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_PREEMPT_RT)) { 235 spin_lock_irqsave(&sd->input_pkt_queue.lock, *flags); 236 } else { 237 local_irq_save(*flags); 238 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 239 spin_lock(&sd->input_pkt_queue.lock); 240 } 241 } 242 243 static inline void backlog_lock_irq_disable(struct softnet_data *sd) 244 { 245 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 246 spin_lock_irq(&sd->input_pkt_queue.lock); 247 else 248 local_irq_disable(); 249 } 250 251 static inline void backlog_unlock_irq_restore(struct softnet_data *sd, 252 unsigned long flags) 253 { 254 if (IS_ENABLED(CONFIG_PREEMPT_RT)) { 255 spin_unlock_irqrestore(&sd->input_pkt_queue.lock, flags); 256 } else { 257 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 258 spin_unlock(&sd->input_pkt_queue.lock); 259 local_irq_restore(flags); 260 } 261 } 262 263 static inline void backlog_unlock_irq_enable(struct softnet_data *sd) 264 { 265 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 266 spin_unlock_irq(&sd->input_pkt_queue.lock); 267 else 268 local_irq_enable(); 269 } 270 271 static struct netdev_name_node *netdev_name_node_alloc(struct net_device *dev, 272 const char *name) 273 { 274 struct netdev_name_node *name_node; 275 276 name_node = kmalloc_obj(*name_node, GFP_KERNEL); 277 if (!name_node) 278 return NULL; 279 INIT_HLIST_NODE(&name_node->hlist); 280 name_node->dev = dev; 281 name_node->name = name; 282 return name_node; 283 } 284 285 static struct netdev_name_node * 286 netdev_name_node_head_alloc(struct net_device *dev) 287 { 288 struct netdev_name_node *name_node; 289 290 name_node = netdev_name_node_alloc(dev, dev->name); 291 if (!name_node) 292 return NULL; 293 INIT_LIST_HEAD(&name_node->list); 294 return name_node; 295 } 296 297 static void netdev_name_node_free(struct netdev_name_node *name_node) 298 { 299 kfree(name_node); 300 } 301 302 static void netdev_name_node_add(struct net *net, 303 struct netdev_name_node *name_node) 304 { 305 hlist_add_head_rcu(&name_node->hlist, 306 dev_name_hash(net, name_node->name)); 307 } 308 309 static void netdev_name_node_del(struct netdev_name_node *name_node) 310 { 311 hlist_del_rcu(&name_node->hlist); 312 } 313 314 static struct netdev_name_node *netdev_name_node_lookup(struct net *net, 315 const char *name) 316 { 317 struct hlist_head *head = dev_name_hash(net, name); 318 struct netdev_name_node *name_node; 319 320 hlist_for_each_entry(name_node, head, hlist) 321 if (!strcmp(name_node->name, name)) 322 return name_node; 323 return NULL; 324 } 325 326 static struct netdev_name_node *netdev_name_node_lookup_rcu(struct net *net, 327 const char *name) 328 { 329 struct hlist_head *head = dev_name_hash(net, name); 330 struct netdev_name_node *name_node; 331 332 hlist_for_each_entry_rcu(name_node, head, hlist) 333 if (!strcmp(name_node->name, name)) 334 return name_node; 335 return NULL; 336 } 337 338 bool netdev_name_in_use(struct net *net, const char *name) 339 { 340 return netdev_name_node_lookup(net, name); 341 } 342 EXPORT_SYMBOL(netdev_name_in_use); 343 344 int netdev_name_node_alt_create(struct net_device *dev, const char *name) 345 { 346 struct netdev_name_node *name_node; 347 struct net *net = dev_net(dev); 348 349 name_node = netdev_name_node_lookup(net, name); 350 if (name_node) 351 return -EEXIST; 352 name_node = netdev_name_node_alloc(dev, name); 353 if (!name_node) 354 return -ENOMEM; 355 netdev_name_node_add(net, name_node); 356 /* The node that holds dev->name acts as a head of per-device list. */ 357 list_add_tail_rcu(&name_node->list, &dev->name_node->list); 358 359 return 0; 360 } 361 362 static void netdev_name_node_alt_free(struct rcu_head *head) 363 { 364 struct netdev_name_node *name_node = 365 container_of(head, struct netdev_name_node, rcu); 366 367 kfree(name_node->name); 368 netdev_name_node_free(name_node); 369 } 370 371 static void __netdev_name_node_alt_destroy(struct netdev_name_node *name_node) 372 { 373 netdev_name_node_del(name_node); 374 list_del(&name_node->list); 375 call_rcu(&name_node->rcu, netdev_name_node_alt_free); 376 } 377 378 int netdev_name_node_alt_destroy(struct net_device *dev, const char *name) 379 { 380 struct netdev_name_node *name_node; 381 struct net *net = dev_net(dev); 382 383 name_node = netdev_name_node_lookup(net, name); 384 if (!name_node) 385 return -ENOENT; 386 /* lookup might have found our primary name or a name belonging 387 * to another device. 388 */ 389 if (name_node == dev->name_node || name_node->dev != dev) 390 return -EINVAL; 391 392 __netdev_name_node_alt_destroy(name_node); 393 return 0; 394 } 395 396 static void netdev_name_node_alt_flush(struct net_device *dev) 397 { 398 struct netdev_name_node *name_node, *tmp; 399 400 list_for_each_entry_safe(name_node, tmp, &dev->name_node->list, list) { 401 list_del(&name_node->list); 402 netdev_name_node_alt_free(&name_node->rcu); 403 } 404 } 405 406 /* Device list insertion */ 407 static void list_netdevice(struct net_device *dev) 408 { 409 struct netdev_name_node *name_node; 410 struct net *net = dev_net(dev); 411 412 ASSERT_RTNL(); 413 414 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head); 415 netdev_name_node_add(net, dev->name_node); 416 hlist_add_head_rcu(&dev->index_hlist, 417 dev_index_hash(net, dev->ifindex)); 418 419 netdev_for_each_altname(dev, name_node) 420 netdev_name_node_add(net, name_node); 421 422 /* We reserved the ifindex, this can't fail */ 423 WARN_ON(xa_store(&net->dev_by_index, dev->ifindex, dev, GFP_KERNEL)); 424 425 dev_base_seq_inc(net); 426 } 427 428 /* Device list removal 429 * caller must respect a RCU grace period before freeing/reusing dev 430 */ 431 static void unlist_netdevice(struct net_device *dev) 432 { 433 struct netdev_name_node *name_node; 434 struct net *net = dev_net(dev); 435 436 ASSERT_RTNL(); 437 438 xa_erase(&net->dev_by_index, dev->ifindex); 439 440 netdev_for_each_altname(dev, name_node) 441 netdev_name_node_del(name_node); 442 443 /* Unlink dev from the device chain */ 444 list_del_rcu(&dev->dev_list); 445 netdev_name_node_del(dev->name_node); 446 hlist_del_rcu(&dev->index_hlist); 447 448 dev_base_seq_inc(dev_net(dev)); 449 } 450 451 /* 452 * Our notifier list 453 */ 454 455 static RAW_NOTIFIER_HEAD(netdev_chain); 456 457 /* 458 * Device drivers call our routines to queue packets here. We empty the 459 * queue in the local softnet handler. 460 */ 461 462 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data) = { 463 .process_queue_bh_lock = INIT_LOCAL_LOCK(process_queue_bh_lock), 464 }; 465 EXPORT_PER_CPU_SYMBOL(softnet_data); 466 467 /* Page_pool has a lockless array/stack to alloc/recycle pages. 468 * PP consumers must pay attention to run APIs in the appropriate context 469 * (e.g. NAPI context). 470 */ 471 DEFINE_PER_CPU(struct page_pool_bh, system_page_pool) = { 472 .bh_lock = INIT_LOCAL_LOCK(bh_lock), 473 }; 474 475 #ifdef CONFIG_LOCKDEP 476 /* 477 * register_netdevice() inits txq->_xmit_lock and sets lockdep class 478 * according to dev->type 479 */ 480 static const unsigned short netdev_lock_type[] = { 481 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25, 482 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET, 483 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM, 484 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP, 485 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD, 486 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25, 487 ARPHRD_CAN, ARPHRD_MCTP, 488 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP, 489 ARPHRD_RAWHDLC, ARPHRD_RAWIP, 490 ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD, 491 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI, 492 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE, 493 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET, 494 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL, 495 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM, 496 ARPHRD_IEEE80211_RADIOTAP, 497 ARPHRD_IEEE802154, ARPHRD_IEEE802154_MONITOR, 498 ARPHRD_PHONET, ARPHRD_PHONET_PIPE, 499 ARPHRD_CAIF, ARPHRD_IP6GRE, ARPHRD_NETLINK, ARPHRD_6LOWPAN, 500 ARPHRD_VSOCKMON, 501 ARPHRD_VOID, ARPHRD_NONE}; 502 503 static const char *const netdev_lock_name[] = { 504 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25", 505 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET", 506 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM", 507 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP", 508 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD", 509 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25", 510 "_xmit_CAN", "_xmit_MCTP", 511 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP", 512 "_xmit_RAWHDLC", "_xmit_RAWIP", 513 "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD", 514 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI", 515 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE", 516 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET", 517 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL", 518 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM", 519 "_xmit_IEEE80211_RADIOTAP", 520 "_xmit_IEEE802154", "_xmit_IEEE802154_MONITOR", 521 "_xmit_PHONET", "_xmit_PHONET_PIPE", 522 "_xmit_CAIF", "_xmit_IP6GRE", "_xmit_NETLINK", "_xmit_6LOWPAN", 523 "_xmit_VSOCKMON", 524 "_xmit_VOID", "_xmit_NONE"}; 525 526 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)]; 527 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)]; 528 529 static inline unsigned short netdev_lock_pos(unsigned short dev_type) 530 { 531 int i; 532 533 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++) 534 if (netdev_lock_type[i] == dev_type) 535 return i; 536 /* the last key is used by default */ 537 WARN_ONCE(1, "netdev_lock_pos() could not find dev_type=%u\n", dev_type); 538 return ARRAY_SIZE(netdev_lock_type) - 1; 539 } 540 541 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 542 unsigned short dev_type) 543 { 544 int i; 545 546 i = netdev_lock_pos(dev_type); 547 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i], 548 netdev_lock_name[i]); 549 } 550 551 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 552 { 553 int i; 554 555 i = netdev_lock_pos(dev->type); 556 lockdep_set_class_and_name(&dev->addr_list_lock, 557 &netdev_addr_lock_key[i], 558 netdev_lock_name[i]); 559 } 560 #else 561 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 562 unsigned short dev_type) 563 { 564 } 565 566 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 567 { 568 } 569 #endif 570 571 /******************************************************************************* 572 * 573 * Protocol management and registration routines 574 * 575 *******************************************************************************/ 576 577 578 /* 579 * Add a protocol ID to the list. Now that the input handler is 580 * smarter we can dispense with all the messy stuff that used to be 581 * here. 582 * 583 * BEWARE!!! Protocol handlers, mangling input packets, 584 * MUST BE last in hash buckets and checking protocol handlers 585 * MUST start from promiscuous ptype_all chain in net_bh. 586 * It is true now, do not change it. 587 * Explanation follows: if protocol handler, mangling packet, will 588 * be the first on list, it is not able to sense, that packet 589 * is cloned and should be copied-on-write, so that it will 590 * change it and subsequent readers will get broken packet. 591 * --ANK (980803) 592 */ 593 594 static inline struct list_head *ptype_head(const struct packet_type *pt) 595 { 596 if (pt->type == htons(ETH_P_ALL)) { 597 if (!pt->af_packet_net && !pt->dev) 598 return NULL; 599 600 return pt->dev ? &pt->dev->ptype_all : 601 &pt->af_packet_net->ptype_all; 602 } 603 604 if (pt->dev) 605 return &pt->dev->ptype_specific; 606 607 return pt->af_packet_net ? &pt->af_packet_net->ptype_specific : 608 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK]; 609 } 610 611 /** 612 * dev_add_pack - add packet handler 613 * @pt: packet type declaration 614 * 615 * Add a protocol handler to the networking stack. The passed &packet_type 616 * is linked into kernel lists and may not be freed until it has been 617 * removed from the kernel lists. 618 * 619 * This call does not sleep therefore it can not 620 * guarantee all CPU's that are in middle of receiving packets 621 * will see the new packet type (until the next received packet). 622 */ 623 624 void dev_add_pack(struct packet_type *pt) 625 { 626 struct list_head *head = ptype_head(pt); 627 628 if (WARN_ON_ONCE(!head)) 629 return; 630 631 spin_lock(&ptype_lock); 632 list_add_rcu(&pt->list, head); 633 spin_unlock(&ptype_lock); 634 } 635 EXPORT_SYMBOL(dev_add_pack); 636 637 /** 638 * __dev_remove_pack - remove packet handler 639 * @pt: packet type declaration 640 * 641 * Remove a protocol handler that was previously added to the kernel 642 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 643 * from the kernel lists and can be freed or reused once this function 644 * returns. 645 * 646 * The packet type might still be in use by receivers 647 * and must not be freed until after all the CPU's have gone 648 * through a quiescent state. 649 */ 650 void __dev_remove_pack(struct packet_type *pt) 651 { 652 struct list_head *head = ptype_head(pt); 653 struct packet_type *pt1; 654 655 if (!head) 656 return; 657 658 spin_lock(&ptype_lock); 659 660 list_for_each_entry(pt1, head, list) { 661 if (pt == pt1) { 662 list_del_rcu(&pt->list); 663 goto out; 664 } 665 } 666 667 pr_warn("dev_remove_pack: %p not found\n", pt); 668 out: 669 spin_unlock(&ptype_lock); 670 } 671 EXPORT_SYMBOL(__dev_remove_pack); 672 673 /** 674 * dev_remove_pack - remove packet handler 675 * @pt: packet type declaration 676 * 677 * Remove a protocol handler that was previously added to the kernel 678 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 679 * from the kernel lists and can be freed or reused once this function 680 * returns. 681 * 682 * This call sleeps to guarantee that no CPU is looking at the packet 683 * type after return. 684 */ 685 void dev_remove_pack(struct packet_type *pt) 686 { 687 __dev_remove_pack(pt); 688 689 synchronize_net(); 690 } 691 EXPORT_SYMBOL(dev_remove_pack); 692 693 694 /******************************************************************************* 695 * 696 * Device Interface Subroutines 697 * 698 *******************************************************************************/ 699 700 /** 701 * dev_get_iflink - get 'iflink' value of a interface 702 * @dev: targeted interface 703 * 704 * Indicates the ifindex the interface is linked to. 705 * Physical interfaces have the same 'ifindex' and 'iflink' values. 706 */ 707 708 int dev_get_iflink(const struct net_device *dev) 709 { 710 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink) 711 return dev->netdev_ops->ndo_get_iflink(dev); 712 713 return READ_ONCE(dev->ifindex); 714 } 715 EXPORT_SYMBOL(dev_get_iflink); 716 717 /** 718 * dev_fill_metadata_dst - Retrieve tunnel egress information. 719 * @dev: targeted interface 720 * @skb: The packet. 721 * 722 * For better visibility of tunnel traffic OVS needs to retrieve 723 * egress tunnel information for a packet. Following API allows 724 * user to get this info. 725 */ 726 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb) 727 { 728 struct ip_tunnel_info *info; 729 730 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst) 731 return -EINVAL; 732 733 info = skb_tunnel_info_unclone(skb); 734 if (!info) 735 return -ENOMEM; 736 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX))) 737 return -EINVAL; 738 739 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb); 740 } 741 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst); 742 743 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack) 744 { 745 int k = stack->num_paths++; 746 747 if (k >= NET_DEVICE_PATH_STACK_MAX) 748 return NULL; 749 750 return &stack->path[k]; 751 } 752 753 int dev_fill_forward_path(const struct net_device *dev, const u8 *daddr, 754 struct net_device_path_stack *stack) 755 { 756 const struct net_device *last_dev; 757 struct net_device_path_ctx ctx = { 758 .dev = dev, 759 }; 760 struct net_device_path *path; 761 int ret = 0; 762 763 memcpy(ctx.daddr, daddr, sizeof(ctx.daddr)); 764 stack->num_paths = 0; 765 while (ctx.dev && ctx.dev->netdev_ops->ndo_fill_forward_path) { 766 last_dev = ctx.dev; 767 path = dev_fwd_path(stack); 768 if (!path) 769 return -1; 770 771 memset(path, 0, sizeof(struct net_device_path)); 772 ret = ctx.dev->netdev_ops->ndo_fill_forward_path(&ctx, path); 773 if (ret < 0) 774 return -1; 775 776 if (WARN_ON_ONCE(last_dev == ctx.dev)) 777 return -1; 778 } 779 780 if (!ctx.dev) 781 return ret; 782 783 path = dev_fwd_path(stack); 784 if (!path) 785 return -1; 786 path->type = DEV_PATH_ETHERNET; 787 path->dev = ctx.dev; 788 789 return ret; 790 } 791 EXPORT_SYMBOL_GPL(dev_fill_forward_path); 792 793 /* must be called under rcu_read_lock(), as we dont take a reference */ 794 static struct napi_struct *napi_by_id(unsigned int napi_id) 795 { 796 unsigned int hash = napi_id % HASH_SIZE(napi_hash); 797 struct napi_struct *napi; 798 799 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node) 800 if (napi->napi_id == napi_id) 801 return napi; 802 803 return NULL; 804 } 805 806 /* must be called under rcu_read_lock(), as we dont take a reference */ 807 static struct napi_struct * 808 netdev_napi_by_id(struct net *net, unsigned int napi_id) 809 { 810 struct napi_struct *napi; 811 812 napi = napi_by_id(napi_id); 813 if (!napi) 814 return NULL; 815 816 if (WARN_ON_ONCE(!napi->dev)) 817 return NULL; 818 if (!net_eq(net, dev_net(napi->dev))) 819 return NULL; 820 821 return napi; 822 } 823 824 /** 825 * netdev_napi_by_id_lock() - find a device by NAPI ID and lock it 826 * @net: the applicable net namespace 827 * @napi_id: ID of a NAPI of a target device 828 * 829 * Find a NAPI instance with @napi_id. Lock its device. 830 * The device must be in %NETREG_REGISTERED state for lookup to succeed. 831 * netdev_unlock() must be called to release it. 832 * 833 * Return: pointer to NAPI, its device with lock held, NULL if not found. 834 */ 835 struct napi_struct * 836 netdev_napi_by_id_lock(struct net *net, unsigned int napi_id) 837 { 838 struct napi_struct *napi; 839 struct net_device *dev; 840 841 rcu_read_lock(); 842 napi = netdev_napi_by_id(net, napi_id); 843 if (!napi || READ_ONCE(napi->dev->reg_state) != NETREG_REGISTERED) { 844 rcu_read_unlock(); 845 return NULL; 846 } 847 848 dev = napi->dev; 849 dev_hold(dev); 850 rcu_read_unlock(); 851 852 dev = __netdev_put_lock(dev, net); 853 if (!dev) 854 return NULL; 855 856 rcu_read_lock(); 857 napi = netdev_napi_by_id(net, napi_id); 858 if (napi && napi->dev != dev) 859 napi = NULL; 860 rcu_read_unlock(); 861 862 if (!napi) 863 netdev_unlock(dev); 864 return napi; 865 } 866 867 /** 868 * __dev_get_by_name - find a device by its name 869 * @net: the applicable net namespace 870 * @name: name to find 871 * 872 * Find an interface by name. Must be called under RTNL semaphore. 873 * If the name is found a pointer to the device is returned. 874 * If the name is not found then %NULL is returned. The 875 * reference counters are not incremented so the caller must be 876 * careful with locks. 877 */ 878 879 struct net_device *__dev_get_by_name(struct net *net, const char *name) 880 { 881 struct netdev_name_node *node_name; 882 883 node_name = netdev_name_node_lookup(net, name); 884 return node_name ? node_name->dev : NULL; 885 } 886 EXPORT_SYMBOL(__dev_get_by_name); 887 888 /** 889 * dev_get_by_name_rcu - find a device by its name 890 * @net: the applicable net namespace 891 * @name: name to find 892 * 893 * Find an interface by name. 894 * If the name is found a pointer to the device is returned. 895 * If the name is not found then %NULL is returned. 896 * The reference counters are not incremented so the caller must be 897 * careful with locks. The caller must hold RCU lock. 898 */ 899 900 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name) 901 { 902 struct netdev_name_node *node_name; 903 904 node_name = netdev_name_node_lookup_rcu(net, name); 905 return node_name ? node_name->dev : NULL; 906 } 907 EXPORT_SYMBOL(dev_get_by_name_rcu); 908 909 /* Deprecated for new users, call netdev_get_by_name() instead */ 910 struct net_device *dev_get_by_name(struct net *net, const char *name) 911 { 912 struct net_device *dev; 913 914 rcu_read_lock(); 915 dev = dev_get_by_name_rcu(net, name); 916 dev_hold(dev); 917 rcu_read_unlock(); 918 return dev; 919 } 920 EXPORT_SYMBOL(dev_get_by_name); 921 922 /** 923 * netdev_get_by_name() - find a device by its name 924 * @net: the applicable net namespace 925 * @name: name to find 926 * @tracker: tracking object for the acquired reference 927 * @gfp: allocation flags for the tracker 928 * 929 * Find an interface by name. This can be called from any 930 * context and does its own locking. The returned handle has 931 * the usage count incremented and the caller must use netdev_put() to 932 * release it when it is no longer needed. %NULL is returned if no 933 * matching device is found. 934 */ 935 struct net_device *netdev_get_by_name(struct net *net, const char *name, 936 netdevice_tracker *tracker, gfp_t gfp) 937 { 938 struct net_device *dev; 939 940 dev = dev_get_by_name(net, name); 941 if (dev) 942 netdev_tracker_alloc(dev, tracker, gfp); 943 return dev; 944 } 945 EXPORT_SYMBOL(netdev_get_by_name); 946 947 /** 948 * __dev_get_by_index - find a device by its ifindex 949 * @net: the applicable net namespace 950 * @ifindex: index of device 951 * 952 * Search for an interface by index. Returns %NULL if the device 953 * is not found or a pointer to the device. The device has not 954 * had its reference counter increased so the caller must be careful 955 * about locking. The caller must hold the RTNL semaphore. 956 */ 957 958 struct net_device *__dev_get_by_index(struct net *net, int ifindex) 959 { 960 struct net_device *dev; 961 struct hlist_head *head = dev_index_hash(net, ifindex); 962 963 hlist_for_each_entry(dev, head, index_hlist) 964 if (dev->ifindex == ifindex) 965 return dev; 966 967 return NULL; 968 } 969 EXPORT_SYMBOL(__dev_get_by_index); 970 971 /** 972 * dev_get_by_index_rcu - find a device by its ifindex 973 * @net: the applicable net namespace 974 * @ifindex: index of device 975 * 976 * Search for an interface by index. Returns %NULL if the device 977 * is not found or a pointer to the device. The device has not 978 * had its reference counter increased so the caller must be careful 979 * about locking. The caller must hold RCU lock. 980 */ 981 982 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex) 983 { 984 struct net_device *dev; 985 struct hlist_head *head = dev_index_hash(net, ifindex); 986 987 hlist_for_each_entry_rcu(dev, head, index_hlist) 988 if (dev->ifindex == ifindex) 989 return dev; 990 991 return NULL; 992 } 993 EXPORT_SYMBOL(dev_get_by_index_rcu); 994 995 /* Deprecated for new users, call netdev_get_by_index() instead */ 996 struct net_device *dev_get_by_index(struct net *net, int ifindex) 997 { 998 struct net_device *dev; 999 1000 rcu_read_lock(); 1001 dev = dev_get_by_index_rcu(net, ifindex); 1002 dev_hold(dev); 1003 rcu_read_unlock(); 1004 return dev; 1005 } 1006 EXPORT_SYMBOL(dev_get_by_index); 1007 1008 /** 1009 * netdev_get_by_index() - find a device by its ifindex 1010 * @net: the applicable net namespace 1011 * @ifindex: index of device 1012 * @tracker: tracking object for the acquired reference 1013 * @gfp: allocation flags for the tracker 1014 * 1015 * Search for an interface by index. Returns NULL if the device 1016 * is not found or a pointer to the device. The device returned has 1017 * had a reference added and the pointer is safe until the user calls 1018 * netdev_put() to indicate they have finished with it. 1019 */ 1020 struct net_device *netdev_get_by_index(struct net *net, int ifindex, 1021 netdevice_tracker *tracker, gfp_t gfp) 1022 { 1023 struct net_device *dev; 1024 1025 dev = dev_get_by_index(net, ifindex); 1026 if (dev) 1027 netdev_tracker_alloc(dev, tracker, gfp); 1028 return dev; 1029 } 1030 EXPORT_SYMBOL(netdev_get_by_index); 1031 1032 /** 1033 * dev_get_by_napi_id - find a device by napi_id 1034 * @napi_id: ID of the NAPI struct 1035 * 1036 * Search for an interface by NAPI ID. Returns %NULL if the device 1037 * is not found or a pointer to the device. The device has not had 1038 * its reference counter increased so the caller must be careful 1039 * about locking. The caller must hold RCU lock. 1040 */ 1041 struct net_device *dev_get_by_napi_id(unsigned int napi_id) 1042 { 1043 struct napi_struct *napi; 1044 1045 WARN_ON_ONCE(!rcu_read_lock_held()); 1046 1047 if (!napi_id_valid(napi_id)) 1048 return NULL; 1049 1050 napi = napi_by_id(napi_id); 1051 1052 return napi ? napi->dev : NULL; 1053 } 1054 1055 /* Release the held reference on the net_device, and if the net_device 1056 * is still registered try to lock the instance lock. If device is being 1057 * unregistered NULL will be returned (but the reference has been released, 1058 * either way!) 1059 * 1060 * This helper is intended for locking net_device after it has been looked up 1061 * using a lockless lookup helper. Lock prevents the instance from going away. 1062 */ 1063 struct net_device *__netdev_put_lock(struct net_device *dev, struct net *net) 1064 { 1065 netdev_lock(dev); 1066 if (dev->reg_state > NETREG_REGISTERED || 1067 dev->moving_ns || !net_eq(dev_net(dev), net)) { 1068 netdev_unlock(dev); 1069 dev_put(dev); 1070 return NULL; 1071 } 1072 dev_put(dev); 1073 return dev; 1074 } 1075 1076 static struct net_device * 1077 __netdev_put_lock_ops_compat(struct net_device *dev, struct net *net) 1078 { 1079 netdev_lock_ops_compat(dev); 1080 if (dev->reg_state > NETREG_REGISTERED || 1081 dev->moving_ns || !net_eq(dev_net(dev), net)) { 1082 netdev_unlock_ops_compat(dev); 1083 dev_put(dev); 1084 return NULL; 1085 } 1086 dev_put(dev); 1087 return dev; 1088 } 1089 1090 /** 1091 * netdev_get_by_index_lock() - find a device by its ifindex 1092 * @net: the applicable net namespace 1093 * @ifindex: index of device 1094 * 1095 * Search for an interface by index. If a valid device 1096 * with @ifindex is found it will be returned with netdev->lock held. 1097 * netdev_unlock() must be called to release it. 1098 * 1099 * Return: pointer to a device with lock held, NULL if not found. 1100 */ 1101 struct net_device *netdev_get_by_index_lock(struct net *net, int ifindex) 1102 { 1103 struct net_device *dev; 1104 1105 dev = dev_get_by_index(net, ifindex); 1106 if (!dev) 1107 return NULL; 1108 1109 return __netdev_put_lock(dev, net); 1110 } 1111 1112 struct net_device * 1113 netdev_get_by_index_lock_ops_compat(struct net *net, int ifindex) 1114 { 1115 struct net_device *dev; 1116 1117 dev = dev_get_by_index(net, ifindex); 1118 if (!dev) 1119 return NULL; 1120 1121 return __netdev_put_lock_ops_compat(dev, net); 1122 } 1123 1124 struct net_device * 1125 netdev_xa_find_lock(struct net *net, struct net_device *dev, 1126 unsigned long *index) 1127 { 1128 if (dev) 1129 netdev_unlock(dev); 1130 1131 do { 1132 rcu_read_lock(); 1133 dev = xa_find(&net->dev_by_index, index, ULONG_MAX, XA_PRESENT); 1134 if (!dev) { 1135 rcu_read_unlock(); 1136 return NULL; 1137 } 1138 dev_hold(dev); 1139 rcu_read_unlock(); 1140 1141 dev = __netdev_put_lock(dev, net); 1142 if (dev) 1143 return dev; 1144 1145 (*index)++; 1146 } while (true); 1147 } 1148 1149 struct net_device * 1150 netdev_xa_find_lock_ops_compat(struct net *net, struct net_device *dev, 1151 unsigned long *index) 1152 { 1153 if (dev) 1154 netdev_unlock_ops_compat(dev); 1155 1156 do { 1157 rcu_read_lock(); 1158 dev = xa_find(&net->dev_by_index, index, ULONG_MAX, XA_PRESENT); 1159 if (!dev) { 1160 rcu_read_unlock(); 1161 return NULL; 1162 } 1163 dev_hold(dev); 1164 rcu_read_unlock(); 1165 1166 dev = __netdev_put_lock_ops_compat(dev, net); 1167 if (dev) 1168 return dev; 1169 1170 (*index)++; 1171 } while (true); 1172 } 1173 1174 static DEFINE_SEQLOCK(netdev_rename_lock); 1175 1176 void netdev_copy_name(struct net_device *dev, char *name) 1177 { 1178 unsigned int seq; 1179 1180 do { 1181 seq = read_seqbegin(&netdev_rename_lock); 1182 strscpy(name, dev->name, IFNAMSIZ); 1183 } while (read_seqretry(&netdev_rename_lock, seq)); 1184 } 1185 EXPORT_IPV6_MOD_GPL(netdev_copy_name); 1186 1187 /** 1188 * netdev_get_name - get a netdevice name, knowing its ifindex. 1189 * @net: network namespace 1190 * @name: a pointer to the buffer where the name will be stored. 1191 * @ifindex: the ifindex of the interface to get the name from. 1192 */ 1193 int netdev_get_name(struct net *net, char *name, int ifindex) 1194 { 1195 struct net_device *dev; 1196 int ret; 1197 1198 rcu_read_lock(); 1199 1200 dev = dev_get_by_index_rcu(net, ifindex); 1201 if (!dev) { 1202 ret = -ENODEV; 1203 goto out; 1204 } 1205 1206 netdev_copy_name(dev, name); 1207 1208 ret = 0; 1209 out: 1210 rcu_read_unlock(); 1211 return ret; 1212 } 1213 1214 static bool dev_addr_cmp(struct net_device *dev, unsigned short type, 1215 const char *ha) 1216 { 1217 return dev->type == type && !memcmp(dev->dev_addr, ha, dev->addr_len); 1218 } 1219 1220 /** 1221 * dev_getbyhwaddr_rcu - find a device by its hardware address 1222 * @net: the applicable net namespace 1223 * @type: media type of device 1224 * @ha: hardware address 1225 * 1226 * Search for an interface by MAC address. Returns NULL if the device 1227 * is not found or a pointer to the device. 1228 * The caller must hold RCU. 1229 * The returned device has not had its ref count increased 1230 * and the caller must therefore be careful about locking 1231 * 1232 */ 1233 1234 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type, 1235 const char *ha) 1236 { 1237 struct net_device *dev; 1238 1239 for_each_netdev_rcu(net, dev) 1240 if (dev_addr_cmp(dev, type, ha)) 1241 return dev; 1242 1243 return NULL; 1244 } 1245 EXPORT_SYMBOL(dev_getbyhwaddr_rcu); 1246 1247 /** 1248 * dev_getbyhwaddr() - find a device by its hardware address 1249 * @net: the applicable net namespace 1250 * @type: media type of device 1251 * @ha: hardware address 1252 * 1253 * Similar to dev_getbyhwaddr_rcu(), but the owner needs to hold 1254 * rtnl_lock. 1255 * 1256 * Context: rtnl_lock() must be held. 1257 * Return: pointer to the net_device, or NULL if not found 1258 */ 1259 struct net_device *dev_getbyhwaddr(struct net *net, unsigned short type, 1260 const char *ha) 1261 { 1262 struct net_device *dev; 1263 1264 ASSERT_RTNL(); 1265 for_each_netdev(net, dev) 1266 if (dev_addr_cmp(dev, type, ha)) 1267 return dev; 1268 1269 return NULL; 1270 } 1271 EXPORT_SYMBOL(dev_getbyhwaddr); 1272 1273 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type) 1274 { 1275 struct net_device *dev, *ret = NULL; 1276 1277 rcu_read_lock(); 1278 for_each_netdev_rcu(net, dev) 1279 if (dev->type == type) { 1280 dev_hold(dev); 1281 ret = dev; 1282 break; 1283 } 1284 rcu_read_unlock(); 1285 return ret; 1286 } 1287 EXPORT_SYMBOL(dev_getfirstbyhwtype); 1288 1289 /** 1290 * netdev_get_by_flags_rcu - find any device with given flags 1291 * @net: the applicable net namespace 1292 * @tracker: tracking object for the acquired reference 1293 * @if_flags: IFF_* values 1294 * @mask: bitmask of bits in if_flags to check 1295 * 1296 * Search for any interface with the given flags. 1297 * 1298 * Context: rcu_read_lock() must be held. 1299 * Returns: NULL if a device is not found or a pointer to the device. 1300 */ 1301 struct net_device *netdev_get_by_flags_rcu(struct net *net, netdevice_tracker *tracker, 1302 unsigned short if_flags, unsigned short mask) 1303 { 1304 struct net_device *dev; 1305 1306 for_each_netdev_rcu(net, dev) { 1307 if (((READ_ONCE(dev->flags) ^ if_flags) & mask) == 0) { 1308 netdev_hold(dev, tracker, GFP_ATOMIC); 1309 return dev; 1310 } 1311 } 1312 1313 return NULL; 1314 } 1315 EXPORT_IPV6_MOD(netdev_get_by_flags_rcu); 1316 1317 /** 1318 * dev_valid_name - check if name is okay for network device 1319 * @name: name string 1320 * 1321 * Network device names need to be valid file names to 1322 * allow sysfs to work. We also disallow any kind of 1323 * whitespace. 1324 */ 1325 bool dev_valid_name(const char *name) 1326 { 1327 if (*name == '\0') 1328 return false; 1329 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ) 1330 return false; 1331 if (!strcmp(name, ".") || !strcmp(name, "..")) 1332 return false; 1333 1334 while (*name) { 1335 if (*name == '/' || *name == ':' || isspace(*name)) 1336 return false; 1337 name++; 1338 } 1339 return true; 1340 } 1341 EXPORT_SYMBOL(dev_valid_name); 1342 1343 /** 1344 * __dev_alloc_name - allocate a name for a device 1345 * @net: network namespace to allocate the device name in 1346 * @name: name format string 1347 * @res: result name string 1348 * 1349 * Passed a format string - eg "lt%d" it will try and find a suitable 1350 * id. It scans list of devices to build up a free map, then chooses 1351 * the first empty slot. The caller must hold the dev_base or rtnl lock 1352 * while allocating the name and adding the device in order to avoid 1353 * duplicates. 1354 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1355 * Returns the number of the unit assigned or a negative errno code. 1356 */ 1357 1358 static int __dev_alloc_name(struct net *net, const char *name, char *res) 1359 { 1360 int i = 0; 1361 const char *p; 1362 const int max_netdevices = 8*PAGE_SIZE; 1363 unsigned long *inuse; 1364 struct net_device *d; 1365 char buf[IFNAMSIZ]; 1366 1367 /* Verify the string as this thing may have come from the user. 1368 * There must be one "%d" and no other "%" characters. 1369 */ 1370 p = strchr(name, '%'); 1371 if (!p || p[1] != 'd' || strchr(p + 2, '%')) 1372 return -EINVAL; 1373 1374 /* Use one page as a bit array of possible slots */ 1375 inuse = bitmap_zalloc(max_netdevices, GFP_ATOMIC); 1376 if (!inuse) 1377 return -ENOMEM; 1378 1379 for_each_netdev(net, d) { 1380 struct netdev_name_node *name_node; 1381 1382 netdev_for_each_altname(d, name_node) { 1383 if (!sscanf(name_node->name, name, &i)) 1384 continue; 1385 if (i < 0 || i >= max_netdevices) 1386 continue; 1387 1388 /* avoid cases where sscanf is not exact inverse of printf */ 1389 snprintf(buf, IFNAMSIZ, name, i); 1390 if (!strncmp(buf, name_node->name, IFNAMSIZ)) 1391 __set_bit(i, inuse); 1392 } 1393 if (!sscanf(d->name, name, &i)) 1394 continue; 1395 if (i < 0 || i >= max_netdevices) 1396 continue; 1397 1398 /* avoid cases where sscanf is not exact inverse of printf */ 1399 snprintf(buf, IFNAMSIZ, name, i); 1400 if (!strncmp(buf, d->name, IFNAMSIZ)) 1401 __set_bit(i, inuse); 1402 } 1403 1404 i = find_first_zero_bit(inuse, max_netdevices); 1405 bitmap_free(inuse); 1406 if (i == max_netdevices) 1407 return -ENFILE; 1408 1409 /* 'res' and 'name' could overlap, use 'buf' as an intermediate buffer */ 1410 strscpy(buf, name, IFNAMSIZ); 1411 snprintf(res, IFNAMSIZ, buf, i); 1412 return i; 1413 } 1414 1415 /* Returns negative errno or allocated unit id (see __dev_alloc_name()) */ 1416 static int dev_prep_valid_name(struct net *net, struct net_device *dev, 1417 const char *want_name, char *out_name, 1418 int dup_errno) 1419 { 1420 if (!dev_valid_name(want_name)) 1421 return -EINVAL; 1422 1423 if (strchr(want_name, '%')) 1424 return __dev_alloc_name(net, want_name, out_name); 1425 1426 if (netdev_name_in_use(net, want_name)) 1427 return -dup_errno; 1428 if (out_name != want_name) 1429 strscpy(out_name, want_name, IFNAMSIZ); 1430 return 0; 1431 } 1432 1433 /** 1434 * dev_alloc_name - allocate a name for a device 1435 * @dev: device 1436 * @name: name format string 1437 * 1438 * Passed a format string - eg "lt%d" it will try and find a suitable 1439 * id. It scans list of devices to build up a free map, then chooses 1440 * the first empty slot. The caller must hold the dev_base or rtnl lock 1441 * while allocating the name and adding the device in order to avoid 1442 * duplicates. 1443 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1444 * Returns the number of the unit assigned or a negative errno code. 1445 */ 1446 1447 int dev_alloc_name(struct net_device *dev, const char *name) 1448 { 1449 return dev_prep_valid_name(dev_net(dev), dev, name, dev->name, ENFILE); 1450 } 1451 EXPORT_SYMBOL(dev_alloc_name); 1452 1453 static int dev_get_valid_name(struct net *net, struct net_device *dev, 1454 const char *name) 1455 { 1456 int ret; 1457 1458 ret = dev_prep_valid_name(net, dev, name, dev->name, EEXIST); 1459 return ret < 0 ? ret : 0; 1460 } 1461 1462 int netif_change_name(struct net_device *dev, const char *newname) 1463 { 1464 struct net *net = dev_net(dev); 1465 unsigned char old_assign_type; 1466 char oldname[IFNAMSIZ]; 1467 int err = 0; 1468 int ret; 1469 1470 ASSERT_RTNL_NET(net); 1471 1472 if (!strncmp(newname, dev->name, IFNAMSIZ)) 1473 return 0; 1474 1475 memcpy(oldname, dev->name, IFNAMSIZ); 1476 1477 write_seqlock_bh(&netdev_rename_lock); 1478 err = dev_get_valid_name(net, dev, newname); 1479 write_sequnlock_bh(&netdev_rename_lock); 1480 1481 if (err < 0) 1482 return err; 1483 1484 if (oldname[0] && !strchr(oldname, '%')) 1485 netdev_info(dev, "renamed from %s%s\n", oldname, 1486 dev->flags & IFF_UP ? " (while UP)" : ""); 1487 1488 old_assign_type = dev->name_assign_type; 1489 WRITE_ONCE(dev->name_assign_type, NET_NAME_RENAMED); 1490 1491 rollback: 1492 ret = device_rename(&dev->dev, dev->name); 1493 if (ret) { 1494 write_seqlock_bh(&netdev_rename_lock); 1495 memcpy(dev->name, oldname, IFNAMSIZ); 1496 write_sequnlock_bh(&netdev_rename_lock); 1497 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1498 return ret; 1499 } 1500 1501 netdev_adjacent_rename_links(dev, oldname); 1502 1503 netdev_name_node_del(dev->name_node); 1504 1505 synchronize_net(); 1506 1507 netdev_name_node_add(net, dev->name_node); 1508 1509 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev); 1510 ret = notifier_to_errno(ret); 1511 1512 if (ret) { 1513 /* err >= 0 after dev_alloc_name() or stores the first errno */ 1514 if (err >= 0) { 1515 err = ret; 1516 write_seqlock_bh(&netdev_rename_lock); 1517 memcpy(dev->name, oldname, IFNAMSIZ); 1518 write_sequnlock_bh(&netdev_rename_lock); 1519 memcpy(oldname, newname, IFNAMSIZ); 1520 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1521 old_assign_type = NET_NAME_RENAMED; 1522 goto rollback; 1523 } else { 1524 netdev_err(dev, "name change rollback failed: %d\n", 1525 ret); 1526 } 1527 } 1528 1529 return err; 1530 } 1531 1532 int netif_set_alias(struct net_device *dev, const char *alias, size_t len) 1533 { 1534 struct dev_ifalias *new_alias = NULL; 1535 1536 if (len >= IFALIASZ) 1537 return -EINVAL; 1538 1539 if (len) { 1540 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL); 1541 if (!new_alias) 1542 return -ENOMEM; 1543 1544 memcpy(new_alias->ifalias, alias, len); 1545 new_alias->ifalias[len] = 0; 1546 } 1547 1548 mutex_lock(&ifalias_mutex); 1549 new_alias = rcu_replace_pointer(dev->ifalias, new_alias, 1550 mutex_is_locked(&ifalias_mutex)); 1551 mutex_unlock(&ifalias_mutex); 1552 1553 if (new_alias) 1554 kfree_rcu(new_alias, rcuhead); 1555 1556 return len; 1557 } 1558 1559 /** 1560 * dev_get_alias - get ifalias of a device 1561 * @dev: device 1562 * @name: buffer to store name of ifalias 1563 * @len: size of buffer 1564 * 1565 * get ifalias for a device. Caller must make sure dev cannot go 1566 * away, e.g. rcu read lock or own a reference count to device. 