_notifiers_info(NETDEV_CHANGE, 1583 &change_info.info); 1584 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL, 0, NULL); 1585 } 1586 } 1587 1588 /** 1589 * __netdev_notify_peers - notify network peers about existence of @dev, 1590 * to be called when rtnl lock is already held. 1591 * @dev: network device 1592 * 1593 * Generate traffic such that interested network peers are aware of 1594 * @dev, such as by generating a gratuitous ARP. This may be used when 1595 * a device wants to inform the rest of the network about some sort of 1596 * reconfiguration such as a failover event or virtual machine 1597 * migration. 1598 */ 1599 void __netdev_notify_peers(struct net_device *dev) 1600 { 1601 ASSERT_RTNL(); 1602 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev); 1603 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev); 1604 } 1605 EXPORT_SYMBOL(__netdev_notify_peers); 1606 1607 /** 1608 * netdev_notify_peers - notify network peers about existence of @dev 1609 * @dev: network device 1610 * 1611 * Generate traffic such that interested network peers are aware of 1612 * @dev, such as by generating a gratuitous ARP. This may be used when 1613 * a device wants to inform the rest of the network about some sort of 1614 * reconfiguration such as a failover event or virtual machine 1615 * migration. 1616 */ 1617 void netdev_notify_peers(struct net_device *dev) 1618 { 1619 rtnl_lock(); 1620 __netdev_notify_peers(dev); 1621 rtnl_unlock(); 1622 } 1623 EXPORT_SYMBOL(netdev_notify_peers); 1624 1625 static int napi_threaded_poll(void *data); 1626 1627 static int napi_kthread_create(struct napi_struct *n) 1628 { 1629 int err = 0; 1630 1631 /* Create and wake up the kthread once to put it in 1632 * TASK_INTERRUPTIBLE mode to avoid the blocked task 1633 * warning and work with loadavg. 1634 */ 1635 n->thread = kthread_run(napi_threaded_poll, n, "napi/%s-%d", 1636 n->dev->name, n->napi_id); 1637 if (IS_ERR(n->thread)) { 1638 err = PTR_ERR(n->thread); 1639 pr_err("kthread_run failed with err %d\n", err); 1640 n->thread = NULL; 1641 } 1642 1643 return err; 1644 } 1645 1646 static int __dev_open(struct net_device *dev, struct netlink_ext_ack *extack) 1647 { 1648 const struct net_device_ops *ops = dev->netdev_ops; 1649 int ret; 1650 1651 ASSERT_RTNL(); 1652 dev_addr_check(dev); 1653 1654 if (!netif_device_present(dev)) { 1655 /* may be detached because parent is runtime-suspended */ 1656 if (dev->dev.parent) 1657 pm_runtime_resume(dev->dev.parent); 1658 if (!netif_device_present(dev)) 1659 return -ENODEV; 1660 } 1661 1662 /* Block netpoll from trying to do any rx path servicing. 1663 * If we don't do this there is a chance ndo_poll_controller 1664 * or ndo_poll may be running while we open the device 1665 */ 1666 netpoll_poll_disable(dev); 1667 1668 ret = call_netdevice_notifiers_extack(NETDEV_PRE_UP, dev, extack); 1669 ret = notifier_to_errno(ret); 1670 if (ret) 1671 return ret; 1672 1673 set_bit(__LINK_STATE_START, &dev->state); 1674 1675 netdev_ops_assert_locked(dev); 1676 1677 if (ops->ndo_validate_addr) 1678 ret = ops->ndo_validate_addr(dev); 1679 1680 if (!ret && ops->ndo_open) 1681 ret = ops->ndo_open(dev); 1682 1683 netpoll_poll_enable(dev); 1684 1685 if (ret) 1686 clear_bit(__LINK_STATE_START, &dev->state); 1687 else { 1688 netif_set_up(dev, true); 1689 dev_set_rx_mode(dev); 1690 dev_activate(dev); 1691 add_device_randomness(dev->dev_addr, dev->addr_len); 1692 } 1693 1694 return ret; 1695 } 1696 1697 int netif_open(struct net_device *dev, struct netlink_ext_ack *extack) 1698 { 1699 int ret; 1700 1701 if (dev->flags & IFF_UP) 1702 return 0; 1703 1704 ret = __dev_open(dev, extack); 1705 if (ret < 0) 1706 return ret; 1707 1708 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1709 call_netdevice_notifiers(NETDEV_UP, dev); 1710 1711 return ret; 1712 } 1713 1714 static void __dev_close_many(struct list_head *head) 1715 { 1716 struct net_device *dev; 1717 1718 ASSERT_RTNL(); 1719 might_sleep(); 1720 1721 list_for_each_entry(dev, head, close_list) { 1722 /* Temporarily disable netpoll until the interface is down */ 1723 netpoll_poll_disable(dev); 1724 1725 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev); 1726 1727 clear_bit(__LINK_STATE_START, &dev->state); 1728 1729 /* Synchronize to scheduled poll. We cannot touch poll list, it 1730 * can be even on different cpu. So just clear netif_running(). 1731 * 1732 * dev->stop() will invoke napi_disable() on all of it's 1733 * napi_struct instances on this device. 1734 */ 1735 smp_mb__after_atomic(); /* Commit netif_running(). */ 1736 } 1737 1738 dev_deactivate_many(head); 1739 1740 list_for_each_entry(dev, head, close_list) { 1741 const struct net_device_ops *ops = dev->netdev_ops; 1742 1743 /* 1744 * Call the device specific close. This cannot fail. 1745 * Only if device is UP 1746 * 1747 * We allow it to be called even after a DETACH hot-plug 1748 * event. 1749 */ 1750 1751 netdev_ops_assert_locked(dev); 1752 1753 if (ops->ndo_stop) 1754 ops->ndo_stop(dev); 1755 1756 netif_set_up(dev, false); 1757 netpoll_poll_enable(dev); 1758 } 1759 } 1760 1761 static void __dev_close(struct net_device *dev) 1762 { 1763 LIST_HEAD(single); 1764 1765 list_add(&dev->close_list, &single); 1766 __dev_close_many(&single); 1767 list_del(&single); 1768 } 1769 1770 void dev_close_many(struct list_head *head, bool unlink) 1771 { 1772 struct net_device *dev, *tmp; 1773 1774 /* Remove the devices that don't need to be closed */ 1775 list_for_each_entry_safe(dev, tmp, head, close_list) 1776 if (!(dev->flags & IFF_UP)) 1777 list_del_init(&dev->close_list); 1778 1779 __dev_close_many(head); 1780 1781 list_for_each_entry_safe(dev, tmp, head, close_list) { 1782 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1783 call_netdevice_notifiers(NETDEV_DOWN, dev); 1784 if (unlink) 1785 list_del_init(&dev->close_list); 1786 } 1787 } 1788 EXPORT_SYMBOL(dev_close_many); 1789 1790 void netif_close(struct net_device *dev) 1791 { 1792 if (dev->flags & IFF_UP) { 1793 LIST_HEAD(single); 1794 1795 list_add(&dev->close_list, &single); 1796 dev_close_many(&single, true); 1797 list_del(&single); 1798 } 1799 } 1800 EXPORT_SYMBOL(netif_close); 1801 1802 void netif_disable_lro(struct net_device *dev) 1803 { 1804 struct net_device *lower_dev; 1805 struct list_head *iter; 1806 1807 dev->wanted_features &= ~NETIF_F_LRO; 1808 netdev_update_features(dev); 1809 1810 if (unlikely(dev->features & NETIF_F_LRO)) 1811 netdev_WARN(dev, "failed to disable LRO!\n"); 1812 1813 netdev_for_each_lower_dev(dev, lower_dev, iter) { 1814 netdev_lock_ops(lower_dev); 1815 netif_disable_lro(lower_dev); 1816 netdev_unlock_ops(lower_dev); 1817 } 1818 } 1819 EXPORT_IPV6_MOD(netif_disable_lro); 1820 1821 /** 1822 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device 1823 * @dev: device 1824 * 1825 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be 1826 * called under RTNL. This is needed if Generic XDP is installed on 1827 * the device. 1828 */ 1829 static void dev_disable_gro_hw(struct net_device *dev) 1830 { 1831 dev->wanted_features &= ~NETIF_F_GRO_HW; 1832 netdev_update_features(dev); 1833 1834 if (unlikely(dev->features & NETIF_F_GRO_HW)) 1835 netdev_WARN(dev, "failed to disable GRO_HW!\n"); 1836 } 1837 1838 const char *netdev_cmd_to_name(enum netdev_cmd cmd) 1839 { 1840 #define N(val) \ 1841 case NETDEV_##val: \ 1842 return "NETDEV_" __stringify(val); 1843 switch (cmd) { 1844 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER) 1845 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE) 1846 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE) 1847 N(POST_INIT) N(PRE_UNINIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) 1848 N(CHANGEUPPER) N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) 1849 N(BONDING_INFO) N(PRECHANGEUPPER) N(CHANGELOWERSTATE) 1850 N(UDP_TUNNEL_PUSH_INFO) N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN) 1851 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO) 1852 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO) 1853 N(PRE_CHANGEADDR) N(OFFLOAD_XSTATS_ENABLE) N(OFFLOAD_XSTATS_DISABLE) 1854 N(OFFLOAD_XSTATS_REPORT_USED) N(OFFLOAD_XSTATS_REPORT_DELTA) 1855 N(XDP_FEAT_CHANGE) 1856 } 1857 #undef N 1858 return "UNKNOWN_NETDEV_EVENT"; 1859 } 1860 EXPORT_SYMBOL_GPL(netdev_cmd_to_name); 1861 1862 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val, 1863 struct net_device *dev) 1864 { 1865 struct netdev_notifier_info info = { 1866 .dev = dev, 1867 }; 1868 1869 return nb->notifier_call(nb, val, &info); 1870 } 1871 1872 static int call_netdevice_register_notifiers(struct notifier_block *nb, 1873 struct net_device *dev) 1874 { 1875 int err; 1876 1877 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev); 1878 err = notifier_to_errno(err); 1879 if (err) 1880 return err; 1881 1882 if (!(dev->flags & IFF_UP)) 1883 return 0; 1884 1885 call_netdevice_notifier(nb, NETDEV_UP, dev); 1886 return 0; 1887 } 1888 1889 static void call_netdevice_unregister_notifiers(struct notifier_block *nb, 1890 struct net_device *dev) 1891 { 1892 if (dev->flags & IFF_UP) { 1893 call_netdevice_notifier(nb, NETDEV_GOING_DOWN, 1894 dev); 1895 call_netdevice_notifier(nb, NETDEV_DOWN, dev); 1896 } 1897 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev); 1898 } 1899 1900 static int call_netdevice_register_net_notifiers(struct notifier_block *nb, 1901 struct net *net) 1902 { 1903 struct net_device *dev; 1904 int err; 1905 1906 for_each_netdev(net, dev) { 1907 netdev_lock_ops(dev); 1908 err = call_netdevice_register_notifiers(nb, dev); 1909 netdev_unlock_ops(dev); 1910 if (err) 1911 goto rollback; 1912 } 1913 return 0; 1914 1915 rollback: 1916 for_each_netdev_continue_reverse(net, dev) 1917 call_netdevice_unregister_notifiers(nb, dev); 1918 return err; 1919 } 1920 1921 static void call_netdevice_unregister_net_notifiers(struct notifier_block *nb, 1922 struct net *net) 1923 { 1924 struct net_device *dev; 1925 1926 for_each_netdev(net, dev) 1927 call_netdevice_unregister_notifiers(nb, dev); 1928 } 1929 1930 static int dev_boot_phase = 1; 1931 1932 /** 1933 * register_netdevice_notifier - register a network notifier block 1934 * @nb: notifier 1935 * 1936 * Register a notifier to be called when network device events occur. 1937 * The notifier passed is linked into the kernel structures and must 1938 * not be reused until it has been unregistered. A negative errno code 1939 * is returned on a failure. 1940 * 1941 * When registered all registration and up events are replayed 1942 * to the new notifier to allow device to have a race free 1943 * view of the network device list. 1944 */ 1945 1946 int register_netdevice_notifier(struct notifier_block *nb) 1947 { 1948 struct net *net; 1949 int err; 1950 1951 /* Close race with setup_net() and cleanup_net() */ 1952 down_write(&pernet_ops_rwsem); 1953 1954 /* When RTNL is removed, we need protection for netdev_chain. */ 1955 rtnl_lock(); 1956 1957 err = raw_notifier_chain_register(&netdev_chain, nb); 1958 if (err) 1959 goto unlock; 1960 if (dev_boot_phase) 1961 goto unlock; 1962 for_each_net(net) { 1963 __rtnl_net_lock(net); 1964 err = call_netdevice_register_net_notifiers(nb, net); 1965 __rtnl_net_unlock(net); 1966 if (err) 1967 goto rollback; 1968 } 1969 1970 unlock: 1971 rtnl_unlock(); 1972 up_write(&pernet_ops_rwsem); 1973 return err; 1974 1975 rollback: 1976 for_each_net_continue_reverse(net) { 1977 __rtnl_net_lock(net); 1978 call_netdevice_unregister_net_notifiers(nb, net); 1979 __rtnl_net_unlock(net); 1980 } 1981 1982 raw_notifier_chain_unregister(&netdev_chain, nb); 1983 goto unlock; 1984 } 1985 EXPORT_SYMBOL(register_netdevice_notifier); 1986 1987 /** 1988 * unregister_netdevice_notifier - unregister a network notifier block 1989 * @nb: notifier 1990 * 1991 * Unregister a notifier previously registered by 1992 * register_netdevice_notifier(). The notifier is unlinked into the 1993 * kernel structures and may then be reused. A negative errno code 1994 * is returned on a failure. 1995 * 1996 * After unregistering unregister and down device events are synthesized 1997 * for all devices on the device list to the removed notifier to remove 1998 * the need for special case cleanup code. 1999 */ 2000 2001 int unregister_netdevice_notifier(struct notifier_block *nb) 2002 { 2003 struct net *net; 2004 int err; 2005 2006 /* Close race with setup_net() and cleanup_net() */ 2007 down_write(&pernet_ops_rwsem); 2008 rtnl_lock(); 2009 err = raw_notifier_chain_unregister(&netdev_chain, nb); 2010 if (err) 2011 goto unlock; 2012 2013 for_each_net(net) { 2014 __rtnl_net_lock(net); 2015 call_netdevice_unregister_net_notifiers(nb, net); 2016 __rtnl_net_unlock(net); 2017 } 2018 2019 unlock: 2020 rtnl_unlock(); 2021 up_write(&pernet_ops_rwsem); 2022 return err; 2023 } 2024 EXPORT_SYMBOL(unregister_netdevice_notifier); 2025 2026 static int __register_netdevice_notifier_net(struct net *net, 2027 struct notifier_block *nb, 2028 bool ignore_call_fail) 2029 { 2030 int err; 2031 2032 err = raw_notifier_chain_register(&net->netdev_chain, nb); 2033 if (err) 2034 return err; 2035 if (dev_boot_phase) 2036 return 0; 2037 2038 err = call_netdevice_register_net_notifiers(nb, net); 2039 if (err && !ignore_call_fail) 2040 goto chain_unregister; 2041 2042 return 0; 2043 2044 chain_unregister: 2045 raw_notifier_chain_unregister(&net->netdev_chain, nb); 2046 return err; 2047 } 2048 2049 static int __unregister_netdevice_notifier_net(struct net *net, 2050 struct notifier_block *nb) 2051 { 2052 int err; 2053 2054 err = raw_notifier_chain_unregister(&net->netdev_chain, nb); 2055 if (err) 2056 return err; 2057 2058 call_netdevice_unregister_net_notifiers(nb, net); 2059 return 0; 2060 } 2061 2062 /** 2063 * register_netdevice_notifier_net - register a per-netns network notifier block 2064 * @net: network namespace 2065 * @nb: notifier 2066 * 2067 * Register a notifier to be called when network device events occur. 2068 * The notifier passed is linked into the kernel structures and must 2069 * not be reused until it has been unregistered. A negative errno code 2070 * is returned on a failure. 2071 * 2072 * When registered all registration and up events are replayed 2073 * to the new notifier to allow device to have a race free 2074 * view of the network device list. 2075 */ 2076 2077 int register_netdevice_notifier_net(struct net *net, struct notifier_block *nb) 2078 { 2079 int err; 2080 2081 rtnl_net_lock(net); 2082 err = __register_netdevice_notifier_net(net, nb, false); 2083 rtnl_net_unlock(net); 2084 2085 return err; 2086 } 2087 EXPORT_SYMBOL(register_netdevice_notifier_net); 2088 2089 /** 2090 * unregister_netdevice_notifier_net - unregister a per-netns 2091 * network notifier block 2092 * @net: network namespace 2093 * @nb: notifier 2094 * 2095 * Unregister a notifier previously registered by 2096 * register_netdevice_notifier_net(). The notifier is unlinked from the 2097 * kernel structures and may then be reused. A negative errno code 2098 * is returned on a failure. 2099 * 2100 * After unregistering unregister and down device events are synthesized 2101 * for all devices on the device list to the removed notifier to remove 2102 * the need for special case cleanup code. 2103 */ 2104 2105 int unregister_netdevice_notifier_net(struct net *net, 2106 struct notifier_block *nb) 2107 { 2108 int err; 2109 2110 rtnl_net_lock(net); 2111 err = __unregister_netdevice_notifier_net(net, nb); 2112 rtnl_net_unlock(net); 2113 2114 return err; 2115 } 2116 EXPORT_SYMBOL(unregister_netdevice_notifier_net); 2117 2118 static void __move_netdevice_notifier_net(struct net *src_net, 2119 struct net *dst_net, 2120 struct notifier_block *nb) 2121 { 2122 __unregister_netdevice_notifier_net(src_net, nb); 2123 __register_netdevice_notifier_net(dst_net, nb, true); 2124 } 2125 2126 static void rtnl_net_dev_lock(struct net_device *dev) 2127 { 2128 bool again; 2129 2130 do { 2131 struct net *net; 2132 2133 again = false; 2134 2135 /* netns might be being dismantled. */ 2136 rcu_read_lock(); 2137 net = dev_net_rcu(dev); 2138 net_passive_inc(net); 2139 rcu_read_unlock(); 2140 2141 rtnl_net_lock(net); 2142 2143 #ifdef CONFIG_NET_NS 2144 /* dev might have been moved to another netns. */ 2145 if (!net_eq(net, rcu_access_pointer(dev->nd_net.net))) { 2146 rtnl_net_unlock(net); 2147 net_passive_dec(net); 2148 again = true; 2149 } 2150 #endif 2151 } while (again); 2152 } 2153 2154 static void rtnl_net_dev_unlock(struct net_device *dev) 2155 { 2156 struct net *net = dev_net(dev); 2157 2158 rtnl_net_unlock(net); 2159 net_passive_dec(net); 2160 } 2161 2162 int register_netdevice_notifier_dev_net(struct net_device *dev, 2163 struct notifier_block *nb, 2164 struct netdev_net_notifier *nn) 2165 { 2166 int err; 2167 2168 rtnl_net_dev_lock(dev); 2169 err = __register_netdevice_notifier_net(dev_net(dev), nb, false); 2170 if (!err) { 2171 nn->nb = nb; 2172 list_add(&nn->list, &dev->net_notifier_list); 2173 } 2174 rtnl_net_dev_unlock(dev); 2175 2176 return err; 2177 } 2178 EXPORT_SYMBOL(register_netdevice_notifier_dev_net); 2179 2180 int unregister_netdevice_notifier_dev_net(struct net_device *dev, 2181 struct notifier_block *nb, 2182 struct netdev_net_notifier *nn) 2183 { 2184 int err; 2185 2186 rtnl_net_dev_lock(dev); 2187 list_del(&nn->list); 2188 err = __unregister_netdevice_notifier_net(dev_net(dev), nb); 2189 rtnl_net_dev_unlock(dev); 2190 2191 return err; 2192 } 2193 EXPORT_SYMBOL(unregister_netdevice_notifier_dev_net); 2194 2195 static void move_netdevice_notifiers_dev_net(struct net_device *dev, 2196 struct net *net) 2197 { 2198 struct netdev_net_notifier *nn; 2199 2200 list_for_each_entry(nn, &dev->net_notifier_list, list) 2201 __move_netdevice_notifier_net(dev_net(dev), net, nn->nb); 2202 } 2203 2204 /** 2205 * call_netdevice_notifiers_info - call all network notifier blocks 2206 * @val: value passed unmodified to notifier function 2207 * @info: notifier information data 2208 * 2209 * Call all network notifier blocks. Parameters and return value 2210 * are as for raw_notifier_call_chain(). 2211 */ 2212 2213 int call_netdevice_notifiers_info(unsigned long val, 2214 struct netdev_notifier_info *info) 2215 { 2216 struct net *net = dev_net(info->dev); 2217 int ret; 2218 2219 ASSERT_RTNL(); 2220 2221 /* Run per-netns notifier block chain first, then run the global one. 2222 * Hopefully, one day, the global one is going to be removed after 2223 * all notifier block registrators get converted to be per-netns. 2224 */ 2225 ret = raw_notifier_call_chain(&net->netdev_chain, val, info); 2226 if (ret & NOTIFY_STOP_MASK) 2227 return ret; 2228 return raw_notifier_call_chain(&netdev_chain, val, info); 2229 } 2230 2231 /** 2232 * call_netdevice_notifiers_info_robust - call per-netns notifier blocks 2233 * for and rollback on error 2234 * @val_up: value passed unmodified to notifier function 2235 * @val_down: value passed unmodified to the notifier function when 2236 * recovering from an error on @val_up 2237 * @info: notifier information data 2238 * 2239 * Call all per-netns network notifier blocks, but not notifier blocks on 2240 * the global notifier chain. Parameters and return value are as for 2241 * raw_notifier_call_chain_robust(). 2242 */ 2243 2244 static int 2245 call_netdevice_notifiers_info_robust(unsigned long val_up, 2246 unsigned long val_down, 2247 struct netdev_notifier_info *info) 2248 { 2249 struct net *net = dev_net(info->dev); 2250 2251 ASSERT_RTNL(); 2252 2253 return raw_notifier_call_chain_robust(&net->netdev_chain, 2254 val_up, val_down, info); 2255 } 2256 2257 static int call_netdevice_notifiers_extack(unsigned long val, 2258 struct net_device *dev, 2259 struct netlink_ext_ack *extack) 2260 { 2261 struct netdev_notifier_info info = { 2262 .dev = dev, 2263 .extack = extack, 2264 }; 2265 2266 return call_netdevice_notifiers_info(val, &info); 2267 } 2268 2269 /** 2270 * call_netdevice_notifiers - call all network notifier blocks 2271 * @val: value passed unmodified to notifier function 2272 * @dev: net_device pointer passed unmodified to notifier function 2273 * 2274 * Call all network notifier blocks. Parameters and return value 2275 * are as for raw_notifier_call_chain(). 2276 */ 2277 2278 int call_netdevice_notifiers(unsigned long val, struct net_device *dev) 2279 { 2280 return call_netdevice_notifiers_extack(val, dev, NULL); 2281 } 2282 EXPORT_SYMBOL(call_netdevice_notifiers); 2283 2284 /** 2285 * call_netdevice_notifiers_mtu - call all network notifier blocks 2286 * @val: value passed unmodified to notifier function 2287 * @dev: net_device pointer passed unmodified to notifier function 2288 * @arg: additional u32 argument passed to the notifier function 2289 * 2290 * Call all network notifier blocks. Parameters and return value 2291 * are as for raw_notifier_call_chain(). 2292 */ 2293 static int call_netdevice_notifiers_mtu(unsigned long val, 2294 struct net_device *dev, u32 arg) 2295 { 2296 struct netdev_notifier_info_ext info = { 2297 .info.dev = dev, 2298 .ext.mtu = arg, 2299 }; 2300 2301 BUILD_BUG_ON(offsetof(struct netdev_notifier_info_ext, info) != 0); 2302 2303 return call_netdevice_notifiers_info(val, &info.info); 2304 } 2305 2306 #ifdef CONFIG_NET_INGRESS 2307 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key); 2308 2309 void net_inc_ingress_queue(void) 2310 { 2311 static_branch_inc(&ingress_needed_key); 2312 } 2313 EXPORT_SYMBOL_GPL(net_inc_ingress_queue); 2314 2315 void net_dec_ingress_queue(void) 2316 { 2317 static_branch_dec(&ingress_needed_key); 2318 } 2319 EXPORT_SYMBOL_GPL(net_dec_ingress_queue); 2320 #endif 2321 2322 #ifdef CONFIG_NET_EGRESS 2323 static DEFINE_STATIC_KEY_FALSE(egress_needed_key); 2324 2325 void net_inc_egress_queue(void) 2326 { 2327 static_branch_inc(&egress_needed_key); 2328 } 2329 EXPORT_SYMBOL_GPL(net_inc_egress_queue); 2330 2331 void net_dec_egress_queue(void) 2332 { 2333 static_branch_dec(&egress_needed_key); 2334 } 2335 EXPORT_SYMBOL_GPL(net_dec_egress_queue); 2336 #endif 2337 2338 #ifdef CONFIG_NET_CLS_ACT 2339 DEFINE_STATIC_KEY_FALSE(tcf_sw_enabled_key); 2340 EXPORT_SYMBOL(tcf_sw_enabled_key); 2341 #endif 2342 2343 DEFINE_STATIC_KEY_FALSE(netstamp_needed_key); 2344 EXPORT_SYMBOL(netstamp_needed_key); 2345 #ifdef CONFIG_JUMP_LABEL 2346 static atomic_t netstamp_needed_deferred; 2347 static atomic_t netstamp_wanted; 2348 static void netstamp_clear(struct work_struct *work) 2349 { 2350 int deferred = atomic_xchg(&netstamp_needed_deferred, 0); 2351 int wanted; 2352 2353 wanted = atomic_add_return(deferred, &netstamp_wanted); 2354 if (wanted > 0) 2355 static_branch_enable(&netstamp_needed_key); 2356 else 2357 static_branch_disable(&netstamp_needed_key); 2358 } 2359 static DECLARE_WORK(netstamp_work, netstamp_clear); 2360 #endif 2361 2362 void net_enable_timestamp(void) 2363 { 2364 #ifdef CONFIG_JUMP_LABEL 2365 int wanted = atomic_read(&netstamp_wanted); 2366 2367 while (wanted > 0) { 2368 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted + 1)) 2369 return; 2370 } 2371 atomic_inc(&netstamp_needed_deferred); 2372 schedule_work(&netstamp_work); 2373 #else 2374 static_branch_inc(&netstamp_needed_key); 2375 #endif 2376 } 2377 EXPORT_SYMBOL(net_enable_timestamp); 2378 2379 void net_disable_timestamp(void) 2380 { 2381 #ifdef CONFIG_JUMP_LABEL 2382 int wanted = atomic_read(&netstamp_wanted); 2383 2384 while (wanted > 1) { 2385 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted - 1)) 2386 return; 2387 } 2388 atomic_dec(&netstamp_needed_deferred); 2389 schedule_work(&netstamp_work); 2390 #else 2391 static_branch_dec(&netstamp_needed_key); 2392 #endif 2393 } 2394 EXPORT_SYMBOL(net_disable_timestamp); 2395 2396 static inline void net_timestamp_set(struct sk_buff *skb) 2397 { 2398 skb->tstamp = 0; 2399 skb->tstamp_type = SKB_CLOCK_REALTIME; 2400 if (static_branch_unlikely(&netstamp_needed_key)) 2401 skb->tstamp = ktime_get_real(); 2402 } 2403 2404 #define net_timestamp_check(COND, SKB) \ 2405 if (static_branch_unlikely(&netstamp_needed_key)) { \ 2406 if ((COND) && !(SKB)->tstamp) \ 2407 (SKB)->tstamp = ktime_get_real(); \ 2408 } \ 2409 2410 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb) 2411 { 2412 return __is_skb_forwardable(dev, skb, true); 2413 } 2414 EXPORT_SYMBOL_GPL(is_skb_forwardable); 2415 2416 static int __dev_forward_skb2(struct net_device *dev, struct sk_buff *skb, 2417 bool check_mtu) 2418 { 2419 int ret = ____dev_forward_skb(dev, skb, check_mtu); 2420 2421 if (likely(!ret)) { 2422 skb->protocol = eth_type_trans(skb, dev); 2423 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN); 2424 } 2425 2426 return ret; 2427 } 2428 2429 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2430 { 2431 return __dev_forward_skb2(dev, skb, true); 2432 } 2433 EXPORT_SYMBOL_GPL(__dev_forward_skb); 2434 2435 /** 2436 * dev_forward_skb - loopback an skb to another netif 2437 * 2438 * @dev: destination network device 2439 * @skb: buffer to forward 2440 * 2441 * return values: 2442 * NET_RX_SUCCESS (no congestion) 2443 * NET_RX_DROP (packet was dropped, but freed) 2444 * 2445 * dev_forward_skb can be used for injecting an skb from the 2446 * start_xmit function of one device into the receive queue 2447 * of another device. 2448 * 2449 * The receiving device may be in another namespace, so 2450 * we have to clear all information in the skb that could 2451 * impact namespace isolation. 2452 */ 2453 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2454 { 2455 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb); 2456 } 2457 EXPORT_SYMBOL_GPL(dev_forward_skb); 2458 2459 int dev_forward_skb_nomtu(struct net_device *dev, struct sk_buff *skb) 2460 { 2461 return __dev_forward_skb2(dev, skb, false) ?: netif_rx_internal(skb); 2462 } 2463 2464 static inline int deliver_skb(struct sk_buff *skb, 2465 struct packet_type *pt_prev, 2466 struct net_device *orig_dev) 2467 { 2468 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 2469 return -ENOMEM; 2470 refcount_inc(&skb->users); 2471 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 2472 } 2473 2474 static inline void deliver_ptype_list_skb(struct sk_buff *skb, 2475 struct packet_type **pt, 2476 struct net_device *orig_dev, 2477 __be16 type, 2478 struct list_head *ptype_list) 2479 { 2480 struct packet_type *ptype, *pt_prev = *pt; 2481 2482 list_for_each_entry_rcu(ptype, ptype_list, list) { 2483 if (ptype->type != type) 2484 continue; 2485 if (pt_prev) 2486 deliver_skb(skb, pt_prev, orig_dev); 2487 pt_prev = ptype; 2488 } 2489 *pt = pt_prev; 2490 } 2491 2492 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb) 2493 { 2494 if (!ptype->af_packet_priv || !skb->sk) 2495 return false; 2496 2497 if (ptype->id_match) 2498 return ptype->id_match(ptype, skb->sk); 2499 else if ((struct sock *)ptype->af_packet_priv == skb->sk) 2500 return true; 2501 2502 return false; 2503 } 2504 2505 /** 2506 * dev_nit_active_rcu - return true if any network interface taps are in use 2507 * 2508 * The caller must hold the RCU lock 2509 * 2510 * @dev: network device to check for the presence of taps 2511 */ 2512 bool dev_nit_active_rcu(const struct net_device *dev) 2513 { 2514 /* Callers may hold either RCU or RCU BH lock */ 2515 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 2516 2517 return !list_empty(&dev_net(dev)->ptype_all) || 2518 !list_empty(&dev->ptype_all); 2519 } 2520 EXPORT_SYMBOL_GPL(dev_nit_active_rcu); 2521 2522 /* 2523 * Support routine. Sends outgoing frames to any network 2524 * taps currently in use. 2525 */ 2526 2527 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev) 2528 { 2529 struct packet_type *ptype, *pt_prev = NULL; 2530 struct list_head *ptype_list; 2531 struct sk_buff *skb2 = NULL; 2532 2533 rcu_read_lock(); 2534 ptype_list = &dev_net_rcu(dev)->ptype_all; 2535 again: 2536 list_for_each_entry_rcu(ptype, ptype_list, list) { 2537 if (READ_ONCE(ptype->ignore_outgoing)) 2538 continue; 2539 2540 /* Never send packets back to the socket 2541 * they originated from - MvS (miquels@drinkel.ow.org) 2542 */ 2543 if (skb_loop_sk(ptype, skb)) 2544 continue; 2545 2546 if (pt_prev) { 2547 deliver_skb(skb2, pt_prev, skb->dev); 2548 pt_prev = ptype; 2549 continue; 2550 } 2551 2552 /* need to clone skb, done only once */ 2553 skb2 = skb_clone(skb, GFP_ATOMIC); 2554 if (!skb2) 2555 goto out_unlock; 2556 2557 net_timestamp_set(skb2); 2558 2559 /* skb->nh should be correctly 2560 * set by sender, so that the second statement is 2561 * just protection against buggy protocols. 2562 */ 2563 skb_reset_mac_header(skb2); 2564 2565 if (skb_network_header(skb2) < skb2->data || 2566 skb_network_header(skb2) > skb_tail_pointer(skb2)) { 2567 net_crit_ratelimited("protocol %04x is buggy, dev %s\n", 2568 ntohs(skb2->protocol), 2569 dev->name); 2570 skb_reset_network_header(skb2); 2571 } 2572 2573 skb2->transport_header = skb2->network_header; 2574 skb2->pkt_type = PACKET_OUTGOING; 2575 pt_prev = ptype; 2576 } 2577 2578 if (ptype_list != &dev->ptype_all) { 2579 ptype_list = &dev->ptype_all; 2580 goto again; 2581 } 2582 out_unlock: 2583 if (pt_prev) { 2584 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC)) 2585 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev); 2586 else 2587 kfree_skb(skb2); 2588 } 2589 rcu_read_unlock(); 2590 } 2591 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit); 2592 2593 /** 2594 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change 2595 * @dev: Network device 2596 * @txq: number of queues available 2597 * 2598 * If real_num_tx_queues is changed the tc mappings may no longer be 2599 * valid. To resolve this verify the tc mapping remains valid and if 2600 * not NULL the mapping. With no priorities mapping to this 2601 * offset/count pair it will no longer be used. In the worst case TC0 2602 * is invalid nothing can be done so disable priority mappings. If is 2603 * expected that drivers will fix this mapping if they can before 2604 * calling netif_set_real_num_tx_queues. 2605 */ 2606 static void netif_setup_tc(struct net_device *dev, unsigned int txq) 2607 { 2608 int i; 2609 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2610 2611 /* If TC0 is invalidated disable TC mapping */ 2612 if (tc->offset + tc->count > txq) { 2613 netdev_warn(dev, "Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n"); 2614 dev->num_tc = 0; 2615 return; 2616 } 2617 2618 /* Invalidated prio to tc mappings set to TC0 */ 2619 for (i = 1; i < TC_BITMASK + 1; i++) { 2620 int q = netdev_get_prio_tc_map(dev, i); 2621 2622 tc = &dev->tc_to_txq[q]; 2623 if (tc->offset + tc->count > txq) { 2624 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", 2625 i, q); 2626 netdev_set_prio_tc_map(dev, i, 0); 2627 } 2628 } 2629 } 2630 2631 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq) 2632 { 2633 if (dev->num_tc) { 2634 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2635 int i; 2636 2637 /* walk through the TCs and see if it falls into any of them */ 2638 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) { 2639 if ((txq - tc->offset) < tc->count) 2640 return i; 2641 } 2642 2643 /* didn't find it, just return -1 to indicate no match */ 2644 return -1; 2645 } 2646 2647 return 0; 2648 } 2649 EXPORT_SYMBOL(netdev_txq_to_tc); 2650 2651 #ifdef CONFIG_XPS 2652 static struct static_key xps_needed __read_mostly; 2653 static struct static_key xps_rxqs_needed __read_mostly; 2654 static DEFINE_MUTEX(xps_map_mutex); 2655 #define xmap_dereference(P) \ 2656 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex)) 2657 2658 static bool remove_xps_queue(struct xps_dev_maps *dev_maps, 2659 struct xps_dev_maps *old_maps, int tci, u16 index) 2660 { 2661 struct xps_map *map = NULL; 2662 int pos; 2663 2664 map = xmap_dereference(dev_maps->attr_map[tci]); 2665 if (!map) 2666 return false; 2667 2668 for (pos = map->len; pos--;) { 2669 if (map->queues[pos] != index) 2670 continue; 2671 2672 if (map->len > 1) { 2673 map->queues[pos] = map->queues[--map->len]; 2674 break; 2675 } 2676 2677 if (old_maps) 2678 RCU_INIT_POINTER(old_maps->attr_map[tci], NULL); 2679 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2680 kfree_rcu(map, rcu); 2681 return false; 2682 } 2683 2684 return true; 2685 } 2686 2687 static bool remove_xps_queue_cpu(struct net_device *dev, 2688 struct xps_dev_maps *dev_maps, 2689 int cpu, u16 offset, u16 count) 2690 { 2691 int num_tc = dev_maps->num_tc; 2692 bool active = false; 2693 int tci; 2694 2695 for (tci = cpu * num_tc; num_tc--; tci++) { 2696 int i, j; 2697 2698 for (i = count, j = offset; i--; j++) { 2699 if (!remove_xps_queue(dev_maps, NULL, tci, j)) 2700 break; 2701 } 2702 2703 active |= i < 0; 2704 } 2705 2706 return active; 2707 } 2708 2709 static void reset_xps_maps(struct net_device *dev, 2710 struct xps_dev_maps *dev_maps, 2711 enum xps_map_type type) 2712 { 2713 static_key_slow_dec_cpuslocked(&xps_needed); 2714 if (type == XPS_RXQS) 2715 static_key_slow_dec_cpuslocked(&xps_rxqs_needed); 2716 2717 RCU_INIT_POINTER(dev->xps_maps[type], NULL); 2718 2719 kfree_rcu(dev_maps, rcu); 2720 } 2721 2722 static void clean_xps_maps(struct net_device *dev, enum xps_map_type type, 2723 u16 offset, u16 count) 2724 { 2725 struct xps_dev_maps *dev_maps; 2726 bool active = false; 2727 int i, j; 2728 2729 dev_maps = xmap_dereference(dev->xps_maps[type]); 2730 if (!dev_maps) 2731 return; 2732 2733 for (j = 0; j < dev_maps->nr_ids; j++) 2734 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, count); 2735 if (!active) 2736 reset_xps_maps(dev, dev_maps, type); 2737 2738 if (type == XPS_CPUS) { 2739 for (i = offset + (count - 1); count--; i--) 2740 netdev_queue_numa_node_write( 2741 netdev_get_tx_queue(dev, i), NUMA_NO_NODE); 2742 } 2743 } 2744 2745 static void netif_reset_xps_queues(struct net_device *dev, u16 offset, 2746 u16 count) 2747 { 2748 if (!static_key_false(&xps_needed)) 2749 return; 2750 2751 cpus_read_lock(); 2752 mutex_lock(&xps_map_mutex); 2753 2754 if (static_key_false(&xps_rxqs_needed)) 2755 clean_xps_maps(dev, XPS_RXQS, offset, count); 2756 2757 clean_xps_maps(dev, XPS_CPUS, offset, count); 2758 2759 mutex_unlock(&xps_map_mutex); 2760 cpus_read_unlock(); 2761 } 2762 2763 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index) 2764 { 2765 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index); 2766 } 2767 2768 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index, 2769 u16 index, bool is_rxqs_map) 2770 { 2771 struct xps_map *new_map; 2772 int alloc_len = XPS_MIN_MAP_ALLOC; 2773 int i, pos; 2774 2775 for (pos = 0; map && pos < map->len; pos++) { 2776 if (map->queues[pos] != index) 2777 continue; 2778 return map; 2779 } 2780 2781 /* Need to add tx-queue to this CPU's/rx-queue's existing map */ 2782 if (map) { 2783 if (pos < map->alloc_len) 2784 return map; 2785 2786 alloc_len = map->alloc_len * 2; 2787 } 2788 2789 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's 2790 * map 2791 */ 2792 if (is_rxqs_map) 2793 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL); 2794 else 2795 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL, 2796 cpu_to_node(attr_index)); 2797 if (!new_map) 2798 return NULL; 2799 2800 for (i = 0; i < pos; i++) 2801 new_map->queues[i] = map->queues[i]; 2802 new_map->alloc_len = alloc_len; 2803 new_map->len = pos; 2804 2805 return new_map; 2806 } 2807 2808 /* Copy xps maps at a given index */ 2809 static void xps_copy_dev_maps(struct xps_dev_maps *dev_maps, 2810 struct xps_dev_maps *new_dev_maps, int index, 2811 int tc, bool skip_tc) 2812 { 2813 int i, tci = index * dev_maps->num_tc; 2814 struct xps_map *map; 2815 2816 /* copy maps belonging to foreign traffic classes */ 2817 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2818 if (i == tc && skip_tc) 2819 continue; 2820 2821 /* fill in the new device map from the old device map */ 2822 map = xmap_dereference(dev_maps->attr_map[tci]); 2823 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2824 } 2825 } 2826 2827 /* Must be called under cpus_read_lock */ 2828 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask, 2829 u16 index, enum xps_map_type type) 2830 { 2831 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL, *old_dev_maps = NULL; 2832 const unsigned long *online_mask = NULL; 2833 bool active = false, copy = false; 2834 int i, j, tci, numa_node_id = -2; 2835 int maps_sz, num_tc = 1, tc = 0; 2836 struct xps_map *map, *new_map; 2837 unsigned int nr_ids; 2838 2839 WARN_ON_ONCE(index >= dev->num_tx_queues); 2840 2841 if (dev->num_tc) { 2842 /* Do not allow XPS on subordinate device directly */ 2843 num_tc = dev->num_tc; 2844 if (num_tc < 0) 2845 return -EINVAL; 2846 2847 /* If queue belongs to subordinate dev use its map */ 2848 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev; 2849 2850 tc = netdev_txq_to_tc(dev, index); 2851 if (tc < 0) 2852 return -EINVAL; 2853 } 2854 2855 mutex_lock(&xps_map_mutex); 2856 2857 dev_maps = xmap_dereference(dev->xps_maps[type]); 2858 if (type == XPS_RXQS) { 2859 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues); 2860 nr_ids = dev->num_rx_queues; 2861 } else { 2862 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc); 2863 if (num_possible_cpus() > 1) 2864 online_mask = cpumask_bits(cpu_online_mask); 2865 nr_ids = nr_cpu_ids; 2866 } 2867 2868 if (maps_sz < L1_CACHE_BYTES) 2869 maps_sz = L1_CACHE_BYTES; 2870 2871 /* The old dev_maps could be larger or smaller than the one we're 2872 * setting up now, as dev->num_tc or nr_ids could have been updated in 2873 * between. We could try to be smart, but let's be safe instead and only 2874 * copy foreign traffic classes if the two map sizes match. 2875 */ 2876 if (dev_maps && 2877 dev_maps->num_tc == num_tc && dev_maps->nr_ids == nr_ids) 2878 copy = true; 2879 2880 /* allocate memory for queue storage */ 2881 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids), 2882 j < nr_ids;) { 2883 if (!new_dev_maps) { 2884 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL); 2885 if (!new_dev_maps) { 2886 mutex_unlock(&xps_map_mutex); 2887 return -ENOMEM; 2888 } 2889 2890 new_dev_maps->nr_ids = nr_ids; 2891 new_dev_maps->num_tc = num_tc; 2892 } 2893 2894 tci = j * num_tc + tc; 2895 map = copy ? xmap_dereference(dev_maps->attr_map[tci]) : NULL; 2896 2897 map = expand_xps_map(map, j, index, type == XPS_RXQS); 2898 if (!map) 2899 goto error; 2900 2901 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2902 } 2903 2904 if (!new_dev_maps) 2905 goto out_no_new_maps; 2906 2907 if (!dev_maps) { 2908 /* Increment static keys at most once per type */ 2909 static_key_slow_inc_cpuslocked(&xps_needed); 2910 if (type == XPS_RXQS) 2911 static_key_slow_inc_cpuslocked(&xps_rxqs_needed); 2912 } 2913 2914 for (j = 0; j < nr_ids; j++) { 2915 bool skip_tc = false; 2916 2917 tci = j * num_tc + tc; 2918 if (netif_attr_test_mask(j, mask, nr_ids) && 2919 netif_attr_test_online(j, online_mask, nr_ids)) { 2920 /* add tx-queue to CPU/rx-queue maps */ 2921 int pos = 0; 2922 2923 skip_tc = true; 2924 2925 map = xmap_dereference(new_dev_maps->attr_map[tci]); 2926 while ((pos < map->len) && (map->queues[pos] != index)) 2927 pos++; 2928 2929 if (pos == map->len) 2930 map->queues[map->len++] = index; 2931 #ifdef CONFIG_NUMA 2932 if (type == XPS_CPUS) { 2933 if (numa_node_id == -2) 2934 numa_node_id = cpu_to_node(j); 2935 else if (numa_node_id != cpu_to_node(j)) 2936 numa_node_id = -1; 2937 } 2938 #endif 2939 } 2940 2941 if (copy) 2942 xps_copy_dev_maps(dev_maps, new_dev_maps, j, tc, 2943 skip_tc); 2944 } 2945 2946 rcu_assign_pointer(dev->xps_maps[type], new_dev_maps); 2947 2948 /* Cleanup old maps */ 2949 if (!dev_maps) 2950 goto out_no_old_maps; 2951 2952 for (j = 0; j < dev_maps->nr_ids; j++) { 2953 for (i = num_tc, tci = j * dev_maps->num_tc; i--; tci++) { 2954 map = xmap_dereference(dev_maps->attr_map[tci]); 2955 if (!map) 2956 continue; 2957 2958 if (copy) { 2959 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2960 if (map == new_map) 2961 continue; 2962 } 2963 2964 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2965 kfree_rcu(map, rcu); 2966 } 2967 } 2968 2969 old_dev_maps = dev_maps; 2970 2971 out_no_old_maps: 2972 dev_maps = new_dev_maps; 2973 active = true; 2974 2975 out_no_new_maps: 2976 if (type == XPS_CPUS) 2977 /* update Tx queue numa node */ 2978 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index), 2979 (numa_node_id >= 0) ? 2980 numa_node_id : NUMA_NO_NODE); 2981 2982 if (!dev_maps) 2983 goto out_no_maps; 2984 2985 /* removes tx-queue from unused CPUs/rx-queues */ 2986 for (j = 0; j < dev_maps->nr_ids; j++) { 2987 tci = j * dev_maps->num_tc; 2988 2989 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2990 if (i == tc && 2991 netif_attr_test_mask(j, mask, dev_maps->nr_ids) && 2992 netif_attr_test_online(j, online_mask, dev_maps->nr_ids)) 2993 continue; 2994 2995 active |= remove_xps_queue(dev_maps, 2996 copy ? old_dev_maps : NULL, 2997 tci, index); 2998 } 2999 } 3000 3001 if (old_dev_maps) 3002 kfree_rcu(old_dev_maps, rcu); 3003 3004 /* free map if not active */ 3005 if (!active) 3006 reset_xps_maps(dev, dev_maps, type); 3007 3008 out_no_maps: 3009 mutex_unlock(&xps_map_mutex); 3010 3011 return 0; 3012 error: 3013 /* remove any maps that we added */ 3014 for (j = 0; j < nr_ids; j++) { 3015 for (i = num_tc, tci = j * num_tc; i--; tci++) { 3016 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 3017 map = copy ? 