1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * NET3 Protocol independent device support routines. 4 * 5 * Derived from the non IP parts of dev.c 1.0.19 6 * Authors: Ross Biro 7 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 8 * Mark Evans, <evansmp@uhura.aston.ac.uk> 9 * 10 * Additional Authors: 11 * Florian la Roche <rzsfl@rz.uni-sb.de> 12 * Alan Cox <gw4pts@gw4pts.ampr.org> 13 * David Hinds <dahinds@users.sourceforge.net> 14 * Alexey Kuznetsov <kuznet@ms2.inr.ac.ru> 15 * Adam Sulmicki <adam@cfar.umd.edu> 16 * Pekka Riikonen <priikone@poesidon.pspt.fi> 17 * 18 * Changes: 19 * D.J. Barrow : Fixed bug where dev->refcnt gets set 20 * to 2 if register_netdev gets called 21 * before net_dev_init & also removed a 22 * few lines of code in the process. 23 * Alan Cox : device private ioctl copies fields back. 24 * Alan Cox : Transmit queue code does relevant 25 * stunts to keep the queue safe. 26 * Alan Cox : Fixed double lock. 27 * Alan Cox : Fixed promisc NULL pointer trap 28 * ???????? : Support the full private ioctl range 29 * Alan Cox : Moved ioctl permission check into 30 * drivers 31 * Tim Kordas : SIOCADDMULTI/SIOCDELMULTI 32 * Alan Cox : 100 backlog just doesn't cut it when 33 * you start doing multicast video 8) 34 * Alan Cox : Rewrote net_bh and list manager. 35 * Alan Cox : Fix ETH_P_ALL echoback lengths. 36 * Alan Cox : Took out transmit every packet pass 37 * Saved a few bytes in the ioctl handler 38 * Alan Cox : Network driver sets packet type before 39 * calling netif_rx. Saves a function 40 * call a packet. 41 * Alan Cox : Hashed net_bh() 42 * Richard Kooijman: Timestamp fixes. 43 * Alan Cox : Wrong field in SIOCGIFDSTADDR 44 * Alan Cox : Device lock protection. 45 * Alan Cox : Fixed nasty side effect of device close 46 * changes. 47 * Rudi Cilibrasi : Pass the right thing to 48 * set_mac_address() 49 * Dave Miller : 32bit quantity for the device lock to 50 * make it work out on a Sparc. 51 * Bjorn Ekwall : Added KERNELD hack. 52 * Alan Cox : Cleaned up the backlog initialise. 53 * Craig Metz : SIOCGIFCONF fix if space for under 54 * 1 device. 55 * Thomas Bogendoerfer : Return ENODEV for dev_open, if there 56 * is no device open function. 57 * Andi Kleen : Fix error reporting for SIOCGIFCONF 58 * Michael Chastain : Fix signed/unsigned for SIOCGIFCONF 59 * Cyrus Durgin : Cleaned for KMOD 60 * Adam Sulmicki : Bug Fix : Network Device Unload 61 * A network device unload needs to purge 62 * the backlog queue. 63 * Paul Rusty Russell : SIOCSIFNAME 64 * Pekka Riikonen : Netdev boot-time settings code 65 * Andrew Morton : Make unregister_netdevice wait 66 * indefinitely on dev->refcnt 67 * J Hadi Salim : - Backlog queue sampling 68 * - netif_rx() feedback 69 */ 70 71 #include <linux/uaccess.h> 72 #include <linux/bitmap.h> 73 #include <linux/capability.h> 74 #include <linux/cpu.h> 75 #include <linux/types.h> 76 #include <linux/kernel.h> 77 #include <linux/hash.h> 78 #include <linux/slab.h> 79 #include <linux/sched.h> 80 #include <linux/sched/isolation.h> 81 #include <linux/sched/mm.h> 82 #include <linux/smpboot.h> 83 #include <linux/mutex.h> 84 #include <linux/rwsem.h> 85 #include <linux/string.h> 86 #include <linux/mm.h> 87 #include <linux/socket.h> 88 #include <linux/sockios.h> 89 #include <linux/errno.h> 90 #include <linux/interrupt.h> 91 #include <linux/if_ether.h> 92 #include <linux/netdevice.h> 93 #include <linux/etherdevice.h> 94 #include <linux/ethtool.h> 95 #include <linux/ethtool_netlink.h> 96 #include <linux/skbuff.h> 97 #include <linux/kthread.h> 98 #include <linux/bpf.h> 99 #include <linux/bpf_trace.h> 100 #include <net/net_namespace.h> 101 #include <net/sock.h> 102 #include <net/busy_poll.h> 103 #include <linux/rtnetlink.h> 104 #include <linux/stat.h> 105 #include <net/dsa.h> 106 #include <net/dst.h> 107 #include <net/dst_metadata.h> 108 #include <net/gro.h> 109 #include <net/netdev_queues.h> 110 #include <net/pkt_sched.h> 111 #include <net/pkt_cls.h> 112 #include <net/checksum.h> 113 #include <net/xfrm.h> 114 #include <net/tcx.h> 115 #include <linux/highmem.h> 116 #include <linux/init.h> 117 #include <linux/module.h> 118 #include <linux/netpoll.h> 119 #include <linux/rcupdate.h> 120 #include <linux/delay.h> 121 #include <net/iw_handler.h> 122 #include <asm/current.h> 123 #include <linux/audit.h> 124 #include <linux/dmaengine.h> 125 #include <linux/err.h> 126 #include <linux/ctype.h> 127 #include <linux/if_arp.h> 128 #include <linux/if_vlan.h> 129 #include <linux/ip.h> 130 #include <net/ip.h> 131 #include <net/mpls.h> 132 #include <linux/ipv6.h> 133 #include <linux/in.h> 134 #include <linux/jhash.h> 135 #include <linux/random.h> 136 #include <trace/events/napi.h> 137 #include <trace/events/net.h> 138 #include <trace/events/skb.h> 139 #include <trace/events/qdisc.h> 140 #include <trace/events/xdp.h> 141 #include <linux/inetdevice.h> 142 #include <linux/cpu_rmap.h> 143 #include <linux/static_key.h> 144 #include <linux/hashtable.h> 145 #include <linux/vmalloc.h> 146 #include <linux/if_macvlan.h> 147 #include <linux/errqueue.h> 148 #include <linux/hrtimer.h> 149 #include <linux/netfilter_netdev.h> 150 #include <linux/crash_dump.h> 151 #include <linux/sctp.h> 152 #include <net/udp_tunnel.h> 153 #include <linux/net_namespace.h> 154 #include <linux/indirect_call_wrapper.h> 155 #include <net/devlink.h> 156 #include <linux/pm_runtime.h> 157 #include <linux/prandom.h> 158 #include <linux/once_lite.h> 159 #include <net/netdev_lock.h> 160 #include <net/netdev_rx_queue.h> 161 #include <net/page_pool/types.h> 162 #include <net/page_pool/helpers.h> 163 #include <net/page_pool/memory_provider.h> 164 #include <net/rps.h> 165 #include <linux/phy_link_topology.h> 166 167 #include "dev.h" 168 #include "devmem.h" 169 #include "net-sysfs.h" 170 171 static DEFINE_SPINLOCK(ptype_lock); 172 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly; 173 174 static int netif_rx_internal(struct sk_buff *skb); 175 static int call_netdevice_notifiers_extack(unsigned long val, 176 struct net_device *dev, 177 struct netlink_ext_ack *extack); 178 179 static DEFINE_MUTEX(ifalias_mutex); 180 181 /* protects napi_hash addition/deletion and napi_gen_id */ 182 static DEFINE_SPINLOCK(napi_hash_lock); 183 184 static unsigned int napi_gen_id = NR_CPUS; 185 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8); 186 187 static inline void dev_base_seq_inc(struct net *net) 188 { 189 unsigned int val = net->dev_base_seq + 1; 190 191 WRITE_ONCE(net->dev_base_seq, val ?: 1); 192 } 193 194 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name) 195 { 196 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ)); 197 198 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)]; 199 } 200 201 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex) 202 { 203 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)]; 204 } 205 206 #ifndef CONFIG_PREEMPT_RT 207 208 static DEFINE_STATIC_KEY_FALSE(use_backlog_threads_key); 209 210 static int __init setup_backlog_napi_threads(char *arg) 211 { 212 static_branch_enable(&use_backlog_threads_key); 213 return 0; 214 } 215 early_param("thread_backlog_napi", setup_backlog_napi_threads); 216 217 static bool use_backlog_threads(void) 218 { 219 return static_branch_unlikely(&use_backlog_threads_key); 220 } 221 222 #else 223 224 static bool use_backlog_threads(void) 225 { 226 return true; 227 } 228 229 #endif 230 231 static inline void backlog_lock_irq_save(struct softnet_data *sd, 232 unsigned long *flags) 233 { 234 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 235 spin_lock_irqsave(&sd->input_pkt_queue.lock, *flags); 236 else 237 local_irq_save(*flags); 238 } 239 240 static inline void backlog_lock_irq_disable(struct softnet_data *sd) 241 { 242 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 243 spin_lock_irq(&sd->input_pkt_queue.lock); 244 else 245 local_irq_disable(); 246 } 247 248 static inline void backlog_unlock_irq_restore(struct softnet_data *sd, 249 unsigned long *flags) 250 { 251 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 252 spin_unlock_irqrestore(&sd->input_pkt_queue.lock, *flags); 253 else 254 local_irq_restore(*flags); 255 } 256 257 static inline void backlog_unlock_irq_enable(struct softnet_data *sd) 258 { 259 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 260 spin_unlock_irq(&sd->input_pkt_queue.lock); 261 else 262 local_irq_enable(); 263 } 264 265 static struct netdev_name_node *netdev_name_node_alloc(struct net_device *dev, 266 const char *name) 267 { 268 struct netdev_name_node *name_node; 269 270 name_node = kmalloc(sizeof(*name_node), GFP_KERNEL); 271 if (!name_node) 272 return NULL; 273 INIT_HLIST_NODE(&name_node->hlist); 274 name_node->dev = dev; 275 name_node->name = name; 276 return name_node; 277 } 278 279 static struct netdev_name_node * 280 netdev_name_node_head_alloc(struct net_device *dev) 281 { 282 struct netdev_name_node *name_node; 283 284 name_node = netdev_name_node_alloc(dev, dev->name); 285 if (!name_node) 286 return NULL; 287 INIT_LIST_HEAD(&name_node->list); 288 return name_node; 289 } 290 291 static void netdev_name_node_free(struct netdev_name_node *name_node) 292 { 293 kfree(name_node); 294 } 295 296 static void netdev_name_node_add(struct net *net, 297 struct netdev_name_node *name_node) 298 { 299 hlist_add_head_rcu(&name_node->hlist, 300 dev_name_hash(net, name_node->name)); 301 } 302 303 static void netdev_name_node_del(struct netdev_name_node *name_node) 304 { 305 hlist_del_rcu(&name_node->hlist); 306 } 307 308 static struct netdev_name_node *netdev_name_node_lookup(struct net *net, 309 const char *name) 310 { 311 struct hlist_head *head = dev_name_hash(net, name); 312 struct netdev_name_node *name_node; 313 314 hlist_for_each_entry(name_node, head, hlist) 315 if (!strcmp(name_node->name, name)) 316 return name_node; 317 return NULL; 318 } 319 320 static struct netdev_name_node *netdev_name_node_lookup_rcu(struct net *net, 321 const char *name) 322 { 323 struct hlist_head *head = dev_name_hash(net, name); 324 struct netdev_name_node *name_node; 325 326 hlist_for_each_entry_rcu(name_node, head, hlist) 327 if (!strcmp(name_node->name, name)) 328 return name_node; 329 return NULL; 330 } 331 332 bool netdev_name_in_use(struct net *net, const char *name) 333 { 334 return netdev_name_node_lookup(net, name); 335 } 336 EXPORT_SYMBOL(netdev_name_in_use); 337 338 int netdev_name_node_alt_create(struct net_device *dev, const char *name) 339 { 340 struct netdev_name_node *name_node; 341 struct net *net = dev_net(dev); 342 343 name_node = netdev_name_node_lookup(net, name); 344 if (name_node) 345 return -EEXIST; 346 name_node = netdev_name_node_alloc(dev, name); 347 if (!name_node) 348 return -ENOMEM; 349 netdev_name_node_add(net, name_node); 350 /* The node that holds dev->name acts as a head of per-device list. */ 351 list_add_tail_rcu(&name_node->list, &dev->name_node->list); 352 353 return 0; 354 } 355 356 static void netdev_name_node_alt_free(struct rcu_head *head) 357 { 358 struct netdev_name_node *name_node = 359 container_of(head, struct netdev_name_node, rcu); 360 361 kfree(name_node->name); 362 netdev_name_node_free(name_node); 363 } 364 365 static void __netdev_name_node_alt_destroy(struct netdev_name_node *name_node) 366 { 367 netdev_name_node_del(name_node); 368 list_del(&name_node->list); 369 call_rcu(&name_node->rcu, netdev_name_node_alt_free); 370 } 371 372 int netdev_name_node_alt_destroy(struct net_device *dev, const char *name) 373 { 374 struct netdev_name_node *name_node; 375 struct net *net = dev_net(dev); 376 377 name_node = netdev_name_node_lookup(net, name); 378 if (!name_node) 379 return -ENOENT; 380 /* lookup might have found our primary name or a name belonging 381 * to another device. 382 */ 383 if (name_node == dev->name_node || name_node->dev != dev) 384 return -EINVAL; 385 386 __netdev_name_node_alt_destroy(name_node); 387 return 0; 388 } 389 390 static void netdev_name_node_alt_flush(struct net_device *dev) 391 { 392 struct netdev_name_node *name_node, *tmp; 393 394 list_for_each_entry_safe(name_node, tmp, &dev->name_node->list, list) { 395 list_del(&name_node->list); 396 netdev_name_node_alt_free(&name_node->rcu); 397 } 398 } 399 400 /* Device list insertion */ 401 static void list_netdevice(struct net_device *dev) 402 { 403 struct netdev_name_node *name_node; 404 struct net *net = dev_net(dev); 405 406 ASSERT_RTNL(); 407 408 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head); 409 netdev_name_node_add(net, dev->name_node); 410 hlist_add_head_rcu(&dev->index_hlist, 411 dev_index_hash(net, dev->ifindex)); 412 413 netdev_for_each_altname(dev, name_node) 414 netdev_name_node_add(net, name_node); 415 416 /* We reserved the ifindex, this can't fail */ 417 WARN_ON(xa_store(&net->dev_by_index, dev->ifindex, dev, GFP_KERNEL)); 418 419 dev_base_seq_inc(net); 420 } 421 422 /* Device list removal 423 * caller must respect a RCU grace period before freeing/reusing dev 424 */ 425 static void unlist_netdevice(struct net_device *dev) 426 { 427 struct netdev_name_node *name_node; 428 struct net *net = dev_net(dev); 429 430 ASSERT_RTNL(); 431 432 xa_erase(&net->dev_by_index, dev->ifindex); 433 434 netdev_for_each_altname(dev, name_node) 435 netdev_name_node_del(name_node); 436 437 /* Unlink dev from the device chain */ 438 list_del_rcu(&dev->dev_list); 439 netdev_name_node_del(dev->name_node); 440 hlist_del_rcu(&dev->index_hlist); 441 442 dev_base_seq_inc(dev_net(dev)); 443 } 444 445 /* 446 * Our notifier list 447 */ 448 449 static RAW_NOTIFIER_HEAD(netdev_chain); 450 451 /* 452 * Device drivers call our routines to queue packets here. We empty the 453 * queue in the local softnet handler. 454 */ 455 456 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data) = { 457 .process_queue_bh_lock = INIT_LOCAL_LOCK(process_queue_bh_lock), 458 }; 459 EXPORT_PER_CPU_SYMBOL(softnet_data); 460 461 /* Page_pool has a lockless array/stack to alloc/recycle pages. 462 * PP consumers must pay attention to run APIs in the appropriate context 463 * (e.g. NAPI context). 464 */ 465 DEFINE_PER_CPU(struct page_pool *, system_page_pool); 466 467 #ifdef CONFIG_LOCKDEP 468 /* 469 * register_netdevice() inits txq->_xmit_lock and sets lockdep class 470 * according to dev->type 471 */ 472 static const unsigned short netdev_lock_type[] = { 473 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25, 474 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET, 475 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM, 476 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP, 477 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD, 478 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25, 479 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP, 480 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD, 481 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI, 482 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE, 483 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET, 484 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL, 485 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM, 486 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE, 487 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE}; 488 489 static const char *const netdev_lock_name[] = { 490 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25", 491 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET", 492 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM", 493 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP", 494 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD", 495 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25", 496 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP", 497 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD", 498 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI", 499 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE", 500 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET", 501 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL", 502 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM", 503 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE", 504 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"}; 505 506 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)]; 507 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)]; 508 509 static inline unsigned short netdev_lock_pos(unsigned short dev_type) 510 { 511 int i; 512 513 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++) 514 if (netdev_lock_type[i] == dev_type) 515 return i; 516 /* the last key is used by default */ 517 return ARRAY_SIZE(netdev_lock_type) - 1; 518 } 519 520 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 521 unsigned short dev_type) 522 { 523 int i; 524 525 i = netdev_lock_pos(dev_type); 526 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i], 527 netdev_lock_name[i]); 528 } 529 530 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 531 { 532 int i; 533 534 i = netdev_lock_pos(dev->type); 535 lockdep_set_class_and_name(&dev->addr_list_lock, 536 &netdev_addr_lock_key[i], 537 netdev_lock_name[i]); 538 } 539 #else 540 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 541 unsigned short dev_type) 542 { 543 } 544 545 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 546 { 547 } 548 #endif 549 550 /******************************************************************************* 551 * 552 * Protocol management and registration routines 553 * 554 *******************************************************************************/ 555 556 557 /* 558 * Add a protocol ID to the list. Now that the input handler is 559 * smarter we can dispense with all the messy stuff that used to be 560 * here. 561 * 562 * BEWARE!!! Protocol handlers, mangling input packets, 563 * MUST BE last in hash buckets and checking protocol handlers 564 * MUST start from promiscuous ptype_all chain in net_bh. 565 * It is true now, do not change it. 566 * Explanation follows: if protocol handler, mangling packet, will 567 * be the first on list, it is not able to sense, that packet 568 * is cloned and should be copied-on-write, so that it will 569 * change it and subsequent readers will get broken packet. 570 * --ANK (980803) 571 */ 572 573 static inline struct list_head *ptype_head(const struct packet_type *pt) 574 { 575 if (pt->type == htons(ETH_P_ALL)) { 576 if (!pt->af_packet_net && !pt->dev) 577 return NULL; 578 579 return pt->dev ? &pt->dev->ptype_all : 580 &pt->af_packet_net->ptype_all; 581 } 582 583 if (pt->dev) 584 return &pt->dev->ptype_specific; 585 586 return pt->af_packet_net ? &pt->af_packet_net->ptype_specific : 587 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK]; 588 } 589 590 /** 591 * dev_add_pack - add packet handler 592 * @pt: packet type declaration 593 * 594 * Add a protocol handler to the networking stack. The passed &packet_type 595 * is linked into kernel lists and may not be freed until it has been 596 * removed from the kernel lists. 597 * 598 * This call does not sleep therefore it can not 599 * guarantee all CPU's that are in middle of receiving packets 600 * will see the new packet type (until the next received packet). 601 */ 602 603 void dev_add_pack(struct packet_type *pt) 604 { 605 struct list_head *head = ptype_head(pt); 606 607 if (WARN_ON_ONCE(!head)) 608 return; 609 610 spin_lock(&ptype_lock); 611 list_add_rcu(&pt->list, head); 612 spin_unlock(&ptype_lock); 613 } 614 EXPORT_SYMBOL(dev_add_pack); 615 616 /** 617 * __dev_remove_pack - remove packet handler 618 * @pt: packet type declaration 619 * 620 * Remove a protocol handler that was previously added to the kernel 621 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 622 * from the kernel lists and can be freed or reused once this function 623 * returns. 624 * 625 * The packet type might still be in use by receivers 626 * and must not be freed until after all the CPU's have gone 627 * through a quiescent state. 628 */ 629 void __dev_remove_pack(struct packet_type *pt) 630 { 631 struct list_head *head = ptype_head(pt); 632 struct packet_type *pt1; 633 634 if (!head) 635 return; 636 637 spin_lock(&ptype_lock); 638 639 list_for_each_entry(pt1, head, list) { 640 if (pt == pt1) { 641 list_del_rcu(&pt->list); 642 goto out; 643 } 644 } 645 646 pr_warn("dev_remove_pack: %p not found\n", pt); 647 out: 648 spin_unlock(&ptype_lock); 649 } 650 EXPORT_SYMBOL(__dev_remove_pack); 651 652 /** 653 * dev_remove_pack - remove packet handler 654 * @pt: packet type declaration 655 * 656 * Remove a protocol handler that was previously added to the kernel 657 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 658 * from the kernel lists and can be freed or reused once this function 659 * returns. 660 * 661 * This call sleeps to guarantee that no CPU is looking at the packet 662 * type after return. 663 */ 664 void dev_remove_pack(struct packet_type *pt) 665 { 666 __dev_remove_pack(pt); 667 668 synchronize_net(); 669 } 670 EXPORT_SYMBOL(dev_remove_pack); 671 672 673 /******************************************************************************* 674 * 675 * Device Interface Subroutines 676 * 677 *******************************************************************************/ 678 679 /** 680 * dev_get_iflink - get 'iflink' value of a interface 681 * @dev: targeted interface 682 * 683 * Indicates the ifindex the interface is linked to. 684 * Physical interfaces have the same 'ifindex' and 'iflink' values. 685 */ 686 687 int dev_get_iflink(const struct net_device *dev) 688 { 689 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink) 690 return dev->netdev_ops->ndo_get_iflink(dev); 691 692 return READ_ONCE(dev->ifindex); 693 } 694 EXPORT_SYMBOL(dev_get_iflink); 695 696 /** 697 * dev_fill_metadata_dst - Retrieve tunnel egress information. 698 * @dev: targeted interface 699 * @skb: The packet. 700 * 701 * For better visibility of tunnel traffic OVS needs to retrieve 702 * egress tunnel information for a packet. Following API allows 703 * user to get this info. 704 */ 705 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb) 706 { 707 struct ip_tunnel_info *info; 708 709 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst) 710 return -EINVAL; 711 712 info = skb_tunnel_info_unclone(skb); 713 if (!info) 714 return -ENOMEM; 715 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX))) 716 return -EINVAL; 717 718 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb); 719 } 720 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst); 721 722 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack) 723 { 724 int k = stack->num_paths++; 725 726 if (WARN_ON_ONCE(k >= NET_DEVICE_PATH_STACK_MAX)) 727 return NULL; 728 729 return &stack->path[k]; 730 } 731 732 int dev_fill_forward_path(const struct net_device *dev, const u8 *daddr, 733 struct net_device_path_stack *stack) 734 { 735 const struct net_device *last_dev; 736 struct net_device_path_ctx ctx = { 737 .dev = dev, 738 }; 739 struct net_device_path *path; 740 int ret = 0; 741 742 memcpy(ctx.daddr, daddr, sizeof(ctx.daddr)); 743 stack->num_paths = 0; 744 while (ctx.dev && ctx.dev->netdev_ops->ndo_fill_forward_path) { 745 last_dev = ctx.dev; 746 path = dev_fwd_path(stack); 747 if (!path) 748 return -1; 749 750 memset(path, 0, sizeof(struct net_device_path)); 751 ret = ctx.dev->netdev_ops->ndo_fill_forward_path(&ctx, path); 752 if (ret < 0) 753 return -1; 754 755 if (WARN_ON_ONCE(last_dev == ctx.dev)) 756 return -1; 757 } 758 759 if (!ctx.dev) 760 return ret; 761 762 path = dev_fwd_path(stack); 763 if (!path) 764 return -1; 765 path->type = DEV_PATH_ETHERNET; 766 path->dev = ctx.dev; 767 768 return ret; 769 } 770 EXPORT_SYMBOL_GPL(dev_fill_forward_path); 771 772 /* must be called under rcu_read_lock(), as we dont take a reference */ 773 static struct napi_struct *napi_by_id(unsigned int napi_id) 774 { 775 unsigned int hash = napi_id % HASH_SIZE(napi_hash); 776 struct napi_struct *napi; 777 778 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node) 779 if (napi->napi_id == napi_id) 780 return napi; 781 782 return NULL; 783 } 784 785 /* must be called under rcu_read_lock(), as we dont take a reference */ 786 static struct napi_struct * 787 netdev_napi_by_id(struct net *net, unsigned int napi_id) 788 { 789 struct napi_struct *napi; 790 791 napi = napi_by_id(napi_id); 792 if (!napi) 793 return NULL; 794 795 if (WARN_ON_ONCE(!napi->dev)) 796 return NULL; 797 if (!net_eq(net, dev_net(napi->dev))) 798 return NULL; 799 800 return napi; 801 } 802 803 /** 804 * netdev_napi_by_id_lock() - find a device by NAPI ID and lock it 805 * @net: the applicable net namespace 806 * @napi_id: ID of a NAPI of a target device 807 * 808 * Find a NAPI instance with @napi_id. Lock its device. 809 * The device must be in %NETREG_REGISTERED state for lookup to succeed. 810 * netdev_unlock() must be called to release it. 811 * 812 * Return: pointer to NAPI, its device with lock held, NULL if not found. 813 */ 814 struct napi_struct * 815 netdev_napi_by_id_lock(struct net *net, unsigned int napi_id) 816 { 817 struct napi_struct *napi; 818 struct net_device *dev; 819 820 rcu_read_lock(); 821 napi = netdev_napi_by_id(net, napi_id); 822 if (!napi || READ_ONCE(napi->dev->reg_state) != NETREG_REGISTERED) { 823 rcu_read_unlock(); 824 return NULL; 825 } 826 827 dev = napi->dev; 828 dev_hold(dev); 829 rcu_read_unlock(); 830 831 dev = __netdev_put_lock(dev); 832 if (!dev) 833 return NULL; 834 835 rcu_read_lock(); 836 napi = netdev_napi_by_id(net, napi_id); 837 if (napi && napi->dev != dev) 838 napi = NULL; 839 rcu_read_unlock(); 840 841 if (!napi) 842 netdev_unlock(dev); 843 return napi; 844 } 845 846 /** 847 * __dev_get_by_name - find a device by its name 848 * @net: the applicable net namespace 849 * @name: name to find 850 * 851 * Find an interface by name. Must be called under RTNL semaphore. 852 * If the name is found a pointer to the device is returned. 853 * If the name is not found then %NULL is returned. The 854 * reference counters are not incremented so the caller must be 855 * careful with locks. 856 */ 857 858 struct net_device *__dev_get_by_name(struct net *net, const char *name) 859 { 860 struct netdev_name_node *node_name; 861 862 node_name = netdev_name_node_lookup(net, name); 863 return node_name ? node_name->dev : NULL; 864 } 865 EXPORT_SYMBOL(__dev_get_by_name); 866 867 /** 868 * dev_get_by_name_rcu - find a device by its name 869 * @net: the applicable net namespace 870 * @name: name to find 871 * 872 * Find an interface by name. 873 * If the name is found a pointer to the device is returned. 874 * If the name is not found then %NULL is returned. 875 * The reference counters are not incremented so the caller must be 876 * careful with locks. The caller must hold RCU lock. 877 */ 878 879 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name) 880 { 881 struct netdev_name_node *node_name; 882 883 node_name = netdev_name_node_lookup_rcu(net, name); 884 return node_name ? node_name->dev : NULL; 885 } 886 EXPORT_SYMBOL(dev_get_by_name_rcu); 887 888 /* Deprecated for new users, call netdev_get_by_name() instead */ 889 struct net_device *dev_get_by_name(struct net *net, const char *name) 890 { 891 struct net_device *dev; 892 893 rcu_read_lock(); 894 dev = dev_get_by_name_rcu(net, name); 895 dev_hold(dev); 896 rcu_read_unlock(); 897 return dev; 898 } 899 EXPORT_SYMBOL(dev_get_by_name); 900 901 /** 902 * netdev_get_by_name() - find a device by its name 903 * @net: the applicable net namespace 904 * @name: name to find 905 * @tracker: tracking object for the acquired reference 906 * @gfp: allocation flags for the tracker 907 * 908 * Find an interface by name. This can be called from any 909 * context and does its own locking. The returned handle has 910 * the usage count incremented and the caller must use netdev_put() to 911 * release it when it is no longer needed. %NULL is returned if no 912 * matching device is found. 913 */ 914 struct net_device *netdev_get_by_name(struct net *net, const char *name, 915 netdevice_tracker *tracker, gfp_t gfp) 916 { 917 struct net_device *dev; 918 919 dev = dev_get_by_name(net, name); 920 if (dev) 921 netdev_tracker_alloc(dev, tracker, gfp); 922 return dev; 923 } 924 EXPORT_SYMBOL(netdev_get_by_name); 925 926 /** 927 * __dev_get_by_index - find a device by its ifindex 928 * @net: the applicable net namespace 929 * @ifindex: index of device 930 * 931 * Search for an interface by index. Returns %NULL if the device 932 * is not found or a pointer to the device. The device has not 933 * had its reference counter increased so the caller must be careful 934 * about locking. The caller must hold the RTNL semaphore. 935 */ 936 937 struct net_device *__dev_get_by_index(struct net *net, int ifindex) 938 { 939 struct net_device *dev; 940 struct hlist_head *head = dev_index_hash(net, ifindex); 941 942 hlist_for_each_entry(dev, head, index_hlist) 943 if (dev->ifindex == ifindex) 944 return dev; 945 946 return NULL; 947 } 948 EXPORT_SYMBOL(__dev_get_by_index); 949 950 /** 951 * dev_get_by_index_rcu - find a device by its ifindex 952 * @net: the applicable net namespace 953 * @ifindex: index of device 954 * 955 * Search for an interface by index. Returns %NULL if the device 956 * is not found or a pointer to the device. The device has not 957 * had its reference counter increased so the caller must be careful 958 * about locking. The caller must hold RCU lock. 959 */ 960 961 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex) 962 { 963 struct net_device *dev; 964 struct hlist_head *head = dev_index_hash(net, ifindex); 965 966 hlist_for_each_entry_rcu(dev, head, index_hlist) 967 if (dev->ifindex == ifindex) 968 return dev; 969 970 return NULL; 971 } 972 EXPORT_SYMBOL(dev_get_by_index_rcu); 973 974 /* Deprecated for new users, call netdev_get_by_index() instead */ 975 struct net_device *dev_get_by_index(struct net *net, int ifindex) 976 { 977 struct net_device *dev; 978 979 rcu_read_lock(); 980 dev = dev_get_by_index_rcu(net, ifindex); 981 dev_hold(dev); 982 rcu_read_unlock(); 983 return dev; 984 } 985 EXPORT_SYMBOL(dev_get_by_index); 986 987 /** 988 * netdev_get_by_index() - find a device by its ifindex 989 * @net: the applicable net namespace 990 * @ifindex: index of device 991 * @tracker: tracking object for the acquired reference 992 * @gfp: allocation flags for the tracker 993 * 994 * Search for an interface by index. Returns NULL if the device 995 * is not found or a pointer to the device. The device returned has 996 * had a reference added and the pointer is safe until the user calls 997 * netdev_put() to indicate they have finished with it. 998 */ 999 struct net_device *netdev_get_by_index(struct net *net, int ifindex, 1000 netdevice_tracker *tracker, gfp_t gfp) 1001 { 1002 struct net_device *dev; 1003 1004 dev = dev_get_by_index(net, ifindex); 1005 if (dev) 1006 netdev_tracker_alloc(dev, tracker, gfp); 1007 return dev; 1008 } 1009 EXPORT_SYMBOL(netdev_get_by_index); 1010 1011 /** 1012 * dev_get_by_napi_id - find a device by napi_id 1013 * @napi_id: ID of the NAPI struct 1014 * 1015 * Search for an interface by NAPI ID. Returns %NULL if the device 1016 * is not found or a pointer to the device. The device has not had 1017 * its reference counter increased so the caller must be careful 1018 * about locking. The caller must hold RCU lock. 1019 */ 1020 struct net_device *dev_get_by_napi_id(unsigned int napi_id) 1021 { 1022 struct napi_struct *napi; 1023 1024 WARN_ON_ONCE(!rcu_read_lock_held()); 1025 1026 if (!napi_id_valid(napi_id)) 1027 return NULL; 1028 1029 napi = napi_by_id(napi_id); 1030 1031 return napi ? napi->dev : NULL; 1032 } 1033 1034 /* Release the held reference on the net_device, and if the net_device 1035 * is still registered try to lock the instance lock. If device is being 1036 * unregistered NULL will be returned (but the reference has been released, 1037 * either way!) 1038 * 1039 * This helper is intended for locking net_device after it has been looked up 1040 * using a lockless lookup helper. Lock prevents the instance from going away. 1041 */ 1042 struct net_device *__netdev_put_lock(struct net_device *dev) 1043 { 1044 netdev_lock(dev); 1045 if (dev->reg_state > NETREG_REGISTERED) { 1046 netdev_unlock(dev); 1047 dev_put(dev); 1048 return NULL; 1049 } 1050 dev_put(dev); 1051 return dev; 1052 } 1053 1054 /** 1055 * netdev_get_by_index_lock() - find a device by its ifindex 1056 * @net: the applicable net namespace 1057 * @ifindex: index of device 1058 * 1059 * Search for an interface by index. If a valid device 1060 * with @ifindex is found it will be returned with netdev->lock held. 1061 * netdev_unlock() must be called to release it. 1062 * 1063 * Return: pointer to a device with lock held, NULL if not found. 1064 */ 1065 struct net_device *netdev_get_by_index_lock(struct net *net, int ifindex) 1066 { 1067 struct net_device *dev; 1068 1069 dev = dev_get_by_index(net, ifindex); 1070 if (!dev) 1071 return NULL; 1072 1073 return __netdev_put_lock(dev); 1074 } 1075 1076 struct net_device * 1077 netdev_xa_find_lock(struct net *net, struct net_device *dev, 1078 unsigned long *index) 1079 { 1080 if (dev) 1081 netdev_unlock(dev); 1082 1083 do { 1084 rcu_read_lock(); 1085 dev = xa_find(&net->dev_by_index, index, ULONG_MAX, XA_PRESENT); 1086 if (!dev) { 1087 rcu_read_unlock(); 1088 return NULL; 1089 } 1090 dev_hold(dev); 1091 rcu_read_unlock(); 1092 1093 dev = __netdev_put_lock(dev); 1094 if (dev) 1095 return dev; 1096 1097 (*index)++; 1098 } while (true); 1099 } 1100 1101 static DEFINE_SEQLOCK(netdev_rename_lock); 1102 1103 void netdev_copy_name(struct net_device *dev, char *name) 1104 { 1105 unsigned int seq; 1106 1107 do { 1108 seq = read_seqbegin(&netdev_rename_lock); 1109 strscpy(name, dev->name, IFNAMSIZ); 1110 } while (read_seqretry(&netdev_rename_lock, seq)); 1111 } 1112 1113 /** 1114 * netdev_get_name - get a netdevice name, knowing its ifindex. 1115 * @net: network namespace 1116 * @name: a pointer to the buffer where the name will be stored. 1117 * @ifindex: the ifindex of the interface to get the name from. 1118 */ 1119 int netdev_get_name(struct net *net, char *name, int ifindex) 1120 { 1121 struct net_device *dev; 1122 int ret; 1123 1124 rcu_read_lock(); 1125 1126 dev = dev_get_by_index_rcu(net, ifindex); 1127 if (!dev) { 1128 ret = -ENODEV; 1129 goto out; 1130 } 1131 1132 netdev_copy_name(dev, name); 1133 1134 ret = 0; 1135 out: 1136 rcu_read_unlock(); 1137 return ret; 1138 } 1139 1140 static bool dev_addr_cmp(struct net_device *dev, unsigned short type, 1141 const char *ha) 1142 { 1143 return dev->type == type && !memcmp(dev->dev_addr, ha, dev->addr_len); 1144 } 1145 1146 /** 1147 * dev_getbyhwaddr_rcu - find a device by its hardware address 1148 * @net: the applicable net namespace 1149 * @type: media type of device 1150 * @ha: hardware address 1151 * 1152 * Search for an interface by MAC address. Returns NULL if the device 1153 * is not found or a pointer to the device. 1154 * The caller must hold RCU. 1155 * The returned device has not had its ref count increased 1156 * and the caller must therefore be careful about locking 1157 * 1158 */ 1159 1160 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type, 1161 const char *ha) 1162 { 1163 struct net_device *dev; 1164 1165 for_each_netdev_rcu(net, dev) 1166 if (dev_addr_cmp(dev, type, ha)) 1167 return dev; 1168 1169 return NULL; 1170 } 1171 EXPORT_SYMBOL(dev_getbyhwaddr_rcu); 1172 1173 /** 1174 * dev_getbyhwaddr() - find a device by its hardware address 1175 * @net: the applicable net namespace 1176 * @type: media type of device 1177 * @ha: hardware address 1178 * 1179 * Similar to dev_getbyhwaddr_rcu(), but the owner needs to hold 1180 * rtnl_lock. 1181 * 1182 * Context: rtnl_lock() must be held. 1183 * Return: pointer to the net_device, or NULL if not found 1184 */ 1185 struct net_device *dev_getbyhwaddr(struct net *net, unsigned short type, 1186 const char *ha) 1187 { 1188 struct net_device *dev; 1189 1190 ASSERT_RTNL(); 1191 for_each_netdev(net, dev) 1192 if (dev_addr_cmp(dev, type, ha)) 1193 return dev; 1194 1195 return NULL; 1196 } 1197 EXPORT_SYMBOL(dev_getbyhwaddr); 1198 1199 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type) 1200 { 1201 struct net_device *dev, *ret = NULL; 1202 1203 rcu_read_lock(); 1204 for_each_netdev_rcu(net, dev) 1205 if (dev->type == type) { 1206 dev_hold(dev); 1207 ret = dev; 1208 break; 1209 } 1210 rcu_read_unlock(); 1211 return ret; 1212 } 1213 EXPORT_SYMBOL(dev_getfirstbyhwtype); 1214 1215 /** 1216 * __dev_get_by_flags - find any device with given flags 1217 * @net: the applicable net namespace 1218 * @if_flags: IFF_* values 1219 * @mask: bitmask of bits in if_flags to check 1220 * 1221 * Search for any interface with the given flags. Returns NULL if a device 1222 * is not found or a pointer to the device. Must be called inside 1223 * rtnl_lock(), and result refcount is unchanged. 1224 */ 1225 1226 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags, 1227 unsigned short mask) 1228 { 1229 struct net_device *dev, *ret; 1230 1231 ASSERT_RTNL(); 1232 1233 ret = NULL; 1234 for_each_netdev(net, dev) { 1235 if (((dev->flags ^ if_flags) & mask) == 0) { 1236 ret = dev; 1237 break; 1238 } 1239 } 1240 return ret; 1241 } 1242 EXPORT_SYMBOL(__dev_get_by_flags); 1243 1244 /** 1245 * dev_valid_name - check if name is okay for network device 1246 * @name: name string 1247 * 1248 * Network device names need to be valid file names to 1249 * allow sysfs to work. We also disallow any kind of 1250 * whitespace. 1251 */ 1252 bool dev_valid_name(const char *name) 1253 { 1254 if (*name == '\0') 1255 return false; 1256 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ) 1257 return false; 1258 if (!strcmp(name, ".") || !strcmp(name, "..")) 1259 return false; 1260 1261 while (*name) { 1262 if (*name == '/' || *name == ':' || isspace(*name)) 1263 return false; 1264 name++; 1265 } 1266 return true; 1267 } 1268 EXPORT_SYMBOL(dev_valid_name); 1269 1270 /** 1271 * __dev_alloc_name - allocate a name for a device 1272 * @net: network namespace to allocate the device name in 1273 * @name: name format string 1274 * @res: result name string 1275 * 1276 * Passed a format string - eg "lt%d" it will try and find a suitable 1277 * id. It scans list of devices to build up a free map, then chooses 1278 * the first empty slot. The caller must hold the dev_base or rtnl lock 1279 * while allocating the name and adding the device in order to avoid 1280 * duplicates. 1281 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1282 * Returns the number of the unit assigned or a negative errno code. 1283 */ 1284 1285 static int __dev_alloc_name(struct net *net, const char *name, char *res) 1286 { 1287 int i = 0; 1288 const char *p; 1289 const int max_netdevices = 8*PAGE_SIZE; 1290 unsigned long *inuse; 1291 struct net_device *d; 1292 char buf[IFNAMSIZ]; 1293 1294 /* Verify the string as this thing may have come from the user. 1295 * There must be one "%d" and no other "%" characters. 1296 */ 1297 p = strchr(name, '%'); 1298 if (!p || p[1] != 'd' || strchr(p + 2, '%')) 1299 return -EINVAL; 1300 1301 /* Use one page as a bit array of possible slots */ 1302 inuse = bitmap_zalloc(max_netdevices, GFP_ATOMIC); 1303 if (!inuse) 1304 return -ENOMEM; 1305 1306 for_each_netdev(net, d) { 1307 struct netdev_name_node *name_node; 1308 1309 netdev_for_each_altname(d, name_node) { 1310 if (!sscanf(name_node->name, name, &i)) 1311 continue; 1312 if (i < 0 || i >= max_netdevices) 1313 continue; 1314 1315 /* avoid cases where sscanf is not exact inverse of printf */ 1316 snprintf(buf, IFNAMSIZ, name, i); 1317 if (!strncmp(buf, name_node->name, IFNAMSIZ)) 1318 __set_bit(i, inuse); 1319 } 1320 if (!sscanf(d->name, name, &i)) 1321 continue; 1322 if (i < 0 || i >= max_netdevices) 1323 continue; 1324 1325 /* avoid cases where sscanf is not exact inverse of printf */ 1326 snprintf(buf, IFNAMSIZ, name, i); 1327 if (!strncmp(buf, d->name, IFNAMSIZ)) 1328 __set_bit(i, inuse); 1329 } 1330 1331 i = find_first_zero_bit(inuse, max_netdevices); 1332 bitmap_free(inuse); 1333 if (i == max_netdevices) 1334 return -ENFILE; 1335 1336 /* 'res' and 'name' could overlap, use 'buf' as an intermediate buffer */ 1337 strscpy(buf, name, IFNAMSIZ); 1338 snprintf(res, IFNAMSIZ, buf, i); 1339 return i; 1340 } 1341 1342 /* Returns negative errno or allocated unit id (see __dev_alloc_name()) */ 1343 static int dev_prep_valid_name(struct net *net, struct net_device *dev, 1344 const char *want_name, char *out_name, 1345 int dup_errno) 1346 { 1347 if (!dev_valid_name(want_name)) 1348 return -EINVAL; 1349 1350 if (strchr(want_name, '%')) 1351 return __dev_alloc_name(net, want_name, out_name); 1352 1353 if (netdev_name_in_use(net, want_name)) 1354 return -dup_errno; 1355 if (out_name != want_name) 1356 strscpy(out_name, want_name, IFNAMSIZ); 1357 return 0; 1358 } 1359 1360 /** 1361 * dev_alloc_name - allocate a name for a device 1362 * @dev: device 1363 * @name: name format string 1364 * 1365 * Passed a format string - eg "lt%d" it will try and find a suitable 1366 * id. It scans list of devices to build up a free map, then chooses 1367 * the first empty slot. The caller must hold the dev_base or rtnl lock 1368 * while allocating the name and adding the device in order to avoid 1369 * duplicates. 1370 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1371 * Returns the number of the unit assigned or a negative errno code. 1372 */ 1373 1374 int dev_alloc_name(struct net_device *dev, const char *name) 1375 { 1376 return dev_prep_valid_name(dev_net(dev), dev, name, dev->name, ENFILE); 1377 } 1378 EXPORT_SYMBOL(dev_alloc_name); 1379 1380 static int dev_get_valid_name(struct net *net, struct net_device *dev, 1381 const char *name) 1382 { 1383 int ret; 1384 1385 ret = dev_prep_valid_name(net, dev, name, dev->name, EEXIST); 1386 return ret < 0 ? ret : 0; 1387 } 1388 1389 int netif_change_name(struct net_device *dev, const char *newname) 1390 { 1391 struct net *net = dev_net(dev); 1392 unsigned char old_assign_type; 1393 char oldname[IFNAMSIZ]; 1394 int err = 0; 1395 int ret; 1396 1397 ASSERT_RTNL_NET(net); 1398 1399 if (!strncmp(newname, dev->name, IFNAMSIZ)) 1400 return 0; 1401 1402 memcpy(oldname, dev->name, IFNAMSIZ); 1403 1404 write_seqlock_bh(&netdev_rename_lock); 1405 err = dev_get_valid_name(net, dev, newname); 1406 write_sequnlock_bh(&netdev_rename_lock); 1407 1408 if (err < 0) 1409 return err; 1410 1411 if (oldname[0] && !strchr(oldname, '%')) 1412 netdev_info(dev, "renamed from %s%s\n", oldname, 1413 dev->flags & IFF_UP ? " (while UP)" : ""); 1414 1415 old_assign_type = dev->name_assign_type; 1416 WRITE_ONCE(dev->name_assign_type, NET_NAME_RENAMED); 1417 1418 rollback: 1419 ret = device_rename(&dev->dev, dev->name); 1420 if (ret) { 1421 write_seqlock_bh(&netdev_rename_lock); 1422 memcpy(dev->name, oldname, IFNAMSIZ); 1423 write_sequnlock_bh(&netdev_rename_lock); 1424 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1425 return ret; 1426 } 1427 1428 netdev_adjacent_rename_links(dev, oldname); 1429 1430 netdev_name_node_del(dev->name_node); 1431 1432 synchronize_net(); 1433 1434 netdev_name_node_add(net, dev->name_node); 1435 1436 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev); 1437 ret = notifier_to_errno(ret); 1438 1439 if (ret) { 1440 /* err >= 0 after dev_alloc_name() or stores the first errno */ 1441 if (err >= 0) { 1442 err = ret; 1443 write_seqlock_bh(&netdev_rename_lock); 1444 memcpy(dev->name, oldname, IFNAMSIZ); 1445 write_sequnlock_bh(&netdev_rename_lock); 1446 memcpy(oldname, newname, IFNAMSIZ); 1447 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1448 old_assign_type = NET_NAME_RENAMED; 1449 goto rollback; 1450 } else { 1451 netdev_err(dev, "name change rollback failed: %d\n", 1452 ret); 1453 } 1454 } 1455 1456 return err; 1457 } 1458 1459 int netif_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 1486 /** 1487 * dev_get_alias - get ifalias of a device 1488 * @dev: device 1489 * @name: buffer to store name of ifalias 1490 * @len: size of buffer 1491 * 1492 * get ifalias for a device. Caller must make sure dev cannot go 1493 * away, e.g. rcu read lock or own a reference count to device. 1494 */ 1495 int dev_get_alias(const struct net_device *dev, char *name, size_t len) 1496 { 1497 const struct dev_ifalias *alias; 1498 int ret = 0; 1499 1500 rcu_read_lock(); 1501 alias = rcu_dereference(dev->ifalias); 1502 if (alias) 1503 ret = snprintf(name, len, "%s", alias->ifalias); 1504 rcu_read_unlock(); 1505 1506 return ret; 1507 } 1508 1509 /** 1510 * netdev_features_change - device changes features 1511 * @dev: device to cause notification 1512 * 1513 * Called to indicate a device has changed features. 1514 */ 1515 void netdev_features_change(struct net_device *dev) 1516 { 1517 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev); 1518 } 1519 EXPORT_SYMBOL(netdev_features_change); 1520 1521 void netif_state_change(struct net_device *dev) 1522 { 1523 if (dev->flags & IFF_UP) { 1524 struct netdev_notifier_change_info change_info = { 1525 .info.dev = dev, 1526 }; 1527 1528 call_netdevice_notifiers_info(NETDEV_CHANGE, 1529 &change_info.info); 1530 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL, 0, NULL); 1531 } 1532 } 1533 1534 /** 1535 * __netdev_notify_peers - notify network peers about existence of @dev, 1536 * to be called when rtnl lock is already held. 1537 * @dev: network device 1538 * 1539 * Generate traffic such that interested network peers are aware of 1540 * @dev, such as by generating a gratuitous ARP. This may be used when 1541 * a device wants to inform the rest of the network about some sort of 1542 * reconfiguration such as a failover event or virtual machine 1543 * migration. 1544 */ 1545 void __netdev_notify_peers(struct net_device *dev) 1546 { 1547 ASSERT_RTNL(); 1548 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev); 1549 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev); 1550 } 1551 EXPORT_SYMBOL(__netdev_notify_peers); 1552 1553 /** 1554 * netdev_notify_peers - notify network peers about existence of @dev 1555 * @dev: network device 1556 * 1557 * Generate traffic such that interested network peers are aware of 1558 * @dev, such as by generating a gratuitous ARP. This may be used when 1559 * a device wants to inform the rest of the network about some sort of 1560 * reconfiguration such as a failover event or virtual machine 1561 * migration. 