1567 */ 1568 int dev_get_alias(const struct net_device *dev, char *name, size_t len) 1569 { 1570 const struct dev_ifalias *alias; 1571 int ret = 0; 1572 1573 rcu_read_lock(); 1574 alias = rcu_dereference(dev->ifalias); 1575 if (alias) 1576 ret = snprintf(name, len, "%s", alias->ifalias); 1577 rcu_read_unlock(); 1578 1579 return ret; 1580 } 1581 1582 /** 1583 * netdev_features_change - device changes features 1584 * @dev: device to cause notification 1585 * 1586 * Called to indicate a device has changed features. 1587 */ 1588 void netdev_features_change(struct net_device *dev) 1589 { 1590 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev); 1591 } 1592 EXPORT_SYMBOL(netdev_features_change); 1593 1594 void netif_state_change(struct net_device *dev) 1595 { 1596 netdev_ops_assert_locked_or_invisible(dev); 1597 1598 if (dev->flags & IFF_UP) { 1599 struct netdev_notifier_change_info change_info = { 1600 .info.dev = dev, 1601 }; 1602 1603 call_netdevice_notifiers_info(NETDEV_CHANGE, 1604 &change_info.info); 1605 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL, 0, NULL); 1606 } 1607 } 1608 1609 /** 1610 * __netdev_notify_peers - notify network peers about existence of @dev, 1611 * to be called when rtnl lock is already held. 1612 * @dev: network device 1613 * 1614 * Generate traffic such that interested network peers are aware of 1615 * @dev, such as by generating a gratuitous ARP. This may be used when 1616 * a device wants to inform the rest of the network about some sort of 1617 * reconfiguration such as a failover event or virtual machine 1618 * migration. 1619 */ 1620 void __netdev_notify_peers(struct net_device *dev) 1621 { 1622 ASSERT_RTNL(); 1623 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev); 1624 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev); 1625 } 1626 EXPORT_SYMBOL(__netdev_notify_peers); 1627 1628 /** 1629 * netdev_notify_peers - notify network peers about existence of @dev 1630 * @dev: network device 1631 * 1632 * Generate traffic such that interested network peers are aware of 1633 * @dev, such as by generating a gratuitous ARP. This may be used when 1634 * a device wants to inform the rest of the network about some sort of 1635 * reconfiguration such as a failover event or virtual machine 1636 * migration. 1637 */ 1638 void netdev_notify_peers(struct net_device *dev) 1639 { 1640 rtnl_lock(); 1641 __netdev_notify_peers(dev); 1642 rtnl_unlock(); 1643 } 1644 EXPORT_SYMBOL(netdev_notify_peers); 1645 1646 static int napi_threaded_poll(void *data); 1647 1648 static int napi_kthread_create(struct napi_struct *n) 1649 { 1650 int err = 0; 1651 1652 /* Create and wake up the kthread once to put it in 1653 * TASK_INTERRUPTIBLE mode to avoid the blocked task 1654 * warning and work with loadavg. 1655 */ 1656 n->thread = kthread_run(napi_threaded_poll, n, "napi/%s-%d", 1657 n->dev->name, n->napi_id); 1658 if (IS_ERR(n->thread)) { 1659 err = PTR_ERR(n->thread); 1660 pr_err("kthread_run failed with err %d\n", err); 1661 n->thread = NULL; 1662 } 1663 1664 return err; 1665 } 1666 1667 static int __dev_open(struct net_device *dev, struct netlink_ext_ack *extack) 1668 { 1669 const struct net_device_ops *ops = dev->netdev_ops; 1670 int ret; 1671 1672 ASSERT_RTNL(); 1673 dev_addr_check(dev); 1674 1675 if (!netif_device_present(dev)) { 1676 /* may be detached because parent is runtime-suspended */ 1677 if (dev->dev.parent) 1678 pm_runtime_resume(dev->dev.parent); 1679 if (!netif_device_present(dev)) 1680 return -ENODEV; 1681 } 1682 1683 /* Block netpoll from trying to do any rx path servicing. 1684 * If we don't do this there is a chance ndo_poll_controller 1685 * or ndo_poll may be running while we open the device 1686 */ 1687 netpoll_poll_disable(dev); 1688 1689 ret = call_netdevice_notifiers_extack(NETDEV_PRE_UP, dev, extack); 1690 ret = notifier_to_errno(ret); 1691 if (ret) 1692 return ret; 1693 1694 set_bit(__LINK_STATE_START, &dev->state); 1695 1696 netdev_ops_assert_locked(dev); 1697 1698 if (ops->ndo_validate_addr) 1699 ret = ops->ndo_validate_addr(dev); 1700 1701 if (!ret && ops->ndo_open) 1702 ret = ops->ndo_open(dev); 1703 1704 netpoll_poll_enable(dev); 1705 1706 if (ret) 1707 clear_bit(__LINK_STATE_START, &dev->state); 1708 else { 1709 netif_set_up(dev, true); 1710 dev_set_rx_mode(dev); 1711 dev_activate(dev); 1712 add_device_randomness(dev->dev_addr, dev->addr_len); 1713 } 1714 1715 return ret; 1716 } 1717 1718 int netif_open(struct net_device *dev, struct netlink_ext_ack *extack) 1719 { 1720 int ret; 1721 1722 if (dev->flags & IFF_UP) 1723 return 0; 1724 1725 ret = __dev_open(dev, extack); 1726 if (ret < 0) 1727 return ret; 1728 1729 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1730 call_netdevice_notifiers(NETDEV_UP, dev); 1731 1732 return ret; 1733 } 1734 1735 static void __dev_close_many(struct list_head *head) 1736 { 1737 struct net_device *dev; 1738 1739 ASSERT_RTNL(); 1740 might_sleep(); 1741 1742 list_for_each_entry(dev, head, close_list) { 1743 /* Temporarily disable netpoll until the interface is down */ 1744 netpoll_poll_disable(dev); 1745 1746 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev); 1747 1748 clear_bit(__LINK_STATE_START, &dev->state); 1749 1750 /* Synchronize to scheduled poll. We cannot touch poll list, it 1751 * can be even on different cpu. So just clear netif_running(). 1752 * 1753 * dev->stop() will invoke napi_disable() on all of it's 1754 * napi_struct instances on this device. 1755 */ 1756 smp_mb__after_atomic(); /* Commit netif_running(). */ 1757 } 1758 1759 dev_deactivate_many(head); 1760 1761 list_for_each_entry(dev, head, close_list) { 1762 const struct net_device_ops *ops = dev->netdev_ops; 1763 1764 /* 1765 * Call the device specific close. This cannot fail. 1766 * Only if device is UP 1767 * 1768 * We allow it to be called even after a DETACH hot-plug 1769 * event. 1770 */ 1771 1772 netdev_ops_assert_locked(dev); 1773 1774 if (ops->ndo_stop) 1775 ops->ndo_stop(dev); 1776 1777 netif_set_up(dev, false); 1778 netpoll_poll_enable(dev); 1779 } 1780 } 1781 1782 static void __dev_close(struct net_device *dev) 1783 { 1784 LIST_HEAD(single); 1785 1786 list_add(&dev->close_list, &single); 1787 __dev_close_many(&single); 1788 list_del(&single); 1789 } 1790 1791 void netif_close_many(struct list_head *head, bool unlink) 1792 { 1793 struct net_device *dev, *tmp; 1794 1795 /* Remove the devices that don't need to be closed */ 1796 list_for_each_entry_safe(dev, tmp, head, close_list) 1797 if (!(dev->flags & IFF_UP)) 1798 list_del_init(&dev->close_list); 1799 1800 __dev_close_many(head); 1801 1802 list_for_each_entry_safe(dev, tmp, head, close_list) { 1803 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1804 call_netdevice_notifiers(NETDEV_DOWN, dev); 1805 if (unlink) 1806 list_del_init(&dev->close_list); 1807 } 1808 } 1809 EXPORT_SYMBOL_NS_GPL(netif_close_many, "NETDEV_INTERNAL"); 1810 1811 void netif_close(struct net_device *dev) 1812 { 1813 if (dev->flags & IFF_UP) { 1814 LIST_HEAD(single); 1815 1816 list_add(&dev->close_list, &single); 1817 netif_close_many(&single, true); 1818 list_del(&single); 1819 } 1820 } 1821 EXPORT_SYMBOL(netif_close); 1822 1823 void netif_disable_lro(struct net_device *dev) 1824 { 1825 struct net_device *lower_dev; 1826 struct list_head *iter; 1827 1828 dev->wanted_features &= ~NETIF_F_LRO; 1829 netdev_update_features(dev); 1830 1831 if (unlikely(dev->features & NETIF_F_LRO)) 1832 netdev_WARN(dev, "failed to disable LRO!\n"); 1833 1834 netdev_for_each_lower_dev(dev, lower_dev, iter) { 1835 netdev_lock_ops(lower_dev); 1836 netif_disable_lro(lower_dev); 1837 netdev_unlock_ops(lower_dev); 1838 } 1839 } 1840 EXPORT_IPV6_MOD(netif_disable_lro); 1841 1842 /** 1843 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device 1844 * @dev: device 1845 * 1846 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be 1847 * called under RTNL. This is needed if Generic XDP is installed on 1848 * the device. 1849 */ 1850 static void dev_disable_gro_hw(struct net_device *dev) 1851 { 1852 dev->wanted_features &= ~NETIF_F_GRO_HW; 1853 netdev_update_features(dev); 1854 1855 if (unlikely(dev->features & NETIF_F_GRO_HW)) 1856 netdev_WARN(dev, "failed to disable GRO_HW!\n"); 1857 } 1858 1859 const char *netdev_cmd_to_name(enum netdev_cmd cmd) 1860 { 1861 #define N(val) \ 1862 case NETDEV_##val: \ 1863 return "NETDEV_" __stringify(val); 1864 switch (cmd) { 1865 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER) 1866 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE) 1867 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE) 1868 N(POST_INIT) N(PRE_UNINIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) 1869 N(CHANGEUPPER) N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) 1870 N(BONDING_INFO) N(PRECHANGEUPPER) N(CHANGELOWERSTATE) 1871 N(UDP_TUNNEL_PUSH_INFO) N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN) 1872 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO) 1873 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO) 1874 N(PRE_CHANGEADDR) N(OFFLOAD_XSTATS_ENABLE) N(OFFLOAD_XSTATS_DISABLE) 1875 N(OFFLOAD_XSTATS_REPORT_USED) N(OFFLOAD_XSTATS_REPORT_DELTA) 1876 N(XDP_FEAT_CHANGE) 1877 } 1878 #undef N 1879 return "UNKNOWN_NETDEV_EVENT"; 1880 } 1881 EXPORT_SYMBOL_GPL(netdev_cmd_to_name); 1882 1883 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val, 1884 struct net_device *dev) 1885 { 1886 struct netdev_notifier_info info = { 1887 .dev = dev, 1888 }; 1889 1890 return nb->notifier_call(nb, val, &info); 1891 } 1892 1893 static int call_netdevice_register_notifiers(struct notifier_block *nb, 1894 struct net_device *dev) 1895 { 1896 int err; 1897 1898 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev); 1899 err = notifier_to_errno(err); 1900 if (err) 1901 return err; 1902 1903 if (!(dev->flags & IFF_UP)) 1904 return 0; 1905 1906 call_netdevice_notifier(nb, NETDEV_UP, dev); 1907 return 0; 1908 } 1909 1910 static void call_netdevice_unregister_notifiers(struct notifier_block *nb, 1911 struct net_device *dev) 1912 { 1913 if (dev->flags & IFF_UP) { 1914 call_netdevice_notifier(nb, NETDEV_GOING_DOWN, 1915 dev); 1916 call_netdevice_notifier(nb, NETDEV_DOWN, dev); 1917 } 1918 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev); 1919 } 1920 1921 static int call_netdevice_register_net_notifiers(struct notifier_block *nb, 1922 struct net *net) 1923 { 1924 struct net_device *dev; 1925 int err; 1926 1927 for_each_netdev(net, dev) { 1928 netdev_lock_ops(dev); 1929 err = call_netdevice_register_notifiers(nb, dev); 1930 netdev_unlock_ops(dev); 1931 if (err) 1932 goto rollback; 1933 } 1934 return 0; 1935 1936 rollback: 1937 for_each_netdev_continue_reverse(net, dev) 1938 call_netdevice_unregister_notifiers(nb, dev); 1939 return err; 1940 } 1941 1942 static void call_netdevice_unregister_net_notifiers(struct notifier_block *nb, 1943 struct net *net) 1944 { 1945 struct net_device *dev; 1946 1947 for_each_netdev(net, dev) 1948 call_netdevice_unregister_notifiers(nb, dev); 1949 } 1950 1951 static int dev_boot_phase = 1; 1952 1953 /** 1954 * register_netdevice_notifier - register a network notifier block 1955 * @nb: notifier 1956 * 1957 * Register a notifier to be called when network device events occur. 1958 * The notifier passed is linked into the kernel structures and must 1959 * not be reused until it has been unregistered. A negative errno code 1960 * is returned on a failure. 1961 * 1962 * When registered all registration and up events are replayed 1963 * to the new notifier to allow device to have a race free 1964 * view of the network device list. 1965 */ 1966 1967 int register_netdevice_notifier(struct notifier_block *nb) 1968 { 1969 struct net *net; 1970 int err; 1971 1972 /* Close race with setup_net() and cleanup_net() */ 1973 down_write(&pernet_ops_rwsem); 1974 1975 /* When RTNL is removed, we need protection for netdev_chain. */ 1976 rtnl_lock(); 1977 1978 err = raw_notifier_chain_register(&netdev_chain, nb); 1979 if (err) 1980 goto unlock; 1981 if (dev_boot_phase) 1982 goto unlock; 1983 for_each_net(net) { 1984 __rtnl_net_lock(net); 1985 err = call_netdevice_register_net_notifiers(nb, net); 1986 __rtnl_net_unlock(net); 1987 if (err) 1988 goto rollback; 1989 } 1990 1991 unlock: 1992 rtnl_unlock(); 1993 up_write(&pernet_ops_rwsem); 1994 return err; 1995 1996 rollback: 1997 for_each_net_continue_reverse(net) { 1998 __rtnl_net_lock(net); 1999 call_netdevice_unregister_net_notifiers(nb, net); 2000 __rtnl_net_unlock(net); 2001 } 2002 2003 raw_notifier_chain_unregister(&netdev_chain, nb); 2004 goto unlock; 2005 } 2006 EXPORT_SYMBOL(register_netdevice_notifier); 2007 2008 /** 2009 * unregister_netdevice_notifier - unregister a network notifier block 2010 * @nb: notifier 2011 * 2012 * Unregister a notifier previously registered by 2013 * register_netdevice_notifier(). The notifier is unlinked into the 2014 * kernel structures and may then be reused. A negative errno code 2015 * is returned on a failure. 2016 * 2017 * After unregistering unregister and down device events are synthesized 2018 * for all devices on the device list to the removed notifier to remove 2019 * the need for special case cleanup code. 2020 */ 2021 2022 int unregister_netdevice_notifier(struct notifier_block *nb) 2023 { 2024 struct net *net; 2025 int err; 2026 2027 /* Close race with setup_net() and cleanup_net() */ 2028 down_write(&pernet_ops_rwsem); 2029 rtnl_lock(); 2030 err = raw_notifier_chain_unregister(&netdev_chain, nb); 2031 if (err) 2032 goto unlock; 2033 2034 for_each_net(net) { 2035 __rtnl_net_lock(net); 2036 call_netdevice_unregister_net_notifiers(nb, net); 2037 __rtnl_net_unlock(net); 2038 } 2039 2040 unlock: 2041 rtnl_unlock(); 2042 up_write(&pernet_ops_rwsem); 2043 return err; 2044 } 2045 EXPORT_SYMBOL(unregister_netdevice_notifier); 2046 2047 static int __register_netdevice_notifier_net(struct net *net, 2048 struct notifier_block *nb, 2049 bool ignore_call_fail) 2050 { 2051 int err; 2052 2053 err = raw_notifier_chain_register(&net->netdev_chain, nb); 2054 if (err) 2055 return err; 2056 if (dev_boot_phase) 2057 return 0; 2058 2059 err = call_netdevice_register_net_notifiers(nb, net); 2060 if (err && !ignore_call_fail) 2061 goto chain_unregister; 2062 2063 return 0; 2064 2065 chain_unregister: 2066 raw_notifier_chain_unregister(&net->netdev_chain, nb); 2067 return err; 2068 } 2069 2070 static int __unregister_netdevice_notifier_net(struct net *net, 2071 struct notifier_block *nb) 2072 { 2073 int err; 2074 2075 err = raw_notifier_chain_unregister(&net->netdev_chain, nb); 2076 if (err) 2077 return err; 2078 2079 call_netdevice_unregister_net_notifiers(nb, net); 2080 return 0; 2081 } 2082 2083 /** 2084 * register_netdevice_notifier_net - register a per-netns network notifier block 2085 * @net: network namespace 2086 * @nb: notifier 2087 * 2088 * Register a notifier to be called when network device events occur. 2089 * The notifier passed is linked into the kernel structures and must 2090 * not be reused until it has been unregistered. A negative errno code 2091 * is returned on a failure. 2092 * 2093 * When registered all registration and up events are replayed 2094 * to the new notifier to allow device to have a race free 2095 * view of the network device list. 2096 */ 2097 2098 int register_netdevice_notifier_net(struct net *net, struct notifier_block *nb) 2099 { 2100 int err; 2101 2102 rtnl_net_lock(net); 2103 err = __register_netdevice_notifier_net(net, nb, false); 2104 rtnl_net_unlock(net); 2105 2106 return err; 2107 } 2108 EXPORT_SYMBOL(register_netdevice_notifier_net); 2109 2110 /** 2111 * unregister_netdevice_notifier_net - unregister a per-netns 2112 * network notifier block 2113 * @net: network namespace 2114 * @nb: notifier 2115 * 2116 * Unregister a notifier previously registered by 2117 * register_netdevice_notifier_net(). The notifier is unlinked from the 2118 * kernel structures and may then be reused. A negative errno code 2119 * is returned on a failure. 2120 * 2121 * After unregistering unregister and down device events are synthesized 2122 * for all devices on the device list to the removed notifier to remove 2123 * the need for special case cleanup code. 2124 */ 2125 2126 int unregister_netdevice_notifier_net(struct net *net, 2127 struct notifier_block *nb) 2128 { 2129 int err; 2130 2131 rtnl_net_lock(net); 2132 err = __unregister_netdevice_notifier_net(net, nb); 2133 rtnl_net_unlock(net); 2134 2135 return err; 2136 } 2137 EXPORT_SYMBOL(unregister_netdevice_notifier_net); 2138 2139 static void __move_netdevice_notifier_net(struct net *src_net, 2140 struct net *dst_net, 2141 struct notifier_block *nb) 2142 { 2143 __unregister_netdevice_notifier_net(src_net, nb); 2144 __register_netdevice_notifier_net(dst_net, nb, true); 2145 } 2146 2147 static void rtnl_net_dev_lock(struct net_device *dev) 2148 { 2149 bool again; 2150 2151 do { 2152 struct net *net; 2153 2154 again = false; 2155 2156 /* netns might be being dismantled. */ 2157 rcu_read_lock(); 2158 net = dev_net_rcu(dev); 2159 net_passive_inc(net); 2160 rcu_read_unlock(); 2161 2162 rtnl_net_lock(net); 2163 2164 #ifdef CONFIG_NET_NS 2165 /* dev might have been moved to another netns. */ 2166 if (!net_eq(net, rcu_access_pointer(dev->nd_net.net))) { 2167 rtnl_net_unlock(net); 2168 net_passive_dec(net); 2169 again = true; 2170 } 2171 #endif 2172 } while (again); 2173 } 2174 2175 static void rtnl_net_dev_unlock(struct net_device *dev) 2176 { 2177 struct net *net = dev_net(dev); 2178 2179 rtnl_net_unlock(net); 2180 net_passive_dec(net); 2181 } 2182 2183 int register_netdevice_notifier_dev_net(struct net_device *dev, 2184 struct notifier_block *nb, 2185 struct netdev_net_notifier *nn) 2186 { 2187 int err; 2188 2189 rtnl_net_dev_lock(dev); 2190 err = __register_netdevice_notifier_net(dev_net(dev), nb, false); 2191 if (!err) { 2192 nn->nb = nb; 2193 list_add(&nn->list, &dev->net_notifier_list); 2194 } 2195 rtnl_net_dev_unlock(dev); 2196 2197 return err; 2198 } 2199 EXPORT_SYMBOL(register_netdevice_notifier_dev_net); 2200 2201 int unregister_netdevice_notifier_dev_net(struct net_device *dev, 2202 struct notifier_block *nb, 2203 struct netdev_net_notifier *nn) 2204 { 2205 int err; 2206 2207 rtnl_net_dev_lock(dev); 2208 list_del(&nn->list); 2209 err = __unregister_netdevice_notifier_net(dev_net(dev), nb); 2210 rtnl_net_dev_unlock(dev); 2211 2212 return err; 2213 } 2214 EXPORT_SYMBOL(unregister_netdevice_notifier_dev_net); 2215 2216 static void move_netdevice_notifiers_dev_net(struct net_device *dev, 2217 struct net *net) 2218 { 2219 struct netdev_net_notifier *nn; 2220 2221 list_for_each_entry(nn, &dev->net_notifier_list, list) 2222 __move_netdevice_notifier_net(dev_net(dev), net, nn->nb); 2223 } 2224 2225 /** 2226 * call_netdevice_notifiers_info - call all network notifier blocks 2227 * @val: value passed unmodified to notifier function 2228 * @info: notifier information data 2229 * 2230 * Call all network notifier blocks. Parameters and return value 2231 * are as for raw_notifier_call_chain(). 2232 */ 2233 2234 int call_netdevice_notifiers_info(unsigned long val, 2235 struct netdev_notifier_info *info) 2236 { 2237 struct net *net = dev_net(info->dev); 2238 int ret; 2239 2240 ASSERT_RTNL(); 2241 2242 /* Run per-netns notifier block chain first, then run the global one. 2243 * Hopefully, one day, the global one is going to be removed after 2244 * all notifier block registrators get converted to be per-netns. 2245 */ 2246 ret = raw_notifier_call_chain(&net->netdev_chain, val, info); 2247 if (ret & NOTIFY_STOP_MASK) 2248 return ret; 2249 return raw_notifier_call_chain(&netdev_chain, val, info); 2250 } 2251 2252 /** 2253 * call_netdevice_notifiers_info_robust - call per-netns notifier blocks 2254 * for and rollback on error 2255 * @val_up: value passed unmodified to notifier function 2256 * @val_down: value passed unmodified to the notifier function when 2257 * recovering from an error on @val_up 2258 * @info: notifier information data 2259 * 2260 * Call all per-netns network notifier blocks, but not notifier blocks on 2261 * the global notifier chain. Parameters and return value are as for 2262 * raw_notifier_call_chain_robust(). 2263 */ 2264 2265 static int 2266 call_netdevice_notifiers_info_robust(unsigned long val_up, 2267 unsigned long val_down, 2268 struct netdev_notifier_info *info) 2269 { 2270 struct net *net = dev_net(info->dev); 2271 2272 ASSERT_RTNL(); 2273 2274 return raw_notifier_call_chain_robust(&net->netdev_chain, 2275 val_up, val_down, info); 2276 } 2277 2278 static int call_netdevice_notifiers_extack(unsigned long val, 2279 struct net_device *dev, 2280 struct netlink_ext_ack *extack) 2281 { 2282 struct netdev_notifier_info info = { 2283 .dev = dev, 2284 .extack = extack, 2285 }; 2286 2287 return call_netdevice_notifiers_info(val, &info); 2288 } 2289 2290 /** 2291 * call_netdevice_notifiers - call all network notifier blocks 2292 * @val: value passed unmodified to notifier function 2293 * @dev: net_device pointer passed unmodified to notifier function 2294 * 2295 * Call all network notifier blocks. Parameters and return value 2296 * are as for raw_notifier_call_chain(). 2297 */ 2298 2299 int call_netdevice_notifiers(unsigned long val, struct net_device *dev) 2300 { 2301 return call_netdevice_notifiers_extack(val, dev, NULL); 2302 } 2303 EXPORT_SYMBOL(call_netdevice_notifiers); 2304 2305 /** 2306 * call_netdevice_notifiers_mtu - call all network notifier blocks 2307 * @val: value passed unmodified to notifier function 2308 * @dev: net_device pointer passed unmodified to notifier function 2309 * @arg: additional u32 argument passed to the notifier function 2310 * 2311 * Call all network notifier blocks. Parameters and return value 2312 * are as for raw_notifier_call_chain(). 2313 */ 2314 static int call_netdevice_notifiers_mtu(unsigned long val, 2315 struct net_device *dev, u32 arg) 2316 { 2317 struct netdev_notifier_info_ext info = { 2318 .info.dev = dev, 2319 .ext.mtu = arg, 2320 }; 2321 2322 BUILD_BUG_ON(offsetof(struct netdev_notifier_info_ext, info) != 0); 2323 2324 return call_netdevice_notifiers_info(val, &info.info); 2325 } 2326 2327 #ifdef CONFIG_NET_INGRESS 2328 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key); 2329 2330 void net_inc_ingress_queue(void) 2331 { 2332 static_branch_inc(&ingress_needed_key); 2333 } 2334 EXPORT_SYMBOL_GPL(net_inc_ingress_queue); 2335 2336 void net_dec_ingress_queue(void) 2337 { 2338 static_branch_dec(&ingress_needed_key); 2339 } 2340 EXPORT_SYMBOL_GPL(net_dec_ingress_queue); 2341 #endif 2342 2343 #ifdef CONFIG_NET_EGRESS 2344 static DEFINE_STATIC_KEY_FALSE(egress_needed_key); 2345 2346 void net_inc_egress_queue(void) 2347 { 2348 static_branch_inc(&egress_needed_key); 2349 } 2350 EXPORT_SYMBOL_GPL(net_inc_egress_queue); 2351 2352 void net_dec_egress_queue(void) 2353 { 2354 static_branch_dec(&egress_needed_key); 2355 } 2356 EXPORT_SYMBOL_GPL(net_dec_egress_queue); 2357 #endif 2358 2359 #ifdef CONFIG_NET_CLS_ACT 2360 DEFINE_STATIC_KEY_FALSE(tcf_sw_enabled_key); 2361 EXPORT_SYMBOL(tcf_sw_enabled_key); 2362 #endif 2363 2364 DEFINE_STATIC_KEY_FALSE(netstamp_needed_key); 2365 EXPORT_SYMBOL(netstamp_needed_key); 2366 #ifdef CONFIG_JUMP_LABEL 2367 static atomic_t netstamp_needed_deferred; 2368 static atomic_t netstamp_wanted; 2369 static void netstamp_clear(struct work_struct *work) 2370 { 2371 int deferred = atomic_xchg(&netstamp_needed_deferred, 0); 2372 int wanted; 2373 2374 wanted = atomic_add_return(deferred, &netstamp_wanted); 2375 if (wanted > 0) 2376 static_branch_enable(&netstamp_needed_key); 2377 else 2378 static_branch_disable(&netstamp_needed_key); 2379 } 2380 static DECLARE_WORK(netstamp_work, netstamp_clear); 2381 #endif 2382 2383 void net_enable_timestamp(void) 2384 { 2385 #ifdef CONFIG_JUMP_LABEL 2386 int wanted = atomic_read(&netstamp_wanted); 2387 2388 while (wanted > 0) { 2389 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted + 1)) 2390 return; 2391 } 2392 atomic_inc(&netstamp_needed_deferred); 2393 schedule_work(&netstamp_work); 2394 #else 2395 static_branch_inc(&netstamp_needed_key); 2396 #endif 2397 } 2398 EXPORT_SYMBOL(net_enable_timestamp); 2399 2400 void net_disable_timestamp(void) 2401 { 2402 #ifdef CONFIG_JUMP_LABEL 2403 int wanted = atomic_read(&netstamp_wanted); 2404 2405 while (wanted > 1) { 2406 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted - 1)) 2407 return; 2408 } 2409 atomic_dec(&netstamp_needed_deferred); 2410 schedule_work(&netstamp_work); 2411 #else 2412 static_branch_dec(&netstamp_needed_key); 2413 #endif 2414 } 2415 EXPORT_SYMBOL(net_disable_timestamp); 2416 2417 static inline void net_timestamp_set(struct sk_buff *skb) 2418 { 2419 skb->tstamp = 0; 2420 skb->tstamp_type = SKB_CLOCK_REALTIME; 2421 if (static_branch_unlikely(&netstamp_needed_key)) 2422 skb->tstamp = ktime_get_real(); 2423 } 2424 2425 #define net_timestamp_check(COND, SKB) \ 2426 if (static_branch_unlikely(&netstamp_needed_key)) { \ 2427 if ((COND) && !(SKB)->tstamp) \ 2428 (SKB)->tstamp = ktime_get_real(); \ 2429 } \ 2430 2431 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb) 2432 { 2433 return __is_skb_forwardable(dev, skb, true); 2434 } 2435 EXPORT_SYMBOL_GPL(is_skb_forwardable); 2436 2437 static int __dev_forward_skb2(struct net_device *dev, struct sk_buff *skb, 2438 bool check_mtu) 2439 { 2440 int ret = ____dev_forward_skb(dev, skb, check_mtu); 2441 2442 if (likely(!ret)) { 2443 skb->protocol = eth_type_trans(skb, dev); 2444 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN); 2445 } 2446 2447 return ret; 2448 } 2449 2450 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2451 { 2452 return __dev_forward_skb2(dev, skb, true); 2453 } 2454 EXPORT_SYMBOL_GPL(__dev_forward_skb); 2455 2456 /** 2457 * dev_forward_skb - loopback an skb to another netif 2458 * 2459 * @dev: destination network device 2460 * @skb: buffer to forward 2461 * 2462 * return values: 2463 * NET_RX_SUCCESS (no congestion) 2464 * NET_RX_DROP (packet was dropped, but freed) 2465 * 2466 * dev_forward_skb can be used for injecting an skb from the 2467 * start_xmit function of one device into the receive queue 2468 * of another device. 2469 * 2470 * The receiving device may be in another namespace, so 2471 * we have to clear all information in the skb that could 2472 * impact namespace isolation. 2473 */ 2474 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2475 { 2476 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb); 2477 } 2478 EXPORT_SYMBOL_GPL(dev_forward_skb); 2479 2480 int dev_forward_skb_nomtu(struct net_device *dev, struct sk_buff *skb) 2481 { 2482 return __dev_forward_skb2(dev, skb, false) ?: netif_rx_internal(skb); 2483 } 2484 2485 static int deliver_skb(struct sk_buff *skb, 2486 struct packet_type *pt_prev, 2487 struct net_device *orig_dev) 2488 { 2489 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 2490 return -ENOMEM; 2491 refcount_inc(&skb->users); 2492 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 2493 } 2494 2495 static inline void deliver_ptype_list_skb(struct sk_buff *skb, 2496 struct packet_type **pt, 2497 struct net_device *orig_dev, 2498 __be16 type, 2499 struct list_head *ptype_list) 2500 { 2501 struct packet_type *ptype, *pt_prev = *pt; 2502 2503 list_for_each_entry_rcu(ptype, ptype_list, list) { 2504 if (ptype->type != type) 2505 continue; 2506 if (unlikely(pt_prev)) 2507 deliver_skb(skb, pt_prev, orig_dev); 2508 pt_prev = ptype; 2509 } 2510 *pt = pt_prev; 2511 } 2512 2513 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb) 2514 { 2515 if (!ptype->af_packet_priv || !skb->sk) 2516 return false; 2517 2518 if (ptype->id_match) 2519 return ptype->id_match(ptype, skb->sk); 2520 else if ((struct sock *)ptype->af_packet_priv == skb->sk) 2521 return true; 2522 2523 return false; 2524 } 2525 2526 /** 2527 * dev_nit_active_rcu - return true if any network interface taps are in use 2528 * 2529 * The caller must hold the RCU lock 2530 * 2531 * @dev: network device to check for the presence of taps 2532 */ 2533 bool dev_nit_active_rcu(const struct net_device *dev) 2534 { 2535 /* Callers may hold either RCU or RCU BH lock */ 2536 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 2537 2538 return !list_empty(&dev_net(dev)->ptype_all) || 2539 !list_empty(&dev->ptype_all); 2540 } 2541 EXPORT_SYMBOL_GPL(dev_nit_active_rcu); 2542 2543 /* 2544 * Support routine. Sends outgoing frames to any network 2545 * taps currently in use. 2546 */ 2547 2548 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev) 2549 { 2550 struct packet_type *ptype, *pt_prev = NULL; 2551 struct list_head *ptype_list; 2552 struct sk_buff *skb2 = NULL; 2553 2554 rcu_read_lock(); 2555 ptype_list = &dev_net_rcu(dev)->ptype_all; 2556 again: 2557 list_for_each_entry_rcu(ptype, ptype_list, list) { 2558 if (READ_ONCE(ptype->ignore_outgoing)) 2559 continue; 2560 2561 /* Never send packets back to the socket 2562 * they originated from - MvS (miquels@drinkel.ow.org) 2563 */ 2564 if (skb_loop_sk(ptype, skb)) 2565 continue; 2566 2567 if (unlikely(pt_prev)) { 2568 deliver_skb(skb2, pt_prev, skb->dev); 2569 pt_prev = ptype; 2570 continue; 2571 } 2572 2573 /* need to clone skb, done only once */ 2574 skb2 = skb_clone(skb, GFP_ATOMIC); 2575 if (!skb2) 2576 goto out_unlock; 2577 2578 net_timestamp_set(skb2); 2579 2580 /* skb->nh should be correctly 2581 * set by sender, so that the second statement is 2582 * just protection against buggy protocols. 2583 */ 2584 skb_reset_mac_header(skb2); 2585 2586 if (skb_network_header(skb2) < skb2->data || 2587 skb_network_header(skb2) > skb_tail_pointer(skb2)) { 2588 net_crit_ratelimited("protocol %04x is buggy, dev %s\n", 2589 ntohs(skb2->protocol), 2590 dev->name); 2591 skb_reset_network_header(skb2); 2592 } 2593 2594 skb2->transport_header = skb2->network_header; 2595 skb2->pkt_type = PACKET_OUTGOING; 2596 pt_prev = ptype; 2597 } 2598 2599 if (ptype_list != &dev->ptype_all) { 2600 ptype_list = &dev->ptype_all; 2601 goto again; 2602 } 2603 out_unlock: 2604 if (pt_prev) { 2605 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC)) 2606 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev); 2607 else 2608 kfree_skb(skb2); 2609 } 2610 rcu_read_unlock(); 2611 } 2612 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit); 2613 2614 /** 2615 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change 2616 * @dev: Network device 2617 * @txq: number of queues available 2618 * 2619 * If real_num_tx_queues is changed the tc mappings may no longer be 2620 * valid. To resolve this verify the tc mapping remains valid and if 2621 * not NULL the mapping. With no priorities mapping to this 2622 * offset/count pair it will no longer be used. In the worst case TC0 2623 * is invalid nothing can be done so disable priority mappings. If is 2624 * expected that drivers will fix this mapping if they can before 2625 * calling netif_set_real_num_tx_queues. 2626 */ 2627 static void netif_setup_tc(struct net_device *dev, unsigned int txq) 2628 { 2629 int i; 2630 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2631 2632 /* If TC0 is invalidated disable TC mapping */ 2633 if (tc->offset + tc->count > txq) { 2634 netdev_warn(dev, "Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n"); 2635 dev->num_tc = 0; 2636 return; 2637 } 2638 2639 /* Invalidated prio to tc mappings set to TC0 */ 2640 for (i = 1; i < TC_BITMASK + 1; i++) { 2641 int q = netdev_get_prio_tc_map(dev, i); 2642 2643 tc = &dev->tc_to_txq[q]; 2644 if (tc->offset + tc->count > txq) { 2645 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", 2646 i, q); 2647 netdev_set_prio_tc_map(dev, i, 0); 2648 } 2649 } 2650 } 2651 2652 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq) 2653 { 2654 if (dev->num_tc) { 2655 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2656 int i; 2657 2658 /* walk through the TCs and see if it falls into any of them */ 2659 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) { 2660 if ((txq - tc->offset) < tc->count) 2661 return i; 2662 } 2663 2664 /* didn't find it, just return -1 to indicate no match */ 2665 return -1; 2666 } 2667 2668 return 0; 2669 } 2670 EXPORT_SYMBOL(netdev_txq_to_tc); 2671 2672 #ifdef CONFIG_XPS 2673 static struct static_key xps_needed __read_mostly; 2674 static struct static_key xps_rxqs_needed __read_mostly; 2675 static DEFINE_MUTEX(xps_map_mutex); 2676 #define xmap_dereference(P) \ 2677 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex)) 2678 2679 static bool remove_xps_queue(struct xps_dev_maps *dev_maps, 2680 struct xps_dev_maps *old_maps, int tci, u16 index) 2681 { 2682 struct xps_map *map = NULL; 2683 int pos; 2684 2685 map = xmap_dereference(dev_maps->attr_map[tci]); 2686 if (!map) 2687 return false; 2688 2689 for (pos = map->len; pos--;) { 2690 if (map->queues[pos] != index) 2691 continue; 2692 2693 if (map->len > 1) { 2694 map->queues[pos] = map->queues[--map->len]; 2695 break; 2696 } 2697 2698 if (old_maps) 2699 RCU_INIT_POINTER(old_maps->attr_map[tci], NULL); 2700 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2701 kfree_rcu(map, rcu); 2702 return false; 2703 } 2704 2705 return true; 2706 } 2707 2708 static bool remove_xps_queue_cpu(struct net_device *dev, 2709 struct xps_dev_maps *dev_maps, 2710 int cpu, u16 offset, u16 count) 2711 { 2712 int num_tc = dev_maps->num_tc; 2713 bool active = false; 2714 int tci; 2715 2716 for (tci = cpu * num_tc; num_tc--; tci++) { 2717 int i, j; 2718 2719 for (i = count, j = offset; i--; j++) { 2720 if (!remove_xps_queue(dev_maps, NULL, tci, j)) 2721 break; 2722 } 2723 2724 active |= i < 0; 2725 } 2726 2727 return active; 2728 } 2729 2730 static void reset_xps_maps(struct net_device *dev, 2731 struct xps_dev_maps *dev_maps, 2732 enum xps_map_type type) 2733 { 2734 static_key_slow_dec_cpuslocked(&xps_needed); 2735 if (type == XPS_RXQS) 2736 static_key_slow_dec_cpuslocked(&xps_rxqs_needed); 2737 2738 RCU_INIT_POINTER(dev->xps_maps[type], NULL); 2739 2740 kfree_rcu(dev_maps, rcu); 2741 } 2742 2743 static void clean_xps_maps(struct net_device *dev, enum xps_map_type type, 2744 u16 offset, u16 count) 2745 { 2746 struct xps_dev_maps *dev_maps; 2747 bool active = false; 2748 int i, j; 2749 2750 dev_maps = xmap_dereference(dev->xps_maps[type]); 2751 if (!dev_maps) 2752 return; 2753 2754 for (j = 0; j < dev_maps->nr_ids; j++) 2755 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, count); 2756 if (!active) 2757 reset_xps_maps(dev, dev_maps, type); 2758 2759 if (type == XPS_CPUS) { 2760 for (i = offset + (count - 1); count--; i--) 2761 netdev_queue_numa_node_write( 2762 netdev_get_tx_queue(dev, i), NUMA_NO_NODE); 2763 } 2764 } 2765 2766 static void netif_reset_xps_queues(struct net_device *dev, u16 offset, 2767 u16 count) 2768 { 2769 if (!static_key_false(&xps_needed)) 2770 return; 2771 2772 cpus_read_lock(); 2773 mutex_lock(&xps_map_mutex); 2774 2775 if (static_key_false(&xps_rxqs_needed)) 2776 clean_xps_maps(dev, XPS_RXQS, offset, count); 2777 2778 clean_xps_maps(dev, XPS_CPUS, offset, count); 2779 2780 mutex_unlock(&xps_map_mutex); 2781 cpus_read_unlock(); 2782 } 2783 2784 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index) 2785 { 2786 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index); 2787 } 2788 2789 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index, 2790 u16 index, bool is_rxqs_map) 2791 { 2792 struct xps_map *new_map; 2793 int alloc_len = XPS_MIN_MAP_ALLOC; 2794 int i, pos; 2795 2796 for (pos = 0; map && pos < map->len; pos++) { 2797 if (map->queues[pos] != index) 2798 continue; 2799 return map; 2800 } 2801 2802 /* Need to add tx-queue to this CPU's/rx-queue's existing map */ 2803 if (map) { 2804 if (pos < map->alloc_len) 2805 return map; 2806 2807 alloc_len = map->alloc_len * 2; 2808 } 2809 2810 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's 2811 * map 2812 */ 2813 if (is_rxqs_map) 2814 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL); 2815 else 2816 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL, 2817 cpu_to_node(attr_index)); 2818 if (!new_map) 2819 return NULL; 2820 2821 for (i = 0; i < pos; i++) 2822 new_map->queues[i] = map->queues[i]; 2823 new_map->alloc_len = alloc_len; 2824 new_map->len = pos; 2825 2826 return new_map; 2827 } 2828 2829 /* Copy xps maps at a given index */ 2830 static void xps_copy_dev_maps(struct xps_dev_maps *dev_maps, 2831 struct xps_dev_maps *new_dev_maps, int index, 2832 int tc, bool skip_tc) 2833 { 2834 int i, tci = index * dev_maps->num_tc; 2835 struct xps_map *map; 2836 2837 /* copy maps belonging to foreign traffic classes */ 2838 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2839 if (i == tc && skip_tc) 2840 continue; 2841 2842 /* fill in the new device map from the old device map */ 2843 map = xmap_dereference(dev_maps->attr_map[tci]); 2844 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2845 } 2846 } 2847 2848 /* Must be called under cpus_read_lock */ 2849 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask, 2850 u16 index, enum xps_map_type type) 2851 { 2852 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL, *old_dev_maps = NULL; 2853 const unsigned long *online_mask = NULL; 2854 bool active = false, copy = false; 2855 int i, j, tci, numa_node_id = -2; 2856 int maps_sz, num_tc = 1, tc = 0; 2857 struct xps_map *map, *new_map; 2858 unsigned int nr_ids; 2859 2860 WARN_ON_ONCE(index >= dev->num_tx_queues); 2861 2862 if (dev->num_tc) { 2863 /* Do not allow XPS on subordinate device directly */ 2864 num_tc = dev->num_tc; 2865 if (num_tc < 0) 2866 return -EINVAL; 2867 2868 /* If queue belongs to subordinate dev use its map */ 2869 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev; 2870 2871 tc = netdev_txq_to_tc(dev, index); 2872 if (tc < 0) 2873 return -EINVAL; 2874 } 2875 2876 mutex_lock(&xps_map_mutex); 2877 2878 dev_maps = xmap_dereference(dev->xps_maps[type]); 2879 if (type == XPS_RXQS) { 2880 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues); 2881 nr_ids = dev->num_rx_queues; 2882 } else { 2883 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc); 2884 if (num_possible_cpus() > 1) 2885 online_mask = cpumask_bits(cpu_online_mask); 2886 nr_ids = nr_cpu_ids; 2887 } 2888 2889 if (maps_sz < L1_CACHE_BYTES) 2890 maps_sz = L1_CACHE_BYTES; 2891 2892 /* The old dev_maps could be larger or smaller than the one we're 2893 * setting up now, as dev->num_tc or nr_ids could have been updated in 2894 * between. We could try to be smart, but let's be safe instead and only 2895 * copy foreign traffic classes if the two map sizes match. 2896 */ 2897 if (dev_maps && 2898 dev_maps->num_tc == num_tc && dev_maps->nr_ids == nr_ids) 2899 copy = true; 2900 2901 /* allocate memory for queue storage */ 2902 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids), 2903 j < nr_ids;) { 2904 if (!new_dev_maps) { 2905 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL); 2906 if (!new_dev_maps) { 2907 mutex_unlock(&xps_map_mutex); 2908 return -ENOMEM; 2909 } 2910 2911 new_dev_maps->nr_ids = nr_ids; 2912 new_dev_maps->num_tc = num_tc; 2913 } 2914 2915 tci = j * num_tc + tc; 2916 map = copy ? xmap_dereference(dev_maps->attr_map[tci]) : NULL; 2917 2918 map = expand_xps_map(map, j, index, type == XPS_RXQS); 2919 if (!map) 2920 goto error; 2921 2922 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2923 } 2924 2925 if (!new_dev_maps) 2926 goto out_no_new_maps; 2927 2928 if (!dev_maps) { 2929 /* Increment static keys at most once per type */ 2930 static_key_slow_inc_cpuslocked(&xps_needed); 2931 if (type == XPS_RXQS) 2932 static_key_slow_inc_cpuslocked(&xps_rxqs_needed); 2933 } 2934 2935 for (j = 0; j < nr_ids; j++) { 2936 bool skip_tc = false; 2937 2938 tci = j * num_tc + tc; 2939 if (netif_attr_test_mask(j, mask, nr_ids) && 2940 netif_attr_test_online(j, online_mask, nr_ids)) { 2941 /* add tx-queue to CPU/rx-queue maps */ 2942 int pos = 0; 2943 2944 skip_tc = true; 2945 2946 map = xmap_dereference(new_dev_maps->attr_map[tci]); 2947 while ((pos < map->len) && (map->queues[pos] != index)) 2948 pos++; 2949 2950 if (pos == map->len) 2951 map->queues[map->len++] = index; 2952 #ifdef CONFIG_NUMA 2953 if (type == XPS_CPUS) { 2954 if (numa_node_id == -2) 2955 numa_node_id = cpu_to_node(j); 2956 else if (numa_node_id != cpu_to_node(j)) 2957 numa_node_id = -1; 2958 } 2959 #endif 2960 } 2961 2962 if (copy) 2963 xps_copy_dev_maps(dev_maps, new_dev_maps, j, tc, 2964 skip_tc); 2965 } 2966 2967 rcu_assign_pointer(dev->xps_maps[type], new_dev_maps); 2968 2969 /* Cleanup old maps */ 2970 if (!dev_maps) 2971 goto out_no_old_maps; 2972 2973 for (j = 0; j < dev_maps->nr_ids; j++) { 2974 for (i = num_tc, tci = j * dev_maps->num_tc; i--; tci++) { 2975 map = xmap_dereference(dev_maps->attr_map[tci]); 2976 if (!map) 2977 continue; 2978 2979 if (copy) { 2980 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2981 if (map == new_map) 2982 continue; 2983 } 2984 2985 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2986 kfree_rcu(map, rcu); 2987 } 2988 } 2989 2990 old_dev_maps = dev_maps; 2991 2992 out_no_old_maps: 2993 dev_maps = new_dev_maps; 2994 active = true; 2995 2996 out_no_new_maps: 2997 if (type == XPS_CPUS) 2998 /* update Tx queue numa node */ 2999 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index), 3000 (numa_node_id >= 0) ? 