3018 xmap_dereference(dev_maps->attr_map[tci]) : 3019 NULL; 3020 if (new_map && new_map != map) 3021 kfree(new_map); 3022 } 3023 } 3024 3025 mutex_unlock(&xps_map_mutex); 3026 3027 kfree(new_dev_maps); 3028 return -ENOMEM; 3029 } 3030 EXPORT_SYMBOL_GPL(__netif_set_xps_queue); 3031 3032 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, 3033 u16 index) 3034 { 3035 int ret; 3036 3037 cpus_read_lock(); 3038 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, XPS_CPUS); 3039 cpus_read_unlock(); 3040 3041 return ret; 3042 } 3043 EXPORT_SYMBOL(netif_set_xps_queue); 3044 3045 #endif 3046 static void netdev_unbind_all_sb_channels(struct net_device *dev) 3047 { 3048 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 3049 3050 /* Unbind any subordinate channels */ 3051 while (txq-- != &dev->_tx[0]) { 3052 if (txq->sb_dev) 3053 netdev_unbind_sb_channel(dev, txq->sb_dev); 3054 } 3055 } 3056 3057 void netdev_reset_tc(struct net_device *dev) 3058 { 3059 #ifdef CONFIG_XPS 3060 netif_reset_xps_queues_gt(dev, 0); 3061 #endif 3062 netdev_unbind_all_sb_channels(dev); 3063 3064 /* Reset TC configuration of device */ 3065 dev->num_tc = 0; 3066 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq)); 3067 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map)); 3068 } 3069 EXPORT_SYMBOL(netdev_reset_tc); 3070 3071 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset) 3072 { 3073 if (tc >= dev->num_tc) 3074 return -EINVAL; 3075 3076 #ifdef CONFIG_XPS 3077 netif_reset_xps_queues(dev, offset, count); 3078 #endif 3079 dev->tc_to_txq[tc].count = count; 3080 dev->tc_to_txq[tc].offset = offset; 3081 return 0; 3082 } 3083 EXPORT_SYMBOL(netdev_set_tc_queue); 3084 3085 int netdev_set_num_tc(struct net_device *dev, u8 num_tc) 3086 { 3087 if (num_tc > TC_MAX_QUEUE) 3088 return -EINVAL; 3089 3090 #ifdef CONFIG_XPS 3091 netif_reset_xps_queues_gt(dev, 0); 3092 #endif 3093 netdev_unbind_all_sb_channels(dev); 3094 3095 dev->num_tc = num_tc; 3096 return 0; 3097 } 3098 EXPORT_SYMBOL(netdev_set_num_tc); 3099 3100 void netdev_unbind_sb_channel(struct net_device *dev, 3101 struct net_device *sb_dev) 3102 { 3103 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 3104 3105 #ifdef CONFIG_XPS 3106 netif_reset_xps_queues_gt(sb_dev, 0); 3107 #endif 3108 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq)); 3109 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map)); 3110 3111 while (txq-- != &dev->_tx[0]) { 3112 if (txq->sb_dev == sb_dev) 3113 txq->sb_dev = NULL; 3114 } 3115 } 3116 EXPORT_SYMBOL(netdev_unbind_sb_channel); 3117 3118 int netdev_bind_sb_channel_queue(struct net_device *dev, 3119 struct net_device *sb_dev, 3120 u8 tc, u16 count, u16 offset) 3121 { 3122 /* Make certain the sb_dev and dev are already configured */ 3123 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc) 3124 return -EINVAL; 3125 3126 /* We cannot hand out queues we don't have */ 3127 if ((offset + count) > dev->real_num_tx_queues) 3128 return -EINVAL; 3129 3130 /* Record the mapping */ 3131 sb_dev->tc_to_txq[tc].count = count; 3132 sb_dev->tc_to_txq[tc].offset = offset; 3133 3134 /* Provide a way for Tx queue to find the tc_to_txq map or 3135 * XPS map for itself. 3136 */ 3137 while (count--) 3138 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev; 3139 3140 return 0; 3141 } 3142 EXPORT_SYMBOL(netdev_bind_sb_channel_queue); 3143 3144 int netdev_set_sb_channel(struct net_device *dev, u16 channel) 3145 { 3146 /* Do not use a multiqueue device to represent a subordinate channel */ 3147 if (netif_is_multiqueue(dev)) 3148 return -ENODEV; 3149 3150 /* We allow channels 1 - 32767 to be used for subordinate channels. 3151 * Channel 0 is meant to be "native" mode and used only to represent 3152 * the main root device. We allow writing 0 to reset the device back 3153 * to normal mode after being used as a subordinate channel. 3154 */ 3155 if (channel > S16_MAX) 3156 return -EINVAL; 3157 3158 dev->num_tc = -channel; 3159 3160 return 0; 3161 } 3162 EXPORT_SYMBOL(netdev_set_sb_channel); 3163 3164 /* 3165 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues 3166 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed. 3167 */ 3168 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq) 3169 { 3170 bool disabling; 3171 int rc; 3172 3173 disabling = txq < dev->real_num_tx_queues; 3174 3175 if (txq < 1 || txq > dev->num_tx_queues) 3176 return -EINVAL; 3177 3178 if (dev->reg_state == NETREG_REGISTERED || 3179 dev->reg_state == NETREG_UNREGISTERING) { 3180 ASSERT_RTNL(); 3181 netdev_ops_assert_locked(dev); 3182 3183 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues, 3184 txq); 3185 if (rc) 3186 return rc; 3187 3188 if (dev->num_tc) 3189 netif_setup_tc(dev, txq); 3190 3191 net_shaper_set_real_num_tx_queues(dev, txq); 3192 3193 dev_qdisc_change_real_num_tx(dev, txq); 3194 3195 dev->real_num_tx_queues = txq; 3196 3197 if (disabling) { 3198 synchronize_net(); 3199 qdisc_reset_all_tx_gt(dev, txq); 3200 #ifdef CONFIG_XPS 3201 netif_reset_xps_queues_gt(dev, txq); 3202 #endif 3203 } 3204 } else { 3205 dev->real_num_tx_queues = txq; 3206 } 3207 3208 return 0; 3209 } 3210 EXPORT_SYMBOL(netif_set_real_num_tx_queues); 3211 3212 /** 3213 * netif_set_real_num_rx_queues - set actual number of RX queues used 3214 * @dev: Network device 3215 * @rxq: Actual number of RX queues 3216 * 3217 * This must be called either with the rtnl_lock held or before 3218 * registration of the net device. Returns 0 on success, or a 3219 * negative error code. If called before registration, it always 3220 * succeeds. 3221 */ 3222 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq) 3223 { 3224 int rc; 3225 3226 if (rxq < 1 || rxq > dev->num_rx_queues) 3227 return -EINVAL; 3228 3229 if (dev->reg_state == NETREG_REGISTERED) { 3230 ASSERT_RTNL(); 3231 netdev_ops_assert_locked(dev); 3232 3233 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues, 3234 rxq); 3235 if (rc) 3236 return rc; 3237 } 3238 3239 dev->real_num_rx_queues = rxq; 3240 return 0; 3241 } 3242 EXPORT_SYMBOL(netif_set_real_num_rx_queues); 3243 3244 /** 3245 * netif_set_real_num_queues - set actual number of RX and TX queues used 3246 * @dev: Network device 3247 * @txq: Actual number of TX queues 3248 * @rxq: Actual number of RX queues 3249 * 3250 * Set the real number of both TX and RX queues. 3251 * Does nothing if the number of queues is already correct. 3252 */ 3253 int netif_set_real_num_queues(struct net_device *dev, 3254 unsigned int txq, unsigned int rxq) 3255 { 3256 unsigned int old_rxq = dev->real_num_rx_queues; 3257 int err; 3258 3259 if (txq < 1 || txq > dev->num_tx_queues || 3260 rxq < 1 || rxq > dev->num_rx_queues) 3261 return -EINVAL; 3262 3263 /* Start from increases, so the error path only does decreases - 3264 * decreases can't fail. 3265 */ 3266 if (rxq > dev->real_num_rx_queues) { 3267 err = netif_set_real_num_rx_queues(dev, rxq); 3268 if (err) 3269 return err; 3270 } 3271 if (txq > dev->real_num_tx_queues) { 3272 err = netif_set_real_num_tx_queues(dev, txq); 3273 if (err) 3274 goto undo_rx; 3275 } 3276 if (rxq < dev->real_num_rx_queues) 3277 WARN_ON(netif_set_real_num_rx_queues(dev, rxq)); 3278 if (txq < dev->real_num_tx_queues) 3279 WARN_ON(netif_set_real_num_tx_queues(dev, txq)); 3280 3281 return 0; 3282 undo_rx: 3283 WARN_ON(netif_set_real_num_rx_queues(dev, old_rxq)); 3284 return err; 3285 } 3286 EXPORT_SYMBOL(netif_set_real_num_queues); 3287 3288 /** 3289 * netif_set_tso_max_size() - set the max size of TSO frames supported 3290 * @dev: netdev to update 3291 * @size: max skb->len of a TSO frame 3292 * 3293 * Set the limit on the size of TSO super-frames the device can handle. 3294 * Unless explicitly set the stack will assume the value of 3295 * %GSO_LEGACY_MAX_SIZE. 3296 */ 3297 void netif_set_tso_max_size(struct net_device *dev, unsigned int size) 3298 { 3299 dev->tso_max_size = min(GSO_MAX_SIZE, size); 3300 if (size < READ_ONCE(dev->gso_max_size)) 3301 netif_set_gso_max_size(dev, size); 3302 if (size < READ_ONCE(dev->gso_ipv4_max_size)) 3303 netif_set_gso_ipv4_max_size(dev, size); 3304 } 3305 EXPORT_SYMBOL(netif_set_tso_max_size); 3306 3307 /** 3308 * netif_set_tso_max_segs() - set the max number of segs supported for TSO 3309 * @dev: netdev to update 3310 * @segs: max number of TCP segments 3311 * 3312 * Set the limit on the number of TCP segments the device can generate from 3313 * a single TSO super-frame. 3314 * Unless explicitly set the stack will assume the value of %GSO_MAX_SEGS. 3315 */ 3316 void netif_set_tso_max_segs(struct net_device *dev, unsigned int segs) 3317 { 3318 dev->tso_max_segs = segs; 3319 if (segs < READ_ONCE(dev->gso_max_segs)) 3320 netif_set_gso_max_segs(dev, segs); 3321 } 3322 EXPORT_SYMBOL(netif_set_tso_max_segs); 3323 3324 /** 3325 * netif_inherit_tso_max() - copy all TSO limits from a lower device to an upper 3326 * @to: netdev to update 3327 * @from: netdev from which to copy the limits 3328 */ 3329 void netif_inherit_tso_max(struct net_device *to, const struct net_device *from) 3330 { 3331 netif_set_tso_max_size(to, from->tso_max_size); 3332 netif_set_tso_max_segs(to, from->tso_max_segs); 3333 } 3334 EXPORT_SYMBOL(netif_inherit_tso_max); 3335 3336 /** 3337 * netif_get_num_default_rss_queues - default number of RSS queues 3338 * 3339 * Default value is the number of physical cores if there are only 1 or 2, or 3340 * divided by 2 if there are more. 3341 */ 3342 int netif_get_num_default_rss_queues(void) 3343 { 3344 cpumask_var_t cpus; 3345 int cpu, count = 0; 3346 3347 if (unlikely(is_kdump_kernel() || !zalloc_cpumask_var(&cpus, GFP_KERNEL))) 3348 return 1; 3349 3350 cpumask_copy(cpus, cpu_online_mask); 3351 for_each_cpu(cpu, cpus) { 3352 ++count; 3353 cpumask_andnot(cpus, cpus, topology_sibling_cpumask(cpu)); 3354 } 3355 free_cpumask_var(cpus); 3356 3357 return count > 2 ? DIV_ROUND_UP(count, 2) : count; 3358 } 3359 EXPORT_SYMBOL(netif_get_num_default_rss_queues); 3360 3361 static void __netif_reschedule(struct Qdisc *q) 3362 { 3363 struct softnet_data *sd; 3364 unsigned long flags; 3365 3366 local_irq_save(flags); 3367 sd = this_cpu_ptr(&softnet_data); 3368 q->next_sched = NULL; 3369 *sd->output_queue_tailp = q; 3370 sd->output_queue_tailp = &q->next_sched; 3371 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3372 local_irq_restore(flags); 3373 } 3374 3375 void __netif_schedule(struct Qdisc *q) 3376 { 3377 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state)) 3378 __netif_reschedule(q); 3379 } 3380 EXPORT_SYMBOL(__netif_schedule); 3381 3382 struct dev_kfree_skb_cb { 3383 enum skb_drop_reason reason; 3384 }; 3385 3386 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb) 3387 { 3388 return (struct dev_kfree_skb_cb *)skb->cb; 3389 } 3390 3391 void netif_schedule_queue(struct netdev_queue *txq) 3392 { 3393 rcu_read_lock(); 3394 if (!netif_xmit_stopped(txq)) { 3395 struct Qdisc *q = rcu_dereference(txq->qdisc); 3396 3397 __netif_schedule(q); 3398 } 3399 rcu_read_unlock(); 3400 } 3401 EXPORT_SYMBOL(netif_schedule_queue); 3402 3403 void netif_tx_wake_queue(struct netdev_queue *dev_queue) 3404 { 3405 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) { 3406 struct Qdisc *q; 3407 3408 rcu_read_lock(); 3409 q = rcu_dereference(dev_queue->qdisc); 3410 __netif_schedule(q); 3411 rcu_read_unlock(); 3412 } 3413 } 3414 EXPORT_SYMBOL(netif_tx_wake_queue); 3415 3416 void dev_kfree_skb_irq_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3417 { 3418 unsigned long flags; 3419 3420 if (unlikely(!skb)) 3421 return; 3422 3423 if (likely(refcount_read(&skb->users) == 1)) { 3424 smp_rmb(); 3425 refcount_set(&skb->users, 0); 3426 } else if (likely(!refcount_dec_and_test(&skb->users))) { 3427 return; 3428 } 3429 get_kfree_skb_cb(skb)->reason = reason; 3430 local_irq_save(flags); 3431 skb->next = __this_cpu_read(softnet_data.completion_queue); 3432 __this_cpu_write(softnet_data.completion_queue, skb); 3433 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3434 local_irq_restore(flags); 3435 } 3436 EXPORT_SYMBOL(dev_kfree_skb_irq_reason); 3437 3438 void dev_kfree_skb_any_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3439 { 3440 if (in_hardirq() || irqs_disabled()) 3441 dev_kfree_skb_irq_reason(skb, reason); 3442 else 3443 kfree_skb_reason(skb, reason); 3444 } 3445 EXPORT_SYMBOL(dev_kfree_skb_any_reason); 3446 3447 3448 /** 3449 * netif_device_detach - mark device as removed 3450 * @dev: network device 3451 * 3452 * Mark device as removed from system and therefore no longer available. 3453 */ 3454 void netif_device_detach(struct net_device *dev) 3455 { 3456 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) && 3457 netif_running(dev)) { 3458 netif_tx_stop_all_queues(dev); 3459 } 3460 } 3461 EXPORT_SYMBOL(netif_device_detach); 3462 3463 /** 3464 * netif_device_attach - mark device as attached 3465 * @dev: network device 3466 * 3467 * Mark device as attached from system and restart if needed. 3468 */ 3469 void netif_device_attach(struct net_device *dev) 3470 { 3471 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) && 3472 netif_running(dev)) { 3473 netif_tx_wake_all_queues(dev); 3474 netdev_watchdog_up(dev); 3475 } 3476 } 3477 EXPORT_SYMBOL(netif_device_attach); 3478 3479 /* 3480 * Returns a Tx hash based on the given packet descriptor a Tx queues' number 3481 * to be used as a distribution range. 3482 */ 3483 static u16 skb_tx_hash(const struct net_device *dev, 3484 const struct net_device *sb_dev, 3485 struct sk_buff *skb) 3486 { 3487 u32 hash; 3488 u16 qoffset = 0; 3489 u16 qcount = dev->real_num_tx_queues; 3490 3491 if (dev->num_tc) { 3492 u8 tc = netdev_get_prio_tc_map(dev, skb->priority); 3493 3494 qoffset = sb_dev->tc_to_txq[tc].offset; 3495 qcount = sb_dev->tc_to_txq[tc].count; 3496 if (unlikely(!qcount)) { 3497 net_warn_ratelimited("%s: invalid qcount, qoffset %u for tc %u\n", 3498 sb_dev->name, qoffset, tc); 3499 qoffset = 0; 3500 qcount = dev->real_num_tx_queues; 3501 } 3502 } 3503 3504 if (skb_rx_queue_recorded(skb)) { 3505 DEBUG_NET_WARN_ON_ONCE(qcount == 0); 3506 hash = skb_get_rx_queue(skb); 3507 if (hash >= qoffset) 3508 hash -= qoffset; 3509 while (unlikely(hash >= qcount)) 3510 hash -= qcount; 3511 return hash + qoffset; 3512 } 3513 3514 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset; 3515 } 3516 3517 void skb_warn_bad_offload(const struct sk_buff *skb) 3518 { 3519 static const netdev_features_t null_features; 3520 struct net_device *dev = skb->dev; 3521 const char *name = ""; 3522 3523 if (!net_ratelimit()) 3524 return; 3525 3526 if (dev) { 3527 if (dev->dev.parent) 3528 name = dev_driver_string(dev->dev.parent); 3529 else 3530 name = netdev_name(dev); 3531 } 3532 skb_dump(KERN_WARNING, skb, false); 3533 WARN(1, "%s: caps=(%pNF, %pNF)\n", 3534 name, dev ? &dev->features : &null_features, 3535 skb->sk ? &skb->sk->sk_route_caps : &null_features); 3536 } 3537 3538 /* 3539 * Invalidate hardware checksum when packet is to be mangled, and 3540 * complete checksum manually on outgoing path. 3541 */ 3542 int skb_checksum_help(struct sk_buff *skb) 3543 { 3544 __wsum csum; 3545 int ret = 0, offset; 3546 3547 if (skb->ip_summed == CHECKSUM_COMPLETE) 3548 goto out_set_summed; 3549 3550 if (unlikely(skb_is_gso(skb))) { 3551 skb_warn_bad_offload(skb); 3552 return -EINVAL; 3553 } 3554 3555 if (!skb_frags_readable(skb)) { 3556 return -EFAULT; 3557 } 3558 3559 /* Before computing a checksum, we should make sure no frag could 3560 * be modified by an external entity : checksum could be wrong. 3561 */ 3562 if (skb_has_shared_frag(skb)) { 3563 ret = __skb_linearize(skb); 3564 if (ret) 3565 goto out; 3566 } 3567 3568 offset = skb_checksum_start_offset(skb); 3569 ret = -EINVAL; 3570 if (unlikely(offset >= skb_headlen(skb))) { 3571 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3572 WARN_ONCE(true, "offset (%d) >= skb_headlen() (%u)\n", 3573 offset, skb_headlen(skb)); 3574 goto out; 3575 } 3576 csum = skb_checksum(skb, offset, skb->len - offset, 0); 3577 3578 offset += skb->csum_offset; 3579 if (unlikely(offset + sizeof(__sum16) > skb_headlen(skb))) { 3580 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3581 WARN_ONCE(true, "offset+2 (%zu) > skb_headlen() (%u)\n", 3582 offset + sizeof(__sum16), skb_headlen(skb)); 3583 goto out; 3584 } 3585 ret = skb_ensure_writable(skb, offset + sizeof(__sum16)); 3586 if (ret) 3587 goto out; 3588 3589 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0; 3590 out_set_summed: 3591 skb->ip_summed = CHECKSUM_NONE; 3592 out: 3593 return ret; 3594 } 3595 EXPORT_SYMBOL(skb_checksum_help); 3596 3597 int skb_crc32c_csum_help(struct sk_buff *skb) 3598 { 3599 __le32 crc32c_csum; 3600 int ret = 0, offset, start; 3601 3602 if (skb->ip_summed != CHECKSUM_PARTIAL) 3603 goto out; 3604 3605 if (unlikely(skb_is_gso(skb))) 3606 goto out; 3607 3608 /* Before computing a checksum, we should make sure no frag could 3609 * be modified by an external entity : checksum could be wrong. 3610 */ 3611 if (unlikely(skb_has_shared_frag(skb))) { 3612 ret = __skb_linearize(skb); 3613 if (ret) 3614 goto out; 3615 } 3616 start = skb_checksum_start_offset(skb); 3617 offset = start + offsetof(struct sctphdr, checksum); 3618 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) { 3619 ret = -EINVAL; 3620 goto out; 3621 } 3622 3623 ret = skb_ensure_writable(skb, offset + sizeof(__le32)); 3624 if (ret) 3625 goto out; 3626 3627 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start, 3628 skb->len - start, ~(__u32)0, 3629 crc32c_csum_stub)); 3630 *(__le32 *)(skb->data + offset) = crc32c_csum; 3631 skb_reset_csum_not_inet(skb); 3632 out: 3633 return ret; 3634 } 3635 EXPORT_SYMBOL(skb_crc32c_csum_help); 3636 3637 __be16 skb_network_protocol(struct sk_buff *skb, int *depth) 3638 { 3639 __be16 type = skb->protocol; 3640 3641 /* Tunnel gso handlers can set protocol to ethernet. */ 3642 if (type == htons(ETH_P_TEB)) { 3643 struct ethhdr *eth; 3644 3645 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr)))) 3646 return 0; 3647 3648 eth = (struct ethhdr *)skb->data; 3649 type = eth->h_proto; 3650 } 3651 3652 return vlan_get_protocol_and_depth(skb, type, depth); 3653 } 3654 3655 3656 /* Take action when hardware reception checksum errors are detected. */ 3657 #ifdef CONFIG_BUG 3658 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3659 { 3660 netdev_err(dev, "hw csum failure\n"); 3661 skb_dump(KERN_ERR, skb, true); 3662 dump_stack(); 3663 } 3664 3665 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3666 { 3667 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb); 3668 } 3669 EXPORT_SYMBOL(netdev_rx_csum_fault); 3670 #endif 3671 3672 /* XXX: check that highmem exists at all on the given machine. */ 3673 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb) 3674 { 3675 #ifdef CONFIG_HIGHMEM 3676 int i; 3677 3678 if (!(dev->features & NETIF_F_HIGHDMA)) { 3679 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 3680 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 3681 struct page *page = skb_frag_page(frag); 3682 3683 if (page && PageHighMem(page)) 3684 return 1; 3685 } 3686 } 3687 #endif 3688 return 0; 3689 } 3690 3691 /* If MPLS offload request, verify we are testing hardware MPLS features 3692 * instead of standard features for the netdev. 3693 */ 3694 #if IS_ENABLED(CONFIG_NET_MPLS_GSO) 3695 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3696 netdev_features_t features, 3697 __be16 type) 3698 { 3699 if (eth_p_mpls(type)) 3700 features &= skb->dev->mpls_features; 3701 3702 return features; 3703 } 3704 #else 3705 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3706 netdev_features_t features, 3707 __be16 type) 3708 { 3709 return features; 3710 } 3711 #endif 3712 3713 static netdev_features_t harmonize_features(struct sk_buff *skb, 3714 netdev_features_t features) 3715 { 3716 __be16 type; 3717 3718 type = skb_network_protocol(skb, NULL); 3719 features = net_mpls_features(skb, features, type); 3720 3721 if (skb->ip_summed != CHECKSUM_NONE && 3722 !can_checksum_protocol(features, type)) { 3723 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3724 } 3725 if (illegal_highdma(skb->dev, skb)) 3726 features &= ~NETIF_F_SG; 3727 3728 return features; 3729 } 3730 3731 netdev_features_t passthru_features_check(struct sk_buff *skb, 3732 struct net_device *dev, 3733 netdev_features_t features) 3734 { 3735 return features; 3736 } 3737 EXPORT_SYMBOL(passthru_features_check); 3738 3739 static netdev_features_t dflt_features_check(struct sk_buff *skb, 3740 struct net_device *dev, 3741 netdev_features_t features) 3742 { 3743 return vlan_features_check(skb, features); 3744 } 3745 3746 static netdev_features_t gso_features_check(const struct sk_buff *skb, 3747 struct net_device *dev, 3748 netdev_features_t features) 3749 { 3750 u16 gso_segs = skb_shinfo(skb)->gso_segs; 3751 3752 if (gso_segs > READ_ONCE(dev->gso_max_segs)) 3753 return features & ~NETIF_F_GSO_MASK; 3754 3755 if (unlikely(skb->len >= netif_get_gso_max_size(dev, skb))) 3756 return features & ~NETIF_F_GSO_MASK; 3757 3758 if (!skb_shinfo(skb)->gso_type) { 3759 skb_warn_bad_offload(skb); 3760 return features & ~NETIF_F_GSO_MASK; 3761 } 3762 3763 /* Support for GSO partial features requires software 3764 * intervention before we can actually process the packets 3765 * so we need to strip support for any partial features now 3766 * and we can pull them back in after we have partially 3767 * segmented the frame. 3768 */ 3769 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)) 3770 features &= ~dev->gso_partial_features; 3771 3772 /* Make sure to clear the IPv4 ID mangling feature if the 3773 * IPv4 header has the potential to be fragmented. 3774 */ 3775 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) { 3776 struct iphdr *iph = skb->encapsulation ? 3777 inner_ip_hdr(skb) : ip_hdr(skb); 3778 3779 if (!(iph->frag_off & htons(IP_DF))) 3780 features &= ~NETIF_F_TSO_MANGLEID; 3781 } 3782 3783 return features; 3784 } 3785 3786 netdev_features_t netif_skb_features(struct sk_buff *skb) 3787 { 3788 struct net_device *dev = skb->dev; 3789 netdev_features_t features = dev->features; 3790 3791 if (skb_is_gso(skb)) 3792 features = gso_features_check(skb, dev, features); 3793 3794 /* If encapsulation offload request, verify we are testing 3795 * hardware encapsulation features instead of standard 3796 * features for the netdev 3797 */ 3798 if (skb->encapsulation) 3799 features &= dev->hw_enc_features; 3800 3801 if (skb_vlan_tagged(skb)) 3802 features = netdev_intersect_features(features, 3803 dev->vlan_features | 3804 NETIF_F_HW_VLAN_CTAG_TX | 3805 NETIF_F_HW_VLAN_STAG_TX); 3806 3807 if (dev->netdev_ops->ndo_features_check) 3808 features &= dev->netdev_ops->ndo_features_check(skb, dev, 3809 features); 3810 else 3811 features &= dflt_features_check(skb, dev, features); 3812 3813 return harmonize_features(skb, features); 3814 } 3815 EXPORT_SYMBOL(netif_skb_features); 3816 3817 static int xmit_one(struct sk_buff *skb, struct net_device *dev, 3818 struct netdev_queue *txq, bool more) 3819 { 3820 unsigned int len; 3821 int rc; 3822 3823 if (dev_nit_active_rcu(dev)) 3824 dev_queue_xmit_nit(skb, dev); 3825 3826 len = skb->len; 3827 trace_net_dev_start_xmit(skb, dev); 3828 rc = netdev_start_xmit(skb, dev, txq, more); 3829 trace_net_dev_xmit(skb, rc, dev, len); 3830 3831 return rc; 3832 } 3833 3834 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev, 3835 struct netdev_queue *txq, int *ret) 3836 { 3837 struct sk_buff *skb = first; 3838 int rc = NETDEV_TX_OK; 3839 3840 while (skb) { 3841 struct sk_buff *next = skb->next; 3842 3843 skb_mark_not_on_list(skb); 3844 rc = xmit_one(skb, dev, txq, next != NULL); 3845 if (unlikely(!dev_xmit_complete(rc))) { 3846 skb->next = next; 3847 goto out; 3848 } 3849 3850 skb = next; 3851 if (netif_tx_queue_stopped(txq) && skb) { 3852 rc = NETDEV_TX_BUSY; 3853 break; 3854 } 3855 } 3856 3857 out: 3858 *ret = rc; 3859 return skb; 3860 } 3861 3862 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb, 3863 netdev_features_t features) 3864 { 3865 if (skb_vlan_tag_present(skb) && 3866 !vlan_hw_offload_capable(features, skb->vlan_proto)) 3867 skb = __vlan_hwaccel_push_inside(skb); 3868 return skb; 3869 } 3870 3871 int skb_csum_hwoffload_help(struct sk_buff *skb, 3872 const netdev_features_t features) 3873 { 3874 if (unlikely(skb_csum_is_sctp(skb))) 3875 return !!(features & NETIF_F_SCTP_CRC) ? 0 : 3876 skb_crc32c_csum_help(skb); 3877 3878 if (features & NETIF_F_HW_CSUM) 3879 return 0; 3880 3881 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) { 3882 if (vlan_get_protocol(skb) == htons(ETH_P_IPV6) && 3883 skb_network_header_len(skb) != sizeof(struct ipv6hdr) && 3884 !ipv6_has_hopopt_jumbo(skb)) 3885 goto sw_checksum; 3886 3887 switch (skb->csum_offset) { 3888 case offsetof(struct tcphdr, check): 3889 case offsetof(struct udphdr, check): 3890 return 0; 3891 } 3892 } 3893 3894 sw_checksum: 3895 return skb_checksum_help(skb); 3896 } 3897 EXPORT_SYMBOL(skb_csum_hwoffload_help); 3898 3899 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again) 3900 { 3901 netdev_features_t features; 3902 3903 if (!skb_frags_readable(skb)) 3904 goto out_kfree_skb; 3905 3906 features = netif_skb_features(skb); 3907 skb = validate_xmit_vlan(skb, features); 3908 if (unlikely(!skb)) 3909 goto out_null; 3910 3911 skb = sk_validate_xmit_skb(skb, dev); 3912 if (unlikely(!skb)) 3913 goto out_null; 3914 3915 if (netif_needs_gso(skb, features)) { 3916 struct sk_buff *segs; 3917 3918 segs = skb_gso_segment(skb, features); 3919 if (IS_ERR(segs)) { 3920 goto out_kfree_skb; 3921 } else if (segs) { 3922 consume_skb(skb); 3923 skb = segs; 3924 } 3925 } else { 3926 if (skb_needs_linearize(skb, features) && 3927 __skb_linearize(skb)) 3928 goto out_kfree_skb; 3929 3930 /* If packet is not checksummed and device does not 3931 * support checksumming for this protocol, complete 3932 * checksumming here. 3933 */ 3934 if (skb->ip_summed == CHECKSUM_PARTIAL) { 3935 if (skb->encapsulation) 3936 skb_set_inner_transport_header(skb, 3937 skb_checksum_start_offset(skb)); 3938 else 3939 skb_set_transport_header(skb, 3940 skb_checksum_start_offset(skb)); 3941 if (skb_csum_hwoffload_help(skb, features)) 3942 goto out_kfree_skb; 3943 } 3944 } 3945 3946 skb = validate_xmit_xfrm(skb, features, again); 3947 3948 return skb; 3949 3950 out_kfree_skb: 3951 kfree_skb(skb); 3952 out_null: 3953 dev_core_stats_tx_dropped_inc(dev); 3954 return NULL; 3955 } 3956 3957 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again) 3958 { 3959 struct sk_buff *next, *head = NULL, *tail; 3960 3961 for (; skb != NULL; skb = next) { 3962 next = skb->next; 3963 skb_mark_not_on_list(skb); 3964 3965 /* in case skb won't be segmented, point to itself */ 3966 skb->prev = skb; 3967 3968 skb = validate_xmit_skb(skb, dev, again); 3969 if (!skb) 3970 continue; 3971 3972 if (!head) 3973 head = skb; 3974 else 3975 tail->next = skb; 3976 /* If skb was segmented, skb->prev points to 3977 * the last segment. If not, it still contains skb. 3978 */ 3979 tail = skb->prev; 3980 } 3981 return head; 3982 } 3983 EXPORT_SYMBOL_GPL(validate_xmit_skb_list); 3984 3985 static void qdisc_pkt_len_init(struct sk_buff *skb) 3986 { 3987 const struct skb_shared_info *shinfo = skb_shinfo(skb); 3988 3989 qdisc_skb_cb(skb)->pkt_len = skb->len; 3990 3991 /* To get more precise estimation of bytes sent on wire, 3992 * we add to pkt_len the headers size of all segments 3993 */ 3994 if (shinfo->gso_size && skb_transport_header_was_set(skb)) { 3995 u16 gso_segs = shinfo->gso_segs; 3996 unsigned int hdr_len; 3997 3998 /* mac layer + network layer */ 3999 hdr_len = skb_transport_offset(skb); 4000 4001 /* + transport layer */ 4002 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) { 4003 const struct tcphdr *th; 4004 struct tcphdr _tcphdr; 4005 4006 th = skb_header_pointer(skb, hdr_len, 4007 sizeof(_tcphdr), &_tcphdr); 4008 if (likely(th)) 4009 hdr_len += __tcp_hdrlen(th); 4010 } else if (shinfo->gso_type & SKB_GSO_UDP_L4) { 4011 struct udphdr _udphdr; 4012 4013 if (skb_header_pointer(skb, hdr_len, 4014 sizeof(_udphdr), &_udphdr)) 4015 hdr_len += sizeof(struct udphdr); 4016 } 4017 4018 if (unlikely(shinfo->gso_type & SKB_GSO_DODGY)) { 4019 int payload = skb->len - hdr_len; 4020 4021 /* Malicious packet. */ 4022 if (payload <= 0) 4023 return; 4024 gso_segs = DIV_ROUND_UP(payload, shinfo->gso_size); 4025 } 4026 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len; 4027 } 4028 } 4029 4030 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q, 4031 struct sk_buff **to_free, 4032 struct netdev_queue *txq) 4033 { 4034 int rc; 4035 4036 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK; 4037 if (rc == NET_XMIT_SUCCESS) 4038 trace_qdisc_enqueue(q, txq, skb); 4039 return rc; 4040 } 4041 4042 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q, 4043 struct net_device *dev, 4044 struct netdev_queue *txq) 4045 { 4046 spinlock_t *root_lock = qdisc_lock(q); 4047 struct sk_buff *to_free = NULL; 4048 bool contended; 4049 int rc; 4050 4051 qdisc_calculate_pkt_len(skb, q); 4052 4053 tcf_set_drop_reason(skb, SKB_DROP_REASON_QDISC_DROP); 4054 4055 if (q->flags & TCQ_F_NOLOCK) { 4056 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) && 4057 qdisc_run_begin(q)) { 4058 /* Retest nolock_qdisc_is_empty() within the protection 4059 * of q->seqlock to protect from racing with requeuing. 4060 */ 4061 if (unlikely(!nolock_qdisc_is_empty(q))) { 4062 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4063 __qdisc_run(q); 4064 qdisc_run_end(q); 4065 4066 goto no_lock_out; 4067 } 4068 4069 qdisc_bstats_cpu_update(q, skb); 4070 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) && 4071 !nolock_qdisc_is_empty(q)) 4072 __qdisc_run(q); 4073 4074 qdisc_run_end(q); 4075 return NET_XMIT_SUCCESS; 4076 } 4077 4078 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4079 qdisc_run(q); 4080 4081 no_lock_out: 4082 if (unlikely(to_free)) 4083 kfree_skb_list_reason(to_free, 4084 tcf_get_drop_reason(to_free)); 4085 return rc; 4086 } 4087 4088 if (unlikely(READ_ONCE(q->owner) == smp_processor_id())) { 4089 kfree_skb_reason(skb, SKB_DROP_REASON_TC_RECLASSIFY_LOOP); 4090 return NET_XMIT_DROP; 4091 } 4092 /* 4093 * Heuristic to force contended enqueues to serialize on a 4094 * separate lock before trying to get qdisc main lock. 4095 * This permits qdisc->running owner to get the lock more 4096 * often and dequeue packets faster. 4097 * On PREEMPT_RT it is possible to preempt the qdisc owner during xmit 4098 * and then other tasks will only enqueue packets. The packets will be 4099 * sent after the qdisc owner is scheduled again. To prevent this 4100 * scenario the task always serialize on the lock. 4101 */ 4102 contended = qdisc_is_running(q) || IS_ENABLED(CONFIG_PREEMPT_RT); 4103 if (unlikely(contended)) 4104 spin_lock(&q->busylock); 4105 4106 spin_lock(root_lock); 4107 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) { 4108 __qdisc_drop(skb, &to_free); 4109 rc = NET_XMIT_DROP; 4110 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && 4111 qdisc_run_begin(q)) { 4112 /* 4113 * This is a work-conserving queue; there are no old skbs 4114 * waiting to be sent out; and the qdisc is not running - 4115 * xmit the skb directly. 4116 */ 4117 4118 qdisc_bstats_update(q, skb); 4119 4120 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) { 4121 if (unlikely(contended)) { 4122 spin_unlock(&q->busylock); 4123 contended = false; 4124 } 4125 __qdisc_run(q); 4126 } 4127 4128 qdisc_run_end(q); 4129 rc = NET_XMIT_SUCCESS; 4130 } else { 4131 WRITE_ONCE(q->owner, smp_processor_id()); 4132 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4133 WRITE_ONCE(q->owner, -1); 4134 if (qdisc_run_begin(q)) { 4135 if (unlikely(contended)) { 4136 spin_unlock(&q->busylock); 4137 contended = false; 4138 } 4139 __qdisc_run(q); 4140 qdisc_run_end(q); 4141 } 4142 } 4143 spin_unlock(root_lock); 4144 if (unlikely(to_free)) 4145 kfree_skb_list_reason(to_free, 4146 tcf_get_drop_reason(to_free)); 4147 if (unlikely(contended)) 4148 spin_unlock(&q->busylock); 4149 return rc; 4150 } 4151 4152 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO) 4153 static void skb_update_prio(struct sk_buff *skb) 4154 { 4155 const struct netprio_map *map; 4156 const struct sock *sk; 4157 unsigned int prioidx; 4158 4159 if (skb->priority) 4160 return; 4161 map = rcu_dereference_bh(skb->dev->priomap); 4162 if (!map) 4163 return; 4164 sk = skb_to_full_sk(skb); 4165 if (!sk) 4166 return; 4167 4168 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data); 4169 4170 if (prioidx < map->priomap_len) 4171 skb->priority = map->priomap[prioidx]; 4172 } 4173 #else 4174 #define skb_update_prio(skb) 4175 #endif 4176 4177 /** 4178 * dev_loopback_xmit - loop back @skb 4179 * @net: network namespace this loopback is happening in 4180 * @sk: sk needed to be a netfilter okfn 4181 * @skb: buffer to transmit 4182 */ 4183 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb) 4184 { 4185 skb_reset_mac_header(skb); 4186 __skb_pull(skb, skb_network_offset(skb)); 4187 skb->pkt_type = PACKET_LOOPBACK; 4188 if (skb->ip_summed == CHECKSUM_NONE) 4189 skb->ip_summed = CHECKSUM_UNNECESSARY; 4190 DEBUG_NET_WARN_ON_ONCE(!skb_dst(skb)); 4191 skb_dst_force(skb); 4192 netif_rx(skb); 4193 return 0; 4194 } 4195 EXPORT_SYMBOL(dev_loopback_xmit); 4196 4197 #ifdef CONFIG_NET_EGRESS 4198 static struct netdev_queue * 4199 netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb) 4200 { 4201 int qm = skb_get_queue_mapping(skb); 4202 4203 return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm)); 4204 } 4205 4206 #ifndef CONFIG_PREEMPT_RT 4207 static bool netdev_xmit_txqueue_skipped(void) 4208 { 4209 return __this_cpu_read(softnet_data.xmit.skip_txqueue); 4210 } 4211 4212 void netdev_xmit_skip_txqueue(bool skip) 4213 { 4214 __this_cpu_write(softnet_data.xmit.skip_txqueue, skip); 4215 } 4216 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4217 4218 #else 4219 static bool netdev_xmit_txqueue_skipped(void) 4220 { 4221 return current->net_xmit.skip_txqueue; 4222 } 4223 4224 void netdev_xmit_skip_txqueue(bool skip) 4225 { 4226 current->net_xmit.skip_txqueue = skip; 4227 } 4228 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4229 #endif 4230 #endif /* CONFIG_NET_EGRESS */ 4231 4232 #ifdef CONFIG_NET_XGRESS 4233 static int tc_run(struct tcx_entry *entry, struct sk_buff *skb, 4234 enum skb_drop_reason *drop_reason) 4235 { 4236 int ret = TC_ACT_UNSPEC; 4237 #ifdef CONFIG_NET_CLS_ACT 4238 struct mini_Qdisc *miniq = rcu_dereference_bh(entry->miniq); 4239 struct tcf_result res; 4240 4241 if (!miniq) 4242 return ret; 4243 4244 /* Global bypass */ 4245 if (!static_branch_likely(&tcf_sw_enabled_key)) 4246 return ret; 4247 4248 /* Block-wise bypass */ 4249 if (tcf_block_bypass_sw(miniq->block)) 4250 return ret; 4251 4252 tc_skb_cb(skb)->mru = 0; 4253 tc_skb_cb(skb)->post_ct = false; 4254 tcf_set_drop_reason(skb, *drop_reason); 4255 4256 mini_qdisc_bstats_cpu_update(miniq, skb); 4257 ret = tcf_classify(skb, miniq->block, miniq->filter_list, &res, false); 4258 /* Only tcf related quirks below. */ 4259 switch (ret) { 4260 case TC_ACT_SHOT: 4261 *drop_reason = tcf_get_drop_reason(skb); 4262 mini_qdisc_qstats_cpu_drop(miniq); 4263 break; 4264 case TC_ACT_OK: 4265 case TC_ACT_RECLASSIFY: 4266 skb->tc_index = TC_H_MIN(res.classid); 4267 break; 4268 } 4269 #endif /* CONFIG_NET_CLS_ACT */ 4270 return ret; 4271 } 4272 4273 static DEFINE_STATIC_KEY_FALSE(tcx_needed_key); 4274 4275 void tcx_inc(void) 4276 { 4277 static_branch_inc(&tcx_needed_key); 4278 } 4279 4280 void tcx_dec(void) 4281 { 4282 static_branch_dec(&tcx_needed_key); 4283 } 4284 4285 static __always_inline enum tcx_action_base 4286 tcx_run(const struct bpf_mprog_entry *entry, struct sk_buff *skb, 4287 const bool needs_mac) 4288 { 4289 const struct bpf_mprog_fp *fp; 4290 const struct bpf_prog *prog; 4291 int ret = TCX_NEXT; 4292 4293 if (needs_mac) 4294 __skb_push(skb, skb->mac_len); 4295 bpf_mprog_foreach_prog(entry, fp, prog) { 4296 bpf_compute_data_pointers(skb); 4297 ret = bpf_prog_run(prog, skb); 4298 if (ret != TCX_NEXT) 4299 break; 4300 } 4301 if (needs_mac) 4302 __skb_pull(skb, skb->mac_len); 4303 return tcx_action_code(skb, ret); 4304 } 4305 4306 static __always_inline struct sk_buff * 4307 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4308 struct net_device *orig_dev, bool *another) 4309 { 4310 struct bpf_mprog_entry *entry = rcu_dereference_bh(skb->dev->tcx_ingress); 4311 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_INGRESS; 4312 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4313 int sch_ret; 4314 4315 if (!entry) 4316 return skb; 4317 4318 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4319 if (*pt_prev) { 4320 *ret = deliver_skb(skb, *pt_prev, orig_dev); 4321 *pt_prev = NULL; 4322 } 4323 4324 qdisc_skb_cb(skb)->pkt_len = skb->len; 4325 tcx_set_ingress(skb, true); 4326 4327 if (static_branch_unlikely(&tcx_needed_key)) { 4328 sch_ret = tcx_run(entry, skb, true); 4329 if (sch_ret != TC_ACT_UNSPEC) 4330 goto ingress_verdict; 4331 } 4332 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4333 ingress_verdict: 4334 switch (sch_ret) { 4335 case TC_ACT_REDIRECT: 4336 /* skb_mac_header check was done by BPF, so we can safely 4337 * push the L2 header back before redirecting to another 4338 * netdev. 4339 */ 4340 __skb_push(skb, skb->mac_len); 4341 if (skb_do_redirect(skb) == -EAGAIN) { 4342 __skb_pull(skb, skb->mac_len); 4343 *another = true; 4344 break; 4345 } 4346 *ret = NET_RX_SUCCESS; 4347 bpf_net_ctx_clear(bpf_net_ctx); 4348 return NULL; 4349 case TC_ACT_SHOT: 4350 kfree_skb_reason(skb, drop_reason); 4351 *ret = NET_RX_DROP; 4352 bpf_net_ctx_clear(bpf_net_ctx); 4353 return NULL; 4354 /* used by tc_run */ 4355 case TC_ACT_STOLEN: 4356 case TC_ACT_QUEUED: 4357 case TC_ACT_TRAP: 4358 consume_skb(skb); 4359 fallthrough; 4360 case TC_ACT_CONSUMED: 4361 *ret = NET_RX_SUCCESS; 4362 bpf_net_ctx_clear(bpf_net_ctx); 4363 return NULL; 4364 } 4365 bpf_net_ctx_clear(bpf_net_ctx); 4366 4367 return skb; 4368 } 4369 4370 static __always_inline struct sk_buff * 4371 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4372 { 4373 struct bpf_mprog_entry *entry = rcu_dereference_bh(dev->tcx_egress); 4374 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_EGRESS; 4375 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4376 int sch_ret; 4377 4378 if (!entry) 4379 return skb; 4380 4381 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4382 4383 /* qdisc_skb_cb(skb)->pkt_len & tcx_set_ingress() was 4384 * already set by the caller. 