1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * NET3 Protocol independent device support routines. 4 * 5 * Derived from the non IP parts of dev.c 1.0.19 6 * Authors: Ross Biro 7 * Fred N. van Kempen, <[email protected]> 8 * Mark Evans, <[email protected]> 9 * 10 * Additional Authors: 11 * Florian la Roche <[email protected]> 12 * Alan Cox <[email protected]> 13 * David Hinds <[email protected]> 14 * Alexey Kuznetsov <[email protected]> 15 * Adam Sulmicki <[email protected]> 16 * Pekka Riikonen <[email protected]> 17 * 18 * Changes: 19 * D.J. Barrow : Fixed bug where dev->refcnt gets set 20 * to 2 if register_netdev gets called 21 * before net_dev_init & also removed a 22 * few lines of code in the process. 23 * Alan Cox : device private ioctl copies fields back. 24 * Alan Cox : Transmit queue code does relevant 25 * stunts to keep the queue safe. 26 * Alan Cox : Fixed double lock. 27 * Alan Cox : Fixed promisc NULL pointer trap 28 * ???????? : Support the full private ioctl range 29 * Alan Cox : Moved ioctl permission check into 30 * drivers 31 * Tim Kordas : SIOCADDMULTI/SIOCDELMULTI 32 * Alan Cox : 100 backlog just doesn't cut it when 33 * you start doing multicast video 8) 34 * Alan Cox : Rewrote net_bh and list manager. 35 * Alan Cox : Fix ETH_P_ALL echoback lengths. 36 * Alan Cox : Took out transmit every packet pass 37 * Saved a few bytes in the ioctl handler 38 * Alan Cox : Network driver sets packet type before 39 * calling netif_rx. Saves a function 40 * call a packet. 41 * Alan Cox : Hashed net_bh() 42 * Richard Kooijman: Timestamp fixes. 43 * Alan Cox : Wrong field in SIOCGIFDSTADDR 44 * Alan Cox : Device lock protection. 45 * Alan Cox : Fixed nasty side effect of device close 46 * changes. 47 * Rudi Cilibrasi : Pass the right thing to 48 * set_mac_address() 49 * Dave Miller : 32bit quantity for the device lock to 50 * make it work out on a Sparc. 51 * Bjorn Ekwall : Added KERNELD hack. 52 * Alan Cox : Cleaned up the backlog initialise. 53 * Craig Metz : SIOCGIFCONF fix if space for under 54 * 1 device. 55 * Thomas Bogendoerfer : Return ENODEV for dev_open, if there 56 * is no device open function. 57 * Andi Kleen : Fix error reporting for SIOCGIFCONF 58 * Michael Chastain : Fix signed/unsigned for SIOCGIFCONF 59 * Cyrus Durgin : Cleaned for KMOD 60 * Adam Sulmicki : Bug Fix : Network Device Unload 61 * A network device unload needs to purge 62 * the backlog queue. 63 * Paul Rusty Russell : SIOCSIFNAME 64 * Pekka Riikonen : Netdev boot-time settings code 65 * Andrew Morton : Make unregister_netdevice wait 66 * indefinitely on dev->refcnt 67 * J Hadi Salim : - Backlog queue sampling 68 * - netif_rx() feedback 69 */ 70 71 #include <linux/uaccess.h> 72 #include <linux/bitmap.h> 73 #include <linux/capability.h> 74 #include <linux/cpu.h> 75 #include <linux/types.h> 76 #include <linux/kernel.h> 77 #include <linux/hash.h> 78 #include <linux/slab.h> 79 #include <linux/sched.h> 80 #include <linux/sched/isolation.h> 81 #include <linux/sched/mm.h> 82 #include <linux/smpboot.h> 83 #include <linux/mutex.h> 84 #include <linux/rwsem.h> 85 #include <linux/string.h> 86 #include <linux/mm.h> 87 #include <linux/socket.h> 88 #include <linux/sockios.h> 89 #include <linux/errno.h> 90 #include <linux/interrupt.h> 91 #include <linux/if_ether.h> 92 #include <linux/netdevice.h> 93 #include <linux/etherdevice.h> 94 #include <linux/ethtool.h> 95 #include <linux/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 return pt->dev ? &pt->dev->ptype_all : &net_hotdata.ptype_all; 577 else 578 return pt->dev ? &pt->dev->ptype_specific : 579 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK]; 580 } 581 582 /** 583 * dev_add_pack - add packet handler 584 * @pt: packet type declaration 585 * 586 * Add a protocol handler to the networking stack. The passed &packet_type 587 * is linked into kernel lists and may not be freed until it has been 588 * removed from the kernel lists. 589 * 590 * This call does not sleep therefore it can not 591 * guarantee all CPU's that are in middle of receiving packets 592 * will see the new packet type (until the next received packet). 593 */ 594 595 void dev_add_pack(struct packet_type *pt) 596 { 597 struct list_head *head = ptype_head(pt); 598 599 spin_lock(&ptype_lock); 600 list_add_rcu(&pt->list, head); 601 spin_unlock(&ptype_lock); 602 } 603 EXPORT_SYMBOL(dev_add_pack); 604 605 /** 606 * __dev_remove_pack - remove packet handler 607 * @pt: packet type declaration 608 * 609 * Remove a protocol handler that was previously added to the kernel 610 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 611 * from the kernel lists and can be freed or reused once this function 612 * returns. 613 * 614 * The packet type might still be in use by receivers 615 * and must not be freed until after all the CPU's have gone 616 * through a quiescent state. 617 */ 618 void __dev_remove_pack(struct packet_type *pt) 619 { 620 struct list_head *head = ptype_head(pt); 621 struct packet_type *pt1; 622 623 spin_lock(&ptype_lock); 624 625 list_for_each_entry(pt1, head, list) { 626 if (pt == pt1) { 627 list_del_rcu(&pt->list); 628 goto out; 629 } 630 } 631 632 pr_warn("dev_remove_pack: %p not found\n", pt); 633 out: 634 spin_unlock(&ptype_lock); 635 } 636 EXPORT_SYMBOL(__dev_remove_pack); 637 638 /** 639 * dev_remove_pack - remove packet handler 640 * @pt: packet type declaration 641 * 642 * Remove a protocol handler that was previously added to the kernel 643 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 644 * from the kernel lists and can be freed or reused once this function 645 * returns. 646 * 647 * This call sleeps to guarantee that no CPU is looking at the packet 648 * type after return. 649 */ 650 void dev_remove_pack(struct packet_type *pt) 651 { 652 __dev_remove_pack(pt); 653 654 synchronize_net(); 655 } 656 EXPORT_SYMBOL(dev_remove_pack); 657 658 659 /******************************************************************************* 660 * 661 * Device Interface Subroutines 662 * 663 *******************************************************************************/ 664 665 /** 666 * dev_get_iflink - get 'iflink' value of a interface 667 * @dev: targeted interface 668 * 669 * Indicates the ifindex the interface is linked to. 670 * Physical interfaces have the same 'ifindex' and 'iflink' values. 671 */ 672 673 int dev_get_iflink(const struct net_device *dev) 674 { 675 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink) 676 return dev->netdev_ops->ndo_get_iflink(dev); 677 678 return READ_ONCE(dev->ifindex); 679 } 680 EXPORT_SYMBOL(dev_get_iflink); 681 682 /** 683 * dev_fill_metadata_dst - Retrieve tunnel egress information. 684 * @dev: targeted interface 685 * @skb: The packet. 686 * 687 * For better visibility of tunnel traffic OVS needs to retrieve 688 * egress tunnel information for a packet. Following API allows 689 * user to get this info. 690 */ 691 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb) 692 { 693 struct ip_tunnel_info *info; 694 695 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst) 696 return -EINVAL; 697 698 info = skb_tunnel_info_unclone(skb); 699 if (!info) 700 return -ENOMEM; 701 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX))) 702 return -EINVAL; 703 704 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb); 705 } 706 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst); 707 708 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack) 709 { 710 int k = stack->num_paths++; 711 712 if (WARN_ON_ONCE(k >= NET_DEVICE_PATH_STACK_MAX)) 713 return NULL; 714 715 return &stack->path[k]; 716 } 717 718 int dev_fill_forward_path(const struct net_device *dev, const u8 *daddr, 719 struct net_device_path_stack *stack) 720 { 721 const struct net_device *last_dev; 722 struct net_device_path_ctx ctx = { 723 .dev = dev, 724 }; 725 struct net_device_path *path; 726 int ret = 0; 727 728 memcpy(ctx.daddr, daddr, sizeof(ctx.daddr)); 729 stack->num_paths = 0; 730 while (ctx.dev && ctx.dev->netdev_ops->ndo_fill_forward_path) { 731 last_dev = ctx.dev; 732 path = dev_fwd_path(stack); 733 if (!path) 734 return -1; 735 736 memset(path, 0, sizeof(struct net_device_path)); 737 ret = ctx.dev->netdev_ops->ndo_fill_forward_path(&ctx, path); 738 if (ret < 0) 739 return -1; 740 741 if (WARN_ON_ONCE(last_dev == ctx.dev)) 742 return -1; 743 } 744 745 if (!ctx.dev) 746 return ret; 747 748 path = dev_fwd_path(stack); 749 if (!path) 750 return -1; 751 path->type = DEV_PATH_ETHERNET; 752 path->dev = ctx.dev; 753 754 return ret; 755 } 756 EXPORT_SYMBOL_GPL(dev_fill_forward_path); 757 758 /* must be called under rcu_read_lock(), as we dont take a reference */ 759 static struct napi_struct *napi_by_id(unsigned int napi_id) 760 { 761 unsigned int hash = napi_id % HASH_SIZE(napi_hash); 762 struct napi_struct *napi; 763 764 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node) 765 if (napi->napi_id == napi_id) 766 return napi; 767 768 return NULL; 769 } 770 771 /* must be called under rcu_read_lock(), as we dont take a reference */ 772 static struct napi_struct * 773 netdev_napi_by_id(struct net *net, unsigned int napi_id) 774 { 775 struct napi_struct *napi; 776 777 napi = napi_by_id(napi_id); 778 if (!napi) 779 return NULL; 780 781 if (WARN_ON_ONCE(!napi->dev)) 782 return NULL; 783 if (!net_eq(net, dev_net(napi->dev))) 784 return NULL; 785 786 return napi; 787 } 788 789 /** 790 * netdev_napi_by_id_lock() - find a device by NAPI ID and lock it 791 * @net: the applicable net namespace 792 * @napi_id: ID of a NAPI of a target device 793 * 794 * Find a NAPI instance with @napi_id. Lock its device. 795 * The device must be in %NETREG_REGISTERED state for lookup to succeed. 796 * netdev_unlock() must be called to release it. 797 * 798 * Return: pointer to NAPI, its device with lock held, NULL if not found. 799 */ 800 struct napi_struct * 801 netdev_napi_by_id_lock(struct net *net, unsigned int napi_id) 802 { 803 struct napi_struct *napi; 804 struct net_device *dev; 805 806 rcu_read_lock(); 807 napi = netdev_napi_by_id(net, napi_id); 808 if (!napi || READ_ONCE(napi->dev->reg_state) != NETREG_REGISTERED) { 809 rcu_read_unlock(); 810 return NULL; 811 } 812 813 dev = napi->dev; 814 dev_hold(dev); 815 rcu_read_unlock(); 816 817 dev = __netdev_put_lock(dev); 818 if (!dev) 819 return NULL; 820 821 rcu_read_lock(); 822 napi = netdev_napi_by_id(net, napi_id); 823 if (napi && napi->dev != dev) 824 napi = NULL; 825 rcu_read_unlock(); 826 827 if (!napi) 828 netdev_unlock(dev); 829 return napi; 830 } 831 832 /** 833 * __dev_get_by_name - find a device by its name 834 * @net: the applicable net namespace 835 * @name: name to find 836 * 837 * Find an interface by name. Must be called under RTNL semaphore. 838 * If the name is found a pointer to the device is returned. 839 * If the name is not found then %NULL is returned. The 840 * reference counters are not incremented so the caller must be 841 * careful with locks. 842 */ 843 844 struct net_device *__dev_get_by_name(struct net *net, const char *name) 845 { 846 struct netdev_name_node *node_name; 847 848 node_name = netdev_name_node_lookup(net, name); 849 return node_name ? node_name->dev : NULL; 850 } 851 EXPORT_SYMBOL(__dev_get_by_name); 852 853 /** 854 * dev_get_by_name_rcu - find a device by its name 855 * @net: the applicable net namespace 856 * @name: name to find 857 * 858 * Find an interface by name. 859 * If the name is found a pointer to the device is returned. 860 * If the name is not found then %NULL is returned. 861 * The reference counters are not incremented so the caller must be 862 * careful with locks. The caller must hold RCU lock. 863 */ 864 865 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name) 866 { 867 struct netdev_name_node *node_name; 868 869 node_name = netdev_name_node_lookup_rcu(net, name); 870 return node_name ? node_name->dev : NULL; 871 } 872 EXPORT_SYMBOL(dev_get_by_name_rcu); 873 874 /* Deprecated for new users, call netdev_get_by_name() instead */ 875 struct net_device *dev_get_by_name(struct net *net, const char *name) 876 { 877 struct net_device *dev; 878 879 rcu_read_lock(); 880 dev = dev_get_by_name_rcu(net, name); 881 dev_hold(dev); 882 rcu_read_unlock(); 883 return dev; 884 } 885 EXPORT_SYMBOL(dev_get_by_name); 886 887 /** 888 * netdev_get_by_name() - find a device by its name 889 * @net: the applicable net namespace 890 * @name: name to find 891 * @tracker: tracking object for the acquired reference 892 * @gfp: allocation flags for the tracker 893 * 894 * Find an interface by name. This can be called from any 895 * context and does its own locking. The returned handle has 896 * the usage count incremented and the caller must use netdev_put() to 897 * release it when it is no longer needed. %NULL is returned if no 898 * matching device is found. 899 */ 900 struct net_device *netdev_get_by_name(struct net *net, const char *name, 901 netdevice_tracker *tracker, gfp_t gfp) 902 { 903 struct net_device *dev; 904 905 dev = dev_get_by_name(net, name); 906 if (dev) 907 netdev_tracker_alloc(dev, tracker, gfp); 908 return dev; 909 } 910 EXPORT_SYMBOL(netdev_get_by_name); 911 912 /** 913 * __dev_get_by_index - find a device by its ifindex 914 * @net: the applicable net namespace 915 * @ifindex: index of device 916 * 917 * Search for an interface by index. Returns %NULL if the device 918 * is not found or a pointer to the device. The device has not 919 * had its reference counter increased so the caller must be careful 920 * about locking. The caller must hold the RTNL semaphore. 921 */ 922 923 struct net_device *__dev_get_by_index(struct net *net, int ifindex) 924 { 925 struct net_device *dev; 926 struct hlist_head *head = dev_index_hash(net, ifindex); 927 928 hlist_for_each_entry(dev, head, index_hlist) 929 if (dev->ifindex == ifindex) 930 return dev; 931 932 return NULL; 933 } 934 EXPORT_SYMBOL(__dev_get_by_index); 935 936 /** 937 * dev_get_by_index_rcu - find a device by its ifindex 938 * @net: the applicable net namespace 939 * @ifindex: index of device 940 * 941 * Search for an interface by index. Returns %NULL if the device 942 * is not found or a pointer to the device. The device has not 943 * had its reference counter increased so the caller must be careful 944 * about locking. The caller must hold RCU lock. 945 */ 946 947 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex) 948 { 949 struct net_device *dev; 950 struct hlist_head *head = dev_index_hash(net, ifindex); 951 952 hlist_for_each_entry_rcu(dev, head, index_hlist) 953 if (dev->ifindex == ifindex) 954 return dev; 955 956 return NULL; 957 } 958 EXPORT_SYMBOL(dev_get_by_index_rcu); 959 960 /* Deprecated for new users, call netdev_get_by_index() instead */ 961 struct net_device *dev_get_by_index(struct net *net, int ifindex) 962 { 963 struct net_device *dev; 964 965 rcu_read_lock(); 966 dev = dev_get_by_index_rcu(net, ifindex); 967 dev_hold(dev); 968 rcu_read_unlock(); 969 return dev; 970 } 971 EXPORT_SYMBOL(dev_get_by_index); 972 973 /** 974 * netdev_get_by_index() - find a device by its ifindex 975 * @net: the applicable net namespace 976 * @ifindex: index of device 977 * @tracker: tracking object for the acquired reference 978 * @gfp: allocation flags for the tracker 979 * 980 * Search for an interface by index. Returns NULL if the device 981 * is not found or a pointer to the device. The device returned has 982 * had a reference added and the pointer is safe until the user calls 983 * netdev_put() to indicate they have finished with it. 984 */ 985 struct net_device *netdev_get_by_index(struct net *net, int ifindex, 986 netdevice_tracker *tracker, gfp_t gfp) 987 { 988 struct net_device *dev; 989 990 dev = dev_get_by_index(net, ifindex); 991 if (dev) 992 netdev_tracker_alloc(dev, tracker, gfp); 993 return dev; 994 } 995 EXPORT_SYMBOL(netdev_get_by_index); 996 997 /** 998 * dev_get_by_napi_id - find a device by napi_id 999 * @napi_id: ID of the NAPI struct 1000 * 1001 * Search for an interface by NAPI ID. Returns %NULL if the device 1002 * is not found or a pointer to the device. The device has not had 1003 * its reference counter increased so the caller must be careful 1004 * about locking. The caller must hold RCU lock. 1005 */ 1006 struct net_device *dev_get_by_napi_id(unsigned int napi_id) 1007 { 1008 struct napi_struct *napi; 1009 1010 WARN_ON_ONCE(!rcu_read_lock_held()); 1011 1012 if (!napi_id_valid(napi_id)) 1013 return NULL; 1014 1015 napi = napi_by_id(napi_id); 1016 1017 return napi ? napi->dev : NULL; 1018 } 1019 1020 /* Release the held reference on the net_device, and if the net_device 1021 * is still registered try to lock the instance lock. If device is being 1022 * unregistered NULL will be returned (but the reference has been released, 1023 * either way!) 1024 * 1025 * This helper is intended for locking net_device after it has been looked up 1026 * using a lockless lookup helper. Lock prevents the instance from going away. 1027 */ 1028 struct net_device *__netdev_put_lock(struct net_device *dev) 1029 { 1030 netdev_lock(dev); 1031 if (dev->reg_state > NETREG_REGISTERED) { 1032 netdev_unlock(dev); 1033 dev_put(dev); 1034 return NULL; 1035 } 1036 dev_put(dev); 1037 return dev; 1038 } 1039 1040 /** 1041 * netdev_get_by_index_lock() - find a device by its ifindex 1042 * @net: the applicable net namespace 1043 * @ifindex: index of device 1044 * 1045 * Search for an interface by index. If a valid device 1046 * with @ifindex is found it will be returned with netdev->lock held. 1047 * netdev_unlock() must be called to release it. 1048 * 1049 * Return: pointer to a device with lock held, NULL if not found. 1050 */ 1051 struct net_device *netdev_get_by_index_lock(struct net *net, int ifindex) 1052 { 1053 struct net_device *dev; 1054 1055 dev = dev_get_by_index(net, ifindex); 1056 if (!dev) 1057 return NULL; 1058 1059 return __netdev_put_lock(dev); 1060 } 1061 1062 /** 1063 * netdev_get_by_name_lock() - find a device by its name 1064 * @net: the applicable net namespace 1065 * @name: name of device 1066 * 1067 * Search for an interface by name. If a valid device 1068 * with @name is found it will be returned with netdev->lock held. 1069 * netdev_unlock() must be called to release it. 1070 * 1071 * Return: pointer to a device with lock held, NULL if not found. 1072 */ 1073 struct net_device *netdev_get_by_name_lock(struct net *net, const char *name) 1074 { 1075 struct net_device *dev; 1076 1077 dev = dev_get_by_name(net, name); 1078 if (!dev) 1079 return NULL; 1080 1081 return __netdev_put_lock(dev); 1082 } 1083 1084 struct net_device * 1085 netdev_xa_find_lock(struct net *net, struct net_device *dev, 1086 unsigned long *index) 1087 { 1088 if (dev) 1089 netdev_unlock(dev); 1090 1091 do { 1092 rcu_read_lock(); 1093 dev = xa_find(&net->dev_by_index, index, ULONG_MAX, XA_PRESENT); 1094 if (!dev) { 1095 rcu_read_unlock(); 1096 return NULL; 1097 } 1098 dev_hold(dev); 1099 rcu_read_unlock(); 1100 1101 dev = __netdev_put_lock(dev); 1102 if (dev) 1103 return dev; 1104 1105 (*index)++; 1106 } while (true); 1107 } 1108 1109 static DEFINE_SEQLOCK(netdev_rename_lock); 1110 1111 void netdev_copy_name(struct net_device *dev, char *name) 1112 { 1113 unsigned int seq; 1114 1115 do { 1116 seq = read_seqbegin(&netdev_rename_lock); 1117 strscpy(name, dev->name, IFNAMSIZ); 1118 } while (read_seqretry(&netdev_rename_lock, seq)); 1119 } 1120 1121 /** 1122 * netdev_get_name - get a netdevice name, knowing its ifindex. 1123 * @net: network namespace 1124 * @name: a pointer to the buffer where the name will be stored. 1125 * @ifindex: the ifindex of the interface to get the name from. 1126 */ 1127 int netdev_get_name(struct net *net, char *name, int ifindex) 1128 { 1129 struct net_device *dev; 1130 int ret; 1131 1132 rcu_read_lock(); 1133 1134 dev = dev_get_by_index_rcu(net, ifindex); 1135 if (!dev) { 1136 ret = -ENODEV; 1137 goto out; 1138 } 1139 1140 netdev_copy_name(dev, name); 1141 1142 ret = 0; 1143 out: 1144 rcu_read_unlock(); 1145 return ret; 1146 } 1147 1148 static bool dev_addr_cmp(struct net_device *dev, unsigned short type, 1149 const char *ha) 1150 { 1151 return dev->type == type && !memcmp(dev->dev_addr, ha, dev->addr_len); 1152 } 1153 1154 /** 1155 * dev_getbyhwaddr_rcu - find a device by its hardware address 1156 * @net: the applicable net namespace 1157 * @type: media type of device 1158 * @ha: hardware address 1159 * 1160 * Search for an interface by MAC address. Returns NULL if the device 1161 * is not found or a pointer to the device. 1162 * The caller must hold RCU. 1163 * The returned device has not had its ref count increased 1164 * and the caller must therefore be careful about locking 1165 * 1166 */ 1167 1168 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type, 1169 const char *ha) 1170 { 1171 struct net_device *dev; 1172 1173 for_each_netdev_rcu(net, dev) 1174 if (dev_addr_cmp(dev, type, ha)) 1175 return dev; 1176 1177 return NULL; 1178 } 1179 EXPORT_SYMBOL(dev_getbyhwaddr_rcu); 1180 1181 /** 1182 * dev_getbyhwaddr() - find a device by its hardware address 1183 * @net: the applicable net namespace 1184 * @type: media type of device 1185 * @ha: hardware address 1186 * 1187 * Similar to dev_getbyhwaddr_rcu(), but the owner needs to hold 1188 * rtnl_lock. 1189 * 1190 * Context: rtnl_lock() must be held. 1191 * Return: pointer to the net_device, or NULL if not found 1192 */ 1193 struct net_device *dev_getbyhwaddr(struct net *net, unsigned short type, 1194 const char *ha) 1195 { 1196 struct net_device *dev; 1197 1198 ASSERT_RTNL(); 1199 for_each_netdev(net, dev) 1200 if (dev_addr_cmp(dev, type, ha)) 1201 return dev; 1202 1203 return NULL; 1204 } 1205 EXPORT_SYMBOL(dev_getbyhwaddr); 1206 1207 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type) 1208 { 1209 struct net_device *dev, *ret = NULL; 1210 1211 rcu_read_lock(); 1212 for_each_netdev_rcu(net, dev) 1213 if (dev->type == type) { 1214 dev_hold(dev); 1215 ret = dev; 1216 break; 1217 } 1218 rcu_read_unlock(); 1219 return ret; 1220 } 1221 EXPORT_SYMBOL(dev_getfirstbyhwtype); 1222 1223 /** 1224 * __dev_get_by_flags - find any device with given flags 1225 * @net: the applicable net namespace 1226 * @if_flags: IFF_* values 1227 * @mask: bitmask of bits in if_flags to check 1228 * 1229 * Search for any interface with the given flags. Returns NULL if a device 1230 * is not found or a pointer to the device. Must be called inside 1231 * rtnl_lock(), and result refcount is unchanged. 1232 */ 1233 1234 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags, 1235 unsigned short mask) 1236 { 1237 struct net_device *dev, *ret; 1238 1239 ASSERT_RTNL(); 1240 1241 ret = NULL; 1242 for_each_netdev(net, dev) { 1243 if (((dev->flags ^ if_flags) & mask) == 0) { 1244 ret = dev; 1245 break; 1246 } 1247 } 1248 return ret; 1249 } 1250 EXPORT_SYMBOL(__dev_get_by_flags); 1251 1252 /** 1253 * dev_valid_name - check if name is okay for network device 1254 * @name: name string 1255 * 1256 * Network device names need to be valid file names to 1257 * allow sysfs to work. We also disallow any kind of 1258 * whitespace. 1259 */ 1260 bool dev_valid_name(const char *name) 1261 { 1262 if (*name == '\0') 1263 return false; 1264 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ) 1265 return false; 1266 if (!strcmp(name, ".") || !strcmp(name, "..")) 1267 return false; 1268 1269 while (*name) { 1270 if (*name == '/' || *name == ':' || isspace(*name)) 1271 return false; 1272 name++; 1273 } 1274 return true; 1275 } 1276 EXPORT_SYMBOL(dev_valid_name); 1277 1278 /** 1279 * __dev_alloc_name - allocate a name for a device 1280 * @net: network namespace to allocate the device name in 1281 * @name: name format string 1282 * @res: result name string 1283 * 1284 * Passed a format string - eg "lt%d" it will try and find a suitable 1285 * id. It scans list of devices to build up a free map, then chooses 1286 * the first empty slot. The caller must hold the dev_base or rtnl lock 1287 * while allocating the name and adding the device in order to avoid 1288 * duplicates. 1289 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1290 * Returns the number of the unit assigned or a negative errno code. 1291 */ 1292 1293 static int __dev_alloc_name(struct net *net, const char *name, char *res) 1294 { 1295 int i = 0; 1296 const char *p; 1297 const int max_netdevices = 8*PAGE_SIZE; 1298 unsigned long *inuse; 1299 struct net_device *d; 1300 char buf[IFNAMSIZ]; 1301 1302 /* Verify the string as this thing may have come from the user. 1303 * There must be one "%d" and no other "%" characters. 1304 */ 1305 p = strchr(name, '%'); 1306 if (!p || p[1] != 'd' || strchr(p + 2, '%')) 1307 return -EINVAL; 1308 1309 /* Use one page as a bit array of possible slots */ 1310 inuse = bitmap_zalloc(max_netdevices, GFP_ATOMIC); 1311 if (!inuse) 1312 return -ENOMEM; 1313 1314 for_each_netdev(net, d) { 1315 struct netdev_name_node *name_node; 1316 1317 netdev_for_each_altname(d, name_node) { 1318 if (!sscanf(name_node->name, name, &i)) 1319 continue; 1320 if (i < 0 || i >= max_netdevices) 1321 continue; 1322 1323 /* avoid cases where sscanf is not exact inverse of printf */ 1324 snprintf(buf, IFNAMSIZ, name, i); 1325 if (!strncmp(buf, name_node->name, IFNAMSIZ)) 1326 __set_bit(i, inuse); 1327 } 1328 if (!sscanf(d->name, name, &i)) 1329 continue; 1330 if (i < 0 || i >= max_netdevices) 1331 continue; 1332 1333 /* avoid cases where sscanf is not exact inverse of printf */ 1334 snprintf(buf, IFNAMSIZ, name, i); 1335 if (!strncmp(buf, d->name, IFNAMSIZ)) 1336 __set_bit(i, inuse); 1337 } 1338 1339 i = find_first_zero_bit(inuse, max_netdevices); 1340 bitmap_free(inuse); 1341 if (i == max_netdevices) 1342 return -ENFILE; 1343 1344 /* 'res' and 'name' could overlap, use 'buf' as an intermediate buffer */ 1345 strscpy(buf, name, IFNAMSIZ); 1346 snprintf(res, IFNAMSIZ, buf, i); 1347 return i; 1348 } 1349 1350 /* Returns negative errno or allocated unit id (see __dev_alloc_name()) */ 1351 static int dev_prep_valid_name(struct net *net, struct net_device *dev, 1352 const char *want_name, char *out_name, 1353 int dup_errno) 1354 { 1355 if (!dev_valid_name(want_name)) 1356 return -EINVAL; 1357 1358 if (strchr(want_name, '%')) 1359 return __dev_alloc_name(net, want_name, out_name); 1360 1361 if (netdev_name_in_use(net, want_name)) 1362 return -dup_errno; 1363 if (out_name != want_name) 1364 strscpy(out_name, want_name, IFNAMSIZ); 1365 return 0; 1366 } 1367 1368 /** 1369 * dev_alloc_name - allocate a name for a device 1370 * @dev: device 1371 * @name: name format string 1372 * 1373 * Passed a format string - eg "lt%d" it will try and find a suitable 1374 * id. It scans list of devices to build up a free map, then chooses 1375 * the first empty slot. The caller must hold the dev_base or rtnl lock 1376 * while allocating the name and adding the device in order to avoid 1377 * duplicates. 1378 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1379 * Returns the number of the unit assigned or a negative errno code. 1380 */ 1381 1382 int dev_alloc_name(struct net_device *dev, const char *name) 1383 { 1384 return dev_prep_valid_name(dev_net(dev), dev, name, dev->name, ENFILE); 1385 } 1386 EXPORT_SYMBOL(dev_alloc_name); 1387 1388 static int dev_get_valid_name(struct net *net, struct net_device *dev, 1389 const char *name) 1390 { 1391 int ret; 1392 1393 ret = dev_prep_valid_name(net, dev, name, dev->name, EEXIST); 1394 return ret < 0 ? ret : 0; 1395 } 1396 1397 int netif_change_name(struct net_device *dev, const char *newname) 1398 { 1399 struct net *net = dev_net(dev); 1400 unsigned char old_assign_type; 1401 char oldname[IFNAMSIZ]; 1402 int err = 0; 1403 int ret; 1404 1405 ASSERT_RTNL_NET(net); 1406 1407 if (!strncmp(newname, dev->name, IFNAMSIZ)) 1408 return 0; 1409 1410 memcpy(oldname, dev->name, IFNAMSIZ); 1411 1412 write_seqlock_bh(&netdev_rename_lock); 1413 err = dev_get_valid_name(net, dev, newname); 1414 write_sequnlock_bh(&netdev_rename_lock); 1415 1416 if (err < 0) 1417 return err; 1418 1419 if (oldname[0] && !strchr(oldname, '%')) 1420 netdev_info(dev, "renamed from %s%s\n", oldname, 1421 dev->flags & IFF_UP ? " (while UP)" : ""); 1422 1423 old_assign_type = dev->name_assign_type; 1424 WRITE_ONCE(dev->name_assign_type, NET_NAME_RENAMED); 1425 1426 rollback: 1427 ret = device_rename(&dev->dev, dev->name); 1428 if (ret) { 1429 write_seqlock_bh(&netdev_rename_lock); 1430 memcpy(dev->name, oldname, IFNAMSIZ); 1431 write_sequnlock_bh(&netdev_rename_lock); 1432 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1433 return ret; 1434 } 1435 1436 netdev_adjacent_rename_links(dev, oldname); 1437 1438 netdev_name_node_del(dev->name_node); 1439 1440 synchronize_net(); 1441 1442 netdev_name_node_add(net, dev->name_node); 1443 1444 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev); 1445 ret = notifier_to_errno(ret); 1446 1447 if (ret) { 1448 /* err >= 0 after dev_alloc_name() or stores the first errno */ 1449 if (err >= 0) { 1450 err = ret; 1451 write_seqlock_bh(&netdev_rename_lock); 1452 memcpy(dev->name, oldname, IFNAMSIZ); 1453 write_sequnlock_bh(&netdev_rename_lock); 1454 memcpy(oldname, newname, IFNAMSIZ); 1455 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1456 old_assign_type = NET_NAME_RENAMED; 1457 goto rollback; 1458 } else { 1459 netdev_err(dev, "name change rollback failed: %d\n", 1460 ret); 1461 } 1462 } 1463 1464 return err; 1465 } 1466 1467 int netif_set_alias(struct net_device *dev, const char *alias, size_t len) 1468 { 1469 struct dev_ifalias *new_alias = NULL; 1470 1471 if (len >= IFALIASZ) 1472 return -EINVAL; 1473 1474 if (len) { 1475 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL); 1476 if (!new_alias) 1477 return -ENOMEM; 1478 1479 memcpy(new_alias->ifalias, alias, len); 1480 new_alias->ifalias[len] = 0; 1481 } 1482 1483 mutex_lock(&ifalias_mutex); 1484 new_alias = rcu_replace_pointer(dev->ifalias, new_alias, 1485 mutex_is_locked(&ifalias_mutex)); 1486 mutex_unlock(&ifalias_mutex); 1487 1488 if (new_alias) 1489 kfree_rcu(new_alias, rcuhead); 1490 1491 return len; 1492 } 1493 1494 /** 1495 * dev_get_alias - get ifalias of a device 1496 * @dev: device 1497 * @name: buffer to store name of ifalias 1498 * @len: size of buffer 1499 * 1500 * get ifalias for a device. Caller must make sure dev cannot go 1501 * away, e.g. rcu read lock or own a reference count to device. 1502 */ 1503 int dev_get_alias(const struct net_device *dev, char *name, size_t len) 1504 { 1505 const struct dev_ifalias *alias; 1506 int ret = 0; 1507 1508 rcu_read_lock(); 1509 alias = rcu_dereference(dev->ifalias); 1510 if (alias) 1511 ret = snprintf(name, len, "%s", alias->ifalias); 1512 rcu_read_unlock(); 1513 1514 return ret; 1515 } 1516 1517 /** 1518 * netdev_features_change - device changes features 1519 * @dev: device to cause notification 1520 * 1521 * Called to indicate a device has changed features. 1522 */ 1523 void netdev_features_change(struct net_device *dev) 1524 { 1525 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev); 1526 } 1527 EXPORT_SYMBOL(netdev_features_change); 1528 1529 /** 1530 * netdev_state_change - device changes state 1531 * @dev: device to cause notification 1532 * 1533 * Called to indicate a device has changed state. This function calls 1534 * the notifier chains for netdev_chain and sends a NEWLINK message 1535 * to the routing socket. 1536 */ 1537 void netdev_state_change(struct net_device *dev) 1538 { 1539 if (dev->flags & IFF_UP) { 1540 struct netdev_notifier_change_info change_info = { 1541 .info.dev = dev, 1542 }; 1543 1544 call_netdevice_notifiers_info(NETDEV_CHANGE, 1545 &change_info.info); 1546 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL, 0, NULL); 1547 } 1548 } 1549 EXPORT_SYMBOL(netdev_state_change); 1550 1551 /** 1552 * __netdev_notify_peers - notify network peers about existence of @dev, 1553 * to be called when rtnl lock is already held. 1554 * @dev: network device 1555 * 1556 * Generate traffic such that interested network peers are aware of 1557 * @dev, such as by generating a gratuitous ARP. This may be used when 1558 * a device wants to inform the rest of the network about some sort of 1559 * reconfiguration such as a failover event or virtual machine 1560 * migration. 1561 */ 1562 void __netdev_notify_peers(struct net_device *dev) 1563 { 1564 ASSERT_RTNL(); 1565 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev); 1566 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev); 1567 } 1568 EXPORT_SYMBOL(__netdev_notify_peers); 1569 1570 /** 1571 * netdev_notify_peers - notify network peers about existence of @dev 1572 * @dev: network device 1573 * 1574 * Generate traffic such that interested network peers are aware of 1575 * @dev, such as by generating a gratuitous ARP. This may be used when 1576 * a device wants to inform the rest of the network about some sort of 1577 * reconfiguration such as a failover event or virtual machine 1578 * migration. 1579 */ 1580 void netdev_notify_peers(struct net_device *dev) 1581 { 1582 rtnl_lock(); 1583 __netdev_notify_peers(dev); 1584 rtnl_unlock(); 1585 } 1586 EXPORT_SYMBOL(netdev_notify_peers); 1587 1588 static int napi_threaded_poll(void *data); 1589 1590 static int napi_kthread_create(struct napi_struct *n) 1591 { 1592 int err = 0; 1593 1594 /* Create and wake up the kthread once to put it in 1595 * TASK_INTERRUPTIBLE mode to avoid the blocked task 1596 * warning and work with loadavg. 1597 */ 1598 n->thread = kthread_run(napi_threaded_poll, n, "napi/%s-%d", 1599 n->dev->name, n->napi_id); 1600 if (IS_ERR(n->thread)) { 1601 err = PTR_ERR(n->thread); 1602 pr_err("kthread_run failed with err %d\n", err); 1603 n->thread = NULL; 1604 } 1605 1606 return err; 1607 } 1608 1609 static int __dev_open(struct net_device *dev, struct netlink_ext_ack *extack) 1610 { 1611 const struct net_device_ops *ops = dev->netdev_ops; 1612 int ret; 1613 1614 ASSERT_RTNL(); 1615 dev_addr_check(dev); 1616 1617 if (!netif_device_present(dev)) { 1618 /* may be detached because parent is runtime-suspended */ 1619 if (dev->dev.parent) 1620 pm_runtime_resume(dev->dev.parent); 1621 if (!netif_device_present(dev)) 1622 return -ENODEV; 1623 } 1624 1625 /* Block netpoll from trying to do any rx path servicing. 1626 * If we don't do this there is a chance ndo_poll_controller 1627 * or ndo_poll may be running while we open the device 1628 */ 1629 netpoll_poll_disable(dev); 1630 1631 ret = call_netdevice_notifiers_extack(NETDEV_PRE_UP, dev, extack); 1632 ret = notifier_to_errno(ret); 1633 if (ret) 1634 return ret; 1635 1636 set_bit(__LINK_STATE_START, &dev->state); 1637 1638 netdev_ops_assert_locked(dev); 1639 1640 if (ops->ndo_validate_addr) 1641 ret = ops->ndo_validate_addr(dev); 1642 1643 if (!ret && ops->ndo_open) 1644 ret = ops->ndo_open(dev); 1645 1646 netpoll_poll_enable(dev); 1647 1648 if (ret) 1649 clear_bit(__LINK_STATE_START, &dev->state); 1650 else { 1651 netif_set_up(dev, true); 1652 dev_set_rx_mode(dev); 1653 dev_activate(dev); 1654 add_device_randomness(dev->dev_addr, dev->addr_len); 1655 } 1656 1657 return ret; 1658 } 1659 1660 int netif_open(struct net_device *dev, struct netlink_ext_ack *extack) 1661 { 1662 int ret; 1663 1664 if (dev->flags & IFF_UP) 1665 return 0; 1666 1667 ret = __dev_open(dev, extack); 1668 if (ret < 0) 1669 return ret; 1670 1671 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1672 call_netdevice_notifiers(NETDEV_UP, dev); 1673 1674 return ret; 1675 } 1676 1677 static void __dev_close_many(struct list_head *head) 1678 { 1679 struct net_device *dev; 1680 1681 ASSERT_RTNL(); 1682 might_sleep(); 1683 1684 list_for_each_entry(dev, head, close_list) { 1685 /* Temporarily disable netpoll until the interface is down */ 1686 netpoll_poll_disable(dev); 1687 1688 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev); 1689 1690 clear_bit(__LINK_STATE_START, &dev->state); 1691 1692 /* Synchronize to scheduled poll. We cannot touch poll list, it 1693 * can be even on different cpu. So just clear netif_running(). 1694 * 1695 * dev->stop() will invoke napi_disable() on all of it's 1696 * napi_struct instances on this device. 1697 */ 1698 smp_mb__after_atomic(); /* Commit netif_running(). */ 1699 } 1700 1701 dev_deactivate_many(head); 1702 1703 list_for_each_entry(dev, head, close_list) { 1704 const struct net_device_ops *ops = dev->netdev_ops; 1705 1706 /* 1707 * Call the device specific close. This cannot fail. 1708 * Only if device is UP 1709 * 1710 * We allow it to be called even after a DETACH hot-plug 1711 * event. 1712 */ 1713 1714 netdev_ops_assert_locked(dev); 1715 1716 if (ops->ndo_stop) 1717 ops->ndo_stop(dev); 1718 1719 netif_set_up(dev, false); 1720 netpoll_poll_enable(dev); 1721 } 1722 } 1723 1724 static void __dev_close(struct net_device *dev) 1725 { 1726 LIST_HEAD(single); 1727 1728 list_add(&dev->close_list, &single); 1729 __dev_close_many(&single); 1730 list_del(&single); 1731 } 1732 1733 void dev_close_many(struct list_head *head, bool unlink) 1734 { 1735 struct net_device *dev, *tmp; 1736 1737 /* Remove the devices that don't need to be closed */ 1738 list_for_each_entry_safe(dev, tmp, head, close_list) 1739 if (!(dev->flags & IFF_UP)) 1740 list_del_init(&dev->close_list); 1741 1742 __dev_close_many(head); 1743 1744 list_for_each_entry_safe(dev, tmp, head, close_list) { 1745 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1746 call_netdevice_notifiers(NETDEV_DOWN, dev); 1747 if (unlink) 1748 list_del_init(&dev->close_list); 1749 } 1750 } 1751 EXPORT_SYMBOL(dev_close_many); 1752 1753 void netif_close(struct net_device *dev) 1754 { 1755 if (dev->flags & IFF_UP) { 1756 LIST_HEAD(single); 1757 1758 list_add(&dev->close_list, &single); 1759 dev_close_many(&single, true); 1760 list_del(&single); 1761 } 1762 } 1763 EXPORT_SYMBOL(netif_close); 1764 1765 void netif_disable_lro(struct net_device *dev) 1766 { 1767 struct net_device *lower_dev; 1768 struct list_head *iter; 1769 1770 dev->wanted_features &= ~NETIF_F_LRO; 1771 netdev_update_features(dev); 1772 1773 if (unlikely(dev->features & NETIF_F_LRO)) 1774 netdev_WARN(dev, "failed to disable LRO!\n"); 1775 1776 netdev_for_each_lower_dev(dev, lower_dev, iter) { 1777 netdev_lock_ops(lower_dev); 1778 netif_disable_lro(lower_dev); 1779 netdev_unlock_ops(lower_dev); 1780 } 1781 } 1782 1783 /** 1784 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device 1785 * @dev: device 1786 * 1787 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be 1788 * called under RTNL. This is needed if Generic XDP is installed on 1789 * the device. 1790 */ 1791 static void dev_disable_gro_hw(struct net_device *dev) 1792 { 1793 dev->wanted_features &= ~NETIF_F_GRO_HW; 1794 netdev_update_features(dev); 1795 1796 if (unlikely(dev->features & NETIF_F_GRO_HW)) 1797 netdev_WARN(dev, "failed to disable GRO_HW!\n"); 1798 } 1799 1800 const char *netdev_cmd_to_name(enum netdev_cmd cmd) 1801 { 1802 #define N(val) \ 1803 case NETDEV_##val: \ 1804 return "NETDEV_" __stringify(val); 1805 switch (cmd) { 1806 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER) 1807 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE) 1808 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE) 1809 N(POST_INIT) N(PRE_UNINIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) 1810 N(CHANGEUPPER) N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) 1811 N(BONDING_INFO) N(PRECHANGEUPPER) N(CHANGELOWERSTATE) 1812 N(UDP_TUNNEL_PUSH_INFO) N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN) 1813 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO) 1814 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO) 1815 N(PRE_CHANGEADDR) N(OFFLOAD_XSTATS_ENABLE) N(OFFLOAD_XSTATS_DISABLE) 1816 N(OFFLOAD_XSTATS_REPORT_USED) N(OFFLOAD_XSTATS_REPORT_DELTA) 1817 N(XDP_FEAT_CHANGE) 1818 } 1819 #undef N 1820 return "UNKNOWN_NETDEV_EVENT"; 1821 } 1822 EXPORT_SYMBOL_GPL(netdev_cmd_to_name); 1823 1824 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val, 1825 struct net_device *dev) 1826 { 1827 struct netdev_notifier_info info = { 1828 .dev = dev, 1829 }; 1830 1831 return nb->notifier_call(nb, val, &info); 1832 } 1833 1834 static int call_netdevice_register_notifiers(struct notifier_block *nb, 1835 struct net_device *dev) 1836 { 1837 int err; 1838 1839 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev); 1840 err = notifier_to_errno(err); 1841 if (err) 1842 return err; 1843 1844 if (!(dev->flags & IFF_UP)) 1845 return 0; 1846 1847 call_netdevice_notifier(nb, NETDEV_UP, dev); 1848 return 0; 1849 } 1850 1851 static void call_netdevice_unregister_notifiers(struct notifier_block *nb, 1852 struct net_device *dev) 1853 { 1854 if (dev->flags & IFF_UP) { 1855 call_netdevice_notifier(nb, NETDEV_GOING_DOWN, 1856 dev); 1857 call_netdevice_notifier(nb, NETDEV_DOWN, dev); 1858 } 1859 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev); 1860 } 1861 1862 static int call_netdevice_register_net_notifiers(struct notifier_block *nb, 1863 struct net *net) 1864 { 1865 struct net_device *dev; 1866 int err; 1867 1868 for_each_netdev(net, dev) { 1869 err = call_netdevice_register_notifiers(nb, dev); 1870 if (err) 1871 goto rollback; 1872 } 1873 return 0; 1874 1875 rollback: 1876 for_each_netdev_continue_reverse(net, dev) 1877 call_netdevice_unregister_notifiers(nb, dev); 1878 return err; 1879 } 1880 1881 static void call_netdevice_unregister_net_notifiers(struct notifier_block *nb, 1882 struct net *net) 1883 { 1884 struct net_device *dev; 1885 1886 for_each_netdev(net, dev) 1887 call_netdevice_unregister_notifiers(nb, dev); 1888 } 1889 1890 static int dev_boot_phase = 1; 1891 1892 /** 1893 * register_netdevice_notifier - register a network notifier block 1894 * @nb: notifier 1895 * 1896 * Register a notifier to be called when network device events occur. 1897 * The notifier passed is linked into the kernel structures and must 1898 * not be reused until it has been unregistered. A negative errno code 1899 * is returned on a failure. 1900 * 1901 * When registered all registration and up events are replayed 1902 * to the new notifier to allow device to have a race free 1903 * view of the network device list. 1904 */ 1905 1906 int register_netdevice_notifier(struct notifier_block *nb) 1907 { 1908 struct net *net; 1909 int err; 1910 1911 /* Close race with setup_net() and cleanup_net() */ 1912 down_write(&pernet_ops_rwsem); 1913 1914 /* When RTNL is removed, we need protection for netdev_chain. */ 1915 rtnl_lock(); 1916 1917 err = raw_notifier_chain_register(&netdev_chain, nb); 1918 if (err) 1919 goto unlock; 1920 if (dev_boot_phase) 1921 goto unlock; 1922 for_each_net(net) { 1923 __rtnl_net_lock(net); 1924 err = call_netdevice_register_net_notifiers(nb, net); 1925 __rtnl_net_unlock(net); 1926 if (err) 1927 goto rollback; 1928 } 1929 1930 unlock: 1931 rtnl_unlock(); 1932 up_write(&pernet_ops_rwsem); 1933 return err; 1934 1935 rollback: 1936 for_each_net_continue_reverse(net) { 1937 __rtnl_net_lock(net); 1938 call_netdevice_unregister_net_notifiers(nb, net); 1939 __rtnl_net_unlock(net); 1940 } 1941 1942 raw_notifier_chain_unregister(&netdev_chain, nb); 1943 goto unlock; 1944 } 1945 EXPORT_SYMBOL(register_netdevice_notifier); 1946 1947 /** 1948 * unregister_netdevice_notifier - unregister a network notifier block 1949 * @nb: notifier 1950 * 1951 * Unregister a notifier previously registered by 1952 * register_netdevice_notifier(). The notifier is unlinked into the 1953 * kernel structures and may then be reused. A negative errno code 1954 * is returned on a failure. 1955 * 1956 * After unregistering unregister and down device events are synthesized 1957 * for all devices on the device list to the removed notifier to remove 1958 * the need for special case cleanup code. 1959 */ 1960 1961 int unregister_netdevice_notifier(struct notifier_block *nb) 1962 { 1963 struct net *net; 1964 int err; 1965 1966 /* Close race with setup_net() and cleanup_net() */ 1967 down_write(&pernet_ops_rwsem); 1968 rtnl_lock(); 1969 err = raw_notifier_chain_unregister(&netdev_chain, nb); 1970 if (err) 1971 goto unlock; 1972 1973 for_each_net(net) { 1974 __rtnl_net_lock(net); 1975 call_netdevice_unregister_net_notifiers(nb, net); 1976 __rtnl_net_unlock(net); 1977 } 1978 1979 unlock: 1980 rtnl_unlock(); 1981 up_write(&pernet_ops_rwsem); 1982 return err; 1983 } 1984 EXPORT_SYMBOL(unregister_netdevice_notifier); 1985 1986 static int __register_netdevice_notifier_net(struct net *net, 1987 struct notifier_block *nb, 1988 bool ignore_call_fail) 1989 { 1990 int err; 1991 1992 err = raw_notifier_chain_register(&net->netdev_chain, nb); 1993 if (err) 1994 return err; 1995 if (dev_boot_phase) 1996 return 0; 1997 1998 err = call_netdevice_register_net_notifiers(nb, net); 1999 if (err && !ignore_call_fail) 2000 goto chain_unregister; 2001 2002 return 0; 2003 2004 chain_unregister: 2005 raw_notifier_chain_unregister(&net->netdev_chain, nb); 2006 return err; 2007 } 2008 2009 static int __unregister_netdevice_notifier_net(struct net *net, 2010 struct notifier_block *nb) 2011 { 2012 int err; 2013 2014 err = raw_notifier_chain_unregister(&net->netdev_chain, nb); 2015 if (err) 2016 return err; 2017 2018 call_netdevice_unregister_net_notifiers(nb, net); 2019 return 0; 2020 } 2021 2022 /** 2023 * register_netdevice_notifier_net - register a per-netns network notifier block 2024 * @net: network namespace 2025 * @nb: notifier 2026 * 2027 * Register a notifier to be called when network device events occur. 