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