3001 numa_node_id : NUMA_NO_NODE); 3002 3003 if (!dev_maps) 3004 goto out_no_maps; 3005 3006 /* removes tx-queue from unused CPUs/rx-queues */ 3007 for (j = 0; j < dev_maps->nr_ids; j++) { 3008 tci = j * dev_maps->num_tc; 3009 3010 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 3011 if (i == tc && 3012 netif_attr_test_mask(j, mask, dev_maps->nr_ids) && 3013 netif_attr_test_online(j, online_mask, dev_maps->nr_ids)) 3014 continue; 3015 3016 active |= remove_xps_queue(dev_maps, 3017 copy ? old_dev_maps : NULL, 3018 tci, index); 3019 } 3020 } 3021 3022 if (old_dev_maps) 3023 kfree_rcu(old_dev_maps, rcu); 3024 3025 /* free map if not active */ 3026 if (!active) 3027 reset_xps_maps(dev, dev_maps, type); 3028 3029 out_no_maps: 3030 mutex_unlock(&xps_map_mutex); 3031 3032 return 0; 3033 error: 3034 /* remove any maps that we added */ 3035 for (j = 0; j < nr_ids; j++) { 3036 for (i = num_tc, tci = j * num_tc; i--; tci++) { 3037 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 3038 map = copy ? 3039 xmap_dereference(dev_maps->attr_map[tci]) : 3040 NULL; 3041 if (new_map && new_map != map) 3042 kfree(new_map); 3043 } 3044 } 3045 3046 mutex_unlock(&xps_map_mutex); 3047 3048 kfree(new_dev_maps); 3049 return -ENOMEM; 3050 } 3051 EXPORT_SYMBOL_GPL(__netif_set_xps_queue); 3052 3053 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, 3054 u16 index) 3055 { 3056 int ret; 3057 3058 cpus_read_lock(); 3059 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, XPS_CPUS); 3060 cpus_read_unlock(); 3061 3062 return ret; 3063 } 3064 EXPORT_SYMBOL(netif_set_xps_queue); 3065 3066 #endif 3067 static void netdev_unbind_all_sb_channels(struct net_device *dev) 3068 { 3069 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 3070 3071 /* Unbind any subordinate channels */ 3072 while (txq-- != &dev->_tx[0]) { 3073 if (txq->sb_dev) 3074 netdev_unbind_sb_channel(dev, txq->sb_dev); 3075 } 3076 } 3077 3078 void netdev_reset_tc(struct net_device *dev) 3079 { 3080 #ifdef CONFIG_XPS 3081 netif_reset_xps_queues_gt(dev, 0); 3082 #endif 3083 netdev_unbind_all_sb_channels(dev); 3084 3085 /* Reset TC configuration of device */ 3086 dev->num_tc = 0; 3087 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq)); 3088 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map)); 3089 } 3090 EXPORT_SYMBOL(netdev_reset_tc); 3091 3092 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset) 3093 { 3094 if (tc >= dev->num_tc) 3095 return -EINVAL; 3096 3097 #ifdef CONFIG_XPS 3098 netif_reset_xps_queues(dev, offset, count); 3099 #endif 3100 dev->tc_to_txq[tc].count = count; 3101 dev->tc_to_txq[tc].offset = offset; 3102 return 0; 3103 } 3104 EXPORT_SYMBOL(netdev_set_tc_queue); 3105 3106 int netdev_set_num_tc(struct net_device *dev, u8 num_tc) 3107 { 3108 if (num_tc > TC_MAX_QUEUE) 3109 return -EINVAL; 3110 3111 #ifdef CONFIG_XPS 3112 netif_reset_xps_queues_gt(dev, 0); 3113 #endif 3114 netdev_unbind_all_sb_channels(dev); 3115 3116 dev->num_tc = num_tc; 3117 return 0; 3118 } 3119 EXPORT_SYMBOL(netdev_set_num_tc); 3120 3121 void netdev_unbind_sb_channel(struct net_device *dev, 3122 struct net_device *sb_dev) 3123 { 3124 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 3125 3126 #ifdef CONFIG_XPS 3127 netif_reset_xps_queues_gt(sb_dev, 0); 3128 #endif 3129 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq)); 3130 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map)); 3131 3132 while (txq-- != &dev->_tx[0]) { 3133 if (txq->sb_dev == sb_dev) 3134 txq->sb_dev = NULL; 3135 } 3136 } 3137 EXPORT_SYMBOL(netdev_unbind_sb_channel); 3138 3139 int netdev_bind_sb_channel_queue(struct net_device *dev, 3140 struct net_device *sb_dev, 3141 u8 tc, u16 count, u16 offset) 3142 { 3143 /* Make certain the sb_dev and dev are already configured */ 3144 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc) 3145 return -EINVAL; 3146 3147 /* We cannot hand out queues we don't have */ 3148 if ((offset + count) > dev->real_num_tx_queues) 3149 return -EINVAL; 3150 3151 /* Record the mapping */ 3152 sb_dev->tc_to_txq[tc].count = count; 3153 sb_dev->tc_to_txq[tc].offset = offset; 3154 3155 /* Provide a way for Tx queue to find the tc_to_txq map or 3156 * XPS map for itself. 3157 */ 3158 while (count--) 3159 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev; 3160 3161 return 0; 3162 } 3163 EXPORT_SYMBOL(netdev_bind_sb_channel_queue); 3164 3165 int netdev_set_sb_channel(struct net_device *dev, u16 channel) 3166 { 3167 /* Do not use a multiqueue device to represent a subordinate channel */ 3168 if (netif_is_multiqueue(dev)) 3169 return -ENODEV; 3170 3171 /* We allow channels 1 - 32767 to be used for subordinate channels. 3172 * Channel 0 is meant to be "native" mode and used only to represent 3173 * the main root device. We allow writing 0 to reset the device back 3174 * to normal mode after being used as a subordinate channel. 3175 */ 3176 if (channel > S16_MAX) 3177 return -EINVAL; 3178 3179 dev->num_tc = -channel; 3180 3181 return 0; 3182 } 3183 EXPORT_SYMBOL(netdev_set_sb_channel); 3184 3185 /* 3186 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues 3187 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed. 3188 */ 3189 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq) 3190 { 3191 bool disabling; 3192 int rc; 3193 3194 disabling = txq < dev->real_num_tx_queues; 3195 3196 if (txq < 1 || txq > dev->num_tx_queues) 3197 return -EINVAL; 3198 3199 if (dev->reg_state == NETREG_REGISTERED || 3200 dev->reg_state == NETREG_UNREGISTERING) { 3201 netdev_ops_assert_locked(dev); 3202 3203 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues, 3204 txq); 3205 if (rc) 3206 return rc; 3207 3208 if (dev->num_tc) 3209 netif_setup_tc(dev, txq); 3210 3211 net_shaper_set_real_num_tx_queues(dev, txq); 3212 3213 dev_qdisc_change_real_num_tx(dev, txq); 3214 3215 dev->real_num_tx_queues = txq; 3216 3217 if (disabling) { 3218 synchronize_net(); 3219 qdisc_reset_all_tx_gt(dev, txq); 3220 #ifdef CONFIG_XPS 3221 netif_reset_xps_queues_gt(dev, txq); 3222 #endif 3223 } 3224 } else { 3225 dev->real_num_tx_queues = txq; 3226 } 3227 3228 return 0; 3229 } 3230 EXPORT_SYMBOL(netif_set_real_num_tx_queues); 3231 3232 /** 3233 * netif_set_real_num_rx_queues - set actual number of RX queues used 3234 * @dev: Network device 3235 * @rxq: Actual number of RX queues 3236 * 3237 * This must be called either with the rtnl_lock held or before 3238 * registration of the net device. Returns 0 on success, or a 3239 * negative error code. If called before registration, it always 3240 * succeeds. 3241 */ 3242 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq) 3243 { 3244 int rc; 3245 3246 if (rxq < 1 || rxq > dev->num_rx_queues) 3247 return -EINVAL; 3248 3249 if (dev->reg_state == NETREG_REGISTERED) { 3250 netdev_ops_assert_locked(dev); 3251 3252 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues, 3253 rxq); 3254 if (rc) 3255 return rc; 3256 } 3257 3258 dev->real_num_rx_queues = rxq; 3259 return 0; 3260 } 3261 EXPORT_SYMBOL(netif_set_real_num_rx_queues); 3262 3263 /** 3264 * netif_set_real_num_queues - set actual number of RX and TX queues used 3265 * @dev: Network device 3266 * @txq: Actual number of TX queues 3267 * @rxq: Actual number of RX queues 3268 * 3269 * Set the real number of both TX and RX queues. 3270 * Does nothing if the number of queues is already correct. 3271 */ 3272 int netif_set_real_num_queues(struct net_device *dev, 3273 unsigned int txq, unsigned int rxq) 3274 { 3275 unsigned int old_rxq = dev->real_num_rx_queues; 3276 int err; 3277 3278 if (txq < 1 || txq > dev->num_tx_queues || 3279 rxq < 1 || rxq > dev->num_rx_queues) 3280 return -EINVAL; 3281 3282 /* Start from increases, so the error path only does decreases - 3283 * decreases can't fail. 3284 */ 3285 if (rxq > dev->real_num_rx_queues) { 3286 err = netif_set_real_num_rx_queues(dev, rxq); 3287 if (err) 3288 return err; 3289 } 3290 if (txq > dev->real_num_tx_queues) { 3291 err = netif_set_real_num_tx_queues(dev, txq); 3292 if (err) 3293 goto undo_rx; 3294 } 3295 if (rxq < dev->real_num_rx_queues) 3296 WARN_ON(netif_set_real_num_rx_queues(dev, rxq)); 3297 if (txq < dev->real_num_tx_queues) 3298 WARN_ON(netif_set_real_num_tx_queues(dev, txq)); 3299 3300 return 0; 3301 undo_rx: 3302 WARN_ON(netif_set_real_num_rx_queues(dev, old_rxq)); 3303 return err; 3304 } 3305 EXPORT_SYMBOL(netif_set_real_num_queues); 3306 3307 /** 3308 * netif_set_tso_max_size() - set the max size of TSO frames supported 3309 * @dev: netdev to update 3310 * @size: max skb->len of a TSO frame 3311 * 3312 * Set the limit on the size of TSO super-frames the device can handle. 3313 * Unless explicitly set the stack will assume the value of 3314 * %GSO_LEGACY_MAX_SIZE. 3315 */ 3316 void netif_set_tso_max_size(struct net_device *dev, unsigned int size) 3317 { 3318 dev->tso_max_size = min(GSO_MAX_SIZE, size); 3319 if (size < READ_ONCE(dev->gso_max_size)) 3320 netif_set_gso_max_size(dev, size); 3321 if (size < READ_ONCE(dev->gso_ipv4_max_size)) 3322 netif_set_gso_ipv4_max_size(dev, size); 3323 } 3324 EXPORT_SYMBOL(netif_set_tso_max_size); 3325 3326 /** 3327 * netif_set_tso_max_segs() - set the max number of segs supported for TSO 3328 * @dev: netdev to update 3329 * @segs: max number of TCP segments 3330 * 3331 * Set the limit on the number of TCP segments the device can generate from 3332 * a single TSO super-frame. 3333 * Unless explicitly set the stack will assume the value of %GSO_MAX_SEGS. 3334 */ 3335 void netif_set_tso_max_segs(struct net_device *dev, unsigned int segs) 3336 { 3337 dev->tso_max_segs = segs; 3338 if (segs < READ_ONCE(dev->gso_max_segs)) 3339 netif_set_gso_max_segs(dev, segs); 3340 } 3341 EXPORT_SYMBOL(netif_set_tso_max_segs); 3342 3343 /** 3344 * netif_inherit_tso_max() - copy all TSO limits from a lower device to an upper 3345 * @to: netdev to update 3346 * @from: netdev from which to copy the limits 3347 */ 3348 void netif_inherit_tso_max(struct net_device *to, const struct net_device *from) 3349 { 3350 netif_set_tso_max_size(to, from->tso_max_size); 3351 netif_set_tso_max_segs(to, from->tso_max_segs); 3352 } 3353 EXPORT_SYMBOL(netif_inherit_tso_max); 3354 3355 /** 3356 * netif_get_num_default_rss_queues - default number of RSS queues 3357 * 3358 * Default value is the number of physical cores if there are only 1 or 2, or 3359 * divided by 2 if there are more. 3360 */ 3361 int netif_get_num_default_rss_queues(void) 3362 { 3363 cpumask_var_t cpus; 3364 int cpu, count = 0; 3365 3366 if (unlikely(is_kdump_kernel() || !zalloc_cpumask_var(&cpus, GFP_KERNEL))) 3367 return 1; 3368 3369 cpumask_copy(cpus, cpu_online_mask); 3370 for_each_cpu(cpu, cpus) { 3371 ++count; 3372 cpumask_andnot(cpus, cpus, topology_sibling_cpumask(cpu)); 3373 } 3374 free_cpumask_var(cpus); 3375 3376 return count > 2 ? DIV_ROUND_UP(count, 2) : count; 3377 } 3378 EXPORT_SYMBOL(netif_get_num_default_rss_queues); 3379 3380 static void __netif_reschedule(struct Qdisc *q) 3381 { 3382 struct softnet_data *sd; 3383 unsigned long flags; 3384 3385 local_irq_save(flags); 3386 sd = this_cpu_ptr(&softnet_data); 3387 q->next_sched = NULL; 3388 *sd->output_queue_tailp = q; 3389 sd->output_queue_tailp = &q->next_sched; 3390 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3391 local_irq_restore(flags); 3392 } 3393 3394 void __netif_schedule(struct Qdisc *q) 3395 { 3396 /* If q->defer_list is not empty, at least one thread is 3397 * in __dev_xmit_skb() before llist_del_all(&q->defer_list). 3398 * This thread will attempt to run the queue. 3399 */ 3400 if (!llist_empty(&q->defer_list)) 3401 return; 3402 3403 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state)) 3404 __netif_reschedule(q); 3405 } 3406 EXPORT_SYMBOL(__netif_schedule); 3407 3408 struct dev_kfree_skb_cb { 3409 enum skb_drop_reason reason; 3410 }; 3411 3412 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb) 3413 { 3414 return (struct dev_kfree_skb_cb *)skb->cb; 3415 } 3416 3417 void netif_schedule_queue(struct netdev_queue *txq) 3418 { 3419 rcu_read_lock(); 3420 if (!netif_xmit_stopped(txq)) { 3421 struct Qdisc *q = rcu_dereference(txq->qdisc); 3422 3423 __netif_schedule(q); 3424 } 3425 rcu_read_unlock(); 3426 } 3427 EXPORT_SYMBOL(netif_schedule_queue); 3428 3429 void netif_tx_wake_queue(struct netdev_queue *dev_queue) 3430 { 3431 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) { 3432 struct Qdisc *q; 3433 3434 rcu_read_lock(); 3435 q = rcu_dereference(dev_queue->qdisc); 3436 __netif_schedule(q); 3437 rcu_read_unlock(); 3438 } 3439 } 3440 EXPORT_SYMBOL(netif_tx_wake_queue); 3441 3442 void dev_kfree_skb_irq_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3443 { 3444 unsigned long flags; 3445 3446 if (unlikely(!skb)) 3447 return; 3448 3449 if (likely(refcount_read(&skb->users) == 1)) { 3450 smp_rmb(); 3451 refcount_set(&skb->users, 0); 3452 } else if (likely(!refcount_dec_and_test(&skb->users))) { 3453 return; 3454 } 3455 get_kfree_skb_cb(skb)->reason = reason; 3456 local_irq_save(flags); 3457 skb->next = __this_cpu_read(softnet_data.completion_queue); 3458 __this_cpu_write(softnet_data.completion_queue, skb); 3459 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3460 local_irq_restore(flags); 3461 } 3462 EXPORT_SYMBOL(dev_kfree_skb_irq_reason); 3463 3464 void dev_kfree_skb_any_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3465 { 3466 if (in_hardirq() || irqs_disabled()) 3467 dev_kfree_skb_irq_reason(skb, reason); 3468 else 3469 kfree_skb_reason(skb, reason); 3470 } 3471 EXPORT_SYMBOL(dev_kfree_skb_any_reason); 3472 3473 3474 /** 3475 * netif_device_detach - mark device as removed 3476 * @dev: network device 3477 * 3478 * Mark device as removed from system and therefore no longer available. 3479 */ 3480 void netif_device_detach(struct net_device *dev) 3481 { 3482 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) && 3483 netif_running(dev)) { 3484 netif_tx_stop_all_queues(dev); 3485 } 3486 } 3487 EXPORT_SYMBOL(netif_device_detach); 3488 3489 /** 3490 * netif_device_attach - mark device as attached 3491 * @dev: network device 3492 * 3493 * Mark device as attached from system and restart if needed. 3494 */ 3495 void netif_device_attach(struct net_device *dev) 3496 { 3497 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) && 3498 netif_running(dev)) { 3499 netif_tx_wake_all_queues(dev); 3500 netdev_watchdog_up(dev); 3501 } 3502 } 3503 EXPORT_SYMBOL(netif_device_attach); 3504 3505 /* 3506 * Returns a Tx hash based on the given packet descriptor a Tx queues' number 3507 * to be used as a distribution range. 3508 */ 3509 static u16 skb_tx_hash(const struct net_device *dev, 3510 const struct net_device *sb_dev, 3511 struct sk_buff *skb) 3512 { 3513 u32 hash; 3514 u16 qoffset = 0; 3515 u16 qcount = dev->real_num_tx_queues; 3516 3517 if (dev->num_tc) { 3518 u8 tc = netdev_get_prio_tc_map(dev, skb->priority); 3519 3520 qoffset = sb_dev->tc_to_txq[tc].offset; 3521 qcount = sb_dev->tc_to_txq[tc].count; 3522 if (unlikely(!qcount)) { 3523 net_warn_ratelimited("%s: invalid qcount, qoffset %u for tc %u\n", 3524 sb_dev->name, qoffset, tc); 3525 qoffset = 0; 3526 qcount = dev->real_num_tx_queues; 3527 } 3528 } 3529 3530 if (skb_rx_queue_recorded(skb)) { 3531 DEBUG_NET_WARN_ON_ONCE(qcount == 0); 3532 hash = skb_get_rx_queue(skb); 3533 if (hash >= qoffset) 3534 hash -= qoffset; 3535 while (unlikely(hash >= qcount)) 3536 hash -= qcount; 3537 return hash + qoffset; 3538 } 3539 3540 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset; 3541 } 3542 3543 void skb_warn_bad_offload(const struct sk_buff *skb) 3544 { 3545 static const netdev_features_t null_features; 3546 struct net_device *dev = skb->dev; 3547 const char *name = ""; 3548 3549 if (!net_ratelimit()) 3550 return; 3551 3552 if (dev) { 3553 if (dev->dev.parent) 3554 name = dev_driver_string(dev->dev.parent); 3555 else 3556 name = netdev_name(dev); 3557 } 3558 skb_dump(KERN_WARNING, skb, false); 3559 WARN(1, "%s: caps=(%pNF, %pNF)\n", 3560 name, dev ? &dev->features : &null_features, 3561 skb->sk ? &skb->sk->sk_route_caps : &null_features); 3562 } 3563 3564 /* 3565 * Invalidate hardware checksum when packet is to be mangled, and 3566 * complete checksum manually on outgoing path. 3567 */ 3568 int skb_checksum_help(struct sk_buff *skb) 3569 { 3570 __wsum csum; 3571 int ret = 0, offset; 3572 3573 if (skb->ip_summed == CHECKSUM_COMPLETE) 3574 goto out_set_summed; 3575 3576 if (unlikely(skb_is_gso(skb))) { 3577 skb_warn_bad_offload(skb); 3578 return -EINVAL; 3579 } 3580 3581 if (!skb_frags_readable(skb)) { 3582 return -EFAULT; 3583 } 3584 3585 /* Before computing a checksum, we should make sure no frag could 3586 * be modified by an external entity : checksum could be wrong. 3587 */ 3588 if (skb_has_shared_frag(skb)) { 3589 ret = __skb_linearize(skb); 3590 if (ret) 3591 goto out; 3592 } 3593 3594 offset = skb_checksum_start_offset(skb); 3595 ret = -EINVAL; 3596 if (unlikely(offset >= skb_headlen(skb))) { 3597 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3598 WARN_ONCE(true, "offset (%d) >= skb_headlen() (%u)\n", 3599 offset, skb_headlen(skb)); 3600 goto out; 3601 } 3602 csum = skb_checksum(skb, offset, skb->len - offset, 0); 3603 3604 offset += skb->csum_offset; 3605 if (unlikely(offset + sizeof(__sum16) > skb_headlen(skb))) { 3606 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3607 WARN_ONCE(true, "offset+2 (%zu) > skb_headlen() (%u)\n", 3608 offset + sizeof(__sum16), skb_headlen(skb)); 3609 goto out; 3610 } 3611 ret = skb_ensure_writable(skb, offset + sizeof(__sum16)); 3612 if (ret) 3613 goto out; 3614 3615 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0; 3616 out_set_summed: 3617 skb->ip_summed = CHECKSUM_NONE; 3618 out: 3619 return ret; 3620 } 3621 EXPORT_SYMBOL(skb_checksum_help); 3622 3623 #ifdef CONFIG_NET_CRC32C 3624 int skb_crc32c_csum_help(struct sk_buff *skb) 3625 { 3626 u32 crc; 3627 int ret = 0, offset, start; 3628 3629 if (skb->ip_summed != CHECKSUM_PARTIAL) 3630 goto out; 3631 3632 if (unlikely(skb_is_gso(skb))) 3633 goto out; 3634 3635 /* Before computing a checksum, we should make sure no frag could 3636 * be modified by an external entity : checksum could be wrong. 3637 */ 3638 if (unlikely(skb_has_shared_frag(skb))) { 3639 ret = __skb_linearize(skb); 3640 if (ret) 3641 goto out; 3642 } 3643 start = skb_checksum_start_offset(skb); 3644 offset = start + offsetof(struct sctphdr, checksum); 3645 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) { 3646 ret = -EINVAL; 3647 goto out; 3648 } 3649 3650 ret = skb_ensure_writable(skb, offset + sizeof(__le32)); 3651 if (ret) 3652 goto out; 3653 3654 crc = ~skb_crc32c(skb, start, skb->len - start, ~0); 3655 *(__le32 *)(skb->data + offset) = cpu_to_le32(crc); 3656 skb_reset_csum_not_inet(skb); 3657 out: 3658 return ret; 3659 } 3660 EXPORT_SYMBOL(skb_crc32c_csum_help); 3661 #endif /* CONFIG_NET_CRC32C */ 3662 3663 __be16 skb_network_protocol(struct sk_buff *skb, int *depth) 3664 { 3665 __be16 type = skb->protocol; 3666 3667 /* Tunnel gso handlers can set protocol to ethernet. */ 3668 if (type == htons(ETH_P_TEB)) { 3669 struct ethhdr *eth; 3670 3671 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr)))) 3672 return 0; 3673 3674 eth = (struct ethhdr *)skb->data; 3675 type = eth->h_proto; 3676 } 3677 3678 return vlan_get_protocol_and_depth(skb, type, depth); 3679 } 3680 3681 3682 /* Take action when hardware reception checksum errors are detected. */ 3683 #ifdef CONFIG_BUG 3684 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3685 { 3686 netdev_err(dev, "hw csum failure\n"); 3687 skb_dump(KERN_ERR, skb, true); 3688 dump_stack(); 3689 } 3690 3691 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3692 { 3693 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb); 3694 } 3695 EXPORT_SYMBOL(netdev_rx_csum_fault); 3696 #endif 3697 3698 /* XXX: check that highmem exists at all on the given machine. */ 3699 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb) 3700 { 3701 #ifdef CONFIG_HIGHMEM 3702 int i; 3703 3704 if (!(dev->features & NETIF_F_HIGHDMA)) { 3705 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 3706 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 3707 struct page *page = skb_frag_page(frag); 3708 3709 if (page && PageHighMem(page)) 3710 return 1; 3711 } 3712 } 3713 #endif 3714 return 0; 3715 } 3716 3717 /* If MPLS offload request, verify we are testing hardware MPLS features 3718 * instead of standard features for the netdev. 3719 */ 3720 #if IS_ENABLED(CONFIG_NET_MPLS_GSO) 3721 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3722 netdev_features_t features, 3723 __be16 type) 3724 { 3725 if (eth_p_mpls(type)) 3726 features &= skb->dev->mpls_features; 3727 3728 return features; 3729 } 3730 #else 3731 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3732 netdev_features_t features, 3733 __be16 type) 3734 { 3735 return features; 3736 } 3737 #endif 3738 3739 static netdev_features_t harmonize_features(struct sk_buff *skb, 3740 netdev_features_t features) 3741 { 3742 __be16 type; 3743 3744 type = skb_network_protocol(skb, NULL); 3745 features = net_mpls_features(skb, features, type); 3746 3747 if (skb->ip_summed != CHECKSUM_NONE && 3748 !can_checksum_protocol(features, type)) { 3749 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3750 } 3751 if (illegal_highdma(skb->dev, skb)) 3752 features &= ~NETIF_F_SG; 3753 3754 return features; 3755 } 3756 3757 netdev_features_t passthru_features_check(struct sk_buff *skb, 3758 struct net_device *dev, 3759 netdev_features_t features) 3760 { 3761 return features; 3762 } 3763 EXPORT_SYMBOL(passthru_features_check); 3764 3765 static netdev_features_t dflt_features_check(struct sk_buff *skb, 3766 struct net_device *dev, 3767 netdev_features_t features) 3768 { 3769 return vlan_features_check(skb, features); 3770 } 3771 3772 static netdev_features_t gso_features_check(const struct sk_buff *skb, 3773 struct net_device *dev, 3774 netdev_features_t features) 3775 { 3776 u16 gso_segs = skb_shinfo(skb)->gso_segs; 3777 3778 if (gso_segs > READ_ONCE(dev->gso_max_segs)) 3779 return features & ~NETIF_F_GSO_MASK; 3780 3781 if (unlikely(skb->len >= netif_get_gso_max_size(dev, skb))) 3782 return features & ~NETIF_F_GSO_MASK; 3783 3784 if (!skb_shinfo(skb)->gso_type) { 3785 skb_warn_bad_offload(skb); 3786 return features & ~NETIF_F_GSO_MASK; 3787 } 3788 3789 /* Support for GSO partial features requires software 3790 * intervention before we can actually process the packets 3791 * so we need to strip support for any partial features now 3792 * and we can pull them back in after we have partially 3793 * segmented the frame. 3794 */ 3795 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)) 3796 features &= ~dev->gso_partial_features; 3797 3798 /* Make sure to clear the IPv4 ID mangling feature if the IPv4 header 3799 * has the potential to be fragmented so that TSO does not generate 3800 * segments with the same ID. For encapsulated packets, the ID mangling 3801 * feature is guaranteed not to use the same ID for the outer IPv4 3802 * headers of the generated segments if the headers have the potential 3803 * to be fragmented, so there is no need to clear the IPv4 ID mangling 3804 * feature (see the section about NETIF_F_TSO_MANGLEID in 3805 * segmentation-offloads.rst). 3806 */ 3807 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) { 3808 struct iphdr *iph = skb->encapsulation ? 3809 inner_ip_hdr(skb) : ip_hdr(skb); 3810 3811 if (!(iph->frag_off & htons(IP_DF))) 3812 features &= ~dev->mangleid_features; 3813 } 3814 3815 /* NETIF_F_IPV6_CSUM does not support IPv6 extension headers, 3816 * so neither does TSO that depends on it. 3817 */ 3818 if (features & NETIF_F_IPV6_CSUM && 3819 (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6 || 3820 (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4 && 3821 vlan_get_protocol(skb) == htons(ETH_P_IPV6))) && 3822 skb_transport_header_was_set(skb) && 3823 skb_network_header_len(skb) != sizeof(struct ipv6hdr)) 3824 features &= ~(NETIF_F_IPV6_CSUM | NETIF_F_TSO6 | NETIF_F_GSO_UDP_L4); 3825 3826 return features; 3827 } 3828 3829 netdev_features_t netif_skb_features(struct sk_buff *skb) 3830 { 3831 struct net_device *dev = skb->dev; 3832 netdev_features_t features = dev->features; 3833 3834 if (skb_is_gso(skb)) 3835 features = gso_features_check(skb, dev, features); 3836 3837 /* If encapsulation offload request, verify we are testing 3838 * hardware encapsulation features instead of standard 3839 * features for the netdev 3840 */ 3841 if (skb->encapsulation) 3842 features &= dev->hw_enc_features; 3843 3844 if (skb_vlan_tagged(skb)) 3845 features = netdev_intersect_features(features, 3846 dev->vlan_features | 3847 NETIF_F_HW_VLAN_CTAG_TX | 3848 NETIF_F_HW_VLAN_STAG_TX); 3849 3850 if (dev->netdev_ops->ndo_features_check) 3851 features &= dev->netdev_ops->ndo_features_check(skb, dev, 3852 features); 3853 else 3854 features &= dflt_features_check(skb, dev, features); 3855 3856 return harmonize_features(skb, features); 3857 } 3858 EXPORT_SYMBOL(netif_skb_features); 3859 3860 static int xmit_one(struct sk_buff *skb, struct net_device *dev, 3861 struct netdev_queue *txq, bool more) 3862 { 3863 unsigned int len; 3864 int rc; 3865 3866 if (dev_nit_active_rcu(dev)) 3867 dev_queue_xmit_nit(skb, dev); 3868 3869 len = skb->len; 3870 trace_net_dev_start_xmit(skb, dev); 3871 rc = netdev_start_xmit(skb, dev, txq, more); 3872 trace_net_dev_xmit(skb, rc, dev, len); 3873 3874 return rc; 3875 } 3876 3877 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev, 3878 struct netdev_queue *txq, int *ret) 3879 { 3880 struct sk_buff *skb = first; 3881 int rc = NETDEV_TX_OK; 3882 3883 while (skb) { 3884 struct sk_buff *next = skb->next; 3885 3886 skb_mark_not_on_list(skb); 3887 rc = xmit_one(skb, dev, txq, next != NULL); 3888 if (unlikely(!dev_xmit_complete(rc))) { 3889 skb->next = next; 3890 goto out; 3891 } 3892 3893 skb = next; 3894 if (netif_tx_queue_stopped(txq) && skb) { 3895 rc = NETDEV_TX_BUSY; 3896 break; 3897 } 3898 } 3899 3900 out: 3901 *ret = rc; 3902 return skb; 3903 } 3904 3905 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb, 3906 netdev_features_t features) 3907 { 3908 if (skb_vlan_tag_present(skb) && 3909 !vlan_hw_offload_capable(features, skb->vlan_proto)) 3910 skb = __vlan_hwaccel_push_inside(skb); 3911 return skb; 3912 } 3913 3914 int skb_csum_hwoffload_help(struct sk_buff *skb, 3915 const netdev_features_t features) 3916 { 3917 if (unlikely(skb_csum_is_sctp(skb))) 3918 return !!(features & NETIF_F_SCTP_CRC) ? 0 : 3919 skb_crc32c_csum_help(skb); 3920 3921 if (features & NETIF_F_HW_CSUM) 3922 return 0; 3923 3924 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) { 3925 if (vlan_get_protocol(skb) == htons(ETH_P_IPV6) && 3926 skb_network_header_len(skb) != sizeof(struct ipv6hdr)) 3927 goto sw_checksum; 3928 3929 switch (skb->csum_offset) { 3930 case offsetof(struct tcphdr, check): 3931 case offsetof(struct udphdr, check): 3932 return 0; 3933 } 3934 } 3935 3936 sw_checksum: 3937 return skb_checksum_help(skb); 3938 } 3939 EXPORT_SYMBOL(skb_csum_hwoffload_help); 3940 3941 /* Checks if this SKB belongs to an HW offloaded socket 3942 * and whether any SW fallbacks are required based on dev. 3943 * Check decrypted mark in case skb_orphan() cleared socket. 3944 */ 3945 static struct sk_buff *sk_validate_xmit_skb(struct sk_buff *skb, 3946 struct net_device *dev) 3947 { 3948 #ifdef CONFIG_SOCK_VALIDATE_XMIT 3949 struct sk_buff *(*sk_validate)(struct sock *sk, struct net_device *dev, 3950 struct sk_buff *skb); 3951 struct sock *sk = skb->sk; 3952 3953 sk_validate = NULL; 3954 if (sk) { 3955 if (sk_fullsock(sk)) 3956 sk_validate = sk->sk_validate_xmit_skb; 3957 else if (sk_is_inet(sk) && sk->sk_state == TCP_TIME_WAIT) 3958 sk_validate = inet_twsk(sk)->tw_validate_xmit_skb; 3959 } 3960 3961 if (sk_validate) { 3962 skb = sk_validate(sk, dev, skb); 3963 } else if (unlikely(skb_is_decrypted(skb))) { 3964 pr_warn_ratelimited("unencrypted skb with no associated socket - dropping\n"); 3965 kfree_skb(skb); 3966 skb = NULL; 3967 } 3968 #endif 3969 3970 return skb; 3971 } 3972 3973 static struct sk_buff *validate_xmit_unreadable_skb(struct sk_buff *skb, 3974 struct net_device *dev) 3975 { 3976 struct skb_shared_info *shinfo; 3977 struct net_iov *niov; 3978 3979 if (likely(skb_frags_readable(skb))) 3980 goto out; 3981 3982 if (!dev->netmem_tx) 3983 goto out_free; 3984 3985 shinfo = skb_shinfo(skb); 3986 3987 if (shinfo->nr_frags > 0) { 3988 niov = netmem_to_net_iov(skb_frag_netmem(&shinfo->frags[0])); 3989 if (net_is_devmem_iov(niov) && 3990 net_devmem_iov_binding(niov)->dev != dev) 3991 goto out_free; 3992 } 3993 3994 out: 3995 return skb; 3996 3997 out_free: 3998 kfree_skb(skb); 3999 return NULL; 4000 } 4001 4002 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again) 4003 { 4004 netdev_features_t features; 4005 4006 skb = validate_xmit_unreadable_skb(skb, dev); 4007 if (unlikely(!skb)) 4008 goto out_null; 4009 4010 features = netif_skb_features(skb); 4011 skb = validate_xmit_vlan(skb, features); 4012 if (unlikely(!skb)) 4013 goto out_null; 4014 4015 skb = sk_validate_xmit_skb(skb, dev); 4016 if (unlikely(!skb)) 4017 goto out_null; 4018 4019 if (netif_needs_gso(skb, features)) { 4020 struct sk_buff *segs; 4021 4022 segs = skb_gso_segment(skb, features); 4023 if (IS_ERR(segs)) { 4024 goto out_kfree_skb; 4025 } else if (segs) { 4026 consume_skb(skb); 4027 skb = segs; 4028 } 4029 } else { 4030 if (skb_needs_linearize(skb, features) && 4031 __skb_linearize(skb)) 4032 goto out_kfree_skb; 4033 4034 /* If packet is not checksummed and device does not 4035 * support checksumming for this protocol, complete 4036 * checksumming here. 4037 */ 4038 if (skb->ip_summed == CHECKSUM_PARTIAL) { 4039 if (skb->encapsulation) 4040 skb_set_inner_transport_header(skb, 4041 skb_checksum_start_offset(skb)); 4042 else 4043 skb_set_transport_header(skb, 4044 skb_checksum_start_offset(skb)); 4045 if (skb_csum_hwoffload_help(skb, features)) 4046 goto out_kfree_skb; 4047 } 4048 } 4049 4050 skb = validate_xmit_xfrm(skb, features, again); 4051 4052 return skb; 4053 4054 out_kfree_skb: 4055 kfree_skb(skb); 4056 out_null: 4057 dev_core_stats_tx_dropped_inc(dev); 4058 return NULL; 4059 } 4060 4061 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again) 4062 { 4063 struct sk_buff *next, *head = NULL, *tail; 4064 4065 for (; skb != NULL; skb = next) { 4066 next = skb->next; 4067 skb_mark_not_on_list(skb); 4068 4069 /* in case skb won't be segmented, point to itself */ 4070 skb->prev = skb; 4071 4072 skb = validate_xmit_skb(skb, dev, again); 4073 if (!skb) 4074 continue; 4075 4076 if (!head) 4077 head = skb; 4078 else 4079 tail->next = skb; 4080 /* If skb was segmented, skb->prev points to 4081 * the last segment. If not, it still contains skb. 4082 */ 4083 tail = skb->prev; 4084 } 4085 return head; 4086 } 4087 EXPORT_SYMBOL_GPL(validate_xmit_skb_list); 4088 4089 static void qdisc_pkt_len_segs_init(struct sk_buff *skb) 4090 { 4091 struct skb_shared_info *shinfo = skb_shinfo(skb); 4092 u16 gso_segs; 4093 4094 qdisc_skb_cb(skb)->pkt_len = skb->len; 4095 if (!shinfo->gso_size) { 4096 qdisc_skb_cb(skb)->pkt_segs = 1; 4097 return; 4098 } 4099 4100 qdisc_skb_cb(skb)->pkt_segs = gso_segs = shinfo->gso_segs; 4101 4102 /* To get more precise estimation of bytes sent on wire, 4103 * we add to pkt_len the headers size of all segments 4104 */ 4105 if (skb_transport_header_was_set(skb)) { 4106 unsigned int hdr_len; 4107 4108 /* mac layer + network layer */ 4109 if (!skb->encapsulation) 4110 hdr_len = skb_transport_offset(skb); 4111 else 4112 hdr_len = skb_inner_transport_offset(skb); 4113 4114 /* + transport layer */ 4115 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) { 4116 const struct tcphdr *th; 4117 struct tcphdr _tcphdr; 4118 4119 th = skb_header_pointer(skb, hdr_len, 4120 sizeof(_tcphdr), &_tcphdr); 4121 if (likely(th)) 4122 hdr_len += __tcp_hdrlen(th); 4123 } else if (shinfo->gso_type & SKB_GSO_UDP_L4) { 4124 struct udphdr _udphdr; 4125 4126 if (skb_header_pointer(skb, hdr_len, 4127 sizeof(_udphdr), &_udphdr)) 4128 hdr_len += sizeof(struct udphdr); 4129 } 4130 4131 if (unlikely(shinfo->gso_type & SKB_GSO_DODGY)) { 4132 int payload = skb->len - hdr_len; 4133 4134 /* Malicious packet. */ 4135 if (payload <= 0) 4136 return; 4137 gso_segs = DIV_ROUND_UP(payload, shinfo->gso_size); 4138 shinfo->gso_segs = gso_segs; 4139 qdisc_skb_cb(skb)->pkt_segs = gso_segs; 4140 } 4141 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len; 4142 } 4143 } 4144 4145 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q, 4146 struct sk_buff **to_free, 4147 struct netdev_queue *txq) 4148 { 4149 int rc; 4150 4151 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK; 4152 if (rc == NET_XMIT_SUCCESS) 4153 trace_qdisc_enqueue(q, txq, skb); 4154 return rc; 4155 } 4156 4157 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q, 4158 struct net_device *dev, 4159 struct netdev_queue *txq) 4160 { 4161 struct sk_buff *next, *to_free = NULL, *to_free2 = NULL; 4162 spinlock_t *root_lock = qdisc_lock(q); 4163 struct llist_node *ll_list, *first_n; 4164 unsigned long defer_count = 0; 4165 int rc; 4166 4167 qdisc_calculate_pkt_len(skb, q); 4168 4169 tcf_set_drop_reason(skb, SKB_DROP_REASON_QDISC_DROP); 4170 4171 if (q->flags & TCQ_F_NOLOCK) { 4172 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) && 4173 qdisc_run_begin(q)) { 4174 /* Retest nolock_qdisc_is_empty() within the protection 4175 * of q->seqlock to protect from racing with requeuing. 4176 */ 4177 if (unlikely(!nolock_qdisc_is_empty(q))) { 4178 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4179 __qdisc_run(q); 4180 to_free2 = qdisc_run_end(q); 4181 4182 goto free_skbs; 4183 } 4184 4185 qdisc_bstats_cpu_update(q, skb); 4186 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) && 4187 !nolock_qdisc_is_empty(q)) 4188 __qdisc_run(q); 4189 4190 to_free2 = qdisc_run_end(q); 4191 rc = NET_XMIT_SUCCESS; 4192 goto free_skbs; 4193 } 4194 4195 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4196 to_free2 = qdisc_run(q); 4197 goto free_skbs; 4198 } 4199 4200 /* Open code llist_add(&skb->ll_node, &q->defer_list) + queue limit. 4201 * In the try_cmpxchg() loop, we want to increment q->defer_count 4202 * at most once to limit the number of skbs in defer_list. 4203 * We perform the defer_count increment only if the list is not empty, 4204 * because some arches have slow atomic_long_inc_return(). 4205 */ 4206 first_n = READ_ONCE(q->defer_list.first); 4207 do { 4208 if (first_n && !defer_count) { 4209 defer_count = atomic_long_inc_return(&q->defer_count); 4210 if (unlikely(defer_count > READ_ONCE(net_hotdata.qdisc_max_burst))) { 4211 kfree_skb_reason(skb, SKB_DROP_REASON_QDISC_BURST_DROP); 4212 return NET_XMIT_DROP; 4213 } 4214 } 4215 skb->ll_node.next = first_n; 4216 } while (!try_cmpxchg(&q->defer_list.first, &first_n, &skb->ll_node)); 4217 4218 /* If defer_list was not empty, we know the cpu which queued 4219 * the first skb will process the whole list for us. 4220 */ 4221 if (first_n) 4222 return NET_XMIT_SUCCESS; 4223 4224 spin_lock(root_lock); 4225 4226 ll_list = llist_del_all(&q->defer_list); 4227 /* There is a small race because we clear defer_count not atomically 4228 * with the prior llist_del_all(). This means defer_list could grow 4229 * over qdisc_max_burst. 4230 */ 4231 atomic_long_set(&q->defer_count, 0); 4232 4233 ll_list = llist_reverse_order(ll_list); 4234 4235 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) { 4236 llist_for_each_entry_safe(skb, next, ll_list, ll_node) 4237 __qdisc_drop(skb, &to_free); 4238 rc = NET_XMIT_DROP; 4239 goto unlock; 4240 } 4241 if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && 4242 !llist_next(ll_list) && qdisc_run_begin(q)) { 4243 /* 4244 * This is a work-conserving queue; there are no old skbs 4245 * waiting to be sent out; and the qdisc is not running - 4246 * xmit the skb directly. 