2028 * The notifier passed is linked into the kernel structures and must 2029 * not be reused until it has been unregistered. A negative errno code 2030 * is returned on a failure. 2031 * 2032 * When registered all registration and up events are replayed 2033 * to the new notifier to allow device to have a race free 2034 * view of the network device list. 2035 */ 2036 2037 int register_netdevice_notifier_net(struct net *net, struct notifier_block *nb) 2038 { 2039 int err; 2040 2041 rtnl_net_lock(net); 2042 err = __register_netdevice_notifier_net(net, nb, false); 2043 rtnl_net_unlock(net); 2044 2045 return err; 2046 } 2047 EXPORT_SYMBOL(register_netdevice_notifier_net); 2048 2049 /** 2050 * unregister_netdevice_notifier_net - unregister a per-netns 2051 * network notifier block 2052 * @net: network namespace 2053 * @nb: notifier 2054 * 2055 * Unregister a notifier previously registered by 2056 * register_netdevice_notifier_net(). The notifier is unlinked from the 2057 * kernel structures and may then be reused. A negative errno code 2058 * is returned on a failure. 2059 * 2060 * After unregistering unregister and down device events are synthesized 2061 * for all devices on the device list to the removed notifier to remove 2062 * the need for special case cleanup code. 2063 */ 2064 2065 int unregister_netdevice_notifier_net(struct net *net, 2066 struct notifier_block *nb) 2067 { 2068 int err; 2069 2070 rtnl_net_lock(net); 2071 err = __unregister_netdevice_notifier_net(net, nb); 2072 rtnl_net_unlock(net); 2073 2074 return err; 2075 } 2076 EXPORT_SYMBOL(unregister_netdevice_notifier_net); 2077 2078 static void __move_netdevice_notifier_net(struct net *src_net, 2079 struct net *dst_net, 2080 struct notifier_block *nb) 2081 { 2082 __unregister_netdevice_notifier_net(src_net, nb); 2083 __register_netdevice_notifier_net(dst_net, nb, true); 2084 } 2085 2086 static void rtnl_net_dev_lock(struct net_device *dev) 2087 { 2088 bool again; 2089 2090 do { 2091 struct net *net; 2092 2093 again = false; 2094 2095 /* netns might be being dismantled. */ 2096 rcu_read_lock(); 2097 net = dev_net_rcu(dev); 2098 net_passive_inc(net); 2099 rcu_read_unlock(); 2100 2101 rtnl_net_lock(net); 2102 2103 #ifdef CONFIG_NET_NS 2104 /* dev might have been moved to another netns. */ 2105 if (!net_eq(net, rcu_access_pointer(dev->nd_net.net))) { 2106 rtnl_net_unlock(net); 2107 net_passive_dec(net); 2108 again = true; 2109 } 2110 #endif 2111 } while (again); 2112 } 2113 2114 static void rtnl_net_dev_unlock(struct net_device *dev) 2115 { 2116 struct net *net = dev_net(dev); 2117 2118 rtnl_net_unlock(net); 2119 net_passive_dec(net); 2120 } 2121 2122 int register_netdevice_notifier_dev_net(struct net_device *dev, 2123 struct notifier_block *nb, 2124 struct netdev_net_notifier *nn) 2125 { 2126 int err; 2127 2128 rtnl_net_dev_lock(dev); 2129 err = __register_netdevice_notifier_net(dev_net(dev), nb, false); 2130 if (!err) { 2131 nn->nb = nb; 2132 list_add(&nn->list, &dev->net_notifier_list); 2133 } 2134 rtnl_net_dev_unlock(dev); 2135 2136 return err; 2137 } 2138 EXPORT_SYMBOL(register_netdevice_notifier_dev_net); 2139 2140 int unregister_netdevice_notifier_dev_net(struct net_device *dev, 2141 struct notifier_block *nb, 2142 struct netdev_net_notifier *nn) 2143 { 2144 int err; 2145 2146 rtnl_net_dev_lock(dev); 2147 list_del(&nn->list); 2148 err = __unregister_netdevice_notifier_net(dev_net(dev), nb); 2149 rtnl_net_dev_unlock(dev); 2150 2151 return err; 2152 } 2153 EXPORT_SYMBOL(unregister_netdevice_notifier_dev_net); 2154 2155 static void move_netdevice_notifiers_dev_net(struct net_device *dev, 2156 struct net *net) 2157 { 2158 struct netdev_net_notifier *nn; 2159 2160 list_for_each_entry(nn, &dev->net_notifier_list, list) 2161 __move_netdevice_notifier_net(dev_net(dev), net, nn->nb); 2162 } 2163 2164 /** 2165 * call_netdevice_notifiers_info - call all network notifier blocks 2166 * @val: value passed unmodified to notifier function 2167 * @info: notifier information data 2168 * 2169 * Call all network notifier blocks. Parameters and return value 2170 * are as for raw_notifier_call_chain(). 2171 */ 2172 2173 int call_netdevice_notifiers_info(unsigned long val, 2174 struct netdev_notifier_info *info) 2175 { 2176 struct net *net = dev_net(info->dev); 2177 int ret; 2178 2179 ASSERT_RTNL(); 2180 2181 /* Run per-netns notifier block chain first, then run the global one. 2182 * Hopefully, one day, the global one is going to be removed after 2183 * all notifier block registrators get converted to be per-netns. 2184 */ 2185 ret = raw_notifier_call_chain(&net->netdev_chain, val, info); 2186 if (ret & NOTIFY_STOP_MASK) 2187 return ret; 2188 return raw_notifier_call_chain(&netdev_chain, val, info); 2189 } 2190 2191 /** 2192 * call_netdevice_notifiers_info_robust - call per-netns notifier blocks 2193 * for and rollback on error 2194 * @val_up: value passed unmodified to notifier function 2195 * @val_down: value passed unmodified to the notifier function when 2196 * recovering from an error on @val_up 2197 * @info: notifier information data 2198 * 2199 * Call all per-netns network notifier blocks, but not notifier blocks on 2200 * the global notifier chain. Parameters and return value are as for 2201 * raw_notifier_call_chain_robust(). 2202 */ 2203 2204 static int 2205 call_netdevice_notifiers_info_robust(unsigned long val_up, 2206 unsigned long val_down, 2207 struct netdev_notifier_info *info) 2208 { 2209 struct net *net = dev_net(info->dev); 2210 2211 ASSERT_RTNL(); 2212 2213 return raw_notifier_call_chain_robust(&net->netdev_chain, 2214 val_up, val_down, info); 2215 } 2216 2217 static int call_netdevice_notifiers_extack(unsigned long val, 2218 struct net_device *dev, 2219 struct netlink_ext_ack *extack) 2220 { 2221 struct netdev_notifier_info info = { 2222 .dev = dev, 2223 .extack = extack, 2224 }; 2225 2226 return call_netdevice_notifiers_info(val, &info); 2227 } 2228 2229 /** 2230 * call_netdevice_notifiers - call all network notifier blocks 2231 * @val: value passed unmodified to notifier function 2232 * @dev: net_device pointer passed unmodified to notifier function 2233 * 2234 * Call all network notifier blocks. Parameters and return value 2235 * are as for raw_notifier_call_chain(). 2236 */ 2237 2238 int call_netdevice_notifiers(unsigned long val, struct net_device *dev) 2239 { 2240 return call_netdevice_notifiers_extack(val, dev, NULL); 2241 } 2242 EXPORT_SYMBOL(call_netdevice_notifiers); 2243 2244 /** 2245 * call_netdevice_notifiers_mtu - call all network notifier blocks 2246 * @val: value passed unmodified to notifier function 2247 * @dev: net_device pointer passed unmodified to notifier function 2248 * @arg: additional u32 argument passed to the notifier function 2249 * 2250 * Call all network notifier blocks. Parameters and return value 2251 * are as for raw_notifier_call_chain(). 2252 */ 2253 static int call_netdevice_notifiers_mtu(unsigned long val, 2254 struct net_device *dev, u32 arg) 2255 { 2256 struct netdev_notifier_info_ext info = { 2257 .info.dev = dev, 2258 .ext.mtu = arg, 2259 }; 2260 2261 BUILD_BUG_ON(offsetof(struct netdev_notifier_info_ext, info) != 0); 2262 2263 return call_netdevice_notifiers_info(val, &info.info); 2264 } 2265 2266 #ifdef CONFIG_NET_INGRESS 2267 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key); 2268 2269 void net_inc_ingress_queue(void) 2270 { 2271 static_branch_inc(&ingress_needed_key); 2272 } 2273 EXPORT_SYMBOL_GPL(net_inc_ingress_queue); 2274 2275 void net_dec_ingress_queue(void) 2276 { 2277 static_branch_dec(&ingress_needed_key); 2278 } 2279 EXPORT_SYMBOL_GPL(net_dec_ingress_queue); 2280 #endif 2281 2282 #ifdef CONFIG_NET_EGRESS 2283 static DEFINE_STATIC_KEY_FALSE(egress_needed_key); 2284 2285 void net_inc_egress_queue(void) 2286 { 2287 static_branch_inc(&egress_needed_key); 2288 } 2289 EXPORT_SYMBOL_GPL(net_inc_egress_queue); 2290 2291 void net_dec_egress_queue(void) 2292 { 2293 static_branch_dec(&egress_needed_key); 2294 } 2295 EXPORT_SYMBOL_GPL(net_dec_egress_queue); 2296 #endif 2297 2298 #ifdef CONFIG_NET_CLS_ACT 2299 DEFINE_STATIC_KEY_FALSE(tcf_sw_enabled_key); 2300 EXPORT_SYMBOL(tcf_sw_enabled_key); 2301 #endif 2302 2303 DEFINE_STATIC_KEY_FALSE(netstamp_needed_key); 2304 EXPORT_SYMBOL(netstamp_needed_key); 2305 #ifdef CONFIG_JUMP_LABEL 2306 static atomic_t netstamp_needed_deferred; 2307 static atomic_t netstamp_wanted; 2308 static void netstamp_clear(struct work_struct *work) 2309 { 2310 int deferred = atomic_xchg(&netstamp_needed_deferred, 0); 2311 int wanted; 2312 2313 wanted = atomic_add_return(deferred, &netstamp_wanted); 2314 if (wanted > 0) 2315 static_branch_enable(&netstamp_needed_key); 2316 else 2317 static_branch_disable(&netstamp_needed_key); 2318 } 2319 static DECLARE_WORK(netstamp_work, netstamp_clear); 2320 #endif 2321 2322 void net_enable_timestamp(void) 2323 { 2324 #ifdef CONFIG_JUMP_LABEL 2325 int wanted = atomic_read(&netstamp_wanted); 2326 2327 while (wanted > 0) { 2328 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted + 1)) 2329 return; 2330 } 2331 atomic_inc(&netstamp_needed_deferred); 2332 schedule_work(&netstamp_work); 2333 #else 2334 static_branch_inc(&netstamp_needed_key); 2335 #endif 2336 } 2337 EXPORT_SYMBOL(net_enable_timestamp); 2338 2339 void net_disable_timestamp(void) 2340 { 2341 #ifdef CONFIG_JUMP_LABEL 2342 int wanted = atomic_read(&netstamp_wanted); 2343 2344 while (wanted > 1) { 2345 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted - 1)) 2346 return; 2347 } 2348 atomic_dec(&netstamp_needed_deferred); 2349 schedule_work(&netstamp_work); 2350 #else 2351 static_branch_dec(&netstamp_needed_key); 2352 #endif 2353 } 2354 EXPORT_SYMBOL(net_disable_timestamp); 2355 2356 static inline void net_timestamp_set(struct sk_buff *skb) 2357 { 2358 skb->tstamp = 0; 2359 skb->tstamp_type = SKB_CLOCK_REALTIME; 2360 if (static_branch_unlikely(&netstamp_needed_key)) 2361 skb->tstamp = ktime_get_real(); 2362 } 2363 2364 #define net_timestamp_check(COND, SKB) \ 2365 if (static_branch_unlikely(&netstamp_needed_key)) { \ 2366 if ((COND) && !(SKB)->tstamp) \ 2367 (SKB)->tstamp = ktime_get_real(); \ 2368 } \ 2369 2370 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb) 2371 { 2372 return __is_skb_forwardable(dev, skb, true); 2373 } 2374 EXPORT_SYMBOL_GPL(is_skb_forwardable); 2375 2376 static int __dev_forward_skb2(struct net_device *dev, struct sk_buff *skb, 2377 bool check_mtu) 2378 { 2379 int ret = ____dev_forward_skb(dev, skb, check_mtu); 2380 2381 if (likely(!ret)) { 2382 skb->protocol = eth_type_trans(skb, dev); 2383 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN); 2384 } 2385 2386 return ret; 2387 } 2388 2389 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2390 { 2391 return __dev_forward_skb2(dev, skb, true); 2392 } 2393 EXPORT_SYMBOL_GPL(__dev_forward_skb); 2394 2395 /** 2396 * dev_forward_skb - loopback an skb to another netif 2397 * 2398 * @dev: destination network device 2399 * @skb: buffer to forward 2400 * 2401 * return values: 2402 * NET_RX_SUCCESS (no congestion) 2403 * NET_RX_DROP (packet was dropped, but freed) 2404 * 2405 * dev_forward_skb can be used for injecting an skb from the 2406 * start_xmit function of one device into the receive queue 2407 * of another device. 2408 * 2409 * The receiving device may be in another namespace, so 2410 * we have to clear all information in the skb that could 2411 * impact namespace isolation. 2412 */ 2413 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2414 { 2415 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb); 2416 } 2417 EXPORT_SYMBOL_GPL(dev_forward_skb); 2418 2419 int dev_forward_skb_nomtu(struct net_device *dev, struct sk_buff *skb) 2420 { 2421 return __dev_forward_skb2(dev, skb, false) ?: netif_rx_internal(skb); 2422 } 2423 2424 static inline int deliver_skb(struct sk_buff *skb, 2425 struct packet_type *pt_prev, 2426 struct net_device *orig_dev) 2427 { 2428 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 2429 return -ENOMEM; 2430 refcount_inc(&skb->users); 2431 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 2432 } 2433 2434 static inline void deliver_ptype_list_skb(struct sk_buff *skb, 2435 struct packet_type **pt, 2436 struct net_device *orig_dev, 2437 __be16 type, 2438 struct list_head *ptype_list) 2439 { 2440 struct packet_type *ptype, *pt_prev = *pt; 2441 2442 list_for_each_entry_rcu(ptype, ptype_list, list) { 2443 if (ptype->type != type) 2444 continue; 2445 if (pt_prev) 2446 deliver_skb(skb, pt_prev, orig_dev); 2447 pt_prev = ptype; 2448 } 2449 *pt = pt_prev; 2450 } 2451 2452 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb) 2453 { 2454 if (!ptype->af_packet_priv || !skb->sk) 2455 return false; 2456 2457 if (ptype->id_match) 2458 return ptype->id_match(ptype, skb->sk); 2459 else if ((struct sock *)ptype->af_packet_priv == skb->sk) 2460 return true; 2461 2462 return false; 2463 } 2464 2465 /** 2466 * dev_nit_active - return true if any network interface taps are in use 2467 * 2468 * @dev: network device to check for the presence of taps 2469 */ 2470 bool dev_nit_active(struct net_device *dev) 2471 { 2472 return !list_empty(&net_hotdata.ptype_all) || 2473 !list_empty(&dev->ptype_all); 2474 } 2475 EXPORT_SYMBOL_GPL(dev_nit_active); 2476 2477 /* 2478 * Support routine. Sends outgoing frames to any network 2479 * taps currently in use. 2480 */ 2481 2482 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev) 2483 { 2484 struct list_head *ptype_list = &net_hotdata.ptype_all; 2485 struct packet_type *ptype, *pt_prev = NULL; 2486 struct sk_buff *skb2 = NULL; 2487 2488 rcu_read_lock(); 2489 again: 2490 list_for_each_entry_rcu(ptype, ptype_list, list) { 2491 if (READ_ONCE(ptype->ignore_outgoing)) 2492 continue; 2493 2494 /* Never send packets back to the socket 2495 * they originated from - MvS ([email protected]) 2496 */ 2497 if (skb_loop_sk(ptype, skb)) 2498 continue; 2499 2500 if (pt_prev) { 2501 deliver_skb(skb2, pt_prev, skb->dev); 2502 pt_prev = ptype; 2503 continue; 2504 } 2505 2506 /* need to clone skb, done only once */ 2507 skb2 = skb_clone(skb, GFP_ATOMIC); 2508 if (!skb2) 2509 goto out_unlock; 2510 2511 net_timestamp_set(skb2); 2512 2513 /* skb->nh should be correctly 2514 * set by sender, so that the second statement is 2515 * just protection against buggy protocols. 2516 */ 2517 skb_reset_mac_header(skb2); 2518 2519 if (skb_network_header(skb2) < skb2->data || 2520 skb_network_header(skb2) > skb_tail_pointer(skb2)) { 2521 net_crit_ratelimited("protocol %04x is buggy, dev %s\n", 2522 ntohs(skb2->protocol), 2523 dev->name); 2524 skb_reset_network_header(skb2); 2525 } 2526 2527 skb2->transport_header = skb2->network_header; 2528 skb2->pkt_type = PACKET_OUTGOING; 2529 pt_prev = ptype; 2530 } 2531 2532 if (ptype_list == &net_hotdata.ptype_all) { 2533 ptype_list = &dev->ptype_all; 2534 goto again; 2535 } 2536 out_unlock: 2537 if (pt_prev) { 2538 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC)) 2539 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev); 2540 else 2541 kfree_skb(skb2); 2542 } 2543 rcu_read_unlock(); 2544 } 2545 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit); 2546 2547 /** 2548 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change 2549 * @dev: Network device 2550 * @txq: number of queues available 2551 * 2552 * If real_num_tx_queues is changed the tc mappings may no longer be 2553 * valid. To resolve this verify the tc mapping remains valid and if 2554 * not NULL the mapping. With no priorities mapping to this 2555 * offset/count pair it will no longer be used. In the worst case TC0 2556 * is invalid nothing can be done so disable priority mappings. If is 2557 * expected that drivers will fix this mapping if they can before 2558 * calling netif_set_real_num_tx_queues. 2559 */ 2560 static void netif_setup_tc(struct net_device *dev, unsigned int txq) 2561 { 2562 int i; 2563 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2564 2565 /* If TC0 is invalidated disable TC mapping */ 2566 if (tc->offset + tc->count > txq) { 2567 netdev_warn(dev, "Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n"); 2568 dev->num_tc = 0; 2569 return; 2570 } 2571 2572 /* Invalidated prio to tc mappings set to TC0 */ 2573 for (i = 1; i < TC_BITMASK + 1; i++) { 2574 int q = netdev_get_prio_tc_map(dev, i); 2575 2576 tc = &dev->tc_to_txq[q]; 2577 if (tc->offset + tc->count > txq) { 2578 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", 2579 i, q); 2580 netdev_set_prio_tc_map(dev, i, 0); 2581 } 2582 } 2583 } 2584 2585 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq) 2586 { 2587 if (dev->num_tc) { 2588 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2589 int i; 2590 2591 /* walk through the TCs and see if it falls into any of them */ 2592 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) { 2593 if ((txq - tc->offset) < tc->count) 2594 return i; 2595 } 2596 2597 /* didn't find it, just return -1 to indicate no match */ 2598 return -1; 2599 } 2600 2601 return 0; 2602 } 2603 EXPORT_SYMBOL(netdev_txq_to_tc); 2604 2605 #ifdef CONFIG_XPS 2606 static struct static_key xps_needed __read_mostly; 2607 static struct static_key xps_rxqs_needed __read_mostly; 2608 static DEFINE_MUTEX(xps_map_mutex); 2609 #define xmap_dereference(P) \ 2610 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex)) 2611 2612 static bool remove_xps_queue(struct xps_dev_maps *dev_maps, 2613 struct xps_dev_maps *old_maps, int tci, u16 index) 2614 { 2615 struct xps_map *map = NULL; 2616 int pos; 2617 2618 map = xmap_dereference(dev_maps->attr_map[tci]); 2619 if (!map) 2620 return false; 2621 2622 for (pos = map->len; pos--;) { 2623 if (map->queues[pos] != index) 2624 continue; 2625 2626 if (map->len > 1) { 2627 map->queues[pos] = map->queues[--map->len]; 2628 break; 2629 } 2630 2631 if (old_maps) 2632 RCU_INIT_POINTER(old_maps->attr_map[tci], NULL); 2633 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2634 kfree_rcu(map, rcu); 2635 return false; 2636 } 2637 2638 return true; 2639 } 2640 2641 static bool remove_xps_queue_cpu(struct net_device *dev, 2642 struct xps_dev_maps *dev_maps, 2643 int cpu, u16 offset, u16 count) 2644 { 2645 int num_tc = dev_maps->num_tc; 2646 bool active = false; 2647 int tci; 2648 2649 for (tci = cpu * num_tc; num_tc--; tci++) { 2650 int i, j; 2651 2652 for (i = count, j = offset; i--; j++) { 2653 if (!remove_xps_queue(dev_maps, NULL, tci, j)) 2654 break; 2655 } 2656 2657 active |= i < 0; 2658 } 2659 2660 return active; 2661 } 2662 2663 static void reset_xps_maps(struct net_device *dev, 2664 struct xps_dev_maps *dev_maps, 2665 enum xps_map_type type) 2666 { 2667 static_key_slow_dec_cpuslocked(&xps_needed); 2668 if (type == XPS_RXQS) 2669 static_key_slow_dec_cpuslocked(&xps_rxqs_needed); 2670 2671 RCU_INIT_POINTER(dev->xps_maps[type], NULL); 2672 2673 kfree_rcu(dev_maps, rcu); 2674 } 2675 2676 static void clean_xps_maps(struct net_device *dev, enum xps_map_type type, 2677 u16 offset, u16 count) 2678 { 2679 struct xps_dev_maps *dev_maps; 2680 bool active = false; 2681 int i, j; 2682 2683 dev_maps = xmap_dereference(dev->xps_maps[type]); 2684 if (!dev_maps) 2685 return; 2686 2687 for (j = 0; j < dev_maps->nr_ids; j++) 2688 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, count); 2689 if (!active) 2690 reset_xps_maps(dev, dev_maps, type); 2691 2692 if (type == XPS_CPUS) { 2693 for (i = offset + (count - 1); count--; i--) 2694 netdev_queue_numa_node_write( 2695 netdev_get_tx_queue(dev, i), NUMA_NO_NODE); 2696 } 2697 } 2698 2699 static void netif_reset_xps_queues(struct net_device *dev, u16 offset, 2700 u16 count) 2701 { 2702 if (!static_key_false(&xps_needed)) 2703 return; 2704 2705 cpus_read_lock(); 2706 mutex_lock(&xps_map_mutex); 2707 2708 if (static_key_false(&xps_rxqs_needed)) 2709 clean_xps_maps(dev, XPS_RXQS, offset, count); 2710 2711 clean_xps_maps(dev, XPS_CPUS, offset, count); 2712 2713 mutex_unlock(&xps_map_mutex); 2714 cpus_read_unlock(); 2715 } 2716 2717 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index) 2718 { 2719 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index); 2720 } 2721 2722 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index, 2723 u16 index, bool is_rxqs_map) 2724 { 2725 struct xps_map *new_map; 2726 int alloc_len = XPS_MIN_MAP_ALLOC; 2727 int i, pos; 2728 2729 for (pos = 0; map && pos < map->len; pos++) { 2730 if (map->queues[pos] != index) 2731 continue; 2732 return map; 2733 } 2734 2735 /* Need to add tx-queue to this CPU's/rx-queue's existing map */ 2736 if (map) { 2737 if (pos < map->alloc_len) 2738 return map; 2739 2740 alloc_len = map->alloc_len * 2; 2741 } 2742 2743 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's 2744 * map 2745 */ 2746 if (is_rxqs_map) 2747 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL); 2748 else 2749 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL, 2750 cpu_to_node(attr_index)); 2751 if (!new_map) 2752 return NULL; 2753 2754 for (i = 0; i < pos; i++) 2755 new_map->queues[i] = map->queues[i]; 2756 new_map->alloc_len = alloc_len; 2757 new_map->len = pos; 2758 2759 return new_map; 2760 } 2761 2762 /* Copy xps maps at a given index */ 2763 static void xps_copy_dev_maps(struct xps_dev_maps *dev_maps, 2764 struct xps_dev_maps *new_dev_maps, int index, 2765 int tc, bool skip_tc) 2766 { 2767 int i, tci = index * dev_maps->num_tc; 2768 struct xps_map *map; 2769 2770 /* copy maps belonging to foreign traffic classes */ 2771 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2772 if (i == tc && skip_tc) 2773 continue; 2774 2775 /* fill in the new device map from the old device map */ 2776 map = xmap_dereference(dev_maps->attr_map[tci]); 2777 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2778 } 2779 } 2780 2781 /* Must be called under cpus_read_lock */ 2782 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask, 2783 u16 index, enum xps_map_type type) 2784 { 2785 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL, *old_dev_maps = NULL; 2786 const unsigned long *online_mask = NULL; 2787 bool active = false, copy = false; 2788 int i, j, tci, numa_node_id = -2; 2789 int maps_sz, num_tc = 1, tc = 0; 2790 struct xps_map *map, *new_map; 2791 unsigned int nr_ids; 2792 2793 WARN_ON_ONCE(index >= dev->num_tx_queues); 2794 2795 if (dev->num_tc) { 2796 /* Do not allow XPS on subordinate device directly */ 2797 num_tc = dev->num_tc; 2798 if (num_tc < 0) 2799 return -EINVAL; 2800 2801 /* If queue belongs to subordinate dev use its map */ 2802 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev; 2803 2804 tc = netdev_txq_to_tc(dev, index); 2805 if (tc < 0) 2806 return -EINVAL; 2807 } 2808 2809 mutex_lock(&xps_map_mutex); 2810 2811 dev_maps = xmap_dereference(dev->xps_maps[type]); 2812 if (type == XPS_RXQS) { 2813 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues); 2814 nr_ids = dev->num_rx_queues; 2815 } else { 2816 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc); 2817 if (num_possible_cpus() > 1) 2818 online_mask = cpumask_bits(cpu_online_mask); 2819 nr_ids = nr_cpu_ids; 2820 } 2821 2822 if (maps_sz < L1_CACHE_BYTES) 2823 maps_sz = L1_CACHE_BYTES; 2824 2825 /* The old dev_maps could be larger or smaller than the one we're 2826 * setting up now, as dev->num_tc or nr_ids could have been updated in 2827 * between. We could try to be smart, but let's be safe instead and only 2828 * copy foreign traffic classes if the two map sizes match. 2829 */ 2830 if (dev_maps && 2831 dev_maps->num_tc == num_tc && dev_maps->nr_ids == nr_ids) 2832 copy = true; 2833 2834 /* allocate memory for queue storage */ 2835 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids), 2836 j < nr_ids;) { 2837 if (!new_dev_maps) { 2838 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL); 2839 if (!new_dev_maps) { 2840 mutex_unlock(&xps_map_mutex); 2841 return -ENOMEM; 2842 } 2843 2844 new_dev_maps->nr_ids = nr_ids; 2845 new_dev_maps->num_tc = num_tc; 2846 } 2847 2848 tci = j * num_tc + tc; 2849 map = copy ? xmap_dereference(dev_maps->attr_map[tci]) : NULL; 2850 2851 map = expand_xps_map(map, j, index, type == XPS_RXQS); 2852 if (!map) 2853 goto error; 2854 2855 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2856 } 2857 2858 if (!new_dev_maps) 2859 goto out_no_new_maps; 2860 2861 if (!dev_maps) { 2862 /* Increment static keys at most once per type */ 2863 static_key_slow_inc_cpuslocked(&xps_needed); 2864 if (type == XPS_RXQS) 2865 static_key_slow_inc_cpuslocked(&xps_rxqs_needed); 2866 } 2867 2868 for (j = 0; j < nr_ids; j++) { 2869 bool skip_tc = false; 2870 2871 tci = j * num_tc + tc; 2872 if (netif_attr_test_mask(j, mask, nr_ids) && 2873 netif_attr_test_online(j, online_mask, nr_ids)) { 2874 /* add tx-queue to CPU/rx-queue maps */ 2875 int pos = 0; 2876 2877 skip_tc = true; 2878 2879 map = xmap_dereference(new_dev_maps->attr_map[tci]); 2880 while ((pos < map->len) && (map->queues[pos] != index)) 2881 pos++; 2882 2883 if (pos == map->len) 2884 map->queues[map->len++] = index; 2885 #ifdef CONFIG_NUMA 2886 if (type == XPS_CPUS) { 2887 if (numa_node_id == -2) 2888 numa_node_id = cpu_to_node(j); 2889 else if (numa_node_id != cpu_to_node(j)) 2890 numa_node_id = -1; 2891 } 2892 #endif 2893 } 2894 2895 if (copy) 2896 xps_copy_dev_maps(dev_maps, new_dev_maps, j, tc, 2897 skip_tc); 2898 } 2899 2900 rcu_assign_pointer(dev->xps_maps[type], new_dev_maps); 2901 2902 /* Cleanup old maps */ 2903 if (!dev_maps) 2904 goto out_no_old_maps; 2905 2906 for (j = 0; j < dev_maps->nr_ids; j++) { 2907 for (i = num_tc, tci = j * dev_maps->num_tc; i--; tci++) { 2908 map = xmap_dereference(dev_maps->attr_map[tci]); 2909 if (!map) 2910 continue; 2911 2912 if (copy) { 2913 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2914 if (map == new_map) 2915 continue; 2916 } 2917 2918 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2919 kfree_rcu(map, rcu); 2920 } 2921 } 2922 2923 old_dev_maps = dev_maps; 2924 2925 out_no_old_maps: 2926 dev_maps = new_dev_maps; 2927 active = true; 2928 2929 out_no_new_maps: 2930 if (type == XPS_CPUS) 2931 /* update Tx queue numa node */ 2932 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index), 2933 (numa_node_id >= 0) ? 2934 numa_node_id : NUMA_NO_NODE); 2935 2936 if (!dev_maps) 2937 goto out_no_maps; 2938 2939 /* removes tx-queue from unused CPUs/rx-queues */ 2940 for (j = 0; j < dev_maps->nr_ids; j++) { 2941 tci = j * dev_maps->num_tc; 2942 2943 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2944 if (i == tc && 2945 netif_attr_test_mask(j, mask, dev_maps->nr_ids) && 2946 netif_attr_test_online(j, online_mask, dev_maps->nr_ids)) 2947 continue; 2948 2949 active |= remove_xps_queue(dev_maps, 2950 copy ? old_dev_maps : NULL, 2951 tci, index); 2952 } 2953 } 2954 2955 if (old_dev_maps) 2956 kfree_rcu(old_dev_maps, rcu); 2957 2958 /* free map if not active */ 2959 if (!active) 2960 reset_xps_maps(dev, dev_maps, type); 2961 2962 out_no_maps: 2963 mutex_unlock(&xps_map_mutex); 2964 2965 return 0; 2966 error: 2967 /* remove any maps that we added */ 2968 for (j = 0; j < nr_ids; j++) { 2969 for (i = num_tc, tci = j * num_tc; i--; tci++) { 2970 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2971 map = copy ? 2972 xmap_dereference(dev_maps->attr_map[tci]) : 2973 NULL; 2974 if (new_map && new_map != map) 2975 kfree(new_map); 2976 } 2977 } 2978 2979 mutex_unlock(&xps_map_mutex); 2980 2981 kfree(new_dev_maps); 2982 return -ENOMEM; 2983 } 2984 EXPORT_SYMBOL_GPL(__netif_set_xps_queue); 2985 2986 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, 2987 u16 index) 2988 { 2989 int ret; 2990 2991 cpus_read_lock(); 2992 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, XPS_CPUS); 2993 cpus_read_unlock(); 2994 2995 return ret; 2996 } 2997 EXPORT_SYMBOL(netif_set_xps_queue); 2998 2999 #endif 3000 static void netdev_unbind_all_sb_channels(struct net_device *dev) 3001 { 3002 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 3003 3004 /* Unbind any subordinate channels */ 3005 while (txq-- != &dev->_tx[0]) { 3006 if (txq->sb_dev) 3007 netdev_unbind_sb_channel(dev, txq->sb_dev); 3008 } 3009 } 3010 3011 void netdev_reset_tc(struct net_device *dev) 3012 { 3013 #ifdef CONFIG_XPS 3014 netif_reset_xps_queues_gt(dev, 0); 3015 #endif 3016 netdev_unbind_all_sb_channels(dev); 3017 3018 /* Reset TC configuration of device */ 3019 dev->num_tc = 0; 3020 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq)); 3021 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map)); 3022 } 3023 EXPORT_SYMBOL(netdev_reset_tc); 3024 3025 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset) 3026 { 3027 if (tc >= dev->num_tc) 3028 return -EINVAL; 3029 3030 #ifdef CONFIG_XPS 3031 netif_reset_xps_queues(dev, offset, count); 3032 #endif 3033 dev->tc_to_txq[tc].count = count; 3034 dev->tc_to_txq[tc].offset = offset; 3035 return 0; 3036 } 3037 EXPORT_SYMBOL(netdev_set_tc_queue); 3038 3039 int netdev_set_num_tc(struct net_device *dev, u8 num_tc) 3040 { 3041 if (num_tc > TC_MAX_QUEUE) 3042 return -EINVAL; 3043 3044 #ifdef CONFIG_XPS 3045 netif_reset_xps_queues_gt(dev, 0); 3046 #endif 3047 netdev_unbind_all_sb_channels(dev); 3048 3049 dev->num_tc = num_tc; 3050 return 0; 3051 } 3052 EXPORT_SYMBOL(netdev_set_num_tc); 3053 3054 void netdev_unbind_sb_channel(struct net_device *dev, 3055 struct net_device *sb_dev) 3056 { 3057 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 3058 3059 #ifdef CONFIG_XPS 3060 netif_reset_xps_queues_gt(sb_dev, 0); 3061 #endif 3062 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq)); 3063 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map)); 3064 3065 while (txq-- != &dev->_tx[0]) { 3066 if (txq->sb_dev == sb_dev) 3067 txq->sb_dev = NULL; 3068 } 3069 } 3070 EXPORT_SYMBOL(netdev_unbind_sb_channel); 3071 3072 int netdev_bind_sb_channel_queue(struct net_device *dev, 3073 struct net_device *sb_dev, 3074 u8 tc, u16 count, u16 offset) 3075 { 3076 /* Make certain the sb_dev and dev are already configured */ 3077 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc) 3078 return -EINVAL; 3079 3080 /* We cannot hand out queues we don't have */ 3081 if ((offset + count) > dev->real_num_tx_queues) 3082 return -EINVAL; 3083 3084 /* Record the mapping */ 3085 sb_dev->tc_to_txq[tc].count = count; 3086 sb_dev->tc_to_txq[tc].offset = offset; 3087 3088 /* Provide a way for Tx queue to find the tc_to_txq map or 3089 * XPS map for itself. 3090 */ 3091 while (count--) 3092 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev; 3093 3094 return 0; 3095 } 3096 EXPORT_SYMBOL(netdev_bind_sb_channel_queue); 3097 3098 int netdev_set_sb_channel(struct net_device *dev, u16 channel) 3099 { 3100 /* Do not use a multiqueue device to represent a subordinate channel */ 3101 if (netif_is_multiqueue(dev)) 3102 return -ENODEV; 3103 3104 /* We allow channels 1 - 32767 to be used for subordinate channels. 3105 * Channel 0 is meant to be "native" mode and used only to represent 3106 * the main root device. We allow writing 0 to reset the device back 3107 * to normal mode after being used as a subordinate channel. 3108 */ 3109 if (channel > S16_MAX) 3110 return -EINVAL; 3111 3112 dev->num_tc = -channel; 3113 3114 return 0; 3115 } 3116 EXPORT_SYMBOL(netdev_set_sb_channel); 3117 3118 /* 3119 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues 3120 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed. 3121 */ 3122 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq) 3123 { 3124 bool disabling; 3125 int rc; 3126 3127 disabling = txq < dev->real_num_tx_queues; 3128 3129 if (txq < 1 || txq > dev->num_tx_queues) 3130 return -EINVAL; 3131 3132 if (dev->reg_state == NETREG_REGISTERED || 3133 dev->reg_state == NETREG_UNREGISTERING) { 3134 ASSERT_RTNL(); 3135 3136 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues, 3137 txq); 3138 if (rc) 3139 return rc; 3140 3141 if (dev->num_tc) 3142 netif_setup_tc(dev, txq); 3143 3144 net_shaper_set_real_num_tx_queues(dev, txq); 3145 3146 dev_qdisc_change_real_num_tx(dev, txq); 3147 3148 dev->real_num_tx_queues = txq; 3149 3150 if (disabling) { 3151 synchronize_net(); 3152 qdisc_reset_all_tx_gt(dev, txq); 3153 #ifdef CONFIG_XPS 3154 netif_reset_xps_queues_gt(dev, txq); 3155 #endif 3156 } 3157 } else { 3158 dev->real_num_tx_queues = txq; 3159 } 3160 3161 return 0; 3162 } 3163 EXPORT_SYMBOL(netif_set_real_num_tx_queues); 3164 3165 #ifdef CONFIG_SYSFS 3166 /** 3167 * netif_set_real_num_rx_queues - set actual number of RX queues used 3168 * @dev: Network device 3169 * @rxq: Actual number of RX queues 3170 * 3171 * This must be called either with the rtnl_lock held or before 3172 * registration of the net device. Returns 0 on success, or a 3173 * negative error code. If called before registration, it always 3174 * succeeds. 3175 */ 3176 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq) 3177 { 3178 int rc; 3179 3180 if (rxq < 1 || rxq > dev->num_rx_queues) 3181 return -EINVAL; 3182 3183 if (dev->reg_state == NETREG_REGISTERED) { 3184 ASSERT_RTNL(); 3185 3186 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues, 3187 rxq); 3188 if (rc) 3189 return rc; 3190 } 3191 3192 dev->real_num_rx_queues = rxq; 3193 return 0; 3194 } 3195 EXPORT_SYMBOL(netif_set_real_num_rx_queues); 3196 #endif 3197 3198 /** 3199 * netif_set_real_num_queues - set actual number of RX and TX queues used 3200 * @dev: Network device 3201 * @txq: Actual number of TX queues 3202 * @rxq: Actual number of RX queues 3203 * 3204 * Set the real number of both TX and RX queues. 3205 * Does nothing if the number of queues is already correct. 3206 */ 3207 int netif_set_real_num_queues(struct net_device *dev, 3208 unsigned int txq, unsigned int rxq) 3209 { 3210 unsigned int old_rxq = dev->real_num_rx_queues; 3211 int err; 3212 3213 if (txq < 1 || txq > dev->num_tx_queues || 3214 rxq < 1 || rxq > dev->num_rx_queues) 3215 return -EINVAL; 3216 3217 /* Start from increases, so the error path only does decreases - 3218 * decreases can't fail. 3219 */ 3220 if (rxq > dev->real_num_rx_queues) { 3221 err = netif_set_real_num_rx_queues(dev, rxq); 3222 if (err) 3223 return err; 3224 } 3225 if (txq > dev->real_num_tx_queues) { 3226 err = netif_set_real_num_tx_queues(dev, txq); 3227 if (err) 3228 goto undo_rx; 3229 } 3230 if (rxq < dev->real_num_rx_queues) 3231 WARN_ON(netif_set_real_num_rx_queues(dev, rxq)); 3232 if (txq < dev->real_num_tx_queues) 3233 WARN_ON(netif_set_real_num_tx_queues(dev, txq)); 3234 3235 return 0; 3236 undo_rx: 3237 WARN_ON(netif_set_real_num_rx_queues(dev, old_rxq)); 3238 return err; 3239 } 3240 EXPORT_SYMBOL(netif_set_real_num_queues); 3241 3242 /** 3243 * netif_set_tso_max_size() - set the max size of TSO frames supported 3244 * @dev: netdev to update 3245 * @size: max skb->len of a TSO frame 3246 * 3247 * Set the limit on the size of TSO super-frames the device can handle. 3248 * Unless explicitly set the stack will assume the value of 3249 * %GSO_LEGACY_MAX_SIZE. 3250 */ 3251 void netif_set_tso_max_size(struct net_device *dev, unsigned int size) 3252 { 3253 dev->tso_max_size = min(GSO_MAX_SIZE, size); 3254 if (size < READ_ONCE(dev->gso_max_size)) 3255 netif_set_gso_max_size(dev, size); 3256 if (size < READ_ONCE(dev->gso_ipv4_max_size)) 3257 netif_set_gso_ipv4_max_size(dev, size); 3258 } 3259 EXPORT_SYMBOL(netif_set_tso_max_size); 3260 3261 /** 3262 * netif_set_tso_max_segs() - set the max number of segs supported for TSO 3263 * @dev: netdev to update 3264 * @segs: max number of TCP segments 3265 * 3266 * Set the limit on the number of TCP segments the device can generate from 3267 * a single TSO super-frame. 3268 * Unless explicitly set the stack will assume the value of %GSO_MAX_SEGS. 3269 */ 3270 void netif_set_tso_max_segs(struct net_device *dev, unsigned int segs) 3271 { 3272 dev->tso_max_segs = segs; 3273 if (segs < READ_ONCE(dev->gso_max_segs)) 3274 netif_set_gso_max_segs(dev, segs); 3275 } 3276 EXPORT_SYMBOL(netif_set_tso_max_segs); 3277 3278 /** 3279 * netif_inherit_tso_max() - copy all TSO limits from a lower device to an upper 3280 * @to: netdev to update 3281 * @from: netdev from which to copy the limits 3282 */ 3283 void netif_inherit_tso_max(struct net_device *to, const struct net_device *from) 3284 { 3285 netif_set_tso_max_size(to, from->tso_max_size); 3286 netif_set_tso_max_segs(to, from->tso_max_segs); 3287 } 3288 EXPORT_SYMBOL(netif_inherit_tso_max); 3289 3290 /** 3291 * netif_get_num_default_rss_queues - default number of RSS queues 3292 * 3293 * Default value is the number of physical cores if there are only 1 or 2, or 3294 * divided by 2 if there are more. 3295 */ 3296 int netif_get_num_default_rss_queues(void) 3297 { 3298 cpumask_var_t cpus; 3299 int cpu, count = 0; 3300 3301 if (unlikely(is_kdump_kernel() || !zalloc_cpumask_var(&cpus, GFP_KERNEL))) 3302 return 1; 3303 3304 cpumask_copy(cpus, cpu_online_mask); 3305 for_each_cpu(cpu, cpus) { 3306 ++count; 3307 cpumask_andnot(cpus, cpus, topology_sibling_cpumask(cpu)); 3308 } 3309 free_cpumask_var(cpus); 3310 3311 return count > 2 ? DIV_ROUND_UP(count, 2) : count; 3312 } 3313 EXPORT_SYMBOL(netif_get_num_default_rss_queues); 3314 3315 static void __netif_reschedule(struct Qdisc *q) 3316 { 3317 struct softnet_data *sd; 3318 unsigned long flags; 3319 3320 local_irq_save(flags); 3321 sd = this_cpu_ptr(&softnet_data); 3322 q->next_sched = NULL; 3323 *sd->output_queue_tailp = q; 3324 sd->output_queue_tailp = &q->next_sched; 3325 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3326 local_irq_restore(flags); 3327 } 3328 3329 void __netif_schedule(struct Qdisc *q) 3330 { 3331 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state)) 3332 __netif_reschedule(q); 3333 } 3334 EXPORT_SYMBOL(__netif_schedule); 3335 3336 struct dev_kfree_skb_cb { 3337 enum skb_drop_reason reason; 3338 }; 3339 3340 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb) 3341 { 3342 return (struct dev_kfree_skb_cb *)skb->cb; 3343 } 3344 3345 void netif_schedule_queue(struct netdev_queue *txq) 3346 { 3347 rcu_read_lock(); 3348 if (!netif_xmit_stopped(txq)) { 3349 struct Qdisc *q = rcu_dereference(txq->qdisc); 3350 3351 __netif_schedule(q); 3352 } 3353 rcu_read_unlock(); 3354 } 3355 EXPORT_SYMBOL(netif_schedule_queue); 3356 3357 void netif_tx_wake_queue(struct netdev_queue *dev_queue) 3358 { 3359 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) { 3360 struct Qdisc *q; 3361 3362 rcu_read_lock(); 3363 q = rcu_dereference(dev_queue->qdisc); 3364 __netif_schedule(q); 3365 rcu_read_unlock(); 3366 } 3367 } 3368 EXPORT_SYMBOL(netif_tx_wake_queue); 3369 3370 void dev_kfree_skb_irq_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3371 { 3372 unsigned long flags; 3373 3374 if (unlikely(!skb)) 3375 return; 3376 3377 if (likely(refcount_read(&skb->users) == 1)) { 3378 smp_rmb(); 3379 refcount_set(&skb->users, 0); 3380 } else if (likely(!refcount_dec_and_test(&skb->users))) { 3381 return; 3382 } 3383 get_kfree_skb_cb(skb)->reason = reason; 3384 local_irq_save(flags); 3385 skb->next = __this_cpu_read(softnet_data.completion_queue); 3386 __this_cpu_write(softnet_data.completion_queue, skb); 3387 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3388 local_irq_restore(flags); 3389 } 3390 EXPORT_SYMBOL(dev_kfree_skb_irq_reason); 3391 3392 void dev_kfree_skb_any_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3393 { 3394 if (in_hardirq() || irqs_disabled()) 3395 dev_kfree_skb_irq_reason(skb, reason); 3396 else 3397 kfree_skb_reason(skb, reason); 3398 } 3399 EXPORT_SYMBOL(dev_kfree_skb_any_reason); 3400 3401 3402 /** 3403 * netif_device_detach - mark device as removed 3404 * @dev: network device 3405 * 3406 * Mark device as removed from system and therefore no longer available. 3407 */ 3408 void netif_device_detach(struct net_device *dev) 3409 { 3410 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) && 3411 netif_running(dev)) { 3412 netif_tx_stop_all_queues(dev); 3413 } 3414 } 3415 EXPORT_SYMBOL(netif_device_detach); 3416 3417 /** 3418 * netif_device_attach - mark device as attached 3419 * @dev: network device 3420 * 3421 * Mark device as attached from system and restart if needed. 3422 */ 3423 void netif_device_attach(struct net_device *dev) 3424 { 3425 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) && 3426 netif_running(dev)) { 3427 netif_tx_wake_all_queues(dev); 3428 netdev_watchdog_up(dev); 3429 } 3430 } 3431 EXPORT_SYMBOL(netif_device_attach); 3432 3433 /* 3434 * Returns a Tx hash based on the given packet descriptor a Tx queues' number 3435 * to be used as a distribution range. 3436 */ 3437 static u16 skb_tx_hash(const struct net_device *dev, 3438 const struct net_device *sb_dev, 3439 struct sk_buff *skb) 3440 { 3441 u32 hash; 3442 u16 qoffset = 0; 3443 u16 qcount = dev->real_num_tx_queues; 3444 3445 if (dev->num_tc) { 3446 u8 tc = netdev_get_prio_tc_map(dev, skb->priority); 3447 3448 qoffset = sb_dev->tc_to_txq[tc].offset; 3449 qcount = sb_dev->tc_to_txq[tc].count; 3450 if (unlikely(!qcount)) { 3451 net_warn_ratelimited("%s: invalid qcount, qoffset %u for tc %u\n", 3452 sb_dev->name, qoffset, tc); 3453 qoffset = 0; 3454 qcount = dev->real_num_tx_queues; 3455 } 3456 } 3457 3458 if (skb_rx_queue_recorded(skb)) { 3459 DEBUG_NET_WARN_ON_ONCE(qcount == 0); 3460 hash = skb_get_rx_queue(skb); 3461 if (hash >= qoffset) 3462 hash -= qoffset; 3463 while (unlikely(hash >= qcount)) 3464 hash -= qcount; 3465 return hash + qoffset; 3466 } 3467 3468 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset; 3469 } 3470 3471 void skb_warn_bad_offload(const struct sk_buff *skb) 3472 { 3473 static const netdev_features_t null_features; 3474 struct net_device *dev = skb->dev; 3475 const char *name = ""; 3476 3477 if (!net_ratelimit()) 3478 return; 3479 3480 if (dev) { 3481 if (dev->dev.parent) 3482 name = dev_driver_string(dev->dev.parent); 3483 else 3484 name = netdev_name(dev); 3485 } 3486 skb_dump(KERN_WARNING, skb, false); 3487 WARN(1, "%s: caps=(%pNF, %pNF)\n", 3488 name, dev ? &dev->features : &null_features, 3489 skb->sk ? &skb->sk->sk_route_caps : &null_features); 3490 } 3491 3492 /* 3493 * Invalidate hardware checksum when packet is to be mangled, and 3494 * complete checksum manually on outgoing path. 3495 */ 3496 int skb_checksum_help(struct sk_buff *skb) 3497 { 3498 __wsum csum; 3499 int ret = 0, offset; 3500 3501 if (skb->ip_summed == CHECKSUM_COMPLETE) 3502 goto out_set_summed; 3503 3504 if (unlikely(skb_is_gso(skb))) { 3505 skb_warn_bad_offload(skb); 3506 return -EINVAL; 3507 } 3508 3509 if (!skb_frags_readable(skb)) { 3510 return -EFAULT; 3511 } 3512 3513 /* Before computing a checksum, we should make sure no frag could 3514 * be modified by an external entity : checksum could be wrong. 3515 */ 3516 if (skb_has_shared_frag(skb)) { 3517 ret = __skb_linearize(skb); 3518 if (ret) 3519 goto out; 3520 } 3521 3522 offset = skb_checksum_start_offset(skb); 3523 ret = -EINVAL; 3524 if (unlikely(offset >= skb_headlen(skb))) { 3525 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3526 WARN_ONCE(true, "offset (%d) >= skb_headlen() (%u)\n", 3527 offset, skb_headlen(skb)); 3528 goto out; 3529 } 3530 csum = skb_checksum(skb, offset, skb->len - offset, 0); 3531 3532 offset += skb->csum_offset; 3533 if (unlikely(offset + sizeof(__sum16) > skb_headlen(skb))) { 3534 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3535 WARN_ONCE(true, "offset+2 (%zu) > skb_headlen() (%u)\n", 3536 offset + sizeof(__sum16), skb_headlen(skb)); 3537 goto out; 3538 } 3539 ret = skb_ensure_writable(skb, offset + sizeof(__sum16)); 3540 if (ret) 3541 goto out; 3542 3543 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0; 3544 out_set_summed: 3545 skb->ip_summed = CHECKSUM_NONE; 3546 out: 3547 return ret; 3548 } 3549 EXPORT_SYMBOL(skb_checksum_help); 3550 3551 int skb_crc32c_csum_help(struct sk_buff *skb) 3552 { 3553 __le32 crc32c_csum; 3554 int ret = 0, offset, start; 3555 3556 if (skb->ip_summed != CHECKSUM_PARTIAL) 3557 goto out; 3558 3559 if (unlikely(skb_is_gso(skb))) 3560 goto out; 3561 3562 /* Before computing a checksum, we should make sure no frag could 3563 * be modified by an external entity : checksum could be wrong. 