4385 */ 4386 if (static_branch_unlikely(&tcx_needed_key)) { 4387 sch_ret = tcx_run(entry, skb, false); 4388 if (sch_ret != TC_ACT_UNSPEC) 4389 goto egress_verdict; 4390 } 4391 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4392 egress_verdict: 4393 switch (sch_ret) { 4394 case TC_ACT_REDIRECT: 4395 /* No need to push/pop skb's mac_header here on egress! */ 4396 skb_do_redirect(skb); 4397 *ret = NET_XMIT_SUCCESS; 4398 bpf_net_ctx_clear(bpf_net_ctx); 4399 return NULL; 4400 case TC_ACT_SHOT: 4401 kfree_skb_reason(skb, drop_reason); 4402 *ret = NET_XMIT_DROP; 4403 bpf_net_ctx_clear(bpf_net_ctx); 4404 return NULL; 4405 /* used by tc_run */ 4406 case TC_ACT_STOLEN: 4407 case TC_ACT_QUEUED: 4408 case TC_ACT_TRAP: 4409 consume_skb(skb); 4410 fallthrough; 4411 case TC_ACT_CONSUMED: 4412 *ret = NET_XMIT_SUCCESS; 4413 bpf_net_ctx_clear(bpf_net_ctx); 4414 return NULL; 4415 } 4416 bpf_net_ctx_clear(bpf_net_ctx); 4417 4418 return skb; 4419 } 4420 #else 4421 static __always_inline struct sk_buff * 4422 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4423 struct net_device *orig_dev, bool *another) 4424 { 4425 return skb; 4426 } 4427 4428 static __always_inline struct sk_buff * 4429 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4430 { 4431 return skb; 4432 } 4433 #endif /* CONFIG_NET_XGRESS */ 4434 4435 #ifdef CONFIG_XPS 4436 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb, 4437 struct xps_dev_maps *dev_maps, unsigned int tci) 4438 { 4439 int tc = netdev_get_prio_tc_map(dev, skb->priority); 4440 struct xps_map *map; 4441 int queue_index = -1; 4442 4443 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids) 4444 return queue_index; 4445 4446 tci *= dev_maps->num_tc; 4447 tci += tc; 4448 4449 map = rcu_dereference(dev_maps->attr_map[tci]); 4450 if (map) { 4451 if (map->len == 1) 4452 queue_index = map->queues[0]; 4453 else 4454 queue_index = map->queues[reciprocal_scale( 4455 skb_get_hash(skb), map->len)]; 4456 if (unlikely(queue_index >= dev->real_num_tx_queues)) 4457 queue_index = -1; 4458 } 4459 return queue_index; 4460 } 4461 #endif 4462 4463 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev, 4464 struct sk_buff *skb) 4465 { 4466 #ifdef CONFIG_XPS 4467 struct xps_dev_maps *dev_maps; 4468 struct sock *sk = skb->sk; 4469 int queue_index = -1; 4470 4471 if (!static_key_false(&xps_needed)) 4472 return -1; 4473 4474 rcu_read_lock(); 4475 if (!static_key_false(&xps_rxqs_needed)) 4476 goto get_cpus_map; 4477 4478 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]); 4479 if (dev_maps) { 4480 int tci = sk_rx_queue_get(sk); 4481 4482 if (tci >= 0) 4483 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4484 tci); 4485 } 4486 4487 get_cpus_map: 4488 if (queue_index < 0) { 4489 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]); 4490 if (dev_maps) { 4491 unsigned int tci = skb->sender_cpu - 1; 4492 4493 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4494 tci); 4495 } 4496 } 4497 rcu_read_unlock(); 4498 4499 return queue_index; 4500 #else 4501 return -1; 4502 #endif 4503 } 4504 4505 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb, 4506 struct net_device *sb_dev) 4507 { 4508 return 0; 4509 } 4510 EXPORT_SYMBOL(dev_pick_tx_zero); 4511 4512 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb, 4513 struct net_device *sb_dev) 4514 { 4515 struct sock *sk = skb->sk; 4516 int queue_index = sk_tx_queue_get(sk); 4517 4518 sb_dev = sb_dev ? : dev; 4519 4520 if (queue_index < 0 || skb->ooo_okay || 4521 queue_index >= dev->real_num_tx_queues) { 4522 int new_index = get_xps_queue(dev, sb_dev, skb); 4523 4524 if (new_index < 0) 4525 new_index = skb_tx_hash(dev, sb_dev, skb); 4526 4527 if (queue_index != new_index && sk && 4528 sk_fullsock(sk) && 4529 rcu_access_pointer(sk->sk_dst_cache)) 4530 sk_tx_queue_set(sk, new_index); 4531 4532 queue_index = new_index; 4533 } 4534 4535 return queue_index; 4536 } 4537 EXPORT_SYMBOL(netdev_pick_tx); 4538 4539 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev, 4540 struct sk_buff *skb, 4541 struct net_device *sb_dev) 4542 { 4543 int queue_index = 0; 4544 4545 #ifdef CONFIG_XPS 4546 u32 sender_cpu = skb->sender_cpu - 1; 4547 4548 if (sender_cpu >= (u32)NR_CPUS) 4549 skb->sender_cpu = raw_smp_processor_id() + 1; 4550 #endif 4551 4552 if (dev->real_num_tx_queues != 1) { 4553 const struct net_device_ops *ops = dev->netdev_ops; 4554 4555 if (ops->ndo_select_queue) 4556 queue_index = ops->ndo_select_queue(dev, skb, sb_dev); 4557 else 4558 queue_index = netdev_pick_tx(dev, skb, sb_dev); 4559 4560 queue_index = netdev_cap_txqueue(dev, queue_index); 4561 } 4562 4563 skb_set_queue_mapping(skb, queue_index); 4564 return netdev_get_tx_queue(dev, queue_index); 4565 } 4566 4567 /** 4568 * __dev_queue_xmit() - transmit a buffer 4569 * @skb: buffer to transmit 4570 * @sb_dev: suboordinate device used for L2 forwarding offload 4571 * 4572 * Queue a buffer for transmission to a network device. The caller must 4573 * have set the device and priority and built the buffer before calling 4574 * this function. The function can be called from an interrupt. 4575 * 4576 * When calling this method, interrupts MUST be enabled. This is because 4577 * the BH enable code must have IRQs enabled so that it will not deadlock. 4578 * 4579 * Regardless of the return value, the skb is consumed, so it is currently 4580 * difficult to retry a send to this method. (You can bump the ref count 4581 * before sending to hold a reference for retry if you are careful.) 4582 * 4583 * Return: 4584 * * 0 - buffer successfully transmitted 4585 * * positive qdisc return code - NET_XMIT_DROP etc. 4586 * * negative errno - other errors 4587 */ 4588 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev) 4589 { 4590 struct net_device *dev = skb->dev; 4591 struct netdev_queue *txq = NULL; 4592 struct Qdisc *q; 4593 int rc = -ENOMEM; 4594 bool again = false; 4595 4596 skb_reset_mac_header(skb); 4597 skb_assert_len(skb); 4598 4599 if (unlikely(skb_shinfo(skb)->tx_flags & 4600 (SKBTX_SCHED_TSTAMP | SKBTX_BPF))) 4601 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED); 4602 4603 /* Disable soft irqs for various locks below. Also 4604 * stops preemption for RCU. 4605 */ 4606 rcu_read_lock_bh(); 4607 4608 skb_update_prio(skb); 4609 4610 qdisc_pkt_len_init(skb); 4611 tcx_set_ingress(skb, false); 4612 #ifdef CONFIG_NET_EGRESS 4613 if (static_branch_unlikely(&egress_needed_key)) { 4614 if (nf_hook_egress_active()) { 4615 skb = nf_hook_egress(skb, &rc, dev); 4616 if (!skb) 4617 goto out; 4618 } 4619 4620 netdev_xmit_skip_txqueue(false); 4621 4622 nf_skip_egress(skb, true); 4623 skb = sch_handle_egress(skb, &rc, dev); 4624 if (!skb) 4625 goto out; 4626 nf_skip_egress(skb, false); 4627 4628 if (netdev_xmit_txqueue_skipped()) 4629 txq = netdev_tx_queue_mapping(dev, skb); 4630 } 4631 #endif 4632 /* If device/qdisc don't need skb->dst, release it right now while 4633 * its hot in this cpu cache. 4634 */ 4635 if (dev->priv_flags & IFF_XMIT_DST_RELEASE) 4636 skb_dst_drop(skb); 4637 else 4638 skb_dst_force(skb); 4639 4640 if (!txq) 4641 txq = netdev_core_pick_tx(dev, skb, sb_dev); 4642 4643 q = rcu_dereference_bh(txq->qdisc); 4644 4645 trace_net_dev_queue(skb); 4646 if (q->enqueue) { 4647 rc = __dev_xmit_skb(skb, q, dev, txq); 4648 goto out; 4649 } 4650 4651 /* The device has no queue. Common case for software devices: 4652 * loopback, all the sorts of tunnels... 4653 4654 * Really, it is unlikely that netif_tx_lock protection is necessary 4655 * here. (f.e. loopback and IP tunnels are clean ignoring statistics 4656 * counters.) 4657 * However, it is possible, that they rely on protection 4658 * made by us here. 4659 4660 * Check this and shot the lock. It is not prone from deadlocks. 4661 *Either shot noqueue qdisc, it is even simpler 8) 4662 */ 4663 if (dev->flags & IFF_UP) { 4664 int cpu = smp_processor_id(); /* ok because BHs are off */ 4665 4666 /* Other cpus might concurrently change txq->xmit_lock_owner 4667 * to -1 or to their cpu id, but not to our id. 4668 */ 4669 if (READ_ONCE(txq->xmit_lock_owner) != cpu) { 4670 if (dev_xmit_recursion()) 4671 goto recursion_alert; 4672 4673 skb = validate_xmit_skb(skb, dev, &again); 4674 if (!skb) 4675 goto out; 4676 4677 HARD_TX_LOCK(dev, txq, cpu); 4678 4679 if (!netif_xmit_stopped(txq)) { 4680 dev_xmit_recursion_inc(); 4681 skb = dev_hard_start_xmit(skb, dev, txq, &rc); 4682 dev_xmit_recursion_dec(); 4683 if (dev_xmit_complete(rc)) { 4684 HARD_TX_UNLOCK(dev, txq); 4685 goto out; 4686 } 4687 } 4688 HARD_TX_UNLOCK(dev, txq); 4689 net_crit_ratelimited("Virtual device %s asks to queue packet!\n", 4690 dev->name); 4691 } else { 4692 /* Recursion is detected! It is possible, 4693 * unfortunately 4694 */ 4695 recursion_alert: 4696 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n", 4697 dev->name); 4698 } 4699 } 4700 4701 rc = -ENETDOWN; 4702 rcu_read_unlock_bh(); 4703 4704 dev_core_stats_tx_dropped_inc(dev); 4705 kfree_skb_list(skb); 4706 return rc; 4707 out: 4708 rcu_read_unlock_bh(); 4709 return rc; 4710 } 4711 EXPORT_SYMBOL(__dev_queue_xmit); 4712 4713 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id) 4714 { 4715 struct net_device *dev = skb->dev; 4716 struct sk_buff *orig_skb = skb; 4717 struct netdev_queue *txq; 4718 int ret = NETDEV_TX_BUSY; 4719 bool again = false; 4720 4721 if (unlikely(!netif_running(dev) || 4722 !netif_carrier_ok(dev))) 4723 goto drop; 4724 4725 skb = validate_xmit_skb_list(skb, dev, &again); 4726 if (skb != orig_skb) 4727 goto drop; 4728 4729 skb_set_queue_mapping(skb, queue_id); 4730 txq = skb_get_tx_queue(dev, skb); 4731 4732 local_bh_disable(); 4733 4734 dev_xmit_recursion_inc(); 4735 HARD_TX_LOCK(dev, txq, smp_processor_id()); 4736 if (!netif_xmit_frozen_or_drv_stopped(txq)) 4737 ret = netdev_start_xmit(skb, dev, txq, false); 4738 HARD_TX_UNLOCK(dev, txq); 4739 dev_xmit_recursion_dec(); 4740 4741 local_bh_enable(); 4742 return ret; 4743 drop: 4744 dev_core_stats_tx_dropped_inc(dev); 4745 kfree_skb_list(skb); 4746 return NET_XMIT_DROP; 4747 } 4748 EXPORT_SYMBOL(__dev_direct_xmit); 4749 4750 /************************************************************************* 4751 * Receiver routines 4752 *************************************************************************/ 4753 static DEFINE_PER_CPU(struct task_struct *, backlog_napi); 4754 4755 int weight_p __read_mostly = 64; /* old backlog weight */ 4756 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */ 4757 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */ 4758 4759 /* Called with irq disabled */ 4760 static inline void ____napi_schedule(struct softnet_data *sd, 4761 struct napi_struct *napi) 4762 { 4763 struct task_struct *thread; 4764 4765 lockdep_assert_irqs_disabled(); 4766 4767 if (test_bit(NAPI_STATE_THREADED, &napi->state)) { 4768 /* Paired with smp_mb__before_atomic() in 4769 * napi_enable()/dev_set_threaded(). 4770 * Use READ_ONCE() to guarantee a complete 4771 * read on napi->thread. Only call 4772 * wake_up_process() when it's not NULL. 4773 */ 4774 thread = READ_ONCE(napi->thread); 4775 if (thread) { 4776 if (use_backlog_threads() && thread == raw_cpu_read(backlog_napi)) 4777 goto use_local_napi; 4778 4779 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 4780 wake_up_process(thread); 4781 return; 4782 } 4783 } 4784 4785 use_local_napi: 4786 list_add_tail(&napi->poll_list, &sd->poll_list); 4787 WRITE_ONCE(napi->list_owner, smp_processor_id()); 4788 /* If not called from net_rx_action() 4789 * we have to raise NET_RX_SOFTIRQ. 4790 */ 4791 if (!sd->in_net_rx_action) 4792 raise_softirq_irqoff(NET_RX_SOFTIRQ); 4793 } 4794 4795 #ifdef CONFIG_RPS 4796 4797 struct static_key_false rps_needed __read_mostly; 4798 EXPORT_SYMBOL(rps_needed); 4799 struct static_key_false rfs_needed __read_mostly; 4800 EXPORT_SYMBOL(rfs_needed); 4801 4802 static u32 rfs_slot(u32 hash, const struct rps_dev_flow_table *flow_table) 4803 { 4804 return hash_32(hash, flow_table->log); 4805 } 4806 4807 static struct rps_dev_flow * 4808 set_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4809 struct rps_dev_flow *rflow, u16 next_cpu) 4810 { 4811 if (next_cpu < nr_cpu_ids) { 4812 u32 head; 4813 #ifdef CONFIG_RFS_ACCEL 4814 struct netdev_rx_queue *rxqueue; 4815 struct rps_dev_flow_table *flow_table; 4816 struct rps_dev_flow *old_rflow; 4817 u16 rxq_index; 4818 u32 flow_id; 4819 int rc; 4820 4821 /* Should we steer this flow to a different hardware queue? */ 4822 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap || 4823 !(dev->features & NETIF_F_NTUPLE)) 4824 goto out; 4825 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu); 4826 if (rxq_index == skb_get_rx_queue(skb)) 4827 goto out; 4828 4829 rxqueue = dev->_rx + rxq_index; 4830 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4831 if (!flow_table) 4832 goto out; 4833 flow_id = rfs_slot(skb_get_hash(skb), flow_table); 4834 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb, 4835 rxq_index, flow_id); 4836 if (rc < 0) 4837 goto out; 4838 old_rflow = rflow; 4839 rflow = &flow_table->flows[flow_id]; 4840 WRITE_ONCE(rflow->filter, rc); 4841 if (old_rflow->filter == rc) 4842 WRITE_ONCE(old_rflow->filter, RPS_NO_FILTER); 4843 out: 4844 #endif 4845 head = READ_ONCE(per_cpu(softnet_data, next_cpu).input_queue_head); 4846 rps_input_queue_tail_save(&rflow->last_qtail, head); 4847 } 4848 4849 WRITE_ONCE(rflow->cpu, next_cpu); 4850 return rflow; 4851 } 4852 4853 /* 4854 * get_rps_cpu is called from netif_receive_skb and returns the target 4855 * CPU from the RPS map of the receiving queue for a given skb. 4856 * rcu_read_lock must be held on entry. 4857 */ 4858 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4859 struct rps_dev_flow **rflowp) 4860 { 4861 const struct rps_sock_flow_table *sock_flow_table; 4862 struct netdev_rx_queue *rxqueue = dev->_rx; 4863 struct rps_dev_flow_table *flow_table; 4864 struct rps_map *map; 4865 int cpu = -1; 4866 u32 tcpu; 4867 u32 hash; 4868 4869 if (skb_rx_queue_recorded(skb)) { 4870 u16 index = skb_get_rx_queue(skb); 4871 4872 if (unlikely(index >= dev->real_num_rx_queues)) { 4873 WARN_ONCE(dev->real_num_rx_queues > 1, 4874 "%s received packet on queue %u, but number " 4875 "of RX queues is %u\n", 4876 dev->name, index, dev->real_num_rx_queues); 4877 goto done; 4878 } 4879 rxqueue += index; 4880 } 4881 4882 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */ 4883 4884 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4885 map = rcu_dereference(rxqueue->rps_map); 4886 if (!flow_table && !map) 4887 goto done; 4888 4889 skb_reset_network_header(skb); 4890 hash = skb_get_hash(skb); 4891 if (!hash) 4892 goto done; 4893 4894 sock_flow_table = rcu_dereference(net_hotdata.rps_sock_flow_table); 4895 if (flow_table && sock_flow_table) { 4896 struct rps_dev_flow *rflow; 4897 u32 next_cpu; 4898 u32 ident; 4899 4900 /* First check into global flow table if there is a match. 4901 * This READ_ONCE() pairs with WRITE_ONCE() from rps_record_sock_flow(). 4902 */ 4903 ident = READ_ONCE(sock_flow_table->ents[hash & sock_flow_table->mask]); 4904 if ((ident ^ hash) & ~net_hotdata.rps_cpu_mask) 4905 goto try_rps; 4906 4907 next_cpu = ident & net_hotdata.rps_cpu_mask; 4908 4909 /* OK, now we know there is a match, 4910 * we can look at the local (per receive queue) flow table 4911 */ 4912 rflow = &flow_table->flows[rfs_slot(hash, flow_table)]; 4913 tcpu = rflow->cpu; 4914 4915 /* 4916 * If the desired CPU (where last recvmsg was done) is 4917 * different from current CPU (one in the rx-queue flow 4918 * table entry), switch if one of the following holds: 4919 * - Current CPU is unset (>= nr_cpu_ids). 4920 * - Current CPU is offline. 4921 * - The current CPU's queue tail has advanced beyond the 4922 * last packet that was enqueued using this table entry. 4923 * This guarantees that all previous packets for the flow 4924 * have been dequeued, thus preserving in order delivery. 4925 */ 4926 if (unlikely(tcpu != next_cpu) && 4927 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) || 4928 ((int)(READ_ONCE(per_cpu(softnet_data, tcpu).input_queue_head) - 4929 rflow->last_qtail)) >= 0)) { 4930 tcpu = next_cpu; 4931 rflow = set_rps_cpu(dev, skb, rflow, next_cpu); 4932 } 4933 4934 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) { 4935 *rflowp = rflow; 4936 cpu = tcpu; 4937 goto done; 4938 } 4939 } 4940 4941 try_rps: 4942 4943 if (map) { 4944 tcpu = map->cpus[reciprocal_scale(hash, map->len)]; 4945 if (cpu_online(tcpu)) { 4946 cpu = tcpu; 4947 goto done; 4948 } 4949 } 4950 4951 done: 4952 return cpu; 4953 } 4954 4955 #ifdef CONFIG_RFS_ACCEL 4956 4957 /** 4958 * rps_may_expire_flow - check whether an RFS hardware filter may be removed 4959 * @dev: Device on which the filter was set 4960 * @rxq_index: RX queue index 4961 * @flow_id: Flow ID passed to ndo_rx_flow_steer() 4962 * @filter_id: Filter ID returned by ndo_rx_flow_steer() 4963 * 4964 * Drivers that implement ndo_rx_flow_steer() should periodically call 4965 * this function for each installed filter and remove the filters for 4966 * which it returns %true. 4967 */ 4968 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index, 4969 u32 flow_id, u16 filter_id) 4970 { 4971 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index; 4972 struct rps_dev_flow_table *flow_table; 4973 struct rps_dev_flow *rflow; 4974 bool expire = true; 4975 unsigned int cpu; 4976 4977 rcu_read_lock(); 4978 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4979 if (flow_table && flow_id < (1UL << flow_table->log)) { 4980 rflow = &flow_table->flows[flow_id]; 4981 cpu = READ_ONCE(rflow->cpu); 4982 if (READ_ONCE(rflow->filter) == filter_id && cpu < nr_cpu_ids && 4983 ((int)(READ_ONCE(per_cpu(softnet_data, cpu).input_queue_head) - 4984 READ_ONCE(rflow->last_qtail)) < 4985 (int)(10 << flow_table->log))) 4986 expire = false; 4987 } 4988 rcu_read_unlock(); 4989 return expire; 4990 } 4991 EXPORT_SYMBOL(rps_may_expire_flow); 4992 4993 #endif /* CONFIG_RFS_ACCEL */ 4994 4995 /* Called from hardirq (IPI) context */ 4996 static void rps_trigger_softirq(void *data) 4997 { 4998 struct softnet_data *sd = data; 4999 5000 ____napi_schedule(sd, &sd->backlog); 5001 /* Pairs with READ_ONCE() in softnet_seq_show() */ 5002 WRITE_ONCE(sd->received_rps, sd->received_rps + 1); 5003 } 5004 5005 #endif /* CONFIG_RPS */ 5006 5007 /* Called from hardirq (IPI) context */ 5008 static void trigger_rx_softirq(void *data) 5009 { 5010 struct softnet_data *sd = data; 5011 5012 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 5013 smp_store_release(&sd->defer_ipi_scheduled, 0); 5014 } 5015 5016 /* 5017 * After we queued a packet into sd->input_pkt_queue, 5018 * we need to make sure this queue is serviced soon. 5019 * 5020 * - If this is another cpu queue, link it to our rps_ipi_list, 5021 * and make sure we will process rps_ipi_list from net_rx_action(). 5022 * 5023 * - If this is our own queue, NAPI schedule our backlog. 5024 * Note that this also raises NET_RX_SOFTIRQ. 5025 */ 5026 static void napi_schedule_rps(struct softnet_data *sd) 5027 { 5028 struct softnet_data *mysd = this_cpu_ptr(&softnet_data); 5029 5030 #ifdef CONFIG_RPS 5031 if (sd != mysd) { 5032 if (use_backlog_threads()) { 5033 __napi_schedule_irqoff(&sd->backlog); 5034 return; 5035 } 5036 5037 sd->rps_ipi_next = mysd->rps_ipi_list; 5038 mysd->rps_ipi_list = sd; 5039 5040 /* If not called from net_rx_action() or napi_threaded_poll() 5041 * we have to raise NET_RX_SOFTIRQ. 5042 */ 5043 if (!mysd->in_net_rx_action && !mysd->in_napi_threaded_poll) 5044 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 5045 return; 5046 } 5047 #endif /* CONFIG_RPS */ 5048 __napi_schedule_irqoff(&mysd->backlog); 5049 } 5050 5051 void kick_defer_list_purge(struct softnet_data *sd, unsigned int cpu) 5052 { 5053 unsigned long flags; 5054 5055 if (use_backlog_threads()) { 5056 backlog_lock_irq_save(sd, &flags); 5057 5058 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) 5059 __napi_schedule_irqoff(&sd->backlog); 5060 5061 backlog_unlock_irq_restore(sd, &flags); 5062 5063 } else if (!cmpxchg(&sd->defer_ipi_scheduled, 0, 1)) { 5064 smp_call_function_single_async(cpu, &sd->defer_csd); 5065 } 5066 } 5067 5068 #ifdef CONFIG_NET_FLOW_LIMIT 5069 int netdev_flow_limit_table_len __read_mostly = (1 << 12); 5070 #endif 5071 5072 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen) 5073 { 5074 #ifdef CONFIG_NET_FLOW_LIMIT 5075 struct sd_flow_limit *fl; 5076 struct softnet_data *sd; 5077 unsigned int old_flow, new_flow; 5078 5079 if (qlen < (READ_ONCE(net_hotdata.max_backlog) >> 1)) 5080 return false; 5081 5082 sd = this_cpu_ptr(&softnet_data); 5083 5084 rcu_read_lock(); 5085 fl = rcu_dereference(sd->flow_limit); 5086 if (fl) { 5087 new_flow = hash_32(skb_get_hash(skb), fl->log_buckets); 5088 old_flow = fl->history[fl->history_head]; 5089 fl->history[fl->history_head] = new_flow; 5090 5091 fl->history_head++; 5092 fl->history_head &= FLOW_LIMIT_HISTORY - 1; 5093 5094 if (likely(fl->buckets[old_flow])) 5095 fl->buckets[old_flow]--; 5096 5097 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) { 5098 /* Pairs with READ_ONCE() in softnet_seq_show() */ 5099 WRITE_ONCE(fl->count, fl->count + 1); 5100 rcu_read_unlock(); 5101 return true; 5102 } 5103 } 5104 rcu_read_unlock(); 5105 #endif 5106 return false; 5107 } 5108 5109 /* 5110 * enqueue_to_backlog is called to queue an skb to a per CPU backlog 5111 * queue (may be a remote CPU queue). 5112 */ 5113 static int enqueue_to_backlog(struct sk_buff *skb, int cpu, 5114 unsigned int *qtail) 5115 { 5116 enum skb_drop_reason reason; 5117 struct softnet_data *sd; 5118 unsigned long flags; 5119 unsigned int qlen; 5120 int max_backlog; 5121 u32 tail; 5122 5123 reason = SKB_DROP_REASON_DEV_READY; 5124 if (!netif_running(skb->dev)) 5125 goto bad_dev; 5126 5127 reason = SKB_DROP_REASON_CPU_BACKLOG; 5128 sd = &per_cpu(softnet_data, cpu); 5129 5130 qlen = skb_queue_len_lockless(&sd->input_pkt_queue); 5131 max_backlog = READ_ONCE(net_hotdata.max_backlog); 5132 if (unlikely(qlen > max_backlog)) 5133 goto cpu_backlog_drop; 5134 backlog_lock_irq_save(sd, &flags); 5135 qlen = skb_queue_len(&sd->input_pkt_queue); 5136 if (qlen <= max_backlog && !skb_flow_limit(skb, qlen)) { 5137 if (!qlen) { 5138 /* Schedule NAPI for backlog device. We can use 5139 * non atomic operation as we own the queue lock. 5140 */ 5141 if (!__test_and_set_bit(NAPI_STATE_SCHED, 5142 &sd->backlog.state)) 5143 napi_schedule_rps(sd); 5144 } 5145 __skb_queue_tail(&sd->input_pkt_queue, skb); 5146 tail = rps_input_queue_tail_incr(sd); 5147 backlog_unlock_irq_restore(sd, &flags); 5148 5149 /* save the tail outside of the critical section */ 5150 rps_input_queue_tail_save(qtail, tail); 5151 return NET_RX_SUCCESS; 5152 } 5153 5154 backlog_unlock_irq_restore(sd, &flags); 5155 5156 cpu_backlog_drop: 5157 atomic_inc(&sd->dropped); 5158 bad_dev: 5159 dev_core_stats_rx_dropped_inc(skb->dev); 5160 kfree_skb_reason(skb, reason); 5161 return NET_RX_DROP; 5162 } 5163 5164 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb) 5165 { 5166 struct net_device *dev = skb->dev; 5167 struct netdev_rx_queue *rxqueue; 5168 5169 rxqueue = dev->_rx; 5170 5171 if (skb_rx_queue_recorded(skb)) { 5172 u16 index = skb_get_rx_queue(skb); 5173 5174 if (unlikely(index >= dev->real_num_rx_queues)) { 5175 WARN_ONCE(dev->real_num_rx_queues > 1, 5176 "%s received packet on queue %u, but number " 5177 "of RX queues is %u\n", 5178 dev->name, index, dev->real_num_rx_queues); 5179 5180 return rxqueue; /* Return first rxqueue */ 5181 } 5182 rxqueue += index; 5183 } 5184 return rxqueue; 5185 } 5186 5187 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, 5188 const struct bpf_prog *xdp_prog) 5189 { 5190 void *orig_data, *orig_data_end, *hard_start; 5191 struct netdev_rx_queue *rxqueue; 5192 bool orig_bcast, orig_host; 5193 u32 mac_len, frame_sz; 5194 __be16 orig_eth_type; 5195 struct ethhdr *eth; 5196 u32 metalen, act; 5197 int off; 5198 5199 /* The XDP program wants to see the packet starting at the MAC 5200 * header. 5201 */ 5202 mac_len = skb->data - skb_mac_header(skb); 5203 hard_start = skb->data - skb_headroom(skb); 5204 5205 /* SKB "head" area always have tailroom for skb_shared_info */ 5206 frame_sz = (void *)skb_end_pointer(skb) - hard_start; 5207 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 5208 5209 rxqueue = netif_get_rxqueue(skb); 5210 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq); 5211 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len, 5212 skb_headlen(skb) + mac_len, true); 5213 if (skb_is_nonlinear(skb)) { 5214 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 5215 xdp_buff_set_frags_flag(xdp); 5216 } else { 5217 xdp_buff_clear_frags_flag(xdp); 5218 } 5219 5220 orig_data_end = xdp->data_end; 5221 orig_data = xdp->data; 5222 eth = (struct ethhdr *)xdp->data; 5223 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr); 5224 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest); 5225 orig_eth_type = eth->h_proto; 5226 5227 act = bpf_prog_run_xdp(xdp_prog, xdp); 5228 5229 /* check if bpf_xdp_adjust_head was used */ 5230 off = xdp->data - orig_data; 5231 if (off) { 5232 if (off > 0) 5233 __skb_pull(skb, off); 5234 else if (off < 0) 5235 __skb_push(skb, -off); 5236 5237 skb->mac_header += off; 5238 skb_reset_network_header(skb); 5239 } 5240 5241 /* check if bpf_xdp_adjust_tail was used */ 5242 off = xdp->data_end - orig_data_end; 5243 if (off != 0) { 5244 skb_set_tail_pointer(skb, xdp->data_end - xdp->data); 5245 skb->len += off; /* positive on grow, negative on shrink */ 5246 } 5247 5248 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 5249 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here. 5250 */ 5251 if (xdp_buff_has_frags(xdp)) 5252 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 5253 else 5254 skb->data_len = 0; 5255 5256 /* check if XDP changed eth hdr such SKB needs update */ 5257 eth = (struct ethhdr *)xdp->data; 5258 if ((orig_eth_type != eth->h_proto) || 5259 (orig_host != ether_addr_equal_64bits(eth->h_dest, 5260 skb->dev->dev_addr)) || 5261 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) { 5262 __skb_push(skb, ETH_HLEN); 5263 skb->pkt_type = PACKET_HOST; 5264 skb->protocol = eth_type_trans(skb, skb->dev); 5265 } 5266 5267 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull 5268 * before calling us again on redirect path. We do not call do_redirect 5269 * as we leave that up to the caller. 5270 * 5271 * Caller is responsible for managing lifetime of skb (i.e. calling 5272 * kfree_skb in response to actions it cannot handle/XDP_DROP). 5273 */ 5274 switch (act) { 5275 case XDP_REDIRECT: 5276 case XDP_TX: 5277 __skb_push(skb, mac_len); 5278 break; 5279 case XDP_PASS: 5280 metalen = xdp->data - xdp->data_meta; 5281 if (metalen) 5282 skb_metadata_set(skb, metalen); 5283 break; 5284 } 5285 5286 return act; 5287 } 5288 5289 static int 5290 netif_skb_check_for_xdp(struct sk_buff **pskb, const struct bpf_prog *prog) 5291 { 5292 struct sk_buff *skb = *pskb; 5293 int err, hroom, troom; 5294 5295 if (!skb_cow_data_for_xdp(this_cpu_read(system_page_pool), pskb, prog)) 5296 return 0; 5297 5298 /* In case we have to go down the path and also linearize, 5299 * then lets do the pskb_expand_head() work just once here. 5300 */ 5301 hroom = XDP_PACKET_HEADROOM - skb_headroom(skb); 5302 troom = skb->tail + skb->data_len - skb->end; 5303 err = pskb_expand_head(skb, 5304 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0, 5305 troom > 0 ? troom + 128 : 0, GFP_ATOMIC); 5306 if (err) 5307 return err; 5308 5309 return skb_linearize(skb); 5310 } 5311 5312 static u32 netif_receive_generic_xdp(struct sk_buff **pskb, 5313 struct xdp_buff *xdp, 5314 const struct bpf_prog *xdp_prog) 5315 { 5316 struct sk_buff *skb = *pskb; 5317 u32 mac_len, act = XDP_DROP; 5318 5319 /* Reinjected packets coming from act_mirred or similar should 5320 * not get XDP generic processing. 5321 */ 5322 if (skb_is_redirected(skb)) 5323 return XDP_PASS; 5324 5325 /* XDP packets must have sufficient headroom of XDP_PACKET_HEADROOM 5326 * bytes. This is the guarantee that also native XDP provides, 5327 * thus we need to do it here as well. 5328 */ 5329 mac_len = skb->data - skb_mac_header(skb); 5330 __skb_push(skb, mac_len); 5331 5332 if (skb_cloned(skb) || skb_is_nonlinear(skb) || 5333 skb_headroom(skb) < XDP_PACKET_HEADROOM) { 5334 if (netif_skb_check_for_xdp(pskb, xdp_prog)) 5335 goto do_drop; 5336 } 5337 5338 __skb_pull(*pskb, mac_len); 5339 5340 act = bpf_prog_run_generic_xdp(*pskb, xdp, xdp_prog); 5341 switch (act) { 5342 case XDP_REDIRECT: 5343 case XDP_TX: 5344 case XDP_PASS: 5345 break; 5346 default: 5347 bpf_warn_invalid_xdp_action((*pskb)->dev, xdp_prog, act); 5348 fallthrough; 5349 case XDP_ABORTED: 5350 trace_xdp_exception((*pskb)->dev, xdp_prog, act); 5351 fallthrough; 5352 case XDP_DROP: 5353 do_drop: 5354 kfree_skb(*pskb); 5355 break; 5356 } 5357 5358 return act; 5359 } 5360 5361 /* When doing generic XDP we have to bypass the qdisc layer and the 5362 * network taps in order to match in-driver-XDP behavior. This also means 5363 * that XDP packets are able to starve other packets going through a qdisc, 5364 * and DDOS attacks will be more effective. In-driver-XDP use dedicated TX 5365 * queues, so they do not have this starvation issue. 5366 */ 5367 void generic_xdp_tx(struct sk_buff *skb, const struct bpf_prog *xdp_prog) 5368 { 5369 struct net_device *dev = skb->dev; 5370 struct netdev_queue *txq; 5371 bool free_skb = true; 5372 int cpu, rc; 5373 5374 txq = netdev_core_pick_tx(dev, skb, NULL); 5375 cpu = smp_processor_id(); 5376 HARD_TX_LOCK(dev, txq, cpu); 5377 if (!netif_xmit_frozen_or_drv_stopped(txq)) { 5378 rc = netdev_start_xmit(skb, dev, txq, 0); 5379 if (dev_xmit_complete(rc)) 5380 free_skb = false; 5381 } 5382 HARD_TX_UNLOCK(dev, txq); 5383 if (free_skb) { 5384 trace_xdp_exception(dev, xdp_prog, XDP_TX); 5385 dev_core_stats_tx_dropped_inc(dev); 5386 kfree_skb(skb); 5387 } 5388 } 5389 5390 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key); 5391 5392 int do_xdp_generic(const struct bpf_prog *xdp_prog, struct sk_buff **pskb) 5393 { 5394 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 5395 5396 if (xdp_prog) { 5397 struct xdp_buff xdp; 5398 u32 act; 5399 int err; 5400 5401 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 5402 act = netif_receive_generic_xdp(pskb, &xdp, xdp_prog); 5403 if (act != XDP_PASS) { 5404 switch (act) { 5405 case XDP_REDIRECT: 5406 err = xdp_do_generic_redirect((*pskb)->dev, *pskb, 5407 &xdp, xdp_prog); 5408 if (err) 5409 goto out_redir; 5410 break; 5411 case XDP_TX: 5412 generic_xdp_tx(*pskb, xdp_prog); 5413 break; 5414 } 5415 bpf_net_ctx_clear(bpf_net_ctx); 5416 return XDP_DROP; 5417 } 5418 bpf_net_ctx_clear(bpf_net_ctx); 5419 } 5420 return XDP_PASS; 5421 out_redir: 5422 bpf_net_ctx_clear(bpf_net_ctx); 5423 kfree_skb_reason(*pskb, SKB_DROP_REASON_XDP); 5424 return XDP_DROP; 5425 } 5426 EXPORT_SYMBOL_GPL(do_xdp_generic); 5427 5428 static int netif_rx_internal(struct sk_buff *skb) 5429 { 5430 int ret; 5431 5432 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5433 5434 trace_netif_rx(skb); 5435 5436 #ifdef CONFIG_RPS 5437 if (static_branch_unlikely(&rps_needed)) { 5438 struct rps_dev_flow voidflow, *rflow = &voidflow; 5439 int cpu; 5440 5441 rcu_read_lock(); 5442 5443 cpu = get_rps_cpu(skb->dev, skb, &rflow); 5444 if (cpu < 0) 5445 cpu = smp_processor_id(); 5446 5447 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5448 5449 rcu_read_unlock(); 5450 } else 5451 #endif 5452 { 5453 unsigned int qtail; 5454 5455 ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail); 5456 } 5457 return ret; 5458 } 5459 5460 /** 5461 * __netif_rx - Slightly optimized version of netif_rx 5462 * @skb: buffer to post 5463 * 5464 * This behaves as netif_rx except that it does not disable bottom halves. 5465 * As a result this function may only be invoked from the interrupt context 5466 * (either hard or soft interrupt). 5467 */ 5468 int __netif_rx(struct sk_buff *skb) 5469 { 5470 int ret; 5471 5472 lockdep_assert_once(hardirq_count() | softirq_count()); 5473 5474 trace_netif_rx_entry(skb); 5475 ret = netif_rx_internal(skb); 5476 trace_netif_rx_exit(ret); 5477 return ret; 5478 } 5479 EXPORT_SYMBOL(__netif_rx); 5480 5481 /** 5482 * netif_rx - post buffer to the network code 5483 * @skb: buffer to post 5484 * 5485 * This function receives a packet from a device driver and queues it for 5486 * the upper (protocol) levels to process via the backlog NAPI device. It 5487 * always succeeds. The buffer may be dropped during processing for 5488 * congestion control or by the protocol layers. 5489 * The network buffer is passed via the backlog NAPI device. Modern NIC 5490 * driver should use NAPI and GRO. 5491 * This function can used from interrupt and from process context. The 5492 * caller from process context must not disable interrupts before invoking 5493 * this function. 5494 * 5495 * return values: 5496 * NET_RX_SUCCESS (no congestion) 5497 * NET_RX_DROP (packet was dropped) 5498 * 5499 */ 5500 int netif_rx(struct sk_buff *skb) 5501 { 5502 bool need_bh_off = !(hardirq_count() | softirq_count()); 5503 int ret; 5504 5505 if (need_bh_off) 5506 local_bh_disable(); 5507 trace_netif_rx_entry(skb); 5508 ret = netif_rx_internal(skb); 5509 trace_netif_rx_exit(ret); 5510 if (need_bh_off) 5511 local_bh_enable(); 5512 return ret; 5513 } 5514 EXPORT_SYMBOL(netif_rx); 5515 5516 static __latent_entropy void net_tx_action(void) 5517 { 5518 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 5519 5520 if (sd->completion_queue) { 5521 struct sk_buff *clist; 5522 5523 local_irq_disable(); 5524 clist = sd->completion_queue; 5525 sd->completion_queue = NULL; 5526 local_irq_enable(); 5527 5528 while (clist) { 5529 struct sk_buff *skb = clist; 5530 5531 clist = clist->next; 5532 5533 WARN_ON(refcount_read(&skb->users)); 5534 if (likely(get_kfree_skb_cb(skb)->reason == SKB_CONSUMED)) 5535 trace_consume_skb(skb, net_tx_action); 5536 else 5537 trace_kfree_skb(skb, net_tx_action, 5538 get_kfree_skb_cb(skb)->reason, NULL); 5539 5540 if (skb->fclone != SKB_FCLONE_UNAVAILABLE) 5541 __kfree_skb(skb); 5542 else 5543 __napi_kfree_skb(skb, 5544 get_kfree_skb_cb(skb)->reason); 5545 } 5546 } 5547 5548 if (sd->output_queue) { 5549 struct Qdisc *head; 5550 5551 local_irq_disable(); 5552 head = sd->output_queue; 5553 sd->output_queue = NULL; 5554 sd->output_queue_tailp = &sd->output_queue; 5555 local_irq_enable(); 5556 5557 rcu_read_lock(); 5558 5559 while (head) { 5560 struct Qdisc *q = head; 5561 spinlock_t *root_lock = NULL; 5562 5563 head = head->next_sched; 5564 5565 /* We need to make sure head->next_sched is read 5566 * before clearing __QDISC_STATE_SCHED 5567 */ 5568 smp_mb__before_atomic(); 5569 5570 if (!(q->flags & TCQ_F_NOLOCK)) { 5571 root_lock = qdisc_lock(q); 5572 spin_lock(root_lock); 5573 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, 5574 &q->state))) { 5575 /* There is a synchronize_net() between 5576 * STATE_DEACTIVATED flag being set and 5577 * qdisc_reset()/some_qdisc_is_busy() in 5578 * dev_deactivate(), so we can safely bail out 5579 * early here to avoid data race between 5580 * qdisc_deactivate() and some_qdisc_is_busy() 5581 * for lockless qdisc. 5582 */ 5583 clear_bit(__QDISC_STATE_SCHED, &q->state); 5584 continue; 5585 } 5586 5587 clear_bit(__QDISC_STATE_SCHED, &q->state); 5588 qdisc_run(q); 5589 if (root_lock) 5590 spin_unlock(root_lock); 5591 } 5592 5593 rcu_read_unlock(); 5594 } 5595 5596 xfrm_dev_backlog(sd); 5597 } 5598 5599 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE) 5600 /* This hook is defined here for ATM LANE */ 5601 int (*br_fdb_test_addr_hook)(struct net_device *dev, 5602 unsigned char *addr) __read_mostly; 5603 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook); 5604 #endif 5605 5606 /** 5607 * netdev_is_rx_handler_busy - check if receive handler is registered 5608 * @dev: device to check 5609 * 5610 * Check if a receive handler is already registered for a given device. 5611 * Return true if there one. 5612 * 5613 * The caller must hold the rtnl_mutex. 5614 */ 5615 bool netdev_is_rx_handler_busy(struct net_device *dev) 5616 { 5617 ASSERT_RTNL(); 5618 return dev && rtnl_dereference(dev->rx_handler); 5619 } 5620 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy); 5621 5622 /** 5623 * netdev_rx_handler_register - register receive handler 5624 * @dev: device to register a handler for 5625 * @rx_handler: receive handler to register 5626 * @rx_handler_data: data pointer that is used by rx handler 5627 * 5628 * Register a receive handler for a device. This handler will then be 5629 * called from __netif_receive_skb. A negative errno code is returned 5630 * on a failure. 5631 * 5632 * The caller must hold the rtnl_mutex. 5633 * 5634 * For a general description of rx_handler, see enum rx_handler_result. 5635 */ 5636 int netdev_rx_handler_register(struct net_device *dev, 5637 rx_handler_func_t *rx_handler, 5638 void *rx_handler_data) 5639 { 5640 if (netdev_is_rx_handler_busy(dev)) 5641 return -EBUSY; 5642 5643 if (dev->priv_flags & IFF_NO_RX_HANDLER) 5644 return -EINVAL; 5645 5646 /* Note: rx_handler_data must be set before rx_handler */ 5647 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data); 5648 rcu_assign_pointer(dev->rx_handler, rx_handler); 5649 5650 return 0; 5651 } 5652 EXPORT_SYMBOL_GPL(netdev_rx_handler_register); 5653 5654 /** 5655 * netdev_rx_handler_unregister - unregister receive handler 5656 * @dev: device to unregister a handler from 5657 * 5658 * Unregister a receive handler from a device. 5659 * 5660 * The caller must hold the rtnl_mutex. 5661 */ 5662 void netdev_rx_handler_unregister(struct net_device *dev) 5663 { 5664 5665 ASSERT_RTNL(); 5666 RCU_INIT_POINTER(dev->rx_handler, NULL); 5667 /* a reader seeing a non NULL rx_handler in a rcu_read_lock() 5668 * section has a guarantee to see a non NULL rx_handler_data 5669 * as well. 5670 */ 5671 synchronize_net(); 5672 RCU_INIT_POINTER(dev->rx_handler_data, NULL); 5673 } 5674 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister); 5675 5676 /* 5677 * Limit the use of PFMEMALLOC reserves to those protocols that implement 5678 * the special handling of PFMEMALLOC skbs. 5679 */ 5680 static bool skb_pfmemalloc_protocol(struct sk_buff *skb) 5681 { 5682 switch (skb->protocol) { 5683 case htons(ETH_P_ARP): 5684 case htons(ETH_P_IP): 5685 case htons(ETH_P_IPV6): 5686 case htons(ETH_P_8021Q): 5687 case htons(ETH_P_8021AD): 5688 return true; 5689 default: 5690 return false; 5691 } 5692 } 5693 5694 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev, 5695 int *ret, struct net_device *orig_dev) 5696 { 5697 if (nf_hook_ingress_active(skb)) { 5698 int ingress_retval; 5699 5700 if (*pt_prev) { 5701 *ret = deliver_skb(skb, *pt_prev, orig_dev); 5702 *pt_prev = NULL; 5703 } 5704 5705 rcu_read_lock(); 5706 ingress_retval = nf_hook_ingress(skb); 5707 rcu_read_unlock(); 5708 return ingress_retval; 5709 } 5710 return 0; 5711 } 5712 5713 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc, 5714 struct packet_type **ppt_prev) 5715 { 5716 struct packet_type *ptype, *pt_prev; 5717 rx_handler_func_t *rx_handler; 5718 struct sk_buff *skb = *pskb; 5719 struct net_device *orig_dev; 5720 bool deliver_exact = false; 5721 int ret = NET_RX_DROP; 5722 __be16 type; 5723 5724 net_timestamp_check(!READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5725 5726 trace_netif_receive_skb(skb); 5727 5728 orig_dev = skb->dev; 5729 5730 skb_reset_network_header(skb); 5731 #if !defined(CONFIG_DEBUG_NET) 5732 /* We plan to no longer reset the transport header here. 5733 * Give some time to fuzzers and dev build to catch bugs 5734 * in network stacks. 5735 */ 5736 if (!skb_transport_header_was_set(skb)) 5737 skb_reset_transport_header(skb); 5738 #endif 5739 skb_reset_mac_len(skb); 5740 5741 pt_prev = NULL; 5742 5743 another_round: 5744 skb->skb_iif = skb->dev->ifindex; 5745 5746 __this_cpu_inc(softnet_data.processed); 5747 5748 if (static_branch_unlikely(&generic_xdp_needed_key)) { 5749 int ret2; 5750 5751 migrate_disable(); 5752 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), 5753 &skb); 5754 migrate_enable(); 5755 5756 if (ret2 != XDP_PASS) { 5757 ret = NET_RX_DROP; 5758 goto out; 5759 } 5760 } 5761 5762 if (eth_type_vlan(skb->protocol)) { 5763 skb = skb_vlan_untag(skb); 5764 if (unlikely(!skb)) 5765 goto out; 5766 } 5767 5768 if (skb_skip_tc_classify(skb)) 5769 goto skip_classify; 5770 5771 if (pfmemalloc) 5772 goto skip_taps; 5773 5774 list_for_each_entry_rcu(ptype, &dev_net_rcu(skb->dev)->ptype_all, 5775 list) { 5776 if (pt_prev) 5777 ret = deliver_skb(skb, pt_prev, orig_dev); 5778 pt_prev = ptype; 5779 } 5780 5781 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) { 5782 if (pt_prev) 5783 ret = deliver_skb(skb, pt_prev, orig_dev); 5784 pt_prev = ptype; 5785 } 5786 5787 skip_taps: 5788 #ifdef CONFIG_NET_INGRESS 5789 if (static_branch_unlikely(&ingress_needed_key)) { 5790 bool another = false; 5791 5792 nf_skip_egress(skb, true); 5793 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev, 5794 &another); 5795 if (another) 5796 goto another_round; 5797 if (!skb) 5798 goto out; 5799 5800 nf_skip_egress(skb, false); 5801 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0) 5802 goto out; 5803 } 5804 #endif 5805 skb_reset_redirect(skb); 5806 skip_classify: 5807 if (pfmemalloc && !skb_pfmemalloc_protocol(skb)) 5808 goto drop; 5809 5810 if (skb_vlan_tag_present(skb)) { 5811 if (pt_prev) { 5812 ret = deliver_skb(skb, pt_prev, orig_dev); 5813 pt_prev = NULL; 5814 } 5815 if (vlan_do_receive(&skb)) 5816 goto another_round; 5817 else if (unlikely(!skb)) 5818 goto out; 5819 } 5820 5821 rx_handler = rcu_dereference(skb->dev->rx_handler); 5822 if (rx_handler) { 5823 if (pt_prev) { 5824 ret = deliver_skb(skb, pt_prev, orig_dev); 5825 pt_prev = NULL; 5826 } 5827 switch (rx_handler(&skb)) { 5828 case RX_HANDLER_CONSUMED: 5829 ret = NET_RX_SUCCESS; 5830 goto out; 5831 case RX_HANDLER_ANOTHER: 5832 goto another_round; 5833 case RX_HANDLER_EXACT: 5834 deliver_exact = true; 5835 break; 5836 case RX_HANDLER_PASS: 5837 break; 5838 default: 5839 BUG(); 5840 } 5841 } 5842 5843 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) { 5844 check_vlan_id: 5845 if (skb_vlan_tag_get_id(skb)) { 5846 /* Vlan id is non 0 and vlan_do_receive() above couldn't 5847 * find vlan device. 5848 */ 5849 skb->pkt_type = PACKET_OTHERHOST; 5850 } else if (eth_type_vlan(skb->protocol)) { 5851 /* Outer header is 802.1P with vlan 0, inner header is 5852 * 802.1Q or 802.1AD and vlan_do_receive() above could 5853 * not find vlan dev for vlan id 0. 5854 */ 5855 __vlan_hwaccel_clear_tag(skb); 5856 skb = skb_vlan_untag(skb); 5857 if (unlikely(!skb)) 5858 goto out; 5859 if (vlan_do_receive(&skb)) 5860 /* After stripping off 802.1P header with vlan 0 5861 * vlan dev is found for inner header. 5862 */ 5863 goto another_round; 5864 else if (unlikely(!skb)) 5865 goto out; 5866 else 5867 /* We have stripped outer 802.1P vlan 0 header. 5868 * But could not find vlan dev. 5869 * check again for vlan id to set OTHERHOST. 