1562 */ 1563 void netdev_notify_peers(struct net_device *dev) 1564 { 1565 rtnl_lock(); 1566 __netdev_notify_peers(dev); 1567 rtnl_unlock(); 1568 } 1569 EXPORT_SYMBOL(netdev_notify_peers); 1570 1571 static int napi_threaded_poll(void *data); 1572 1573 static int napi_kthread_create(struct napi_struct *n) 1574 { 1575 int err = 0; 1576 1577 /* Create and wake up the kthread once to put it in 1578 * TASK_INTERRUPTIBLE mode to avoid the blocked task 1579 * warning and work with loadavg. 1580 */ 1581 n->thread = kthread_run(napi_threaded_poll, n, "napi/%s-%d", 1582 n->dev->name, n->napi_id); 1583 if (IS_ERR(n->thread)) { 1584 err = PTR_ERR(n->thread); 1585 pr_err("kthread_run failed with err %d\n", err); 1586 n->thread = NULL; 1587 } 1588 1589 return err; 1590 } 1591 1592 static int __dev_open(struct net_device *dev, struct netlink_ext_ack *extack) 1593 { 1594 const struct net_device_ops *ops = dev->netdev_ops; 1595 int ret; 1596 1597 ASSERT_RTNL(); 1598 dev_addr_check(dev); 1599 1600 if (!netif_device_present(dev)) { 1601 /* may be detached because parent is runtime-suspended */ 1602 if (dev->dev.parent) 1603 pm_runtime_resume(dev->dev.parent); 1604 if (!netif_device_present(dev)) 1605 return -ENODEV; 1606 } 1607 1608 /* Block netpoll from trying to do any rx path servicing. 1609 * If we don't do this there is a chance ndo_poll_controller 1610 * or ndo_poll may be running while we open the device 1611 */ 1612 netpoll_poll_disable(dev); 1613 1614 ret = call_netdevice_notifiers_extack(NETDEV_PRE_UP, dev, extack); 1615 ret = notifier_to_errno(ret); 1616 if (ret) 1617 return ret; 1618 1619 set_bit(__LINK_STATE_START, &dev->state); 1620 1621 netdev_ops_assert_locked(dev); 1622 1623 if (ops->ndo_validate_addr) 1624 ret = ops->ndo_validate_addr(dev); 1625 1626 if (!ret && ops->ndo_open) 1627 ret = ops->ndo_open(dev); 1628 1629 netpoll_poll_enable(dev); 1630 1631 if (ret) 1632 clear_bit(__LINK_STATE_START, &dev->state); 1633 else { 1634 netif_set_up(dev, true); 1635 dev_set_rx_mode(dev); 1636 dev_activate(dev); 1637 add_device_randomness(dev->dev_addr, dev->addr_len); 1638 } 1639 1640 return ret; 1641 } 1642 1643 int netif_open(struct net_device *dev, struct netlink_ext_ack *extack) 1644 { 1645 int ret; 1646 1647 if (dev->flags & IFF_UP) 1648 return 0; 1649 1650 ret = __dev_open(dev, extack); 1651 if (ret < 0) 1652 return ret; 1653 1654 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1655 call_netdevice_notifiers(NETDEV_UP, dev); 1656 1657 return ret; 1658 } 1659 1660 static void __dev_close_many(struct list_head *head) 1661 { 1662 struct net_device *dev; 1663 1664 ASSERT_RTNL(); 1665 might_sleep(); 1666 1667 list_for_each_entry(dev, head, close_list) { 1668 /* Temporarily disable netpoll until the interface is down */ 1669 netpoll_poll_disable(dev); 1670 1671 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev); 1672 1673 clear_bit(__LINK_STATE_START, &dev->state); 1674 1675 /* Synchronize to scheduled poll. We cannot touch poll list, it 1676 * can be even on different cpu. So just clear netif_running(). 1677 * 1678 * dev->stop() will invoke napi_disable() on all of it's 1679 * napi_struct instances on this device. 1680 */ 1681 smp_mb__after_atomic(); /* Commit netif_running(). */ 1682 } 1683 1684 dev_deactivate_many(head); 1685 1686 list_for_each_entry(dev, head, close_list) { 1687 const struct net_device_ops *ops = dev->netdev_ops; 1688 1689 /* 1690 * Call the device specific close. This cannot fail. 1691 * Only if device is UP 1692 * 1693 * We allow it to be called even after a DETACH hot-plug 1694 * event. 1695 */ 1696 1697 netdev_ops_assert_locked(dev); 1698 1699 if (ops->ndo_stop) 1700 ops->ndo_stop(dev); 1701 1702 netif_set_up(dev, false); 1703 netpoll_poll_enable(dev); 1704 } 1705 } 1706 1707 static void __dev_close(struct net_device *dev) 1708 { 1709 LIST_HEAD(single); 1710 1711 list_add(&dev->close_list, &single); 1712 __dev_close_many(&single); 1713 list_del(&single); 1714 } 1715 1716 void dev_close_many(struct list_head *head, bool unlink) 1717 { 1718 struct net_device *dev, *tmp; 1719 1720 /* Remove the devices that don't need to be closed */ 1721 list_for_each_entry_safe(dev, tmp, head, close_list) 1722 if (!(dev->flags & IFF_UP)) 1723 list_del_init(&dev->close_list); 1724 1725 __dev_close_many(head); 1726 1727 list_for_each_entry_safe(dev, tmp, head, close_list) { 1728 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1729 call_netdevice_notifiers(NETDEV_DOWN, dev); 1730 if (unlink) 1731 list_del_init(&dev->close_list); 1732 } 1733 } 1734 EXPORT_SYMBOL(dev_close_many); 1735 1736 void netif_close(struct net_device *dev) 1737 { 1738 if (dev->flags & IFF_UP) { 1739 LIST_HEAD(single); 1740 1741 list_add(&dev->close_list, &single); 1742 dev_close_many(&single, true); 1743 list_del(&single); 1744 } 1745 } 1746 EXPORT_SYMBOL(netif_close); 1747 1748 void netif_disable_lro(struct net_device *dev) 1749 { 1750 struct net_device *lower_dev; 1751 struct list_head *iter; 1752 1753 dev->wanted_features &= ~NETIF_F_LRO; 1754 netdev_update_features(dev); 1755 1756 if (unlikely(dev->features & NETIF_F_LRO)) 1757 netdev_WARN(dev, "failed to disable LRO!\n"); 1758 1759 netdev_for_each_lower_dev(dev, lower_dev, iter) { 1760 netdev_lock_ops(lower_dev); 1761 netif_disable_lro(lower_dev); 1762 netdev_unlock_ops(lower_dev); 1763 } 1764 } 1765 EXPORT_IPV6_MOD(netif_disable_lro); 1766 1767 /** 1768 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device 1769 * @dev: device 1770 * 1771 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be 1772 * called under RTNL. This is needed if Generic XDP is installed on 1773 * the device. 1774 */ 1775 static void dev_disable_gro_hw(struct net_device *dev) 1776 { 1777 dev->wanted_features &= ~NETIF_F_GRO_HW; 1778 netdev_update_features(dev); 1779 1780 if (unlikely(dev->features & NETIF_F_GRO_HW)) 1781 netdev_WARN(dev, "failed to disable GRO_HW!\n"); 1782 } 1783 1784 const char *netdev_cmd_to_name(enum netdev_cmd cmd) 1785 { 1786 #define N(val) \ 1787 case NETDEV_##val: \ 1788 return "NETDEV_" __stringify(val); 1789 switch (cmd) { 1790 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER) 1791 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE) 1792 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE) 1793 N(POST_INIT) N(PRE_UNINIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) 1794 N(CHANGEUPPER) N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) 1795 N(BONDING_INFO) N(PRECHANGEUPPER) N(CHANGELOWERSTATE) 1796 N(UDP_TUNNEL_PUSH_INFO) N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN) 1797 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO) 1798 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO) 1799 N(PRE_CHANGEADDR) N(OFFLOAD_XSTATS_ENABLE) N(OFFLOAD_XSTATS_DISABLE) 1800 N(OFFLOAD_XSTATS_REPORT_USED) N(OFFLOAD_XSTATS_REPORT_DELTA) 1801 N(XDP_FEAT_CHANGE) 1802 } 1803 #undef N 1804 return "UNKNOWN_NETDEV_EVENT"; 1805 } 1806 EXPORT_SYMBOL_GPL(netdev_cmd_to_name); 1807 1808 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val, 1809 struct net_device *dev) 1810 { 1811 struct netdev_notifier_info info = { 1812 .dev = dev, 1813 }; 1814 1815 return nb->notifier_call(nb, val, &info); 1816 } 1817 1818 static int call_netdevice_register_notifiers(struct notifier_block *nb, 1819 struct net_device *dev) 1820 { 1821 int err; 1822 1823 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev); 1824 err = notifier_to_errno(err); 1825 if (err) 1826 return err; 1827 1828 if (!(dev->flags & IFF_UP)) 1829 return 0; 1830 1831 call_netdevice_notifier(nb, NETDEV_UP, dev); 1832 return 0; 1833 } 1834 1835 static void call_netdevice_unregister_notifiers(struct notifier_block *nb, 1836 struct net_device *dev) 1837 { 1838 if (dev->flags & IFF_UP) { 1839 call_netdevice_notifier(nb, NETDEV_GOING_DOWN, 1840 dev); 1841 call_netdevice_notifier(nb, NETDEV_DOWN, dev); 1842 } 1843 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev); 1844 } 1845 1846 static int call_netdevice_register_net_notifiers(struct notifier_block *nb, 1847 struct net *net) 1848 { 1849 struct net_device *dev; 1850 int err; 1851 1852 for_each_netdev(net, dev) { 1853 netdev_lock_ops(dev); 1854 err = call_netdevice_register_notifiers(nb, dev); 1855 netdev_unlock_ops(dev); 1856 if (err) 1857 goto rollback; 1858 } 1859 return 0; 1860 1861 rollback: 1862 for_each_netdev_continue_reverse(net, dev) 1863 call_netdevice_unregister_notifiers(nb, dev); 1864 return err; 1865 } 1866 1867 static void call_netdevice_unregister_net_notifiers(struct notifier_block *nb, 1868 struct net *net) 1869 { 1870 struct net_device *dev; 1871 1872 for_each_netdev(net, dev) 1873 call_netdevice_unregister_notifiers(nb, dev); 1874 } 1875 1876 static int dev_boot_phase = 1; 1877 1878 /** 1879 * register_netdevice_notifier - register a network notifier block 1880 * @nb: notifier 1881 * 1882 * Register a notifier to be called when network device events occur. 1883 * The notifier passed is linked into the kernel structures and must 1884 * not be reused until it has been unregistered. A negative errno code 1885 * is returned on a failure. 1886 * 1887 * When registered all registration and up events are replayed 1888 * to the new notifier to allow device to have a race free 1889 * view of the network device list. 1890 */ 1891 1892 int register_netdevice_notifier(struct notifier_block *nb) 1893 { 1894 struct net *net; 1895 int err; 1896 1897 /* Close race with setup_net() and cleanup_net() */ 1898 down_write(&pernet_ops_rwsem); 1899 1900 /* When RTNL is removed, we need protection for netdev_chain. */ 1901 rtnl_lock(); 1902 1903 err = raw_notifier_chain_register(&netdev_chain, nb); 1904 if (err) 1905 goto unlock; 1906 if (dev_boot_phase) 1907 goto unlock; 1908 for_each_net(net) { 1909 __rtnl_net_lock(net); 1910 err = call_netdevice_register_net_notifiers(nb, net); 1911 __rtnl_net_unlock(net); 1912 if (err) 1913 goto rollback; 1914 } 1915 1916 unlock: 1917 rtnl_unlock(); 1918 up_write(&pernet_ops_rwsem); 1919 return err; 1920 1921 rollback: 1922 for_each_net_continue_reverse(net) { 1923 __rtnl_net_lock(net); 1924 call_netdevice_unregister_net_notifiers(nb, net); 1925 __rtnl_net_unlock(net); 1926 } 1927 1928 raw_notifier_chain_unregister(&netdev_chain, nb); 1929 goto unlock; 1930 } 1931 EXPORT_SYMBOL(register_netdevice_notifier); 1932 1933 /** 1934 * unregister_netdevice_notifier - unregister a network notifier block 1935 * @nb: notifier 1936 * 1937 * Unregister a notifier previously registered by 1938 * register_netdevice_notifier(). The notifier is unlinked into the 1939 * kernel structures and may then be reused. A negative errno code 1940 * is returned on a failure. 1941 * 1942 * After unregistering unregister and down device events are synthesized 1943 * for all devices on the device list to the removed notifier to remove 1944 * the need for special case cleanup code. 1945 */ 1946 1947 int unregister_netdevice_notifier(struct notifier_block *nb) 1948 { 1949 struct net *net; 1950 int err; 1951 1952 /* Close race with setup_net() and cleanup_net() */ 1953 down_write(&pernet_ops_rwsem); 1954 rtnl_lock(); 1955 err = raw_notifier_chain_unregister(&netdev_chain, nb); 1956 if (err) 1957 goto unlock; 1958 1959 for_each_net(net) { 1960 __rtnl_net_lock(net); 1961 call_netdevice_unregister_net_notifiers(nb, net); 1962 __rtnl_net_unlock(net); 1963 } 1964 1965 unlock: 1966 rtnl_unlock(); 1967 up_write(&pernet_ops_rwsem); 1968 return err; 1969 } 1970 EXPORT_SYMBOL(unregister_netdevice_notifier); 1971 1972 static int __register_netdevice_notifier_net(struct net *net, 1973 struct notifier_block *nb, 1974 bool ignore_call_fail) 1975 { 1976 int err; 1977 1978 err = raw_notifier_chain_register(&net->netdev_chain, nb); 1979 if (err) 1980 return err; 1981 if (dev_boot_phase) 1982 return 0; 1983 1984 err = call_netdevice_register_net_notifiers(nb, net); 1985 if (err && !ignore_call_fail) 1986 goto chain_unregister; 1987 1988 return 0; 1989 1990 chain_unregister: 1991 raw_notifier_chain_unregister(&net->netdev_chain, nb); 1992 return err; 1993 } 1994 1995 static int __unregister_netdevice_notifier_net(struct net *net, 1996 struct notifier_block *nb) 1997 { 1998 int err; 1999 2000 err = raw_notifier_chain_unregister(&net->netdev_chain, nb); 2001 if (err) 2002 return err; 2003 2004 call_netdevice_unregister_net_notifiers(nb, net); 2005 return 0; 2006 } 2007 2008 /** 2009 * register_netdevice_notifier_net - register a per-netns network notifier block 2010 * @net: network namespace 2011 * @nb: notifier 2012 * 2013 * Register a notifier to be called when network device events occur. 2014 * The notifier passed is linked into the kernel structures and must 2015 * not be reused until it has been unregistered. A negative errno code 2016 * is returned on a failure. 2017 * 2018 * When registered all registration and up events are replayed 2019 * to the new notifier to allow device to have a race free 2020 * view of the network device list. 2021 */ 2022 2023 int register_netdevice_notifier_net(struct net *net, struct notifier_block *nb) 2024 { 2025 int err; 2026 2027 rtnl_net_lock(net); 2028 err = __register_netdevice_notifier_net(net, nb, false); 2029 rtnl_net_unlock(net); 2030 2031 return err; 2032 } 2033 EXPORT_SYMBOL(register_netdevice_notifier_net); 2034 2035 /** 2036 * unregister_netdevice_notifier_net - unregister a per-netns 2037 * network notifier block 2038 * @net: network namespace 2039 * @nb: notifier 2040 * 2041 * Unregister a notifier previously registered by 2042 * register_netdevice_notifier_net(). The notifier is unlinked from the 2043 * kernel structures and may then be reused. A negative errno code 2044 * is returned on a failure. 2045 * 2046 * After unregistering unregister and down device events are synthesized 2047 * for all devices on the device list to the removed notifier to remove 2048 * the need for special case cleanup code. 2049 */ 2050 2051 int unregister_netdevice_notifier_net(struct net *net, 2052 struct notifier_block *nb) 2053 { 2054 int err; 2055 2056 rtnl_net_lock(net); 2057 err = __unregister_netdevice_notifier_net(net, nb); 2058 rtnl_net_unlock(net); 2059 2060 return err; 2061 } 2062 EXPORT_SYMBOL(unregister_netdevice_notifier_net); 2063 2064 static void __move_netdevice_notifier_net(struct net *src_net, 2065 struct net *dst_net, 2066 struct notifier_block *nb) 2067 { 2068 __unregister_netdevice_notifier_net(src_net, nb); 2069 __register_netdevice_notifier_net(dst_net, nb, true); 2070 } 2071 2072 static void rtnl_net_dev_lock(struct net_device *dev) 2073 { 2074 bool again; 2075 2076 do { 2077 struct net *net; 2078 2079 again = false; 2080 2081 /* netns might be being dismantled. */ 2082 rcu_read_lock(); 2083 net = dev_net_rcu(dev); 2084 net_passive_inc(net); 2085 rcu_read_unlock(); 2086 2087 rtnl_net_lock(net); 2088 2089 #ifdef CONFIG_NET_NS 2090 /* dev might have been moved to another netns. */ 2091 if (!net_eq(net, rcu_access_pointer(dev->nd_net.net))) { 2092 rtnl_net_unlock(net); 2093 net_passive_dec(net); 2094 again = true; 2095 } 2096 #endif 2097 } while (again); 2098 } 2099 2100 static void rtnl_net_dev_unlock(struct net_device *dev) 2101 { 2102 struct net *net = dev_net(dev); 2103 2104 rtnl_net_unlock(net); 2105 net_passive_dec(net); 2106 } 2107 2108 int register_netdevice_notifier_dev_net(struct net_device *dev, 2109 struct notifier_block *nb, 2110 struct netdev_net_notifier *nn) 2111 { 2112 int err; 2113 2114 rtnl_net_dev_lock(dev); 2115 err = __register_netdevice_notifier_net(dev_net(dev), nb, false); 2116 if (!err) { 2117 nn->nb = nb; 2118 list_add(&nn->list, &dev->net_notifier_list); 2119 } 2120 rtnl_net_dev_unlock(dev); 2121 2122 return err; 2123 } 2124 EXPORT_SYMBOL(register_netdevice_notifier_dev_net); 2125 2126 int unregister_netdevice_notifier_dev_net(struct net_device *dev, 2127 struct notifier_block *nb, 2128 struct netdev_net_notifier *nn) 2129 { 2130 int err; 2131 2132 rtnl_net_dev_lock(dev); 2133 list_del(&nn->list); 2134 err = __unregister_netdevice_notifier_net(dev_net(dev), nb); 2135 rtnl_net_dev_unlock(dev); 2136 2137 return err; 2138 } 2139 EXPORT_SYMBOL(unregister_netdevice_notifier_dev_net); 2140 2141 static void move_netdevice_notifiers_dev_net(struct net_device *dev, 2142 struct net *net) 2143 { 2144 struct netdev_net_notifier *nn; 2145 2146 list_for_each_entry(nn, &dev->net_notifier_list, list) 2147 __move_netdevice_notifier_net(dev_net(dev), net, nn->nb); 2148 } 2149 2150 /** 2151 * call_netdevice_notifiers_info - call all network notifier blocks 2152 * @val: value passed unmodified to notifier function 2153 * @info: notifier information data 2154 * 2155 * Call all network notifier blocks. Parameters and return value 2156 * are as for raw_notifier_call_chain(). 2157 */ 2158 2159 int call_netdevice_notifiers_info(unsigned long val, 2160 struct netdev_notifier_info *info) 2161 { 2162 struct net *net = dev_net(info->dev); 2163 int ret; 2164 2165 ASSERT_RTNL(); 2166 2167 /* Run per-netns notifier block chain first, then run the global one. 2168 * Hopefully, one day, the global one is going to be removed after 2169 * all notifier block registrators get converted to be per-netns. 2170 */ 2171 ret = raw_notifier_call_chain(&net->netdev_chain, val, info); 2172 if (ret & NOTIFY_STOP_MASK) 2173 return ret; 2174 return raw_notifier_call_chain(&netdev_chain, val, info); 2175 } 2176 2177 /** 2178 * call_netdevice_notifiers_info_robust - call per-netns notifier blocks 2179 * for and rollback on error 2180 * @val_up: value passed unmodified to notifier function 2181 * @val_down: value passed unmodified to the notifier function when 2182 * recovering from an error on @val_up 2183 * @info: notifier information data 2184 * 2185 * Call all per-netns network notifier blocks, but not notifier blocks on 2186 * the global notifier chain. Parameters and return value are as for 2187 * raw_notifier_call_chain_robust(). 2188 */ 2189 2190 static int 2191 call_netdevice_notifiers_info_robust(unsigned long val_up, 2192 unsigned long val_down, 2193 struct netdev_notifier_info *info) 2194 { 2195 struct net *net = dev_net(info->dev); 2196 2197 ASSERT_RTNL(); 2198 2199 return raw_notifier_call_chain_robust(&net->netdev_chain, 2200 val_up, val_down, info); 2201 } 2202 2203 static int call_netdevice_notifiers_extack(unsigned long val, 2204 struct net_device *dev, 2205 struct netlink_ext_ack *extack) 2206 { 2207 struct netdev_notifier_info info = { 2208 .dev = dev, 2209 .extack = extack, 2210 }; 2211 2212 return call_netdevice_notifiers_info(val, &info); 2213 } 2214 2215 /** 2216 * call_netdevice_notifiers - call all network notifier blocks 2217 * @val: value passed unmodified to notifier function 2218 * @dev: net_device pointer passed unmodified to notifier function 2219 * 2220 * Call all network notifier blocks. Parameters and return value 2221 * are as for raw_notifier_call_chain(). 2222 */ 2223 2224 int call_netdevice_notifiers(unsigned long val, struct net_device *dev) 2225 { 2226 return call_netdevice_notifiers_extack(val, dev, NULL); 2227 } 2228 EXPORT_SYMBOL(call_netdevice_notifiers); 2229 2230 /** 2231 * call_netdevice_notifiers_mtu - call all network notifier blocks 2232 * @val: value passed unmodified to notifier function 2233 * @dev: net_device pointer passed unmodified to notifier function 2234 * @arg: additional u32 argument passed to the notifier function 2235 * 2236 * Call all network notifier blocks. Parameters and return value 2237 * are as for raw_notifier_call_chain(). 2238 */ 2239 static int call_netdevice_notifiers_mtu(unsigned long val, 2240 struct net_device *dev, u32 arg) 2241 { 2242 struct netdev_notifier_info_ext info = { 2243 .info.dev = dev, 2244 .ext.mtu = arg, 2245 }; 2246 2247 BUILD_BUG_ON(offsetof(struct netdev_notifier_info_ext, info) != 0); 2248 2249 return call_netdevice_notifiers_info(val, &info.info); 2250 } 2251 2252 #ifdef CONFIG_NET_INGRESS 2253 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key); 2254 2255 void net_inc_ingress_queue(void) 2256 { 2257 static_branch_inc(&ingress_needed_key); 2258 } 2259 EXPORT_SYMBOL_GPL(net_inc_ingress_queue); 2260 2261 void net_dec_ingress_queue(void) 2262 { 2263 static_branch_dec(&ingress_needed_key); 2264 } 2265 EXPORT_SYMBOL_GPL(net_dec_ingress_queue); 2266 #endif 2267 2268 #ifdef CONFIG_NET_EGRESS 2269 static DEFINE_STATIC_KEY_FALSE(egress_needed_key); 2270 2271 void net_inc_egress_queue(void) 2272 { 2273 static_branch_inc(&egress_needed_key); 2274 } 2275 EXPORT_SYMBOL_GPL(net_inc_egress_queue); 2276 2277 void net_dec_egress_queue(void) 2278 { 2279 static_branch_dec(&egress_needed_key); 2280 } 2281 EXPORT_SYMBOL_GPL(net_dec_egress_queue); 2282 #endif 2283 2284 #ifdef CONFIG_NET_CLS_ACT 2285 DEFINE_STATIC_KEY_FALSE(tcf_sw_enabled_key); 2286 EXPORT_SYMBOL(tcf_sw_enabled_key); 2287 #endif 2288 2289 DEFINE_STATIC_KEY_FALSE(netstamp_needed_key); 2290 EXPORT_SYMBOL(netstamp_needed_key); 2291 #ifdef CONFIG_JUMP_LABEL 2292 static atomic_t netstamp_needed_deferred; 2293 static atomic_t netstamp_wanted; 2294 static void netstamp_clear(struct work_struct *work) 2295 { 2296 int deferred = atomic_xchg(&netstamp_needed_deferred, 0); 2297 int wanted; 2298 2299 wanted = atomic_add_return(deferred, &netstamp_wanted); 2300 if (wanted > 0) 2301 static_branch_enable(&netstamp_needed_key); 2302 else 2303 static_branch_disable(&netstamp_needed_key); 2304 } 2305 static DECLARE_WORK(netstamp_work, netstamp_clear); 2306 #endif 2307 2308 void net_enable_timestamp(void) 2309 { 2310 #ifdef CONFIG_JUMP_LABEL 2311 int wanted = atomic_read(&netstamp_wanted); 2312 2313 while (wanted > 0) { 2314 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted + 1)) 2315 return; 2316 } 2317 atomic_inc(&netstamp_needed_deferred); 2318 schedule_work(&netstamp_work); 2319 #else 2320 static_branch_inc(&netstamp_needed_key); 2321 #endif 2322 } 2323 EXPORT_SYMBOL(net_enable_timestamp); 2324 2325 void net_disable_timestamp(void) 2326 { 2327 #ifdef CONFIG_JUMP_LABEL 2328 int wanted = atomic_read(&netstamp_wanted); 2329 2330 while (wanted > 1) { 2331 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted - 1)) 2332 return; 2333 } 2334 atomic_dec(&netstamp_needed_deferred); 2335 schedule_work(&netstamp_work); 2336 #else 2337 static_branch_dec(&netstamp_needed_key); 2338 #endif 2339 } 2340 EXPORT_SYMBOL(net_disable_timestamp); 2341 2342 static inline void net_timestamp_set(struct sk_buff *skb) 2343 { 2344 skb->tstamp = 0; 2345 skb->tstamp_type = SKB_CLOCK_REALTIME; 2346 if (static_branch_unlikely(&netstamp_needed_key)) 2347 skb->tstamp = ktime_get_real(); 2348 } 2349 2350 #define net_timestamp_check(COND, SKB) \ 2351 if (static_branch_unlikely(&netstamp_needed_key)) { \ 2352 if ((COND) && !(SKB)->tstamp) \ 2353 (SKB)->tstamp = ktime_get_real(); \ 2354 } \ 2355 2356 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb) 2357 { 2358 return __is_skb_forwardable(dev, skb, true); 2359 } 2360 EXPORT_SYMBOL_GPL(is_skb_forwardable); 2361 2362 static int __dev_forward_skb2(struct net_device *dev, struct sk_buff *skb, 2363 bool check_mtu) 2364 { 2365 int ret = ____dev_forward_skb(dev, skb, check_mtu); 2366 2367 if (likely(!ret)) { 2368 skb->protocol = eth_type_trans(skb, dev); 2369 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN); 2370 } 2371 2372 return ret; 2373 } 2374 2375 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2376 { 2377 return __dev_forward_skb2(dev, skb, true); 2378 } 2379 EXPORT_SYMBOL_GPL(__dev_forward_skb); 2380 2381 /** 2382 * dev_forward_skb - loopback an skb to another netif 2383 * 2384 * @dev: destination network device 2385 * @skb: buffer to forward 2386 * 2387 * return values: 2388 * NET_RX_SUCCESS (no congestion) 2389 * NET_RX_DROP (packet was dropped, but freed) 2390 * 2391 * dev_forward_skb can be used for injecting an skb from the 2392 * start_xmit function of one device into the receive queue 2393 * of another device. 2394 * 2395 * The receiving device may be in another namespace, so 2396 * we have to clear all information in the skb that could 2397 * impact namespace isolation. 2398 */ 2399 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2400 { 2401 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb); 2402 } 2403 EXPORT_SYMBOL_GPL(dev_forward_skb); 2404 2405 int dev_forward_skb_nomtu(struct net_device *dev, struct sk_buff *skb) 2406 { 2407 return __dev_forward_skb2(dev, skb, false) ?: netif_rx_internal(skb); 2408 } 2409 2410 static inline int deliver_skb(struct sk_buff *skb, 2411 struct packet_type *pt_prev, 2412 struct net_device *orig_dev) 2413 { 2414 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 2415 return -ENOMEM; 2416 refcount_inc(&skb->users); 2417 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 2418 } 2419 2420 static inline void deliver_ptype_list_skb(struct sk_buff *skb, 2421 struct packet_type **pt, 2422 struct net_device *orig_dev, 2423 __be16 type, 2424 struct list_head *ptype_list) 2425 { 2426 struct packet_type *ptype, *pt_prev = *pt; 2427 2428 list_for_each_entry_rcu(ptype, ptype_list, list) { 2429 if (ptype->type != type) 2430 continue; 2431 if (pt_prev) 2432 deliver_skb(skb, pt_prev, orig_dev); 2433 pt_prev = ptype; 2434 } 2435 *pt = pt_prev; 2436 } 2437 2438 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb) 2439 { 2440 if (!ptype->af_packet_priv || !skb->sk) 2441 return false; 2442 2443 if (ptype->id_match) 2444 return ptype->id_match(ptype, skb->sk); 2445 else if ((struct sock *)ptype->af_packet_priv == skb->sk) 2446 return true; 2447 2448 return false; 2449 } 2450 2451 /** 2452 * dev_nit_active_rcu - return true if any network interface taps are in use 2453 * 2454 * The caller must hold the RCU lock 2455 * 2456 * @dev: network device to check for the presence of taps 2457 */ 2458 bool dev_nit_active_rcu(const struct net_device *dev) 2459 { 2460 /* Callers may hold either RCU or RCU BH lock */ 2461 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 2462 2463 return !list_empty(&dev_net(dev)->ptype_all) || 2464 !list_empty(&dev->ptype_all); 2465 } 2466 EXPORT_SYMBOL_GPL(dev_nit_active_rcu); 2467 2468 /* 2469 * Support routine. Sends outgoing frames to any network 2470 * taps currently in use. 2471 */ 2472 2473 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev) 2474 { 2475 struct packet_type *ptype, *pt_prev = NULL; 2476 struct list_head *ptype_list; 2477 struct sk_buff *skb2 = NULL; 2478 2479 rcu_read_lock(); 2480 ptype_list = &dev_net_rcu(dev)->ptype_all; 2481 again: 2482 list_for_each_entry_rcu(ptype, ptype_list, list) { 2483 if (READ_ONCE(ptype->ignore_outgoing)) 2484 continue; 2485 2486 /* Never send packets back to the socket 2487 * they originated from - MvS (miquels@drinkel.ow.org) 2488 */ 2489 if (skb_loop_sk(ptype, skb)) 2490 continue; 2491 2492 if (pt_prev) { 2493 deliver_skb(skb2, pt_prev, skb->dev); 2494 pt_prev = ptype; 2495 continue; 2496 } 2497 2498 /* need to clone skb, done only once */ 2499 skb2 = skb_clone(skb, GFP_ATOMIC); 2500 if (!skb2) 2501 goto out_unlock; 2502 2503 net_timestamp_set(skb2); 2504 2505 /* skb->nh should be correctly 2506 * set by sender, so that the second statement is 2507 * just protection against buggy protocols. 2508 */ 2509 skb_reset_mac_header(skb2); 2510 2511 if (skb_network_header(skb2) < skb2->data || 2512 skb_network_header(skb2) > skb_tail_pointer(skb2)) { 2513 net_crit_ratelimited("protocol %04x is buggy, dev %s\n", 2514 ntohs(skb2->protocol), 2515 dev->name); 2516 skb_reset_network_header(skb2); 2517 } 2518 2519 skb2->transport_header = skb2->network_header; 2520 skb2->pkt_type = PACKET_OUTGOING; 2521 pt_prev = ptype; 2522 } 2523 2524 if (ptype_list != &dev->ptype_all) { 2525 ptype_list = &dev->ptype_all; 2526 goto again; 2527 } 2528 out_unlock: 2529 if (pt_prev) { 2530 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC)) 2531 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev); 2532 else 2533 kfree_skb(skb2); 2534 } 2535 rcu_read_unlock(); 2536 } 2537 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit); 2538 2539 /** 2540 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change 2541 * @dev: Network device 2542 * @txq: number of queues available 2543 * 2544 * If real_num_tx_queues is changed the tc mappings may no longer be 2545 * valid. To resolve this verify the tc mapping remains valid and if 2546 * not NULL the mapping. With no priorities mapping to this 2547 * offset/count pair it will no longer be used. In the worst case TC0 2548 * is invalid nothing can be done so disable priority mappings. If is 2549 * expected that drivers will fix this mapping if they can before 2550 * calling netif_set_real_num_tx_queues. 2551 */ 2552 static void netif_setup_tc(struct net_device *dev, unsigned int txq) 2553 { 2554 int i; 2555 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2556 2557 /* If TC0 is invalidated disable TC mapping */ 2558 if (tc->offset + tc->count > txq) { 2559 netdev_warn(dev, "Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n"); 2560 dev->num_tc = 0; 2561 return; 2562 } 2563 2564 /* Invalidated prio to tc mappings set to TC0 */ 2565 for (i = 1; i < TC_BITMASK + 1; i++) { 2566 int q = netdev_get_prio_tc_map(dev, i); 2567 2568 tc = &dev->tc_to_txq[q]; 2569 if (tc->offset + tc->count > txq) { 2570 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", 2571 i, q); 2572 netdev_set_prio_tc_map(dev, i, 0); 2573 } 2574 } 2575 } 2576 2577 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq) 2578 { 2579 if (dev->num_tc) { 2580 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2581 int i; 2582 2583 /* walk through the TCs and see if it falls into any of them */ 2584 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) { 2585 if ((txq - tc->offset) < tc->count) 2586 return i; 2587 } 2588 2589 /* didn't find it, just return -1 to indicate no match */ 2590 return -1; 2591 } 2592 2593 return 0; 2594 } 2595 EXPORT_SYMBOL(netdev_txq_to_tc); 2596 2597 #ifdef CONFIG_XPS 2598 static struct static_key xps_needed __read_mostly; 2599 static struct static_key xps_rxqs_needed __read_mostly; 2600 static DEFINE_MUTEX(xps_map_mutex); 2601 #define xmap_dereference(P) \ 2602 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex)) 2603 2604 static bool remove_xps_queue(struct xps_dev_maps *dev_maps, 2605 struct xps_dev_maps *old_maps, int tci, u16 index) 2606 { 2607 struct xps_map *map = NULL; 2608 int pos; 2609 2610 map = xmap_dereference(dev_maps->attr_map[tci]); 2611 if (!map) 2612 return false; 2613 2614 for (pos = map->len; pos--;) { 2615 if (map->queues[pos] != index) 2616 continue; 2617 2618 if (map->len > 1) { 2619 map->queues[pos] = map->queues[--map->len]; 2620 break; 2621 } 2622 2623 if (old_maps) 2624 RCU_INIT_POINTER(old_maps->attr_map[tci], NULL); 2625 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2626 kfree_rcu(map, rcu); 2627 return false; 2628 } 2629 2630 return true; 2631 } 2632 2633 static bool remove_xps_queue_cpu(struct net_device *dev, 2634 struct xps_dev_maps *dev_maps, 2635 int cpu, u16 offset, u16 count) 2636 { 2637 int num_tc = dev_maps->num_tc; 2638 bool active = false; 2639 int tci; 2640 2641 for (tci = cpu * num_tc; num_tc--; tci++) { 2642 int i, j; 2643 2644 for (i = count, j = offset; i--; j++) { 2645 if (!remove_xps_queue(dev_maps, NULL, tci, j)) 2646 break; 2647 } 2648 2649 active |= i < 0; 2650 } 2651 2652 return active; 2653 } 2654 2655 static void reset_xps_maps(struct net_device *dev, 2656 struct xps_dev_maps *dev_maps, 2657 enum xps_map_type type) 2658 { 2659 static_key_slow_dec_cpuslocked(&xps_needed); 2660 if (type == XPS_RXQS) 2661 static_key_slow_dec_cpuslocked(&xps_rxqs_needed); 2662 2663 RCU_INIT_POINTER(dev->xps_maps[type], NULL); 2664 2665 kfree_rcu(dev_maps, rcu); 2666 } 2667 2668 static void clean_xps_maps(struct net_device *dev, enum xps_map_type type, 2669 u16 offset, u16 count) 2670 { 2671 struct xps_dev_maps *dev_maps; 2672 bool active = false; 2673 int i, j; 2674 2675 dev_maps = xmap_dereference(dev->xps_maps[type]); 2676 if (!dev_maps) 2677 return; 2678 2679 for (j = 0; j < dev_maps->nr_ids; j++) 2680 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, count); 2681 if (!active) 2682 reset_xps_maps(dev, dev_maps, type); 2683 2684 if (type == XPS_CPUS) { 2685 for (i = offset + (count - 1); count--; i--) 2686 netdev_queue_numa_node_write( 2687 netdev_get_tx_queue(dev, i), NUMA_NO_NODE); 2688 } 2689 } 2690 2691 static void netif_reset_xps_queues(struct net_device *dev, u16 offset, 2692 u16 count) 2693 { 2694 if (!static_key_false(&xps_needed)) 2695 return; 2696 2697 cpus_read_lock(); 2698 mutex_lock(&xps_map_mutex); 2699 2700 if (static_key_false(&xps_rxqs_needed)) 2701 clean_xps_maps(dev, XPS_RXQS, offset, count); 2702 2703 clean_xps_maps(dev, XPS_CPUS, offset, count); 2704 2705 mutex_unlock(&xps_map_mutex); 2706 cpus_read_unlock(); 2707 } 2708 2709 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index) 2710 { 2711 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index); 2712 } 2713 2714 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index, 2715 u16 index, bool is_rxqs_map) 2716 { 2717 struct xps_map *new_map; 2718 int alloc_len = XPS_MIN_MAP_ALLOC; 2719 int i, pos; 2720 2721 for (pos = 0; map && pos < map->len; pos++) { 2722 if (map->queues[pos] != index) 2723 continue; 2724 return map; 2725 } 2726 2727 /* Need to add tx-queue to this CPU's/rx-queue's existing map */ 2728 if (map) { 2729 if (pos < map->alloc_len) 2730 return map; 2731 2732 alloc_len = map->alloc_len * 2; 2733 } 2734 2735 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's 2736 * map 2737 */ 2738 if (is_rxqs_map) 2739 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL); 2740 else 2741 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL, 2742 cpu_to_node(attr_index)); 2743 if (!new_map) 2744 return NULL; 2745 2746 for (i = 0; i < pos; i++) 2747 new_map->queues[i] = map->queues[i]; 2748 new_map->alloc_len = alloc_len; 2749 new_map->len = pos; 2750 2751 return new_map; 2752 } 2753 2754 /* Copy xps maps at a given index */ 2755 static void xps_copy_dev_maps(struct xps_dev_maps *dev_maps, 2756 struct xps_dev_maps *new_dev_maps, int index, 2757 int tc, bool skip_tc) 2758 { 2759 int i, tci = index * dev_maps->num_tc; 2760 struct xps_map *map; 2761 2762 /* copy maps belonging to foreign traffic classes */ 2763 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2764 if (i == tc && skip_tc) 2765 continue; 2766 2767 /* fill in the new device map from the old device map */ 2768 map = xmap_dereference(dev_maps->attr_map[tci]); 2769 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2770 } 2771 } 2772 2773 /* Must be called under cpus_read_lock */ 2774 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask, 2775 u16 index, enum xps_map_type type) 2776 { 2777 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL, *old_dev_maps = NULL; 2778 const unsigned long *online_mask = NULL; 2779 bool active = false, copy = false; 2780 int i, j, tci, numa_node_id = -2; 2781 int maps_sz, num_tc = 1, tc = 0; 2782 struct xps_map *map, *new_map; 2783 unsigned int nr_ids; 2784 2785 WARN_ON_ONCE(index >= dev->num_tx_queues); 2786 2787 if (dev->num_tc) { 2788 /* Do not allow XPS on subordinate device directly */ 2789 num_tc = dev->num_tc; 2790 if (num_tc < 0) 2791 return -EINVAL; 2792 2793 /* If queue belongs to subordinate dev use its map */ 2794 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev; 2795 2796 tc = netdev_txq_to_tc(dev, index); 2797 if (tc < 0) 2798 return -EINVAL; 2799 } 2800 2801 mutex_lock(&xps_map_mutex); 2802 2803 dev_maps = xmap_dereference(dev->xps_maps[type]); 2804 if (type == XPS_RXQS) { 2805 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues); 2806 nr_ids = dev->num_rx_queues; 2807 } else { 2808 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc); 2809 if (num_possible_cpus() > 1) 2810 online_mask = cpumask_bits(cpu_online_mask); 2811 nr_ids = nr_cpu_ids; 2812 } 2813 2814 if (maps_sz < L1_CACHE_BYTES) 2815 maps_sz = L1_CACHE_BYTES; 2816 2817 /* The old dev_maps could be larger or smaller than the one we're 2818 * setting up now, as dev->num_tc or nr_ids could have been updated in 2819 * between. We could try to be smart, but let's be safe instead and only 2820 * copy foreign traffic classes if the two map sizes match. 2821 */ 2822 if (dev_maps && 2823 dev_maps->num_tc == num_tc && dev_maps->nr_ids == nr_ids) 2824 copy = true; 2825 2826 /* allocate memory for queue storage */ 2827 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids), 2828 j < nr_ids;) { 2829 if (!new_dev_maps) { 2830 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL); 2831 if (!new_dev_maps) { 2832 mutex_unlock(&xps_map_mutex); 2833 return -ENOMEM; 2834 } 2835 2836 new_dev_maps->nr_ids = nr_ids; 2837 new_dev_maps->num_tc = num_tc; 2838 } 2839 2840 tci = j * num_tc + tc; 2841 map = copy ? xmap_dereference(dev_maps->attr_map[tci]) : NULL; 2842 2843 map = expand_xps_map(map, j, index, type == XPS_RXQS); 2844 if (!map) 2845 goto error; 2846 2847 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2848 } 2849 2850 if (!new_dev_maps) 2851 goto out_no_new_maps; 2852 2853 if (!dev_maps) { 2854 /* Increment static keys at most once per type */ 2855 static_key_slow_inc_cpuslocked(&xps_needed); 2856 if (type == XPS_RXQS) 2857 static_key_slow_inc_cpuslocked(&xps_rxqs_needed); 2858 } 2859 2860 for (j = 0; j < nr_ids; j++) { 2861 bool skip_tc = false; 2862 2863 tci = j * num_tc + tc; 2864 if (netif_attr_test_mask(j, mask, nr_ids) && 2865 netif_attr_test_online(j, online_mask, nr_ids)) { 2866 /* add tx-queue to CPU/rx-queue maps */ 2867 int pos = 0; 2868 2869 skip_tc = true; 2870 2871 map = xmap_dereference(new_dev_maps->attr_map[tci]); 2872 while ((pos < map->len) && (map->queues[pos] != index)) 2873 pos++; 2874 2875 if (pos == map->len) 2876 map->queues[map->len++] = index; 2877 #ifdef CONFIG_NUMA 2878 if (type == XPS_CPUS) { 2879 if (numa_node_id == -2) 2880 numa_node_id = cpu_to_node(j); 2881 else if (numa_node_id != cpu_to_node(j)) 2882 numa_node_id = -1; 2883 } 2884 #endif 2885 } 2886 2887 if (copy) 2888 xps_copy_dev_maps(dev_maps, new_dev_maps, j, tc, 2889 skip_tc); 2890 } 2891 2892 rcu_assign_pointer(dev->xps_maps[type], new_dev_maps); 2893 2894 /* Cleanup old maps */ 2895 if (!dev_maps) 2896 goto out_no_old_maps; 2897 2898 for (j = 0; j < dev_maps->nr_ids; j++) { 2899 for (i = num_tc, tci = j * dev_maps->num_tc; i--; tci++) { 2900 map = xmap_dereference(dev_maps->attr_map[tci]); 2901 if (!map) 2902 continue; 2903 2904 if (copy) { 2905 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2906 if (map == new_map) 2907 continue; 2908 } 2909 2910 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2911 kfree_rcu(map, rcu); 2912 } 2913 } 2914 2915 old_dev_maps = dev_maps; 2916 2917 out_no_old_maps: 2918 dev_maps = new_dev_maps; 2919 active = true; 2920 2921 out_no_new_maps: 2922 if (type == XPS_CPUS) 2923 /* update Tx queue numa node */ 2924 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index), 2925 (numa_node_id >= 0) ? 2926 numa_node_id : NUMA_NO_NODE); 2927 2928 if (!dev_maps) 2929 goto out_no_maps; 2930 2931 /* removes tx-queue from unused CPUs/rx-queues */ 2932 for (j = 0; j < dev_maps->nr_ids; j++) { 2933 tci = j * dev_maps->num_tc; 2934 2935 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2936 if (i == tc && 2937 netif_attr_test_mask(j, mask, dev_maps->nr_ids) && 2938 netif_attr_test_online(j, online_mask, dev_maps->nr_ids)) 2939 continue; 2940 2941 active |= remove_xps_queue(dev_maps, 2942 copy ? old_dev_maps : NULL, 2943 tci, index); 2944 } 2945 } 2946 2947 if (old_dev_maps) 2948 kfree_rcu(old_dev_maps, rcu); 2949 2950 /* free map if not active */ 2951 if (!active) 2952 reset_xps_maps(dev, dev_maps, type); 2953 2954 out_no_maps: 2955 mutex_unlock(&xps_map_mutex); 2956 2957 return 0; 2958 error: 2959 /* remove any maps that we added */ 2960 for (j = 0; j < nr_ids; j++) { 2961 for (i = num_tc, tci = j * num_tc; i--; tci++) { 2962 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2963 map = copy ? 2964 xmap_dereference(dev_maps->attr_map[tci]) : 2965 NULL; 2966 if (new_map && new_map != map) 2967 kfree(new_map); 2968 } 2969 } 2970 2971 mutex_unlock(&xps_map_mutex); 2972 2973 kfree(new_dev_maps); 2974 return -ENOMEM; 2975 } 2976 EXPORT_SYMBOL_GPL(__netif_set_xps_queue); 2977 2978 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, 2979 u16 index) 2980 { 2981 int ret; 2982 2983 cpus_read_lock(); 2984 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, XPS_CPUS); 2985 cpus_read_unlock(); 2986 2987 return ret; 2988 } 2989 EXPORT_SYMBOL(netif_set_xps_queue); 2990 2991 #endif 2992 static void netdev_unbind_all_sb_channels(struct net_device *dev) 2993 { 2994 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 2995 2996 /* Unbind any subordinate channels */ 2997 while (txq-- != &dev->_tx[0]) { 2998 if (txq->sb_dev) 2999 netdev_unbind_sb_channel(dev, txq->sb_dev); 3000 } 3001 } 3002 3003 void netdev_reset_tc(struct net_device *dev) 3004 { 3005 #ifdef CONFIG_XPS 3006 netif_reset_xps_queues_gt(dev, 0); 3007 #endif 3008 netdev_unbind_all_sb_channels(dev); 3009 3010 /* Reset TC configuration of device */ 3011 dev->num_tc = 0; 3012 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq)); 3013 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map)); 3014 } 3015 EXPORT_SYMBOL(netdev_reset_tc); 3016 3017 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset) 3018 { 3019 if (tc >= dev->num_tc) 3020 return -EINVAL; 3021 3022 #ifdef CONFIG_XPS 3023 netif_reset_xps_queues(dev, offset, count); 3024 #endif 3025 dev->tc_to_txq[tc].count = count; 3026 dev->tc_to_txq[tc].offset = offset; 3027 return 0; 3028 } 3029 EXPORT_SYMBOL(netdev_set_tc_queue); 3030 3031 int netdev_set_num_tc(struct net_device *dev, u8 num_tc) 3032 { 3033 if (num_tc > TC_MAX_QUEUE) 3034 return -EINVAL; 3035 3036 #ifdef CONFIG_XPS 3037 netif_reset_xps_queues_gt(dev, 0); 3038 #endif 3039 netdev_unbind_all_sb_channels(dev); 3040 3041 dev->num_tc = num_tc; 3042 return 0; 3043 } 3044 EXPORT_SYMBOL(netdev_set_num_tc); 3045 3046 void netdev_unbind_sb_channel(struct net_device *dev, 3047 struct net_device *sb_dev) 3048 { 3049 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 3050 3051 #ifdef CONFIG_XPS 3052 netif_reset_xps_queues_gt(sb_dev, 0); 3053 #endif 3054 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq)); 3055 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map)); 3056 3057 while (txq-- != &dev->_tx[0]) { 3058 if (txq->sb_dev == sb_dev) 3059 txq->sb_dev = NULL; 3060 } 3061 } 3062 EXPORT_SYMBOL(netdev_unbind_sb_channel); 3063 3064 int netdev_bind_sb_channel_queue(struct net_device *dev, 3065 struct net_device *sb_dev, 3066 u8 tc, u16 count, u16 offset) 3067 { 3068 /* Make certain the sb_dev and dev are already configured */ 3069 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc) 3070 return -EINVAL; 3071 3072 /* We cannot hand out queues we don't have */ 3073 if ((offset + count) > dev->real_num_tx_queues) 3074 return -EINVAL; 3075 3076 /* Record the mapping */ 3077 sb_dev->tc_to_txq[tc].count = count; 3078 sb_dev->tc_to_txq[tc].offset = offset; 3079 3080 /* Provide a way for Tx queue to find the tc_to_txq map or 3081 * XPS map for itself. 3082 */ 3083 while (count--) 3084 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev; 3085 3086 return 0; 3087 } 3088 EXPORT_SYMBOL(netdev_bind_sb_channel_queue); 3089 3090 int netdev_set_sb_channel(struct net_device *dev, u16 channel) 3091 { 3092 /* Do not use a multiqueue device to represent a subordinate channel */ 3093 if (netif_is_multiqueue(dev)) 3094 return -ENODEV; 3095 3096 /* We allow channels 1 - 32767 to be used for subordinate channels. 3097 * Channel 0 is meant to be "native" mode and used only to represent 3098 * the main root device. We allow writing 0 to reset the device back 3099 * to normal mode after being used as a subordinate channel. 3100 */ 3101 if (channel > S16_MAX) 3102 return -EINVAL; 3103 3104 dev->num_tc = -channel; 3105 3106 return 0; 3107 } 3108 EXPORT_SYMBOL(netdev_set_sb_channel); 3109 3110 /* 3111 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues 3112 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed. 3113 */ 3114 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq) 3115 { 3116 bool disabling; 3117 int rc; 3118 3119 disabling = txq < dev->real_num_tx_queues; 3120 3121 if (txq < 1 || txq > dev->num_tx_queues) 3122 return -EINVAL; 3123 3124 if (dev->reg_state == NETREG_REGISTERED || 3125 dev->reg_state == NETREG_UNREGISTERING) { 3126 ASSERT_RTNL(); 3127 netdev_ops_assert_locked(dev); 3128 3129 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues, 3130 txq); 3131 if (rc) 3132 return rc; 3133 3134 if (dev->num_tc) 3135 netif_setup_tc(dev, txq); 3136 3137 net_shaper_set_real_num_tx_queues(dev, txq); 3138 3139 dev_qdisc_change_real_num_tx(dev, txq); 3140 3141 dev->real_num_tx_queues = txq; 3142 3143 if (disabling) { 3144 synchronize_net(); 3145 qdisc_reset_all_tx_gt(dev, txq); 3146 #ifdef CONFIG_XPS 3147 netif_reset_xps_queues_gt(dev, txq); 3148 #endif 3149 } 3150 } else { 3151 dev->real_num_tx_queues = txq; 3152 } 3153 3154 return 0; 3155 } 3156 EXPORT_SYMBOL(netif_set_real_num_tx_queues); 3157 3158 /** 3159 * netif_set_real_num_rx_queues - set actual number of RX queues used 3160 * @dev: Network device 3161 * @rxq: Actual number of RX queues 3162 * 3163 * This must be called either with the rtnl_lock held or before 3164 * registration of the net device. Returns 0 on success, or a 3165 * negative error code. If called before registration, it always 3166 * succeeds. 3167 */ 3168 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq) 3169 { 3170 int rc; 3171 3172 if (rxq < 1 || rxq > dev->num_rx_queues) 3173 return -EINVAL; 3174 3175 if (dev->reg_state == NETREG_REGISTERED) { 3176 ASSERT_RTNL(); 3177 netdev_ops_assert_locked(dev); 3178 3179 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues, 3180 rxq); 3181 if (rc) 3182 return rc; 3183 } 3184 3185 dev->real_num_rx_queues = rxq; 3186 return 0; 3187 } 3188 EXPORT_SYMBOL(netif_set_real_num_rx_queues); 3189 3190 /** 3191 * netif_set_real_num_queues - set actual number of RX and TX queues used 3192 * @dev: Network device 3193 * @txq: Actual number of TX queues 3194 * @rxq: Actual number of RX queues 3195 * 3196 * Set the real number of both TX and RX queues. 3197 * Does nothing if the number of queues is already correct. 