3564 */ 3565 if (unlikely(skb_has_shared_frag(skb))) { 3566 ret = __skb_linearize(skb); 3567 if (ret) 3568 goto out; 3569 } 3570 start = skb_checksum_start_offset(skb); 3571 offset = start + offsetof(struct sctphdr, checksum); 3572 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) { 3573 ret = -EINVAL; 3574 goto out; 3575 } 3576 3577 ret = skb_ensure_writable(skb, offset + sizeof(__le32)); 3578 if (ret) 3579 goto out; 3580 3581 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start, 3582 skb->len - start, ~(__u32)0, 3583 crc32c_csum_stub)); 3584 *(__le32 *)(skb->data + offset) = crc32c_csum; 3585 skb_reset_csum_not_inet(skb); 3586 out: 3587 return ret; 3588 } 3589 EXPORT_SYMBOL(skb_crc32c_csum_help); 3590 3591 __be16 skb_network_protocol(struct sk_buff *skb, int *depth) 3592 { 3593 __be16 type = skb->protocol; 3594 3595 /* Tunnel gso handlers can set protocol to ethernet. */ 3596 if (type == htons(ETH_P_TEB)) { 3597 struct ethhdr *eth; 3598 3599 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr)))) 3600 return 0; 3601 3602 eth = (struct ethhdr *)skb->data; 3603 type = eth->h_proto; 3604 } 3605 3606 return vlan_get_protocol_and_depth(skb, type, depth); 3607 } 3608 3609 3610 /* Take action when hardware reception checksum errors are detected. */ 3611 #ifdef CONFIG_BUG 3612 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3613 { 3614 netdev_err(dev, "hw csum failure\n"); 3615 skb_dump(KERN_ERR, skb, true); 3616 dump_stack(); 3617 } 3618 3619 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3620 { 3621 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb); 3622 } 3623 EXPORT_SYMBOL(netdev_rx_csum_fault); 3624 #endif 3625 3626 /* XXX: check that highmem exists at all on the given machine. */ 3627 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb) 3628 { 3629 #ifdef CONFIG_HIGHMEM 3630 int i; 3631 3632 if (!(dev->features & NETIF_F_HIGHDMA)) { 3633 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 3634 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 3635 struct page *page = skb_frag_page(frag); 3636 3637 if (page && PageHighMem(page)) 3638 return 1; 3639 } 3640 } 3641 #endif 3642 return 0; 3643 } 3644 3645 /* If MPLS offload request, verify we are testing hardware MPLS features 3646 * instead of standard features for the netdev. 3647 */ 3648 #if IS_ENABLED(CONFIG_NET_MPLS_GSO) 3649 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3650 netdev_features_t features, 3651 __be16 type) 3652 { 3653 if (eth_p_mpls(type)) 3654 features &= skb->dev->mpls_features; 3655 3656 return features; 3657 } 3658 #else 3659 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3660 netdev_features_t features, 3661 __be16 type) 3662 { 3663 return features; 3664 } 3665 #endif 3666 3667 static netdev_features_t harmonize_features(struct sk_buff *skb, 3668 netdev_features_t features) 3669 { 3670 __be16 type; 3671 3672 type = skb_network_protocol(skb, NULL); 3673 features = net_mpls_features(skb, features, type); 3674 3675 if (skb->ip_summed != CHECKSUM_NONE && 3676 !can_checksum_protocol(features, type)) { 3677 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3678 } 3679 if (illegal_highdma(skb->dev, skb)) 3680 features &= ~NETIF_F_SG; 3681 3682 return features; 3683 } 3684 3685 netdev_features_t passthru_features_check(struct sk_buff *skb, 3686 struct net_device *dev, 3687 netdev_features_t features) 3688 { 3689 return features; 3690 } 3691 EXPORT_SYMBOL(passthru_features_check); 3692 3693 static netdev_features_t dflt_features_check(struct sk_buff *skb, 3694 struct net_device *dev, 3695 netdev_features_t features) 3696 { 3697 return vlan_features_check(skb, features); 3698 } 3699 3700 static netdev_features_t gso_features_check(const struct sk_buff *skb, 3701 struct net_device *dev, 3702 netdev_features_t features) 3703 { 3704 u16 gso_segs = skb_shinfo(skb)->gso_segs; 3705 3706 if (gso_segs > READ_ONCE(dev->gso_max_segs)) 3707 return features & ~NETIF_F_GSO_MASK; 3708 3709 if (unlikely(skb->len >= netif_get_gso_max_size(dev, skb))) 3710 return features & ~NETIF_F_GSO_MASK; 3711 3712 if (!skb_shinfo(skb)->gso_type) { 3713 skb_warn_bad_offload(skb); 3714 return features & ~NETIF_F_GSO_MASK; 3715 } 3716 3717 /* Support for GSO partial features requires software 3718 * intervention before we can actually process the packets 3719 * so we need to strip support for any partial features now 3720 * and we can pull them back in after we have partially 3721 * segmented the frame. 3722 */ 3723 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)) 3724 features &= ~dev->gso_partial_features; 3725 3726 /* Make sure to clear the IPv4 ID mangling feature if the 3727 * IPv4 header has the potential to be fragmented. 3728 */ 3729 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) { 3730 struct iphdr *iph = skb->encapsulation ? 3731 inner_ip_hdr(skb) : ip_hdr(skb); 3732 3733 if (!(iph->frag_off & htons(IP_DF))) 3734 features &= ~NETIF_F_TSO_MANGLEID; 3735 } 3736 3737 return features; 3738 } 3739 3740 netdev_features_t netif_skb_features(struct sk_buff *skb) 3741 { 3742 struct net_device *dev = skb->dev; 3743 netdev_features_t features = dev->features; 3744 3745 if (skb_is_gso(skb)) 3746 features = gso_features_check(skb, dev, features); 3747 3748 /* If encapsulation offload request, verify we are testing 3749 * hardware encapsulation features instead of standard 3750 * features for the netdev 3751 */ 3752 if (skb->encapsulation) 3753 features &= dev->hw_enc_features; 3754 3755 if (skb_vlan_tagged(skb)) 3756 features = netdev_intersect_features(features, 3757 dev->vlan_features | 3758 NETIF_F_HW_VLAN_CTAG_TX | 3759 NETIF_F_HW_VLAN_STAG_TX); 3760 3761 if (dev->netdev_ops->ndo_features_check) 3762 features &= dev->netdev_ops->ndo_features_check(skb, dev, 3763 features); 3764 else 3765 features &= dflt_features_check(skb, dev, features); 3766 3767 return harmonize_features(skb, features); 3768 } 3769 EXPORT_SYMBOL(netif_skb_features); 3770 3771 static int xmit_one(struct sk_buff *skb, struct net_device *dev, 3772 struct netdev_queue *txq, bool more) 3773 { 3774 unsigned int len; 3775 int rc; 3776 3777 if (dev_nit_active(dev)) 3778 dev_queue_xmit_nit(skb, dev); 3779 3780 len = skb->len; 3781 trace_net_dev_start_xmit(skb, dev); 3782 rc = netdev_start_xmit(skb, dev, txq, more); 3783 trace_net_dev_xmit(skb, rc, dev, len); 3784 3785 return rc; 3786 } 3787 3788 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev, 3789 struct netdev_queue *txq, int *ret) 3790 { 3791 struct sk_buff *skb = first; 3792 int rc = NETDEV_TX_OK; 3793 3794 while (skb) { 3795 struct sk_buff *next = skb->next; 3796 3797 skb_mark_not_on_list(skb); 3798 rc = xmit_one(skb, dev, txq, next != NULL); 3799 if (unlikely(!dev_xmit_complete(rc))) { 3800 skb->next = next; 3801 goto out; 3802 } 3803 3804 skb = next; 3805 if (netif_tx_queue_stopped(txq) && skb) { 3806 rc = NETDEV_TX_BUSY; 3807 break; 3808 } 3809 } 3810 3811 out: 3812 *ret = rc; 3813 return skb; 3814 } 3815 3816 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb, 3817 netdev_features_t features) 3818 { 3819 if (skb_vlan_tag_present(skb) && 3820 !vlan_hw_offload_capable(features, skb->vlan_proto)) 3821 skb = __vlan_hwaccel_push_inside(skb); 3822 return skb; 3823 } 3824 3825 int skb_csum_hwoffload_help(struct sk_buff *skb, 3826 const netdev_features_t features) 3827 { 3828 if (unlikely(skb_csum_is_sctp(skb))) 3829 return !!(features & NETIF_F_SCTP_CRC) ? 0 : 3830 skb_crc32c_csum_help(skb); 3831 3832 if (features & NETIF_F_HW_CSUM) 3833 return 0; 3834 3835 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) { 3836 if (vlan_get_protocol(skb) == htons(ETH_P_IPV6) && 3837 skb_network_header_len(skb) != sizeof(struct ipv6hdr) && 3838 !ipv6_has_hopopt_jumbo(skb)) 3839 goto sw_checksum; 3840 3841 switch (skb->csum_offset) { 3842 case offsetof(struct tcphdr, check): 3843 case offsetof(struct udphdr, check): 3844 return 0; 3845 } 3846 } 3847 3848 sw_checksum: 3849 return skb_checksum_help(skb); 3850 } 3851 EXPORT_SYMBOL(skb_csum_hwoffload_help); 3852 3853 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again) 3854 { 3855 netdev_features_t features; 3856 3857 features = netif_skb_features(skb); 3858 skb = validate_xmit_vlan(skb, features); 3859 if (unlikely(!skb)) 3860 goto out_null; 3861 3862 skb = sk_validate_xmit_skb(skb, dev); 3863 if (unlikely(!skb)) 3864 goto out_null; 3865 3866 if (netif_needs_gso(skb, features)) { 3867 struct sk_buff *segs; 3868 3869 segs = skb_gso_segment(skb, features); 3870 if (IS_ERR(segs)) { 3871 goto out_kfree_skb; 3872 } else if (segs) { 3873 consume_skb(skb); 3874 skb = segs; 3875 } 3876 } else { 3877 if (skb_needs_linearize(skb, features) && 3878 __skb_linearize(skb)) 3879 goto out_kfree_skb; 3880 3881 /* If packet is not checksummed and device does not 3882 * support checksumming for this protocol, complete 3883 * checksumming here. 3884 */ 3885 if (skb->ip_summed == CHECKSUM_PARTIAL) { 3886 if (skb->encapsulation) 3887 skb_set_inner_transport_header(skb, 3888 skb_checksum_start_offset(skb)); 3889 else 3890 skb_set_transport_header(skb, 3891 skb_checksum_start_offset(skb)); 3892 if (skb_csum_hwoffload_help(skb, features)) 3893 goto out_kfree_skb; 3894 } 3895 } 3896 3897 skb = validate_xmit_xfrm(skb, features, again); 3898 3899 return skb; 3900 3901 out_kfree_skb: 3902 kfree_skb(skb); 3903 out_null: 3904 dev_core_stats_tx_dropped_inc(dev); 3905 return NULL; 3906 } 3907 3908 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again) 3909 { 3910 struct sk_buff *next, *head = NULL, *tail; 3911 3912 for (; skb != NULL; skb = next) { 3913 next = skb->next; 3914 skb_mark_not_on_list(skb); 3915 3916 /* in case skb won't be segmented, point to itself */ 3917 skb->prev = skb; 3918 3919 skb = validate_xmit_skb(skb, dev, again); 3920 if (!skb) 3921 continue; 3922 3923 if (!head) 3924 head = skb; 3925 else 3926 tail->next = skb; 3927 /* If skb was segmented, skb->prev points to 3928 * the last segment. If not, it still contains skb. 3929 */ 3930 tail = skb->prev; 3931 } 3932 return head; 3933 } 3934 EXPORT_SYMBOL_GPL(validate_xmit_skb_list); 3935 3936 static void qdisc_pkt_len_init(struct sk_buff *skb) 3937 { 3938 const struct skb_shared_info *shinfo = skb_shinfo(skb); 3939 3940 qdisc_skb_cb(skb)->pkt_len = skb->len; 3941 3942 /* To get more precise estimation of bytes sent on wire, 3943 * we add to pkt_len the headers size of all segments 3944 */ 3945 if (shinfo->gso_size && skb_transport_header_was_set(skb)) { 3946 u16 gso_segs = shinfo->gso_segs; 3947 unsigned int hdr_len; 3948 3949 /* mac layer + network layer */ 3950 hdr_len = skb_transport_offset(skb); 3951 3952 /* + transport layer */ 3953 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) { 3954 const struct tcphdr *th; 3955 struct tcphdr _tcphdr; 3956 3957 th = skb_header_pointer(skb, hdr_len, 3958 sizeof(_tcphdr), &_tcphdr); 3959 if (likely(th)) 3960 hdr_len += __tcp_hdrlen(th); 3961 } else if (shinfo->gso_type & SKB_GSO_UDP_L4) { 3962 struct udphdr _udphdr; 3963 3964 if (skb_header_pointer(skb, hdr_len, 3965 sizeof(_udphdr), &_udphdr)) 3966 hdr_len += sizeof(struct udphdr); 3967 } 3968 3969 if (unlikely(shinfo->gso_type & SKB_GSO_DODGY)) { 3970 int payload = skb->len - hdr_len; 3971 3972 /* Malicious packet. */ 3973 if (payload <= 0) 3974 return; 3975 gso_segs = DIV_ROUND_UP(payload, shinfo->gso_size); 3976 } 3977 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len; 3978 } 3979 } 3980 3981 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q, 3982 struct sk_buff **to_free, 3983 struct netdev_queue *txq) 3984 { 3985 int rc; 3986 3987 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK; 3988 if (rc == NET_XMIT_SUCCESS) 3989 trace_qdisc_enqueue(q, txq, skb); 3990 return rc; 3991 } 3992 3993 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q, 3994 struct net_device *dev, 3995 struct netdev_queue *txq) 3996 { 3997 spinlock_t *root_lock = qdisc_lock(q); 3998 struct sk_buff *to_free = NULL; 3999 bool contended; 4000 int rc; 4001 4002 qdisc_calculate_pkt_len(skb, q); 4003 4004 tcf_set_drop_reason(skb, SKB_DROP_REASON_QDISC_DROP); 4005 4006 if (q->flags & TCQ_F_NOLOCK) { 4007 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) && 4008 qdisc_run_begin(q)) { 4009 /* Retest nolock_qdisc_is_empty() within the protection 4010 * of q->seqlock to protect from racing with requeuing. 4011 */ 4012 if (unlikely(!nolock_qdisc_is_empty(q))) { 4013 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4014 __qdisc_run(q); 4015 qdisc_run_end(q); 4016 4017 goto no_lock_out; 4018 } 4019 4020 qdisc_bstats_cpu_update(q, skb); 4021 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) && 4022 !nolock_qdisc_is_empty(q)) 4023 __qdisc_run(q); 4024 4025 qdisc_run_end(q); 4026 return NET_XMIT_SUCCESS; 4027 } 4028 4029 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4030 qdisc_run(q); 4031 4032 no_lock_out: 4033 if (unlikely(to_free)) 4034 kfree_skb_list_reason(to_free, 4035 tcf_get_drop_reason(to_free)); 4036 return rc; 4037 } 4038 4039 if (unlikely(READ_ONCE(q->owner) == smp_processor_id())) { 4040 kfree_skb_reason(skb, SKB_DROP_REASON_TC_RECLASSIFY_LOOP); 4041 return NET_XMIT_DROP; 4042 } 4043 /* 4044 * Heuristic to force contended enqueues to serialize on a 4045 * separate lock before trying to get qdisc main lock. 4046 * This permits qdisc->running owner to get the lock more 4047 * often and dequeue packets faster. 4048 * On PREEMPT_RT it is possible to preempt the qdisc owner during xmit 4049 * and then other tasks will only enqueue packets. The packets will be 4050 * sent after the qdisc owner is scheduled again. To prevent this 4051 * scenario the task always serialize on the lock. 4052 */ 4053 contended = qdisc_is_running(q) || IS_ENABLED(CONFIG_PREEMPT_RT); 4054 if (unlikely(contended)) 4055 spin_lock(&q->busylock); 4056 4057 spin_lock(root_lock); 4058 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) { 4059 __qdisc_drop(skb, &to_free); 4060 rc = NET_XMIT_DROP; 4061 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && 4062 qdisc_run_begin(q)) { 4063 /* 4064 * This is a work-conserving queue; there are no old skbs 4065 * waiting to be sent out; and the qdisc is not running - 4066 * xmit the skb directly. 4067 */ 4068 4069 qdisc_bstats_update(q, skb); 4070 4071 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) { 4072 if (unlikely(contended)) { 4073 spin_unlock(&q->busylock); 4074 contended = false; 4075 } 4076 __qdisc_run(q); 4077 } 4078 4079 qdisc_run_end(q); 4080 rc = NET_XMIT_SUCCESS; 4081 } else { 4082 WRITE_ONCE(q->owner, smp_processor_id()); 4083 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4084 WRITE_ONCE(q->owner, -1); 4085 if (qdisc_run_begin(q)) { 4086 if (unlikely(contended)) { 4087 spin_unlock(&q->busylock); 4088 contended = false; 4089 } 4090 __qdisc_run(q); 4091 qdisc_run_end(q); 4092 } 4093 } 4094 spin_unlock(root_lock); 4095 if (unlikely(to_free)) 4096 kfree_skb_list_reason(to_free, 4097 tcf_get_drop_reason(to_free)); 4098 if (unlikely(contended)) 4099 spin_unlock(&q->busylock); 4100 return rc; 4101 } 4102 4103 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO) 4104 static void skb_update_prio(struct sk_buff *skb) 4105 { 4106 const struct netprio_map *map; 4107 const struct sock *sk; 4108 unsigned int prioidx; 4109 4110 if (skb->priority) 4111 return; 4112 map = rcu_dereference_bh(skb->dev->priomap); 4113 if (!map) 4114 return; 4115 sk = skb_to_full_sk(skb); 4116 if (!sk) 4117 return; 4118 4119 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data); 4120 4121 if (prioidx < map->priomap_len) 4122 skb->priority = map->priomap[prioidx]; 4123 } 4124 #else 4125 #define skb_update_prio(skb) 4126 #endif 4127 4128 /** 4129 * dev_loopback_xmit - loop back @skb 4130 * @net: network namespace this loopback is happening in 4131 * @sk: sk needed to be a netfilter okfn 4132 * @skb: buffer to transmit 4133 */ 4134 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb) 4135 { 4136 skb_reset_mac_header(skb); 4137 __skb_pull(skb, skb_network_offset(skb)); 4138 skb->pkt_type = PACKET_LOOPBACK; 4139 if (skb->ip_summed == CHECKSUM_NONE) 4140 skb->ip_summed = CHECKSUM_UNNECESSARY; 4141 DEBUG_NET_WARN_ON_ONCE(!skb_dst(skb)); 4142 skb_dst_force(skb); 4143 netif_rx(skb); 4144 return 0; 4145 } 4146 EXPORT_SYMBOL(dev_loopback_xmit); 4147 4148 #ifdef CONFIG_NET_EGRESS 4149 static struct netdev_queue * 4150 netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb) 4151 { 4152 int qm = skb_get_queue_mapping(skb); 4153 4154 return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm)); 4155 } 4156 4157 #ifndef CONFIG_PREEMPT_RT 4158 static bool netdev_xmit_txqueue_skipped(void) 4159 { 4160 return __this_cpu_read(softnet_data.xmit.skip_txqueue); 4161 } 4162 4163 void netdev_xmit_skip_txqueue(bool skip) 4164 { 4165 __this_cpu_write(softnet_data.xmit.skip_txqueue, skip); 4166 } 4167 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4168 4169 #else 4170 static bool netdev_xmit_txqueue_skipped(void) 4171 { 4172 return current->net_xmit.skip_txqueue; 4173 } 4174 4175 void netdev_xmit_skip_txqueue(bool skip) 4176 { 4177 current->net_xmit.skip_txqueue = skip; 4178 } 4179 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4180 #endif 4181 #endif /* CONFIG_NET_EGRESS */ 4182 4183 #ifdef CONFIG_NET_XGRESS 4184 static int tc_run(struct tcx_entry *entry, struct sk_buff *skb, 4185 enum skb_drop_reason *drop_reason) 4186 { 4187 int ret = TC_ACT_UNSPEC; 4188 #ifdef CONFIG_NET_CLS_ACT 4189 struct mini_Qdisc *miniq = rcu_dereference_bh(entry->miniq); 4190 struct tcf_result res; 4191 4192 if (!miniq) 4193 return ret; 4194 4195 /* Global bypass */ 4196 if (!static_branch_likely(&tcf_sw_enabled_key)) 4197 return ret; 4198 4199 /* Block-wise bypass */ 4200 if (tcf_block_bypass_sw(miniq->block)) 4201 return ret; 4202 4203 tc_skb_cb(skb)->mru = 0; 4204 tc_skb_cb(skb)->post_ct = false; 4205 tcf_set_drop_reason(skb, *drop_reason); 4206 4207 mini_qdisc_bstats_cpu_update(miniq, skb); 4208 ret = tcf_classify(skb, miniq->block, miniq->filter_list, &res, false); 4209 /* Only tcf related quirks below. */ 4210 switch (ret) { 4211 case TC_ACT_SHOT: 4212 *drop_reason = tcf_get_drop_reason(skb); 4213 mini_qdisc_qstats_cpu_drop(miniq); 4214 break; 4215 case TC_ACT_OK: 4216 case TC_ACT_RECLASSIFY: 4217 skb->tc_index = TC_H_MIN(res.classid); 4218 break; 4219 } 4220 #endif /* CONFIG_NET_CLS_ACT */ 4221 return ret; 4222 } 4223 4224 static DEFINE_STATIC_KEY_FALSE(tcx_needed_key); 4225 4226 void tcx_inc(void) 4227 { 4228 static_branch_inc(&tcx_needed_key); 4229 } 4230 4231 void tcx_dec(void) 4232 { 4233 static_branch_dec(&tcx_needed_key); 4234 } 4235 4236 static __always_inline enum tcx_action_base 4237 tcx_run(const struct bpf_mprog_entry *entry, struct sk_buff *skb, 4238 const bool needs_mac) 4239 { 4240 const struct bpf_mprog_fp *fp; 4241 const struct bpf_prog *prog; 4242 int ret = TCX_NEXT; 4243 4244 if (needs_mac) 4245 __skb_push(skb, skb->mac_len); 4246 bpf_mprog_foreach_prog(entry, fp, prog) { 4247 bpf_compute_data_pointers(skb); 4248 ret = bpf_prog_run(prog, skb); 4249 if (ret != TCX_NEXT) 4250 break; 4251 } 4252 if (needs_mac) 4253 __skb_pull(skb, skb->mac_len); 4254 return tcx_action_code(skb, ret); 4255 } 4256 4257 static __always_inline struct sk_buff * 4258 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4259 struct net_device *orig_dev, bool *another) 4260 { 4261 struct bpf_mprog_entry *entry = rcu_dereference_bh(skb->dev->tcx_ingress); 4262 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_INGRESS; 4263 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4264 int sch_ret; 4265 4266 if (!entry) 4267 return skb; 4268 4269 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4270 if (*pt_prev) { 4271 *ret = deliver_skb(skb, *pt_prev, orig_dev); 4272 *pt_prev = NULL; 4273 } 4274 4275 qdisc_skb_cb(skb)->pkt_len = skb->len; 4276 tcx_set_ingress(skb, true); 4277 4278 if (static_branch_unlikely(&tcx_needed_key)) { 4279 sch_ret = tcx_run(entry, skb, true); 4280 if (sch_ret != TC_ACT_UNSPEC) 4281 goto ingress_verdict; 4282 } 4283 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4284 ingress_verdict: 4285 switch (sch_ret) { 4286 case TC_ACT_REDIRECT: 4287 /* skb_mac_header check was done by BPF, so we can safely 4288 * push the L2 header back before redirecting to another 4289 * netdev. 4290 */ 4291 __skb_push(skb, skb->mac_len); 4292 if (skb_do_redirect(skb) == -EAGAIN) { 4293 __skb_pull(skb, skb->mac_len); 4294 *another = true; 4295 break; 4296 } 4297 *ret = NET_RX_SUCCESS; 4298 bpf_net_ctx_clear(bpf_net_ctx); 4299 return NULL; 4300 case TC_ACT_SHOT: 4301 kfree_skb_reason(skb, drop_reason); 4302 *ret = NET_RX_DROP; 4303 bpf_net_ctx_clear(bpf_net_ctx); 4304 return NULL; 4305 /* used by tc_run */ 4306 case TC_ACT_STOLEN: 4307 case TC_ACT_QUEUED: 4308 case TC_ACT_TRAP: 4309 consume_skb(skb); 4310 fallthrough; 4311 case TC_ACT_CONSUMED: 4312 *ret = NET_RX_SUCCESS; 4313 bpf_net_ctx_clear(bpf_net_ctx); 4314 return NULL; 4315 } 4316 bpf_net_ctx_clear(bpf_net_ctx); 4317 4318 return skb; 4319 } 4320 4321 static __always_inline struct sk_buff * 4322 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4323 { 4324 struct bpf_mprog_entry *entry = rcu_dereference_bh(dev->tcx_egress); 4325 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_EGRESS; 4326 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4327 int sch_ret; 4328 4329 if (!entry) 4330 return skb; 4331 4332 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4333 4334 /* qdisc_skb_cb(skb)->pkt_len & tcx_set_ingress() was 4335 * already set by the caller. 4336 */ 4337 if (static_branch_unlikely(&tcx_needed_key)) { 4338 sch_ret = tcx_run(entry, skb, false); 4339 if (sch_ret != TC_ACT_UNSPEC) 4340 goto egress_verdict; 4341 } 4342 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4343 egress_verdict: 4344 switch (sch_ret) { 4345 case TC_ACT_REDIRECT: 4346 /* No need to push/pop skb's mac_header here on egress! */ 4347 skb_do_redirect(skb); 4348 *ret = NET_XMIT_SUCCESS; 4349 bpf_net_ctx_clear(bpf_net_ctx); 4350 return NULL; 4351 case TC_ACT_SHOT: 4352 kfree_skb_reason(skb, drop_reason); 4353 *ret = NET_XMIT_DROP; 4354 bpf_net_ctx_clear(bpf_net_ctx); 4355 return NULL; 4356 /* used by tc_run */ 4357 case TC_ACT_STOLEN: 4358 case TC_ACT_QUEUED: 4359 case TC_ACT_TRAP: 4360 consume_skb(skb); 4361 fallthrough; 4362 case TC_ACT_CONSUMED: 4363 *ret = NET_XMIT_SUCCESS; 4364 bpf_net_ctx_clear(bpf_net_ctx); 4365 return NULL; 4366 } 4367 bpf_net_ctx_clear(bpf_net_ctx); 4368 4369 return skb; 4370 } 4371 #else 4372 static __always_inline struct sk_buff * 4373 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4374 struct net_device *orig_dev, bool *another) 4375 { 4376 return skb; 4377 } 4378 4379 static __always_inline struct sk_buff * 4380 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4381 { 4382 return skb; 4383 } 4384 #endif /* CONFIG_NET_XGRESS */ 4385 4386 #ifdef CONFIG_XPS 4387 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb, 4388 struct xps_dev_maps *dev_maps, unsigned int tci) 4389 { 4390 int tc = netdev_get_prio_tc_map(dev, skb->priority); 4391 struct xps_map *map; 4392 int queue_index = -1; 4393 4394 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids) 4395 return queue_index; 4396 4397 tci *= dev_maps->num_tc; 4398 tci += tc; 4399 4400 map = rcu_dereference(dev_maps->attr_map[tci]); 4401 if (map) { 4402 if (map->len == 1) 4403 queue_index = map->queues[0]; 4404 else 4405 queue_index = map->queues[reciprocal_scale( 4406 skb_get_hash(skb), map->len)]; 4407 if (unlikely(queue_index >= dev->real_num_tx_queues)) 4408 queue_index = -1; 4409 } 4410 return queue_index; 4411 } 4412 #endif 4413 4414 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev, 4415 struct sk_buff *skb) 4416 { 4417 #ifdef CONFIG_XPS 4418 struct xps_dev_maps *dev_maps; 4419 struct sock *sk = skb->sk; 4420 int queue_index = -1; 4421 4422 if (!static_key_false(&xps_needed)) 4423 return -1; 4424 4425 rcu_read_lock(); 4426 if (!static_key_false(&xps_rxqs_needed)) 4427 goto get_cpus_map; 4428 4429 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]); 4430 if (dev_maps) { 4431 int tci = sk_rx_queue_get(sk); 4432 4433 if (tci >= 0) 4434 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4435 tci); 4436 } 4437 4438 get_cpus_map: 4439 if (queue_index < 0) { 4440 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]); 4441 if (dev_maps) { 4442 unsigned int tci = skb->sender_cpu - 1; 4443 4444 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4445 tci); 4446 } 4447 } 4448 rcu_read_unlock(); 4449 4450 return queue_index; 4451 #else 4452 return -1; 4453 #endif 4454 } 4455 4456 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb, 4457 struct net_device *sb_dev) 4458 { 4459 return 0; 4460 } 4461 EXPORT_SYMBOL(dev_pick_tx_zero); 4462 4463 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb, 4464 struct net_device *sb_dev) 4465 { 4466 struct sock *sk = skb->sk; 4467 int queue_index = sk_tx_queue_get(sk); 4468 4469 sb_dev = sb_dev ? : dev; 4470 4471 if (queue_index < 0 || skb->ooo_okay || 4472 queue_index >= dev->real_num_tx_queues) { 4473 int new_index = get_xps_queue(dev, sb_dev, skb); 4474 4475 if (new_index < 0) 4476 new_index = skb_tx_hash(dev, sb_dev, skb); 4477 4478 if (queue_index != new_index && sk && 4479 sk_fullsock(sk) && 4480 rcu_access_pointer(sk->sk_dst_cache)) 4481 sk_tx_queue_set(sk, new_index); 4482 4483 queue_index = new_index; 4484 } 4485 4486 return queue_index; 4487 } 4488 EXPORT_SYMBOL(netdev_pick_tx); 4489 4490 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev, 4491 struct sk_buff *skb, 4492 struct net_device *sb_dev) 4493 { 4494 int queue_index = 0; 4495 4496 #ifdef CONFIG_XPS 4497 u32 sender_cpu = skb->sender_cpu - 1; 4498 4499 if (sender_cpu >= (u32)NR_CPUS) 4500 skb->sender_cpu = raw_smp_processor_id() + 1; 4501 #endif 4502 4503 if (dev->real_num_tx_queues != 1) { 4504 const struct net_device_ops *ops = dev->netdev_ops; 4505 4506 if (ops->ndo_select_queue) 4507 queue_index = ops->ndo_select_queue(dev, skb, sb_dev); 4508 else 4509 queue_index = netdev_pick_tx(dev, skb, sb_dev); 4510 4511 queue_index = netdev_cap_txqueue(dev, queue_index); 4512 } 4513 4514 skb_set_queue_mapping(skb, queue_index); 4515 return netdev_get_tx_queue(dev, queue_index); 4516 } 4517 4518 /** 4519 * __dev_queue_xmit() - transmit a buffer 4520 * @skb: buffer to transmit 4521 * @sb_dev: suboordinate device used for L2 forwarding offload 4522 * 4523 * Queue a buffer for transmission to a network device. The caller must 4524 * have set the device and priority and built the buffer before calling 4525 * this function. The function can be called from an interrupt. 4526 * 4527 * When calling this method, interrupts MUST be enabled. This is because 4528 * the BH enable code must have IRQs enabled so that it will not deadlock. 4529 * 4530 * Regardless of the return value, the skb is consumed, so it is currently 4531 * difficult to retry a send to this method. (You can bump the ref count 4532 * before sending to hold a reference for retry if you are careful.) 4533 * 4534 * Return: 4535 * * 0 - buffer successfully transmitted 4536 * * positive qdisc return code - NET_XMIT_DROP etc. 4537 * * negative errno - other errors 4538 */ 4539 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev) 4540 { 4541 struct net_device *dev = skb->dev; 4542 struct netdev_queue *txq = NULL; 4543 struct Qdisc *q; 4544 int rc = -ENOMEM; 4545 bool again = false; 4546 4547 skb_reset_mac_header(skb); 4548 skb_assert_len(skb); 4549 4550 if (unlikely(skb_shinfo(skb)->tx_flags & 4551 (SKBTX_SCHED_TSTAMP | SKBTX_BPF))) 4552 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED); 4553 4554 /* Disable soft irqs for various locks below. Also 4555 * stops preemption for RCU. 4556 */ 4557 rcu_read_lock_bh(); 4558 4559 skb_update_prio(skb); 4560 4561 qdisc_pkt_len_init(skb); 4562 tcx_set_ingress(skb, false); 4563 #ifdef CONFIG_NET_EGRESS 4564 if (static_branch_unlikely(&egress_needed_key)) { 4565 if (nf_hook_egress_active()) { 4566 skb = nf_hook_egress(skb, &rc, dev); 4567 if (!skb) 4568 goto out; 4569 } 4570 4571 netdev_xmit_skip_txqueue(false); 4572 4573 nf_skip_egress(skb, true); 4574 skb = sch_handle_egress(skb, &rc, dev); 4575 if (!skb) 4576 goto out; 4577 nf_skip_egress(skb, false); 4578 4579 if (netdev_xmit_txqueue_skipped()) 4580 txq = netdev_tx_queue_mapping(dev, skb); 4581 } 4582 #endif 4583 /* If device/qdisc don't need skb->dst, release it right now while 4584 * its hot in this cpu cache. 4585 */ 4586 if (dev->priv_flags & IFF_XMIT_DST_RELEASE) 4587 skb_dst_drop(skb); 4588 else 4589 skb_dst_force(skb); 4590 4591 if (!txq) 4592 txq = netdev_core_pick_tx(dev, skb, sb_dev); 4593 4594 q = rcu_dereference_bh(txq->qdisc); 4595 4596 trace_net_dev_queue(skb); 4597 if (q->enqueue) { 4598 rc = __dev_xmit_skb(skb, q, dev, txq); 4599 goto out; 4600 } 4601 4602 /* The device has no queue. Common case for software devices: 4603 * loopback, all the sorts of tunnels... 4604 4605 * Really, it is unlikely that netif_tx_lock protection is necessary 4606 * here. (f.e. loopback and IP tunnels are clean ignoring statistics 4607 * counters.) 4608 * However, it is possible, that they rely on protection 4609 * made by us here. 4610 4611 * Check this and shot the lock. It is not prone from deadlocks. 4612 *Either shot noqueue qdisc, it is even simpler 8) 4613 */ 4614 if (dev->flags & IFF_UP) { 4615 int cpu = smp_processor_id(); /* ok because BHs are off */ 4616 4617 /* Other cpus might concurrently change txq->xmit_lock_owner 4618 * to -1 or to their cpu id, but not to our id. 4619 */ 4620 if (READ_ONCE(txq->xmit_lock_owner) != cpu) { 4621 if (dev_xmit_recursion()) 4622 goto recursion_alert; 4623 4624 skb = validate_xmit_skb(skb, dev, &again); 4625 if (!skb) 4626 goto out; 4627 4628 HARD_TX_LOCK(dev, txq, cpu); 4629 4630 if (!netif_xmit_stopped(txq)) { 4631 dev_xmit_recursion_inc(); 4632 skb = dev_hard_start_xmit(skb, dev, txq, &rc); 4633 dev_xmit_recursion_dec(); 4634 if (dev_xmit_complete(rc)) { 4635 HARD_TX_UNLOCK(dev, txq); 4636 goto out; 4637 } 4638 } 4639 HARD_TX_UNLOCK(dev, txq); 4640 net_crit_ratelimited("Virtual device %s asks to queue packet!\n", 4641 dev->name); 4642 } else { 4643 /* Recursion is detected! It is possible, 4644 * unfortunately 4645 */ 4646 recursion_alert: 4647 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n", 4648 dev->name); 4649 } 4650 } 4651 4652 rc = -ENETDOWN; 4653 rcu_read_unlock_bh(); 4654 4655 dev_core_stats_tx_dropped_inc(dev); 4656 kfree_skb_list(skb); 4657 return rc; 4658 out: 4659 rcu_read_unlock_bh(); 4660 return rc; 4661 } 4662 EXPORT_SYMBOL(__dev_queue_xmit); 4663 4664 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id) 4665 { 4666 struct net_device *dev = skb->dev; 4667 struct sk_buff *orig_skb = skb; 4668 struct netdev_queue *txq; 4669 int ret = NETDEV_TX_BUSY; 4670 bool again = false; 4671 4672 if (unlikely(!netif_running(dev) || 4673 !netif_carrier_ok(dev))) 4674 goto drop; 4675 4676 skb = validate_xmit_skb_list(skb, dev, &again); 4677 if (skb != orig_skb) 4678 goto drop; 4679 4680 skb_set_queue_mapping(skb, queue_id); 4681 txq = skb_get_tx_queue(dev, skb); 4682 4683 local_bh_disable(); 4684 4685 dev_xmit_recursion_inc(); 4686 HARD_TX_LOCK(dev, txq, smp_processor_id()); 4687 if (!netif_xmit_frozen_or_drv_stopped(txq)) 4688 ret = netdev_start_xmit(skb, dev, txq, false); 4689 HARD_TX_UNLOCK(dev, txq); 4690 dev_xmit_recursion_dec(); 4691 4692 local_bh_enable(); 4693 return ret; 4694 drop: 4695 dev_core_stats_tx_dropped_inc(dev); 4696 kfree_skb_list(skb); 4697 return NET_XMIT_DROP; 4698 } 4699 EXPORT_SYMBOL(__dev_direct_xmit); 4700 4701 /************************************************************************* 4702 * Receiver routines 4703 *************************************************************************/ 4704 static DEFINE_PER_CPU(struct task_struct *, backlog_napi); 4705 4706 int weight_p __read_mostly = 64; /* old backlog weight */ 4707 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */ 4708 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */ 4709 4710 /* Called with irq disabled */ 4711 static inline void ____napi_schedule(struct softnet_data *sd, 4712 struct napi_struct *napi) 4713 { 4714 struct task_struct *thread; 4715 4716 lockdep_assert_irqs_disabled(); 4717 4718 if (test_bit(NAPI_STATE_THREADED, &napi->state)) { 4719 /* Paired with smp_mb__before_atomic() in 4720 * napi_enable()/dev_set_threaded(). 4721 * Use READ_ONCE() to guarantee a complete 4722 * read on napi->thread. Only call 4723 * wake_up_process() when it's not NULL. 4724 */ 4725 thread = READ_ONCE(napi->thread); 4726 if (thread) { 4727 if (use_backlog_threads() && thread == raw_cpu_read(backlog_napi)) 4728 goto use_local_napi; 4729 4730 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 4731 wake_up_process(thread); 4732 return; 4733 } 4734 } 4735 4736 use_local_napi: 4737 list_add_tail(&napi->poll_list, &sd->poll_list); 4738 WRITE_ONCE(napi->list_owner, smp_processor_id()); 4739 /* If not called from net_rx_action() 4740 * we have to raise NET_RX_SOFTIRQ. 4741 */ 4742 if (!sd->in_net_rx_action) 4743 raise_softirq_irqoff(NET_RX_SOFTIRQ); 4744 } 4745 4746 #ifdef CONFIG_RPS 4747 4748 struct static_key_false rps_needed __read_mostly; 4749 EXPORT_SYMBOL(rps_needed); 4750 struct static_key_false rfs_needed __read_mostly; 4751 EXPORT_SYMBOL(rfs_needed); 4752 4753 static struct rps_dev_flow * 4754 set_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4755 struct rps_dev_flow *rflow, u16 next_cpu) 4756 { 4757 if (next_cpu < nr_cpu_ids) { 4758 u32 head; 4759 #ifdef CONFIG_RFS_ACCEL 4760 struct netdev_rx_queue *rxqueue; 4761 struct rps_dev_flow_table *flow_table; 4762 struct rps_dev_flow *old_rflow; 4763 u16 rxq_index; 4764 u32 flow_id; 4765 int rc; 4766 4767 /* Should we steer this flow to a different hardware queue? */ 4768 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap || 4769 !(dev->features & NETIF_F_NTUPLE)) 4770 goto out; 4771 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu); 4772 if (rxq_index == skb_get_rx_queue(skb)) 4773 goto out; 4774 4775 rxqueue = dev->_rx + rxq_index; 4776 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4777 if (!flow_table) 4778 goto out; 4779 flow_id = skb_get_hash(skb) & flow_table->mask; 4780 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb, 4781 rxq_index, flow_id); 4782 if (rc < 0) 4783 goto out; 4784 old_rflow = rflow; 4785 rflow = &flow_table->flows[flow_id]; 4786 WRITE_ONCE(rflow->filter, rc); 4787 if (old_rflow->filter == rc) 4788 WRITE_ONCE(old_rflow->filter, RPS_NO_FILTER); 4789 out: 4790 #endif 4791 head = READ_ONCE(per_cpu(softnet_data, next_cpu).input_queue_head); 4792 rps_input_queue_tail_save(&rflow->last_qtail, head); 4793 } 4794 4795 WRITE_ONCE(rflow->cpu, next_cpu); 4796 return rflow; 4797 } 4798 4799 /* 4800 * get_rps_cpu is called from netif_receive_skb and returns the target 4801 * CPU from the RPS map of the receiving queue for a given skb. 4802 * rcu_read_lock must be held on entry. 4803 */ 4804 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4805 struct rps_dev_flow **rflowp) 4806 { 4807 const struct rps_sock_flow_table *sock_flow_table; 4808 struct netdev_rx_queue *rxqueue = dev->_rx; 4809 struct rps_dev_flow_table *flow_table; 4810 struct rps_map *map; 4811 int cpu = -1; 4812 u32 tcpu; 4813 u32 hash; 4814 4815 if (skb_rx_queue_recorded(skb)) { 4816 u16 index = skb_get_rx_queue(skb); 4817 4818 if (unlikely(index >= dev->real_num_rx_queues)) { 4819 WARN_ONCE(dev->real_num_rx_queues > 1, 4820 "%s received packet on queue %u, but number " 4821 "of RX queues is %u\n", 4822 dev->name, index, dev->real_num_rx_queues); 4823 goto done; 4824 } 4825 rxqueue += index; 4826 } 4827 4828 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */ 4829 4830 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4831 map = rcu_dereference(rxqueue->rps_map); 4832 if (!flow_table && !map) 4833 goto done; 4834 4835 skb_reset_network_header(skb); 4836 hash = skb_get_hash(skb); 4837 if (!hash) 4838 goto done; 4839 4840 sock_flow_table = rcu_dereference(net_hotdata.rps_sock_flow_table); 4841 if (flow_table && sock_flow_table) { 4842 struct rps_dev_flow *rflow; 4843 u32 next_cpu; 4844 u32 ident; 4845 4846 /* First check into global flow table if there is a match. 4847 * This READ_ONCE() pairs with WRITE_ONCE() from rps_record_sock_flow(). 4848 */ 4849 ident = READ_ONCE(sock_flow_table->ents[hash & sock_flow_table->mask]); 4850 if ((ident ^ hash) & ~net_hotdata.rps_cpu_mask) 4851 goto try_rps; 4852 4853 next_cpu = ident & net_hotdata.rps_cpu_mask; 4854 4855 /* OK, now we know there is a match, 4856 * we can look at the local (per receive queue) flow table 4857 */ 4858 rflow = &flow_table->flows[hash & flow_table->mask]; 4859 tcpu = rflow->cpu; 4860 4861 /* 4862 * If the desired CPU (where last recvmsg was done) is 4863 * different from current CPU (one in the rx-queue flow 4864 * table entry), switch if one of the following holds: 4865 * - Current CPU is unset (>= nr_cpu_ids). 4866 * - Current CPU is offline. 4867 * - The current CPU's queue tail has advanced beyond the 4868 * last packet that was enqueued using this table entry. 4869 * This guarantees that all previous packets for the flow 4870 * have been dequeued, thus preserving in order delivery. 4871 */ 4872 if (unlikely(tcpu != next_cpu) && 4873 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) || 4874 ((int)(READ_ONCE(per_cpu(softnet_data, tcpu).input_queue_head) - 4875 rflow->last_qtail)) >= 0)) { 4876 tcpu = next_cpu; 4877 rflow = set_rps_cpu(dev, skb, rflow, next_cpu); 4878 } 4879 4880 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) { 4881 *rflowp = rflow; 4882 cpu = tcpu; 4883 goto done; 4884 } 4885 } 4886 4887 try_rps: 4888 4889 if (map) { 4890 tcpu = map->cpus[reciprocal_scale(hash, map->len)]; 4891 if (cpu_online(tcpu)) { 4892 cpu = tcpu; 4893 goto done; 4894 } 4895 } 4896 4897 done: 4898 return cpu; 4899 } 4900 4901 #ifdef CONFIG_RFS_ACCEL 4902 4903 /** 4904 * rps_may_expire_flow - check whether an RFS hardware filter may be removed 4905 * @dev: Device on which the filter was set 4906 * @rxq_index: RX queue index 4907 * @flow_id: Flow ID passed to ndo_rx_flow_steer() 4908 * @filter_id: Filter ID returned by ndo_rx_flow_steer() 4909 * 4910 * Drivers that implement ndo_rx_flow_steer() should periodically call 4911 * this function for each installed filter and remove the filters for 4912 * which it returns %true. 4913 */ 4914 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index, 4915 u32 flow_id, u16 filter_id) 4916 { 4917 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index; 4918 struct rps_dev_flow_table *flow_table; 4919 struct rps_dev_flow *rflow; 4920 bool expire = true; 4921 unsigned int cpu; 4922 4923 rcu_read_lock(); 4924 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4925 if (flow_table && flow_id <= flow_table->mask) { 4926 rflow = &flow_table->flows[flow_id]; 4927 cpu = READ_ONCE(rflow->cpu); 4928 if (READ_ONCE(rflow->filter) == filter_id && cpu < nr_cpu_ids && 4929 ((int)(READ_ONCE(per_cpu(softnet_data, cpu).input_queue_head) - 4930 READ_ONCE(rflow->last_qtail)) < 4931 (int)(10 * flow_table->mask))) 4932 expire = false; 4933 } 4934 rcu_read_unlock(); 4935 return expire; 4936 } 4937 EXPORT_SYMBOL(rps_may_expire_flow); 4938 4939 #endif /* CONFIG_RFS_ACCEL */ 4940 4941 /* Called from hardirq (IPI) context */ 4942 static void rps_trigger_softirq(void *data) 4943 { 4944 struct softnet_data *sd = data; 4945 4946 ____napi_schedule(sd, &sd->backlog); 4947 sd->received_rps++; 4948 } 4949 4950 #endif /* CONFIG_RPS */ 4951 4952 /* Called from hardirq (IPI) context */ 4953 static void trigger_rx_softirq(void *data) 4954 { 4955 struct softnet_data *sd = data; 4956 4957 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4958 smp_store_release(&sd->defer_ipi_scheduled, 0); 4959 } 4960 4961 /* 4962 * After we queued a packet into sd->input_pkt_queue, 4963 * we need to make sure this queue is serviced soon. 4964 * 4965 * - If this is another cpu queue, link it to our rps_ipi_list, 4966 * and make sure we will process rps_ipi_list from net_rx_action(). 4967 * 4968 * - If this is our own queue, NAPI schedule our backlog. 4969 * Note that this also raises NET_RX_SOFTIRQ. 4970 */ 4971 static void napi_schedule_rps(struct softnet_data *sd) 4972 { 4973 struct softnet_data *mysd = this_cpu_ptr(&softnet_data); 4974 4975 #ifdef CONFIG_RPS 4976 if (sd != mysd) { 4977 if (use_backlog_threads()) { 4978 __napi_schedule_irqoff(&sd->backlog); 4979 return; 4980 } 4981 4982 sd->rps_ipi_next = mysd->rps_ipi_list; 4983 mysd->rps_ipi_list = sd; 4984 4985 /* If not called from net_rx_action() or napi_threaded_poll() 4986 * we have to raise NET_RX_SOFTIRQ. 4987 */ 4988 if (!mysd->in_net_rx_action && !mysd->in_napi_threaded_poll) 4989 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4990 return; 4991 } 4992 #endif /* CONFIG_RPS */ 4993 __napi_schedule_irqoff(&mysd->backlog); 4994 } 4995 4996 void kick_defer_list_purge(struct softnet_data *sd, unsigned int cpu) 4997 { 4998 unsigned long flags; 4999 5000 if (use_backlog_threads()) { 5001 backlog_lock_irq_save(sd, &flags); 5002 5003 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) 5004 __napi_schedule_irqoff(&sd->backlog); 5005 5006 backlog_unlock_irq_restore(sd, &flags); 5007 5008 } else if (!cmpxchg(&sd->defer_ipi_scheduled, 0, 1)) { 5009 smp_call_function_single_async(cpu, &sd->defer_csd); 5010 } 5011 } 5012 5013 #ifdef CONFIG_NET_FLOW_LIMIT 5014 int netdev_flow_limit_table_len __read_mostly = (1 << 12); 5015 #endif 5016 5017 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen) 5018 { 5019 #ifdef CONFIG_NET_FLOW_LIMIT 5020 struct sd_flow_limit *fl; 5021 struct softnet_data *sd; 5022 unsigned int old_flow, new_flow; 5023 5024 if (qlen < (READ_ONCE(net_hotdata.max_backlog) >> 1)) 5025 return false; 5026 5027 sd = this_cpu_ptr(&softnet_data); 5028 5029 rcu_read_lock(); 5030 fl = rcu_dereference(sd->flow_limit); 5031 if (fl) { 5032 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1); 5033 old_flow = fl->history[fl->history_head]; 5034 fl->history[fl->history_head] = new_flow; 5035 5036 fl->history_head++; 5037 fl->history_head &= FLOW_LIMIT_HISTORY - 1; 5038 5039 if (likely(fl->buckets[old_flow])) 5040 fl->buckets[old_flow]--; 5041 5042 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) { 5043 fl->count++; 5044 rcu_read_unlock(); 5045 return true; 5046 } 5047 } 5048 rcu_read_unlock(); 5049 #endif 5050 return false; 5051 } 5052 5053 /* 5054 * enqueue_to_backlog is called to queue an skb to a per CPU backlog 5055 * queue (may be a remote CPU queue). 5056 */ 5057 static int enqueue_to_backlog(struct sk_buff *skb, int cpu, 5058 unsigned int *qtail) 5059 { 5060 enum skb_drop_reason reason; 5061 struct softnet_data *sd; 5062 unsigned long flags; 5063 unsigned int qlen; 5064 int max_backlog; 5065 u32 tail; 5066 5067 reason = SKB_DROP_REASON_DEV_READY; 5068 if (!netif_running(skb->dev)) 5069 goto bad_dev; 5070 5071 reason = SKB_DROP_REASON_CPU_BACKLOG; 5072 sd = &per_cpu(softnet_data, cpu); 5073 5074 qlen = skb_queue_len_lockless(&sd->input_pkt_queue); 5075 max_backlog = READ_ONCE(net_hotdata.max_backlog); 5076 if (unlikely(qlen > max_backlog)) 5077 goto cpu_backlog_drop; 5078 backlog_lock_irq_save(sd, &flags); 5079 qlen = skb_queue_len(&sd->input_pkt_queue); 5080 if (qlen <= max_backlog && !skb_flow_limit(skb, qlen)) { 5081 if (!qlen) { 5082 /* Schedule NAPI for backlog device. We can use 5083 * non atomic operation as we own the queue lock. 5084 */ 5085 if (!__test_and_set_bit(NAPI_STATE_SCHED, 5086 &sd->backlog.state)) 5087 napi_schedule_rps(sd); 5088 } 5089 __skb_queue_tail(&sd->input_pkt_queue, skb); 5090 tail = rps_input_queue_tail_incr(sd); 5091 backlog_unlock_irq_restore(sd, &flags); 5092 5093 /* save the tail outside of the critical section */ 5094 rps_input_queue_tail_save(qtail, tail); 5095 return NET_RX_SUCCESS; 5096 } 5097 5098 backlog_unlock_irq_restore(sd, &flags); 5099 5100 cpu_backlog_drop: 5101 atomic_inc(&sd->dropped); 5102 bad_dev: 5103 dev_core_stats_rx_dropped_inc(skb->dev); 5104 kfree_skb_reason(skb, reason); 5105 return NET_RX_DROP; 5106 } 5107 5108 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb) 5109 { 5110 struct net_device *dev = skb->dev; 5111 struct netdev_rx_queue *rxqueue; 5112 5113 rxqueue = dev->_rx; 5114 5115 if (skb_rx_queue_recorded(skb)) { 5116 u16 index = skb_get_rx_queue(skb); 5117 5118 if (unlikely(index >= dev->real_num_rx_queues)) { 5119 WARN_ONCE(dev->real_num_rx_queues > 1, 5120 "%s received packet on queue %u, but number " 5121 "of RX queues is %u\n", 5122 dev->name, index, dev->real_num_rx_queues); 5123 5124 return rxqueue; /* Return first rxqueue */ 5125 } 5126 rxqueue += index; 5127 } 5128 return rxqueue; 5129 } 5130 5131 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, 5132 const struct bpf_prog *xdp_prog) 5133 { 5134 void *orig_data, *orig_data_end, *hard_start; 5135 struct netdev_rx_queue *rxqueue; 5136 bool orig_bcast, orig_host; 5137 u32 mac_len, frame_sz; 5138 __be16 orig_eth_type; 5139 struct ethhdr *eth; 5140 u32 metalen, act; 5141 int off; 5142 5143 /* The XDP program wants to see the packet starting at the MAC 5144 * header. 