5870 */ 5871 goto check_vlan_id; 5872 } 5873 /* Note: we might in the future use prio bits 5874 * and set skb->priority like in vlan_do_receive() 5875 * For the time being, just ignore Priority Code Point 5876 */ 5877 __vlan_hwaccel_clear_tag(skb); 5878 } 5879 5880 type = skb->protocol; 5881 5882 /* deliver only exact match when indicated */ 5883 if (likely(!deliver_exact)) { 5884 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5885 &ptype_base[ntohs(type) & 5886 PTYPE_HASH_MASK]); 5887 5888 /* orig_dev and skb->dev could belong to different netns; 5889 * Even in such case we need to traverse only the list 5890 * coming from skb->dev, as the ptype owner (packet socket) 5891 * will use dev_net(skb->dev) to do namespace filtering. 5892 */ 5893 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5894 &dev_net_rcu(skb->dev)->ptype_specific); 5895 } 5896 5897 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5898 &orig_dev->ptype_specific); 5899 5900 if (unlikely(skb->dev != orig_dev)) { 5901 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5902 &skb->dev->ptype_specific); 5903 } 5904 5905 if (pt_prev) { 5906 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 5907 goto drop; 5908 *ppt_prev = pt_prev; 5909 } else { 5910 drop: 5911 if (!deliver_exact) 5912 dev_core_stats_rx_dropped_inc(skb->dev); 5913 else 5914 dev_core_stats_rx_nohandler_inc(skb->dev); 5915 kfree_skb_reason(skb, SKB_DROP_REASON_UNHANDLED_PROTO); 5916 /* Jamal, now you will not able to escape explaining 5917 * me how you were going to use this. :-) 5918 */ 5919 ret = NET_RX_DROP; 5920 } 5921 5922 out: 5923 /* The invariant here is that if *ppt_prev is not NULL 5924 * then skb should also be non-NULL. 5925 * 5926 * Apparently *ppt_prev assignment above holds this invariant due to 5927 * skb dereferencing near it. 5928 */ 5929 *pskb = skb; 5930 return ret; 5931 } 5932 5933 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc) 5934 { 5935 struct net_device *orig_dev = skb->dev; 5936 struct packet_type *pt_prev = NULL; 5937 int ret; 5938 5939 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5940 if (pt_prev) 5941 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb, 5942 skb->dev, pt_prev, orig_dev); 5943 return ret; 5944 } 5945 5946 /** 5947 * netif_receive_skb_core - special purpose version of netif_receive_skb 5948 * @skb: buffer to process 5949 * 5950 * More direct receive version of netif_receive_skb(). It should 5951 * only be used by callers that have a need to skip RPS and Generic XDP. 5952 * Caller must also take care of handling if ``(page_is_)pfmemalloc``. 5953 * 5954 * This function may only be called from softirq context and interrupts 5955 * should be enabled. 5956 * 5957 * Return values (usually ignored): 5958 * NET_RX_SUCCESS: no congestion 5959 * NET_RX_DROP: packet was dropped 5960 */ 5961 int netif_receive_skb_core(struct sk_buff *skb) 5962 { 5963 int ret; 5964 5965 rcu_read_lock(); 5966 ret = __netif_receive_skb_one_core(skb, false); 5967 rcu_read_unlock(); 5968 5969 return ret; 5970 } 5971 EXPORT_SYMBOL(netif_receive_skb_core); 5972 5973 static inline void __netif_receive_skb_list_ptype(struct list_head *head, 5974 struct packet_type *pt_prev, 5975 struct net_device *orig_dev) 5976 { 5977 struct sk_buff *skb, *next; 5978 5979 if (!pt_prev) 5980 return; 5981 if (list_empty(head)) 5982 return; 5983 if (pt_prev->list_func != NULL) 5984 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv, 5985 ip_list_rcv, head, pt_prev, orig_dev); 5986 else 5987 list_for_each_entry_safe(skb, next, head, list) { 5988 skb_list_del_init(skb); 5989 pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 5990 } 5991 } 5992 5993 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc) 5994 { 5995 /* Fast-path assumptions: 5996 * - There is no RX handler. 5997 * - Only one packet_type matches. 5998 * If either of these fails, we will end up doing some per-packet 5999 * processing in-line, then handling the 'last ptype' for the whole 6000 * sublist. This can't cause out-of-order delivery to any single ptype, 6001 * because the 'last ptype' must be constant across the sublist, and all 6002 * other ptypes are handled per-packet. 6003 */ 6004 /* Current (common) ptype of sublist */ 6005 struct packet_type *pt_curr = NULL; 6006 /* Current (common) orig_dev of sublist */ 6007 struct net_device *od_curr = NULL; 6008 struct sk_buff *skb, *next; 6009 LIST_HEAD(sublist); 6010 6011 list_for_each_entry_safe(skb, next, head, list) { 6012 struct net_device *orig_dev = skb->dev; 6013 struct packet_type *pt_prev = NULL; 6014 6015 skb_list_del_init(skb); 6016 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 6017 if (!pt_prev) 6018 continue; 6019 if (pt_curr != pt_prev || od_curr != orig_dev) { 6020 /* dispatch old sublist */ 6021 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 6022 /* start new sublist */ 6023 INIT_LIST_HEAD(&sublist); 6024 pt_curr = pt_prev; 6025 od_curr = orig_dev; 6026 } 6027 list_add_tail(&skb->list, &sublist); 6028 } 6029 6030 /* dispatch final sublist */ 6031 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 6032 } 6033 6034 static int __netif_receive_skb(struct sk_buff *skb) 6035 { 6036 int ret; 6037 6038 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) { 6039 unsigned int noreclaim_flag; 6040 6041 /* 6042 * PFMEMALLOC skbs are special, they should 6043 * - be delivered to SOCK_MEMALLOC sockets only 6044 * - stay away from userspace 6045 * - have bounded memory usage 6046 * 6047 * Use PF_MEMALLOC as this saves us from propagating the allocation 6048 * context down to all allocation sites. 6049 */ 6050 noreclaim_flag = memalloc_noreclaim_save(); 6051 ret = __netif_receive_skb_one_core(skb, true); 6052 memalloc_noreclaim_restore(noreclaim_flag); 6053 } else 6054 ret = __netif_receive_skb_one_core(skb, false); 6055 6056 return ret; 6057 } 6058 6059 static void __netif_receive_skb_list(struct list_head *head) 6060 { 6061 unsigned long noreclaim_flag = 0; 6062 struct sk_buff *skb, *next; 6063 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */ 6064 6065 list_for_each_entry_safe(skb, next, head, list) { 6066 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) { 6067 struct list_head sublist; 6068 6069 /* Handle the previous sublist */ 6070 list_cut_before(&sublist, head, &skb->list); 6071 if (!list_empty(&sublist)) 6072 __netif_receive_skb_list_core(&sublist, pfmemalloc); 6073 pfmemalloc = !pfmemalloc; 6074 /* See comments in __netif_receive_skb */ 6075 if (pfmemalloc) 6076 noreclaim_flag = memalloc_noreclaim_save(); 6077 else 6078 memalloc_noreclaim_restore(noreclaim_flag); 6079 } 6080 } 6081 /* Handle the remaining sublist */ 6082 if (!list_empty(head)) 6083 __netif_receive_skb_list_core(head, pfmemalloc); 6084 /* Restore pflags */ 6085 if (pfmemalloc) 6086 memalloc_noreclaim_restore(noreclaim_flag); 6087 } 6088 6089 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp) 6090 { 6091 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog); 6092 struct bpf_prog *new = xdp->prog; 6093 int ret = 0; 6094 6095 switch (xdp->command) { 6096 case XDP_SETUP_PROG: 6097 rcu_assign_pointer(dev->xdp_prog, new); 6098 if (old) 6099 bpf_prog_put(old); 6100 6101 if (old && !new) { 6102 static_branch_dec(&generic_xdp_needed_key); 6103 } else if (new && !old) { 6104 static_branch_inc(&generic_xdp_needed_key); 6105 netif_disable_lro(dev); 6106 dev_disable_gro_hw(dev); 6107 } 6108 break; 6109 6110 default: 6111 ret = -EINVAL; 6112 break; 6113 } 6114 6115 return ret; 6116 } 6117 6118 static int netif_receive_skb_internal(struct sk_buff *skb) 6119 { 6120 int ret; 6121 6122 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 6123 6124 if (skb_defer_rx_timestamp(skb)) 6125 return NET_RX_SUCCESS; 6126 6127 rcu_read_lock(); 6128 #ifdef CONFIG_RPS 6129 if (static_branch_unlikely(&rps_needed)) { 6130 struct rps_dev_flow voidflow, *rflow = &voidflow; 6131 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6132 6133 if (cpu >= 0) { 6134 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6135 rcu_read_unlock(); 6136 return ret; 6137 } 6138 } 6139 #endif 6140 ret = __netif_receive_skb(skb); 6141 rcu_read_unlock(); 6142 return ret; 6143 } 6144 6145 void netif_receive_skb_list_internal(struct list_head *head) 6146 { 6147 struct sk_buff *skb, *next; 6148 LIST_HEAD(sublist); 6149 6150 list_for_each_entry_safe(skb, next, head, list) { 6151 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), 6152 skb); 6153 skb_list_del_init(skb); 6154 if (!skb_defer_rx_timestamp(skb)) 6155 list_add_tail(&skb->list, &sublist); 6156 } 6157 list_splice_init(&sublist, head); 6158 6159 rcu_read_lock(); 6160 #ifdef CONFIG_RPS 6161 if (static_branch_unlikely(&rps_needed)) { 6162 list_for_each_entry_safe(skb, next, head, list) { 6163 struct rps_dev_flow voidflow, *rflow = &voidflow; 6164 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6165 6166 if (cpu >= 0) { 6167 /* Will be handled, remove from list */ 6168 skb_list_del_init(skb); 6169 enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6170 } 6171 } 6172 } 6173 #endif 6174 __netif_receive_skb_list(head); 6175 rcu_read_unlock(); 6176 } 6177 6178 /** 6179 * netif_receive_skb - process receive buffer from network 6180 * @skb: buffer to process 6181 * 6182 * netif_receive_skb() is the main receive data processing function. 6183 * It always succeeds. The buffer may be dropped during processing 6184 * for congestion control or by the protocol layers. 6185 * 6186 * This function may only be called from softirq context and interrupts 6187 * should be enabled. 6188 * 6189 * Return values (usually ignored): 6190 * NET_RX_SUCCESS: no congestion 6191 * NET_RX_DROP: packet was dropped 6192 */ 6193 int netif_receive_skb(struct sk_buff *skb) 6194 { 6195 int ret; 6196 6197 trace_netif_receive_skb_entry(skb); 6198 6199 ret = netif_receive_skb_internal(skb); 6200 trace_netif_receive_skb_exit(ret); 6201 6202 return ret; 6203 } 6204 EXPORT_SYMBOL(netif_receive_skb); 6205 6206 /** 6207 * netif_receive_skb_list - process many receive buffers from network 6208 * @head: list of skbs to process. 6209 * 6210 * Since return value of netif_receive_skb() is normally ignored, and 6211 * wouldn't be meaningful for a list, this function returns void. 6212 * 6213 * This function may only be called from softirq context and interrupts 6214 * should be enabled. 6215 */ 6216 void netif_receive_skb_list(struct list_head *head) 6217 { 6218 struct sk_buff *skb; 6219 6220 if (list_empty(head)) 6221 return; 6222 if (trace_netif_receive_skb_list_entry_enabled()) { 6223 list_for_each_entry(skb, head, list) 6224 trace_netif_receive_skb_list_entry(skb); 6225 } 6226 netif_receive_skb_list_internal(head); 6227 trace_netif_receive_skb_list_exit(0); 6228 } 6229 EXPORT_SYMBOL(netif_receive_skb_list); 6230 6231 /* Network device is going away, flush any packets still pending */ 6232 static void flush_backlog(struct work_struct *work) 6233 { 6234 struct sk_buff *skb, *tmp; 6235 struct sk_buff_head list; 6236 struct softnet_data *sd; 6237 6238 __skb_queue_head_init(&list); 6239 local_bh_disable(); 6240 sd = this_cpu_ptr(&softnet_data); 6241 6242 backlog_lock_irq_disable(sd); 6243 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) { 6244 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6245 __skb_unlink(skb, &sd->input_pkt_queue); 6246 __skb_queue_tail(&list, skb); 6247 rps_input_queue_head_incr(sd); 6248 } 6249 } 6250 backlog_unlock_irq_enable(sd); 6251 6252 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6253 skb_queue_walk_safe(&sd->process_queue, skb, tmp) { 6254 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6255 __skb_unlink(skb, &sd->process_queue); 6256 __skb_queue_tail(&list, skb); 6257 rps_input_queue_head_incr(sd); 6258 } 6259 } 6260 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6261 local_bh_enable(); 6262 6263 __skb_queue_purge_reason(&list, SKB_DROP_REASON_DEV_READY); 6264 } 6265 6266 static bool flush_required(int cpu) 6267 { 6268 #if IS_ENABLED(CONFIG_RPS) 6269 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 6270 bool do_flush; 6271 6272 backlog_lock_irq_disable(sd); 6273 6274 /* as insertion into process_queue happens with the rps lock held, 6275 * process_queue access may race only with dequeue 6276 */ 6277 do_flush = !skb_queue_empty(&sd->input_pkt_queue) || 6278 !skb_queue_empty_lockless(&sd->process_queue); 6279 backlog_unlock_irq_enable(sd); 6280 6281 return do_flush; 6282 #endif 6283 /* without RPS we can't safely check input_pkt_queue: during a 6284 * concurrent remote skb_queue_splice() we can detect as empty both 6285 * input_pkt_queue and process_queue even if the latter could end-up 6286 * containing a lot of packets. 6287 */ 6288 return true; 6289 } 6290 6291 struct flush_backlogs { 6292 cpumask_t flush_cpus; 6293 struct work_struct w[]; 6294 }; 6295 6296 static struct flush_backlogs *flush_backlogs_alloc(void) 6297 { 6298 return kmalloc(struct_size_t(struct flush_backlogs, w, nr_cpu_ids), 6299 GFP_KERNEL); 6300 } 6301 6302 static struct flush_backlogs *flush_backlogs_fallback; 6303 static DEFINE_MUTEX(flush_backlogs_mutex); 6304 6305 static void flush_all_backlogs(void) 6306 { 6307 struct flush_backlogs *ptr = flush_backlogs_alloc(); 6308 unsigned int cpu; 6309 6310 if (!ptr) { 6311 mutex_lock(&flush_backlogs_mutex); 6312 ptr = flush_backlogs_fallback; 6313 } 6314 cpumask_clear(&ptr->flush_cpus); 6315 6316 cpus_read_lock(); 6317 6318 for_each_online_cpu(cpu) { 6319 if (flush_required(cpu)) { 6320 INIT_WORK(&ptr->w[cpu], flush_backlog); 6321 queue_work_on(cpu, system_highpri_wq, &ptr->w[cpu]); 6322 __cpumask_set_cpu(cpu, &ptr->flush_cpus); 6323 } 6324 } 6325 6326 /* we can have in flight packet[s] on the cpus we are not flushing, 6327 * synchronize_net() in unregister_netdevice_many() will take care of 6328 * them. 6329 */ 6330 for_each_cpu(cpu, &ptr->flush_cpus) 6331 flush_work(&ptr->w[cpu]); 6332 6333 cpus_read_unlock(); 6334 6335 if (ptr != flush_backlogs_fallback) 6336 kfree(ptr); 6337 else 6338 mutex_unlock(&flush_backlogs_mutex); 6339 } 6340 6341 static void net_rps_send_ipi(struct softnet_data *remsd) 6342 { 6343 #ifdef CONFIG_RPS 6344 while (remsd) { 6345 struct softnet_data *next = remsd->rps_ipi_next; 6346 6347 if (cpu_online(remsd->cpu)) 6348 smp_call_function_single_async(remsd->cpu, &remsd->csd); 6349 remsd = next; 6350 } 6351 #endif 6352 } 6353 6354 /* 6355 * net_rps_action_and_irq_enable sends any pending IPI's for rps. 6356 * Note: called with local irq disabled, but exits with local irq enabled. 6357 */ 6358 static void net_rps_action_and_irq_enable(struct softnet_data *sd) 6359 { 6360 #ifdef CONFIG_RPS 6361 struct softnet_data *remsd = sd->rps_ipi_list; 6362 6363 if (!use_backlog_threads() && remsd) { 6364 sd->rps_ipi_list = NULL; 6365 6366 local_irq_enable(); 6367 6368 /* Send pending IPI's to kick RPS processing on remote cpus. */ 6369 net_rps_send_ipi(remsd); 6370 } else 6371 #endif 6372 local_irq_enable(); 6373 } 6374 6375 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd) 6376 { 6377 #ifdef CONFIG_RPS 6378 return !use_backlog_threads() && sd->rps_ipi_list; 6379 #else 6380 return false; 6381 #endif 6382 } 6383 6384 static int process_backlog(struct napi_struct *napi, int quota) 6385 { 6386 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog); 6387 bool again = true; 6388 int work = 0; 6389 6390 /* Check if we have pending ipi, its better to send them now, 6391 * not waiting net_rx_action() end. 6392 */ 6393 if (sd_has_rps_ipi_waiting(sd)) { 6394 local_irq_disable(); 6395 net_rps_action_and_irq_enable(sd); 6396 } 6397 6398 napi->weight = READ_ONCE(net_hotdata.dev_rx_weight); 6399 while (again) { 6400 struct sk_buff *skb; 6401 6402 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6403 while ((skb = __skb_dequeue(&sd->process_queue))) { 6404 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6405 rcu_read_lock(); 6406 __netif_receive_skb(skb); 6407 rcu_read_unlock(); 6408 if (++work >= quota) { 6409 rps_input_queue_head_add(sd, work); 6410 return work; 6411 } 6412 6413 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6414 } 6415 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6416 6417 backlog_lock_irq_disable(sd); 6418 if (skb_queue_empty(&sd->input_pkt_queue)) { 6419 /* 6420 * Inline a custom version of __napi_complete(). 6421 * only current cpu owns and manipulates this napi, 6422 * and NAPI_STATE_SCHED is the only possible flag set 6423 * on backlog. 6424 * We can use a plain write instead of clear_bit(), 6425 * and we dont need an smp_mb() memory barrier. 6426 */ 6427 napi->state &= NAPIF_STATE_THREADED; 6428 again = false; 6429 } else { 6430 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6431 skb_queue_splice_tail_init(&sd->input_pkt_queue, 6432 &sd->process_queue); 6433 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6434 } 6435 backlog_unlock_irq_enable(sd); 6436 } 6437 6438 if (work) 6439 rps_input_queue_head_add(sd, work); 6440 return work; 6441 } 6442 6443 /** 6444 * __napi_schedule - schedule for receive 6445 * @n: entry to schedule 6446 * 6447 * The entry's receive function will be scheduled to run. 6448 * Consider using __napi_schedule_irqoff() if hard irqs are masked. 6449 */ 6450 void __napi_schedule(struct napi_struct *n) 6451 { 6452 unsigned long flags; 6453 6454 local_irq_save(flags); 6455 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6456 local_irq_restore(flags); 6457 } 6458 EXPORT_SYMBOL(__napi_schedule); 6459 6460 /** 6461 * napi_schedule_prep - check if napi can be scheduled 6462 * @n: napi context 6463 * 6464 * Test if NAPI routine is already running, and if not mark 6465 * it as running. This is used as a condition variable to 6466 * insure only one NAPI poll instance runs. We also make 6467 * sure there is no pending NAPI disable. 6468 */ 6469 bool napi_schedule_prep(struct napi_struct *n) 6470 { 6471 unsigned long new, val = READ_ONCE(n->state); 6472 6473 do { 6474 if (unlikely(val & NAPIF_STATE_DISABLE)) 6475 return false; 6476 new = val | NAPIF_STATE_SCHED; 6477 6478 /* Sets STATE_MISSED bit if STATE_SCHED was already set 6479 * This was suggested by Alexander Duyck, as compiler 6480 * emits better code than : 6481 * if (val & NAPIF_STATE_SCHED) 6482 * new |= NAPIF_STATE_MISSED; 6483 */ 6484 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED * 6485 NAPIF_STATE_MISSED; 6486 } while (!try_cmpxchg(&n->state, &val, new)); 6487 6488 return !(val & NAPIF_STATE_SCHED); 6489 } 6490 EXPORT_SYMBOL(napi_schedule_prep); 6491 6492 /** 6493 * __napi_schedule_irqoff - schedule for receive 6494 * @n: entry to schedule 6495 * 6496 * Variant of __napi_schedule() assuming hard irqs are masked. 6497 * 6498 * On PREEMPT_RT enabled kernels this maps to __napi_schedule() 6499 * because the interrupt disabled assumption might not be true 6500 * due to force-threaded interrupts and spinlock substitution. 6501 */ 6502 void __napi_schedule_irqoff(struct napi_struct *n) 6503 { 6504 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6505 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6506 else 6507 __napi_schedule(n); 6508 } 6509 EXPORT_SYMBOL(__napi_schedule_irqoff); 6510 6511 bool napi_complete_done(struct napi_struct *n, int work_done) 6512 { 6513 unsigned long flags, val, new, timeout = 0; 6514 bool ret = true; 6515 6516 /* 6517 * 1) Don't let napi dequeue from the cpu poll list 6518 * just in case its running on a different cpu. 6519 * 2) If we are busy polling, do nothing here, we have 6520 * the guarantee we will be called later. 6521 */ 6522 if (unlikely(n->state & (NAPIF_STATE_NPSVC | 6523 NAPIF_STATE_IN_BUSY_POLL))) 6524 return false; 6525 6526 if (work_done) { 6527 if (n->gro.bitmask) 6528 timeout = napi_get_gro_flush_timeout(n); 6529 n->defer_hard_irqs_count = napi_get_defer_hard_irqs(n); 6530 } 6531 if (n->defer_hard_irqs_count > 0) { 6532 n->defer_hard_irqs_count--; 6533 timeout = napi_get_gro_flush_timeout(n); 6534 if (timeout) 6535 ret = false; 6536 } 6537 6538 /* 6539 * When the NAPI instance uses a timeout and keeps postponing 6540 * it, we need to bound somehow the time packets are kept in 6541 * the GRO layer. 6542 */ 6543 gro_flush(&n->gro, !!timeout); 6544 gro_normal_list(&n->gro); 6545 6546 if (unlikely(!list_empty(&n->poll_list))) { 6547 /* If n->poll_list is not empty, we need to mask irqs */ 6548 local_irq_save(flags); 6549 list_del_init(&n->poll_list); 6550 local_irq_restore(flags); 6551 } 6552 WRITE_ONCE(n->list_owner, -1); 6553 6554 val = READ_ONCE(n->state); 6555 do { 6556 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED)); 6557 6558 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED | 6559 NAPIF_STATE_SCHED_THREADED | 6560 NAPIF_STATE_PREFER_BUSY_POLL); 6561 6562 /* If STATE_MISSED was set, leave STATE_SCHED set, 6563 * because we will call napi->poll() one more time. 6564 * This C code was suggested by Alexander Duyck to help gcc. 6565 */ 6566 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED * 6567 NAPIF_STATE_SCHED; 6568 } while (!try_cmpxchg(&n->state, &val, new)); 6569 6570 if (unlikely(val & NAPIF_STATE_MISSED)) { 6571 __napi_schedule(n); 6572 return false; 6573 } 6574 6575 if (timeout) 6576 hrtimer_start(&n->timer, ns_to_ktime(timeout), 6577 HRTIMER_MODE_REL_PINNED); 6578 return ret; 6579 } 6580 EXPORT_SYMBOL(napi_complete_done); 6581 6582 static void skb_defer_free_flush(struct softnet_data *sd) 6583 { 6584 struct sk_buff *skb, *next; 6585 6586 /* Paired with WRITE_ONCE() in skb_attempt_defer_free() */ 6587 if (!READ_ONCE(sd->defer_list)) 6588 return; 6589 6590 spin_lock(&sd->defer_lock); 6591 skb = sd->defer_list; 6592 sd->defer_list = NULL; 6593 sd->defer_count = 0; 6594 spin_unlock(&sd->defer_lock); 6595 6596 while (skb != NULL) { 6597 next = skb->next; 6598 napi_consume_skb(skb, 1); 6599 skb = next; 6600 } 6601 } 6602 6603 #if defined(CONFIG_NET_RX_BUSY_POLL) 6604 6605 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule) 6606 { 6607 if (!skip_schedule) { 6608 gro_normal_list(&napi->gro); 6609 __napi_schedule(napi); 6610 return; 6611 } 6612 6613 /* Flush too old packets. If HZ < 1000, flush all packets */ 6614 gro_flush(&napi->gro, HZ >= 1000); 6615 gro_normal_list(&napi->gro); 6616 6617 clear_bit(NAPI_STATE_SCHED, &napi->state); 6618 } 6619 6620 enum { 6621 NAPI_F_PREFER_BUSY_POLL = 1, 6622 NAPI_F_END_ON_RESCHED = 2, 6623 }; 6624 6625 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, 6626 unsigned flags, u16 budget) 6627 { 6628 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6629 bool skip_schedule = false; 6630 unsigned long timeout; 6631 int rc; 6632 6633 /* Busy polling means there is a high chance device driver hard irq 6634 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was 6635 * set in napi_schedule_prep(). 6636 * Since we are about to call napi->poll() once more, we can safely 6637 * clear NAPI_STATE_MISSED. 6638 * 6639 * Note: x86 could use a single "lock and ..." instruction 6640 * to perform these two clear_bit() 6641 */ 6642 clear_bit(NAPI_STATE_MISSED, &napi->state); 6643 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state); 6644 6645 local_bh_disable(); 6646 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6647 6648 if (flags & NAPI_F_PREFER_BUSY_POLL) { 6649 napi->defer_hard_irqs_count = napi_get_defer_hard_irqs(napi); 6650 timeout = napi_get_gro_flush_timeout(napi); 6651 if (napi->defer_hard_irqs_count && timeout) { 6652 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED); 6653 skip_schedule = true; 6654 } 6655 } 6656 6657 /* All we really want here is to re-enable device interrupts. 6658 * Ideally, a new ndo_busy_poll_stop() could avoid another round. 6659 */ 6660 rc = napi->poll(napi, budget); 6661 /* We can't gro_normal_list() here, because napi->poll() might have 6662 * rearmed the napi (napi_complete_done()) in which case it could 6663 * already be running on another CPU. 6664 */ 6665 trace_napi_poll(napi, rc, budget); 6666 netpoll_poll_unlock(have_poll_lock); 6667 if (rc == budget) 6668 __busy_poll_stop(napi, skip_schedule); 6669 bpf_net_ctx_clear(bpf_net_ctx); 6670 local_bh_enable(); 6671 } 6672 6673 static void __napi_busy_loop(unsigned int napi_id, 6674 bool (*loop_end)(void *, unsigned long), 6675 void *loop_end_arg, unsigned flags, u16 budget) 6676 { 6677 unsigned long start_time = loop_end ? busy_loop_current_time() : 0; 6678 int (*napi_poll)(struct napi_struct *napi, int budget); 6679 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6680 void *have_poll_lock = NULL; 6681 struct napi_struct *napi; 6682 6683 WARN_ON_ONCE(!rcu_read_lock_held()); 6684 6685 restart: 6686 napi_poll = NULL; 6687 6688 napi = napi_by_id(napi_id); 6689 if (!napi) 6690 return; 6691 6692 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6693 preempt_disable(); 6694 for (;;) { 6695 int work = 0; 6696 6697 local_bh_disable(); 6698 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6699 if (!napi_poll) { 6700 unsigned long val = READ_ONCE(napi->state); 6701 6702 /* If multiple threads are competing for this napi, 6703 * we avoid dirtying napi->state as much as we can. 6704 */ 6705 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED | 6706 NAPIF_STATE_IN_BUSY_POLL)) { 6707 if (flags & NAPI_F_PREFER_BUSY_POLL) 6708 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6709 goto count; 6710 } 6711 if (cmpxchg(&napi->state, val, 6712 val | NAPIF_STATE_IN_BUSY_POLL | 6713 NAPIF_STATE_SCHED) != val) { 6714 if (flags & NAPI_F_PREFER_BUSY_POLL) 6715 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6716 goto count; 6717 } 6718 have_poll_lock = netpoll_poll_lock(napi); 6719 napi_poll = napi->poll; 6720 } 6721 work = napi_poll(napi, budget); 6722 trace_napi_poll(napi, work, budget); 6723 gro_normal_list(&napi->gro); 6724 count: 6725 if (work > 0) 6726 __NET_ADD_STATS(dev_net(napi->dev), 6727 LINUX_MIB_BUSYPOLLRXPACKETS, work); 6728 skb_defer_free_flush(this_cpu_ptr(&softnet_data)); 6729 bpf_net_ctx_clear(bpf_net_ctx); 6730 local_bh_enable(); 6731 6732 if (!loop_end || loop_end(loop_end_arg, start_time)) 6733 break; 6734 6735 if (unlikely(need_resched())) { 6736 if (flags & NAPI_F_END_ON_RESCHED) 6737 break; 6738 if (napi_poll) 6739 busy_poll_stop(napi, have_poll_lock, flags, budget); 6740 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6741 preempt_enable(); 6742 rcu_read_unlock(); 6743 cond_resched(); 6744 rcu_read_lock(); 6745 if (loop_end(loop_end_arg, start_time)) 6746 return; 6747 goto restart; 6748 } 6749 cpu_relax(); 6750 } 6751 if (napi_poll) 6752 busy_poll_stop(napi, have_poll_lock, flags, budget); 6753 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6754 preempt_enable(); 6755 } 6756 6757 void napi_busy_loop_rcu(unsigned int napi_id, 6758 bool (*loop_end)(void *, unsigned long), 6759 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6760 { 6761 unsigned flags = NAPI_F_END_ON_RESCHED; 6762 6763 if (prefer_busy_poll) 6764 flags |= NAPI_F_PREFER_BUSY_POLL; 6765 6766 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6767 } 6768 6769 void napi_busy_loop(unsigned int napi_id, 6770 bool (*loop_end)(void *, unsigned long), 6771 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6772 { 6773 unsigned flags = prefer_busy_poll ? NAPI_F_PREFER_BUSY_POLL : 0; 6774 6775 rcu_read_lock(); 6776 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6777 rcu_read_unlock(); 6778 } 6779 EXPORT_SYMBOL(napi_busy_loop); 6780 6781 void napi_suspend_irqs(unsigned int napi_id) 6782 { 6783 struct napi_struct *napi; 6784 6785 rcu_read_lock(); 6786 napi = napi_by_id(napi_id); 6787 if (napi) { 6788 unsigned long timeout = napi_get_irq_suspend_timeout(napi); 6789 6790 if (timeout) 6791 hrtimer_start(&napi->timer, ns_to_ktime(timeout), 6792 HRTIMER_MODE_REL_PINNED); 6793 } 6794 rcu_read_unlock(); 6795 } 6796 6797 void napi_resume_irqs(unsigned int napi_id) 6798 { 6799 struct napi_struct *napi; 6800 6801 rcu_read_lock(); 6802 napi = napi_by_id(napi_id); 6803 if (napi) { 6804 /* If irq_suspend_timeout is set to 0 between the call to 6805 * napi_suspend_irqs and now, the original value still 6806 * determines the safety timeout as intended and napi_watchdog 6807 * will resume irq processing. 6808 */ 6809 if (napi_get_irq_suspend_timeout(napi)) { 6810 local_bh_disable(); 6811 napi_schedule(napi); 6812 local_bh_enable(); 6813 } 6814 } 6815 rcu_read_unlock(); 6816 } 6817 6818 #endif /* CONFIG_NET_RX_BUSY_POLL */ 6819 6820 static void __napi_hash_add_with_id(struct napi_struct *napi, 6821 unsigned int napi_id) 6822 { 6823 napi->gro.cached_napi_id = napi_id; 6824 6825 WRITE_ONCE(napi->napi_id, napi_id); 6826 hlist_add_head_rcu(&napi->napi_hash_node, 6827 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]); 6828 } 6829 6830 static void napi_hash_add_with_id(struct napi_struct *napi, 6831 unsigned int napi_id) 6832 { 6833 unsigned long flags; 6834 6835 spin_lock_irqsave(&napi_hash_lock, flags); 6836 WARN_ON_ONCE(napi_by_id(napi_id)); 6837 __napi_hash_add_with_id(napi, napi_id); 6838 spin_unlock_irqrestore(&napi_hash_lock, flags); 6839 } 6840 6841 static void napi_hash_add(struct napi_struct *napi) 6842 { 6843 unsigned long flags; 6844 6845 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state)) 6846 return; 6847 6848 spin_lock_irqsave(&napi_hash_lock, flags); 6849 6850 /* 0..NR_CPUS range is reserved for sender_cpu use */ 6851 do { 6852 if (unlikely(!napi_id_valid(++napi_gen_id))) 6853 napi_gen_id = MIN_NAPI_ID; 6854 } while (napi_by_id(napi_gen_id)); 6855 6856 __napi_hash_add_with_id(napi, napi_gen_id); 6857 6858 spin_unlock_irqrestore(&napi_hash_lock, flags); 6859 } 6860 6861 /* Warning : caller is responsible to make sure rcu grace period 6862 * is respected before freeing memory containing @napi 6863 */ 6864 static void napi_hash_del(struct napi_struct *napi) 6865 { 6866 unsigned long flags; 6867 6868 spin_lock_irqsave(&napi_hash_lock, flags); 6869 6870 hlist_del_init_rcu(&napi->napi_hash_node); 6871 6872 spin_unlock_irqrestore(&napi_hash_lock, flags); 6873 } 6874 6875 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer) 6876 { 6877 struct napi_struct *napi; 6878 6879 napi = container_of(timer, struct napi_struct, timer); 6880 6881 /* Note : we use a relaxed variant of napi_schedule_prep() not setting 6882 * NAPI_STATE_MISSED, since we do not react to a device IRQ. 6883 */ 6884 if (!napi_disable_pending(napi) && 6885 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) { 6886 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6887 __napi_schedule_irqoff(napi); 6888 } 6889 6890 return HRTIMER_NORESTART; 6891 } 6892 6893 int dev_set_threaded(struct net_device *dev, bool threaded) 6894 { 6895 struct napi_struct *napi; 6896 int err = 0; 6897 6898 netdev_assert_locked_or_invisible(dev); 6899 6900 if (dev->threaded == threaded) 6901 return 0; 6902 6903 if (threaded) { 6904 list_for_each_entry(napi, &dev->napi_list, dev_list) { 6905 if (!napi->thread) { 6906 err = napi_kthread_create(napi); 6907 if (err) { 6908 threaded = false; 6909 break; 6910 } 6911 } 6912 } 6913 } 6914 6915 WRITE_ONCE(dev->threaded, threaded); 6916 6917 /* Make sure kthread is created before THREADED bit 6918 * is set. 6919 */ 6920 smp_mb__before_atomic(); 6921 6922 /* Setting/unsetting threaded mode on a napi might not immediately 6923 * take effect, if the current napi instance is actively being 6924 * polled. In this case, the switch between threaded mode and 6925 * softirq mode will happen in the next round of napi_schedule(). 6926 * This should not cause hiccups/stalls to the live traffic. 6927 */ 6928 list_for_each_entry(napi, &dev->napi_list, dev_list) 6929 assign_bit(NAPI_STATE_THREADED, &napi->state, threaded); 6930 6931 return err; 6932 } 6933 EXPORT_SYMBOL(dev_set_threaded); 6934 6935 /** 6936 * netif_queue_set_napi - Associate queue with the napi 6937 * @dev: device to which NAPI and queue belong 6938 * @queue_index: Index of queue 6939 * @type: queue type as RX or TX 6940 * @napi: NAPI context, pass NULL to clear previously set NAPI 6941 * 6942 * Set queue with its corresponding napi context. This should be done after 6943 * registering the NAPI handler for the queue-vector and the queues have been 6944 * mapped to the corresponding interrupt vector. 6945 */ 6946 void netif_queue_set_napi(struct net_device *dev, unsigned int queue_index, 6947 enum netdev_queue_type type, struct napi_struct *napi) 6948 { 6949 struct netdev_rx_queue *rxq; 6950 struct netdev_queue *txq; 6951 6952 if (WARN_ON_ONCE(napi && !napi->dev)) 6953 return; 6954 netdev_ops_assert_locked_or_invisible(dev); 6955 6956 switch (type) { 6957 case NETDEV_QUEUE_TYPE_RX: 6958 rxq = __netif_get_rx_queue(dev, queue_index); 6959 rxq->napi = napi; 6960 return; 6961 case NETDEV_QUEUE_TYPE_TX: 6962 txq = netdev_get_tx_queue(dev, queue_index); 6963 txq->napi = napi; 6964 return; 6965 default: 6966 return; 6967 } 6968 } 6969 EXPORT_SYMBOL(netif_queue_set_napi); 6970 6971 static void 6972 netif_napi_irq_notify(struct irq_affinity_notify *notify, 6973 const cpumask_t *mask) 6974 { 6975 struct napi_struct *napi = 6976 container_of(notify, struct napi_struct, notify); 6977 #ifdef CONFIG_RFS_ACCEL 6978 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 6979 int err; 6980 #endif 6981 6982 if (napi->config && napi->dev->irq_affinity_auto) 6983 cpumask_copy(&napi->config->affinity_mask, mask); 6984 6985 #ifdef CONFIG_RFS_ACCEL 6986 if (napi->dev->rx_cpu_rmap_auto) { 6987 err = cpu_rmap_update(rmap, napi->napi_rmap_idx, mask); 6988 if (err) 6989 netdev_warn(napi->dev, "RMAP update failed (%d)\n", 6990 err); 6991 } 6992 #endif 6993 } 6994 6995 #ifdef CONFIG_RFS_ACCEL 6996 static void netif_napi_affinity_release(struct kref *ref) 6997 { 6998 struct napi_struct *napi = 6999 container_of(ref, struct napi_struct, notify.kref); 7000 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 7001 7002 netdev_assert_locked(napi->dev); 7003 WARN_ON(test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, 7004 &napi->state)); 7005 7006 if (!napi->dev->rx_cpu_rmap_auto) 7007 return; 7008 rmap->obj[napi->napi_rmap_idx] = NULL; 7009 napi->napi_rmap_idx = -1; 7010 cpu_rmap_put(rmap); 7011 } 7012 7013 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 7014 { 7015 if (dev->rx_cpu_rmap_auto) 7016 return 0; 7017 7018 dev->rx_cpu_rmap = alloc_irq_cpu_rmap(num_irqs); 7019 if (!dev->rx_cpu_rmap) 7020 return -ENOMEM; 7021 7022 dev->rx_cpu_rmap_auto = true; 7023 return 0; 7024 } 7025 EXPORT_SYMBOL(netif_enable_cpu_rmap); 7026 7027 static void netif_del_cpu_rmap(struct net_device *dev) 7028 { 7029 struct cpu_rmap *rmap = dev->rx_cpu_rmap; 7030 7031 if (!dev->rx_cpu_rmap_auto) 7032 return; 7033 7034 /* Free the rmap */ 7035 cpu_rmap_put(rmap); 7036 dev->rx_cpu_rmap = NULL; 7037 dev->rx_cpu_rmap_auto = false; 7038 } 7039 7040 #else 7041 static void netif_napi_affinity_release(struct kref *ref) 7042 { 7043 } 7044 7045 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 7046 { 7047 return 0; 7048 } 7049 EXPORT_SYMBOL(netif_enable_cpu_rmap); 7050 7051 static void netif_del_cpu_rmap(struct net_device *dev) 7052 { 7053 } 7054 #endif 7055 7056 void netif_set_affinity_auto(struct net_device *dev) 7057 { 7058 unsigned int i, maxqs, numa; 7059 7060 maxqs = max(dev->num_tx_queues, dev->num_rx_queues); 7061 numa = dev_to_node(&dev->dev); 7062 7063 for (i = 0; i < maxqs; i++) 7064 cpumask_set_cpu(cpumask_local_spread(i, numa), 7065 &dev->napi_config[i].affinity_mask); 7066 7067 dev->irq_affinity_auto = true; 7068 } 7069 EXPORT_SYMBOL(netif_set_affinity_auto); 7070 7071 void netif_napi_set_irq_locked(struct napi_struct *napi, int irq) 7072 { 7073 int rc; 7074 7075 netdev_assert_locked_or_invisible(napi->dev); 7076 7077 if (napi->irq == irq) 7078 return; 7079 7080 /* Remove existing resources */ 7081 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7082 irq_set_affinity_notifier(napi->irq, NULL); 7083 7084 napi->irq = irq; 7085 if (irq < 0 || 7086 (!napi->dev->rx_cpu_rmap_auto && !napi->dev->irq_affinity_auto)) 7087 return; 7088 7089 /* Abort for buggy drivers */ 7090 if (napi->dev->irq_affinity_auto && WARN_ON_ONCE(!napi->config)) 7091 return; 7092 7093 #ifdef CONFIG_RFS_ACCEL 7094 if (napi->dev->rx_cpu_rmap_auto) { 7095 rc = cpu_rmap_add(napi->dev->rx_cpu_rmap, napi); 7096 if (rc < 0) 7097 return; 7098 7099 cpu_rmap_get(napi->dev->rx_cpu_rmap); 7100 napi->napi_rmap_idx = rc; 7101 } 7102 #endif 7103 7104 /* Use core IRQ notifier */ 7105 napi->notify.notify = netif_napi_irq_notify; 7106 napi->notify.release = netif_napi_affinity_release; 7107 rc = irq_set_affinity_notifier(irq, &napi->notify); 7108 if (rc) { 7109 netdev_warn(napi->dev, "Unable to set IRQ notifier (%d)\n", 7110 rc); 7111 goto put_rmap; 7112 } 7113 7114 set_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state); 7115 return; 7116 7117 put_rmap: 7118 #ifdef CONFIG_RFS_ACCEL 7119 if (napi->dev->rx_cpu_rmap_auto) { 7120 napi->dev->rx_cpu_rmap->obj[napi->napi_rmap_idx] = NULL; 7121 cpu_rmap_put(napi->dev->rx_cpu_rmap); 7122 napi->napi_rmap_idx = -1; 7123 } 7124 #endif 7125 napi->notify.notify = NULL; 7126 napi->notify.release = NULL; 7127 } 7128 EXPORT_SYMBOL(netif_napi_set_irq_locked); 7129 7130 static void napi_restore_config(struct napi_struct *n) 7131 { 7132 n->defer_hard_irqs = n->config->defer_hard_irqs; 7133 n->gro_flush_timeout = n->config->gro_flush_timeout; 7134 n->irq_suspend_timeout = n->config->irq_suspend_timeout; 7135 7136 if (n->dev->irq_affinity_auto && 7137 test_bit(NAPI_STATE_HAS_NOTIFIER, &n->state)) 7138 irq_set_affinity(n->irq, &n->config->affinity_mask); 7139 7140 /* a NAPI ID might be stored in the config, if so use it. if not, use 7141 * napi_hash_add to generate one for us. 7142 */ 7143 if (n->config->napi_id) { 7144 napi_hash_add_with_id(n, n->config->napi_id); 7145 } else { 7146 napi_hash_add(n); 7147 n->config->napi_id = n->napi_id; 7148 } 7149 } 7150 7151 static void napi_save_config(struct napi_struct *n) 7152 { 7153 n->config->defer_hard_irqs = n->defer_hard_irqs; 7154 n->config->gro_flush_timeout = n->gro_flush_timeout; 7155 n->config->irq_suspend_timeout = n->irq_suspend_timeout; 7156 napi_hash_del(n); 7157 } 7158 7159 /* Netlink wants the NAPI list to be sorted by ID, if adding a NAPI which will 7160 * inherit an existing ID try to insert it at the right position. 7161 */ 7162 static void 7163 netif_napi_dev_list_add(struct net_device *dev, struct napi_struct *napi) 7164 { 7165 unsigned int new_id, pos_id; 7166 struct list_head *higher; 7167 struct napi_struct *pos; 7168 7169 new_id = UINT_MAX; 7170 if (napi->config && napi->config->napi_id) 7171 new_id = napi->config->napi_id; 7172 7173 higher = &dev->napi_list; 7174 list_for_each_entry(pos, &dev->napi_list, dev_list) { 7175 if (napi_id_valid(pos->napi_id)) 7176 pos_id = pos->napi_id; 7177 else if (pos->config) 7178 pos_id = pos->config->napi_id; 7179 else 7180 pos_id = UINT_MAX; 7181 7182 if (pos_id <= new_id) 7183 break; 7184 higher = &pos->dev_list; 7185 } 7186 list_add_rcu(&napi->dev_list, higher); /* adds after higher */ 7187 } 7188 7189 /* Double check that napi_get_frags() allocates skbs with 7190 * skb->head being backed by slab, not a page fragment. 7191 * This is to make sure bug fixed in 3226b158e67c 7192 * ("net: avoid 32 x truesize under-estimation for tiny skbs") 7193 * does not accidentally come back. 7194 */ 7195 static void napi_get_frags_check(struct napi_struct *napi) 7196 { 7197 struct sk_buff *skb; 7198 7199 local_bh_disable(); 7200 skb = napi_get_frags(napi); 7201 WARN_ON_ONCE(skb && skb->head_frag); 7202 napi_free_frags(napi); 7203 local_bh_enable(); 7204 } 7205 7206 void netif_napi_add_weight_locked(struct net_device *dev, 7207 struct napi_struct *napi, 7208 int (*poll)(struct napi_struct *, int), 7209 int weight) 7210 { 7211 netdev_assert_locked(dev); 7212 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state))) 7213 return; 7214 7215 INIT_LIST_HEAD(&napi->poll_list); 7216 INIT_HLIST_NODE(&napi->napi_hash_node); 7217 hrtimer_setup(&napi->timer, napi_watchdog, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); 7218 gro_init(&napi->gro); 7219 napi->skb = NULL; 7220 napi->poll = poll; 7221 if (weight > NAPI_POLL_WEIGHT) 7222 netdev_err_once(dev, "%s() called with weight %d\n", __func__, 7223 weight); 7224 napi->weight = weight; 7225 napi->dev = dev; 7226 #ifdef CONFIG_NETPOLL 7227 napi->poll_owner = -1; 7228 #endif 7229 napi->list_owner = -1; 7230 set_bit(NAPI_STATE_SCHED, &napi->state); 7231 set_bit(NAPI_STATE_NPSVC, &napi->state); 7232 netif_napi_dev_list_add(dev, napi); 7233 7234 /* default settings from sysfs are applied to all NAPIs. any per-NAPI 7235 * configuration will be loaded in napi_enable 7236 */ 7237 napi_set_defer_hard_irqs(napi, READ_ONCE(dev->napi_defer_hard_irqs)); 7238 napi_set_gro_flush_timeout(napi, READ_ONCE(dev->gro_flush_timeout)); 7239 7240 napi_get_frags_check(napi); 7241 /* Create kthread for this napi if dev->threaded is set. 7242 * Clear dev->threaded if kthread creation failed so that 7243 * threaded mode will not be enabled in napi_enable(). 