4247 */ 4248 4249 DEBUG_NET_WARN_ON_ONCE(skb != llist_entry(ll_list, 4250 struct sk_buff, 4251 ll_node)); 4252 qdisc_bstats_update(q, skb); 4253 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) 4254 __qdisc_run(q); 4255 to_free2 = qdisc_run_end(q); 4256 rc = NET_XMIT_SUCCESS; 4257 } else { 4258 int count = 0; 4259 4260 llist_for_each_entry_safe(skb, next, ll_list, ll_node) { 4261 if (next) { 4262 prefetch(next); 4263 prefetch(&next->priority); 4264 skb_mark_not_on_list(skb); 4265 } 4266 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4267 count++; 4268 } 4269 to_free2 = qdisc_run(q); 4270 if (count != 1) 4271 rc = NET_XMIT_SUCCESS; 4272 } 4273 unlock: 4274 spin_unlock(root_lock); 4275 4276 free_skbs: 4277 tcf_kfree_skb_list(to_free); 4278 tcf_kfree_skb_list(to_free2); 4279 return rc; 4280 } 4281 4282 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO) 4283 static void skb_update_prio(struct sk_buff *skb) 4284 { 4285 const struct netprio_map *map; 4286 const struct sock *sk; 4287 unsigned int prioidx; 4288 4289 if (skb->priority) 4290 return; 4291 map = rcu_dereference_bh(skb->dev->priomap); 4292 if (!map) 4293 return; 4294 sk = skb_to_full_sk(skb); 4295 if (!sk) 4296 return; 4297 4298 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data); 4299 4300 if (prioidx < map->priomap_len) 4301 skb->priority = map->priomap[prioidx]; 4302 } 4303 #else 4304 #define skb_update_prio(skb) 4305 #endif 4306 4307 /** 4308 * dev_loopback_xmit - loop back @skb 4309 * @net: network namespace this loopback is happening in 4310 * @sk: sk needed to be a netfilter okfn 4311 * @skb: buffer to transmit 4312 */ 4313 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb) 4314 { 4315 skb_reset_mac_header(skb); 4316 __skb_pull(skb, skb_network_offset(skb)); 4317 skb->pkt_type = PACKET_LOOPBACK; 4318 if (skb->ip_summed == CHECKSUM_NONE) 4319 skb->ip_summed = CHECKSUM_UNNECESSARY; 4320 DEBUG_NET_WARN_ON_ONCE(!skb_dst(skb)); 4321 skb_dst_force(skb); 4322 netif_rx(skb); 4323 return 0; 4324 } 4325 EXPORT_SYMBOL(dev_loopback_xmit); 4326 4327 #ifdef CONFIG_NET_EGRESS 4328 static struct netdev_queue * 4329 netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb) 4330 { 4331 int qm = skb_get_queue_mapping(skb); 4332 4333 return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm)); 4334 } 4335 4336 #ifndef CONFIG_PREEMPT_RT 4337 static bool netdev_xmit_txqueue_skipped(void) 4338 { 4339 return __this_cpu_read(softnet_data.xmit.skip_txqueue); 4340 } 4341 4342 void netdev_xmit_skip_txqueue(bool skip) 4343 { 4344 __this_cpu_write(softnet_data.xmit.skip_txqueue, skip); 4345 } 4346 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4347 4348 #else 4349 static bool netdev_xmit_txqueue_skipped(void) 4350 { 4351 return current->net_xmit.skip_txqueue; 4352 } 4353 4354 void netdev_xmit_skip_txqueue(bool skip) 4355 { 4356 current->net_xmit.skip_txqueue = skip; 4357 } 4358 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4359 #endif 4360 #endif /* CONFIG_NET_EGRESS */ 4361 4362 #ifdef CONFIG_NET_XGRESS 4363 static int tc_run(struct tcx_entry *entry, struct sk_buff *skb, 4364 enum skb_drop_reason *drop_reason) 4365 { 4366 int ret = TC_ACT_UNSPEC; 4367 #ifdef CONFIG_NET_CLS_ACT 4368 struct mini_Qdisc *miniq = rcu_dereference_bh(entry->miniq); 4369 struct tcf_result res; 4370 4371 if (!miniq) 4372 return ret; 4373 4374 /* Global bypass */ 4375 if (!static_branch_likely(&tcf_sw_enabled_key)) 4376 return ret; 4377 4378 /* Block-wise bypass */ 4379 if (tcf_block_bypass_sw(miniq->block)) 4380 return ret; 4381 4382 tc_skb_cb(skb)->mru = 0; 4383 qdisc_skb_cb(skb)->post_ct = false; 4384 tcf_set_drop_reason(skb, *drop_reason); 4385 4386 mini_qdisc_bstats_cpu_update(miniq, skb); 4387 ret = tcf_classify(skb, miniq->block, miniq->filter_list, &res, false); 4388 /* Only tcf related quirks below. */ 4389 switch (ret) { 4390 case TC_ACT_SHOT: 4391 *drop_reason = tcf_get_drop_reason(skb); 4392 mini_qdisc_qstats_cpu_drop(miniq); 4393 break; 4394 case TC_ACT_OK: 4395 case TC_ACT_RECLASSIFY: 4396 skb->tc_index = TC_H_MIN(res.classid); 4397 break; 4398 } 4399 #endif /* CONFIG_NET_CLS_ACT */ 4400 return ret; 4401 } 4402 4403 static DEFINE_STATIC_KEY_FALSE(tcx_needed_key); 4404 4405 void tcx_inc(void) 4406 { 4407 static_branch_inc(&tcx_needed_key); 4408 } 4409 4410 void tcx_dec(void) 4411 { 4412 static_branch_dec(&tcx_needed_key); 4413 } 4414 4415 static __always_inline enum tcx_action_base 4416 tcx_run(const struct bpf_mprog_entry *entry, struct sk_buff *skb, 4417 const bool needs_mac) 4418 { 4419 const struct bpf_mprog_fp *fp; 4420 const struct bpf_prog *prog; 4421 int ret = TCX_NEXT; 4422 4423 if (needs_mac) 4424 __skb_push(skb, skb->mac_len); 4425 bpf_mprog_foreach_prog(entry, fp, prog) { 4426 bpf_compute_data_pointers(skb); 4427 ret = bpf_prog_run(prog, skb); 4428 if (ret != TCX_NEXT) 4429 break; 4430 } 4431 if (needs_mac) 4432 __skb_pull(skb, skb->mac_len); 4433 return tcx_action_code(skb, ret); 4434 } 4435 4436 static __always_inline struct sk_buff * 4437 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4438 struct net_device *orig_dev, bool *another) 4439 { 4440 struct bpf_mprog_entry *entry = rcu_dereference_bh(skb->dev->tcx_ingress); 4441 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_INGRESS; 4442 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4443 int sch_ret; 4444 4445 if (!entry) 4446 return skb; 4447 4448 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4449 if (unlikely(*pt_prev)) { 4450 *ret = deliver_skb(skb, *pt_prev, orig_dev); 4451 *pt_prev = NULL; 4452 } 4453 4454 qdisc_pkt_len_segs_init(skb); 4455 tcx_set_ingress(skb, true); 4456 4457 if (static_branch_unlikely(&tcx_needed_key)) { 4458 sch_ret = tcx_run(entry, skb, true); 4459 if (sch_ret != TC_ACT_UNSPEC) 4460 goto ingress_verdict; 4461 } 4462 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4463 ingress_verdict: 4464 switch (sch_ret) { 4465 case TC_ACT_REDIRECT: 4466 /* skb_mac_header check was done by BPF, so we can safely 4467 * push the L2 header back before redirecting to another 4468 * netdev. 4469 */ 4470 __skb_push(skb, skb->mac_len); 4471 if (skb_do_redirect(skb) == -EAGAIN) { 4472 __skb_pull(skb, skb->mac_len); 4473 *another = true; 4474 break; 4475 } 4476 *ret = NET_RX_SUCCESS; 4477 bpf_net_ctx_clear(bpf_net_ctx); 4478 return NULL; 4479 case TC_ACT_SHOT: 4480 kfree_skb_reason(skb, drop_reason); 4481 *ret = NET_RX_DROP; 4482 bpf_net_ctx_clear(bpf_net_ctx); 4483 return NULL; 4484 /* used by tc_run */ 4485 case TC_ACT_STOLEN: 4486 case TC_ACT_QUEUED: 4487 case TC_ACT_TRAP: 4488 consume_skb(skb); 4489 fallthrough; 4490 case TC_ACT_CONSUMED: 4491 *ret = NET_RX_SUCCESS; 4492 bpf_net_ctx_clear(bpf_net_ctx); 4493 return NULL; 4494 } 4495 bpf_net_ctx_clear(bpf_net_ctx); 4496 4497 return skb; 4498 } 4499 4500 static __always_inline struct sk_buff * 4501 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4502 { 4503 struct bpf_mprog_entry *entry = rcu_dereference_bh(dev->tcx_egress); 4504 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_EGRESS; 4505 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4506 int sch_ret; 4507 4508 if (!entry) 4509 return skb; 4510 4511 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4512 4513 /* qdisc_skb_cb(skb)->pkt_len & tcx_set_ingress() was 4514 * already set by the caller. 4515 */ 4516 if (static_branch_unlikely(&tcx_needed_key)) { 4517 sch_ret = tcx_run(entry, skb, false); 4518 if (sch_ret != TC_ACT_UNSPEC) 4519 goto egress_verdict; 4520 } 4521 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4522 egress_verdict: 4523 switch (sch_ret) { 4524 case TC_ACT_REDIRECT: 4525 /* No need to push/pop skb's mac_header here on egress! */ 4526 skb_do_redirect(skb); 4527 *ret = NET_XMIT_SUCCESS; 4528 bpf_net_ctx_clear(bpf_net_ctx); 4529 return NULL; 4530 case TC_ACT_SHOT: 4531 kfree_skb_reason(skb, drop_reason); 4532 *ret = NET_XMIT_DROP; 4533 bpf_net_ctx_clear(bpf_net_ctx); 4534 return NULL; 4535 /* used by tc_run */ 4536 case TC_ACT_STOLEN: 4537 case TC_ACT_QUEUED: 4538 case TC_ACT_TRAP: 4539 consume_skb(skb); 4540 fallthrough; 4541 case TC_ACT_CONSUMED: 4542 *ret = NET_XMIT_SUCCESS; 4543 bpf_net_ctx_clear(bpf_net_ctx); 4544 return NULL; 4545 } 4546 bpf_net_ctx_clear(bpf_net_ctx); 4547 4548 return skb; 4549 } 4550 #else 4551 static __always_inline struct sk_buff * 4552 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4553 struct net_device *orig_dev, bool *another) 4554 { 4555 return skb; 4556 } 4557 4558 static __always_inline struct sk_buff * 4559 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4560 { 4561 return skb; 4562 } 4563 #endif /* CONFIG_NET_XGRESS */ 4564 4565 #ifdef CONFIG_XPS 4566 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb, 4567 struct xps_dev_maps *dev_maps, unsigned int tci) 4568 { 4569 int tc = netdev_get_prio_tc_map(dev, skb->priority); 4570 struct xps_map *map; 4571 int queue_index = -1; 4572 4573 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids) 4574 return queue_index; 4575 4576 tci *= dev_maps->num_tc; 4577 tci += tc; 4578 4579 map = rcu_dereference(dev_maps->attr_map[tci]); 4580 if (map) { 4581 if (map->len == 1) 4582 queue_index = map->queues[0]; 4583 else 4584 queue_index = map->queues[reciprocal_scale( 4585 skb_get_hash(skb), map->len)]; 4586 if (unlikely(queue_index >= dev->real_num_tx_queues)) 4587 queue_index = -1; 4588 } 4589 return queue_index; 4590 } 4591 #endif 4592 4593 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev, 4594 struct sk_buff *skb) 4595 { 4596 #ifdef CONFIG_XPS 4597 struct xps_dev_maps *dev_maps; 4598 struct sock *sk = skb->sk; 4599 int queue_index = -1; 4600 4601 if (!static_key_false(&xps_needed)) 4602 return -1; 4603 4604 rcu_read_lock(); 4605 if (!static_key_false(&xps_rxqs_needed)) 4606 goto get_cpus_map; 4607 4608 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]); 4609 if (dev_maps) { 4610 int tci = sk_rx_queue_get(sk); 4611 4612 if (tci >= 0) 4613 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4614 tci); 4615 } 4616 4617 get_cpus_map: 4618 if (queue_index < 0) { 4619 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]); 4620 if (dev_maps) { 4621 unsigned int tci = skb->sender_cpu - 1; 4622 4623 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4624 tci); 4625 } 4626 } 4627 rcu_read_unlock(); 4628 4629 return queue_index; 4630 #else 4631 return -1; 4632 #endif 4633 } 4634 4635 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb, 4636 struct net_device *sb_dev) 4637 { 4638 return 0; 4639 } 4640 EXPORT_SYMBOL(dev_pick_tx_zero); 4641 4642 int sk_tx_queue_get(const struct sock *sk) 4643 { 4644 int resel, val; 4645 4646 if (!sk) 4647 return -1; 4648 /* Paired with WRITE_ONCE() in sk_tx_queue_clear() 4649 * and sk_tx_queue_set(). 4650 */ 4651 val = READ_ONCE(sk->sk_tx_queue_mapping); 4652 4653 if (val == NO_QUEUE_MAPPING) 4654 return -1; 4655 4656 if (!sk_fullsock(sk)) 4657 return val; 4658 4659 resel = READ_ONCE(sock_net(sk)->core.sysctl_txq_reselection); 4660 if (resel && time_is_before_jiffies( 4661 READ_ONCE(sk->sk_tx_queue_mapping_jiffies) + resel)) 4662 return -1; 4663 4664 return val; 4665 } 4666 EXPORT_SYMBOL(sk_tx_queue_get); 4667 4668 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb, 4669 struct net_device *sb_dev) 4670 { 4671 struct sock *sk = skb->sk; 4672 int queue_index = sk_tx_queue_get(sk); 4673 4674 sb_dev = sb_dev ? : dev; 4675 4676 if (queue_index < 0 || skb->ooo_okay || 4677 queue_index >= dev->real_num_tx_queues) { 4678 int new_index = get_xps_queue(dev, sb_dev, skb); 4679 4680 if (new_index < 0) 4681 new_index = skb_tx_hash(dev, sb_dev, skb); 4682 4683 if (sk && sk_fullsock(sk) && 4684 rcu_access_pointer(sk->sk_dst_cache)) 4685 sk_tx_queue_set(sk, new_index); 4686 4687 queue_index = new_index; 4688 } 4689 4690 return queue_index; 4691 } 4692 EXPORT_SYMBOL(netdev_pick_tx); 4693 4694 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev, 4695 struct sk_buff *skb, 4696 struct net_device *sb_dev) 4697 { 4698 int queue_index = 0; 4699 4700 #ifdef CONFIG_XPS 4701 u32 sender_cpu = skb->sender_cpu - 1; 4702 4703 if (sender_cpu >= (u32)NR_CPUS) 4704 skb->sender_cpu = raw_smp_processor_id() + 1; 4705 #endif 4706 4707 if (dev->real_num_tx_queues != 1) { 4708 const struct net_device_ops *ops = dev->netdev_ops; 4709 4710 if (ops->ndo_select_queue) 4711 queue_index = ops->ndo_select_queue(dev, skb, sb_dev); 4712 else 4713 queue_index = netdev_pick_tx(dev, skb, sb_dev); 4714 4715 queue_index = netdev_cap_txqueue(dev, queue_index); 4716 } 4717 4718 skb_set_queue_mapping(skb, queue_index); 4719 return netdev_get_tx_queue(dev, queue_index); 4720 } 4721 4722 /** 4723 * __dev_queue_xmit() - transmit a buffer 4724 * @skb: buffer to transmit 4725 * @sb_dev: suboordinate device used for L2 forwarding offload 4726 * 4727 * Queue a buffer for transmission to a network device. The caller must 4728 * have set the device and priority and built the buffer before calling 4729 * this function. The function can be called from an interrupt. 4730 * 4731 * When calling this method, interrupts MUST be enabled. This is because 4732 * the BH enable code must have IRQs enabled so that it will not deadlock. 4733 * 4734 * Regardless of the return value, the skb is consumed, so it is currently 4735 * difficult to retry a send to this method. (You can bump the ref count 4736 * before sending to hold a reference for retry if you are careful.) 4737 * 4738 * Return: 4739 * * 0 - buffer successfully transmitted 4740 * * positive qdisc return code - NET_XMIT_DROP etc. 4741 * * negative errno - other errors 4742 */ 4743 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev) 4744 { 4745 struct net_device *dev = skb->dev; 4746 struct netdev_queue *txq = NULL; 4747 struct Qdisc *q; 4748 int rc = -ENOMEM; 4749 bool again = false; 4750 4751 skb_reset_mac_header(skb); 4752 skb_assert_len(skb); 4753 4754 if (unlikely(skb_shinfo(skb)->tx_flags & 4755 (SKBTX_SCHED_TSTAMP | SKBTX_BPF))) 4756 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED); 4757 4758 /* Disable soft irqs for various locks below. Also 4759 * stops preemption for RCU. 4760 */ 4761 rcu_read_lock_bh(); 4762 4763 skb_update_prio(skb); 4764 4765 qdisc_pkt_len_segs_init(skb); 4766 tcx_set_ingress(skb, false); 4767 #ifdef CONFIG_NET_EGRESS 4768 if (static_branch_unlikely(&egress_needed_key)) { 4769 if (nf_hook_egress_active()) { 4770 skb = nf_hook_egress(skb, &rc, dev); 4771 if (!skb) 4772 goto out; 4773 } 4774 4775 netdev_xmit_skip_txqueue(false); 4776 4777 nf_skip_egress(skb, true); 4778 skb = sch_handle_egress(skb, &rc, dev); 4779 if (!skb) 4780 goto out; 4781 nf_skip_egress(skb, false); 4782 4783 if (netdev_xmit_txqueue_skipped()) 4784 txq = netdev_tx_queue_mapping(dev, skb); 4785 } 4786 #endif 4787 /* If device/qdisc don't need skb->dst, release it right now while 4788 * its hot in this cpu cache. 4789 */ 4790 if (dev->priv_flags & IFF_XMIT_DST_RELEASE) 4791 skb_dst_drop(skb); 4792 else 4793 skb_dst_force(skb); 4794 4795 if (!txq) 4796 txq = netdev_core_pick_tx(dev, skb, sb_dev); 4797 4798 q = rcu_dereference_bh(txq->qdisc); 4799 4800 trace_net_dev_queue(skb); 4801 if (q->enqueue) { 4802 rc = __dev_xmit_skb(skb, q, dev, txq); 4803 goto out; 4804 } 4805 4806 /* The device has no queue. Common case for software devices: 4807 * loopback, all the sorts of tunnels... 4808 4809 * Really, it is unlikely that netif_tx_lock protection is necessary 4810 * here. (f.e. loopback and IP tunnels are clean ignoring statistics 4811 * counters.) 4812 * However, it is possible, that they rely on protection 4813 * made by us here. 4814 4815 * Check this and shot the lock. It is not prone from deadlocks. 4816 *Either shot noqueue qdisc, it is even simpler 8) 4817 */ 4818 if (dev->flags & IFF_UP) { 4819 int cpu = smp_processor_id(); /* ok because BHs are off */ 4820 4821 /* Other cpus might concurrently change txq->xmit_lock_owner 4822 * to -1 or to their cpu id, but not to our id. 4823 */ 4824 if (READ_ONCE(txq->xmit_lock_owner) != cpu) { 4825 if (dev_xmit_recursion()) 4826 goto recursion_alert; 4827 4828 skb = validate_xmit_skb(skb, dev, &again); 4829 if (!skb) 4830 goto out; 4831 4832 HARD_TX_LOCK(dev, txq, cpu); 4833 4834 if (!netif_xmit_stopped(txq)) { 4835 dev_xmit_recursion_inc(); 4836 skb = dev_hard_start_xmit(skb, dev, txq, &rc); 4837 dev_xmit_recursion_dec(); 4838 if (dev_xmit_complete(rc)) { 4839 HARD_TX_UNLOCK(dev, txq); 4840 goto out; 4841 } 4842 } 4843 HARD_TX_UNLOCK(dev, txq); 4844 net_crit_ratelimited("Virtual device %s asks to queue packet!\n", 4845 dev->name); 4846 } else { 4847 /* Recursion is detected! It is possible, 4848 * unfortunately 4849 */ 4850 recursion_alert: 4851 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n", 4852 dev->name); 4853 } 4854 } 4855 4856 rc = -ENETDOWN; 4857 rcu_read_unlock_bh(); 4858 4859 dev_core_stats_tx_dropped_inc(dev); 4860 kfree_skb_list(skb); 4861 return rc; 4862 out: 4863 rcu_read_unlock_bh(); 4864 return rc; 4865 } 4866 EXPORT_SYMBOL(__dev_queue_xmit); 4867 4868 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id) 4869 { 4870 struct net_device *dev = skb->dev; 4871 struct sk_buff *orig_skb = skb; 4872 struct netdev_queue *txq; 4873 int ret = NETDEV_TX_BUSY; 4874 bool again = false; 4875 4876 if (unlikely(!netif_running(dev) || 4877 !netif_carrier_ok(dev))) 4878 goto drop; 4879 4880 skb = validate_xmit_skb_list(skb, dev, &again); 4881 if (skb != orig_skb) 4882 goto drop; 4883 4884 skb_set_queue_mapping(skb, queue_id); 4885 txq = skb_get_tx_queue(dev, skb); 4886 4887 local_bh_disable(); 4888 4889 dev_xmit_recursion_inc(); 4890 HARD_TX_LOCK(dev, txq, smp_processor_id()); 4891 if (!netif_xmit_frozen_or_drv_stopped(txq)) 4892 ret = netdev_start_xmit(skb, dev, txq, false); 4893 HARD_TX_UNLOCK(dev, txq); 4894 dev_xmit_recursion_dec(); 4895 4896 local_bh_enable(); 4897 return ret; 4898 drop: 4899 dev_core_stats_tx_dropped_inc(dev); 4900 kfree_skb_list(skb); 4901 return NET_XMIT_DROP; 4902 } 4903 EXPORT_SYMBOL(__dev_direct_xmit); 4904 4905 /************************************************************************* 4906 * Receiver routines 4907 *************************************************************************/ 4908 static DEFINE_PER_CPU(struct task_struct *, backlog_napi); 4909 4910 int weight_p __read_mostly = 64; /* old backlog weight */ 4911 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */ 4912 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */ 4913 4914 /* Called with irq disabled */ 4915 static inline void ____napi_schedule(struct softnet_data *sd, 4916 struct napi_struct *napi) 4917 { 4918 struct task_struct *thread; 4919 4920 lockdep_assert_irqs_disabled(); 4921 4922 if (test_bit(NAPI_STATE_THREADED, &napi->state)) { 4923 /* Paired with smp_mb__before_atomic() in 4924 * napi_enable()/netif_set_threaded(). 4925 * Use READ_ONCE() to guarantee a complete 4926 * read on napi->thread. Only call 4927 * wake_up_process() when it's not NULL. 4928 */ 4929 thread = READ_ONCE(napi->thread); 4930 if (thread) { 4931 if (use_backlog_threads() && thread == raw_cpu_read(backlog_napi)) 4932 goto use_local_napi; 4933 4934 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 4935 wake_up_process(thread); 4936 return; 4937 } 4938 } 4939 4940 use_local_napi: 4941 DEBUG_NET_WARN_ON_ONCE(!list_empty(&napi->poll_list)); 4942 list_add_tail(&napi->poll_list, &sd->poll_list); 4943 WRITE_ONCE(napi->list_owner, smp_processor_id()); 4944 /* If not called from net_rx_action() 4945 * we have to raise NET_RX_SOFTIRQ. 4946 */ 4947 if (!sd->in_net_rx_action) 4948 raise_softirq_irqoff(NET_RX_SOFTIRQ); 4949 } 4950 4951 #ifdef CONFIG_RPS 4952 4953 struct static_key_false rps_needed __read_mostly; 4954 EXPORT_SYMBOL(rps_needed); 4955 struct static_key_false rfs_needed __read_mostly; 4956 EXPORT_SYMBOL(rfs_needed); 4957 4958 static u32 rfs_slot(u32 hash, const struct rps_dev_flow_table *flow_table) 4959 { 4960 return hash_32(hash, flow_table->log); 4961 } 4962 4963 #ifdef CONFIG_RFS_ACCEL 4964 /** 4965 * rps_flow_is_active - check whether the flow is recently active. 4966 * @rflow: Specific flow to check activity. 4967 * @flow_table: per-queue flowtable that @rflow belongs to. 4968 * @cpu: CPU saved in @rflow. 4969 * 4970 * If the CPU has processed many packets since the flow's last activity 4971 * (beyond 10 times the table size), the flow is considered stale. 4972 * 4973 * Return: true if flow was recently active. 4974 */ 4975 static bool rps_flow_is_active(struct rps_dev_flow *rflow, 4976 struct rps_dev_flow_table *flow_table, 4977 unsigned int cpu) 4978 { 4979 unsigned int flow_last_active; 4980 unsigned int sd_input_head; 4981 4982 if (cpu >= nr_cpu_ids) 4983 return false; 4984 4985 sd_input_head = READ_ONCE(per_cpu(softnet_data, cpu).input_queue_head); 4986 flow_last_active = READ_ONCE(rflow->last_qtail); 4987 4988 return (int)(sd_input_head - flow_last_active) < 4989 (int)(10 << flow_table->log); 4990 } 4991 #endif 4992 4993 static struct rps_dev_flow * 4994 set_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4995 struct rps_dev_flow *rflow, u16 next_cpu, u32 hash, 4996 u32 flow_id) 4997 { 4998 if (next_cpu < nr_cpu_ids) { 4999 u32 head; 5000 #ifdef CONFIG_RFS_ACCEL 5001 struct netdev_rx_queue *rxqueue; 5002 struct rps_dev_flow_table *flow_table; 5003 struct rps_dev_flow *old_rflow; 5004 struct rps_dev_flow *tmp_rflow; 5005 unsigned int tmp_cpu; 5006 u16 rxq_index; 5007 int rc; 5008 5009 /* Should we steer this flow to a different hardware queue? */ 5010 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap || 5011 !(dev->features & NETIF_F_NTUPLE)) 5012 goto out; 5013 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu); 5014 if (rxq_index == skb_get_rx_queue(skb)) 5015 goto out; 5016 5017 rxqueue = dev->_rx + rxq_index; 5018 flow_table = rcu_dereference(rxqueue->rps_flow_table); 5019 if (!flow_table) 5020 goto out; 5021 5022 tmp_rflow = &flow_table->flows[flow_id]; 5023 tmp_cpu = READ_ONCE(tmp_rflow->cpu); 5024 5025 if (READ_ONCE(tmp_rflow->filter) != RPS_NO_FILTER) { 5026 if (rps_flow_is_active(tmp_rflow, flow_table, 5027 tmp_cpu)) { 5028 if (hash != READ_ONCE(tmp_rflow->hash) || 5029 next_cpu == tmp_cpu) 5030 goto out; 5031 } 5032 } 5033 5034 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb, 5035 rxq_index, flow_id); 5036 if (rc < 0) 5037 goto out; 5038 5039 old_rflow = rflow; 5040 rflow = tmp_rflow; 5041 WRITE_ONCE(rflow->filter, rc); 5042 WRITE_ONCE(rflow->hash, hash); 5043 5044 if (old_rflow->filter == rc) 5045 WRITE_ONCE(old_rflow->filter, RPS_NO_FILTER); 5046 out: 5047 #endif 5048 head = READ_ONCE(per_cpu(softnet_data, next_cpu).input_queue_head); 5049 rps_input_queue_tail_save(&rflow->last_qtail, head); 5050 } 5051 5052 WRITE_ONCE(rflow->cpu, next_cpu); 5053 return rflow; 5054 } 5055 5056 /* 5057 * get_rps_cpu is called from netif_receive_skb and returns the target 5058 * CPU from the RPS map of the receiving queue for a given skb. 5059 * rcu_read_lock must be held on entry. 5060 */ 5061 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb, 5062 struct rps_dev_flow **rflowp) 5063 { 5064 const struct rps_sock_flow_table *sock_flow_table; 5065 struct netdev_rx_queue *rxqueue = dev->_rx; 5066 struct rps_dev_flow_table *flow_table; 5067 struct rps_map *map; 5068 int cpu = -1; 5069 u32 flow_id; 5070 u32 tcpu; 5071 u32 hash; 5072 5073 if (skb_rx_queue_recorded(skb)) { 5074 u16 index = skb_get_rx_queue(skb); 5075 5076 if (unlikely(index >= dev->real_num_rx_queues)) { 5077 WARN_ONCE(dev->real_num_rx_queues > 1, 5078 "%s received packet on queue %u, but number " 5079 "of RX queues is %u\n", 5080 dev->name, index, dev->real_num_rx_queues); 5081 goto done; 5082 } 5083 rxqueue += index; 5084 } 5085 5086 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */ 5087 5088 flow_table = rcu_dereference(rxqueue->rps_flow_table); 5089 map = rcu_dereference(rxqueue->rps_map); 5090 if (!flow_table && !map) 5091 goto done; 5092 5093 skb_reset_network_header(skb); 5094 hash = skb_get_hash(skb); 5095 if (!hash) 5096 goto done; 5097 5098 sock_flow_table = rcu_dereference(net_hotdata.rps_sock_flow_table); 5099 if (flow_table && sock_flow_table) { 5100 struct rps_dev_flow *rflow; 5101 u32 next_cpu; 5102 u32 ident; 5103 5104 /* First check into global flow table if there is a match. 5105 * This READ_ONCE() pairs with WRITE_ONCE() from rps_record_sock_flow(). 5106 */ 5107 ident = READ_ONCE(sock_flow_table->ents[hash & sock_flow_table->mask]); 5108 if ((ident ^ hash) & ~net_hotdata.rps_cpu_mask) 5109 goto try_rps; 5110 5111 next_cpu = ident & net_hotdata.rps_cpu_mask; 5112 5113 /* OK, now we know there is a match, 5114 * we can look at the local (per receive queue) flow table 5115 */ 5116 flow_id = rfs_slot(hash, flow_table); 5117 rflow = &flow_table->flows[flow_id]; 5118 tcpu = rflow->cpu; 5119 5120 /* 5121 * If the desired CPU (where last recvmsg was done) is 5122 * different from current CPU (one in the rx-queue flow 5123 * table entry), switch if one of the following holds: 5124 * - Current CPU is unset (>= nr_cpu_ids). 5125 * - Current CPU is offline. 5126 * - The current CPU's queue tail has advanced beyond the 5127 * last packet that was enqueued using this table entry. 5128 * This guarantees that all previous packets for the flow 5129 * have been dequeued, thus preserving in order delivery. 5130 */ 5131 if (unlikely(tcpu != next_cpu) && 5132 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) || 5133 ((int)(READ_ONCE(per_cpu(softnet_data, tcpu).input_queue_head) - 5134 rflow->last_qtail)) >= 0)) { 5135 tcpu = next_cpu; 5136 rflow = set_rps_cpu(dev, skb, rflow, next_cpu, hash, 5137 flow_id); 5138 } 5139 5140 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) { 5141 *rflowp = rflow; 5142 cpu = tcpu; 5143 goto done; 5144 } 5145 } 5146 5147 try_rps: 5148 5149 if (map) { 5150 tcpu = map->cpus[reciprocal_scale(hash, map->len)]; 5151 if (cpu_online(tcpu)) { 5152 cpu = tcpu; 5153 goto done; 5154 } 5155 } 5156 5157 done: 5158 return cpu; 5159 } 5160 5161 #ifdef CONFIG_RFS_ACCEL 5162 5163 /** 5164 * rps_may_expire_flow - check whether an RFS hardware filter may be removed 5165 * @dev: Device on which the filter was set 5166 * @rxq_index: RX queue index 5167 * @flow_id: Flow ID passed to ndo_rx_flow_steer() 5168 * @filter_id: Filter ID returned by ndo_rx_flow_steer() 5169 * 5170 * Drivers that implement ndo_rx_flow_steer() should periodically call 5171 * this function for each installed filter and remove the filters for 5172 * which it returns %true. 5173 */ 5174 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index, 5175 u32 flow_id, u16 filter_id) 5176 { 5177 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index; 5178 struct rps_dev_flow_table *flow_table; 5179 struct rps_dev_flow *rflow; 5180 bool expire = true; 5181 5182 rcu_read_lock(); 5183 flow_table = rcu_dereference(rxqueue->rps_flow_table); 5184 if (flow_table && flow_id < (1UL << flow_table->log)) { 5185 unsigned int cpu; 5186 5187 rflow = &flow_table->flows[flow_id]; 5188 cpu = READ_ONCE(rflow->cpu); 5189 if (READ_ONCE(rflow->filter) == filter_id && 5190 rps_flow_is_active(rflow, flow_table, cpu)) 5191 expire = false; 5192 } 5193 rcu_read_unlock(); 5194 return expire; 5195 } 5196 EXPORT_SYMBOL(rps_may_expire_flow); 5197 5198 #endif /* CONFIG_RFS_ACCEL */ 5199 5200 /* Called from hardirq (IPI) context */ 5201 static void rps_trigger_softirq(void *data) 5202 { 5203 struct softnet_data *sd = data; 5204 5205 ____napi_schedule(sd, &sd->backlog); 5206 /* Pairs with READ_ONCE() in softnet_seq_show() */ 5207 WRITE_ONCE(sd->received_rps, sd->received_rps + 1); 5208 } 5209 5210 #endif /* CONFIG_RPS */ 5211 5212 /* Called from hardirq (IPI) context */ 5213 static void trigger_rx_softirq(void *data) 5214 { 5215 struct softnet_data *sd = data; 5216 5217 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 5218 smp_store_release(&sd->defer_ipi_scheduled, 0); 5219 } 5220 5221 /* 5222 * After we queued a packet into sd->input_pkt_queue, 5223 * we need to make sure this queue is serviced soon. 5224 * 5225 * - If this is another cpu queue, link it to our rps_ipi_list, 5226 * and make sure we will process rps_ipi_list from net_rx_action(). 5227 * 5228 * - If this is our own queue, NAPI schedule our backlog. 5229 * Note that this also raises NET_RX_SOFTIRQ. 5230 */ 5231 static void napi_schedule_rps(struct softnet_data *sd) 5232 { 5233 struct softnet_data *mysd = this_cpu_ptr(&softnet_data); 5234 5235 #ifdef CONFIG_RPS 5236 if (sd != mysd) { 5237 if (use_backlog_threads()) { 5238 __napi_schedule_irqoff(&sd->backlog); 5239 return; 5240 } 5241 5242 sd->rps_ipi_next = mysd->rps_ipi_list; 5243 mysd->rps_ipi_list = sd; 5244 5245 /* If not called from net_rx_action() or napi_threaded_poll() 5246 * we have to raise NET_RX_SOFTIRQ. 5247 */ 5248 if (!mysd->in_net_rx_action && !mysd->in_napi_threaded_poll) 5249 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 5250 return; 5251 } 5252 #endif /* CONFIG_RPS */ 5253 __napi_schedule_irqoff(&mysd->backlog); 5254 } 5255 5256 void kick_defer_list_purge(unsigned int cpu) 5257 { 5258 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 5259 unsigned long flags; 5260 5261 if (use_backlog_threads()) { 5262 backlog_lock_irq_save(sd, &flags); 5263 5264 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) 5265 __napi_schedule_irqoff(&sd->backlog); 5266 5267 backlog_unlock_irq_restore(sd, flags); 5268 5269 } else if (!cmpxchg(&sd->defer_ipi_scheduled, 0, 1)) { 5270 smp_call_function_single_async(cpu, &sd->defer_csd); 5271 } 5272 } 5273 5274 #ifdef CONFIG_NET_FLOW_LIMIT 5275 int netdev_flow_limit_table_len __read_mostly = (1 << 12); 5276 #endif 5277 5278 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen, 5279 int max_backlog) 5280 { 5281 #ifdef CONFIG_NET_FLOW_LIMIT 5282 unsigned int old_flow, new_flow; 5283 const struct softnet_data *sd; 5284 struct sd_flow_limit *fl; 5285 5286 if (likely(qlen < (max_backlog >> 1))) 5287 return false; 5288 5289 sd = this_cpu_ptr(&softnet_data); 5290 5291 rcu_read_lock(); 5292 fl = rcu_dereference(sd->flow_limit); 5293 if (fl) { 5294 new_flow = hash_32(skb_get_hash(skb), fl->log_buckets); 5295 old_flow = fl->history[fl->history_head]; 5296 fl->history[fl->history_head] = new_flow; 5297 5298 fl->history_head++; 5299 fl->history_head &= FLOW_LIMIT_HISTORY - 1; 5300 5301 if (likely(fl->buckets[old_flow])) 5302 fl->buckets[old_flow]--; 5303 5304 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) { 5305 /* Pairs with READ_ONCE() in softnet_seq_show() */ 5306 WRITE_ONCE(fl->count, fl->count + 1); 5307 rcu_read_unlock(); 5308 return true; 5309 } 5310 } 5311 rcu_read_unlock(); 5312 #endif 5313 return false; 5314 } 5315 5316 /* 5317 * enqueue_to_backlog is called to queue an skb to a per CPU backlog 5318 * queue (may be a remote CPU queue). 5319 */ 5320 static int enqueue_to_backlog(struct sk_buff *skb, int cpu, 5321 unsigned int *qtail) 5322 { 5323 enum skb_drop_reason reason; 5324 struct softnet_data *sd; 5325 unsigned long flags; 5326 unsigned int qlen; 5327 int max_backlog; 5328 u32 tail; 5329 5330 reason = SKB_DROP_REASON_DEV_READY; 5331 if (unlikely(!netif_running(skb->dev))) 5332 goto bad_dev; 5333 5334 sd = &per_cpu(softnet_data, cpu); 5335 5336 qlen = skb_queue_len_lockless(&sd->input_pkt_queue); 5337 max_backlog = READ_ONCE(net_hotdata.max_backlog); 5338 if (unlikely(qlen > max_backlog) || 5339 skb_flow_limit(skb, qlen, max_backlog)) 5340 goto cpu_backlog_drop; 5341 backlog_lock_irq_save(sd, &flags); 5342 qlen = skb_queue_len(&sd->input_pkt_queue); 5343 if (likely(qlen <= max_backlog)) { 5344 if (!qlen) { 5345 /* Schedule NAPI for backlog device. We can use 5346 * non atomic operation as we own the queue lock. 5347 */ 5348 if (!__test_and_set_bit(NAPI_STATE_SCHED, 5349 &sd->backlog.state)) 5350 napi_schedule_rps(sd); 5351 } 5352 __skb_queue_tail(&sd->input_pkt_queue, skb); 5353 tail = rps_input_queue_tail_incr(sd); 5354 backlog_unlock_irq_restore(sd, flags); 5355 5356 /* save the tail outside of the critical section */ 5357 rps_input_queue_tail_save(qtail, tail); 5358 return NET_RX_SUCCESS; 5359 } 5360 5361 backlog_unlock_irq_restore(sd, flags); 5362 5363 cpu_backlog_drop: 5364 reason = SKB_DROP_REASON_CPU_BACKLOG; 5365 numa_drop_add(&sd->drop_counters, 1); 5366 bad_dev: 5367 dev_core_stats_rx_dropped_inc(skb->dev); 5368 kfree_skb_reason(skb, reason); 5369 return NET_RX_DROP; 5370 } 5371 5372 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb) 5373 { 5374 struct net_device *dev = skb->dev; 5375 struct netdev_rx_queue *rxqueue; 5376 5377 rxqueue = dev->_rx; 5378 5379 if (skb_rx_queue_recorded(skb)) { 5380 u16 index = skb_get_rx_queue(skb); 5381 5382 if (unlikely(index >= dev->real_num_rx_queues)) { 5383 WARN_ONCE(dev->real_num_rx_queues > 1, 5384 "%s received packet on queue %u, but number " 5385 "of RX queues is %u\n", 5386 dev->name, index, dev->real_num_rx_queues); 5387 5388 return rxqueue; /* Return first rxqueue */ 5389 } 5390 rxqueue += index; 5391 } 5392 return rxqueue; 5393 } 5394 5395 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, 5396 const struct bpf_prog *xdp_prog) 5397 { 5398 void *orig_data, *orig_data_end, *hard_start; 5399 struct netdev_rx_queue *rxqueue; 5400 bool orig_bcast, orig_host; 5401 u32 mac_len, frame_sz; 5402 __be16 orig_eth_type; 5403 struct ethhdr *eth; 5404 u32 metalen, act; 5405 int off; 5406 5407 /* The XDP program wants to see the packet starting at the MAC 5408 * header. 5409 */ 5410 mac_len = skb->data - skb_mac_header(skb); 5411 hard_start = skb->data - skb_headroom(skb); 5412 5413 /* SKB "head" area always have tailroom for skb_shared_info */ 5414 frame_sz = (void *)skb_end_pointer(skb) - hard_start; 5415 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 5416 5417 rxqueue = netif_get_rxqueue(skb); 5418 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq); 5419 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len, 5420 skb_headlen(skb) + mac_len, true); 5421 if (skb_is_nonlinear(skb)) { 5422 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 5423 xdp_buff_set_frags_flag(xdp); 5424 } else { 5425 xdp_buff_clear_frags_flag(xdp); 5426 } 5427 5428 orig_data_end = xdp->data_end; 5429 orig_data = xdp->data; 5430 eth = (struct ethhdr *)xdp->data; 5431 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr); 5432 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest); 5433 orig_eth_type = eth->h_proto; 5434 5435 act = bpf_prog_run_xdp(xdp_prog, xdp); 5436 5437 /* check if bpf_xdp_adjust_head was used */ 5438 off = xdp->data - orig_data; 5439 if (off) { 5440 if (off > 0) 5441 __skb_pull(skb, off); 5442 else if (off < 0) 5443 __skb_push(skb, -off); 5444 5445 skb->mac_header += off; 5446 skb_reset_network_header(skb); 5447 } 5448 5449 /* check if bpf_xdp_adjust_tail was used */ 5450 off = xdp->data_end - orig_data_end; 5451 if (off != 0) { 5452 skb_set_tail_pointer(skb, xdp->data_end - xdp->data); 5453 skb->len += off; /* positive on grow, negative on shrink */ 5454 } 5455 5456 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 5457 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here. 5458 */ 5459 if (xdp_buff_has_frags(xdp)) 5460 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 5461 else 5462 skb->data_len = 0; 5463 5464 /* check if XDP changed eth hdr such SKB needs update */ 5465 eth = (struct ethhdr *)xdp->data; 5466 if ((orig_eth_type != eth->h_proto) || 5467 (orig_host != ether_addr_equal_64bits(eth->h_dest, 5468 skb->dev->dev_addr)) || 5469 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) { 5470 __skb_push(skb, ETH_HLEN); 5471 skb->pkt_type = PACKET_HOST; 5472 skb->protocol = eth_type_trans(skb, skb->dev); 5473 } 5474 5475 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull 5476 * before calling us again on redirect path. We do not call do_redirect 5477 * as we leave that up to the caller. 5478 * 5479 * Caller is responsible for managing lifetime of skb (i.e. calling 5480 * kfree_skb in response to actions it cannot handle/XDP_DROP). 5481 */ 5482 switch (act) { 5483 case XDP_REDIRECT: 5484 case XDP_TX: 5485 __skb_push(skb, mac_len); 5486 break; 5487 case XDP_PASS: 5488 metalen = xdp->data - xdp->data_meta; 5489 if (metalen) 5490 skb_metadata_set(skb, metalen); 5491 break; 5492 } 5493 5494 return act; 5495 } 5496 5497 static int 5498 netif_skb_check_for_xdp(struct sk_buff **pskb, const struct bpf_prog *prog) 5499 { 5500 struct sk_buff *skb = *pskb; 5501 int err, hroom, troom; 5502 5503 local_lock_nested_bh(&system_page_pool.bh_lock); 5504 err = skb_cow_data_for_xdp(this_cpu_read(system_page_pool.pool), pskb, prog); 5505 local_unlock_nested_bh(&system_page_pool.bh_lock); 5506 if (!err) 5507 return 0; 5508 5509 /* In case we have to go down the path and also linearize, 5510 * then lets do the pskb_expand_head() work just once here. 5511 */ 5512 hroom = XDP_PACKET_HEADROOM - skb_headroom(skb); 5513 troom = skb->tail + skb->data_len - skb->end; 5514 err = pskb_expand_head(skb, 5515 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0, 5516 troom > 0 ? troom + 128 : 0, GFP_ATOMIC); 5517 if (err) 5518 return err; 5519 5520 return skb_linearize(skb); 5521 } 5522 5523 static u32 netif_receive_generic_xdp(struct sk_buff **pskb, 5524 struct xdp_buff *xdp, 5525 const struct bpf_prog *xdp_prog) 5526 { 5527 struct sk_buff *skb = *pskb; 5528 u32 mac_len, act = XDP_DROP; 5529 5530 /* Reinjected packets coming from act_mirred or similar should 5531 * not get XDP generic processing. 5532 */ 5533 if (skb_is_redirected(skb)) 5534 return XDP_PASS; 5535 5536 /* XDP packets must have sufficient headroom of XDP_PACKET_HEADROOM 5537 * bytes. This is the guarantee that also native XDP provides, 5538 * thus we need to do it here as well. 5539 */ 5540 mac_len = skb->data - skb_mac_header(skb); 5541 __skb_push(skb, mac_len); 5542 5543 if (skb_cloned(skb) || skb_is_nonlinear(skb) || 5544 skb_headroom(skb) < XDP_PACKET_HEADROOM) { 5545 if (netif_skb_check_for_xdp(pskb, xdp_prog)) 5546 goto do_drop; 5547 } 5548 5549 __skb_pull(*pskb, mac_len); 5550 5551 act = bpf_prog_run_generic_xdp(*pskb, xdp, xdp_prog); 5552 switch (act) { 5553 case XDP_REDIRECT: 5554 case XDP_TX: 5555 case XDP_PASS: 5556 break; 5557 default: 5558 bpf_warn_invalid_xdp_action((*pskb)->dev, xdp_prog, act); 5559 fallthrough; 5560 case XDP_ABORTED: 5561 trace_xdp_exception((*pskb)->dev, xdp_prog, act); 5562 fallthrough; 5563 case XDP_DROP: 5564 do_drop: 5565 kfree_skb(*pskb); 5566 break; 5567 } 5568 5569 return act; 5570 } 5571 5572 /* When doing generic XDP we have to bypass the qdisc layer and the 5573 * network taps in order to match in-driver-XDP behavior. This also means 5574 * that XDP packets are able to starve other packets going through a qdisc, 5575 * and DDOS attacks will be more effective. In-driver-XDP use dedicated TX 5576 * queues, so they do not have this starvation issue. 5577 */ 5578 void generic_xdp_tx(struct sk_buff *skb, const struct bpf_prog *xdp_prog) 5579 { 5580 struct net_device *dev = skb->dev; 5581 struct netdev_queue *txq; 5582 bool free_skb = true; 5583 int cpu, rc; 5584 5585 txq = netdev_core_pick_tx(dev, skb, NULL); 5586 cpu = smp_processor_id(); 5587 HARD_TX_LOCK(dev, txq, cpu); 5588 if (!netif_xmit_frozen_or_drv_stopped(txq)) { 5589 rc = netdev_start_xmit(skb, dev, txq, 0); 5590 if (dev_xmit_complete(rc)) 5591 free_skb = false; 5592 } 5593 HARD_TX_UNLOCK(dev, txq); 5594 if (free_skb) { 5595 trace_xdp_exception(dev, xdp_prog, XDP_TX); 5596 dev_core_stats_tx_dropped_inc(dev); 5597 kfree_skb(skb); 5598 } 5599 } 5600 5601 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key); 5602 5603 int do_xdp_generic(const struct bpf_prog *xdp_prog, struct sk_buff **pskb) 5604 { 5605 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 5606 5607 if (xdp_prog) { 5608 struct xdp_buff xdp; 5609 u32 act; 5610 int err; 5611 5612 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 5613 act = netif_receive_generic_xdp(pskb, &xdp, xdp_prog); 5614 if (act != XDP_PASS) { 5615 switch (act) { 5616 case XDP_REDIRECT: 5617 err = xdp_do_generic_redirect((*pskb)->dev, *pskb, 5618 &xdp, xdp_prog); 5619 if (err) 5620 goto out_redir; 5621 break; 5622 case XDP_TX: 5623 generic_xdp_tx(*pskb, xdp_prog); 5624 break; 5625 } 5626 bpf_net_ctx_clear(bpf_net_ctx); 5627 return XDP_DROP; 5628 } 5629 bpf_net_ctx_clear(bpf_net_ctx); 5630 } 5631 return XDP_PASS; 5632 out_redir: 5633 bpf_net_ctx_clear(bpf_net_ctx); 5634 kfree_skb_reason(*pskb, SKB_DROP_REASON_XDP); 5635 return XDP_DROP; 5636 } 5637 EXPORT_SYMBOL_GPL(do_xdp_generic); 5638 5639 static int netif_rx_internal(struct sk_buff *skb) 5640 { 5641 int ret; 5642 5643 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5644 5645 trace_netif_rx(skb); 5646 5647 #ifdef CONFIG_RPS 5648 if (static_branch_unlikely(&rps_needed)) { 5649 struct rps_dev_flow voidflow, *rflow = &voidflow; 5650 int cpu; 5651 5652 rcu_read_lock(); 5653 5654 cpu = get_rps_cpu(skb->dev, skb, &rflow); 5655 if (cpu < 0) 5656 cpu = smp_processor_id(); 5657 5658 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5659 5660 rcu_read_unlock(); 5661 } else 5662 #endif 5663 { 5664 unsigned int qtail; 5665 5666 ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail); 5667 } 5668 return ret; 5669 } 5670 5671 /** 5672 * __netif_rx - Slightly optimized version of netif_rx 5673 * @skb: buffer to post 5674 * 5675 * This behaves as netif_rx except that it does not disable bottom halves. 5676 * As a result this function may only be invoked from the interrupt context 5677 * (either hard or soft interrupt). 5678 */ 5679 int __netif_rx(struct sk_buff *skb) 5680 { 5681 int ret; 5682 5683 lockdep_assert_once(hardirq_count() | softirq_count()); 5684 5685 trace_netif_rx_entry(skb); 5686 ret = netif_rx_internal(skb); 5687 trace_netif_rx_exit(ret); 5688 return ret; 5689 } 5690 EXPORT_SYMBOL(__netif_rx); 5691 5692 /** 5693 * netif_rx - post buffer to the network code 5694 * @skb: buffer to post 5695 * 5696 * This function receives a packet from a device driver and queues it for 5697 * the upper (protocol) levels to process via the backlog NAPI device. It 5698 * always succeeds. The buffer may be dropped during processing for 5699 * congestion control or by the protocol layers. 