3198 */ 3199 int netif_set_real_num_queues(struct net_device *dev, 3200 unsigned int txq, unsigned int rxq) 3201 { 3202 unsigned int old_rxq = dev->real_num_rx_queues; 3203 int err; 3204 3205 if (txq < 1 || txq > dev->num_tx_queues || 3206 rxq < 1 || rxq > dev->num_rx_queues) 3207 return -EINVAL; 3208 3209 /* Start from increases, so the error path only does decreases - 3210 * decreases can't fail. 3211 */ 3212 if (rxq > dev->real_num_rx_queues) { 3213 err = netif_set_real_num_rx_queues(dev, rxq); 3214 if (err) 3215 return err; 3216 } 3217 if (txq > dev->real_num_tx_queues) { 3218 err = netif_set_real_num_tx_queues(dev, txq); 3219 if (err) 3220 goto undo_rx; 3221 } 3222 if (rxq < dev->real_num_rx_queues) 3223 WARN_ON(netif_set_real_num_rx_queues(dev, rxq)); 3224 if (txq < dev->real_num_tx_queues) 3225 WARN_ON(netif_set_real_num_tx_queues(dev, txq)); 3226 3227 return 0; 3228 undo_rx: 3229 WARN_ON(netif_set_real_num_rx_queues(dev, old_rxq)); 3230 return err; 3231 } 3232 EXPORT_SYMBOL(netif_set_real_num_queues); 3233 3234 /** 3235 * netif_set_tso_max_size() - set the max size of TSO frames supported 3236 * @dev: netdev to update 3237 * @size: max skb->len of a TSO frame 3238 * 3239 * Set the limit on the size of TSO super-frames the device can handle. 3240 * Unless explicitly set the stack will assume the value of 3241 * %GSO_LEGACY_MAX_SIZE. 3242 */ 3243 void netif_set_tso_max_size(struct net_device *dev, unsigned int size) 3244 { 3245 dev->tso_max_size = min(GSO_MAX_SIZE, size); 3246 if (size < READ_ONCE(dev->gso_max_size)) 3247 netif_set_gso_max_size(dev, size); 3248 if (size < READ_ONCE(dev->gso_ipv4_max_size)) 3249 netif_set_gso_ipv4_max_size(dev, size); 3250 } 3251 EXPORT_SYMBOL(netif_set_tso_max_size); 3252 3253 /** 3254 * netif_set_tso_max_segs() - set the max number of segs supported for TSO 3255 * @dev: netdev to update 3256 * @segs: max number of TCP segments 3257 * 3258 * Set the limit on the number of TCP segments the device can generate from 3259 * a single TSO super-frame. 3260 * Unless explicitly set the stack will assume the value of %GSO_MAX_SEGS. 3261 */ 3262 void netif_set_tso_max_segs(struct net_device *dev, unsigned int segs) 3263 { 3264 dev->tso_max_segs = segs; 3265 if (segs < READ_ONCE(dev->gso_max_segs)) 3266 netif_set_gso_max_segs(dev, segs); 3267 } 3268 EXPORT_SYMBOL(netif_set_tso_max_segs); 3269 3270 /** 3271 * netif_inherit_tso_max() - copy all TSO limits from a lower device to an upper 3272 * @to: netdev to update 3273 * @from: netdev from which to copy the limits 3274 */ 3275 void netif_inherit_tso_max(struct net_device *to, const struct net_device *from) 3276 { 3277 netif_set_tso_max_size(to, from->tso_max_size); 3278 netif_set_tso_max_segs(to, from->tso_max_segs); 3279 } 3280 EXPORT_SYMBOL(netif_inherit_tso_max); 3281 3282 /** 3283 * netif_get_num_default_rss_queues - default number of RSS queues 3284 * 3285 * Default value is the number of physical cores if there are only 1 or 2, or 3286 * divided by 2 if there are more. 3287 */ 3288 int netif_get_num_default_rss_queues(void) 3289 { 3290 cpumask_var_t cpus; 3291 int cpu, count = 0; 3292 3293 if (unlikely(is_kdump_kernel() || !zalloc_cpumask_var(&cpus, GFP_KERNEL))) 3294 return 1; 3295 3296 cpumask_copy(cpus, cpu_online_mask); 3297 for_each_cpu(cpu, cpus) { 3298 ++count; 3299 cpumask_andnot(cpus, cpus, topology_sibling_cpumask(cpu)); 3300 } 3301 free_cpumask_var(cpus); 3302 3303 return count > 2 ? DIV_ROUND_UP(count, 2) : count; 3304 } 3305 EXPORT_SYMBOL(netif_get_num_default_rss_queues); 3306 3307 static void __netif_reschedule(struct Qdisc *q) 3308 { 3309 struct softnet_data *sd; 3310 unsigned long flags; 3311 3312 local_irq_save(flags); 3313 sd = this_cpu_ptr(&softnet_data); 3314 q->next_sched = NULL; 3315 *sd->output_queue_tailp = q; 3316 sd->output_queue_tailp = &q->next_sched; 3317 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3318 local_irq_restore(flags); 3319 } 3320 3321 void __netif_schedule(struct Qdisc *q) 3322 { 3323 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state)) 3324 __netif_reschedule(q); 3325 } 3326 EXPORT_SYMBOL(__netif_schedule); 3327 3328 struct dev_kfree_skb_cb { 3329 enum skb_drop_reason reason; 3330 }; 3331 3332 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb) 3333 { 3334 return (struct dev_kfree_skb_cb *)skb->cb; 3335 } 3336 3337 void netif_schedule_queue(struct netdev_queue *txq) 3338 { 3339 rcu_read_lock(); 3340 if (!netif_xmit_stopped(txq)) { 3341 struct Qdisc *q = rcu_dereference(txq->qdisc); 3342 3343 __netif_schedule(q); 3344 } 3345 rcu_read_unlock(); 3346 } 3347 EXPORT_SYMBOL(netif_schedule_queue); 3348 3349 void netif_tx_wake_queue(struct netdev_queue *dev_queue) 3350 { 3351 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) { 3352 struct Qdisc *q; 3353 3354 rcu_read_lock(); 3355 q = rcu_dereference(dev_queue->qdisc); 3356 __netif_schedule(q); 3357 rcu_read_unlock(); 3358 } 3359 } 3360 EXPORT_SYMBOL(netif_tx_wake_queue); 3361 3362 void dev_kfree_skb_irq_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3363 { 3364 unsigned long flags; 3365 3366 if (unlikely(!skb)) 3367 return; 3368 3369 if (likely(refcount_read(&skb->users) == 1)) { 3370 smp_rmb(); 3371 refcount_set(&skb->users, 0); 3372 } else if (likely(!refcount_dec_and_test(&skb->users))) { 3373 return; 3374 } 3375 get_kfree_skb_cb(skb)->reason = reason; 3376 local_irq_save(flags); 3377 skb->next = __this_cpu_read(softnet_data.completion_queue); 3378 __this_cpu_write(softnet_data.completion_queue, skb); 3379 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3380 local_irq_restore(flags); 3381 } 3382 EXPORT_SYMBOL(dev_kfree_skb_irq_reason); 3383 3384 void dev_kfree_skb_any_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3385 { 3386 if (in_hardirq() || irqs_disabled()) 3387 dev_kfree_skb_irq_reason(skb, reason); 3388 else 3389 kfree_skb_reason(skb, reason); 3390 } 3391 EXPORT_SYMBOL(dev_kfree_skb_any_reason); 3392 3393 3394 /** 3395 * netif_device_detach - mark device as removed 3396 * @dev: network device 3397 * 3398 * Mark device as removed from system and therefore no longer available. 3399 */ 3400 void netif_device_detach(struct net_device *dev) 3401 { 3402 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) && 3403 netif_running(dev)) { 3404 netif_tx_stop_all_queues(dev); 3405 } 3406 } 3407 EXPORT_SYMBOL(netif_device_detach); 3408 3409 /** 3410 * netif_device_attach - mark device as attached 3411 * @dev: network device 3412 * 3413 * Mark device as attached from system and restart if needed. 3414 */ 3415 void netif_device_attach(struct net_device *dev) 3416 { 3417 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) && 3418 netif_running(dev)) { 3419 netif_tx_wake_all_queues(dev); 3420 netdev_watchdog_up(dev); 3421 } 3422 } 3423 EXPORT_SYMBOL(netif_device_attach); 3424 3425 /* 3426 * Returns a Tx hash based on the given packet descriptor a Tx queues' number 3427 * to be used as a distribution range. 3428 */ 3429 static u16 skb_tx_hash(const struct net_device *dev, 3430 const struct net_device *sb_dev, 3431 struct sk_buff *skb) 3432 { 3433 u32 hash; 3434 u16 qoffset = 0; 3435 u16 qcount = dev->real_num_tx_queues; 3436 3437 if (dev->num_tc) { 3438 u8 tc = netdev_get_prio_tc_map(dev, skb->priority); 3439 3440 qoffset = sb_dev->tc_to_txq[tc].offset; 3441 qcount = sb_dev->tc_to_txq[tc].count; 3442 if (unlikely(!qcount)) { 3443 net_warn_ratelimited("%s: invalid qcount, qoffset %u for tc %u\n", 3444 sb_dev->name, qoffset, tc); 3445 qoffset = 0; 3446 qcount = dev->real_num_tx_queues; 3447 } 3448 } 3449 3450 if (skb_rx_queue_recorded(skb)) { 3451 DEBUG_NET_WARN_ON_ONCE(qcount == 0); 3452 hash = skb_get_rx_queue(skb); 3453 if (hash >= qoffset) 3454 hash -= qoffset; 3455 while (unlikely(hash >= qcount)) 3456 hash -= qcount; 3457 return hash + qoffset; 3458 } 3459 3460 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset; 3461 } 3462 3463 void skb_warn_bad_offload(const struct sk_buff *skb) 3464 { 3465 static const netdev_features_t null_features; 3466 struct net_device *dev = skb->dev; 3467 const char *name = ""; 3468 3469 if (!net_ratelimit()) 3470 return; 3471 3472 if (dev) { 3473 if (dev->dev.parent) 3474 name = dev_driver_string(dev->dev.parent); 3475 else 3476 name = netdev_name(dev); 3477 } 3478 skb_dump(KERN_WARNING, skb, false); 3479 WARN(1, "%s: caps=(%pNF, %pNF)\n", 3480 name, dev ? &dev->features : &null_features, 3481 skb->sk ? &skb->sk->sk_route_caps : &null_features); 3482 } 3483 3484 /* 3485 * Invalidate hardware checksum when packet is to be mangled, and 3486 * complete checksum manually on outgoing path. 3487 */ 3488 int skb_checksum_help(struct sk_buff *skb) 3489 { 3490 __wsum csum; 3491 int ret = 0, offset; 3492 3493 if (skb->ip_summed == CHECKSUM_COMPLETE) 3494 goto out_set_summed; 3495 3496 if (unlikely(skb_is_gso(skb))) { 3497 skb_warn_bad_offload(skb); 3498 return -EINVAL; 3499 } 3500 3501 if (!skb_frags_readable(skb)) { 3502 return -EFAULT; 3503 } 3504 3505 /* Before computing a checksum, we should make sure no frag could 3506 * be modified by an external entity : checksum could be wrong. 3507 */ 3508 if (skb_has_shared_frag(skb)) { 3509 ret = __skb_linearize(skb); 3510 if (ret) 3511 goto out; 3512 } 3513 3514 offset = skb_checksum_start_offset(skb); 3515 ret = -EINVAL; 3516 if (unlikely(offset >= skb_headlen(skb))) { 3517 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3518 WARN_ONCE(true, "offset (%d) >= skb_headlen() (%u)\n", 3519 offset, skb_headlen(skb)); 3520 goto out; 3521 } 3522 csum = skb_checksum(skb, offset, skb->len - offset, 0); 3523 3524 offset += skb->csum_offset; 3525 if (unlikely(offset + sizeof(__sum16) > skb_headlen(skb))) { 3526 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3527 WARN_ONCE(true, "offset+2 (%zu) > skb_headlen() (%u)\n", 3528 offset + sizeof(__sum16), skb_headlen(skb)); 3529 goto out; 3530 } 3531 ret = skb_ensure_writable(skb, offset + sizeof(__sum16)); 3532 if (ret) 3533 goto out; 3534 3535 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0; 3536 out_set_summed: 3537 skb->ip_summed = CHECKSUM_NONE; 3538 out: 3539 return ret; 3540 } 3541 EXPORT_SYMBOL(skb_checksum_help); 3542 3543 int skb_crc32c_csum_help(struct sk_buff *skb) 3544 { 3545 __le32 crc32c_csum; 3546 int ret = 0, offset, start; 3547 3548 if (skb->ip_summed != CHECKSUM_PARTIAL) 3549 goto out; 3550 3551 if (unlikely(skb_is_gso(skb))) 3552 goto out; 3553 3554 /* Before computing a checksum, we should make sure no frag could 3555 * be modified by an external entity : checksum could be wrong. 3556 */ 3557 if (unlikely(skb_has_shared_frag(skb))) { 3558 ret = __skb_linearize(skb); 3559 if (ret) 3560 goto out; 3561 } 3562 start = skb_checksum_start_offset(skb); 3563 offset = start + offsetof(struct sctphdr, checksum); 3564 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) { 3565 ret = -EINVAL; 3566 goto out; 3567 } 3568 3569 ret = skb_ensure_writable(skb, offset + sizeof(__le32)); 3570 if (ret) 3571 goto out; 3572 3573 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start, 3574 skb->len - start, ~(__u32)0, 3575 crc32c_csum_stub)); 3576 *(__le32 *)(skb->data + offset) = crc32c_csum; 3577 skb_reset_csum_not_inet(skb); 3578 out: 3579 return ret; 3580 } 3581 EXPORT_SYMBOL(skb_crc32c_csum_help); 3582 3583 __be16 skb_network_protocol(struct sk_buff *skb, int *depth) 3584 { 3585 __be16 type = skb->protocol; 3586 3587 /* Tunnel gso handlers can set protocol to ethernet. */ 3588 if (type == htons(ETH_P_TEB)) { 3589 struct ethhdr *eth; 3590 3591 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr)))) 3592 return 0; 3593 3594 eth = (struct ethhdr *)skb->data; 3595 type = eth->h_proto; 3596 } 3597 3598 return vlan_get_protocol_and_depth(skb, type, depth); 3599 } 3600 3601 3602 /* Take action when hardware reception checksum errors are detected. */ 3603 #ifdef CONFIG_BUG 3604 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3605 { 3606 netdev_err(dev, "hw csum failure\n"); 3607 skb_dump(KERN_ERR, skb, true); 3608 dump_stack(); 3609 } 3610 3611 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3612 { 3613 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb); 3614 } 3615 EXPORT_SYMBOL(netdev_rx_csum_fault); 3616 #endif 3617 3618 /* XXX: check that highmem exists at all on the given machine. */ 3619 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb) 3620 { 3621 #ifdef CONFIG_HIGHMEM 3622 int i; 3623 3624 if (!(dev->features & NETIF_F_HIGHDMA)) { 3625 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 3626 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 3627 struct page *page = skb_frag_page(frag); 3628 3629 if (page && PageHighMem(page)) 3630 return 1; 3631 } 3632 } 3633 #endif 3634 return 0; 3635 } 3636 3637 /* If MPLS offload request, verify we are testing hardware MPLS features 3638 * instead of standard features for the netdev. 3639 */ 3640 #if IS_ENABLED(CONFIG_NET_MPLS_GSO) 3641 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3642 netdev_features_t features, 3643 __be16 type) 3644 { 3645 if (eth_p_mpls(type)) 3646 features &= skb->dev->mpls_features; 3647 3648 return features; 3649 } 3650 #else 3651 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3652 netdev_features_t features, 3653 __be16 type) 3654 { 3655 return features; 3656 } 3657 #endif 3658 3659 static netdev_features_t harmonize_features(struct sk_buff *skb, 3660 netdev_features_t features) 3661 { 3662 __be16 type; 3663 3664 type = skb_network_protocol(skb, NULL); 3665 features = net_mpls_features(skb, features, type); 3666 3667 if (skb->ip_summed != CHECKSUM_NONE && 3668 !can_checksum_protocol(features, type)) { 3669 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3670 } 3671 if (illegal_highdma(skb->dev, skb)) 3672 features &= ~NETIF_F_SG; 3673 3674 return features; 3675 } 3676 3677 netdev_features_t passthru_features_check(struct sk_buff *skb, 3678 struct net_device *dev, 3679 netdev_features_t features) 3680 { 3681 return features; 3682 } 3683 EXPORT_SYMBOL(passthru_features_check); 3684 3685 static netdev_features_t dflt_features_check(struct sk_buff *skb, 3686 struct net_device *dev, 3687 netdev_features_t features) 3688 { 3689 return vlan_features_check(skb, features); 3690 } 3691 3692 static netdev_features_t gso_features_check(const struct sk_buff *skb, 3693 struct net_device *dev, 3694 netdev_features_t features) 3695 { 3696 u16 gso_segs = skb_shinfo(skb)->gso_segs; 3697 3698 if (gso_segs > READ_ONCE(dev->gso_max_segs)) 3699 return features & ~NETIF_F_GSO_MASK; 3700 3701 if (unlikely(skb->len >= netif_get_gso_max_size(dev, skb))) 3702 return features & ~NETIF_F_GSO_MASK; 3703 3704 if (!skb_shinfo(skb)->gso_type) { 3705 skb_warn_bad_offload(skb); 3706 return features & ~NETIF_F_GSO_MASK; 3707 } 3708 3709 /* Support for GSO partial features requires software 3710 * intervention before we can actually process the packets 3711 * so we need to strip support for any partial features now 3712 * and we can pull them back in after we have partially 3713 * segmented the frame. 3714 */ 3715 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)) 3716 features &= ~dev->gso_partial_features; 3717 3718 /* Make sure to clear the IPv4 ID mangling feature if the 3719 * IPv4 header has the potential to be fragmented. 3720 */ 3721 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) { 3722 struct iphdr *iph = skb->encapsulation ? 3723 inner_ip_hdr(skb) : ip_hdr(skb); 3724 3725 if (!(iph->frag_off & htons(IP_DF))) 3726 features &= ~NETIF_F_TSO_MANGLEID; 3727 } 3728 3729 return features; 3730 } 3731 3732 netdev_features_t netif_skb_features(struct sk_buff *skb) 3733 { 3734 struct net_device *dev = skb->dev; 3735 netdev_features_t features = dev->features; 3736 3737 if (skb_is_gso(skb)) 3738 features = gso_features_check(skb, dev, features); 3739 3740 /* If encapsulation offload request, verify we are testing 3741 * hardware encapsulation features instead of standard 3742 * features for the netdev 3743 */ 3744 if (skb->encapsulation) 3745 features &= dev->hw_enc_features; 3746 3747 if (skb_vlan_tagged(skb)) 3748 features = netdev_intersect_features(features, 3749 dev->vlan_features | 3750 NETIF_F_HW_VLAN_CTAG_TX | 3751 NETIF_F_HW_VLAN_STAG_TX); 3752 3753 if (dev->netdev_ops->ndo_features_check) 3754 features &= dev->netdev_ops->ndo_features_check(skb, dev, 3755 features); 3756 else 3757 features &= dflt_features_check(skb, dev, features); 3758 3759 return harmonize_features(skb, features); 3760 } 3761 EXPORT_SYMBOL(netif_skb_features); 3762 3763 static int xmit_one(struct sk_buff *skb, struct net_device *dev, 3764 struct netdev_queue *txq, bool more) 3765 { 3766 unsigned int len; 3767 int rc; 3768 3769 if (dev_nit_active_rcu(dev)) 3770 dev_queue_xmit_nit(skb, dev); 3771 3772 len = skb->len; 3773 trace_net_dev_start_xmit(skb, dev); 3774 rc = netdev_start_xmit(skb, dev, txq, more); 3775 trace_net_dev_xmit(skb, rc, dev, len); 3776 3777 return rc; 3778 } 3779 3780 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev, 3781 struct netdev_queue *txq, int *ret) 3782 { 3783 struct sk_buff *skb = first; 3784 int rc = NETDEV_TX_OK; 3785 3786 while (skb) { 3787 struct sk_buff *next = skb->next; 3788 3789 skb_mark_not_on_list(skb); 3790 rc = xmit_one(skb, dev, txq, next != NULL); 3791 if (unlikely(!dev_xmit_complete(rc))) { 3792 skb->next = next; 3793 goto out; 3794 } 3795 3796 skb = next; 3797 if (netif_tx_queue_stopped(txq) && skb) { 3798 rc = NETDEV_TX_BUSY; 3799 break; 3800 } 3801 } 3802 3803 out: 3804 *ret = rc; 3805 return skb; 3806 } 3807 3808 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb, 3809 netdev_features_t features) 3810 { 3811 if (skb_vlan_tag_present(skb) && 3812 !vlan_hw_offload_capable(features, skb->vlan_proto)) 3813 skb = __vlan_hwaccel_push_inside(skb); 3814 return skb; 3815 } 3816 3817 int skb_csum_hwoffload_help(struct sk_buff *skb, 3818 const netdev_features_t features) 3819 { 3820 if (unlikely(skb_csum_is_sctp(skb))) 3821 return !!(features & NETIF_F_SCTP_CRC) ? 0 : 3822 skb_crc32c_csum_help(skb); 3823 3824 if (features & NETIF_F_HW_CSUM) 3825 return 0; 3826 3827 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) { 3828 if (vlan_get_protocol(skb) == htons(ETH_P_IPV6) && 3829 skb_network_header_len(skb) != sizeof(struct ipv6hdr) && 3830 !ipv6_has_hopopt_jumbo(skb)) 3831 goto sw_checksum; 3832 3833 switch (skb->csum_offset) { 3834 case offsetof(struct tcphdr, check): 3835 case offsetof(struct udphdr, check): 3836 return 0; 3837 } 3838 } 3839 3840 sw_checksum: 3841 return skb_checksum_help(skb); 3842 } 3843 EXPORT_SYMBOL(skb_csum_hwoffload_help); 3844 3845 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again) 3846 { 3847 netdev_features_t features; 3848 3849 if (!skb_frags_readable(skb)) 3850 goto out_kfree_skb; 3851 3852 features = netif_skb_features(skb); 3853 skb = validate_xmit_vlan(skb, features); 3854 if (unlikely(!skb)) 3855 goto out_null; 3856 3857 skb = sk_validate_xmit_skb(skb, dev); 3858 if (unlikely(!skb)) 3859 goto out_null; 3860 3861 if (netif_needs_gso(skb, features)) { 3862 struct sk_buff *segs; 3863 3864 segs = skb_gso_segment(skb, features); 3865 if (IS_ERR(segs)) { 3866 goto out_kfree_skb; 3867 } else if (segs) { 3868 consume_skb(skb); 3869 skb = segs; 3870 } 3871 } else { 3872 if (skb_needs_linearize(skb, features) && 3873 __skb_linearize(skb)) 3874 goto out_kfree_skb; 3875 3876 /* If packet is not checksummed and device does not 3877 * support checksumming for this protocol, complete 3878 * checksumming here. 3879 */ 3880 if (skb->ip_summed == CHECKSUM_PARTIAL) { 3881 if (skb->encapsulation) 3882 skb_set_inner_transport_header(skb, 3883 skb_checksum_start_offset(skb)); 3884 else 3885 skb_set_transport_header(skb, 3886 skb_checksum_start_offset(skb)); 3887 if (skb_csum_hwoffload_help(skb, features)) 3888 goto out_kfree_skb; 3889 } 3890 } 3891 3892 skb = validate_xmit_xfrm(skb, features, again); 3893 3894 return skb; 3895 3896 out_kfree_skb: 3897 kfree_skb(skb); 3898 out_null: 3899 dev_core_stats_tx_dropped_inc(dev); 3900 return NULL; 3901 } 3902 3903 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again) 3904 { 3905 struct sk_buff *next, *head = NULL, *tail; 3906 3907 for (; skb != NULL; skb = next) { 3908 next = skb->next; 3909 skb_mark_not_on_list(skb); 3910 3911 /* in case skb won't be segmented, point to itself */ 3912 skb->prev = skb; 3913 3914 skb = validate_xmit_skb(skb, dev, again); 3915 if (!skb) 3916 continue; 3917 3918 if (!head) 3919 head = skb; 3920 else 3921 tail->next = skb; 3922 /* If skb was segmented, skb->prev points to 3923 * the last segment. If not, it still contains skb. 3924 */ 3925 tail = skb->prev; 3926 } 3927 return head; 3928 } 3929 EXPORT_SYMBOL_GPL(validate_xmit_skb_list); 3930 3931 static void qdisc_pkt_len_init(struct sk_buff *skb) 3932 { 3933 const struct skb_shared_info *shinfo = skb_shinfo(skb); 3934 3935 qdisc_skb_cb(skb)->pkt_len = skb->len; 3936 3937 /* To get more precise estimation of bytes sent on wire, 3938 * we add to pkt_len the headers size of all segments 3939 */ 3940 if (shinfo->gso_size && skb_transport_header_was_set(skb)) { 3941 u16 gso_segs = shinfo->gso_segs; 3942 unsigned int hdr_len; 3943 3944 /* mac layer + network layer */ 3945 hdr_len = skb_transport_offset(skb); 3946 3947 /* + transport layer */ 3948 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) { 3949 const struct tcphdr *th; 3950 struct tcphdr _tcphdr; 3951 3952 th = skb_header_pointer(skb, hdr_len, 3953 sizeof(_tcphdr), &_tcphdr); 3954 if (likely(th)) 3955 hdr_len += __tcp_hdrlen(th); 3956 } else if (shinfo->gso_type & SKB_GSO_UDP_L4) { 3957 struct udphdr _udphdr; 3958 3959 if (skb_header_pointer(skb, hdr_len, 3960 sizeof(_udphdr), &_udphdr)) 3961 hdr_len += sizeof(struct udphdr); 3962 } 3963 3964 if (unlikely(shinfo->gso_type & SKB_GSO_DODGY)) { 3965 int payload = skb->len - hdr_len; 3966 3967 /* Malicious packet. */ 3968 if (payload <= 0) 3969 return; 3970 gso_segs = DIV_ROUND_UP(payload, shinfo->gso_size); 3971 } 3972 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len; 3973 } 3974 } 3975 3976 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q, 3977 struct sk_buff **to_free, 3978 struct netdev_queue *txq) 3979 { 3980 int rc; 3981 3982 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK; 3983 if (rc == NET_XMIT_SUCCESS) 3984 trace_qdisc_enqueue(q, txq, skb); 3985 return rc; 3986 } 3987 3988 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q, 3989 struct net_device *dev, 3990 struct netdev_queue *txq) 3991 { 3992 spinlock_t *root_lock = qdisc_lock(q); 3993 struct sk_buff *to_free = NULL; 3994 bool contended; 3995 int rc; 3996 3997 qdisc_calculate_pkt_len(skb, q); 3998 3999 tcf_set_drop_reason(skb, SKB_DROP_REASON_QDISC_DROP); 4000 4001 if (q->flags & TCQ_F_NOLOCK) { 4002 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) && 4003 qdisc_run_begin(q)) { 4004 /* Retest nolock_qdisc_is_empty() within the protection 4005 * of q->seqlock to protect from racing with requeuing. 4006 */ 4007 if (unlikely(!nolock_qdisc_is_empty(q))) { 4008 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4009 __qdisc_run(q); 4010 qdisc_run_end(q); 4011 4012 goto no_lock_out; 4013 } 4014 4015 qdisc_bstats_cpu_update(q, skb); 4016 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) && 4017 !nolock_qdisc_is_empty(q)) 4018 __qdisc_run(q); 4019 4020 qdisc_run_end(q); 4021 return NET_XMIT_SUCCESS; 4022 } 4023 4024 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4025 qdisc_run(q); 4026 4027 no_lock_out: 4028 if (unlikely(to_free)) 4029 kfree_skb_list_reason(to_free, 4030 tcf_get_drop_reason(to_free)); 4031 return rc; 4032 } 4033 4034 if (unlikely(READ_ONCE(q->owner) == smp_processor_id())) { 4035 kfree_skb_reason(skb, SKB_DROP_REASON_TC_RECLASSIFY_LOOP); 4036 return NET_XMIT_DROP; 4037 } 4038 /* 4039 * Heuristic to force contended enqueues to serialize on a 4040 * separate lock before trying to get qdisc main lock. 4041 * This permits qdisc->running owner to get the lock more 4042 * often and dequeue packets faster. 4043 * On PREEMPT_RT it is possible to preempt the qdisc owner during xmit 4044 * and then other tasks will only enqueue packets. The packets will be 4045 * sent after the qdisc owner is scheduled again. To prevent this 4046 * scenario the task always serialize on the lock. 4047 */ 4048 contended = qdisc_is_running(q) || IS_ENABLED(CONFIG_PREEMPT_RT); 4049 if (unlikely(contended)) 4050 spin_lock(&q->busylock); 4051 4052 spin_lock(root_lock); 4053 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) { 4054 __qdisc_drop(skb, &to_free); 4055 rc = NET_XMIT_DROP; 4056 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && 4057 qdisc_run_begin(q)) { 4058 /* 4059 * This is a work-conserving queue; there are no old skbs 4060 * waiting to be sent out; and the qdisc is not running - 4061 * xmit the skb directly. 4062 */ 4063 4064 qdisc_bstats_update(q, skb); 4065 4066 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) { 4067 if (unlikely(contended)) { 4068 spin_unlock(&q->busylock); 4069 contended = false; 4070 } 4071 __qdisc_run(q); 4072 } 4073 4074 qdisc_run_end(q); 4075 rc = NET_XMIT_SUCCESS; 4076 } else { 4077 WRITE_ONCE(q->owner, smp_processor_id()); 4078 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4079 WRITE_ONCE(q->owner, -1); 4080 if (qdisc_run_begin(q)) { 4081 if (unlikely(contended)) { 4082 spin_unlock(&q->busylock); 4083 contended = false; 4084 } 4085 __qdisc_run(q); 4086 qdisc_run_end(q); 4087 } 4088 } 4089 spin_unlock(root_lock); 4090 if (unlikely(to_free)) 4091 kfree_skb_list_reason(to_free, 4092 tcf_get_drop_reason(to_free)); 4093 if (unlikely(contended)) 4094 spin_unlock(&q->busylock); 4095 return rc; 4096 } 4097 4098 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO) 4099 static void skb_update_prio(struct sk_buff *skb) 4100 { 4101 const struct netprio_map *map; 4102 const struct sock *sk; 4103 unsigned int prioidx; 4104 4105 if (skb->priority) 4106 return; 4107 map = rcu_dereference_bh(skb->dev->priomap); 4108 if (!map) 4109 return; 4110 sk = skb_to_full_sk(skb); 4111 if (!sk) 4112 return; 4113 4114 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data); 4115 4116 if (prioidx < map->priomap_len) 4117 skb->priority = map->priomap[prioidx]; 4118 } 4119 #else 4120 #define skb_update_prio(skb) 4121 #endif 4122 4123 /** 4124 * dev_loopback_xmit - loop back @skb 4125 * @net: network namespace this loopback is happening in 4126 * @sk: sk needed to be a netfilter okfn 4127 * @skb: buffer to transmit 4128 */ 4129 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb) 4130 { 4131 skb_reset_mac_header(skb); 4132 __skb_pull(skb, skb_network_offset(skb)); 4133 skb->pkt_type = PACKET_LOOPBACK; 4134 if (skb->ip_summed == CHECKSUM_NONE) 4135 skb->ip_summed = CHECKSUM_UNNECESSARY; 4136 DEBUG_NET_WARN_ON_ONCE(!skb_dst(skb)); 4137 skb_dst_force(skb); 4138 netif_rx(skb); 4139 return 0; 4140 } 4141 EXPORT_SYMBOL(dev_loopback_xmit); 4142 4143 #ifdef CONFIG_NET_EGRESS 4144 static struct netdev_queue * 4145 netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb) 4146 { 4147 int qm = skb_get_queue_mapping(skb); 4148 4149 return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm)); 4150 } 4151 4152 #ifndef CONFIG_PREEMPT_RT 4153 static bool netdev_xmit_txqueue_skipped(void) 4154 { 4155 return __this_cpu_read(softnet_data.xmit.skip_txqueue); 4156 } 4157 4158 void netdev_xmit_skip_txqueue(bool skip) 4159 { 4160 __this_cpu_write(softnet_data.xmit.skip_txqueue, skip); 4161 } 4162 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4163 4164 #else 4165 static bool netdev_xmit_txqueue_skipped(void) 4166 { 4167 return current->net_xmit.skip_txqueue; 4168 } 4169 4170 void netdev_xmit_skip_txqueue(bool skip) 4171 { 4172 current->net_xmit.skip_txqueue = skip; 4173 } 4174 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4175 #endif 4176 #endif /* CONFIG_NET_EGRESS */ 4177 4178 #ifdef CONFIG_NET_XGRESS 4179 static int tc_run(struct tcx_entry *entry, struct sk_buff *skb, 4180 enum skb_drop_reason *drop_reason) 4181 { 4182 int ret = TC_ACT_UNSPEC; 4183 #ifdef CONFIG_NET_CLS_ACT 4184 struct mini_Qdisc *miniq = rcu_dereference_bh(entry->miniq); 4185 struct tcf_result res; 4186 4187 if (!miniq) 4188 return ret; 4189 4190 /* Global bypass */ 4191 if (!static_branch_likely(&tcf_sw_enabled_key)) 4192 return ret; 4193 4194 /* Block-wise bypass */ 4195 if (tcf_block_bypass_sw(miniq->block)) 4196 return ret; 4197 4198 tc_skb_cb(skb)->mru = 0; 4199 tc_skb_cb(skb)->post_ct = false; 4200 tcf_set_drop_reason(skb, *drop_reason); 4201 4202 mini_qdisc_bstats_cpu_update(miniq, skb); 4203 ret = tcf_classify(skb, miniq->block, miniq->filter_list, &res, false); 4204 /* Only tcf related quirks below. */ 4205 switch (ret) { 4206 case TC_ACT_SHOT: 4207 *drop_reason = tcf_get_drop_reason(skb); 4208 mini_qdisc_qstats_cpu_drop(miniq); 4209 break; 4210 case TC_ACT_OK: 4211 case TC_ACT_RECLASSIFY: 4212 skb->tc_index = TC_H_MIN(res.classid); 4213 break; 4214 } 4215 #endif /* CONFIG_NET_CLS_ACT */ 4216 return ret; 4217 } 4218 4219 static DEFINE_STATIC_KEY_FALSE(tcx_needed_key); 4220 4221 void tcx_inc(void) 4222 { 4223 static_branch_inc(&tcx_needed_key); 4224 } 4225 4226 void tcx_dec(void) 4227 { 4228 static_branch_dec(&tcx_needed_key); 4229 } 4230 4231 static __always_inline enum tcx_action_base 4232 tcx_run(const struct bpf_mprog_entry *entry, struct sk_buff *skb, 4233 const bool needs_mac) 4234 { 4235 const struct bpf_mprog_fp *fp; 4236 const struct bpf_prog *prog; 4237 int ret = TCX_NEXT; 4238 4239 if (needs_mac) 4240 __skb_push(skb, skb->mac_len); 4241 bpf_mprog_foreach_prog(entry, fp, prog) { 4242 bpf_compute_data_pointers(skb); 4243 ret = bpf_prog_run(prog, skb); 4244 if (ret != TCX_NEXT) 4245 break; 4246 } 4247 if (needs_mac) 4248 __skb_pull(skb, skb->mac_len); 4249 return tcx_action_code(skb, ret); 4250 } 4251 4252 static __always_inline struct sk_buff * 4253 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4254 struct net_device *orig_dev, bool *another) 4255 { 4256 struct bpf_mprog_entry *entry = rcu_dereference_bh(skb->dev->tcx_ingress); 4257 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_INGRESS; 4258 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4259 int sch_ret; 4260 4261 if (!entry) 4262 return skb; 4263 4264 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4265 if (*pt_prev) { 4266 *ret = deliver_skb(skb, *pt_prev, orig_dev); 4267 *pt_prev = NULL; 4268 } 4269 4270 qdisc_skb_cb(skb)->pkt_len = skb->len; 4271 tcx_set_ingress(skb, true); 4272 4273 if (static_branch_unlikely(&tcx_needed_key)) { 4274 sch_ret = tcx_run(entry, skb, true); 4275 if (sch_ret != TC_ACT_UNSPEC) 4276 goto ingress_verdict; 4277 } 4278 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4279 ingress_verdict: 4280 switch (sch_ret) { 4281 case TC_ACT_REDIRECT: 4282 /* skb_mac_header check was done by BPF, so we can safely 4283 * push the L2 header back before redirecting to another 4284 * netdev. 4285 */ 4286 __skb_push(skb, skb->mac_len); 4287 if (skb_do_redirect(skb) == -EAGAIN) { 4288 __skb_pull(skb, skb->mac_len); 4289 *another = true; 4290 break; 4291 } 4292 *ret = NET_RX_SUCCESS; 4293 bpf_net_ctx_clear(bpf_net_ctx); 4294 return NULL; 4295 case TC_ACT_SHOT: 4296 kfree_skb_reason(skb, drop_reason); 4297 *ret = NET_RX_DROP; 4298 bpf_net_ctx_clear(bpf_net_ctx); 4299 return NULL; 4300 /* used by tc_run */ 4301 case TC_ACT_STOLEN: 4302 case TC_ACT_QUEUED: 4303 case TC_ACT_TRAP: 4304 consume_skb(skb); 4305 fallthrough; 4306 case TC_ACT_CONSUMED: 4307 *ret = NET_RX_SUCCESS; 4308 bpf_net_ctx_clear(bpf_net_ctx); 4309 return NULL; 4310 } 4311 bpf_net_ctx_clear(bpf_net_ctx); 4312 4313 return skb; 4314 } 4315 4316 static __always_inline struct sk_buff * 4317 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4318 { 4319 struct bpf_mprog_entry *entry = rcu_dereference_bh(dev->tcx_egress); 4320 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_EGRESS; 4321 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4322 int sch_ret; 4323 4324 if (!entry) 4325 return skb; 4326 4327 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4328 4329 /* qdisc_skb_cb(skb)->pkt_len & tcx_set_ingress() was 4330 * already set by the caller. 4331 */ 4332 if (static_branch_unlikely(&tcx_needed_key)) { 4333 sch_ret = tcx_run(entry, skb, false); 4334 if (sch_ret != TC_ACT_UNSPEC) 4335 goto egress_verdict; 4336 } 4337 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4338 egress_verdict: 4339 switch (sch_ret) { 4340 case TC_ACT_REDIRECT: 4341 /* No need to push/pop skb's mac_header here on egress! */ 4342 skb_do_redirect(skb); 4343 *ret = NET_XMIT_SUCCESS; 4344 bpf_net_ctx_clear(bpf_net_ctx); 4345 return NULL; 4346 case TC_ACT_SHOT: 4347 kfree_skb_reason(skb, drop_reason); 4348 *ret = NET_XMIT_DROP; 4349 bpf_net_ctx_clear(bpf_net_ctx); 4350 return NULL; 4351 /* used by tc_run */ 4352 case TC_ACT_STOLEN: 4353 case TC_ACT_QUEUED: 4354 case TC_ACT_TRAP: 4355 consume_skb(skb); 4356 fallthrough; 4357 case TC_ACT_CONSUMED: 4358 *ret = NET_XMIT_SUCCESS; 4359 bpf_net_ctx_clear(bpf_net_ctx); 4360 return NULL; 4361 } 4362 bpf_net_ctx_clear(bpf_net_ctx); 4363 4364 return skb; 4365 } 4366 #else 4367 static __always_inline struct sk_buff * 4368 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4369 struct net_device *orig_dev, bool *another) 4370 { 4371 return skb; 4372 } 4373 4374 static __always_inline struct sk_buff * 4375 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4376 { 4377 return skb; 4378 } 4379 #endif /* CONFIG_NET_XGRESS */ 4380 4381 #ifdef CONFIG_XPS 4382 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb, 4383 struct xps_dev_maps *dev_maps, unsigned int tci) 4384 { 4385 int tc = netdev_get_prio_tc_map(dev, skb->priority); 4386 struct xps_map *map; 4387 int queue_index = -1; 4388 4389 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids) 4390 return queue_index; 4391 4392 tci *= dev_maps->num_tc; 4393 tci += tc; 4394 4395 map = rcu_dereference(dev_maps->attr_map[tci]); 4396 if (map) { 4397 if (map->len == 1) 4398 queue_index = map->queues[0]; 4399 else 4400 queue_index = map->queues[reciprocal_scale( 4401 skb_get_hash(skb), map->len)]; 4402 if (unlikely(queue_index >= dev->real_num_tx_queues)) 4403 queue_index = -1; 4404 } 4405 return queue_index; 4406 } 4407 #endif 4408 4409 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev, 4410 struct sk_buff *skb) 4411 { 4412 #ifdef CONFIG_XPS 4413 struct xps_dev_maps *dev_maps; 4414 struct sock *sk = skb->sk; 4415 int queue_index = -1; 4416 4417 if (!static_key_false(&xps_needed)) 4418 return -1; 4419 4420 rcu_read_lock(); 4421 if (!static_key_false(&xps_rxqs_needed)) 4422 goto get_cpus_map; 4423 4424 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]); 4425 if (dev_maps) { 4426 int tci = sk_rx_queue_get(sk); 4427 4428 if (tci >= 0) 4429 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4430 tci); 4431 } 4432 4433 get_cpus_map: 4434 if (queue_index < 0) { 4435 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]); 4436 if (dev_maps) { 4437 unsigned int tci = skb->sender_cpu - 1; 4438 4439 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4440 tci); 4441 } 4442 } 4443 rcu_read_unlock(); 4444 4445 return queue_index; 4446 #else 4447 return -1; 4448 #endif 4449 } 4450 4451 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb, 4452 struct net_device *sb_dev) 4453 { 4454 return 0; 4455 } 4456 EXPORT_SYMBOL(dev_pick_tx_zero); 4457 4458 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb, 4459 struct net_device *sb_dev) 4460 { 4461 struct sock *sk = skb->sk; 4462 int queue_index = sk_tx_queue_get(sk); 4463 4464 sb_dev = sb_dev ? : dev; 4465 4466 if (queue_index < 0 || skb->ooo_okay || 4467 queue_index >= dev->real_num_tx_queues) { 4468 int new_index = get_xps_queue(dev, sb_dev, skb); 4469 4470 if (new_index < 0) 4471 new_index = skb_tx_hash(dev, sb_dev, skb); 4472 4473 if (queue_index != new_index && sk && 4474 sk_fullsock(sk) && 4475 rcu_access_pointer(sk->sk_dst_cache)) 4476 sk_tx_queue_set(sk, new_index); 4477 4478 queue_index = new_index; 4479 } 4480 4481 return queue_index; 4482 } 4483 EXPORT_SYMBOL(netdev_pick_tx); 4484 4485 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev, 4486 struct sk_buff *skb, 4487 struct net_device *sb_dev) 4488 { 4489 int queue_index = 0; 4490 4491 #ifdef CONFIG_XPS 4492 u32 sender_cpu = skb->sender_cpu - 1; 4493 4494 if (sender_cpu >= (u32)NR_CPUS) 4495 skb->sender_cpu = raw_smp_processor_id() + 1; 4496 #endif 4497 4498 if (dev->real_num_tx_queues != 1) { 4499 const struct net_device_ops *ops = dev->netdev_ops; 4500 4501 if (ops->ndo_select_queue) 4502 queue_index = ops->ndo_select_queue(dev, skb, sb_dev); 4503 else 4504 queue_index = netdev_pick_tx(dev, skb, sb_dev); 4505 4506 queue_index = netdev_cap_txqueue(dev, queue_index); 4507 } 4508 4509 skb_set_queue_mapping(skb, queue_index); 4510 return netdev_get_tx_queue(dev, queue_index); 4511 } 4512 4513 /** 4514 * __dev_queue_xmit() - transmit a buffer 4515 * @skb: buffer to transmit 4516 * @sb_dev: suboordinate device used for L2 forwarding offload 4517 * 4518 * Queue a buffer for transmission to a network device. The caller must 4519 * have set the device and priority and built the buffer before calling 4520 * this function. The function can be called from an interrupt. 4521 * 4522 * When calling this method, interrupts MUST be enabled. This is because 4523 * the BH enable code must have IRQs enabled so that it will not deadlock. 4524 * 4525 * Regardless of the return value, the skb is consumed, so it is currently 4526 * difficult to retry a send to this method. (You can bump the ref count 4527 * before sending to hold a reference for retry if you are careful.) 4528 * 4529 * Return: 4530 * * 0 - buffer successfully transmitted 4531 * * positive qdisc return code - NET_XMIT_DROP etc. 4532 * * negative errno - other errors 4533 */ 4534 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev) 4535 { 4536 struct net_device *dev = skb->dev; 4537 struct netdev_queue *txq = NULL; 4538 struct Qdisc *q; 4539 int rc = -ENOMEM; 4540 bool again = false; 4541 4542 skb_reset_mac_header(skb); 4543 skb_assert_len(skb); 4544 4545 if (unlikely(skb_shinfo(skb)->tx_flags & 4546 (SKBTX_SCHED_TSTAMP | SKBTX_BPF))) 4547 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED); 4548 4549 /* Disable soft irqs for various locks below. Also 4550 * stops preemption for RCU. 4551 */ 4552 rcu_read_lock_bh(); 4553 4554 skb_update_prio(skb); 4555 4556 qdisc_pkt_len_init(skb); 4557 tcx_set_ingress(skb, false); 4558 #ifdef CONFIG_NET_EGRESS 4559 if (static_branch_unlikely(&egress_needed_key)) { 4560 if (nf_hook_egress_active()) { 4561 skb = nf_hook_egress(skb, &rc, dev); 4562 if (!skb) 4563 goto out; 4564 } 4565 4566 netdev_xmit_skip_txqueue(false); 4567 4568 nf_skip_egress(skb, true); 4569 skb = sch_handle_egress(skb, &rc, dev); 4570 if (!skb) 4571 goto out; 4572 nf_skip_egress(skb, false); 4573 4574 if (netdev_xmit_txqueue_skipped()) 4575 txq = netdev_tx_queue_mapping(dev, skb); 4576 } 4577 #endif 4578 /* If device/qdisc don't need skb->dst, release it right now while 4579 * its hot in this cpu cache. 4580 */ 4581 if (dev->priv_flags & IFF_XMIT_DST_RELEASE) 4582 skb_dst_drop(skb); 4583 else 4584 skb_dst_force(skb); 4585 4586 if (!txq) 4587 txq = netdev_core_pick_tx(dev, skb, sb_dev); 4588 4589 q = rcu_dereference_bh(txq->qdisc); 4590 4591 trace_net_dev_queue(skb); 4592 if (q->enqueue) { 4593 rc = __dev_xmit_skb(skb, q, dev, txq); 4594 goto out; 4595 } 4596 4597 /* The device has no queue. Common case for software devices: 4598 * loopback, all the sorts of tunnels... 4599 4600 * Really, it is unlikely that netif_tx_lock protection is necessary 4601 * here. (f.e. loopback and IP tunnels are clean ignoring statistics 4602 * counters.) 4603 * However, it is possible, that they rely on protection 4604 * made by us here. 4605 4606 * Check this and shot the lock. It is not prone from deadlocks. 4607 *Either shot noqueue qdisc, it is even simpler 8) 4608 */ 4609 if (dev->flags & IFF_UP) { 4610 int cpu = smp_processor_id(); /* ok because BHs are off */ 4611 4612 /* Other cpus might concurrently change txq->xmit_lock_owner 4613 * to -1 or to their cpu id, but not to our id. 4614 */ 4615 if (READ_ONCE(txq->xmit_lock_owner) != cpu) { 4616 if (dev_xmit_recursion()) 4617 goto recursion_alert; 4618 4619 skb = validate_xmit_skb(skb, dev, &again); 4620 if (!skb) 4621 goto out; 4622 4623 HARD_TX_LOCK(dev, txq, cpu); 4624 4625 if (!netif_xmit_stopped(txq)) { 4626 dev_xmit_recursion_inc(); 4627 skb = dev_hard_start_xmit(skb, dev, txq, &rc); 4628 dev_xmit_recursion_dec(); 4629 if (dev_xmit_complete(rc)) { 4630 HARD_TX_UNLOCK(dev, txq); 4631 goto out; 4632 } 4633 } 4634 HARD_TX_UNLOCK(dev, txq); 4635 net_crit_ratelimited("Virtual device %s asks to queue packet!\n", 4636 dev->name); 4637 } else { 4638 /* Recursion is detected! It is possible, 4639 * unfortunately 4640 */ 4641 recursion_alert: 4642 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n", 4643 dev->name); 4644 } 4645 } 4646 4647 rc = -ENETDOWN; 4648 rcu_read_unlock_bh(); 4649 4650 dev_core_stats_tx_dropped_inc(dev); 4651 kfree_skb_list(skb); 4652 return rc; 4653 out: 4654 rcu_read_unlock_bh(); 4655 return rc; 4656 } 4657 EXPORT_SYMBOL(__dev_queue_xmit); 4658 4659 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id) 4660 { 4661 struct net_device *dev = skb->dev; 4662 struct sk_buff *orig_skb = skb; 4663 struct netdev_queue *txq; 4664 int ret = NETDEV_TX_BUSY; 4665 bool again = false; 4666 4667 if (unlikely(!netif_running(dev) || 4668 !netif_carrier_ok(dev))) 4669 goto drop; 4670 4671 skb = validate_xmit_skb_list(skb, dev, &again); 4672 if (skb != orig_skb) 4673 goto drop; 4674 4675 skb_set_queue_mapping(skb, queue_id); 4676 txq = skb_get_tx_queue(dev, skb); 4677 4678 local_bh_disable(); 4679 4680 dev_xmit_recursion_inc(); 4681 HARD_TX_LOCK(dev, txq, smp_processor_id()); 4682 if (!netif_xmit_frozen_or_drv_stopped(txq)) 4683 ret = netdev_start_xmit(skb, dev, txq, false); 4684 HARD_TX_UNLOCK(dev, txq); 4685 dev_xmit_recursion_dec(); 4686 4687 local_bh_enable(); 4688 return ret; 4689 drop: 4690 dev_core_stats_tx_dropped_inc(dev); 4691 kfree_skb_list(skb); 4692 return NET_XMIT_DROP; 4693 } 4694 EXPORT_SYMBOL(__dev_direct_xmit); 4695 4696 /************************************************************************* 4697 * Receiver routines 4698 *************************************************************************/ 4699 static DEFINE_PER_CPU(struct task_struct *, backlog_napi); 4700 4701 int weight_p __read_mostly = 64; /* old backlog weight */ 4702 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */ 4703 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */ 4704 4705 /* Called with irq disabled */ 4706 static inline void ____napi_schedule(struct softnet_data *sd, 4707 struct napi_struct *napi) 4708 { 4709 struct task_struct *thread; 4710 4711 lockdep_assert_irqs_disabled(); 4712 4713 if (test_bit(NAPI_STATE_THREADED, &napi->state)) { 4714 /* Paired with smp_mb__before_atomic() in 4715 * napi_enable()/dev_set_threaded(). 4716 * Use READ_ONCE() to guarantee a complete 4717 * read on napi->thread. Only call 4718 * wake_up_process() when it's not NULL. 4719 */ 4720 thread = READ_ONCE(napi->thread); 4721 if (thread) { 4722 if (use_backlog_threads() && thread == raw_cpu_read(backlog_napi)) 4723 goto use_local_napi; 4724 4725 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 4726 wake_up_process(thread); 4727 return; 4728 } 4729 } 4730 4731 use_local_napi: 4732 list_add_tail(&napi->poll_list, &sd->poll_list); 4733 WRITE_ONCE(napi->list_owner, smp_processor_id()); 4734 /* If not called from net_rx_action() 4735 * we have to raise NET_RX_SOFTIRQ. 4736 */ 4737 if (!sd->in_net_rx_action) 4738 raise_softirq_irqoff(NET_RX_SOFTIRQ); 4739 } 4740 4741 #ifdef CONFIG_RPS 4742 4743 struct static_key_false rps_needed __read_mostly; 4744 EXPORT_SYMBOL(rps_needed); 4745 struct static_key_false rfs_needed __read_mostly; 4746 EXPORT_SYMBOL(rfs_needed); 4747 4748 static u32 rfs_slot(u32 hash, const struct rps_dev_flow_table *flow_table) 4749 { 4750 return hash_32(hash, flow_table->log); 4751 } 4752 4753 static struct rps_dev_flow * 4754 set_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4755 struct rps_dev_flow *rflow, u16 next_cpu) 4756 { 4757 if (next_cpu < nr_cpu_ids) { 4758 u32 head; 4759 #ifdef CONFIG_RFS_ACCEL 4760 struct netdev_rx_queue *rxqueue; 4761 struct rps_dev_flow_table *flow_table; 4762 struct rps_dev_flow *old_rflow; 4763 u16 rxq_index; 4764 u32 flow_id; 4765 int rc; 4766 4767 /* Should we steer this flow to a different hardware queue? */ 4768 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap || 4769 !(dev->features & NETIF_F_NTUPLE)) 4770 goto out; 4771 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu); 4772 if (rxq_index == skb_get_rx_queue(skb)) 4773 goto out; 4774 4775 rxqueue = dev->_rx + rxq_index; 4776 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4777 if (!flow_table) 4778 goto out; 4779 flow_id = rfs_slot(skb_get_hash(skb), flow_table); 4780 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb, 4781 rxq_index, flow_id); 4782 if (rc < 0) 4783 goto out; 4784 old_rflow = rflow; 4785 rflow = &flow_table->flows[flow_id]; 4786 WRITE_ONCE(rflow->filter, rc); 4787 if (old_rflow->filter == rc) 4788 WRITE_ONCE(old_rflow->filter, RPS_NO_FILTER); 4789 out: 4790 #endif 4791 head = READ_ONCE(per_cpu(softnet_data, next_cpu).input_queue_head); 4792 rps_input_queue_tail_save(&rflow->last_qtail, head); 4793 } 4794 4795 WRITE_ONCE(rflow->cpu, next_cpu); 4796 return rflow; 4797 } 4798 4799 /* 4800 * get_rps_cpu is called from netif_receive_skb and returns the target 4801 * CPU from the RPS map of the receiving queue for a given skb. 4802 * rcu_read_lock must be held on entry. 4803 */ 4804 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4805 struct rps_dev_flow **rflowp) 4806 { 4807 const struct rps_sock_flow_table *sock_flow_table; 4808 struct netdev_rx_queue *rxqueue = dev->_rx; 4809 struct rps_dev_flow_table *flow_table; 4810 struct rps_map *map; 4811 int cpu = -1; 4812 u32 tcpu; 4813 u32 hash; 4814 4815 if (skb_rx_queue_recorded(skb)) { 4816 u16 index = skb_get_rx_queue(skb); 4817 4818 if (unlikely(index >= dev->real_num_rx_queues)) { 4819 WARN_ONCE(dev->real_num_rx_queues > 1, 4820 "%s received packet on queue %u, but number " 4821 "of RX queues is %u\n", 4822 dev->name, index, dev->real_num_rx_queues); 4823 goto done; 4824 } 4825 rxqueue += index; 4826 } 4827 4828 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */ 4829 4830 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4831 map = rcu_dereference(rxqueue->rps_map); 4832 if (!flow_table && !map) 4833 goto done; 4834 4835 skb_reset_network_header(skb); 4836 hash = skb_get_hash(skb); 4837 if (!hash) 4838 goto done; 4839 4840 sock_flow_table = rcu_dereference(net_hotdata.rps_sock_flow_table); 4841 if (flow_table && sock_flow_table) { 4842 struct rps_dev_flow *rflow; 4843 u32 next_cpu; 4844 u32 ident; 4845 4846 /* First check into global flow table if there is a match. 