7244 */ 7245 if (dev->threaded && napi_kthread_create(napi)) 7246 dev->threaded = false; 7247 netif_napi_set_irq_locked(napi, -1); 7248 } 7249 EXPORT_SYMBOL(netif_napi_add_weight_locked); 7250 7251 void napi_disable_locked(struct napi_struct *n) 7252 { 7253 unsigned long val, new; 7254 7255 might_sleep(); 7256 netdev_assert_locked(n->dev); 7257 7258 set_bit(NAPI_STATE_DISABLE, &n->state); 7259 7260 val = READ_ONCE(n->state); 7261 do { 7262 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) { 7263 usleep_range(20, 200); 7264 val = READ_ONCE(n->state); 7265 } 7266 7267 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC; 7268 new &= ~(NAPIF_STATE_THREADED | NAPIF_STATE_PREFER_BUSY_POLL); 7269 } while (!try_cmpxchg(&n->state, &val, new)); 7270 7271 hrtimer_cancel(&n->timer); 7272 7273 if (n->config) 7274 napi_save_config(n); 7275 else 7276 napi_hash_del(n); 7277 7278 clear_bit(NAPI_STATE_DISABLE, &n->state); 7279 } 7280 EXPORT_SYMBOL(napi_disable_locked); 7281 7282 /** 7283 * napi_disable() - prevent NAPI from scheduling 7284 * @n: NAPI context 7285 * 7286 * Stop NAPI from being scheduled on this context. 7287 * Waits till any outstanding processing completes. 7288 * Takes netdev_lock() for associated net_device. 7289 */ 7290 void napi_disable(struct napi_struct *n) 7291 { 7292 netdev_lock(n->dev); 7293 napi_disable_locked(n); 7294 netdev_unlock(n->dev); 7295 } 7296 EXPORT_SYMBOL(napi_disable); 7297 7298 void napi_enable_locked(struct napi_struct *n) 7299 { 7300 unsigned long new, val = READ_ONCE(n->state); 7301 7302 if (n->config) 7303 napi_restore_config(n); 7304 else 7305 napi_hash_add(n); 7306 7307 do { 7308 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val)); 7309 7310 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC); 7311 if (n->dev->threaded && n->thread) 7312 new |= NAPIF_STATE_THREADED; 7313 } while (!try_cmpxchg(&n->state, &val, new)); 7314 } 7315 EXPORT_SYMBOL(napi_enable_locked); 7316 7317 /** 7318 * napi_enable() - enable NAPI scheduling 7319 * @n: NAPI context 7320 * 7321 * Enable scheduling of a NAPI instance. 7322 * Must be paired with napi_disable(). 7323 * Takes netdev_lock() for associated net_device. 7324 */ 7325 void napi_enable(struct napi_struct *n) 7326 { 7327 netdev_lock(n->dev); 7328 napi_enable_locked(n); 7329 netdev_unlock(n->dev); 7330 } 7331 EXPORT_SYMBOL(napi_enable); 7332 7333 /* Must be called in process context */ 7334 void __netif_napi_del_locked(struct napi_struct *napi) 7335 { 7336 netdev_assert_locked(napi->dev); 7337 7338 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state)) 7339 return; 7340 7341 /* Make sure NAPI is disabled (or was never enabled). */ 7342 WARN_ON(!test_bit(NAPI_STATE_SCHED, &napi->state)); 7343 7344 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7345 irq_set_affinity_notifier(napi->irq, NULL); 7346 7347 if (napi->config) { 7348 napi->index = -1; 7349 napi->config = NULL; 7350 } 7351 7352 list_del_rcu(&napi->dev_list); 7353 napi_free_frags(napi); 7354 7355 gro_cleanup(&napi->gro); 7356 7357 if (napi->thread) { 7358 kthread_stop(napi->thread); 7359 napi->thread = NULL; 7360 } 7361 } 7362 EXPORT_SYMBOL(__netif_napi_del_locked); 7363 7364 static int __napi_poll(struct napi_struct *n, bool *repoll) 7365 { 7366 int work, weight; 7367 7368 weight = n->weight; 7369 7370 /* This NAPI_STATE_SCHED test is for avoiding a race 7371 * with netpoll's poll_napi(). Only the entity which 7372 * obtains the lock and sees NAPI_STATE_SCHED set will 7373 * actually make the ->poll() call. Therefore we avoid 7374 * accidentally calling ->poll() when NAPI is not scheduled. 7375 */ 7376 work = 0; 7377 if (napi_is_scheduled(n)) { 7378 work = n->poll(n, weight); 7379 trace_napi_poll(n, work, weight); 7380 7381 xdp_do_check_flushed(n); 7382 } 7383 7384 if (unlikely(work > weight)) 7385 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n", 7386 n->poll, work, weight); 7387 7388 if (likely(work < weight)) 7389 return work; 7390 7391 /* Drivers must not modify the NAPI state if they 7392 * consume the entire weight. In such cases this code 7393 * still "owns" the NAPI instance and therefore can 7394 * move the instance around on the list at-will. 7395 */ 7396 if (unlikely(napi_disable_pending(n))) { 7397 napi_complete(n); 7398 return work; 7399 } 7400 7401 /* The NAPI context has more processing work, but busy-polling 7402 * is preferred. Exit early. 7403 */ 7404 if (napi_prefer_busy_poll(n)) { 7405 if (napi_complete_done(n, work)) { 7406 /* If timeout is not set, we need to make sure 7407 * that the NAPI is re-scheduled. 7408 */ 7409 napi_schedule(n); 7410 } 7411 return work; 7412 } 7413 7414 /* Flush too old packets. If HZ < 1000, flush all packets */ 7415 gro_flush(&n->gro, HZ >= 1000); 7416 gro_normal_list(&n->gro); 7417 7418 /* Some drivers may have called napi_schedule 7419 * prior to exhausting their budget. 7420 */ 7421 if (unlikely(!list_empty(&n->poll_list))) { 7422 pr_warn_once("%s: Budget exhausted after napi rescheduled\n", 7423 n->dev ? n->dev->name : "backlog"); 7424 return work; 7425 } 7426 7427 *repoll = true; 7428 7429 return work; 7430 } 7431 7432 static int napi_poll(struct napi_struct *n, struct list_head *repoll) 7433 { 7434 bool do_repoll = false; 7435 void *have; 7436 int work; 7437 7438 list_del_init(&n->poll_list); 7439 7440 have = netpoll_poll_lock(n); 7441 7442 work = __napi_poll(n, &do_repoll); 7443 7444 if (do_repoll) 7445 list_add_tail(&n->poll_list, repoll); 7446 7447 netpoll_poll_unlock(have); 7448 7449 return work; 7450 } 7451 7452 static int napi_thread_wait(struct napi_struct *napi) 7453 { 7454 set_current_state(TASK_INTERRUPTIBLE); 7455 7456 while (!kthread_should_stop()) { 7457 /* Testing SCHED_THREADED bit here to make sure the current 7458 * kthread owns this napi and could poll on this napi. 7459 * Testing SCHED bit is not enough because SCHED bit might be 7460 * set by some other busy poll thread or by napi_disable(). 7461 */ 7462 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) { 7463 WARN_ON(!list_empty(&napi->poll_list)); 7464 __set_current_state(TASK_RUNNING); 7465 return 0; 7466 } 7467 7468 schedule(); 7469 set_current_state(TASK_INTERRUPTIBLE); 7470 } 7471 __set_current_state(TASK_RUNNING); 7472 7473 return -1; 7474 } 7475 7476 static void napi_threaded_poll_loop(struct napi_struct *napi) 7477 { 7478 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7479 struct softnet_data *sd; 7480 unsigned long last_qs = jiffies; 7481 7482 for (;;) { 7483 bool repoll = false; 7484 void *have; 7485 7486 local_bh_disable(); 7487 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7488 7489 sd = this_cpu_ptr(&softnet_data); 7490 sd->in_napi_threaded_poll = true; 7491 7492 have = netpoll_poll_lock(napi); 7493 __napi_poll(napi, &repoll); 7494 netpoll_poll_unlock(have); 7495 7496 sd->in_napi_threaded_poll = false; 7497 barrier(); 7498 7499 if (sd_has_rps_ipi_waiting(sd)) { 7500 local_irq_disable(); 7501 net_rps_action_and_irq_enable(sd); 7502 } 7503 skb_defer_free_flush(sd); 7504 bpf_net_ctx_clear(bpf_net_ctx); 7505 local_bh_enable(); 7506 7507 if (!repoll) 7508 break; 7509 7510 rcu_softirq_qs_periodic(last_qs); 7511 cond_resched(); 7512 } 7513 } 7514 7515 static int napi_threaded_poll(void *data) 7516 { 7517 struct napi_struct *napi = data; 7518 7519 while (!napi_thread_wait(napi)) 7520 napi_threaded_poll_loop(napi); 7521 7522 return 0; 7523 } 7524 7525 static __latent_entropy void net_rx_action(void) 7526 { 7527 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 7528 unsigned long time_limit = jiffies + 7529 usecs_to_jiffies(READ_ONCE(net_hotdata.netdev_budget_usecs)); 7530 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7531 int budget = READ_ONCE(net_hotdata.netdev_budget); 7532 LIST_HEAD(list); 7533 LIST_HEAD(repoll); 7534 7535 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7536 start: 7537 sd->in_net_rx_action = true; 7538 local_irq_disable(); 7539 list_splice_init(&sd->poll_list, &list); 7540 local_irq_enable(); 7541 7542 for (;;) { 7543 struct napi_struct *n; 7544 7545 skb_defer_free_flush(sd); 7546 7547 if (list_empty(&list)) { 7548 if (list_empty(&repoll)) { 7549 sd->in_net_rx_action = false; 7550 barrier(); 7551 /* We need to check if ____napi_schedule() 7552 * had refilled poll_list while 7553 * sd->in_net_rx_action was true. 7554 */ 7555 if (!list_empty(&sd->poll_list)) 7556 goto start; 7557 if (!sd_has_rps_ipi_waiting(sd)) 7558 goto end; 7559 } 7560 break; 7561 } 7562 7563 n = list_first_entry(&list, struct napi_struct, poll_list); 7564 budget -= napi_poll(n, &repoll); 7565 7566 /* If softirq window is exhausted then punt. 7567 * Allow this to run for 2 jiffies since which will allow 7568 * an average latency of 1.5/HZ. 7569 */ 7570 if (unlikely(budget <= 0 || 7571 time_after_eq(jiffies, time_limit))) { 7572 /* Pairs with READ_ONCE() in softnet_seq_show() */ 7573 WRITE_ONCE(sd->time_squeeze, sd->time_squeeze + 1); 7574 break; 7575 } 7576 } 7577 7578 local_irq_disable(); 7579 7580 list_splice_tail_init(&sd->poll_list, &list); 7581 list_splice_tail(&repoll, &list); 7582 list_splice(&list, &sd->poll_list); 7583 if (!list_empty(&sd->poll_list)) 7584 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 7585 else 7586 sd->in_net_rx_action = false; 7587 7588 net_rps_action_and_irq_enable(sd); 7589 end: 7590 bpf_net_ctx_clear(bpf_net_ctx); 7591 } 7592 7593 struct netdev_adjacent { 7594 struct net_device *dev; 7595 netdevice_tracker dev_tracker; 7596 7597 /* upper master flag, there can only be one master device per list */ 7598 bool master; 7599 7600 /* lookup ignore flag */ 7601 bool ignore; 7602 7603 /* counter for the number of times this device was added to us */ 7604 u16 ref_nr; 7605 7606 /* private field for the users */ 7607 void *private; 7608 7609 struct list_head list; 7610 struct rcu_head rcu; 7611 }; 7612 7613 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev, 7614 struct list_head *adj_list) 7615 { 7616 struct netdev_adjacent *adj; 7617 7618 list_for_each_entry(adj, adj_list, list) { 7619 if (adj->dev == adj_dev) 7620 return adj; 7621 } 7622 return NULL; 7623 } 7624 7625 static int ____netdev_has_upper_dev(struct net_device *upper_dev, 7626 struct netdev_nested_priv *priv) 7627 { 7628 struct net_device *dev = (struct net_device *)priv->data; 7629 7630 return upper_dev == dev; 7631 } 7632 7633 /** 7634 * netdev_has_upper_dev - Check if device is linked to an upper device 7635 * @dev: device 7636 * @upper_dev: upper device to check 7637 * 7638 * Find out if a device is linked to specified upper device and return true 7639 * in case it is. Note that this checks only immediate upper device, 7640 * not through a complete stack of devices. The caller must hold the RTNL lock. 7641 */ 7642 bool netdev_has_upper_dev(struct net_device *dev, 7643 struct net_device *upper_dev) 7644 { 7645 struct netdev_nested_priv priv = { 7646 .data = (void *)upper_dev, 7647 }; 7648 7649 ASSERT_RTNL(); 7650 7651 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7652 &priv); 7653 } 7654 EXPORT_SYMBOL(netdev_has_upper_dev); 7655 7656 /** 7657 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device 7658 * @dev: device 7659 * @upper_dev: upper device to check 7660 * 7661 * Find out if a device is linked to specified upper device and return true 7662 * in case it is. Note that this checks the entire upper device chain. 7663 * The caller must hold rcu lock. 7664 */ 7665 7666 bool netdev_has_upper_dev_all_rcu(struct net_device *dev, 7667 struct net_device *upper_dev) 7668 { 7669 struct netdev_nested_priv priv = { 7670 .data = (void *)upper_dev, 7671 }; 7672 7673 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7674 &priv); 7675 } 7676 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu); 7677 7678 /** 7679 * netdev_has_any_upper_dev - Check if device is linked to some device 7680 * @dev: device 7681 * 7682 * Find out if a device is linked to an upper device and return true in case 7683 * it is. The caller must hold the RTNL lock. 7684 */ 7685 bool netdev_has_any_upper_dev(struct net_device *dev) 7686 { 7687 ASSERT_RTNL(); 7688 7689 return !list_empty(&dev->adj_list.upper); 7690 } 7691 EXPORT_SYMBOL(netdev_has_any_upper_dev); 7692 7693 /** 7694 * netdev_master_upper_dev_get - Get master upper device 7695 * @dev: device 7696 * 7697 * Find a master upper device and return pointer to it or NULL in case 7698 * it's not there. The caller must hold the RTNL lock. 7699 */ 7700 struct net_device *netdev_master_upper_dev_get(struct net_device *dev) 7701 { 7702 struct netdev_adjacent *upper; 7703 7704 ASSERT_RTNL(); 7705 7706 if (list_empty(&dev->adj_list.upper)) 7707 return NULL; 7708 7709 upper = list_first_entry(&dev->adj_list.upper, 7710 struct netdev_adjacent, list); 7711 if (likely(upper->master)) 7712 return upper->dev; 7713 return NULL; 7714 } 7715 EXPORT_SYMBOL(netdev_master_upper_dev_get); 7716 7717 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev) 7718 { 7719 struct netdev_adjacent *upper; 7720 7721 ASSERT_RTNL(); 7722 7723 if (list_empty(&dev->adj_list.upper)) 7724 return NULL; 7725 7726 upper = list_first_entry(&dev->adj_list.upper, 7727 struct netdev_adjacent, list); 7728 if (likely(upper->master) && !upper->ignore) 7729 return upper->dev; 7730 return NULL; 7731 } 7732 7733 /** 7734 * netdev_has_any_lower_dev - Check if device is linked to some device 7735 * @dev: device 7736 * 7737 * Find out if a device is linked to a lower device and return true in case 7738 * it is. The caller must hold the RTNL lock. 7739 */ 7740 static bool netdev_has_any_lower_dev(struct net_device *dev) 7741 { 7742 ASSERT_RTNL(); 7743 7744 return !list_empty(&dev->adj_list.lower); 7745 } 7746 7747 void *netdev_adjacent_get_private(struct list_head *adj_list) 7748 { 7749 struct netdev_adjacent *adj; 7750 7751 adj = list_entry(adj_list, struct netdev_adjacent, list); 7752 7753 return adj->private; 7754 } 7755 EXPORT_SYMBOL(netdev_adjacent_get_private); 7756 7757 /** 7758 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list 7759 * @dev: device 7760 * @iter: list_head ** of the current position 7761 * 7762 * Gets the next device from the dev's upper list, starting from iter 7763 * position. The caller must hold RCU read lock. 7764 */ 7765 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev, 7766 struct list_head **iter) 7767 { 7768 struct netdev_adjacent *upper; 7769 7770 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7771 7772 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7773 7774 if (&upper->list == &dev->adj_list.upper) 7775 return NULL; 7776 7777 *iter = &upper->list; 7778 7779 return upper->dev; 7780 } 7781 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu); 7782 7783 static struct net_device *__netdev_next_upper_dev(struct net_device *dev, 7784 struct list_head **iter, 7785 bool *ignore) 7786 { 7787 struct netdev_adjacent *upper; 7788 7789 upper = list_entry((*iter)->next, struct netdev_adjacent, list); 7790 7791 if (&upper->list == &dev->adj_list.upper) 7792 return NULL; 7793 7794 *iter = &upper->list; 7795 *ignore = upper->ignore; 7796 7797 return upper->dev; 7798 } 7799 7800 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev, 7801 struct list_head **iter) 7802 { 7803 struct netdev_adjacent *upper; 7804 7805 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7806 7807 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7808 7809 if (&upper->list == &dev->adj_list.upper) 7810 return NULL; 7811 7812 *iter = &upper->list; 7813 7814 return upper->dev; 7815 } 7816 7817 static int __netdev_walk_all_upper_dev(struct net_device *dev, 7818 int (*fn)(struct net_device *dev, 7819 struct netdev_nested_priv *priv), 7820 struct netdev_nested_priv *priv) 7821 { 7822 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7823 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7824 int ret, cur = 0; 7825 bool ignore; 7826 7827 now = dev; 7828 iter = &dev->adj_list.upper; 7829 7830 while (1) { 7831 if (now != dev) { 7832 ret = fn(now, priv); 7833 if (ret) 7834 return ret; 7835 } 7836 7837 next = NULL; 7838 while (1) { 7839 udev = __netdev_next_upper_dev(now, &iter, &ignore); 7840 if (!udev) 7841 break; 7842 if (ignore) 7843 continue; 7844 7845 next = udev; 7846 niter = &udev->adj_list.upper; 7847 dev_stack[cur] = now; 7848 iter_stack[cur++] = iter; 7849 break; 7850 } 7851 7852 if (!next) { 7853 if (!cur) 7854 return 0; 7855 next = dev_stack[--cur]; 7856 niter = iter_stack[cur]; 7857 } 7858 7859 now = next; 7860 iter = niter; 7861 } 7862 7863 return 0; 7864 } 7865 7866 int netdev_walk_all_upper_dev_rcu(struct net_device *dev, 7867 int (*fn)(struct net_device *dev, 7868 struct netdev_nested_priv *priv), 7869 struct netdev_nested_priv *priv) 7870 { 7871 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7872 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7873 int ret, cur = 0; 7874 7875 now = dev; 7876 iter = &dev->adj_list.upper; 7877 7878 while (1) { 7879 if (now != dev) { 7880 ret = fn(now, priv); 7881 if (ret) 7882 return ret; 7883 } 7884 7885 next = NULL; 7886 while (1) { 7887 udev = netdev_next_upper_dev_rcu(now, &iter); 7888 if (!udev) 7889 break; 7890 7891 next = udev; 7892 niter = &udev->adj_list.upper; 7893 dev_stack[cur] = now; 7894 iter_stack[cur++] = iter; 7895 break; 7896 } 7897 7898 if (!next) { 7899 if (!cur) 7900 return 0; 7901 next = dev_stack[--cur]; 7902 niter = iter_stack[cur]; 7903 } 7904 7905 now = next; 7906 iter = niter; 7907 } 7908 7909 return 0; 7910 } 7911 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu); 7912 7913 static bool __netdev_has_upper_dev(struct net_device *dev, 7914 struct net_device *upper_dev) 7915 { 7916 struct netdev_nested_priv priv = { 7917 .flags = 0, 7918 .data = (void *)upper_dev, 7919 }; 7920 7921 ASSERT_RTNL(); 7922 7923 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev, 7924 &priv); 7925 } 7926 7927 /** 7928 * netdev_lower_get_next_private - Get the next ->private from the 7929 * lower neighbour list 7930 * @dev: device 7931 * @iter: list_head ** of the current position 7932 * 7933 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7934 * list, starting from iter position. The caller must hold either hold the 7935 * RTNL lock or its own locking that guarantees that the neighbour lower 7936 * list will remain unchanged. 7937 */ 7938 void *netdev_lower_get_next_private(struct net_device *dev, 7939 struct list_head **iter) 7940 { 7941 struct netdev_adjacent *lower; 7942 7943 lower = list_entry(*iter, struct netdev_adjacent, list); 7944 7945 if (&lower->list == &dev->adj_list.lower) 7946 return NULL; 7947 7948 *iter = lower->list.next; 7949 7950 return lower->private; 7951 } 7952 EXPORT_SYMBOL(netdev_lower_get_next_private); 7953 7954 /** 7955 * netdev_lower_get_next_private_rcu - Get the next ->private from the 7956 * lower neighbour list, RCU 7957 * variant 7958 * @dev: device 7959 * @iter: list_head ** of the current position 7960 * 7961 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7962 * list, starting from iter position. The caller must hold RCU read lock. 7963 */ 7964 void *netdev_lower_get_next_private_rcu(struct net_device *dev, 7965 struct list_head **iter) 7966 { 7967 struct netdev_adjacent *lower; 7968 7969 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 7970 7971 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7972 7973 if (&lower->list == &dev->adj_list.lower) 7974 return NULL; 7975 7976 *iter = &lower->list; 7977 7978 return lower->private; 7979 } 7980 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu); 7981 7982 /** 7983 * netdev_lower_get_next - Get the next device from the lower neighbour 7984 * list 7985 * @dev: device 7986 * @iter: list_head ** of the current position 7987 * 7988 * Gets the next netdev_adjacent from the dev's lower neighbour 7989 * list, starting from iter position. The caller must hold RTNL lock or 7990 * its own locking that guarantees that the neighbour lower 7991 * list will remain unchanged. 7992 */ 7993 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter) 7994 { 7995 struct netdev_adjacent *lower; 7996 7997 lower = list_entry(*iter, struct netdev_adjacent, list); 7998 7999 if (&lower->list == &dev->adj_list.lower) 8000 return NULL; 8001 8002 *iter = lower->list.next; 8003 8004 return lower->dev; 8005 } 8006 EXPORT_SYMBOL(netdev_lower_get_next); 8007 8008 static struct net_device *netdev_next_lower_dev(struct net_device *dev, 8009 struct list_head **iter) 8010 { 8011 struct netdev_adjacent *lower; 8012 8013 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 8014 8015 if (&lower->list == &dev->adj_list.lower) 8016 return NULL; 8017 8018 *iter = &lower->list; 8019 8020 return lower->dev; 8021 } 8022 8023 static struct net_device *__netdev_next_lower_dev(struct net_device *dev, 8024 struct list_head **iter, 8025 bool *ignore) 8026 { 8027 struct netdev_adjacent *lower; 8028 8029 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 8030 8031 if (&lower->list == &dev->adj_list.lower) 8032 return NULL; 8033 8034 *iter = &lower->list; 8035 *ignore = lower->ignore; 8036 8037 return lower->dev; 8038 } 8039 8040 int netdev_walk_all_lower_dev(struct net_device *dev, 8041 int (*fn)(struct net_device *dev, 8042 struct netdev_nested_priv *priv), 8043 struct netdev_nested_priv *priv) 8044 { 8045 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8046 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8047 int ret, cur = 0; 8048 8049 now = dev; 8050 iter = &dev->adj_list.lower; 8051 8052 while (1) { 8053 if (now != dev) { 8054 ret = fn(now, priv); 8055 if (ret) 8056 return ret; 8057 } 8058 8059 next = NULL; 8060 while (1) { 8061 ldev = netdev_next_lower_dev(now, &iter); 8062 if (!ldev) 8063 break; 8064 8065 next = ldev; 8066 niter = &ldev->adj_list.lower; 8067 dev_stack[cur] = now; 8068 iter_stack[cur++] = iter; 8069 break; 8070 } 8071 8072 if (!next) { 8073 if (!cur) 8074 return 0; 8075 next = dev_stack[--cur]; 8076 niter = iter_stack[cur]; 8077 } 8078 8079 now = next; 8080 iter = niter; 8081 } 8082 8083 return 0; 8084 } 8085 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev); 8086 8087 static int __netdev_walk_all_lower_dev(struct net_device *dev, 8088 int (*fn)(struct net_device *dev, 8089 struct netdev_nested_priv *priv), 8090 struct netdev_nested_priv *priv) 8091 { 8092 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8093 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8094 int ret, cur = 0; 8095 bool ignore; 8096 8097 now = dev; 8098 iter = &dev->adj_list.lower; 8099 8100 while (1) { 8101 if (now != dev) { 8102 ret = fn(now, priv); 8103 if (ret) 8104 return ret; 8105 } 8106 8107 next = NULL; 8108 while (1) { 8109 ldev = __netdev_next_lower_dev(now, &iter, &ignore); 8110 if (!ldev) 8111 break; 8112 if (ignore) 8113 continue; 8114 8115 next = ldev; 8116 niter = &ldev->adj_list.lower; 8117 dev_stack[cur] = now; 8118 iter_stack[cur++] = iter; 8119 break; 8120 } 8121 8122 if (!next) { 8123 if (!cur) 8124 return 0; 8125 next = dev_stack[--cur]; 8126 niter = iter_stack[cur]; 8127 } 8128 8129 now = next; 8130 iter = niter; 8131 } 8132 8133 return 0; 8134 } 8135 8136 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev, 8137 struct list_head **iter) 8138 { 8139 struct netdev_adjacent *lower; 8140 8141 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8142 if (&lower->list == &dev->adj_list.lower) 8143 return NULL; 8144 8145 *iter = &lower->list; 8146 8147 return lower->dev; 8148 } 8149 EXPORT_SYMBOL(netdev_next_lower_dev_rcu); 8150 8151 static u8 __netdev_upper_depth(struct net_device *dev) 8152 { 8153 struct net_device *udev; 8154 struct list_head *iter; 8155 u8 max_depth = 0; 8156 bool ignore; 8157 8158 for (iter = &dev->adj_list.upper, 8159 udev = __netdev_next_upper_dev(dev, &iter, &ignore); 8160 udev; 8161 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) { 8162 if (ignore) 8163 continue; 8164 if (max_depth < udev->upper_level) 8165 max_depth = udev->upper_level; 8166 } 8167 8168 return max_depth; 8169 } 8170 8171 static u8 __netdev_lower_depth(struct net_device *dev) 8172 { 8173 struct net_device *ldev; 8174 struct list_head *iter; 8175 u8 max_depth = 0; 8176 bool ignore; 8177 8178 for (iter = &dev->adj_list.lower, 8179 ldev = __netdev_next_lower_dev(dev, &iter, &ignore); 8180 ldev; 8181 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) { 8182 if (ignore) 8183 continue; 8184 if (max_depth < ldev->lower_level) 8185 max_depth = ldev->lower_level; 8186 } 8187 8188 return max_depth; 8189 } 8190 8191 static int __netdev_update_upper_level(struct net_device *dev, 8192 struct netdev_nested_priv *__unused) 8193 { 8194 dev->upper_level = __netdev_upper_depth(dev) + 1; 8195 return 0; 8196 } 8197 8198 #ifdef CONFIG_LOCKDEP 8199 static LIST_HEAD(net_unlink_list); 8200 8201 static void net_unlink_todo(struct net_device *dev) 8202 { 8203 if (list_empty(&dev->unlink_list)) 8204 list_add_tail(&dev->unlink_list, &net_unlink_list); 8205 } 8206 #endif 8207 8208 static int __netdev_update_lower_level(struct net_device *dev, 8209 struct netdev_nested_priv *priv) 8210 { 8211 dev->lower_level = __netdev_lower_depth(dev) + 1; 8212 8213 #ifdef CONFIG_LOCKDEP 8214 if (!priv) 8215 return 0; 8216 8217 if (priv->flags & NESTED_SYNC_IMM) 8218 dev->nested_level = dev->lower_level - 1; 8219 if (priv->flags & NESTED_SYNC_TODO) 8220 net_unlink_todo(dev); 8221 #endif 8222 return 0; 8223 } 8224 8225 int netdev_walk_all_lower_dev_rcu(struct net_device *dev, 8226 int (*fn)(struct net_device *dev, 8227 struct netdev_nested_priv *priv), 8228 struct netdev_nested_priv *priv) 8229 { 8230 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8231 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8232 int ret, cur = 0; 8233 8234 now = dev; 8235 iter = &dev->adj_list.lower; 8236 8237 while (1) { 8238 if (now != dev) { 8239 ret = fn(now, priv); 8240 if (ret) 8241 return ret; 8242 } 8243 8244 next = NULL; 8245 while (1) { 8246 ldev = netdev_next_lower_dev_rcu(now, &iter); 8247 if (!ldev) 8248 break; 8249 8250 next = ldev; 8251 niter = &ldev->adj_list.lower; 8252 dev_stack[cur] = now; 8253 iter_stack[cur++] = iter; 8254 break; 8255 } 8256 8257 if (!next) { 8258 if (!cur) 8259 return 0; 8260 next = dev_stack[--cur]; 8261 niter = iter_stack[cur]; 8262 } 8263 8264 now = next; 8265 iter = niter; 8266 } 8267 8268 return 0; 8269 } 8270 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu); 8271 8272 /** 8273 * netdev_lower_get_first_private_rcu - Get the first ->private from the 8274 * lower neighbour list, RCU 8275 * variant 8276 * @dev: device 8277 * 8278 * Gets the first netdev_adjacent->private from the dev's lower neighbour 8279 * list. The caller must hold RCU read lock. 8280 */ 8281 void *netdev_lower_get_first_private_rcu(struct net_device *dev) 8282 { 8283 struct netdev_adjacent *lower; 8284 8285 lower = list_first_or_null_rcu(&dev->adj_list.lower, 8286 struct netdev_adjacent, list); 8287 if (lower) 8288 return lower->private; 8289 return NULL; 8290 } 8291 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu); 8292 8293 /** 8294 * netdev_master_upper_dev_get_rcu - Get master upper device 8295 * @dev: device 8296 * 8297 * Find a master upper device and return pointer to it or NULL in case 8298 * it's not there. The caller must hold the RCU read lock. 8299 */ 8300 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev) 8301 { 8302 struct netdev_adjacent *upper; 8303 8304 upper = list_first_or_null_rcu(&dev->adj_list.upper, 8305 struct netdev_adjacent, list); 8306 if (upper && likely(upper->master)) 8307 return upper->dev; 8308 return NULL; 8309 } 8310 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu); 8311 8312 static int netdev_adjacent_sysfs_add(struct net_device *dev, 8313 struct net_device *adj_dev, 8314 struct list_head *dev_list) 8315 { 8316 char linkname[IFNAMSIZ+7]; 8317 8318 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8319 "upper_%s" : "lower_%s", adj_dev->name); 8320 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj), 8321 linkname); 8322 } 8323 static void netdev_adjacent_sysfs_del(struct net_device *dev, 8324 char *name, 8325 struct list_head *dev_list) 8326 { 8327 char linkname[IFNAMSIZ+7]; 8328 8329 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8330 "upper_%s" : "lower_%s", name); 8331 sysfs_remove_link(&(dev->dev.kobj), linkname); 8332 } 8333 8334 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev, 8335 struct net_device *adj_dev, 8336 struct list_head *dev_list) 8337 { 8338 return (dev_list == &dev->adj_list.upper || 8339 dev_list == &dev->adj_list.lower) && 8340 net_eq(dev_net(dev), dev_net(adj_dev)); 8341 } 8342 8343 static int __netdev_adjacent_dev_insert(struct net_device *dev, 8344 struct net_device *adj_dev, 8345 struct list_head *dev_list, 8346 void *private, bool master) 8347 { 8348 struct netdev_adjacent *adj; 8349 int ret; 8350 8351 adj = __netdev_find_adj(adj_dev, dev_list); 8352 8353 if (adj) { 8354 adj->ref_nr += 1; 8355 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n", 8356 dev->name, adj_dev->name, adj->ref_nr); 8357 8358 return 0; 8359 } 8360 8361 adj = kmalloc(sizeof(*adj), GFP_KERNEL); 8362 if (!adj) 8363 return -ENOMEM; 8364 8365 adj->dev = adj_dev; 8366 adj->master = master; 8367 adj->ref_nr = 1; 8368 adj->private = private; 8369 adj->ignore = false; 8370 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL); 8371 8372 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n", 8373 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name); 8374 8375 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) { 8376 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list); 8377 if (ret) 8378 goto free_adj; 8379 } 8380 8381 /* Ensure that master link is always the first item in list. */ 8382 if (master) { 8383 ret = sysfs_create_link(&(dev->dev.kobj), 8384 &(adj_dev->dev.kobj), "master"); 8385 if (ret) 8386 goto remove_symlinks; 8387 8388 list_add_rcu(&adj->list, dev_list); 8389 } else { 8390 list_add_tail_rcu(&adj->list, dev_list); 8391 } 8392 8393 return 0; 8394 8395 remove_symlinks: 8396 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8397 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8398 free_adj: 8399 netdev_put(adj_dev, &adj->dev_tracker); 8400 kfree(adj); 8401 8402 return ret; 8403 } 8404 8405 static void __netdev_adjacent_dev_remove(struct net_device *dev, 8406 struct net_device *adj_dev, 8407 u16 ref_nr, 8408 struct list_head *dev_list) 8409 { 8410 struct netdev_adjacent *adj; 8411 8412 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n", 8413 dev->name, adj_dev->name, ref_nr); 8414 8415 adj = __netdev_find_adj(adj_dev, dev_list); 8416 8417 if (!adj) { 8418 pr_err("Adjacency does not exist for device %s from %s\n", 8419 dev->name, adj_dev->name); 8420 WARN_ON(1); 8421 return; 8422 } 8423 8424 if (adj->ref_nr > ref_nr) { 8425 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n", 8426 dev->name, adj_dev->name, ref_nr, 8427 adj->ref_nr - ref_nr); 8428 adj->ref_nr -= ref_nr; 8429 return; 8430 } 8431 8432 if (adj->master) 8433 sysfs_remove_link(&(dev->dev.kobj), "master"); 8434 8435 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8436 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8437 8438 list_del_rcu(&adj->list); 8439 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n", 8440 adj_dev->name, dev->name, adj_dev->name); 8441 netdev_put(adj_dev, &adj->dev_tracker); 8442 kfree_rcu(adj, rcu); 8443 } 8444 8445 static int __netdev_adjacent_dev_link_lists(struct net_device *dev, 8446 struct net_device *upper_dev, 8447 struct list_head *up_list, 8448 struct list_head *down_list, 8449 void *private, bool master) 8450 { 8451 int ret; 8452 8453 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list, 8454 private, master); 8455 if (ret) 8456 return ret; 8457 8458 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list, 8459 private, false); 8460 if (ret) { 8461 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list); 8462 return ret; 8463 } 8464 8465 return 0; 8466 } 8467 8468 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev, 8469 struct net_device *upper_dev, 8470 u16 ref_nr, 8471 struct list_head *up_list, 8472 struct list_head *down_list) 8473 { 8474 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list); 8475 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list); 8476 } 8477 8478 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev, 8479 struct net_device *upper_dev, 8480 void *private, bool master) 8481 { 8482 return __netdev_adjacent_dev_link_lists(dev, upper_dev, 8483 &dev->adj_list.upper, 8484 &upper_dev->adj_list.lower, 8485 private, master); 8486 } 8487 8488 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev, 8489 struct net_device *upper_dev) 8490 { 8491 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1, 8492 &dev->adj_list.upper, 8493 &upper_dev->adj_list.lower); 8494 } 8495 8496 static int __netdev_upper_dev_link(struct net_device *dev, 8497 struct net_device *upper_dev, bool master, 8498 void *upper_priv, void *upper_info, 8499 struct netdev_nested_priv *priv, 8500 struct netlink_ext_ack *extack) 8501 { 8502 struct netdev_notifier_changeupper_info changeupper_info = { 8503 .info = { 8504 .dev = dev, 8505 .extack = extack, 8506 }, 8507 .upper_dev = upper_dev, 8508 .master = master, 8509 .linking = true, 8510 .upper_info = upper_info, 8511 }; 8512 struct net_device *master_dev; 8513 int ret = 0; 8514 8515 ASSERT_RTNL(); 8516 8517 if (dev == upper_dev) 8518 return -EBUSY; 8519 8520 /* To prevent loops, check if dev is not upper device to upper_dev. */ 8521 if (__netdev_has_upper_dev(upper_dev, dev)) 8522 return -EBUSY; 8523 8524 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV) 8525 return -EMLINK; 8526 8527 if (!master) { 8528 if (__netdev_has_upper_dev(dev, upper_dev)) 8529 return -EEXIST; 8530 } else { 8531 master_dev = __netdev_master_upper_dev_get(dev); 8532 if (master_dev) 8533 return master_dev == upper_dev ? -EEXIST : -EBUSY; 8534 } 8535 8536 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8537 &changeupper_info.info); 8538 ret = notifier_to_errno(ret); 8539 if (ret) 8540 return ret; 8541 8542 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv, 8543 master); 8544 if (ret) 8545 return ret; 8546 8547 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8548 &changeupper_info.info); 8549 ret = notifier_to_errno(ret); 8550 if (ret) 8551 goto rollback; 8552 8553 __netdev_update_upper_level(dev, NULL); 8554 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8555 8556 __netdev_update_lower_level(upper_dev, priv); 8557 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8558 priv); 8559 8560 return 0; 8561 8562 rollback: 8563 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8564 8565 return ret; 8566 } 8567 8568 /** 8569 * netdev_upper_dev_link - Add a link to the upper device 8570 * @dev: device 8571 * @upper_dev: new upper device 8572 * @extack: netlink extended ack 8573 * 8574 * Adds a link to device which is upper to this one. The caller must hold 8575 * the RTNL lock. On a failure a negative errno code is returned. 