5700 * The network buffer is passed via the backlog NAPI device. Modern NIC 5701 * driver should use NAPI and GRO. 5702 * This function can used from interrupt and from process context. The 5703 * caller from process context must not disable interrupts before invoking 5704 * this function. 5705 * 5706 * return values: 5707 * NET_RX_SUCCESS (no congestion) 5708 * NET_RX_DROP (packet was dropped) 5709 * 5710 */ 5711 int netif_rx(struct sk_buff *skb) 5712 { 5713 bool need_bh_off = !(hardirq_count() | softirq_count()); 5714 int ret; 5715 5716 if (need_bh_off) 5717 local_bh_disable(); 5718 trace_netif_rx_entry(skb); 5719 ret = netif_rx_internal(skb); 5720 trace_netif_rx_exit(ret); 5721 if (need_bh_off) 5722 local_bh_enable(); 5723 return ret; 5724 } 5725 EXPORT_SYMBOL(netif_rx); 5726 5727 static __latent_entropy void net_tx_action(void) 5728 { 5729 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 5730 5731 if (sd->completion_queue) { 5732 struct sk_buff *clist; 5733 5734 local_irq_disable(); 5735 clist = sd->completion_queue; 5736 sd->completion_queue = NULL; 5737 local_irq_enable(); 5738 5739 while (clist) { 5740 struct sk_buff *skb = clist; 5741 5742 clist = clist->next; 5743 5744 WARN_ON(refcount_read(&skb->users)); 5745 if (likely(get_kfree_skb_cb(skb)->reason == SKB_CONSUMED)) 5746 trace_consume_skb(skb, net_tx_action); 5747 else 5748 trace_kfree_skb(skb, net_tx_action, 5749 get_kfree_skb_cb(skb)->reason, NULL); 5750 5751 if (skb->fclone != SKB_FCLONE_UNAVAILABLE) 5752 __kfree_skb(skb); 5753 else 5754 __napi_kfree_skb(skb, 5755 get_kfree_skb_cb(skb)->reason); 5756 } 5757 } 5758 5759 if (sd->output_queue) { 5760 struct Qdisc *head; 5761 5762 local_irq_disable(); 5763 head = sd->output_queue; 5764 sd->output_queue = NULL; 5765 sd->output_queue_tailp = &sd->output_queue; 5766 local_irq_enable(); 5767 5768 rcu_read_lock(); 5769 5770 while (head) { 5771 spinlock_t *root_lock = NULL; 5772 struct sk_buff *to_free; 5773 struct Qdisc *q = head; 5774 5775 head = head->next_sched; 5776 5777 /* We need to make sure head->next_sched is read 5778 * before clearing __QDISC_STATE_SCHED 5779 */ 5780 smp_mb__before_atomic(); 5781 5782 if (!(q->flags & TCQ_F_NOLOCK)) { 5783 root_lock = qdisc_lock(q); 5784 spin_lock(root_lock); 5785 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, 5786 &q->state))) { 5787 /* There is a synchronize_net() between 5788 * STATE_DEACTIVATED flag being set and 5789 * qdisc_reset()/some_qdisc_is_busy() in 5790 * dev_deactivate(), so we can safely bail out 5791 * early here to avoid data race between 5792 * qdisc_deactivate() and some_qdisc_is_busy() 5793 * for lockless qdisc. 5794 */ 5795 clear_bit(__QDISC_STATE_SCHED, &q->state); 5796 continue; 5797 } 5798 5799 clear_bit(__QDISC_STATE_SCHED, &q->state); 5800 to_free = qdisc_run(q); 5801 if (root_lock) 5802 spin_unlock(root_lock); 5803 tcf_kfree_skb_list(to_free); 5804 } 5805 5806 rcu_read_unlock(); 5807 } 5808 5809 xfrm_dev_backlog(sd); 5810 } 5811 5812 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE) 5813 /* This hook is defined here for ATM LANE */ 5814 int (*br_fdb_test_addr_hook)(struct net_device *dev, 5815 unsigned char *addr) __read_mostly; 5816 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook); 5817 #endif 5818 5819 /** 5820 * netdev_is_rx_handler_busy - check if receive handler is registered 5821 * @dev: device to check 5822 * 5823 * Check if a receive handler is already registered for a given device. 5824 * Return true if there one. 5825 * 5826 * The caller must hold the rtnl_mutex. 5827 */ 5828 bool netdev_is_rx_handler_busy(struct net_device *dev) 5829 { 5830 ASSERT_RTNL(); 5831 return dev && rtnl_dereference(dev->rx_handler); 5832 } 5833 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy); 5834 5835 /** 5836 * netdev_rx_handler_register - register receive handler 5837 * @dev: device to register a handler for 5838 * @rx_handler: receive handler to register 5839 * @rx_handler_data: data pointer that is used by rx handler 5840 * 5841 * Register a receive handler for a device. This handler will then be 5842 * called from __netif_receive_skb. A negative errno code is returned 5843 * on a failure. 5844 * 5845 * The caller must hold the rtnl_mutex. 5846 * 5847 * For a general description of rx_handler, see enum rx_handler_result. 5848 */ 5849 int netdev_rx_handler_register(struct net_device *dev, 5850 rx_handler_func_t *rx_handler, 5851 void *rx_handler_data) 5852 { 5853 if (netdev_is_rx_handler_busy(dev)) 5854 return -EBUSY; 5855 5856 if (dev->priv_flags & IFF_NO_RX_HANDLER) 5857 return -EINVAL; 5858 5859 /* Note: rx_handler_data must be set before rx_handler */ 5860 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data); 5861 rcu_assign_pointer(dev->rx_handler, rx_handler); 5862 5863 return 0; 5864 } 5865 EXPORT_SYMBOL_GPL(netdev_rx_handler_register); 5866 5867 /** 5868 * netdev_rx_handler_unregister - unregister receive handler 5869 * @dev: device to unregister a handler from 5870 * 5871 * Unregister a receive handler from a device. 5872 * 5873 * The caller must hold the rtnl_mutex. 5874 */ 5875 void netdev_rx_handler_unregister(struct net_device *dev) 5876 { 5877 5878 ASSERT_RTNL(); 5879 RCU_INIT_POINTER(dev->rx_handler, NULL); 5880 /* a reader seeing a non NULL rx_handler in a rcu_read_lock() 5881 * section has a guarantee to see a non NULL rx_handler_data 5882 * as well. 5883 */ 5884 synchronize_net(); 5885 RCU_INIT_POINTER(dev->rx_handler_data, NULL); 5886 } 5887 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister); 5888 5889 /* 5890 * Limit the use of PFMEMALLOC reserves to those protocols that implement 5891 * the special handling of PFMEMALLOC skbs. 5892 */ 5893 static bool skb_pfmemalloc_protocol(struct sk_buff *skb) 5894 { 5895 switch (skb->protocol) { 5896 case htons(ETH_P_ARP): 5897 case htons(ETH_P_IP): 5898 case htons(ETH_P_IPV6): 5899 case htons(ETH_P_8021Q): 5900 case htons(ETH_P_8021AD): 5901 return true; 5902 default: 5903 return false; 5904 } 5905 } 5906 5907 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev, 5908 int *ret, struct net_device *orig_dev) 5909 { 5910 if (nf_hook_ingress_active(skb)) { 5911 int ingress_retval; 5912 5913 if (unlikely(*pt_prev)) { 5914 *ret = deliver_skb(skb, *pt_prev, orig_dev); 5915 *pt_prev = NULL; 5916 } 5917 5918 rcu_read_lock(); 5919 ingress_retval = nf_hook_ingress(skb); 5920 rcu_read_unlock(); 5921 return ingress_retval; 5922 } 5923 return 0; 5924 } 5925 5926 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc, 5927 struct packet_type **ppt_prev) 5928 { 5929 enum skb_drop_reason drop_reason = SKB_DROP_REASON_UNHANDLED_PROTO; 5930 struct packet_type *ptype, *pt_prev; 5931 rx_handler_func_t *rx_handler; 5932 struct sk_buff *skb = *pskb; 5933 struct net_device *orig_dev; 5934 bool deliver_exact = false; 5935 int ret = NET_RX_DROP; 5936 __be16 type; 5937 5938 net_timestamp_check(!READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5939 5940 trace_netif_receive_skb(skb); 5941 5942 orig_dev = skb->dev; 5943 5944 skb_reset_network_header(skb); 5945 #if !defined(CONFIG_DEBUG_NET) 5946 /* We plan to no longer reset the transport header here. 5947 * Give some time to fuzzers and dev build to catch bugs 5948 * in network stacks. 5949 */ 5950 if (!skb_transport_header_was_set(skb)) 5951 skb_reset_transport_header(skb); 5952 #endif 5953 skb_reset_mac_len(skb); 5954 5955 pt_prev = NULL; 5956 5957 another_round: 5958 skb->skb_iif = skb->dev->ifindex; 5959 5960 __this_cpu_inc(softnet_data.processed); 5961 5962 if (static_branch_unlikely(&generic_xdp_needed_key)) { 5963 int ret2; 5964 5965 migrate_disable(); 5966 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), 5967 &skb); 5968 migrate_enable(); 5969 5970 if (ret2 != XDP_PASS) { 5971 ret = NET_RX_DROP; 5972 goto out; 5973 } 5974 } 5975 5976 if (eth_type_vlan(skb->protocol)) { 5977 skb = skb_vlan_untag(skb); 5978 if (unlikely(!skb)) 5979 goto out; 5980 } 5981 5982 if (skb_skip_tc_classify(skb)) 5983 goto skip_classify; 5984 5985 if (pfmemalloc) 5986 goto skip_taps; 5987 5988 list_for_each_entry_rcu(ptype, &dev_net_rcu(skb->dev)->ptype_all, 5989 list) { 5990 if (unlikely(pt_prev)) 5991 ret = deliver_skb(skb, pt_prev, orig_dev); 5992 pt_prev = ptype; 5993 } 5994 5995 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) { 5996 if (unlikely(pt_prev)) 5997 ret = deliver_skb(skb, pt_prev, orig_dev); 5998 pt_prev = ptype; 5999 } 6000 6001 skip_taps: 6002 #ifdef CONFIG_NET_INGRESS 6003 if (static_branch_unlikely(&ingress_needed_key)) { 6004 bool another = false; 6005 6006 nf_skip_egress(skb, true); 6007 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev, 6008 &another); 6009 if (another) 6010 goto another_round; 6011 if (!skb) 6012 goto out; 6013 6014 nf_skip_egress(skb, false); 6015 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0) 6016 goto out; 6017 } 6018 #endif 6019 skb_reset_redirect(skb); 6020 skip_classify: 6021 if (pfmemalloc && !skb_pfmemalloc_protocol(skb)) { 6022 drop_reason = SKB_DROP_REASON_PFMEMALLOC; 6023 goto drop; 6024 } 6025 6026 if (skb_vlan_tag_present(skb)) { 6027 if (unlikely(pt_prev)) { 6028 ret = deliver_skb(skb, pt_prev, orig_dev); 6029 pt_prev = NULL; 6030 } 6031 if (vlan_do_receive(&skb)) 6032 goto another_round; 6033 else if (unlikely(!skb)) 6034 goto out; 6035 } 6036 6037 rx_handler = rcu_dereference(skb->dev->rx_handler); 6038 if (rx_handler) { 6039 if (unlikely(pt_prev)) { 6040 ret = deliver_skb(skb, pt_prev, orig_dev); 6041 pt_prev = NULL; 6042 } 6043 switch (rx_handler(&skb)) { 6044 case RX_HANDLER_CONSUMED: 6045 ret = NET_RX_SUCCESS; 6046 goto out; 6047 case RX_HANDLER_ANOTHER: 6048 goto another_round; 6049 case RX_HANDLER_EXACT: 6050 deliver_exact = true; 6051 break; 6052 case RX_HANDLER_PASS: 6053 break; 6054 default: 6055 BUG(); 6056 } 6057 } 6058 6059 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) { 6060 check_vlan_id: 6061 if (skb_vlan_tag_get_id(skb)) { 6062 /* Vlan id is non 0 and vlan_do_receive() above couldn't 6063 * find vlan device. 6064 */ 6065 skb->pkt_type = PACKET_OTHERHOST; 6066 } else if (eth_type_vlan(skb->protocol)) { 6067 /* Outer header is 802.1P with vlan 0, inner header is 6068 * 802.1Q or 802.1AD and vlan_do_receive() above could 6069 * not find vlan dev for vlan id 0. 6070 */ 6071 __vlan_hwaccel_clear_tag(skb); 6072 skb = skb_vlan_untag(skb); 6073 if (unlikely(!skb)) 6074 goto out; 6075 if (vlan_do_receive(&skb)) 6076 /* After stripping off 802.1P header with vlan 0 6077 * vlan dev is found for inner header. 6078 */ 6079 goto another_round; 6080 else if (unlikely(!skb)) 6081 goto out; 6082 else 6083 /* We have stripped outer 802.1P vlan 0 header. 6084 * But could not find vlan dev. 6085 * check again for vlan id to set OTHERHOST. 6086 */ 6087 goto check_vlan_id; 6088 } 6089 /* Note: we might in the future use prio bits 6090 * and set skb->priority like in vlan_do_receive() 6091 * For the time being, just ignore Priority Code Point 6092 */ 6093 __vlan_hwaccel_clear_tag(skb); 6094 } 6095 6096 type = skb->protocol; 6097 6098 /* deliver only exact match when indicated */ 6099 if (likely(!deliver_exact)) { 6100 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6101 &ptype_base[ntohs(type) & 6102 PTYPE_HASH_MASK]); 6103 6104 /* orig_dev and skb->dev could belong to different netns; 6105 * Even in such case we need to traverse only the list 6106 * coming from skb->dev, as the ptype owner (packet socket) 6107 * will use dev_net(skb->dev) to do namespace filtering. 6108 */ 6109 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6110 &dev_net_rcu(skb->dev)->ptype_specific); 6111 } 6112 6113 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6114 &orig_dev->ptype_specific); 6115 6116 if (unlikely(skb->dev != orig_dev)) { 6117 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6118 &skb->dev->ptype_specific); 6119 } 6120 6121 if (pt_prev) { 6122 *ppt_prev = pt_prev; 6123 } else { 6124 drop: 6125 if (!deliver_exact) 6126 dev_core_stats_rx_dropped_inc(skb->dev); 6127 else 6128 dev_core_stats_rx_nohandler_inc(skb->dev); 6129 6130 kfree_skb_reason(skb, drop_reason); 6131 /* Jamal, now you will not able to escape explaining 6132 * me how you were going to use this. :-) 6133 */ 6134 ret = NET_RX_DROP; 6135 } 6136 6137 out: 6138 /* The invariant here is that if *ppt_prev is not NULL 6139 * then skb should also be non-NULL. 6140 * 6141 * Apparently *ppt_prev assignment above holds this invariant due to 6142 * skb dereferencing near it. 6143 */ 6144 *pskb = skb; 6145 return ret; 6146 } 6147 6148 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc) 6149 { 6150 struct net_device *orig_dev = skb->dev; 6151 struct packet_type *pt_prev = NULL; 6152 int ret; 6153 6154 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 6155 if (pt_prev) 6156 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb, 6157 skb->dev, pt_prev, orig_dev); 6158 return ret; 6159 } 6160 6161 /** 6162 * netif_receive_skb_core - special purpose version of netif_receive_skb 6163 * @skb: buffer to process 6164 * 6165 * More direct receive version of netif_receive_skb(). It should 6166 * only be used by callers that have a need to skip RPS and Generic XDP. 6167 * Caller must also take care of handling if ``(page_is_)pfmemalloc``. 6168 * 6169 * This function may only be called from softirq context and interrupts 6170 * should be enabled. 6171 * 6172 * Return values (usually ignored): 6173 * NET_RX_SUCCESS: no congestion 6174 * NET_RX_DROP: packet was dropped 6175 */ 6176 int netif_receive_skb_core(struct sk_buff *skb) 6177 { 6178 int ret; 6179 6180 rcu_read_lock(); 6181 ret = __netif_receive_skb_one_core(skb, false); 6182 rcu_read_unlock(); 6183 6184 return ret; 6185 } 6186 EXPORT_SYMBOL(netif_receive_skb_core); 6187 6188 static inline void __netif_receive_skb_list_ptype(struct list_head *head, 6189 struct packet_type *pt_prev, 6190 struct net_device *orig_dev) 6191 { 6192 struct sk_buff *skb, *next; 6193 6194 if (!pt_prev) 6195 return; 6196 if (list_empty(head)) 6197 return; 6198 if (pt_prev->list_func != NULL) 6199 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv, 6200 ip_list_rcv, head, pt_prev, orig_dev); 6201 else 6202 list_for_each_entry_safe(skb, next, head, list) { 6203 skb_list_del_init(skb); 6204 pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 6205 } 6206 } 6207 6208 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc) 6209 { 6210 /* Fast-path assumptions: 6211 * - There is no RX handler. 6212 * - Only one packet_type matches. 6213 * If either of these fails, we will end up doing some per-packet 6214 * processing in-line, then handling the 'last ptype' for the whole 6215 * sublist. This can't cause out-of-order delivery to any single ptype, 6216 * because the 'last ptype' must be constant across the sublist, and all 6217 * other ptypes are handled per-packet. 6218 */ 6219 /* Current (common) ptype of sublist */ 6220 struct packet_type *pt_curr = NULL; 6221 /* Current (common) orig_dev of sublist */ 6222 struct net_device *od_curr = NULL; 6223 struct sk_buff *skb, *next; 6224 LIST_HEAD(sublist); 6225 6226 list_for_each_entry_safe(skb, next, head, list) { 6227 struct net_device *orig_dev = skb->dev; 6228 struct packet_type *pt_prev = NULL; 6229 6230 skb_list_del_init(skb); 6231 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 6232 if (!pt_prev) 6233 continue; 6234 if (pt_curr != pt_prev || od_curr != orig_dev) { 6235 /* dispatch old sublist */ 6236 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 6237 /* start new sublist */ 6238 INIT_LIST_HEAD(&sublist); 6239 pt_curr = pt_prev; 6240 od_curr = orig_dev; 6241 } 6242 list_add_tail(&skb->list, &sublist); 6243 } 6244 6245 /* dispatch final sublist */ 6246 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 6247 } 6248 6249 static int __netif_receive_skb(struct sk_buff *skb) 6250 { 6251 int ret; 6252 6253 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) { 6254 unsigned int noreclaim_flag; 6255 6256 /* 6257 * PFMEMALLOC skbs are special, they should 6258 * - be delivered to SOCK_MEMALLOC sockets only 6259 * - stay away from userspace 6260 * - have bounded memory usage 6261 * 6262 * Use PF_MEMALLOC as this saves us from propagating the allocation 6263 * context down to all allocation sites. 6264 */ 6265 noreclaim_flag = memalloc_noreclaim_save(); 6266 ret = __netif_receive_skb_one_core(skb, true); 6267 memalloc_noreclaim_restore(noreclaim_flag); 6268 } else 6269 ret = __netif_receive_skb_one_core(skb, false); 6270 6271 return ret; 6272 } 6273 6274 static void __netif_receive_skb_list(struct list_head *head) 6275 { 6276 unsigned long noreclaim_flag = 0; 6277 struct sk_buff *skb, *next; 6278 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */ 6279 6280 list_for_each_entry_safe(skb, next, head, list) { 6281 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) { 6282 struct list_head sublist; 6283 6284 /* Handle the previous sublist */ 6285 list_cut_before(&sublist, head, &skb->list); 6286 if (!list_empty(&sublist)) 6287 __netif_receive_skb_list_core(&sublist, pfmemalloc); 6288 pfmemalloc = !pfmemalloc; 6289 /* See comments in __netif_receive_skb */ 6290 if (pfmemalloc) 6291 noreclaim_flag = memalloc_noreclaim_save(); 6292 else 6293 memalloc_noreclaim_restore(noreclaim_flag); 6294 } 6295 } 6296 /* Handle the remaining sublist */ 6297 if (!list_empty(head)) 6298 __netif_receive_skb_list_core(head, pfmemalloc); 6299 /* Restore pflags */ 6300 if (pfmemalloc) 6301 memalloc_noreclaim_restore(noreclaim_flag); 6302 } 6303 6304 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp) 6305 { 6306 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog); 6307 struct bpf_prog *new = xdp->prog; 6308 int ret = 0; 6309 6310 switch (xdp->command) { 6311 case XDP_SETUP_PROG: 6312 rcu_assign_pointer(dev->xdp_prog, new); 6313 if (old) 6314 bpf_prog_put(old); 6315 6316 if (old && !new) { 6317 static_branch_dec(&generic_xdp_needed_key); 6318 } else if (new && !old) { 6319 static_branch_inc(&generic_xdp_needed_key); 6320 netif_disable_lro(dev); 6321 dev_disable_gro_hw(dev); 6322 } 6323 break; 6324 6325 default: 6326 ret = -EINVAL; 6327 break; 6328 } 6329 6330 return ret; 6331 } 6332 6333 static int netif_receive_skb_internal(struct sk_buff *skb) 6334 { 6335 int ret; 6336 6337 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 6338 6339 if (skb_defer_rx_timestamp(skb)) 6340 return NET_RX_SUCCESS; 6341 6342 rcu_read_lock(); 6343 #ifdef CONFIG_RPS 6344 if (static_branch_unlikely(&rps_needed)) { 6345 struct rps_dev_flow voidflow, *rflow = &voidflow; 6346 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6347 6348 if (cpu >= 0) { 6349 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6350 rcu_read_unlock(); 6351 return ret; 6352 } 6353 } 6354 #endif 6355 ret = __netif_receive_skb(skb); 6356 rcu_read_unlock(); 6357 return ret; 6358 } 6359 6360 void netif_receive_skb_list_internal(struct list_head *head) 6361 { 6362 struct sk_buff *skb, *next; 6363 LIST_HEAD(sublist); 6364 6365 list_for_each_entry_safe(skb, next, head, list) { 6366 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), 6367 skb); 6368 skb_list_del_init(skb); 6369 if (!skb_defer_rx_timestamp(skb)) 6370 list_add_tail(&skb->list, &sublist); 6371 } 6372 list_splice_init(&sublist, head); 6373 6374 rcu_read_lock(); 6375 #ifdef CONFIG_RPS 6376 if (static_branch_unlikely(&rps_needed)) { 6377 list_for_each_entry_safe(skb, next, head, list) { 6378 struct rps_dev_flow voidflow, *rflow = &voidflow; 6379 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6380 6381 if (cpu >= 0) { 6382 /* Will be handled, remove from list */ 6383 skb_list_del_init(skb); 6384 enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6385 } 6386 } 6387 } 6388 #endif 6389 __netif_receive_skb_list(head); 6390 rcu_read_unlock(); 6391 } 6392 6393 /** 6394 * netif_receive_skb - process receive buffer from network 6395 * @skb: buffer to process 6396 * 6397 * netif_receive_skb() is the main receive data processing function. 6398 * It always succeeds. The buffer may be dropped during processing 6399 * for congestion control or by the protocol layers. 6400 * 6401 * This function may only be called from softirq context and interrupts 6402 * should be enabled. 6403 * 6404 * Return values (usually ignored): 6405 * NET_RX_SUCCESS: no congestion 6406 * NET_RX_DROP: packet was dropped 6407 */ 6408 int netif_receive_skb(struct sk_buff *skb) 6409 { 6410 int ret; 6411 6412 trace_netif_receive_skb_entry(skb); 6413 6414 ret = netif_receive_skb_internal(skb); 6415 trace_netif_receive_skb_exit(ret); 6416 6417 return ret; 6418 } 6419 EXPORT_SYMBOL(netif_receive_skb); 6420 6421 /** 6422 * netif_receive_skb_list - process many receive buffers from network 6423 * @head: list of skbs to process. 6424 * 6425 * Since return value of netif_receive_skb() is normally ignored, and 6426 * wouldn't be meaningful for a list, this function returns void. 6427 * 6428 * This function may only be called from softirq context and interrupts 6429 * should be enabled. 6430 */ 6431 void netif_receive_skb_list(struct list_head *head) 6432 { 6433 struct sk_buff *skb; 6434 6435 if (list_empty(head)) 6436 return; 6437 if (trace_netif_receive_skb_list_entry_enabled()) { 6438 list_for_each_entry(skb, head, list) 6439 trace_netif_receive_skb_list_entry(skb); 6440 } 6441 netif_receive_skb_list_internal(head); 6442 trace_netif_receive_skb_list_exit(0); 6443 } 6444 EXPORT_SYMBOL(netif_receive_skb_list); 6445 6446 /* Network device is going away, flush any packets still pending */ 6447 static void flush_backlog(struct work_struct *work) 6448 { 6449 struct sk_buff *skb, *tmp; 6450 struct sk_buff_head list; 6451 struct softnet_data *sd; 6452 6453 __skb_queue_head_init(&list); 6454 local_bh_disable(); 6455 sd = this_cpu_ptr(&softnet_data); 6456 6457 backlog_lock_irq_disable(sd); 6458 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) { 6459 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6460 __skb_unlink(skb, &sd->input_pkt_queue); 6461 __skb_queue_tail(&list, skb); 6462 rps_input_queue_head_incr(sd); 6463 } 6464 } 6465 backlog_unlock_irq_enable(sd); 6466 6467 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6468 skb_queue_walk_safe(&sd->process_queue, skb, tmp) { 6469 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6470 __skb_unlink(skb, &sd->process_queue); 6471 __skb_queue_tail(&list, skb); 6472 rps_input_queue_head_incr(sd); 6473 } 6474 } 6475 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6476 local_bh_enable(); 6477 6478 __skb_queue_purge_reason(&list, SKB_DROP_REASON_DEV_READY); 6479 } 6480 6481 static bool flush_required(int cpu) 6482 { 6483 #if IS_ENABLED(CONFIG_RPS) 6484 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 6485 bool do_flush; 6486 6487 backlog_lock_irq_disable(sd); 6488 6489 /* as insertion into process_queue happens with the rps lock held, 6490 * process_queue access may race only with dequeue 6491 */ 6492 do_flush = !skb_queue_empty(&sd->input_pkt_queue) || 6493 !skb_queue_empty_lockless(&sd->process_queue); 6494 backlog_unlock_irq_enable(sd); 6495 6496 return do_flush; 6497 #endif 6498 /* without RPS we can't safely check input_pkt_queue: during a 6499 * concurrent remote skb_queue_splice() we can detect as empty both 6500 * input_pkt_queue and process_queue even if the latter could end-up 6501 * containing a lot of packets. 6502 */ 6503 return true; 6504 } 6505 6506 struct flush_backlogs { 6507 cpumask_t flush_cpus; 6508 struct work_struct w[]; 6509 }; 6510 6511 static struct flush_backlogs *flush_backlogs_alloc(void) 6512 { 6513 return kmalloc_flex(struct flush_backlogs, w, nr_cpu_ids, GFP_KERNEL); 6514 } 6515 6516 static struct flush_backlogs *flush_backlogs_fallback; 6517 static DEFINE_MUTEX(flush_backlogs_mutex); 6518 6519 static void flush_all_backlogs(void) 6520 { 6521 struct flush_backlogs *ptr = flush_backlogs_alloc(); 6522 unsigned int cpu; 6523 6524 if (!ptr) { 6525 mutex_lock(&flush_backlogs_mutex); 6526 ptr = flush_backlogs_fallback; 6527 } 6528 cpumask_clear(&ptr->flush_cpus); 6529 6530 cpus_read_lock(); 6531 6532 for_each_online_cpu(cpu) { 6533 if (flush_required(cpu)) { 6534 INIT_WORK(&ptr->w[cpu], flush_backlog); 6535 queue_work_on(cpu, system_highpri_wq, &ptr->w[cpu]); 6536 __cpumask_set_cpu(cpu, &ptr->flush_cpus); 6537 } 6538 } 6539 6540 /* we can have in flight packet[s] on the cpus we are not flushing, 6541 * synchronize_net() in unregister_netdevice_many() will take care of 6542 * them. 6543 */ 6544 for_each_cpu(cpu, &ptr->flush_cpus) 6545 flush_work(&ptr->w[cpu]); 6546 6547 cpus_read_unlock(); 6548 6549 if (ptr != flush_backlogs_fallback) 6550 kfree(ptr); 6551 else 6552 mutex_unlock(&flush_backlogs_mutex); 6553 } 6554 6555 static void net_rps_send_ipi(struct softnet_data *remsd) 6556 { 6557 #ifdef CONFIG_RPS 6558 while (remsd) { 6559 struct softnet_data *next = remsd->rps_ipi_next; 6560 6561 if (cpu_online(remsd->cpu)) 6562 smp_call_function_single_async(remsd->cpu, &remsd->csd); 6563 remsd = next; 6564 } 6565 #endif 6566 } 6567 6568 /* 6569 * net_rps_action_and_irq_enable sends any pending IPI's for rps. 6570 * Note: called with local irq disabled, but exits with local irq enabled. 6571 */ 6572 static void net_rps_action_and_irq_enable(struct softnet_data *sd) 6573 { 6574 #ifdef CONFIG_RPS 6575 struct softnet_data *remsd = sd->rps_ipi_list; 6576 6577 if (!use_backlog_threads() && remsd) { 6578 sd->rps_ipi_list = NULL; 6579 6580 local_irq_enable(); 6581 6582 /* Send pending IPI's to kick RPS processing on remote cpus. */ 6583 net_rps_send_ipi(remsd); 6584 } else 6585 #endif 6586 local_irq_enable(); 6587 } 6588 6589 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd) 6590 { 6591 #ifdef CONFIG_RPS 6592 return !use_backlog_threads() && sd->rps_ipi_list; 6593 #else 6594 return false; 6595 #endif 6596 } 6597 6598 static int process_backlog(struct napi_struct *napi, int quota) 6599 { 6600 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog); 6601 bool again = true; 6602 int work = 0; 6603 6604 /* Check if we have pending ipi, its better to send them now, 6605 * not waiting net_rx_action() end. 6606 */ 6607 if (sd_has_rps_ipi_waiting(sd)) { 6608 local_irq_disable(); 6609 net_rps_action_and_irq_enable(sd); 6610 } 6611 6612 napi->weight = READ_ONCE(net_hotdata.dev_rx_weight); 6613 while (again) { 6614 struct sk_buff *skb; 6615 6616 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6617 while ((skb = __skb_dequeue(&sd->process_queue))) { 6618 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6619 rcu_read_lock(); 6620 __netif_receive_skb(skb); 6621 rcu_read_unlock(); 6622 if (++work >= quota) { 6623 rps_input_queue_head_add(sd, work); 6624 return work; 6625 } 6626 6627 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6628 } 6629 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6630 6631 backlog_lock_irq_disable(sd); 6632 if (skb_queue_empty(&sd->input_pkt_queue)) { 6633 /* 6634 * Inline a custom version of __napi_complete(). 6635 * only current cpu owns and manipulates this napi, 6636 * and NAPI_STATE_SCHED is the only possible flag set 6637 * on backlog. 6638 * We can use a plain write instead of clear_bit(), 6639 * and we dont need an smp_mb() memory barrier. 6640 */ 6641 napi->state &= NAPIF_STATE_THREADED; 6642 again = false; 6643 } else { 6644 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6645 skb_queue_splice_tail_init(&sd->input_pkt_queue, 6646 &sd->process_queue); 6647 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6648 } 6649 backlog_unlock_irq_enable(sd); 6650 } 6651 6652 if (work) 6653 rps_input_queue_head_add(sd, work); 6654 return work; 6655 } 6656 6657 /** 6658 * __napi_schedule - schedule for receive 6659 * @n: entry to schedule 6660 * 6661 * The entry's receive function will be scheduled to run. 6662 * Consider using __napi_schedule_irqoff() if hard irqs are masked. 6663 */ 6664 void __napi_schedule(struct napi_struct *n) 6665 { 6666 unsigned long flags; 6667 6668 local_irq_save(flags); 6669 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6670 local_irq_restore(flags); 6671 } 6672 EXPORT_SYMBOL(__napi_schedule); 6673 6674 /** 6675 * napi_schedule_prep - check if napi can be scheduled 6676 * @n: napi context 6677 * 6678 * Test if NAPI routine is already running, and if not mark 6679 * it as running. This is used as a condition variable to 6680 * insure only one NAPI poll instance runs. We also make 6681 * sure there is no pending NAPI disable. 6682 */ 6683 bool napi_schedule_prep(struct napi_struct *n) 6684 { 6685 unsigned long new, val = READ_ONCE(n->state); 6686 6687 do { 6688 if (unlikely(val & NAPIF_STATE_DISABLE)) 6689 return false; 6690 new = val | NAPIF_STATE_SCHED; 6691 6692 /* Sets STATE_MISSED bit if STATE_SCHED was already set 6693 * This was suggested by Alexander Duyck, as compiler 6694 * emits better code than : 6695 * if (val & NAPIF_STATE_SCHED) 6696 * new |= NAPIF_STATE_MISSED; 6697 */ 6698 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED * 6699 NAPIF_STATE_MISSED; 6700 } while (!try_cmpxchg(&n->state, &val, new)); 6701 6702 return !(val & NAPIF_STATE_SCHED); 6703 } 6704 EXPORT_SYMBOL(napi_schedule_prep); 6705 6706 /** 6707 * __napi_schedule_irqoff - schedule for receive 6708 * @n: entry to schedule 6709 * 6710 * Variant of __napi_schedule() assuming hard irqs are masked. 6711 * 6712 * On PREEMPT_RT enabled kernels this maps to __napi_schedule() 6713 * because the interrupt disabled assumption might not be true 6714 * due to force-threaded interrupts and spinlock substitution. 6715 */ 6716 void __napi_schedule_irqoff(struct napi_struct *n) 6717 { 6718 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6719 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6720 else 6721 __napi_schedule(n); 6722 } 6723 EXPORT_SYMBOL(__napi_schedule_irqoff); 6724 6725 bool napi_complete_done(struct napi_struct *n, int work_done) 6726 { 6727 unsigned long flags, val, new, timeout = 0; 6728 bool ret = true; 6729 6730 /* 6731 * 1) Don't let napi dequeue from the cpu poll list 6732 * just in case its running on a different cpu. 6733 * 2) If we are busy polling, do nothing here, we have 6734 * the guarantee we will be called later. 6735 */ 6736 if (unlikely(n->state & (NAPIF_STATE_NPSVC | 6737 NAPIF_STATE_IN_BUSY_POLL))) 6738 return false; 6739 6740 if (work_done) { 6741 if (n->gro.bitmask) 6742 timeout = napi_get_gro_flush_timeout(n); 6743 n->defer_hard_irqs_count = napi_get_defer_hard_irqs(n); 6744 } 6745 if (n->defer_hard_irqs_count > 0) { 6746 n->defer_hard_irqs_count--; 6747 timeout = napi_get_gro_flush_timeout(n); 6748 if (timeout) 6749 ret = false; 6750 } 6751 6752 /* 6753 * When the NAPI instance uses a timeout and keeps postponing 6754 * it, we need to bound somehow the time packets are kept in 6755 * the GRO layer. 6756 */ 6757 gro_flush_normal(&n->gro, !!timeout); 6758 6759 if (unlikely(!list_empty(&n->poll_list))) { 6760 /* If n->poll_list is not empty, we need to mask irqs */ 6761 local_irq_save(flags); 6762 list_del_init(&n->poll_list); 6763 local_irq_restore(flags); 6764 } 6765 WRITE_ONCE(n->list_owner, -1); 6766 6767 val = READ_ONCE(n->state); 6768 do { 6769 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED)); 6770 6771 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED | 6772 NAPIF_STATE_SCHED_THREADED | 6773 NAPIF_STATE_PREFER_BUSY_POLL); 6774 6775 /* If STATE_MISSED was set, leave STATE_SCHED set, 6776 * because we will call napi->poll() one more time. 6777 * This C code was suggested by Alexander Duyck to help gcc. 6778 */ 6779 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED * 6780 NAPIF_STATE_SCHED; 6781 } while (!try_cmpxchg(&n->state, &val, new)); 6782 6783 if (unlikely(val & NAPIF_STATE_MISSED)) { 6784 __napi_schedule(n); 6785 return false; 6786 } 6787 6788 if (timeout) 6789 hrtimer_start(&n->timer, ns_to_ktime(timeout), 6790 HRTIMER_MODE_REL_PINNED); 6791 return ret; 6792 } 6793 EXPORT_SYMBOL(napi_complete_done); 6794 6795 static void skb_defer_free_flush(void) 6796 { 6797 struct llist_node *free_list; 6798 struct sk_buff *skb, *next; 6799 struct skb_defer_node *sdn; 6800 int node; 6801 6802 for_each_node(node) { 6803 sdn = this_cpu_ptr(net_hotdata.skb_defer_nodes) + node; 6804 6805 if (llist_empty(&sdn->defer_list)) 6806 continue; 6807 atomic_long_set(&sdn->defer_count, 0); 6808 free_list = llist_del_all(&sdn->defer_list); 6809 6810 llist_for_each_entry_safe(skb, next, free_list, ll_node) { 6811 prefetch(next); 6812 napi_consume_skb(skb, 1); 6813 } 6814 } 6815 } 6816 6817 #if defined(CONFIG_NET_RX_BUSY_POLL) 6818 6819 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule) 6820 { 6821 if (!skip_schedule) { 6822 gro_normal_list(&napi->gro); 6823 __napi_schedule(napi); 6824 return; 6825 } 6826 6827 /* Flush too old packets. If HZ < 1000, flush all packets */ 6828 gro_flush_normal(&napi->gro, HZ >= 1000); 6829 6830 clear_bit(NAPI_STATE_SCHED, &napi->state); 6831 } 6832 6833 enum { 6834 NAPI_F_PREFER_BUSY_POLL = 1, 6835 NAPI_F_END_ON_RESCHED = 2, 6836 }; 6837 6838 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, 6839 unsigned flags, u16 budget) 6840 { 6841 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6842 bool skip_schedule = false; 6843 unsigned long timeout; 6844 int rc; 6845 6846 /* Busy polling means there is a high chance device driver hard irq 6847 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was 6848 * set in napi_schedule_prep(). 6849 * Since we are about to call napi->poll() once more, we can safely 6850 * clear NAPI_STATE_MISSED. 6851 * 6852 * Note: x86 could use a single "lock and ..." instruction 6853 * to perform these two clear_bit() 6854 */ 6855 clear_bit(NAPI_STATE_MISSED, &napi->state); 6856 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state); 6857 6858 local_bh_disable(); 6859 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6860 6861 if (flags & NAPI_F_PREFER_BUSY_POLL) { 6862 napi->defer_hard_irqs_count = napi_get_defer_hard_irqs(napi); 6863 timeout = napi_get_gro_flush_timeout(napi); 6864 if (napi->defer_hard_irqs_count && timeout) { 6865 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED); 6866 skip_schedule = true; 6867 } 6868 } 6869 6870 /* All we really want here is to re-enable device interrupts. 6871 * Ideally, a new ndo_busy_poll_stop() could avoid another round. 6872 */ 6873 rc = napi->poll(napi, budget); 6874 /* We can't gro_normal_list() here, because napi->poll() might have 6875 * rearmed the napi (napi_complete_done()) in which case it could 6876 * already be running on another CPU. 6877 */ 6878 trace_napi_poll(napi, rc, budget); 6879 netpoll_poll_unlock(have_poll_lock); 6880 if (rc == budget) 6881 __busy_poll_stop(napi, skip_schedule); 6882 bpf_net_ctx_clear(bpf_net_ctx); 6883 local_bh_enable(); 6884 } 6885 6886 static void __napi_busy_loop(unsigned int napi_id, 6887 bool (*loop_end)(void *, unsigned long), 6888 void *loop_end_arg, unsigned flags, u16 budget) 6889 { 6890 unsigned long start_time = loop_end ? busy_loop_current_time() : 0; 6891 int (*napi_poll)(struct napi_struct *napi, int budget); 6892 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6893 void *have_poll_lock = NULL; 6894 struct napi_struct *napi; 6895 6896 WARN_ON_ONCE(!rcu_read_lock_held()); 6897 6898 restart: 6899 napi_poll = NULL; 6900 6901 napi = napi_by_id(napi_id); 6902 if (!napi) 6903 return; 6904 6905 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6906 preempt_disable(); 6907 for (;;) { 6908 int work = 0; 6909 6910 local_bh_disable(); 6911 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6912 if (!napi_poll) { 6913 unsigned long val = READ_ONCE(napi->state); 6914 6915 /* If multiple threads are competing for this napi, 6916 * we avoid dirtying napi->state as much as we can. 6917 */ 6918 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED | 6919 NAPIF_STATE_IN_BUSY_POLL)) { 6920 if (flags & NAPI_F_PREFER_BUSY_POLL) 6921 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6922 goto count; 6923 } 6924 if (cmpxchg(&napi->state, val, 6925 val | NAPIF_STATE_IN_BUSY_POLL | 6926 NAPIF_STATE_SCHED) != val) { 6927 if (flags & NAPI_F_PREFER_BUSY_POLL) 6928 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6929 goto count; 6930 } 6931 have_poll_lock = netpoll_poll_lock(napi); 6932 napi_poll = napi->poll; 6933 } 6934 work = napi_poll(napi, budget); 6935 trace_napi_poll(napi, work, budget); 6936 gro_normal_list(&napi->gro); 6937 count: 6938 if (work > 0) 6939 __NET_ADD_STATS(dev_net(napi->dev), 6940 LINUX_MIB_BUSYPOLLRXPACKETS, work); 6941 skb_defer_free_flush(); 6942 bpf_net_ctx_clear(bpf_net_ctx); 6943 local_bh_enable(); 6944 6945 if (!loop_end || loop_end(loop_end_arg, start_time)) 6946 break; 6947 6948 if (unlikely(need_resched())) { 6949 if (flags & NAPI_F_END_ON_RESCHED) 6950 break; 6951 if (napi_poll) 6952 busy_poll_stop(napi, have_poll_lock, flags, budget); 6953 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6954 preempt_enable(); 6955 rcu_read_unlock(); 6956 cond_resched(); 6957 rcu_read_lock(); 6958 if (loop_end(loop_end_arg, start_time)) 6959 return; 6960 goto restart; 6961 } 6962 cpu_relax(); 6963 } 6964 if (napi_poll) 6965 busy_poll_stop(napi, have_poll_lock, flags, budget); 6966 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6967 preempt_enable(); 6968 } 6969 6970 void napi_busy_loop_rcu(unsigned int napi_id, 6971 bool (*loop_end)(void *, unsigned long), 6972 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6973 { 6974 unsigned flags = NAPI_F_END_ON_RESCHED; 6975 6976 if (prefer_busy_poll) 6977 flags |= NAPI_F_PREFER_BUSY_POLL; 6978 6979 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6980 } 6981 6982 void napi_busy_loop(unsigned int napi_id, 6983 bool (*loop_end)(void *, unsigned long), 6984 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6985 { 6986 unsigned flags = prefer_busy_poll ? NAPI_F_PREFER_BUSY_POLL : 0; 6987 6988 rcu_read_lock(); 6989 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6990 rcu_read_unlock(); 6991 } 6992 EXPORT_SYMBOL(napi_busy_loop); 6993 6994 void napi_suspend_irqs(unsigned int napi_id) 6995 { 6996 struct napi_struct *napi; 6997 6998 rcu_read_lock(); 6999 napi = napi_by_id(napi_id); 7000 if (napi) { 7001 unsigned long timeout = napi_get_irq_suspend_timeout(napi); 7002 7003 if (timeout) 7004 hrtimer_start(&napi->timer, ns_to_ktime(timeout), 7005 HRTIMER_MODE_REL_PINNED); 7006 } 7007 rcu_read_unlock(); 7008 } 7009 7010 void napi_resume_irqs(unsigned int napi_id) 7011 { 7012 struct napi_struct *napi; 7013 7014 rcu_read_lock(); 7015 napi = napi_by_id(napi_id); 7016 if (napi) { 7017 /* If irq_suspend_timeout is set to 0 between the call to 7018 * napi_suspend_irqs and now, the original value still 7019 * determines the safety timeout as intended and napi_watchdog 7020 * will resume irq processing. 