5145 */ 5146 mac_len = skb->data - skb_mac_header(skb); 5147 hard_start = skb->data - skb_headroom(skb); 5148 5149 /* SKB "head" area always have tailroom for skb_shared_info */ 5150 frame_sz = (void *)skb_end_pointer(skb) - hard_start; 5151 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 5152 5153 rxqueue = netif_get_rxqueue(skb); 5154 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq); 5155 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len, 5156 skb_headlen(skb) + mac_len, true); 5157 if (skb_is_nonlinear(skb)) { 5158 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 5159 xdp_buff_set_frags_flag(xdp); 5160 } else { 5161 xdp_buff_clear_frags_flag(xdp); 5162 } 5163 5164 orig_data_end = xdp->data_end; 5165 orig_data = xdp->data; 5166 eth = (struct ethhdr *)xdp->data; 5167 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr); 5168 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest); 5169 orig_eth_type = eth->h_proto; 5170 5171 act = bpf_prog_run_xdp(xdp_prog, xdp); 5172 5173 /* check if bpf_xdp_adjust_head was used */ 5174 off = xdp->data - orig_data; 5175 if (off) { 5176 if (off > 0) 5177 __skb_pull(skb, off); 5178 else if (off < 0) 5179 __skb_push(skb, -off); 5180 5181 skb->mac_header += off; 5182 skb_reset_network_header(skb); 5183 } 5184 5185 /* check if bpf_xdp_adjust_tail was used */ 5186 off = xdp->data_end - orig_data_end; 5187 if (off != 0) { 5188 skb_set_tail_pointer(skb, xdp->data_end - xdp->data); 5189 skb->len += off; /* positive on grow, negative on shrink */ 5190 } 5191 5192 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 5193 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here. 5194 */ 5195 if (xdp_buff_has_frags(xdp)) 5196 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 5197 else 5198 skb->data_len = 0; 5199 5200 /* check if XDP changed eth hdr such SKB needs update */ 5201 eth = (struct ethhdr *)xdp->data; 5202 if ((orig_eth_type != eth->h_proto) || 5203 (orig_host != ether_addr_equal_64bits(eth->h_dest, 5204 skb->dev->dev_addr)) || 5205 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) { 5206 __skb_push(skb, ETH_HLEN); 5207 skb->pkt_type = PACKET_HOST; 5208 skb->protocol = eth_type_trans(skb, skb->dev); 5209 } 5210 5211 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull 5212 * before calling us again on redirect path. We do not call do_redirect 5213 * as we leave that up to the caller. 5214 * 5215 * Caller is responsible for managing lifetime of skb (i.e. calling 5216 * kfree_skb in response to actions it cannot handle/XDP_DROP). 5217 */ 5218 switch (act) { 5219 case XDP_REDIRECT: 5220 case XDP_TX: 5221 __skb_push(skb, mac_len); 5222 break; 5223 case XDP_PASS: 5224 metalen = xdp->data - xdp->data_meta; 5225 if (metalen) 5226 skb_metadata_set(skb, metalen); 5227 break; 5228 } 5229 5230 return act; 5231 } 5232 5233 static int 5234 netif_skb_check_for_xdp(struct sk_buff **pskb, const struct bpf_prog *prog) 5235 { 5236 struct sk_buff *skb = *pskb; 5237 int err, hroom, troom; 5238 5239 if (!skb_cow_data_for_xdp(this_cpu_read(system_page_pool), pskb, prog)) 5240 return 0; 5241 5242 /* In case we have to go down the path and also linearize, 5243 * then lets do the pskb_expand_head() work just once here. 5244 */ 5245 hroom = XDP_PACKET_HEADROOM - skb_headroom(skb); 5246 troom = skb->tail + skb->data_len - skb->end; 5247 err = pskb_expand_head(skb, 5248 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0, 5249 troom > 0 ? troom + 128 : 0, GFP_ATOMIC); 5250 if (err) 5251 return err; 5252 5253 return skb_linearize(skb); 5254 } 5255 5256 static u32 netif_receive_generic_xdp(struct sk_buff **pskb, 5257 struct xdp_buff *xdp, 5258 const struct bpf_prog *xdp_prog) 5259 { 5260 struct sk_buff *skb = *pskb; 5261 u32 mac_len, act = XDP_DROP; 5262 5263 /* Reinjected packets coming from act_mirred or similar should 5264 * not get XDP generic processing. 5265 */ 5266 if (skb_is_redirected(skb)) 5267 return XDP_PASS; 5268 5269 /* XDP packets must have sufficient headroom of XDP_PACKET_HEADROOM 5270 * bytes. This is the guarantee that also native XDP provides, 5271 * thus we need to do it here as well. 5272 */ 5273 mac_len = skb->data - skb_mac_header(skb); 5274 __skb_push(skb, mac_len); 5275 5276 if (skb_cloned(skb) || skb_is_nonlinear(skb) || 5277 skb_headroom(skb) < XDP_PACKET_HEADROOM) { 5278 if (netif_skb_check_for_xdp(pskb, xdp_prog)) 5279 goto do_drop; 5280 } 5281 5282 __skb_pull(*pskb, mac_len); 5283 5284 act = bpf_prog_run_generic_xdp(*pskb, xdp, xdp_prog); 5285 switch (act) { 5286 case XDP_REDIRECT: 5287 case XDP_TX: 5288 case XDP_PASS: 5289 break; 5290 default: 5291 bpf_warn_invalid_xdp_action((*pskb)->dev, xdp_prog, act); 5292 fallthrough; 5293 case XDP_ABORTED: 5294 trace_xdp_exception((*pskb)->dev, xdp_prog, act); 5295 fallthrough; 5296 case XDP_DROP: 5297 do_drop: 5298 kfree_skb(*pskb); 5299 break; 5300 } 5301 5302 return act; 5303 } 5304 5305 /* When doing generic XDP we have to bypass the qdisc layer and the 5306 * network taps in order to match in-driver-XDP behavior. This also means 5307 * that XDP packets are able to starve other packets going through a qdisc, 5308 * and DDOS attacks will be more effective. In-driver-XDP use dedicated TX 5309 * queues, so they do not have this starvation issue. 5310 */ 5311 void generic_xdp_tx(struct sk_buff *skb, const struct bpf_prog *xdp_prog) 5312 { 5313 struct net_device *dev = skb->dev; 5314 struct netdev_queue *txq; 5315 bool free_skb = true; 5316 int cpu, rc; 5317 5318 txq = netdev_core_pick_tx(dev, skb, NULL); 5319 cpu = smp_processor_id(); 5320 HARD_TX_LOCK(dev, txq, cpu); 5321 if (!netif_xmit_frozen_or_drv_stopped(txq)) { 5322 rc = netdev_start_xmit(skb, dev, txq, 0); 5323 if (dev_xmit_complete(rc)) 5324 free_skb = false; 5325 } 5326 HARD_TX_UNLOCK(dev, txq); 5327 if (free_skb) { 5328 trace_xdp_exception(dev, xdp_prog, XDP_TX); 5329 dev_core_stats_tx_dropped_inc(dev); 5330 kfree_skb(skb); 5331 } 5332 } 5333 5334 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key); 5335 5336 int do_xdp_generic(const struct bpf_prog *xdp_prog, struct sk_buff **pskb) 5337 { 5338 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 5339 5340 if (xdp_prog) { 5341 struct xdp_buff xdp; 5342 u32 act; 5343 int err; 5344 5345 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 5346 act = netif_receive_generic_xdp(pskb, &xdp, xdp_prog); 5347 if (act != XDP_PASS) { 5348 switch (act) { 5349 case XDP_REDIRECT: 5350 err = xdp_do_generic_redirect((*pskb)->dev, *pskb, 5351 &xdp, xdp_prog); 5352 if (err) 5353 goto out_redir; 5354 break; 5355 case XDP_TX: 5356 generic_xdp_tx(*pskb, xdp_prog); 5357 break; 5358 } 5359 bpf_net_ctx_clear(bpf_net_ctx); 5360 return XDP_DROP; 5361 } 5362 bpf_net_ctx_clear(bpf_net_ctx); 5363 } 5364 return XDP_PASS; 5365 out_redir: 5366 bpf_net_ctx_clear(bpf_net_ctx); 5367 kfree_skb_reason(*pskb, SKB_DROP_REASON_XDP); 5368 return XDP_DROP; 5369 } 5370 EXPORT_SYMBOL_GPL(do_xdp_generic); 5371 5372 static int netif_rx_internal(struct sk_buff *skb) 5373 { 5374 int ret; 5375 5376 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5377 5378 trace_netif_rx(skb); 5379 5380 #ifdef CONFIG_RPS 5381 if (static_branch_unlikely(&rps_needed)) { 5382 struct rps_dev_flow voidflow, *rflow = &voidflow; 5383 int cpu; 5384 5385 rcu_read_lock(); 5386 5387 cpu = get_rps_cpu(skb->dev, skb, &rflow); 5388 if (cpu < 0) 5389 cpu = smp_processor_id(); 5390 5391 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5392 5393 rcu_read_unlock(); 5394 } else 5395 #endif 5396 { 5397 unsigned int qtail; 5398 5399 ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail); 5400 } 5401 return ret; 5402 } 5403 5404 /** 5405 * __netif_rx - Slightly optimized version of netif_rx 5406 * @skb: buffer to post 5407 * 5408 * This behaves as netif_rx except that it does not disable bottom halves. 5409 * As a result this function may only be invoked from the interrupt context 5410 * (either hard or soft interrupt). 5411 */ 5412 int __netif_rx(struct sk_buff *skb) 5413 { 5414 int ret; 5415 5416 lockdep_assert_once(hardirq_count() | softirq_count()); 5417 5418 trace_netif_rx_entry(skb); 5419 ret = netif_rx_internal(skb); 5420 trace_netif_rx_exit(ret); 5421 return ret; 5422 } 5423 EXPORT_SYMBOL(__netif_rx); 5424 5425 /** 5426 * netif_rx - post buffer to the network code 5427 * @skb: buffer to post 5428 * 5429 * This function receives a packet from a device driver and queues it for 5430 * the upper (protocol) levels to process via the backlog NAPI device. It 5431 * always succeeds. The buffer may be dropped during processing for 5432 * congestion control or by the protocol layers. 5433 * The network buffer is passed via the backlog NAPI device. Modern NIC 5434 * driver should use NAPI and GRO. 5435 * This function can used from interrupt and from process context. The 5436 * caller from process context must not disable interrupts before invoking 5437 * this function. 5438 * 5439 * return values: 5440 * NET_RX_SUCCESS (no congestion) 5441 * NET_RX_DROP (packet was dropped) 5442 * 5443 */ 5444 int netif_rx(struct sk_buff *skb) 5445 { 5446 bool need_bh_off = !(hardirq_count() | softirq_count()); 5447 int ret; 5448 5449 if (need_bh_off) 5450 local_bh_disable(); 5451 trace_netif_rx_entry(skb); 5452 ret = netif_rx_internal(skb); 5453 trace_netif_rx_exit(ret); 5454 if (need_bh_off) 5455 local_bh_enable(); 5456 return ret; 5457 } 5458 EXPORT_SYMBOL(netif_rx); 5459 5460 static __latent_entropy void net_tx_action(void) 5461 { 5462 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 5463 5464 if (sd->completion_queue) { 5465 struct sk_buff *clist; 5466 5467 local_irq_disable(); 5468 clist = sd->completion_queue; 5469 sd->completion_queue = NULL; 5470 local_irq_enable(); 5471 5472 while (clist) { 5473 struct sk_buff *skb = clist; 5474 5475 clist = clist->next; 5476 5477 WARN_ON(refcount_read(&skb->users)); 5478 if (likely(get_kfree_skb_cb(skb)->reason == SKB_CONSUMED)) 5479 trace_consume_skb(skb, net_tx_action); 5480 else 5481 trace_kfree_skb(skb, net_tx_action, 5482 get_kfree_skb_cb(skb)->reason, NULL); 5483 5484 if (skb->fclone != SKB_FCLONE_UNAVAILABLE) 5485 __kfree_skb(skb); 5486 else 5487 __napi_kfree_skb(skb, 5488 get_kfree_skb_cb(skb)->reason); 5489 } 5490 } 5491 5492 if (sd->output_queue) { 5493 struct Qdisc *head; 5494 5495 local_irq_disable(); 5496 head = sd->output_queue; 5497 sd->output_queue = NULL; 5498 sd->output_queue_tailp = &sd->output_queue; 5499 local_irq_enable(); 5500 5501 rcu_read_lock(); 5502 5503 while (head) { 5504 struct Qdisc *q = head; 5505 spinlock_t *root_lock = NULL; 5506 5507 head = head->next_sched; 5508 5509 /* We need to make sure head->next_sched is read 5510 * before clearing __QDISC_STATE_SCHED 5511 */ 5512 smp_mb__before_atomic(); 5513 5514 if (!(q->flags & TCQ_F_NOLOCK)) { 5515 root_lock = qdisc_lock(q); 5516 spin_lock(root_lock); 5517 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, 5518 &q->state))) { 5519 /* There is a synchronize_net() between 5520 * STATE_DEACTIVATED flag being set and 5521 * qdisc_reset()/some_qdisc_is_busy() in 5522 * dev_deactivate(), so we can safely bail out 5523 * early here to avoid data race between 5524 * qdisc_deactivate() and some_qdisc_is_busy() 5525 * for lockless qdisc. 5526 */ 5527 clear_bit(__QDISC_STATE_SCHED, &q->state); 5528 continue; 5529 } 5530 5531 clear_bit(__QDISC_STATE_SCHED, &q->state); 5532 qdisc_run(q); 5533 if (root_lock) 5534 spin_unlock(root_lock); 5535 } 5536 5537 rcu_read_unlock(); 5538 } 5539 5540 xfrm_dev_backlog(sd); 5541 } 5542 5543 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE) 5544 /* This hook is defined here for ATM LANE */ 5545 int (*br_fdb_test_addr_hook)(struct net_device *dev, 5546 unsigned char *addr) __read_mostly; 5547 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook); 5548 #endif 5549 5550 /** 5551 * netdev_is_rx_handler_busy - check if receive handler is registered 5552 * @dev: device to check 5553 * 5554 * Check if a receive handler is already registered for a given device. 5555 * Return true if there one. 5556 * 5557 * The caller must hold the rtnl_mutex. 5558 */ 5559 bool netdev_is_rx_handler_busy(struct net_device *dev) 5560 { 5561 ASSERT_RTNL(); 5562 return dev && rtnl_dereference(dev->rx_handler); 5563 } 5564 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy); 5565 5566 /** 5567 * netdev_rx_handler_register - register receive handler 5568 * @dev: device to register a handler for 5569 * @rx_handler: receive handler to register 5570 * @rx_handler_data: data pointer that is used by rx handler 5571 * 5572 * Register a receive handler for a device. This handler will then be 5573 * called from __netif_receive_skb. A negative errno code is returned 5574 * on a failure. 5575 * 5576 * The caller must hold the rtnl_mutex. 5577 * 5578 * For a general description of rx_handler, see enum rx_handler_result. 5579 */ 5580 int netdev_rx_handler_register(struct net_device *dev, 5581 rx_handler_func_t *rx_handler, 5582 void *rx_handler_data) 5583 { 5584 if (netdev_is_rx_handler_busy(dev)) 5585 return -EBUSY; 5586 5587 if (dev->priv_flags & IFF_NO_RX_HANDLER) 5588 return -EINVAL; 5589 5590 /* Note: rx_handler_data must be set before rx_handler */ 5591 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data); 5592 rcu_assign_pointer(dev->rx_handler, rx_handler); 5593 5594 return 0; 5595 } 5596 EXPORT_SYMBOL_GPL(netdev_rx_handler_register); 5597 5598 /** 5599 * netdev_rx_handler_unregister - unregister receive handler 5600 * @dev: device to unregister a handler from 5601 * 5602 * Unregister a receive handler from a device. 5603 * 5604 * The caller must hold the rtnl_mutex. 5605 */ 5606 void netdev_rx_handler_unregister(struct net_device *dev) 5607 { 5608 5609 ASSERT_RTNL(); 5610 RCU_INIT_POINTER(dev->rx_handler, NULL); 5611 /* a reader seeing a non NULL rx_handler in a rcu_read_lock() 5612 * section has a guarantee to see a non NULL rx_handler_data 5613 * as well. 5614 */ 5615 synchronize_net(); 5616 RCU_INIT_POINTER(dev->rx_handler_data, NULL); 5617 } 5618 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister); 5619 5620 /* 5621 * Limit the use of PFMEMALLOC reserves to those protocols that implement 5622 * the special handling of PFMEMALLOC skbs. 5623 */ 5624 static bool skb_pfmemalloc_protocol(struct sk_buff *skb) 5625 { 5626 switch (skb->protocol) { 5627 case htons(ETH_P_ARP): 5628 case htons(ETH_P_IP): 5629 case htons(ETH_P_IPV6): 5630 case htons(ETH_P_8021Q): 5631 case htons(ETH_P_8021AD): 5632 return true; 5633 default: 5634 return false; 5635 } 5636 } 5637 5638 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev, 5639 int *ret, struct net_device *orig_dev) 5640 { 5641 if (nf_hook_ingress_active(skb)) { 5642 int ingress_retval; 5643 5644 if (*pt_prev) { 5645 *ret = deliver_skb(skb, *pt_prev, orig_dev); 5646 *pt_prev = NULL; 5647 } 5648 5649 rcu_read_lock(); 5650 ingress_retval = nf_hook_ingress(skb); 5651 rcu_read_unlock(); 5652 return ingress_retval; 5653 } 5654 return 0; 5655 } 5656 5657 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc, 5658 struct packet_type **ppt_prev) 5659 { 5660 struct packet_type *ptype, *pt_prev; 5661 rx_handler_func_t *rx_handler; 5662 struct sk_buff *skb = *pskb; 5663 struct net_device *orig_dev; 5664 bool deliver_exact = false; 5665 int ret = NET_RX_DROP; 5666 __be16 type; 5667 5668 net_timestamp_check(!READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5669 5670 trace_netif_receive_skb(skb); 5671 5672 orig_dev = skb->dev; 5673 5674 skb_reset_network_header(skb); 5675 #if !defined(CONFIG_DEBUG_NET) 5676 /* We plan to no longer reset the transport header here. 5677 * Give some time to fuzzers and dev build to catch bugs 5678 * in network stacks. 5679 */ 5680 if (!skb_transport_header_was_set(skb)) 5681 skb_reset_transport_header(skb); 5682 #endif 5683 skb_reset_mac_len(skb); 5684 5685 pt_prev = NULL; 5686 5687 another_round: 5688 skb->skb_iif = skb->dev->ifindex; 5689 5690 __this_cpu_inc(softnet_data.processed); 5691 5692 if (static_branch_unlikely(&generic_xdp_needed_key)) { 5693 int ret2; 5694 5695 migrate_disable(); 5696 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), 5697 &skb); 5698 migrate_enable(); 5699 5700 if (ret2 != XDP_PASS) { 5701 ret = NET_RX_DROP; 5702 goto out; 5703 } 5704 } 5705 5706 if (eth_type_vlan(skb->protocol)) { 5707 skb = skb_vlan_untag(skb); 5708 if (unlikely(!skb)) 5709 goto out; 5710 } 5711 5712 if (skb_skip_tc_classify(skb)) 5713 goto skip_classify; 5714 5715 if (pfmemalloc) 5716 goto skip_taps; 5717 5718 list_for_each_entry_rcu(ptype, &net_hotdata.ptype_all, list) { 5719 if (pt_prev) 5720 ret = deliver_skb(skb, pt_prev, orig_dev); 5721 pt_prev = ptype; 5722 } 5723 5724 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) { 5725 if (pt_prev) 5726 ret = deliver_skb(skb, pt_prev, orig_dev); 5727 pt_prev = ptype; 5728 } 5729 5730 skip_taps: 5731 #ifdef CONFIG_NET_INGRESS 5732 if (static_branch_unlikely(&ingress_needed_key)) { 5733 bool another = false; 5734 5735 nf_skip_egress(skb, true); 5736 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev, 5737 &another); 5738 if (another) 5739 goto another_round; 5740 if (!skb) 5741 goto out; 5742 5743 nf_skip_egress(skb, false); 5744 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0) 5745 goto out; 5746 } 5747 #endif 5748 skb_reset_redirect(skb); 5749 skip_classify: 5750 if (pfmemalloc && !skb_pfmemalloc_protocol(skb)) 5751 goto drop; 5752 5753 if (skb_vlan_tag_present(skb)) { 5754 if (pt_prev) { 5755 ret = deliver_skb(skb, pt_prev, orig_dev); 5756 pt_prev = NULL; 5757 } 5758 if (vlan_do_receive(&skb)) 5759 goto another_round; 5760 else if (unlikely(!skb)) 5761 goto out; 5762 } 5763 5764 rx_handler = rcu_dereference(skb->dev->rx_handler); 5765 if (rx_handler) { 5766 if (pt_prev) { 5767 ret = deliver_skb(skb, pt_prev, orig_dev); 5768 pt_prev = NULL; 5769 } 5770 switch (rx_handler(&skb)) { 5771 case RX_HANDLER_CONSUMED: 5772 ret = NET_RX_SUCCESS; 5773 goto out; 5774 case RX_HANDLER_ANOTHER: 5775 goto another_round; 5776 case RX_HANDLER_EXACT: 5777 deliver_exact = true; 5778 break; 5779 case RX_HANDLER_PASS: 5780 break; 5781 default: 5782 BUG(); 5783 } 5784 } 5785 5786 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) { 5787 check_vlan_id: 5788 if (skb_vlan_tag_get_id(skb)) { 5789 /* Vlan id is non 0 and vlan_do_receive() above couldn't 5790 * find vlan device. 5791 */ 5792 skb->pkt_type = PACKET_OTHERHOST; 5793 } else if (eth_type_vlan(skb->protocol)) { 5794 /* Outer header is 802.1P with vlan 0, inner header is 5795 * 802.1Q or 802.1AD and vlan_do_receive() above could 5796 * not find vlan dev for vlan id 0. 5797 */ 5798 __vlan_hwaccel_clear_tag(skb); 5799 skb = skb_vlan_untag(skb); 5800 if (unlikely(!skb)) 5801 goto out; 5802 if (vlan_do_receive(&skb)) 5803 /* After stripping off 802.1P header with vlan 0 5804 * vlan dev is found for inner header. 5805 */ 5806 goto another_round; 5807 else if (unlikely(!skb)) 5808 goto out; 5809 else 5810 /* We have stripped outer 802.1P vlan 0 header. 5811 * But could not find vlan dev. 5812 * check again for vlan id to set OTHERHOST. 5813 */ 5814 goto check_vlan_id; 5815 } 5816 /* Note: we might in the future use prio bits 5817 * and set skb->priority like in vlan_do_receive() 5818 * For the time being, just ignore Priority Code Point 5819 */ 5820 __vlan_hwaccel_clear_tag(skb); 5821 } 5822 5823 type = skb->protocol; 5824 5825 /* deliver only exact match when indicated */ 5826 if (likely(!deliver_exact)) { 5827 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5828 &ptype_base[ntohs(type) & 5829 PTYPE_HASH_MASK]); 5830 } 5831 5832 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5833 &orig_dev->ptype_specific); 5834 5835 if (unlikely(skb->dev != orig_dev)) { 5836 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5837 &skb->dev->ptype_specific); 5838 } 5839 5840 if (pt_prev) { 5841 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 5842 goto drop; 5843 *ppt_prev = pt_prev; 5844 } else { 5845 drop: 5846 if (!deliver_exact) 5847 dev_core_stats_rx_dropped_inc(skb->dev); 5848 else 5849 dev_core_stats_rx_nohandler_inc(skb->dev); 5850 kfree_skb_reason(skb, SKB_DROP_REASON_UNHANDLED_PROTO); 5851 /* Jamal, now you will not able to escape explaining 5852 * me how you were going to use this. :-) 5853 */ 5854 ret = NET_RX_DROP; 5855 } 5856 5857 out: 5858 /* The invariant here is that if *ppt_prev is not NULL 5859 * then skb should also be non-NULL. 5860 * 5861 * Apparently *ppt_prev assignment above holds this invariant due to 5862 * skb dereferencing near it. 5863 */ 5864 *pskb = skb; 5865 return ret; 5866 } 5867 5868 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc) 5869 { 5870 struct net_device *orig_dev = skb->dev; 5871 struct packet_type *pt_prev = NULL; 5872 int ret; 5873 5874 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5875 if (pt_prev) 5876 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb, 5877 skb->dev, pt_prev, orig_dev); 5878 return ret; 5879 } 5880 5881 /** 5882 * netif_receive_skb_core - special purpose version of netif_receive_skb 5883 * @skb: buffer to process 5884 * 5885 * More direct receive version of netif_receive_skb(). It should 5886 * only be used by callers that have a need to skip RPS and Generic XDP. 5887 * Caller must also take care of handling if ``(page_is_)pfmemalloc``. 5888 * 5889 * This function may only be called from softirq context and interrupts 5890 * should be enabled. 5891 * 5892 * Return values (usually ignored): 5893 * NET_RX_SUCCESS: no congestion 5894 * NET_RX_DROP: packet was dropped 5895 */ 5896 int netif_receive_skb_core(struct sk_buff *skb) 5897 { 5898 int ret; 5899 5900 rcu_read_lock(); 5901 ret = __netif_receive_skb_one_core(skb, false); 5902 rcu_read_unlock(); 5903 5904 return ret; 5905 } 5906 EXPORT_SYMBOL(netif_receive_skb_core); 5907 5908 static inline void __netif_receive_skb_list_ptype(struct list_head *head, 5909 struct packet_type *pt_prev, 5910 struct net_device *orig_dev) 5911 { 5912 struct sk_buff *skb, *next; 5913 5914 if (!pt_prev) 5915 return; 5916 if (list_empty(head)) 5917 return; 5918 if (pt_prev->list_func != NULL) 5919 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv, 5920 ip_list_rcv, head, pt_prev, orig_dev); 5921 else 5922 list_for_each_entry_safe(skb, next, head, list) { 5923 skb_list_del_init(skb); 5924 pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 5925 } 5926 } 5927 5928 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc) 5929 { 5930 /* Fast-path assumptions: 5931 * - There is no RX handler. 5932 * - Only one packet_type matches. 5933 * If either of these fails, we will end up doing some per-packet 5934 * processing in-line, then handling the 'last ptype' for the whole 5935 * sublist. This can't cause out-of-order delivery to any single ptype, 5936 * because the 'last ptype' must be constant across the sublist, and all 5937 * other ptypes are handled per-packet. 5938 */ 5939 /* Current (common) ptype of sublist */ 5940 struct packet_type *pt_curr = NULL; 5941 /* Current (common) orig_dev of sublist */ 5942 struct net_device *od_curr = NULL; 5943 struct sk_buff *skb, *next; 5944 LIST_HEAD(sublist); 5945 5946 list_for_each_entry_safe(skb, next, head, list) { 5947 struct net_device *orig_dev = skb->dev; 5948 struct packet_type *pt_prev = NULL; 5949 5950 skb_list_del_init(skb); 5951 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5952 if (!pt_prev) 5953 continue; 5954 if (pt_curr != pt_prev || od_curr != orig_dev) { 5955 /* dispatch old sublist */ 5956 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5957 /* start new sublist */ 5958 INIT_LIST_HEAD(&sublist); 5959 pt_curr = pt_prev; 5960 od_curr = orig_dev; 5961 } 5962 list_add_tail(&skb->list, &sublist); 5963 } 5964 5965 /* dispatch final sublist */ 5966 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5967 } 5968 5969 static int __netif_receive_skb(struct sk_buff *skb) 5970 { 5971 int ret; 5972 5973 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) { 5974 unsigned int noreclaim_flag; 5975 5976 /* 5977 * PFMEMALLOC skbs are special, they should 5978 * - be delivered to SOCK_MEMALLOC sockets only 5979 * - stay away from userspace 5980 * - have bounded memory usage 5981 * 5982 * Use PF_MEMALLOC as this saves us from propagating the allocation 5983 * context down to all allocation sites. 5984 */ 5985 noreclaim_flag = memalloc_noreclaim_save(); 5986 ret = __netif_receive_skb_one_core(skb, true); 5987 memalloc_noreclaim_restore(noreclaim_flag); 5988 } else 5989 ret = __netif_receive_skb_one_core(skb, false); 5990 5991 return ret; 5992 } 5993 5994 static void __netif_receive_skb_list(struct list_head *head) 5995 { 5996 unsigned long noreclaim_flag = 0; 5997 struct sk_buff *skb, *next; 5998 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */ 5999 6000 list_for_each_entry_safe(skb, next, head, list) { 6001 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) { 6002 struct list_head sublist; 6003 6004 /* Handle the previous sublist */ 6005 list_cut_before(&sublist, head, &skb->list); 6006 if (!list_empty(&sublist)) 6007 __netif_receive_skb_list_core(&sublist, pfmemalloc); 6008 pfmemalloc = !pfmemalloc; 6009 /* See comments in __netif_receive_skb */ 6010 if (pfmemalloc) 6011 noreclaim_flag = memalloc_noreclaim_save(); 6012 else 6013 memalloc_noreclaim_restore(noreclaim_flag); 6014 } 6015 } 6016 /* Handle the remaining sublist */ 6017 if (!list_empty(head)) 6018 __netif_receive_skb_list_core(head, pfmemalloc); 6019 /* Restore pflags */ 6020 if (pfmemalloc) 6021 memalloc_noreclaim_restore(noreclaim_flag); 6022 } 6023 6024 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp) 6025 { 6026 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog); 6027 struct bpf_prog *new = xdp->prog; 6028 int ret = 0; 6029 6030 switch (xdp->command) { 6031 case XDP_SETUP_PROG: 6032 rcu_assign_pointer(dev->xdp_prog, new); 6033 if (old) 6034 bpf_prog_put(old); 6035 6036 if (old && !new) { 6037 static_branch_dec(&generic_xdp_needed_key); 6038 } else if (new && !old) { 6039 static_branch_inc(&generic_xdp_needed_key); 6040 netif_disable_lro(dev); 6041 dev_disable_gro_hw(dev); 6042 } 6043 break; 6044 6045 default: 6046 ret = -EINVAL; 6047 break; 6048 } 6049 6050 return ret; 6051 } 6052 6053 static int netif_receive_skb_internal(struct sk_buff *skb) 6054 { 6055 int ret; 6056 6057 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 6058 6059 if (skb_defer_rx_timestamp(skb)) 6060 return NET_RX_SUCCESS; 6061 6062 rcu_read_lock(); 6063 #ifdef CONFIG_RPS 6064 if (static_branch_unlikely(&rps_needed)) { 6065 struct rps_dev_flow voidflow, *rflow = &voidflow; 6066 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6067 6068 if (cpu >= 0) { 6069 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6070 rcu_read_unlock(); 6071 return ret; 6072 } 6073 } 6074 #endif 6075 ret = __netif_receive_skb(skb); 6076 rcu_read_unlock(); 6077 return ret; 6078 } 6079 6080 void netif_receive_skb_list_internal(struct list_head *head) 6081 { 6082 struct sk_buff *skb, *next; 6083 LIST_HEAD(sublist); 6084 6085 list_for_each_entry_safe(skb, next, head, list) { 6086 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), 6087 skb); 6088 skb_list_del_init(skb); 6089 if (!skb_defer_rx_timestamp(skb)) 6090 list_add_tail(&skb->list, &sublist); 6091 } 6092 list_splice_init(&sublist, head); 6093 6094 rcu_read_lock(); 6095 #ifdef CONFIG_RPS 6096 if (static_branch_unlikely(&rps_needed)) { 6097 list_for_each_entry_safe(skb, next, head, list) { 6098 struct rps_dev_flow voidflow, *rflow = &voidflow; 6099 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6100 6101 if (cpu >= 0) { 6102 /* Will be handled, remove from list */ 6103 skb_list_del_init(skb); 6104 enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6105 } 6106 } 6107 } 6108 #endif 6109 __netif_receive_skb_list(head); 6110 rcu_read_unlock(); 6111 } 6112 6113 /** 6114 * netif_receive_skb - process receive buffer from network 6115 * @skb: buffer to process 6116 * 6117 * netif_receive_skb() is the main receive data processing function. 6118 * It always succeeds. The buffer may be dropped during processing 6119 * for congestion control or by the protocol layers. 6120 * 6121 * This function may only be called from softirq context and interrupts 6122 * should be enabled. 6123 * 6124 * Return values (usually ignored): 6125 * NET_RX_SUCCESS: no congestion 6126 * NET_RX_DROP: packet was dropped 6127 */ 6128 int netif_receive_skb(struct sk_buff *skb) 6129 { 6130 int ret; 6131 6132 trace_netif_receive_skb_entry(skb); 6133 6134 ret = netif_receive_skb_internal(skb); 6135 trace_netif_receive_skb_exit(ret); 6136 6137 return ret; 6138 } 6139 EXPORT_SYMBOL(netif_receive_skb); 6140 6141 /** 6142 * netif_receive_skb_list - process many receive buffers from network 6143 * @head: list of skbs to process. 6144 * 6145 * Since return value of netif_receive_skb() is normally ignored, and 6146 * wouldn't be meaningful for a list, this function returns void. 6147 * 6148 * This function may only be called from softirq context and interrupts 6149 * should be enabled. 6150 */ 6151 void netif_receive_skb_list(struct list_head *head) 6152 { 6153 struct sk_buff *skb; 6154 6155 if (list_empty(head)) 6156 return; 6157 if (trace_netif_receive_skb_list_entry_enabled()) { 6158 list_for_each_entry(skb, head, list) 6159 trace_netif_receive_skb_list_entry(skb); 6160 } 6161 netif_receive_skb_list_internal(head); 6162 trace_netif_receive_skb_list_exit(0); 6163 } 6164 EXPORT_SYMBOL(netif_receive_skb_list); 6165 6166 /* Network device is going away, flush any packets still pending */ 6167 static void flush_backlog(struct work_struct *work) 6168 { 6169 struct sk_buff *skb, *tmp; 6170 struct sk_buff_head list; 6171 struct softnet_data *sd; 6172 6173 __skb_queue_head_init(&list); 6174 local_bh_disable(); 6175 sd = this_cpu_ptr(&softnet_data); 6176 6177 backlog_lock_irq_disable(sd); 6178 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) { 6179 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6180 __skb_unlink(skb, &sd->input_pkt_queue); 6181 __skb_queue_tail(&list, skb); 6182 rps_input_queue_head_incr(sd); 6183 } 6184 } 6185 backlog_unlock_irq_enable(sd); 6186 6187 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6188 skb_queue_walk_safe(&sd->process_queue, skb, tmp) { 6189 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6190 __skb_unlink(skb, &sd->process_queue); 6191 __skb_queue_tail(&list, skb); 6192 rps_input_queue_head_incr(sd); 6193 } 6194 } 6195 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6196 local_bh_enable(); 6197 6198 __skb_queue_purge_reason(&list, SKB_DROP_REASON_DEV_READY); 6199 } 6200 6201 static bool flush_required(int cpu) 6202 { 6203 #if IS_ENABLED(CONFIG_RPS) 6204 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 6205 bool do_flush; 6206 6207 backlog_lock_irq_disable(sd); 6208 6209 /* as insertion into process_queue happens with the rps lock held, 6210 * process_queue access may race only with dequeue 6211 */ 6212 do_flush = !skb_queue_empty(&sd->input_pkt_queue) || 6213 !skb_queue_empty_lockless(&sd->process_queue); 6214 backlog_unlock_irq_enable(sd); 6215 6216 return do_flush; 6217 #endif 6218 /* without RPS we can't safely check input_pkt_queue: during a 6219 * concurrent remote skb_queue_splice() we can detect as empty both 6220 * input_pkt_queue and process_queue even if the latter could end-up 6221 * containing a lot of packets. 6222 */ 6223 return true; 6224 } 6225 6226 struct flush_backlogs { 6227 cpumask_t flush_cpus; 6228 struct work_struct w[]; 6229 }; 6230 6231 static struct flush_backlogs *flush_backlogs_alloc(void) 6232 { 6233 return kmalloc(struct_size_t(struct flush_backlogs, w, nr_cpu_ids), 6234 GFP_KERNEL); 6235 } 6236 6237 static struct flush_backlogs *flush_backlogs_fallback; 6238 static DEFINE_MUTEX(flush_backlogs_mutex); 6239 6240 static void flush_all_backlogs(void) 6241 { 6242 struct flush_backlogs *ptr = flush_backlogs_alloc(); 6243 unsigned int cpu; 6244 6245 if (!ptr) { 6246 mutex_lock(&flush_backlogs_mutex); 6247 ptr = flush_backlogs_fallback; 6248 } 6249 cpumask_clear(&ptr->flush_cpus); 6250 6251 cpus_read_lock(); 6252 6253 for_each_online_cpu(cpu) { 6254 if (flush_required(cpu)) { 6255 INIT_WORK(&ptr->w[cpu], flush_backlog); 6256 queue_work_on(cpu, system_highpri_wq, &ptr->w[cpu]); 6257 __cpumask_set_cpu(cpu, &ptr->flush_cpus); 6258 } 6259 } 6260 6261 /* we can have in flight packet[s] on the cpus we are not flushing, 6262 * synchronize_net() in unregister_netdevice_many() will take care of 6263 * them. 6264 */ 6265 for_each_cpu(cpu, &ptr->flush_cpus) 6266 flush_work(&ptr->w[cpu]); 6267 6268 cpus_read_unlock(); 6269 6270 if (ptr != flush_backlogs_fallback) 6271 kfree(ptr); 6272 else 6273 mutex_unlock(&flush_backlogs_mutex); 6274 } 6275 6276 static void net_rps_send_ipi(struct softnet_data *remsd) 6277 { 6278 #ifdef CONFIG_RPS 6279 while (remsd) { 6280 struct softnet_data *next = remsd->rps_ipi_next; 6281 6282 if (cpu_online(remsd->cpu)) 6283 smp_call_function_single_async(remsd->cpu, &remsd->csd); 6284 remsd = next; 6285 } 6286 #endif 6287 } 6288 6289 /* 6290 * net_rps_action_and_irq_enable sends any pending IPI's for rps. 6291 * Note: called with local irq disabled, but exits with local irq enabled. 6292 */ 6293 static void net_rps_action_and_irq_enable(struct softnet_data *sd) 6294 { 6295 #ifdef CONFIG_RPS 6296 struct softnet_data *remsd = sd->rps_ipi_list; 6297 6298 if (!use_backlog_threads() && remsd) { 6299 sd->rps_ipi_list = NULL; 6300 6301 local_irq_enable(); 6302 6303 /* Send pending IPI's to kick RPS processing on remote cpus. */ 6304 net_rps_send_ipi(remsd); 6305 } else 6306 #endif 6307 local_irq_enable(); 6308 } 6309 6310 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd) 6311 { 6312 #ifdef CONFIG_RPS 6313 return !use_backlog_threads() && sd->rps_ipi_list; 6314 #else 6315 return false; 6316 #endif 6317 } 6318 6319 static int process_backlog(struct napi_struct *napi, int quota) 6320 { 6321 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog); 6322 bool again = true; 6323 int work = 0; 6324 6325 /* Check if we have pending ipi, its better to send them now, 6326 * not waiting net_rx_action() end. 6327 */ 6328 if (sd_has_rps_ipi_waiting(sd)) { 6329 local_irq_disable(); 6330 net_rps_action_and_irq_enable(sd); 6331 } 6332 6333 napi->weight = READ_ONCE(net_hotdata.dev_rx_weight); 6334 while (again) { 6335 struct sk_buff *skb; 6336 6337 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6338 while ((skb = __skb_dequeue(&sd->process_queue))) { 6339 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6340 rcu_read_lock(); 6341 __netif_receive_skb(skb); 6342 rcu_read_unlock(); 6343 if (++work >= quota) { 6344 rps_input_queue_head_add(sd, work); 6345 return work; 6346 } 6347 6348 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6349 } 6350 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6351 6352 backlog_lock_irq_disable(sd); 6353 if (skb_queue_empty(&sd->input_pkt_queue)) { 6354 /* 6355 * Inline a custom version of __napi_complete(). 6356 * only current cpu owns and manipulates this napi, 6357 * and NAPI_STATE_SCHED is the only possible flag set 6358 * on backlog. 6359 * We can use a plain write instead of clear_bit(), 6360 * and we dont need an smp_mb() memory barrier. 6361 */ 6362 napi->state &= NAPIF_STATE_THREADED; 6363 again = false; 6364 } else { 6365 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6366 skb_queue_splice_tail_init(&sd->input_pkt_queue, 6367 &sd->process_queue); 6368 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6369 } 6370 backlog_unlock_irq_enable(sd); 6371 } 6372 6373 if (work) 6374 rps_input_queue_head_add(sd, work); 6375 return work; 6376 } 6377 6378 /** 6379 * __napi_schedule - schedule for receive 6380 * @n: entry to schedule 6381 * 6382 * The entry's receive function will be scheduled to run. 6383 * Consider using __napi_schedule_irqoff() if hard irqs are masked. 6384 */ 6385 void __napi_schedule(struct napi_struct *n) 6386 { 6387 unsigned long flags; 6388 6389 local_irq_save(flags); 6390 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6391 local_irq_restore(flags); 6392 } 6393 EXPORT_SYMBOL(__napi_schedule); 6394 6395 /** 6396 * napi_schedule_prep - check if napi can be scheduled 6397 * @n: napi context 6398 * 6399 * Test if NAPI routine is already running, and if not mark 6400 * it as running. This is used as a condition variable to 6401 * insure only one NAPI poll instance runs. We also make 6402 * sure there is no pending NAPI disable. 6403 */ 6404 bool napi_schedule_prep(struct napi_struct *n) 6405 { 6406 unsigned long new, val = READ_ONCE(n->state); 6407 6408 do { 6409 if (unlikely(val & NAPIF_STATE_DISABLE)) 6410 return false; 6411 new = val | NAPIF_STATE_SCHED; 6412 6413 /* Sets STATE_MISSED bit if STATE_SCHED was already set 6414 * This was suggested by Alexander Duyck, as compiler 6415 * emits better code than : 6416 * if (val & NAPIF_STATE_SCHED) 6417 * new |= NAPIF_STATE_MISSED; 6418 */ 6419 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED * 6420 NAPIF_STATE_MISSED; 6421 } while (!try_cmpxchg(&n->state, &val, new)); 6422 6423 return !(val & NAPIF_STATE_SCHED); 6424 } 6425 EXPORT_SYMBOL(napi_schedule_prep); 6426 6427 /** 6428 * __napi_schedule_irqoff - schedule for receive 6429 * @n: entry to schedule 6430 * 6431 * Variant of __napi_schedule() assuming hard irqs are masked. 6432 * 6433 * On PREEMPT_RT enabled kernels this maps to __napi_schedule() 6434 * because the interrupt disabled assumption might not be true 6435 * due to force-threaded interrupts and spinlock substitution. 6436 */ 6437 void __napi_schedule_irqoff(struct napi_struct *n) 6438 { 6439 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6440 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6441 else 6442 __napi_schedule(n); 6443 } 6444 EXPORT_SYMBOL(__napi_schedule_irqoff); 6445 6446 bool napi_complete_done(struct napi_struct *n, int work_done) 6447 { 6448 unsigned long flags, val, new, timeout = 0; 6449 bool ret = true; 6450 6451 /* 6452 * 1) Don't let napi dequeue from the cpu poll list 6453 * just in case its running on a different cpu. 6454 * 2) If we are busy polling, do nothing here, we have 6455 * the guarantee we will be called later. 6456 */ 6457 if (unlikely(n->state & (NAPIF_STATE_NPSVC | 6458 NAPIF_STATE_IN_BUSY_POLL))) 6459 return false; 6460 6461 if (work_done) { 6462 if (n->gro.bitmask) 6463 timeout = napi_get_gro_flush_timeout(n); 6464 n->defer_hard_irqs_count = napi_get_defer_hard_irqs(n); 6465 } 6466 if (n->defer_hard_irqs_count > 0) { 6467 n->defer_hard_irqs_count--; 6468 timeout = napi_get_gro_flush_timeout(n); 6469 if (timeout) 6470 ret = false; 6471 } 6472 6473 /* 6474 * When the NAPI instance uses a timeout and keeps postponing 6475 * it, we need to bound somehow the time packets are kept in 6476 * the GRO layer. 6477 */ 6478 gro_flush(&n->gro, !!timeout); 6479 gro_normal_list(&n->gro); 6480 6481 if (unlikely(!list_empty(&n->poll_list))) { 6482 /* If n->poll_list is not empty, we need to mask irqs */ 6483 local_irq_save(flags); 6484 list_del_init(&n->poll_list); 6485 local_irq_restore(flags); 6486 } 6487 WRITE_ONCE(n->list_owner, -1); 6488 6489 val = READ_ONCE(n->state); 6490 do { 6491 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED)); 6492 6493 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED | 6494 NAPIF_STATE_SCHED_THREADED | 6495 NAPIF_STATE_PREFER_BUSY_POLL); 6496 6497 /* If STATE_MISSED was set, leave STATE_SCHED set, 6498 * because we will call napi->poll() one more time. 6499 * This C code was suggested by Alexander Duyck to help gcc. 6500 */ 6501 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED * 6502 NAPIF_STATE_SCHED; 6503 } while (!try_cmpxchg(&n->state, &val, new)); 6504 6505 if (unlikely(val & NAPIF_STATE_MISSED)) { 6506 __napi_schedule(n); 6507 return false; 6508 } 6509 6510 if (timeout) 6511 hrtimer_start(&n->timer, ns_to_ktime(timeout), 6512 HRTIMER_MODE_REL_PINNED); 6513 return ret; 6514 } 6515 EXPORT_SYMBOL(napi_complete_done); 6516 6517 static void skb_defer_free_flush(struct softnet_data *sd) 6518 { 6519 struct sk_buff *skb, *next; 6520 6521 /* Paired with WRITE_ONCE() in skb_attempt_defer_free() */ 6522 if (!READ_ONCE(sd->defer_list)) 6523 return; 6524 6525 spin_lock(&sd->defer_lock); 6526 skb = sd->defer_list; 6527 sd->defer_list = NULL; 6528 sd->defer_count = 0; 6529 spin_unlock(&sd->defer_lock); 6530 6531 while (skb != NULL) { 6532 next = skb->next; 6533 napi_consume_skb(skb, 1); 6534 skb = next; 6535 } 6536 } 6537 6538 #if defined(CONFIG_NET_RX_BUSY_POLL) 6539 6540 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule) 6541 { 6542 if (!skip_schedule) { 6543 gro_normal_list(&napi->gro); 6544 __napi_schedule(napi); 6545 return; 6546 } 6547 6548 /* Flush too old packets. If HZ < 1000, flush all packets */ 6549 gro_flush(&napi->gro, HZ >= 1000); 6550 gro_normal_list(&napi->gro); 6551 6552 clear_bit(NAPI_STATE_SCHED, &napi->state); 6553 } 6554 6555 enum { 6556 NAPI_F_PREFER_BUSY_POLL = 1, 6557 NAPI_F_END_ON_RESCHED = 2, 6558 }; 6559 6560 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, 6561 unsigned flags, u16 budget) 6562 { 6563 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6564 bool skip_schedule = false; 6565 unsigned long timeout; 6566 int rc; 6567 6568 /* Busy polling means there is a high chance device driver hard irq 6569 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was 6570 * set in napi_schedule_prep(). 6571 * Since we are about to call napi->poll() once more, we can safely 6572 * clear NAPI_STATE_MISSED. 6573 * 6574 * Note: x86 could use a single "lock and ..." instruction 6575 * to perform these two clear_bit() 6576 */ 6577 clear_bit(NAPI_STATE_MISSED, &napi->state); 6578 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state); 6579 6580 local_bh_disable(); 6581 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6582 6583 if (flags & NAPI_F_PREFER_BUSY_POLL) { 6584 napi->defer_hard_irqs_count = napi_get_defer_hard_irqs(napi); 6585 timeout = napi_get_gro_flush_timeout(napi); 6586 if (napi->defer_hard_irqs_count && timeout) { 6587 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED); 6588 skip_schedule = true; 6589 } 6590 } 6591 6592 /* All we really want here is to re-enable device interrupts. 6593 * Ideally, a new ndo_busy_poll_stop() could avoid another round. 