8576 * On success the reference counts are adjusted and the function 8577 * returns zero. 8578 */ 8579 int netdev_upper_dev_link(struct net_device *dev, 8580 struct net_device *upper_dev, 8581 struct netlink_ext_ack *extack) 8582 { 8583 struct netdev_nested_priv priv = { 8584 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8585 .data = NULL, 8586 }; 8587 8588 return __netdev_upper_dev_link(dev, upper_dev, false, 8589 NULL, NULL, &priv, extack); 8590 } 8591 EXPORT_SYMBOL(netdev_upper_dev_link); 8592 8593 /** 8594 * netdev_master_upper_dev_link - Add a master link to the upper device 8595 * @dev: device 8596 * @upper_dev: new upper device 8597 * @upper_priv: upper device private 8598 * @upper_info: upper info to be passed down via notifier 8599 * @extack: netlink extended ack 8600 * 8601 * Adds a link to device which is upper to this one. In this case, only 8602 * one master upper device can be linked, although other non-master devices 8603 * might be linked as well. The caller must hold the RTNL lock. 8604 * On a failure a negative errno code is returned. On success the reference 8605 * counts are adjusted and the function returns zero. 8606 */ 8607 int netdev_master_upper_dev_link(struct net_device *dev, 8608 struct net_device *upper_dev, 8609 void *upper_priv, void *upper_info, 8610 struct netlink_ext_ack *extack) 8611 { 8612 struct netdev_nested_priv priv = { 8613 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8614 .data = NULL, 8615 }; 8616 8617 return __netdev_upper_dev_link(dev, upper_dev, true, 8618 upper_priv, upper_info, &priv, extack); 8619 } 8620 EXPORT_SYMBOL(netdev_master_upper_dev_link); 8621 8622 static void __netdev_upper_dev_unlink(struct net_device *dev, 8623 struct net_device *upper_dev, 8624 struct netdev_nested_priv *priv) 8625 { 8626 struct netdev_notifier_changeupper_info changeupper_info = { 8627 .info = { 8628 .dev = dev, 8629 }, 8630 .upper_dev = upper_dev, 8631 .linking = false, 8632 }; 8633 8634 ASSERT_RTNL(); 8635 8636 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev; 8637 8638 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8639 &changeupper_info.info); 8640 8641 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8642 8643 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8644 &changeupper_info.info); 8645 8646 __netdev_update_upper_level(dev, NULL); 8647 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8648 8649 __netdev_update_lower_level(upper_dev, priv); 8650 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8651 priv); 8652 } 8653 8654 /** 8655 * netdev_upper_dev_unlink - Removes a link to upper device 8656 * @dev: device 8657 * @upper_dev: new upper device 8658 * 8659 * Removes a link to device which is upper to this one. The caller must hold 8660 * the RTNL lock. 8661 */ 8662 void netdev_upper_dev_unlink(struct net_device *dev, 8663 struct net_device *upper_dev) 8664 { 8665 struct netdev_nested_priv priv = { 8666 .flags = NESTED_SYNC_TODO, 8667 .data = NULL, 8668 }; 8669 8670 __netdev_upper_dev_unlink(dev, upper_dev, &priv); 8671 } 8672 EXPORT_SYMBOL(netdev_upper_dev_unlink); 8673 8674 static void __netdev_adjacent_dev_set(struct net_device *upper_dev, 8675 struct net_device *lower_dev, 8676 bool val) 8677 { 8678 struct netdev_adjacent *adj; 8679 8680 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower); 8681 if (adj) 8682 adj->ignore = val; 8683 8684 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper); 8685 if (adj) 8686 adj->ignore = val; 8687 } 8688 8689 static void netdev_adjacent_dev_disable(struct net_device *upper_dev, 8690 struct net_device *lower_dev) 8691 { 8692 __netdev_adjacent_dev_set(upper_dev, lower_dev, true); 8693 } 8694 8695 static void netdev_adjacent_dev_enable(struct net_device *upper_dev, 8696 struct net_device *lower_dev) 8697 { 8698 __netdev_adjacent_dev_set(upper_dev, lower_dev, false); 8699 } 8700 8701 int netdev_adjacent_change_prepare(struct net_device *old_dev, 8702 struct net_device *new_dev, 8703 struct net_device *dev, 8704 struct netlink_ext_ack *extack) 8705 { 8706 struct netdev_nested_priv priv = { 8707 .flags = 0, 8708 .data = NULL, 8709 }; 8710 int err; 8711 8712 if (!new_dev) 8713 return 0; 8714 8715 if (old_dev && new_dev != old_dev) 8716 netdev_adjacent_dev_disable(dev, old_dev); 8717 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv, 8718 extack); 8719 if (err) { 8720 if (old_dev && new_dev != old_dev) 8721 netdev_adjacent_dev_enable(dev, old_dev); 8722 return err; 8723 } 8724 8725 return 0; 8726 } 8727 EXPORT_SYMBOL(netdev_adjacent_change_prepare); 8728 8729 void netdev_adjacent_change_commit(struct net_device *old_dev, 8730 struct net_device *new_dev, 8731 struct net_device *dev) 8732 { 8733 struct netdev_nested_priv priv = { 8734 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8735 .data = NULL, 8736 }; 8737 8738 if (!new_dev || !old_dev) 8739 return; 8740 8741 if (new_dev == old_dev) 8742 return; 8743 8744 netdev_adjacent_dev_enable(dev, old_dev); 8745 __netdev_upper_dev_unlink(old_dev, dev, &priv); 8746 } 8747 EXPORT_SYMBOL(netdev_adjacent_change_commit); 8748 8749 void netdev_adjacent_change_abort(struct net_device *old_dev, 8750 struct net_device *new_dev, 8751 struct net_device *dev) 8752 { 8753 struct netdev_nested_priv priv = { 8754 .flags = 0, 8755 .data = NULL, 8756 }; 8757 8758 if (!new_dev) 8759 return; 8760 8761 if (old_dev && new_dev != old_dev) 8762 netdev_adjacent_dev_enable(dev, old_dev); 8763 8764 __netdev_upper_dev_unlink(new_dev, dev, &priv); 8765 } 8766 EXPORT_SYMBOL(netdev_adjacent_change_abort); 8767 8768 /** 8769 * netdev_bonding_info_change - Dispatch event about slave change 8770 * @dev: device 8771 * @bonding_info: info to dispatch 8772 * 8773 * Send NETDEV_BONDING_INFO to netdev notifiers with info. 8774 * The caller must hold the RTNL lock. 8775 */ 8776 void netdev_bonding_info_change(struct net_device *dev, 8777 struct netdev_bonding_info *bonding_info) 8778 { 8779 struct netdev_notifier_bonding_info info = { 8780 .info.dev = dev, 8781 }; 8782 8783 memcpy(&info.bonding_info, bonding_info, 8784 sizeof(struct netdev_bonding_info)); 8785 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, 8786 &info.info); 8787 } 8788 EXPORT_SYMBOL(netdev_bonding_info_change); 8789 8790 static int netdev_offload_xstats_enable_l3(struct net_device *dev, 8791 struct netlink_ext_ack *extack) 8792 { 8793 struct netdev_notifier_offload_xstats_info info = { 8794 .info.dev = dev, 8795 .info.extack = extack, 8796 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8797 }; 8798 int err; 8799 int rc; 8800 8801 dev->offload_xstats_l3 = kzalloc(sizeof(*dev->offload_xstats_l3), 8802 GFP_KERNEL); 8803 if (!dev->offload_xstats_l3) 8804 return -ENOMEM; 8805 8806 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE, 8807 NETDEV_OFFLOAD_XSTATS_DISABLE, 8808 &info.info); 8809 err = notifier_to_errno(rc); 8810 if (err) 8811 goto free_stats; 8812 8813 return 0; 8814 8815 free_stats: 8816 kfree(dev->offload_xstats_l3); 8817 dev->offload_xstats_l3 = NULL; 8818 return err; 8819 } 8820 8821 int netdev_offload_xstats_enable(struct net_device *dev, 8822 enum netdev_offload_xstats_type type, 8823 struct netlink_ext_ack *extack) 8824 { 8825 ASSERT_RTNL(); 8826 8827 if (netdev_offload_xstats_enabled(dev, type)) 8828 return -EALREADY; 8829 8830 switch (type) { 8831 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8832 return netdev_offload_xstats_enable_l3(dev, extack); 8833 } 8834 8835 WARN_ON(1); 8836 return -EINVAL; 8837 } 8838 EXPORT_SYMBOL(netdev_offload_xstats_enable); 8839 8840 static void netdev_offload_xstats_disable_l3(struct net_device *dev) 8841 { 8842 struct netdev_notifier_offload_xstats_info info = { 8843 .info.dev = dev, 8844 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8845 }; 8846 8847 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE, 8848 &info.info); 8849 kfree(dev->offload_xstats_l3); 8850 dev->offload_xstats_l3 = NULL; 8851 } 8852 8853 int netdev_offload_xstats_disable(struct net_device *dev, 8854 enum netdev_offload_xstats_type type) 8855 { 8856 ASSERT_RTNL(); 8857 8858 if (!netdev_offload_xstats_enabled(dev, type)) 8859 return -EALREADY; 8860 8861 switch (type) { 8862 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8863 netdev_offload_xstats_disable_l3(dev); 8864 return 0; 8865 } 8866 8867 WARN_ON(1); 8868 return -EINVAL; 8869 } 8870 EXPORT_SYMBOL(netdev_offload_xstats_disable); 8871 8872 static void netdev_offload_xstats_disable_all(struct net_device *dev) 8873 { 8874 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3); 8875 } 8876 8877 static struct rtnl_hw_stats64 * 8878 netdev_offload_xstats_get_ptr(const struct net_device *dev, 8879 enum netdev_offload_xstats_type type) 8880 { 8881 switch (type) { 8882 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8883 return dev->offload_xstats_l3; 8884 } 8885 8886 WARN_ON(1); 8887 return NULL; 8888 } 8889 8890 bool netdev_offload_xstats_enabled(const struct net_device *dev, 8891 enum netdev_offload_xstats_type type) 8892 { 8893 ASSERT_RTNL(); 8894 8895 return netdev_offload_xstats_get_ptr(dev, type); 8896 } 8897 EXPORT_SYMBOL(netdev_offload_xstats_enabled); 8898 8899 struct netdev_notifier_offload_xstats_ru { 8900 bool used; 8901 }; 8902 8903 struct netdev_notifier_offload_xstats_rd { 8904 struct rtnl_hw_stats64 stats; 8905 bool used; 8906 }; 8907 8908 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest, 8909 const struct rtnl_hw_stats64 *src) 8910 { 8911 dest->rx_packets += src->rx_packets; 8912 dest->tx_packets += src->tx_packets; 8913 dest->rx_bytes += src->rx_bytes; 8914 dest->tx_bytes += src->tx_bytes; 8915 dest->rx_errors += src->rx_errors; 8916 dest->tx_errors += src->tx_errors; 8917 dest->rx_dropped += src->rx_dropped; 8918 dest->tx_dropped += src->tx_dropped; 8919 dest->multicast += src->multicast; 8920 } 8921 8922 static int netdev_offload_xstats_get_used(struct net_device *dev, 8923 enum netdev_offload_xstats_type type, 8924 bool *p_used, 8925 struct netlink_ext_ack *extack) 8926 { 8927 struct netdev_notifier_offload_xstats_ru report_used = {}; 8928 struct netdev_notifier_offload_xstats_info info = { 8929 .info.dev = dev, 8930 .info.extack = extack, 8931 .type = type, 8932 .report_used = &report_used, 8933 }; 8934 int rc; 8935 8936 WARN_ON(!netdev_offload_xstats_enabled(dev, type)); 8937 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED, 8938 &info.info); 8939 *p_used = report_used.used; 8940 return notifier_to_errno(rc); 8941 } 8942 8943 static int netdev_offload_xstats_get_stats(struct net_device *dev, 8944 enum netdev_offload_xstats_type type, 8945 struct rtnl_hw_stats64 *p_stats, 8946 bool *p_used, 8947 struct netlink_ext_ack *extack) 8948 { 8949 struct netdev_notifier_offload_xstats_rd report_delta = {}; 8950 struct netdev_notifier_offload_xstats_info info = { 8951 .info.dev = dev, 8952 .info.extack = extack, 8953 .type = type, 8954 .report_delta = &report_delta, 8955 }; 8956 struct rtnl_hw_stats64 *stats; 8957 int rc; 8958 8959 stats = netdev_offload_xstats_get_ptr(dev, type); 8960 if (WARN_ON(!stats)) 8961 return -EINVAL; 8962 8963 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA, 8964 &info.info); 8965 8966 /* Cache whatever we got, even if there was an error, otherwise the 8967 * successful stats retrievals would get lost. 8968 */ 8969 netdev_hw_stats64_add(stats, &report_delta.stats); 8970 8971 if (p_stats) 8972 *p_stats = *stats; 8973 *p_used = report_delta.used; 8974 8975 return notifier_to_errno(rc); 8976 } 8977 8978 int netdev_offload_xstats_get(struct net_device *dev, 8979 enum netdev_offload_xstats_type type, 8980 struct rtnl_hw_stats64 *p_stats, bool *p_used, 8981 struct netlink_ext_ack *extack) 8982 { 8983 ASSERT_RTNL(); 8984 8985 if (p_stats) 8986 return netdev_offload_xstats_get_stats(dev, type, p_stats, 8987 p_used, extack); 8988 else 8989 return netdev_offload_xstats_get_used(dev, type, p_used, 8990 extack); 8991 } 8992 EXPORT_SYMBOL(netdev_offload_xstats_get); 8993 8994 void 8995 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta, 8996 const struct rtnl_hw_stats64 *stats) 8997 { 8998 report_delta->used = true; 8999 netdev_hw_stats64_add(&report_delta->stats, stats); 9000 } 9001 EXPORT_SYMBOL(netdev_offload_xstats_report_delta); 9002 9003 void 9004 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used) 9005 { 9006 report_used->used = true; 9007 } 9008 EXPORT_SYMBOL(netdev_offload_xstats_report_used); 9009 9010 void netdev_offload_xstats_push_delta(struct net_device *dev, 9011 enum netdev_offload_xstats_type type, 9012 const struct rtnl_hw_stats64 *p_stats) 9013 { 9014 struct rtnl_hw_stats64 *stats; 9015 9016 ASSERT_RTNL(); 9017 9018 stats = netdev_offload_xstats_get_ptr(dev, type); 9019 if (WARN_ON(!stats)) 9020 return; 9021 9022 netdev_hw_stats64_add(stats, p_stats); 9023 } 9024 EXPORT_SYMBOL(netdev_offload_xstats_push_delta); 9025 9026 /** 9027 * netdev_get_xmit_slave - Get the xmit slave of master device 9028 * @dev: device 9029 * @skb: The packet 9030 * @all_slaves: assume all the slaves are active 9031 * 9032 * The reference counters are not incremented so the caller must be 9033 * careful with locks. The caller must hold RCU lock. 9034 * %NULL is returned if no slave is found. 9035 */ 9036 9037 struct net_device *netdev_get_xmit_slave(struct net_device *dev, 9038 struct sk_buff *skb, 9039 bool all_slaves) 9040 { 9041 const struct net_device_ops *ops = dev->netdev_ops; 9042 9043 if (!ops->ndo_get_xmit_slave) 9044 return NULL; 9045 return ops->ndo_get_xmit_slave(dev, skb, all_slaves); 9046 } 9047 EXPORT_SYMBOL(netdev_get_xmit_slave); 9048 9049 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev, 9050 struct sock *sk) 9051 { 9052 const struct net_device_ops *ops = dev->netdev_ops; 9053 9054 if (!ops->ndo_sk_get_lower_dev) 9055 return NULL; 9056 return ops->ndo_sk_get_lower_dev(dev, sk); 9057 } 9058 9059 /** 9060 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket 9061 * @dev: device 9062 * @sk: the socket 9063 * 9064 * %NULL is returned if no lower device is found. 9065 */ 9066 9067 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev, 9068 struct sock *sk) 9069 { 9070 struct net_device *lower; 9071 9072 lower = netdev_sk_get_lower_dev(dev, sk); 9073 while (lower) { 9074 dev = lower; 9075 lower = netdev_sk_get_lower_dev(dev, sk); 9076 } 9077 9078 return dev; 9079 } 9080 EXPORT_SYMBOL(netdev_sk_get_lowest_dev); 9081 9082 static void netdev_adjacent_add_links(struct net_device *dev) 9083 { 9084 struct netdev_adjacent *iter; 9085 9086 struct net *net = dev_net(dev); 9087 9088 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9089 if (!net_eq(net, dev_net(iter->dev))) 9090 continue; 9091 netdev_adjacent_sysfs_add(iter->dev, dev, 9092 &iter->dev->adj_list.lower); 9093 netdev_adjacent_sysfs_add(dev, iter->dev, 9094 &dev->adj_list.upper); 9095 } 9096 9097 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9098 if (!net_eq(net, dev_net(iter->dev))) 9099 continue; 9100 netdev_adjacent_sysfs_add(iter->dev, dev, 9101 &iter->dev->adj_list.upper); 9102 netdev_adjacent_sysfs_add(dev, iter->dev, 9103 &dev->adj_list.lower); 9104 } 9105 } 9106 9107 static void netdev_adjacent_del_links(struct net_device *dev) 9108 { 9109 struct netdev_adjacent *iter; 9110 9111 struct net *net = dev_net(dev); 9112 9113 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9114 if (!net_eq(net, dev_net(iter->dev))) 9115 continue; 9116 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9117 &iter->dev->adj_list.lower); 9118 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9119 &dev->adj_list.upper); 9120 } 9121 9122 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9123 if (!net_eq(net, dev_net(iter->dev))) 9124 continue; 9125 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9126 &iter->dev->adj_list.upper); 9127 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9128 &dev->adj_list.lower); 9129 } 9130 } 9131 9132 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname) 9133 { 9134 struct netdev_adjacent *iter; 9135 9136 struct net *net = dev_net(dev); 9137 9138 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9139 if (!net_eq(net, dev_net(iter->dev))) 9140 continue; 9141 netdev_adjacent_sysfs_del(iter->dev, oldname, 9142 &iter->dev->adj_list.lower); 9143 netdev_adjacent_sysfs_add(iter->dev, dev, 9144 &iter->dev->adj_list.lower); 9145 } 9146 9147 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9148 if (!net_eq(net, dev_net(iter->dev))) 9149 continue; 9150 netdev_adjacent_sysfs_del(iter->dev, oldname, 9151 &iter->dev->adj_list.upper); 9152 netdev_adjacent_sysfs_add(iter->dev, dev, 9153 &iter->dev->adj_list.upper); 9154 } 9155 } 9156 9157 void *netdev_lower_dev_get_private(struct net_device *dev, 9158 struct net_device *lower_dev) 9159 { 9160 struct netdev_adjacent *lower; 9161 9162 if (!lower_dev) 9163 return NULL; 9164 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower); 9165 if (!lower) 9166 return NULL; 9167 9168 return lower->private; 9169 } 9170 EXPORT_SYMBOL(netdev_lower_dev_get_private); 9171 9172 9173 /** 9174 * netdev_lower_state_changed - Dispatch event about lower device state change 9175 * @lower_dev: device 9176 * @lower_state_info: state to dispatch 9177 * 9178 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info. 9179 * The caller must hold the RTNL lock. 9180 */ 9181 void netdev_lower_state_changed(struct net_device *lower_dev, 9182 void *lower_state_info) 9183 { 9184 struct netdev_notifier_changelowerstate_info changelowerstate_info = { 9185 .info.dev = lower_dev, 9186 }; 9187 9188 ASSERT_RTNL(); 9189 changelowerstate_info.lower_state_info = lower_state_info; 9190 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, 9191 &changelowerstate_info.info); 9192 } 9193 EXPORT_SYMBOL(netdev_lower_state_changed); 9194 9195 static void dev_change_rx_flags(struct net_device *dev, int flags) 9196 { 9197 const struct net_device_ops *ops = dev->netdev_ops; 9198 9199 if (ops->ndo_change_rx_flags) 9200 ops->ndo_change_rx_flags(dev, flags); 9201 } 9202 9203 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify) 9204 { 9205 unsigned int old_flags = dev->flags; 9206 unsigned int promiscuity, flags; 9207 kuid_t uid; 9208 kgid_t gid; 9209 9210 ASSERT_RTNL(); 9211 9212 promiscuity = dev->promiscuity + inc; 9213 if (promiscuity == 0) { 9214 /* 9215 * Avoid overflow. 9216 * If inc causes overflow, untouch promisc and return error. 9217 */ 9218 if (unlikely(inc > 0)) { 9219 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n"); 9220 return -EOVERFLOW; 9221 } 9222 flags = old_flags & ~IFF_PROMISC; 9223 } else { 9224 flags = old_flags | IFF_PROMISC; 9225 } 9226 WRITE_ONCE(dev->promiscuity, promiscuity); 9227 if (flags != old_flags) { 9228 WRITE_ONCE(dev->flags, flags); 9229 netdev_info(dev, "%s promiscuous mode\n", 9230 dev->flags & IFF_PROMISC ? "entered" : "left"); 9231 if (audit_enabled) { 9232 current_uid_gid(&uid, &gid); 9233 audit_log(audit_context(), GFP_ATOMIC, 9234 AUDIT_ANOM_PROMISCUOUS, 9235 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u", 9236 dev->name, (dev->flags & IFF_PROMISC), 9237 (old_flags & IFF_PROMISC), 9238 from_kuid(&init_user_ns, audit_get_loginuid(current)), 9239 from_kuid(&init_user_ns, uid), 9240 from_kgid(&init_user_ns, gid), 9241 audit_get_sessionid(current)); 9242 } 9243 9244 dev_change_rx_flags(dev, IFF_PROMISC); 9245 } 9246 if (notify) 9247 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL); 9248 return 0; 9249 } 9250 9251 /** 9252 * dev_set_promiscuity - update promiscuity count on a device 9253 * @dev: device 9254 * @inc: modifier 9255 * 9256 * Add or remove promiscuity from a device. While the count in the device 9257 * remains above zero the interface remains promiscuous. Once it hits zero 9258 * the device reverts back to normal filtering operation. A negative inc 9259 * value is used to drop promiscuity on the device. 9260 * Return 0 if successful or a negative errno code on error. 9261 */ 9262 int dev_set_promiscuity(struct net_device *dev, int inc) 9263 { 9264 unsigned int old_flags = dev->flags; 9265 int err; 9266 9267 err = __dev_set_promiscuity(dev, inc, true); 9268 if (err < 0) 9269 return err; 9270 if (dev->flags != old_flags) 9271 dev_set_rx_mode(dev); 9272 return err; 9273 } 9274 EXPORT_SYMBOL(dev_set_promiscuity); 9275 9276 int netif_set_allmulti(struct net_device *dev, int inc, bool notify) 9277 { 9278 unsigned int old_flags = dev->flags, old_gflags = dev->gflags; 9279 unsigned int allmulti, flags; 9280 9281 ASSERT_RTNL(); 9282 9283 allmulti = dev->allmulti + inc; 9284 if (allmulti == 0) { 9285 /* 9286 * Avoid overflow. 9287 * If inc causes overflow, untouch allmulti and return error. 9288 */ 9289 if (unlikely(inc > 0)) { 9290 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n"); 9291 return -EOVERFLOW; 9292 } 9293 flags = old_flags & ~IFF_ALLMULTI; 9294 } else { 9295 flags = old_flags | IFF_ALLMULTI; 9296 } 9297 WRITE_ONCE(dev->allmulti, allmulti); 9298 if (flags != old_flags) { 9299 WRITE_ONCE(dev->flags, flags); 9300 netdev_info(dev, "%s allmulticast mode\n", 9301 dev->flags & IFF_ALLMULTI ? "entered" : "left"); 9302 dev_change_rx_flags(dev, IFF_ALLMULTI); 9303 dev_set_rx_mode(dev); 9304 if (notify) 9305 __dev_notify_flags(dev, old_flags, 9306 dev->gflags ^ old_gflags, 0, NULL); 9307 } 9308 return 0; 9309 } 9310 9311 /* 9312 * Upload unicast and multicast address lists to device and 9313 * configure RX filtering. When the device doesn't support unicast 9314 * filtering it is put in promiscuous mode while unicast addresses 9315 * are present. 9316 */ 9317 void __dev_set_rx_mode(struct net_device *dev) 9318 { 9319 const struct net_device_ops *ops = dev->netdev_ops; 9320 9321 /* dev_open will call this function so the list will stay sane. */ 9322 if (!(dev->flags&IFF_UP)) 9323 return; 9324 9325 if (!netif_device_present(dev)) 9326 return; 9327 9328 if (!(dev->priv_flags & IFF_UNICAST_FLT)) { 9329 /* Unicast addresses changes may only happen under the rtnl, 9330 * therefore calling __dev_set_promiscuity here is safe. 9331 */ 9332 if (!netdev_uc_empty(dev) && !dev->uc_promisc) { 9333 __dev_set_promiscuity(dev, 1, false); 9334 dev->uc_promisc = true; 9335 } else if (netdev_uc_empty(dev) && dev->uc_promisc) { 9336 __dev_set_promiscuity(dev, -1, false); 9337 dev->uc_promisc = false; 9338 } 9339 } 9340 9341 if (ops->ndo_set_rx_mode) 9342 ops->ndo_set_rx_mode(dev); 9343 } 9344 9345 void dev_set_rx_mode(struct net_device *dev) 9346 { 9347 netif_addr_lock_bh(dev); 9348 __dev_set_rx_mode(dev); 9349 netif_addr_unlock_bh(dev); 9350 } 9351 9352 /** 9353 * dev_get_flags - get flags reported to userspace 9354 * @dev: device 9355 * 9356 * Get the combination of flag bits exported through APIs to userspace. 9357 */ 9358 unsigned int dev_get_flags(const struct net_device *dev) 9359 { 9360 unsigned int flags; 9361 9362 flags = (READ_ONCE(dev->flags) & ~(IFF_PROMISC | 9363 IFF_ALLMULTI | 9364 IFF_RUNNING | 9365 IFF_LOWER_UP | 9366 IFF_DORMANT)) | 9367 (READ_ONCE(dev->gflags) & (IFF_PROMISC | 9368 IFF_ALLMULTI)); 9369 9370 if (netif_running(dev)) { 9371 if (netif_oper_up(dev)) 9372 flags |= IFF_RUNNING; 9373 if (netif_carrier_ok(dev)) 9374 flags |= IFF_LOWER_UP; 9375 if (netif_dormant(dev)) 9376 flags |= IFF_DORMANT; 9377 } 9378 9379 return flags; 9380 } 9381 EXPORT_SYMBOL(dev_get_flags); 9382 9383 int __dev_change_flags(struct net_device *dev, unsigned int flags, 9384 struct netlink_ext_ack *extack) 9385 { 9386 unsigned int old_flags = dev->flags; 9387 int ret; 9388 9389 ASSERT_RTNL(); 9390 9391 /* 9392 * Set the flags on our device. 9393 */ 9394 9395 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP | 9396 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL | 9397 IFF_AUTOMEDIA)) | 9398 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC | 9399 IFF_ALLMULTI)); 9400 9401 /* 9402 * Load in the correct multicast list now the flags have changed. 9403 */ 9404 9405 if ((old_flags ^ flags) & IFF_MULTICAST) 9406 dev_change_rx_flags(dev, IFF_MULTICAST); 9407 9408 dev_set_rx_mode(dev); 9409 9410 /* 9411 * Have we downed the interface. We handle IFF_UP ourselves 9412 * according to user attempts to set it, rather than blindly 9413 * setting it. 9414 */ 9415 9416 ret = 0; 9417 if ((old_flags ^ flags) & IFF_UP) { 9418 if (old_flags & IFF_UP) 9419 __dev_close(dev); 9420 else 9421 ret = __dev_open(dev, extack); 9422 } 9423 9424 if ((flags ^ dev->gflags) & IFF_PROMISC) { 9425 int inc = (flags & IFF_PROMISC) ? 1 : -1; 9426 old_flags = dev->flags; 9427 9428 dev->gflags ^= IFF_PROMISC; 9429 9430 if (__dev_set_promiscuity(dev, inc, false) >= 0) 9431 if (dev->flags != old_flags) 9432 dev_set_rx_mode(dev); 9433 } 9434 9435 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI 9436 * is important. Some (broken) drivers set IFF_PROMISC, when 9437 * IFF_ALLMULTI is requested not asking us and not reporting. 9438 */ 9439 if ((flags ^ dev->gflags) & IFF_ALLMULTI) { 9440 int inc = (flags & IFF_ALLMULTI) ? 1 : -1; 9441 9442 dev->gflags ^= IFF_ALLMULTI; 9443 netif_set_allmulti(dev, inc, false); 9444 } 9445 9446 return ret; 9447 } 9448 9449 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags, 9450 unsigned int gchanges, u32 portid, 9451 const struct nlmsghdr *nlh) 9452 { 9453 unsigned int changes = dev->flags ^ old_flags; 9454 9455 if (gchanges) 9456 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh); 9457 9458 if (changes & IFF_UP) { 9459 if (dev->flags & IFF_UP) 9460 call_netdevice_notifiers(NETDEV_UP, dev); 9461 else 9462 call_netdevice_notifiers(NETDEV_DOWN, dev); 9463 } 9464 9465 if (dev->flags & IFF_UP && 9466 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) { 9467 struct netdev_notifier_change_info change_info = { 9468 .info = { 9469 .dev = dev, 9470 }, 9471 .flags_changed = changes, 9472 }; 9473 9474 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info); 9475 } 9476 } 9477 9478 int netif_change_flags(struct net_device *dev, unsigned int flags, 9479 struct netlink_ext_ack *extack) 9480 { 9481 int ret; 9482 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags; 9483 9484 ret = __dev_change_flags(dev, flags, extack); 9485 if (ret < 0) 9486 return ret; 9487 9488 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags); 9489 __dev_notify_flags(dev, old_flags, changes, 0, NULL); 9490 return ret; 9491 } 9492 9493 int __dev_set_mtu(struct net_device *dev, int new_mtu) 9494 { 9495 const struct net_device_ops *ops = dev->netdev_ops; 9496 9497 if (ops->ndo_change_mtu) 9498 return ops->ndo_change_mtu(dev, new_mtu); 9499 9500 /* Pairs with all the lockless reads of dev->mtu in the stack */ 9501 WRITE_ONCE(dev->mtu, new_mtu); 9502 return 0; 9503 } 9504 EXPORT_SYMBOL(__dev_set_mtu); 9505 9506 int dev_validate_mtu(struct net_device *dev, int new_mtu, 9507 struct netlink_ext_ack *extack) 9508 { 9509 /* MTU must be positive, and in range */ 9510 if (new_mtu < 0 || new_mtu < dev->min_mtu) { 9511 NL_SET_ERR_MSG(extack, "mtu less than device minimum"); 9512 return -EINVAL; 9513 } 9514 9515 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) { 9516 NL_SET_ERR_MSG(extack, "mtu greater than device maximum"); 9517 return -EINVAL; 9518 } 9519 return 0; 9520 } 9521 9522 /** 9523 * netif_set_mtu_ext - Change maximum transfer unit 9524 * @dev: device 9525 * @new_mtu: new transfer unit 9526 * @extack: netlink extended ack 9527 * 9528 * Change the maximum transfer size of the network device. 9529 */ 9530 int netif_set_mtu_ext(struct net_device *dev, int new_mtu, 9531 struct netlink_ext_ack *extack) 9532 { 9533 int err, orig_mtu; 9534 9535 if (new_mtu == dev->mtu) 9536 return 0; 9537 9538 err = dev_validate_mtu(dev, new_mtu, extack); 9539 if (err) 9540 return err; 9541 9542 if (!netif_device_present(dev)) 9543 return -ENODEV; 9544 9545 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev); 9546 err = notifier_to_errno(err); 9547 if (err) 9548 return err; 9549 9550 orig_mtu = dev->mtu; 9551 err = __dev_set_mtu(dev, new_mtu); 9552 9553 if (!err) { 9554 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9555 orig_mtu); 9556 err = notifier_to_errno(err); 9557 if (err) { 9558 /* setting mtu back and notifying everyone again, 9559 * so that they have a chance to revert changes. 9560 */ 9561 __dev_set_mtu(dev, orig_mtu); 9562 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9563 new_mtu); 9564 } 9565 } 9566 return err; 9567 } 9568 9569 int netif_set_mtu(struct net_device *dev, int new_mtu) 9570 { 9571 struct netlink_ext_ack extack; 9572 int err; 9573 9574 memset(&extack, 0, sizeof(extack)); 9575 err = netif_set_mtu_ext(dev, new_mtu, &extack); 9576 if (err && extack._msg) 9577 net_err_ratelimited("%s: %s\n", dev->name, extack._msg); 9578 return err; 9579 } 9580 EXPORT_SYMBOL(netif_set_mtu); 9581 9582 int netif_change_tx_queue_len(struct net_device *dev, unsigned long new_len) 9583 { 9584 unsigned int orig_len = dev->tx_queue_len; 9585 int res; 9586 9587 if (new_len != (unsigned int)new_len) 9588 return -ERANGE; 9589 9590 if (new_len != orig_len) { 9591 WRITE_ONCE(dev->tx_queue_len, new_len); 9592 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev); 9593 res = notifier_to_errno(res); 9594 if (res) 9595 goto err_rollback; 9596 res = dev_qdisc_change_tx_queue_len(dev); 9597 if (res) 9598 goto err_rollback; 9599 } 9600 9601 return 0; 9602 9603 err_rollback: 9604 netdev_err(dev, "refused to change device tx_queue_len\n"); 9605 WRITE_ONCE(dev->tx_queue_len, orig_len); 9606 return res; 9607 } 9608 9609 void netif_set_group(struct net_device *dev, int new_group) 9610 { 9611 dev->group = new_group; 9612 } 9613 9614 /** 9615 * dev_pre_changeaddr_notify - Call NETDEV_PRE_CHANGEADDR. 9616 * @dev: device 9617 * @addr: new address 9618 * @extack: netlink extended ack 9619 */ 9620 int dev_pre_changeaddr_notify(struct net_device *dev, const char *addr, 9621 struct netlink_ext_ack *extack) 9622 { 9623 struct netdev_notifier_pre_changeaddr_info info = { 9624 .info.dev = dev, 9625 .info.extack = extack, 9626 .dev_addr = addr, 9627 }; 9628 int rc; 9629 9630 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info); 9631 return notifier_to_errno(rc); 9632 } 9633 EXPORT_SYMBOL(dev_pre_changeaddr_notify); 9634 9635 int netif_set_mac_address(struct net_device *dev, struct sockaddr *sa, 9636 struct netlink_ext_ack *extack) 9637 { 9638 const struct net_device_ops *ops = dev->netdev_ops; 9639 int err; 9640 9641 if (!ops->ndo_set_mac_address) 9642 return -EOPNOTSUPP; 9643 if (sa->sa_family != dev->type) 9644 return -EINVAL; 9645 if (!netif_device_present(dev)) 9646 return -ENODEV; 9647 err = dev_pre_changeaddr_notify(dev, sa->sa_data, extack); 9648 if (err) 9649 return err; 9650 if (memcmp(dev->dev_addr, sa->sa_data, dev->addr_len)) { 9651 err = ops->ndo_set_mac_address(dev, sa); 9652 if (err) 9653 return err; 9654 } 9655 dev->addr_assign_type = NET_ADDR_SET; 9656 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev); 9657 add_device_randomness(dev->dev_addr, dev->addr_len); 9658 return 0; 9659 } 9660 9661 DECLARE_RWSEM(dev_addr_sem); 9662 9663 int dev_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name) 9664 { 9665 size_t size = sizeof(sa->sa_data_min); 9666 struct net_device *dev; 9667 int ret = 0; 9668 9669 down_read(&dev_addr_sem); 9670 rcu_read_lock(); 9671 9672 dev = dev_get_by_name_rcu(net, dev_name); 9673 if (!dev) { 9674 ret = -ENODEV; 9675 goto unlock; 9676 } 9677 if (!dev->addr_len) 9678 memset(sa->sa_data, 0, size); 9679 else 9680 memcpy(sa->sa_data, dev->dev_addr, 9681 min_t(size_t, size, dev->addr_len)); 9682 sa->sa_family = dev->type; 9683 9684 unlock: 9685 rcu_read_unlock(); 9686 up_read(&dev_addr_sem); 9687 return ret; 9688 } 9689 EXPORT_SYMBOL(dev_get_mac_address); 9690 9691 int netif_change_carrier(struct net_device *dev, bool new_carrier) 9692 { 9693 const struct net_device_ops *ops = dev->netdev_ops; 9694 9695 if (!ops->ndo_change_carrier) 9696 return -EOPNOTSUPP; 9697 if (!netif_device_present(dev)) 9698 return -ENODEV; 9699 return ops->ndo_change_carrier(dev, new_carrier); 9700 } 9701 9702 /** 9703 * dev_get_phys_port_id - Get device physical port ID 9704 * @dev: device 9705 * @ppid: port ID 9706 * 9707 * Get device physical port ID 9708 */ 9709 int dev_get_phys_port_id(struct net_device *dev, 9710 struct netdev_phys_item_id *ppid) 9711 { 9712 const struct net_device_ops *ops = dev->netdev_ops; 9713 9714 if (!ops->ndo_get_phys_port_id) 9715 return -EOPNOTSUPP; 9716 return ops->ndo_get_phys_port_id(dev, ppid); 9717 } 9718 9719 /** 9720 * dev_get_phys_port_name - Get device physical port name 9721 * @dev: device 9722 * @name: port name 9723 * @len: limit of bytes to copy to name 9724 * 9725 * Get device physical port name 9726 */ 9727 int dev_get_phys_port_name(struct net_device *dev, 9728 char *name, size_t len) 9729 { 9730 const struct net_device_ops *ops = dev->netdev_ops; 9731 int err; 9732 9733 if (ops->ndo_get_phys_port_name) { 9734 err = ops->ndo_get_phys_port_name(dev, name, len); 9735 if (err != -EOPNOTSUPP) 9736 return err; 9737 } 9738 return devlink_compat_phys_port_name_get(dev, name, len); 9739 } 9740 9741 /** 9742 * dev_get_port_parent_id - Get the device's port parent identifier 9743 * @dev: network device 9744 * @ppid: pointer to a storage for the port's parent identifier 9745 * @recurse: allow/disallow recursion to lower devices 9746 * 9747 * Get the devices's port parent identifier 9748 */ 9749 int dev_get_port_parent_id(struct net_device *dev, 9750 struct netdev_phys_item_id *ppid, 9751 bool recurse) 9752 { 9753 const struct net_device_ops *ops = dev->netdev_ops; 9754 struct netdev_phys_item_id first = { }; 9755 struct net_device *lower_dev; 9756 struct list_head *iter; 9757 int err; 9758 9759 if (ops->ndo_get_port_parent_id) { 9760 err = ops->ndo_get_port_parent_id(dev, ppid); 9761 if (err != -EOPNOTSUPP) 9762 return err; 9763 } 9764 9765 err = devlink_compat_switch_id_get(dev, ppid); 9766 if (!recurse || err != -EOPNOTSUPP) 9767 return err; 9768 9769 netdev_for_each_lower_dev(dev, lower_dev, iter) { 9770 err = dev_get_port_parent_id(lower_dev, ppid, true); 9771 if (err) 9772 break; 9773 if (!first.id_len) 9774 first = *ppid; 9775 else if (memcmp(&first, ppid, sizeof(*ppid))) 9776 return -EOPNOTSUPP; 9777 } 9778 9779 return err; 9780 } 9781 EXPORT_SYMBOL(dev_get_port_parent_id); 9782 9783 /** 9784 * netdev_port_same_parent_id - Indicate if two network devices have 9785 * the same port parent identifier 9786 * @a: first network device 9787 * @b: second network device 9788 */ 9789 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b) 9790 { 9791 struct netdev_phys_item_id a_id = { }; 9792 struct netdev_phys_item_id b_id = { }; 9793 9794 if (dev_get_port_parent_id(a, &a_id, true) || 9795 dev_get_port_parent_id(b, &b_id, true)) 9796 return false; 9797 9798 return netdev_phys_item_id_same(&a_id, &b_id); 9799 } 9800 EXPORT_SYMBOL(netdev_port_same_parent_id); 9801 9802 int netif_change_proto_down(struct net_device *dev, bool proto_down) 9803 { 9804 if (!dev->change_proto_down) 9805 return -EOPNOTSUPP; 9806 if (!netif_device_present(dev)) 9807 return -ENODEV; 9808 if (proto_down) 9809 netif_carrier_off(dev); 9810 else 9811 netif_carrier_on(dev); 9812 WRITE_ONCE(dev->proto_down, proto_down); 9813 return 0; 9814 } 9815 9816 /** 9817 * netdev_change_proto_down_reason_locked - proto down reason 9818 * 9819 * @dev: device 9820 * @mask: proto down mask 9821 * @value: proto down value 9822 */ 9823 void netdev_change_proto_down_reason_locked(struct net_device *dev, 9824 unsigned long mask, u32 value) 9825 { 9826 u32 proto_down_reason; 9827 int b; 9828 9829 if (!mask) { 9830 proto_down_reason = value; 9831 } else { 9832 proto_down_reason = dev->proto_down_reason; 9833 for_each_set_bit(b, &mask, 32) { 9834 if (value & (1 << b)) 9835 proto_down_reason |= BIT(b); 9836 else 9837 proto_down_reason &= ~BIT(b); 9838 } 9839 } 9840 WRITE_ONCE(dev->proto_down_reason, proto_down_reason); 9841 } 9842 9843 struct bpf_xdp_link { 9844 struct bpf_link link; 9845 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */ 9846 int flags; 9847 }; 9848 9849 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags) 9850 { 9851 if (flags & XDP_FLAGS_HW_MODE) 9852 return XDP_MODE_HW; 9853 if (flags & XDP_FLAGS_DRV_MODE) 9854 return XDP_MODE_DRV; 9855 if (flags & XDP_FLAGS_SKB_MODE) 9856 return XDP_MODE_SKB; 9857 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB; 9858 } 9859 9860 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode) 9861 { 9862 switch (mode) { 9863 case XDP_MODE_SKB: 9864 return generic_xdp_install; 9865 case XDP_MODE_DRV: 9866 case XDP_MODE_HW: 9867 return dev->netdev_ops->ndo_bpf; 9868 default: 9869 return NULL; 9870 } 9871 } 9872 9873 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev, 9874 enum bpf_xdp_mode mode) 9875 { 9876 return dev->xdp_state[mode].link; 9877 } 9878 9879 static struct bpf_prog *dev_xdp_prog(struct net_device *dev, 9880 enum bpf_xdp_mode mode) 9881 { 9882 struct bpf_xdp_link *link = dev_xdp_link(dev, mode); 9883 9884 if (link) 9885 return link->link.prog; 9886 return dev->xdp_state[mode].prog; 9887 } 9888 9889 u8 dev_xdp_prog_count(struct net_device *dev) 9890 { 9891 u8 count = 0; 9892 int i; 9893 9894 for (i = 0; i < __MAX_XDP_MODE; i++) 9895 if (dev->xdp_state[i].prog || dev->xdp_state[i].link) 9896 count++; 9897 return count; 9898 } 9899 EXPORT_SYMBOL_GPL(dev_xdp_prog_count); 9900 9901 u8 dev_xdp_sb_prog_count(struct net_device *dev) 9902 { 9903 u8 count = 0; 9904 int i; 9905 9906 for (i = 0; i < __MAX_XDP_MODE; i++) 9907 if (dev->xdp_state[i].prog && 9908 !dev->xdp_state[i].prog->aux->xdp_has_frags) 9909 count++; 9910 return count; 9911 } 9912 9913 int netif_xdp_propagate(struct net_device *dev, struct netdev_bpf *bpf) 9914 { 9915 if (!dev->netdev_ops->ndo_bpf) 9916 return -EOPNOTSUPP; 9917 9918 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 9919 bpf->command == XDP_SETUP_PROG && 9920 bpf->prog && !bpf->prog->aux->xdp_has_frags) { 9921 NL_SET_ERR_MSG(bpf->extack, 9922 "unable to propagate XDP to device using tcp-data-split"); 9923 return -EBUSY; 9924 } 9925 9926 if (dev_get_min_mp_channel_count(dev)) { 9927 NL_SET_ERR_MSG(bpf->extack, "unable to propagate XDP to device using memory provider"); 9928 return -EBUSY; 9929 } 9930 9931 return dev->netdev_ops->ndo_bpf(dev, bpf); 9932 } 9933 9934 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode) 9935 { 9936 struct bpf_prog *prog = dev_xdp_prog(dev, mode); 9937 9938 return prog ? prog->aux->id : 0; 9939 } 9940 9941 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode, 9942 struct bpf_xdp_link *link) 9943 { 9944 dev->xdp_state[mode].link = link; 9945 dev->xdp_state[mode].prog = NULL; 9946 } 9947 9948 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode, 9949 struct bpf_prog *prog) 9950 { 9951 dev->xdp_state[mode].link = NULL; 9952 dev->xdp_state[mode].prog = prog; 9953 } 9954 9955 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode, 9956 bpf_op_t bpf_op, struct netlink_ext_ack *extack, 9957 u32 flags, struct bpf_prog *prog) 9958 { 9959 struct netdev_bpf xdp; 9960 int err; 9961 9962 netdev_ops_assert_locked(dev); 9963 9964 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 9965 prog && !prog->aux->xdp_has_frags) { 9966 NL_SET_ERR_MSG(extack, "unable to install XDP to device using tcp-data-split"); 9967 return -EBUSY; 9968 } 9969 9970 if (dev_get_min_mp_channel_count(dev)) { 9971 NL_SET_ERR_MSG(extack, "unable to install XDP to device using memory provider"); 9972 return -EBUSY; 9973 } 9974 9975 memset(&xdp, 0, sizeof(xdp)); 9976 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG; 9977 xdp.extack = extack; 9978 xdp.flags = flags; 9979 xdp.prog = prog; 9980 9981 /* Drivers assume refcnt is already incremented (i.e, prog pointer is 9982 * "moved" into driver), so they don't increment it on their own, but 9983 * they do decrement refcnt when program is detached or replaced. 9984 * Given net_device also owns link/prog, we need to bump refcnt here 9985 * to prevent drivers from underflowing it. 9986 */ 9987 if (prog) 9988 bpf_prog_inc(prog); 9989 err = bpf_op(dev, &xdp); 9990 if (err) { 9991 if (prog) 9992 bpf_prog_put(prog); 9993 return err; 9994 } 9995 9996 if (mode != XDP_MODE_HW) 9997 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog); 9998 9999 return 0; 10000 } 10001 10002 static void dev_xdp_uninstall(struct net_device *dev) 10003 { 10004 struct bpf_xdp_link *link; 10005 struct bpf_prog *prog; 10006 enum bpf_xdp_mode mode; 10007 bpf_op_t bpf_op; 10008 10009 ASSERT_RTNL(); 10010 10011 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) { 10012 prog = dev_xdp_prog(dev, mode); 10013 if (!prog) 10014 continue; 10015 10016 bpf_op = dev_xdp_bpf_op(dev, mode); 10017 if (!bpf_op) 10018 continue; 10019 10020 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 10021 10022 /* auto-detach link from net device */ 10023 link = dev_xdp_link(dev, mode); 10024 if (link) 10025 link->dev = NULL; 10026 else 10027 bpf_prog_put(prog); 10028 10029 dev_xdp_set_link(dev, mode, NULL); 10030 } 10031 } 10032 10033 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack, 10034 struct bpf_xdp_link *link, struct bpf_prog *new_prog, 10035 struct bpf_prog *old_prog, u32 flags) 10036 { 10037 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES); 10038 struct bpf_prog *cur_prog; 10039 struct net_device *upper; 10040 struct list_head *iter; 10041 enum bpf_xdp_mode mode; 10042 bpf_op_t bpf_op; 10043 int err; 10044 10045 ASSERT_RTNL(); 10046 10047 /* either link or prog attachment, never both */ 10048 if (link && (new_prog || old_prog)) 10049 return -EINVAL; 10050 /* link supports only XDP mode flags */ 10051 if (link && (flags & ~XDP_FLAGS_MODES)) { 10052 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment"); 10053 return -EINVAL; 10054 } 10055 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */ 10056 if (num_modes > 1) { 10057 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set"); 10058 return -EINVAL; 10059 } 10060 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */ 10061 if (!num_modes && dev_xdp_prog_count(dev) > 1) { 10062 NL_SET_ERR_MSG(extack, 10063 "More than one program loaded, unset mode is ambiguous"); 10064 return -EINVAL; 10065 } 10066 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */ 10067 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) { 10068 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified"); 10069 return -EINVAL; 10070 } 10071 10072 mode = dev_xdp_mode(dev, flags); 10073 /* can't replace attached link */ 10074 if (dev_xdp_link(dev, mode)) { 10075 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link"); 10076 return -EBUSY; 10077 } 10078 10079 /* don't allow if an upper device already has a program */ 10080 netdev_for_each_upper_dev_rcu(dev, upper, iter) { 10081 if (dev_xdp_prog_count(upper) > 0) { 10082 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program"); 10083 return -EEXIST; 10084 } 10085 } 10086 10087 cur_prog = dev_xdp_prog(dev, mode); 10088 /* can't replace attached prog with link */ 10089 if (link && cur_prog) { 10090 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link"); 10091 return -EBUSY; 10092 } 10093 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) { 10094 NL_SET_ERR_MSG(extack, "Active program does not match expected"); 10095 return -EEXIST; 10096 } 10097 10098 /* put effective new program into new_prog */ 10099 if (link) 10100 new_prog = link->link.prog; 10101 10102 if (new_prog) { 10103 bool offload = mode == XDP_MODE_HW; 10104 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB 10105 ? XDP_MODE_DRV : XDP_MODE_SKB; 10106 10107 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) { 10108 NL_SET_ERR_MSG(extack, "XDP program already attached"); 10109 return -EBUSY; 10110 } 10111 if (!offload && dev_xdp_prog(dev, other_mode)) { 10112 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time"); 10113 return -EEXIST; 10114 } 10115 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) { 10116 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported"); 10117 return -EINVAL; 10118 } 10119 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) { 10120 NL_SET_ERR_MSG(extack, "Program bound to different device"); 10121 return -EINVAL; 10122 } 10123 if (bpf_prog_is_dev_bound(new_prog->aux) && mode == XDP_MODE_SKB) { 10124 NL_SET_ERR_MSG(extack, "Can't attach device-bound programs in generic mode"); 10125 return -EINVAL; 10126 } 10127 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) { 10128 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device"); 10129 return -EINVAL; 10130 } 10131 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) { 10132 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device"); 10133 return -EINVAL; 10134 } 10135 } 10136 10137 /* don't call drivers if the effective program didn't change */ 10138 if (new_prog != cur_prog) { 10139 bpf_op = dev_xdp_bpf_op(dev, mode); 10140 if (!bpf_op) { 10141 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode"); 10142 return -EOPNOTSUPP; 10143 } 10144 10145 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog); 10146 if (err) 10147 return err; 10148 } 10149 10150 if (link) 10151 dev_xdp_set_link(dev, mode, link); 10152 else 10153 dev_xdp_set_prog(dev, mode, new_prog); 10154 if (cur_prog) 10155 bpf_prog_put(cur_prog); 10156 10157 return 0; 10158 } 10159 10160 static int dev_xdp_attach_link(struct net_device *dev, 10161 struct netlink_ext_ack *extack, 10162 struct bpf_xdp_link *link) 10163 { 10164 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags); 10165 } 10166 10167 static int dev_xdp_detach_link(struct net_device *dev, 10168 struct netlink_ext_ack *extack, 10169 struct bpf_xdp_link *link) 10170 { 10171 enum bpf_xdp_mode mode; 10172 bpf_op_t bpf_op; 10173 10174 ASSERT_RTNL(); 10175 10176 mode = dev_xdp_mode(dev, link->flags); 10177 if (dev_xdp_link(dev, mode) != link) 10178 return -EINVAL; 10179 10180 bpf_op = dev_xdp_bpf_op(dev, mode); 10181 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 10182 dev_xdp_set_link(dev, mode, NULL); 10183 return 0; 10184 } 10185 10186 static void bpf_xdp_link_release(struct bpf_link *link) 10187 { 10188 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10189 10190 rtnl_lock(); 10191 10192 /* if racing with net_device's tear down, xdp_link->dev might be 10193 * already NULL, in which case link was already auto-detached 10194 */ 10195 if (xdp_link->dev) { 10196 netdev_lock_ops(xdp_link->dev); 10197 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link)); 10198 netdev_unlock_ops(xdp_link->dev); 10199 xdp_link->dev = NULL; 10200 } 10201 10202 rtnl_unlock(); 10203 } 10204 10205 static int bpf_xdp_link_detach(struct bpf_link *link) 10206 { 10207 bpf_xdp_link_release(link); 10208 return 0; 10209 } 10210 10211 static void bpf_xdp_link_dealloc(struct bpf_link *link) 10212 { 10213 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10214 10215 kfree(xdp_link); 10216 } 10217 10218 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link, 10219 struct seq_file *seq) 10220 { 10221 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10222 u32 ifindex = 0; 10223 10224 rtnl_lock(); 10225 if (xdp_link->dev) 10226 ifindex = xdp_link->dev->ifindex; 10227 rtnl_unlock(); 10228 10229 seq_printf(seq, "ifindex:\t%u\n", ifindex); 10230 } 10231 10232 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link, 10233 struct bpf_link_info *info) 10234 { 10235 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10236 u32 ifindex = 0; 10237 10238 rtnl_lock(); 10239 if (xdp_link->dev) 10240 ifindex = xdp_link->dev->ifindex; 10241 rtnl_unlock(); 10242 10243 info->xdp.ifindex = ifindex; 10244 return 0; 10245 } 10246 10247 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog, 10248 struct bpf_prog *old_prog) 10249 { 10250 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10251 enum bpf_xdp_mode mode; 10252 bpf_op_t bpf_op; 10253 int err = 0; 10254 10255 rtnl_lock(); 10256 10257 /* link might have been auto-released already, so fail */ 10258 if (!xdp_link->dev) { 10259 err = -ENOLINK; 10260 goto out_unlock; 10261 } 10262 10263 if (old_prog && link->prog != old_prog) { 10264 err = -EPERM; 10265 goto out_unlock; 10266 } 10267 old_prog = link->prog; 10268 if (old_prog->type != new_prog->type || 10269 old_prog->expected_attach_type != new_prog->expected_attach_type) { 10270 err = -EINVAL; 10271 goto out_unlock; 10272 } 10273 10274 if (old_prog == new_prog) { 10275 /* no-op, don't disturb drivers */ 10276 bpf_prog_put(new_prog); 10277 goto out_unlock; 10278 } 10279 10280 netdev_lock_ops(xdp_link->dev); 10281 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags); 10282 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode); 10283 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL, 10284 xdp_link->flags, new_prog); 10285 netdev_unlock_ops(xdp_link->dev); 10286 if (err) 10287 goto out_unlock; 10288 10289 old_prog = xchg(&link->prog, new_prog); 10290 bpf_prog_put(old_prog); 10291 10292 out_unlock: 10293 rtnl_unlock(); 10294 return err; 10295 } 10296 10297 static const struct bpf_link_ops bpf_xdp_link_lops = { 10298 .release = bpf_xdp_link_release, 10299 .dealloc = bpf_xdp_link_dealloc, 10300 .detach = bpf_xdp_link_detach, 10301 .show_fdinfo = bpf_xdp_link_show_fdinfo, 10302 .fill_link_info = bpf_xdp_link_fill_link_info, 10303 .update_prog = bpf_xdp_link_update, 10304 }; 10305 10306 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog) 10307 { 10308 struct net *net = current->nsproxy->net_ns; 10309 struct bpf_link_primer link_primer; 10310 struct netlink_ext_ack extack = {}; 10311 struct bpf_xdp_link *link; 10312 struct net_device *dev; 10313 int err, fd; 10314 10315 rtnl_lock(); 10316 dev = dev_get_by_index(net, attr->link_create.target_ifindex); 10317 if (!dev) { 10318 rtnl_unlock(); 10319 return -EINVAL; 10320 } 10321 10322 link = kzalloc(sizeof(*link), GFP_USER); 10323 if (!link) { 10324 err = -ENOMEM; 10325 goto unlock; 10326 } 10327 10328 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog); 10329 link->dev = dev; 10330 link->flags = attr->link_create.flags; 10331 10332 err = bpf_link_prime(&link->link, &link_primer); 10333 if (err) { 10334 kfree(link); 10335 goto unlock; 10336 } 10337 10338 netdev_lock_ops(dev); 10339 err = dev_xdp_attach_link(dev, &extack, link); 10340 netdev_unlock_ops(dev); 10341 rtnl_unlock(); 10342 10343 if (err) { 10344 link->dev = NULL; 10345 bpf_link_cleanup(&link_primer); 10346 trace_bpf_xdp_link_attach_failed(extack._msg); 10347 goto out_put_dev; 10348 } 10349 10350 fd = bpf_link_settle(&link_primer); 10351 /* link itself doesn't hold dev's refcnt to not complicate shutdown */ 10352 dev_put(dev); 10353 return fd; 10354 10355 unlock: 10356 rtnl_unlock(); 10357 10358 out_put_dev: 10359 dev_put(dev); 10360 return err; 10361 } 10362 10363 /** 10364 * dev_change_xdp_fd - set or clear a bpf program for a device rx path 10365 * @dev: device 10366 * @extack: netlink extended ack 10367 * @fd: new program fd or negative value to clear 10368 * @expected_fd: old program fd that userspace expects to replace or clear 10369 * @flags: xdp-related flags 10370 * 10371 * Set or clear a bpf program for a device 10372 */ 10373 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, 10374 int fd, int expected_fd, u32 flags) 10375 { 10376 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags); 10377 struct bpf_prog *new_prog = NULL, *old_prog = NULL; 10378 int err; 10379 10380 ASSERT_RTNL(); 10381 10382 if (fd >= 0) { 10383 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP, 10384 mode != XDP_MODE_SKB); 10385 if (IS_ERR(new_prog)) 10386 return PTR_ERR(new_prog); 10387 } 10388 10389 if (expected_fd >= 0) { 10390 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP, 10391 mode != XDP_MODE_SKB); 10392 if (IS_ERR(old_prog)) { 10393 err = PTR_ERR(old_prog); 10394 old_prog = NULL; 10395 goto err_out; 10396 } 10397 } 10398 10399 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags); 10400 10401 err_out: 10402 if (err && new_prog) 10403 bpf_prog_put(new_prog); 10404 if (old_prog) 10405 bpf_prog_put(old_prog); 10406 return err; 10407 } 10408 10409 u32 dev_get_min_mp_channel_count(const struct net_device *dev) 10410 { 10411 int i; 10412 10413 netdev_ops_assert_locked(dev); 10414 10415 for (i = dev->real_num_rx_queues - 1; i >= 0; i--) 10416 if (dev->_rx[i].mp_params.mp_priv) 10417 /* The channel count is the idx plus 1. */ 10418 return i + 1; 10419 10420 return 0; 10421 } 10422 10423 /** 10424 * dev_index_reserve() - allocate an ifindex in a namespace 10425 * @net: the applicable net namespace 10426 * @ifindex: requested ifindex, pass %0 to get one allocated 10427 * 10428 * Allocate a ifindex for a new device. Caller must either use the ifindex 10429 * to store the device (via list_netdevice()) or call dev_index_release() 10430 * to give the index up. 10431 * 10432 * Return: a suitable unique value for a new device interface number or -errno. 