4847 * This READ_ONCE() pairs with WRITE_ONCE() from rps_record_sock_flow(). 4848 */ 4849 ident = READ_ONCE(sock_flow_table->ents[hash & sock_flow_table->mask]); 4850 if ((ident ^ hash) & ~net_hotdata.rps_cpu_mask) 4851 goto try_rps; 4852 4853 next_cpu = ident & net_hotdata.rps_cpu_mask; 4854 4855 /* OK, now we know there is a match, 4856 * we can look at the local (per receive queue) flow table 4857 */ 4858 rflow = &flow_table->flows[rfs_slot(hash, flow_table)]; 4859 tcpu = rflow->cpu; 4860 4861 /* 4862 * If the desired CPU (where last recvmsg was done) is 4863 * different from current CPU (one in the rx-queue flow 4864 * table entry), switch if one of the following holds: 4865 * - Current CPU is unset (>= nr_cpu_ids). 4866 * - Current CPU is offline. 4867 * - The current CPU's queue tail has advanced beyond the 4868 * last packet that was enqueued using this table entry. 4869 * This guarantees that all previous packets for the flow 4870 * have been dequeued, thus preserving in order delivery. 4871 */ 4872 if (unlikely(tcpu != next_cpu) && 4873 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) || 4874 ((int)(READ_ONCE(per_cpu(softnet_data, tcpu).input_queue_head) - 4875 rflow->last_qtail)) >= 0)) { 4876 tcpu = next_cpu; 4877 rflow = set_rps_cpu(dev, skb, rflow, next_cpu); 4878 } 4879 4880 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) { 4881 *rflowp = rflow; 4882 cpu = tcpu; 4883 goto done; 4884 } 4885 } 4886 4887 try_rps: 4888 4889 if (map) { 4890 tcpu = map->cpus[reciprocal_scale(hash, map->len)]; 4891 if (cpu_online(tcpu)) { 4892 cpu = tcpu; 4893 goto done; 4894 } 4895 } 4896 4897 done: 4898 return cpu; 4899 } 4900 4901 #ifdef CONFIG_RFS_ACCEL 4902 4903 /** 4904 * rps_may_expire_flow - check whether an RFS hardware filter may be removed 4905 * @dev: Device on which the filter was set 4906 * @rxq_index: RX queue index 4907 * @flow_id: Flow ID passed to ndo_rx_flow_steer() 4908 * @filter_id: Filter ID returned by ndo_rx_flow_steer() 4909 * 4910 * Drivers that implement ndo_rx_flow_steer() should periodically call 4911 * this function for each installed filter and remove the filters for 4912 * which it returns %true. 4913 */ 4914 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index, 4915 u32 flow_id, u16 filter_id) 4916 { 4917 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index; 4918 struct rps_dev_flow_table *flow_table; 4919 struct rps_dev_flow *rflow; 4920 bool expire = true; 4921 unsigned int cpu; 4922 4923 rcu_read_lock(); 4924 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4925 if (flow_table && flow_id < (1UL << flow_table->log)) { 4926 rflow = &flow_table->flows[flow_id]; 4927 cpu = READ_ONCE(rflow->cpu); 4928 if (READ_ONCE(rflow->filter) == filter_id && cpu < nr_cpu_ids && 4929 ((int)(READ_ONCE(per_cpu(softnet_data, cpu).input_queue_head) - 4930 READ_ONCE(rflow->last_qtail)) < 4931 (int)(10 << flow_table->log))) 4932 expire = false; 4933 } 4934 rcu_read_unlock(); 4935 return expire; 4936 } 4937 EXPORT_SYMBOL(rps_may_expire_flow); 4938 4939 #endif /* CONFIG_RFS_ACCEL */ 4940 4941 /* Called from hardirq (IPI) context */ 4942 static void rps_trigger_softirq(void *data) 4943 { 4944 struct softnet_data *sd = data; 4945 4946 ____napi_schedule(sd, &sd->backlog); 4947 sd->received_rps++; 4948 } 4949 4950 #endif /* CONFIG_RPS */ 4951 4952 /* Called from hardirq (IPI) context */ 4953 static void trigger_rx_softirq(void *data) 4954 { 4955 struct softnet_data *sd = data; 4956 4957 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4958 smp_store_release(&sd->defer_ipi_scheduled, 0); 4959 } 4960 4961 /* 4962 * After we queued a packet into sd->input_pkt_queue, 4963 * we need to make sure this queue is serviced soon. 4964 * 4965 * - If this is another cpu queue, link it to our rps_ipi_list, 4966 * and make sure we will process rps_ipi_list from net_rx_action(). 4967 * 4968 * - If this is our own queue, NAPI schedule our backlog. 4969 * Note that this also raises NET_RX_SOFTIRQ. 4970 */ 4971 static void napi_schedule_rps(struct softnet_data *sd) 4972 { 4973 struct softnet_data *mysd = this_cpu_ptr(&softnet_data); 4974 4975 #ifdef CONFIG_RPS 4976 if (sd != mysd) { 4977 if (use_backlog_threads()) { 4978 __napi_schedule_irqoff(&sd->backlog); 4979 return; 4980 } 4981 4982 sd->rps_ipi_next = mysd->rps_ipi_list; 4983 mysd->rps_ipi_list = sd; 4984 4985 /* If not called from net_rx_action() or napi_threaded_poll() 4986 * we have to raise NET_RX_SOFTIRQ. 4987 */ 4988 if (!mysd->in_net_rx_action && !mysd->in_napi_threaded_poll) 4989 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4990 return; 4991 } 4992 #endif /* CONFIG_RPS */ 4993 __napi_schedule_irqoff(&mysd->backlog); 4994 } 4995 4996 void kick_defer_list_purge(struct softnet_data *sd, unsigned int cpu) 4997 { 4998 unsigned long flags; 4999 5000 if (use_backlog_threads()) { 5001 backlog_lock_irq_save(sd, &flags); 5002 5003 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) 5004 __napi_schedule_irqoff(&sd->backlog); 5005 5006 backlog_unlock_irq_restore(sd, &flags); 5007 5008 } else if (!cmpxchg(&sd->defer_ipi_scheduled, 0, 1)) { 5009 smp_call_function_single_async(cpu, &sd->defer_csd); 5010 } 5011 } 5012 5013 #ifdef CONFIG_NET_FLOW_LIMIT 5014 int netdev_flow_limit_table_len __read_mostly = (1 << 12); 5015 #endif 5016 5017 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen) 5018 { 5019 #ifdef CONFIG_NET_FLOW_LIMIT 5020 struct sd_flow_limit *fl; 5021 struct softnet_data *sd; 5022 unsigned int old_flow, new_flow; 5023 5024 if (qlen < (READ_ONCE(net_hotdata.max_backlog) >> 1)) 5025 return false; 5026 5027 sd = this_cpu_ptr(&softnet_data); 5028 5029 rcu_read_lock(); 5030 fl = rcu_dereference(sd->flow_limit); 5031 if (fl) { 5032 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1); 5033 old_flow = fl->history[fl->history_head]; 5034 fl->history[fl->history_head] = new_flow; 5035 5036 fl->history_head++; 5037 fl->history_head &= FLOW_LIMIT_HISTORY - 1; 5038 5039 if (likely(fl->buckets[old_flow])) 5040 fl->buckets[old_flow]--; 5041 5042 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) { 5043 fl->count++; 5044 rcu_read_unlock(); 5045 return true; 5046 } 5047 } 5048 rcu_read_unlock(); 5049 #endif 5050 return false; 5051 } 5052 5053 /* 5054 * enqueue_to_backlog is called to queue an skb to a per CPU backlog 5055 * queue (may be a remote CPU queue). 5056 */ 5057 static int enqueue_to_backlog(struct sk_buff *skb, int cpu, 5058 unsigned int *qtail) 5059 { 5060 enum skb_drop_reason reason; 5061 struct softnet_data *sd; 5062 unsigned long flags; 5063 unsigned int qlen; 5064 int max_backlog; 5065 u32 tail; 5066 5067 reason = SKB_DROP_REASON_DEV_READY; 5068 if (!netif_running(skb->dev)) 5069 goto bad_dev; 5070 5071 reason = SKB_DROP_REASON_CPU_BACKLOG; 5072 sd = &per_cpu(softnet_data, cpu); 5073 5074 qlen = skb_queue_len_lockless(&sd->input_pkt_queue); 5075 max_backlog = READ_ONCE(net_hotdata.max_backlog); 5076 if (unlikely(qlen > max_backlog)) 5077 goto cpu_backlog_drop; 5078 backlog_lock_irq_save(sd, &flags); 5079 qlen = skb_queue_len(&sd->input_pkt_queue); 5080 if (qlen <= max_backlog && !skb_flow_limit(skb, qlen)) { 5081 if (!qlen) { 5082 /* Schedule NAPI for backlog device. We can use 5083 * non atomic operation as we own the queue lock. 5084 */ 5085 if (!__test_and_set_bit(NAPI_STATE_SCHED, 5086 &sd->backlog.state)) 5087 napi_schedule_rps(sd); 5088 } 5089 __skb_queue_tail(&sd->input_pkt_queue, skb); 5090 tail = rps_input_queue_tail_incr(sd); 5091 backlog_unlock_irq_restore(sd, &flags); 5092 5093 /* save the tail outside of the critical section */ 5094 rps_input_queue_tail_save(qtail, tail); 5095 return NET_RX_SUCCESS; 5096 } 5097 5098 backlog_unlock_irq_restore(sd, &flags); 5099 5100 cpu_backlog_drop: 5101 atomic_inc(&sd->dropped); 5102 bad_dev: 5103 dev_core_stats_rx_dropped_inc(skb->dev); 5104 kfree_skb_reason(skb, reason); 5105 return NET_RX_DROP; 5106 } 5107 5108 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb) 5109 { 5110 struct net_device *dev = skb->dev; 5111 struct netdev_rx_queue *rxqueue; 5112 5113 rxqueue = dev->_rx; 5114 5115 if (skb_rx_queue_recorded(skb)) { 5116 u16 index = skb_get_rx_queue(skb); 5117 5118 if (unlikely(index >= dev->real_num_rx_queues)) { 5119 WARN_ONCE(dev->real_num_rx_queues > 1, 5120 "%s received packet on queue %u, but number " 5121 "of RX queues is %u\n", 5122 dev->name, index, dev->real_num_rx_queues); 5123 5124 return rxqueue; /* Return first rxqueue */ 5125 } 5126 rxqueue += index; 5127 } 5128 return rxqueue; 5129 } 5130 5131 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, 5132 const struct bpf_prog *xdp_prog) 5133 { 5134 void *orig_data, *orig_data_end, *hard_start; 5135 struct netdev_rx_queue *rxqueue; 5136 bool orig_bcast, orig_host; 5137 u32 mac_len, frame_sz; 5138 __be16 orig_eth_type; 5139 struct ethhdr *eth; 5140 u32 metalen, act; 5141 int off; 5142 5143 /* The XDP program wants to see the packet starting at the MAC 5144 * header. 5145 */ 5146 mac_len = skb->data - skb_mac_header(skb); 5147 hard_start = skb->data - skb_headroom(skb); 5148 5149 /* SKB "head" area always have tailroom for skb_shared_info */ 5150 frame_sz = (void *)skb_end_pointer(skb) - hard_start; 5151 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 5152 5153 rxqueue = netif_get_rxqueue(skb); 5154 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq); 5155 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len, 5156 skb_headlen(skb) + mac_len, true); 5157 if (skb_is_nonlinear(skb)) { 5158 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 5159 xdp_buff_set_frags_flag(xdp); 5160 } else { 5161 xdp_buff_clear_frags_flag(xdp); 5162 } 5163 5164 orig_data_end = xdp->data_end; 5165 orig_data = xdp->data; 5166 eth = (struct ethhdr *)xdp->data; 5167 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr); 5168 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest); 5169 orig_eth_type = eth->h_proto; 5170 5171 act = bpf_prog_run_xdp(xdp_prog, xdp); 5172 5173 /* check if bpf_xdp_adjust_head was used */ 5174 off = xdp->data - orig_data; 5175 if (off) { 5176 if (off > 0) 5177 __skb_pull(skb, off); 5178 else if (off < 0) 5179 __skb_push(skb, -off); 5180 5181 skb->mac_header += off; 5182 skb_reset_network_header(skb); 5183 } 5184 5185 /* check if bpf_xdp_adjust_tail was used */ 5186 off = xdp->data_end - orig_data_end; 5187 if (off != 0) { 5188 skb_set_tail_pointer(skb, xdp->data_end - xdp->data); 5189 skb->len += off; /* positive on grow, negative on shrink */ 5190 } 5191 5192 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 5193 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here. 5194 */ 5195 if (xdp_buff_has_frags(xdp)) 5196 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 5197 else 5198 skb->data_len = 0; 5199 5200 /* check if XDP changed eth hdr such SKB needs update */ 5201 eth = (struct ethhdr *)xdp->data; 5202 if ((orig_eth_type != eth->h_proto) || 5203 (orig_host != ether_addr_equal_64bits(eth->h_dest, 5204 skb->dev->dev_addr)) || 5205 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) { 5206 __skb_push(skb, ETH_HLEN); 5207 skb->pkt_type = PACKET_HOST; 5208 skb->protocol = eth_type_trans(skb, skb->dev); 5209 } 5210 5211 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull 5212 * before calling us again on redirect path. We do not call do_redirect 5213 * as we leave that up to the caller. 5214 * 5215 * Caller is responsible for managing lifetime of skb (i.e. calling 5216 * kfree_skb in response to actions it cannot handle/XDP_DROP). 5217 */ 5218 switch (act) { 5219 case XDP_REDIRECT: 5220 case XDP_TX: 5221 __skb_push(skb, mac_len); 5222 break; 5223 case XDP_PASS: 5224 metalen = xdp->data - xdp->data_meta; 5225 if (metalen) 5226 skb_metadata_set(skb, metalen); 5227 break; 5228 } 5229 5230 return act; 5231 } 5232 5233 static int 5234 netif_skb_check_for_xdp(struct sk_buff **pskb, const struct bpf_prog *prog) 5235 { 5236 struct sk_buff *skb = *pskb; 5237 int err, hroom, troom; 5238 5239 if (!skb_cow_data_for_xdp(this_cpu_read(system_page_pool), pskb, prog)) 5240 return 0; 5241 5242 /* In case we have to go down the path and also linearize, 5243 * then lets do the pskb_expand_head() work just once here. 5244 */ 5245 hroom = XDP_PACKET_HEADROOM - skb_headroom(skb); 5246 troom = skb->tail + skb->data_len - skb->end; 5247 err = pskb_expand_head(skb, 5248 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0, 5249 troom > 0 ? troom + 128 : 0, GFP_ATOMIC); 5250 if (err) 5251 return err; 5252 5253 return skb_linearize(skb); 5254 } 5255 5256 static u32 netif_receive_generic_xdp(struct sk_buff **pskb, 5257 struct xdp_buff *xdp, 5258 const struct bpf_prog *xdp_prog) 5259 { 5260 struct sk_buff *skb = *pskb; 5261 u32 mac_len, act = XDP_DROP; 5262 5263 /* Reinjected packets coming from act_mirred or similar should 5264 * not get XDP generic processing. 5265 */ 5266 if (skb_is_redirected(skb)) 5267 return XDP_PASS; 5268 5269 /* XDP packets must have sufficient headroom of XDP_PACKET_HEADROOM 5270 * bytes. This is the guarantee that also native XDP provides, 5271 * thus we need to do it here as well. 5272 */ 5273 mac_len = skb->data - skb_mac_header(skb); 5274 __skb_push(skb, mac_len); 5275 5276 if (skb_cloned(skb) || skb_is_nonlinear(skb) || 5277 skb_headroom(skb) < XDP_PACKET_HEADROOM) { 5278 if (netif_skb_check_for_xdp(pskb, xdp_prog)) 5279 goto do_drop; 5280 } 5281 5282 __skb_pull(*pskb, mac_len); 5283 5284 act = bpf_prog_run_generic_xdp(*pskb, xdp, xdp_prog); 5285 switch (act) { 5286 case XDP_REDIRECT: 5287 case XDP_TX: 5288 case XDP_PASS: 5289 break; 5290 default: 5291 bpf_warn_invalid_xdp_action((*pskb)->dev, xdp_prog, act); 5292 fallthrough; 5293 case XDP_ABORTED: 5294 trace_xdp_exception((*pskb)->dev, xdp_prog, act); 5295 fallthrough; 5296 case XDP_DROP: 5297 do_drop: 5298 kfree_skb(*pskb); 5299 break; 5300 } 5301 5302 return act; 5303 } 5304 5305 /* When doing generic XDP we have to bypass the qdisc layer and the 5306 * network taps in order to match in-driver-XDP behavior. This also means 5307 * that XDP packets are able to starve other packets going through a qdisc, 5308 * and DDOS attacks will be more effective. In-driver-XDP use dedicated TX 5309 * queues, so they do not have this starvation issue. 5310 */ 5311 void generic_xdp_tx(struct sk_buff *skb, const struct bpf_prog *xdp_prog) 5312 { 5313 struct net_device *dev = skb->dev; 5314 struct netdev_queue *txq; 5315 bool free_skb = true; 5316 int cpu, rc; 5317 5318 txq = netdev_core_pick_tx(dev, skb, NULL); 5319 cpu = smp_processor_id(); 5320 HARD_TX_LOCK(dev, txq, cpu); 5321 if (!netif_xmit_frozen_or_drv_stopped(txq)) { 5322 rc = netdev_start_xmit(skb, dev, txq, 0); 5323 if (dev_xmit_complete(rc)) 5324 free_skb = false; 5325 } 5326 HARD_TX_UNLOCK(dev, txq); 5327 if (free_skb) { 5328 trace_xdp_exception(dev, xdp_prog, XDP_TX); 5329 dev_core_stats_tx_dropped_inc(dev); 5330 kfree_skb(skb); 5331 } 5332 } 5333 5334 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key); 5335 5336 int do_xdp_generic(const struct bpf_prog *xdp_prog, struct sk_buff **pskb) 5337 { 5338 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 5339 5340 if (xdp_prog) { 5341 struct xdp_buff xdp; 5342 u32 act; 5343 int err; 5344 5345 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 5346 act = netif_receive_generic_xdp(pskb, &xdp, xdp_prog); 5347 if (act != XDP_PASS) { 5348 switch (act) { 5349 case XDP_REDIRECT: 5350 err = xdp_do_generic_redirect((*pskb)->dev, *pskb, 5351 &xdp, xdp_prog); 5352 if (err) 5353 goto out_redir; 5354 break; 5355 case XDP_TX: 5356 generic_xdp_tx(*pskb, xdp_prog); 5357 break; 5358 } 5359 bpf_net_ctx_clear(bpf_net_ctx); 5360 return XDP_DROP; 5361 } 5362 bpf_net_ctx_clear(bpf_net_ctx); 5363 } 5364 return XDP_PASS; 5365 out_redir: 5366 bpf_net_ctx_clear(bpf_net_ctx); 5367 kfree_skb_reason(*pskb, SKB_DROP_REASON_XDP); 5368 return XDP_DROP; 5369 } 5370 EXPORT_SYMBOL_GPL(do_xdp_generic); 5371 5372 static int netif_rx_internal(struct sk_buff *skb) 5373 { 5374 int ret; 5375 5376 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5377 5378 trace_netif_rx(skb); 5379 5380 #ifdef CONFIG_RPS 5381 if (static_branch_unlikely(&rps_needed)) { 5382 struct rps_dev_flow voidflow, *rflow = &voidflow; 5383 int cpu; 5384 5385 rcu_read_lock(); 5386 5387 cpu = get_rps_cpu(skb->dev, skb, &rflow); 5388 if (cpu < 0) 5389 cpu = smp_processor_id(); 5390 5391 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5392 5393 rcu_read_unlock(); 5394 } else 5395 #endif 5396 { 5397 unsigned int qtail; 5398 5399 ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail); 5400 } 5401 return ret; 5402 } 5403 5404 /** 5405 * __netif_rx - Slightly optimized version of netif_rx 5406 * @skb: buffer to post 5407 * 5408 * This behaves as netif_rx except that it does not disable bottom halves. 5409 * As a result this function may only be invoked from the interrupt context 5410 * (either hard or soft interrupt). 5411 */ 5412 int __netif_rx(struct sk_buff *skb) 5413 { 5414 int ret; 5415 5416 lockdep_assert_once(hardirq_count() | softirq_count()); 5417 5418 trace_netif_rx_entry(skb); 5419 ret = netif_rx_internal(skb); 5420 trace_netif_rx_exit(ret); 5421 return ret; 5422 } 5423 EXPORT_SYMBOL(__netif_rx); 5424 5425 /** 5426 * netif_rx - post buffer to the network code 5427 * @skb: buffer to post 5428 * 5429 * This function receives a packet from a device driver and queues it for 5430 * the upper (protocol) levels to process via the backlog NAPI device. It 5431 * always succeeds. The buffer may be dropped during processing for 5432 * congestion control or by the protocol layers. 5433 * The network buffer is passed via the backlog NAPI device. Modern NIC 5434 * driver should use NAPI and GRO. 5435 * This function can used from interrupt and from process context. The 5436 * caller from process context must not disable interrupts before invoking 5437 * this function. 5438 * 5439 * return values: 5440 * NET_RX_SUCCESS (no congestion) 5441 * NET_RX_DROP (packet was dropped) 5442 * 5443 */ 5444 int netif_rx(struct sk_buff *skb) 5445 { 5446 bool need_bh_off = !(hardirq_count() | softirq_count()); 5447 int ret; 5448 5449 if (need_bh_off) 5450 local_bh_disable(); 5451 trace_netif_rx_entry(skb); 5452 ret = netif_rx_internal(skb); 5453 trace_netif_rx_exit(ret); 5454 if (need_bh_off) 5455 local_bh_enable(); 5456 return ret; 5457 } 5458 EXPORT_SYMBOL(netif_rx); 5459 5460 static __latent_entropy void net_tx_action(void) 5461 { 5462 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 5463 5464 if (sd->completion_queue) { 5465 struct sk_buff *clist; 5466 5467 local_irq_disable(); 5468 clist = sd->completion_queue; 5469 sd->completion_queue = NULL; 5470 local_irq_enable(); 5471 5472 while (clist) { 5473 struct sk_buff *skb = clist; 5474 5475 clist = clist->next; 5476 5477 WARN_ON(refcount_read(&skb->users)); 5478 if (likely(get_kfree_skb_cb(skb)->reason == SKB_CONSUMED)) 5479 trace_consume_skb(skb, net_tx_action); 5480 else 5481 trace_kfree_skb(skb, net_tx_action, 5482 get_kfree_skb_cb(skb)->reason, NULL); 5483 5484 if (skb->fclone != SKB_FCLONE_UNAVAILABLE) 5485 __kfree_skb(skb); 5486 else 5487 __napi_kfree_skb(skb, 5488 get_kfree_skb_cb(skb)->reason); 5489 } 5490 } 5491 5492 if (sd->output_queue) { 5493 struct Qdisc *head; 5494 5495 local_irq_disable(); 5496 head = sd->output_queue; 5497 sd->output_queue = NULL; 5498 sd->output_queue_tailp = &sd->output_queue; 5499 local_irq_enable(); 5500 5501 rcu_read_lock(); 5502 5503 while (head) { 5504 struct Qdisc *q = head; 5505 spinlock_t *root_lock = NULL; 5506 5507 head = head->next_sched; 5508 5509 /* We need to make sure head->next_sched is read 5510 * before clearing __QDISC_STATE_SCHED 5511 */ 5512 smp_mb__before_atomic(); 5513 5514 if (!(q->flags & TCQ_F_NOLOCK)) { 5515 root_lock = qdisc_lock(q); 5516 spin_lock(root_lock); 5517 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, 5518 &q->state))) { 5519 /* There is a synchronize_net() between 5520 * STATE_DEACTIVATED flag being set and 5521 * qdisc_reset()/some_qdisc_is_busy() in 5522 * dev_deactivate(), so we can safely bail out 5523 * early here to avoid data race between 5524 * qdisc_deactivate() and some_qdisc_is_busy() 5525 * for lockless qdisc. 5526 */ 5527 clear_bit(__QDISC_STATE_SCHED, &q->state); 5528 continue; 5529 } 5530 5531 clear_bit(__QDISC_STATE_SCHED, &q->state); 5532 qdisc_run(q); 5533 if (root_lock) 5534 spin_unlock(root_lock); 5535 } 5536 5537 rcu_read_unlock(); 5538 } 5539 5540 xfrm_dev_backlog(sd); 5541 } 5542 5543 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE) 5544 /* This hook is defined here for ATM LANE */ 5545 int (*br_fdb_test_addr_hook)(struct net_device *dev, 5546 unsigned char *addr) __read_mostly; 5547 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook); 5548 #endif 5549 5550 /** 5551 * netdev_is_rx_handler_busy - check if receive handler is registered 5552 * @dev: device to check 5553 * 5554 * Check if a receive handler is already registered for a given device. 5555 * Return true if there one. 5556 * 5557 * The caller must hold the rtnl_mutex. 5558 */ 5559 bool netdev_is_rx_handler_busy(struct net_device *dev) 5560 { 5561 ASSERT_RTNL(); 5562 return dev && rtnl_dereference(dev->rx_handler); 5563 } 5564 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy); 5565 5566 /** 5567 * netdev_rx_handler_register - register receive handler 5568 * @dev: device to register a handler for 5569 * @rx_handler: receive handler to register 5570 * @rx_handler_data: data pointer that is used by rx handler 5571 * 5572 * Register a receive handler for a device. This handler will then be 5573 * called from __netif_receive_skb. A negative errno code is returned 5574 * on a failure. 5575 * 5576 * The caller must hold the rtnl_mutex. 5577 * 5578 * For a general description of rx_handler, see enum rx_handler_result. 5579 */ 5580 int netdev_rx_handler_register(struct net_device *dev, 5581 rx_handler_func_t *rx_handler, 5582 void *rx_handler_data) 5583 { 5584 if (netdev_is_rx_handler_busy(dev)) 5585 return -EBUSY; 5586 5587 if (dev->priv_flags & IFF_NO_RX_HANDLER) 5588 return -EINVAL; 5589 5590 /* Note: rx_handler_data must be set before rx_handler */ 5591 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data); 5592 rcu_assign_pointer(dev->rx_handler, rx_handler); 5593 5594 return 0; 5595 } 5596 EXPORT_SYMBOL_GPL(netdev_rx_handler_register); 5597 5598 /** 5599 * netdev_rx_handler_unregister - unregister receive handler 5600 * @dev: device to unregister a handler from 5601 * 5602 * Unregister a receive handler from a device. 5603 * 5604 * The caller must hold the rtnl_mutex. 5605 */ 5606 void netdev_rx_handler_unregister(struct net_device *dev) 5607 { 5608 5609 ASSERT_RTNL(); 5610 RCU_INIT_POINTER(dev->rx_handler, NULL); 5611 /* a reader seeing a non NULL rx_handler in a rcu_read_lock() 5612 * section has a guarantee to see a non NULL rx_handler_data 5613 * as well. 5614 */ 5615 synchronize_net(); 5616 RCU_INIT_POINTER(dev->rx_handler_data, NULL); 5617 } 5618 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister); 5619 5620 /* 5621 * Limit the use of PFMEMALLOC reserves to those protocols that implement 5622 * the special handling of PFMEMALLOC skbs. 5623 */ 5624 static bool skb_pfmemalloc_protocol(struct sk_buff *skb) 5625 { 5626 switch (skb->protocol) { 5627 case htons(ETH_P_ARP): 5628 case htons(ETH_P_IP): 5629 case htons(ETH_P_IPV6): 5630 case htons(ETH_P_8021Q): 5631 case htons(ETH_P_8021AD): 5632 return true; 5633 default: 5634 return false; 5635 } 5636 } 5637 5638 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev, 5639 int *ret, struct net_device *orig_dev) 5640 { 5641 if (nf_hook_ingress_active(skb)) { 5642 int ingress_retval; 5643 5644 if (*pt_prev) { 5645 *ret = deliver_skb(skb, *pt_prev, orig_dev); 5646 *pt_prev = NULL; 5647 } 5648 5649 rcu_read_lock(); 5650 ingress_retval = nf_hook_ingress(skb); 5651 rcu_read_unlock(); 5652 return ingress_retval; 5653 } 5654 return 0; 5655 } 5656 5657 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc, 5658 struct packet_type **ppt_prev) 5659 { 5660 struct packet_type *ptype, *pt_prev; 5661 rx_handler_func_t *rx_handler; 5662 struct sk_buff *skb = *pskb; 5663 struct net_device *orig_dev; 5664 bool deliver_exact = false; 5665 int ret = NET_RX_DROP; 5666 __be16 type; 5667 5668 net_timestamp_check(!READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5669 5670 trace_netif_receive_skb(skb); 5671 5672 orig_dev = skb->dev; 5673 5674 skb_reset_network_header(skb); 5675 #if !defined(CONFIG_DEBUG_NET) 5676 /* We plan to no longer reset the transport header here. 5677 * Give some time to fuzzers and dev build to catch bugs 5678 * in network stacks. 5679 */ 5680 if (!skb_transport_header_was_set(skb)) 5681 skb_reset_transport_header(skb); 5682 #endif 5683 skb_reset_mac_len(skb); 5684 5685 pt_prev = NULL; 5686 5687 another_round: 5688 skb->skb_iif = skb->dev->ifindex; 5689 5690 __this_cpu_inc(softnet_data.processed); 5691 5692 if (static_branch_unlikely(&generic_xdp_needed_key)) { 5693 int ret2; 5694 5695 migrate_disable(); 5696 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), 5697 &skb); 5698 migrate_enable(); 5699 5700 if (ret2 != XDP_PASS) { 5701 ret = NET_RX_DROP; 5702 goto out; 5703 } 5704 } 5705 5706 if (eth_type_vlan(skb->protocol)) { 5707 skb = skb_vlan_untag(skb); 5708 if (unlikely(!skb)) 5709 goto out; 5710 } 5711 5712 if (skb_skip_tc_classify(skb)) 5713 goto skip_classify; 5714 5715 if (pfmemalloc) 5716 goto skip_taps; 5717 5718 list_for_each_entry_rcu(ptype, &dev_net_rcu(skb->dev)->ptype_all, 5719 list) { 5720 if (pt_prev) 5721 ret = deliver_skb(skb, pt_prev, orig_dev); 5722 pt_prev = ptype; 5723 } 5724 5725 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) { 5726 if (pt_prev) 5727 ret = deliver_skb(skb, pt_prev, orig_dev); 5728 pt_prev = ptype; 5729 } 5730 5731 skip_taps: 5732 #ifdef CONFIG_NET_INGRESS 5733 if (static_branch_unlikely(&ingress_needed_key)) { 5734 bool another = false; 5735 5736 nf_skip_egress(skb, true); 5737 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev, 5738 &another); 5739 if (another) 5740 goto another_round; 5741 if (!skb) 5742 goto out; 5743 5744 nf_skip_egress(skb, false); 5745 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0) 5746 goto out; 5747 } 5748 #endif 5749 skb_reset_redirect(skb); 5750 skip_classify: 5751 if (pfmemalloc && !skb_pfmemalloc_protocol(skb)) 5752 goto drop; 5753 5754 if (skb_vlan_tag_present(skb)) { 5755 if (pt_prev) { 5756 ret = deliver_skb(skb, pt_prev, orig_dev); 5757 pt_prev = NULL; 5758 } 5759 if (vlan_do_receive(&skb)) 5760 goto another_round; 5761 else if (unlikely(!skb)) 5762 goto out; 5763 } 5764 5765 rx_handler = rcu_dereference(skb->dev->rx_handler); 5766 if (rx_handler) { 5767 if (pt_prev) { 5768 ret = deliver_skb(skb, pt_prev, orig_dev); 5769 pt_prev = NULL; 5770 } 5771 switch (rx_handler(&skb)) { 5772 case RX_HANDLER_CONSUMED: 5773 ret = NET_RX_SUCCESS; 5774 goto out; 5775 case RX_HANDLER_ANOTHER: 5776 goto another_round; 5777 case RX_HANDLER_EXACT: 5778 deliver_exact = true; 5779 break; 5780 case RX_HANDLER_PASS: 5781 break; 5782 default: 5783 BUG(); 5784 } 5785 } 5786 5787 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) { 5788 check_vlan_id: 5789 if (skb_vlan_tag_get_id(skb)) { 5790 /* Vlan id is non 0 and vlan_do_receive() above couldn't 5791 * find vlan device. 5792 */ 5793 skb->pkt_type = PACKET_OTHERHOST; 5794 } else if (eth_type_vlan(skb->protocol)) { 5795 /* Outer header is 802.1P with vlan 0, inner header is 5796 * 802.1Q or 802.1AD and vlan_do_receive() above could 5797 * not find vlan dev for vlan id 0. 5798 */ 5799 __vlan_hwaccel_clear_tag(skb); 5800 skb = skb_vlan_untag(skb); 5801 if (unlikely(!skb)) 5802 goto out; 5803 if (vlan_do_receive(&skb)) 5804 /* After stripping off 802.1P header with vlan 0 5805 * vlan dev is found for inner header. 5806 */ 5807 goto another_round; 5808 else if (unlikely(!skb)) 5809 goto out; 5810 else 5811 /* We have stripped outer 802.1P vlan 0 header. 5812 * But could not find vlan dev. 5813 * check again for vlan id to set OTHERHOST. 5814 */ 5815 goto check_vlan_id; 5816 } 5817 /* Note: we might in the future use prio bits 5818 * and set skb->priority like in vlan_do_receive() 5819 * For the time being, just ignore Priority Code Point 5820 */ 5821 __vlan_hwaccel_clear_tag(skb); 5822 } 5823 5824 type = skb->protocol; 5825 5826 /* deliver only exact match when indicated */ 5827 if (likely(!deliver_exact)) { 5828 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5829 &ptype_base[ntohs(type) & 5830 PTYPE_HASH_MASK]); 5831 5832 /* orig_dev and skb->dev could belong to different netns; 5833 * Even in such case we need to traverse only the list 5834 * coming from skb->dev, as the ptype owner (packet socket) 5835 * will use dev_net(skb->dev) to do namespace filtering. 5836 */ 5837 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5838 &dev_net_rcu(skb->dev)->ptype_specific); 5839 } 5840 5841 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5842 &orig_dev->ptype_specific); 5843 5844 if (unlikely(skb->dev != orig_dev)) { 5845 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5846 &skb->dev->ptype_specific); 5847 } 5848 5849 if (pt_prev) { 5850 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 5851 goto drop; 5852 *ppt_prev = pt_prev; 5853 } else { 5854 drop: 5855 if (!deliver_exact) 5856 dev_core_stats_rx_dropped_inc(skb->dev); 5857 else 5858 dev_core_stats_rx_nohandler_inc(skb->dev); 5859 kfree_skb_reason(skb, SKB_DROP_REASON_UNHANDLED_PROTO); 5860 /* Jamal, now you will not able to escape explaining 5861 * me how you were going to use this. :-) 5862 */ 5863 ret = NET_RX_DROP; 5864 } 5865 5866 out: 5867 /* The invariant here is that if *ppt_prev is not NULL 5868 * then skb should also be non-NULL. 5869 * 5870 * Apparently *ppt_prev assignment above holds this invariant due to 5871 * skb dereferencing near it. 5872 */ 5873 *pskb = skb; 5874 return ret; 5875 } 5876 5877 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc) 5878 { 5879 struct net_device *orig_dev = skb->dev; 5880 struct packet_type *pt_prev = NULL; 5881 int ret; 5882 5883 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5884 if (pt_prev) 5885 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb, 5886 skb->dev, pt_prev, orig_dev); 5887 return ret; 5888 } 5889 5890 /** 5891 * netif_receive_skb_core - special purpose version of netif_receive_skb 5892 * @skb: buffer to process 5893 * 5894 * More direct receive version of netif_receive_skb(). It should 5895 * only be used by callers that have a need to skip RPS and Generic XDP. 5896 * Caller must also take care of handling if ``(page_is_)pfmemalloc``. 5897 * 5898 * This function may only be called from softirq context and interrupts 5899 * should be enabled. 5900 * 5901 * Return values (usually ignored): 5902 * NET_RX_SUCCESS: no congestion 5903 * NET_RX_DROP: packet was dropped 5904 */ 5905 int netif_receive_skb_core(struct sk_buff *skb) 5906 { 5907 int ret; 5908 5909 rcu_read_lock(); 5910 ret = __netif_receive_skb_one_core(skb, false); 5911 rcu_read_unlock(); 5912 5913 return ret; 5914 } 5915 EXPORT_SYMBOL(netif_receive_skb_core); 5916 5917 static inline void __netif_receive_skb_list_ptype(struct list_head *head, 5918 struct packet_type *pt_prev, 5919 struct net_device *orig_dev) 5920 { 5921 struct sk_buff *skb, *next; 5922 5923 if (!pt_prev) 5924 return; 5925 if (list_empty(head)) 5926 return; 5927 if (pt_prev->list_func != NULL) 5928 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv, 5929 ip_list_rcv, head, pt_prev, orig_dev); 5930 else 5931 list_for_each_entry_safe(skb, next, head, list) { 5932 skb_list_del_init(skb); 5933 pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 5934 } 5935 } 5936 5937 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc) 5938 { 5939 /* Fast-path assumptions: 5940 * - There is no RX handler. 5941 * - Only one packet_type matches. 5942 * If either of these fails, we will end up doing some per-packet 5943 * processing in-line, then handling the 'last ptype' for the whole 5944 * sublist. This can't cause out-of-order delivery to any single ptype, 5945 * because the 'last ptype' must be constant across the sublist, and all 5946 * other ptypes are handled per-packet. 5947 */ 5948 /* Current (common) ptype of sublist */ 5949 struct packet_type *pt_curr = NULL; 5950 /* Current (common) orig_dev of sublist */ 5951 struct net_device *od_curr = NULL; 5952 struct sk_buff *skb, *next; 5953 LIST_HEAD(sublist); 5954 5955 list_for_each_entry_safe(skb, next, head, list) { 5956 struct net_device *orig_dev = skb->dev; 5957 struct packet_type *pt_prev = NULL; 5958 5959 skb_list_del_init(skb); 5960 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5961 if (!pt_prev) 5962 continue; 5963 if (pt_curr != pt_prev || od_curr != orig_dev) { 5964 /* dispatch old sublist */ 5965 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5966 /* start new sublist */ 5967 INIT_LIST_HEAD(&sublist); 5968 pt_curr = pt_prev; 5969 od_curr = orig_dev; 5970 } 5971 list_add_tail(&skb->list, &sublist); 5972 } 5973 5974 /* dispatch final sublist */ 5975 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5976 } 5977 5978 static int __netif_receive_skb(struct sk_buff *skb) 5979 { 5980 int ret; 5981 5982 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) { 5983 unsigned int noreclaim_flag; 5984 5985 /* 5986 * PFMEMALLOC skbs are special, they should 5987 * - be delivered to SOCK_MEMALLOC sockets only 5988 * - stay away from userspace 5989 * - have bounded memory usage 5990 * 5991 * Use PF_MEMALLOC as this saves us from propagating the allocation 5992 * context down to all allocation sites. 5993 */ 5994 noreclaim_flag = memalloc_noreclaim_save(); 5995 ret = __netif_receive_skb_one_core(skb, true); 5996 memalloc_noreclaim_restore(noreclaim_flag); 5997 } else 5998 ret = __netif_receive_skb_one_core(skb, false); 5999 6000 return ret; 6001 } 6002 6003 static void __netif_receive_skb_list(struct list_head *head) 6004 { 6005 unsigned long noreclaim_flag = 0; 6006 struct sk_buff *skb, *next; 6007 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */ 6008 6009 list_for_each_entry_safe(skb, next, head, list) { 6010 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) { 6011 struct list_head sublist; 6012 6013 /* Handle the previous sublist */ 6014 list_cut_before(&sublist, head, &skb->list); 6015 if (!list_empty(&sublist)) 6016 __netif_receive_skb_list_core(&sublist, pfmemalloc); 6017 pfmemalloc = !pfmemalloc; 6018 /* See comments in __netif_receive_skb */ 6019 if (pfmemalloc) 6020 noreclaim_flag = memalloc_noreclaim_save(); 6021 else 6022 memalloc_noreclaim_restore(noreclaim_flag); 6023 } 6024 } 6025 /* Handle the remaining sublist */ 6026 if (!list_empty(head)) 6027 __netif_receive_skb_list_core(head, pfmemalloc); 6028 /* Restore pflags */ 6029 if (pfmemalloc) 6030 memalloc_noreclaim_restore(noreclaim_flag); 6031 } 6032 6033 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp) 6034 { 6035 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog); 6036 struct bpf_prog *new = xdp->prog; 6037 int ret = 0; 6038 6039 switch (xdp->command) { 6040 case XDP_SETUP_PROG: 6041 rcu_assign_pointer(dev->xdp_prog, new); 6042 if (old) 6043 bpf_prog_put(old); 6044 6045 if (old && !new) { 6046 static_branch_dec(&generic_xdp_needed_key); 6047 } else if (new && !old) { 6048 static_branch_inc(&generic_xdp_needed_key); 6049 netif_disable_lro(dev); 6050 dev_disable_gro_hw(dev); 6051 } 6052 break; 6053 6054 default: 6055 ret = -EINVAL; 6056 break; 6057 } 6058 6059 return ret; 6060 } 6061 6062 static int netif_receive_skb_internal(struct sk_buff *skb) 6063 { 6064 int ret; 6065 6066 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 6067 6068 if (skb_defer_rx_timestamp(skb)) 6069 return NET_RX_SUCCESS; 6070 6071 rcu_read_lock(); 6072 #ifdef CONFIG_RPS 6073 if (static_branch_unlikely(&rps_needed)) { 6074 struct rps_dev_flow voidflow, *rflow = &voidflow; 6075 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6076 6077 if (cpu >= 0) { 6078 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6079 rcu_read_unlock(); 6080 return ret; 6081 } 6082 } 6083 #endif 6084 ret = __netif_receive_skb(skb); 6085 rcu_read_unlock(); 6086 return ret; 6087 } 6088 6089 void netif_receive_skb_list_internal(struct list_head *head) 6090 { 6091 struct sk_buff *skb, *next; 6092 LIST_HEAD(sublist); 6093 6094 list_for_each_entry_safe(skb, next, head, list) { 6095 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), 6096 skb); 6097 skb_list_del_init(skb); 6098 if (!skb_defer_rx_timestamp(skb)) 6099 list_add_tail(&skb->list, &sublist); 6100 } 6101 list_splice_init(&sublist, head); 6102 6103 rcu_read_lock(); 6104 #ifdef CONFIG_RPS 6105 if (static_branch_unlikely(&rps_needed)) { 6106 list_for_each_entry_safe(skb, next, head, list) { 6107 struct rps_dev_flow voidflow, *rflow = &voidflow; 6108 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6109 6110 if (cpu >= 0) { 6111 /* Will be handled, remove from list */ 6112 skb_list_del_init(skb); 6113 enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6114 } 6115 } 6116 } 6117 #endif 6118 __netif_receive_skb_list(head); 6119 rcu_read_unlock(); 6120 } 6121 6122 /** 6123 * netif_receive_skb - process receive buffer from network 6124 * @skb: buffer to process 6125 * 6126 * netif_receive_skb() is the main receive data processing function. 6127 * It always succeeds. The buffer may be dropped during processing 6128 * for congestion control or by the protocol layers. 6129 * 6130 * This function may only be called from softirq context and interrupts 6131 * should be enabled. 6132 * 6133 * Return values (usually ignored): 6134 * NET_RX_SUCCESS: no congestion 6135 * NET_RX_DROP: packet was dropped 6136 */ 6137 int netif_receive_skb(struct sk_buff *skb) 6138 { 6139 int ret; 6140 6141 trace_netif_receive_skb_entry(skb); 6142 6143 ret = netif_receive_skb_internal(skb); 6144 trace_netif_receive_skb_exit(ret); 6145 6146 return ret; 6147 } 6148 EXPORT_SYMBOL(netif_receive_skb); 6149 6150 /** 6151 * netif_receive_skb_list - process many receive buffers from network 6152 * @head: list of skbs to process. 6153 * 6154 * Since return value of netif_receive_skb() is normally ignored, and 6155 * wouldn't be meaningful for a list, this function returns void. 6156 * 6157 * This function may only be called from softirq context and interrupts 6158 * should be enabled. 6159 */ 6160 void netif_receive_skb_list(struct list_head *head) 6161 { 6162 struct sk_buff *skb; 6163 6164 if (list_empty(head)) 6165 return; 6166 if (trace_netif_receive_skb_list_entry_enabled()) { 6167 list_for_each_entry(skb, head, list) 6168 trace_netif_receive_skb_list_entry(skb); 6169 } 6170 netif_receive_skb_list_internal(head); 6171 trace_netif_receive_skb_list_exit(0); 6172 } 6173 EXPORT_SYMBOL(netif_receive_skb_list); 6174 6175 /* Network device is going away, flush any packets still pending */ 6176 static void flush_backlog(struct work_struct *work) 6177 { 6178 struct sk_buff *skb, *tmp; 6179 struct sk_buff_head list; 6180 struct softnet_data *sd; 6181 6182 __skb_queue_head_init(&list); 6183 local_bh_disable(); 6184 sd = this_cpu_ptr(&softnet_data); 6185 6186 backlog_lock_irq_disable(sd); 6187 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) { 6188 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6189 __skb_unlink(skb, &sd->input_pkt_queue); 6190 __skb_queue_tail(&list, skb); 6191 rps_input_queue_head_incr(sd); 6192 } 6193 } 6194 backlog_unlock_irq_enable(sd); 6195 6196 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6197 skb_queue_walk_safe(&sd->process_queue, skb, tmp) { 6198 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6199 __skb_unlink(skb, &sd->process_queue); 6200 __skb_queue_tail(&list, skb); 6201 rps_input_queue_head_incr(sd); 6202 } 6203 } 6204 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6205 local_bh_enable(); 6206 6207 __skb_queue_purge_reason(&list, SKB_DROP_REASON_DEV_READY); 6208 } 6209 6210 static bool flush_required(int cpu) 6211 { 6212 #if IS_ENABLED(CONFIG_RPS) 6213 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 6214 bool do_flush; 6215 6216 backlog_lock_irq_disable(sd); 6217 6218 /* as insertion into process_queue happens with the rps lock held, 6219 * process_queue access may race only with dequeue 6220 */ 6221 do_flush = !skb_queue_empty(&sd->input_pkt_queue) || 6222 !skb_queue_empty_lockless(&sd->process_queue); 6223 backlog_unlock_irq_enable(sd); 6224 6225 return do_flush; 6226 #endif 6227 /* without RPS we can't safely check input_pkt_queue: during a 6228 * concurrent remote skb_queue_splice() we can detect as empty both 6229 * input_pkt_queue and process_queue even if the latter could end-up 6230 * containing a lot of packets. 6231 */ 6232 return true; 6233 } 6234 6235 struct flush_backlogs { 6236 cpumask_t flush_cpus; 6237 struct work_struct w[]; 6238 }; 6239 6240 static struct flush_backlogs *flush_backlogs_alloc(void) 6241 { 6242 return kmalloc(struct_size_t(struct flush_backlogs, w, nr_cpu_ids), 6243 GFP_KERNEL); 6244 } 6245 6246 static struct flush_backlogs *flush_backlogs_fallback; 6247 static DEFINE_MUTEX(flush_backlogs_mutex); 6248 6249 static void flush_all_backlogs(void) 6250 { 6251 struct flush_backlogs *ptr = flush_backlogs_alloc(); 6252 unsigned int cpu; 6253 6254 if (!ptr) { 6255 mutex_lock(&flush_backlogs_mutex); 6256 ptr = flush_backlogs_fallback; 6257 } 6258 cpumask_clear(&ptr->flush_cpus); 6259 6260 cpus_read_lock(); 6261 6262 for_each_online_cpu(cpu) { 6263 if (flush_required(cpu)) { 6264 INIT_WORK(&ptr->w[cpu], flush_backlog); 6265 queue_work_on(cpu, system_highpri_wq, &ptr->w[cpu]); 6266 __cpumask_set_cpu(cpu, &ptr->flush_cpus); 6267 } 6268 } 6269 6270 /* we can have in flight packet[s] on the cpus we are not flushing, 6271 * synchronize_net() in unregister_netdevice_many() will take care of 6272 * them. 6273 */ 6274 for_each_cpu(cpu, &ptr->flush_cpus) 6275 flush_work(&ptr->w[cpu]); 6276 6277 cpus_read_unlock(); 6278 6279 if (ptr != flush_backlogs_fallback) 6280 kfree(ptr); 6281 else 6282 mutex_unlock(&flush_backlogs_mutex); 6283 } 6284 6285 static void net_rps_send_ipi(struct softnet_data *remsd) 6286 { 6287 #ifdef CONFIG_RPS 6288 while (remsd) { 6289 struct softnet_data *next = remsd->rps_ipi_next; 6290 6291 if (cpu_online(remsd->cpu)) 6292 smp_call_function_single_async(remsd->cpu, &remsd->csd); 6293 remsd = next; 6294 } 6295 #endif 6296 } 6297 6298 /* 6299 * net_rps_action_and_irq_enable sends any pending IPI's for rps. 6300 * Note: called with local irq disabled, but exits with local irq enabled. 6301 */ 6302 static void net_rps_action_and_irq_enable(struct softnet_data *sd) 6303 { 6304 #ifdef CONFIG_RPS 6305 struct softnet_data *remsd = sd->rps_ipi_list; 6306 6307 if (!use_backlog_threads() && remsd) { 6308 sd->rps_ipi_list = NULL; 6309 6310 local_irq_enable(); 6311 6312 /* Send pending IPI's to kick RPS processing on remote cpus. */ 6313 net_rps_send_ipi(remsd); 6314 } else 6315 #endif 6316 local_irq_enable(); 6317 } 6318 6319 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd) 6320 { 6321 #ifdef CONFIG_RPS 6322 return !use_backlog_threads() && sd->rps_ipi_list; 6323 #else 6324 return false; 6325 #endif 6326 } 6327 6328 static int process_backlog(struct napi_struct *napi, int quota) 6329 { 6330 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog); 6331 bool again = true; 6332 int work = 0; 6333 6334 /* Check if we have pending ipi, its better to send them now, 6335 * not waiting net_rx_action() end. 6336 */ 6337 if (sd_has_rps_ipi_waiting(sd)) { 6338 local_irq_disable(); 6339 net_rps_action_and_irq_enable(sd); 6340 } 6341 6342 napi->weight = READ_ONCE(net_hotdata.dev_rx_weight); 6343 while (again) { 6344 struct sk_buff *skb; 6345 6346 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6347 while ((skb = __skb_dequeue(&sd->process_queue))) { 6348 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6349 rcu_read_lock(); 6350 __netif_receive_skb(skb); 6351 rcu_read_unlock(); 6352 if (++work >= quota) { 6353 rps_input_queue_head_add(sd, work); 6354 return work; 6355 } 6356 6357 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6358 } 6359 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6360 6361 backlog_lock_irq_disable(sd); 6362 if (skb_queue_empty(&sd->input_pkt_queue)) { 6363 /* 6364 * Inline a custom version of __napi_complete(). 6365 * only current cpu owns and manipulates this napi, 6366 * and NAPI_STATE_SCHED is the only possible flag set 6367 * on backlog. 6368 * We can use a plain write instead of clear_bit(), 6369 * and we dont need an smp_mb() memory barrier. 6370 */ 6371 napi->state &= NAPIF_STATE_THREADED; 6372 again = false; 6373 } else { 6374 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6375 skb_queue_splice_tail_init(&sd->input_pkt_queue, 6376 &sd->process_queue); 6377 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6378 } 6379 backlog_unlock_irq_enable(sd); 6380 } 6381 6382 if (work) 6383 rps_input_queue_head_add(sd, work); 6384 return work; 6385 } 6386 6387 /** 6388 * __napi_schedule - schedule for receive 6389 * @n: entry to schedule 6390 * 6391 * The entry's receive function will be scheduled to run. 6392 * Consider using __napi_schedule_irqoff() if hard irqs are masked. 6393 */ 6394 void __napi_schedule(struct napi_struct *n) 6395 { 6396 unsigned long flags; 6397 6398 local_irq_save(flags); 6399 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6400 local_irq_restore(flags); 6401 } 6402 EXPORT_SYMBOL(__napi_schedule); 6403 6404 /** 6405 * napi_schedule_prep - check if napi can be scheduled 6406 * @n: napi context 6407 * 6408 * Test if NAPI routine is already running, and if not mark 6409 * it as running. This is used as a condition variable to 6410 * insure only one NAPI poll instance runs. We also make 6411 * sure there is no pending NAPI disable. 6412 */ 6413 bool napi_schedule_prep(struct napi_struct *n) 6414 { 6415 unsigned long new, val = READ_ONCE(n->state); 6416 6417 do { 6418 if (unlikely(val & NAPIF_STATE_DISABLE)) 6419 return false; 6420 new = val | NAPIF_STATE_SCHED; 6421 6422 /* Sets STATE_MISSED bit if STATE_SCHED was already set 6423 * This was suggested by Alexander Duyck, as compiler 6424 * emits better code than : 6425 * if (val & NAPIF_STATE_SCHED) 6426 * new |= NAPIF_STATE_MISSED; 6427 */ 6428 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED * 6429 NAPIF_STATE_MISSED; 6430 } while (!try_cmpxchg(&n->state, &val, new)); 6431 6432 return !