6594 */ 6595 rc = napi->poll(napi, budget); 6596 /* We can't gro_normal_list() here, because napi->poll() might have 6597 * rearmed the napi (napi_complete_done()) in which case it could 6598 * already be running on another CPU. 6599 */ 6600 trace_napi_poll(napi, rc, budget); 6601 netpoll_poll_unlock(have_poll_lock); 6602 if (rc == budget) 6603 __busy_poll_stop(napi, skip_schedule); 6604 bpf_net_ctx_clear(bpf_net_ctx); 6605 local_bh_enable(); 6606 } 6607 6608 static void __napi_busy_loop(unsigned int napi_id, 6609 bool (*loop_end)(void *, unsigned long), 6610 void *loop_end_arg, unsigned flags, u16 budget) 6611 { 6612 unsigned long start_time = loop_end ? busy_loop_current_time() : 0; 6613 int (*napi_poll)(struct napi_struct *napi, int budget); 6614 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6615 void *have_poll_lock = NULL; 6616 struct napi_struct *napi; 6617 6618 WARN_ON_ONCE(!rcu_read_lock_held()); 6619 6620 restart: 6621 napi_poll = NULL; 6622 6623 napi = napi_by_id(napi_id); 6624 if (!napi) 6625 return; 6626 6627 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6628 preempt_disable(); 6629 for (;;) { 6630 int work = 0; 6631 6632 local_bh_disable(); 6633 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6634 if (!napi_poll) { 6635 unsigned long val = READ_ONCE(napi->state); 6636 6637 /* If multiple threads are competing for this napi, 6638 * we avoid dirtying napi->state as much as we can. 6639 */ 6640 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED | 6641 NAPIF_STATE_IN_BUSY_POLL)) { 6642 if (flags & NAPI_F_PREFER_BUSY_POLL) 6643 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6644 goto count; 6645 } 6646 if (cmpxchg(&napi->state, val, 6647 val | NAPIF_STATE_IN_BUSY_POLL | 6648 NAPIF_STATE_SCHED) != val) { 6649 if (flags & NAPI_F_PREFER_BUSY_POLL) 6650 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6651 goto count; 6652 } 6653 have_poll_lock = netpoll_poll_lock(napi); 6654 napi_poll = napi->poll; 6655 } 6656 work = napi_poll(napi, budget); 6657 trace_napi_poll(napi, work, budget); 6658 gro_normal_list(&napi->gro); 6659 count: 6660 if (work > 0) 6661 __NET_ADD_STATS(dev_net(napi->dev), 6662 LINUX_MIB_BUSYPOLLRXPACKETS, work); 6663 skb_defer_free_flush(this_cpu_ptr(&softnet_data)); 6664 bpf_net_ctx_clear(bpf_net_ctx); 6665 local_bh_enable(); 6666 6667 if (!loop_end || loop_end(loop_end_arg, start_time)) 6668 break; 6669 6670 if (unlikely(need_resched())) { 6671 if (flags & NAPI_F_END_ON_RESCHED) 6672 break; 6673 if (napi_poll) 6674 busy_poll_stop(napi, have_poll_lock, flags, budget); 6675 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6676 preempt_enable(); 6677 rcu_read_unlock(); 6678 cond_resched(); 6679 rcu_read_lock(); 6680 if (loop_end(loop_end_arg, start_time)) 6681 return; 6682 goto restart; 6683 } 6684 cpu_relax(); 6685 } 6686 if (napi_poll) 6687 busy_poll_stop(napi, have_poll_lock, flags, budget); 6688 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6689 preempt_enable(); 6690 } 6691 6692 void napi_busy_loop_rcu(unsigned int napi_id, 6693 bool (*loop_end)(void *, unsigned long), 6694 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6695 { 6696 unsigned flags = NAPI_F_END_ON_RESCHED; 6697 6698 if (prefer_busy_poll) 6699 flags |= NAPI_F_PREFER_BUSY_POLL; 6700 6701 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6702 } 6703 6704 void napi_busy_loop(unsigned int napi_id, 6705 bool (*loop_end)(void *, unsigned long), 6706 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6707 { 6708 unsigned flags = prefer_busy_poll ? NAPI_F_PREFER_BUSY_POLL : 0; 6709 6710 rcu_read_lock(); 6711 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6712 rcu_read_unlock(); 6713 } 6714 EXPORT_SYMBOL(napi_busy_loop); 6715 6716 void napi_suspend_irqs(unsigned int napi_id) 6717 { 6718 struct napi_struct *napi; 6719 6720 rcu_read_lock(); 6721 napi = napi_by_id(napi_id); 6722 if (napi) { 6723 unsigned long timeout = napi_get_irq_suspend_timeout(napi); 6724 6725 if (timeout) 6726 hrtimer_start(&napi->timer, ns_to_ktime(timeout), 6727 HRTIMER_MODE_REL_PINNED); 6728 } 6729 rcu_read_unlock(); 6730 } 6731 6732 void napi_resume_irqs(unsigned int napi_id) 6733 { 6734 struct napi_struct *napi; 6735 6736 rcu_read_lock(); 6737 napi = napi_by_id(napi_id); 6738 if (napi) { 6739 /* If irq_suspend_timeout is set to 0 between the call to 6740 * napi_suspend_irqs and now, the original value still 6741 * determines the safety timeout as intended and napi_watchdog 6742 * will resume irq processing. 6743 */ 6744 if (napi_get_irq_suspend_timeout(napi)) { 6745 local_bh_disable(); 6746 napi_schedule(napi); 6747 local_bh_enable(); 6748 } 6749 } 6750 rcu_read_unlock(); 6751 } 6752 6753 #endif /* CONFIG_NET_RX_BUSY_POLL */ 6754 6755 static void __napi_hash_add_with_id(struct napi_struct *napi, 6756 unsigned int napi_id) 6757 { 6758 napi->gro.cached_napi_id = napi_id; 6759 6760 WRITE_ONCE(napi->napi_id, napi_id); 6761 hlist_add_head_rcu(&napi->napi_hash_node, 6762 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]); 6763 } 6764 6765 static void napi_hash_add_with_id(struct napi_struct *napi, 6766 unsigned int napi_id) 6767 { 6768 unsigned long flags; 6769 6770 spin_lock_irqsave(&napi_hash_lock, flags); 6771 WARN_ON_ONCE(napi_by_id(napi_id)); 6772 __napi_hash_add_with_id(napi, napi_id); 6773 spin_unlock_irqrestore(&napi_hash_lock, flags); 6774 } 6775 6776 static void napi_hash_add(struct napi_struct *napi) 6777 { 6778 unsigned long flags; 6779 6780 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state)) 6781 return; 6782 6783 spin_lock_irqsave(&napi_hash_lock, flags); 6784 6785 /* 0..NR_CPUS range is reserved for sender_cpu use */ 6786 do { 6787 if (unlikely(!napi_id_valid(++napi_gen_id))) 6788 napi_gen_id = MIN_NAPI_ID; 6789 } while (napi_by_id(napi_gen_id)); 6790 6791 __napi_hash_add_with_id(napi, napi_gen_id); 6792 6793 spin_unlock_irqrestore(&napi_hash_lock, flags); 6794 } 6795 6796 /* Warning : caller is responsible to make sure rcu grace period 6797 * is respected before freeing memory containing @napi 6798 */ 6799 static void napi_hash_del(struct napi_struct *napi) 6800 { 6801 unsigned long flags; 6802 6803 spin_lock_irqsave(&napi_hash_lock, flags); 6804 6805 hlist_del_init_rcu(&napi->napi_hash_node); 6806 6807 spin_unlock_irqrestore(&napi_hash_lock, flags); 6808 } 6809 6810 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer) 6811 { 6812 struct napi_struct *napi; 6813 6814 napi = container_of(timer, struct napi_struct, timer); 6815 6816 /* Note : we use a relaxed variant of napi_schedule_prep() not setting 6817 * NAPI_STATE_MISSED, since we do not react to a device IRQ. 6818 */ 6819 if (!napi_disable_pending(napi) && 6820 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) { 6821 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6822 __napi_schedule_irqoff(napi); 6823 } 6824 6825 return HRTIMER_NORESTART; 6826 } 6827 6828 int dev_set_threaded(struct net_device *dev, bool threaded) 6829 { 6830 struct napi_struct *napi; 6831 int err = 0; 6832 6833 netdev_assert_locked_or_invisible(dev); 6834 6835 if (dev->threaded == threaded) 6836 return 0; 6837 6838 if (threaded) { 6839 list_for_each_entry(napi, &dev->napi_list, dev_list) { 6840 if (!napi->thread) { 6841 err = napi_kthread_create(napi); 6842 if (err) { 6843 threaded = false; 6844 break; 6845 } 6846 } 6847 } 6848 } 6849 6850 WRITE_ONCE(dev->threaded, threaded); 6851 6852 /* Make sure kthread is created before THREADED bit 6853 * is set. 6854 */ 6855 smp_mb__before_atomic(); 6856 6857 /* Setting/unsetting threaded mode on a napi might not immediately 6858 * take effect, if the current napi instance is actively being 6859 * polled. In this case, the switch between threaded mode and 6860 * softirq mode will happen in the next round of napi_schedule(). 6861 * This should not cause hiccups/stalls to the live traffic. 6862 */ 6863 list_for_each_entry(napi, &dev->napi_list, dev_list) 6864 assign_bit(NAPI_STATE_THREADED, &napi->state, threaded); 6865 6866 return err; 6867 } 6868 EXPORT_SYMBOL(dev_set_threaded); 6869 6870 /** 6871 * netif_queue_set_napi - Associate queue with the napi 6872 * @dev: device to which NAPI and queue belong 6873 * @queue_index: Index of queue 6874 * @type: queue type as RX or TX 6875 * @napi: NAPI context, pass NULL to clear previously set NAPI 6876 * 6877 * Set queue with its corresponding napi context. This should be done after 6878 * registering the NAPI handler for the queue-vector and the queues have been 6879 * mapped to the corresponding interrupt vector. 6880 */ 6881 void netif_queue_set_napi(struct net_device *dev, unsigned int queue_index, 6882 enum netdev_queue_type type, struct napi_struct *napi) 6883 { 6884 struct netdev_rx_queue *rxq; 6885 struct netdev_queue *txq; 6886 6887 if (WARN_ON_ONCE(napi && !napi->dev)) 6888 return; 6889 if (dev->reg_state >= NETREG_REGISTERED) 6890 ASSERT_RTNL(); 6891 6892 switch (type) { 6893 case NETDEV_QUEUE_TYPE_RX: 6894 rxq = __netif_get_rx_queue(dev, queue_index); 6895 rxq->napi = napi; 6896 return; 6897 case NETDEV_QUEUE_TYPE_TX: 6898 txq = netdev_get_tx_queue(dev, queue_index); 6899 txq->napi = napi; 6900 return; 6901 default: 6902 return; 6903 } 6904 } 6905 EXPORT_SYMBOL(netif_queue_set_napi); 6906 6907 static void 6908 netif_napi_irq_notify(struct irq_affinity_notify *notify, 6909 const cpumask_t *mask) 6910 { 6911 struct napi_struct *napi = 6912 container_of(notify, struct napi_struct, notify); 6913 #ifdef CONFIG_RFS_ACCEL 6914 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 6915 int err; 6916 #endif 6917 6918 if (napi->config && napi->dev->irq_affinity_auto) 6919 cpumask_copy(&napi->config->affinity_mask, mask); 6920 6921 #ifdef CONFIG_RFS_ACCEL 6922 if (napi->dev->rx_cpu_rmap_auto) { 6923 err = cpu_rmap_update(rmap, napi->napi_rmap_idx, mask); 6924 if (err) 6925 netdev_warn(napi->dev, "RMAP update failed (%d)\n", 6926 err); 6927 } 6928 #endif 6929 } 6930 6931 #ifdef CONFIG_RFS_ACCEL 6932 static void netif_napi_affinity_release(struct kref *ref) 6933 { 6934 struct napi_struct *napi = 6935 container_of(ref, struct napi_struct, notify.kref); 6936 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 6937 6938 netdev_assert_locked(napi->dev); 6939 WARN_ON(test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, 6940 &napi->state)); 6941 6942 if (!napi->dev->rx_cpu_rmap_auto) 6943 return; 6944 rmap->obj[napi->napi_rmap_idx] = NULL; 6945 napi->napi_rmap_idx = -1; 6946 cpu_rmap_put(rmap); 6947 } 6948 6949 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 6950 { 6951 if (dev->rx_cpu_rmap_auto) 6952 return 0; 6953 6954 dev->rx_cpu_rmap = alloc_irq_cpu_rmap(num_irqs); 6955 if (!dev->rx_cpu_rmap) 6956 return -ENOMEM; 6957 6958 dev->rx_cpu_rmap_auto = true; 6959 return 0; 6960 } 6961 EXPORT_SYMBOL(netif_enable_cpu_rmap); 6962 6963 static void netif_del_cpu_rmap(struct net_device *dev) 6964 { 6965 struct cpu_rmap *rmap = dev->rx_cpu_rmap; 6966 6967 if (!dev->rx_cpu_rmap_auto) 6968 return; 6969 6970 /* Free the rmap */ 6971 cpu_rmap_put(rmap); 6972 dev->rx_cpu_rmap = NULL; 6973 dev->rx_cpu_rmap_auto = false; 6974 } 6975 6976 #else 6977 static void netif_napi_affinity_release(struct kref *ref) 6978 { 6979 } 6980 6981 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 6982 { 6983 return 0; 6984 } 6985 EXPORT_SYMBOL(netif_enable_cpu_rmap); 6986 6987 static void netif_del_cpu_rmap(struct net_device *dev) 6988 { 6989 } 6990 #endif 6991 6992 void netif_set_affinity_auto(struct net_device *dev) 6993 { 6994 unsigned int i, maxqs, numa; 6995 6996 maxqs = max(dev->num_tx_queues, dev->num_rx_queues); 6997 numa = dev_to_node(&dev->dev); 6998 6999 for (i = 0; i < maxqs; i++) 7000 cpumask_set_cpu(cpumask_local_spread(i, numa), 7001 &dev->napi_config[i].affinity_mask); 7002 7003 dev->irq_affinity_auto = true; 7004 } 7005 EXPORT_SYMBOL(netif_set_affinity_auto); 7006 7007 void netif_napi_set_irq_locked(struct napi_struct *napi, int irq) 7008 { 7009 int rc; 7010 7011 netdev_assert_locked_or_invisible(napi->dev); 7012 7013 if (napi->irq == irq) 7014 return; 7015 7016 /* Remove existing resources */ 7017 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7018 irq_set_affinity_notifier(napi->irq, NULL); 7019 7020 napi->irq = irq; 7021 if (irq < 0 || 7022 (!napi->dev->rx_cpu_rmap_auto && !napi->dev->irq_affinity_auto)) 7023 return; 7024 7025 /* Abort for buggy drivers */ 7026 if (napi->dev->irq_affinity_auto && WARN_ON_ONCE(!napi->config)) 7027 return; 7028 7029 #ifdef CONFIG_RFS_ACCEL 7030 if (napi->dev->rx_cpu_rmap_auto) { 7031 rc = cpu_rmap_add(napi->dev->rx_cpu_rmap, napi); 7032 if (rc < 0) 7033 return; 7034 7035 cpu_rmap_get(napi->dev->rx_cpu_rmap); 7036 napi->napi_rmap_idx = rc; 7037 } 7038 #endif 7039 7040 /* Use core IRQ notifier */ 7041 napi->notify.notify = netif_napi_irq_notify; 7042 napi->notify.release = netif_napi_affinity_release; 7043 rc = irq_set_affinity_notifier(irq, &napi->notify); 7044 if (rc) { 7045 netdev_warn(napi->dev, "Unable to set IRQ notifier (%d)\n", 7046 rc); 7047 goto put_rmap; 7048 } 7049 7050 set_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state); 7051 return; 7052 7053 put_rmap: 7054 #ifdef CONFIG_RFS_ACCEL 7055 if (napi->dev->rx_cpu_rmap_auto) { 7056 napi->dev->rx_cpu_rmap->obj[napi->napi_rmap_idx] = NULL; 7057 cpu_rmap_put(napi->dev->rx_cpu_rmap); 7058 napi->napi_rmap_idx = -1; 7059 } 7060 #endif 7061 napi->notify.notify = NULL; 7062 napi->notify.release = NULL; 7063 } 7064 EXPORT_SYMBOL(netif_napi_set_irq_locked); 7065 7066 static void napi_restore_config(struct napi_struct *n) 7067 { 7068 n->defer_hard_irqs = n->config->defer_hard_irqs; 7069 n->gro_flush_timeout = n->config->gro_flush_timeout; 7070 n->irq_suspend_timeout = n->config->irq_suspend_timeout; 7071 7072 if (n->dev->irq_affinity_auto && 7073 test_bit(NAPI_STATE_HAS_NOTIFIER, &n->state)) 7074 irq_set_affinity(n->irq, &n->config->affinity_mask); 7075 7076 /* a NAPI ID might be stored in the config, if so use it. if not, use 7077 * napi_hash_add to generate one for us. 7078 */ 7079 if (n->config->napi_id) { 7080 napi_hash_add_with_id(n, n->config->napi_id); 7081 } else { 7082 napi_hash_add(n); 7083 n->config->napi_id = n->napi_id; 7084 } 7085 } 7086 7087 static void napi_save_config(struct napi_struct *n) 7088 { 7089 n->config->defer_hard_irqs = n->defer_hard_irqs; 7090 n->config->gro_flush_timeout = n->gro_flush_timeout; 7091 n->config->irq_suspend_timeout = n->irq_suspend_timeout; 7092 napi_hash_del(n); 7093 } 7094 7095 /* Netlink wants the NAPI list to be sorted by ID, if adding a NAPI which will 7096 * inherit an existing ID try to insert it at the right position. 7097 */ 7098 static void 7099 netif_napi_dev_list_add(struct net_device *dev, struct napi_struct *napi) 7100 { 7101 unsigned int new_id, pos_id; 7102 struct list_head *higher; 7103 struct napi_struct *pos; 7104 7105 new_id = UINT_MAX; 7106 if (napi->config && napi->config->napi_id) 7107 new_id = napi->config->napi_id; 7108 7109 higher = &dev->napi_list; 7110 list_for_each_entry(pos, &dev->napi_list, dev_list) { 7111 if (napi_id_valid(pos->napi_id)) 7112 pos_id = pos->napi_id; 7113 else if (pos->config) 7114 pos_id = pos->config->napi_id; 7115 else 7116 pos_id = UINT_MAX; 7117 7118 if (pos_id <= new_id) 7119 break; 7120 higher = &pos->dev_list; 7121 } 7122 list_add_rcu(&napi->dev_list, higher); /* adds after higher */ 7123 } 7124 7125 /* Double check that napi_get_frags() allocates skbs with 7126 * skb->head being backed by slab, not a page fragment. 7127 * This is to make sure bug fixed in 3226b158e67c 7128 * ("net: avoid 32 x truesize under-estimation for tiny skbs") 7129 * does not accidentally come back. 7130 */ 7131 static void napi_get_frags_check(struct napi_struct *napi) 7132 { 7133 struct sk_buff *skb; 7134 7135 local_bh_disable(); 7136 skb = napi_get_frags(napi); 7137 WARN_ON_ONCE(skb && skb->head_frag); 7138 napi_free_frags(napi); 7139 local_bh_enable(); 7140 } 7141 7142 void netif_napi_add_weight_locked(struct net_device *dev, 7143 struct napi_struct *napi, 7144 int (*poll)(struct napi_struct *, int), 7145 int weight) 7146 { 7147 netdev_assert_locked(dev); 7148 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state))) 7149 return; 7150 7151 INIT_LIST_HEAD(&napi->poll_list); 7152 INIT_HLIST_NODE(&napi->napi_hash_node); 7153 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); 7154 napi->timer.function = napi_watchdog; 7155 gro_init(&napi->gro); 7156 napi->skb = NULL; 7157 napi->poll = poll; 7158 if (weight > NAPI_POLL_WEIGHT) 7159 netdev_err_once(dev, "%s() called with weight %d\n", __func__, 7160 weight); 7161 napi->weight = weight; 7162 napi->dev = dev; 7163 #ifdef CONFIG_NETPOLL 7164 napi->poll_owner = -1; 7165 #endif 7166 napi->list_owner = -1; 7167 set_bit(NAPI_STATE_SCHED, &napi->state); 7168 set_bit(NAPI_STATE_NPSVC, &napi->state); 7169 netif_napi_dev_list_add(dev, napi); 7170 7171 /* default settings from sysfs are applied to all NAPIs. any per-NAPI 7172 * configuration will be loaded in napi_enable 7173 */ 7174 napi_set_defer_hard_irqs(napi, READ_ONCE(dev->napi_defer_hard_irqs)); 7175 napi_set_gro_flush_timeout(napi, READ_ONCE(dev->gro_flush_timeout)); 7176 7177 napi_get_frags_check(napi); 7178 /* Create kthread for this napi if dev->threaded is set. 7179 * Clear dev->threaded if kthread creation failed so that 7180 * threaded mode will not be enabled in napi_enable(). 7181 */ 7182 if (dev->threaded && napi_kthread_create(napi)) 7183 dev->threaded = false; 7184 netif_napi_set_irq_locked(napi, -1); 7185 } 7186 EXPORT_SYMBOL(netif_napi_add_weight_locked); 7187 7188 void napi_disable_locked(struct napi_struct *n) 7189 { 7190 unsigned long val, new; 7191 7192 might_sleep(); 7193 netdev_assert_locked(n->dev); 7194 7195 set_bit(NAPI_STATE_DISABLE, &n->state); 7196 7197 val = READ_ONCE(n->state); 7198 do { 7199 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) { 7200 usleep_range(20, 200); 7201 val = READ_ONCE(n->state); 7202 } 7203 7204 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC; 7205 new &= ~(NAPIF_STATE_THREADED | NAPIF_STATE_PREFER_BUSY_POLL); 7206 } while (!try_cmpxchg(&n->state, &val, new)); 7207 7208 hrtimer_cancel(&n->timer); 7209 7210 if (n->config) 7211 napi_save_config(n); 7212 else 7213 napi_hash_del(n); 7214 7215 clear_bit(NAPI_STATE_DISABLE, &n->state); 7216 } 7217 EXPORT_SYMBOL(napi_disable_locked); 7218 7219 /** 7220 * napi_disable() - prevent NAPI from scheduling 7221 * @n: NAPI context 7222 * 7223 * Stop NAPI from being scheduled on this context. 7224 * Waits till any outstanding processing completes. 7225 * Takes netdev_lock() for associated net_device. 7226 */ 7227 void napi_disable(struct napi_struct *n) 7228 { 7229 netdev_lock(n->dev); 7230 napi_disable_locked(n); 7231 netdev_unlock(n->dev); 7232 } 7233 EXPORT_SYMBOL(napi_disable); 7234 7235 void napi_enable_locked(struct napi_struct *n) 7236 { 7237 unsigned long new, val = READ_ONCE(n->state); 7238 7239 if (n->config) 7240 napi_restore_config(n); 7241 else 7242 napi_hash_add(n); 7243 7244 do { 7245 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val)); 7246 7247 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC); 7248 if (n->dev->threaded && n->thread) 7249 new |= NAPIF_STATE_THREADED; 7250 } while (!try_cmpxchg(&n->state, &val, new)); 7251 } 7252 EXPORT_SYMBOL(napi_enable_locked); 7253 7254 /** 7255 * napi_enable() - enable NAPI scheduling 7256 * @n: NAPI context 7257 * 7258 * Enable scheduling of a NAPI instance. 7259 * Must be paired with napi_disable(). 7260 * Takes netdev_lock() for associated net_device. 7261 */ 7262 void napi_enable(struct napi_struct *n) 7263 { 7264 netdev_lock(n->dev); 7265 napi_enable_locked(n); 7266 netdev_unlock(n->dev); 7267 } 7268 EXPORT_SYMBOL(napi_enable); 7269 7270 /* Must be called in process context */ 7271 void __netif_napi_del_locked(struct napi_struct *napi) 7272 { 7273 netdev_assert_locked(napi->dev); 7274 7275 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state)) 7276 return; 7277 7278 /* Make sure NAPI is disabled (or was never enabled). */ 7279 WARN_ON(!test_bit(NAPI_STATE_SCHED, &napi->state)); 7280 7281 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7282 irq_set_affinity_notifier(napi->irq, NULL); 7283 7284 if (napi->config) { 7285 napi->index = -1; 7286 napi->config = NULL; 7287 } 7288 7289 list_del_rcu(&napi->dev_list); 7290 napi_free_frags(napi); 7291 7292 gro_cleanup(&napi->gro); 7293 7294 if (napi->thread) { 7295 kthread_stop(napi->thread); 7296 napi->thread = NULL; 7297 } 7298 } 7299 EXPORT_SYMBOL(__netif_napi_del_locked); 7300 7301 static int __napi_poll(struct napi_struct *n, bool *repoll) 7302 { 7303 int work, weight; 7304 7305 weight = n->weight; 7306 7307 /* This NAPI_STATE_SCHED test is for avoiding a race 7308 * with netpoll's poll_napi(). Only the entity which 7309 * obtains the lock and sees NAPI_STATE_SCHED set will 7310 * actually make the ->poll() call. Therefore we avoid 7311 * accidentally calling ->poll() when NAPI is not scheduled. 7312 */ 7313 work = 0; 7314 if (napi_is_scheduled(n)) { 7315 work = n->poll(n, weight); 7316 trace_napi_poll(n, work, weight); 7317 7318 xdp_do_check_flushed(n); 7319 } 7320 7321 if (unlikely(work > weight)) 7322 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n", 7323 n->poll, work, weight); 7324 7325 if (likely(work < weight)) 7326 return work; 7327 7328 /* Drivers must not modify the NAPI state if they 7329 * consume the entire weight. In such cases this code 7330 * still "owns" the NAPI instance and therefore can 7331 * move the instance around on the list at-will. 7332 */ 7333 if (unlikely(napi_disable_pending(n))) { 7334 napi_complete(n); 7335 return work; 7336 } 7337 7338 /* The NAPI context has more processing work, but busy-polling 7339 * is preferred. Exit early. 7340 */ 7341 if (napi_prefer_busy_poll(n)) { 7342 if (napi_complete_done(n, work)) { 7343 /* If timeout is not set, we need to make sure 7344 * that the NAPI is re-scheduled. 7345 */ 7346 napi_schedule(n); 7347 } 7348 return work; 7349 } 7350 7351 /* Flush too old packets. If HZ < 1000, flush all packets */ 7352 gro_flush(&n->gro, HZ >= 1000); 7353 gro_normal_list(&n->gro); 7354 7355 /* Some drivers may have called napi_schedule 7356 * prior to exhausting their budget. 7357 */ 7358 if (unlikely(!list_empty(&n->poll_list))) { 7359 pr_warn_once("%s: Budget exhausted after napi rescheduled\n", 7360 n->dev ? n->dev->name : "backlog"); 7361 return work; 7362 } 7363 7364 *repoll = true; 7365 7366 return work; 7367 } 7368 7369 static int napi_poll(struct napi_struct *n, struct list_head *repoll) 7370 { 7371 bool do_repoll = false; 7372 void *have; 7373 int work; 7374 7375 list_del_init(&n->poll_list); 7376 7377 have = netpoll_poll_lock(n); 7378 7379 work = __napi_poll(n, &do_repoll); 7380 7381 if (do_repoll) 7382 list_add_tail(&n->poll_list, repoll); 7383 7384 netpoll_poll_unlock(have); 7385 7386 return work; 7387 } 7388 7389 static int napi_thread_wait(struct napi_struct *napi) 7390 { 7391 set_current_state(TASK_INTERRUPTIBLE); 7392 7393 while (!kthread_should_stop()) { 7394 /* Testing SCHED_THREADED bit here to make sure the current 7395 * kthread owns this napi and could poll on this napi. 7396 * Testing SCHED bit is not enough because SCHED bit might be 7397 * set by some other busy poll thread or by napi_disable(). 7398 */ 7399 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) { 7400 WARN_ON(!list_empty(&napi->poll_list)); 7401 __set_current_state(TASK_RUNNING); 7402 return 0; 7403 } 7404 7405 schedule(); 7406 set_current_state(TASK_INTERRUPTIBLE); 7407 } 7408 __set_current_state(TASK_RUNNING); 7409 7410 return -1; 7411 } 7412 7413 static void napi_threaded_poll_loop(struct napi_struct *napi) 7414 { 7415 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7416 struct softnet_data *sd; 7417 unsigned long last_qs = jiffies; 7418 7419 for (;;) { 7420 bool repoll = false; 7421 void *have; 7422 7423 local_bh_disable(); 7424 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7425 7426 sd = this_cpu_ptr(&softnet_data); 7427 sd->in_napi_threaded_poll = true; 7428 7429 have = netpoll_poll_lock(napi); 7430 __napi_poll(napi, &repoll); 7431 netpoll_poll_unlock(have); 7432 7433 sd->in_napi_threaded_poll = false; 7434 barrier(); 7435 7436 if (sd_has_rps_ipi_waiting(sd)) { 7437 local_irq_disable(); 7438 net_rps_action_and_irq_enable(sd); 7439 } 7440 skb_defer_free_flush(sd); 7441 bpf_net_ctx_clear(bpf_net_ctx); 7442 local_bh_enable(); 7443 7444 if (!repoll) 7445 break; 7446 7447 rcu_softirq_qs_periodic(last_qs); 7448 cond_resched(); 7449 } 7450 } 7451 7452 static int napi_threaded_poll(void *data) 7453 { 7454 struct napi_struct *napi = data; 7455 7456 while (!napi_thread_wait(napi)) 7457 napi_threaded_poll_loop(napi); 7458 7459 return 0; 7460 } 7461 7462 static __latent_entropy void net_rx_action(void) 7463 { 7464 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 7465 unsigned long time_limit = jiffies + 7466 usecs_to_jiffies(READ_ONCE(net_hotdata.netdev_budget_usecs)); 7467 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7468 int budget = READ_ONCE(net_hotdata.netdev_budget); 7469 LIST_HEAD(list); 7470 LIST_HEAD(repoll); 7471 7472 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7473 start: 7474 sd->in_net_rx_action = true; 7475 local_irq_disable(); 7476 list_splice_init(&sd->poll_list, &list); 7477 local_irq_enable(); 7478 7479 for (;;) { 7480 struct napi_struct *n; 7481 7482 skb_defer_free_flush(sd); 7483 7484 if (list_empty(&list)) { 7485 if (list_empty(&repoll)) { 7486 sd->in_net_rx_action = false; 7487 barrier(); 7488 /* We need to check if ____napi_schedule() 7489 * had refilled poll_list while 7490 * sd->in_net_rx_action was true. 7491 */ 7492 if (!list_empty(&sd->poll_list)) 7493 goto start; 7494 if (!sd_has_rps_ipi_waiting(sd)) 7495 goto end; 7496 } 7497 break; 7498 } 7499 7500 n = list_first_entry(&list, struct napi_struct, poll_list); 7501 budget -= napi_poll(n, &repoll); 7502 7503 /* If softirq window is exhausted then punt. 7504 * Allow this to run for 2 jiffies since which will allow 7505 * an average latency of 1.5/HZ. 7506 */ 7507 if (unlikely(budget <= 0 || 7508 time_after_eq(jiffies, time_limit))) { 7509 sd->time_squeeze++; 7510 break; 7511 } 7512 } 7513 7514 local_irq_disable(); 7515 7516 list_splice_tail_init(&sd->poll_list, &list); 7517 list_splice_tail(&repoll, &list); 7518 list_splice(&list, &sd->poll_list); 7519 if (!list_empty(&sd->poll_list)) 7520 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 7521 else 7522 sd->in_net_rx_action = false; 7523 7524 net_rps_action_and_irq_enable(sd); 7525 end: 7526 bpf_net_ctx_clear(bpf_net_ctx); 7527 } 7528 7529 struct netdev_adjacent { 7530 struct net_device *dev; 7531 netdevice_tracker dev_tracker; 7532 7533 /* upper master flag, there can only be one master device per list */ 7534 bool master; 7535 7536 /* lookup ignore flag */ 7537 bool ignore; 7538 7539 /* counter for the number of times this device was added to us */ 7540 u16 ref_nr; 7541 7542 /* private field for the users */ 7543 void *private; 7544 7545 struct list_head list; 7546 struct rcu_head rcu; 7547 }; 7548 7549 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev, 7550 struct list_head *adj_list) 7551 { 7552 struct netdev_adjacent *adj; 7553 7554 list_for_each_entry(adj, adj_list, list) { 7555 if (adj->dev == adj_dev) 7556 return adj; 7557 } 7558 return NULL; 7559 } 7560 7561 static int ____netdev_has_upper_dev(struct net_device *upper_dev, 7562 struct netdev_nested_priv *priv) 7563 { 7564 struct net_device *dev = (struct net_device *)priv->data; 7565 7566 return upper_dev == dev; 7567 } 7568 7569 /** 7570 * netdev_has_upper_dev - Check if device is linked to an upper device 7571 * @dev: device 7572 * @upper_dev: upper device to check 7573 * 7574 * Find out if a device is linked to specified upper device and return true 7575 * in case it is. Note that this checks only immediate upper device, 7576 * not through a complete stack of devices. The caller must hold the RTNL lock. 7577 */ 7578 bool netdev_has_upper_dev(struct net_device *dev, 7579 struct net_device *upper_dev) 7580 { 7581 struct netdev_nested_priv priv = { 7582 .data = (void *)upper_dev, 7583 }; 7584 7585 ASSERT_RTNL(); 7586 7587 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7588 &priv); 7589 } 7590 EXPORT_SYMBOL(netdev_has_upper_dev); 7591 7592 /** 7593 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device 7594 * @dev: device 7595 * @upper_dev: upper device to check 7596 * 7597 * Find out if a device is linked to specified upper device and return true 7598 * in case it is. Note that this checks the entire upper device chain. 7599 * The caller must hold rcu lock. 7600 */ 7601 7602 bool netdev_has_upper_dev_all_rcu(struct net_device *dev, 7603 struct net_device *upper_dev) 7604 { 7605 struct netdev_nested_priv priv = { 7606 .data = (void *)upper_dev, 7607 }; 7608 7609 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7610 &priv); 7611 } 7612 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu); 7613 7614 /** 7615 * netdev_has_any_upper_dev - Check if device is linked to some device 7616 * @dev: device 7617 * 7618 * Find out if a device is linked to an upper device and return true in case 7619 * it is. The caller must hold the RTNL lock. 7620 */ 7621 bool netdev_has_any_upper_dev(struct net_device *dev) 7622 { 7623 ASSERT_RTNL(); 7624 7625 return !list_empty(&dev->adj_list.upper); 7626 } 7627 EXPORT_SYMBOL(netdev_has_any_upper_dev); 7628 7629 /** 7630 * netdev_master_upper_dev_get - Get master upper device 7631 * @dev: device 7632 * 7633 * Find a master upper device and return pointer to it or NULL in case 7634 * it's not there. The caller must hold the RTNL lock. 7635 */ 7636 struct net_device *netdev_master_upper_dev_get(struct net_device *dev) 7637 { 7638 struct netdev_adjacent *upper; 7639 7640 ASSERT_RTNL(); 7641 7642 if (list_empty(&dev->adj_list.upper)) 7643 return NULL; 7644 7645 upper = list_first_entry(&dev->adj_list.upper, 7646 struct netdev_adjacent, list); 7647 if (likely(upper->master)) 7648 return upper->dev; 7649 return NULL; 7650 } 7651 EXPORT_SYMBOL(netdev_master_upper_dev_get); 7652 7653 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev) 7654 { 7655 struct netdev_adjacent *upper; 7656 7657 ASSERT_RTNL(); 7658 7659 if (list_empty(&dev->adj_list.upper)) 7660 return NULL; 7661 7662 upper = list_first_entry(&dev->adj_list.upper, 7663 struct netdev_adjacent, list); 7664 if (likely(upper->master) && !upper->ignore) 7665 return upper->dev; 7666 return NULL; 7667 } 7668 7669 /** 7670 * netdev_has_any_lower_dev - Check if device is linked to some device 7671 * @dev: device 7672 * 7673 * Find out if a device is linked to a lower device and return true in case 7674 * it is. The caller must hold the RTNL lock. 7675 */ 7676 static bool netdev_has_any_lower_dev(struct net_device *dev) 7677 { 7678 ASSERT_RTNL(); 7679 7680 return !list_empty(&dev->adj_list.lower); 7681 } 7682 7683 void *netdev_adjacent_get_private(struct list_head *adj_list) 7684 { 7685 struct netdev_adjacent *adj; 7686 7687 adj = list_entry(adj_list, struct netdev_adjacent, list); 7688 7689 return adj->private; 7690 } 7691 EXPORT_SYMBOL(netdev_adjacent_get_private); 7692 7693 /** 7694 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list 7695 * @dev: device 7696 * @iter: list_head ** of the current position 7697 * 7698 * Gets the next device from the dev's upper list, starting from iter 7699 * position. The caller must hold RCU read lock. 7700 */ 7701 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev, 7702 struct list_head **iter) 7703 { 7704 struct netdev_adjacent *upper; 7705 7706 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7707 7708 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7709 7710 if (&upper->list == &dev->adj_list.upper) 7711 return NULL; 7712 7713 *iter = &upper->list; 7714 7715 return upper->dev; 7716 } 7717 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu); 7718 7719 static struct net_device *__netdev_next_upper_dev(struct net_device *dev, 7720 struct list_head **iter, 7721 bool *ignore) 7722 { 7723 struct netdev_adjacent *upper; 7724 7725 upper = list_entry((*iter)->next, struct netdev_adjacent, list); 7726 7727 if (&upper->list == &dev->adj_list.upper) 7728 return NULL; 7729 7730 *iter = &upper->list; 7731 *ignore = upper->ignore; 7732 7733 return upper->dev; 7734 } 7735 7736 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev, 7737 struct list_head **iter) 7738 { 7739 struct netdev_adjacent *upper; 7740 7741 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7742 7743 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7744 7745 if (&upper->list == &dev->adj_list.upper) 7746 return NULL; 7747 7748 *iter = &upper->list; 7749 7750 return upper->dev; 7751 } 7752 7753 static int __netdev_walk_all_upper_dev(struct net_device *dev, 7754 int (*fn)(struct net_device *dev, 7755 struct netdev_nested_priv *priv), 7756 struct netdev_nested_priv *priv) 7757 { 7758 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7759 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7760 int ret, cur = 0; 7761 bool ignore; 7762 7763 now = dev; 7764 iter = &dev->adj_list.upper; 7765 7766 while (1) { 7767 if (now != dev) { 7768 ret = fn(now, priv); 7769 if (ret) 7770 return ret; 7771 } 7772 7773 next = NULL; 7774 while (1) { 7775 udev = __netdev_next_upper_dev(now, &iter, &ignore); 7776 if (!udev) 7777 break; 7778 if (ignore) 7779 continue; 7780 7781 next = udev; 7782 niter = &udev->adj_list.upper; 7783 dev_stack[cur] = now; 7784 iter_stack[cur++] = iter; 7785 break; 7786 } 7787 7788 if (!next) { 7789 if (!cur) 7790 return 0; 7791 next = dev_stack[--cur]; 7792 niter = iter_stack[cur]; 7793 } 7794 7795 now = next; 7796 iter = niter; 7797 } 7798 7799 return 0; 7800 } 7801 7802 int netdev_walk_all_upper_dev_rcu(struct net_device *dev, 7803 int (*fn)(struct net_device *dev, 7804 struct netdev_nested_priv *priv), 7805 struct netdev_nested_priv *priv) 7806 { 7807 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7808 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7809 int ret, cur = 0; 7810 7811 now = dev; 7812 iter = &dev->adj_list.upper; 7813 7814 while (1) { 7815 if (now != dev) { 7816 ret = fn(now, priv); 7817 if (ret) 7818 return ret; 7819 } 7820 7821 next = NULL; 7822 while (1) { 7823 udev = netdev_next_upper_dev_rcu(now, &iter); 7824 if (!udev) 7825 break; 7826 7827 next = udev; 7828 niter = &udev->adj_list.upper; 7829 dev_stack[cur] = now; 7830 iter_stack[cur++] = iter; 7831 break; 7832 } 7833 7834 if (!next) { 7835 if (!cur) 7836 return 0; 7837 next = dev_stack[--cur]; 7838 niter = iter_stack[cur]; 7839 } 7840 7841 now = next; 7842 iter = niter; 7843 } 7844 7845 return 0; 7846 } 7847 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu); 7848 7849 static bool __netdev_has_upper_dev(struct net_device *dev, 7850 struct net_device *upper_dev) 7851 { 7852 struct netdev_nested_priv priv = { 7853 .flags = 0, 7854 .data = (void *)upper_dev, 7855 }; 7856 7857 ASSERT_RTNL(); 7858 7859 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev, 7860 &priv); 7861 } 7862 7863 /** 7864 * netdev_lower_get_next_private - Get the next ->private from the 7865 * lower neighbour list 7866 * @dev: device 7867 * @iter: list_head ** of the current position 7868 * 7869 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7870 * list, starting from iter position. The caller must hold either hold the 7871 * RTNL lock or its own locking that guarantees that the neighbour lower 7872 * list will remain unchanged. 7873 */ 7874 void *netdev_lower_get_next_private(struct net_device *dev, 7875 struct list_head **iter) 7876 { 7877 struct netdev_adjacent *lower; 7878 7879 lower = list_entry(*iter, struct netdev_adjacent, list); 7880 7881 if (&lower->list == &dev->adj_list.lower) 7882 return NULL; 7883 7884 *iter = lower->list.next; 7885 7886 return lower->private; 7887 } 7888 EXPORT_SYMBOL(netdev_lower_get_next_private); 7889 7890 /** 7891 * netdev_lower_get_next_private_rcu - Get the next ->private from the 7892 * lower neighbour list, RCU 7893 * variant 7894 * @dev: device 7895 * @iter: list_head ** of the current position 7896 * 7897 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7898 * list, starting from iter position. The caller must hold RCU read lock. 7899 */ 7900 void *netdev_lower_get_next_private_rcu(struct net_device *dev, 7901 struct list_head **iter) 7902 { 7903 struct netdev_adjacent *lower; 7904 7905 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 7906 7907 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7908 7909 if (&lower->list == &dev->adj_list.lower) 7910 return NULL; 7911 7912 *iter = &lower->list; 7913 7914 return lower->private; 7915 } 7916 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu); 7917 7918 /** 7919 * netdev_lower_get_next - Get the next device from the lower neighbour 7920 * list 7921 * @dev: device 7922 * @iter: list_head ** of the current position 7923 * 7924 * Gets the next netdev_adjacent from the dev's lower neighbour 7925 * list, starting from iter position. The caller must hold RTNL lock or 7926 * its own locking that guarantees that the neighbour lower 7927 * list will remain unchanged. 7928 */ 7929 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter) 7930 { 7931 struct netdev_adjacent *lower; 7932 7933 lower = list_entry(*iter, struct netdev_adjacent, list); 7934 7935 if (&lower->list == &dev->adj_list.lower) 7936 return NULL; 7937 7938 *iter = lower->list.next; 7939 7940 return lower->dev; 7941 } 7942 EXPORT_SYMBOL(netdev_lower_get_next); 7943 7944 static struct net_device *netdev_next_lower_dev(struct net_device *dev, 7945 struct list_head **iter) 7946 { 7947 struct netdev_adjacent *lower; 7948 7949 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7950 7951 if (&lower->list == &dev->adj_list.lower) 7952 return NULL; 7953 7954 *iter = &lower->list; 7955 7956 return lower->dev; 7957 } 7958 7959 static struct net_device *__netdev_next_lower_dev(struct net_device *dev, 7960 struct list_head **iter, 7961 bool *ignore) 7962 { 7963 struct netdev_adjacent *lower; 7964 7965 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7966 7967 if (&lower->list == &dev->adj_list.lower) 7968 return NULL; 7969 7970 *iter = &lower->list; 7971 *ignore = lower->ignore; 7972 7973 return lower->dev; 7974 } 7975 7976 int netdev_walk_all_lower_dev(struct net_device *dev, 7977 int (*fn)(struct net_device *dev, 7978 struct netdev_nested_priv *priv), 7979 struct netdev_nested_priv *priv) 7980 { 7981 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7982 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7983 int ret, cur = 0; 7984 7985 now = dev; 7986 iter = &dev->adj_list.lower; 7987 7988 while (1) { 7989 if (now != dev) { 7990 ret = fn(now, priv); 7991 if (ret) 7992 return ret; 7993 } 7994 7995 next = NULL; 7996 while (1) { 7997 ldev = netdev_next_lower_dev(now, &iter); 7998 if (!ldev) 7999 break; 8000 8001 next = ldev; 8002 niter = &ldev->adj_list.lower; 8003 dev_stack[cur] = now; 8004 iter_stack[cur++] = iter; 8005 break; 8006 } 8007 8008 if (!next) { 8009 if (!cur) 8010 return 0; 8011 next = dev_stack[--cur]; 8012 niter = iter_stack[cur]; 8013 } 8014 8015 now = next; 8016 iter = niter; 8017 } 8018 8019 return 0; 8020 } 8021 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev); 8022 8023 static int __netdev_walk_all_lower_dev(struct net_device *dev, 8024 int (*fn)(struct net_device *dev, 8025 struct netdev_nested_priv *priv), 8026 struct netdev_nested_priv *priv) 8027 { 8028 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8029 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8030 int ret, cur = 0; 8031 bool ignore; 8032 8033 now = dev; 8034 iter = &dev->adj_list.lower; 8035 8036 while (1) { 8037 if (now != dev) { 8038 ret = fn(now, priv); 8039 if (ret) 8040 return ret; 8041 } 8042 8043 next = NULL; 8044 while (1) { 8045 ldev = __netdev_next_lower_dev(now, &iter, &ignore); 8046 if (!ldev) 8047 break; 8048 if (ignore) 8049 continue; 8050 8051 next = ldev; 8052 niter = &ldev->adj_list.lower; 8053 dev_stack[cur] = now; 8054 iter_stack[cur++] = iter; 8055 break; 8056 } 8057 8058 if (!next) { 8059 if (!cur) 8060 return 0; 8061 next = dev_stack[--cur]; 8062 niter = iter_stack[cur]; 8063 } 8064 8065 now = next; 8066 iter = niter; 8067 } 8068 8069 return 0; 8070 } 8071 8072 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev, 8073 struct list_head **iter) 8074 { 8075 struct netdev_adjacent *lower; 8076 8077 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8078 if (&lower->list == &dev->adj_list.lower) 8079 return NULL; 8080 8081 *iter = &lower->list; 8082 8083 return lower->dev; 8084 } 8085 EXPORT_SYMBOL(netdev_next_lower_dev_rcu); 8086 8087 static u8 __netdev_upper_depth(struct net_device *dev) 8088 { 8089 struct net_device *udev; 8090 struct list_head *iter; 8091 u8 max_depth = 0; 8092 bool ignore; 8093 8094 for (iter = &dev->adj_list.upper, 8095 udev = __netdev_next_upper_dev(dev, &iter, &ignore); 8096 udev; 8097 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) { 8098 if (ignore) 8099 continue; 8100 if (max_depth < udev->upper_level) 8101 max_depth = udev->upper_level; 8102 } 8103 8104 return max_depth; 8105 } 8106 8107 static u8 __netdev_lower_depth(struct net_device *dev) 8108 { 8109 struct net_device *ldev; 8110 struct list_head *iter; 8111 u8 max_depth = 0; 8112 bool ignore; 8113 8114 for (iter = &dev->adj_list.lower, 8115 ldev = __netdev_next_lower_dev(dev, &iter, &ignore); 8116 ldev; 8117 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) { 8118 if (ignore) 8119 continue; 8120 if (max_depth < ldev->lower_level) 8121 max_depth = ldev->lower_level; 8122 } 8123 8124 return max_depth; 8125 } 8126 8127 static int __netdev_update_upper_level(struct net_device *dev, 8128 struct netdev_nested_priv *__unused) 8129 { 8130 dev->upper_level = __netdev_upper_depth(dev) + 1; 8131 return 0; 8132 } 8133 8134 #ifdef CONFIG_LOCKDEP 8135 static LIST_HEAD(net_unlink_list); 8136 8137 static void net_unlink_todo(struct net_device *dev) 8138 { 8139 if (list_empty(&dev->unlink_list)) 8140 list_add_tail(&dev->unlink_list, &net_unlink_list); 8141 } 8142 #endif 8143 8144 static int __netdev_update_lower_level(struct net_device *dev, 8145 struct netdev_nested_priv *priv) 8146 { 8147 dev->lower_level = __netdev_lower_depth(dev) + 1; 8148 8149 #ifdef CONFIG_LOCKDEP 8150 if (!priv) 8151 return 0; 8152 8153 if (priv->flags & NESTED_SYNC_IMM) 8154 dev->nested_level = dev->lower_level - 1; 8155 if (priv->flags & NESTED_SYNC_TODO) 8156 net_unlink_todo(dev); 8157 #endif 8158 return 0; 8159 } 8160 8161 int netdev_walk_all_lower_dev_rcu(struct net_device *dev, 8162 int (*fn)(struct net_device *dev, 8163 struct netdev_nested_priv *priv), 8164 struct netdev_nested_priv *priv) 8165 { 8166 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8167 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8168 int ret, cur = 0; 8169 8170 now = dev; 8171 iter = &dev->adj_list.lower; 8172 8173 while (1) { 8174 if (now != dev) { 8175 ret = fn(now, priv); 8176 if (ret) 8177 return ret; 8178 } 8179 8180 next = NULL; 8181 while (1) { 8182 ldev = netdev_next_lower_dev_rcu(now, &iter); 8183 if (!ldev) 8184 break; 8185 8186 next = ldev; 8187 niter = &ldev->adj_list.lower; 8188 dev_stack[cur] = now; 8189 iter_stack[cur++] = iter; 8190 break; 8191 } 8192 8193 if (!next) { 8194 if (!cur) 8195 return 0; 8196 next = dev_stack[--cur]; 8197 niter = iter_stack[cur]; 8198 } 8199 8200 now = next; 8201 iter = niter; 8202 } 8203 8204 return 0; 8205 } 8206 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu); 8207 8208 /** 8209 * netdev_lower_get_first_private_rcu - Get the first ->private from the 8210 * lower neighbour list, RCU 8211 * variant 8212 * @dev: device 8213 * 8214 * Gets the first netdev_adjacent->private from the dev's lower neighbour 8215 * list. The caller must hold RCU read lock. 8216 */ 8217 void *netdev_lower_get_first_private_rcu(struct net_device *dev) 8218 { 8219 struct netdev_adjacent *lower; 8220 8221 lower = list_first_or_null_rcu(&dev->adj_list.lower, 8222 struct netdev_adjacent, list); 8223 if (lower) 8224 return lower->private; 8225 return NULL; 8226 } 8227 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu); 8228 8229 /** 8230 * netdev_master_upper_dev_get_rcu - Get master upper device 8231 * @dev: device 8232 * 8233 * Find a master upper device and return pointer to it or NULL in case 8234 * it's not there. The caller must hold the RCU read lock. 8235 */ 8236 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev) 8237 { 8238 struct netdev_adjacent *upper; 8239 8240 upper = list_first_or_null_rcu(&dev->adj_list.upper, 8241 struct netdev_adjacent, list); 8242 if (upper && likely(upper->master)) 8243 return upper->dev; 8244 return NULL; 8245 } 8246 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu); 8247 8248 static int netdev_adjacent_sysfs_add(struct net_device *dev, 8249 struct net_device *adj_dev, 8250 struct list_head *dev_list) 8251 { 8252 char linkname[IFNAMSIZ+7]; 8253 8254 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8255 "upper_%s" : "lower_%s", adj_dev->name); 8256 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj), 8257 linkname); 8258 } 8259 static void netdev_adjacent_sysfs_del(struct net_device *dev, 8260 char *name, 8261 struct list_head *dev_list) 8262 { 8263 char linkname[IFNAMSIZ+7]; 8264 8265 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8266 "upper_%s" : "lower_%s", name); 8267 sysfs_remove_link(&(dev->dev.kobj), linkname); 8268 } 8269 8270 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev, 8271 struct net_device *adj_dev, 8272 struct list_head *dev_list) 8273 { 8274 return (dev_list == &dev->adj_list.upper || 8275 dev_list == &dev->adj_list.lower) && 8276 net_eq(dev_net(dev), dev_net(adj_dev)); 8277 } 8278 8279 static int __netdev_adjacent_dev_insert(struct net_device *dev, 8280 struct net_device *adj_dev, 8281 struct list_head *dev_list, 8282 void *private, bool master) 8283 { 8284 struct netdev_adjacent *adj; 8285 int ret; 8286 8287 adj = __netdev_find_adj(adj_dev, dev_list); 8288 8289 if (adj) { 8290 adj->ref_nr += 1; 8291 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n", 8292 dev->name, adj_dev->name, adj->ref_nr); 8293 8294 return 0; 8295 } 8296 8297 adj = kmalloc(sizeof(*adj), GFP_KERNEL); 8298 if (!adj) 8299 return -ENOMEM; 8300 8301 adj->dev = adj_dev; 8302 adj->master = master; 8303 adj->ref_nr = 1; 8304 adj->private = private; 8305 adj->ignore = false; 8306 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL); 8307 8308 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n", 8309 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name); 8310 8311 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) { 8312 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list); 8313 if (ret) 8314 goto free_adj; 8315 } 8316 8317 /* Ensure that master link is always the first item in list. */ 8318 if (master) { 8319 ret = sysfs_create_link(&(dev->dev.kobj), 8320 &(adj_dev->dev.kobj), "master"); 8321 if (ret) 8322 goto remove_symlinks; 8323 8324 list_add_rcu(&adj->list, dev_list); 8325 } else { 8326 list_add_tail_rcu(&adj->list, dev_list); 8327 } 8328 8329 return 0; 8330 8331 remove_symlinks: 8332 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8333 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8334 free_adj: 8335 netdev_put(adj_dev, &adj->dev_tracker); 8336 kfree(adj); 8337 8338 return ret; 8339 } 8340 8341 static void __netdev_adjacent_dev_remove(struct net_device *dev, 8342 struct net_device *adj_dev, 8343 u16 ref_nr, 8344 struct list_head *dev_list) 8345 { 8346 struct netdev_adjacent *adj; 8347 8348 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n", 8349 dev->name, adj_dev->name, ref_nr); 8350 8351 adj = __netdev_find_adj(adj_dev, dev_list); 8352 8353 if (!adj) { 8354 pr_err("Adjacency does not exist for device %s from %s\n", 8355 dev->name, adj_dev->name); 8356 WARN_ON(1); 8357 return; 8358 } 8359 8360 if (adj->ref_nr > ref_nr) { 8361 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n", 8362 dev->name, adj_dev->name, ref_nr, 8363 adj->ref_nr - ref_nr); 8364 adj->ref_nr -= ref_nr; 8365 return; 8366 } 8367 8368 if (adj->master) 8369 sysfs_remove_link(&(dev->dev.kobj), "master"); 8370 8371 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8372 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8373 8374 list_del_rcu(&adj->list); 8375 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n", 8376 adj_dev->name, dev->name, adj_dev->name); 8377 netdev_put(adj_dev, &adj->dev_tracker); 8378 kfree_rcu(adj, rcu); 8379 } 8380 8381 static int __netdev_adjacent_dev_link_lists(struct net_device *dev, 8382 struct net_device *upper_dev, 8383 struct list_head *up_list, 8384 struct list_head *down_list, 8385 void *private, bool master) 8386 { 8387 int ret; 8388 8389 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list, 8390 private, master); 8391 if (ret) 8392 return ret; 8393 8394 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list, 8395 private, false); 8396 if (ret) { 8397 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list); 8398 return ret; 8399 } 8400 8401 return 0; 8402 } 8403 8404 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev, 8405 struct net_device *upper_dev, 8406 u16 ref_nr, 8407 struct list_head *up_list, 8408 struct list_head *down_list) 8409 { 8410 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list); 8411 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list); 8412 } 8413 8414 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev, 8415 struct net_device *upper_dev, 8416 void *private, bool master) 8417 { 8418 return __netdev_adjacent_dev_link_lists(dev, upper_dev, 8419 &dev->adj_list.upper, 8420 &upper_dev->adj_list.lower, 8421 private, master); 8422 } 8423 8424 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev, 8425 struct net_device *upper_dev) 8426 { 8427 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1, 8428 &dev->adj_list.upper, 8429 &upper_dev->adj_list.lower); 8430 } 8431 8432 static int __netdev_upper_dev_link(struct net_device *dev, 8433 struct net_device *upper_dev, bool master, 8434 void *upper_priv, void *upper_info, 8435 struct netdev_nested_priv *priv, 8436 struct netlink_ext_ack *extack) 8437 { 8438 struct netdev_notifier_changeupper_info changeupper_info = { 8439 .info = { 8440 .dev = dev, 8441 .extack = extack, 8442 }, 8443 .upper_dev = upper_dev, 8444 .master = master, 8445 .linking = true, 8446 .upper_info = upper_info, 8447 }; 8448 struct net_device *master_dev; 8449 int ret = 0; 8450 8451 ASSERT_RTNL(); 8452 8453 if (dev == upper_dev) 8454 return -EBUSY; 8455 8456 /* To prevent loops, check if dev is not upper device to upper_dev. */ 8457 if (__netdev_has_upper_dev(upper_dev, dev)) 8458 return -EBUSY; 8459 8460 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV) 8461 return -EMLINK; 8462 8463 if (!master) { 8464 if (__netdev_has_upper_dev(dev, upper_dev)) 8465 return -EEXIST; 8466 } else { 8467 master_dev = __netdev_master_upper_dev_get(dev); 8468 if (master_dev) 8469 return master_dev == upper_dev ? -EEXIST : -EBUSY; 8470 } 8471 8472 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8473 &changeupper_info.info); 8474 ret = notifier_to_errno(ret); 8475 if (ret) 8476 return ret; 8477 8478 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv, 8479 master); 8480 if (ret) 8481 return ret; 8482 8483 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8484 &changeupper_info.info); 8485 ret = notifier_to_errno(ret); 8486 if (ret) 8487 goto rollback; 8488 8489 __netdev_update_upper_level(dev, NULL); 8490 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8491 8492 __netdev_update_lower_level(upper_dev, priv); 8493 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8494 priv); 8495 8496 return 0; 8497 8498 rollback: 8499 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8500 8501 return ret; 8502 } 8503 8504 /** 8505 * netdev_upper_dev_link - Add a link to the upper device 8506 * @dev: device 8507 * @upper_dev: new upper device 8508 * @extack: netlink extended ack 8509 * 8510 * Adds a link to device which is upper to this one. The caller must hold 8511 * the RTNL lock. On a failure a negative errno code is returned. 