10433 */ 10434 static int dev_index_reserve(struct net *net, u32 ifindex) 10435 { 10436 int err; 10437 10438 if (ifindex > INT_MAX) { 10439 DEBUG_NET_WARN_ON_ONCE(1); 10440 return -EINVAL; 10441 } 10442 10443 if (!ifindex) 10444 err = xa_alloc_cyclic(&net->dev_by_index, &ifindex, NULL, 10445 xa_limit_31b, &net->ifindex, GFP_KERNEL); 10446 else 10447 err = xa_insert(&net->dev_by_index, ifindex, NULL, GFP_KERNEL); 10448 if (err < 0) 10449 return err; 10450 10451 return ifindex; 10452 } 10453 10454 static void dev_index_release(struct net *net, int ifindex) 10455 { 10456 /* Expect only unused indexes, unlist_netdevice() removes the used */ 10457 WARN_ON(xa_erase(&net->dev_by_index, ifindex)); 10458 } 10459 10460 static bool from_cleanup_net(void) 10461 { 10462 #ifdef CONFIG_NET_NS 10463 return current == cleanup_net_task; 10464 #else 10465 return false; 10466 #endif 10467 } 10468 10469 /* Delayed registration/unregisteration */ 10470 LIST_HEAD(net_todo_list); 10471 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq); 10472 atomic_t dev_unreg_count = ATOMIC_INIT(0); 10473 10474 static void net_set_todo(struct net_device *dev) 10475 { 10476 list_add_tail(&dev->todo_list, &net_todo_list); 10477 } 10478 10479 static netdev_features_t netdev_sync_upper_features(struct net_device *lower, 10480 struct net_device *upper, netdev_features_t features) 10481 { 10482 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10483 netdev_features_t feature; 10484 int feature_bit; 10485 10486 for_each_netdev_feature(upper_disables, feature_bit) { 10487 feature = __NETIF_F_BIT(feature_bit); 10488 if (!(upper->wanted_features & feature) 10489 && (features & feature)) { 10490 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n", 10491 &feature, upper->name); 10492 features &= ~feature; 10493 } 10494 } 10495 10496 return features; 10497 } 10498 10499 static void netdev_sync_lower_features(struct net_device *upper, 10500 struct net_device *lower, netdev_features_t features) 10501 { 10502 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10503 netdev_features_t feature; 10504 int feature_bit; 10505 10506 for_each_netdev_feature(upper_disables, feature_bit) { 10507 feature = __NETIF_F_BIT(feature_bit); 10508 if (!(features & feature) && (lower->features & feature)) { 10509 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n", 10510 &feature, lower->name); 10511 lower->wanted_features &= ~feature; 10512 __netdev_update_features(lower); 10513 10514 if (unlikely(lower->features & feature)) 10515 netdev_WARN(upper, "failed to disable %pNF on %s!\n", 10516 &feature, lower->name); 10517 else 10518 netdev_features_change(lower); 10519 } 10520 } 10521 } 10522 10523 static bool netdev_has_ip_or_hw_csum(netdev_features_t features) 10524 { 10525 netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 10526 bool ip_csum = (features & ip_csum_mask) == ip_csum_mask; 10527 bool hw_csum = features & NETIF_F_HW_CSUM; 10528 10529 return ip_csum || hw_csum; 10530 } 10531 10532 static netdev_features_t netdev_fix_features(struct net_device *dev, 10533 netdev_features_t features) 10534 { 10535 /* Fix illegal checksum combinations */ 10536 if ((features & NETIF_F_HW_CSUM) && 10537 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) { 10538 netdev_warn(dev, "mixed HW and IP checksum settings.\n"); 10539 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM); 10540 } 10541 10542 /* TSO requires that SG is present as well. */ 10543 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) { 10544 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n"); 10545 features &= ~NETIF_F_ALL_TSO; 10546 } 10547 10548 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) && 10549 !(features & NETIF_F_IP_CSUM)) { 10550 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n"); 10551 features &= ~NETIF_F_TSO; 10552 features &= ~NETIF_F_TSO_ECN; 10553 } 10554 10555 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) && 10556 !(features & NETIF_F_IPV6_CSUM)) { 10557 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n"); 10558 features &= ~NETIF_F_TSO6; 10559 } 10560 10561 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */ 10562 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO)) 10563 features &= ~NETIF_F_TSO_MANGLEID; 10564 10565 /* TSO ECN requires that TSO is present as well. */ 10566 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN) 10567 features &= ~NETIF_F_TSO_ECN; 10568 10569 /* Software GSO depends on SG. */ 10570 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) { 10571 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n"); 10572 features &= ~NETIF_F_GSO; 10573 } 10574 10575 /* GSO partial features require GSO partial be set */ 10576 if ((features & dev->gso_partial_features) && 10577 !(features & NETIF_F_GSO_PARTIAL)) { 10578 netdev_dbg(dev, 10579 "Dropping partially supported GSO features since no GSO partial.\n"); 10580 features &= ~dev->gso_partial_features; 10581 } 10582 10583 if (!(features & NETIF_F_RXCSUM)) { 10584 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet 10585 * successfully merged by hardware must also have the 10586 * checksum verified by hardware. If the user does not 10587 * want to enable RXCSUM, logically, we should disable GRO_HW. 10588 */ 10589 if (features & NETIF_F_GRO_HW) { 10590 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n"); 10591 features &= ~NETIF_F_GRO_HW; 10592 } 10593 } 10594 10595 /* LRO/HW-GRO features cannot be combined with RX-FCS */ 10596 if (features & NETIF_F_RXFCS) { 10597 if (features & NETIF_F_LRO) { 10598 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n"); 10599 features &= ~NETIF_F_LRO; 10600 } 10601 10602 if (features & NETIF_F_GRO_HW) { 10603 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n"); 10604 features &= ~NETIF_F_GRO_HW; 10605 } 10606 } 10607 10608 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) { 10609 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n"); 10610 features &= ~NETIF_F_LRO; 10611 } 10612 10613 if ((features & NETIF_F_HW_TLS_TX) && !netdev_has_ip_or_hw_csum(features)) { 10614 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n"); 10615 features &= ~NETIF_F_HW_TLS_TX; 10616 } 10617 10618 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) { 10619 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n"); 10620 features &= ~NETIF_F_HW_TLS_RX; 10621 } 10622 10623 if ((features & NETIF_F_GSO_UDP_L4) && !netdev_has_ip_or_hw_csum(features)) { 10624 netdev_dbg(dev, "Dropping USO feature since no CSUM feature.\n"); 10625 features &= ~NETIF_F_GSO_UDP_L4; 10626 } 10627 10628 return features; 10629 } 10630 10631 int __netdev_update_features(struct net_device *dev) 10632 { 10633 struct net_device *upper, *lower; 10634 netdev_features_t features; 10635 struct list_head *iter; 10636 int err = -1; 10637 10638 ASSERT_RTNL(); 10639 netdev_ops_assert_locked(dev); 10640 10641 features = netdev_get_wanted_features(dev); 10642 10643 if (dev->netdev_ops->ndo_fix_features) 10644 features = dev->netdev_ops->ndo_fix_features(dev, features); 10645 10646 /* driver might be less strict about feature dependencies */ 10647 features = netdev_fix_features(dev, features); 10648 10649 /* some features can't be enabled if they're off on an upper device */ 10650 netdev_for_each_upper_dev_rcu(dev, upper, iter) 10651 features = netdev_sync_upper_features(dev, upper, features); 10652 10653 if (dev->features == features) 10654 goto sync_lower; 10655 10656 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n", 10657 &dev->features, &features); 10658 10659 if (dev->netdev_ops->ndo_set_features) 10660 err = dev->netdev_ops->ndo_set_features(dev, features); 10661 else 10662 err = 0; 10663 10664 if (unlikely(err < 0)) { 10665 netdev_err(dev, 10666 "set_features() failed (%d); wanted %pNF, left %pNF\n", 10667 err, &features, &dev->features); 10668 /* return non-0 since some features might have changed and 10669 * it's better to fire a spurious notification than miss it 10670 */ 10671 return -1; 10672 } 10673 10674 sync_lower: 10675 /* some features must be disabled on lower devices when disabled 10676 * on an upper device (think: bonding master or bridge) 10677 */ 10678 netdev_for_each_lower_dev(dev, lower, iter) 10679 netdev_sync_lower_features(dev, lower, features); 10680 10681 if (!err) { 10682 netdev_features_t diff = features ^ dev->features; 10683 10684 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) { 10685 /* udp_tunnel_{get,drop}_rx_info both need 10686 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the 10687 * device, or they won't do anything. 10688 * Thus we need to update dev->features 10689 * *before* calling udp_tunnel_get_rx_info, 10690 * but *after* calling udp_tunnel_drop_rx_info. 10691 */ 10692 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) { 10693 dev->features = features; 10694 udp_tunnel_get_rx_info(dev); 10695 } else { 10696 udp_tunnel_drop_rx_info(dev); 10697 } 10698 } 10699 10700 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) { 10701 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) { 10702 dev->features = features; 10703 err |= vlan_get_rx_ctag_filter_info(dev); 10704 } else { 10705 vlan_drop_rx_ctag_filter_info(dev); 10706 } 10707 } 10708 10709 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) { 10710 if (features & NETIF_F_HW_VLAN_STAG_FILTER) { 10711 dev->features = features; 10712 err |= vlan_get_rx_stag_filter_info(dev); 10713 } else { 10714 vlan_drop_rx_stag_filter_info(dev); 10715 } 10716 } 10717 10718 dev->features = features; 10719 } 10720 10721 return err < 0 ? 0 : 1; 10722 } 10723 10724 /** 10725 * netdev_update_features - recalculate device features 10726 * @dev: the device to check 10727 * 10728 * Recalculate dev->features set and send notifications if it 10729 * has changed. Should be called after driver or hardware dependent 10730 * conditions might have changed that influence the features. 10731 */ 10732 void netdev_update_features(struct net_device *dev) 10733 { 10734 if (__netdev_update_features(dev)) 10735 netdev_features_change(dev); 10736 } 10737 EXPORT_SYMBOL(netdev_update_features); 10738 10739 /** 10740 * netdev_change_features - recalculate device features 10741 * @dev: the device to check 10742 * 10743 * Recalculate dev->features set and send notifications even 10744 * if they have not changed. Should be called instead of 10745 * netdev_update_features() if also dev->vlan_features might 10746 * have changed to allow the changes to be propagated to stacked 10747 * VLAN devices. 10748 */ 10749 void netdev_change_features(struct net_device *dev) 10750 { 10751 __netdev_update_features(dev); 10752 netdev_features_change(dev); 10753 } 10754 EXPORT_SYMBOL(netdev_change_features); 10755 10756 /** 10757 * netif_stacked_transfer_operstate - transfer operstate 10758 * @rootdev: the root or lower level device to transfer state from 10759 * @dev: the device to transfer operstate to 10760 * 10761 * Transfer operational state from root to device. This is normally 10762 * called when a stacking relationship exists between the root 10763 * device and the device(a leaf device). 10764 */ 10765 void netif_stacked_transfer_operstate(const struct net_device *rootdev, 10766 struct net_device *dev) 10767 { 10768 if (rootdev->operstate == IF_OPER_DORMANT) 10769 netif_dormant_on(dev); 10770 else 10771 netif_dormant_off(dev); 10772 10773 if (rootdev->operstate == IF_OPER_TESTING) 10774 netif_testing_on(dev); 10775 else 10776 netif_testing_off(dev); 10777 10778 if (netif_carrier_ok(rootdev)) 10779 netif_carrier_on(dev); 10780 else 10781 netif_carrier_off(dev); 10782 } 10783 EXPORT_SYMBOL(netif_stacked_transfer_operstate); 10784 10785 static int netif_alloc_rx_queues(struct net_device *dev) 10786 { 10787 unsigned int i, count = dev->num_rx_queues; 10788 struct netdev_rx_queue *rx; 10789 size_t sz = count * sizeof(*rx); 10790 int err = 0; 10791 10792 BUG_ON(count < 1); 10793 10794 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10795 if (!rx) 10796 return -ENOMEM; 10797 10798 dev->_rx = rx; 10799 10800 for (i = 0; i < count; i++) { 10801 rx[i].dev = dev; 10802 10803 /* XDP RX-queue setup */ 10804 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0); 10805 if (err < 0) 10806 goto err_rxq_info; 10807 } 10808 return 0; 10809 10810 err_rxq_info: 10811 /* Rollback successful reg's and free other resources */ 10812 while (i--) 10813 xdp_rxq_info_unreg(&rx[i].xdp_rxq); 10814 kvfree(dev->_rx); 10815 dev->_rx = NULL; 10816 return err; 10817 } 10818 10819 static void netif_free_rx_queues(struct net_device *dev) 10820 { 10821 unsigned int i, count = dev->num_rx_queues; 10822 10823 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */ 10824 if (!dev->_rx) 10825 return; 10826 10827 for (i = 0; i < count; i++) 10828 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq); 10829 10830 kvfree(dev->_rx); 10831 } 10832 10833 static void netdev_init_one_queue(struct net_device *dev, 10834 struct netdev_queue *queue, void *_unused) 10835 { 10836 /* Initialize queue lock */ 10837 spin_lock_init(&queue->_xmit_lock); 10838 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type); 10839 queue->xmit_lock_owner = -1; 10840 netdev_queue_numa_node_write(queue, NUMA_NO_NODE); 10841 queue->dev = dev; 10842 #ifdef CONFIG_BQL 10843 dql_init(&queue->dql, HZ); 10844 #endif 10845 } 10846 10847 static void netif_free_tx_queues(struct net_device *dev) 10848 { 10849 kvfree(dev->_tx); 10850 } 10851 10852 static int netif_alloc_netdev_queues(struct net_device *dev) 10853 { 10854 unsigned int count = dev->num_tx_queues; 10855 struct netdev_queue *tx; 10856 size_t sz = count * sizeof(*tx); 10857 10858 if (count < 1 || count > 0xffff) 10859 return -EINVAL; 10860 10861 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10862 if (!tx) 10863 return -ENOMEM; 10864 10865 dev->_tx = tx; 10866 10867 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL); 10868 spin_lock_init(&dev->tx_global_lock); 10869 10870 return 0; 10871 } 10872 10873 void netif_tx_stop_all_queues(struct net_device *dev) 10874 { 10875 unsigned int i; 10876 10877 for (i = 0; i < dev->num_tx_queues; i++) { 10878 struct netdev_queue *txq = netdev_get_tx_queue(dev, i); 10879 10880 netif_tx_stop_queue(txq); 10881 } 10882 } 10883 EXPORT_SYMBOL(netif_tx_stop_all_queues); 10884 10885 static int netdev_do_alloc_pcpu_stats(struct net_device *dev) 10886 { 10887 void __percpu *v; 10888 10889 /* Drivers implementing ndo_get_peer_dev must support tstat 10890 * accounting, so that skb_do_redirect() can bump the dev's 10891 * RX stats upon network namespace switch. 10892 */ 10893 if (dev->netdev_ops->ndo_get_peer_dev && 10894 dev->pcpu_stat_type != NETDEV_PCPU_STAT_TSTATS) 10895 return -EOPNOTSUPP; 10896 10897 switch (dev->pcpu_stat_type) { 10898 case NETDEV_PCPU_STAT_NONE: 10899 return 0; 10900 case NETDEV_PCPU_STAT_LSTATS: 10901 v = dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats); 10902 break; 10903 case NETDEV_PCPU_STAT_TSTATS: 10904 v = dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats); 10905 break; 10906 case NETDEV_PCPU_STAT_DSTATS: 10907 v = dev->dstats = netdev_alloc_pcpu_stats(struct pcpu_dstats); 10908 break; 10909 default: 10910 return -EINVAL; 10911 } 10912 10913 return v ? 0 : -ENOMEM; 10914 } 10915 10916 static void netdev_do_free_pcpu_stats(struct net_device *dev) 10917 { 10918 switch (dev->pcpu_stat_type) { 10919 case NETDEV_PCPU_STAT_NONE: 10920 return; 10921 case NETDEV_PCPU_STAT_LSTATS: 10922 free_percpu(dev->lstats); 10923 break; 10924 case NETDEV_PCPU_STAT_TSTATS: 10925 free_percpu(dev->tstats); 10926 break; 10927 case NETDEV_PCPU_STAT_DSTATS: 10928 free_percpu(dev->dstats); 10929 break; 10930 } 10931 } 10932 10933 static void netdev_free_phy_link_topology(struct net_device *dev) 10934 { 10935 struct phy_link_topology *topo = dev->link_topo; 10936 10937 if (IS_ENABLED(CONFIG_PHYLIB) && topo) { 10938 xa_destroy(&topo->phys); 10939 kfree(topo); 10940 dev->link_topo = NULL; 10941 } 10942 } 10943 10944 /** 10945 * register_netdevice() - register a network device 10946 * @dev: device to register 10947 * 10948 * Take a prepared network device structure and make it externally accessible. 10949 * A %NETDEV_REGISTER message is sent to the netdev notifier chain. 10950 * Callers must hold the rtnl lock - you may want register_netdev() 10951 * instead of this. 10952 */ 10953 int register_netdevice(struct net_device *dev) 10954 { 10955 int ret; 10956 struct net *net = dev_net(dev); 10957 10958 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE < 10959 NETDEV_FEATURE_COUNT); 10960 BUG_ON(dev_boot_phase); 10961 ASSERT_RTNL(); 10962 10963 might_sleep(); 10964 10965 /* When net_device's are persistent, this will be fatal. */ 10966 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED); 10967 BUG_ON(!net); 10968 10969 ret = ethtool_check_ops(dev->ethtool_ops); 10970 if (ret) 10971 return ret; 10972 10973 /* rss ctx ID 0 is reserved for the default context, start from 1 */ 10974 xa_init_flags(&dev->ethtool->rss_ctx, XA_FLAGS_ALLOC1); 10975 mutex_init(&dev->ethtool->rss_lock); 10976 10977 spin_lock_init(&dev->addr_list_lock); 10978 netdev_set_addr_lockdep_class(dev); 10979 10980 ret = dev_get_valid_name(net, dev, dev->name); 10981 if (ret < 0) 10982 goto out; 10983 10984 ret = -ENOMEM; 10985 dev->name_node = netdev_name_node_head_alloc(dev); 10986 if (!dev->name_node) 10987 goto out; 10988 10989 /* Init, if this function is available */ 10990 if (dev->netdev_ops->ndo_init) { 10991 ret = dev->netdev_ops->ndo_init(dev); 10992 if (ret) { 10993 if (ret > 0) 10994 ret = -EIO; 10995 goto err_free_name; 10996 } 10997 } 10998 10999 if (((dev->hw_features | dev->features) & 11000 NETIF_F_HW_VLAN_CTAG_FILTER) && 11001 (!dev->netdev_ops->ndo_vlan_rx_add_vid || 11002 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) { 11003 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n"); 11004 ret = -EINVAL; 11005 goto err_uninit; 11006 } 11007 11008 ret = netdev_do_alloc_pcpu_stats(dev); 11009 if (ret) 11010 goto err_uninit; 11011 11012 ret = dev_index_reserve(net, dev->ifindex); 11013 if (ret < 0) 11014 goto err_free_pcpu; 11015 dev->ifindex = ret; 11016 11017 /* Transfer changeable features to wanted_features and enable 11018 * software offloads (GSO and GRO). 11019 */ 11020 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF); 11021 dev->features |= NETIF_F_SOFT_FEATURES; 11022 11023 if (dev->udp_tunnel_nic_info) { 11024 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT; 11025 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT; 11026 } 11027 11028 dev->wanted_features = dev->features & dev->hw_features; 11029 11030 if (!(dev->flags & IFF_LOOPBACK)) 11031 dev->hw_features |= NETIF_F_NOCACHE_COPY; 11032 11033 /* If IPv4 TCP segmentation offload is supported we should also 11034 * allow the device to enable segmenting the frame with the option 11035 * of ignoring a static IP ID value. This doesn't enable the 11036 * feature itself but allows the user to enable it later. 11037 */ 11038 if (dev->hw_features & NETIF_F_TSO) 11039 dev->hw_features |= NETIF_F_TSO_MANGLEID; 11040 if (dev->vlan_features & NETIF_F_TSO) 11041 dev->vlan_features |= NETIF_F_TSO_MANGLEID; 11042 if (dev->mpls_features & NETIF_F_TSO) 11043 dev->mpls_features |= NETIF_F_TSO_MANGLEID; 11044 if (dev->hw_enc_features & NETIF_F_TSO) 11045 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID; 11046 11047 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices. 11048 */ 11049 dev->vlan_features |= NETIF_F_HIGHDMA; 11050 11051 /* Make NETIF_F_SG inheritable to tunnel devices. 11052 */ 11053 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL; 11054 11055 /* Make NETIF_F_SG inheritable to MPLS. 11056 */ 11057 dev->mpls_features |= NETIF_F_SG; 11058 11059 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev); 11060 ret = notifier_to_errno(ret); 11061 if (ret) 11062 goto err_ifindex_release; 11063 11064 ret = netdev_register_kobject(dev); 11065 11066 netdev_lock(dev); 11067 WRITE_ONCE(dev->reg_state, ret ? NETREG_UNREGISTERED : NETREG_REGISTERED); 11068 netdev_unlock(dev); 11069 11070 if (ret) 11071 goto err_uninit_notify; 11072 11073 netdev_lock_ops(dev); 11074 __netdev_update_features(dev); 11075 netdev_unlock_ops(dev); 11076 11077 /* 11078 * Default initial state at registry is that the 11079 * device is present. 11080 */ 11081 11082 set_bit(__LINK_STATE_PRESENT, &dev->state); 11083 11084 linkwatch_init_dev(dev); 11085 11086 dev_init_scheduler(dev); 11087 11088 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL); 11089 list_netdevice(dev); 11090 11091 add_device_randomness(dev->dev_addr, dev->addr_len); 11092 11093 /* If the device has permanent device address, driver should 11094 * set dev_addr and also addr_assign_type should be set to 11095 * NET_ADDR_PERM (default value). 11096 */ 11097 if (dev->addr_assign_type == NET_ADDR_PERM) 11098 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len); 11099 11100 /* Notify protocols, that a new device appeared. */ 11101 netdev_lock_ops(dev); 11102 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev); 11103 netdev_unlock_ops(dev); 11104 ret = notifier_to_errno(ret); 11105 if (ret) { 11106 /* Expect explicit free_netdev() on failure */ 11107 dev->needs_free_netdev = false; 11108 unregister_netdevice_queue(dev, NULL); 11109 goto out; 11110 } 11111 /* 11112 * Prevent userspace races by waiting until the network 11113 * device is fully setup before sending notifications. 11114 */ 11115 if (!(dev->rtnl_link_ops && dev->rtnl_link_initializing)) 11116 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 11117 11118 out: 11119 return ret; 11120 11121 err_uninit_notify: 11122 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11123 err_ifindex_release: 11124 dev_index_release(net, dev->ifindex); 11125 err_free_pcpu: 11126 netdev_do_free_pcpu_stats(dev); 11127 err_uninit: 11128 if (dev->netdev_ops->ndo_uninit) 11129 dev->netdev_ops->ndo_uninit(dev); 11130 if (dev->priv_destructor) 11131 dev->priv_destructor(dev); 11132 err_free_name: 11133 netdev_name_node_free(dev->name_node); 11134 goto out; 11135 } 11136 EXPORT_SYMBOL(register_netdevice); 11137 11138 /* Initialize the core of a dummy net device. 11139 * The setup steps dummy netdevs need which normal netdevs get by going 11140 * through register_netdevice(). 11141 */ 11142 static void init_dummy_netdev(struct net_device *dev) 11143 { 11144 /* make sure we BUG if trying to hit standard 11145 * register/unregister code path 11146 */ 11147 dev->reg_state = NETREG_DUMMY; 11148 11149 /* a dummy interface is started by default */ 11150 set_bit(__LINK_STATE_PRESENT, &dev->state); 11151 set_bit(__LINK_STATE_START, &dev->state); 11152 11153 /* Note : We dont allocate pcpu_refcnt for dummy devices, 11154 * because users of this 'device' dont need to change 11155 * its refcount. 11156 */ 11157 } 11158 11159 /** 11160 * register_netdev - register a network device 11161 * @dev: device to register 11162 * 11163 * Take a completed network device structure and add it to the kernel 11164 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier 11165 * chain. 0 is returned on success. A negative errno code is returned 11166 * on a failure to set up the device, or if the name is a duplicate. 11167 * 11168 * This is a wrapper around register_netdevice that takes the rtnl semaphore 11169 * and expands the device name if you passed a format string to 11170 * alloc_netdev. 11171 */ 11172 int register_netdev(struct net_device *dev) 11173 { 11174 struct net *net = dev_net(dev); 11175 int err; 11176 11177 if (rtnl_net_lock_killable(net)) 11178 return -EINTR; 11179 11180 err = register_netdevice(dev); 11181 11182 rtnl_net_unlock(net); 11183 11184 return err; 11185 } 11186 EXPORT_SYMBOL(register_netdev); 11187 11188 int netdev_refcnt_read(const struct net_device *dev) 11189 { 11190 #ifdef CONFIG_PCPU_DEV_REFCNT 11191 int i, refcnt = 0; 11192 11193 for_each_possible_cpu(i) 11194 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i); 11195 return refcnt; 11196 #else 11197 return refcount_read(&dev->dev_refcnt); 11198 #endif 11199 } 11200 EXPORT_SYMBOL(netdev_refcnt_read); 11201 11202 int netdev_unregister_timeout_secs __read_mostly = 10; 11203 11204 #define WAIT_REFS_MIN_MSECS 1 11205 #define WAIT_REFS_MAX_MSECS 250 11206 /** 11207 * netdev_wait_allrefs_any - wait until all references are gone. 11208 * @list: list of net_devices to wait on 11209 * 11210 * This is called when unregistering network devices. 11211 * 11212 * Any protocol or device that holds a reference should register 11213 * for netdevice notification, and cleanup and put back the 11214 * reference if they receive an UNREGISTER event. 11215 * We can get stuck here if buggy protocols don't correctly 11216 * call dev_put. 11217 */ 11218 static struct net_device *netdev_wait_allrefs_any(struct list_head *list) 11219 { 11220 unsigned long rebroadcast_time, warning_time; 11221 struct net_device *dev; 11222 int wait = 0; 11223 11224 rebroadcast_time = warning_time = jiffies; 11225 11226 list_for_each_entry(dev, list, todo_list) 11227 if (netdev_refcnt_read(dev) == 1) 11228 return dev; 11229 11230 while (true) { 11231 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) { 11232 rtnl_lock(); 11233 11234 /* Rebroadcast unregister notification */ 11235 list_for_each_entry(dev, list, todo_list) 11236 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11237 11238 __rtnl_unlock(); 11239 rcu_barrier(); 11240 rtnl_lock(); 11241 11242 list_for_each_entry(dev, list, todo_list) 11243 if (test_bit(__LINK_STATE_LINKWATCH_PENDING, 11244 &dev->state)) { 11245 /* We must not have linkwatch events 11246 * pending on unregister. If this 11247 * happens, we simply run the queue 11248 * unscheduled, resulting in a noop 11249 * for this device. 11250 */ 11251 linkwatch_run_queue(); 11252 break; 11253 } 11254 11255 __rtnl_unlock(); 11256 11257 rebroadcast_time = jiffies; 11258 } 11259 11260 rcu_barrier(); 11261 11262 if (!wait) { 11263 wait = WAIT_REFS_MIN_MSECS; 11264 } else { 11265 msleep(wait); 11266 wait = min(wait << 1, WAIT_REFS_MAX_MSECS); 11267 } 11268 11269 list_for_each_entry(dev, list, todo_list) 11270 if (netdev_refcnt_read(dev) == 1) 11271 return dev; 11272 11273 if (time_after(jiffies, warning_time + 11274 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) { 11275 list_for_each_entry(dev, list, todo_list) { 11276 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n", 11277 dev->name, netdev_refcnt_read(dev)); 11278 ref_tracker_dir_print(&dev->refcnt_tracker, 10); 11279 } 11280 11281 warning_time = jiffies; 11282 } 11283 } 11284 } 11285 11286 /* The sequence is: 11287 * 11288 * rtnl_lock(); 11289 * ... 11290 * register_netdevice(x1); 11291 * register_netdevice(x2); 11292 * ... 11293 * unregister_netdevice(y1); 11294 * unregister_netdevice(y2); 11295 * ... 11296 * rtnl_unlock(); 11297 * free_netdev(y1); 11298 * free_netdev(y2); 11299 * 11300 * We are invoked by rtnl_unlock(). 11301 * This allows us to deal with problems: 11302 * 1) We can delete sysfs objects which invoke hotplug 11303 * without deadlocking with linkwatch via keventd. 11304 * 2) Since we run with the RTNL semaphore not held, we can sleep 11305 * safely in order to wait for the netdev refcnt to drop to zero. 11306 * 11307 * We must not return until all unregister events added during 11308 * the interval the lock was held have been completed. 11309 */ 11310 void netdev_run_todo(void) 11311 { 11312 struct net_device *dev, *tmp; 11313 struct list_head list; 11314 int cnt; 11315 #ifdef CONFIG_LOCKDEP 11316 struct list_head unlink_list; 11317 11318 list_replace_init(&net_unlink_list, &unlink_list); 11319 11320 while (!list_empty(&unlink_list)) { 11321 dev = list_first_entry(&unlink_list, struct net_device, 11322 unlink_list); 11323 list_del_init(&dev->unlink_list); 11324 dev->nested_level = dev->lower_level - 1; 11325 } 11326 #endif 11327 11328 /* Snapshot list, allow later requests */ 11329 list_replace_init(&net_todo_list, &list); 11330 11331 __rtnl_unlock(); 11332 11333 /* Wait for rcu callbacks to finish before next phase */ 11334 if (!list_empty(&list)) 11335 rcu_barrier(); 11336 11337 list_for_each_entry_safe(dev, tmp, &list, todo_list) { 11338 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) { 11339 netdev_WARN(dev, "run_todo but not unregistering\n"); 11340 list_del(&dev->todo_list); 11341 continue; 11342 } 11343 11344 netdev_lock(dev); 11345 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERED); 11346 netdev_unlock(dev); 11347 linkwatch_sync_dev(dev); 11348 } 11349 11350 cnt = 0; 11351 while (!list_empty(&list)) { 11352 dev = netdev_wait_allrefs_any(&list); 11353 list_del(&dev->todo_list); 11354 11355 /* paranoia */ 11356 BUG_ON(netdev_refcnt_read(dev) != 1); 11357 BUG_ON(!list_empty(&dev->ptype_all)); 11358 BUG_ON(!list_empty(&dev->ptype_specific)); 11359 WARN_ON(rcu_access_pointer(dev->ip_ptr)); 11360 WARN_ON(rcu_access_pointer(dev->ip6_ptr)); 11361 11362 netdev_do_free_pcpu_stats(dev); 11363 if (dev->priv_destructor) 11364 dev->priv_destructor(dev); 11365 if (dev->needs_free_netdev) 11366 free_netdev(dev); 11367 11368 cnt++; 11369 11370 /* Free network device */ 11371 kobject_put(&dev->dev.kobj); 11372 } 11373 if (cnt && atomic_sub_and_test(cnt, &dev_unreg_count)) 11374 wake_up(&netdev_unregistering_wq); 11375 } 11376 11377 /* Collate per-cpu network dstats statistics 11378 * 11379 * Read per-cpu network statistics from dev->dstats and populate the related 11380 * fields in @s. 11381 */ 11382 static void dev_fetch_dstats(struct rtnl_link_stats64 *s, 11383 const struct pcpu_dstats __percpu *dstats) 11384 { 11385 int cpu; 11386 11387 for_each_possible_cpu(cpu) { 11388 u64 rx_packets, rx_bytes, rx_drops; 11389 u64 tx_packets, tx_bytes, tx_drops; 11390 const struct pcpu_dstats *stats; 11391 unsigned int start; 11392 11393 stats = per_cpu_ptr(dstats, cpu); 11394 do { 11395 start = u64_stats_fetch_begin(&stats->syncp); 11396 rx_packets = u64_stats_read(&stats->rx_packets); 11397 rx_bytes = u64_stats_read(&stats->rx_bytes); 11398 rx_drops = u64_stats_read(&stats->rx_drops); 11399 tx_packets = u64_stats_read(&stats->tx_packets); 11400 tx_bytes = u64_stats_read(&stats->tx_bytes); 11401 tx_drops = u64_stats_read(&stats->tx_drops); 11402 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11403 11404 s->rx_packets += rx_packets; 11405 s->rx_bytes += rx_bytes; 11406 s->rx_dropped += rx_drops; 11407 s->tx_packets += tx_packets; 11408 s->tx_bytes += tx_bytes; 11409 s->tx_dropped += tx_drops; 11410 } 11411 } 11412 11413 /* ndo_get_stats64 implementation for dtstats-based accounting. 11414 * 11415 * Populate @s from dev->stats and dev->dstats. This is used internally by the 11416 * core for NETDEV_PCPU_STAT_DSTAT-type stats collection. 11417 */ 11418 static void dev_get_dstats64(const struct net_device *dev, 11419 struct rtnl_link_stats64 *s) 11420 { 11421 netdev_stats_to_stats64(s, &dev->stats); 11422 dev_fetch_dstats(s, dev->dstats); 11423 } 11424 11425 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has 11426 * all the same fields in the same order as net_device_stats, with only 11427 * the type differing, but rtnl_link_stats64 may have additional fields 11428 * at the end for newer counters. 11429 */ 11430 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64, 11431 const struct net_device_stats *netdev_stats) 11432 { 11433 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t); 11434 const atomic_long_t *src = (atomic_long_t *)netdev_stats; 11435 u64 *dst = (u64 *)stats64; 11436 11437 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64)); 11438 for (i = 0; i < n; i++) 11439 dst[i] = (unsigned long)atomic_long_read(&src[i]); 11440 /* zero out counters that only exist in rtnl_link_stats64 */ 11441 memset((char *)stats64 + n * sizeof(u64), 0, 11442 sizeof(*stats64) - n * sizeof(u64)); 11443 } 11444 EXPORT_SYMBOL(netdev_stats_to_stats64); 11445 11446 static __cold struct net_device_core_stats __percpu *netdev_core_stats_alloc( 11447 struct net_device *dev) 11448 { 11449 struct net_device_core_stats __percpu *p; 11450 11451 p = alloc_percpu_gfp(struct net_device_core_stats, 11452 GFP_ATOMIC | __GFP_NOWARN); 11453 11454 if (p && cmpxchg(&dev->core_stats, NULL, p)) 11455 free_percpu(p); 11456 11457 /* This READ_ONCE() pairs with the cmpxchg() above */ 11458 return READ_ONCE(dev->core_stats); 11459 } 11460 11461 noinline void netdev_core_stats_inc(struct net_device *dev, u32 offset) 11462 { 11463 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11464 struct net_device_core_stats __percpu *p = READ_ONCE(dev->core_stats); 11465 unsigned long __percpu *field; 11466 11467 if (unlikely(!p)) { 11468 p = netdev_core_stats_alloc(dev); 11469 if (!p) 11470 return; 11471 } 11472 11473 field = (unsigned long __percpu *)((void __percpu *)p + offset); 11474 this_cpu_inc(*field); 11475 } 11476 EXPORT_SYMBOL_GPL(netdev_core_stats_inc); 11477 11478 /** 11479 * dev_get_stats - get network device statistics 11480 * @dev: device to get statistics from 11481 * @storage: place to store stats 11482 * 11483 * Get network statistics from device. Return @storage. 11484 * The device driver may provide its own method by setting 11485 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats; 11486 * otherwise the internal statistics structure is used. 11487 */ 11488 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev, 11489 struct rtnl_link_stats64 *storage) 11490 { 11491 const struct net_device_ops *ops = dev->netdev_ops; 11492 const struct net_device_core_stats __percpu *p; 11493 11494 /* 11495 * IPv{4,6} and udp tunnels share common stat helpers and use 11496 * different stat type (NETDEV_PCPU_STAT_TSTATS vs 11497 * NETDEV_PCPU_STAT_DSTATS). Ensure the accounting is consistent. 