(val & NAPIF_STATE_SCHED); 6433 } 6434 EXPORT_SYMBOL(napi_schedule_prep); 6435 6436 /** 6437 * __napi_schedule_irqoff - schedule for receive 6438 * @n: entry to schedule 6439 * 6440 * Variant of __napi_schedule() assuming hard irqs are masked. 6441 * 6442 * On PREEMPT_RT enabled kernels this maps to __napi_schedule() 6443 * because the interrupt disabled assumption might not be true 6444 * due to force-threaded interrupts and spinlock substitution. 6445 */ 6446 void __napi_schedule_irqoff(struct napi_struct *n) 6447 { 6448 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6449 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6450 else 6451 __napi_schedule(n); 6452 } 6453 EXPORT_SYMBOL(__napi_schedule_irqoff); 6454 6455 bool napi_complete_done(struct napi_struct *n, int work_done) 6456 { 6457 unsigned long flags, val, new, timeout = 0; 6458 bool ret = true; 6459 6460 /* 6461 * 1) Don't let napi dequeue from the cpu poll list 6462 * just in case its running on a different cpu. 6463 * 2) If we are busy polling, do nothing here, we have 6464 * the guarantee we will be called later. 6465 */ 6466 if (unlikely(n->state & (NAPIF_STATE_NPSVC | 6467 NAPIF_STATE_IN_BUSY_POLL))) 6468 return false; 6469 6470 if (work_done) { 6471 if (n->gro.bitmask) 6472 timeout = napi_get_gro_flush_timeout(n); 6473 n->defer_hard_irqs_count = napi_get_defer_hard_irqs(n); 6474 } 6475 if (n->defer_hard_irqs_count > 0) { 6476 n->defer_hard_irqs_count--; 6477 timeout = napi_get_gro_flush_timeout(n); 6478 if (timeout) 6479 ret = false; 6480 } 6481 6482 /* 6483 * When the NAPI instance uses a timeout and keeps postponing 6484 * it, we need to bound somehow the time packets are kept in 6485 * the GRO layer. 6486 */ 6487 gro_flush(&n->gro, !!timeout); 6488 gro_normal_list(&n->gro); 6489 6490 if (unlikely(!list_empty(&n->poll_list))) { 6491 /* If n->poll_list is not empty, we need to mask irqs */ 6492 local_irq_save(flags); 6493 list_del_init(&n->poll_list); 6494 local_irq_restore(flags); 6495 } 6496 WRITE_ONCE(n->list_owner, -1); 6497 6498 val = READ_ONCE(n->state); 6499 do { 6500 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED)); 6501 6502 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED | 6503 NAPIF_STATE_SCHED_THREADED | 6504 NAPIF_STATE_PREFER_BUSY_POLL); 6505 6506 /* If STATE_MISSED was set, leave STATE_SCHED set, 6507 * because we will call napi->poll() one more time. 6508 * This C code was suggested by Alexander Duyck to help gcc. 6509 */ 6510 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED * 6511 NAPIF_STATE_SCHED; 6512 } while (!try_cmpxchg(&n->state, &val, new)); 6513 6514 if (unlikely(val & NAPIF_STATE_MISSED)) { 6515 __napi_schedule(n); 6516 return false; 6517 } 6518 6519 if (timeout) 6520 hrtimer_start(&n->timer, ns_to_ktime(timeout), 6521 HRTIMER_MODE_REL_PINNED); 6522 return ret; 6523 } 6524 EXPORT_SYMBOL(napi_complete_done); 6525 6526 static void skb_defer_free_flush(struct softnet_data *sd) 6527 { 6528 struct sk_buff *skb, *next; 6529 6530 /* Paired with WRITE_ONCE() in skb_attempt_defer_free() */ 6531 if (!READ_ONCE(sd->defer_list)) 6532 return; 6533 6534 spin_lock(&sd->defer_lock); 6535 skb = sd->defer_list; 6536 sd->defer_list = NULL; 6537 sd->defer_count = 0; 6538 spin_unlock(&sd->defer_lock); 6539 6540 while (skb != NULL) { 6541 next = skb->next; 6542 napi_consume_skb(skb, 1); 6543 skb = next; 6544 } 6545 } 6546 6547 #if defined(CONFIG_NET_RX_BUSY_POLL) 6548 6549 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule) 6550 { 6551 if (!skip_schedule) { 6552 gro_normal_list(&napi->gro); 6553 __napi_schedule(napi); 6554 return; 6555 } 6556 6557 /* Flush too old packets. If HZ < 1000, flush all packets */ 6558 gro_flush(&napi->gro, HZ >= 1000); 6559 gro_normal_list(&napi->gro); 6560 6561 clear_bit(NAPI_STATE_SCHED, &napi->state); 6562 } 6563 6564 enum { 6565 NAPI_F_PREFER_BUSY_POLL = 1, 6566 NAPI_F_END_ON_RESCHED = 2, 6567 }; 6568 6569 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, 6570 unsigned flags, u16 budget) 6571 { 6572 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6573 bool skip_schedule = false; 6574 unsigned long timeout; 6575 int rc; 6576 6577 /* Busy polling means there is a high chance device driver hard irq 6578 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was 6579 * set in napi_schedule_prep(). 6580 * Since we are about to call napi->poll() once more, we can safely 6581 * clear NAPI_STATE_MISSED. 6582 * 6583 * Note: x86 could use a single "lock and ..." instruction 6584 * to perform these two clear_bit() 6585 */ 6586 clear_bit(NAPI_STATE_MISSED, &napi->state); 6587 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state); 6588 6589 local_bh_disable(); 6590 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6591 6592 if (flags & NAPI_F_PREFER_BUSY_POLL) { 6593 napi->defer_hard_irqs_count = napi_get_defer_hard_irqs(napi); 6594 timeout = napi_get_gro_flush_timeout(napi); 6595 if (napi->defer_hard_irqs_count && timeout) { 6596 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED); 6597 skip_schedule = true; 6598 } 6599 } 6600 6601 /* All we really want here is to re-enable device interrupts. 6602 * Ideally, a new ndo_busy_poll_stop() could avoid another round. 6603 */ 6604 rc = napi->poll(napi, budget); 6605 /* We can't gro_normal_list() here, because napi->poll() might have 6606 * rearmed the napi (napi_complete_done()) in which case it could 6607 * already be running on another CPU. 6608 */ 6609 trace_napi_poll(napi, rc, budget); 6610 netpoll_poll_unlock(have_poll_lock); 6611 if (rc == budget) 6612 __busy_poll_stop(napi, skip_schedule); 6613 bpf_net_ctx_clear(bpf_net_ctx); 6614 local_bh_enable(); 6615 } 6616 6617 static void __napi_busy_loop(unsigned int napi_id, 6618 bool (*loop_end)(void *, unsigned long), 6619 void *loop_end_arg, unsigned flags, u16 budget) 6620 { 6621 unsigned long start_time = loop_end ? busy_loop_current_time() : 0; 6622 int (*napi_poll)(struct napi_struct *napi, int budget); 6623 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6624 void *have_poll_lock = NULL; 6625 struct napi_struct *napi; 6626 6627 WARN_ON_ONCE(!rcu_read_lock_held()); 6628 6629 restart: 6630 napi_poll = NULL; 6631 6632 napi = napi_by_id(napi_id); 6633 if (!napi) 6634 return; 6635 6636 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6637 preempt_disable(); 6638 for (;;) { 6639 int work = 0; 6640 6641 local_bh_disable(); 6642 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6643 if (!napi_poll) { 6644 unsigned long val = READ_ONCE(napi->state); 6645 6646 /* If multiple threads are competing for this napi, 6647 * we avoid dirtying napi->state as much as we can. 6648 */ 6649 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED | 6650 NAPIF_STATE_IN_BUSY_POLL)) { 6651 if (flags & NAPI_F_PREFER_BUSY_POLL) 6652 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6653 goto count; 6654 } 6655 if (cmpxchg(&napi->state, val, 6656 val | NAPIF_STATE_IN_BUSY_POLL | 6657 NAPIF_STATE_SCHED) != val) { 6658 if (flags & NAPI_F_PREFER_BUSY_POLL) 6659 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6660 goto count; 6661 } 6662 have_poll_lock = netpoll_poll_lock(napi); 6663 napi_poll = napi->poll; 6664 } 6665 work = napi_poll(napi, budget); 6666 trace_napi_poll(napi, work, budget); 6667 gro_normal_list(&napi->gro); 6668 count: 6669 if (work > 0) 6670 __NET_ADD_STATS(dev_net(napi->dev), 6671 LINUX_MIB_BUSYPOLLRXPACKETS, work); 6672 skb_defer_free_flush(this_cpu_ptr(&softnet_data)); 6673 bpf_net_ctx_clear(bpf_net_ctx); 6674 local_bh_enable(); 6675 6676 if (!loop_end || loop_end(loop_end_arg, start_time)) 6677 break; 6678 6679 if (unlikely(need_resched())) { 6680 if (flags & NAPI_F_END_ON_RESCHED) 6681 break; 6682 if (napi_poll) 6683 busy_poll_stop(napi, have_poll_lock, flags, budget); 6684 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6685 preempt_enable(); 6686 rcu_read_unlock(); 6687 cond_resched(); 6688 rcu_read_lock(); 6689 if (loop_end(loop_end_arg, start_time)) 6690 return; 6691 goto restart; 6692 } 6693 cpu_relax(); 6694 } 6695 if (napi_poll) 6696 busy_poll_stop(napi, have_poll_lock, flags, budget); 6697 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6698 preempt_enable(); 6699 } 6700 6701 void napi_busy_loop_rcu(unsigned int napi_id, 6702 bool (*loop_end)(void *, unsigned long), 6703 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6704 { 6705 unsigned flags = NAPI_F_END_ON_RESCHED; 6706 6707 if (prefer_busy_poll) 6708 flags |= NAPI_F_PREFER_BUSY_POLL; 6709 6710 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6711 } 6712 6713 void napi_busy_loop(unsigned int napi_id, 6714 bool (*loop_end)(void *, unsigned long), 6715 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6716 { 6717 unsigned flags = prefer_busy_poll ? NAPI_F_PREFER_BUSY_POLL : 0; 6718 6719 rcu_read_lock(); 6720 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6721 rcu_read_unlock(); 6722 } 6723 EXPORT_SYMBOL(napi_busy_loop); 6724 6725 void napi_suspend_irqs(unsigned int napi_id) 6726 { 6727 struct napi_struct *napi; 6728 6729 rcu_read_lock(); 6730 napi = napi_by_id(napi_id); 6731 if (napi) { 6732 unsigned long timeout = napi_get_irq_suspend_timeout(napi); 6733 6734 if (timeout) 6735 hrtimer_start(&napi->timer, ns_to_ktime(timeout), 6736 HRTIMER_MODE_REL_PINNED); 6737 } 6738 rcu_read_unlock(); 6739 } 6740 6741 void napi_resume_irqs(unsigned int napi_id) 6742 { 6743 struct napi_struct *napi; 6744 6745 rcu_read_lock(); 6746 napi = napi_by_id(napi_id); 6747 if (napi) { 6748 /* If irq_suspend_timeout is set to 0 between the call to 6749 * napi_suspend_irqs and now, the original value still 6750 * determines the safety timeout as intended and napi_watchdog 6751 * will resume irq processing. 6752 */ 6753 if (napi_get_irq_suspend_timeout(napi)) { 6754 local_bh_disable(); 6755 napi_schedule(napi); 6756 local_bh_enable(); 6757 } 6758 } 6759 rcu_read_unlock(); 6760 } 6761 6762 #endif /* CONFIG_NET_RX_BUSY_POLL */ 6763 6764 static void __napi_hash_add_with_id(struct napi_struct *napi, 6765 unsigned int napi_id) 6766 { 6767 napi->gro.cached_napi_id = napi_id; 6768 6769 WRITE_ONCE(napi->napi_id, napi_id); 6770 hlist_add_head_rcu(&napi->napi_hash_node, 6771 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]); 6772 } 6773 6774 static void napi_hash_add_with_id(struct napi_struct *napi, 6775 unsigned int napi_id) 6776 { 6777 unsigned long flags; 6778 6779 spin_lock_irqsave(&napi_hash_lock, flags); 6780 WARN_ON_ONCE(napi_by_id(napi_id)); 6781 __napi_hash_add_with_id(napi, napi_id); 6782 spin_unlock_irqrestore(&napi_hash_lock, flags); 6783 } 6784 6785 static void napi_hash_add(struct napi_struct *napi) 6786 { 6787 unsigned long flags; 6788 6789 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state)) 6790 return; 6791 6792 spin_lock_irqsave(&napi_hash_lock, flags); 6793 6794 /* 0..NR_CPUS range is reserved for sender_cpu use */ 6795 do { 6796 if (unlikely(!napi_id_valid(++napi_gen_id))) 6797 napi_gen_id = MIN_NAPI_ID; 6798 } while (napi_by_id(napi_gen_id)); 6799 6800 __napi_hash_add_with_id(napi, napi_gen_id); 6801 6802 spin_unlock_irqrestore(&napi_hash_lock, flags); 6803 } 6804 6805 /* Warning : caller is responsible to make sure rcu grace period 6806 * is respected before freeing memory containing @napi 6807 */ 6808 static void napi_hash_del(struct napi_struct *napi) 6809 { 6810 unsigned long flags; 6811 6812 spin_lock_irqsave(&napi_hash_lock, flags); 6813 6814 hlist_del_init_rcu(&napi->napi_hash_node); 6815 6816 spin_unlock_irqrestore(&napi_hash_lock, flags); 6817 } 6818 6819 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer) 6820 { 6821 struct napi_struct *napi; 6822 6823 napi = container_of(timer, struct napi_struct, timer); 6824 6825 /* Note : we use a relaxed variant of napi_schedule_prep() not setting 6826 * NAPI_STATE_MISSED, since we do not react to a device IRQ. 6827 */ 6828 if (!napi_disable_pending(napi) && 6829 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) { 6830 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6831 __napi_schedule_irqoff(napi); 6832 } 6833 6834 return HRTIMER_NORESTART; 6835 } 6836 6837 int dev_set_threaded(struct net_device *dev, bool threaded) 6838 { 6839 struct napi_struct *napi; 6840 int err = 0; 6841 6842 netdev_assert_locked_or_invisible(dev); 6843 6844 if (dev->threaded == threaded) 6845 return 0; 6846 6847 if (threaded) { 6848 list_for_each_entry(napi, &dev->napi_list, dev_list) { 6849 if (!napi->thread) { 6850 err = napi_kthread_create(napi); 6851 if (err) { 6852 threaded = false; 6853 break; 6854 } 6855 } 6856 } 6857 } 6858 6859 WRITE_ONCE(dev->threaded, threaded); 6860 6861 /* Make sure kthread is created before THREADED bit 6862 * is set. 6863 */ 6864 smp_mb__before_atomic(); 6865 6866 /* Setting/unsetting threaded mode on a napi might not immediately 6867 * take effect, if the current napi instance is actively being 6868 * polled. In this case, the switch between threaded mode and 6869 * softirq mode will happen in the next round of napi_schedule(). 6870 * This should not cause hiccups/stalls to the live traffic. 6871 */ 6872 list_for_each_entry(napi, &dev->napi_list, dev_list) 6873 assign_bit(NAPI_STATE_THREADED, &napi->state, threaded); 6874 6875 return err; 6876 } 6877 EXPORT_SYMBOL(dev_set_threaded); 6878 6879 /** 6880 * netif_queue_set_napi - Associate queue with the napi 6881 * @dev: device to which NAPI and queue belong 6882 * @queue_index: Index of queue 6883 * @type: queue type as RX or TX 6884 * @napi: NAPI context, pass NULL to clear previously set NAPI 6885 * 6886 * Set queue with its corresponding napi context. This should be done after 6887 * registering the NAPI handler for the queue-vector and the queues have been 6888 * mapped to the corresponding interrupt vector. 6889 */ 6890 void netif_queue_set_napi(struct net_device *dev, unsigned int queue_index, 6891 enum netdev_queue_type type, struct napi_struct *napi) 6892 { 6893 struct netdev_rx_queue *rxq; 6894 struct netdev_queue *txq; 6895 6896 if (WARN_ON_ONCE(napi && !napi->dev)) 6897 return; 6898 netdev_ops_assert_locked_or_invisible(dev); 6899 6900 switch (type) { 6901 case NETDEV_QUEUE_TYPE_RX: 6902 rxq = __netif_get_rx_queue(dev, queue_index); 6903 rxq->napi = napi; 6904 return; 6905 case NETDEV_QUEUE_TYPE_TX: 6906 txq = netdev_get_tx_queue(dev, queue_index); 6907 txq->napi = napi; 6908 return; 6909 default: 6910 return; 6911 } 6912 } 6913 EXPORT_SYMBOL(netif_queue_set_napi); 6914 6915 static void 6916 netif_napi_irq_notify(struct irq_affinity_notify *notify, 6917 const cpumask_t *mask) 6918 { 6919 struct napi_struct *napi = 6920 container_of(notify, struct napi_struct, notify); 6921 #ifdef CONFIG_RFS_ACCEL 6922 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 6923 int err; 6924 #endif 6925 6926 if (napi->config && napi->dev->irq_affinity_auto) 6927 cpumask_copy(&napi->config->affinity_mask, mask); 6928 6929 #ifdef CONFIG_RFS_ACCEL 6930 if (napi->dev->rx_cpu_rmap_auto) { 6931 err = cpu_rmap_update(rmap, napi->napi_rmap_idx, mask); 6932 if (err) 6933 netdev_warn(napi->dev, "RMAP update failed (%d)\n", 6934 err); 6935 } 6936 #endif 6937 } 6938 6939 #ifdef CONFIG_RFS_ACCEL 6940 static void netif_napi_affinity_release(struct kref *ref) 6941 { 6942 struct napi_struct *napi = 6943 container_of(ref, struct napi_struct, notify.kref); 6944 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 6945 6946 netdev_assert_locked(napi->dev); 6947 WARN_ON(test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, 6948 &napi->state)); 6949 6950 if (!napi->dev->rx_cpu_rmap_auto) 6951 return; 6952 rmap->obj[napi->napi_rmap_idx] = NULL; 6953 napi->napi_rmap_idx = -1; 6954 cpu_rmap_put(rmap); 6955 } 6956 6957 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 6958 { 6959 if (dev->rx_cpu_rmap_auto) 6960 return 0; 6961 6962 dev->rx_cpu_rmap = alloc_irq_cpu_rmap(num_irqs); 6963 if (!dev->rx_cpu_rmap) 6964 return -ENOMEM; 6965 6966 dev->rx_cpu_rmap_auto = true; 6967 return 0; 6968 } 6969 EXPORT_SYMBOL(netif_enable_cpu_rmap); 6970 6971 static void netif_del_cpu_rmap(struct net_device *dev) 6972 { 6973 struct cpu_rmap *rmap = dev->rx_cpu_rmap; 6974 6975 if (!dev->rx_cpu_rmap_auto) 6976 return; 6977 6978 /* Free the rmap */ 6979 cpu_rmap_put(rmap); 6980 dev->rx_cpu_rmap = NULL; 6981 dev->rx_cpu_rmap_auto = false; 6982 } 6983 6984 #else 6985 static void netif_napi_affinity_release(struct kref *ref) 6986 { 6987 } 6988 6989 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 6990 { 6991 return 0; 6992 } 6993 EXPORT_SYMBOL(netif_enable_cpu_rmap); 6994 6995 static void netif_del_cpu_rmap(struct net_device *dev) 6996 { 6997 } 6998 #endif 6999 7000 void netif_set_affinity_auto(struct net_device *dev) 7001 { 7002 unsigned int i, maxqs, numa; 7003 7004 maxqs = max(dev->num_tx_queues, dev->num_rx_queues); 7005 numa = dev_to_node(&dev->dev); 7006 7007 for (i = 0; i < maxqs; i++) 7008 cpumask_set_cpu(cpumask_local_spread(i, numa), 7009 &dev->napi_config[i].affinity_mask); 7010 7011 dev->irq_affinity_auto = true; 7012 } 7013 EXPORT_SYMBOL(netif_set_affinity_auto); 7014 7015 void netif_napi_set_irq_locked(struct napi_struct *napi, int irq) 7016 { 7017 int rc; 7018 7019 netdev_assert_locked_or_invisible(napi->dev); 7020 7021 if (napi->irq == irq) 7022 return; 7023 7024 /* Remove existing resources */ 7025 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7026 irq_set_affinity_notifier(napi->irq, NULL); 7027 7028 napi->irq = irq; 7029 if (irq < 0 || 7030 (!napi->dev->rx_cpu_rmap_auto && !napi->dev->irq_affinity_auto)) 7031 return; 7032 7033 /* Abort for buggy drivers */ 7034 if (napi->dev->irq_affinity_auto && WARN_ON_ONCE(!napi->config)) 7035 return; 7036 7037 #ifdef CONFIG_RFS_ACCEL 7038 if (napi->dev->rx_cpu_rmap_auto) { 7039 rc = cpu_rmap_add(napi->dev->rx_cpu_rmap, napi); 7040 if (rc < 0) 7041 return; 7042 7043 cpu_rmap_get(napi->dev->rx_cpu_rmap); 7044 napi->napi_rmap_idx = rc; 7045 } 7046 #endif 7047 7048 /* Use core IRQ notifier */ 7049 napi->notify.notify = netif_napi_irq_notify; 7050 napi->notify.release = netif_napi_affinity_release; 7051 rc = irq_set_affinity_notifier(irq, &napi->notify); 7052 if (rc) { 7053 netdev_warn(napi->dev, "Unable to set IRQ notifier (%d)\n", 7054 rc); 7055 goto put_rmap; 7056 } 7057 7058 set_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state); 7059 return; 7060 7061 put_rmap: 7062 #ifdef CONFIG_RFS_ACCEL 7063 if (napi->dev->rx_cpu_rmap_auto) { 7064 napi->dev->rx_cpu_rmap->obj[napi->napi_rmap_idx] = NULL; 7065 cpu_rmap_put(napi->dev->rx_cpu_rmap); 7066 napi->napi_rmap_idx = -1; 7067 } 7068 #endif 7069 napi->notify.notify = NULL; 7070 napi->notify.release = NULL; 7071 } 7072 EXPORT_SYMBOL(netif_napi_set_irq_locked); 7073 7074 static void napi_restore_config(struct napi_struct *n) 7075 { 7076 n->defer_hard_irqs = n->config->defer_hard_irqs; 7077 n->gro_flush_timeout = n->config->gro_flush_timeout; 7078 n->irq_suspend_timeout = n->config->irq_suspend_timeout; 7079 7080 if (n->dev->irq_affinity_auto && 7081 test_bit(NAPI_STATE_HAS_NOTIFIER, &n->state)) 7082 irq_set_affinity(n->irq, &n->config->affinity_mask); 7083 7084 /* a NAPI ID might be stored in the config, if so use it. if not, use 7085 * napi_hash_add to generate one for us. 7086 */ 7087 if (n->config->napi_id) { 7088 napi_hash_add_with_id(n, n->config->napi_id); 7089 } else { 7090 napi_hash_add(n); 7091 n->config->napi_id = n->napi_id; 7092 } 7093 } 7094 7095 static void napi_save_config(struct napi_struct *n) 7096 { 7097 n->config->defer_hard_irqs = n->defer_hard_irqs; 7098 n->config->gro_flush_timeout = n->gro_flush_timeout; 7099 n->config->irq_suspend_timeout = n->irq_suspend_timeout; 7100 napi_hash_del(n); 7101 } 7102 7103 /* Netlink wants the NAPI list to be sorted by ID, if adding a NAPI which will 7104 * inherit an existing ID try to insert it at the right position. 7105 */ 7106 static void 7107 netif_napi_dev_list_add(struct net_device *dev, struct napi_struct *napi) 7108 { 7109 unsigned int new_id, pos_id; 7110 struct list_head *higher; 7111 struct napi_struct *pos; 7112 7113 new_id = UINT_MAX; 7114 if (napi->config && napi->config->napi_id) 7115 new_id = napi->config->napi_id; 7116 7117 higher = &dev->napi_list; 7118 list_for_each_entry(pos, &dev->napi_list, dev_list) { 7119 if (napi_id_valid(pos->napi_id)) 7120 pos_id = pos->napi_id; 7121 else if (pos->config) 7122 pos_id = pos->config->napi_id; 7123 else 7124 pos_id = UINT_MAX; 7125 7126 if (pos_id <= new_id) 7127 break; 7128 higher = &pos->dev_list; 7129 } 7130 list_add_rcu(&napi->dev_list, higher); /* adds after higher */ 7131 } 7132 7133 /* Double check that napi_get_frags() allocates skbs with 7134 * skb->head being backed by slab, not a page fragment. 7135 * This is to make sure bug fixed in 3226b158e67c 7136 * ("net: avoid 32 x truesize under-estimation for tiny skbs") 7137 * does not accidentally come back. 7138 */ 7139 static void napi_get_frags_check(struct napi_struct *napi) 7140 { 7141 struct sk_buff *skb; 7142 7143 local_bh_disable(); 7144 skb = napi_get_frags(napi); 7145 WARN_ON_ONCE(skb && skb->head_frag); 7146 napi_free_frags(napi); 7147 local_bh_enable(); 7148 } 7149 7150 void netif_napi_add_weight_locked(struct net_device *dev, 7151 struct napi_struct *napi, 7152 int (*poll)(struct napi_struct *, int), 7153 int weight) 7154 { 7155 netdev_assert_locked(dev); 7156 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state))) 7157 return; 7158 7159 INIT_LIST_HEAD(&napi->poll_list); 7160 INIT_HLIST_NODE(&napi->napi_hash_node); 7161 hrtimer_setup(&napi->timer, napi_watchdog, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); 7162 gro_init(&napi->gro); 7163 napi->skb = NULL; 7164 napi->poll = poll; 7165 if (weight > NAPI_POLL_WEIGHT) 7166 netdev_err_once(dev, "%s() called with weight %d\n", __func__, 7167 weight); 7168 napi->weight = weight; 7169 napi->dev = dev; 7170 #ifdef CONFIG_NETPOLL 7171 napi->poll_owner = -1; 7172 #endif 7173 napi->list_owner = -1; 7174 set_bit(NAPI_STATE_SCHED, &napi->state); 7175 set_bit(NAPI_STATE_NPSVC, &napi->state); 7176 netif_napi_dev_list_add(dev, napi); 7177 7178 /* default settings from sysfs are applied to all NAPIs. any per-NAPI 7179 * configuration will be loaded in napi_enable 7180 */ 7181 napi_set_defer_hard_irqs(napi, READ_ONCE(dev->napi_defer_hard_irqs)); 7182 napi_set_gro_flush_timeout(napi, READ_ONCE(dev->gro_flush_timeout)); 7183 7184 napi_get_frags_check(napi); 7185 /* Create kthread for this napi if dev->threaded is set. 7186 * Clear dev->threaded if kthread creation failed so that 7187 * threaded mode will not be enabled in napi_enable(). 7188 */ 7189 if (dev->threaded && napi_kthread_create(napi)) 7190 dev->threaded = false; 7191 netif_napi_set_irq_locked(napi, -1); 7192 } 7193 EXPORT_SYMBOL(netif_napi_add_weight_locked); 7194 7195 void napi_disable_locked(struct napi_struct *n) 7196 { 7197 unsigned long val, new; 7198 7199 might_sleep(); 7200 netdev_assert_locked(n->dev); 7201 7202 set_bit(NAPI_STATE_DISABLE, &n->state); 7203 7204 val = READ_ONCE(n->state); 7205 do { 7206 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) { 7207 usleep_range(20, 200); 7208 val = READ_ONCE(n->state); 7209 } 7210 7211 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC; 7212 new &= ~(NAPIF_STATE_THREADED | NAPIF_STATE_PREFER_BUSY_POLL); 7213 } while (!try_cmpxchg(&n->state, &val, new)); 7214 7215 hrtimer_cancel(&n->timer); 7216 7217 if (n->config) 7218 napi_save_config(n); 7219 else 7220 napi_hash_del(n); 7221 7222 clear_bit(NAPI_STATE_DISABLE, &n->state); 7223 } 7224 EXPORT_SYMBOL(napi_disable_locked); 7225 7226 /** 7227 * napi_disable() - prevent NAPI from scheduling 7228 * @n: NAPI context 7229 * 7230 * Stop NAPI from being scheduled on this context. 7231 * Waits till any outstanding processing completes. 7232 * Takes netdev_lock() for associated net_device. 7233 */ 7234 void napi_disable(struct napi_struct *n) 7235 { 7236 netdev_lock(n->dev); 7237 napi_disable_locked(n); 7238 netdev_unlock(n->dev); 7239 } 7240 EXPORT_SYMBOL(napi_disable); 7241 7242 void napi_enable_locked(struct napi_struct *n) 7243 { 7244 unsigned long new, val = READ_ONCE(n->state); 7245 7246 if (n->config) 7247 napi_restore_config(n); 7248 else 7249 napi_hash_add(n); 7250 7251 do { 7252 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val)); 7253 7254 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC); 7255 if (n->dev->threaded && n->thread) 7256 new |= NAPIF_STATE_THREADED; 7257 } while (!try_cmpxchg(&n->state, &val, new)); 7258 } 7259 EXPORT_SYMBOL(napi_enable_locked); 7260 7261 /** 7262 * napi_enable() - enable NAPI scheduling 7263 * @n: NAPI context 7264 * 7265 * Enable scheduling of a NAPI instance. 7266 * Must be paired with napi_disable(). 7267 * Takes netdev_lock() for associated net_device. 7268 */ 7269 void napi_enable(struct napi_struct *n) 7270 { 7271 netdev_lock(n->dev); 7272 napi_enable_locked(n); 7273 netdev_unlock(n->dev); 7274 } 7275 EXPORT_SYMBOL(napi_enable); 7276 7277 /* Must be called in process context */ 7278 void __netif_napi_del_locked(struct napi_struct *napi) 7279 { 7280 netdev_assert_locked(napi->dev); 7281 7282 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state)) 7283 return; 7284 7285 /* Make sure NAPI is disabled (or was never enabled). */ 7286 WARN_ON(!test_bit(NAPI_STATE_SCHED, &napi->state)); 7287 7288 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7289 irq_set_affinity_notifier(napi->irq, NULL); 7290 7291 if (napi->config) { 7292 napi->index = -1; 7293 napi->config = NULL; 7294 } 7295 7296 list_del_rcu(&napi->dev_list); 7297 napi_free_frags(napi); 7298 7299 gro_cleanup(&napi->gro); 7300 7301 if (napi->thread) { 7302 kthread_stop(napi->thread); 7303 napi->thread = NULL; 7304 } 7305 } 7306 EXPORT_SYMBOL(__netif_napi_del_locked); 7307 7308 static int __napi_poll(struct napi_struct *n, bool *repoll) 7309 { 7310 int work, weight; 7311 7312 weight = n->weight; 7313 7314 /* This NAPI_STATE_SCHED test is for avoiding a race 7315 * with netpoll's poll_napi(). Only the entity which 7316 * obtains the lock and sees NAPI_STATE_SCHED set will 7317 * actually make the ->poll() call. Therefore we avoid 7318 * accidentally calling ->poll() when NAPI is not scheduled. 7319 */ 7320 work = 0; 7321 if (napi_is_scheduled(n)) { 7322 work = n->poll(n, weight); 7323 trace_napi_poll(n, work, weight); 7324 7325 xdp_do_check_flushed(n); 7326 } 7327 7328 if (unlikely(work > weight)) 7329 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n", 7330 n->poll, work, weight); 7331 7332 if (likely(work < weight)) 7333 return work; 7334 7335 /* Drivers must not modify the NAPI state if they 7336 * consume the entire weight. In such cases this code 7337 * still "owns" the NAPI instance and therefore can 7338 * move the instance around on the list at-will. 7339 */ 7340 if (unlikely(napi_disable_pending(n))) { 7341 napi_complete(n); 7342 return work; 7343 } 7344 7345 /* The NAPI context has more processing work, but busy-polling 7346 * is preferred. Exit early. 7347 */ 7348 if (napi_prefer_busy_poll(n)) { 7349 if (napi_complete_done(n, work)) { 7350 /* If timeout is not set, we need to make sure 7351 * that the NAPI is re-scheduled. 7352 */ 7353 napi_schedule(n); 7354 } 7355 return work; 7356 } 7357 7358 /* Flush too old packets. If HZ < 1000, flush all packets */ 7359 gro_flush(&n->gro, HZ >= 1000); 7360 gro_normal_list(&n->gro); 7361 7362 /* Some drivers may have called napi_schedule 7363 * prior to exhausting their budget. 7364 */ 7365 if (unlikely(!list_empty(&n->poll_list))) { 7366 pr_warn_once("%s: Budget exhausted after napi rescheduled\n", 7367 n->dev ? n->dev->name : "backlog"); 7368 return work; 7369 } 7370 7371 *repoll = true; 7372 7373 return work; 7374 } 7375 7376 static int napi_poll(struct napi_struct *n, struct list_head *repoll) 7377 { 7378 bool do_repoll = false; 7379 void *have; 7380 int work; 7381 7382 list_del_init(&n->poll_list); 7383 7384 have = netpoll_poll_lock(n); 7385 7386 work = __napi_poll(n, &do_repoll); 7387 7388 if (do_repoll) 7389 list_add_tail(&n->poll_list, repoll); 7390 7391 netpoll_poll_unlock(have); 7392 7393 return work; 7394 } 7395 7396 static int napi_thread_wait(struct napi_struct *napi) 7397 { 7398 set_current_state(TASK_INTERRUPTIBLE); 7399 7400 while (!kthread_should_stop()) { 7401 /* Testing SCHED_THREADED bit here to make sure the current 7402 * kthread owns this napi and could poll on this napi. 7403 * Testing SCHED bit is not enough because SCHED bit might be 7404 * set by some other busy poll thread or by napi_disable(). 7405 */ 7406 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) { 7407 WARN_ON(!list_empty(&napi->poll_list)); 7408 __set_current_state(TASK_RUNNING); 7409 return 0; 7410 } 7411 7412 schedule(); 7413 set_current_state(TASK_INTERRUPTIBLE); 7414 } 7415 __set_current_state(TASK_RUNNING); 7416 7417 return -1; 7418 } 7419 7420 static void napi_threaded_poll_loop(struct napi_struct *napi) 7421 { 7422 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7423 struct softnet_data *sd; 7424 unsigned long last_qs = jiffies; 7425 7426 for (;;) { 7427 bool repoll = false; 7428 void *have; 7429 7430 local_bh_disable(); 7431 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7432 7433 sd = this_cpu_ptr(&softnet_data); 7434 sd->in_napi_threaded_poll = true; 7435 7436 have = netpoll_poll_lock(napi); 7437 __napi_poll(napi, &repoll); 7438 netpoll_poll_unlock(have); 7439 7440 sd->in_napi_threaded_poll = false; 7441 barrier(); 7442 7443 if (sd_has_rps_ipi_waiting(sd)) { 7444 local_irq_disable(); 7445 net_rps_action_and_irq_enable(sd); 7446 } 7447 skb_defer_free_flush(sd); 7448 bpf_net_ctx_clear(bpf_net_ctx); 7449 local_bh_enable(); 7450 7451 if (!repoll) 7452 break; 7453 7454 rcu_softirq_qs_periodic(last_qs); 7455 cond_resched(); 7456 } 7457 } 7458 7459 static int napi_threaded_poll(void *data) 7460 { 7461 struct napi_struct *napi = data; 7462 7463 while (!napi_thread_wait(napi)) 7464 napi_threaded_poll_loop(napi); 7465 7466 return 0; 7467 } 7468 7469 static __latent_entropy void net_rx_action(void) 7470 { 7471 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 7472 unsigned long time_limit = jiffies + 7473 usecs_to_jiffies(READ_ONCE(net_hotdata.netdev_budget_usecs)); 7474 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7475 int budget = READ_ONCE(net_hotdata.netdev_budget); 7476 LIST_HEAD(list); 7477 LIST_HEAD(repoll); 7478 7479 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7480 start: 7481 sd->in_net_rx_action = true; 7482 local_irq_disable(); 7483 list_splice_init(&sd->poll_list, &list); 7484 local_irq_enable(); 7485 7486 for (;;) { 7487 struct napi_struct *n; 7488 7489 skb_defer_free_flush(sd); 7490 7491 if (list_empty(&list)) { 7492 if (list_empty(&repoll)) { 7493 sd->in_net_rx_action = false; 7494 barrier(); 7495 /* We need to check if ____napi_schedule() 7496 * had refilled poll_list while 7497 * sd->in_net_rx_action was true. 7498 */ 7499 if (!list_empty(&sd->poll_list)) 7500 goto start; 7501 if (!sd_has_rps_ipi_waiting(sd)) 7502 goto end; 7503 } 7504 break; 7505 } 7506 7507 n = list_first_entry(&list, struct napi_struct, poll_list); 7508 budget -= napi_poll(n, &repoll); 7509 7510 /* If softirq window is exhausted then punt. 7511 * Allow this to run for 2 jiffies since which will allow 7512 * an average latency of 1.5/HZ. 7513 */ 7514 if (unlikely(budget <= 0 || 7515 time_after_eq(jiffies, time_limit))) { 7516 sd->time_squeeze++; 7517 break; 7518 } 7519 } 7520 7521 local_irq_disable(); 7522 7523 list_splice_tail_init(&sd->poll_list, &list); 7524 list_splice_tail(&repoll, &list); 7525 list_splice(&list, &sd->poll_list); 7526 if (!list_empty(&sd->poll_list)) 7527 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 7528 else 7529 sd->in_net_rx_action = false; 7530 7531 net_rps_action_and_irq_enable(sd); 7532 end: 7533 bpf_net_ctx_clear(bpf_net_ctx); 7534 } 7535 7536 struct netdev_adjacent { 7537 struct net_device *dev; 7538 netdevice_tracker dev_tracker; 7539 7540 /* upper master flag, there can only be one master device per list */ 7541 bool master; 7542 7543 /* lookup ignore flag */ 7544 bool ignore; 7545 7546 /* counter for the number of times this device was added to us */ 7547 u16 ref_nr; 7548 7549 /* private field for the users */ 7550 void *private; 7551 7552 struct list_head list; 7553 struct rcu_head rcu; 7554 }; 7555 7556 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev, 7557 struct list_head *adj_list) 7558 { 7559 struct netdev_adjacent *adj; 7560 7561 list_for_each_entry(adj, adj_list, list) { 7562 if (adj->dev == adj_dev) 7563 return adj; 7564 } 7565 return NULL; 7566 } 7567 7568 static int ____netdev_has_upper_dev(struct net_device *upper_dev, 7569 struct netdev_nested_priv *priv) 7570 { 7571 struct net_device *dev = (struct net_device *)priv->data; 7572 7573 return upper_dev == dev; 7574 } 7575 7576 /** 7577 * netdev_has_upper_dev - Check if device is linked to an upper device 7578 * @dev: device 7579 * @upper_dev: upper device to check 7580 * 7581 * Find out if a device is linked to specified upper device and return true 7582 * in case it is. Note that this checks only immediate upper device, 7583 * not through a complete stack of devices. The caller must hold the RTNL lock. 7584 */ 7585 bool netdev_has_upper_dev(struct net_device *dev, 7586 struct net_device *upper_dev) 7587 { 7588 struct netdev_nested_priv priv = { 7589 .data = (void *)upper_dev, 7590 }; 7591 7592 ASSERT_RTNL(); 7593 7594 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7595 &priv); 7596 } 7597 EXPORT_SYMBOL(netdev_has_upper_dev); 7598 7599 /** 7600 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device 7601 * @dev: device 7602 * @upper_dev: upper device to check 7603 * 7604 * Find out if a device is linked to specified upper device and return true 7605 * in case it is. Note that this checks the entire upper device chain. 7606 * The caller must hold rcu lock. 7607 */ 7608 7609 bool netdev_has_upper_dev_all_rcu(struct net_device *dev, 7610 struct net_device *upper_dev) 7611 { 7612 struct netdev_nested_priv priv = { 7613 .data = (void *)upper_dev, 7614 }; 7615 7616 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7617 &priv); 7618 } 7619 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu); 7620 7621 /** 7622 * netdev_has_any_upper_dev - Check if device is linked to some device 7623 * @dev: device 7624 * 7625 * Find out if a device is linked to an upper device and return true in case 7626 * it is. The caller must hold the RTNL lock. 7627 */ 7628 bool netdev_has_any_upper_dev(struct net_device *dev) 7629 { 7630 ASSERT_RTNL(); 7631 7632 return !list_empty(&dev->adj_list.upper); 7633 } 7634 EXPORT_SYMBOL(netdev_has_any_upper_dev); 7635 7636 /** 7637 * netdev_master_upper_dev_get - Get master upper device 7638 * @dev: device 7639 * 7640 * Find a master upper device and return pointer to it or NULL in case 7641 * it's not there. The caller must hold the RTNL lock. 7642 */ 7643 struct net_device *netdev_master_upper_dev_get(struct net_device *dev) 7644 { 7645 struct netdev_adjacent *upper; 7646 7647 ASSERT_RTNL(); 7648 7649 if (list_empty(&dev->adj_list.upper)) 7650 return NULL; 7651 7652 upper = list_first_entry(&dev->adj_list.upper, 7653 struct netdev_adjacent, list); 7654 if (likely(upper->master)) 7655 return upper->dev; 7656 return NULL; 7657 } 7658 EXPORT_SYMBOL(netdev_master_upper_dev_get); 7659 7660 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev) 7661 { 7662 struct netdev_adjacent *upper; 7663 7664 ASSERT_RTNL(); 7665 7666 if (list_empty(&dev->adj_list.upper)) 7667 return NULL; 7668 7669 upper = list_first_entry(&dev->adj_list.upper, 7670 struct netdev_adjacent, list); 7671 if (likely(upper->master) && !upper->ignore) 7672 return upper->dev; 7673 return NULL; 7674 } 7675 7676 /** 7677 * netdev_has_any_lower_dev - Check if device is linked to some device 7678 * @dev: device 7679 * 7680 * Find out if a device is linked to a lower device and return true in case 7681 * it is. The caller must hold the RTNL lock. 7682 */ 7683 static bool netdev_has_any_lower_dev(struct net_device *dev) 7684 { 7685 ASSERT_RTNL(); 7686 7687 return !list_empty(&dev->adj_list.lower); 7688 } 7689 7690 void *netdev_adjacent_get_private(struct list_head *adj_list) 7691 { 7692 struct netdev_adjacent *adj; 7693 7694 adj = list_entry(adj_list, struct netdev_adjacent, list); 7695 7696 return adj->private; 7697 } 7698 EXPORT_SYMBOL(netdev_adjacent_get_private); 7699 7700 /** 7701 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list 7702 * @dev: device 7703 * @iter: list_head ** of the current position 7704 * 7705 * Gets the next device from the dev's upper list, starting from iter 7706 * position. The caller must hold RCU read lock. 7707 */ 7708 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev, 7709 struct list_head **iter) 7710 { 7711 struct netdev_adjacent *upper; 7712 7713 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7714 7715 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7716 7717 if (&upper->list == &dev->adj_list.upper) 7718 return NULL; 7719 7720 *iter = &upper->list; 7721 7722 return upper->dev; 7723 } 7724 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu); 7725 7726 static struct net_device *__netdev_next_upper_dev(struct net_device *dev, 7727 struct list_head **iter, 7728 bool *ignore) 7729 { 7730 struct netdev_adjacent *upper; 7731 7732 upper = list_entry((*iter)->next, struct netdev_adjacent, list); 7733 7734 if (&upper->list == &dev->adj_list.upper) 7735 return NULL; 7736 7737 *iter = &upper->list; 7738 *ignore = upper->ignore; 7739 7740 return upper->dev; 7741 } 7742 7743 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev, 7744 struct list_head **iter) 7745 { 7746 struct netdev_adjacent *upper; 7747 7748 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7749 7750 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7751 7752 if (&upper->list == &dev->adj_list.upper) 7753 return NULL; 7754 7755 *iter = &upper->list; 7756 7757 return upper->dev; 7758 } 7759 7760 static int __netdev_walk_all_upper_dev(struct net_device *dev, 7761 int (*fn)(struct net_device *dev, 7762 struct netdev_nested_priv *priv), 7763 struct netdev_nested_priv *priv) 7764 { 7765 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7766 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7767 int ret, cur = 0; 7768 bool ignore; 7769 7770 now = dev; 7771 iter = &dev->adj_list.upper; 7772 7773 while (1) { 7774 if (now != dev) { 7775 ret = fn(now, priv); 7776 if (ret) 7777 return ret; 7778 } 7779 7780 next = NULL; 7781 while (1) { 7782 udev = __netdev_next_upper_dev(now, &iter, &ignore); 7783 if (!udev) 7784 break; 7785 if (ignore) 7786 continue; 7787 7788 next = udev; 7789 niter = &udev->adj_list.upper; 7790 dev_stack[cur] = now; 7791 iter_stack[cur++] = iter; 7792 break; 7793 } 7794 7795 if (!next) { 7796 if (!cur) 7797 return 0; 7798 next = dev_stack[--cur]; 7799 niter = iter_stack[cur]; 7800 } 7801 7802 now = next; 7803 iter = niter; 7804 } 7805 7806 return 0; 7807 } 7808 7809 int netdev_walk_all_upper_dev_rcu(struct net_device *dev, 7810 int (*fn)(struct net_device *dev, 7811 struct netdev_nested_priv *priv), 7812 struct netdev_nested_priv *priv) 7813 { 7814 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7815 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7816 int ret, cur = 0; 7817 7818 now = dev; 7819 iter = &dev->adj_list.upper; 7820 7821 while (1) { 7822 if (now != dev) { 7823 ret = fn(now, priv); 7824 if (ret) 7825 return ret; 7826 } 7827 7828 next = NULL; 7829 while (1) { 7830 udev = netdev_next_upper_dev_rcu(now, &iter); 7831 if (!udev) 7832 break; 7833 7834 next = udev; 7835 niter = &udev->adj_list.upper; 7836 dev_stack[cur] = now; 7837 iter_stack[cur++] = iter; 7838 break; 7839 } 7840 7841 if (!next) { 7842 if (!cur) 7843 return 0; 7844 next = dev_stack[--cur]; 7845 niter = iter_stack[cur]; 7846 } 7847 7848 now = next; 7849 iter = niter; 7850 } 7851 7852 return 0; 7853 } 7854 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu); 7855 7856 static bool __netdev_has_upper_dev(struct net_device *dev, 7857 struct net_device *upper_dev) 7858 { 7859 struct netdev_nested_priv priv = { 7860 .flags = 0, 7861 .data = (void *)upper_dev, 7862 }; 7863 7864 ASSERT_RTNL(); 7865 7866 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev, 7867 &priv); 7868 } 7869 7870 /** 7871 * netdev_lower_get_next_private - Get the next ->private from the 7872 * lower neighbour list 7873 * @dev: device 7874 * @iter: list_head ** of the current position 7875 * 7876 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7877 * list, starting from iter position. The caller must hold either hold the 7878 * RTNL lock or its own locking that guarantees that the neighbour lower 7879 * list will remain unchanged. 7880 */ 7881 void *netdev_lower_get_next_private(struct net_device *dev, 7882 struct list_head **iter) 7883 { 7884 struct netdev_adjacent *lower; 7885 7886 lower = list_entry(*iter, struct netdev_adjacent, list); 7887 7888 if (&lower->list == &dev->adj_list.lower) 7889 return NULL; 7890 7891 *iter = lower->list.next; 7892 7893 return lower->private; 7894 } 7895 EXPORT_SYMBOL(netdev_lower_get_next_private); 7896 7897 /** 7898 * netdev_lower_get_next_private_rcu - Get the next ->private from the 7899 * lower neighbour list, RCU 7900 * variant 7901 * @dev: device 7902 * @iter: list_head ** of the current position 7903 * 7904 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7905 * list, starting from iter position. The caller must hold RCU read lock. 7906 */ 7907 void *netdev_lower_get_next_private_rcu(struct net_device *dev, 7908 struct list_head **iter) 7909 { 7910 struct netdev_adjacent *lower; 7911 7912 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 7913 7914 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7915 7916 if (&lower->list == &dev->adj_list.lower) 7917 return NULL; 7918 7919 *iter = &lower->list; 7920 7921 return lower->private; 7922 } 7923 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu); 7924 7925 /** 7926 * netdev_lower_get_next - Get the next device from the lower neighbour 7927 * list 7928 * @dev: device 7929 * @iter: list_head ** of the current position 7930 * 7931 * Gets the next netdev_adjacent from the dev's lower neighbour 7932 * list, starting from iter position. The caller must hold RTNL lock or 7933 * its own locking that guarantees that the neighbour lower 7934 * list will remain unchanged. 