8512 * On success the reference counts are adjusted and the function 8513 * returns zero. 8514 */ 8515 int netdev_upper_dev_link(struct net_device *dev, 8516 struct net_device *upper_dev, 8517 struct netlink_ext_ack *extack) 8518 { 8519 struct netdev_nested_priv priv = { 8520 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8521 .data = NULL, 8522 }; 8523 8524 return __netdev_upper_dev_link(dev, upper_dev, false, 8525 NULL, NULL, &priv, extack); 8526 } 8527 EXPORT_SYMBOL(netdev_upper_dev_link); 8528 8529 /** 8530 * netdev_master_upper_dev_link - Add a master link to the upper device 8531 * @dev: device 8532 * @upper_dev: new upper device 8533 * @upper_priv: upper device private 8534 * @upper_info: upper info to be passed down via notifier 8535 * @extack: netlink extended ack 8536 * 8537 * Adds a link to device which is upper to this one. In this case, only 8538 * one master upper device can be linked, although other non-master devices 8539 * might be linked as well. The caller must hold the RTNL lock. 8540 * On a failure a negative errno code is returned. On success the reference 8541 * counts are adjusted and the function returns zero. 8542 */ 8543 int netdev_master_upper_dev_link(struct net_device *dev, 8544 struct net_device *upper_dev, 8545 void *upper_priv, void *upper_info, 8546 struct netlink_ext_ack *extack) 8547 { 8548 struct netdev_nested_priv priv = { 8549 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8550 .data = NULL, 8551 }; 8552 8553 return __netdev_upper_dev_link(dev, upper_dev, true, 8554 upper_priv, upper_info, &priv, extack); 8555 } 8556 EXPORT_SYMBOL(netdev_master_upper_dev_link); 8557 8558 static void __netdev_upper_dev_unlink(struct net_device *dev, 8559 struct net_device *upper_dev, 8560 struct netdev_nested_priv *priv) 8561 { 8562 struct netdev_notifier_changeupper_info changeupper_info = { 8563 .info = { 8564 .dev = dev, 8565 }, 8566 .upper_dev = upper_dev, 8567 .linking = false, 8568 }; 8569 8570 ASSERT_RTNL(); 8571 8572 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev; 8573 8574 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8575 &changeupper_info.info); 8576 8577 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8578 8579 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8580 &changeupper_info.info); 8581 8582 __netdev_update_upper_level(dev, NULL); 8583 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8584 8585 __netdev_update_lower_level(upper_dev, priv); 8586 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8587 priv); 8588 } 8589 8590 /** 8591 * netdev_upper_dev_unlink - Removes a link to upper device 8592 * @dev: device 8593 * @upper_dev: new upper device 8594 * 8595 * Removes a link to device which is upper to this one. The caller must hold 8596 * the RTNL lock. 8597 */ 8598 void netdev_upper_dev_unlink(struct net_device *dev, 8599 struct net_device *upper_dev) 8600 { 8601 struct netdev_nested_priv priv = { 8602 .flags = NESTED_SYNC_TODO, 8603 .data = NULL, 8604 }; 8605 8606 __netdev_upper_dev_unlink(dev, upper_dev, &priv); 8607 } 8608 EXPORT_SYMBOL(netdev_upper_dev_unlink); 8609 8610 static void __netdev_adjacent_dev_set(struct net_device *upper_dev, 8611 struct net_device *lower_dev, 8612 bool val) 8613 { 8614 struct netdev_adjacent *adj; 8615 8616 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower); 8617 if (adj) 8618 adj->ignore = val; 8619 8620 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper); 8621 if (adj) 8622 adj->ignore = val; 8623 } 8624 8625 static void netdev_adjacent_dev_disable(struct net_device *upper_dev, 8626 struct net_device *lower_dev) 8627 { 8628 __netdev_adjacent_dev_set(upper_dev, lower_dev, true); 8629 } 8630 8631 static void netdev_adjacent_dev_enable(struct net_device *upper_dev, 8632 struct net_device *lower_dev) 8633 { 8634 __netdev_adjacent_dev_set(upper_dev, lower_dev, false); 8635 } 8636 8637 int netdev_adjacent_change_prepare(struct net_device *old_dev, 8638 struct net_device *new_dev, 8639 struct net_device *dev, 8640 struct netlink_ext_ack *extack) 8641 { 8642 struct netdev_nested_priv priv = { 8643 .flags = 0, 8644 .data = NULL, 8645 }; 8646 int err; 8647 8648 if (!new_dev) 8649 return 0; 8650 8651 if (old_dev && new_dev != old_dev) 8652 netdev_adjacent_dev_disable(dev, old_dev); 8653 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv, 8654 extack); 8655 if (err) { 8656 if (old_dev && new_dev != old_dev) 8657 netdev_adjacent_dev_enable(dev, old_dev); 8658 return err; 8659 } 8660 8661 return 0; 8662 } 8663 EXPORT_SYMBOL(netdev_adjacent_change_prepare); 8664 8665 void netdev_adjacent_change_commit(struct net_device *old_dev, 8666 struct net_device *new_dev, 8667 struct net_device *dev) 8668 { 8669 struct netdev_nested_priv priv = { 8670 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8671 .data = NULL, 8672 }; 8673 8674 if (!new_dev || !old_dev) 8675 return; 8676 8677 if (new_dev == old_dev) 8678 return; 8679 8680 netdev_adjacent_dev_enable(dev, old_dev); 8681 __netdev_upper_dev_unlink(old_dev, dev, &priv); 8682 } 8683 EXPORT_SYMBOL(netdev_adjacent_change_commit); 8684 8685 void netdev_adjacent_change_abort(struct net_device *old_dev, 8686 struct net_device *new_dev, 8687 struct net_device *dev) 8688 { 8689 struct netdev_nested_priv priv = { 8690 .flags = 0, 8691 .data = NULL, 8692 }; 8693 8694 if (!new_dev) 8695 return; 8696 8697 if (old_dev && new_dev != old_dev) 8698 netdev_adjacent_dev_enable(dev, old_dev); 8699 8700 __netdev_upper_dev_unlink(new_dev, dev, &priv); 8701 } 8702 EXPORT_SYMBOL(netdev_adjacent_change_abort); 8703 8704 /** 8705 * netdev_bonding_info_change - Dispatch event about slave change 8706 * @dev: device 8707 * @bonding_info: info to dispatch 8708 * 8709 * Send NETDEV_BONDING_INFO to netdev notifiers with info. 8710 * The caller must hold the RTNL lock. 8711 */ 8712 void netdev_bonding_info_change(struct net_device *dev, 8713 struct netdev_bonding_info *bonding_info) 8714 { 8715 struct netdev_notifier_bonding_info info = { 8716 .info.dev = dev, 8717 }; 8718 8719 memcpy(&info.bonding_info, bonding_info, 8720 sizeof(struct netdev_bonding_info)); 8721 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, 8722 &info.info); 8723 } 8724 EXPORT_SYMBOL(netdev_bonding_info_change); 8725 8726 static int netdev_offload_xstats_enable_l3(struct net_device *dev, 8727 struct netlink_ext_ack *extack) 8728 { 8729 struct netdev_notifier_offload_xstats_info info = { 8730 .info.dev = dev, 8731 .info.extack = extack, 8732 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8733 }; 8734 int err; 8735 int rc; 8736 8737 dev->offload_xstats_l3 = kzalloc(sizeof(*dev->offload_xstats_l3), 8738 GFP_KERNEL); 8739 if (!dev->offload_xstats_l3) 8740 return -ENOMEM; 8741 8742 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE, 8743 NETDEV_OFFLOAD_XSTATS_DISABLE, 8744 &info.info); 8745 err = notifier_to_errno(rc); 8746 if (err) 8747 goto free_stats; 8748 8749 return 0; 8750 8751 free_stats: 8752 kfree(dev->offload_xstats_l3); 8753 dev->offload_xstats_l3 = NULL; 8754 return err; 8755 } 8756 8757 int netdev_offload_xstats_enable(struct net_device *dev, 8758 enum netdev_offload_xstats_type type, 8759 struct netlink_ext_ack *extack) 8760 { 8761 ASSERT_RTNL(); 8762 8763 if (netdev_offload_xstats_enabled(dev, type)) 8764 return -EALREADY; 8765 8766 switch (type) { 8767 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8768 return netdev_offload_xstats_enable_l3(dev, extack); 8769 } 8770 8771 WARN_ON(1); 8772 return -EINVAL; 8773 } 8774 EXPORT_SYMBOL(netdev_offload_xstats_enable); 8775 8776 static void netdev_offload_xstats_disable_l3(struct net_device *dev) 8777 { 8778 struct netdev_notifier_offload_xstats_info info = { 8779 .info.dev = dev, 8780 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8781 }; 8782 8783 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE, 8784 &info.info); 8785 kfree(dev->offload_xstats_l3); 8786 dev->offload_xstats_l3 = NULL; 8787 } 8788 8789 int netdev_offload_xstats_disable(struct net_device *dev, 8790 enum netdev_offload_xstats_type type) 8791 { 8792 ASSERT_RTNL(); 8793 8794 if (!netdev_offload_xstats_enabled(dev, type)) 8795 return -EALREADY; 8796 8797 switch (type) { 8798 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8799 netdev_offload_xstats_disable_l3(dev); 8800 return 0; 8801 } 8802 8803 WARN_ON(1); 8804 return -EINVAL; 8805 } 8806 EXPORT_SYMBOL(netdev_offload_xstats_disable); 8807 8808 static void netdev_offload_xstats_disable_all(struct net_device *dev) 8809 { 8810 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3); 8811 } 8812 8813 static struct rtnl_hw_stats64 * 8814 netdev_offload_xstats_get_ptr(const struct net_device *dev, 8815 enum netdev_offload_xstats_type type) 8816 { 8817 switch (type) { 8818 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8819 return dev->offload_xstats_l3; 8820 } 8821 8822 WARN_ON(1); 8823 return NULL; 8824 } 8825 8826 bool netdev_offload_xstats_enabled(const struct net_device *dev, 8827 enum netdev_offload_xstats_type type) 8828 { 8829 ASSERT_RTNL(); 8830 8831 return netdev_offload_xstats_get_ptr(dev, type); 8832 } 8833 EXPORT_SYMBOL(netdev_offload_xstats_enabled); 8834 8835 struct netdev_notifier_offload_xstats_ru { 8836 bool used; 8837 }; 8838 8839 struct netdev_notifier_offload_xstats_rd { 8840 struct rtnl_hw_stats64 stats; 8841 bool used; 8842 }; 8843 8844 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest, 8845 const struct rtnl_hw_stats64 *src) 8846 { 8847 dest->rx_packets += src->rx_packets; 8848 dest->tx_packets += src->tx_packets; 8849 dest->rx_bytes += src->rx_bytes; 8850 dest->tx_bytes += src->tx_bytes; 8851 dest->rx_errors += src->rx_errors; 8852 dest->tx_errors += src->tx_errors; 8853 dest->rx_dropped += src->rx_dropped; 8854 dest->tx_dropped += src->tx_dropped; 8855 dest->multicast += src->multicast; 8856 } 8857 8858 static int netdev_offload_xstats_get_used(struct net_device *dev, 8859 enum netdev_offload_xstats_type type, 8860 bool *p_used, 8861 struct netlink_ext_ack *extack) 8862 { 8863 struct netdev_notifier_offload_xstats_ru report_used = {}; 8864 struct netdev_notifier_offload_xstats_info info = { 8865 .info.dev = dev, 8866 .info.extack = extack, 8867 .type = type, 8868 .report_used = &report_used, 8869 }; 8870 int rc; 8871 8872 WARN_ON(!netdev_offload_xstats_enabled(dev, type)); 8873 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED, 8874 &info.info); 8875 *p_used = report_used.used; 8876 return notifier_to_errno(rc); 8877 } 8878 8879 static int netdev_offload_xstats_get_stats(struct net_device *dev, 8880 enum netdev_offload_xstats_type type, 8881 struct rtnl_hw_stats64 *p_stats, 8882 bool *p_used, 8883 struct netlink_ext_ack *extack) 8884 { 8885 struct netdev_notifier_offload_xstats_rd report_delta = {}; 8886 struct netdev_notifier_offload_xstats_info info = { 8887 .info.dev = dev, 8888 .info.extack = extack, 8889 .type = type, 8890 .report_delta = &report_delta, 8891 }; 8892 struct rtnl_hw_stats64 *stats; 8893 int rc; 8894 8895 stats = netdev_offload_xstats_get_ptr(dev, type); 8896 if (WARN_ON(!stats)) 8897 return -EINVAL; 8898 8899 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA, 8900 &info.info); 8901 8902 /* Cache whatever we got, even if there was an error, otherwise the 8903 * successful stats retrievals would get lost. 8904 */ 8905 netdev_hw_stats64_add(stats, &report_delta.stats); 8906 8907 if (p_stats) 8908 *p_stats = *stats; 8909 *p_used = report_delta.used; 8910 8911 return notifier_to_errno(rc); 8912 } 8913 8914 int netdev_offload_xstats_get(struct net_device *dev, 8915 enum netdev_offload_xstats_type type, 8916 struct rtnl_hw_stats64 *p_stats, bool *p_used, 8917 struct netlink_ext_ack *extack) 8918 { 8919 ASSERT_RTNL(); 8920 8921 if (p_stats) 8922 return netdev_offload_xstats_get_stats(dev, type, p_stats, 8923 p_used, extack); 8924 else 8925 return netdev_offload_xstats_get_used(dev, type, p_used, 8926 extack); 8927 } 8928 EXPORT_SYMBOL(netdev_offload_xstats_get); 8929 8930 void 8931 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta, 8932 const struct rtnl_hw_stats64 *stats) 8933 { 8934 report_delta->used = true; 8935 netdev_hw_stats64_add(&report_delta->stats, stats); 8936 } 8937 EXPORT_SYMBOL(netdev_offload_xstats_report_delta); 8938 8939 void 8940 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used) 8941 { 8942 report_used->used = true; 8943 } 8944 EXPORT_SYMBOL(netdev_offload_xstats_report_used); 8945 8946 void netdev_offload_xstats_push_delta(struct net_device *dev, 8947 enum netdev_offload_xstats_type type, 8948 const struct rtnl_hw_stats64 *p_stats) 8949 { 8950 struct rtnl_hw_stats64 *stats; 8951 8952 ASSERT_RTNL(); 8953 8954 stats = netdev_offload_xstats_get_ptr(dev, type); 8955 if (WARN_ON(!stats)) 8956 return; 8957 8958 netdev_hw_stats64_add(stats, p_stats); 8959 } 8960 EXPORT_SYMBOL(netdev_offload_xstats_push_delta); 8961 8962 /** 8963 * netdev_get_xmit_slave - Get the xmit slave of master device 8964 * @dev: device 8965 * @skb: The packet 8966 * @all_slaves: assume all the slaves are active 8967 * 8968 * The reference counters are not incremented so the caller must be 8969 * careful with locks. The caller must hold RCU lock. 8970 * %NULL is returned if no slave is found. 8971 */ 8972 8973 struct net_device *netdev_get_xmit_slave(struct net_device *dev, 8974 struct sk_buff *skb, 8975 bool all_slaves) 8976 { 8977 const struct net_device_ops *ops = dev->netdev_ops; 8978 8979 if (!ops->ndo_get_xmit_slave) 8980 return NULL; 8981 return ops->ndo_get_xmit_slave(dev, skb, all_slaves); 8982 } 8983 EXPORT_SYMBOL(netdev_get_xmit_slave); 8984 8985 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev, 8986 struct sock *sk) 8987 { 8988 const struct net_device_ops *ops = dev->netdev_ops; 8989 8990 if (!ops->ndo_sk_get_lower_dev) 8991 return NULL; 8992 return ops->ndo_sk_get_lower_dev(dev, sk); 8993 } 8994 8995 /** 8996 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket 8997 * @dev: device 8998 * @sk: the socket 8999 * 9000 * %NULL is returned if no lower device is found. 9001 */ 9002 9003 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev, 9004 struct sock *sk) 9005 { 9006 struct net_device *lower; 9007 9008 lower = netdev_sk_get_lower_dev(dev, sk); 9009 while (lower) { 9010 dev = lower; 9011 lower = netdev_sk_get_lower_dev(dev, sk); 9012 } 9013 9014 return dev; 9015 } 9016 EXPORT_SYMBOL(netdev_sk_get_lowest_dev); 9017 9018 static void netdev_adjacent_add_links(struct net_device *dev) 9019 { 9020 struct netdev_adjacent *iter; 9021 9022 struct net *net = dev_net(dev); 9023 9024 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9025 if (!net_eq(net, dev_net(iter->dev))) 9026 continue; 9027 netdev_adjacent_sysfs_add(iter->dev, dev, 9028 &iter->dev->adj_list.lower); 9029 netdev_adjacent_sysfs_add(dev, iter->dev, 9030 &dev->adj_list.upper); 9031 } 9032 9033 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9034 if (!net_eq(net, dev_net(iter->dev))) 9035 continue; 9036 netdev_adjacent_sysfs_add(iter->dev, dev, 9037 &iter->dev->adj_list.upper); 9038 netdev_adjacent_sysfs_add(dev, iter->dev, 9039 &dev->adj_list.lower); 9040 } 9041 } 9042 9043 static void netdev_adjacent_del_links(struct net_device *dev) 9044 { 9045 struct netdev_adjacent *iter; 9046 9047 struct net *net = dev_net(dev); 9048 9049 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9050 if (!net_eq(net, dev_net(iter->dev))) 9051 continue; 9052 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9053 &iter->dev->adj_list.lower); 9054 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9055 &dev->adj_list.upper); 9056 } 9057 9058 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9059 if (!net_eq(net, dev_net(iter->dev))) 9060 continue; 9061 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9062 &iter->dev->adj_list.upper); 9063 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9064 &dev->adj_list.lower); 9065 } 9066 } 9067 9068 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname) 9069 { 9070 struct netdev_adjacent *iter; 9071 9072 struct net *net = dev_net(dev); 9073 9074 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9075 if (!net_eq(net, dev_net(iter->dev))) 9076 continue; 9077 netdev_adjacent_sysfs_del(iter->dev, oldname, 9078 &iter->dev->adj_list.lower); 9079 netdev_adjacent_sysfs_add(iter->dev, dev, 9080 &iter->dev->adj_list.lower); 9081 } 9082 9083 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9084 if (!net_eq(net, dev_net(iter->dev))) 9085 continue; 9086 netdev_adjacent_sysfs_del(iter->dev, oldname, 9087 &iter->dev->adj_list.upper); 9088 netdev_adjacent_sysfs_add(iter->dev, dev, 9089 &iter->dev->adj_list.upper); 9090 } 9091 } 9092 9093 void *netdev_lower_dev_get_private(struct net_device *dev, 9094 struct net_device *lower_dev) 9095 { 9096 struct netdev_adjacent *lower; 9097 9098 if (!lower_dev) 9099 return NULL; 9100 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower); 9101 if (!lower) 9102 return NULL; 9103 9104 return lower->private; 9105 } 9106 EXPORT_SYMBOL(netdev_lower_dev_get_private); 9107 9108 9109 /** 9110 * netdev_lower_state_changed - Dispatch event about lower device state change 9111 * @lower_dev: device 9112 * @lower_state_info: state to dispatch 9113 * 9114 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info. 9115 * The caller must hold the RTNL lock. 9116 */ 9117 void netdev_lower_state_changed(struct net_device *lower_dev, 9118 void *lower_state_info) 9119 { 9120 struct netdev_notifier_changelowerstate_info changelowerstate_info = { 9121 .info.dev = lower_dev, 9122 }; 9123 9124 ASSERT_RTNL(); 9125 changelowerstate_info.lower_state_info = lower_state_info; 9126 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, 9127 &changelowerstate_info.info); 9128 } 9129 EXPORT_SYMBOL(netdev_lower_state_changed); 9130 9131 static void dev_change_rx_flags(struct net_device *dev, int flags) 9132 { 9133 const struct net_device_ops *ops = dev->netdev_ops; 9134 9135 if (ops->ndo_change_rx_flags) 9136 ops->ndo_change_rx_flags(dev, flags); 9137 } 9138 9139 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify) 9140 { 9141 unsigned int old_flags = dev->flags; 9142 unsigned int promiscuity, flags; 9143 kuid_t uid; 9144 kgid_t gid; 9145 9146 ASSERT_RTNL(); 9147 9148 promiscuity = dev->promiscuity + inc; 9149 if (promiscuity == 0) { 9150 /* 9151 * Avoid overflow. 9152 * If inc causes overflow, untouch promisc and return error. 9153 */ 9154 if (unlikely(inc > 0)) { 9155 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n"); 9156 return -EOVERFLOW; 9157 } 9158 flags = old_flags & ~IFF_PROMISC; 9159 } else { 9160 flags = old_flags | IFF_PROMISC; 9161 } 9162 WRITE_ONCE(dev->promiscuity, promiscuity); 9163 if (flags != old_flags) { 9164 WRITE_ONCE(dev->flags, flags); 9165 netdev_info(dev, "%s promiscuous mode\n", 9166 dev->flags & IFF_PROMISC ? "entered" : "left"); 9167 if (audit_enabled) { 9168 current_uid_gid(&uid, &gid); 9169 audit_log(audit_context(), GFP_ATOMIC, 9170 AUDIT_ANOM_PROMISCUOUS, 9171 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u", 9172 dev->name, (dev->flags & IFF_PROMISC), 9173 (old_flags & IFF_PROMISC), 9174 from_kuid(&init_user_ns, audit_get_loginuid(current)), 9175 from_kuid(&init_user_ns, uid), 9176 from_kgid(&init_user_ns, gid), 9177 audit_get_sessionid(current)); 9178 } 9179 9180 dev_change_rx_flags(dev, IFF_PROMISC); 9181 } 9182 if (notify) 9183 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL); 9184 return 0; 9185 } 9186 9187 /** 9188 * dev_set_promiscuity - update promiscuity count on a device 9189 * @dev: device 9190 * @inc: modifier 9191 * 9192 * Add or remove promiscuity from a device. While the count in the device 9193 * remains above zero the interface remains promiscuous. Once it hits zero 9194 * the device reverts back to normal filtering operation. A negative inc 9195 * value is used to drop promiscuity on the device. 9196 * Return 0 if successful or a negative errno code on error. 9197 */ 9198 int dev_set_promiscuity(struct net_device *dev, int inc) 9199 { 9200 unsigned int old_flags = dev->flags; 9201 int err; 9202 9203 err = __dev_set_promiscuity(dev, inc, true); 9204 if (err < 0) 9205 return err; 9206 if (dev->flags != old_flags) 9207 dev_set_rx_mode(dev); 9208 return err; 9209 } 9210 EXPORT_SYMBOL(dev_set_promiscuity); 9211 9212 int netif_set_allmulti(struct net_device *dev, int inc, bool notify) 9213 { 9214 unsigned int old_flags = dev->flags, old_gflags = dev->gflags; 9215 unsigned int allmulti, flags; 9216 9217 ASSERT_RTNL(); 9218 9219 allmulti = dev->allmulti + inc; 9220 if (allmulti == 0) { 9221 /* 9222 * Avoid overflow. 9223 * If inc causes overflow, untouch allmulti and return error. 9224 */ 9225 if (unlikely(inc > 0)) { 9226 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n"); 9227 return -EOVERFLOW; 9228 } 9229 flags = old_flags & ~IFF_ALLMULTI; 9230 } else { 9231 flags = old_flags | IFF_ALLMULTI; 9232 } 9233 WRITE_ONCE(dev->allmulti, allmulti); 9234 if (flags != old_flags) { 9235 WRITE_ONCE(dev->flags, flags); 9236 netdev_info(dev, "%s allmulticast mode\n", 9237 dev->flags & IFF_ALLMULTI ? "entered" : "left"); 9238 dev_change_rx_flags(dev, IFF_ALLMULTI); 9239 dev_set_rx_mode(dev); 9240 if (notify) 9241 __dev_notify_flags(dev, old_flags, 9242 dev->gflags ^ old_gflags, 0, NULL); 9243 } 9244 return 0; 9245 } 9246 9247 /* 9248 * Upload unicast and multicast address lists to device and 9249 * configure RX filtering. When the device doesn't support unicast 9250 * filtering it is put in promiscuous mode while unicast addresses 9251 * are present. 9252 */ 9253 void __dev_set_rx_mode(struct net_device *dev) 9254 { 9255 const struct net_device_ops *ops = dev->netdev_ops; 9256 9257 /* dev_open will call this function so the list will stay sane. */ 9258 if (!(dev->flags&IFF_UP)) 9259 return; 9260 9261 if (!netif_device_present(dev)) 9262 return; 9263 9264 if (!(dev->priv_flags & IFF_UNICAST_FLT)) { 9265 /* Unicast addresses changes may only happen under the rtnl, 9266 * therefore calling __dev_set_promiscuity here is safe. 9267 */ 9268 if (!netdev_uc_empty(dev) && !dev->uc_promisc) { 9269 __dev_set_promiscuity(dev, 1, false); 9270 dev->uc_promisc = true; 9271 } else if (netdev_uc_empty(dev) && dev->uc_promisc) { 9272 __dev_set_promiscuity(dev, -1, false); 9273 dev->uc_promisc = false; 9274 } 9275 } 9276 9277 if (ops->ndo_set_rx_mode) 9278 ops->ndo_set_rx_mode(dev); 9279 } 9280 9281 void dev_set_rx_mode(struct net_device *dev) 9282 { 9283 netif_addr_lock_bh(dev); 9284 __dev_set_rx_mode(dev); 9285 netif_addr_unlock_bh(dev); 9286 } 9287 9288 /** 9289 * dev_get_flags - get flags reported to userspace 9290 * @dev: device 9291 * 9292 * Get the combination of flag bits exported through APIs to userspace. 9293 */ 9294 unsigned int dev_get_flags(const struct net_device *dev) 9295 { 9296 unsigned int flags; 9297 9298 flags = (READ_ONCE(dev->flags) & ~(IFF_PROMISC | 9299 IFF_ALLMULTI | 9300 IFF_RUNNING | 9301 IFF_LOWER_UP | 9302 IFF_DORMANT)) | 9303 (READ_ONCE(dev->gflags) & (IFF_PROMISC | 9304 IFF_ALLMULTI)); 9305 9306 if (netif_running(dev)) { 9307 if (netif_oper_up(dev)) 9308 flags |= IFF_RUNNING; 9309 if (netif_carrier_ok(dev)) 9310 flags |= IFF_LOWER_UP; 9311 if (netif_dormant(dev)) 9312 flags |= IFF_DORMANT; 9313 } 9314 9315 return flags; 9316 } 9317 EXPORT_SYMBOL(dev_get_flags); 9318 9319 int __dev_change_flags(struct net_device *dev, unsigned int flags, 9320 struct netlink_ext_ack *extack) 9321 { 9322 unsigned int old_flags = dev->flags; 9323 int ret; 9324 9325 ASSERT_RTNL(); 9326 9327 /* 9328 * Set the flags on our device. 9329 */ 9330 9331 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP | 9332 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL | 9333 IFF_AUTOMEDIA)) | 9334 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC | 9335 IFF_ALLMULTI)); 9336 9337 /* 9338 * Load in the correct multicast list now the flags have changed. 9339 */ 9340 9341 if ((old_flags ^ flags) & IFF_MULTICAST) 9342 dev_change_rx_flags(dev, IFF_MULTICAST); 9343 9344 dev_set_rx_mode(dev); 9345 9346 /* 9347 * Have we downed the interface. We handle IFF_UP ourselves 9348 * according to user attempts to set it, rather than blindly 9349 * setting it. 9350 */ 9351 9352 ret = 0; 9353 if ((old_flags ^ flags) & IFF_UP) { 9354 if (old_flags & IFF_UP) 9355 __dev_close(dev); 9356 else 9357 ret = __dev_open(dev, extack); 9358 } 9359 9360 if ((flags ^ dev->gflags) & IFF_PROMISC) { 9361 int inc = (flags & IFF_PROMISC) ? 1 : -1; 9362 old_flags = dev->flags; 9363 9364 dev->gflags ^= IFF_PROMISC; 9365 9366 if (__dev_set_promiscuity(dev, inc, false) >= 0) 9367 if (dev->flags != old_flags) 9368 dev_set_rx_mode(dev); 9369 } 9370 9371 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI 9372 * is important. Some (broken) drivers set IFF_PROMISC, when 9373 * IFF_ALLMULTI is requested not asking us and not reporting. 9374 */ 9375 if ((flags ^ dev->gflags) & IFF_ALLMULTI) { 9376 int inc = (flags & IFF_ALLMULTI) ? 1 : -1; 9377 9378 dev->gflags ^= IFF_ALLMULTI; 9379 netif_set_allmulti(dev, inc, false); 9380 } 9381 9382 return ret; 9383 } 9384 9385 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags, 9386 unsigned int gchanges, u32 portid, 9387 const struct nlmsghdr *nlh) 9388 { 9389 unsigned int changes = dev->flags ^ old_flags; 9390 9391 if (gchanges) 9392 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh); 9393 9394 if (changes & IFF_UP) { 9395 if (dev->flags & IFF_UP) 9396 call_netdevice_notifiers(NETDEV_UP, dev); 9397 else 9398 call_netdevice_notifiers(NETDEV_DOWN, dev); 9399 } 9400 9401 if (dev->flags & IFF_UP && 9402 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) { 9403 struct netdev_notifier_change_info change_info = { 9404 .info = { 9405 .dev = dev, 9406 }, 9407 .flags_changed = changes, 9408 }; 9409 9410 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info); 9411 } 9412 } 9413 9414 int netif_change_flags(struct net_device *dev, unsigned int flags, 9415 struct netlink_ext_ack *extack) 9416 { 9417 int ret; 9418 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags; 9419 9420 ret = __dev_change_flags(dev, flags, extack); 9421 if (ret < 0) 9422 return ret; 9423 9424 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags); 9425 __dev_notify_flags(dev, old_flags, changes, 0, NULL); 9426 return ret; 9427 } 9428 9429 int __dev_set_mtu(struct net_device *dev, int new_mtu) 9430 { 9431 const struct net_device_ops *ops = dev->netdev_ops; 9432 9433 if (ops->ndo_change_mtu) 9434 return ops->ndo_change_mtu(dev, new_mtu); 9435 9436 /* Pairs with all the lockless reads of dev->mtu in the stack */ 9437 WRITE_ONCE(dev->mtu, new_mtu); 9438 return 0; 9439 } 9440 EXPORT_SYMBOL(__dev_set_mtu); 9441 9442 int dev_validate_mtu(struct net_device *dev, int new_mtu, 9443 struct netlink_ext_ack *extack) 9444 { 9445 /* MTU must be positive, and in range */ 9446 if (new_mtu < 0 || new_mtu < dev->min_mtu) { 9447 NL_SET_ERR_MSG(extack, "mtu less than device minimum"); 9448 return -EINVAL; 9449 } 9450 9451 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) { 9452 NL_SET_ERR_MSG(extack, "mtu greater than device maximum"); 9453 return -EINVAL; 9454 } 9455 return 0; 9456 } 9457 9458 /** 9459 * netif_set_mtu_ext - Change maximum transfer unit 9460 * @dev: device 9461 * @new_mtu: new transfer unit 9462 * @extack: netlink extended ack 9463 * 9464 * Change the maximum transfer size of the network device. 9465 */ 9466 int netif_set_mtu_ext(struct net_device *dev, int new_mtu, 9467 struct netlink_ext_ack *extack) 9468 { 9469 int err, orig_mtu; 9470 9471 if (new_mtu == dev->mtu) 9472 return 0; 9473 9474 err = dev_validate_mtu(dev, new_mtu, extack); 9475 if (err) 9476 return err; 9477 9478 if (!netif_device_present(dev)) 9479 return -ENODEV; 9480 9481 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev); 9482 err = notifier_to_errno(err); 9483 if (err) 9484 return err; 9485 9486 orig_mtu = dev->mtu; 9487 err = __dev_set_mtu(dev, new_mtu); 9488 9489 if (!err) { 9490 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9491 orig_mtu); 9492 err = notifier_to_errno(err); 9493 if (err) { 9494 /* setting mtu back and notifying everyone again, 9495 * so that they have a chance to revert changes. 9496 */ 9497 __dev_set_mtu(dev, orig_mtu); 9498 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9499 new_mtu); 9500 } 9501 } 9502 return err; 9503 } 9504 9505 int netif_set_mtu(struct net_device *dev, int new_mtu) 9506 { 9507 struct netlink_ext_ack extack; 9508 int err; 9509 9510 memset(&extack, 0, sizeof(extack)); 9511 err = netif_set_mtu_ext(dev, new_mtu, &extack); 9512 if (err && extack._msg) 9513 net_err_ratelimited("%s: %s\n", dev->name, extack._msg); 9514 return err; 9515 } 9516 EXPORT_SYMBOL(netif_set_mtu); 9517 9518 int netif_change_tx_queue_len(struct net_device *dev, unsigned long new_len) 9519 { 9520 unsigned int orig_len = dev->tx_queue_len; 9521 int res; 9522 9523 if (new_len != (unsigned int)new_len) 9524 return -ERANGE; 9525 9526 if (new_len != orig_len) { 9527 WRITE_ONCE(dev->tx_queue_len, new_len); 9528 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev); 9529 res = notifier_to_errno(res); 9530 if (res) 9531 goto err_rollback; 9532 res = dev_qdisc_change_tx_queue_len(dev); 9533 if (res) 9534 goto err_rollback; 9535 } 9536 9537 return 0; 9538 9539 err_rollback: 9540 netdev_err(dev, "refused to change device tx_queue_len\n"); 9541 WRITE_ONCE(dev->tx_queue_len, orig_len); 9542 return res; 9543 } 9544 9545 void netif_set_group(struct net_device *dev, int new_group) 9546 { 9547 dev->group = new_group; 9548 } 9549 9550 /** 9551 * dev_pre_changeaddr_notify - Call NETDEV_PRE_CHANGEADDR. 9552 * @dev: device 9553 * @addr: new address 9554 * @extack: netlink extended ack 9555 */ 9556 int dev_pre_changeaddr_notify(struct net_device *dev, const char *addr, 9557 struct netlink_ext_ack *extack) 9558 { 9559 struct netdev_notifier_pre_changeaddr_info info = { 9560 .info.dev = dev, 9561 .info.extack = extack, 9562 .dev_addr = addr, 9563 }; 9564 int rc; 9565 9566 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info); 9567 return notifier_to_errno(rc); 9568 } 9569 EXPORT_SYMBOL(dev_pre_changeaddr_notify); 9570 9571 int netif_set_mac_address(struct net_device *dev, struct sockaddr *sa, 9572 struct netlink_ext_ack *extack) 9573 { 9574 const struct net_device_ops *ops = dev->netdev_ops; 9575 int err; 9576 9577 if (!ops->ndo_set_mac_address) 9578 return -EOPNOTSUPP; 9579 if (sa->sa_family != dev->type) 9580 return -EINVAL; 9581 if (!netif_device_present(dev)) 9582 return -ENODEV; 9583 err = dev_pre_changeaddr_notify(dev, sa->sa_data, extack); 9584 if (err) 9585 return err; 9586 if (memcmp(dev->dev_addr, sa->sa_data, dev->addr_len)) { 9587 err = ops->ndo_set_mac_address(dev, sa); 9588 if (err) 9589 return err; 9590 } 9591 dev->addr_assign_type = NET_ADDR_SET; 9592 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev); 9593 add_device_randomness(dev->dev_addr, dev->addr_len); 9594 return 0; 9595 } 9596 9597 int dev_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name) 9598 { 9599 size_t size = sizeof(sa->sa_data_min); 9600 struct net_device *dev; 9601 9602 dev = netdev_get_by_name_lock(net, dev_name); 9603 if (!dev) 9604 return -ENODEV; 9605 9606 if (!dev->addr_len) 9607 memset(sa->sa_data, 0, size); 9608 else 9609 memcpy(sa->sa_data, dev->dev_addr, 9610 min_t(size_t, size, dev->addr_len)); 9611 sa->sa_family = dev->type; 9612 netdev_unlock(dev); 9613 9614 return 0; 9615 } 9616 EXPORT_SYMBOL(dev_get_mac_address); 9617 9618 int netif_change_carrier(struct net_device *dev, bool new_carrier) 9619 { 9620 const struct net_device_ops *ops = dev->netdev_ops; 9621 9622 if (!ops->ndo_change_carrier) 9623 return -EOPNOTSUPP; 9624 if (!netif_device_present(dev)) 9625 return -ENODEV; 9626 return ops->ndo_change_carrier(dev, new_carrier); 9627 } 9628 9629 /** 9630 * dev_get_phys_port_id - Get device physical port ID 9631 * @dev: device 9632 * @ppid: port ID 9633 * 9634 * Get device physical port ID 9635 */ 9636 int dev_get_phys_port_id(struct net_device *dev, 9637 struct netdev_phys_item_id *ppid) 9638 { 9639 const struct net_device_ops *ops = dev->netdev_ops; 9640 9641 if (!ops->ndo_get_phys_port_id) 9642 return -EOPNOTSUPP; 9643 return ops->ndo_get_phys_port_id(dev, ppid); 9644 } 9645 9646 /** 9647 * dev_get_phys_port_name - Get device physical port name 9648 * @dev: device 9649 * @name: port name 9650 * @len: limit of bytes to copy to name 9651 * 9652 * Get device physical port name 9653 */ 9654 int dev_get_phys_port_name(struct net_device *dev, 9655 char *name, size_t len) 9656 { 9657 const struct net_device_ops *ops = dev->netdev_ops; 9658 int err; 9659 9660 if (ops->ndo_get_phys_port_name) { 9661 err = ops->ndo_get_phys_port_name(dev, name, len); 9662 if (err != -EOPNOTSUPP) 9663 return err; 9664 } 9665 return devlink_compat_phys_port_name_get(dev, name, len); 9666 } 9667 9668 /** 9669 * dev_get_port_parent_id - Get the device's port parent identifier 9670 * @dev: network device 9671 * @ppid: pointer to a storage for the port's parent identifier 9672 * @recurse: allow/disallow recursion to lower devices 9673 * 9674 * Get the devices's port parent identifier 9675 */ 9676 int dev_get_port_parent_id(struct net_device *dev, 9677 struct netdev_phys_item_id *ppid, 9678 bool recurse) 9679 { 9680 const struct net_device_ops *ops = dev->netdev_ops; 9681 struct netdev_phys_item_id first = { }; 9682 struct net_device *lower_dev; 9683 struct list_head *iter; 9684 int err; 9685 9686 if (ops->ndo_get_port_parent_id) { 9687 err = ops->ndo_get_port_parent_id(dev, ppid); 9688 if (err != -EOPNOTSUPP) 9689 return err; 9690 } 9691 9692 err = devlink_compat_switch_id_get(dev, ppid); 9693 if (!recurse || err != -EOPNOTSUPP) 9694 return err; 9695 9696 netdev_for_each_lower_dev(dev, lower_dev, iter) { 9697 err = dev_get_port_parent_id(lower_dev, ppid, true); 9698 if (err) 9699 break; 9700 if (!first.id_len) 9701 first = *ppid; 9702 else if (memcmp(&first, ppid, sizeof(*ppid))) 9703 return -EOPNOTSUPP; 9704 } 9705 9706 return err; 9707 } 9708 EXPORT_SYMBOL(dev_get_port_parent_id); 9709 9710 /** 9711 * netdev_port_same_parent_id - Indicate if two network devices have 9712 * the same port parent identifier 9713 * @a: first network device 9714 * @b: second network device 9715 */ 9716 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b) 9717 { 9718 struct netdev_phys_item_id a_id = { }; 9719 struct netdev_phys_item_id b_id = { }; 9720 9721 if (dev_get_port_parent_id(a, &a_id, true) || 9722 dev_get_port_parent_id(b, &b_id, true)) 9723 return false; 9724 9725 return netdev_phys_item_id_same(&a_id, &b_id); 9726 } 9727 EXPORT_SYMBOL(netdev_port_same_parent_id); 9728 9729 int netif_change_proto_down(struct net_device *dev, bool proto_down) 9730 { 9731 if (!dev->change_proto_down) 9732 return -EOPNOTSUPP; 9733 if (!netif_device_present(dev)) 9734 return -ENODEV; 9735 if (proto_down) 9736 netif_carrier_off(dev); 9737 else 9738 netif_carrier_on(dev); 9739 WRITE_ONCE(dev->proto_down, proto_down); 9740 return 0; 9741 } 9742 9743 /** 9744 * netdev_change_proto_down_reason_locked - proto down reason 9745 * 9746 * @dev: device 9747 * @mask: proto down mask 9748 * @value: proto down value 9749 */ 9750 void netdev_change_proto_down_reason_locked(struct net_device *dev, 9751 unsigned long mask, u32 value) 9752 { 9753 u32 proto_down_reason; 9754 int b; 9755 9756 if (!mask) { 9757 proto_down_reason = value; 9758 } else { 9759 proto_down_reason = dev->proto_down_reason; 9760 for_each_set_bit(b, &mask, 32) { 9761 if (value & (1 << b)) 9762 proto_down_reason |= BIT(b); 9763 else 9764 proto_down_reason &= ~BIT(b); 9765 } 9766 } 9767 WRITE_ONCE(dev->proto_down_reason, proto_down_reason); 9768 } 9769 9770 struct bpf_xdp_link { 9771 struct bpf_link link; 9772 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */ 9773 int flags; 9774 }; 9775 9776 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags) 9777 { 9778 if (flags & XDP_FLAGS_HW_MODE) 9779 return XDP_MODE_HW; 9780 if (flags & XDP_FLAGS_DRV_MODE) 9781 return XDP_MODE_DRV; 9782 if (flags & XDP_FLAGS_SKB_MODE) 9783 return XDP_MODE_SKB; 9784 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB; 9785 } 9786 9787 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode) 9788 { 9789 switch (mode) { 9790 case XDP_MODE_SKB: 9791 return generic_xdp_install; 9792 case XDP_MODE_DRV: 9793 case XDP_MODE_HW: 9794 return dev->netdev_ops->ndo_bpf; 9795 default: 9796 return NULL; 9797 } 9798 } 9799 9800 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev, 9801 enum bpf_xdp_mode mode) 9802 { 9803 return dev->xdp_state[mode].link; 9804 } 9805 9806 static struct bpf_prog *dev_xdp_prog(struct net_device *dev, 9807 enum bpf_xdp_mode mode) 9808 { 9809 struct bpf_xdp_link *link = dev_xdp_link(dev, mode); 9810 9811 if (link) 9812 return link->link.prog; 9813 return dev->xdp_state[mode].prog; 9814 } 9815 9816 u8 dev_xdp_prog_count(struct net_device *dev) 9817 { 9818 u8 count = 0; 9819 int i; 9820 9821 for (i = 0; i < __MAX_XDP_MODE; i++) 9822 if (dev->xdp_state[i].prog || dev->xdp_state[i].link) 9823 count++; 9824 return count; 9825 } 9826 EXPORT_SYMBOL_GPL(dev_xdp_prog_count); 9827 9828 u8 dev_xdp_sb_prog_count(struct net_device *dev) 9829 { 9830 u8 count = 0; 9831 int i; 9832 9833 for (i = 0; i < __MAX_XDP_MODE; i++) 9834 if (dev->xdp_state[i].prog && 9835 !dev->xdp_state[i].prog->aux->xdp_has_frags) 9836 count++; 9837 return count; 9838 } 9839 9840 int netif_xdp_propagate(struct net_device *dev, struct netdev_bpf *bpf) 9841 { 9842 if (!dev->netdev_ops->ndo_bpf) 9843 return -EOPNOTSUPP; 9844 9845 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 9846 bpf->command == XDP_SETUP_PROG && 9847 bpf->prog && !bpf->prog->aux->xdp_has_frags) { 9848 NL_SET_ERR_MSG(bpf->extack, 9849 "unable to propagate XDP to device using tcp-data-split"); 9850 return -EBUSY; 9851 } 9852 9853 if (dev_get_min_mp_channel_count(dev)) { 9854 NL_SET_ERR_MSG(bpf->extack, "unable to propagate XDP to device using memory provider"); 9855 return -EBUSY; 9856 } 9857 9858 return dev->netdev_ops->ndo_bpf(dev, bpf); 9859 } 9860 9861 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode) 9862 { 9863 struct bpf_prog *prog = dev_xdp_prog(dev, mode); 9864 9865 return prog ? prog->aux->id : 0; 9866 } 9867 9868 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode, 9869 struct bpf_xdp_link *link) 9870 { 9871 dev->xdp_state[mode].link = link; 9872 dev->xdp_state[mode].prog = NULL; 9873 } 9874 9875 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode, 9876 struct bpf_prog *prog) 9877 { 9878 dev->xdp_state[mode].link = NULL; 9879 dev->xdp_state[mode].prog = prog; 9880 } 9881 9882 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode, 9883 bpf_op_t bpf_op, struct netlink_ext_ack *extack, 9884 u32 flags, struct bpf_prog *prog) 9885 { 9886 struct netdev_bpf xdp; 9887 int err; 9888 9889 netdev_ops_assert_locked(dev); 9890 9891 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 9892 prog && !prog->aux->xdp_has_frags) { 9893 NL_SET_ERR_MSG(extack, "unable to install XDP to device using tcp-data-split"); 9894 return -EBUSY; 9895 } 9896 9897 if (dev_get_min_mp_channel_count(dev)) { 9898 NL_SET_ERR_MSG(extack, "unable to install XDP to device using memory provider"); 9899 return -EBUSY; 9900 } 9901 9902 memset(&xdp, 0, sizeof(xdp)); 9903 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG; 9904 xdp.extack = extack; 9905 xdp.flags = flags; 9906 xdp.prog = prog; 9907 9908 /* Drivers assume refcnt is already incremented (i.e, prog pointer is 9909 * "moved" into driver), so they don't increment it on their own, but 9910 * they do decrement refcnt when program is detached or replaced. 