7021 */ 7022 if (napi_get_irq_suspend_timeout(napi)) { 7023 local_bh_disable(); 7024 napi_schedule(napi); 7025 local_bh_enable(); 7026 } 7027 } 7028 rcu_read_unlock(); 7029 } 7030 7031 #endif /* CONFIG_NET_RX_BUSY_POLL */ 7032 7033 static void __napi_hash_add_with_id(struct napi_struct *napi, 7034 unsigned int napi_id) 7035 { 7036 napi->gro.cached_napi_id = napi_id; 7037 7038 WRITE_ONCE(napi->napi_id, napi_id); 7039 hlist_add_head_rcu(&napi->napi_hash_node, 7040 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]); 7041 } 7042 7043 static void napi_hash_add_with_id(struct napi_struct *napi, 7044 unsigned int napi_id) 7045 { 7046 unsigned long flags; 7047 7048 spin_lock_irqsave(&napi_hash_lock, flags); 7049 WARN_ON_ONCE(napi_by_id(napi_id)); 7050 __napi_hash_add_with_id(napi, napi_id); 7051 spin_unlock_irqrestore(&napi_hash_lock, flags); 7052 } 7053 7054 static void napi_hash_add(struct napi_struct *napi) 7055 { 7056 unsigned long flags; 7057 7058 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state)) 7059 return; 7060 7061 spin_lock_irqsave(&napi_hash_lock, flags); 7062 7063 /* 0..NR_CPUS range is reserved for sender_cpu use */ 7064 do { 7065 if (unlikely(!napi_id_valid(++napi_gen_id))) 7066 napi_gen_id = MIN_NAPI_ID; 7067 } while (napi_by_id(napi_gen_id)); 7068 7069 __napi_hash_add_with_id(napi, napi_gen_id); 7070 7071 spin_unlock_irqrestore(&napi_hash_lock, flags); 7072 } 7073 7074 /* Warning : caller is responsible to make sure rcu grace period 7075 * is respected before freeing memory containing @napi 7076 */ 7077 static void napi_hash_del(struct napi_struct *napi) 7078 { 7079 unsigned long flags; 7080 7081 spin_lock_irqsave(&napi_hash_lock, flags); 7082 7083 hlist_del_init_rcu(&napi->napi_hash_node); 7084 7085 spin_unlock_irqrestore(&napi_hash_lock, flags); 7086 } 7087 7088 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer) 7089 { 7090 struct napi_struct *napi; 7091 7092 napi = container_of(timer, struct napi_struct, timer); 7093 7094 /* Note : we use a relaxed variant of napi_schedule_prep() not setting 7095 * NAPI_STATE_MISSED, since we do not react to a device IRQ. 7096 */ 7097 if (!napi_disable_pending(napi) && 7098 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) { 7099 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 7100 __napi_schedule_irqoff(napi); 7101 } 7102 7103 return HRTIMER_NORESTART; 7104 } 7105 7106 static void napi_stop_kthread(struct napi_struct *napi) 7107 { 7108 unsigned long val, new; 7109 7110 /* Wait until the napi STATE_THREADED is unset. */ 7111 while (true) { 7112 val = READ_ONCE(napi->state); 7113 7114 /* If napi kthread own this napi or the napi is idle, 7115 * STATE_THREADED can be unset here. 7116 */ 7117 if ((val & NAPIF_STATE_SCHED_THREADED) || 7118 !(val & NAPIF_STATE_SCHED)) { 7119 new = val & (~(NAPIF_STATE_THREADED | 7120 NAPIF_STATE_THREADED_BUSY_POLL)); 7121 } else { 7122 msleep(20); 7123 continue; 7124 } 7125 7126 if (try_cmpxchg(&napi->state, &val, new)) 7127 break; 7128 } 7129 7130 /* Once STATE_THREADED is unset, wait for SCHED_THREADED to be unset by 7131 * the kthread. 7132 */ 7133 while (true) { 7134 if (!test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) 7135 break; 7136 7137 msleep(20); 7138 } 7139 7140 kthread_stop(napi->thread); 7141 napi->thread = NULL; 7142 } 7143 7144 static void napi_set_threaded_state(struct napi_struct *napi, 7145 enum netdev_napi_threaded threaded_mode) 7146 { 7147 bool threaded = threaded_mode != NETDEV_NAPI_THREADED_DISABLED; 7148 bool busy_poll = threaded_mode == NETDEV_NAPI_THREADED_BUSY_POLL; 7149 7150 assign_bit(NAPI_STATE_THREADED, &napi->state, threaded); 7151 assign_bit(NAPI_STATE_THREADED_BUSY_POLL, &napi->state, busy_poll); 7152 } 7153 7154 int napi_set_threaded(struct napi_struct *napi, 7155 enum netdev_napi_threaded threaded) 7156 { 7157 if (threaded) { 7158 if (!napi->thread) { 7159 int err = napi_kthread_create(napi); 7160 7161 if (err) 7162 return err; 7163 } 7164 } 7165 7166 if (napi->config) 7167 napi->config->threaded = threaded; 7168 7169 /* Setting/unsetting threaded mode on a napi might not immediately 7170 * take effect, if the current napi instance is actively being 7171 * polled. In this case, the switch between threaded mode and 7172 * softirq mode will happen in the next round of napi_schedule(). 7173 * This should not cause hiccups/stalls to the live traffic. 7174 */ 7175 if (!threaded && napi->thread) { 7176 napi_stop_kthread(napi); 7177 } else { 7178 /* Make sure kthread is created before THREADED bit is set. */ 7179 smp_mb__before_atomic(); 7180 napi_set_threaded_state(napi, threaded); 7181 } 7182 7183 return 0; 7184 } 7185 7186 int netif_set_threaded(struct net_device *dev, 7187 enum netdev_napi_threaded threaded) 7188 { 7189 struct napi_struct *napi; 7190 int i, err = 0; 7191 7192 netdev_assert_locked_or_invisible(dev); 7193 7194 if (threaded) { 7195 list_for_each_entry(napi, &dev->napi_list, dev_list) { 7196 if (!napi->thread) { 7197 err = napi_kthread_create(napi); 7198 if (err) { 7199 threaded = NETDEV_NAPI_THREADED_DISABLED; 7200 break; 7201 } 7202 } 7203 } 7204 } 7205 7206 WRITE_ONCE(dev->threaded, threaded); 7207 7208 /* The error should not occur as the kthreads are already created. */ 7209 list_for_each_entry(napi, &dev->napi_list, dev_list) 7210 WARN_ON_ONCE(napi_set_threaded(napi, threaded)); 7211 7212 /* Override the config for all NAPIs even if currently not listed */ 7213 for (i = 0; i < dev->num_napi_configs; i++) 7214 dev->napi_config[i].threaded = threaded; 7215 7216 return err; 7217 } 7218 7219 /** 7220 * netif_threaded_enable() - enable threaded NAPIs 7221 * @dev: net_device instance 7222 * 7223 * Enable threaded mode for the NAPI instances of the device. This may be useful 7224 * for devices where multiple NAPI instances get scheduled by a single 7225 * interrupt. Threaded NAPI allows moving the NAPI processing to cores other 7226 * than the core where IRQ is mapped. 7227 * 7228 * This function should be called before @dev is registered. 7229 */ 7230 void netif_threaded_enable(struct net_device *dev) 7231 { 7232 WARN_ON_ONCE(netif_set_threaded(dev, NETDEV_NAPI_THREADED_ENABLED)); 7233 } 7234 EXPORT_SYMBOL(netif_threaded_enable); 7235 7236 /** 7237 * netif_queue_set_napi - Associate queue with the napi 7238 * @dev: device to which NAPI and queue belong 7239 * @queue_index: Index of queue 7240 * @type: queue type as RX or TX 7241 * @napi: NAPI context, pass NULL to clear previously set NAPI 7242 * 7243 * Set queue with its corresponding napi context. This should be done after 7244 * registering the NAPI handler for the queue-vector and the queues have been 7245 * mapped to the corresponding interrupt vector. 7246 */ 7247 void netif_queue_set_napi(struct net_device *dev, unsigned int queue_index, 7248 enum netdev_queue_type type, struct napi_struct *napi) 7249 { 7250 struct netdev_rx_queue *rxq; 7251 struct netdev_queue *txq; 7252 7253 if (WARN_ON_ONCE(napi && !napi->dev)) 7254 return; 7255 netdev_ops_assert_locked_or_invisible(dev); 7256 7257 switch (type) { 7258 case NETDEV_QUEUE_TYPE_RX: 7259 rxq = __netif_get_rx_queue(dev, queue_index); 7260 rxq->napi = napi; 7261 return; 7262 case NETDEV_QUEUE_TYPE_TX: 7263 txq = netdev_get_tx_queue(dev, queue_index); 7264 txq->napi = napi; 7265 return; 7266 default: 7267 return; 7268 } 7269 } 7270 EXPORT_SYMBOL(netif_queue_set_napi); 7271 7272 static void 7273 netif_napi_irq_notify(struct irq_affinity_notify *notify, 7274 const cpumask_t *mask) 7275 { 7276 struct napi_struct *napi = 7277 container_of(notify, struct napi_struct, notify); 7278 #ifdef CONFIG_RFS_ACCEL 7279 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 7280 int err; 7281 #endif 7282 7283 if (napi->config && napi->dev->irq_affinity_auto) 7284 cpumask_copy(&napi->config->affinity_mask, mask); 7285 7286 #ifdef CONFIG_RFS_ACCEL 7287 if (napi->dev->rx_cpu_rmap_auto) { 7288 err = cpu_rmap_update(rmap, napi->napi_rmap_idx, mask); 7289 if (err) 7290 netdev_warn(napi->dev, "RMAP update failed (%d)\n", 7291 err); 7292 } 7293 #endif 7294 } 7295 7296 #ifdef CONFIG_RFS_ACCEL 7297 static void netif_napi_affinity_release(struct kref *ref) 7298 { 7299 struct napi_struct *napi = 7300 container_of(ref, struct napi_struct, notify.kref); 7301 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 7302 7303 netdev_assert_locked(napi->dev); 7304 WARN_ON(test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, 7305 &napi->state)); 7306 7307 if (!napi->dev->rx_cpu_rmap_auto) 7308 return; 7309 rmap->obj[napi->napi_rmap_idx] = NULL; 7310 napi->napi_rmap_idx = -1; 7311 cpu_rmap_put(rmap); 7312 } 7313 7314 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 7315 { 7316 if (dev->rx_cpu_rmap_auto) 7317 return 0; 7318 7319 dev->rx_cpu_rmap = alloc_irq_cpu_rmap(num_irqs); 7320 if (!dev->rx_cpu_rmap) 7321 return -ENOMEM; 7322 7323 dev->rx_cpu_rmap_auto = true; 7324 return 0; 7325 } 7326 EXPORT_SYMBOL(netif_enable_cpu_rmap); 7327 7328 static void netif_del_cpu_rmap(struct net_device *dev) 7329 { 7330 struct cpu_rmap *rmap = dev->rx_cpu_rmap; 7331 7332 if (!dev->rx_cpu_rmap_auto) 7333 return; 7334 7335 /* Free the rmap */ 7336 cpu_rmap_put(rmap); 7337 dev->rx_cpu_rmap = NULL; 7338 dev->rx_cpu_rmap_auto = false; 7339 } 7340 7341 #else 7342 static void netif_napi_affinity_release(struct kref *ref) 7343 { 7344 } 7345 7346 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 7347 { 7348 return 0; 7349 } 7350 EXPORT_SYMBOL(netif_enable_cpu_rmap); 7351 7352 static void netif_del_cpu_rmap(struct net_device *dev) 7353 { 7354 } 7355 #endif 7356 7357 void netif_set_affinity_auto(struct net_device *dev) 7358 { 7359 unsigned int i, maxqs, numa; 7360 7361 maxqs = max(dev->num_tx_queues, dev->num_rx_queues); 7362 numa = dev_to_node(&dev->dev); 7363 7364 for (i = 0; i < maxqs; i++) 7365 cpumask_set_cpu(cpumask_local_spread(i, numa), 7366 &dev->napi_config[i].affinity_mask); 7367 7368 dev->irq_affinity_auto = true; 7369 } 7370 EXPORT_SYMBOL(netif_set_affinity_auto); 7371 7372 void netif_napi_set_irq_locked(struct napi_struct *napi, int irq) 7373 { 7374 int rc; 7375 7376 netdev_assert_locked_or_invisible(napi->dev); 7377 7378 if (napi->irq == irq) 7379 return; 7380 7381 /* Remove existing resources */ 7382 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7383 irq_set_affinity_notifier(napi->irq, NULL); 7384 7385 napi->irq = irq; 7386 if (irq < 0 || 7387 (!napi->dev->rx_cpu_rmap_auto && !napi->dev->irq_affinity_auto)) 7388 return; 7389 7390 /* Abort for buggy drivers */ 7391 if (napi->dev->irq_affinity_auto && WARN_ON_ONCE(!napi->config)) 7392 return; 7393 7394 #ifdef CONFIG_RFS_ACCEL 7395 if (napi->dev->rx_cpu_rmap_auto) { 7396 rc = cpu_rmap_add(napi->dev->rx_cpu_rmap, napi); 7397 if (rc < 0) 7398 return; 7399 7400 cpu_rmap_get(napi->dev->rx_cpu_rmap); 7401 napi->napi_rmap_idx = rc; 7402 } 7403 #endif 7404 7405 /* Use core IRQ notifier */ 7406 napi->notify.notify = netif_napi_irq_notify; 7407 napi->notify.release = netif_napi_affinity_release; 7408 rc = irq_set_affinity_notifier(irq, &napi->notify); 7409 if (rc) { 7410 netdev_warn(napi->dev, "Unable to set IRQ notifier (%d)\n", 7411 rc); 7412 goto put_rmap; 7413 } 7414 7415 set_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state); 7416 return; 7417 7418 put_rmap: 7419 #ifdef CONFIG_RFS_ACCEL 7420 if (napi->dev->rx_cpu_rmap_auto) { 7421 napi->dev->rx_cpu_rmap->obj[napi->napi_rmap_idx] = NULL; 7422 cpu_rmap_put(napi->dev->rx_cpu_rmap); 7423 napi->napi_rmap_idx = -1; 7424 } 7425 #endif 7426 napi->notify.notify = NULL; 7427 napi->notify.release = NULL; 7428 } 7429 EXPORT_SYMBOL(netif_napi_set_irq_locked); 7430 7431 static void napi_restore_config(struct napi_struct *n) 7432 { 7433 n->defer_hard_irqs = n->config->defer_hard_irqs; 7434 n->gro_flush_timeout = n->config->gro_flush_timeout; 7435 n->irq_suspend_timeout = n->config->irq_suspend_timeout; 7436 7437 if (n->dev->irq_affinity_auto && 7438 test_bit(NAPI_STATE_HAS_NOTIFIER, &n->state)) 7439 irq_set_affinity(n->irq, &n->config->affinity_mask); 7440 7441 /* a NAPI ID might be stored in the config, if so use it. if not, use 7442 * napi_hash_add to generate one for us. 7443 */ 7444 if (n->config->napi_id) { 7445 napi_hash_add_with_id(n, n->config->napi_id); 7446 } else { 7447 napi_hash_add(n); 7448 n->config->napi_id = n->napi_id; 7449 } 7450 7451 WARN_ON_ONCE(napi_set_threaded(n, n->config->threaded)); 7452 } 7453 7454 static void napi_save_config(struct napi_struct *n) 7455 { 7456 n->config->defer_hard_irqs = n->defer_hard_irqs; 7457 n->config->gro_flush_timeout = n->gro_flush_timeout; 7458 n->config->irq_suspend_timeout = n->irq_suspend_timeout; 7459 napi_hash_del(n); 7460 } 7461 7462 /* Netlink wants the NAPI list to be sorted by ID, if adding a NAPI which will 7463 * inherit an existing ID try to insert it at the right position. 7464 */ 7465 static void 7466 netif_napi_dev_list_add(struct net_device *dev, struct napi_struct *napi) 7467 { 7468 unsigned int new_id, pos_id; 7469 struct list_head *higher; 7470 struct napi_struct *pos; 7471 7472 new_id = UINT_MAX; 7473 if (napi->config && napi->config->napi_id) 7474 new_id = napi->config->napi_id; 7475 7476 higher = &dev->napi_list; 7477 list_for_each_entry(pos, &dev->napi_list, dev_list) { 7478 if (napi_id_valid(pos->napi_id)) 7479 pos_id = pos->napi_id; 7480 else if (pos->config) 7481 pos_id = pos->config->napi_id; 7482 else 7483 pos_id = UINT_MAX; 7484 7485 if (pos_id <= new_id) 7486 break; 7487 higher = &pos->dev_list; 7488 } 7489 list_add_rcu(&napi->dev_list, higher); /* adds after higher */ 7490 } 7491 7492 /* Double check that napi_get_frags() allocates skbs with 7493 * skb->head being backed by slab, not a page fragment. 7494 * This is to make sure bug fixed in 3226b158e67c 7495 * ("net: avoid 32 x truesize under-estimation for tiny skbs") 7496 * does not accidentally come back. 7497 */ 7498 static void napi_get_frags_check(struct napi_struct *napi) 7499 { 7500 struct sk_buff *skb; 7501 7502 local_bh_disable(); 7503 skb = napi_get_frags(napi); 7504 WARN_ON_ONCE(skb && skb->head_frag); 7505 napi_free_frags(napi); 7506 local_bh_enable(); 7507 } 7508 7509 void netif_napi_add_weight_locked(struct net_device *dev, 7510 struct napi_struct *napi, 7511 int (*poll)(struct napi_struct *, int), 7512 int weight) 7513 { 7514 netdev_assert_locked(dev); 7515 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state))) 7516 return; 7517 7518 INIT_LIST_HEAD(&napi->poll_list); 7519 INIT_HLIST_NODE(&napi->napi_hash_node); 7520 hrtimer_setup(&napi->timer, napi_watchdog, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); 7521 gro_init(&napi->gro); 7522 napi->skb = NULL; 7523 napi->poll = poll; 7524 if (weight > NAPI_POLL_WEIGHT) 7525 netdev_err_once(dev, "%s() called with weight %d\n", __func__, 7526 weight); 7527 napi->weight = weight; 7528 napi->dev = dev; 7529 #ifdef CONFIG_NETPOLL 7530 napi->poll_owner = -1; 7531 #endif 7532 napi->list_owner = -1; 7533 set_bit(NAPI_STATE_SCHED, &napi->state); 7534 set_bit(NAPI_STATE_NPSVC, &napi->state); 7535 netif_napi_dev_list_add(dev, napi); 7536 7537 /* default settings from sysfs are applied to all NAPIs. any per-NAPI 7538 * configuration will be loaded in napi_enable 7539 */ 7540 napi_set_defer_hard_irqs(napi, READ_ONCE(dev->napi_defer_hard_irqs)); 7541 napi_set_gro_flush_timeout(napi, READ_ONCE(dev->gro_flush_timeout)); 7542 7543 napi_get_frags_check(napi); 7544 /* Create kthread for this napi if dev->threaded is set. 7545 * Clear dev->threaded if kthread creation failed so that 7546 * threaded mode will not be enabled in napi_enable(). 7547 */ 7548 if (napi_get_threaded_config(dev, napi)) 7549 if (napi_kthread_create(napi)) 7550 dev->threaded = NETDEV_NAPI_THREADED_DISABLED; 7551 netif_napi_set_irq_locked(napi, -1); 7552 } 7553 EXPORT_SYMBOL(netif_napi_add_weight_locked); 7554 7555 void napi_disable_locked(struct napi_struct *n) 7556 { 7557 unsigned long val, new; 7558 7559 might_sleep(); 7560 netdev_assert_locked(n->dev); 7561 7562 set_bit(NAPI_STATE_DISABLE, &n->state); 7563 7564 val = READ_ONCE(n->state); 7565 do { 7566 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) { 7567 usleep_range(20, 200); 7568 val = READ_ONCE(n->state); 7569 } 7570 7571 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC; 7572 new &= ~(NAPIF_STATE_THREADED | 7573 NAPIF_STATE_THREADED_BUSY_POLL | 7574 NAPIF_STATE_PREFER_BUSY_POLL); 7575 } while (!try_cmpxchg(&n->state, &val, new)); 7576 7577 hrtimer_cancel(&n->timer); 7578 7579 if (n->config) 7580 napi_save_config(n); 7581 else 7582 napi_hash_del(n); 7583 7584 clear_bit(NAPI_STATE_DISABLE, &n->state); 7585 } 7586 EXPORT_SYMBOL(napi_disable_locked); 7587 7588 /** 7589 * napi_disable() - prevent NAPI from scheduling 7590 * @n: NAPI context 7591 * 7592 * Stop NAPI from being scheduled on this context. 7593 * Waits till any outstanding processing completes. 7594 * Takes netdev_lock() for associated net_device. 7595 */ 7596 void napi_disable(struct napi_struct *n) 7597 { 7598 netdev_lock(n->dev); 7599 napi_disable_locked(n); 7600 netdev_unlock(n->dev); 7601 } 7602 EXPORT_SYMBOL(napi_disable); 7603 7604 void napi_enable_locked(struct napi_struct *n) 7605 { 7606 unsigned long new, val = READ_ONCE(n->state); 7607 7608 if (n->config) 7609 napi_restore_config(n); 7610 else 7611 napi_hash_add(n); 7612 7613 do { 7614 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val)); 7615 7616 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC); 7617 if (n->dev->threaded && n->thread) 7618 new |= NAPIF_STATE_THREADED; 7619 } while (!try_cmpxchg(&n->state, &val, new)); 7620 } 7621 EXPORT_SYMBOL(napi_enable_locked); 7622 7623 /** 7624 * napi_enable() - enable NAPI scheduling 7625 * @n: NAPI context 7626 * 7627 * Enable scheduling of a NAPI instance. 7628 * Must be paired with napi_disable(). 7629 * Takes netdev_lock() for associated net_device. 7630 */ 7631 void napi_enable(struct napi_struct *n) 7632 { 7633 netdev_lock(n->dev); 7634 napi_enable_locked(n); 7635 netdev_unlock(n->dev); 7636 } 7637 EXPORT_SYMBOL(napi_enable); 7638 7639 /* Must be called in process context */ 7640 void __netif_napi_del_locked(struct napi_struct *napi) 7641 { 7642 netdev_assert_locked(napi->dev); 7643 7644 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state)) 7645 return; 7646 7647 /* Make sure NAPI is disabled (or was never enabled). */ 7648 WARN_ON(!test_bit(NAPI_STATE_SCHED, &napi->state)); 7649 7650 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7651 irq_set_affinity_notifier(napi->irq, NULL); 7652 7653 if (napi->config) { 7654 napi->index = -1; 7655 napi->config = NULL; 7656 } 7657 7658 list_del_rcu(&napi->dev_list); 7659 napi_free_frags(napi); 7660 7661 gro_cleanup(&napi->gro); 7662 7663 if (napi->thread) { 7664 kthread_stop(napi->thread); 7665 napi->thread = NULL; 7666 } 7667 } 7668 EXPORT_SYMBOL(__netif_napi_del_locked); 7669 7670 static int __napi_poll(struct napi_struct *n, bool *repoll) 7671 { 7672 int work, weight; 7673 7674 weight = n->weight; 7675 7676 /* This NAPI_STATE_SCHED test is for avoiding a race 7677 * with netpoll's poll_napi(). Only the entity which 7678 * obtains the lock and sees NAPI_STATE_SCHED set will 7679 * actually make the ->poll() call. Therefore we avoid 7680 * accidentally calling ->poll() when NAPI is not scheduled. 7681 */ 7682 work = 0; 7683 if (napi_is_scheduled(n)) { 7684 work = n->poll(n, weight); 7685 trace_napi_poll(n, work, weight); 7686 7687 xdp_do_check_flushed(n); 7688 } 7689 7690 if (unlikely(work > weight)) 7691 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n", 7692 n->poll, work, weight); 7693 7694 if (likely(work < weight)) 7695 return work; 7696 7697 /* Drivers must not modify the NAPI state if they 7698 * consume the entire weight. In such cases this code 7699 * still "owns" the NAPI instance and therefore can 7700 * move the instance around on the list at-will. 7701 */ 7702 if (unlikely(napi_disable_pending(n))) { 7703 napi_complete(n); 7704 return work; 7705 } 7706 7707 /* The NAPI context has more processing work, but busy-polling 7708 * is preferred. Exit early. 7709 */ 7710 if (napi_prefer_busy_poll(n)) { 7711 if (napi_complete_done(n, work)) { 7712 /* If timeout is not set, we need to make sure 7713 * that the NAPI is re-scheduled. 7714 */ 7715 napi_schedule(n); 7716 } 7717 return work; 7718 } 7719 7720 /* Flush too old packets. If HZ < 1000, flush all packets */ 7721 gro_flush_normal(&n->gro, HZ >= 1000); 7722 7723 /* Some drivers may have called napi_schedule 7724 * prior to exhausting their budget. 7725 */ 7726 if (unlikely(!list_empty(&n->poll_list))) { 7727 pr_warn_once("%s: Budget exhausted after napi rescheduled\n", 7728 n->dev ? n->dev->name : "backlog"); 7729 return work; 7730 } 7731 7732 *repoll = true; 7733 7734 return work; 7735 } 7736 7737 static int napi_poll(struct napi_struct *n, struct list_head *repoll) 7738 { 7739 bool do_repoll = false; 7740 void *have; 7741 int work; 7742 7743 list_del_init(&n->poll_list); 7744 7745 have = netpoll_poll_lock(n); 7746 7747 work = __napi_poll(n, &do_repoll); 7748 7749 if (do_repoll) { 7750 #if defined(CONFIG_DEBUG_NET) 7751 if (unlikely(!napi_is_scheduled(n))) 7752 pr_crit("repoll requested for device %s %ps but napi is not scheduled.\n", 7753 n->dev->name, n->poll); 7754 #endif 7755 list_add_tail(&n->poll_list, repoll); 7756 } 7757 netpoll_poll_unlock(have); 7758 7759 return work; 7760 } 7761 7762 static int napi_thread_wait(struct napi_struct *napi) 7763 { 7764 set_current_state(TASK_INTERRUPTIBLE); 7765 7766 while (!kthread_should_stop()) { 7767 /* Testing SCHED_THREADED bit here to make sure the current 7768 * kthread owns this napi and could poll on this napi. 7769 * Testing SCHED bit is not enough because SCHED bit might be 7770 * set by some other busy poll thread or by napi_disable(). 7771 */ 7772 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) { 7773 WARN_ON(!list_empty(&napi->poll_list)); 7774 __set_current_state(TASK_RUNNING); 7775 return 0; 7776 } 7777 7778 schedule(); 7779 set_current_state(TASK_INTERRUPTIBLE); 7780 } 7781 __set_current_state(TASK_RUNNING); 7782 7783 return -1; 7784 } 7785 7786 static void napi_threaded_poll_loop(struct napi_struct *napi, bool busy_poll) 7787 { 7788 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7789 struct softnet_data *sd; 7790 unsigned long last_qs = jiffies; 7791 7792 for (;;) { 7793 bool repoll = false; 7794 void *have; 7795 7796 local_bh_disable(); 7797 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7798 7799 sd = this_cpu_ptr(&softnet_data); 7800 sd->in_napi_threaded_poll = true; 7801 7802 have = netpoll_poll_lock(napi); 7803 __napi_poll(napi, &repoll); 7804 netpoll_poll_unlock(have); 7805 7806 sd->in_napi_threaded_poll = false; 7807 barrier(); 7808 7809 if (sd_has_rps_ipi_waiting(sd)) { 7810 local_irq_disable(); 7811 net_rps_action_and_irq_enable(sd); 7812 } 7813 skb_defer_free_flush(); 7814 bpf_net_ctx_clear(bpf_net_ctx); 7815 7816 /* When busy poll is enabled, the old packets are not flushed in 7817 * napi_complete_done. So flush them here. 7818 */ 7819 if (busy_poll) 7820 gro_flush_normal(&napi->gro, HZ >= 1000); 7821 local_bh_enable(); 7822 7823 /* Call cond_resched here to avoid watchdog warnings. */ 7824 if (repoll || busy_poll) { 7825 rcu_softirq_qs_periodic(last_qs); 7826 cond_resched(); 7827 } 7828 7829 if (!repoll) 7830 break; 7831 } 7832 } 7833 7834 static int napi_threaded_poll(void *data) 7835 { 7836 struct napi_struct *napi = data; 7837 bool want_busy_poll; 7838 bool in_busy_poll; 7839 unsigned long val; 7840 7841 while (!napi_thread_wait(napi)) { 7842 val = READ_ONCE(napi->state); 7843 7844 want_busy_poll = val & NAPIF_STATE_THREADED_BUSY_POLL; 7845 in_busy_poll = val & NAPIF_STATE_IN_BUSY_POLL; 7846 7847 if (unlikely(val & NAPIF_STATE_DISABLE)) 7848 want_busy_poll = false; 7849 7850 if (want_busy_poll != in_busy_poll) 7851 assign_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state, 7852 want_busy_poll); 7853 7854 napi_threaded_poll_loop(napi, want_busy_poll); 7855 } 7856 7857 return 0; 7858 } 7859 7860 static __latent_entropy void net_rx_action(void) 7861 { 7862 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 7863 unsigned long time_limit = jiffies + 7864 usecs_to_jiffies(READ_ONCE(net_hotdata.netdev_budget_usecs)); 7865 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7866 int budget = READ_ONCE(net_hotdata.netdev_budget); 7867 LIST_HEAD(list); 7868 LIST_HEAD(repoll); 7869 7870 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7871 start: 7872 sd->in_net_rx_action = true; 7873 local_irq_disable(); 7874 list_splice_init(&sd->poll_list, &list); 7875 local_irq_enable(); 7876 7877 for (;;) { 7878 struct napi_struct *n; 7879 7880 skb_defer_free_flush(); 7881 7882 if (list_empty(&list)) { 7883 if (list_empty(&repoll)) { 7884 sd->in_net_rx_action = false; 7885 barrier(); 7886 /* We need to check if ____napi_schedule() 7887 * had refilled poll_list while 7888 * sd->in_net_rx_action was true. 7889 */ 7890 if (!list_empty(&sd->poll_list)) 7891 goto start; 7892 if (!sd_has_rps_ipi_waiting(sd)) 7893 goto end; 7894 } 7895 break; 7896 } 7897 7898 n = list_first_entry(&list, struct napi_struct, poll_list); 7899 budget -= napi_poll(n, &repoll); 7900 7901 /* If softirq window is exhausted then punt. 7902 * Allow this to run for 2 jiffies since which will allow 7903 * an average latency of 1.5/HZ. 7904 */ 7905 if (unlikely(budget <= 0 || 7906 time_after_eq(jiffies, time_limit))) { 7907 /* Pairs with READ_ONCE() in softnet_seq_show() */ 7908 WRITE_ONCE(sd->time_squeeze, sd->time_squeeze + 1); 7909 break; 7910 } 7911 } 7912 7913 local_irq_disable(); 7914 7915 list_splice_tail_init(&sd->poll_list, &list); 7916 list_splice_tail(&repoll, &list); 7917 list_splice(&list, &sd->poll_list); 7918 if (!list_empty(&sd->poll_list)) 7919 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 7920 else 7921 sd->in_net_rx_action = false; 7922 7923 net_rps_action_and_irq_enable(sd); 7924 end: 7925 bpf_net_ctx_clear(bpf_net_ctx); 7926 } 7927 7928 struct netdev_adjacent { 7929 struct net_device *dev; 7930 netdevice_tracker dev_tracker; 7931 7932 /* upper master flag, there can only be one master device per list */ 7933 bool master; 7934 7935 /* lookup ignore flag */ 7936 bool ignore; 7937 7938 /* counter for the number of times this device was added to us */ 7939 u16 ref_nr; 7940 7941 /* private field for the users */ 7942 void *private; 7943 7944 struct list_head list; 7945 struct rcu_head rcu; 7946 }; 7947 7948 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev, 7949 struct list_head *adj_list) 7950 { 7951 struct netdev_adjacent *adj; 7952 7953 list_for_each_entry(adj, adj_list, list) { 7954 if (adj->dev == adj_dev) 7955 return adj; 7956 } 7957 return NULL; 7958 } 7959 7960 static int ____netdev_has_upper_dev(struct net_device *upper_dev, 7961 struct netdev_nested_priv *priv) 7962 { 7963 struct net_device *dev = (struct net_device *)priv->data; 7964 7965 return upper_dev == dev; 7966 } 7967 7968 /** 7969 * netdev_has_upper_dev - Check if device is linked to an upper device 7970 * @dev: device 7971 * @upper_dev: upper device to check 7972 * 7973 * Find out if a device is linked to specified upper device and return true 7974 * in case it is. Note that this checks only immediate upper device, 7975 * not through a complete stack of devices. The caller must hold the RTNL lock. 7976 */ 7977 bool netdev_has_upper_dev(struct net_device *dev, 7978 struct net_device *upper_dev) 7979 { 7980 struct netdev_nested_priv priv = { 7981 .data = (void *)upper_dev, 7982 }; 7983 7984 ASSERT_RTNL(); 7985 7986 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7987 &priv); 7988 } 7989 EXPORT_SYMBOL(netdev_has_upper_dev); 7990 7991 /** 7992 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device 7993 * @dev: device 7994 * @upper_dev: upper device to check 7995 * 7996 * Find out if a device is linked to specified upper device and return true 7997 * in case it is. Note that this checks the entire upper device chain. 7998 * The caller must hold rcu lock. 7999 */ 8000 8001 bool netdev_has_upper_dev_all_rcu(struct net_device *dev, 8002 struct net_device *upper_dev) 8003 { 8004 struct netdev_nested_priv priv = { 8005 .data = (void *)upper_dev, 8006 }; 8007 8008 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 8009 &priv); 8010 } 8011 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu); 8012 8013 /** 8014 * netdev_has_any_upper_dev - Check if device is linked to some device 8015 * @dev: device 8016 * 8017 * Find out if a device is linked to an upper device and return true in case 8018 * it is. The caller must hold the RTNL lock. 8019 */ 8020 bool netdev_has_any_upper_dev(struct net_device *dev) 8021 { 8022 ASSERT_RTNL(); 8023 8024 return !list_empty(&dev->adj_list.upper); 8025 } 8026 EXPORT_SYMBOL(netdev_has_any_upper_dev); 8027 8028 /** 8029 * netdev_master_upper_dev_get - Get master upper device 8030 * @dev: device 8031 * 8032 * Find a master upper device and return pointer to it or NULL in case 8033 * it's not there. The caller must hold the RTNL lock. 8034 */ 8035 struct net_device *netdev_master_upper_dev_get(struct net_device *dev) 8036 { 8037 struct netdev_adjacent *upper; 8038 8039 ASSERT_RTNL(); 8040 8041 if (list_empty(&dev->adj_list.upper)) 8042 return NULL; 8043 8044 upper = list_first_entry(&dev->adj_list.upper, 8045 struct netdev_adjacent, list); 8046 if (likely(upper->master)) 8047 return upper->dev; 8048 return NULL; 8049 } 8050 EXPORT_SYMBOL(netdev_master_upper_dev_get); 8051 8052 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev) 8053 { 8054 struct netdev_adjacent *upper; 8055 8056 ASSERT_RTNL(); 8057 8058 if (list_empty(&dev->adj_list.upper)) 8059 return NULL; 8060 8061 upper = list_first_entry(&dev->adj_list.upper, 8062 struct netdev_adjacent, list); 8063 if (likely(upper->master) && !upper->ignore) 8064 return upper->dev; 8065 return NULL; 8066 } 8067 8068 /** 8069 * netdev_has_any_lower_dev - Check if device is linked to some device 8070 * @dev: device 8071 * 8072 * Find out if a device is linked to a lower device and return true in case 8073 * it is. The caller must hold the RTNL lock. 8074 */ 8075 static bool netdev_has_any_lower_dev(struct net_device *dev) 8076 { 8077 ASSERT_RTNL(); 8078 8079 return !list_empty(&dev->adj_list.lower); 8080 } 8081 8082 void *netdev_adjacent_get_private(struct list_head *adj_list) 8083 { 8084 struct netdev_adjacent *adj; 8085 8086 adj = list_entry(adj_list, struct netdev_adjacent, list); 8087 8088 return adj->private; 8089 } 8090 EXPORT_SYMBOL(netdev_adjacent_get_private); 8091 8092 /** 8093 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list 8094 * @dev: device 8095 * @iter: list_head ** of the current position 8096 * 8097 * Gets the next device from the dev's upper list, starting from iter 8098 * position. The caller must hold RCU read lock. 8099 */ 8100 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev, 8101 struct list_head **iter) 8102 { 8103 struct netdev_adjacent *upper; 8104 8105 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 8106 8107 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8108 8109 if (&upper->list == &dev->adj_list.upper) 8110 return NULL; 8111 8112 *iter = &upper->list; 8113 8114 return upper->dev; 8115 } 8116 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu); 8117 8118 static struct net_device *__netdev_next_upper_dev(struct net_device *dev, 8119 struct list_head **iter, 8120 bool *ignore) 8121 { 8122 struct netdev_adjacent *upper; 8123 8124 upper = list_entry((*iter)->next, struct netdev_adjacent, list); 8125 8126 if (&upper->list == &dev->adj_list.upper) 8127 return NULL; 8128 8129 *iter = &upper->list; 8130 *ignore = upper->ignore; 8131 8132 return upper->dev; 8133 } 8134 8135 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev, 8136 struct list_head **iter) 8137 { 8138 struct netdev_adjacent *upper; 8139 8140 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 8141 8142 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8143 8144 if (&upper->list == &dev->adj_list.upper) 8145 return NULL; 8146 8147 *iter = &upper->list; 8148 8149 return upper->dev; 8150 } 8151 8152 static int __netdev_walk_all_upper_dev(struct net_device *dev, 8153 int (*fn)(struct net_device *dev, 8154 struct netdev_nested_priv *priv), 8155 struct netdev_nested_priv *priv) 8156 { 8157 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8158 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8159 int ret, cur = 0; 8160 bool ignore; 8161 8162 now = dev; 8163 iter = &dev->adj_list.upper; 8164 8165 while (1) { 8166 if (now != dev) { 8167 ret = fn(now, priv); 8168 if (ret) 8169 return ret; 8170 } 8171 8172 next = NULL; 8173 while (1) { 8174 udev = __netdev_next_upper_dev(now, &iter, &ignore); 8175 if (!udev) 8176 break; 8177 if (ignore) 8178 continue; 8179 8180 next = udev; 8181 niter = &udev->adj_list.upper; 8182 dev_stack[cur] = now; 8183 iter_stack[cur++] = iter; 8184 break; 8185 } 8186 8187 if (!next) { 8188 if (!cur) 8189 return 0; 8190 next = dev_stack[--cur]; 8191 niter = iter_stack[cur]; 8192 } 8193 8194 now = next; 8195 iter = niter; 8196 } 8197 8198 return 0; 8199 } 8200 8201 int netdev_walk_all_upper_dev_rcu(struct net_device *dev, 8202 int (*fn)(struct net_device *dev, 8203 struct netdev_nested_priv *priv), 8204 struct netdev_nested_priv *priv) 8205 { 8206 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8207 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8208 int ret, cur = 0; 8209 8210 now = dev; 8211 iter = &dev->adj_list.upper; 8212 8213 while (1) { 8214 if (now != dev) { 8215 ret = fn(now, priv); 8216 if (ret) 8217 return ret; 8218 } 8219 8220 next = NULL; 8221 while (1) { 8222 udev = netdev_next_upper_dev_rcu(now, &iter); 8223 if (!udev) 8224 break; 8225 8226 next = udev; 8227 niter = &udev->adj_list.upper; 8228 dev_stack[cur] = now; 8229 iter_stack[cur++] = iter; 8230 break; 8231 } 8232 8233 if (!next) { 8234 if (!cur) 8235 return 0; 8236 next = dev_stack[--cur]; 8237 niter = iter_stack[cur]; 8238 } 8239 8240 now = next; 8241 iter = niter; 8242 } 8243 8244 return 0; 8245 } 8246 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu); 8247 8248 static bool __netdev_has_upper_dev(struct net_device *dev, 8249 struct net_device *upper_dev) 8250 { 8251 struct netdev_nested_priv priv = { 8252 .flags = 0, 8253 .data = (void *)upper_dev, 8254 }; 8255 8256 ASSERT_RTNL(); 8257 8258 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev, 8259 &priv); 8260 } 8261 8262 /** 8263 * netdev_lower_get_next_private - Get the next ->private from the 8264 * lower neighbour list 8265 * @dev: device 8266 * @iter: list_head ** of the current position 8267 * 8268 * Gets the next netdev_adjacent->private from the dev's lower neighbour 8269 * list, starting from iter position. The caller must hold either hold the 8270 * RTNL lock or its own locking that guarantees that the neighbour lower 8271 * list will remain unchanged. 8272 */ 8273 void *netdev_lower_get_next_private(struct net_device *dev, 8274 struct list_head **iter) 8275 { 8276 struct netdev_adjacent *lower; 8277 8278 lower = list_entry(*iter, struct netdev_adjacent, list); 8279 8280 if (&lower->list == &dev->adj_list.lower) 8281 return NULL; 8282 8283 *iter = lower->list.next; 8284 8285 return lower->private; 8286 } 8287 EXPORT_SYMBOL(netdev_lower_get_next_private); 8288 8289 /** 8290 * netdev_lower_get_next_private_rcu - Get the next ->private from the 8291 * lower neighbour list, RCU 8292 * variant 8293 * @dev: device 8294 * @iter: list_head ** of the current position 8295 * 8296 * Gets the next netdev_adjacent->private from the dev's lower neighbour 8297 * list, starting from iter position. The caller must hold RCU read lock. 8298 */ 8299 void *netdev_lower_get_next_private_rcu(struct net_device *dev, 8300 struct list_head **iter) 8301 { 8302 struct netdev_adjacent *lower; 8303 8304 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 8305 8306 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8307 8308 if (&lower->list == &dev->adj_list.lower) 8309 return NULL; 8310 8311 *iter = &lower->list; 8312 8313 return lower->private; 8314 } 8315 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu); 8316 8317 /** 8318 * netdev_lower_get_next - Get the next device from the lower neighbour 8319 * list 8320 * @dev: device 8321 * @iter: list_head ** of the current position 8322 * 8323 * Gets the next netdev_adjacent from the dev's lower neighbour 8324 * list, starting from iter position. The caller must hold RTNL lock or 8325 * its own locking that guarantees that the neighbour lower 8326 * list will remain unchanged. 