11498 */ 11499 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_bytes) != 11500 offsetof(struct pcpu_dstats, rx_bytes)); 11501 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_packets) != 11502 offsetof(struct pcpu_dstats, rx_packets)); 11503 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_bytes) != 11504 offsetof(struct pcpu_dstats, tx_bytes)); 11505 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_packets) != 11506 offsetof(struct pcpu_dstats, tx_packets)); 11507 11508 if (ops->ndo_get_stats64) { 11509 memset(storage, 0, sizeof(*storage)); 11510 ops->ndo_get_stats64(dev, storage); 11511 } else if (ops->ndo_get_stats) { 11512 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev)); 11513 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_TSTATS) { 11514 dev_get_tstats64(dev, storage); 11515 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_DSTATS) { 11516 dev_get_dstats64(dev, storage); 11517 } else { 11518 netdev_stats_to_stats64(storage, &dev->stats); 11519 } 11520 11521 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11522 p = READ_ONCE(dev->core_stats); 11523 if (p) { 11524 const struct net_device_core_stats *core_stats; 11525 int i; 11526 11527 for_each_possible_cpu(i) { 11528 core_stats = per_cpu_ptr(p, i); 11529 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped); 11530 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped); 11531 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler); 11532 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped); 11533 } 11534 } 11535 return storage; 11536 } 11537 EXPORT_SYMBOL(dev_get_stats); 11538 11539 /** 11540 * dev_fetch_sw_netstats - get per-cpu network device statistics 11541 * @s: place to store stats 11542 * @netstats: per-cpu network stats to read from 11543 * 11544 * Read per-cpu network statistics and populate the related fields in @s. 11545 */ 11546 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s, 11547 const struct pcpu_sw_netstats __percpu *netstats) 11548 { 11549 int cpu; 11550 11551 for_each_possible_cpu(cpu) { 11552 u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 11553 const struct pcpu_sw_netstats *stats; 11554 unsigned int start; 11555 11556 stats = per_cpu_ptr(netstats, cpu); 11557 do { 11558 start = u64_stats_fetch_begin(&stats->syncp); 11559 rx_packets = u64_stats_read(&stats->rx_packets); 11560 rx_bytes = u64_stats_read(&stats->rx_bytes); 11561 tx_packets = u64_stats_read(&stats->tx_packets); 11562 tx_bytes = u64_stats_read(&stats->tx_bytes); 11563 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11564 11565 s->rx_packets += rx_packets; 11566 s->rx_bytes += rx_bytes; 11567 s->tx_packets += tx_packets; 11568 s->tx_bytes += tx_bytes; 11569 } 11570 } 11571 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats); 11572 11573 /** 11574 * dev_get_tstats64 - ndo_get_stats64 implementation 11575 * @dev: device to get statistics from 11576 * @s: place to store stats 11577 * 11578 * Populate @s from dev->stats and dev->tstats. Can be used as 11579 * ndo_get_stats64() callback. 11580 */ 11581 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s) 11582 { 11583 netdev_stats_to_stats64(s, &dev->stats); 11584 dev_fetch_sw_netstats(s, dev->tstats); 11585 } 11586 EXPORT_SYMBOL_GPL(dev_get_tstats64); 11587 11588 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev) 11589 { 11590 struct netdev_queue *queue = dev_ingress_queue(dev); 11591 11592 #ifdef CONFIG_NET_CLS_ACT 11593 if (queue) 11594 return queue; 11595 queue = kzalloc(sizeof(*queue), GFP_KERNEL); 11596 if (!queue) 11597 return NULL; 11598 netdev_init_one_queue(dev, queue, NULL); 11599 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc); 11600 RCU_INIT_POINTER(queue->qdisc_sleeping, &noop_qdisc); 11601 rcu_assign_pointer(dev->ingress_queue, queue); 11602 #endif 11603 return queue; 11604 } 11605 11606 static const struct ethtool_ops default_ethtool_ops; 11607 11608 void netdev_set_default_ethtool_ops(struct net_device *dev, 11609 const struct ethtool_ops *ops) 11610 { 11611 if (dev->ethtool_ops == &default_ethtool_ops) 11612 dev->ethtool_ops = ops; 11613 } 11614 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops); 11615 11616 /** 11617 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default 11618 * @dev: netdev to enable the IRQ coalescing on 11619 * 11620 * Sets a conservative default for SW IRQ coalescing. Users can use 11621 * sysfs attributes to override the default values. 11622 */ 11623 void netdev_sw_irq_coalesce_default_on(struct net_device *dev) 11624 { 11625 WARN_ON(dev->reg_state == NETREG_REGISTERED); 11626 11627 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) { 11628 netdev_set_gro_flush_timeout(dev, 20000); 11629 netdev_set_defer_hard_irqs(dev, 1); 11630 } 11631 } 11632 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on); 11633 11634 /** 11635 * alloc_netdev_mqs - allocate network device 11636 * @sizeof_priv: size of private data to allocate space for 11637 * @name: device name format string 11638 * @name_assign_type: origin of device name 11639 * @setup: callback to initialize device 11640 * @txqs: the number of TX subqueues to allocate 11641 * @rxqs: the number of RX subqueues to allocate 11642 * 11643 * Allocates a struct net_device with private data area for driver use 11644 * and performs basic initialization. Also allocates subqueue structs 11645 * for each queue on the device. 11646 */ 11647 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name, 11648 unsigned char name_assign_type, 11649 void (*setup)(struct net_device *), 11650 unsigned int txqs, unsigned int rxqs) 11651 { 11652 struct net_device *dev; 11653 size_t napi_config_sz; 11654 unsigned int maxqs; 11655 11656 BUG_ON(strlen(name) >= sizeof(dev->name)); 11657 11658 if (txqs < 1) { 11659 pr_err("alloc_netdev: Unable to allocate device with zero queues\n"); 11660 return NULL; 11661 } 11662 11663 if (rxqs < 1) { 11664 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n"); 11665 return NULL; 11666 } 11667 11668 maxqs = max(txqs, rxqs); 11669 11670 dev = kvzalloc(struct_size(dev, priv, sizeof_priv), 11671 GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11672 if (!dev) 11673 return NULL; 11674 11675 dev->priv_len = sizeof_priv; 11676 11677 ref_tracker_dir_init(&dev->refcnt_tracker, 128, name); 11678 #ifdef CONFIG_PCPU_DEV_REFCNT 11679 dev->pcpu_refcnt = alloc_percpu(int); 11680 if (!dev->pcpu_refcnt) 11681 goto free_dev; 11682 __dev_hold(dev); 11683 #else 11684 refcount_set(&dev->dev_refcnt, 1); 11685 #endif 11686 11687 if (dev_addr_init(dev)) 11688 goto free_pcpu; 11689 11690 dev_mc_init(dev); 11691 dev_uc_init(dev); 11692 11693 dev_net_set(dev, &init_net); 11694 11695 dev->gso_max_size = GSO_LEGACY_MAX_SIZE; 11696 dev->xdp_zc_max_segs = 1; 11697 dev->gso_max_segs = GSO_MAX_SEGS; 11698 dev->gro_max_size = GRO_LEGACY_MAX_SIZE; 11699 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE; 11700 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE; 11701 dev->tso_max_size = TSO_LEGACY_MAX_SIZE; 11702 dev->tso_max_segs = TSO_MAX_SEGS; 11703 dev->upper_level = 1; 11704 dev->lower_level = 1; 11705 #ifdef CONFIG_LOCKDEP 11706 dev->nested_level = 0; 11707 INIT_LIST_HEAD(&dev->unlink_list); 11708 #endif 11709 11710 INIT_LIST_HEAD(&dev->napi_list); 11711 INIT_LIST_HEAD(&dev->unreg_list); 11712 INIT_LIST_HEAD(&dev->close_list); 11713 INIT_LIST_HEAD(&dev->link_watch_list); 11714 INIT_LIST_HEAD(&dev->adj_list.upper); 11715 INIT_LIST_HEAD(&dev->adj_list.lower); 11716 INIT_LIST_HEAD(&dev->ptype_all); 11717 INIT_LIST_HEAD(&dev->ptype_specific); 11718 INIT_LIST_HEAD(&dev->net_notifier_list); 11719 #ifdef CONFIG_NET_SCHED 11720 hash_init(dev->qdisc_hash); 11721 #endif 11722 11723 mutex_init(&dev->lock); 11724 11725 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 11726 setup(dev); 11727 11728 if (!dev->tx_queue_len) { 11729 dev->priv_flags |= IFF_NO_QUEUE; 11730 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN; 11731 } 11732 11733 dev->num_tx_queues = txqs; 11734 dev->real_num_tx_queues = txqs; 11735 if (netif_alloc_netdev_queues(dev)) 11736 goto free_all; 11737 11738 dev->num_rx_queues = rxqs; 11739 dev->real_num_rx_queues = rxqs; 11740 if (netif_alloc_rx_queues(dev)) 11741 goto free_all; 11742 dev->ethtool = kzalloc(sizeof(*dev->ethtool), GFP_KERNEL_ACCOUNT); 11743 if (!dev->ethtool) 11744 goto free_all; 11745 11746 dev->cfg = kzalloc(sizeof(*dev->cfg), GFP_KERNEL_ACCOUNT); 11747 if (!dev->cfg) 11748 goto free_all; 11749 dev->cfg_pending = dev->cfg; 11750 11751 napi_config_sz = array_size(maxqs, sizeof(*dev->napi_config)); 11752 dev->napi_config = kvzalloc(napi_config_sz, GFP_KERNEL_ACCOUNT); 11753 if (!dev->napi_config) 11754 goto free_all; 11755 11756 strscpy(dev->name, name); 11757 dev->name_assign_type = name_assign_type; 11758 dev->group = INIT_NETDEV_GROUP; 11759 if (!dev->ethtool_ops) 11760 dev->ethtool_ops = &default_ethtool_ops; 11761 11762 nf_hook_netdev_init(dev); 11763 11764 return dev; 11765 11766 free_all: 11767 free_netdev(dev); 11768 return NULL; 11769 11770 free_pcpu: 11771 #ifdef CONFIG_PCPU_DEV_REFCNT 11772 free_percpu(dev->pcpu_refcnt); 11773 free_dev: 11774 #endif 11775 kvfree(dev); 11776 return NULL; 11777 } 11778 EXPORT_SYMBOL(alloc_netdev_mqs); 11779 11780 static void netdev_napi_exit(struct net_device *dev) 11781 { 11782 if (!list_empty(&dev->napi_list)) { 11783 struct napi_struct *p, *n; 11784 11785 netdev_lock(dev); 11786 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list) 11787 __netif_napi_del_locked(p); 11788 netdev_unlock(dev); 11789 11790 synchronize_net(); 11791 } 11792 11793 kvfree(dev->napi_config); 11794 } 11795 11796 /** 11797 * free_netdev - free network device 11798 * @dev: device 11799 * 11800 * This function does the last stage of destroying an allocated device 11801 * interface. The reference to the device object is released. If this 11802 * is the last reference then it will be freed.Must be called in process 11803 * context. 11804 */ 11805 void free_netdev(struct net_device *dev) 11806 { 11807 might_sleep(); 11808 11809 /* When called immediately after register_netdevice() failed the unwind 11810 * handling may still be dismantling the device. Handle that case by 11811 * deferring the free. 11812 */ 11813 if (dev->reg_state == NETREG_UNREGISTERING) { 11814 ASSERT_RTNL(); 11815 dev->needs_free_netdev = true; 11816 return; 11817 } 11818 11819 WARN_ON(dev->cfg != dev->cfg_pending); 11820 kfree(dev->cfg); 11821 kfree(dev->ethtool); 11822 netif_free_tx_queues(dev); 11823 netif_free_rx_queues(dev); 11824 11825 kfree(rcu_dereference_protected(dev->ingress_queue, 1)); 11826 11827 /* Flush device addresses */ 11828 dev_addr_flush(dev); 11829 11830 netdev_napi_exit(dev); 11831 11832 netif_del_cpu_rmap(dev); 11833 11834 ref_tracker_dir_exit(&dev->refcnt_tracker); 11835 #ifdef CONFIG_PCPU_DEV_REFCNT 11836 free_percpu(dev->pcpu_refcnt); 11837 dev->pcpu_refcnt = NULL; 11838 #endif 11839 free_percpu(dev->core_stats); 11840 dev->core_stats = NULL; 11841 free_percpu(dev->xdp_bulkq); 11842 dev->xdp_bulkq = NULL; 11843 11844 netdev_free_phy_link_topology(dev); 11845 11846 mutex_destroy(&dev->lock); 11847 11848 /* Compatibility with error handling in drivers */ 11849 if (dev->reg_state == NETREG_UNINITIALIZED || 11850 dev->reg_state == NETREG_DUMMY) { 11851 kvfree(dev); 11852 return; 11853 } 11854 11855 BUG_ON(dev->reg_state != NETREG_UNREGISTERED); 11856 WRITE_ONCE(dev->reg_state, NETREG_RELEASED); 11857 11858 /* will free via device release */ 11859 put_device(&dev->dev); 11860 } 11861 EXPORT_SYMBOL(free_netdev); 11862 11863 /** 11864 * alloc_netdev_dummy - Allocate and initialize a dummy net device. 11865 * @sizeof_priv: size of private data to allocate space for 11866 * 11867 * Return: the allocated net_device on success, NULL otherwise 11868 */ 11869 struct net_device *alloc_netdev_dummy(int sizeof_priv) 11870 { 11871 return alloc_netdev(sizeof_priv, "dummy#", NET_NAME_UNKNOWN, 11872 init_dummy_netdev); 11873 } 11874 EXPORT_SYMBOL_GPL(alloc_netdev_dummy); 11875 11876 /** 11877 * synchronize_net - Synchronize with packet receive processing 11878 * 11879 * Wait for packets currently being received to be done. 11880 * Does not block later packets from starting. 11881 */ 11882 void synchronize_net(void) 11883 { 11884 might_sleep(); 11885 if (from_cleanup_net() || rtnl_is_locked()) 11886 synchronize_rcu_expedited(); 11887 else 11888 synchronize_rcu(); 11889 } 11890 EXPORT_SYMBOL(synchronize_net); 11891 11892 static void netdev_rss_contexts_free(struct net_device *dev) 11893 { 11894 struct ethtool_rxfh_context *ctx; 11895 unsigned long context; 11896 11897 mutex_lock(&dev->ethtool->rss_lock); 11898 xa_for_each(&dev->ethtool->rss_ctx, context, ctx) { 11899 struct ethtool_rxfh_param rxfh; 11900 11901 rxfh.indir = ethtool_rxfh_context_indir(ctx); 11902 rxfh.key = ethtool_rxfh_context_key(ctx); 11903 rxfh.hfunc = ctx->hfunc; 11904 rxfh.input_xfrm = ctx->input_xfrm; 11905 rxfh.rss_context = context; 11906 rxfh.rss_delete = true; 11907 11908 xa_erase(&dev->ethtool->rss_ctx, context); 11909 if (dev->ethtool_ops->create_rxfh_context) 11910 dev->ethtool_ops->remove_rxfh_context(dev, ctx, 11911 context, NULL); 11912 else 11913 dev->ethtool_ops->set_rxfh(dev, &rxfh, NULL); 11914 kfree(ctx); 11915 } 11916 xa_destroy(&dev->ethtool->rss_ctx); 11917 mutex_unlock(&dev->ethtool->rss_lock); 11918 } 11919 11920 /** 11921 * unregister_netdevice_queue - remove device from the kernel 11922 * @dev: device 11923 * @head: list 11924 * 11925 * This function shuts down a device interface and removes it 11926 * from the kernel tables. 11927 * If head not NULL, device is queued to be unregistered later. 11928 * 11929 * Callers must hold the rtnl semaphore. You may want 11930 * unregister_netdev() instead of this. 11931 */ 11932 11933 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head) 11934 { 11935 ASSERT_RTNL(); 11936 11937 if (head) { 11938 list_move_tail(&dev->unreg_list, head); 11939 } else { 11940 LIST_HEAD(single); 11941 11942 list_add(&dev->unreg_list, &single); 11943 unregister_netdevice_many(&single); 11944 } 11945 } 11946 EXPORT_SYMBOL(unregister_netdevice_queue); 11947 11948 static void dev_memory_provider_uninstall(struct net_device *dev) 11949 { 11950 unsigned int i; 11951 11952 for (i = 0; i < dev->real_num_rx_queues; i++) { 11953 struct netdev_rx_queue *rxq = &dev->_rx[i]; 11954 struct pp_memory_provider_params *p = &rxq->mp_params; 11955 11956 if (p->mp_ops && p->mp_ops->uninstall) 11957 p->mp_ops->uninstall(rxq->mp_params.mp_priv, rxq); 11958 } 11959 } 11960 11961 void unregister_netdevice_many_notify(struct list_head *head, 11962 u32 portid, const struct nlmsghdr *nlh) 11963 { 11964 struct net_device *dev, *tmp; 11965 LIST_HEAD(close_head); 11966 int cnt = 0; 11967 11968 BUG_ON(dev_boot_phase); 11969 ASSERT_RTNL(); 11970 11971 if (list_empty(head)) 11972 return; 11973 11974 list_for_each_entry_safe(dev, tmp, head, unreg_list) { 11975 /* Some devices call without registering 11976 * for initialization unwind. Remove those 11977 * devices and proceed with the remaining. 11978 */ 11979 if (dev->reg_state == NETREG_UNINITIALIZED) { 11980 pr_debug("unregister_netdevice: device %s/%p never was registered\n", 11981 dev->name, dev); 11982 11983 WARN_ON(1); 11984 list_del(&dev->unreg_list); 11985 continue; 11986 } 11987 dev->dismantle = true; 11988 BUG_ON(dev->reg_state != NETREG_REGISTERED); 11989 } 11990 11991 /* If device is running, close it first. Start with ops locked... */ 11992 list_for_each_entry(dev, head, unreg_list) { 11993 if (netdev_need_ops_lock(dev)) { 11994 list_add_tail(&dev->close_list, &close_head); 11995 netdev_lock(dev); 11996 } 11997 } 11998 dev_close_many(&close_head, true); 11999 /* ... now unlock them and go over the rest. */ 12000 list_for_each_entry(dev, head, unreg_list) { 12001 if (netdev_need_ops_lock(dev)) 12002 netdev_unlock(dev); 12003 else 12004 list_add_tail(&dev->close_list, &close_head); 12005 } 12006 dev_close_many(&close_head, true); 12007 12008 list_for_each_entry(dev, head, unreg_list) { 12009 /* And unlink it from device chain. */ 12010 unlist_netdevice(dev); 12011 netdev_lock(dev); 12012 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERING); 12013 netdev_unlock(dev); 12014 } 12015 flush_all_backlogs(); 12016 12017 synchronize_net(); 12018 12019 list_for_each_entry(dev, head, unreg_list) { 12020 struct sk_buff *skb = NULL; 12021 12022 /* Shutdown queueing discipline. */ 12023 dev_shutdown(dev); 12024 dev_tcx_uninstall(dev); 12025 netdev_lock_ops(dev); 12026 dev_xdp_uninstall(dev); 12027 dev_memory_provider_uninstall(dev); 12028 netdev_unlock_ops(dev); 12029 bpf_dev_bound_netdev_unregister(dev); 12030 12031 netdev_offload_xstats_disable_all(dev); 12032 12033 /* Notify protocols, that we are about to destroy 12034 * this device. They should clean all the things. 12035 */ 12036 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 12037 12038 if (!(dev->rtnl_link_ops && dev->rtnl_link_initializing)) 12039 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0, 12040 GFP_KERNEL, NULL, 0, 12041 portid, nlh); 12042 12043 /* 12044 * Flush the unicast and multicast chains 12045 */ 12046 dev_uc_flush(dev); 12047 dev_mc_flush(dev); 12048 12049 netdev_name_node_alt_flush(dev); 12050 netdev_name_node_free(dev->name_node); 12051 12052 netdev_rss_contexts_free(dev); 12053 12054 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 12055 12056 if (dev->netdev_ops->ndo_uninit) 12057 dev->netdev_ops->ndo_uninit(dev); 12058 12059 mutex_destroy(&dev->ethtool->rss_lock); 12060 12061 net_shaper_flush_netdev(dev); 12062 12063 if (skb) 12064 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh); 12065 12066 /* Notifier chain MUST detach us all upper devices. */ 12067 WARN_ON(netdev_has_any_upper_dev(dev)); 12068 WARN_ON(netdev_has_any_lower_dev(dev)); 12069 12070 /* Remove entries from kobject tree */ 12071 netdev_unregister_kobject(dev); 12072 #ifdef CONFIG_XPS 12073 /* Remove XPS queueing entries */ 12074 netif_reset_xps_queues_gt(dev, 0); 12075 #endif 12076 } 12077 12078 synchronize_net(); 12079 12080 list_for_each_entry(dev, head, unreg_list) { 12081 netdev_put(dev, &dev->dev_registered_tracker); 12082 net_set_todo(dev); 12083 cnt++; 12084 } 12085 atomic_add(cnt, &dev_unreg_count); 12086 12087 list_del(head); 12088 } 12089 12090 /** 12091 * unregister_netdevice_many - unregister many devices 12092 * @head: list of devices 12093 * 12094 * Note: As most callers use a stack allocated list_head, 12095 * we force a list_del() to make sure stack won't be corrupted later. 12096 */ 12097 void unregister_netdevice_many(struct list_head *head) 12098 { 12099 unregister_netdevice_many_notify(head, 0, NULL); 12100 } 12101 EXPORT_SYMBOL(unregister_netdevice_many); 12102 12103 /** 12104 * unregister_netdev - remove device from the kernel 12105 * @dev: device 12106 * 12107 * This function shuts down a device interface and removes it 12108 * from the kernel tables. 12109 * 12110 * This is just a wrapper for unregister_netdevice that takes 12111 * the rtnl semaphore. In general you want to use this and not 12112 * unregister_netdevice. 12113 */ 12114 void unregister_netdev(struct net_device *dev) 12115 { 12116 rtnl_net_dev_lock(dev); 12117 unregister_netdevice(dev); 12118 rtnl_net_dev_unlock(dev); 12119 } 12120 EXPORT_SYMBOL(unregister_netdev); 12121 12122 int __dev_change_net_namespace(struct net_device *dev, struct net *net, 12123 const char *pat, int new_ifindex, 12124 struct netlink_ext_ack *extack) 12125 { 12126 struct netdev_name_node *name_node; 12127 struct net *net_old = dev_net(dev); 12128 char new_name[IFNAMSIZ] = {}; 12129 int err, new_nsid; 12130 12131 ASSERT_RTNL(); 12132 12133 /* Don't allow namespace local devices to be moved. */ 12134 err = -EINVAL; 12135 if (dev->netns_immutable) { 12136 NL_SET_ERR_MSG(extack, "The interface netns is immutable"); 12137 goto out; 12138 } 12139 12140 /* Ensure the device has been registered */ 12141 if (dev->reg_state != NETREG_REGISTERED) { 12142 NL_SET_ERR_MSG(extack, "The interface isn't registered"); 12143 goto out; 12144 } 12145 12146 /* Get out if there is nothing todo */ 12147 err = 0; 12148 if (net_eq(net_old, net)) 12149 goto out; 12150 12151 /* Pick the destination device name, and ensure 12152 * we can use it in the destination network namespace. 12153 */ 12154 err = -EEXIST; 12155 if (netdev_name_in_use(net, dev->name)) { 12156 /* We get here if we can't use the current device name */ 12157 if (!pat) { 12158 NL_SET_ERR_MSG(extack, 12159 "An interface with the same name exists in the target netns"); 12160 goto out; 12161 } 12162 err = dev_prep_valid_name(net, dev, pat, new_name, EEXIST); 12163 if (err < 0) { 12164 NL_SET_ERR_MSG_FMT(extack, 12165 "Unable to use '%s' for the new interface name in the target netns", 12166 pat); 12167 goto out; 12168 } 12169 } 12170 /* Check that none of the altnames conflicts. */ 12171 err = -EEXIST; 12172 netdev_for_each_altname(dev, name_node) { 12173 if (netdev_name_in_use(net, name_node->name)) { 12174 NL_SET_ERR_MSG_FMT(extack, 12175 "An interface with the altname %s exists in the target netns", 12176 name_node->name); 12177 goto out; 12178 } 12179 } 12180 12181 /* Check that new_ifindex isn't used yet. */ 12182 if (new_ifindex) { 12183 err = dev_index_reserve(net, new_ifindex); 12184 if (err < 0) { 12185 NL_SET_ERR_MSG_FMT(extack, 12186 "The ifindex %d is not available in the target netns", 12187 new_ifindex); 12188 goto out; 12189 } 12190 } else { 12191 /* If there is an ifindex conflict assign a new one */ 12192 err = dev_index_reserve(net, dev->ifindex); 12193 if (err == -EBUSY) 12194 err = dev_index_reserve(net, 0); 12195 if (err < 0) { 12196 NL_SET_ERR_MSG(extack, 12197 "Unable to allocate a new ifindex in the target netns"); 12198 goto out; 12199 } 12200 new_ifindex = err; 12201 } 12202 12203 /* 12204 * And now a mini version of register_netdevice unregister_netdevice. 12205 */ 12206 12207 netdev_lock_ops(dev); 12208 /* If device is running close it first. */ 12209 netif_close(dev); 12210 /* And unlink it from device chain */ 12211 unlist_netdevice(dev); 12212 12213 if (!netdev_need_ops_lock(dev)) 12214 netdev_lock(dev); 12215 dev->moving_ns = true; 12216 netdev_unlock(dev); 12217 12218 synchronize_net(); 12219 12220 /* Shutdown queueing discipline. */ 12221 dev_shutdown(dev); 12222 12223 /* Notify protocols, that we are about to destroy 12224 * this device. They should clean all the things. 12225 * 12226 * Note that dev->reg_state stays at NETREG_REGISTERED. 12227 * This is wanted because this way 8021q and macvlan know 12228 * the device is just moving and can keep their slaves up. 12229 */ 12230 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 12231 rcu_barrier(); 12232 12233 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL); 12234 12235 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid, 12236 new_ifindex); 12237 12238 /* 12239 * Flush the unicast and multicast chains 12240 */ 12241 dev_uc_flush(dev); 12242 dev_mc_flush(dev); 12243 12244 /* Send a netdev-removed uevent to the old namespace */ 12245 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE); 12246 netdev_adjacent_del_links(dev); 12247 12248 /* Move per-net netdevice notifiers that are following the netdevice */ 12249 move_netdevice_notifiers_dev_net(dev, net); 12250 12251 /* Actually switch the network namespace */ 12252 netdev_lock(dev); 12253 dev_net_set(dev, net); 12254 netdev_unlock(dev); 12255 dev->ifindex = new_ifindex; 12256 12257 if (new_name[0]) { 12258 /* Rename the netdev to prepared name */ 12259 write_seqlock_bh(&netdev_rename_lock); 12260 strscpy(dev->name, new_name, IFNAMSIZ); 12261 write_sequnlock_bh(&netdev_rename_lock); 12262 } 12263 12264 /* Fixup kobjects */ 12265 dev_set_uevent_suppress(&dev->dev, 1); 12266 err = device_rename(&dev->dev, dev->name); 12267 dev_set_uevent_suppress(&dev->dev, 0); 12268 WARN_ON(err); 12269 12270 /* Send a netdev-add uevent to the new namespace */ 12271 kobject_uevent(&dev->dev.kobj, KOBJ_ADD); 12272 netdev_adjacent_add_links(dev); 12273 12274 /* Adapt owner in case owning user namespace of target network 12275 * namespace is different from the original one. 12276 */ 12277 err = netdev_change_owner(dev, net_old, net); 12278 WARN_ON(err); 12279 12280 netdev_lock(dev); 12281 dev->moving_ns = false; 12282 if (!netdev_need_ops_lock(dev)) 12283 netdev_unlock(dev); 12284 12285 /* Add the device back in the hashes */ 12286 list_netdevice(dev); 12287 /* Notify protocols, that a new device appeared. */ 12288 call_netdevice_notifiers(NETDEV_REGISTER, dev); 12289 netdev_unlock_ops(dev); 12290 12291 /* 12292 * Prevent userspace races by waiting until the network 12293 * device is fully setup before sending notifications. 12294 */ 12295 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 12296 12297 synchronize_net(); 12298 err = 0; 12299 out: 12300 return err; 12301 } 12302 12303 static int dev_cpu_dead(unsigned int oldcpu) 12304 { 12305 struct sk_buff **list_skb; 12306 struct sk_buff *skb; 12307 unsigned int cpu; 12308 struct softnet_data *sd, *oldsd, *remsd = NULL; 12309 12310 local_irq_disable(); 12311 cpu = smp_processor_id(); 12312 sd = &per_cpu(softnet_data, cpu); 12313 oldsd = &per_cpu(softnet_data, oldcpu); 12314 12315 /* Find end of our completion_queue. */ 12316 list_skb = &sd->completion_queue; 12317 while (*list_skb) 12318 list_skb = &(*list_skb)->next; 12319 /* Append completion queue from offline CPU. */ 12320 *list_skb = oldsd->completion_queue; 12321 oldsd->completion_queue = NULL; 12322 12323 /* Append output queue from offline CPU. */ 12324 if (oldsd->output_queue) { 12325 *sd->output_queue_tailp = oldsd->output_queue; 12326 sd->output_queue_tailp = oldsd->output_queue_tailp; 12327 oldsd->output_queue = NULL; 12328 oldsd->output_queue_tailp = &oldsd->output_queue; 12329 } 12330 /* Append NAPI poll list from offline CPU, with one exception : 12331 * process_backlog() must be called by cpu owning percpu backlog. 12332 * We properly handle process_queue & input_pkt_queue later. 12333 */ 12334 while (!list_empty(&oldsd->poll_list)) { 12335 struct napi_struct *napi = list_first_entry(&oldsd->poll_list, 12336 struct napi_struct, 12337 poll_list); 12338 12339 list_del_init(&napi->poll_list); 12340 if (napi->poll == process_backlog) 12341 napi->state &= NAPIF_STATE_THREADED; 12342 else 12343 ____napi_schedule(sd, napi); 12344 } 12345 12346 raise_softirq_irqoff(NET_TX_SOFTIRQ); 12347 local_irq_enable(); 12348 12349 if (!use_backlog_threads()) { 12350 #ifdef CONFIG_RPS 12351 remsd = oldsd->rps_ipi_list; 12352 oldsd->rps_ipi_list = NULL; 12353 #endif 12354 /* send out pending IPI's on offline CPU */ 12355 net_rps_send_ipi(remsd); 12356 } 12357 12358 /* Process offline CPU's input_pkt_queue */ 12359 while ((skb = __skb_dequeue(&oldsd->process_queue))) { 12360 netif_rx(skb); 12361 rps_input_queue_head_incr(oldsd); 12362 } 12363 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) { 12364 netif_rx(skb); 12365 rps_input_queue_head_incr(oldsd); 12366 } 12367 12368 return 0; 12369 } 12370 12371 /** 12372 * netdev_increment_features - increment feature set by one 12373 * @all: current feature set 12374 * @one: new feature set 12375 * @mask: mask feature set 12376 * 12377 * Computes a new feature set after adding a device with feature set 12378 * @one to the master device with current feature set @all. Will not 12379 * enable anything that is off in @mask. Returns the new feature set. 12380 */ 12381 netdev_features_t netdev_increment_features(netdev_features_t all, 12382 netdev_features_t one, netdev_features_t mask) 12383 { 12384 if (mask & NETIF_F_HW_CSUM) 12385 mask |= NETIF_F_CSUM_MASK; 12386 mask |= NETIF_F_VLAN_CHALLENGED; 12387 12388 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask; 12389 all &= one | ~NETIF_F_ALL_FOR_ALL; 12390 12391 /* If one device supports hw checksumming, set for all. */ 12392 if (all & NETIF_F_HW_CSUM) 12393 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM); 12394 12395 return all; 12396 } 12397 EXPORT_SYMBOL(netdev_increment_features); 12398 12399 static struct hlist_head * __net_init netdev_create_hash(void) 12400 { 12401 int i; 12402 struct hlist_head *hash; 12403 12404 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL); 12405 if (hash != NULL) 12406 for (i = 0; i < NETDEV_HASHENTRIES; i++) 12407 INIT_HLIST_HEAD(&hash[i]); 12408 12409 return hash; 12410 } 12411 12412 /* Initialize per network namespace state */ 12413 static int __net_init netdev_init(struct net *net) 12414 { 12415 BUILD_BUG_ON(GRO_HASH_BUCKETS > 12416 BITS_PER_BYTE * sizeof_field(struct gro_node, bitmask)); 12417 12418 INIT_LIST_HEAD(&net->dev_base_head); 12419 12420 net->dev_name_head = netdev_create_hash(); 12421 if (net->dev_name_head == NULL) 12422 goto err_name; 12423 12424 net->dev_index_head = netdev_create_hash(); 12425 if (net->dev_index_head == NULL) 12426 goto err_idx; 12427 12428 xa_init_flags(&net->dev_by_index, XA_FLAGS_ALLOC1); 12429 12430 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain); 12431 12432 return 0; 12433 12434 err_idx: 12435 kfree(net->dev_name_head); 12436 err_name: 12437 return -ENOMEM; 12438 } 12439 12440 /** 12441 * netdev_drivername - network driver for the device 12442 * @dev: network device 12443 * 12444 * Determine network driver for device. 12445 */ 12446 const char *netdev_drivername(const struct net_device *dev) 12447 { 12448 const struct device_driver *driver; 12449 const struct device *parent; 12450 const char *empty = ""; 12451 12452 parent = dev->dev.parent; 12453 if (!parent) 12454 return empty; 12455 12456 driver = parent->driver; 12457 if (driver && driver->name) 12458 return driver->name; 12459 return empty; 12460 } 12461 12462 static void __netdev_printk(const char *level, const struct net_device *dev, 12463 struct va_format *vaf) 12464 { 12465 if (dev && dev->dev.parent) { 12466 dev_printk_emit(level[1] - '0', 12467 dev->dev.parent, 12468 "%s %s %s%s: %pV", 12469 dev_driver_string(dev->dev.parent), 12470 dev_name(dev->dev.parent), 12471 netdev_name(dev), netdev_reg_state(dev), 12472 vaf); 12473 } else if (dev) { 12474 printk("%s%s%s: %pV", 12475 level, netdev_name(dev), netdev_reg_state(dev), vaf); 12476 } else { 12477 printk("%s(NULL net_device): %pV", level, vaf); 12478 } 12479 } 12480 12481 void netdev_printk(const char *level, const struct net_device *dev, 12482 const char *format, ...) 12483 { 12484 struct va_format vaf; 12485 va_list args; 12486 12487 va_start(args, format); 12488 12489 vaf.fmt = format; 12490 vaf.va = &args; 12491 12492 __netdev_printk(level, dev, &vaf); 12493 12494 va_end(args); 12495 } 12496 EXPORT_SYMBOL(netdev_printk); 12497 12498 #define define_netdev_printk_level(func, level) \ 12499 void func(const struct net_device *dev, const char *fmt, ...) \ 12500 { \ 12501 struct va_format vaf; \ 12502 va_list args; \ 12503 \ 12504 va_start(args, fmt); \ 12505 \ 12506 vaf.fmt = fmt; \ 12507 vaf.va = &args; \ 12508 \ 12509 __netdev_printk(level, dev, &vaf); \ 12510 \ 12511 va_end(args); \ 12512 } \ 12513 EXPORT_SYMBOL(func); 12514 12515 define_netdev_printk_level(netdev_emerg, KERN_EMERG); 12516 define_netdev_printk_level(netdev_alert, KERN_ALERT); 12517 define_netdev_printk_level(netdev_crit, KERN_CRIT); 12518 define_netdev_printk_level(netdev_err, KERN_ERR); 12519 define_netdev_printk_level(netdev_warn, KERN_WARNING); 12520 define_netdev_printk_level(netdev_notice, KERN_NOTICE); 12521 define_netdev_printk_level(netdev_info, KERN_INFO); 12522 12523 static void __net_exit netdev_exit(struct net *net) 12524 { 12525 kfree(net->dev_name_head); 12526 kfree(net->dev_index_head); 12527 xa_destroy(&net->dev_by_index); 12528 if (net != &init_net) 12529 WARN_ON_ONCE(!list_empty(&net->dev_base_head)); 12530 } 12531 12532 static struct pernet_operations __net_initdata netdev_net_ops = { 12533 .init = netdev_init, 12534 .exit = netdev_exit, 12535 }; 12536 12537 static void __net_exit default_device_exit_net(struct net *net) 12538 { 12539 struct netdev_name_node *name_node, *tmp; 12540 struct net_device *dev, *aux; 12541 /* 12542 * Push all migratable network devices back to the 12543 * initial network namespace 12544 */ 12545 ASSERT_RTNL(); 12546 for_each_netdev_safe(net, dev, aux) { 12547 int err; 12548 char fb_name[IFNAMSIZ]; 12549 12550 /* Ignore unmoveable devices (i.e. loopback) */ 12551 if (dev->netns_immutable) 12552 continue; 12553 12554 /* Leave virtual devices for the generic cleanup */ 12555 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund) 12556 continue; 12557 12558 /* Push remaining network devices to init_net */ 12559 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex); 12560 if (netdev_name_in_use(&init_net, fb_name)) 12561 snprintf(fb_name, IFNAMSIZ, "dev%%d"); 12562 12563 netdev_for_each_altname_safe(dev, name_node, tmp) 12564 if (netdev_name_in_use(&init_net, name_node->name)) 12565 __netdev_name_node_alt_destroy(name_node); 12566 12567 err = dev_change_net_namespace(dev, &init_net, fb_name); 12568 if (err) { 12569 pr_emerg("%s: failed to move %s to init_net: %d\n", 12570 __func__, dev->name, err); 12571 BUG(); 12572 } 12573 } 12574 } 12575 12576 static void __net_exit default_device_exit_batch(struct list_head *net_list) 12577 { 12578 /* At exit all network devices most be removed from a network 12579 * namespace. Do this in the reverse order of registration. 12580 * Do this across as many network namespaces as possible to 12581 * improve batching efficiency. 12582 */ 12583 struct net_device *dev; 12584 struct net *net; 12585 LIST_HEAD(dev_kill_list); 12586 12587 rtnl_lock(); 12588 list_for_each_entry(net, net_list, exit_list) { 12589 default_device_exit_net(net); 12590 cond_resched(); 12591 } 12592 12593 list_for_each_entry(net, net_list, exit_list) { 12594 for_each_netdev_reverse(net, dev) { 12595 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink) 12596 dev->rtnl_link_ops->dellink(dev, &dev_kill_list); 12597 else 12598 unregister_netdevice_queue(dev, &dev_kill_list); 12599 } 12600 } 12601 unregister_netdevice_many(&dev_kill_list); 12602 rtnl_unlock(); 12603 } 12604 12605 static struct pernet_operations __net_initdata default_device_ops = { 12606 .exit_batch = default_device_exit_batch, 12607 }; 12608 12609 static void __init net_dev_struct_check(void) 12610 { 12611 /* TX read-mostly hotpath */ 12612 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, priv_flags_fast); 12613 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, netdev_ops); 12614 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, header_ops); 12615 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, _tx); 12616 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, real_num_tx_queues); 12617 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_size); 12618 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_ipv4_max_size); 12619 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_segs); 12620 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_partial_features); 12621 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, num_tc); 12622 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, mtu); 12623 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, needed_headroom); 12624 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tc_to_txq); 12625 #ifdef CONFIG_XPS 12626 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, xps_maps); 12627 #endif 12628 #ifdef CONFIG_NETFILTER_EGRESS 12629 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, nf_hooks_egress); 12630 #endif 12631 #ifdef CONFIG_NET_XGRESS 12632 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tcx_egress); 12633 #endif 12634 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_tx, 160); 12635 12636 /* TXRX read-mostly hotpath */ 12637 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, lstats); 12638 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, state); 12639 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, flags); 12640 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, hard_header_len); 12641 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, features); 12642 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, ip6_ptr); 12643 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_txrx, 46); 12644 12645 /* RX read-mostly hotpath */ 12646 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ptype_specific); 12647 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ifindex); 12648 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, real_num_rx_queues); 12649 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, _rx); 12650 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_max_size); 12651 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_ipv4_max_size); 12652 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler); 12653 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler_data); 12654 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, nd_net); 12655 #ifdef CONFIG_NETPOLL 12656 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, npinfo); 12657 #endif 12658 #ifdef CONFIG_NET_XGRESS 12659 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, tcx_ingress); 12660 #endif 12661 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_rx, 92); 12662 } 12663 12664 /* 12665 * Initialize the DEV module. At boot time this walks the device list and 12666 * unhooks any devices that fail to initialise (normally hardware not 12667 * present) and leaves us with a valid list of present and active devices. 12668 * 12669 */ 12670 12671 /* We allocate 256 pages for each CPU if PAGE_SHIFT is 12 */ 12672 #define SYSTEM_PERCPU_PAGE_POOL_SIZE ((1 << 20) / PAGE_SIZE) 12673 12674 static int net_page_pool_create(int cpuid) 12675 { 12676 #if IS_ENABLED(CONFIG_PAGE_POOL) 12677 struct page_pool_params page_pool_params = { 12678 .pool_size = SYSTEM_PERCPU_PAGE_POOL_SIZE, 12679 .flags = PP_FLAG_SYSTEM_POOL, 12680 .nid = cpu_to_mem(cpuid), 12681 }; 12682 struct page_pool *pp_ptr; 12683 int err; 12684 12685 pp_ptr = page_pool_create_percpu(&page_pool_params, cpuid); 12686 if (IS_ERR(pp_ptr)) 12687 return -ENOMEM; 12688 12689 err = xdp_reg_page_pool(pp_ptr); 12690 if (err) { 12691 page_pool_destroy(pp_ptr); 12692 return err; 12693 } 12694 12695 per_cpu(system_page_pool, cpuid) = pp_ptr; 12696 #endif 12697 return 0; 12698 } 12699 12700 static int backlog_napi_should_run(unsigned int cpu) 12701 { 12702 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12703 struct napi_struct *napi = &sd->backlog; 12704 12705 return test_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 12706 } 12707 12708 static void run_backlog_napi(unsigned int cpu) 12709 { 12710 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12711 12712 napi_threaded_poll_loop(&sd->backlog); 12713 } 12714 12715 static void backlog_napi_setup(unsigned int cpu) 12716 { 12717 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12718 struct napi_struct *napi = &sd->backlog; 12719 12720 napi->thread = this_cpu_read(backlog_napi); 12721 set_bit(NAPI_STATE_THREADED, &napi->state); 12722 } 12723 12724 static struct smp_hotplug_thread backlog_threads = { 12725 .store = &backlog_napi, 12726 .thread_should_run = backlog_napi_should_run, 12727 .thread_fn = run_backlog_napi, 12728 .thread_comm = "backlog_napi/%u", 12729 .setup = backlog_napi_setup, 12730 }; 12731 12732 /* 12733 * This is called single threaded during boot, so no need 12734 * to take the rtnl semaphore. 12735 */ 12736 static int __init net_dev_init(void) 12737 { 12738 int i, rc = -ENOMEM; 12739 12740 BUG_ON(!dev_boot_phase); 12741 12742 net_dev_struct_check(); 12743 12744 if (dev_proc_init()) 12745 goto out; 12746 12747 if (netdev_kobject_init()) 12748 goto out; 12749 12750 for (i = 0; i < PTYPE_HASH_SIZE; i++) 12751 INIT_LIST_HEAD(&ptype_base[i]); 12752 12753 if (register_pernet_subsys(&netdev_net_ops)) 12754 goto out; 12755 12756 /* 12757 * Initialise the packet receive queues. 12758 */ 12759 12760 flush_backlogs_fallback = flush_backlogs_alloc(); 12761 if (!flush_backlogs_fallback) 12762 goto out; 12763 12764 for_each_possible_cpu(i) { 12765 struct softnet_data *sd = &per_cpu(softnet_data, i); 12766 12767 skb_queue_head_init(&sd->input_pkt_queue); 12768 skb_queue_head_init(&sd->process_queue); 12769 #ifdef CONFIG_XFRM_OFFLOAD 12770 skb_queue_head_init(&sd->xfrm_backlog); 12771 #endif 12772 INIT_LIST_HEAD(&sd->poll_list); 12773 sd->output_queue_tailp = &sd->output_queue; 12774 #ifdef CONFIG_RPS 12775 INIT_CSD(&sd->csd, rps_trigger_softirq, sd); 12776 sd->cpu = i; 12777 #endif 12778 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd); 12779 spin_lock_init(&sd->defer_lock); 12780 12781 gro_init(&sd->backlog.gro); 12782 sd->backlog.poll = process_backlog; 12783 sd->backlog.weight = weight_p; 12784 INIT_LIST_HEAD(&sd->backlog.poll_list); 12785 12786 if (net_page_pool_create(i)) 12787 goto out; 12788 } 12789 if (use_backlog_threads()) 12790 smpboot_register_percpu_thread(&backlog_threads); 12791 12792 dev_boot_phase = 0; 12793 12794 /* The loopback device is special if any other network devices 12795 * is present in a network namespace the loopback device must 12796 * be present. Since we now dynamically allocate and free the 12797 * loopback device ensure this invariant is maintained by 12798 * keeping the loopback device as the first device on the 12799 * list of network devices. Ensuring the loopback devices 12800 * is the first device that appears and the last network device 12801 * that disappears. 12802 */ 12803 if (register_pernet_device(&loopback_net_ops)) 12804 goto out; 12805 12806 if (register_pernet_device(&default_device_ops)) 12807 goto out; 12808 12809 open_softirq(NET_TX_SOFTIRQ, net_tx_action); 12810 open_softirq(NET_RX_SOFTIRQ, net_rx_action); 12811 12812 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead", 12813 NULL, dev_cpu_dead); 12814 WARN_ON(rc < 0); 12815 rc = 0; 12816 12817 /* avoid static key IPIs to isolated CPUs */ 12818 if (housekeeping_enabled(HK_TYPE_MISC)) 12819 net_enable_timestamp(); 12820 out: 12821 if (rc < 0) { 12822 for_each_possible_cpu(i) { 12823 struct page_pool *pp_ptr; 12824 12825 pp_ptr = per_cpu(system_page_pool, i); 12826 if (!pp_ptr) 12827 continue; 12828 12829 xdp_unreg_page_pool(pp_ptr); 12830 page_pool_destroy(pp_ptr); 12831 per_cpu(system_page_pool, i) = NULL; 12832 } 12833 } 12834 12835 return rc; 12836 } 12837 12838 subsys_initcall(net_dev_init); 12839