9911 * Given net_device also owns link/prog, we need to bump refcnt here 9912 * to prevent drivers from underflowing it. 9913 */ 9914 if (prog) 9915 bpf_prog_inc(prog); 9916 err = bpf_op(dev, &xdp); 9917 if (err) { 9918 if (prog) 9919 bpf_prog_put(prog); 9920 return err; 9921 } 9922 9923 if (mode != XDP_MODE_HW) 9924 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog); 9925 9926 return 0; 9927 } 9928 9929 static void dev_xdp_uninstall(struct net_device *dev) 9930 { 9931 struct bpf_xdp_link *link; 9932 struct bpf_prog *prog; 9933 enum bpf_xdp_mode mode; 9934 bpf_op_t bpf_op; 9935 9936 ASSERT_RTNL(); 9937 9938 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) { 9939 prog = dev_xdp_prog(dev, mode); 9940 if (!prog) 9941 continue; 9942 9943 bpf_op = dev_xdp_bpf_op(dev, mode); 9944 if (!bpf_op) 9945 continue; 9946 9947 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 9948 9949 /* auto-detach link from net device */ 9950 link = dev_xdp_link(dev, mode); 9951 if (link) 9952 link->dev = NULL; 9953 else 9954 bpf_prog_put(prog); 9955 9956 dev_xdp_set_link(dev, mode, NULL); 9957 } 9958 } 9959 9960 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack, 9961 struct bpf_xdp_link *link, struct bpf_prog *new_prog, 9962 struct bpf_prog *old_prog, u32 flags) 9963 { 9964 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES); 9965 struct bpf_prog *cur_prog; 9966 struct net_device *upper; 9967 struct list_head *iter; 9968 enum bpf_xdp_mode mode; 9969 bpf_op_t bpf_op; 9970 int err; 9971 9972 ASSERT_RTNL(); 9973 9974 /* either link or prog attachment, never both */ 9975 if (link && (new_prog || old_prog)) 9976 return -EINVAL; 9977 /* link supports only XDP mode flags */ 9978 if (link && (flags & ~XDP_FLAGS_MODES)) { 9979 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment"); 9980 return -EINVAL; 9981 } 9982 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */ 9983 if (num_modes > 1) { 9984 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set"); 9985 return -EINVAL; 9986 } 9987 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */ 9988 if (!num_modes && dev_xdp_prog_count(dev) > 1) { 9989 NL_SET_ERR_MSG(extack, 9990 "More than one program loaded, unset mode is ambiguous"); 9991 return -EINVAL; 9992 } 9993 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */ 9994 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) { 9995 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified"); 9996 return -EINVAL; 9997 } 9998 9999 mode = dev_xdp_mode(dev, flags); 10000 /* can't replace attached link */ 10001 if (dev_xdp_link(dev, mode)) { 10002 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link"); 10003 return -EBUSY; 10004 } 10005 10006 /* don't allow if an upper device already has a program */ 10007 netdev_for_each_upper_dev_rcu(dev, upper, iter) { 10008 if (dev_xdp_prog_count(upper) > 0) { 10009 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program"); 10010 return -EEXIST; 10011 } 10012 } 10013 10014 cur_prog = dev_xdp_prog(dev, mode); 10015 /* can't replace attached prog with link */ 10016 if (link && cur_prog) { 10017 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link"); 10018 return -EBUSY; 10019 } 10020 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) { 10021 NL_SET_ERR_MSG(extack, "Active program does not match expected"); 10022 return -EEXIST; 10023 } 10024 10025 /* put effective new program into new_prog */ 10026 if (link) 10027 new_prog = link->link.prog; 10028 10029 if (new_prog) { 10030 bool offload = mode == XDP_MODE_HW; 10031 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB 10032 ? XDP_MODE_DRV : XDP_MODE_SKB; 10033 10034 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) { 10035 NL_SET_ERR_MSG(extack, "XDP program already attached"); 10036 return -EBUSY; 10037 } 10038 if (!offload && dev_xdp_prog(dev, other_mode)) { 10039 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time"); 10040 return -EEXIST; 10041 } 10042 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) { 10043 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported"); 10044 return -EINVAL; 10045 } 10046 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) { 10047 NL_SET_ERR_MSG(extack, "Program bound to different device"); 10048 return -EINVAL; 10049 } 10050 if (bpf_prog_is_dev_bound(new_prog->aux) && mode == XDP_MODE_SKB) { 10051 NL_SET_ERR_MSG(extack, "Can't attach device-bound programs in generic mode"); 10052 return -EINVAL; 10053 } 10054 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) { 10055 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device"); 10056 return -EINVAL; 10057 } 10058 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) { 10059 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device"); 10060 return -EINVAL; 10061 } 10062 } 10063 10064 /* don't call drivers if the effective program didn't change */ 10065 if (new_prog != cur_prog) { 10066 bpf_op = dev_xdp_bpf_op(dev, mode); 10067 if (!bpf_op) { 10068 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode"); 10069 return -EOPNOTSUPP; 10070 } 10071 10072 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog); 10073 if (err) 10074 return err; 10075 } 10076 10077 if (link) 10078 dev_xdp_set_link(dev, mode, link); 10079 else 10080 dev_xdp_set_prog(dev, mode, new_prog); 10081 if (cur_prog) 10082 bpf_prog_put(cur_prog); 10083 10084 return 0; 10085 } 10086 10087 static int dev_xdp_attach_link(struct net_device *dev, 10088 struct netlink_ext_ack *extack, 10089 struct bpf_xdp_link *link) 10090 { 10091 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags); 10092 } 10093 10094 static int dev_xdp_detach_link(struct net_device *dev, 10095 struct netlink_ext_ack *extack, 10096 struct bpf_xdp_link *link) 10097 { 10098 enum bpf_xdp_mode mode; 10099 bpf_op_t bpf_op; 10100 10101 ASSERT_RTNL(); 10102 10103 mode = dev_xdp_mode(dev, link->flags); 10104 if (dev_xdp_link(dev, mode) != link) 10105 return -EINVAL; 10106 10107 bpf_op = dev_xdp_bpf_op(dev, mode); 10108 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 10109 dev_xdp_set_link(dev, mode, NULL); 10110 return 0; 10111 } 10112 10113 static void bpf_xdp_link_release(struct bpf_link *link) 10114 { 10115 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10116 10117 rtnl_lock(); 10118 10119 /* if racing with net_device's tear down, xdp_link->dev might be 10120 * already NULL, in which case link was already auto-detached 10121 */ 10122 if (xdp_link->dev) { 10123 netdev_lock_ops(xdp_link->dev); 10124 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link)); 10125 netdev_unlock_ops(xdp_link->dev); 10126 xdp_link->dev = NULL; 10127 } 10128 10129 rtnl_unlock(); 10130 } 10131 10132 static int bpf_xdp_link_detach(struct bpf_link *link) 10133 { 10134 bpf_xdp_link_release(link); 10135 return 0; 10136 } 10137 10138 static void bpf_xdp_link_dealloc(struct bpf_link *link) 10139 { 10140 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10141 10142 kfree(xdp_link); 10143 } 10144 10145 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link, 10146 struct seq_file *seq) 10147 { 10148 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10149 u32 ifindex = 0; 10150 10151 rtnl_lock(); 10152 if (xdp_link->dev) 10153 ifindex = xdp_link->dev->ifindex; 10154 rtnl_unlock(); 10155 10156 seq_printf(seq, "ifindex:\t%u\n", ifindex); 10157 } 10158 10159 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link, 10160 struct bpf_link_info *info) 10161 { 10162 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10163 u32 ifindex = 0; 10164 10165 rtnl_lock(); 10166 if (xdp_link->dev) 10167 ifindex = xdp_link->dev->ifindex; 10168 rtnl_unlock(); 10169 10170 info->xdp.ifindex = ifindex; 10171 return 0; 10172 } 10173 10174 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog, 10175 struct bpf_prog *old_prog) 10176 { 10177 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10178 enum bpf_xdp_mode mode; 10179 bpf_op_t bpf_op; 10180 int err = 0; 10181 10182 rtnl_lock(); 10183 10184 /* link might have been auto-released already, so fail */ 10185 if (!xdp_link->dev) { 10186 err = -ENOLINK; 10187 goto out_unlock; 10188 } 10189 10190 if (old_prog && link->prog != old_prog) { 10191 err = -EPERM; 10192 goto out_unlock; 10193 } 10194 old_prog = link->prog; 10195 if (old_prog->type != new_prog->type || 10196 old_prog->expected_attach_type != new_prog->expected_attach_type) { 10197 err = -EINVAL; 10198 goto out_unlock; 10199 } 10200 10201 if (old_prog == new_prog) { 10202 /* no-op, don't disturb drivers */ 10203 bpf_prog_put(new_prog); 10204 goto out_unlock; 10205 } 10206 10207 netdev_lock_ops(xdp_link->dev); 10208 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags); 10209 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode); 10210 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL, 10211 xdp_link->flags, new_prog); 10212 netdev_unlock_ops(xdp_link->dev); 10213 if (err) 10214 goto out_unlock; 10215 10216 old_prog = xchg(&link->prog, new_prog); 10217 bpf_prog_put(old_prog); 10218 10219 out_unlock: 10220 rtnl_unlock(); 10221 return err; 10222 } 10223 10224 static const struct bpf_link_ops bpf_xdp_link_lops = { 10225 .release = bpf_xdp_link_release, 10226 .dealloc = bpf_xdp_link_dealloc, 10227 .detach = bpf_xdp_link_detach, 10228 .show_fdinfo = bpf_xdp_link_show_fdinfo, 10229 .fill_link_info = bpf_xdp_link_fill_link_info, 10230 .update_prog = bpf_xdp_link_update, 10231 }; 10232 10233 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog) 10234 { 10235 struct net *net = current->nsproxy->net_ns; 10236 struct bpf_link_primer link_primer; 10237 struct netlink_ext_ack extack = {}; 10238 struct bpf_xdp_link *link; 10239 struct net_device *dev; 10240 int err, fd; 10241 10242 rtnl_lock(); 10243 dev = dev_get_by_index(net, attr->link_create.target_ifindex); 10244 if (!dev) { 10245 rtnl_unlock(); 10246 return -EINVAL; 10247 } 10248 10249 link = kzalloc(sizeof(*link), GFP_USER); 10250 if (!link) { 10251 err = -ENOMEM; 10252 goto unlock; 10253 } 10254 10255 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog); 10256 link->dev = dev; 10257 link->flags = attr->link_create.flags; 10258 10259 err = bpf_link_prime(&link->link, &link_primer); 10260 if (err) { 10261 kfree(link); 10262 goto unlock; 10263 } 10264 10265 err = dev_xdp_attach_link(dev, &extack, link); 10266 rtnl_unlock(); 10267 10268 if (err) { 10269 link->dev = NULL; 10270 bpf_link_cleanup(&link_primer); 10271 trace_bpf_xdp_link_attach_failed(extack._msg); 10272 goto out_put_dev; 10273 } 10274 10275 fd = bpf_link_settle(&link_primer); 10276 /* link itself doesn't hold dev's refcnt to not complicate shutdown */ 10277 dev_put(dev); 10278 return fd; 10279 10280 unlock: 10281 rtnl_unlock(); 10282 10283 out_put_dev: 10284 dev_put(dev); 10285 return err; 10286 } 10287 10288 /** 10289 * dev_change_xdp_fd - set or clear a bpf program for a device rx path 10290 * @dev: device 10291 * @extack: netlink extended ack 10292 * @fd: new program fd or negative value to clear 10293 * @expected_fd: old program fd that userspace expects to replace or clear 10294 * @flags: xdp-related flags 10295 * 10296 * Set or clear a bpf program for a device 10297 */ 10298 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, 10299 int fd, int expected_fd, u32 flags) 10300 { 10301 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags); 10302 struct bpf_prog *new_prog = NULL, *old_prog = NULL; 10303 int err; 10304 10305 ASSERT_RTNL(); 10306 10307 if (fd >= 0) { 10308 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP, 10309 mode != XDP_MODE_SKB); 10310 if (IS_ERR(new_prog)) 10311 return PTR_ERR(new_prog); 10312 } 10313 10314 if (expected_fd >= 0) { 10315 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP, 10316 mode != XDP_MODE_SKB); 10317 if (IS_ERR(old_prog)) { 10318 err = PTR_ERR(old_prog); 10319 old_prog = NULL; 10320 goto err_out; 10321 } 10322 } 10323 10324 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags); 10325 10326 err_out: 10327 if (err && new_prog) 10328 bpf_prog_put(new_prog); 10329 if (old_prog) 10330 bpf_prog_put(old_prog); 10331 return err; 10332 } 10333 10334 u32 dev_get_min_mp_channel_count(const struct net_device *dev) 10335 { 10336 int i; 10337 10338 ASSERT_RTNL(); 10339 10340 for (i = dev->real_num_rx_queues - 1; i >= 0; i--) 10341 if (dev->_rx[i].mp_params.mp_priv) 10342 /* The channel count is the idx plus 1. */ 10343 return i + 1; 10344 10345 return 0; 10346 } 10347 10348 /** 10349 * dev_index_reserve() - allocate an ifindex in a namespace 10350 * @net: the applicable net namespace 10351 * @ifindex: requested ifindex, pass %0 to get one allocated 10352 * 10353 * Allocate a ifindex for a new device. Caller must either use the ifindex 10354 * to store the device (via list_netdevice()) or call dev_index_release() 10355 * to give the index up. 10356 * 10357 * Return: a suitable unique value for a new device interface number or -errno. 10358 */ 10359 static int dev_index_reserve(struct net *net, u32 ifindex) 10360 { 10361 int err; 10362 10363 if (ifindex > INT_MAX) { 10364 DEBUG_NET_WARN_ON_ONCE(1); 10365 return -EINVAL; 10366 } 10367 10368 if (!ifindex) 10369 err = xa_alloc_cyclic(&net->dev_by_index, &ifindex, NULL, 10370 xa_limit_31b, &net->ifindex, GFP_KERNEL); 10371 else 10372 err = xa_insert(&net->dev_by_index, ifindex, NULL, GFP_KERNEL); 10373 if (err < 0) 10374 return err; 10375 10376 return ifindex; 10377 } 10378 10379 static void dev_index_release(struct net *net, int ifindex) 10380 { 10381 /* Expect only unused indexes, unlist_netdevice() removes the used */ 10382 WARN_ON(xa_erase(&net->dev_by_index, ifindex)); 10383 } 10384 10385 static bool from_cleanup_net(void) 10386 { 10387 #ifdef CONFIG_NET_NS 10388 return current == cleanup_net_task; 10389 #else 10390 return false; 10391 #endif 10392 } 10393 10394 /* Delayed registration/unregisteration */ 10395 LIST_HEAD(net_todo_list); 10396 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq); 10397 atomic_t dev_unreg_count = ATOMIC_INIT(0); 10398 10399 static void net_set_todo(struct net_device *dev) 10400 { 10401 list_add_tail(&dev->todo_list, &net_todo_list); 10402 } 10403 10404 static netdev_features_t netdev_sync_upper_features(struct net_device *lower, 10405 struct net_device *upper, netdev_features_t features) 10406 { 10407 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10408 netdev_features_t feature; 10409 int feature_bit; 10410 10411 for_each_netdev_feature(upper_disables, feature_bit) { 10412 feature = __NETIF_F_BIT(feature_bit); 10413 if (!(upper->wanted_features & feature) 10414 && (features & feature)) { 10415 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n", 10416 &feature, upper->name); 10417 features &= ~feature; 10418 } 10419 } 10420 10421 return features; 10422 } 10423 10424 static void netdev_sync_lower_features(struct net_device *upper, 10425 struct net_device *lower, netdev_features_t features) 10426 { 10427 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10428 netdev_features_t feature; 10429 int feature_bit; 10430 10431 for_each_netdev_feature(upper_disables, feature_bit) { 10432 feature = __NETIF_F_BIT(feature_bit); 10433 if (!(features & feature) && (lower->features & feature)) { 10434 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n", 10435 &feature, lower->name); 10436 lower->wanted_features &= ~feature; 10437 __netdev_update_features(lower); 10438 10439 if (unlikely(lower->features & feature)) 10440 netdev_WARN(upper, "failed to disable %pNF on %s!\n", 10441 &feature, lower->name); 10442 else 10443 netdev_features_change(lower); 10444 } 10445 } 10446 } 10447 10448 static bool netdev_has_ip_or_hw_csum(netdev_features_t features) 10449 { 10450 netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 10451 bool ip_csum = (features & ip_csum_mask) == ip_csum_mask; 10452 bool hw_csum = features & NETIF_F_HW_CSUM; 10453 10454 return ip_csum || hw_csum; 10455 } 10456 10457 static netdev_features_t netdev_fix_features(struct net_device *dev, 10458 netdev_features_t features) 10459 { 10460 /* Fix illegal checksum combinations */ 10461 if ((features & NETIF_F_HW_CSUM) && 10462 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) { 10463 netdev_warn(dev, "mixed HW and IP checksum settings.\n"); 10464 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM); 10465 } 10466 10467 /* TSO requires that SG is present as well. */ 10468 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) { 10469 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n"); 10470 features &= ~NETIF_F_ALL_TSO; 10471 } 10472 10473 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) && 10474 !(features & NETIF_F_IP_CSUM)) { 10475 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n"); 10476 features &= ~NETIF_F_TSO; 10477 features &= ~NETIF_F_TSO_ECN; 10478 } 10479 10480 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) && 10481 !(features & NETIF_F_IPV6_CSUM)) { 10482 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n"); 10483 features &= ~NETIF_F_TSO6; 10484 } 10485 10486 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */ 10487 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO)) 10488 features &= ~NETIF_F_TSO_MANGLEID; 10489 10490 /* TSO ECN requires that TSO is present as well. */ 10491 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN) 10492 features &= ~NETIF_F_TSO_ECN; 10493 10494 /* Software GSO depends on SG. */ 10495 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) { 10496 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n"); 10497 features &= ~NETIF_F_GSO; 10498 } 10499 10500 /* GSO partial features require GSO partial be set */ 10501 if ((features & dev->gso_partial_features) && 10502 !(features & NETIF_F_GSO_PARTIAL)) { 10503 netdev_dbg(dev, 10504 "Dropping partially supported GSO features since no GSO partial.\n"); 10505 features &= ~dev->gso_partial_features; 10506 } 10507 10508 if (!(features & NETIF_F_RXCSUM)) { 10509 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet 10510 * successfully merged by hardware must also have the 10511 * checksum verified by hardware. If the user does not 10512 * want to enable RXCSUM, logically, we should disable GRO_HW. 10513 */ 10514 if (features & NETIF_F_GRO_HW) { 10515 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n"); 10516 features &= ~NETIF_F_GRO_HW; 10517 } 10518 } 10519 10520 /* LRO/HW-GRO features cannot be combined with RX-FCS */ 10521 if (features & NETIF_F_RXFCS) { 10522 if (features & NETIF_F_LRO) { 10523 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n"); 10524 features &= ~NETIF_F_LRO; 10525 } 10526 10527 if (features & NETIF_F_GRO_HW) { 10528 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n"); 10529 features &= ~NETIF_F_GRO_HW; 10530 } 10531 } 10532 10533 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) { 10534 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n"); 10535 features &= ~NETIF_F_LRO; 10536 } 10537 10538 if ((features & NETIF_F_HW_TLS_TX) && !netdev_has_ip_or_hw_csum(features)) { 10539 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n"); 10540 features &= ~NETIF_F_HW_TLS_TX; 10541 } 10542 10543 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) { 10544 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n"); 10545 features &= ~NETIF_F_HW_TLS_RX; 10546 } 10547 10548 if ((features & NETIF_F_GSO_UDP_L4) && !netdev_has_ip_or_hw_csum(features)) { 10549 netdev_dbg(dev, "Dropping USO feature since no CSUM feature.\n"); 10550 features &= ~NETIF_F_GSO_UDP_L4; 10551 } 10552 10553 return features; 10554 } 10555 10556 int __netdev_update_features(struct net_device *dev) 10557 { 10558 struct net_device *upper, *lower; 10559 netdev_features_t features; 10560 struct list_head *iter; 10561 int err = -1; 10562 10563 ASSERT_RTNL(); 10564 netdev_ops_assert_locked(dev); 10565 10566 features = netdev_get_wanted_features(dev); 10567 10568 if (dev->netdev_ops->ndo_fix_features) 10569 features = dev->netdev_ops->ndo_fix_features(dev, features); 10570 10571 /* driver might be less strict about feature dependencies */ 10572 features = netdev_fix_features(dev, features); 10573 10574 /* some features can't be enabled if they're off on an upper device */ 10575 netdev_for_each_upper_dev_rcu(dev, upper, iter) 10576 features = netdev_sync_upper_features(dev, upper, features); 10577 10578 if (dev->features == features) 10579 goto sync_lower; 10580 10581 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n", 10582 &dev->features, &features); 10583 10584 if (dev->netdev_ops->ndo_set_features) 10585 err = dev->netdev_ops->ndo_set_features(dev, features); 10586 else 10587 err = 0; 10588 10589 if (unlikely(err < 0)) { 10590 netdev_err(dev, 10591 "set_features() failed (%d); wanted %pNF, left %pNF\n", 10592 err, &features, &dev->features); 10593 /* return non-0 since some features might have changed and 10594 * it's better to fire a spurious notification than miss it 10595 */ 10596 return -1; 10597 } 10598 10599 sync_lower: 10600 /* some features must be disabled on lower devices when disabled 10601 * on an upper device (think: bonding master or bridge) 10602 */ 10603 netdev_for_each_lower_dev(dev, lower, iter) 10604 netdev_sync_lower_features(dev, lower, features); 10605 10606 if (!err) { 10607 netdev_features_t diff = features ^ dev->features; 10608 10609 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) { 10610 /* udp_tunnel_{get,drop}_rx_info both need 10611 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the 10612 * device, or they won't do anything. 10613 * Thus we need to update dev->features 10614 * *before* calling udp_tunnel_get_rx_info, 10615 * but *after* calling udp_tunnel_drop_rx_info. 10616 */ 10617 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) { 10618 dev->features = features; 10619 udp_tunnel_get_rx_info(dev); 10620 } else { 10621 udp_tunnel_drop_rx_info(dev); 10622 } 10623 } 10624 10625 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) { 10626 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) { 10627 dev->features = features; 10628 err |= vlan_get_rx_ctag_filter_info(dev); 10629 } else { 10630 vlan_drop_rx_ctag_filter_info(dev); 10631 } 10632 } 10633 10634 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) { 10635 if (features & NETIF_F_HW_VLAN_STAG_FILTER) { 10636 dev->features = features; 10637 err |= vlan_get_rx_stag_filter_info(dev); 10638 } else { 10639 vlan_drop_rx_stag_filter_info(dev); 10640 } 10641 } 10642 10643 dev->features = features; 10644 } 10645 10646 return err < 0 ? 0 : 1; 10647 } 10648 10649 /** 10650 * netdev_update_features - recalculate device features 10651 * @dev: the device to check 10652 * 10653 * Recalculate dev->features set and send notifications if it 10654 * has changed. Should be called after driver or hardware dependent 10655 * conditions might have changed that influence the features. 10656 */ 10657 void netdev_update_features(struct net_device *dev) 10658 { 10659 if (__netdev_update_features(dev)) 10660 netdev_features_change(dev); 10661 } 10662 EXPORT_SYMBOL(netdev_update_features); 10663 10664 /** 10665 * netdev_change_features - recalculate device features 10666 * @dev: the device to check 10667 * 10668 * Recalculate dev->features set and send notifications even 10669 * if they have not changed. Should be called instead of 10670 * netdev_update_features() if also dev->vlan_features might 10671 * have changed to allow the changes to be propagated to stacked 10672 * VLAN devices. 10673 */ 10674 void netdev_change_features(struct net_device *dev) 10675 { 10676 __netdev_update_features(dev); 10677 netdev_features_change(dev); 10678 } 10679 EXPORT_SYMBOL(netdev_change_features); 10680 10681 /** 10682 * netif_stacked_transfer_operstate - transfer operstate 10683 * @rootdev: the root or lower level device to transfer state from 10684 * @dev: the device to transfer operstate to 10685 * 10686 * Transfer operational state from root to device. This is normally 10687 * called when a stacking relationship exists between the root 10688 * device and the device(a leaf device). 10689 */ 10690 void netif_stacked_transfer_operstate(const struct net_device *rootdev, 10691 struct net_device *dev) 10692 { 10693 if (rootdev->operstate == IF_OPER_DORMANT) 10694 netif_dormant_on(dev); 10695 else 10696 netif_dormant_off(dev); 10697 10698 if (rootdev->operstate == IF_OPER_TESTING) 10699 netif_testing_on(dev); 10700 else 10701 netif_testing_off(dev); 10702 10703 if (netif_carrier_ok(rootdev)) 10704 netif_carrier_on(dev); 10705 else 10706 netif_carrier_off(dev); 10707 } 10708 EXPORT_SYMBOL(netif_stacked_transfer_operstate); 10709 10710 static int netif_alloc_rx_queues(struct net_device *dev) 10711 { 10712 unsigned int i, count = dev->num_rx_queues; 10713 struct netdev_rx_queue *rx; 10714 size_t sz = count * sizeof(*rx); 10715 int err = 0; 10716 10717 BUG_ON(count < 1); 10718 10719 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10720 if (!rx) 10721 return -ENOMEM; 10722 10723 dev->_rx = rx; 10724 10725 for (i = 0; i < count; i++) { 10726 rx[i].dev = dev; 10727 10728 /* XDP RX-queue setup */ 10729 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0); 10730 if (err < 0) 10731 goto err_rxq_info; 10732 } 10733 return 0; 10734 10735 err_rxq_info: 10736 /* Rollback successful reg's and free other resources */ 10737 while (i--) 10738 xdp_rxq_info_unreg(&rx[i].xdp_rxq); 10739 kvfree(dev->_rx); 10740 dev->_rx = NULL; 10741 return err; 10742 } 10743 10744 static void netif_free_rx_queues(struct net_device *dev) 10745 { 10746 unsigned int i, count = dev->num_rx_queues; 10747 10748 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */ 10749 if (!dev->_rx) 10750 return; 10751 10752 for (i = 0; i < count; i++) 10753 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq); 10754 10755 kvfree(dev->_rx); 10756 } 10757 10758 static void netdev_init_one_queue(struct net_device *dev, 10759 struct netdev_queue *queue, void *_unused) 10760 { 10761 /* Initialize queue lock */ 10762 spin_lock_init(&queue->_xmit_lock); 10763 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type); 10764 queue->xmit_lock_owner = -1; 10765 netdev_queue_numa_node_write(queue, NUMA_NO_NODE); 10766 queue->dev = dev; 10767 #ifdef CONFIG_BQL 10768 dql_init(&queue->dql, HZ); 10769 #endif 10770 } 10771 10772 static void netif_free_tx_queues(struct net_device *dev) 10773 { 10774 kvfree(dev->_tx); 10775 } 10776 10777 static int netif_alloc_netdev_queues(struct net_device *dev) 10778 { 10779 unsigned int count = dev->num_tx_queues; 10780 struct netdev_queue *tx; 10781 size_t sz = count * sizeof(*tx); 10782 10783 if (count < 1 || count > 0xffff) 10784 return -EINVAL; 10785 10786 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10787 if (!tx) 10788 return -ENOMEM; 10789 10790 dev->_tx = tx; 10791 10792 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL); 10793 spin_lock_init(&dev->tx_global_lock); 10794 10795 return 0; 10796 } 10797 10798 void netif_tx_stop_all_queues(struct net_device *dev) 10799 { 10800 unsigned int i; 10801 10802 for (i = 0; i < dev->num_tx_queues; i++) { 10803 struct netdev_queue *txq = netdev_get_tx_queue(dev, i); 10804 10805 netif_tx_stop_queue(txq); 10806 } 10807 } 10808 EXPORT_SYMBOL(netif_tx_stop_all_queues); 10809 10810 static int netdev_do_alloc_pcpu_stats(struct net_device *dev) 10811 { 10812 void __percpu *v; 10813 10814 /* Drivers implementing ndo_get_peer_dev must support tstat 10815 * accounting, so that skb_do_redirect() can bump the dev's 10816 * RX stats upon network namespace switch. 10817 */ 10818 if (dev->netdev_ops->ndo_get_peer_dev && 10819 dev->pcpu_stat_type != NETDEV_PCPU_STAT_TSTATS) 10820 return -EOPNOTSUPP; 10821 10822 switch (dev->pcpu_stat_type) { 10823 case NETDEV_PCPU_STAT_NONE: 10824 return 0; 10825 case NETDEV_PCPU_STAT_LSTATS: 10826 v = dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats); 10827 break; 10828 case NETDEV_PCPU_STAT_TSTATS: 10829 v = dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats); 10830 break; 10831 case NETDEV_PCPU_STAT_DSTATS: 10832 v = dev->dstats = netdev_alloc_pcpu_stats(struct pcpu_dstats); 10833 break; 10834 default: 10835 return -EINVAL; 10836 } 10837 10838 return v ? 0 : -ENOMEM; 10839 } 10840 10841 static void netdev_do_free_pcpu_stats(struct net_device *dev) 10842 { 10843 switch (dev->pcpu_stat_type) { 10844 case NETDEV_PCPU_STAT_NONE: 10845 return; 10846 case NETDEV_PCPU_STAT_LSTATS: 10847 free_percpu(dev->lstats); 10848 break; 10849 case NETDEV_PCPU_STAT_TSTATS: 10850 free_percpu(dev->tstats); 10851 break; 10852 case NETDEV_PCPU_STAT_DSTATS: 10853 free_percpu(dev->dstats); 10854 break; 10855 } 10856 } 10857 10858 static void netdev_free_phy_link_topology(struct net_device *dev) 10859 { 10860 struct phy_link_topology *topo = dev->link_topo; 10861 10862 if (IS_ENABLED(CONFIG_PHYLIB) && topo) { 10863 xa_destroy(&topo->phys); 10864 kfree(topo); 10865 dev->link_topo = NULL; 10866 } 10867 } 10868 10869 /** 10870 * register_netdevice() - register a network device 10871 * @dev: device to register 10872 * 10873 * Take a prepared network device structure and make it externally accessible. 10874 * A %NETDEV_REGISTER message is sent to the netdev notifier chain. 10875 * Callers must hold the rtnl lock - you may want register_netdev() 10876 * instead of this. 10877 */ 10878 int register_netdevice(struct net_device *dev) 10879 { 10880 int ret; 10881 struct net *net = dev_net(dev); 10882 10883 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE < 10884 NETDEV_FEATURE_COUNT); 10885 BUG_ON(dev_boot_phase); 10886 ASSERT_RTNL(); 10887 10888 might_sleep(); 10889 10890 /* When net_device's are persistent, this will be fatal. */ 10891 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED); 10892 BUG_ON(!net); 10893 10894 ret = ethtool_check_ops(dev->ethtool_ops); 10895 if (ret) 10896 return ret; 10897 10898 /* rss ctx ID 0 is reserved for the default context, start from 1 */ 10899 xa_init_flags(&dev->ethtool->rss_ctx, XA_FLAGS_ALLOC1); 10900 mutex_init(&dev->ethtool->rss_lock); 10901 10902 spin_lock_init(&dev->addr_list_lock); 10903 netdev_set_addr_lockdep_class(dev); 10904 10905 ret = dev_get_valid_name(net, dev, dev->name); 10906 if (ret < 0) 10907 goto out; 10908 10909 ret = -ENOMEM; 10910 dev->name_node = netdev_name_node_head_alloc(dev); 10911 if (!dev->name_node) 10912 goto out; 10913 10914 /* Init, if this function is available */ 10915 if (dev->netdev_ops->ndo_init) { 10916 ret = dev->netdev_ops->ndo_init(dev); 10917 if (ret) { 10918 if (ret > 0) 10919 ret = -EIO; 10920 goto err_free_name; 10921 } 10922 } 10923 10924 if (((dev->hw_features | dev->features) & 10925 NETIF_F_HW_VLAN_CTAG_FILTER) && 10926 (!dev->netdev_ops->ndo_vlan_rx_add_vid || 10927 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) { 10928 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n"); 10929 ret = -EINVAL; 10930 goto err_uninit; 10931 } 10932 10933 ret = netdev_do_alloc_pcpu_stats(dev); 10934 if (ret) 10935 goto err_uninit; 10936 10937 ret = dev_index_reserve(net, dev->ifindex); 10938 if (ret < 0) 10939 goto err_free_pcpu; 10940 dev->ifindex = ret; 10941 10942 /* Transfer changeable features to wanted_features and enable 10943 * software offloads (GSO and GRO). 10944 */ 10945 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF); 10946 dev->features |= NETIF_F_SOFT_FEATURES; 10947 10948 if (dev->udp_tunnel_nic_info) { 10949 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10950 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10951 } 10952 10953 dev->wanted_features = dev->features & dev->hw_features; 10954 10955 if (!(dev->flags & IFF_LOOPBACK)) 10956 dev->hw_features |= NETIF_F_NOCACHE_COPY; 10957 10958 /* If IPv4 TCP segmentation offload is supported we should also 10959 * allow the device to enable segmenting the frame with the option 10960 * of ignoring a static IP ID value. This doesn't enable the 10961 * feature itself but allows the user to enable it later. 10962 */ 10963 if (dev->hw_features & NETIF_F_TSO) 10964 dev->hw_features |= NETIF_F_TSO_MANGLEID; 10965 if (dev->vlan_features & NETIF_F_TSO) 10966 dev->vlan_features |= NETIF_F_TSO_MANGLEID; 10967 if (dev->mpls_features & NETIF_F_TSO) 10968 dev->mpls_features |= NETIF_F_TSO_MANGLEID; 10969 if (dev->hw_enc_features & NETIF_F_TSO) 10970 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID; 10971 10972 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices. 10973 */ 10974 dev->vlan_features |= NETIF_F_HIGHDMA; 10975 10976 /* Make NETIF_F_SG inheritable to tunnel devices. 10977 */ 10978 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL; 10979 10980 /* Make NETIF_F_SG inheritable to MPLS. 10981 */ 10982 dev->mpls_features |= NETIF_F_SG; 10983 10984 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev); 10985 ret = notifier_to_errno(ret); 10986 if (ret) 10987 goto err_ifindex_release; 10988 10989 ret = netdev_register_kobject(dev); 10990 10991 netdev_lock(dev); 10992 WRITE_ONCE(dev->reg_state, ret ? NETREG_UNREGISTERED : NETREG_REGISTERED); 10993 netdev_unlock(dev); 10994 10995 if (ret) 10996 goto err_uninit_notify; 10997 10998 netdev_lock_ops(dev); 10999 __netdev_update_features(dev); 11000 netdev_unlock_ops(dev); 11001 11002 /* 11003 * Default initial state at registry is that the 11004 * device is present. 11005 */ 11006 11007 set_bit(__LINK_STATE_PRESENT, &dev->state); 11008 11009 linkwatch_init_dev(dev); 11010 11011 dev_init_scheduler(dev); 11012 11013 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL); 11014 list_netdevice(dev); 11015 11016 add_device_randomness(dev->dev_addr, dev->addr_len); 11017 11018 /* If the device has permanent device address, driver should 11019 * set dev_addr and also addr_assign_type should be set to 11020 * NET_ADDR_PERM (default value). 11021 */ 11022 if (dev->addr_assign_type == NET_ADDR_PERM) 11023 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len); 11024 11025 /* Notify protocols, that a new device appeared. */ 11026 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev); 11027 ret = notifier_to_errno(ret); 11028 if (ret) { 11029 /* Expect explicit free_netdev() on failure */ 11030 dev->needs_free_netdev = false; 11031 unregister_netdevice_queue(dev, NULL); 11032 goto out; 11033 } 11034 /* 11035 * Prevent userspace races by waiting until the network 11036 * device is fully setup before sending notifications. 11037 */ 11038 if (!dev->rtnl_link_ops || 11039 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 11040 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 11041 11042 out: 11043 return ret; 11044 11045 err_uninit_notify: 11046 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11047 err_ifindex_release: 11048 dev_index_release(net, dev->ifindex); 11049 err_free_pcpu: 11050 netdev_do_free_pcpu_stats(dev); 11051 err_uninit: 11052 if (dev->netdev_ops->ndo_uninit) 11053 dev->netdev_ops->ndo_uninit(dev); 11054 if (dev->priv_destructor) 11055 dev->priv_destructor(dev); 11056 err_free_name: 11057 netdev_name_node_free(dev->name_node); 11058 goto out; 11059 } 11060 EXPORT_SYMBOL(register_netdevice); 11061 11062 /* Initialize the core of a dummy net device. 11063 * The setup steps dummy netdevs need which normal netdevs get by going 11064 * through register_netdevice(). 11065 */ 11066 static void init_dummy_netdev(struct net_device *dev) 11067 { 11068 /* make sure we BUG if trying to hit standard 11069 * register/unregister code path 11070 */ 11071 dev->reg_state = NETREG_DUMMY; 11072 11073 /* a dummy interface is started by default */ 11074 set_bit(__LINK_STATE_PRESENT, &dev->state); 11075 set_bit(__LINK_STATE_START, &dev->state); 11076 11077 /* Note : We dont allocate pcpu_refcnt for dummy devices, 11078 * because users of this 'device' dont need to change 11079 * its refcount. 11080 */ 11081 } 11082 11083 /** 11084 * register_netdev - register a network device 11085 * @dev: device to register 11086 * 11087 * Take a completed network device structure and add it to the kernel 11088 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier 11089 * chain. 0 is returned on success. A negative errno code is returned 11090 * on a failure to set up the device, or if the name is a duplicate. 11091 * 11092 * This is a wrapper around register_netdevice that takes the rtnl semaphore 11093 * and expands the device name if you passed a format string to 11094 * alloc_netdev. 11095 */ 11096 int register_netdev(struct net_device *dev) 11097 { 11098 struct net *net = dev_net(dev); 11099 int err; 11100 11101 if (rtnl_net_lock_killable(net)) 11102 return -EINTR; 11103 11104 err = register_netdevice(dev); 11105 11106 rtnl_net_unlock(net); 11107 11108 return err; 11109 } 11110 EXPORT_SYMBOL(register_netdev); 11111 11112 int netdev_refcnt_read(const struct net_device *dev) 11113 { 11114 #ifdef CONFIG_PCPU_DEV_REFCNT 11115 int i, refcnt = 0; 11116 11117 for_each_possible_cpu(i) 11118 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i); 11119 return refcnt; 11120 #else 11121 return refcount_read(&dev->dev_refcnt); 11122 #endif 11123 } 11124 EXPORT_SYMBOL(netdev_refcnt_read); 11125 11126 int netdev_unregister_timeout_secs __read_mostly = 10; 11127 11128 #define WAIT_REFS_MIN_MSECS 1 11129 #define WAIT_REFS_MAX_MSECS 250 11130 /** 11131 * netdev_wait_allrefs_any - wait until all references are gone. 11132 * @list: list of net_devices to wait on 11133 * 11134 * This is called when unregistering network devices. 11135 * 11136 * Any protocol or device that holds a reference should register 11137 * for netdevice notification, and cleanup and put back the 11138 * reference if they receive an UNREGISTER event. 11139 * We can get stuck here if buggy protocols don't correctly 11140 * call dev_put. 11141 */ 11142 static struct net_device *netdev_wait_allrefs_any(struct list_head *list) 11143 { 11144 unsigned long rebroadcast_time, warning_time; 11145 struct net_device *dev; 11146 int wait = 0; 11147 11148 rebroadcast_time = warning_time = jiffies; 11149 11150 list_for_each_entry(dev, list, todo_list) 11151 if (netdev_refcnt_read(dev) == 1) 11152 return dev; 11153 11154 while (true) { 11155 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) { 11156 rtnl_lock(); 11157 11158 /* Rebroadcast unregister notification */ 11159 list_for_each_entry(dev, list, todo_list) 11160 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11161 11162 __rtnl_unlock(); 11163 rcu_barrier(); 11164 rtnl_lock(); 11165 11166 list_for_each_entry(dev, list, todo_list) 11167 if (test_bit(__LINK_STATE_LINKWATCH_PENDING, 11168 &dev->state)) { 11169 /* We must not have linkwatch events 11170 * pending on unregister. If this 11171 * happens, we simply run the queue 11172 * unscheduled, resulting in a noop 11173 * for this device. 11174 */ 11175 linkwatch_run_queue(); 11176 break; 11177 } 11178 11179 __rtnl_unlock(); 11180 11181 rebroadcast_time = jiffies; 11182 } 11183 11184 rcu_barrier(); 11185 11186 if (!wait) { 11187 wait = WAIT_REFS_MIN_MSECS; 11188 } else { 11189 msleep(wait); 11190 wait = min(wait << 1, WAIT_REFS_MAX_MSECS); 11191 } 11192 11193 list_for_each_entry(dev, list, todo_list) 11194 if (netdev_refcnt_read(dev) == 1) 11195 return dev; 11196 11197 if (time_after(jiffies, warning_time + 11198 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) { 11199 list_for_each_entry(dev, list, todo_list) { 11200 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n", 11201 dev->name, netdev_refcnt_read(dev)); 11202 ref_tracker_dir_print(&dev->refcnt_tracker, 10); 11203 } 11204 11205 warning_time = jiffies; 11206 } 11207 } 11208 } 11209 11210 /* The sequence is: 11211 * 11212 * rtnl_lock(); 11213 * ... 11214 * register_netdevice(x1); 11215 * register_netdevice(x2); 11216 * ... 11217 * unregister_netdevice(y1); 11218 * unregister_netdevice(y2); 11219 * ... 11220 * rtnl_unlock(); 11221 * free_netdev(y1); 11222 * free_netdev(y2); 11223 * 11224 * We are invoked by rtnl_unlock(). 11225 * This allows us to deal with problems: 11226 * 1) We can delete sysfs objects which invoke hotplug 11227 * without deadlocking with linkwatch via keventd. 11228 * 2) Since we run with the RTNL semaphore not held, we can sleep 11229 * safely in order to wait for the netdev refcnt to drop to zero. 11230 * 11231 * We must not return until all unregister events added during 11232 * the interval the lock was held have been completed. 11233 */ 11234 void netdev_run_todo(void) 11235 { 11236 struct net_device *dev, *tmp; 11237 struct list_head list; 11238 int cnt; 11239 #ifdef CONFIG_LOCKDEP 11240 struct list_head unlink_list; 11241 11242 list_replace_init(&net_unlink_list, &unlink_list); 11243 11244 while (!list_empty(&unlink_list)) { 11245 dev = list_first_entry(&unlink_list, struct net_device, 11246 unlink_list); 11247 list_del_init(&dev->unlink_list); 11248 dev->nested_level = dev->lower_level - 1; 11249 } 11250 #endif 11251 11252 /* Snapshot list, allow later requests */ 11253 list_replace_init(&net_todo_list, &list); 11254 11255 __rtnl_unlock(); 11256 11257 /* Wait for rcu callbacks to finish before next phase */ 11258 if (!list_empty(&list)) 11259 rcu_barrier(); 11260 11261 list_for_each_entry_safe(dev, tmp, &list, todo_list) { 11262 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) { 11263 netdev_WARN(dev, "run_todo but not unregistering\n"); 11264 list_del(&dev->todo_list); 11265 continue; 11266 } 11267 11268 netdev_lock(dev); 11269 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERED); 11270 netdev_unlock(dev); 11271 linkwatch_sync_dev(dev); 11272 } 11273 11274 cnt = 0; 11275 while (!list_empty(&list)) { 11276 dev = netdev_wait_allrefs_any(&list); 11277 list_del(&dev->todo_list); 11278 11279 /* paranoia */ 11280 BUG_ON(netdev_refcnt_read(dev) != 1); 11281 BUG_ON(!list_empty(&dev->ptype_all)); 11282 BUG_ON(!list_empty(&dev->ptype_specific)); 11283 WARN_ON(rcu_access_pointer(dev->ip_ptr)); 11284 WARN_ON(rcu_access_pointer(dev->ip6_ptr)); 11285 11286 netdev_do_free_pcpu_stats(dev); 11287 if (dev->priv_destructor) 11288 dev->priv_destructor(dev); 11289 if (dev->needs_free_netdev) 11290 free_netdev(dev); 11291 11292 cnt++; 11293 11294 /* Free network device */ 11295 kobject_put(&dev->dev.kobj); 11296 } 11297 if (cnt && atomic_sub_and_test(cnt, &dev_unreg_count)) 11298 wake_up(&netdev_unregistering_wq); 11299 } 11300 11301 /* Collate per-cpu network dstats statistics 11302 * 11303 * Read per-cpu network statistics from dev->dstats and populate the related 11304 * fields in @s. 11305 */ 11306 static void dev_fetch_dstats(struct rtnl_link_stats64 *s, 11307 const struct pcpu_dstats __percpu *dstats) 11308 { 11309 int cpu; 11310 11311 for_each_possible_cpu(cpu) { 11312 u64 rx_packets, rx_bytes, rx_drops; 11313 u64 tx_packets, tx_bytes, tx_drops; 11314 const struct pcpu_dstats *stats; 11315 unsigned int start; 11316 11317 stats = per_cpu_ptr(dstats, cpu); 11318 do { 11319 start = u64_stats_fetch_begin(&stats->syncp); 11320 rx_packets = u64_stats_read(&stats->rx_packets); 11321 rx_bytes = u64_stats_read(&stats->rx_bytes); 11322 rx_drops = u64_stats_read(&stats->rx_drops); 11323 tx_packets = u64_stats_read(&stats->tx_packets); 11324 tx_bytes = u64_stats_read(&stats->tx_bytes); 11325 tx_drops = u64_stats_read(&stats->tx_drops); 11326 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11327 11328 s->rx_packets += rx_packets; 11329 s->rx_bytes += rx_bytes; 11330 s->rx_dropped += rx_drops; 11331 s->tx_packets += tx_packets; 11332 s->tx_bytes += tx_bytes; 11333 s->tx_dropped += tx_drops; 11334 } 11335 } 11336 11337 /* ndo_get_stats64 implementation for dtstats-based accounting. 11338 * 11339 * Populate @s from dev->stats and dev->dstats. This is used internally by the 11340 * core for NETDEV_PCPU_STAT_DSTAT-type stats collection. 