8327 */ 8328 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter) 8329 { 8330 struct netdev_adjacent *lower; 8331 8332 lower = list_entry(*iter, struct netdev_adjacent, list); 8333 8334 if (&lower->list == &dev->adj_list.lower) 8335 return NULL; 8336 8337 *iter = lower->list.next; 8338 8339 return lower->dev; 8340 } 8341 EXPORT_SYMBOL(netdev_lower_get_next); 8342 8343 static struct net_device *netdev_next_lower_dev(struct net_device *dev, 8344 struct list_head **iter) 8345 { 8346 struct netdev_adjacent *lower; 8347 8348 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 8349 8350 if (&lower->list == &dev->adj_list.lower) 8351 return NULL; 8352 8353 *iter = &lower->list; 8354 8355 return lower->dev; 8356 } 8357 8358 static struct net_device *__netdev_next_lower_dev(struct net_device *dev, 8359 struct list_head **iter, 8360 bool *ignore) 8361 { 8362 struct netdev_adjacent *lower; 8363 8364 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 8365 8366 if (&lower->list == &dev->adj_list.lower) 8367 return NULL; 8368 8369 *iter = &lower->list; 8370 *ignore = lower->ignore; 8371 8372 return lower->dev; 8373 } 8374 8375 int netdev_walk_all_lower_dev(struct net_device *dev, 8376 int (*fn)(struct net_device *dev, 8377 struct netdev_nested_priv *priv), 8378 struct netdev_nested_priv *priv) 8379 { 8380 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8381 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8382 int ret, cur = 0; 8383 8384 now = dev; 8385 iter = &dev->adj_list.lower; 8386 8387 while (1) { 8388 if (now != dev) { 8389 ret = fn(now, priv); 8390 if (ret) 8391 return ret; 8392 } 8393 8394 next = NULL; 8395 while (1) { 8396 ldev = netdev_next_lower_dev(now, &iter); 8397 if (!ldev) 8398 break; 8399 8400 next = ldev; 8401 niter = &ldev->adj_list.lower; 8402 dev_stack[cur] = now; 8403 iter_stack[cur++] = iter; 8404 break; 8405 } 8406 8407 if (!next) { 8408 if (!cur) 8409 return 0; 8410 next = dev_stack[--cur]; 8411 niter = iter_stack[cur]; 8412 } 8413 8414 now = next; 8415 iter = niter; 8416 } 8417 8418 return 0; 8419 } 8420 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev); 8421 8422 static int __netdev_walk_all_lower_dev(struct net_device *dev, 8423 int (*fn)(struct net_device *dev, 8424 struct netdev_nested_priv *priv), 8425 struct netdev_nested_priv *priv) 8426 { 8427 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8428 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8429 int ret, cur = 0; 8430 bool ignore; 8431 8432 now = dev; 8433 iter = &dev->adj_list.lower; 8434 8435 while (1) { 8436 if (now != dev) { 8437 ret = fn(now, priv); 8438 if (ret) 8439 return ret; 8440 } 8441 8442 next = NULL; 8443 while (1) { 8444 ldev = __netdev_next_lower_dev(now, &iter, &ignore); 8445 if (!ldev) 8446 break; 8447 if (ignore) 8448 continue; 8449 8450 next = ldev; 8451 niter = &ldev->adj_list.lower; 8452 dev_stack[cur] = now; 8453 iter_stack[cur++] = iter; 8454 break; 8455 } 8456 8457 if (!next) { 8458 if (!cur) 8459 return 0; 8460 next = dev_stack[--cur]; 8461 niter = iter_stack[cur]; 8462 } 8463 8464 now = next; 8465 iter = niter; 8466 } 8467 8468 return 0; 8469 } 8470 8471 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev, 8472 struct list_head **iter) 8473 { 8474 struct netdev_adjacent *lower; 8475 8476 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8477 if (&lower->list == &dev->adj_list.lower) 8478 return NULL; 8479 8480 *iter = &lower->list; 8481 8482 return lower->dev; 8483 } 8484 EXPORT_SYMBOL(netdev_next_lower_dev_rcu); 8485 8486 static u8 __netdev_upper_depth(struct net_device *dev) 8487 { 8488 struct net_device *udev; 8489 struct list_head *iter; 8490 u8 max_depth = 0; 8491 bool ignore; 8492 8493 for (iter = &dev->adj_list.upper, 8494 udev = __netdev_next_upper_dev(dev, &iter, &ignore); 8495 udev; 8496 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) { 8497 if (ignore) 8498 continue; 8499 if (max_depth < udev->upper_level) 8500 max_depth = udev->upper_level; 8501 } 8502 8503 return max_depth; 8504 } 8505 8506 static u8 __netdev_lower_depth(struct net_device *dev) 8507 { 8508 struct net_device *ldev; 8509 struct list_head *iter; 8510 u8 max_depth = 0; 8511 bool ignore; 8512 8513 for (iter = &dev->adj_list.lower, 8514 ldev = __netdev_next_lower_dev(dev, &iter, &ignore); 8515 ldev; 8516 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) { 8517 if (ignore) 8518 continue; 8519 if (max_depth < ldev->lower_level) 8520 max_depth = ldev->lower_level; 8521 } 8522 8523 return max_depth; 8524 } 8525 8526 static int __netdev_update_upper_level(struct net_device *dev, 8527 struct netdev_nested_priv *__unused) 8528 { 8529 dev->upper_level = __netdev_upper_depth(dev) + 1; 8530 return 0; 8531 } 8532 8533 #ifdef CONFIG_LOCKDEP 8534 static LIST_HEAD(net_unlink_list); 8535 8536 static void net_unlink_todo(struct net_device *dev) 8537 { 8538 if (list_empty(&dev->unlink_list)) 8539 list_add_tail(&dev->unlink_list, &net_unlink_list); 8540 } 8541 #endif 8542 8543 static int __netdev_update_lower_level(struct net_device *dev, 8544 struct netdev_nested_priv *priv) 8545 { 8546 dev->lower_level = __netdev_lower_depth(dev) + 1; 8547 8548 #ifdef CONFIG_LOCKDEP 8549 if (!priv) 8550 return 0; 8551 8552 if (priv->flags & NESTED_SYNC_IMM) 8553 dev->nested_level = dev->lower_level - 1; 8554 if (priv->flags & NESTED_SYNC_TODO) 8555 net_unlink_todo(dev); 8556 #endif 8557 return 0; 8558 } 8559 8560 int netdev_walk_all_lower_dev_rcu(struct net_device *dev, 8561 int (*fn)(struct net_device *dev, 8562 struct netdev_nested_priv *priv), 8563 struct netdev_nested_priv *priv) 8564 { 8565 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8566 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8567 int ret, cur = 0; 8568 8569 now = dev; 8570 iter = &dev->adj_list.lower; 8571 8572 while (1) { 8573 if (now != dev) { 8574 ret = fn(now, priv); 8575 if (ret) 8576 return ret; 8577 } 8578 8579 next = NULL; 8580 while (1) { 8581 ldev = netdev_next_lower_dev_rcu(now, &iter); 8582 if (!ldev) 8583 break; 8584 8585 next = ldev; 8586 niter = &ldev->adj_list.lower; 8587 dev_stack[cur] = now; 8588 iter_stack[cur++] = iter; 8589 break; 8590 } 8591 8592 if (!next) { 8593 if (!cur) 8594 return 0; 8595 next = dev_stack[--cur]; 8596 niter = iter_stack[cur]; 8597 } 8598 8599 now = next; 8600 iter = niter; 8601 } 8602 8603 return 0; 8604 } 8605 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu); 8606 8607 /** 8608 * netdev_lower_get_first_private_rcu - Get the first ->private from the 8609 * lower neighbour list, RCU 8610 * variant 8611 * @dev: device 8612 * 8613 * Gets the first netdev_adjacent->private from the dev's lower neighbour 8614 * list. The caller must hold RCU read lock. 8615 */ 8616 void *netdev_lower_get_first_private_rcu(struct net_device *dev) 8617 { 8618 struct netdev_adjacent *lower; 8619 8620 lower = list_first_or_null_rcu(&dev->adj_list.lower, 8621 struct netdev_adjacent, list); 8622 if (lower) 8623 return lower->private; 8624 return NULL; 8625 } 8626 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu); 8627 8628 /** 8629 * netdev_master_upper_dev_get_rcu - Get master upper device 8630 * @dev: device 8631 * 8632 * Find a master upper device and return pointer to it or NULL in case 8633 * it's not there. The caller must hold the RCU read lock. 8634 */ 8635 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev) 8636 { 8637 struct netdev_adjacent *upper; 8638 8639 upper = list_first_or_null_rcu(&dev->adj_list.upper, 8640 struct netdev_adjacent, list); 8641 if (upper && likely(upper->master)) 8642 return upper->dev; 8643 return NULL; 8644 } 8645 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu); 8646 8647 static int netdev_adjacent_sysfs_add(struct net_device *dev, 8648 struct net_device *adj_dev, 8649 struct list_head *dev_list) 8650 { 8651 char linkname[IFNAMSIZ+7]; 8652 8653 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8654 "upper_%s" : "lower_%s", adj_dev->name); 8655 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj), 8656 linkname); 8657 } 8658 static void netdev_adjacent_sysfs_del(struct net_device *dev, 8659 char *name, 8660 struct list_head *dev_list) 8661 { 8662 char linkname[IFNAMSIZ+7]; 8663 8664 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8665 "upper_%s" : "lower_%s", name); 8666 sysfs_remove_link(&(dev->dev.kobj), linkname); 8667 } 8668 8669 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev, 8670 struct net_device *adj_dev, 8671 struct list_head *dev_list) 8672 { 8673 return (dev_list == &dev->adj_list.upper || 8674 dev_list == &dev->adj_list.lower) && 8675 net_eq(dev_net(dev), dev_net(adj_dev)); 8676 } 8677 8678 static int __netdev_adjacent_dev_insert(struct net_device *dev, 8679 struct net_device *adj_dev, 8680 struct list_head *dev_list, 8681 void *private, bool master) 8682 { 8683 struct netdev_adjacent *adj; 8684 int ret; 8685 8686 adj = __netdev_find_adj(adj_dev, dev_list); 8687 8688 if (adj) { 8689 adj->ref_nr += 1; 8690 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n", 8691 dev->name, adj_dev->name, adj->ref_nr); 8692 8693 return 0; 8694 } 8695 8696 adj = kmalloc_obj(*adj, GFP_KERNEL); 8697 if (!adj) 8698 return -ENOMEM; 8699 8700 adj->dev = adj_dev; 8701 adj->master = master; 8702 adj->ref_nr = 1; 8703 adj->private = private; 8704 adj->ignore = false; 8705 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL); 8706 8707 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n", 8708 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name); 8709 8710 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) { 8711 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list); 8712 if (ret) 8713 goto free_adj; 8714 } 8715 8716 /* Ensure that master link is always the first item in list. */ 8717 if (master) { 8718 ret = sysfs_create_link(&(dev->dev.kobj), 8719 &(adj_dev->dev.kobj), "master"); 8720 if (ret) 8721 goto remove_symlinks; 8722 8723 list_add_rcu(&adj->list, dev_list); 8724 } else { 8725 list_add_tail_rcu(&adj->list, dev_list); 8726 } 8727 8728 return 0; 8729 8730 remove_symlinks: 8731 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8732 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8733 free_adj: 8734 netdev_put(adj_dev, &adj->dev_tracker); 8735 kfree(adj); 8736 8737 return ret; 8738 } 8739 8740 static void __netdev_adjacent_dev_remove(struct net_device *dev, 8741 struct net_device *adj_dev, 8742 u16 ref_nr, 8743 struct list_head *dev_list) 8744 { 8745 struct netdev_adjacent *adj; 8746 8747 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n", 8748 dev->name, adj_dev->name, ref_nr); 8749 8750 adj = __netdev_find_adj(adj_dev, dev_list); 8751 8752 if (!adj) { 8753 pr_err("Adjacency does not exist for device %s from %s\n", 8754 dev->name, adj_dev->name); 8755 WARN_ON(1); 8756 return; 8757 } 8758 8759 if (adj->ref_nr > ref_nr) { 8760 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n", 8761 dev->name, adj_dev->name, ref_nr, 8762 adj->ref_nr - ref_nr); 8763 adj->ref_nr -= ref_nr; 8764 return; 8765 } 8766 8767 if (adj->master) 8768 sysfs_remove_link(&(dev->dev.kobj), "master"); 8769 8770 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8771 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8772 8773 list_del_rcu(&adj->list); 8774 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n", 8775 adj_dev->name, dev->name, adj_dev->name); 8776 netdev_put(adj_dev, &adj->dev_tracker); 8777 kfree_rcu(adj, rcu); 8778 } 8779 8780 static int __netdev_adjacent_dev_link_lists(struct net_device *dev, 8781 struct net_device *upper_dev, 8782 struct list_head *up_list, 8783 struct list_head *down_list, 8784 void *private, bool master) 8785 { 8786 int ret; 8787 8788 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list, 8789 private, master); 8790 if (ret) 8791 return ret; 8792 8793 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list, 8794 private, false); 8795 if (ret) { 8796 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list); 8797 return ret; 8798 } 8799 8800 return 0; 8801 } 8802 8803 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev, 8804 struct net_device *upper_dev, 8805 u16 ref_nr, 8806 struct list_head *up_list, 8807 struct list_head *down_list) 8808 { 8809 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list); 8810 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list); 8811 } 8812 8813 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev, 8814 struct net_device *upper_dev, 8815 void *private, bool master) 8816 { 8817 return __netdev_adjacent_dev_link_lists(dev, upper_dev, 8818 &dev->adj_list.upper, 8819 &upper_dev->adj_list.lower, 8820 private, master); 8821 } 8822 8823 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev, 8824 struct net_device *upper_dev) 8825 { 8826 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1, 8827 &dev->adj_list.upper, 8828 &upper_dev->adj_list.lower); 8829 } 8830 8831 static int __netdev_upper_dev_link(struct net_device *dev, 8832 struct net_device *upper_dev, bool master, 8833 void *upper_priv, void *upper_info, 8834 struct netdev_nested_priv *priv, 8835 struct netlink_ext_ack *extack) 8836 { 8837 struct netdev_notifier_changeupper_info changeupper_info = { 8838 .info = { 8839 .dev = dev, 8840 .extack = extack, 8841 }, 8842 .upper_dev = upper_dev, 8843 .master = master, 8844 .linking = true, 8845 .upper_info = upper_info, 8846 }; 8847 struct net_device *master_dev; 8848 int ret = 0; 8849 8850 ASSERT_RTNL(); 8851 8852 if (dev == upper_dev) 8853 return -EBUSY; 8854 8855 /* To prevent loops, check if dev is not upper device to upper_dev. */ 8856 if (__netdev_has_upper_dev(upper_dev, dev)) 8857 return -EBUSY; 8858 8859 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV) 8860 return -EMLINK; 8861 8862 if (!master) { 8863 if (__netdev_has_upper_dev(dev, upper_dev)) 8864 return -EEXIST; 8865 } else { 8866 master_dev = __netdev_master_upper_dev_get(dev); 8867 if (master_dev) 8868 return master_dev == upper_dev ? -EEXIST : -EBUSY; 8869 } 8870 8871 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8872 &changeupper_info.info); 8873 ret = notifier_to_errno(ret); 8874 if (ret) 8875 return ret; 8876 8877 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv, 8878 master); 8879 if (ret) 8880 return ret; 8881 8882 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8883 &changeupper_info.info); 8884 ret = notifier_to_errno(ret); 8885 if (ret) 8886 goto rollback; 8887 8888 __netdev_update_upper_level(dev, NULL); 8889 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8890 8891 __netdev_update_lower_level(upper_dev, priv); 8892 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8893 priv); 8894 8895 return 0; 8896 8897 rollback: 8898 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8899 8900 return ret; 8901 } 8902 8903 /** 8904 * netdev_upper_dev_link - Add a link to the upper device 8905 * @dev: device 8906 * @upper_dev: new upper device 8907 * @extack: netlink extended ack 8908 * 8909 * Adds a link to device which is upper to this one. The caller must hold 8910 * the RTNL lock. On a failure a negative errno code is returned. 8911 * On success the reference counts are adjusted and the function 8912 * returns zero. 8913 */ 8914 int netdev_upper_dev_link(struct net_device *dev, 8915 struct net_device *upper_dev, 8916 struct netlink_ext_ack *extack) 8917 { 8918 struct netdev_nested_priv priv = { 8919 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8920 .data = NULL, 8921 }; 8922 8923 return __netdev_upper_dev_link(dev, upper_dev, false, 8924 NULL, NULL, &priv, extack); 8925 } 8926 EXPORT_SYMBOL(netdev_upper_dev_link); 8927 8928 /** 8929 * netdev_master_upper_dev_link - Add a master link to the upper device 8930 * @dev: device 8931 * @upper_dev: new upper device 8932 * @upper_priv: upper device private 8933 * @upper_info: upper info to be passed down via notifier 8934 * @extack: netlink extended ack 8935 * 8936 * Adds a link to device which is upper to this one. In this case, only 8937 * one master upper device can be linked, although other non-master devices 8938 * might be linked as well. The caller must hold the RTNL lock. 8939 * On a failure a negative errno code is returned. On success the reference 8940 * counts are adjusted and the function returns zero. 8941 */ 8942 int netdev_master_upper_dev_link(struct net_device *dev, 8943 struct net_device *upper_dev, 8944 void *upper_priv, void *upper_info, 8945 struct netlink_ext_ack *extack) 8946 { 8947 struct netdev_nested_priv priv = { 8948 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8949 .data = NULL, 8950 }; 8951 8952 return __netdev_upper_dev_link(dev, upper_dev, true, 8953 upper_priv, upper_info, &priv, extack); 8954 } 8955 EXPORT_SYMBOL(netdev_master_upper_dev_link); 8956 8957 static void __netdev_upper_dev_unlink(struct net_device *dev, 8958 struct net_device *upper_dev, 8959 struct netdev_nested_priv *priv) 8960 { 8961 struct netdev_notifier_changeupper_info changeupper_info = { 8962 .info = { 8963 .dev = dev, 8964 }, 8965 .upper_dev = upper_dev, 8966 .linking = false, 8967 }; 8968 8969 ASSERT_RTNL(); 8970 8971 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev; 8972 8973 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8974 &changeupper_info.info); 8975 8976 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8977 8978 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8979 &changeupper_info.info); 8980 8981 __netdev_update_upper_level(dev, NULL); 8982 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8983 8984 __netdev_update_lower_level(upper_dev, priv); 8985 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8986 priv); 8987 } 8988 8989 /** 8990 * netdev_upper_dev_unlink - Removes a link to upper device 8991 * @dev: device 8992 * @upper_dev: new upper device 8993 * 8994 * Removes a link to device which is upper to this one. The caller must hold 8995 * the RTNL lock. 8996 */ 8997 void netdev_upper_dev_unlink(struct net_device *dev, 8998 struct net_device *upper_dev) 8999 { 9000 struct netdev_nested_priv priv = { 9001 .flags = NESTED_SYNC_TODO, 9002 .data = NULL, 9003 }; 9004 9005 __netdev_upper_dev_unlink(dev, upper_dev, &priv); 9006 } 9007 EXPORT_SYMBOL(netdev_upper_dev_unlink); 9008 9009 static void __netdev_adjacent_dev_set(struct net_device *upper_dev, 9010 struct net_device *lower_dev, 9011 bool val) 9012 { 9013 struct netdev_adjacent *adj; 9014 9015 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower); 9016 if (adj) 9017 adj->ignore = val; 9018 9019 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper); 9020 if (adj) 9021 adj->ignore = val; 9022 } 9023 9024 static void netdev_adjacent_dev_disable(struct net_device *upper_dev, 9025 struct net_device *lower_dev) 9026 { 9027 __netdev_adjacent_dev_set(upper_dev, lower_dev, true); 9028 } 9029 9030 static void netdev_adjacent_dev_enable(struct net_device *upper_dev, 9031 struct net_device *lower_dev) 9032 { 9033 __netdev_adjacent_dev_set(upper_dev, lower_dev, false); 9034 } 9035 9036 int netdev_adjacent_change_prepare(struct net_device *old_dev, 9037 struct net_device *new_dev, 9038 struct net_device *dev, 9039 struct netlink_ext_ack *extack) 9040 { 9041 struct netdev_nested_priv priv = { 9042 .flags = 0, 9043 .data = NULL, 9044 }; 9045 int err; 9046 9047 if (!new_dev) 9048 return 0; 9049 9050 if (old_dev && new_dev != old_dev) 9051 netdev_adjacent_dev_disable(dev, old_dev); 9052 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv, 9053 extack); 9054 if (err) { 9055 if (old_dev && new_dev != old_dev) 9056 netdev_adjacent_dev_enable(dev, old_dev); 9057 return err; 9058 } 9059 9060 return 0; 9061 } 9062 EXPORT_SYMBOL(netdev_adjacent_change_prepare); 9063 9064 void netdev_adjacent_change_commit(struct net_device *old_dev, 9065 struct net_device *new_dev, 9066 struct net_device *dev) 9067 { 9068 struct netdev_nested_priv priv = { 9069 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 9070 .data = NULL, 9071 }; 9072 9073 if (!new_dev || !old_dev) 9074 return; 9075 9076 if (new_dev == old_dev) 9077 return; 9078 9079 netdev_adjacent_dev_enable(dev, old_dev); 9080 __netdev_upper_dev_unlink(old_dev, dev, &priv); 9081 } 9082 EXPORT_SYMBOL(netdev_adjacent_change_commit); 9083 9084 void netdev_adjacent_change_abort(struct net_device *old_dev, 9085 struct net_device *new_dev, 9086 struct net_device *dev) 9087 { 9088 struct netdev_nested_priv priv = { 9089 .flags = 0, 9090 .data = NULL, 9091 }; 9092 9093 if (!new_dev) 9094 return; 9095 9096 if (old_dev && new_dev != old_dev) 9097 netdev_adjacent_dev_enable(dev, old_dev); 9098 9099 __netdev_upper_dev_unlink(new_dev, dev, &priv); 9100 } 9101 EXPORT_SYMBOL(netdev_adjacent_change_abort); 9102 9103 /** 9104 * netdev_bonding_info_change - Dispatch event about slave change 9105 * @dev: device 9106 * @bonding_info: info to dispatch 9107 * 9108 * Send NETDEV_BONDING_INFO to netdev notifiers with info. 9109 * The caller must hold the RTNL lock. 9110 */ 9111 void netdev_bonding_info_change(struct net_device *dev, 9112 struct netdev_bonding_info *bonding_info) 9113 { 9114 struct netdev_notifier_bonding_info info = { 9115 .info.dev = dev, 9116 }; 9117 9118 memcpy(&info.bonding_info, bonding_info, 9119 sizeof(struct netdev_bonding_info)); 9120 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, 9121 &info.info); 9122 } 9123 EXPORT_SYMBOL(netdev_bonding_info_change); 9124 9125 static int netdev_offload_xstats_enable_l3(struct net_device *dev, 9126 struct netlink_ext_ack *extack) 9127 { 9128 struct netdev_notifier_offload_xstats_info info = { 9129 .info.dev = dev, 9130 .info.extack = extack, 9131 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 9132 }; 9133 int err; 9134 int rc; 9135 9136 dev->offload_xstats_l3 = kzalloc_obj(*dev->offload_xstats_l3, 9137 GFP_KERNEL); 9138 if (!dev->offload_xstats_l3) 9139 return -ENOMEM; 9140 9141 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE, 9142 NETDEV_OFFLOAD_XSTATS_DISABLE, 9143 &info.info); 9144 err = notifier_to_errno(rc); 9145 if (err) 9146 goto free_stats; 9147 9148 return 0; 9149 9150 free_stats: 9151 kfree(dev->offload_xstats_l3); 9152 dev->offload_xstats_l3 = NULL; 9153 return err; 9154 } 9155 9156 int netdev_offload_xstats_enable(struct net_device *dev, 9157 enum netdev_offload_xstats_type type, 9158 struct netlink_ext_ack *extack) 9159 { 9160 ASSERT_RTNL(); 9161 9162 if (netdev_offload_xstats_enabled(dev, type)) 9163 return -EALREADY; 9164 9165 switch (type) { 9166 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 9167 return netdev_offload_xstats_enable_l3(dev, extack); 9168 } 9169 9170 WARN_ON(1); 9171 return -EINVAL; 9172 } 9173 EXPORT_SYMBOL(netdev_offload_xstats_enable); 9174 9175 static void netdev_offload_xstats_disable_l3(struct net_device *dev) 9176 { 9177 struct netdev_notifier_offload_xstats_info info = { 9178 .info.dev = dev, 9179 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 9180 }; 9181 9182 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE, 9183 &info.info); 9184 kfree(dev->offload_xstats_l3); 9185 dev->offload_xstats_l3 = NULL; 9186 } 9187 9188 int netdev_offload_xstats_disable(struct net_device *dev, 9189 enum netdev_offload_xstats_type type) 9190 { 9191 ASSERT_RTNL(); 9192 9193 if (!netdev_offload_xstats_enabled(dev, type)) 9194 return -EALREADY; 9195 9196 switch (type) { 9197 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 9198 netdev_offload_xstats_disable_l3(dev); 9199 return 0; 9200 } 9201 9202 WARN_ON(1); 9203 return -EINVAL; 9204 } 9205 EXPORT_SYMBOL(netdev_offload_xstats_disable); 9206 9207 static void netdev_offload_xstats_disable_all(struct net_device *dev) 9208 { 9209 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3); 9210 } 9211 9212 static struct rtnl_hw_stats64 * 9213 netdev_offload_xstats_get_ptr(const struct net_device *dev, 9214 enum netdev_offload_xstats_type type) 9215 { 9216 switch (type) { 9217 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 9218 return dev->offload_xstats_l3; 9219 } 9220 9221 WARN_ON(1); 9222 return NULL; 9223 } 9224 9225 bool netdev_offload_xstats_enabled(const struct net_device *dev, 9226 enum netdev_offload_xstats_type type) 9227 { 9228 ASSERT_RTNL(); 9229 9230 return netdev_offload_xstats_get_ptr(dev, type); 9231 } 9232 EXPORT_SYMBOL(netdev_offload_xstats_enabled); 9233 9234 struct netdev_notifier_offload_xstats_ru { 9235 bool used; 9236 }; 9237 9238 struct netdev_notifier_offload_xstats_rd { 9239 struct rtnl_hw_stats64 stats; 9240 bool used; 9241 }; 9242 9243 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest, 9244 const struct rtnl_hw_stats64 *src) 9245 { 9246 dest->rx_packets += src->rx_packets; 9247 dest->tx_packets += src->tx_packets; 9248 dest->rx_bytes += src->rx_bytes; 9249 dest->tx_bytes += src->tx_bytes; 9250 dest->rx_errors += src->rx_errors; 9251 dest->tx_errors += src->tx_errors; 9252 dest->rx_dropped += src->rx_dropped; 9253 dest->tx_dropped += src->tx_dropped; 9254 dest->multicast += src->multicast; 9255 } 9256 9257 static int netdev_offload_xstats_get_used(struct net_device *dev, 9258 enum netdev_offload_xstats_type type, 9259 bool *p_used, 9260 struct netlink_ext_ack *extack) 9261 { 9262 struct netdev_notifier_offload_xstats_ru report_used = {}; 9263 struct netdev_notifier_offload_xstats_info info = { 9264 .info.dev = dev, 9265 .info.extack = extack, 9266 .type = type, 9267 .report_used = &report_used, 9268 }; 9269 int rc; 9270 9271 WARN_ON(!netdev_offload_xstats_enabled(dev, type)); 9272 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED, 9273 &info.info); 9274 *p_used = report_used.used; 9275 return notifier_to_errno(rc); 9276 } 9277 9278 static int netdev_offload_xstats_get_stats(struct net_device *dev, 9279 enum netdev_offload_xstats_type type, 9280 struct rtnl_hw_stats64 *p_stats, 9281 bool *p_used, 9282 struct netlink_ext_ack *extack) 9283 { 9284 struct netdev_notifier_offload_xstats_rd report_delta = {}; 9285 struct netdev_notifier_offload_xstats_info info = { 9286 .info.dev = dev, 9287 .info.extack = extack, 9288 .type = type, 9289 .report_delta = &report_delta, 9290 }; 9291 struct rtnl_hw_stats64 *stats; 9292 int rc; 9293 9294 stats = netdev_offload_xstats_get_ptr(dev, type); 9295 if (WARN_ON(!stats)) 9296 return -EINVAL; 9297 9298 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA, 9299 &info.info); 9300 9301 /* Cache whatever we got, even if there was an error, otherwise the 9302 * successful stats retrievals would get lost. 9303 */ 9304 netdev_hw_stats64_add(stats, &report_delta.stats); 9305 9306 if (p_stats) 9307 *p_stats = *stats; 9308 *p_used = report_delta.used; 9309 9310 return notifier_to_errno(rc); 9311 } 9312 9313 int netdev_offload_xstats_get(struct net_device *dev, 9314 enum netdev_offload_xstats_type type, 9315 struct rtnl_hw_stats64 *p_stats, bool *p_used, 9316 struct netlink_ext_ack *extack) 9317 { 9318 ASSERT_RTNL(); 9319 9320 if (p_stats) 9321 return netdev_offload_xstats_get_stats(dev, type, p_stats, 9322 p_used, extack); 9323 else 9324 return netdev_offload_xstats_get_used(dev, type, p_used, 9325 extack); 9326 } 9327 EXPORT_SYMBOL(netdev_offload_xstats_get); 9328 9329 void 9330 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta, 9331 const struct rtnl_hw_stats64 *stats) 9332 { 9333 report_delta->used = true; 9334 netdev_hw_stats64_add(&report_delta->stats, stats); 9335 } 9336 EXPORT_SYMBOL(netdev_offload_xstats_report_delta); 9337 9338 void 9339 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used) 9340 { 9341 report_used->used = true; 9342 } 9343 EXPORT_SYMBOL(netdev_offload_xstats_report_used); 9344 9345 void netdev_offload_xstats_push_delta(struct net_device *dev, 9346 enum netdev_offload_xstats_type type, 9347 const struct rtnl_hw_stats64 *p_stats) 9348 { 9349 struct rtnl_hw_stats64 *stats; 9350 9351 ASSERT_RTNL(); 9352 9353 stats = netdev_offload_xstats_get_ptr(dev, type); 9354 if (WARN_ON(!stats)) 9355 return; 9356 9357 netdev_hw_stats64_add(stats, p_stats); 9358 } 9359 EXPORT_SYMBOL(netdev_offload_xstats_push_delta); 9360 9361 /** 9362 * netdev_get_xmit_slave - Get the xmit slave of master device 9363 * @dev: device 9364 * @skb: The packet 9365 * @all_slaves: assume all the slaves are active 9366 * 9367 * The reference counters are not incremented so the caller must be 9368 * careful with locks. The caller must hold RCU lock. 9369 * %NULL is returned if no slave is found. 9370 */ 9371 9372 struct net_device *netdev_get_xmit_slave(struct net_device *dev, 9373 struct sk_buff *skb, 9374 bool all_slaves) 9375 { 9376 const struct net_device_ops *ops = dev->netdev_ops; 9377 9378 if (!ops->ndo_get_xmit_slave) 9379 return NULL; 9380 return ops->ndo_get_xmit_slave(dev, skb, all_slaves); 9381 } 9382 EXPORT_SYMBOL(netdev_get_xmit_slave); 9383 9384 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev, 9385 struct sock *sk) 9386 { 9387 const struct net_device_ops *ops = dev->netdev_ops; 9388 9389 if (!ops->ndo_sk_get_lower_dev) 9390 return NULL; 9391 return ops->ndo_sk_get_lower_dev(dev, sk); 9392 } 9393 9394 /** 9395 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket 9396 * @dev: device 9397 * @sk: the socket 9398 * 9399 * %NULL is returned if no lower device is found. 9400 */ 9401 9402 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev, 9403 struct sock *sk) 9404 { 9405 struct net_device *lower; 9406 9407 lower = netdev_sk_get_lower_dev(dev, sk); 9408 while (lower) { 9409 dev = lower; 9410 lower = netdev_sk_get_lower_dev(dev, sk); 9411 } 9412 9413 return dev; 9414 } 9415 EXPORT_SYMBOL(netdev_sk_get_lowest_dev); 9416 9417 static void netdev_adjacent_add_links(struct net_device *dev) 9418 { 9419 struct netdev_adjacent *iter; 9420 9421 struct net *net = dev_net(dev); 9422 9423 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9424 if (!net_eq(net, dev_net(iter->dev))) 9425 continue; 9426 netdev_adjacent_sysfs_add(iter->dev, dev, 9427 &iter->dev->adj_list.lower); 9428 netdev_adjacent_sysfs_add(dev, iter->dev, 9429 &dev->adj_list.upper); 9430 } 9431 9432 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9433 if (!net_eq(net, dev_net(iter->dev))) 9434 continue; 9435 netdev_adjacent_sysfs_add(iter->dev, dev, 9436 &iter->dev->adj_list.upper); 9437 netdev_adjacent_sysfs_add(dev, iter->dev, 9438 &dev->adj_list.lower); 9439 } 9440 } 9441 9442 static void netdev_adjacent_del_links(struct net_device *dev) 9443 { 9444 struct netdev_adjacent *iter; 9445 9446 struct net *net = dev_net(dev); 9447 9448 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9449 if (!net_eq(net, dev_net(iter->dev))) 9450 continue; 9451 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9452 &iter->dev->adj_list.lower); 9453 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9454 &dev->adj_list.upper); 9455 } 9456 9457 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9458 if (!net_eq(net, dev_net(iter->dev))) 9459 continue; 9460 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9461 &iter->dev->adj_list.upper); 9462 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9463 &dev->adj_list.lower); 9464 } 9465 } 9466 9467 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname) 9468 { 9469 struct netdev_adjacent *iter; 9470 9471 struct net *net = dev_net(dev); 9472 9473 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9474 if (!net_eq(net, dev_net(iter->dev))) 9475 continue; 9476 netdev_adjacent_sysfs_del(iter->dev, oldname, 9477 &iter->dev->adj_list.lower); 9478 netdev_adjacent_sysfs_add(iter->dev, dev, 9479 &iter->dev->adj_list.lower); 9480 } 9481 9482 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9483 if (!net_eq(net, dev_net(iter->dev))) 9484 continue; 9485 netdev_adjacent_sysfs_del(iter->dev, oldname, 9486 &iter->dev->adj_list.upper); 9487 netdev_adjacent_sysfs_add(iter->dev, dev, 9488 &iter->dev->adj_list.upper); 9489 } 9490 } 9491 9492 void *netdev_lower_dev_get_private(struct net_device *dev, 9493 struct net_device *lower_dev) 9494 { 9495 struct netdev_adjacent *lower; 9496 9497 if (!lower_dev) 9498 return NULL; 9499 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower); 9500 if (!lower) 9501 return NULL; 9502 9503 return lower->private; 9504 } 9505 EXPORT_SYMBOL(netdev_lower_dev_get_private); 9506 9507 9508 /** 9509 * netdev_lower_state_changed - Dispatch event about lower device state change 9510 * @lower_dev: device 9511 * @lower_state_info: state to dispatch 9512 * 9513 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info. 9514 * The caller must hold the RTNL lock. 9515 */ 9516 void netdev_lower_state_changed(struct net_device *lower_dev, 9517 void *lower_state_info) 9518 { 9519 struct netdev_notifier_changelowerstate_info changelowerstate_info = { 9520 .info.dev = lower_dev, 9521 }; 9522 9523 ASSERT_RTNL(); 9524 changelowerstate_info.lower_state_info = lower_state_info; 9525 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, 9526 &changelowerstate_info.info); 9527 } 9528 EXPORT_SYMBOL(netdev_lower_state_changed); 9529 9530 static void dev_change_rx_flags(struct net_device *dev, int flags) 9531 { 9532 const struct net_device_ops *ops = dev->netdev_ops; 9533 9534 if (ops->ndo_change_rx_flags) 9535 ops->ndo_change_rx_flags(dev, flags); 9536 } 9537 9538 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify) 9539 { 9540 unsigned int old_flags = dev->flags; 9541 unsigned int promiscuity, flags; 9542 kuid_t uid; 9543 kgid_t gid; 9544 9545 ASSERT_RTNL(); 9546 9547 promiscuity = dev->promiscuity + inc; 9548 if (promiscuity == 0) { 9549 /* 9550 * Avoid overflow. 9551 * If inc causes overflow, untouch promisc and return error. 9552 */ 9553 if (unlikely(inc > 0)) { 9554 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n"); 9555 return -EOVERFLOW; 9556 } 9557 flags = old_flags & ~IFF_PROMISC; 9558 } else { 9559 flags = old_flags | IFF_PROMISC; 9560 } 9561 WRITE_ONCE(dev->promiscuity, promiscuity); 9562 if (flags != old_flags) { 9563 WRITE_ONCE(dev->flags, flags); 9564 netdev_info(dev, "%s promiscuous mode\n", 9565 dev->flags & IFF_PROMISC ? "entered" : "left"); 9566 if (audit_enabled) { 9567 current_uid_gid(&uid, &gid); 9568 audit_log(audit_context(), GFP_ATOMIC, 9569 AUDIT_ANOM_PROMISCUOUS, 9570 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u", 9571 dev->name, (dev->flags & IFF_PROMISC), 9572 (old_flags & IFF_PROMISC), 9573 from_kuid(&init_user_ns, audit_get_loginuid(current)), 9574 from_kuid(&init_user_ns, uid), 9575 from_kgid(&init_user_ns, gid), 9576 audit_get_sessionid(current)); 9577 } 9578 9579 dev_change_rx_flags(dev, IFF_PROMISC); 9580 } 9581 if (notify) { 9582 /* The ops lock is only required to ensure consistent locking 9583 * for `NETDEV_CHANGE` notifiers. This function is sometimes 9584 * called without the lock, even for devices that are ops 9585 * locked, such as in `dev_uc_sync_multiple` when using 9586 * bonding or teaming. 9587 */ 9588 netdev_ops_assert_locked(dev); 9589 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL); 9590 } 9591 return 0; 9592 } 9593 9594 int netif_set_promiscuity(struct net_device *dev, int inc) 9595 { 9596 unsigned int old_flags = dev->flags; 9597 int err; 9598 9599 err = __dev_set_promiscuity(dev, inc, true); 9600 if (err < 0) 9601 return err; 9602 if (dev->flags != old_flags) 9603 dev_set_rx_mode(dev); 9604 return err; 9605 } 9606 9607 int netif_set_allmulti(struct net_device *dev, int inc, bool notify) 9608 { 9609 unsigned int old_flags = dev->flags, old_gflags = dev->gflags; 9610 unsigned int allmulti, flags; 9611 9612 ASSERT_RTNL(); 9613 9614 allmulti = dev->allmulti + inc; 9615 if (allmulti == 0) { 9616 /* 9617 * Avoid overflow. 9618 * If inc causes overflow, untouch allmulti and return error. 9619 */ 9620 if (unlikely(inc > 0)) { 9621 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n"); 9622 return -EOVERFLOW; 9623 } 9624 flags = old_flags & ~IFF_ALLMULTI; 9625 } else { 9626 flags = old_flags | IFF_ALLMULTI; 9627 } 9628 WRITE_ONCE(dev->allmulti, allmulti); 9629 if (flags != old_flags) { 9630 WRITE_ONCE(dev->flags, flags); 9631 netdev_info(dev, "%s allmulticast mode\n", 9632 dev->flags & IFF_ALLMULTI ? "entered" : "left"); 9633 dev_change_rx_flags(dev, IFF_ALLMULTI); 9634 dev_set_rx_mode(dev); 9635 if (notify) 9636 __dev_notify_flags(dev, old_flags, 9637 dev->gflags ^ old_gflags, 0, NULL); 9638 } 9639 return 0; 9640 } 9641 9642 /* 9643 * Upload unicast and multicast address lists to device and 9644 * configure RX filtering. When the device doesn't support unicast 9645 * filtering it is put in promiscuous mode while unicast addresses 9646 * are present. 9647 */ 9648 void __dev_set_rx_mode(struct net_device *dev) 9649 { 9650 const struct net_device_ops *ops = dev->netdev_ops; 9651 9652 /* dev_open will call this function so the list will stay sane. */ 9653 if (!(dev->flags&IFF_UP)) 9654 return; 9655 9656 if (!netif_device_present(dev)) 9657 return; 9658 9659 if (!(dev->priv_flags & IFF_UNICAST_FLT)) { 9660 /* Unicast addresses changes may only happen under the rtnl, 9661 * therefore calling __dev_set_promiscuity here is safe. 9662 */ 9663 if (!netdev_uc_empty(dev) && !dev->uc_promisc) { 9664 __dev_set_promiscuity(dev, 1, false); 9665 dev->uc_promisc = true; 9666 } else if (netdev_uc_empty(dev) && dev->uc_promisc) { 9667 __dev_set_promiscuity(dev, -1, false); 9668 dev->uc_promisc = false; 9669 } 9670 } 9671 9672 if (ops->ndo_set_rx_mode) 9673 ops->ndo_set_rx_mode(dev); 9674 } 9675 9676 void dev_set_rx_mode(struct net_device *dev) 9677 { 9678 netif_addr_lock_bh(dev); 9679 __dev_set_rx_mode(dev); 9680 netif_addr_unlock_bh(dev); 9681 } 9682 9683 /** 9684 * netif_get_flags() - get flags reported to userspace 9685 * @dev: device 9686 * 9687 * Get the combination of flag bits exported through APIs to userspace. 9688 */ 9689 unsigned int netif_get_flags(const struct net_device *dev) 9690 { 9691 unsigned int flags; 9692 9693 flags = (READ_ONCE(dev->flags) & ~(IFF_PROMISC | 9694 IFF_ALLMULTI | 9695 IFF_RUNNING | 9696 IFF_LOWER_UP | 9697 IFF_DORMANT)) | 9698 (READ_ONCE(dev->gflags) & (IFF_PROMISC | 9699 IFF_ALLMULTI)); 9700 9701 if (netif_running(dev)) { 9702 if (netif_oper_up(dev)) 9703 flags |= IFF_RUNNING; 9704 if (netif_carrier_ok(dev)) 9705 flags |= IFF_LOWER_UP; 9706 if (netif_dormant(dev)) 9707 flags |= IFF_DORMANT; 9708 } 9709 9710 return flags; 9711 } 9712 EXPORT_SYMBOL(netif_get_flags); 9713 9714 int __dev_change_flags(struct net_device *dev, unsigned int flags, 9715 struct netlink_ext_ack *extack) 9716 { 9717 unsigned int old_flags = dev->flags; 9718 int ret; 9719 9720 ASSERT_RTNL(); 9721 9722 /* 9723 * Set the flags on our device. 9724 */ 9725 9726 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP | 9727 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL | 9728 IFF_AUTOMEDIA)) | 9729 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC | 9730 IFF_ALLMULTI)); 9731 9732 /* 9733 * Load in the correct multicast list now the flags have changed. 9734 */ 9735 9736 if ((old_flags ^ flags) & IFF_MULTICAST) 9737 dev_change_rx_flags(dev, IFF_MULTICAST); 9738 9739 dev_set_rx_mode(dev); 9740 9741 /* 9742 * Have we downed the interface. We handle IFF_UP ourselves 9743 * according to user attempts to set it, rather than blindly 9744 * setting it. 9745 */ 9746 9747 ret = 0; 9748 if ((old_flags ^ flags) & IFF_UP) { 9749 if (old_flags & IFF_UP) 9750 __dev_close(dev); 9751 else 9752 ret = __dev_open(dev, extack); 9753 } 9754 9755 if ((flags ^ dev->gflags) & IFF_PROMISC) { 9756 int inc = (flags & IFF_PROMISC) ? 1 : -1; 9757 old_flags = dev->flags; 9758 9759 dev->gflags ^= IFF_PROMISC; 9760 9761 if (__dev_set_promiscuity(dev, inc, false) >= 0) 9762 if (dev->flags != old_flags) 9763 dev_set_rx_mode(dev); 9764 } 9765 9766 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI 9767 * is important. Some (broken) drivers set IFF_PROMISC, when 9768 * IFF_ALLMULTI is requested not asking us and not reporting. 9769 */ 9770 if ((flags ^ dev->gflags) & IFF_ALLMULTI) { 9771 int inc = (flags & IFF_ALLMULTI) ? 1 : -1; 9772 9773 dev->gflags ^= IFF_ALLMULTI; 9774 netif_set_allmulti(dev, inc, false); 9775 } 9776 9777 return ret; 9778 } 9779 9780 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags, 9781 unsigned int gchanges, u32 portid, 9782 const struct nlmsghdr *nlh) 9783 { 9784 unsigned int changes = dev->flags ^ old_flags; 9785 9786 if (gchanges) 9787 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh); 9788 9789 if (changes & IFF_UP) { 9790 if (dev->flags & IFF_UP) 9791 call_netdevice_notifiers(NETDEV_UP, dev); 9792 else 9793 call_netdevice_notifiers(NETDEV_DOWN, dev); 9794 } 9795 9796 if (dev->flags & IFF_UP && 9797 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) { 9798 struct netdev_notifier_change_info change_info = { 9799 .info = { 9800 .dev = dev, 9801 }, 9802 .flags_changed = changes, 9803 }; 9804 9805 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info); 9806 } 9807 } 9808 9809 int netif_change_flags(struct net_device *dev, unsigned int flags, 9810 struct netlink_ext_ack *extack) 9811 { 9812 int ret; 9813 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags; 9814 9815 ret = __dev_change_flags(dev, flags, extack); 9816 if (ret < 0) 9817 return ret; 9818 9819 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags); 9820 __dev_notify_flags(dev, old_flags, changes, 0, NULL); 9821 return ret; 9822 } 9823 9824 int __netif_set_mtu(struct net_device *dev, int new_mtu) 9825 { 9826 const struct net_device_ops *ops = dev->netdev_ops; 9827 9828 if (ops->ndo_change_mtu) 9829 return ops->ndo_change_mtu(dev, new_mtu); 9830 9831 /* Pairs with all the lockless reads of dev->mtu in the stack */ 9832 WRITE_ONCE(dev->mtu, new_mtu); 9833 return 0; 9834 } 9835 EXPORT_SYMBOL_NS_GPL(__netif_set_mtu, "NETDEV_INTERNAL"); 9836 9837 int dev_validate_mtu(struct net_device *dev, int new_mtu, 9838 struct netlink_ext_ack *extack) 9839 { 9840 /* MTU must be positive, and in range */ 9841 if (new_mtu < 0 || new_mtu < dev->min_mtu) { 9842 NL_SET_ERR_MSG(extack, "mtu less than device minimum"); 9843 return -EINVAL; 9844 } 9845 9846 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) { 9847 NL_SET_ERR_MSG(extack, "mtu greater than device maximum"); 9848 return -EINVAL; 9849 } 9850 return 0; 9851 } 9852 9853 /** 9854 * netif_set_mtu_ext() - Change maximum transfer unit 9855 * @dev: device 9856 * @new_mtu: new transfer unit 9857 * @extack: netlink extended ack 9858 * 9859 * Change the maximum transfer size of the network device. 9860 * 9861 * Return: 0 on success, -errno on failure. 9862 */ 9863 int netif_set_mtu_ext(struct net_device *dev, int new_mtu, 9864 struct netlink_ext_ack *extack) 9865 { 9866 int err, orig_mtu; 9867 9868 netdev_ops_assert_locked(dev); 9869 9870 if (new_mtu == dev->mtu) 9871 return 0; 9872 9873 err = dev_validate_mtu(dev, new_mtu, extack); 9874 if (err) 9875 return err; 9876 9877 if (!netif_device_present(dev)) 9878 return -ENODEV; 9879 9880 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev); 9881 err = notifier_to_errno(err); 9882 if (err) 9883 return err; 9884 9885 orig_mtu = dev->mtu; 9886 err = __netif_set_mtu(dev, new_mtu); 9887 9888 if (!err) { 9889 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9890 orig_mtu); 9891 err = notifier_to_errno(err); 9892 if (err) { 9893 /* setting mtu back and notifying everyone again, 9894 * so that they have a chance to revert changes. 9895 */ 9896 __netif_set_mtu(dev, orig_mtu); 9897 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9898 new_mtu); 9899 } 9900 } 9901 return err; 9902 } 9903 9904 int netif_set_mtu(struct net_device *dev, int new_mtu) 9905 { 9906 struct netlink_ext_ack extack; 9907 int err; 9908 9909 memset(&extack, 0, sizeof(extack)); 9910 err = netif_set_mtu_ext(dev, new_mtu, &extack); 9911 if (err && extack._msg) 9912 net_err_ratelimited("%s: %s\n", dev->name, extack._msg); 9913 return err; 9914 } 9915 EXPORT_SYMBOL(netif_set_mtu); 9916 9917 int netif_change_tx_queue_len(struct net_device *dev, unsigned long new_len) 9918 { 9919 unsigned int orig_len = dev->tx_queue_len; 9920 int res; 9921 9922 if (new_len != (unsigned int)new_len) 9923 return -ERANGE; 9924 9925 if (new_len != orig_len) { 9926 WRITE_ONCE(dev->tx_queue_len, new_len); 9927 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev); 9928 res = notifier_to_errno(res); 9929 if (res) 9930 goto err_rollback; 9931 res = dev_qdisc_change_tx_queue_len(dev); 9932 if (res) 9933 goto err_rollback; 9934 } 9935 9936 return 0; 9937 9938 err_rollback: 9939 netdev_err(dev, "refused to change device tx_queue_len\n"); 9940 WRITE_ONCE(dev->tx_queue_len, orig_len); 9941 return res; 9942 } 9943 9944 void netif_set_group(struct net_device *dev, int new_group) 9945 { 9946 dev->group = new_group; 9947 } 9948 9949 /** 9950 * netif_pre_changeaddr_notify() - Call NETDEV_PRE_CHANGEADDR. 