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