7935 */ 7936 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter) 7937 { 7938 struct netdev_adjacent *lower; 7939 7940 lower = list_entry(*iter, struct netdev_adjacent, list); 7941 7942 if (&lower->list == &dev->adj_list.lower) 7943 return NULL; 7944 7945 *iter = lower->list.next; 7946 7947 return lower->dev; 7948 } 7949 EXPORT_SYMBOL(netdev_lower_get_next); 7950 7951 static struct net_device *netdev_next_lower_dev(struct net_device *dev, 7952 struct list_head **iter) 7953 { 7954 struct netdev_adjacent *lower; 7955 7956 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7957 7958 if (&lower->list == &dev->adj_list.lower) 7959 return NULL; 7960 7961 *iter = &lower->list; 7962 7963 return lower->dev; 7964 } 7965 7966 static struct net_device *__netdev_next_lower_dev(struct net_device *dev, 7967 struct list_head **iter, 7968 bool *ignore) 7969 { 7970 struct netdev_adjacent *lower; 7971 7972 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7973 7974 if (&lower->list == &dev->adj_list.lower) 7975 return NULL; 7976 7977 *iter = &lower->list; 7978 *ignore = lower->ignore; 7979 7980 return lower->dev; 7981 } 7982 7983 int netdev_walk_all_lower_dev(struct net_device *dev, 7984 int (*fn)(struct net_device *dev, 7985 struct netdev_nested_priv *priv), 7986 struct netdev_nested_priv *priv) 7987 { 7988 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7989 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7990 int ret, cur = 0; 7991 7992 now = dev; 7993 iter = &dev->adj_list.lower; 7994 7995 while (1) { 7996 if (now != dev) { 7997 ret = fn(now, priv); 7998 if (ret) 7999 return ret; 8000 } 8001 8002 next = NULL; 8003 while (1) { 8004 ldev = netdev_next_lower_dev(now, &iter); 8005 if (!ldev) 8006 break; 8007 8008 next = ldev; 8009 niter = &ldev->adj_list.lower; 8010 dev_stack[cur] = now; 8011 iter_stack[cur++] = iter; 8012 break; 8013 } 8014 8015 if (!next) { 8016 if (!cur) 8017 return 0; 8018 next = dev_stack[--cur]; 8019 niter = iter_stack[cur]; 8020 } 8021 8022 now = next; 8023 iter = niter; 8024 } 8025 8026 return 0; 8027 } 8028 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev); 8029 8030 static int __netdev_walk_all_lower_dev(struct net_device *dev, 8031 int (*fn)(struct net_device *dev, 8032 struct netdev_nested_priv *priv), 8033 struct netdev_nested_priv *priv) 8034 { 8035 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8036 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8037 int ret, cur = 0; 8038 bool ignore; 8039 8040 now = dev; 8041 iter = &dev->adj_list.lower; 8042 8043 while (1) { 8044 if (now != dev) { 8045 ret = fn(now, priv); 8046 if (ret) 8047 return ret; 8048 } 8049 8050 next = NULL; 8051 while (1) { 8052 ldev = __netdev_next_lower_dev(now, &iter, &ignore); 8053 if (!ldev) 8054 break; 8055 if (ignore) 8056 continue; 8057 8058 next = ldev; 8059 niter = &ldev->adj_list.lower; 8060 dev_stack[cur] = now; 8061 iter_stack[cur++] = iter; 8062 break; 8063 } 8064 8065 if (!next) { 8066 if (!cur) 8067 return 0; 8068 next = dev_stack[--cur]; 8069 niter = iter_stack[cur]; 8070 } 8071 8072 now = next; 8073 iter = niter; 8074 } 8075 8076 return 0; 8077 } 8078 8079 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev, 8080 struct list_head **iter) 8081 { 8082 struct netdev_adjacent *lower; 8083 8084 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8085 if (&lower->list == &dev->adj_list.lower) 8086 return NULL; 8087 8088 *iter = &lower->list; 8089 8090 return lower->dev; 8091 } 8092 EXPORT_SYMBOL(netdev_next_lower_dev_rcu); 8093 8094 static u8 __netdev_upper_depth(struct net_device *dev) 8095 { 8096 struct net_device *udev; 8097 struct list_head *iter; 8098 u8 max_depth = 0; 8099 bool ignore; 8100 8101 for (iter = &dev->adj_list.upper, 8102 udev = __netdev_next_upper_dev(dev, &iter, &ignore); 8103 udev; 8104 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) { 8105 if (ignore) 8106 continue; 8107 if (max_depth < udev->upper_level) 8108 max_depth = udev->upper_level; 8109 } 8110 8111 return max_depth; 8112 } 8113 8114 static u8 __netdev_lower_depth(struct net_device *dev) 8115 { 8116 struct net_device *ldev; 8117 struct list_head *iter; 8118 u8 max_depth = 0; 8119 bool ignore; 8120 8121 for (iter = &dev->adj_list.lower, 8122 ldev = __netdev_next_lower_dev(dev, &iter, &ignore); 8123 ldev; 8124 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) { 8125 if (ignore) 8126 continue; 8127 if (max_depth < ldev->lower_level) 8128 max_depth = ldev->lower_level; 8129 } 8130 8131 return max_depth; 8132 } 8133 8134 static int __netdev_update_upper_level(struct net_device *dev, 8135 struct netdev_nested_priv *__unused) 8136 { 8137 dev->upper_level = __netdev_upper_depth(dev) + 1; 8138 return 0; 8139 } 8140 8141 #ifdef CONFIG_LOCKDEP 8142 static LIST_HEAD(net_unlink_list); 8143 8144 static void net_unlink_todo(struct net_device *dev) 8145 { 8146 if (list_empty(&dev->unlink_list)) 8147 list_add_tail(&dev->unlink_list, &net_unlink_list); 8148 } 8149 #endif 8150 8151 static int __netdev_update_lower_level(struct net_device *dev, 8152 struct netdev_nested_priv *priv) 8153 { 8154 dev->lower_level = __netdev_lower_depth(dev) + 1; 8155 8156 #ifdef CONFIG_LOCKDEP 8157 if (!priv) 8158 return 0; 8159 8160 if (priv->flags & NESTED_SYNC_IMM) 8161 dev->nested_level = dev->lower_level - 1; 8162 if (priv->flags & NESTED_SYNC_TODO) 8163 net_unlink_todo(dev); 8164 #endif 8165 return 0; 8166 } 8167 8168 int netdev_walk_all_lower_dev_rcu(struct net_device *dev, 8169 int (*fn)(struct net_device *dev, 8170 struct netdev_nested_priv *priv), 8171 struct netdev_nested_priv *priv) 8172 { 8173 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8174 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8175 int ret, cur = 0; 8176 8177 now = dev; 8178 iter = &dev->adj_list.lower; 8179 8180 while (1) { 8181 if (now != dev) { 8182 ret = fn(now, priv); 8183 if (ret) 8184 return ret; 8185 } 8186 8187 next = NULL; 8188 while (1) { 8189 ldev = netdev_next_lower_dev_rcu(now, &iter); 8190 if (!ldev) 8191 break; 8192 8193 next = ldev; 8194 niter = &ldev->adj_list.lower; 8195 dev_stack[cur] = now; 8196 iter_stack[cur++] = iter; 8197 break; 8198 } 8199 8200 if (!next) { 8201 if (!cur) 8202 return 0; 8203 next = dev_stack[--cur]; 8204 niter = iter_stack[cur]; 8205 } 8206 8207 now = next; 8208 iter = niter; 8209 } 8210 8211 return 0; 8212 } 8213 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu); 8214 8215 /** 8216 * netdev_lower_get_first_private_rcu - Get the first ->private from the 8217 * lower neighbour list, RCU 8218 * variant 8219 * @dev: device 8220 * 8221 * Gets the first netdev_adjacent->private from the dev's lower neighbour 8222 * list. The caller must hold RCU read lock. 8223 */ 8224 void *netdev_lower_get_first_private_rcu(struct net_device *dev) 8225 { 8226 struct netdev_adjacent *lower; 8227 8228 lower = list_first_or_null_rcu(&dev->adj_list.lower, 8229 struct netdev_adjacent, list); 8230 if (lower) 8231 return lower->private; 8232 return NULL; 8233 } 8234 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu); 8235 8236 /** 8237 * netdev_master_upper_dev_get_rcu - Get master upper device 8238 * @dev: device 8239 * 8240 * Find a master upper device and return pointer to it or NULL in case 8241 * it's not there. The caller must hold the RCU read lock. 8242 */ 8243 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev) 8244 { 8245 struct netdev_adjacent *upper; 8246 8247 upper = list_first_or_null_rcu(&dev->adj_list.upper, 8248 struct netdev_adjacent, list); 8249 if (upper && likely(upper->master)) 8250 return upper->dev; 8251 return NULL; 8252 } 8253 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu); 8254 8255 static int netdev_adjacent_sysfs_add(struct net_device *dev, 8256 struct net_device *adj_dev, 8257 struct list_head *dev_list) 8258 { 8259 char linkname[IFNAMSIZ+7]; 8260 8261 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8262 "upper_%s" : "lower_%s", adj_dev->name); 8263 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj), 8264 linkname); 8265 } 8266 static void netdev_adjacent_sysfs_del(struct net_device *dev, 8267 char *name, 8268 struct list_head *dev_list) 8269 { 8270 char linkname[IFNAMSIZ+7]; 8271 8272 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8273 "upper_%s" : "lower_%s", name); 8274 sysfs_remove_link(&(dev->dev.kobj), linkname); 8275 } 8276 8277 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev, 8278 struct net_device *adj_dev, 8279 struct list_head *dev_list) 8280 { 8281 return (dev_list == &dev->adj_list.upper || 8282 dev_list == &dev->adj_list.lower) && 8283 net_eq(dev_net(dev), dev_net(adj_dev)); 8284 } 8285 8286 static int __netdev_adjacent_dev_insert(struct net_device *dev, 8287 struct net_device *adj_dev, 8288 struct list_head *dev_list, 8289 void *private, bool master) 8290 { 8291 struct netdev_adjacent *adj; 8292 int ret; 8293 8294 adj = __netdev_find_adj(adj_dev, dev_list); 8295 8296 if (adj) { 8297 adj->ref_nr += 1; 8298 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n", 8299 dev->name, adj_dev->name, adj->ref_nr); 8300 8301 return 0; 8302 } 8303 8304 adj = kmalloc(sizeof(*adj), GFP_KERNEL); 8305 if (!adj) 8306 return -ENOMEM; 8307 8308 adj->dev = adj_dev; 8309 adj->master = master; 8310 adj->ref_nr = 1; 8311 adj->private = private; 8312 adj->ignore = false; 8313 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL); 8314 8315 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n", 8316 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name); 8317 8318 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) { 8319 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list); 8320 if (ret) 8321 goto free_adj; 8322 } 8323 8324 /* Ensure that master link is always the first item in list. */ 8325 if (master) { 8326 ret = sysfs_create_link(&(dev->dev.kobj), 8327 &(adj_dev->dev.kobj), "master"); 8328 if (ret) 8329 goto remove_symlinks; 8330 8331 list_add_rcu(&adj->list, dev_list); 8332 } else { 8333 list_add_tail_rcu(&adj->list, dev_list); 8334 } 8335 8336 return 0; 8337 8338 remove_symlinks: 8339 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8340 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8341 free_adj: 8342 netdev_put(adj_dev, &adj->dev_tracker); 8343 kfree(adj); 8344 8345 return ret; 8346 } 8347 8348 static void __netdev_adjacent_dev_remove(struct net_device *dev, 8349 struct net_device *adj_dev, 8350 u16 ref_nr, 8351 struct list_head *dev_list) 8352 { 8353 struct netdev_adjacent *adj; 8354 8355 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n", 8356 dev->name, adj_dev->name, ref_nr); 8357 8358 adj = __netdev_find_adj(adj_dev, dev_list); 8359 8360 if (!adj) { 8361 pr_err("Adjacency does not exist for device %s from %s\n", 8362 dev->name, adj_dev->name); 8363 WARN_ON(1); 8364 return; 8365 } 8366 8367 if (adj->ref_nr > ref_nr) { 8368 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n", 8369 dev->name, adj_dev->name, ref_nr, 8370 adj->ref_nr - ref_nr); 8371 adj->ref_nr -= ref_nr; 8372 return; 8373 } 8374 8375 if (adj->master) 8376 sysfs_remove_link(&(dev->dev.kobj), "master"); 8377 8378 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8379 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8380 8381 list_del_rcu(&adj->list); 8382 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n", 8383 adj_dev->name, dev->name, adj_dev->name); 8384 netdev_put(adj_dev, &adj->dev_tracker); 8385 kfree_rcu(adj, rcu); 8386 } 8387 8388 static int __netdev_adjacent_dev_link_lists(struct net_device *dev, 8389 struct net_device *upper_dev, 8390 struct list_head *up_list, 8391 struct list_head *down_list, 8392 void *private, bool master) 8393 { 8394 int ret; 8395 8396 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list, 8397 private, master); 8398 if (ret) 8399 return ret; 8400 8401 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list, 8402 private, false); 8403 if (ret) { 8404 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list); 8405 return ret; 8406 } 8407 8408 return 0; 8409 } 8410 8411 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev, 8412 struct net_device *upper_dev, 8413 u16 ref_nr, 8414 struct list_head *up_list, 8415 struct list_head *down_list) 8416 { 8417 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list); 8418 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list); 8419 } 8420 8421 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev, 8422 struct net_device *upper_dev, 8423 void *private, bool master) 8424 { 8425 return __netdev_adjacent_dev_link_lists(dev, upper_dev, 8426 &dev->adj_list.upper, 8427 &upper_dev->adj_list.lower, 8428 private, master); 8429 } 8430 8431 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev, 8432 struct net_device *upper_dev) 8433 { 8434 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1, 8435 &dev->adj_list.upper, 8436 &upper_dev->adj_list.lower); 8437 } 8438 8439 static int __netdev_upper_dev_link(struct net_device *dev, 8440 struct net_device *upper_dev, bool master, 8441 void *upper_priv, void *upper_info, 8442 struct netdev_nested_priv *priv, 8443 struct netlink_ext_ack *extack) 8444 { 8445 struct netdev_notifier_changeupper_info changeupper_info = { 8446 .info = { 8447 .dev = dev, 8448 .extack = extack, 8449 }, 8450 .upper_dev = upper_dev, 8451 .master = master, 8452 .linking = true, 8453 .upper_info = upper_info, 8454 }; 8455 struct net_device *master_dev; 8456 int ret = 0; 8457 8458 ASSERT_RTNL(); 8459 8460 if (dev == upper_dev) 8461 return -EBUSY; 8462 8463 /* To prevent loops, check if dev is not upper device to upper_dev. */ 8464 if (__netdev_has_upper_dev(upper_dev, dev)) 8465 return -EBUSY; 8466 8467 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV) 8468 return -EMLINK; 8469 8470 if (!master) { 8471 if (__netdev_has_upper_dev(dev, upper_dev)) 8472 return -EEXIST; 8473 } else { 8474 master_dev = __netdev_master_upper_dev_get(dev); 8475 if (master_dev) 8476 return master_dev == upper_dev ? -EEXIST : -EBUSY; 8477 } 8478 8479 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8480 &changeupper_info.info); 8481 ret = notifier_to_errno(ret); 8482 if (ret) 8483 return ret; 8484 8485 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv, 8486 master); 8487 if (ret) 8488 return ret; 8489 8490 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8491 &changeupper_info.info); 8492 ret = notifier_to_errno(ret); 8493 if (ret) 8494 goto rollback; 8495 8496 __netdev_update_upper_level(dev, NULL); 8497 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8498 8499 __netdev_update_lower_level(upper_dev, priv); 8500 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8501 priv); 8502 8503 return 0; 8504 8505 rollback: 8506 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8507 8508 return ret; 8509 } 8510 8511 /** 8512 * netdev_upper_dev_link - Add a link to the upper device 8513 * @dev: device 8514 * @upper_dev: new upper device 8515 * @extack: netlink extended ack 8516 * 8517 * Adds a link to device which is upper to this one. The caller must hold 8518 * the RTNL lock. On a failure a negative errno code is returned. 8519 * On success the reference counts are adjusted and the function 8520 * returns zero. 8521 */ 8522 int netdev_upper_dev_link(struct net_device *dev, 8523 struct net_device *upper_dev, 8524 struct netlink_ext_ack *extack) 8525 { 8526 struct netdev_nested_priv priv = { 8527 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8528 .data = NULL, 8529 }; 8530 8531 return __netdev_upper_dev_link(dev, upper_dev, false, 8532 NULL, NULL, &priv, extack); 8533 } 8534 EXPORT_SYMBOL(netdev_upper_dev_link); 8535 8536 /** 8537 * netdev_master_upper_dev_link - Add a master link to the upper device 8538 * @dev: device 8539 * @upper_dev: new upper device 8540 * @upper_priv: upper device private 8541 * @upper_info: upper info to be passed down via notifier 8542 * @extack: netlink extended ack 8543 * 8544 * Adds a link to device which is upper to this one. In this case, only 8545 * one master upper device can be linked, although other non-master devices 8546 * might be linked as well. The caller must hold the RTNL lock. 8547 * On a failure a negative errno code is returned. On success the reference 8548 * counts are adjusted and the function returns zero. 8549 */ 8550 int netdev_master_upper_dev_link(struct net_device *dev, 8551 struct net_device *upper_dev, 8552 void *upper_priv, void *upper_info, 8553 struct netlink_ext_ack *extack) 8554 { 8555 struct netdev_nested_priv priv = { 8556 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8557 .data = NULL, 8558 }; 8559 8560 return __netdev_upper_dev_link(dev, upper_dev, true, 8561 upper_priv, upper_info, &priv, extack); 8562 } 8563 EXPORT_SYMBOL(netdev_master_upper_dev_link); 8564 8565 static void __netdev_upper_dev_unlink(struct net_device *dev, 8566 struct net_device *upper_dev, 8567 struct netdev_nested_priv *priv) 8568 { 8569 struct netdev_notifier_changeupper_info changeupper_info = { 8570 .info = { 8571 .dev = dev, 8572 }, 8573 .upper_dev = upper_dev, 8574 .linking = false, 8575 }; 8576 8577 ASSERT_RTNL(); 8578 8579 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev; 8580 8581 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8582 &changeupper_info.info); 8583 8584 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8585 8586 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8587 &changeupper_info.info); 8588 8589 __netdev_update_upper_level(dev, NULL); 8590 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8591 8592 __netdev_update_lower_level(upper_dev, priv); 8593 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8594 priv); 8595 } 8596 8597 /** 8598 * netdev_upper_dev_unlink - Removes a link to upper device 8599 * @dev: device 8600 * @upper_dev: new upper device 8601 * 8602 * Removes a link to device which is upper to this one. The caller must hold 8603 * the RTNL lock. 8604 */ 8605 void netdev_upper_dev_unlink(struct net_device *dev, 8606 struct net_device *upper_dev) 8607 { 8608 struct netdev_nested_priv priv = { 8609 .flags = NESTED_SYNC_TODO, 8610 .data = NULL, 8611 }; 8612 8613 __netdev_upper_dev_unlink(dev, upper_dev, &priv); 8614 } 8615 EXPORT_SYMBOL(netdev_upper_dev_unlink); 8616 8617 static void __netdev_adjacent_dev_set(struct net_device *upper_dev, 8618 struct net_device *lower_dev, 8619 bool val) 8620 { 8621 struct netdev_adjacent *adj; 8622 8623 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower); 8624 if (adj) 8625 adj->ignore = val; 8626 8627 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper); 8628 if (adj) 8629 adj->ignore = val; 8630 } 8631 8632 static void netdev_adjacent_dev_disable(struct net_device *upper_dev, 8633 struct net_device *lower_dev) 8634 { 8635 __netdev_adjacent_dev_set(upper_dev, lower_dev, true); 8636 } 8637 8638 static void netdev_adjacent_dev_enable(struct net_device *upper_dev, 8639 struct net_device *lower_dev) 8640 { 8641 __netdev_adjacent_dev_set(upper_dev, lower_dev, false); 8642 } 8643 8644 int netdev_adjacent_change_prepare(struct net_device *old_dev, 8645 struct net_device *new_dev, 8646 struct net_device *dev, 8647 struct netlink_ext_ack *extack) 8648 { 8649 struct netdev_nested_priv priv = { 8650 .flags = 0, 8651 .data = NULL, 8652 }; 8653 int err; 8654 8655 if (!new_dev) 8656 return 0; 8657 8658 if (old_dev && new_dev != old_dev) 8659 netdev_adjacent_dev_disable(dev, old_dev); 8660 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv, 8661 extack); 8662 if (err) { 8663 if (old_dev && new_dev != old_dev) 8664 netdev_adjacent_dev_enable(dev, old_dev); 8665 return err; 8666 } 8667 8668 return 0; 8669 } 8670 EXPORT_SYMBOL(netdev_adjacent_change_prepare); 8671 8672 void netdev_adjacent_change_commit(struct net_device *old_dev, 8673 struct net_device *new_dev, 8674 struct net_device *dev) 8675 { 8676 struct netdev_nested_priv priv = { 8677 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8678 .data = NULL, 8679 }; 8680 8681 if (!new_dev || !old_dev) 8682 return; 8683 8684 if (new_dev == old_dev) 8685 return; 8686 8687 netdev_adjacent_dev_enable(dev, old_dev); 8688 __netdev_upper_dev_unlink(old_dev, dev, &priv); 8689 } 8690 EXPORT_SYMBOL(netdev_adjacent_change_commit); 8691 8692 void netdev_adjacent_change_abort(struct net_device *old_dev, 8693 struct net_device *new_dev, 8694 struct net_device *dev) 8695 { 8696 struct netdev_nested_priv priv = { 8697 .flags = 0, 8698 .data = NULL, 8699 }; 8700 8701 if (!new_dev) 8702 return; 8703 8704 if (old_dev && new_dev != old_dev) 8705 netdev_adjacent_dev_enable(dev, old_dev); 8706 8707 __netdev_upper_dev_unlink(new_dev, dev, &priv); 8708 } 8709 EXPORT_SYMBOL(netdev_adjacent_change_abort); 8710 8711 /** 8712 * netdev_bonding_info_change - Dispatch event about slave change 8713 * @dev: device 8714 * @bonding_info: info to dispatch 8715 * 8716 * Send NETDEV_BONDING_INFO to netdev notifiers with info. 8717 * The caller must hold the RTNL lock. 8718 */ 8719 void netdev_bonding_info_change(struct net_device *dev, 8720 struct netdev_bonding_info *bonding_info) 8721 { 8722 struct netdev_notifier_bonding_info info = { 8723 .info.dev = dev, 8724 }; 8725 8726 memcpy(&info.bonding_info, bonding_info, 8727 sizeof(struct netdev_bonding_info)); 8728 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, 8729 &info.info); 8730 } 8731 EXPORT_SYMBOL(netdev_bonding_info_change); 8732 8733 static int netdev_offload_xstats_enable_l3(struct net_device *dev, 8734 struct netlink_ext_ack *extack) 8735 { 8736 struct netdev_notifier_offload_xstats_info info = { 8737 .info.dev = dev, 8738 .info.extack = extack, 8739 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8740 }; 8741 int err; 8742 int rc; 8743 8744 dev->offload_xstats_l3 = kzalloc(sizeof(*dev->offload_xstats_l3), 8745 GFP_KERNEL); 8746 if (!dev->offload_xstats_l3) 8747 return -ENOMEM; 8748 8749 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE, 8750 NETDEV_OFFLOAD_XSTATS_DISABLE, 8751 &info.info); 8752 err = notifier_to_errno(rc); 8753 if (err) 8754 goto free_stats; 8755 8756 return 0; 8757 8758 free_stats: 8759 kfree(dev->offload_xstats_l3); 8760 dev->offload_xstats_l3 = NULL; 8761 return err; 8762 } 8763 8764 int netdev_offload_xstats_enable(struct net_device *dev, 8765 enum netdev_offload_xstats_type type, 8766 struct netlink_ext_ack *extack) 8767 { 8768 ASSERT_RTNL(); 8769 8770 if (netdev_offload_xstats_enabled(dev, type)) 8771 return -EALREADY; 8772 8773 switch (type) { 8774 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8775 return netdev_offload_xstats_enable_l3(dev, extack); 8776 } 8777 8778 WARN_ON(1); 8779 return -EINVAL; 8780 } 8781 EXPORT_SYMBOL(netdev_offload_xstats_enable); 8782 8783 static void netdev_offload_xstats_disable_l3(struct net_device *dev) 8784 { 8785 struct netdev_notifier_offload_xstats_info info = { 8786 .info.dev = dev, 8787 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8788 }; 8789 8790 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE, 8791 &info.info); 8792 kfree(dev->offload_xstats_l3); 8793 dev->offload_xstats_l3 = NULL; 8794 } 8795 8796 int netdev_offload_xstats_disable(struct net_device *dev, 8797 enum netdev_offload_xstats_type type) 8798 { 8799 ASSERT_RTNL(); 8800 8801 if (!netdev_offload_xstats_enabled(dev, type)) 8802 return -EALREADY; 8803 8804 switch (type) { 8805 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8806 netdev_offload_xstats_disable_l3(dev); 8807 return 0; 8808 } 8809 8810 WARN_ON(1); 8811 return -EINVAL; 8812 } 8813 EXPORT_SYMBOL(netdev_offload_xstats_disable); 8814 8815 static void netdev_offload_xstats_disable_all(struct net_device *dev) 8816 { 8817 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3); 8818 } 8819 8820 static struct rtnl_hw_stats64 * 8821 netdev_offload_xstats_get_ptr(const struct net_device *dev, 8822 enum netdev_offload_xstats_type type) 8823 { 8824 switch (type) { 8825 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8826 return dev->offload_xstats_l3; 8827 } 8828 8829 WARN_ON(1); 8830 return NULL; 8831 } 8832 8833 bool netdev_offload_xstats_enabled(const struct net_device *dev, 8834 enum netdev_offload_xstats_type type) 8835 { 8836 ASSERT_RTNL(); 8837 8838 return netdev_offload_xstats_get_ptr(dev, type); 8839 } 8840 EXPORT_SYMBOL(netdev_offload_xstats_enabled); 8841 8842 struct netdev_notifier_offload_xstats_ru { 8843 bool used; 8844 }; 8845 8846 struct netdev_notifier_offload_xstats_rd { 8847 struct rtnl_hw_stats64 stats; 8848 bool used; 8849 }; 8850 8851 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest, 8852 const struct rtnl_hw_stats64 *src) 8853 { 8854 dest->rx_packets += src->rx_packets; 8855 dest->tx_packets += src->tx_packets; 8856 dest->rx_bytes += src->rx_bytes; 8857 dest->tx_bytes += src->tx_bytes; 8858 dest->rx_errors += src->rx_errors; 8859 dest->tx_errors += src->tx_errors; 8860 dest->rx_dropped += src->rx_dropped; 8861 dest->tx_dropped += src->tx_dropped; 8862 dest->multicast += src->multicast; 8863 } 8864 8865 static int netdev_offload_xstats_get_used(struct net_device *dev, 8866 enum netdev_offload_xstats_type type, 8867 bool *p_used, 8868 struct netlink_ext_ack *extack) 8869 { 8870 struct netdev_notifier_offload_xstats_ru report_used = {}; 8871 struct netdev_notifier_offload_xstats_info info = { 8872 .info.dev = dev, 8873 .info.extack = extack, 8874 .type = type, 8875 .report_used = &report_used, 8876 }; 8877 int rc; 8878 8879 WARN_ON(!netdev_offload_xstats_enabled(dev, type)); 8880 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED, 8881 &info.info); 8882 *p_used = report_used.used; 8883 return notifier_to_errno(rc); 8884 } 8885 8886 static int netdev_offload_xstats_get_stats(struct net_device *dev, 8887 enum netdev_offload_xstats_type type, 8888 struct rtnl_hw_stats64 *p_stats, 8889 bool *p_used, 8890 struct netlink_ext_ack *extack) 8891 { 8892 struct netdev_notifier_offload_xstats_rd report_delta = {}; 8893 struct netdev_notifier_offload_xstats_info info = { 8894 .info.dev = dev, 8895 .info.extack = extack, 8896 .type = type, 8897 .report_delta = &report_delta, 8898 }; 8899 struct rtnl_hw_stats64 *stats; 8900 int rc; 8901 8902 stats = netdev_offload_xstats_get_ptr(dev, type); 8903 if (WARN_ON(!stats)) 8904 return -EINVAL; 8905 8906 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA, 8907 &info.info); 8908 8909 /* Cache whatever we got, even if there was an error, otherwise the 8910 * successful stats retrievals would get lost. 8911 */ 8912 netdev_hw_stats64_add(stats, &report_delta.stats); 8913 8914 if (p_stats) 8915 *p_stats = *stats; 8916 *p_used = report_delta.used; 8917 8918 return notifier_to_errno(rc); 8919 } 8920 8921 int netdev_offload_xstats_get(struct net_device *dev, 8922 enum netdev_offload_xstats_type type, 8923 struct rtnl_hw_stats64 *p_stats, bool *p_used, 8924 struct netlink_ext_ack *extack) 8925 { 8926 ASSERT_RTNL(); 8927 8928 if (p_stats) 8929 return netdev_offload_xstats_get_stats(dev, type, p_stats, 8930 p_used, extack); 8931 else 8932 return netdev_offload_xstats_get_used(dev, type, p_used, 8933 extack); 8934 } 8935 EXPORT_SYMBOL(netdev_offload_xstats_get); 8936 8937 void 8938 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta, 8939 const struct rtnl_hw_stats64 *stats) 8940 { 8941 report_delta->used = true; 8942 netdev_hw_stats64_add(&report_delta->stats, stats); 8943 } 8944 EXPORT_SYMBOL(netdev_offload_xstats_report_delta); 8945 8946 void 8947 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used) 8948 { 8949 report_used->used = true; 8950 } 8951 EXPORT_SYMBOL(netdev_offload_xstats_report_used); 8952 8953 void netdev_offload_xstats_push_delta(struct net_device *dev, 8954 enum netdev_offload_xstats_type type, 8955 const struct rtnl_hw_stats64 *p_stats) 8956 { 8957 struct rtnl_hw_stats64 *stats; 8958 8959 ASSERT_RTNL(); 8960 8961 stats = netdev_offload_xstats_get_ptr(dev, type); 8962 if (WARN_ON(!stats)) 8963 return; 8964 8965 netdev_hw_stats64_add(stats, p_stats); 8966 } 8967 EXPORT_SYMBOL(netdev_offload_xstats_push_delta); 8968 8969 /** 8970 * netdev_get_xmit_slave - Get the xmit slave of master device 8971 * @dev: device 8972 * @skb: The packet 8973 * @all_slaves: assume all the slaves are active 8974 * 8975 * The reference counters are not incremented so the caller must be 8976 * careful with locks. The caller must hold RCU lock. 8977 * %NULL is returned if no slave is found. 8978 */ 8979 8980 struct net_device *netdev_get_xmit_slave(struct net_device *dev, 8981 struct sk_buff *skb, 8982 bool all_slaves) 8983 { 8984 const struct net_device_ops *ops = dev->netdev_ops; 8985 8986 if (!ops->ndo_get_xmit_slave) 8987 return NULL; 8988 return ops->ndo_get_xmit_slave(dev, skb, all_slaves); 8989 } 8990 EXPORT_SYMBOL(netdev_get_xmit_slave); 8991 8992 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev, 8993 struct sock *sk) 8994 { 8995 const struct net_device_ops *ops = dev->netdev_ops; 8996 8997 if (!ops->ndo_sk_get_lower_dev) 8998 return NULL; 8999 return ops->ndo_sk_get_lower_dev(dev, sk); 9000 } 9001 9002 /** 9003 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket 9004 * @dev: device 9005 * @sk: the socket 9006 * 9007 * %NULL is returned if no lower device is found. 9008 */ 9009 9010 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev, 9011 struct sock *sk) 9012 { 9013 struct net_device *lower; 9014 9015 lower = netdev_sk_get_lower_dev(dev, sk); 9016 while (lower) { 9017 dev = lower; 9018 lower = netdev_sk_get_lower_dev(dev, sk); 9019 } 9020 9021 return dev; 9022 } 9023 EXPORT_SYMBOL(netdev_sk_get_lowest_dev); 9024 9025 static void netdev_adjacent_add_links(struct net_device *dev) 9026 { 9027 struct netdev_adjacent *iter; 9028 9029 struct net *net = dev_net(dev); 9030 9031 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9032 if (!net_eq(net, dev_net(iter->dev))) 9033 continue; 9034 netdev_adjacent_sysfs_add(iter->dev, dev, 9035 &iter->dev->adj_list.lower); 9036 netdev_adjacent_sysfs_add(dev, iter->dev, 9037 &dev->adj_list.upper); 9038 } 9039 9040 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9041 if (!net_eq(net, dev_net(iter->dev))) 9042 continue; 9043 netdev_adjacent_sysfs_add(iter->dev, dev, 9044 &iter->dev->adj_list.upper); 9045 netdev_adjacent_sysfs_add(dev, iter->dev, 9046 &dev->adj_list.lower); 9047 } 9048 } 9049 9050 static void netdev_adjacent_del_links(struct net_device *dev) 9051 { 9052 struct netdev_adjacent *iter; 9053 9054 struct net *net = dev_net(dev); 9055 9056 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9057 if (!net_eq(net, dev_net(iter->dev))) 9058 continue; 9059 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9060 &iter->dev->adj_list.lower); 9061 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9062 &dev->adj_list.upper); 9063 } 9064 9065 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9066 if (!net_eq(net, dev_net(iter->dev))) 9067 continue; 9068 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9069 &iter->dev->adj_list.upper); 9070 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9071 &dev->adj_list.lower); 9072 } 9073 } 9074 9075 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname) 9076 { 9077 struct netdev_adjacent *iter; 9078 9079 struct net *net = dev_net(dev); 9080 9081 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9082 if (!net_eq(net, dev_net(iter->dev))) 9083 continue; 9084 netdev_adjacent_sysfs_del(iter->dev, oldname, 9085 &iter->dev->adj_list.lower); 9086 netdev_adjacent_sysfs_add(iter->dev, dev, 9087 &iter->dev->adj_list.lower); 9088 } 9089 9090 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9091 if (!net_eq(net, dev_net(iter->dev))) 9092 continue; 9093 netdev_adjacent_sysfs_del(iter->dev, oldname, 9094 &iter->dev->adj_list.upper); 9095 netdev_adjacent_sysfs_add(iter->dev, dev, 9096 &iter->dev->adj_list.upper); 9097 } 9098 } 9099 9100 void *netdev_lower_dev_get_private(struct net_device *dev, 9101 struct net_device *lower_dev) 9102 { 9103 struct netdev_adjacent *lower; 9104 9105 if (!lower_dev) 9106 return NULL; 9107 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower); 9108 if (!lower) 9109 return NULL; 9110 9111 return lower->private; 9112 } 9113 EXPORT_SYMBOL(netdev_lower_dev_get_private); 9114 9115 9116 /** 9117 * netdev_lower_state_changed - Dispatch event about lower device state change 9118 * @lower_dev: device 9119 * @lower_state_info: state to dispatch 9120 * 9121 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info. 9122 * The caller must hold the RTNL lock. 9123 */ 9124 void netdev_lower_state_changed(struct net_device *lower_dev, 9125 void *lower_state_info) 9126 { 9127 struct netdev_notifier_changelowerstate_info changelowerstate_info = { 9128 .info.dev = lower_dev, 9129 }; 9130 9131 ASSERT_RTNL(); 9132 changelowerstate_info.lower_state_info = lower_state_info; 9133 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, 9134 &changelowerstate_info.info); 9135 } 9136 EXPORT_SYMBOL(netdev_lower_state_changed); 9137 9138 static void dev_change_rx_flags(struct net_device *dev, int flags) 9139 { 9140 const struct net_device_ops *ops = dev->netdev_ops; 9141 9142 if (ops->ndo_change_rx_flags) 9143 ops->ndo_change_rx_flags(dev, flags); 9144 } 9145 9146 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify) 9147 { 9148 unsigned int old_flags = dev->flags; 9149 unsigned int promiscuity, flags; 9150 kuid_t uid; 9151 kgid_t gid; 9152 9153 ASSERT_RTNL(); 9154 9155 promiscuity = dev->promiscuity + inc; 9156 if (promiscuity == 0) { 9157 /* 9158 * Avoid overflow. 9159 * If inc causes overflow, untouch promisc and return error. 9160 */ 9161 if (unlikely(inc > 0)) { 9162 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n"); 9163 return -EOVERFLOW; 9164 } 9165 flags = old_flags & ~IFF_PROMISC; 9166 } else { 9167 flags = old_flags | IFF_PROMISC; 9168 } 9169 WRITE_ONCE(dev->promiscuity, promiscuity); 9170 if (flags != old_flags) { 9171 WRITE_ONCE(dev->flags, flags); 9172 netdev_info(dev, "%s promiscuous mode\n", 9173 dev->flags & IFF_PROMISC ? "entered" : "left"); 9174 if (audit_enabled) { 9175 current_uid_gid(&uid, &gid); 9176 audit_log(audit_context(), GFP_ATOMIC, 9177 AUDIT_ANOM_PROMISCUOUS, 9178 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u", 9179 dev->name, (dev->flags & IFF_PROMISC), 9180 (old_flags & IFF_PROMISC), 9181 from_kuid(&init_user_ns, audit_get_loginuid(current)), 9182 from_kuid(&init_user_ns, uid), 9183 from_kgid(&init_user_ns, gid), 9184 audit_get_sessionid(current)); 9185 } 9186 9187 dev_change_rx_flags(dev, IFF_PROMISC); 9188 } 9189 if (notify) 9190 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL); 9191 return 0; 9192 } 9193 9194 /** 9195 * dev_set_promiscuity - update promiscuity count on a device 9196 * @dev: device 9197 * @inc: modifier 9198 * 9199 * Add or remove promiscuity from a device. While the count in the device 9200 * remains above zero the interface remains promiscuous. Once it hits zero 9201 * the device reverts back to normal filtering operation. A negative inc 9202 * value is used to drop promiscuity on the device. 9203 * Return 0 if successful or a negative errno code on error. 9204 */ 9205 int dev_set_promiscuity(struct net_device *dev, int inc) 9206 { 9207 unsigned int old_flags = dev->flags; 9208 int err; 9209 9210 err = __dev_set_promiscuity(dev, inc, true); 9211 if (err < 0) 9212 return err; 9213 if (dev->flags != old_flags) 9214 dev_set_rx_mode(dev); 9215 return err; 9216 } 9217 EXPORT_SYMBOL(dev_set_promiscuity); 9218 9219 int netif_set_allmulti(struct net_device *dev, int inc, bool notify) 9220 { 9221 unsigned int old_flags = dev->flags, old_gflags = dev->gflags; 9222 unsigned int allmulti, flags; 9223 9224 ASSERT_RTNL(); 9225 9226 allmulti = dev->allmulti + inc; 9227 if (allmulti == 0) { 9228 /* 9229 * Avoid overflow. 9230 * If inc causes overflow, untouch allmulti and return error. 9231 */ 9232 if (unlikely(inc > 0)) { 9233 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n"); 9234 return -EOVERFLOW; 9235 } 9236 flags = old_flags & ~IFF_ALLMULTI; 9237 } else { 9238 flags = old_flags | IFF_ALLMULTI; 9239 } 9240 WRITE_ONCE(dev->allmulti, allmulti); 9241 if (flags != old_flags) { 9242 WRITE_ONCE(dev->flags, flags); 9243 netdev_info(dev, "%s allmulticast mode\n", 9244 dev->flags & IFF_ALLMULTI ? "entered" : "left"); 9245 dev_change_rx_flags(dev, IFF_ALLMULTI); 9246 dev_set_rx_mode(dev); 9247 if (notify) 9248 __dev_notify_flags(dev, old_flags, 9249 dev->gflags ^ old_gflags, 0, NULL); 9250 } 9251 return 0; 9252 } 9253 9254 /* 9255 * Upload unicast and multicast address lists to device and 9256 * configure RX filtering. When the device doesn't support unicast 9257 * filtering it is put in promiscuous mode while unicast addresses 9258 * are present. 9259 */ 9260 void __dev_set_rx_mode(struct net_device *dev) 9261 { 9262 const struct net_device_ops *ops = dev->netdev_ops; 9263 9264 /* dev_open will call this function so the list will stay sane. */ 9265 if (!(dev->flags&IFF_UP)) 9266 return; 9267 9268 if (!netif_device_present(dev)) 9269 return; 9270 9271 if (!(dev->priv_flags & IFF_UNICAST_FLT)) { 9272 /* Unicast addresses changes may only happen under the rtnl, 9273 * therefore calling __dev_set_promiscuity here is safe. 9274 */ 9275 if (!netdev_uc_empty(dev) && !dev->uc_promisc) { 9276 __dev_set_promiscuity(dev, 1, false); 9277 dev->uc_promisc = true; 9278 } else if (netdev_uc_empty(dev) && dev->uc_promisc) { 9279 __dev_set_promiscuity(dev, -1, false); 9280 dev->uc_promisc = false; 9281 } 9282 } 9283 9284 if (ops->ndo_set_rx_mode) 9285 ops->ndo_set_rx_mode(dev); 9286 } 9287 9288 void dev_set_rx_mode(struct net_device *dev) 9289 { 9290 netif_addr_lock_bh(dev); 9291 __dev_set_rx_mode(dev); 9292 netif_addr_unlock_bh(dev); 9293 } 9294 9295 /** 9296 * dev_get_flags - get flags reported to userspace 9297 * @dev: device 9298 * 9299 * Get the combination of flag bits exported through APIs to userspace. 9300 */ 9301 unsigned int dev_get_flags(const struct net_device *dev) 9302 { 9303 unsigned int flags; 9304 9305 flags = (READ_ONCE(dev->flags) & ~(IFF_PROMISC | 9306 IFF_ALLMULTI | 9307 IFF_RUNNING | 9308 IFF_LOWER_UP | 9309 IFF_DORMANT)) | 9310 (READ_ONCE(dev->gflags) & (IFF_PROMISC | 9311 IFF_ALLMULTI)); 9312 9313 if (netif_running(dev)) { 9314 if (netif_oper_up(dev)) 9315 flags |= IFF_RUNNING; 9316 if (netif_carrier_ok(dev)) 9317 flags |= IFF_LOWER_UP; 9318 if (netif_dormant(dev)) 9319 flags |= IFF_DORMANT; 9320 } 9321 9322 return flags; 9323 } 9324 EXPORT_SYMBOL(dev_get_flags); 9325 9326 int __dev_change_flags(struct net_device *dev, unsigned int flags, 9327 struct netlink_ext_ack *extack) 9328 { 9329 unsigned int old_flags = dev->flags; 9330 int ret; 9331 9332 ASSERT_RTNL(); 9333 9334 /* 9335 * Set the flags on our device. 9336 */ 9337 9338 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP | 9339 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL | 9340 IFF_AUTOMEDIA)) | 9341 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC | 9342 IFF_ALLMULTI)); 9343 9344 /* 9345 * Load in the correct multicast list now the flags have changed. 9346 */ 9347 9348 if ((old_flags ^ flags) & IFF_MULTICAST) 9349 dev_change_rx_flags(dev, IFF_MULTICAST); 9350 9351 dev_set_rx_mode(dev); 9352 9353 /* 9354 * Have we downed the interface. We handle IFF_UP ourselves 9355 * according to user attempts to set it, rather than blindly 9356 * setting it. 9357 */ 9358 9359 ret = 0; 9360 if ((old_flags ^ flags) & IFF_UP) { 9361 if (old_flags & IFF_UP) 9362 __dev_close(dev); 9363 else 9364 ret = __dev_open(dev, extack); 9365 } 9366 9367 if ((flags ^ dev->gflags) & IFF_PROMISC) { 9368 int inc = (flags & IFF_PROMISC) ? 1 : -1; 9369 old_flags = dev->flags; 9370 9371 dev->gflags ^= IFF_PROMISC; 9372 9373 if (__dev_set_promiscuity(dev, inc, false) >= 0) 9374 if (dev->flags != old_flags) 9375 dev_set_rx_mode(dev); 9376 } 9377 9378 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI 9379 * is important. Some (broken) drivers set IFF_PROMISC, when 9380 * IFF_ALLMULTI is requested not asking us and not reporting. 9381 */ 9382 if ((flags ^ dev->gflags) & IFF_ALLMULTI) { 9383 int inc = (flags & IFF_ALLMULTI) ? 1 : -1; 9384 9385 dev->gflags ^= IFF_ALLMULTI; 9386 netif_set_allmulti(dev, inc, false); 9387 } 9388 9389 return ret; 9390 } 9391 9392 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags, 9393 unsigned int gchanges, u32 portid, 9394 const struct nlmsghdr *nlh) 9395 { 9396 unsigned int changes = dev->flags ^ old_flags; 9397 9398 if (gchanges) 9399 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh); 9400 9401 if (changes & IFF_UP) { 9402 if (dev->flags & IFF_UP) 9403 call_netdevice_notifiers(NETDEV_UP, dev); 9404 else 9405 call_netdevice_notifiers(NETDEV_DOWN, dev); 9406 } 9407 9408 if (dev->flags & IFF_UP && 9409 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) { 9410 struct netdev_notifier_change_info change_info = { 9411 .info = { 9412 .dev = dev, 9413 }, 9414 .flags_changed = changes, 9415 }; 9416 9417 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info); 9418 } 9419 } 9420 9421 int netif_change_flags(struct net_device *dev, unsigned int flags, 9422 struct netlink_ext_ack *extack) 9423 { 9424 int ret; 9425 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags; 9426 9427 ret = __dev_change_flags(dev, flags, extack); 9428 if (ret < 0) 9429 return ret; 9430 9431 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags); 9432 __dev_notify_flags(dev, old_flags, changes, 0, NULL); 9433 return ret; 9434 } 9435 9436 int __dev_set_mtu(struct net_device *dev, int new_mtu) 9437 { 9438 const struct net_device_ops *ops = dev->netdev_ops; 9439 9440 if (ops->ndo_change_mtu) 9441 return ops->ndo_change_mtu(dev, new_mtu); 9442 9443 /* Pairs with all the lockless reads of dev->mtu in the stack */ 9444 WRITE_ONCE(dev->mtu, new_mtu); 9445 return 0; 9446 } 9447 EXPORT_SYMBOL(__dev_set_mtu); 9448 9449 int dev_validate_mtu(struct net_device *dev, int new_mtu, 9450 struct netlink_ext_ack *extack) 9451 { 9452 /* MTU must be positive, and in range */ 9453 if (new_mtu < 0 || new_mtu < dev->min_mtu) { 9454 NL_SET_ERR_MSG(extack, "mtu less than device minimum"); 9455 return -EINVAL; 9456 } 9457 9458 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) { 9459 NL_SET_ERR_MSG(extack, "mtu greater than device maximum"); 9460 return -EINVAL; 9461 } 9462 return 0; 9463 } 9464 9465 /** 9466 * netif_set_mtu_ext - Change maximum transfer unit 9467 * @dev: device 9468 * @new_mtu: new transfer unit 9469 * @extack: netlink extended ack 9470 * 9471 * Change the maximum transfer size of the network device. 9472 */ 9473 int netif_set_mtu_ext(struct net_device *dev, int new_mtu, 9474 struct netlink_ext_ack *extack) 9475 { 9476 int err, orig_mtu; 9477 9478 if (new_mtu == dev->mtu) 9479 return 0; 9480 9481 err = dev_validate_mtu(dev, new_mtu, extack); 9482 if (err) 9483 return err; 9484 9485 if (!netif_device_present(dev)) 9486 return -ENODEV; 9487 9488 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev); 9489 err = notifier_to_errno(err); 9490 if (err) 9491 return err; 9492 9493 orig_mtu = dev->mtu; 9494 err = __dev_set_mtu(dev, new_mtu); 9495 9496 if (!err) { 9497 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9498 orig_mtu); 9499 err = notifier_to_errno(err); 9500 if (err) { 9501 /* setting mtu back and notifying everyone again, 9502 * so that they have a chance to revert changes. 9503 */ 9504 __dev_set_mtu(dev, orig_mtu); 9505 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9506 new_mtu); 9507 } 9508 } 9509 return err; 9510 } 9511 9512 int netif_set_mtu(struct net_device *dev, int new_mtu) 9513 { 9514 struct netlink_ext_ack extack; 9515 int err; 9516 9517 memset(&extack, 0, sizeof(extack)); 9518 err = netif_set_mtu_ext(dev, new_mtu, &extack); 9519 if (err && extack._msg) 9520 net_err_ratelimited("%s: %s\n", dev->name, extack._msg); 9521 return err; 9522 } 9523 EXPORT_SYMBOL(netif_set_mtu); 9524 9525 int netif_change_tx_queue_len(struct net_device *dev, unsigned long new_len) 9526 { 9527 unsigned int orig_len = dev->tx_queue_len; 9528 int res; 9529 9530 if (new_len != (unsigned int)new_len) 9531 return -ERANGE; 9532 9533 if (new_len != orig_len) { 9534 WRITE_ONCE(dev->tx_queue_len, new_len); 9535 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev); 9536 res = notifier_to_errno(res); 9537 if (res) 9538 goto err_rollback; 9539 res = dev_qdisc_change_tx_queue_len(dev); 9540 if (res) 9541 goto err_rollback; 9542 } 9543 9544 return 0; 9545 9546 err_rollback: 9547 netdev_err(dev, "refused to change device tx_queue_len\n"); 9548 WRITE_ONCE(dev->tx_queue_len, orig_len); 9549 return res; 9550 } 9551 9552 void netif_set_group(struct net_device *dev, int new_group) 9553 { 9554 dev->group = new_group; 9555 } 9556 9557 /** 9558 * dev_pre_changeaddr_notify - Call NETDEV_PRE_CHANGEADDR. 9559 * @dev: device 9560 * @addr: new address 9561 * @extack: netlink extended ack 9562 */ 9563 int dev_pre_changeaddr_notify(struct net_device *dev, const char *addr, 9564 struct netlink_ext_ack *extack) 9565 { 9566 struct netdev_notifier_pre_changeaddr_info info = { 9567 .info.dev = dev, 9568 .info.extack = extack, 9569 .dev_addr = addr, 9570 }; 9571 int rc; 9572 9573 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info); 9574 return notifier_to_errno(rc); 9575 } 9576 EXPORT_SYMBOL(dev_pre_changeaddr_notify); 9577 9578 int netif_set_mac_address(struct net_device *dev, struct sockaddr *sa, 9579 struct netlink_ext_ack *extack) 9580 { 9581 const struct net_device_ops *ops = dev->netdev_ops; 9582 int err; 9583 9584 if (!ops->ndo_set_mac_address) 9585 return -EOPNOTSUPP; 9586 if (sa->sa_family != dev->type) 9587 return -EINVAL; 9588 if (!netif_device_present(dev)) 9589 return -ENODEV; 9590 err = dev_pre_changeaddr_notify(dev, sa->sa_data, extack); 9591 if (err) 9592 return err; 9593 if (memcmp(dev->dev_addr, sa->sa_data, dev->addr_len)) { 9594 err = ops->ndo_set_mac_address(dev, sa); 9595 if (err) 9596 return err; 9597 } 9598 dev->addr_assign_type = NET_ADDR_SET; 9599 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev); 9600 add_device_randomness(dev->dev_addr, dev->addr_len); 9601 return 0; 9602 } 9603 9604 DECLARE_RWSEM(dev_addr_sem); 9605 9606 int dev_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name) 9607 { 9608 size_t size = sizeof(sa->sa_data_min); 9609 struct net_device *dev; 9610 int ret = 0; 9611 9612 down_read(&dev_addr_sem); 9613 rcu_read_lock(); 9614 9615 dev = dev_get_by_name_rcu(net, dev_name); 9616 if (!dev) { 9617 ret = -ENODEV; 9618 goto unlock; 9619 } 9620 if (!dev->addr_len) 9621 memset(sa->sa_data, 0, size); 9622 else 9623 memcpy(sa->sa_data, dev->dev_addr, 9624 min_t(size_t, size, dev->addr_len)); 9625 sa->sa_family = dev->type; 9626 9627 unlock: 9628 rcu_read_unlock(); 9629 up_read(&dev_addr_sem); 9630 return ret; 9631 } 9632 EXPORT_SYMBOL(dev_get_mac_address); 9633 9634 int netif_change_carrier(struct net_device *dev, bool new_carrier) 9635 { 9636 const struct net_device_ops *ops = dev->netdev_ops; 9637 9638 if (!ops->ndo_change_carrier) 9639 return -EOPNOTSUPP; 9640 if (!netif_device_present(dev)) 9641 return -ENODEV; 9642 return ops->ndo_change_carrier(dev, new_carrier); 9643 } 9644 9645 /** 9646 * dev_get_phys_port_id - Get device physical port ID 9647 * @dev: device 9648 * @ppid: port ID 9649 * 9650 * Get device physical port ID 9651 */ 9652 int dev_get_phys_port_id(struct net_device *dev, 9653 struct netdev_phys_item_id *ppid) 9654 { 9655 const struct net_device_ops *ops = dev->netdev_ops; 9656 9657 if (!ops->ndo_get_phys_port_id) 9658 return -EOPNOTSUPP; 9659 return ops->ndo_get_phys_port_id(dev, ppid); 9660 } 9661 9662 /** 9663 * dev_get_phys_port_name - Get device physical port name 9664 * @dev: device 9665 * @name: port name 9666 * @len: limit of bytes to copy to name 9667 * 9668 * Get device physical port name 9669 */ 9670 int dev_get_phys_port_name(struct net_device *dev, 9671 char *name, size_t len) 9672 { 9673 const struct net_device_ops *ops = dev->netdev_ops; 9674 int err; 9675 9676 if (ops->ndo_get_phys_port_name) { 9677 err = ops->ndo_get_phys_port_name(dev, name, len); 9678 if (err != -EOPNOTSUPP) 9679 return err; 9680 } 9681 return devlink_compat_phys_port_name_get(dev, name, len); 9682 } 9683 9684 /** 9685 * dev_get_port_parent_id - Get the device's port parent identifier 9686 * @dev: network device 9687 * @ppid: pointer to a storage for the port's parent identifier 9688 * @recurse: allow/disallow recursion to lower devices 9689 * 9690 * Get the devices's port parent identifier 9691 */ 9692 int dev_get_port_parent_id(struct net_device *dev, 9693 struct netdev_phys_item_id *ppid, 9694 bool recurse) 9695 { 9696 const struct net_device_ops *ops = dev->netdev_ops; 9697 struct netdev_phys_item_id first = { }; 9698 struct net_device *lower_dev; 9699 struct list_head *iter; 9700 int err; 9701 9702 if (ops->ndo_get_port_parent_id) { 9703 err = ops->ndo_get_port_parent_id(dev, ppid); 9704 if (err != -EOPNOTSUPP) 9705 return err; 9706 } 9707 9708 err = devlink_compat_switch_id_get(dev, ppid); 9709 if (!recurse || err != -EOPNOTSUPP) 9710 return err; 9711 9712 netdev_for_each_lower_dev(dev, lower_dev, iter) { 9713 err = dev_get_port_parent_id(lower_dev, ppid, true); 9714 if (err) 9715 break; 9716 if (!first.id_len) 9717 first = *ppid; 9718 else if (memcmp(&first, ppid, sizeof(*ppid))) 9719 return -EOPNOTSUPP; 9720 } 9721 9722 return err; 9723 } 9724 EXPORT_SYMBOL(dev_get_port_parent_id); 9725 9726 /** 9727 * netdev_port_same_parent_id - Indicate if two network devices have 9728 * the same port parent identifier 9729 * @a: first network device 9730 * @b: second network device 9731 */ 9732 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b) 9733 { 9734 struct netdev_phys_item_id a_id = { }; 9735 struct netdev_phys_item_id b_id = { }; 9736 9737 if (dev_get_port_parent_id(a, &a_id, true) || 9738 dev_get_port_parent_id(b, &b_id, true)) 9739 return false; 9740 9741 return netdev_phys_item_id_same(&a_id, &b_id); 9742 } 9743 EXPORT_SYMBOL(netdev_port_same_parent_id); 9744 9745 int netif_change_proto_down(struct net_device *dev, bool proto_down) 9746 { 9747 if (!dev->change_proto_down) 9748 return -EOPNOTSUPP; 9749 if (!netif_device_present(dev)) 9750 return -ENODEV; 9751 if (proto_down) 9752 netif_carrier_off(dev); 9753 else 9754 netif_carrier_on(dev); 9755 WRITE_ONCE(dev->proto_down, proto_down); 9756 return 0; 9757 } 9758 9759 /** 9760 * netdev_change_proto_down_reason_locked - proto down reason 9761 * 9762 * @dev: device 9763 * @mask: proto down mask 9764 * @value: proto down value 9765 */ 9766 void netdev_change_proto_down_reason_locked(struct net_device *dev, 9767 unsigned long mask, u32 value) 9768 { 9769 u32 proto_down_reason; 9770 int b; 9771 9772 if (!mask) { 9773 proto_down_reason = value; 9774 } else { 9775 proto_down_reason = dev->proto_down_reason; 9776 for_each_set_bit(b, &mask, 32) { 9777 if (value & (1 << b)) 9778 proto_down_reason |= BIT(b); 9779 else 9780 proto_down_reason &= ~BIT(b); 9781 } 9782 } 9783 WRITE_ONCE(dev->proto_down_reason, proto_down_reason); 9784 } 9785 9786 struct bpf_xdp_link { 9787 struct bpf_link link; 9788 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */ 9789 int flags; 9790 }; 9791 9792 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags) 9793 { 9794 if (flags & XDP_FLAGS_HW_MODE) 9795 return XDP_MODE_HW; 9796 if (flags & XDP_FLAGS_DRV_MODE) 9797 return XDP_MODE_DRV; 9798 if (flags & XDP_FLAGS_SKB_MODE) 9799 return XDP_MODE_SKB; 9800 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB; 9801 } 9802 9803 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode) 9804 { 9805 switch (mode) { 9806 case XDP_MODE_SKB: 9807 return generic_xdp_install; 9808 case XDP_MODE_DRV: 9809 case XDP_MODE_HW: 9810 return dev->netdev_ops->ndo_bpf; 9811 default: 9812 return NULL; 9813 } 9814 } 9815 9816 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev, 9817 enum bpf_xdp_mode mode) 9818 { 9819 return dev->xdp_state[mode].link; 9820 } 9821 9822 static struct bpf_prog *dev_xdp_prog(struct net_device *dev, 9823 enum bpf_xdp_mode mode) 9824 { 9825 struct bpf_xdp_link *link = dev_xdp_link(dev, mode); 9826 9827 if (link) 9828 return link->link.prog; 9829 return dev->xdp_state[mode].prog; 9830 } 9831 9832 u8 dev_xdp_prog_count(struct net_device *dev) 9833 { 9834 u8 count = 0; 9835 int i; 9836 9837 for (i = 0; i < __MAX_XDP_MODE; i++) 9838 if (dev->xdp_state[i].prog || dev->xdp_state[i].link) 9839 count++; 9840 return count; 9841 } 9842 EXPORT_SYMBOL_GPL(dev_xdp_prog_count); 9843 9844 u8 dev_xdp_sb_prog_count(struct net_device *dev) 9845 { 9846 u8 count = 0; 9847 int i; 9848 9849 for (i = 0; i < __MAX_XDP_MODE; i++) 9850 if (dev->xdp_state[i].prog && 9851 !dev->xdp_state[i].prog->aux->xdp_has_frags) 9852 count++; 9853 return count; 9854 } 9855 9856 int netif_xdp_propagate(struct net_device *dev, struct netdev_bpf *bpf) 9857 { 9858 if (!dev->netdev_ops->ndo_bpf) 9859 return -EOPNOTSUPP; 9860 9861 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 9862 bpf->command == XDP_SETUP_PROG && 9863 bpf->prog && !bpf->prog->aux->xdp_has_frags) { 9864 NL_SET_ERR_MSG(bpf->extack, 9865 "unable to propagate XDP to device using tcp-data-split"); 9866 return -EBUSY; 9867 } 9868 9869 if (dev_get_min_mp_channel_count(dev)) { 9870 NL_SET_ERR_MSG(bpf->extack, "unable to propagate XDP to device using memory provider"); 9871 return -EBUSY; 9872 } 9873 9874 return dev->netdev_ops->ndo_bpf(dev, bpf); 9875 } 9876 9877 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode) 9878 { 9879 struct bpf_prog *prog = dev_xdp_prog(dev, mode); 9880 9881 return prog ? prog->aux->id : 0; 9882 } 9883 9884 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode, 9885 struct bpf_xdp_link *link) 9886 { 9887 dev->xdp_state[mode].link = link; 9888 dev->xdp_state[mode].prog = NULL; 9889 } 9890 9891 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode, 9892 struct bpf_prog *prog) 9893 { 9894 dev->xdp_state[mode].link = NULL; 9895 dev->xdp_state[mode].prog = prog; 9896 } 9897 9898 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode, 9899 bpf_op_t bpf_op, struct netlink_ext_ack *extack, 9900 u32 flags, struct bpf_prog *prog) 9901 { 9902 struct netdev_bpf xdp; 9903 int err; 9904 9905 netdev_ops_assert_locked(dev); 9906 9907 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 9908 prog && !prog->aux->xdp_has_frags) { 9909 NL_SET_ERR_MSG(extack, "unable to install XDP to device using tcp-data-split"); 9910 return -EBUSY; 9911 } 9912 9913 if (dev_get_min_mp_channel_count(dev)) { 9914 NL_SET_ERR_MSG(extack, "unable to install XDP to device using memory provider"); 9915 return -EBUSY; 9916 } 9917 9918 memset(&xdp, 0, sizeof(xdp)); 9919 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG; 9920 xdp.extack = extack; 9921 xdp.flags = flags; 9922 xdp.prog = prog; 9923 9924 /* Drivers assume refcnt is already incremented (i.e, prog pointer is 9925 * "moved" into driver), so they don't increment it on their own, but 9926 * they do decrement refcnt when program is detached or replaced. 