11341 */ 11342 static void dev_get_dstats64(const struct net_device *dev, 11343 struct rtnl_link_stats64 *s) 11344 { 11345 netdev_stats_to_stats64(s, &dev->stats); 11346 dev_fetch_dstats(s, dev->dstats); 11347 } 11348 11349 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has 11350 * all the same fields in the same order as net_device_stats, with only 11351 * the type differing, but rtnl_link_stats64 may have additional fields 11352 * at the end for newer counters. 11353 */ 11354 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64, 11355 const struct net_device_stats *netdev_stats) 11356 { 11357 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t); 11358 const atomic_long_t *src = (atomic_long_t *)netdev_stats; 11359 u64 *dst = (u64 *)stats64; 11360 11361 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64)); 11362 for (i = 0; i < n; i++) 11363 dst[i] = (unsigned long)atomic_long_read(&src[i]); 11364 /* zero out counters that only exist in rtnl_link_stats64 */ 11365 memset((char *)stats64 + n * sizeof(u64), 0, 11366 sizeof(*stats64) - n * sizeof(u64)); 11367 } 11368 EXPORT_SYMBOL(netdev_stats_to_stats64); 11369 11370 static __cold struct net_device_core_stats __percpu *netdev_core_stats_alloc( 11371 struct net_device *dev) 11372 { 11373 struct net_device_core_stats __percpu *p; 11374 11375 p = alloc_percpu_gfp(struct net_device_core_stats, 11376 GFP_ATOMIC | __GFP_NOWARN); 11377 11378 if (p && cmpxchg(&dev->core_stats, NULL, p)) 11379 free_percpu(p); 11380 11381 /* This READ_ONCE() pairs with the cmpxchg() above */ 11382 return READ_ONCE(dev->core_stats); 11383 } 11384 11385 noinline void netdev_core_stats_inc(struct net_device *dev, u32 offset) 11386 { 11387 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11388 struct net_device_core_stats __percpu *p = READ_ONCE(dev->core_stats); 11389 unsigned long __percpu *field; 11390 11391 if (unlikely(!p)) { 11392 p = netdev_core_stats_alloc(dev); 11393 if (!p) 11394 return; 11395 } 11396 11397 field = (unsigned long __percpu *)((void __percpu *)p + offset); 11398 this_cpu_inc(*field); 11399 } 11400 EXPORT_SYMBOL_GPL(netdev_core_stats_inc); 11401 11402 /** 11403 * dev_get_stats - get network device statistics 11404 * @dev: device to get statistics from 11405 * @storage: place to store stats 11406 * 11407 * Get network statistics from device. Return @storage. 11408 * The device driver may provide its own method by setting 11409 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats; 11410 * otherwise the internal statistics structure is used. 11411 */ 11412 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev, 11413 struct rtnl_link_stats64 *storage) 11414 { 11415 const struct net_device_ops *ops = dev->netdev_ops; 11416 const struct net_device_core_stats __percpu *p; 11417 11418 /* 11419 * IPv{4,6} and udp tunnels share common stat helpers and use 11420 * different stat type (NETDEV_PCPU_STAT_TSTATS vs 11421 * NETDEV_PCPU_STAT_DSTATS). Ensure the accounting is consistent. 11422 */ 11423 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_bytes) != 11424 offsetof(struct pcpu_dstats, rx_bytes)); 11425 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_packets) != 11426 offsetof(struct pcpu_dstats, rx_packets)); 11427 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_bytes) != 11428 offsetof(struct pcpu_dstats, tx_bytes)); 11429 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_packets) != 11430 offsetof(struct pcpu_dstats, tx_packets)); 11431 11432 if (ops->ndo_get_stats64) { 11433 memset(storage, 0, sizeof(*storage)); 11434 ops->ndo_get_stats64(dev, storage); 11435 } else if (ops->ndo_get_stats) { 11436 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev)); 11437 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_TSTATS) { 11438 dev_get_tstats64(dev, storage); 11439 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_DSTATS) { 11440 dev_get_dstats64(dev, storage); 11441 } else { 11442 netdev_stats_to_stats64(storage, &dev->stats); 11443 } 11444 11445 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11446 p = READ_ONCE(dev->core_stats); 11447 if (p) { 11448 const struct net_device_core_stats *core_stats; 11449 int i; 11450 11451 for_each_possible_cpu(i) { 11452 core_stats = per_cpu_ptr(p, i); 11453 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped); 11454 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped); 11455 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler); 11456 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped); 11457 } 11458 } 11459 return storage; 11460 } 11461 EXPORT_SYMBOL(dev_get_stats); 11462 11463 /** 11464 * dev_fetch_sw_netstats - get per-cpu network device statistics 11465 * @s: place to store stats 11466 * @netstats: per-cpu network stats to read from 11467 * 11468 * Read per-cpu network statistics and populate the related fields in @s. 11469 */ 11470 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s, 11471 const struct pcpu_sw_netstats __percpu *netstats) 11472 { 11473 int cpu; 11474 11475 for_each_possible_cpu(cpu) { 11476 u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 11477 const struct pcpu_sw_netstats *stats; 11478 unsigned int start; 11479 11480 stats = per_cpu_ptr(netstats, cpu); 11481 do { 11482 start = u64_stats_fetch_begin(&stats->syncp); 11483 rx_packets = u64_stats_read(&stats->rx_packets); 11484 rx_bytes = u64_stats_read(&stats->rx_bytes); 11485 tx_packets = u64_stats_read(&stats->tx_packets); 11486 tx_bytes = u64_stats_read(&stats->tx_bytes); 11487 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11488 11489 s->rx_packets += rx_packets; 11490 s->rx_bytes += rx_bytes; 11491 s->tx_packets += tx_packets; 11492 s->tx_bytes += tx_bytes; 11493 } 11494 } 11495 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats); 11496 11497 /** 11498 * dev_get_tstats64 - ndo_get_stats64 implementation 11499 * @dev: device to get statistics from 11500 * @s: place to store stats 11501 * 11502 * Populate @s from dev->stats and dev->tstats. Can be used as 11503 * ndo_get_stats64() callback. 11504 */ 11505 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s) 11506 { 11507 netdev_stats_to_stats64(s, &dev->stats); 11508 dev_fetch_sw_netstats(s, dev->tstats); 11509 } 11510 EXPORT_SYMBOL_GPL(dev_get_tstats64); 11511 11512 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev) 11513 { 11514 struct netdev_queue *queue = dev_ingress_queue(dev); 11515 11516 #ifdef CONFIG_NET_CLS_ACT 11517 if (queue) 11518 return queue; 11519 queue = kzalloc(sizeof(*queue), GFP_KERNEL); 11520 if (!queue) 11521 return NULL; 11522 netdev_init_one_queue(dev, queue, NULL); 11523 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc); 11524 RCU_INIT_POINTER(queue->qdisc_sleeping, &noop_qdisc); 11525 rcu_assign_pointer(dev->ingress_queue, queue); 11526 #endif 11527 return queue; 11528 } 11529 11530 static const struct ethtool_ops default_ethtool_ops; 11531 11532 void netdev_set_default_ethtool_ops(struct net_device *dev, 11533 const struct ethtool_ops *ops) 11534 { 11535 if (dev->ethtool_ops == &default_ethtool_ops) 11536 dev->ethtool_ops = ops; 11537 } 11538 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops); 11539 11540 /** 11541 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default 11542 * @dev: netdev to enable the IRQ coalescing on 11543 * 11544 * Sets a conservative default for SW IRQ coalescing. Users can use 11545 * sysfs attributes to override the default values. 11546 */ 11547 void netdev_sw_irq_coalesce_default_on(struct net_device *dev) 11548 { 11549 WARN_ON(dev->reg_state == NETREG_REGISTERED); 11550 11551 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) { 11552 netdev_set_gro_flush_timeout(dev, 20000); 11553 netdev_set_defer_hard_irqs(dev, 1); 11554 } 11555 } 11556 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on); 11557 11558 /** 11559 * alloc_netdev_mqs - allocate network device 11560 * @sizeof_priv: size of private data to allocate space for 11561 * @name: device name format string 11562 * @name_assign_type: origin of device name 11563 * @setup: callback to initialize device 11564 * @txqs: the number of TX subqueues to allocate 11565 * @rxqs: the number of RX subqueues to allocate 11566 * 11567 * Allocates a struct net_device with private data area for driver use 11568 * and performs basic initialization. Also allocates subqueue structs 11569 * for each queue on the device. 11570 */ 11571 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name, 11572 unsigned char name_assign_type, 11573 void (*setup)(struct net_device *), 11574 unsigned int txqs, unsigned int rxqs) 11575 { 11576 struct net_device *dev; 11577 size_t napi_config_sz; 11578 unsigned int maxqs; 11579 11580 BUG_ON(strlen(name) >= sizeof(dev->name)); 11581 11582 if (txqs < 1) { 11583 pr_err("alloc_netdev: Unable to allocate device with zero queues\n"); 11584 return NULL; 11585 } 11586 11587 if (rxqs < 1) { 11588 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n"); 11589 return NULL; 11590 } 11591 11592 maxqs = max(txqs, rxqs); 11593 11594 dev = kvzalloc(struct_size(dev, priv, sizeof_priv), 11595 GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11596 if (!dev) 11597 return NULL; 11598 11599 dev->priv_len = sizeof_priv; 11600 11601 ref_tracker_dir_init(&dev->refcnt_tracker, 128, name); 11602 #ifdef CONFIG_PCPU_DEV_REFCNT 11603 dev->pcpu_refcnt = alloc_percpu(int); 11604 if (!dev->pcpu_refcnt) 11605 goto free_dev; 11606 __dev_hold(dev); 11607 #else 11608 refcount_set(&dev->dev_refcnt, 1); 11609 #endif 11610 11611 if (dev_addr_init(dev)) 11612 goto free_pcpu; 11613 11614 dev_mc_init(dev); 11615 dev_uc_init(dev); 11616 11617 dev_net_set(dev, &init_net); 11618 11619 dev->gso_max_size = GSO_LEGACY_MAX_SIZE; 11620 dev->xdp_zc_max_segs = 1; 11621 dev->gso_max_segs = GSO_MAX_SEGS; 11622 dev->gro_max_size = GRO_LEGACY_MAX_SIZE; 11623 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE; 11624 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE; 11625 dev->tso_max_size = TSO_LEGACY_MAX_SIZE; 11626 dev->tso_max_segs = TSO_MAX_SEGS; 11627 dev->upper_level = 1; 11628 dev->lower_level = 1; 11629 #ifdef CONFIG_LOCKDEP 11630 dev->nested_level = 0; 11631 INIT_LIST_HEAD(&dev->unlink_list); 11632 #endif 11633 11634 INIT_LIST_HEAD(&dev->napi_list); 11635 INIT_LIST_HEAD(&dev->unreg_list); 11636 INIT_LIST_HEAD(&dev->close_list); 11637 INIT_LIST_HEAD(&dev->link_watch_list); 11638 INIT_LIST_HEAD(&dev->adj_list.upper); 11639 INIT_LIST_HEAD(&dev->adj_list.lower); 11640 INIT_LIST_HEAD(&dev->ptype_all); 11641 INIT_LIST_HEAD(&dev->ptype_specific); 11642 INIT_LIST_HEAD(&dev->net_notifier_list); 11643 #ifdef CONFIG_NET_SCHED 11644 hash_init(dev->qdisc_hash); 11645 #endif 11646 11647 mutex_init(&dev->lock); 11648 11649 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 11650 setup(dev); 11651 11652 if (!dev->tx_queue_len) { 11653 dev->priv_flags |= IFF_NO_QUEUE; 11654 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN; 11655 } 11656 11657 dev->num_tx_queues = txqs; 11658 dev->real_num_tx_queues = txqs; 11659 if (netif_alloc_netdev_queues(dev)) 11660 goto free_all; 11661 11662 dev->num_rx_queues = rxqs; 11663 dev->real_num_rx_queues = rxqs; 11664 if (netif_alloc_rx_queues(dev)) 11665 goto free_all; 11666 dev->ethtool = kzalloc(sizeof(*dev->ethtool), GFP_KERNEL_ACCOUNT); 11667 if (!dev->ethtool) 11668 goto free_all; 11669 11670 dev->cfg = kzalloc(sizeof(*dev->cfg), GFP_KERNEL_ACCOUNT); 11671 if (!dev->cfg) 11672 goto free_all; 11673 dev->cfg_pending = dev->cfg; 11674 11675 napi_config_sz = array_size(maxqs, sizeof(*dev->napi_config)); 11676 dev->napi_config = kvzalloc(napi_config_sz, GFP_KERNEL_ACCOUNT); 11677 if (!dev->napi_config) 11678 goto free_all; 11679 11680 strscpy(dev->name, name); 11681 dev->name_assign_type = name_assign_type; 11682 dev->group = INIT_NETDEV_GROUP; 11683 if (!dev->ethtool_ops) 11684 dev->ethtool_ops = &default_ethtool_ops; 11685 11686 nf_hook_netdev_init(dev); 11687 11688 return dev; 11689 11690 free_all: 11691 free_netdev(dev); 11692 return NULL; 11693 11694 free_pcpu: 11695 #ifdef CONFIG_PCPU_DEV_REFCNT 11696 free_percpu(dev->pcpu_refcnt); 11697 free_dev: 11698 #endif 11699 kvfree(dev); 11700 return NULL; 11701 } 11702 EXPORT_SYMBOL(alloc_netdev_mqs); 11703 11704 static void netdev_napi_exit(struct net_device *dev) 11705 { 11706 if (!list_empty(&dev->napi_list)) { 11707 struct napi_struct *p, *n; 11708 11709 netdev_lock(dev); 11710 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list) 11711 __netif_napi_del_locked(p); 11712 netdev_unlock(dev); 11713 11714 synchronize_net(); 11715 } 11716 11717 kvfree(dev->napi_config); 11718 } 11719 11720 /** 11721 * free_netdev - free network device 11722 * @dev: device 11723 * 11724 * This function does the last stage of destroying an allocated device 11725 * interface. The reference to the device object is released. If this 11726 * is the last reference then it will be freed.Must be called in process 11727 * context. 11728 */ 11729 void free_netdev(struct net_device *dev) 11730 { 11731 might_sleep(); 11732 11733 /* When called immediately after register_netdevice() failed the unwind 11734 * handling may still be dismantling the device. Handle that case by 11735 * deferring the free. 11736 */ 11737 if (dev->reg_state == NETREG_UNREGISTERING) { 11738 ASSERT_RTNL(); 11739 dev->needs_free_netdev = true; 11740 return; 11741 } 11742 11743 WARN_ON(dev->cfg != dev->cfg_pending); 11744 kfree(dev->cfg); 11745 kfree(dev->ethtool); 11746 netif_free_tx_queues(dev); 11747 netif_free_rx_queues(dev); 11748 11749 kfree(rcu_dereference_protected(dev->ingress_queue, 1)); 11750 11751 /* Flush device addresses */ 11752 dev_addr_flush(dev); 11753 11754 netdev_napi_exit(dev); 11755 11756 netif_del_cpu_rmap(dev); 11757 11758 ref_tracker_dir_exit(&dev->refcnt_tracker); 11759 #ifdef CONFIG_PCPU_DEV_REFCNT 11760 free_percpu(dev->pcpu_refcnt); 11761 dev->pcpu_refcnt = NULL; 11762 #endif 11763 free_percpu(dev->core_stats); 11764 dev->core_stats = NULL; 11765 free_percpu(dev->xdp_bulkq); 11766 dev->xdp_bulkq = NULL; 11767 11768 netdev_free_phy_link_topology(dev); 11769 11770 mutex_destroy(&dev->lock); 11771 11772 /* Compatibility with error handling in drivers */ 11773 if (dev->reg_state == NETREG_UNINITIALIZED || 11774 dev->reg_state == NETREG_DUMMY) { 11775 kvfree(dev); 11776 return; 11777 } 11778 11779 BUG_ON(dev->reg_state != NETREG_UNREGISTERED); 11780 WRITE_ONCE(dev->reg_state, NETREG_RELEASED); 11781 11782 /* will free via device release */ 11783 put_device(&dev->dev); 11784 } 11785 EXPORT_SYMBOL(free_netdev); 11786 11787 /** 11788 * alloc_netdev_dummy - Allocate and initialize a dummy net device. 11789 * @sizeof_priv: size of private data to allocate space for 11790 * 11791 * Return: the allocated net_device on success, NULL otherwise 11792 */ 11793 struct net_device *alloc_netdev_dummy(int sizeof_priv) 11794 { 11795 return alloc_netdev(sizeof_priv, "dummy#", NET_NAME_UNKNOWN, 11796 init_dummy_netdev); 11797 } 11798 EXPORT_SYMBOL_GPL(alloc_netdev_dummy); 11799 11800 /** 11801 * synchronize_net - Synchronize with packet receive processing 11802 * 11803 * Wait for packets currently being received to be done. 11804 * Does not block later packets from starting. 11805 */ 11806 void synchronize_net(void) 11807 { 11808 might_sleep(); 11809 if (from_cleanup_net() || rtnl_is_locked()) 11810 synchronize_rcu_expedited(); 11811 else 11812 synchronize_rcu(); 11813 } 11814 EXPORT_SYMBOL(synchronize_net); 11815 11816 static void netdev_rss_contexts_free(struct net_device *dev) 11817 { 11818 struct ethtool_rxfh_context *ctx; 11819 unsigned long context; 11820 11821 mutex_lock(&dev->ethtool->rss_lock); 11822 xa_for_each(&dev->ethtool->rss_ctx, context, ctx) { 11823 struct ethtool_rxfh_param rxfh; 11824 11825 rxfh.indir = ethtool_rxfh_context_indir(ctx); 11826 rxfh.key = ethtool_rxfh_context_key(ctx); 11827 rxfh.hfunc = ctx->hfunc; 11828 rxfh.input_xfrm = ctx->input_xfrm; 11829 rxfh.rss_context = context; 11830 rxfh.rss_delete = true; 11831 11832 xa_erase(&dev->ethtool->rss_ctx, context); 11833 if (dev->ethtool_ops->create_rxfh_context) 11834 dev->ethtool_ops->remove_rxfh_context(dev, ctx, 11835 context, NULL); 11836 else 11837 dev->ethtool_ops->set_rxfh(dev, &rxfh, NULL); 11838 kfree(ctx); 11839 } 11840 xa_destroy(&dev->ethtool->rss_ctx); 11841 mutex_unlock(&dev->ethtool->rss_lock); 11842 } 11843 11844 /** 11845 * unregister_netdevice_queue - remove device from the kernel 11846 * @dev: device 11847 * @head: list 11848 * 11849 * This function shuts down a device interface and removes it 11850 * from the kernel tables. 11851 * If head not NULL, device is queued to be unregistered later. 11852 * 11853 * Callers must hold the rtnl semaphore. You may want 11854 * unregister_netdev() instead of this. 11855 */ 11856 11857 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head) 11858 { 11859 ASSERT_RTNL(); 11860 11861 if (head) { 11862 list_move_tail(&dev->unreg_list, head); 11863 } else { 11864 LIST_HEAD(single); 11865 11866 list_add(&dev->unreg_list, &single); 11867 unregister_netdevice_many(&single); 11868 } 11869 } 11870 EXPORT_SYMBOL(unregister_netdevice_queue); 11871 11872 static void dev_memory_provider_uninstall(struct net_device *dev) 11873 { 11874 unsigned int i; 11875 11876 for (i = 0; i < dev->real_num_rx_queues; i++) { 11877 struct netdev_rx_queue *rxq = &dev->_rx[i]; 11878 struct pp_memory_provider_params *p = &rxq->mp_params; 11879 11880 if (p->mp_ops && p->mp_ops->uninstall) 11881 p->mp_ops->uninstall(rxq->mp_params.mp_priv, rxq); 11882 } 11883 } 11884 11885 void unregister_netdevice_many_notify(struct list_head *head, 11886 u32 portid, const struct nlmsghdr *nlh) 11887 { 11888 struct net_device *dev, *tmp; 11889 LIST_HEAD(close_head); 11890 int cnt = 0; 11891 11892 BUG_ON(dev_boot_phase); 11893 ASSERT_RTNL(); 11894 11895 if (list_empty(head)) 11896 return; 11897 11898 list_for_each_entry_safe(dev, tmp, head, unreg_list) { 11899 /* Some devices call without registering 11900 * for initialization unwind. Remove those 11901 * devices and proceed with the remaining. 11902 */ 11903 if (dev->reg_state == NETREG_UNINITIALIZED) { 11904 pr_debug("unregister_netdevice: device %s/%p never was registered\n", 11905 dev->name, dev); 11906 11907 WARN_ON(1); 11908 list_del(&dev->unreg_list); 11909 continue; 11910 } 11911 dev->dismantle = true; 11912 BUG_ON(dev->reg_state != NETREG_REGISTERED); 11913 } 11914 11915 /* If device is running, close it first. */ 11916 list_for_each_entry(dev, head, unreg_list) { 11917 list_add_tail(&dev->close_list, &close_head); 11918 netdev_lock_ops(dev); 11919 } 11920 dev_close_many(&close_head, true); 11921 11922 list_for_each_entry(dev, head, unreg_list) { 11923 netdev_unlock_ops(dev); 11924 /* And unlink it from device chain. */ 11925 unlist_netdevice(dev); 11926 netdev_lock(dev); 11927 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERING); 11928 netdev_unlock(dev); 11929 } 11930 flush_all_backlogs(); 11931 11932 synchronize_net(); 11933 11934 list_for_each_entry(dev, head, unreg_list) { 11935 struct sk_buff *skb = NULL; 11936 11937 /* Shutdown queueing discipline. */ 11938 dev_shutdown(dev); 11939 dev_tcx_uninstall(dev); 11940 netdev_lock_ops(dev); 11941 dev_xdp_uninstall(dev); 11942 netdev_unlock_ops(dev); 11943 bpf_dev_bound_netdev_unregister(dev); 11944 dev_memory_provider_uninstall(dev); 11945 11946 netdev_offload_xstats_disable_all(dev); 11947 11948 /* Notify protocols, that we are about to destroy 11949 * this device. They should clean all the things. 11950 */ 11951 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11952 11953 if (!dev->rtnl_link_ops || 11954 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 11955 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0, 11956 GFP_KERNEL, NULL, 0, 11957 portid, nlh); 11958 11959 /* 11960 * Flush the unicast and multicast chains 11961 */ 11962 dev_uc_flush(dev); 11963 dev_mc_flush(dev); 11964 11965 netdev_name_node_alt_flush(dev); 11966 netdev_name_node_free(dev->name_node); 11967 11968 netdev_rss_contexts_free(dev); 11969 11970 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11971 11972 if (dev->netdev_ops->ndo_uninit) 11973 dev->netdev_ops->ndo_uninit(dev); 11974 11975 mutex_destroy(&dev->ethtool->rss_lock); 11976 11977 net_shaper_flush_netdev(dev); 11978 11979 if (skb) 11980 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh); 11981 11982 /* Notifier chain MUST detach us all upper devices. */ 11983 WARN_ON(netdev_has_any_upper_dev(dev)); 11984 WARN_ON(netdev_has_any_lower_dev(dev)); 11985 11986 /* Remove entries from kobject tree */ 11987 netdev_unregister_kobject(dev); 11988 #ifdef CONFIG_XPS 11989 /* Remove XPS queueing entries */ 11990 netif_reset_xps_queues_gt(dev, 0); 11991 #endif 11992 } 11993 11994 synchronize_net(); 11995 11996 list_for_each_entry(dev, head, unreg_list) { 11997 netdev_put(dev, &dev->dev_registered_tracker); 11998 net_set_todo(dev); 11999 cnt++; 12000 } 12001 atomic_add(cnt, &dev_unreg_count); 12002 12003 list_del(head); 12004 } 12005 12006 /** 12007 * unregister_netdevice_many - unregister many devices 12008 * @head: list of devices 12009 * 12010 * Note: As most callers use a stack allocated list_head, 12011 * we force a list_del() to make sure stack won't be corrupted later. 12012 */ 12013 void unregister_netdevice_many(struct list_head *head) 12014 { 12015 unregister_netdevice_many_notify(head, 0, NULL); 12016 } 12017 EXPORT_SYMBOL(unregister_netdevice_many); 12018 12019 /** 12020 * unregister_netdev - remove device from the kernel 12021 * @dev: device 12022 * 12023 * This function shuts down a device interface and removes it 12024 * from the kernel tables. 12025 * 12026 * This is just a wrapper for unregister_netdevice that takes 12027 * the rtnl semaphore. In general you want to use this and not 12028 * unregister_netdevice. 12029 */ 12030 void unregister_netdev(struct net_device *dev) 12031 { 12032 rtnl_net_dev_lock(dev); 12033 unregister_netdevice(dev); 12034 rtnl_net_dev_unlock(dev); 12035 } 12036 EXPORT_SYMBOL(unregister_netdev); 12037 12038 int netif_change_net_namespace(struct net_device *dev, struct net *net, 12039 const char *pat, int new_ifindex, 12040 struct netlink_ext_ack *extack) 12041 { 12042 struct netdev_name_node *name_node; 12043 struct net *net_old = dev_net(dev); 12044 char new_name[IFNAMSIZ] = {}; 12045 int err, new_nsid; 12046 12047 ASSERT_RTNL(); 12048 12049 /* Don't allow namespace local devices to be moved. */ 12050 err = -EINVAL; 12051 if (dev->netns_immutable) { 12052 NL_SET_ERR_MSG(extack, "The interface netns is immutable"); 12053 goto out; 12054 } 12055 12056 /* Ensure the device has been registered */ 12057 if (dev->reg_state != NETREG_REGISTERED) { 12058 NL_SET_ERR_MSG(extack, "The interface isn't registered"); 12059 goto out; 12060 } 12061 12062 /* Get out if there is nothing todo */ 12063 err = 0; 12064 if (net_eq(net_old, net)) 12065 goto out; 12066 12067 /* Pick the destination device name, and ensure 12068 * we can use it in the destination network namespace. 12069 */ 12070 err = -EEXIST; 12071 if (netdev_name_in_use(net, dev->name)) { 12072 /* We get here if we can't use the current device name */ 12073 if (!pat) { 12074 NL_SET_ERR_MSG(extack, 12075 "An interface with the same name exists in the target netns"); 12076 goto out; 12077 } 12078 err = dev_prep_valid_name(net, dev, pat, new_name, EEXIST); 12079 if (err < 0) { 12080 NL_SET_ERR_MSG_FMT(extack, 12081 "Unable to use '%s' for the new interface name in the target netns", 12082 pat); 12083 goto out; 12084 } 12085 } 12086 /* Check that none of the altnames conflicts. */ 12087 err = -EEXIST; 12088 netdev_for_each_altname(dev, name_node) { 12089 if (netdev_name_in_use(net, name_node->name)) { 12090 NL_SET_ERR_MSG_FMT(extack, 12091 "An interface with the altname %s exists in the target netns", 12092 name_node->name); 12093 goto out; 12094 } 12095 } 12096 12097 /* Check that new_ifindex isn't used yet. */ 12098 if (new_ifindex) { 12099 err = dev_index_reserve(net, new_ifindex); 12100 if (err < 0) { 12101 NL_SET_ERR_MSG_FMT(extack, 12102 "The ifindex %d is not available in the target netns", 12103 new_ifindex); 12104 goto out; 12105 } 12106 } else { 12107 /* If there is an ifindex conflict assign a new one */ 12108 err = dev_index_reserve(net, dev->ifindex); 12109 if (err == -EBUSY) 12110 err = dev_index_reserve(net, 0); 12111 if (err < 0) { 12112 NL_SET_ERR_MSG(extack, 12113 "Unable to allocate a new ifindex in the target netns"); 12114 goto out; 12115 } 12116 new_ifindex = err; 12117 } 12118 12119 /* 12120 * And now a mini version of register_netdevice unregister_netdevice. 12121 */ 12122 12123 /* If device is running close it first. */ 12124 netif_close(dev); 12125 12126 /* And unlink it from device chain */ 12127 unlist_netdevice(dev); 12128 12129 synchronize_net(); 12130 12131 /* Shutdown queueing discipline. */ 12132 dev_shutdown(dev); 12133 12134 /* Notify protocols, that we are about to destroy 12135 * this device. They should clean all the things. 12136 * 12137 * Note that dev->reg_state stays at NETREG_REGISTERED. 12138 * This is wanted because this way 8021q and macvlan know 12139 * the device is just moving and can keep their slaves up. 12140 */ 12141 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 12142 rcu_barrier(); 12143 12144 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL); 12145 12146 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid, 12147 new_ifindex); 12148 12149 /* 12150 * Flush the unicast and multicast chains 12151 */ 12152 dev_uc_flush(dev); 12153 dev_mc_flush(dev); 12154 12155 /* Send a netdev-removed uevent to the old namespace */ 12156 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE); 12157 netdev_adjacent_del_links(dev); 12158 12159 /* Move per-net netdevice notifiers that are following the netdevice */ 12160 move_netdevice_notifiers_dev_net(dev, net); 12161 12162 /* Actually switch the network namespace */ 12163 dev_net_set(dev, net); 12164 dev->ifindex = new_ifindex; 12165 12166 if (new_name[0]) { 12167 /* Rename the netdev to prepared name */ 12168 write_seqlock_bh(&netdev_rename_lock); 12169 strscpy(dev->name, new_name, IFNAMSIZ); 12170 write_sequnlock_bh(&netdev_rename_lock); 12171 } 12172 12173 /* Fixup kobjects */ 12174 dev_set_uevent_suppress(&dev->dev, 1); 12175 err = device_rename(&dev->dev, dev->name); 12176 dev_set_uevent_suppress(&dev->dev, 0); 12177 WARN_ON(err); 12178 12179 /* Send a netdev-add uevent to the new namespace */ 12180 kobject_uevent(&dev->dev.kobj, KOBJ_ADD); 12181 netdev_adjacent_add_links(dev); 12182 12183 /* Adapt owner in case owning user namespace of target network 12184 * namespace is different from the original one. 12185 */ 12186 err = netdev_change_owner(dev, net_old, net); 12187 WARN_ON(err); 12188 12189 /* Add the device back in the hashes */ 12190 list_netdevice(dev); 12191 12192 /* Notify protocols, that a new device appeared. */ 12193 call_netdevice_notifiers(NETDEV_REGISTER, dev); 12194 12195 /* 12196 * Prevent userspace races by waiting until the network 12197 * device is fully setup before sending notifications. 12198 */ 12199 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 12200 12201 synchronize_net(); 12202 err = 0; 12203 out: 12204 return err; 12205 } 12206 12207 static int dev_cpu_dead(unsigned int oldcpu) 12208 { 12209 struct sk_buff **list_skb; 12210 struct sk_buff *skb; 12211 unsigned int cpu; 12212 struct softnet_data *sd, *oldsd, *remsd = NULL; 12213 12214 local_irq_disable(); 12215 cpu = smp_processor_id(); 12216 sd = &per_cpu(softnet_data, cpu); 12217 oldsd = &per_cpu(softnet_data, oldcpu); 12218 12219 /* Find end of our completion_queue. */ 12220 list_skb = &sd->completion_queue; 12221 while (*list_skb) 12222 list_skb = &(*list_skb)->next; 12223 /* Append completion queue from offline CPU. */ 12224 *list_skb = oldsd->completion_queue; 12225 oldsd->completion_queue = NULL; 12226 12227 /* Append output queue from offline CPU. */ 12228 if (oldsd->output_queue) { 12229 *sd->output_queue_tailp = oldsd->output_queue; 12230 sd->output_queue_tailp = oldsd->output_queue_tailp; 12231 oldsd->output_queue = NULL; 12232 oldsd->output_queue_tailp = &oldsd->output_queue; 12233 } 12234 /* Append NAPI poll list from offline CPU, with one exception : 12235 * process_backlog() must be called by cpu owning percpu backlog. 12236 * We properly handle process_queue & input_pkt_queue later. 12237 */ 12238 while (!list_empty(&oldsd->poll_list)) { 12239 struct napi_struct *napi = list_first_entry(&oldsd->poll_list, 12240 struct napi_struct, 12241 poll_list); 12242 12243 list_del_init(&napi->poll_list); 12244 if (napi->poll == process_backlog) 12245 napi->state &= NAPIF_STATE_THREADED; 12246 else 12247 ____napi_schedule(sd, napi); 12248 } 12249 12250 raise_softirq_irqoff(NET_TX_SOFTIRQ); 12251 local_irq_enable(); 12252 12253 if (!use_backlog_threads()) { 12254 #ifdef CONFIG_RPS 12255 remsd = oldsd->rps_ipi_list; 12256 oldsd->rps_ipi_list = NULL; 12257 #endif 12258 /* send out pending IPI's on offline CPU */ 12259 net_rps_send_ipi(remsd); 12260 } 12261 12262 /* Process offline CPU's input_pkt_queue */ 12263 while ((skb = __skb_dequeue(&oldsd->process_queue))) { 12264 netif_rx(skb); 12265 rps_input_queue_head_incr(oldsd); 12266 } 12267 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) { 12268 netif_rx(skb); 12269 rps_input_queue_head_incr(oldsd); 12270 } 12271 12272 return 0; 12273 } 12274 12275 /** 12276 * netdev_increment_features - increment feature set by one 12277 * @all: current feature set 12278 * @one: new feature set 12279 * @mask: mask feature set 12280 * 12281 * Computes a new feature set after adding a device with feature set 12282 * @one to the master device with current feature set @all. Will not 12283 * enable anything that is off in @mask. Returns the new feature set. 12284 */ 12285 netdev_features_t netdev_increment_features(netdev_features_t all, 12286 netdev_features_t one, netdev_features_t mask) 12287 { 12288 if (mask & NETIF_F_HW_CSUM) 12289 mask |= NETIF_F_CSUM_MASK; 12290 mask |= NETIF_F_VLAN_CHALLENGED; 12291 12292 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask; 12293 all &= one | ~NETIF_F_ALL_FOR_ALL; 12294 12295 /* If one device supports hw checksumming, set for all. */ 12296 if (all & NETIF_F_HW_CSUM) 12297 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM); 12298 12299 return all; 12300 } 12301 EXPORT_SYMBOL(netdev_increment_features); 12302 12303 static struct hlist_head * __net_init netdev_create_hash(void) 12304 { 12305 int i; 12306 struct hlist_head *hash; 12307 12308 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL); 12309 if (hash != NULL) 12310 for (i = 0; i < NETDEV_HASHENTRIES; i++) 12311 INIT_HLIST_HEAD(&hash[i]); 12312 12313 return hash; 12314 } 12315 12316 /* Initialize per network namespace state */ 12317 static int __net_init netdev_init(struct net *net) 12318 { 12319 BUILD_BUG_ON(GRO_HASH_BUCKETS > 12320 BITS_PER_BYTE * sizeof_field(struct gro_node, bitmask)); 12321 12322 INIT_LIST_HEAD(&net->dev_base_head); 12323 12324 net->dev_name_head = netdev_create_hash(); 12325 if (net->dev_name_head == NULL) 12326 goto err_name; 12327 12328 net->dev_index_head = netdev_create_hash(); 12329 if (net->dev_index_head == NULL) 12330 goto err_idx; 12331 12332 xa_init_flags(&net->dev_by_index, XA_FLAGS_ALLOC1); 12333 12334 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain); 12335 12336 return 0; 12337 12338 err_idx: 12339 kfree(net->dev_name_head); 12340 err_name: 12341 return -ENOMEM; 12342 } 12343 12344 /** 12345 * netdev_drivername - network driver for the device 12346 * @dev: network device 12347 * 12348 * Determine network driver for device. 12349 */ 12350 const char *netdev_drivername(const struct net_device *dev) 12351 { 12352 const struct device_driver *driver; 12353 const struct device *parent; 12354 const char *empty = ""; 12355 12356 parent = dev->dev.parent; 12357 if (!parent) 12358 return empty; 12359 12360 driver = parent->driver; 12361 if (driver && driver->name) 12362 return driver->name; 12363 return empty; 12364 } 12365 12366 static void __netdev_printk(const char *level, const struct net_device *dev, 12367 struct va_format *vaf) 12368 { 12369 if (dev && dev->dev.parent) { 12370 dev_printk_emit(level[1] - '0', 12371 dev->dev.parent, 12372 "%s %s %s%s: %pV", 12373 dev_driver_string(dev->dev.parent), 12374 dev_name(dev->dev.parent), 12375 netdev_name(dev), netdev_reg_state(dev), 12376 vaf); 12377 } else if (dev) { 12378 printk("%s%s%s: %pV", 12379 level, netdev_name(dev), netdev_reg_state(dev), vaf); 12380 } else { 12381 printk("%s(NULL net_device): %pV", level, vaf); 12382 } 12383 } 12384 12385 void netdev_printk(const char *level, const struct net_device *dev, 12386 const char *format, ...) 12387 { 12388 struct va_format vaf; 12389 va_list args; 12390 12391 va_start(args, format); 12392 12393 vaf.fmt = format; 12394 vaf.va = &args; 12395 12396 __netdev_printk(level, dev, &vaf); 12397 12398 va_end(args); 12399 } 12400 EXPORT_SYMBOL(netdev_printk); 12401 12402 #define define_netdev_printk_level(func, level) \ 12403 void func(const struct net_device *dev, const char *fmt, ...) \ 12404 { \ 12405 struct va_format vaf; \ 12406 va_list args; \ 12407 \ 12408 va_start(args, fmt); \ 12409 \ 12410 vaf.fmt = fmt; \ 12411 vaf.va = &args; \ 12412 \ 12413 __netdev_printk(level, dev, &vaf); \ 12414 \ 12415 va_end(args); \ 12416 } \ 12417 EXPORT_SYMBOL(func); 12418 12419 define_netdev_printk_level(netdev_emerg, KERN_EMERG); 12420 define_netdev_printk_level(netdev_alert, KERN_ALERT); 12421 define_netdev_printk_level(netdev_crit, KERN_CRIT); 12422 define_netdev_printk_level(netdev_err, KERN_ERR); 12423 define_netdev_printk_level(netdev_warn, KERN_WARNING); 12424 define_netdev_printk_level(netdev_notice, KERN_NOTICE); 12425 define_netdev_printk_level(netdev_info, KERN_INFO); 12426 12427 static void __net_exit netdev_exit(struct net *net) 12428 { 12429 kfree(net->dev_name_head); 12430 kfree(net->dev_index_head); 12431 xa_destroy(&net->dev_by_index); 12432 if (net != &init_net) 12433 WARN_ON_ONCE(!list_empty(&net->dev_base_head)); 12434 } 12435 12436 static struct pernet_operations __net_initdata netdev_net_ops = { 12437 .init = netdev_init, 12438 .exit = netdev_exit, 12439 }; 12440 12441 static void __net_exit default_device_exit_net(struct net *net) 12442 { 12443 struct netdev_name_node *name_node, *tmp; 12444 struct net_device *dev, *aux; 12445 /* 12446 * Push all migratable network devices back to the 12447 * initial network namespace 12448 */ 12449 ASSERT_RTNL(); 12450 for_each_netdev_safe(net, dev, aux) { 12451 int err; 12452 char fb_name[IFNAMSIZ]; 12453 12454 /* Ignore unmoveable devices (i.e. loopback) */ 12455 if (dev->netns_immutable) 12456 continue; 12457 12458 /* Leave virtual devices for the generic cleanup */ 12459 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund) 12460 continue; 12461 12462 /* Push remaining network devices to init_net */ 12463 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex); 12464 if (netdev_name_in_use(&init_net, fb_name)) 12465 snprintf(fb_name, IFNAMSIZ, "dev%%d"); 12466 12467 netdev_for_each_altname_safe(dev, name_node, tmp) 12468 if (netdev_name_in_use(&init_net, name_node->name)) 12469 __netdev_name_node_alt_destroy(name_node); 12470 12471 err = dev_change_net_namespace(dev, &init_net, fb_name); 12472 if (err) { 12473 pr_emerg("%s: failed to move %s to init_net: %d\n", 12474 __func__, dev->name, err); 12475 BUG(); 12476 } 12477 } 12478 } 12479 12480 static void __net_exit default_device_exit_batch(struct list_head *net_list) 12481 { 12482 /* At exit all network devices most be removed from a network 12483 * namespace. Do this in the reverse order of registration. 12484 * Do this across as many network namespaces as possible to 12485 * improve batching efficiency. 12486 */ 12487 struct net_device *dev; 12488 struct net *net; 12489 LIST_HEAD(dev_kill_list); 12490 12491 rtnl_lock(); 12492 list_for_each_entry(net, net_list, exit_list) { 12493 default_device_exit_net(net); 12494 cond_resched(); 12495 } 12496 12497 list_for_each_entry(net, net_list, exit_list) { 12498 for_each_netdev_reverse(net, dev) { 12499 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink) 12500 dev->rtnl_link_ops->dellink(dev, &dev_kill_list); 12501 else 12502 unregister_netdevice_queue(dev, &dev_kill_list); 12503 } 12504 } 12505 unregister_netdevice_many(&dev_kill_list); 12506 rtnl_unlock(); 12507 } 12508 12509 static struct pernet_operations __net_initdata default_device_ops = { 12510 .exit_batch = default_device_exit_batch, 12511 }; 12512 12513 static void __init net_dev_struct_check(void) 12514 { 12515 /* TX read-mostly hotpath */ 12516 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, priv_flags_fast); 12517 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, netdev_ops); 12518 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, header_ops); 12519 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, _tx); 12520 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, real_num_tx_queues); 12521 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_size); 12522 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_ipv4_max_size); 12523 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_segs); 12524 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_partial_features); 12525 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, num_tc); 12526 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, mtu); 12527 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, needed_headroom); 12528 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tc_to_txq); 12529 #ifdef CONFIG_XPS 12530 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, xps_maps); 12531 #endif 12532 #ifdef CONFIG_NETFILTER_EGRESS 12533 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, nf_hooks_egress); 12534 #endif 12535 #ifdef CONFIG_NET_XGRESS 12536 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tcx_egress); 12537 #endif 12538 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_tx, 160); 12539 12540 /* TXRX read-mostly hotpath */ 12541 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, lstats); 12542 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, state); 12543 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, flags); 12544 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, hard_header_len); 12545 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, features); 12546 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, ip6_ptr); 12547 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_txrx, 46); 12548 12549 /* RX read-mostly hotpath */ 12550 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ptype_specific); 12551 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ifindex); 12552 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, real_num_rx_queues); 12553 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, _rx); 12554 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_max_size); 12555 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_ipv4_max_size); 12556 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler); 12557 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler_data); 12558 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, nd_net); 12559 #ifdef CONFIG_NETPOLL 12560 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, npinfo); 12561 #endif 12562 #ifdef CONFIG_NET_XGRESS 12563 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, tcx_ingress); 12564 #endif 12565 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_rx, 92); 12566 } 12567 12568 /* 12569 * Initialize the DEV module. At boot time this walks the device list and 12570 * unhooks any devices that fail to initialise (normally hardware not 12571 * present) and leaves us with a valid list of present and active devices. 12572 * 12573 */ 12574 12575 /* We allocate 256 pages for each CPU if PAGE_SHIFT is 12 */ 12576 #define SYSTEM_PERCPU_PAGE_POOL_SIZE ((1 << 20) / PAGE_SIZE) 12577 12578 static int net_page_pool_create(int cpuid) 12579 { 12580 #if IS_ENABLED(CONFIG_PAGE_POOL) 12581 struct page_pool_params page_pool_params = { 12582 .pool_size = SYSTEM_PERCPU_PAGE_POOL_SIZE, 12583 .flags = PP_FLAG_SYSTEM_POOL, 12584 .nid = cpu_to_mem(cpuid), 12585 }; 12586 struct page_pool *pp_ptr; 12587 int err; 12588 12589 pp_ptr = page_pool_create_percpu(&page_pool_params, cpuid); 12590 if (IS_ERR(pp_ptr)) 12591 return -ENOMEM; 12592 12593 err = xdp_reg_page_pool(pp_ptr); 12594 if (err) { 12595 page_pool_destroy(pp_ptr); 12596 return err; 12597 } 12598 12599 per_cpu(system_page_pool, cpuid) = pp_ptr; 12600 #endif 12601 return 0; 12602 } 12603 12604 static int backlog_napi_should_run(unsigned int cpu) 12605 { 12606 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12607 struct napi_struct *napi = &sd->backlog; 12608 12609 return test_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 12610 } 12611 12612 static void run_backlog_napi(unsigned int cpu) 12613 { 12614 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12615 12616 napi_threaded_poll_loop(&sd->backlog); 12617 } 12618 12619 static void backlog_napi_setup(unsigned int cpu) 12620 { 12621 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12622 struct napi_struct *napi = &sd->backlog; 12623 12624 napi->thread = this_cpu_read(backlog_napi); 12625 set_bit(NAPI_STATE_THREADED, &napi->state); 12626 } 12627 12628 static struct smp_hotplug_thread backlog_threads = { 12629 .store = &backlog_napi, 12630 .thread_should_run = backlog_napi_should_run, 12631 .thread_fn = run_backlog_napi, 12632 .thread_comm = "backlog_napi/%u", 12633 .setup = backlog_napi_setup, 12634 }; 12635 12636 /* 12637 * This is called single threaded during boot, so no need 12638 * to take the rtnl semaphore. 12639 */ 12640 static int __init net_dev_init(void) 12641 { 12642 int i, rc = -ENOMEM; 12643 12644 BUG_ON(!dev_boot_phase); 12645 12646 net_dev_struct_check(); 12647 12648 if (dev_proc_init()) 12649 goto out; 12650 12651 if (netdev_kobject_init()) 12652 goto out; 12653 12654 for (i = 0; i < PTYPE_HASH_SIZE; i++) 12655 INIT_LIST_HEAD(&ptype_base[i]); 12656 12657 if (register_pernet_subsys(&netdev_net_ops)) 12658 goto out; 12659 12660 /* 12661 * Initialise the packet receive queues. 12662 */ 12663 12664 flush_backlogs_fallback = flush_backlogs_alloc(); 12665 if (!flush_backlogs_fallback) 12666 goto out; 12667 12668 for_each_possible_cpu(i) { 12669 struct softnet_data *sd = &per_cpu(softnet_data, i); 12670 12671 skb_queue_head_init(&sd->input_pkt_queue); 12672 skb_queue_head_init(&sd->process_queue); 12673 #ifdef CONFIG_XFRM_OFFLOAD 12674 skb_queue_head_init(&sd->xfrm_backlog); 12675 #endif 12676 INIT_LIST_HEAD(&sd->poll_list); 12677 sd->output_queue_tailp = &sd->output_queue; 12678 #ifdef CONFIG_RPS 12679 INIT_CSD(&sd->csd, rps_trigger_softirq, sd); 12680 sd->cpu = i; 12681 #endif 12682 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd); 12683 spin_lock_init(&sd->defer_lock); 12684 12685 gro_init(&sd->backlog.gro); 12686 sd->backlog.poll = process_backlog; 12687 sd->backlog.weight = weight_p; 12688 INIT_LIST_HEAD(&sd->backlog.poll_list); 12689 12690 if (net_page_pool_create(i)) 12691 goto out; 12692 } 12693 if (use_backlog_threads()) 12694 smpboot_register_percpu_thread(&backlog_threads); 12695 12696 dev_boot_phase = 0; 12697 12698 /* The loopback device is special if any other network devices 12699 * is present in a network namespace the loopback device must 12700 * be present. Since we now dynamically allocate and free the 12701 * loopback device ensure this invariant is maintained by 12702 * keeping the loopback device as the first device on the 12703 * list of network devices. Ensuring the loopback devices 12704 * is the first device that appears and the last network device 12705 * that disappears. 12706 */ 12707 if (register_pernet_device(&loopback_net_ops)) 12708 goto out; 12709 12710 if (register_pernet_device(&default_device_ops)) 12711 goto out; 12712 12713 open_softirq(NET_TX_SOFTIRQ, net_tx_action); 12714 open_softirq(NET_RX_SOFTIRQ, net_rx_action); 12715 12716 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead", 12717 NULL, dev_cpu_dead); 12718 WARN_ON(rc < 0); 12719 rc = 0; 12720 12721 /* avoid static key IPIs to isolated CPUs */ 12722 if (housekeeping_enabled(HK_TYPE_MISC)) 12723 net_enable_timestamp(); 12724 out: 12725 if (rc < 0) { 12726 for_each_possible_cpu(i) { 12727 struct page_pool *pp_ptr; 12728 12729 pp_ptr = per_cpu(system_page_pool, i); 12730 if (!pp_ptr) 12731 continue; 12732 12733 xdp_unreg_page_pool(pp_ptr); 12734 page_pool_destroy(pp_ptr); 12735 per_cpu(system_page_pool, i) = NULL; 12736 } 12737 } 12738 12739 return rc; 12740 } 12741 12742 subsys_initcall(net_dev_init); 12743