9927 * Given net_device also owns link/prog, we need to bump refcnt here 9928 * to prevent drivers from underflowing it. 9929 */ 9930 if (prog) 9931 bpf_prog_inc(prog); 9932 err = bpf_op(dev, &xdp); 9933 if (err) { 9934 if (prog) 9935 bpf_prog_put(prog); 9936 return err; 9937 } 9938 9939 if (mode != XDP_MODE_HW) 9940 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog); 9941 9942 return 0; 9943 } 9944 9945 static void dev_xdp_uninstall(struct net_device *dev) 9946 { 9947 struct bpf_xdp_link *link; 9948 struct bpf_prog *prog; 9949 enum bpf_xdp_mode mode; 9950 bpf_op_t bpf_op; 9951 9952 ASSERT_RTNL(); 9953 9954 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) { 9955 prog = dev_xdp_prog(dev, mode); 9956 if (!prog) 9957 continue; 9958 9959 bpf_op = dev_xdp_bpf_op(dev, mode); 9960 if (!bpf_op) 9961 continue; 9962 9963 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 9964 9965 /* auto-detach link from net device */ 9966 link = dev_xdp_link(dev, mode); 9967 if (link) 9968 link->dev = NULL; 9969 else 9970 bpf_prog_put(prog); 9971 9972 dev_xdp_set_link(dev, mode, NULL); 9973 } 9974 } 9975 9976 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack, 9977 struct bpf_xdp_link *link, struct bpf_prog *new_prog, 9978 struct bpf_prog *old_prog, u32 flags) 9979 { 9980 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES); 9981 struct bpf_prog *cur_prog; 9982 struct net_device *upper; 9983 struct list_head *iter; 9984 enum bpf_xdp_mode mode; 9985 bpf_op_t bpf_op; 9986 int err; 9987 9988 ASSERT_RTNL(); 9989 9990 /* either link or prog attachment, never both */ 9991 if (link && (new_prog || old_prog)) 9992 return -EINVAL; 9993 /* link supports only XDP mode flags */ 9994 if (link && (flags & ~XDP_FLAGS_MODES)) { 9995 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment"); 9996 return -EINVAL; 9997 } 9998 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */ 9999 if (num_modes > 1) { 10000 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set"); 10001 return -EINVAL; 10002 } 10003 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */ 10004 if (!num_modes && dev_xdp_prog_count(dev) > 1) { 10005 NL_SET_ERR_MSG(extack, 10006 "More than one program loaded, unset mode is ambiguous"); 10007 return -EINVAL; 10008 } 10009 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */ 10010 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) { 10011 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified"); 10012 return -EINVAL; 10013 } 10014 10015 mode = dev_xdp_mode(dev, flags); 10016 /* can't replace attached link */ 10017 if (dev_xdp_link(dev, mode)) { 10018 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link"); 10019 return -EBUSY; 10020 } 10021 10022 /* don't allow if an upper device already has a program */ 10023 netdev_for_each_upper_dev_rcu(dev, upper, iter) { 10024 if (dev_xdp_prog_count(upper) > 0) { 10025 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program"); 10026 return -EEXIST; 10027 } 10028 } 10029 10030 cur_prog = dev_xdp_prog(dev, mode); 10031 /* can't replace attached prog with link */ 10032 if (link && cur_prog) { 10033 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link"); 10034 return -EBUSY; 10035 } 10036 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) { 10037 NL_SET_ERR_MSG(extack, "Active program does not match expected"); 10038 return -EEXIST; 10039 } 10040 10041 /* put effective new program into new_prog */ 10042 if (link) 10043 new_prog = link->link.prog; 10044 10045 if (new_prog) { 10046 bool offload = mode == XDP_MODE_HW; 10047 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB 10048 ? XDP_MODE_DRV : XDP_MODE_SKB; 10049 10050 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) { 10051 NL_SET_ERR_MSG(extack, "XDP program already attached"); 10052 return -EBUSY; 10053 } 10054 if (!offload && dev_xdp_prog(dev, other_mode)) { 10055 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time"); 10056 return -EEXIST; 10057 } 10058 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) { 10059 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported"); 10060 return -EINVAL; 10061 } 10062 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) { 10063 NL_SET_ERR_MSG(extack, "Program bound to different device"); 10064 return -EINVAL; 10065 } 10066 if (bpf_prog_is_dev_bound(new_prog->aux) && mode == XDP_MODE_SKB) { 10067 NL_SET_ERR_MSG(extack, "Can't attach device-bound programs in generic mode"); 10068 return -EINVAL; 10069 } 10070 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) { 10071 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device"); 10072 return -EINVAL; 10073 } 10074 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) { 10075 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device"); 10076 return -EINVAL; 10077 } 10078 } 10079 10080 /* don't call drivers if the effective program didn't change */ 10081 if (new_prog != cur_prog) { 10082 bpf_op = dev_xdp_bpf_op(dev, mode); 10083 if (!bpf_op) { 10084 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode"); 10085 return -EOPNOTSUPP; 10086 } 10087 10088 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog); 10089 if (err) 10090 return err; 10091 } 10092 10093 if (link) 10094 dev_xdp_set_link(dev, mode, link); 10095 else 10096 dev_xdp_set_prog(dev, mode, new_prog); 10097 if (cur_prog) 10098 bpf_prog_put(cur_prog); 10099 10100 return 0; 10101 } 10102 10103 static int dev_xdp_attach_link(struct net_device *dev, 10104 struct netlink_ext_ack *extack, 10105 struct bpf_xdp_link *link) 10106 { 10107 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags); 10108 } 10109 10110 static int dev_xdp_detach_link(struct net_device *dev, 10111 struct netlink_ext_ack *extack, 10112 struct bpf_xdp_link *link) 10113 { 10114 enum bpf_xdp_mode mode; 10115 bpf_op_t bpf_op; 10116 10117 ASSERT_RTNL(); 10118 10119 mode = dev_xdp_mode(dev, link->flags); 10120 if (dev_xdp_link(dev, mode) != link) 10121 return -EINVAL; 10122 10123 bpf_op = dev_xdp_bpf_op(dev, mode); 10124 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 10125 dev_xdp_set_link(dev, mode, NULL); 10126 return 0; 10127 } 10128 10129 static void bpf_xdp_link_release(struct bpf_link *link) 10130 { 10131 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10132 10133 rtnl_lock(); 10134 10135 /* if racing with net_device's tear down, xdp_link->dev might be 10136 * already NULL, in which case link was already auto-detached 10137 */ 10138 if (xdp_link->dev) { 10139 netdev_lock_ops(xdp_link->dev); 10140 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link)); 10141 netdev_unlock_ops(xdp_link->dev); 10142 xdp_link->dev = NULL; 10143 } 10144 10145 rtnl_unlock(); 10146 } 10147 10148 static int bpf_xdp_link_detach(struct bpf_link *link) 10149 { 10150 bpf_xdp_link_release(link); 10151 return 0; 10152 } 10153 10154 static void bpf_xdp_link_dealloc(struct bpf_link *link) 10155 { 10156 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10157 10158 kfree(xdp_link); 10159 } 10160 10161 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link, 10162 struct seq_file *seq) 10163 { 10164 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10165 u32 ifindex = 0; 10166 10167 rtnl_lock(); 10168 if (xdp_link->dev) 10169 ifindex = xdp_link->dev->ifindex; 10170 rtnl_unlock(); 10171 10172 seq_printf(seq, "ifindex:\t%u\n", ifindex); 10173 } 10174 10175 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link, 10176 struct bpf_link_info *info) 10177 { 10178 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10179 u32 ifindex = 0; 10180 10181 rtnl_lock(); 10182 if (xdp_link->dev) 10183 ifindex = xdp_link->dev->ifindex; 10184 rtnl_unlock(); 10185 10186 info->xdp.ifindex = ifindex; 10187 return 0; 10188 } 10189 10190 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog, 10191 struct bpf_prog *old_prog) 10192 { 10193 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10194 enum bpf_xdp_mode mode; 10195 bpf_op_t bpf_op; 10196 int err = 0; 10197 10198 rtnl_lock(); 10199 10200 /* link might have been auto-released already, so fail */ 10201 if (!xdp_link->dev) { 10202 err = -ENOLINK; 10203 goto out_unlock; 10204 } 10205 10206 if (old_prog && link->prog != old_prog) { 10207 err = -EPERM; 10208 goto out_unlock; 10209 } 10210 old_prog = link->prog; 10211 if (old_prog->type != new_prog->type || 10212 old_prog->expected_attach_type != new_prog->expected_attach_type) { 10213 err = -EINVAL; 10214 goto out_unlock; 10215 } 10216 10217 if (old_prog == new_prog) { 10218 /* no-op, don't disturb drivers */ 10219 bpf_prog_put(new_prog); 10220 goto out_unlock; 10221 } 10222 10223 netdev_lock_ops(xdp_link->dev); 10224 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags); 10225 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode); 10226 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL, 10227 xdp_link->flags, new_prog); 10228 netdev_unlock_ops(xdp_link->dev); 10229 if (err) 10230 goto out_unlock; 10231 10232 old_prog = xchg(&link->prog, new_prog); 10233 bpf_prog_put(old_prog); 10234 10235 out_unlock: 10236 rtnl_unlock(); 10237 return err; 10238 } 10239 10240 static const struct bpf_link_ops bpf_xdp_link_lops = { 10241 .release = bpf_xdp_link_release, 10242 .dealloc = bpf_xdp_link_dealloc, 10243 .detach = bpf_xdp_link_detach, 10244 .show_fdinfo = bpf_xdp_link_show_fdinfo, 10245 .fill_link_info = bpf_xdp_link_fill_link_info, 10246 .update_prog = bpf_xdp_link_update, 10247 }; 10248 10249 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog) 10250 { 10251 struct net *net = current->nsproxy->net_ns; 10252 struct bpf_link_primer link_primer; 10253 struct netlink_ext_ack extack = {}; 10254 struct bpf_xdp_link *link; 10255 struct net_device *dev; 10256 int err, fd; 10257 10258 rtnl_lock(); 10259 dev = dev_get_by_index(net, attr->link_create.target_ifindex); 10260 if (!dev) { 10261 rtnl_unlock(); 10262 return -EINVAL; 10263 } 10264 10265 link = kzalloc(sizeof(*link), GFP_USER); 10266 if (!link) { 10267 err = -ENOMEM; 10268 goto unlock; 10269 } 10270 10271 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog); 10272 link->dev = dev; 10273 link->flags = attr->link_create.flags; 10274 10275 err = bpf_link_prime(&link->link, &link_primer); 10276 if (err) { 10277 kfree(link); 10278 goto unlock; 10279 } 10280 10281 netdev_lock_ops(dev); 10282 err = dev_xdp_attach_link(dev, &extack, link); 10283 netdev_unlock_ops(dev); 10284 rtnl_unlock(); 10285 10286 if (err) { 10287 link->dev = NULL; 10288 bpf_link_cleanup(&link_primer); 10289 trace_bpf_xdp_link_attach_failed(extack._msg); 10290 goto out_put_dev; 10291 } 10292 10293 fd = bpf_link_settle(&link_primer); 10294 /* link itself doesn't hold dev's refcnt to not complicate shutdown */ 10295 dev_put(dev); 10296 return fd; 10297 10298 unlock: 10299 rtnl_unlock(); 10300 10301 out_put_dev: 10302 dev_put(dev); 10303 return err; 10304 } 10305 10306 /** 10307 * dev_change_xdp_fd - set or clear a bpf program for a device rx path 10308 * @dev: device 10309 * @extack: netlink extended ack 10310 * @fd: new program fd or negative value to clear 10311 * @expected_fd: old program fd that userspace expects to replace or clear 10312 * @flags: xdp-related flags 10313 * 10314 * Set or clear a bpf program for a device 10315 */ 10316 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, 10317 int fd, int expected_fd, u32 flags) 10318 { 10319 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags); 10320 struct bpf_prog *new_prog = NULL, *old_prog = NULL; 10321 int err; 10322 10323 ASSERT_RTNL(); 10324 10325 if (fd >= 0) { 10326 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP, 10327 mode != XDP_MODE_SKB); 10328 if (IS_ERR(new_prog)) 10329 return PTR_ERR(new_prog); 10330 } 10331 10332 if (expected_fd >= 0) { 10333 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP, 10334 mode != XDP_MODE_SKB); 10335 if (IS_ERR(old_prog)) { 10336 err = PTR_ERR(old_prog); 10337 old_prog = NULL; 10338 goto err_out; 10339 } 10340 } 10341 10342 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags); 10343 10344 err_out: 10345 if (err && new_prog) 10346 bpf_prog_put(new_prog); 10347 if (old_prog) 10348 bpf_prog_put(old_prog); 10349 return err; 10350 } 10351 10352 u32 dev_get_min_mp_channel_count(const struct net_device *dev) 10353 { 10354 int i; 10355 10356 netdev_ops_assert_locked(dev); 10357 10358 for (i = dev->real_num_rx_queues - 1; i >= 0; i--) 10359 if (dev->_rx[i].mp_params.mp_priv) 10360 /* The channel count is the idx plus 1. */ 10361 return i + 1; 10362 10363 return 0; 10364 } 10365 10366 /** 10367 * dev_index_reserve() - allocate an ifindex in a namespace 10368 * @net: the applicable net namespace 10369 * @ifindex: requested ifindex, pass %0 to get one allocated 10370 * 10371 * Allocate a ifindex for a new device. Caller must either use the ifindex 10372 * to store the device (via list_netdevice()) or call dev_index_release() 10373 * to give the index up. 10374 * 10375 * Return: a suitable unique value for a new device interface number or -errno. 10376 */ 10377 static int dev_index_reserve(struct net *net, u32 ifindex) 10378 { 10379 int err; 10380 10381 if (ifindex > INT_MAX) { 10382 DEBUG_NET_WARN_ON_ONCE(1); 10383 return -EINVAL; 10384 } 10385 10386 if (!ifindex) 10387 err = xa_alloc_cyclic(&net->dev_by_index, &ifindex, NULL, 10388 xa_limit_31b, &net->ifindex, GFP_KERNEL); 10389 else 10390 err = xa_insert(&net->dev_by_index, ifindex, NULL, GFP_KERNEL); 10391 if (err < 0) 10392 return err; 10393 10394 return ifindex; 10395 } 10396 10397 static void dev_index_release(struct net *net, int ifindex) 10398 { 10399 /* Expect only unused indexes, unlist_netdevice() removes the used */ 10400 WARN_ON(xa_erase(&net->dev_by_index, ifindex)); 10401 } 10402 10403 static bool from_cleanup_net(void) 10404 { 10405 #ifdef CONFIG_NET_NS 10406 return current == cleanup_net_task; 10407 #else 10408 return false; 10409 #endif 10410 } 10411 10412 /* Delayed registration/unregisteration */ 10413 LIST_HEAD(net_todo_list); 10414 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq); 10415 atomic_t dev_unreg_count = ATOMIC_INIT(0); 10416 10417 static void net_set_todo(struct net_device *dev) 10418 { 10419 list_add_tail(&dev->todo_list, &net_todo_list); 10420 } 10421 10422 static netdev_features_t netdev_sync_upper_features(struct net_device *lower, 10423 struct net_device *upper, netdev_features_t features) 10424 { 10425 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10426 netdev_features_t feature; 10427 int feature_bit; 10428 10429 for_each_netdev_feature(upper_disables, feature_bit) { 10430 feature = __NETIF_F_BIT(feature_bit); 10431 if (!(upper->wanted_features & feature) 10432 && (features & feature)) { 10433 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n", 10434 &feature, upper->name); 10435 features &= ~feature; 10436 } 10437 } 10438 10439 return features; 10440 } 10441 10442 static void netdev_sync_lower_features(struct net_device *upper, 10443 struct net_device *lower, netdev_features_t features) 10444 { 10445 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10446 netdev_features_t feature; 10447 int feature_bit; 10448 10449 for_each_netdev_feature(upper_disables, feature_bit) { 10450 feature = __NETIF_F_BIT(feature_bit); 10451 if (!(features & feature) && (lower->features & feature)) { 10452 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n", 10453 &feature, lower->name); 10454 lower->wanted_features &= ~feature; 10455 __netdev_update_features(lower); 10456 10457 if (unlikely(lower->features & feature)) 10458 netdev_WARN(upper, "failed to disable %pNF on %s!\n", 10459 &feature, lower->name); 10460 else 10461 netdev_features_change(lower); 10462 } 10463 } 10464 } 10465 10466 static bool netdev_has_ip_or_hw_csum(netdev_features_t features) 10467 { 10468 netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 10469 bool ip_csum = (features & ip_csum_mask) == ip_csum_mask; 10470 bool hw_csum = features & NETIF_F_HW_CSUM; 10471 10472 return ip_csum || hw_csum; 10473 } 10474 10475 static netdev_features_t netdev_fix_features(struct net_device *dev, 10476 netdev_features_t features) 10477 { 10478 /* Fix illegal checksum combinations */ 10479 if ((features & NETIF_F_HW_CSUM) && 10480 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) { 10481 netdev_warn(dev, "mixed HW and IP checksum settings.\n"); 10482 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM); 10483 } 10484 10485 /* TSO requires that SG is present as well. */ 10486 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) { 10487 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n"); 10488 features &= ~NETIF_F_ALL_TSO; 10489 } 10490 10491 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) && 10492 !(features & NETIF_F_IP_CSUM)) { 10493 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n"); 10494 features &= ~NETIF_F_TSO; 10495 features &= ~NETIF_F_TSO_ECN; 10496 } 10497 10498 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) && 10499 !(features & NETIF_F_IPV6_CSUM)) { 10500 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n"); 10501 features &= ~NETIF_F_TSO6; 10502 } 10503 10504 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */ 10505 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO)) 10506 features &= ~NETIF_F_TSO_MANGLEID; 10507 10508 /* TSO ECN requires that TSO is present as well. */ 10509 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN) 10510 features &= ~NETIF_F_TSO_ECN; 10511 10512 /* Software GSO depends on SG. */ 10513 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) { 10514 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n"); 10515 features &= ~NETIF_F_GSO; 10516 } 10517 10518 /* GSO partial features require GSO partial be set */ 10519 if ((features & dev->gso_partial_features) && 10520 !(features & NETIF_F_GSO_PARTIAL)) { 10521 netdev_dbg(dev, 10522 "Dropping partially supported GSO features since no GSO partial.\n"); 10523 features &= ~dev->gso_partial_features; 10524 } 10525 10526 if (!(features & NETIF_F_RXCSUM)) { 10527 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet 10528 * successfully merged by hardware must also have the 10529 * checksum verified by hardware. If the user does not 10530 * want to enable RXCSUM, logically, we should disable GRO_HW. 10531 */ 10532 if (features & NETIF_F_GRO_HW) { 10533 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n"); 10534 features &= ~NETIF_F_GRO_HW; 10535 } 10536 } 10537 10538 /* LRO/HW-GRO features cannot be combined with RX-FCS */ 10539 if (features & NETIF_F_RXFCS) { 10540 if (features & NETIF_F_LRO) { 10541 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n"); 10542 features &= ~NETIF_F_LRO; 10543 } 10544 10545 if (features & NETIF_F_GRO_HW) { 10546 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n"); 10547 features &= ~NETIF_F_GRO_HW; 10548 } 10549 } 10550 10551 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) { 10552 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n"); 10553 features &= ~NETIF_F_LRO; 10554 } 10555 10556 if ((features & NETIF_F_HW_TLS_TX) && !netdev_has_ip_or_hw_csum(features)) { 10557 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n"); 10558 features &= ~NETIF_F_HW_TLS_TX; 10559 } 10560 10561 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) { 10562 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n"); 10563 features &= ~NETIF_F_HW_TLS_RX; 10564 } 10565 10566 if ((features & NETIF_F_GSO_UDP_L4) && !netdev_has_ip_or_hw_csum(features)) { 10567 netdev_dbg(dev, "Dropping USO feature since no CSUM feature.\n"); 10568 features &= ~NETIF_F_GSO_UDP_L4; 10569 } 10570 10571 return features; 10572 } 10573 10574 int __netdev_update_features(struct net_device *dev) 10575 { 10576 struct net_device *upper, *lower; 10577 netdev_features_t features; 10578 struct list_head *iter; 10579 int err = -1; 10580 10581 ASSERT_RTNL(); 10582 netdev_ops_assert_locked(dev); 10583 10584 features = netdev_get_wanted_features(dev); 10585 10586 if (dev->netdev_ops->ndo_fix_features) 10587 features = dev->netdev_ops->ndo_fix_features(dev, features); 10588 10589 /* driver might be less strict about feature dependencies */ 10590 features = netdev_fix_features(dev, features); 10591 10592 /* some features can't be enabled if they're off on an upper device */ 10593 netdev_for_each_upper_dev_rcu(dev, upper, iter) 10594 features = netdev_sync_upper_features(dev, upper, features); 10595 10596 if (dev->features == features) 10597 goto sync_lower; 10598 10599 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n", 10600 &dev->features, &features); 10601 10602 if (dev->netdev_ops->ndo_set_features) 10603 err = dev->netdev_ops->ndo_set_features(dev, features); 10604 else 10605 err = 0; 10606 10607 if (unlikely(err < 0)) { 10608 netdev_err(dev, 10609 "set_features() failed (%d); wanted %pNF, left %pNF\n", 10610 err, &features, &dev->features); 10611 /* return non-0 since some features might have changed and 10612 * it's better to fire a spurious notification than miss it 10613 */ 10614 return -1; 10615 } 10616 10617 sync_lower: 10618 /* some features must be disabled on lower devices when disabled 10619 * on an upper device (think: bonding master or bridge) 10620 */ 10621 netdev_for_each_lower_dev(dev, lower, iter) 10622 netdev_sync_lower_features(dev, lower, features); 10623 10624 if (!err) { 10625 netdev_features_t diff = features ^ dev->features; 10626 10627 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) { 10628 /* udp_tunnel_{get,drop}_rx_info both need 10629 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the 10630 * device, or they won't do anything. 10631 * Thus we need to update dev->features 10632 * *before* calling udp_tunnel_get_rx_info, 10633 * but *after* calling udp_tunnel_drop_rx_info. 10634 */ 10635 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) { 10636 dev->features = features; 10637 udp_tunnel_get_rx_info(dev); 10638 } else { 10639 udp_tunnel_drop_rx_info(dev); 10640 } 10641 } 10642 10643 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) { 10644 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) { 10645 dev->features = features; 10646 err |= vlan_get_rx_ctag_filter_info(dev); 10647 } else { 10648 vlan_drop_rx_ctag_filter_info(dev); 10649 } 10650 } 10651 10652 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) { 10653 if (features & NETIF_F_HW_VLAN_STAG_FILTER) { 10654 dev->features = features; 10655 err |= vlan_get_rx_stag_filter_info(dev); 10656 } else { 10657 vlan_drop_rx_stag_filter_info(dev); 10658 } 10659 } 10660 10661 dev->features = features; 10662 } 10663 10664 return err < 0 ? 0 : 1; 10665 } 10666 10667 /** 10668 * netdev_update_features - recalculate device features 10669 * @dev: the device to check 10670 * 10671 * Recalculate dev->features set and send notifications if it 10672 * has changed. Should be called after driver or hardware dependent 10673 * conditions might have changed that influence the features. 10674 */ 10675 void netdev_update_features(struct net_device *dev) 10676 { 10677 if (__netdev_update_features(dev)) 10678 netdev_features_change(dev); 10679 } 10680 EXPORT_SYMBOL(netdev_update_features); 10681 10682 /** 10683 * netdev_change_features - recalculate device features 10684 * @dev: the device to check 10685 * 10686 * Recalculate dev->features set and send notifications even 10687 * if they have not changed. Should be called instead of 10688 * netdev_update_features() if also dev->vlan_features might 10689 * have changed to allow the changes to be propagated to stacked 10690 * VLAN devices. 10691 */ 10692 void netdev_change_features(struct net_device *dev) 10693 { 10694 __netdev_update_features(dev); 10695 netdev_features_change(dev); 10696 } 10697 EXPORT_SYMBOL(netdev_change_features); 10698 10699 /** 10700 * netif_stacked_transfer_operstate - transfer operstate 10701 * @rootdev: the root or lower level device to transfer state from 10702 * @dev: the device to transfer operstate to 10703 * 10704 * Transfer operational state from root to device. This is normally 10705 * called when a stacking relationship exists between the root 10706 * device and the device(a leaf device). 10707 */ 10708 void netif_stacked_transfer_operstate(const struct net_device *rootdev, 10709 struct net_device *dev) 10710 { 10711 if (rootdev->operstate == IF_OPER_DORMANT) 10712 netif_dormant_on(dev); 10713 else 10714 netif_dormant_off(dev); 10715 10716 if (rootdev->operstate == IF_OPER_TESTING) 10717 netif_testing_on(dev); 10718 else 10719 netif_testing_off(dev); 10720 10721 if (netif_carrier_ok(rootdev)) 10722 netif_carrier_on(dev); 10723 else 10724 netif_carrier_off(dev); 10725 } 10726 EXPORT_SYMBOL(netif_stacked_transfer_operstate); 10727 10728 static int netif_alloc_rx_queues(struct net_device *dev) 10729 { 10730 unsigned int i, count = dev->num_rx_queues; 10731 struct netdev_rx_queue *rx; 10732 size_t sz = count * sizeof(*rx); 10733 int err = 0; 10734 10735 BUG_ON(count < 1); 10736 10737 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10738 if (!rx) 10739 return -ENOMEM; 10740 10741 dev->_rx = rx; 10742 10743 for (i = 0; i < count; i++) { 10744 rx[i].dev = dev; 10745 10746 /* XDP RX-queue setup */ 10747 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0); 10748 if (err < 0) 10749 goto err_rxq_info; 10750 } 10751 return 0; 10752 10753 err_rxq_info: 10754 /* Rollback successful reg's and free other resources */ 10755 while (i--) 10756 xdp_rxq_info_unreg(&rx[i].xdp_rxq); 10757 kvfree(dev->_rx); 10758 dev->_rx = NULL; 10759 return err; 10760 } 10761 10762 static void netif_free_rx_queues(struct net_device *dev) 10763 { 10764 unsigned int i, count = dev->num_rx_queues; 10765 10766 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */ 10767 if (!dev->_rx) 10768 return; 10769 10770 for (i = 0; i < count; i++) 10771 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq); 10772 10773 kvfree(dev->_rx); 10774 } 10775 10776 static void netdev_init_one_queue(struct net_device *dev, 10777 struct netdev_queue *queue, void *_unused) 10778 { 10779 /* Initialize queue lock */ 10780 spin_lock_init(&queue->_xmit_lock); 10781 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type); 10782 queue->xmit_lock_owner = -1; 10783 netdev_queue_numa_node_write(queue, NUMA_NO_NODE); 10784 queue->dev = dev; 10785 #ifdef CONFIG_BQL 10786 dql_init(&queue->dql, HZ); 10787 #endif 10788 } 10789 10790 static void netif_free_tx_queues(struct net_device *dev) 10791 { 10792 kvfree(dev->_tx); 10793 } 10794 10795 static int netif_alloc_netdev_queues(struct net_device *dev) 10796 { 10797 unsigned int count = dev->num_tx_queues; 10798 struct netdev_queue *tx; 10799 size_t sz = count * sizeof(*tx); 10800 10801 if (count < 1 || count > 0xffff) 10802 return -EINVAL; 10803 10804 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10805 if (!tx) 10806 return -ENOMEM; 10807 10808 dev->_tx = tx; 10809 10810 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL); 10811 spin_lock_init(&dev->tx_global_lock); 10812 10813 return 0; 10814 } 10815 10816 void netif_tx_stop_all_queues(struct net_device *dev) 10817 { 10818 unsigned int i; 10819 10820 for (i = 0; i < dev->num_tx_queues; i++) { 10821 struct netdev_queue *txq = netdev_get_tx_queue(dev, i); 10822 10823 netif_tx_stop_queue(txq); 10824 } 10825 } 10826 EXPORT_SYMBOL(netif_tx_stop_all_queues); 10827 10828 static int netdev_do_alloc_pcpu_stats(struct net_device *dev) 10829 { 10830 void __percpu *v; 10831 10832 /* Drivers implementing ndo_get_peer_dev must support tstat 10833 * accounting, so that skb_do_redirect() can bump the dev's 10834 * RX stats upon network namespace switch. 10835 */ 10836 if (dev->netdev_ops->ndo_get_peer_dev && 10837 dev->pcpu_stat_type != NETDEV_PCPU_STAT_TSTATS) 10838 return -EOPNOTSUPP; 10839 10840 switch (dev->pcpu_stat_type) { 10841 case NETDEV_PCPU_STAT_NONE: 10842 return 0; 10843 case NETDEV_PCPU_STAT_LSTATS: 10844 v = dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats); 10845 break; 10846 case NETDEV_PCPU_STAT_TSTATS: 10847 v = dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats); 10848 break; 10849 case NETDEV_PCPU_STAT_DSTATS: 10850 v = dev->dstats = netdev_alloc_pcpu_stats(struct pcpu_dstats); 10851 break; 10852 default: 10853 return -EINVAL; 10854 } 10855 10856 return v ? 0 : -ENOMEM; 10857 } 10858 10859 static void netdev_do_free_pcpu_stats(struct net_device *dev) 10860 { 10861 switch (dev->pcpu_stat_type) { 10862 case NETDEV_PCPU_STAT_NONE: 10863 return; 10864 case NETDEV_PCPU_STAT_LSTATS: 10865 free_percpu(dev->lstats); 10866 break; 10867 case NETDEV_PCPU_STAT_TSTATS: 10868 free_percpu(dev->tstats); 10869 break; 10870 case NETDEV_PCPU_STAT_DSTATS: 10871 free_percpu(dev->dstats); 10872 break; 10873 } 10874 } 10875 10876 static void netdev_free_phy_link_topology(struct net_device *dev) 10877 { 10878 struct phy_link_topology *topo = dev->link_topo; 10879 10880 if (IS_ENABLED(CONFIG_PHYLIB) && topo) { 10881 xa_destroy(&topo->phys); 10882 kfree(topo); 10883 dev->link_topo = NULL; 10884 } 10885 } 10886 10887 /** 10888 * register_netdevice() - register a network device 10889 * @dev: device to register 10890 * 10891 * Take a prepared network device structure and make it externally accessible. 10892 * A %NETDEV_REGISTER message is sent to the netdev notifier chain. 10893 * Callers must hold the rtnl lock - you may want register_netdev() 10894 * instead of this. 10895 */ 10896 int register_netdevice(struct net_device *dev) 10897 { 10898 int ret; 10899 struct net *net = dev_net(dev); 10900 10901 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE < 10902 NETDEV_FEATURE_COUNT); 10903 BUG_ON(dev_boot_phase); 10904 ASSERT_RTNL(); 10905 10906 might_sleep(); 10907 10908 /* When net_device's are persistent, this will be fatal. */ 10909 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED); 10910 BUG_ON(!net); 10911 10912 ret = ethtool_check_ops(dev->ethtool_ops); 10913 if (ret) 10914 return ret; 10915 10916 /* rss ctx ID 0 is reserved for the default context, start from 1 */ 10917 xa_init_flags(&dev->ethtool->rss_ctx, XA_FLAGS_ALLOC1); 10918 mutex_init(&dev->ethtool->rss_lock); 10919 10920 spin_lock_init(&dev->addr_list_lock); 10921 netdev_set_addr_lockdep_class(dev); 10922 10923 ret = dev_get_valid_name(net, dev, dev->name); 10924 if (ret < 0) 10925 goto out; 10926 10927 ret = -ENOMEM; 10928 dev->name_node = netdev_name_node_head_alloc(dev); 10929 if (!dev->name_node) 10930 goto out; 10931 10932 /* Init, if this function is available */ 10933 if (dev->netdev_ops->ndo_init) { 10934 ret = dev->netdev_ops->ndo_init(dev); 10935 if (ret) { 10936 if (ret > 0) 10937 ret = -EIO; 10938 goto err_free_name; 10939 } 10940 } 10941 10942 if (((dev->hw_features | dev->features) & 10943 NETIF_F_HW_VLAN_CTAG_FILTER) && 10944 (!dev->netdev_ops->ndo_vlan_rx_add_vid || 10945 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) { 10946 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n"); 10947 ret = -EINVAL; 10948 goto err_uninit; 10949 } 10950 10951 ret = netdev_do_alloc_pcpu_stats(dev); 10952 if (ret) 10953 goto err_uninit; 10954 10955 ret = dev_index_reserve(net, dev->ifindex); 10956 if (ret < 0) 10957 goto err_free_pcpu; 10958 dev->ifindex = ret; 10959 10960 /* Transfer changeable features to wanted_features and enable 10961 * software offloads (GSO and GRO). 10962 */ 10963 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF); 10964 dev->features |= NETIF_F_SOFT_FEATURES; 10965 10966 if (dev->udp_tunnel_nic_info) { 10967 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10968 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10969 } 10970 10971 dev->wanted_features = dev->features & dev->hw_features; 10972 10973 if (!(dev->flags & IFF_LOOPBACK)) 10974 dev->hw_features |= NETIF_F_NOCACHE_COPY; 10975 10976 /* If IPv4 TCP segmentation offload is supported we should also 10977 * allow the device to enable segmenting the frame with the option 10978 * of ignoring a static IP ID value. This doesn't enable the 10979 * feature itself but allows the user to enable it later. 10980 */ 10981 if (dev->hw_features & NETIF_F_TSO) 10982 dev->hw_features |= NETIF_F_TSO_MANGLEID; 10983 if (dev->vlan_features & NETIF_F_TSO) 10984 dev->vlan_features |= NETIF_F_TSO_MANGLEID; 10985 if (dev->mpls_features & NETIF_F_TSO) 10986 dev->mpls_features |= NETIF_F_TSO_MANGLEID; 10987 if (dev->hw_enc_features & NETIF_F_TSO) 10988 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID; 10989 10990 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices. 10991 */ 10992 dev->vlan_features |= NETIF_F_HIGHDMA; 10993 10994 /* Make NETIF_F_SG inheritable to tunnel devices. 10995 */ 10996 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL; 10997 10998 /* Make NETIF_F_SG inheritable to MPLS. 10999 */ 11000 dev->mpls_features |= NETIF_F_SG; 11001 11002 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev); 11003 ret = notifier_to_errno(ret); 11004 if (ret) 11005 goto err_ifindex_release; 11006 11007 ret = netdev_register_kobject(dev); 11008 11009 netdev_lock(dev); 11010 WRITE_ONCE(dev->reg_state, ret ? NETREG_UNREGISTERED : NETREG_REGISTERED); 11011 netdev_unlock(dev); 11012 11013 if (ret) 11014 goto err_uninit_notify; 11015 11016 netdev_lock_ops(dev); 11017 __netdev_update_features(dev); 11018 netdev_unlock_ops(dev); 11019 11020 /* 11021 * Default initial state at registry is that the 11022 * device is present. 11023 */ 11024 11025 set_bit(__LINK_STATE_PRESENT, &dev->state); 11026 11027 linkwatch_init_dev(dev); 11028 11029 dev_init_scheduler(dev); 11030 11031 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL); 11032 list_netdevice(dev); 11033 11034 add_device_randomness(dev->dev_addr, dev->addr_len); 11035 11036 /* If the device has permanent device address, driver should 11037 * set dev_addr and also addr_assign_type should be set to 11038 * NET_ADDR_PERM (default value). 11039 */ 11040 if (dev->addr_assign_type == NET_ADDR_PERM) 11041 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len); 11042 11043 /* Notify protocols, that a new device appeared. */ 11044 netdev_lock_ops(dev); 11045 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev); 11046 netdev_unlock_ops(dev); 11047 ret = notifier_to_errno(ret); 11048 if (ret) { 11049 /* Expect explicit free_netdev() on failure */ 11050 dev->needs_free_netdev = false; 11051 unregister_netdevice_queue(dev, NULL); 11052 goto out; 11053 } 11054 /* 11055 * Prevent userspace races by waiting until the network 11056 * device is fully setup before sending notifications. 11057 */ 11058 if (!dev->rtnl_link_ops || 11059 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 11060 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 11061 11062 out: 11063 return ret; 11064 11065 err_uninit_notify: 11066 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11067 err_ifindex_release: 11068 dev_index_release(net, dev->ifindex); 11069 err_free_pcpu: 11070 netdev_do_free_pcpu_stats(dev); 11071 err_uninit: 11072 if (dev->netdev_ops->ndo_uninit) 11073 dev->netdev_ops->ndo_uninit(dev); 11074 if (dev->priv_destructor) 11075 dev->priv_destructor(dev); 11076 err_free_name: 11077 netdev_name_node_free(dev->name_node); 11078 goto out; 11079 } 11080 EXPORT_SYMBOL(register_netdevice); 11081 11082 /* Initialize the core of a dummy net device. 11083 * The setup steps dummy netdevs need which normal netdevs get by going 11084 * through register_netdevice(). 11085 */ 11086 static void init_dummy_netdev(struct net_device *dev) 11087 { 11088 /* make sure we BUG if trying to hit standard 11089 * register/unregister code path 11090 */ 11091 dev->reg_state = NETREG_DUMMY; 11092 11093 /* a dummy interface is started by default */ 11094 set_bit(__LINK_STATE_PRESENT, &dev->state); 11095 set_bit(__LINK_STATE_START, &dev->state); 11096 11097 /* Note : We dont allocate pcpu_refcnt for dummy devices, 11098 * because users of this 'device' dont need to change 11099 * its refcount. 11100 */ 11101 } 11102 11103 /** 11104 * register_netdev - register a network device 11105 * @dev: device to register 11106 * 11107 * Take a completed network device structure and add it to the kernel 11108 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier 11109 * chain. 0 is returned on success. A negative errno code is returned 11110 * on a failure to set up the device, or if the name is a duplicate. 11111 * 11112 * This is a wrapper around register_netdevice that takes the rtnl semaphore 11113 * and expands the device name if you passed a format string to 11114 * alloc_netdev. 11115 */ 11116 int register_netdev(struct net_device *dev) 11117 { 11118 struct net *net = dev_net(dev); 11119 int err; 11120 11121 if (rtnl_net_lock_killable(net)) 11122 return -EINTR; 11123 11124 err = register_netdevice(dev); 11125 11126 rtnl_net_unlock(net); 11127 11128 return err; 11129 } 11130 EXPORT_SYMBOL(register_netdev); 11131 11132 int netdev_refcnt_read(const struct net_device *dev) 11133 { 11134 #ifdef CONFIG_PCPU_DEV_REFCNT 11135 int i, refcnt = 0; 11136 11137 for_each_possible_cpu(i) 11138 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i); 11139 return refcnt; 11140 #else 11141 return refcount_read(&dev->dev_refcnt); 11142 #endif 11143 } 11144 EXPORT_SYMBOL(netdev_refcnt_read); 11145 11146 int netdev_unregister_timeout_secs __read_mostly = 10; 11147 11148 #define WAIT_REFS_MIN_MSECS 1 11149 #define WAIT_REFS_MAX_MSECS 250 11150 /** 11151 * netdev_wait_allrefs_any - wait until all references are gone. 11152 * @list: list of net_devices to wait on 11153 * 11154 * This is called when unregistering network devices. 11155 * 11156 * Any protocol or device that holds a reference should register 11157 * for netdevice notification, and cleanup and put back the 11158 * reference if they receive an UNREGISTER event. 11159 * We can get stuck here if buggy protocols don't correctly 11160 * call dev_put. 11161 */ 11162 static struct net_device *netdev_wait_allrefs_any(struct list_head *list) 11163 { 11164 unsigned long rebroadcast_time, warning_time; 11165 struct net_device *dev; 11166 int wait = 0; 11167 11168 rebroadcast_time = warning_time = jiffies; 11169 11170 list_for_each_entry(dev, list, todo_list) 11171 if (netdev_refcnt_read(dev) == 1) 11172 return dev; 11173 11174 while (true) { 11175 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) { 11176 rtnl_lock(); 11177 11178 /* Rebroadcast unregister notification */ 11179 list_for_each_entry(dev, list, todo_list) 11180 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11181 11182 __rtnl_unlock(); 11183 rcu_barrier(); 11184 rtnl_lock(); 11185 11186 list_for_each_entry(dev, list, todo_list) 11187 if (test_bit(__LINK_STATE_LINKWATCH_PENDING, 11188 &dev->state)) { 11189 /* We must not have linkwatch events 11190 * pending on unregister. If this 11191 * happens, we simply run the queue 11192 * unscheduled, resulting in a noop 11193 * for this device. 11194 */ 11195 linkwatch_run_queue(); 11196 break; 11197 } 11198 11199 __rtnl_unlock(); 11200 11201 rebroadcast_time = jiffies; 11202 } 11203 11204 rcu_barrier(); 11205 11206 if (!wait) { 11207 wait = WAIT_REFS_MIN_MSECS; 11208 } else { 11209 msleep(wait); 11210 wait = min(wait << 1, WAIT_REFS_MAX_MSECS); 11211 } 11212 11213 list_for_each_entry(dev, list, todo_list) 11214 if (netdev_refcnt_read(dev) == 1) 11215 return dev; 11216 11217 if (time_after(jiffies, warning_time + 11218 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) { 11219 list_for_each_entry(dev, list, todo_list) { 11220 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n", 11221 dev->name, netdev_refcnt_read(dev)); 11222 ref_tracker_dir_print(&dev->refcnt_tracker, 10); 11223 } 11224 11225 warning_time = jiffies; 11226 } 11227 } 11228 } 11229 11230 /* The sequence is: 11231 * 11232 * rtnl_lock(); 11233 * ... 11234 * register_netdevice(x1); 11235 * register_netdevice(x2); 11236 * ... 11237 * unregister_netdevice(y1); 11238 * unregister_netdevice(y2); 11239 * ... 11240 * rtnl_unlock(); 11241 * free_netdev(y1); 11242 * free_netdev(y2); 11243 * 11244 * We are invoked by rtnl_unlock(). 11245 * This allows us to deal with problems: 11246 * 1) We can delete sysfs objects which invoke hotplug 11247 * without deadlocking with linkwatch via keventd. 11248 * 2) Since we run with the RTNL semaphore not held, we can sleep 11249 * safely in order to wait for the netdev refcnt to drop to zero. 11250 * 11251 * We must not return until all unregister events added during 11252 * the interval the lock was held have been completed. 11253 */ 11254 void netdev_run_todo(void) 11255 { 11256 struct net_device *dev, *tmp; 11257 struct list_head list; 11258 int cnt; 11259 #ifdef CONFIG_LOCKDEP 11260 struct list_head unlink_list; 11261 11262 list_replace_init(&net_unlink_list, &unlink_list); 11263 11264 while (!list_empty(&unlink_list)) { 11265 dev = list_first_entry(&unlink_list, struct net_device, 11266 unlink_list); 11267 list_del_init(&dev->unlink_list); 11268 dev->nested_level = dev->lower_level - 1; 11269 } 11270 #endif 11271 11272 /* Snapshot list, allow later requests */ 11273 list_replace_init(&net_todo_list, &list); 11274 11275 __rtnl_unlock(); 11276 11277 /* Wait for rcu callbacks to finish before next phase */ 11278 if (!list_empty(&list)) 11279 rcu_barrier(); 11280 11281 list_for_each_entry_safe(dev, tmp, &list, todo_list) { 11282 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) { 11283 netdev_WARN(dev, "run_todo but not unregistering\n"); 11284 list_del(&dev->todo_list); 11285 continue; 11286 } 11287 11288 netdev_lock(dev); 11289 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERED); 11290 netdev_unlock(dev); 11291 linkwatch_sync_dev(dev); 11292 } 11293 11294 cnt = 0; 11295 while (!list_empty(&list)) { 11296 dev = netdev_wait_allrefs_any(&list); 11297 list_del(&dev->todo_list); 11298 11299 /* paranoia */ 11300 BUG_ON(netdev_refcnt_read(dev) != 1); 11301 BUG_ON(!list_empty(&dev->ptype_all)); 11302 BUG_ON(!list_empty(&dev->ptype_specific)); 11303 WARN_ON(rcu_access_pointer(dev->ip_ptr)); 11304 WARN_ON(rcu_access_pointer(dev->ip6_ptr)); 11305 11306 netdev_do_free_pcpu_stats(dev); 11307 if (dev->priv_destructor) 11308 dev->priv_destructor(dev); 11309 if (dev->needs_free_netdev) 11310 free_netdev(dev); 11311 11312 cnt++; 11313 11314 /* Free network device */ 11315 kobject_put(&dev->dev.kobj); 11316 } 11317 if (cnt && atomic_sub_and_test(cnt, &dev_unreg_count)) 11318 wake_up(&netdev_unregistering_wq); 11319 } 11320 11321 /* Collate per-cpu network dstats statistics 11322 * 11323 * Read per-cpu network statistics from dev->dstats and populate the related 11324 * fields in @s. 11325 */ 11326 static void dev_fetch_dstats(struct rtnl_link_stats64 *s, 11327 const struct pcpu_dstats __percpu *dstats) 11328 { 11329 int cpu; 11330 11331 for_each_possible_cpu(cpu) { 11332 u64 rx_packets, rx_bytes, rx_drops; 11333 u64 tx_packets, tx_bytes, tx_drops; 11334 const struct pcpu_dstats *stats; 11335 unsigned int start; 11336 11337 stats = per_cpu_ptr(dstats, cpu); 11338 do { 11339 start = u64_stats_fetch_begin(&stats->syncp); 11340 rx_packets = u64_stats_read(&stats->rx_packets); 11341 rx_bytes = u64_stats_read(&stats->rx_bytes); 11342 rx_drops = u64_stats_read(&stats->rx_drops); 11343 tx_packets = u64_stats_read(&stats->tx_packets); 11344 tx_bytes = u64_stats_read(&stats->tx_bytes); 11345 tx_drops = u64_stats_read(&stats->tx_drops); 11346 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11347 11348 s->rx_packets += rx_packets; 11349 s->rx_bytes += rx_bytes; 11350 s->rx_dropped += rx_drops; 11351 s->tx_packets += tx_packets; 11352 s->tx_bytes += tx_bytes; 11353 s->tx_dropped += tx_drops; 11354 } 11355 } 11356 11357 /* ndo_get_stats64 implementation for dtstats-based accounting. 11358 * 11359 * Populate @s from dev->stats and dev->dstats. This is used internally by the 11360 * core for NETDEV_PCPU_STAT_DSTAT-type stats collection. 11361 */ 11362 static void dev_get_dstats64(const struct net_device *dev, 11363 struct rtnl_link_stats64 *s) 11364 { 11365 netdev_stats_to_stats64(s, &dev->stats); 11366 dev_fetch_dstats(s, dev->dstats); 11367 } 11368 11369 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has 11370 * all the same fields in the same order as net_device_stats, with only 11371 * the type differing, but rtnl_link_stats64 may have additional fields 11372 * at the end for newer counters. 11373 */ 11374 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64, 11375 const struct net_device_stats *netdev_stats) 11376 { 11377 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t); 11378 const atomic_long_t *src = (atomic_long_t *)netdev_stats; 11379 u64 *dst = (u64 *)stats64; 11380 11381 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64)); 11382 for (i = 0; i < n; i++) 11383 dst[i] = (unsigned long)atomic_long_read(&src[i]); 11384 /* zero out counters that only exist in rtnl_link_stats64 */ 11385 memset((char *)stats64 + n * sizeof(u64), 0, 11386 sizeof(*stats64) - n * sizeof(u64)); 11387 } 11388 EXPORT_SYMBOL(netdev_stats_to_stats64); 11389 11390 static __cold struct net_device_core_stats __percpu *netdev_core_stats_alloc( 11391 struct net_device *dev) 11392 { 11393 struct net_device_core_stats __percpu *p; 11394 11395 p = alloc_percpu_gfp(struct net_device_core_stats, 11396 GFP_ATOMIC | __GFP_NOWARN); 11397 11398 if (p && cmpxchg(&dev->core_stats, NULL, p)) 11399 free_percpu(p); 11400 11401 /* This READ_ONCE() pairs with the cmpxchg() above */ 11402 return READ_ONCE(dev->core_stats); 11403 } 11404 11405 noinline void netdev_core_stats_inc(struct net_device *dev, u32 offset) 11406 { 11407 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11408 struct net_device_core_stats __percpu *p = READ_ONCE(dev->core_stats); 11409 unsigned long __percpu *field; 11410 11411 if (unlikely(!p)) { 11412 p = netdev_core_stats_alloc(dev); 11413 if (!p) 11414 return; 11415 } 11416 11417 field = (unsigned long __percpu *)((void __percpu *)p + offset); 11418 this_cpu_inc(*field); 11419 } 11420 EXPORT_SYMBOL_GPL(netdev_core_stats_inc); 11421 11422 /** 11423 * dev_get_stats - get network device statistics 11424 * @dev: device to get statistics from 11425 * @storage: place to store stats 11426 * 11427 * Get network statistics from device. Return @storage. 11428 * The device driver may provide its own method by setting 11429 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats; 11430 * otherwise the internal statistics structure is used. 11431 */ 11432 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev, 11433 struct rtnl_link_stats64 *storage) 11434 { 11435 const struct net_device_ops *ops = dev->netdev_ops; 11436 const struct net_device_core_stats __percpu *p; 11437 11438 /* 11439 * IPv{4,6} and udp tunnels share common stat helpers and use 11440 * different stat type (NETDEV_PCPU_STAT_TSTATS vs 11441 * NETDEV_PCPU_STAT_DSTATS). Ensure the accounting is consistent. 11442 */ 11443 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_bytes) != 11444 offsetof(struct pcpu_dstats, rx_bytes)); 11445 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_packets) != 11446 offsetof(struct pcpu_dstats, rx_packets)); 11447 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_bytes) != 11448 offsetof(struct pcpu_dstats, tx_bytes)); 11449 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_packets) != 11450 offsetof(struct pcpu_dstats, tx_packets)); 11451 11452 if (ops->ndo_get_stats64) { 11453 memset(storage, 0, sizeof(*storage)); 11454 ops->ndo_get_stats64(dev, storage); 11455 } else if (ops->ndo_get_stats) { 11456 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev)); 11457 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_TSTATS) { 11458 dev_get_tstats64(dev, storage); 11459 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_DSTATS) { 11460 dev_get_dstats64(dev, storage); 11461 } else { 11462 netdev_stats_to_stats64(storage, &dev->stats); 11463 } 11464 11465 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11466 p = READ_ONCE(dev->core_stats); 11467 if (p) { 11468 const struct net_device_core_stats *core_stats; 11469 int i; 11470 11471 for_each_possible_cpu(i) { 11472 core_stats = per_cpu_ptr(p, i); 11473 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped); 11474 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped); 11475 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler); 11476 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped); 11477 } 11478 } 11479 return storage; 11480 } 11481 EXPORT_SYMBOL(dev_get_stats); 11482 11483 /** 11484 * dev_fetch_sw_netstats - get per-cpu network device statistics 11485 * @s: place to store stats 11486 * @netstats: per-cpu network stats to read from 11487 * 11488 * Read per-cpu network statistics and populate the related fields in @s. 11489 */ 11490 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s, 11491 const struct pcpu_sw_netstats __percpu *netstats) 11492 { 11493 int cpu; 11494 11495 for_each_possible_cpu(cpu) { 11496 u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 11497 const struct pcpu_sw_netstats *stats; 11498 unsigned int start; 11499 11500 stats = per_cpu_ptr(netstats, cpu); 11501 do { 11502 start = u64_stats_fetch_begin(&stats->syncp); 11503 rx_packets = u64_stats_read(&stats->rx_packets); 11504 rx_bytes = u64_stats_read(&stats->rx_bytes); 11505 tx_packets = u64_stats_read(&stats->tx_packets); 11506 tx_bytes = u64_stats_read(&stats->tx_bytes); 11507 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11508 11509 s->rx_packets += rx_packets; 11510 s->rx_bytes += rx_bytes; 11511 s->tx_packets += tx_packets; 11512 s->tx_bytes += tx_bytes; 11513 } 11514 } 11515 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats); 11516 11517 /** 11518 * dev_get_tstats64 - ndo_get_stats64 implementation 11519 * @dev: device to get statistics from 11520 * @s: place to store stats 11521 * 11522 * Populate @s from dev->stats and dev->tstats. Can be used as 11523 * ndo_get_stats64() callback. 11524 */ 11525 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s) 11526 { 11527 netdev_stats_to_stats64(s, &dev->stats); 11528 dev_fetch_sw_netstats(s, dev->tstats); 11529 } 11530 EXPORT_SYMBOL_GPL(dev_get_tstats64); 11531 11532 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev) 11533 { 11534 struct netdev_queue *queue = dev_ingress_queue(dev); 11535 11536 #ifdef CONFIG_NET_CLS_ACT 11537 if (queue) 11538 return queue; 11539 queue = kzalloc(sizeof(*queue), GFP_KERNEL); 11540 if (!queue) 11541 return NULL; 11542 netdev_init_one_queue(dev, queue, NULL); 11543 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc); 11544 RCU_INIT_POINTER(queue->qdisc_sleeping, &noop_qdisc); 11545 rcu_assign_pointer(dev->ingress_queue, queue); 11546 #endif 11547 return queue; 11548 } 11549 11550 static const struct ethtool_ops default_ethtool_ops; 11551 11552 void netdev_set_default_ethtool_ops(struct net_device *dev, 11553 const struct ethtool_ops *ops) 11554 { 11555 if (dev->ethtool_ops == &default_ethtool_ops) 11556 dev->ethtool_ops = ops; 11557 } 11558 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops); 11559 11560 /** 11561 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default 11562 * @dev: netdev to enable the IRQ coalescing on 11563 * 11564 * Sets a conservative default for SW IRQ coalescing. Users can use 11565 * sysfs attributes to override the default values. 11566 */ 11567 void netdev_sw_irq_coalesce_default_on(struct net_device *dev) 11568 { 11569 WARN_ON(dev->reg_state == NETREG_REGISTERED); 11570 11571 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) { 11572 netdev_set_gro_flush_timeout(dev, 20000); 11573 netdev_set_defer_hard_irqs(dev, 1); 11574 } 11575 } 11576 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on); 11577 11578 /** 11579 * alloc_netdev_mqs - allocate network device 11580 * @sizeof_priv: size of private data to allocate space for 11581 * @name: device name format string 11582 * @name_assign_type: origin of device name 11583 * @setup: callback to initialize device 11584 * @txqs: the number of TX subqueues to allocate 11585 * @rxqs: the number of RX subqueues to allocate 11586 * 11587 * Allocates a struct net_device with private data area for driver use 11588 * and performs basic initialization. Also allocates subqueue structs 11589 * for each queue on the device. 11590 */ 11591 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name, 11592 unsigned char name_assign_type, 11593 void (*setup)(struct net_device *), 11594 unsigned int txqs, unsigned int rxqs) 11595 { 11596 struct net_device *dev; 11597 size_t napi_config_sz; 11598 unsigned int maxqs; 11599 11600 BUG_ON(strlen(name) >= sizeof(dev->name)); 11601 11602 if (txqs < 1) { 11603 pr_err("alloc_netdev: Unable to allocate device with zero queues\n"); 11604 return NULL; 11605 } 11606 11607 if (rxqs < 1) { 11608 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n"); 11609 return NULL; 11610 } 11611 11612 maxqs = max(txqs, rxqs); 11613 11614 dev = kvzalloc(struct_size(dev, priv, sizeof_priv), 11615 GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11616 if (!dev) 11617 return NULL; 11618 11619 dev->priv_len = sizeof_priv; 11620 11621 ref_tracker_dir_init(&dev->refcnt_tracker, 128, name); 11622 #ifdef CONFIG_PCPU_DEV_REFCNT 11623 dev->pcpu_refcnt = alloc_percpu(int); 11624 if (!dev->pcpu_refcnt) 11625 goto free_dev; 11626 __dev_hold(dev); 11627 #else 11628 refcount_set(&dev->dev_refcnt, 1); 11629 #endif 11630 11631 if (dev_addr_init(dev)) 11632 goto free_pcpu; 11633 11634 dev_mc_init(dev); 11635 dev_uc_init(dev); 11636 11637 dev_net_set(dev, &init_net); 11638 11639 dev->gso_max_size = GSO_LEGACY_MAX_SIZE; 11640 dev->xdp_zc_max_segs = 1; 11641 dev->gso_max_segs = GSO_MAX_SEGS; 11642 dev->gro_max_size = GRO_LEGACY_MAX_SIZE; 11643 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE; 11644 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE; 11645 dev->tso_max_size = TSO_LEGACY_MAX_SIZE; 11646 dev->tso_max_segs = TSO_MAX_SEGS; 11647 dev->upper_level = 1; 11648 dev->lower_level = 1; 11649 #ifdef CONFIG_LOCKDEP 11650 dev->nested_level = 0; 11651 INIT_LIST_HEAD(&dev->unlink_list); 11652 #endif 11653 11654 INIT_LIST_HEAD(&dev->napi_list); 11655 INIT_LIST_HEAD(&dev->unreg_list); 11656 INIT_LIST_HEAD(&dev->close_list); 11657 INIT_LIST_HEAD(&dev->link_watch_list); 11658 INIT_LIST_HEAD(&dev->adj_list.upper); 11659 INIT_LIST_HEAD(&dev->adj_list.lower); 11660 INIT_LIST_HEAD(&dev->ptype_all); 11661 INIT_LIST_HEAD(&dev->ptype_specific); 11662 INIT_LIST_HEAD(&dev->net_notifier_list); 11663 #ifdef CONFIG_NET_SCHED 11664 hash_init(dev->qdisc_hash); 11665 #endif 11666 11667 mutex_init(&dev->lock); 11668 11669 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 11670 setup(dev); 11671 11672 if (!dev->tx_queue_len) { 11673 dev->priv_flags |= IFF_NO_QUEUE; 11674 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN; 11675 } 11676 11677 dev->num_tx_queues = txqs; 11678 dev->real_num_tx_queues = txqs; 11679 if (netif_alloc_netdev_queues(dev)) 11680 goto free_all; 11681 11682 dev->num_rx_queues = rxqs; 11683 dev->real_num_rx_queues = rxqs; 11684 if (netif_alloc_rx_queues(dev)) 11685 goto free_all; 11686 dev->ethtool = kzalloc(sizeof(*dev->ethtool), GFP_KERNEL_ACCOUNT); 11687 if (!dev->ethtool) 11688 goto free_all; 11689 11690 dev->cfg = kzalloc(sizeof(*dev->cfg), GFP_KERNEL_ACCOUNT); 11691 if (!dev->cfg) 11692 goto free_all; 11693 dev->cfg_pending = dev->cfg; 11694 11695 napi_config_sz = array_size(maxqs, sizeof(*dev->napi_config)); 11696 dev->napi_config = kvzalloc(napi_config_sz, GFP_KERNEL_ACCOUNT); 11697 if (!dev->napi_config) 11698 goto free_all; 11699 11700 strscpy(dev->name, name); 11701 dev->name_assign_type = name_assign_type; 11702 dev->group = INIT_NETDEV_GROUP; 11703 if (!dev->ethtool_ops) 11704 dev->ethtool_ops = &default_ethtool_ops; 11705 11706 nf_hook_netdev_init(dev); 11707 11708 return dev; 11709 11710 free_all: 11711 free_netdev(dev); 11712 return NULL; 11713 11714 free_pcpu: 11715 #ifdef CONFIG_PCPU_DEV_REFCNT 11716 free_percpu(dev->pcpu_refcnt); 11717 free_dev: 11718 #endif 11719 kvfree(dev); 11720 return NULL; 11721 } 11722 EXPORT_SYMBOL(alloc_netdev_mqs); 11723 11724 static void netdev_napi_exit(struct net_device *dev) 11725 { 11726 if (!list_empty(&dev->napi_list)) { 11727 struct napi_struct *p, *n; 11728 11729 netdev_lock(dev); 11730 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list) 11731 __netif_napi_del_locked(p); 11732 netdev_unlock(dev); 11733 11734 synchronize_net(); 11735 } 11736 11737 kvfree(dev->napi_config); 11738 } 11739 11740 /** 11741 * free_netdev - free network device 11742 * @dev: device 11743 * 11744 * This function does the last stage of destroying an allocated device 11745 * interface. The reference to the device object is released. If this 11746 * is the last reference then it will be freed.Must be called in process 11747 * context. 11748 */ 11749 void free_netdev(struct net_device *dev) 11750 { 11751 might_sleep(); 11752 11753 /* When called immediately after register_netdevice() failed the unwind 11754 * handling may still be dismantling the device. Handle that case by 11755 * deferring the free. 11756 */ 11757 if (dev->reg_state == NETREG_UNREGISTERING) { 11758 ASSERT_RTNL(); 11759 dev->needs_free_netdev = true; 11760 return; 11761 } 11762 11763 WARN_ON(dev->cfg != dev->cfg_pending); 11764 kfree(dev->cfg); 11765 kfree(dev->ethtool); 11766 netif_free_tx_queues(dev); 11767 netif_free_rx_queues(dev); 11768 11769 kfree(rcu_dereference_protected(dev->ingress_queue, 1)); 11770 11771 /* Flush device addresses */ 11772 dev_addr_flush(dev); 11773 11774 netdev_napi_exit(dev); 11775 11776 netif_del_cpu_rmap(dev); 11777 11778 ref_tracker_dir_exit(&dev->refcnt_tracker); 11779 #ifdef CONFIG_PCPU_DEV_REFCNT 11780 free_percpu(dev->pcpu_refcnt); 11781 dev->pcpu_refcnt = NULL; 11782 #endif 11783 free_percpu(dev->core_stats); 11784 dev->core_stats = NULL; 11785 free_percpu(dev->xdp_bulkq); 11786 dev->xdp_bulkq = NULL; 11787 11788 netdev_free_phy_link_topology(dev); 11789 11790 mutex_destroy(&dev->lock); 11791 11792 /* Compatibility with error handling in drivers */ 11793 if (dev->reg_state == NETREG_UNINITIALIZED || 11794 dev->reg_state == NETREG_DUMMY) { 11795 kvfree(dev); 11796 return; 11797 } 11798 11799 BUG_ON(dev->reg_state != NETREG_UNREGISTERED); 11800 WRITE_ONCE(dev->reg_state, NETREG_RELEASED); 11801 11802 /* will free via device release */ 11803 put_device(&dev->dev); 11804 } 11805 EXPORT_SYMBOL(free_netdev); 11806 11807 /** 11808 * alloc_netdev_dummy - Allocate and initialize a dummy net device. 11809 * @sizeof_priv: size of private data to allocate space for 11810 * 11811 * Return: the allocated net_device on success, NULL otherwise 11812 */ 11813 struct net_device *alloc_netdev_dummy(int sizeof_priv) 11814 { 11815 return alloc_netdev(sizeof_priv, "dummy#", NET_NAME_UNKNOWN, 11816 init_dummy_netdev); 11817 } 11818 EXPORT_SYMBOL_GPL(alloc_netdev_dummy); 11819 11820 /** 11821 * synchronize_net - Synchronize with packet receive processing 11822 * 11823 * Wait for packets currently being received to be done. 11824 * Does not block later packets from starting. 11825 */ 11826 void synchronize_net(void) 11827 { 11828 might_sleep(); 11829 if (from_cleanup_net() || rtnl_is_locked()) 11830 synchronize_rcu_expedited(); 11831 else 11832 synchronize_rcu(); 11833 } 11834 EXPORT_SYMBOL(synchronize_net); 11835 11836 static void netdev_rss_contexts_free(struct net_device *dev) 11837 { 11838 struct ethtool_rxfh_context *ctx; 11839 unsigned long context; 11840 11841 mutex_lock(&dev->ethtool->rss_lock); 11842 xa_for_each(&dev->ethtool->rss_ctx, context, ctx) { 11843 struct ethtool_rxfh_param rxfh; 11844 11845 rxfh.indir = ethtool_rxfh_context_indir(ctx); 11846 rxfh.key = ethtool_rxfh_context_key(ctx); 11847 rxfh.hfunc = ctx->hfunc; 11848 rxfh.input_xfrm = ctx->input_xfrm; 11849 rxfh.rss_context = context; 11850 rxfh.rss_delete = true; 11851 11852 xa_erase(&dev->ethtool->rss_ctx, context); 11853 if (dev->ethtool_ops->create_rxfh_context) 11854 dev->ethtool_ops->remove_rxfh_context(dev, ctx, 11855 context, NULL); 11856 else 11857 dev->ethtool_ops->set_rxfh(dev, &rxfh, NULL); 11858 kfree(ctx); 11859 } 11860 xa_destroy(&dev->ethtool->rss_ctx); 11861 mutex_unlock(&dev->ethtool->rss_lock); 11862 } 11863 11864 /** 11865 * unregister_netdevice_queue - remove device from the kernel 11866 * @dev: device 11867 * @head: list 11868 * 11869 * This function shuts down a device interface and removes it 11870 * from the kernel tables. 11871 * If head not NULL, device is queued to be unregistered later. 11872 * 11873 * Callers must hold the rtnl semaphore. You may want 11874 * unregister_netdev() instead of this. 11875 */ 11876 11877 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head) 11878 { 11879 ASSERT_RTNL(); 11880 11881 if (head) { 11882 list_move_tail(&dev->unreg_list, head); 11883 } else { 11884 LIST_HEAD(single); 11885 11886 list_add(&dev->unreg_list, &single); 11887 unregister_netdevice_many(&single); 11888 } 11889 } 11890 EXPORT_SYMBOL(unregister_netdevice_queue); 11891 11892 static void dev_memory_provider_uninstall(struct net_device *dev) 11893 { 11894 unsigned int i; 11895 11896 for (i = 0; i < dev->real_num_rx_queues; i++) { 11897 struct netdev_rx_queue *rxq = &dev->_rx[i]; 11898 struct pp_memory_provider_params *p = &rxq->mp_params; 11899 11900 if (p->mp_ops && p->mp_ops->uninstall) 11901 p->mp_ops->uninstall(rxq->mp_params.mp_priv, rxq); 11902 } 11903 } 11904 11905 void unregister_netdevice_many_notify(struct list_head *head, 11906 u32 portid, const struct nlmsghdr *nlh) 11907 { 11908 struct net_device *dev, *tmp; 11909 LIST_HEAD(close_head); 11910 int cnt = 0; 11911 11912 BUG_ON(dev_boot_phase); 11913 ASSERT_RTNL(); 11914 11915 if (list_empty(head)) 11916 return; 11917 11918 list_for_each_entry_safe(dev, tmp, head, unreg_list) { 11919 /* Some devices call without registering 11920 * for initialization unwind. Remove those 11921 * devices and proceed with the remaining. 11922 */ 11923 if (dev->reg_state == NETREG_UNINITIALIZED) { 11924 pr_debug("unregister_netdevice: device %s/%p never was registered\n", 11925 dev->name, dev); 11926 11927 WARN_ON(1); 11928 list_del(&dev->unreg_list); 11929 continue; 11930 } 11931 dev->dismantle = true; 11932 BUG_ON(dev->reg_state != NETREG_REGISTERED); 11933 } 11934 11935 /* If device is running, close it first. */ 11936 list_for_each_entry(dev, head, unreg_list) { 11937 list_add_tail(&dev->close_list, &close_head); 11938 netdev_lock_ops(dev); 11939 } 11940 dev_close_many(&close_head, true); 11941 11942 list_for_each_entry(dev, head, unreg_list) { 11943 netdev_unlock_ops(dev); 11944 /* And unlink it from device chain. */ 11945 unlist_netdevice(dev); 11946 netdev_lock(dev); 11947 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERING); 11948 netdev_unlock(dev); 11949 } 11950 flush_all_backlogs(); 11951 11952 synchronize_net(); 11953 11954 list_for_each_entry(dev, head, unreg_list) { 11955 struct sk_buff *skb = NULL; 11956 11957 /* Shutdown queueing discipline. */ 11958 dev_shutdown(dev); 11959 dev_tcx_uninstall(dev); 11960 netdev_lock_ops(dev); 11961 dev_xdp_uninstall(dev); 11962 dev_memory_provider_uninstall(dev); 11963 netdev_unlock_ops(dev); 11964 bpf_dev_bound_netdev_unregister(dev); 11965 11966 netdev_offload_xstats_disable_all(dev); 11967 11968 /* Notify protocols, that we are about to destroy 11969 * this device. They should clean all the things. 11970 */ 11971 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11972 11973 if (!dev->rtnl_link_ops || 11974 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 11975 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0, 11976 GFP_KERNEL, NULL, 0, 11977 portid, nlh); 11978 11979 /* 11980 * Flush the unicast and multicast chains 11981 */ 11982 dev_uc_flush(dev); 11983 dev_mc_flush(dev); 11984 11985 netdev_name_node_alt_flush(dev); 11986 netdev_name_node_free(dev->name_node); 11987 11988 netdev_rss_contexts_free(dev); 11989 11990 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11991 11992 if (dev->netdev_ops->ndo_uninit) 11993 dev->netdev_ops->ndo_uninit(dev); 11994 11995 mutex_destroy(&dev->ethtool->rss_lock); 11996 11997 net_shaper_flush_netdev(dev); 11998 11999 if (skb) 12000 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh); 12001 12002 /* Notifier chain MUST detach us all upper devices. */ 12003 WARN_ON(netdev_has_any_upper_dev(dev)); 12004 WARN_ON(netdev_has_any_lower_dev(dev)); 12005 12006 /* Remove entries from kobject tree */ 12007 netdev_unregister_kobject(dev); 12008 #ifdef CONFIG_XPS 12009 /* Remove XPS queueing entries */ 12010 netif_reset_xps_queues_gt(dev, 0); 12011 #endif 12012 } 12013 12014 synchronize_net(); 12015 12016 list_for_each_entry(dev, head, unreg_list) { 12017 netdev_put(dev, &dev->dev_registered_tracker); 12018 net_set_todo(dev); 12019 cnt++; 12020 } 12021 atomic_add(cnt, &dev_unreg_count); 12022 12023 list_del(head); 12024 } 12025 12026 /** 12027 * unregister_netdevice_many - unregister many devices 12028 * @head: list of devices 12029 * 12030 * Note: As most callers use a stack allocated list_head, 12031 * we force a list_del() to make sure stack won't be corrupted later. 12032 */ 12033 void unregister_netdevice_many(struct list_head *head) 12034 { 12035 unregister_netdevice_many_notify(head, 0, NULL); 12036 } 12037 EXPORT_SYMBOL(unregister_netdevice_many); 12038 12039 /** 12040 * unregister_netdev - remove device from the kernel 12041 * @dev: device 12042 * 12043 * This function shuts down a device interface and removes it 12044 * from the kernel tables. 12045 * 12046 * This is just a wrapper for unregister_netdevice that takes 12047 * the rtnl semaphore. In general you want to use this and not 12048 * unregister_netdevice. 12049 */ 12050 void unregister_netdev(struct net_device *dev) 12051 { 12052 rtnl_net_dev_lock(dev); 12053 unregister_netdevice(dev); 12054 rtnl_net_dev_unlock(dev); 12055 } 12056 EXPORT_SYMBOL(unregister_netdev); 12057 12058 int __dev_change_net_namespace(struct net_device *dev, struct net *net, 12059 const char *pat, int new_ifindex, 12060 struct netlink_ext_ack *extack) 12061 { 12062 struct netdev_name_node *name_node; 12063 struct net *net_old = dev_net(dev); 12064 char new_name[IFNAMSIZ] = {}; 12065 int err, new_nsid; 12066 12067 ASSERT_RTNL(); 12068 12069 /* Don't allow namespace local devices to be moved. */ 12070 err = -EINVAL; 12071 if (dev->netns_immutable) { 12072 NL_SET_ERR_MSG(extack, "The interface netns is immutable"); 12073 goto out; 12074 } 12075 12076 /* Ensure the device has been registered */ 12077 if (dev->reg_state != NETREG_REGISTERED) { 12078 NL_SET_ERR_MSG(extack, "The interface isn't registered"); 12079 goto out; 12080 } 12081 12082 /* Get out if there is nothing todo */ 12083 err = 0; 12084 if (net_eq(net_old, net)) 12085 goto out; 12086 12087 /* Pick the destination device name, and ensure 12088 * we can use it in the destination network namespace. 12089 */ 12090 err = -EEXIST; 12091 if (netdev_name_in_use(net, dev->name)) { 12092 /* We get here if we can't use the current device name */ 12093 if (!pat) { 12094 NL_SET_ERR_MSG(extack, 12095 "An interface with the same name exists in the target netns"); 12096 goto out; 12097 } 12098 err = dev_prep_valid_name(net, dev, pat, new_name, EEXIST); 12099 if (err < 0) { 12100 NL_SET_ERR_MSG_FMT(extack, 12101 "Unable to use '%s' for the new interface name in the target netns", 12102 pat); 12103 goto out; 12104 } 12105 } 12106 /* Check that none of the altnames conflicts. */ 12107 err = -EEXIST; 12108 netdev_for_each_altname(dev, name_node) { 12109 if (netdev_name_in_use(net, name_node->name)) { 12110 NL_SET_ERR_MSG_FMT(extack, 12111 "An interface with the altname %s exists in the target netns", 12112 name_node->name); 12113 goto out; 12114 } 12115 } 12116 12117 /* Check that new_ifindex isn't used yet. */ 12118 if (new_ifindex) { 12119 err = dev_index_reserve(net, new_ifindex); 12120 if (err < 0) { 12121 NL_SET_ERR_MSG_FMT(extack, 12122 "The ifindex %d is not available in the target netns", 12123 new_ifindex); 12124 goto out; 12125 } 12126 } else { 12127 /* If there is an ifindex conflict assign a new one */ 12128 err = dev_index_reserve(net, dev->ifindex); 12129 if (err == -EBUSY) 12130 err = dev_index_reserve(net, 0); 12131 if (err < 0) { 12132 NL_SET_ERR_MSG(extack, 12133 "Unable to allocate a new ifindex in the target netns"); 12134 goto out; 12135 } 12136 new_ifindex = err; 12137 } 12138 12139 /* 12140 * And now a mini version of register_netdevice unregister_netdevice. 12141 */ 12142 12143 netdev_lock_ops(dev); 12144 /* If device is running close it first. */ 12145 netif_close(dev); 12146 /* And unlink it from device chain */ 12147 unlist_netdevice(dev); 12148 netdev_unlock_ops(dev); 12149 12150 synchronize_net(); 12151 12152 /* Shutdown queueing discipline. */ 12153 dev_shutdown(dev); 12154 12155 /* Notify protocols, that we are about to destroy 12156 * this device. They should clean all the things. 12157 * 12158 * Note that dev->reg_state stays at NETREG_REGISTERED. 12159 * This is wanted because this way 8021q and macvlan know 12160 * the device is just moving and can keep their slaves up. 12161 */ 12162 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 12163 rcu_barrier(); 12164 12165 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL); 12166 12167 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid, 12168 new_ifindex); 12169 12170 /* 12171 * Flush the unicast and multicast chains 12172 */ 12173 dev_uc_flush(dev); 12174 dev_mc_flush(dev); 12175 12176 /* Send a netdev-removed uevent to the old namespace */ 12177 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE); 12178 netdev_adjacent_del_links(dev); 12179 12180 /* Move per-net netdevice notifiers that are following the netdevice */ 12181 move_netdevice_notifiers_dev_net(dev, net); 12182 12183 /* Actually switch the network namespace */ 12184 dev_net_set(dev, net); 12185 dev->ifindex = new_ifindex; 12186 12187 if (new_name[0]) { 12188 /* Rename the netdev to prepared name */ 12189 write_seqlock_bh(&netdev_rename_lock); 12190 strscpy(dev->name, new_name, IFNAMSIZ); 12191 write_sequnlock_bh(&netdev_rename_lock); 12192 } 12193 12194 /* Fixup kobjects */ 12195 dev_set_uevent_suppress(&dev->dev, 1); 12196 err = device_rename(&dev->dev, dev->name); 12197 dev_set_uevent_suppress(&dev->dev, 0); 12198 WARN_ON(err); 12199 12200 /* Send a netdev-add uevent to the new namespace */ 12201 kobject_uevent(&dev->dev.kobj, KOBJ_ADD); 12202 netdev_adjacent_add_links(dev); 12203 12204 /* Adapt owner in case owning user namespace of target network 12205 * namespace is different from the original one. 12206 */ 12207 err = netdev_change_owner(dev, net_old, net); 12208 WARN_ON(err); 12209 12210 netdev_lock_ops(dev); 12211 /* Add the device back in the hashes */ 12212 list_netdevice(dev); 12213 /* Notify protocols, that a new device appeared. */ 12214 call_netdevice_notifiers(NETDEV_REGISTER, dev); 12215 netdev_unlock_ops(dev); 12216 12217 /* 12218 * Prevent userspace races by waiting until the network 12219 * device is fully setup before sending notifications. 12220 */ 12221 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 12222 12223 synchronize_net(); 12224 err = 0; 12225 out: 12226 return err; 12227 } 12228 12229 static int dev_cpu_dead(unsigned int oldcpu) 12230 { 12231 struct sk_buff **list_skb; 12232 struct sk_buff *skb; 12233 unsigned int cpu; 12234 struct softnet_data *sd, *oldsd, *remsd = NULL; 12235 12236 local_irq_disable(); 12237 cpu = smp_processor_id(); 12238 sd = &per_cpu(softnet_data, cpu); 12239 oldsd = &per_cpu(softnet_data, oldcpu); 12240 12241 /* Find end of our completion_queue. */ 12242 list_skb = &sd->completion_queue; 12243 while (*list_skb) 12244 list_skb = &(*list_skb)->next; 12245 /* Append completion queue from offline CPU. */ 12246 *list_skb = oldsd->completion_queue; 12247 oldsd->completion_queue = NULL; 12248 12249 /* Append output queue from offline CPU. */ 12250 if (oldsd->output_queue) { 12251 *sd->output_queue_tailp = oldsd->output_queue; 12252 sd->output_queue_tailp = oldsd->output_queue_tailp; 12253 oldsd->output_queue = NULL; 12254 oldsd->output_queue_tailp = &oldsd->output_queue; 12255 } 12256 /* Append NAPI poll list from offline CPU, with one exception : 12257 * process_backlog() must be called by cpu owning percpu backlog. 12258 * We properly handle process_queue & input_pkt_queue later. 12259 */ 12260 while (!list_empty(&oldsd->poll_list)) { 12261 struct napi_struct *napi = list_first_entry(&oldsd->poll_list, 12262 struct napi_struct, 12263 poll_list); 12264 12265 list_del_init(&napi->poll_list); 12266 if (napi->poll == process_backlog) 12267 napi->state &= NAPIF_STATE_THREADED; 12268 else 12269 ____napi_schedule(sd, napi); 12270 } 12271 12272 raise_softirq_irqoff(NET_TX_SOFTIRQ); 12273 local_irq_enable(); 12274 12275 if (!use_backlog_threads()) { 12276 #ifdef CONFIG_RPS 12277 remsd = oldsd->rps_ipi_list; 12278 oldsd->rps_ipi_list = NULL; 12279 #endif 12280 /* send out pending IPI's on offline CPU */ 12281 net_rps_send_ipi(remsd); 12282 } 12283 12284 /* Process offline CPU's input_pkt_queue */ 12285 while ((skb = __skb_dequeue(&oldsd->process_queue))) { 12286 netif_rx(skb); 12287 rps_input_queue_head_incr(oldsd); 12288 } 12289 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) { 12290 netif_rx(skb); 12291 rps_input_queue_head_incr(oldsd); 12292 } 12293 12294 return 0; 12295 } 12296 12297 /** 12298 * netdev_increment_features - increment feature set by one 12299 * @all: current feature set 12300 * @one: new feature set 12301 * @mask: mask feature set 12302 * 12303 * Computes a new feature set after adding a device with feature set 12304 * @one to the master device with current feature set @all. Will not 12305 * enable anything that is off in @mask. Returns the new feature set. 12306 */ 12307 netdev_features_t netdev_increment_features(netdev_features_t all, 12308 netdev_features_t one, netdev_features_t mask) 12309 { 12310 if (mask & NETIF_F_HW_CSUM) 12311 mask |= NETIF_F_CSUM_MASK; 12312 mask |= NETIF_F_VLAN_CHALLENGED; 12313 12314 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask; 12315 all &= one | ~NETIF_F_ALL_FOR_ALL; 12316 12317 /* If one device supports hw checksumming, set for all. */ 12318 if (all & NETIF_F_HW_CSUM) 12319 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM); 12320 12321 return all; 12322 } 12323 EXPORT_SYMBOL(netdev_increment_features); 12324 12325 static struct hlist_head * __net_init netdev_create_hash(void) 12326 { 12327 int i; 12328 struct hlist_head *hash; 12329 12330 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL); 12331 if (hash != NULL) 12332 for (i = 0; i < NETDEV_HASHENTRIES; i++) 12333 INIT_HLIST_HEAD(&hash[i]); 12334 12335 return hash; 12336 } 12337 12338 /* Initialize per network namespace state */ 12339 static int __net_init netdev_init(struct net *net) 12340 { 12341 BUILD_BUG_ON(GRO_HASH_BUCKETS > 12342 BITS_PER_BYTE * sizeof_field(struct gro_node, bitmask)); 12343 12344 INIT_LIST_HEAD(&net->dev_base_head); 12345 12346 net->dev_name_head = netdev_create_hash(); 12347 if (net->dev_name_head == NULL) 12348 goto err_name; 12349 12350 net->dev_index_head = netdev_create_hash(); 12351 if (net->dev_index_head == NULL) 12352 goto err_idx; 12353 12354 xa_init_flags(&net->dev_by_index, XA_FLAGS_ALLOC1); 12355 12356 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain); 12357 12358 return 0; 12359 12360 err_idx: 12361 kfree(net->dev_name_head); 12362 err_name: 12363 return -ENOMEM; 12364 } 12365 12366 /** 12367 * netdev_drivername - network driver for the device 12368 * @dev: network device 12369 * 12370 * Determine network driver for device. 12371 */ 12372 const char *netdev_drivername(const struct net_device *dev) 12373 { 12374 const struct device_driver *driver; 12375 const struct device *parent; 12376 const char *empty = ""; 12377 12378 parent = dev->dev.parent; 12379 if (!parent) 12380 return empty; 12381 12382 driver = parent->driver; 12383 if (driver && driver->name) 12384 return driver->name; 12385 return empty; 12386 } 12387 12388 static void __netdev_printk(const char *level, const struct net_device *dev, 12389 struct va_format *vaf) 12390 { 12391 if (dev && dev->dev.parent) { 12392 dev_printk_emit(level[1] - '0', 12393 dev->dev.parent, 12394 "%s %s %s%s: %pV", 12395 dev_driver_string(dev->dev.parent), 12396 dev_name(dev->dev.parent), 12397 netdev_name(dev), netdev_reg_state(dev), 12398 vaf); 12399 } else if (dev) { 12400 printk("%s%s%s: %pV", 12401 level, netdev_name(dev), netdev_reg_state(dev), vaf); 12402 } else { 12403 printk("%s(NULL net_device): %pV", level, vaf); 12404 } 12405 } 12406 12407 void netdev_printk(const char *level, const struct net_device *dev, 12408 const char *format, ...) 12409 { 12410 struct va_format vaf; 12411 va_list args; 12412 12413 va_start(args, format); 12414 12415 vaf.fmt = format; 12416 vaf.va = &args; 12417 12418 __netdev_printk(level, dev, &vaf); 12419 12420 va_end(args); 12421 } 12422 EXPORT_SYMBOL(netdev_printk); 12423 12424 #define define_netdev_printk_level(func, level) \ 12425 void func(const struct net_device *dev, const char *fmt, ...) \ 12426 { \ 12427 struct va_format vaf; \ 12428 va_list args; \ 12429 \ 12430 va_start(args, fmt); \ 12431 \ 12432 vaf.fmt = fmt; \ 12433 vaf.va = &args; \ 12434 \ 12435 __netdev_printk(level, dev, &vaf); \ 12436 \ 12437 va_end(args); \ 12438 } \ 12439 EXPORT_SYMBOL(func); 12440 12441 define_netdev_printk_level(netdev_emerg, KERN_EMERG); 12442 define_netdev_printk_level(netdev_alert, KERN_ALERT); 12443 define_netdev_printk_level(netdev_crit, KERN_CRIT); 12444 define_netdev_printk_level(netdev_err, KERN_ERR); 12445 define_netdev_printk_level(netdev_warn, KERN_WARNING); 12446 define_netdev_printk_level(netdev_notice, KERN_NOTICE); 12447 define_netdev_printk_level(netdev_info, KERN_INFO); 12448 12449 static void __net_exit netdev_exit(struct net *net) 12450 { 12451 kfree(net->dev_name_head); 12452 kfree(net->dev_index_head); 12453 xa_destroy(&net->dev_by_index); 12454 if (net != &init_net) 12455 WARN_ON_ONCE(!list_empty(&net->dev_base_head)); 12456 } 12457 12458 static struct pernet_operations __net_initdata netdev_net_ops = { 12459 .init = netdev_init, 12460 .exit = netdev_exit, 12461 }; 12462 12463 static void __net_exit default_device_exit_net(struct net *net) 12464 { 12465 struct netdev_name_node *name_node, *tmp; 12466 struct net_device *dev, *aux; 12467 /* 12468 * Push all migratable network devices back to the 12469 * initial network namespace 12470 */ 12471 ASSERT_RTNL(); 12472 for_each_netdev_safe(net, dev, aux) { 12473 int err; 12474 char fb_name[IFNAMSIZ]; 12475 12476 /* Ignore unmoveable devices (i.e. loopback) */ 12477 if (dev->netns_immutable) 12478 continue; 12479 12480 /* Leave virtual devices for the generic cleanup */ 12481 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund) 12482 continue; 12483 12484 /* Push remaining network devices to init_net */ 12485 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex); 12486 if (netdev_name_in_use(&init_net, fb_name)) 12487 snprintf(fb_name, IFNAMSIZ, "dev%%d"); 12488 12489 netdev_for_each_altname_safe(dev, name_node, tmp) 12490 if (netdev_name_in_use(&init_net, name_node->name)) 12491 __netdev_name_node_alt_destroy(name_node); 12492 12493 err = dev_change_net_namespace(dev, &init_net, fb_name); 12494 if (err) { 12495 pr_emerg("%s: failed to move %s to init_net: %d\n", 12496 __func__, dev->name, err); 12497 BUG(); 12498 } 12499 } 12500 } 12501 12502 static void __net_exit default_device_exit_batch(struct list_head *net_list) 12503 { 12504 /* At exit all network devices most be removed from a network 12505 * namespace. Do this in the reverse order of registration. 12506 * Do this across as many network namespaces as possible to 12507 * improve batching efficiency. 12508 */ 12509 struct net_device *dev; 12510 struct net *net; 12511 LIST_HEAD(dev_kill_list); 12512 12513 rtnl_lock(); 12514 list_for_each_entry(net, net_list, exit_list) { 12515 default_device_exit_net(net); 12516 cond_resched(); 12517 } 12518 12519 list_for_each_entry(net, net_list, exit_list) { 12520 for_each_netdev_reverse(net, dev) { 12521 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink) 12522 dev->rtnl_link_ops->dellink(dev, &dev_kill_list); 12523 else 12524 unregister_netdevice_queue(dev, &dev_kill_list); 12525 } 12526 } 12527 unregister_netdevice_many(&dev_kill_list); 12528 rtnl_unlock(); 12529 } 12530 12531 static struct pernet_operations __net_initdata default_device_ops = { 12532 .exit_batch = default_device_exit_batch, 12533 }; 12534 12535 static void __init net_dev_struct_check(void) 12536 { 12537 /* TX read-mostly hotpath */ 12538 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, priv_flags_fast); 12539 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, netdev_ops); 12540 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, header_ops); 12541 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, _tx); 12542 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, real_num_tx_queues); 12543 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_size); 12544 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_ipv4_max_size); 12545 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_segs); 12546 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_partial_features); 12547 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, num_tc); 12548 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, mtu); 12549 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, needed_headroom); 12550 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tc_to_txq); 12551 #ifdef CONFIG_XPS 12552 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, xps_maps); 12553 #endif 12554 #ifdef CONFIG_NETFILTER_EGRESS 12555 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, nf_hooks_egress); 12556 #endif 12557 #ifdef CONFIG_NET_XGRESS 12558 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tcx_egress); 12559 #endif 12560 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_tx, 160); 12561 12562 /* TXRX read-mostly hotpath */ 12563 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, lstats); 12564 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, state); 12565 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, flags); 12566 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, hard_header_len); 12567 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, features); 12568 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, ip6_ptr); 12569 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_txrx, 46); 12570 12571 /* RX read-mostly hotpath */ 12572 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ptype_specific); 12573 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ifindex); 12574 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, real_num_rx_queues); 12575 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, _rx); 12576 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_max_size); 12577 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_ipv4_max_size); 12578 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler); 12579 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler_data); 12580 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, nd_net); 12581 #ifdef CONFIG_NETPOLL 12582 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, npinfo); 12583 #endif 12584 #ifdef CONFIG_NET_XGRESS 12585 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, tcx_ingress); 12586 #endif 12587 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_rx, 92); 12588 } 12589 12590 /* 12591 * Initialize the DEV module. At boot time this walks the device list and 12592 * unhooks any devices that fail to initialise (normally hardware not 12593 * present) and leaves us with a valid list of present and active devices. 12594 * 12595 */ 12596 12597 /* We allocate 256 pages for each CPU if PAGE_SHIFT is 12 */ 12598 #define SYSTEM_PERCPU_PAGE_POOL_SIZE ((1 << 20) / PAGE_SIZE) 12599 12600 static int net_page_pool_create(int cpuid) 12601 { 12602 #if IS_ENABLED(CONFIG_PAGE_POOL) 12603 struct page_pool_params page_pool_params = { 12604 .pool_size = SYSTEM_PERCPU_PAGE_POOL_SIZE, 12605 .flags = PP_FLAG_SYSTEM_POOL, 12606 .nid = cpu_to_mem(cpuid), 12607 }; 12608 struct page_pool *pp_ptr; 12609 int err; 12610 12611 pp_ptr = page_pool_create_percpu(&page_pool_params, cpuid); 12612 if (IS_ERR(pp_ptr)) 12613 return -ENOMEM; 12614 12615 err = xdp_reg_page_pool(pp_ptr); 12616 if (err) { 12617 page_pool_destroy(pp_ptr); 12618 return err; 12619 } 12620 12621 per_cpu(system_page_pool, cpuid) = pp_ptr; 12622 #endif 12623 return 0; 12624 } 12625 12626 static int backlog_napi_should_run(unsigned int cpu) 12627 { 12628 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12629 struct napi_struct *napi = &sd->backlog; 12630 12631 return test_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 12632 } 12633 12634 static void run_backlog_napi(unsigned int cpu) 12635 { 12636 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12637 12638 napi_threaded_poll_loop(&sd->backlog); 12639 } 12640 12641 static void backlog_napi_setup(unsigned int cpu) 12642 { 12643 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12644 struct napi_struct *napi = &sd->backlog; 12645 12646 napi->thread = this_cpu_read(backlog_napi); 12647 set_bit(NAPI_STATE_THREADED, &napi->state); 12648 } 12649 12650 static struct smp_hotplug_thread backlog_threads = { 12651 .store = &backlog_napi, 12652 .thread_should_run = backlog_napi_should_run, 12653 .thread_fn = run_backlog_napi, 12654 .thread_comm = "backlog_napi/%u", 12655 .setup = backlog_napi_setup, 12656 }; 12657 12658 /* 12659 * This is called single threaded during boot, so no need 12660 * to take the rtnl semaphore. 12661 */ 12662 static int __init net_dev_init(void) 12663 { 12664 int i, rc = -ENOMEM; 12665 12666 BUG_ON(!dev_boot_phase); 12667 12668 net_dev_struct_check(); 12669 12670 if (dev_proc_init()) 12671 goto out; 12672 12673 if (netdev_kobject_init()) 12674 goto out; 12675 12676 for (i = 0; i < PTYPE_HASH_SIZE; i++) 12677 INIT_LIST_HEAD(&ptype_base[i]); 12678 12679 if (register_pernet_subsys(&netdev_net_ops)) 12680 goto out; 12681 12682 /* 12683 * Initialise the packet receive queues. 12684 */ 12685 12686 flush_backlogs_fallback = flush_backlogs_alloc(); 12687 if (!flush_backlogs_fallback) 12688 goto out; 12689 12690 for_each_possible_cpu(i) { 12691 struct softnet_data *sd = &per_cpu(softnet_data, i); 12692 12693 skb_queue_head_init(&sd->input_pkt_queue); 12694 skb_queue_head_init(&sd->process_queue); 12695 #ifdef CONFIG_XFRM_OFFLOAD 12696 skb_queue_head_init(&sd->xfrm_backlog); 12697 #endif 12698 INIT_LIST_HEAD(&sd->poll_list); 12699 sd->output_queue_tailp = &sd->output_queue; 12700 #ifdef CONFIG_RPS 12701 INIT_CSD(&sd->csd, rps_trigger_softirq, sd); 12702 sd->cpu = i; 12703 #endif 12704 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd); 12705 spin_lock_init(&sd->defer_lock); 12706 12707 gro_init(&sd->backlog.gro); 12708 sd->backlog.poll = process_backlog; 12709 sd->backlog.weight = weight_p; 12710 INIT_LIST_HEAD(&sd->backlog.poll_list); 12711 12712 if (net_page_pool_create(i)) 12713 goto out; 12714 } 12715 if (use_backlog_threads()) 12716 smpboot_register_percpu_thread(&backlog_threads); 12717 12718 dev_boot_phase = 0; 12719 12720 /* The loopback device is special if any other network devices 12721 * is present in a network namespace the loopback device must 12722 * be present. Since we now dynamically allocate and free the 12723 * loopback device ensure this invariant is maintained by 12724 * keeping the loopback device as the first device on the 12725 * list of network devices. Ensuring the loopback devices 12726 * is the first device that appears and the last network device 12727 * that disappears. 12728 */ 12729 if (register_pernet_device(&loopback_net_ops)) 12730 goto out; 12731 12732 if (register_pernet_device(&default_device_ops)) 12733 goto out; 12734 12735 open_softirq(NET_TX_SOFTIRQ, net_tx_action); 12736 open_softirq(NET_RX_SOFTIRQ, net_rx_action); 12737 12738 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead", 12739 NULL, dev_cpu_dead); 12740 WARN_ON(rc < 0); 12741 rc = 0; 12742 12743 /* avoid static key IPIs to isolated CPUs */ 12744 if (housekeeping_enabled(HK_TYPE_MISC)) 12745 net_enable_timestamp(); 12746 out: 12747 if (rc < 0) { 12748 for_each_possible_cpu(i) { 12749 struct page_pool *pp_ptr; 12750 12751 pp_ptr = per_cpu(system_page_pool, i); 12752 if (!pp_ptr) 12753 continue; 12754 12755 xdp_unreg_page_pool(pp_ptr); 12756 page_pool_destroy(pp_ptr); 12757 per_cpu(system_page_pool, i) = NULL; 12758 } 12759 } 12760 12761 return rc; 12762 } 12763 12764 subsys_initcall(net_dev_init); 12765