9951 * @dev: device 9952 * @addr: new address 9953 * @extack: netlink extended ack 9954 * 9955 * Return: 0 on success, -errno on failure. 9956 */ 9957 int netif_pre_changeaddr_notify(struct net_device *dev, const char *addr, 9958 struct netlink_ext_ack *extack) 9959 { 9960 struct netdev_notifier_pre_changeaddr_info info = { 9961 .info.dev = dev, 9962 .info.extack = extack, 9963 .dev_addr = addr, 9964 }; 9965 int rc; 9966 9967 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info); 9968 return notifier_to_errno(rc); 9969 } 9970 EXPORT_SYMBOL_NS_GPL(netif_pre_changeaddr_notify, "NETDEV_INTERNAL"); 9971 9972 int netif_set_mac_address(struct net_device *dev, struct sockaddr_storage *ss, 9973 struct netlink_ext_ack *extack) 9974 { 9975 const struct net_device_ops *ops = dev->netdev_ops; 9976 int err; 9977 9978 if (!ops->ndo_set_mac_address) 9979 return -EOPNOTSUPP; 9980 if (ss->ss_family != dev->type) 9981 return -EINVAL; 9982 if (!netif_device_present(dev)) 9983 return -ENODEV; 9984 err = netif_pre_changeaddr_notify(dev, ss->__data, extack); 9985 if (err) 9986 return err; 9987 if (memcmp(dev->dev_addr, ss->__data, dev->addr_len)) { 9988 err = ops->ndo_set_mac_address(dev, ss); 9989 if (err) 9990 return err; 9991 } 9992 dev->addr_assign_type = NET_ADDR_SET; 9993 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev); 9994 add_device_randomness(dev->dev_addr, dev->addr_len); 9995 return 0; 9996 } 9997 9998 DECLARE_RWSEM(dev_addr_sem); 9999 10000 /* "sa" is a true struct sockaddr with limited "sa_data" member. */ 10001 int netif_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name) 10002 { 10003 size_t size = sizeof(sa->sa_data); 10004 struct net_device *dev; 10005 int ret = 0; 10006 10007 down_read(&dev_addr_sem); 10008 rcu_read_lock(); 10009 10010 dev = dev_get_by_name_rcu(net, dev_name); 10011 if (!dev) { 10012 ret = -ENODEV; 10013 goto unlock; 10014 } 10015 if (!dev->addr_len) 10016 memset(sa->sa_data, 0, size); 10017 else 10018 memcpy(sa->sa_data, dev->dev_addr, 10019 min_t(size_t, size, dev->addr_len)); 10020 sa->sa_family = dev->type; 10021 10022 unlock: 10023 rcu_read_unlock(); 10024 up_read(&dev_addr_sem); 10025 return ret; 10026 } 10027 EXPORT_SYMBOL_NS_GPL(netif_get_mac_address, "NETDEV_INTERNAL"); 10028 10029 int netif_change_carrier(struct net_device *dev, bool new_carrier) 10030 { 10031 const struct net_device_ops *ops = dev->netdev_ops; 10032 10033 if (!ops->ndo_change_carrier) 10034 return -EOPNOTSUPP; 10035 if (!netif_device_present(dev)) 10036 return -ENODEV; 10037 return ops->ndo_change_carrier(dev, new_carrier); 10038 } 10039 10040 /** 10041 * dev_get_phys_port_id - Get device physical port ID 10042 * @dev: device 10043 * @ppid: port ID 10044 * 10045 * Get device physical port ID 10046 */ 10047 int dev_get_phys_port_id(struct net_device *dev, 10048 struct netdev_phys_item_id *ppid) 10049 { 10050 const struct net_device_ops *ops = dev->netdev_ops; 10051 10052 if (!ops->ndo_get_phys_port_id) 10053 return -EOPNOTSUPP; 10054 return ops->ndo_get_phys_port_id(dev, ppid); 10055 } 10056 10057 /** 10058 * dev_get_phys_port_name - Get device physical port name 10059 * @dev: device 10060 * @name: port name 10061 * @len: limit of bytes to copy to name 10062 * 10063 * Get device physical port name 10064 */ 10065 int dev_get_phys_port_name(struct net_device *dev, 10066 char *name, size_t len) 10067 { 10068 const struct net_device_ops *ops = dev->netdev_ops; 10069 int err; 10070 10071 if (ops->ndo_get_phys_port_name) { 10072 err = ops->ndo_get_phys_port_name(dev, name, len); 10073 if (err != -EOPNOTSUPP) 10074 return err; 10075 } 10076 return devlink_compat_phys_port_name_get(dev, name, len); 10077 } 10078 10079 /** 10080 * netif_get_port_parent_id() - Get the device's port parent identifier 10081 * @dev: network device 10082 * @ppid: pointer to a storage for the port's parent identifier 10083 * @recurse: allow/disallow recursion to lower devices 10084 * 10085 * Get the devices's port parent identifier. 10086 * 10087 * Return: 0 on success, -errno on failure. 10088 */ 10089 int netif_get_port_parent_id(struct net_device *dev, 10090 struct netdev_phys_item_id *ppid, bool recurse) 10091 { 10092 const struct net_device_ops *ops = dev->netdev_ops; 10093 struct netdev_phys_item_id first = { }; 10094 struct net_device *lower_dev; 10095 struct list_head *iter; 10096 int err; 10097 10098 if (ops->ndo_get_port_parent_id) { 10099 err = ops->ndo_get_port_parent_id(dev, ppid); 10100 if (err != -EOPNOTSUPP) 10101 return err; 10102 } 10103 10104 err = devlink_compat_switch_id_get(dev, ppid); 10105 if (!recurse || err != -EOPNOTSUPP) 10106 return err; 10107 10108 netdev_for_each_lower_dev(dev, lower_dev, iter) { 10109 err = netif_get_port_parent_id(lower_dev, ppid, true); 10110 if (err) 10111 break; 10112 if (!first.id_len) 10113 first = *ppid; 10114 else if (memcmp(&first, ppid, sizeof(*ppid))) 10115 return -EOPNOTSUPP; 10116 } 10117 10118 return err; 10119 } 10120 EXPORT_SYMBOL(netif_get_port_parent_id); 10121 10122 /** 10123 * netdev_port_same_parent_id - Indicate if two network devices have 10124 * the same port parent identifier 10125 * @a: first network device 10126 * @b: second network device 10127 */ 10128 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b) 10129 { 10130 struct netdev_phys_item_id a_id = { }; 10131 struct netdev_phys_item_id b_id = { }; 10132 10133 if (netif_get_port_parent_id(a, &a_id, true) || 10134 netif_get_port_parent_id(b, &b_id, true)) 10135 return false; 10136 10137 return netdev_phys_item_id_same(&a_id, &b_id); 10138 } 10139 EXPORT_SYMBOL(netdev_port_same_parent_id); 10140 10141 int netif_change_proto_down(struct net_device *dev, bool proto_down) 10142 { 10143 if (!dev->change_proto_down) 10144 return -EOPNOTSUPP; 10145 if (!netif_device_present(dev)) 10146 return -ENODEV; 10147 if (proto_down) 10148 netif_carrier_off(dev); 10149 else 10150 netif_carrier_on(dev); 10151 WRITE_ONCE(dev->proto_down, proto_down); 10152 return 0; 10153 } 10154 10155 /** 10156 * netdev_change_proto_down_reason_locked - proto down reason 10157 * 10158 * @dev: device 10159 * @mask: proto down mask 10160 * @value: proto down value 10161 */ 10162 void netdev_change_proto_down_reason_locked(struct net_device *dev, 10163 unsigned long mask, u32 value) 10164 { 10165 u32 proto_down_reason; 10166 int b; 10167 10168 if (!mask) { 10169 proto_down_reason = value; 10170 } else { 10171 proto_down_reason = dev->proto_down_reason; 10172 for_each_set_bit(b, &mask, 32) { 10173 if (value & (1 << b)) 10174 proto_down_reason |= BIT(b); 10175 else 10176 proto_down_reason &= ~BIT(b); 10177 } 10178 } 10179 WRITE_ONCE(dev->proto_down_reason, proto_down_reason); 10180 } 10181 10182 struct bpf_xdp_link { 10183 struct bpf_link link; 10184 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */ 10185 int flags; 10186 }; 10187 10188 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags) 10189 { 10190 if (flags & XDP_FLAGS_HW_MODE) 10191 return XDP_MODE_HW; 10192 if (flags & XDP_FLAGS_DRV_MODE) 10193 return XDP_MODE_DRV; 10194 if (flags & XDP_FLAGS_SKB_MODE) 10195 return XDP_MODE_SKB; 10196 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB; 10197 } 10198 10199 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode) 10200 { 10201 switch (mode) { 10202 case XDP_MODE_SKB: 10203 return generic_xdp_install; 10204 case XDP_MODE_DRV: 10205 case XDP_MODE_HW: 10206 return dev->netdev_ops->ndo_bpf; 10207 default: 10208 return NULL; 10209 } 10210 } 10211 10212 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev, 10213 enum bpf_xdp_mode mode) 10214 { 10215 return dev->xdp_state[mode].link; 10216 } 10217 10218 static struct bpf_prog *dev_xdp_prog(struct net_device *dev, 10219 enum bpf_xdp_mode mode) 10220 { 10221 struct bpf_xdp_link *link = dev_xdp_link(dev, mode); 10222 10223 if (link) 10224 return link->link.prog; 10225 return dev->xdp_state[mode].prog; 10226 } 10227 10228 u8 dev_xdp_prog_count(struct net_device *dev) 10229 { 10230 u8 count = 0; 10231 int i; 10232 10233 for (i = 0; i < __MAX_XDP_MODE; i++) 10234 if (dev->xdp_state[i].prog || dev->xdp_state[i].link) 10235 count++; 10236 return count; 10237 } 10238 EXPORT_SYMBOL_GPL(dev_xdp_prog_count); 10239 10240 u8 dev_xdp_sb_prog_count(struct net_device *dev) 10241 { 10242 u8 count = 0; 10243 int i; 10244 10245 for (i = 0; i < __MAX_XDP_MODE; i++) 10246 if (dev->xdp_state[i].prog && 10247 !dev->xdp_state[i].prog->aux->xdp_has_frags) 10248 count++; 10249 return count; 10250 } 10251 10252 int netif_xdp_propagate(struct net_device *dev, struct netdev_bpf *bpf) 10253 { 10254 if (!dev->netdev_ops->ndo_bpf) 10255 return -EOPNOTSUPP; 10256 10257 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 10258 bpf->command == XDP_SETUP_PROG && 10259 bpf->prog && !bpf->prog->aux->xdp_has_frags) { 10260 NL_SET_ERR_MSG(bpf->extack, 10261 "unable to propagate XDP to device using tcp-data-split"); 10262 return -EBUSY; 10263 } 10264 10265 if (dev_get_min_mp_channel_count(dev)) { 10266 NL_SET_ERR_MSG(bpf->extack, "unable to propagate XDP to device using memory provider"); 10267 return -EBUSY; 10268 } 10269 10270 return dev->netdev_ops->ndo_bpf(dev, bpf); 10271 } 10272 EXPORT_SYMBOL_GPL(netif_xdp_propagate); 10273 10274 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode) 10275 { 10276 struct bpf_prog *prog = dev_xdp_prog(dev, mode); 10277 10278 return prog ? prog->aux->id : 0; 10279 } 10280 10281 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode, 10282 struct bpf_xdp_link *link) 10283 { 10284 dev->xdp_state[mode].link = link; 10285 dev->xdp_state[mode].prog = NULL; 10286 } 10287 10288 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode, 10289 struct bpf_prog *prog) 10290 { 10291 dev->xdp_state[mode].link = NULL; 10292 dev->xdp_state[mode].prog = prog; 10293 } 10294 10295 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode, 10296 bpf_op_t bpf_op, struct netlink_ext_ack *extack, 10297 u32 flags, struct bpf_prog *prog) 10298 { 10299 struct netdev_bpf xdp; 10300 int err; 10301 10302 netdev_ops_assert_locked(dev); 10303 10304 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 10305 prog && !prog->aux->xdp_has_frags) { 10306 NL_SET_ERR_MSG(extack, "unable to install XDP to device using tcp-data-split"); 10307 return -EBUSY; 10308 } 10309 10310 if (dev_get_min_mp_channel_count(dev)) { 10311 NL_SET_ERR_MSG(extack, "unable to install XDP to device using memory provider"); 10312 return -EBUSY; 10313 } 10314 10315 memset(&xdp, 0, sizeof(xdp)); 10316 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG; 10317 xdp.extack = extack; 10318 xdp.flags = flags; 10319 xdp.prog = prog; 10320 10321 /* Drivers assume refcnt is already incremented (i.e, prog pointer is 10322 * "moved" into driver), so they don't increment it on their own, but 10323 * they do decrement refcnt when program is detached or replaced. 10324 * Given net_device also owns link/prog, we need to bump refcnt here 10325 * to prevent drivers from underflowing it. 10326 */ 10327 if (prog) 10328 bpf_prog_inc(prog); 10329 err = bpf_op(dev, &xdp); 10330 if (err) { 10331 if (prog) 10332 bpf_prog_put(prog); 10333 return err; 10334 } 10335 10336 if (mode != XDP_MODE_HW) 10337 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog); 10338 10339 return 0; 10340 } 10341 10342 static void dev_xdp_uninstall(struct net_device *dev) 10343 { 10344 struct bpf_xdp_link *link; 10345 struct bpf_prog *prog; 10346 enum bpf_xdp_mode mode; 10347 bpf_op_t bpf_op; 10348 10349 ASSERT_RTNL(); 10350 10351 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) { 10352 prog = dev_xdp_prog(dev, mode); 10353 if (!prog) 10354 continue; 10355 10356 bpf_op = dev_xdp_bpf_op(dev, mode); 10357 if (!bpf_op) 10358 continue; 10359 10360 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 10361 10362 /* auto-detach link from net device */ 10363 link = dev_xdp_link(dev, mode); 10364 if (link) 10365 link->dev = NULL; 10366 else 10367 bpf_prog_put(prog); 10368 10369 dev_xdp_set_link(dev, mode, NULL); 10370 } 10371 } 10372 10373 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack, 10374 struct bpf_xdp_link *link, struct bpf_prog *new_prog, 10375 struct bpf_prog *old_prog, u32 flags) 10376 { 10377 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES); 10378 struct bpf_prog *cur_prog; 10379 struct net_device *upper; 10380 struct list_head *iter; 10381 enum bpf_xdp_mode mode; 10382 bpf_op_t bpf_op; 10383 int err; 10384 10385 ASSERT_RTNL(); 10386 10387 /* either link or prog attachment, never both */ 10388 if (link && (new_prog || old_prog)) 10389 return -EINVAL; 10390 /* link supports only XDP mode flags */ 10391 if (link && (flags & ~XDP_FLAGS_MODES)) { 10392 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment"); 10393 return -EINVAL; 10394 } 10395 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */ 10396 if (num_modes > 1) { 10397 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set"); 10398 return -EINVAL; 10399 } 10400 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */ 10401 if (!num_modes && dev_xdp_prog_count(dev) > 1) { 10402 NL_SET_ERR_MSG(extack, 10403 "More than one program loaded, unset mode is ambiguous"); 10404 return -EINVAL; 10405 } 10406 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */ 10407 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) { 10408 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified"); 10409 return -EINVAL; 10410 } 10411 10412 mode = dev_xdp_mode(dev, flags); 10413 /* can't replace attached link */ 10414 if (dev_xdp_link(dev, mode)) { 10415 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link"); 10416 return -EBUSY; 10417 } 10418 10419 /* don't allow if an upper device already has a program */ 10420 netdev_for_each_upper_dev_rcu(dev, upper, iter) { 10421 if (dev_xdp_prog_count(upper) > 0) { 10422 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program"); 10423 return -EEXIST; 10424 } 10425 } 10426 10427 cur_prog = dev_xdp_prog(dev, mode); 10428 /* can't replace attached prog with link */ 10429 if (link && cur_prog) { 10430 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link"); 10431 return -EBUSY; 10432 } 10433 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) { 10434 NL_SET_ERR_MSG(extack, "Active program does not match expected"); 10435 return -EEXIST; 10436 } 10437 10438 /* put effective new program into new_prog */ 10439 if (link) 10440 new_prog = link->link.prog; 10441 10442 if (new_prog) { 10443 bool offload = mode == XDP_MODE_HW; 10444 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB 10445 ? XDP_MODE_DRV : XDP_MODE_SKB; 10446 10447 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) { 10448 NL_SET_ERR_MSG(extack, "XDP program already attached"); 10449 return -EBUSY; 10450 } 10451 if (!offload && dev_xdp_prog(dev, other_mode)) { 10452 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time"); 10453 return -EEXIST; 10454 } 10455 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) { 10456 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported"); 10457 return -EINVAL; 10458 } 10459 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) { 10460 NL_SET_ERR_MSG(extack, "Program bound to different device"); 10461 return -EINVAL; 10462 } 10463 if (bpf_prog_is_dev_bound(new_prog->aux) && mode == XDP_MODE_SKB) { 10464 NL_SET_ERR_MSG(extack, "Can't attach device-bound programs in generic mode"); 10465 return -EINVAL; 10466 } 10467 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) { 10468 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device"); 10469 return -EINVAL; 10470 } 10471 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) { 10472 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device"); 10473 return -EINVAL; 10474 } 10475 } 10476 10477 /* don't call drivers if the effective program didn't change */ 10478 if (new_prog != cur_prog) { 10479 bpf_op = dev_xdp_bpf_op(dev, mode); 10480 if (!bpf_op) { 10481 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode"); 10482 return -EOPNOTSUPP; 10483 } 10484 10485 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog); 10486 if (err) 10487 return err; 10488 } 10489 10490 if (link) 10491 dev_xdp_set_link(dev, mode, link); 10492 else 10493 dev_xdp_set_prog(dev, mode, new_prog); 10494 if (cur_prog) 10495 bpf_prog_put(cur_prog); 10496 10497 return 0; 10498 } 10499 10500 static int dev_xdp_attach_link(struct net_device *dev, 10501 struct netlink_ext_ack *extack, 10502 struct bpf_xdp_link *link) 10503 { 10504 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags); 10505 } 10506 10507 static int dev_xdp_detach_link(struct net_device *dev, 10508 struct netlink_ext_ack *extack, 10509 struct bpf_xdp_link *link) 10510 { 10511 enum bpf_xdp_mode mode; 10512 bpf_op_t bpf_op; 10513 10514 ASSERT_RTNL(); 10515 10516 mode = dev_xdp_mode(dev, link->flags); 10517 if (dev_xdp_link(dev, mode) != link) 10518 return -EINVAL; 10519 10520 bpf_op = dev_xdp_bpf_op(dev, mode); 10521 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 10522 dev_xdp_set_link(dev, mode, NULL); 10523 return 0; 10524 } 10525 10526 static void bpf_xdp_link_release(struct bpf_link *link) 10527 { 10528 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10529 10530 rtnl_lock(); 10531 10532 /* if racing with net_device's tear down, xdp_link->dev might be 10533 * already NULL, in which case link was already auto-detached 10534 */ 10535 if (xdp_link->dev) { 10536 netdev_lock_ops(xdp_link->dev); 10537 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link)); 10538 netdev_unlock_ops(xdp_link->dev); 10539 xdp_link->dev = NULL; 10540 } 10541 10542 rtnl_unlock(); 10543 } 10544 10545 static int bpf_xdp_link_detach(struct bpf_link *link) 10546 { 10547 bpf_xdp_link_release(link); 10548 return 0; 10549 } 10550 10551 static void bpf_xdp_link_dealloc(struct bpf_link *link) 10552 { 10553 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10554 10555 kfree(xdp_link); 10556 } 10557 10558 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link, 10559 struct seq_file *seq) 10560 { 10561 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10562 u32 ifindex = 0; 10563 10564 rtnl_lock(); 10565 if (xdp_link->dev) 10566 ifindex = xdp_link->dev->ifindex; 10567 rtnl_unlock(); 10568 10569 seq_printf(seq, "ifindex:\t%u\n", ifindex); 10570 } 10571 10572 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link, 10573 struct bpf_link_info *info) 10574 { 10575 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10576 u32 ifindex = 0; 10577 10578 rtnl_lock(); 10579 if (xdp_link->dev) 10580 ifindex = xdp_link->dev->ifindex; 10581 rtnl_unlock(); 10582 10583 info->xdp.ifindex = ifindex; 10584 return 0; 10585 } 10586 10587 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog, 10588 struct bpf_prog *old_prog) 10589 { 10590 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10591 enum bpf_xdp_mode mode; 10592 bpf_op_t bpf_op; 10593 int err = 0; 10594 10595 rtnl_lock(); 10596 10597 /* link might have been auto-released already, so fail */ 10598 if (!xdp_link->dev) { 10599 err = -ENOLINK; 10600 goto out_unlock; 10601 } 10602 10603 if (old_prog && link->prog != old_prog) { 10604 err = -EPERM; 10605 goto out_unlock; 10606 } 10607 old_prog = link->prog; 10608 if (old_prog->type != new_prog->type || 10609 old_prog->expected_attach_type != new_prog->expected_attach_type) { 10610 err = -EINVAL; 10611 goto out_unlock; 10612 } 10613 10614 if (old_prog == new_prog) { 10615 /* no-op, don't disturb drivers */ 10616 bpf_prog_put(new_prog); 10617 goto out_unlock; 10618 } 10619 10620 netdev_lock_ops(xdp_link->dev); 10621 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags); 10622 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode); 10623 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL, 10624 xdp_link->flags, new_prog); 10625 netdev_unlock_ops(xdp_link->dev); 10626 if (err) 10627 goto out_unlock; 10628 10629 old_prog = xchg(&link->prog, new_prog); 10630 bpf_prog_put(old_prog); 10631 10632 out_unlock: 10633 rtnl_unlock(); 10634 return err; 10635 } 10636 10637 static const struct bpf_link_ops bpf_xdp_link_lops = { 10638 .release = bpf_xdp_link_release, 10639 .dealloc = bpf_xdp_link_dealloc, 10640 .detach = bpf_xdp_link_detach, 10641 .show_fdinfo = bpf_xdp_link_show_fdinfo, 10642 .fill_link_info = bpf_xdp_link_fill_link_info, 10643 .update_prog = bpf_xdp_link_update, 10644 }; 10645 10646 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog) 10647 { 10648 struct net *net = current->nsproxy->net_ns; 10649 struct bpf_link_primer link_primer; 10650 struct netlink_ext_ack extack = {}; 10651 struct bpf_xdp_link *link; 10652 struct net_device *dev; 10653 int err, fd; 10654 10655 rtnl_lock(); 10656 dev = dev_get_by_index(net, attr->link_create.target_ifindex); 10657 if (!dev) { 10658 rtnl_unlock(); 10659 return -EINVAL; 10660 } 10661 10662 link = kzalloc_obj(*link, GFP_USER); 10663 if (!link) { 10664 err = -ENOMEM; 10665 goto unlock; 10666 } 10667 10668 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog, 10669 attr->link_create.attach_type); 10670 link->dev = dev; 10671 link->flags = attr->link_create.flags; 10672 10673 err = bpf_link_prime(&link->link, &link_primer); 10674 if (err) { 10675 kfree(link); 10676 goto unlock; 10677 } 10678 10679 netdev_lock_ops(dev); 10680 err = dev_xdp_attach_link(dev, &extack, link); 10681 netdev_unlock_ops(dev); 10682 rtnl_unlock(); 10683 10684 if (err) { 10685 link->dev = NULL; 10686 bpf_link_cleanup(&link_primer); 10687 trace_bpf_xdp_link_attach_failed(extack._msg); 10688 goto out_put_dev; 10689 } 10690 10691 fd = bpf_link_settle(&link_primer); 10692 /* link itself doesn't hold dev's refcnt to not complicate shutdown */ 10693 dev_put(dev); 10694 return fd; 10695 10696 unlock: 10697 rtnl_unlock(); 10698 10699 out_put_dev: 10700 dev_put(dev); 10701 return err; 10702 } 10703 10704 /** 10705 * dev_change_xdp_fd - set or clear a bpf program for a device rx path 10706 * @dev: device 10707 * @extack: netlink extended ack 10708 * @fd: new program fd or negative value to clear 10709 * @expected_fd: old program fd that userspace expects to replace or clear 10710 * @flags: xdp-related flags 10711 * 10712 * Set or clear a bpf program for a device 10713 */ 10714 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, 10715 int fd, int expected_fd, u32 flags) 10716 { 10717 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags); 10718 struct bpf_prog *new_prog = NULL, *old_prog = NULL; 10719 int err; 10720 10721 ASSERT_RTNL(); 10722 10723 if (fd >= 0) { 10724 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP, 10725 mode != XDP_MODE_SKB); 10726 if (IS_ERR(new_prog)) 10727 return PTR_ERR(new_prog); 10728 } 10729 10730 if (expected_fd >= 0) { 10731 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP, 10732 mode != XDP_MODE_SKB); 10733 if (IS_ERR(old_prog)) { 10734 err = PTR_ERR(old_prog); 10735 old_prog = NULL; 10736 goto err_out; 10737 } 10738 } 10739 10740 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags); 10741 10742 err_out: 10743 if (err && new_prog) 10744 bpf_prog_put(new_prog); 10745 if (old_prog) 10746 bpf_prog_put(old_prog); 10747 return err; 10748 } 10749 10750 u32 dev_get_min_mp_channel_count(const struct net_device *dev) 10751 { 10752 int i; 10753 10754 netdev_ops_assert_locked(dev); 10755 10756 for (i = dev->real_num_rx_queues - 1; i >= 0; i--) 10757 if (dev->_rx[i].mp_params.mp_priv) 10758 /* The channel count is the idx plus 1. */ 10759 return i + 1; 10760 10761 return 0; 10762 } 10763 10764 /** 10765 * dev_index_reserve() - allocate an ifindex in a namespace 10766 * @net: the applicable net namespace 10767 * @ifindex: requested ifindex, pass %0 to get one allocated 10768 * 10769 * Allocate a ifindex for a new device. Caller must either use the ifindex 10770 * to store the device (via list_netdevice()) or call dev_index_release() 10771 * to give the index up. 10772 * 10773 * Return: a suitable unique value for a new device interface number or -errno. 10774 */ 10775 static int dev_index_reserve(struct net *net, u32 ifindex) 10776 { 10777 int err; 10778 10779 if (ifindex > INT_MAX) { 10780 DEBUG_NET_WARN_ON_ONCE(1); 10781 return -EINVAL; 10782 } 10783 10784 if (!ifindex) 10785 err = xa_alloc_cyclic(&net->dev_by_index, &ifindex, NULL, 10786 xa_limit_31b, &net->ifindex, GFP_KERNEL); 10787 else 10788 err = xa_insert(&net->dev_by_index, ifindex, NULL, GFP_KERNEL); 10789 if (err < 0) 10790 return err; 10791 10792 return ifindex; 10793 } 10794 10795 static void dev_index_release(struct net *net, int ifindex) 10796 { 10797 /* Expect only unused indexes, unlist_netdevice() removes the used */ 10798 WARN_ON(xa_erase(&net->dev_by_index, ifindex)); 10799 } 10800 10801 static bool from_cleanup_net(void) 10802 { 10803 #ifdef CONFIG_NET_NS 10804 return current == READ_ONCE(cleanup_net_task); 10805 #else 10806 return false; 10807 #endif 10808 } 10809 10810 /* Delayed registration/unregisteration */ 10811 LIST_HEAD(net_todo_list); 10812 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq); 10813 atomic_t dev_unreg_count = ATOMIC_INIT(0); 10814 10815 static void net_set_todo(struct net_device *dev) 10816 { 10817 list_add_tail(&dev->todo_list, &net_todo_list); 10818 } 10819 10820 static netdev_features_t netdev_sync_upper_features(struct net_device *lower, 10821 struct net_device *upper, netdev_features_t features) 10822 { 10823 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10824 netdev_features_t feature; 10825 int feature_bit; 10826 10827 for_each_netdev_feature(upper_disables, feature_bit) { 10828 feature = __NETIF_F_BIT(feature_bit); 10829 if (!(upper->wanted_features & feature) 10830 && (features & feature)) { 10831 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n", 10832 &feature, upper->name); 10833 features &= ~feature; 10834 } 10835 } 10836 10837 return features; 10838 } 10839 10840 static void netdev_sync_lower_features(struct net_device *upper, 10841 struct net_device *lower, netdev_features_t features) 10842 { 10843 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10844 netdev_features_t feature; 10845 int feature_bit; 10846 10847 for_each_netdev_feature(upper_disables, feature_bit) { 10848 feature = __NETIF_F_BIT(feature_bit); 10849 if (!(features & feature) && (lower->features & feature)) { 10850 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n", 10851 &feature, lower->name); 10852 netdev_lock_ops(lower); 10853 lower->wanted_features &= ~feature; 10854 __netdev_update_features(lower); 10855 10856 if (unlikely(lower->features & feature)) 10857 netdev_WARN(upper, "failed to disable %pNF on %s!\n", 10858 &feature, lower->name); 10859 else 10860 netdev_features_change(lower); 10861 netdev_unlock_ops(lower); 10862 } 10863 } 10864 } 10865 10866 static bool netdev_has_ip_or_hw_csum(netdev_features_t features) 10867 { 10868 netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 10869 bool ip_csum = (features & ip_csum_mask) == ip_csum_mask; 10870 bool hw_csum = features & NETIF_F_HW_CSUM; 10871 10872 return ip_csum || hw_csum; 10873 } 10874 10875 static netdev_features_t netdev_fix_features(struct net_device *dev, 10876 netdev_features_t features) 10877 { 10878 /* Fix illegal checksum combinations */ 10879 if ((features & NETIF_F_HW_CSUM) && 10880 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) { 10881 netdev_warn(dev, "mixed HW and IP checksum settings.\n"); 10882 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM); 10883 } 10884 10885 /* TSO requires that SG is present as well. */ 10886 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) { 10887 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n"); 10888 features &= ~NETIF_F_ALL_TSO; 10889 } 10890 10891 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) && 10892 !(features & NETIF_F_IP_CSUM)) { 10893 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n"); 10894 features &= ~NETIF_F_TSO; 10895 features &= ~NETIF_F_TSO_ECN; 10896 } 10897 10898 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) && 10899 !(features & NETIF_F_IPV6_CSUM)) { 10900 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n"); 10901 features &= ~NETIF_F_TSO6; 10902 } 10903 10904 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */ 10905 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO)) 10906 features &= ~NETIF_F_TSO_MANGLEID; 10907 10908 /* TSO ECN requires that TSO is present as well. */ 10909 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN) 10910 features &= ~NETIF_F_TSO_ECN; 10911 10912 /* Software GSO depends on SG. */ 10913 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) { 10914 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n"); 10915 features &= ~NETIF_F_GSO; 10916 } 10917 10918 /* GSO partial features require GSO partial be set */ 10919 if ((features & dev->gso_partial_features) && 10920 !(features & NETIF_F_GSO_PARTIAL)) { 10921 netdev_dbg(dev, 10922 "Dropping partially supported GSO features since no GSO partial.\n"); 10923 features &= ~dev->gso_partial_features; 10924 } 10925 10926 if (!(features & NETIF_F_RXCSUM)) { 10927 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet 10928 * successfully merged by hardware must also have the 10929 * checksum verified by hardware. If the user does not 10930 * want to enable RXCSUM, logically, we should disable GRO_HW. 10931 */ 10932 if (features & NETIF_F_GRO_HW) { 10933 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n"); 10934 features &= ~NETIF_F_GRO_HW; 10935 } 10936 } 10937 10938 /* LRO/HW-GRO features cannot be combined with RX-FCS */ 10939 if (features & NETIF_F_RXFCS) { 10940 if (features & NETIF_F_LRO) { 10941 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n"); 10942 features &= ~NETIF_F_LRO; 10943 } 10944 10945 if (features & NETIF_F_GRO_HW) { 10946 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n"); 10947 features &= ~NETIF_F_GRO_HW; 10948 } 10949 } 10950 10951 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) { 10952 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n"); 10953 features &= ~NETIF_F_LRO; 10954 } 10955 10956 if ((features & NETIF_F_HW_TLS_TX) && !netdev_has_ip_or_hw_csum(features)) { 10957 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n"); 10958 features &= ~NETIF_F_HW_TLS_TX; 10959 } 10960 10961 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) { 10962 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n"); 10963 features &= ~NETIF_F_HW_TLS_RX; 10964 } 10965 10966 if ((features & NETIF_F_GSO_UDP_L4) && !netdev_has_ip_or_hw_csum(features)) { 10967 netdev_dbg(dev, "Dropping USO feature since no CSUM feature.\n"); 10968 features &= ~NETIF_F_GSO_UDP_L4; 10969 } 10970 10971 return features; 10972 } 10973 10974 int __netdev_update_features(struct net_device *dev) 10975 { 10976 struct net_device *upper, *lower; 10977 netdev_features_t features; 10978 struct list_head *iter; 10979 int err = -1; 10980 10981 ASSERT_RTNL(); 10982 netdev_ops_assert_locked(dev); 10983 10984 features = netdev_get_wanted_features(dev); 10985 10986 if (dev->netdev_ops->ndo_fix_features) 10987 features = dev->netdev_ops->ndo_fix_features(dev, features); 10988 10989 /* driver might be less strict about feature dependencies */ 10990 features = netdev_fix_features(dev, features); 10991 10992 /* some features can't be enabled if they're off on an upper device */ 10993 netdev_for_each_upper_dev_rcu(dev, upper, iter) 10994 features = netdev_sync_upper_features(dev, upper, features); 10995 10996 if (dev->features == features) 10997 goto sync_lower; 10998 10999 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n", 11000 &dev->features, &features); 11001 11002 if (dev->netdev_ops->ndo_set_features) 11003 err = dev->netdev_ops->ndo_set_features(dev, features); 11004 else 11005 err = 0; 11006 11007 if (unlikely(err < 0)) { 11008 netdev_err(dev, 11009 "set_features() failed (%d); wanted %pNF, left %pNF\n", 11010 err, &features, &dev->features); 11011 /* return non-0 since some features might have changed and 11012 * it's better to fire a spurious notification than miss it 11013 */ 11014 return -1; 11015 } 11016 11017 sync_lower: 11018 /* some features must be disabled on lower devices when disabled 11019 * on an upper device (think: bonding master or bridge) 11020 */ 11021 netdev_for_each_lower_dev(dev, lower, iter) 11022 netdev_sync_lower_features(dev, lower, features); 11023 11024 if (!err) { 11025 netdev_features_t diff = features ^ dev->features; 11026 11027 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) { 11028 /* udp_tunnel_{get,drop}_rx_info both need 11029 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the 11030 * device, or they won't do anything. 11031 * Thus we need to update dev->features 11032 * *before* calling udp_tunnel_get_rx_info, 11033 * but *after* calling udp_tunnel_drop_rx_info. 11034 */ 11035 udp_tunnel_nic_lock(dev); 11036 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) { 11037 dev->features = features; 11038 udp_tunnel_get_rx_info(dev); 11039 } else { 11040 udp_tunnel_drop_rx_info(dev); 11041 } 11042 udp_tunnel_nic_unlock(dev); 11043 } 11044 11045 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) { 11046 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) { 11047 dev->features = features; 11048 err |= vlan_get_rx_ctag_filter_info(dev); 11049 } else { 11050 vlan_drop_rx_ctag_filter_info(dev); 11051 } 11052 } 11053 11054 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) { 11055 if (features & NETIF_F_HW_VLAN_STAG_FILTER) { 11056 dev->features = features; 11057 err |= vlan_get_rx_stag_filter_info(dev); 11058 } else { 11059 vlan_drop_rx_stag_filter_info(dev); 11060 } 11061 } 11062 11063 dev->features = features; 11064 } 11065 11066 return err < 0 ? 0 : 1; 11067 } 11068 11069 /** 11070 * netdev_update_features - recalculate device features 11071 * @dev: the device to check 11072 * 11073 * Recalculate dev->features set and send notifications if it 11074 * has changed. Should be called after driver or hardware dependent 11075 * conditions might have changed that influence the features. 11076 */ 11077 void netdev_update_features(struct net_device *dev) 11078 { 11079 if (__netdev_update_features(dev)) 11080 netdev_features_change(dev); 11081 } 11082 EXPORT_SYMBOL(netdev_update_features); 11083 11084 /** 11085 * netdev_change_features - recalculate device features 11086 * @dev: the device to check 11087 * 11088 * Recalculate dev->features set and send notifications even 11089 * if they have not changed. Should be called instead of 11090 * netdev_update_features() if also dev->vlan_features might 11091 * have changed to allow the changes to be propagated to stacked 11092 * VLAN devices. 11093 */ 11094 void netdev_change_features(struct net_device *dev) 11095 { 11096 __netdev_update_features(dev); 11097 netdev_features_change(dev); 11098 } 11099 EXPORT_SYMBOL(netdev_change_features); 11100 11101 /** 11102 * netif_stacked_transfer_operstate - transfer operstate 11103 * @rootdev: the root or lower level device to transfer state from 11104 * @dev: the device to transfer operstate to 11105 * 11106 * Transfer operational state from root to device. This is normally 11107 * called when a stacking relationship exists between the root 11108 * device and the device(a leaf device). 11109 */ 11110 void netif_stacked_transfer_operstate(const struct net_device *rootdev, 11111 struct net_device *dev) 11112 { 11113 if (rootdev->operstate == IF_OPER_DORMANT) 11114 netif_dormant_on(dev); 11115 else 11116 netif_dormant_off(dev); 11117 11118 if (rootdev->operstate == IF_OPER_TESTING) 11119 netif_testing_on(dev); 11120 else 11121 netif_testing_off(dev); 11122 11123 if (netif_carrier_ok(rootdev)) 11124 netif_carrier_on(dev); 11125 else 11126 netif_carrier_off(dev); 11127 } 11128 EXPORT_SYMBOL(netif_stacked_transfer_operstate); 11129 11130 static int netif_alloc_rx_queues(struct net_device *dev) 11131 { 11132 unsigned int i, count = dev->num_rx_queues; 11133 struct netdev_rx_queue *rx; 11134 size_t sz = count * sizeof(*rx); 11135 int err = 0; 11136 11137 BUG_ON(count < 1); 11138 11139 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11140 if (!rx) 11141 return -ENOMEM; 11142 11143 dev->_rx = rx; 11144 11145 for (i = 0; i < count; i++) { 11146 rx[i].dev = dev; 11147 11148 /* XDP RX-queue setup */ 11149 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0); 11150 if (err < 0) 11151 goto err_rxq_info; 11152 } 11153 return 0; 11154 11155 err_rxq_info: 11156 /* Rollback successful reg's and free other resources */ 11157 while (i--) 11158 xdp_rxq_info_unreg(&rx[i].xdp_rxq); 11159 kvfree(dev->_rx); 11160 dev->_rx = NULL; 11161 return err; 11162 } 11163 11164 static void netif_free_rx_queues(struct net_device *dev) 11165 { 11166 unsigned int i, count = dev->num_rx_queues; 11167 11168 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */ 11169 if (!dev->_rx) 11170 return; 11171 11172 for (i = 0; i < count; i++) 11173 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq); 11174 11175 kvfree(dev->_rx); 11176 } 11177 11178 static void netdev_init_one_queue(struct net_device *dev, 11179 struct netdev_queue *queue, void *_unused) 11180 { 11181 /* Initialize queue lock */ 11182 spin_lock_init(&queue->_xmit_lock); 11183 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type); 11184 queue->xmit_lock_owner = -1; 11185 netdev_queue_numa_node_write(queue, NUMA_NO_NODE); 11186 queue->dev = dev; 11187 #ifdef CONFIG_BQL 11188 dql_init(&queue->dql, HZ); 11189 #endif 11190 } 11191 11192 static void netif_free_tx_queues(struct net_device *dev) 11193 { 11194 kvfree(dev->_tx); 11195 } 11196 11197 static int netif_alloc_netdev_queues(struct net_device *dev) 11198 { 11199 unsigned int count = dev->num_tx_queues; 11200 struct netdev_queue *tx; 11201 size_t sz = count * sizeof(*tx); 11202 11203 if (count < 1 || count > 0xffff) 11204 return -EINVAL; 11205 11206 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11207 if (!tx) 11208 return -ENOMEM; 11209 11210 dev->_tx = tx; 11211 11212 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL); 11213 spin_lock_init(&dev->tx_global_lock); 11214 11215 return 0; 11216 } 11217 11218 void netif_tx_stop_all_queues(struct net_device *dev) 11219 { 11220 unsigned int i; 11221 11222 for (i = 0; i < dev->num_tx_queues; i++) { 11223 struct netdev_queue *txq = netdev_get_tx_queue(dev, i); 11224 11225 netif_tx_stop_queue(txq); 11226 } 11227 } 11228 EXPORT_SYMBOL(netif_tx_stop_all_queues); 11229 11230 static int netdev_do_alloc_pcpu_stats(struct net_device *dev) 11231 { 11232 void __percpu *v; 11233 11234 /* Drivers implementing ndo_get_peer_dev must support tstat 11235 * accounting, so that skb_do_redirect() can bump the dev's 11236 * RX stats upon network namespace switch. 11237 */ 11238 if (dev->netdev_ops->ndo_get_peer_dev && 11239 dev->pcpu_stat_type != NETDEV_PCPU_STAT_TSTATS) 11240 return -EOPNOTSUPP; 11241 11242 switch (dev->pcpu_stat_type) { 11243 case NETDEV_PCPU_STAT_NONE: 11244 return 0; 11245 case NETDEV_PCPU_STAT_LSTATS: 11246 v = dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats); 11247 break; 11248 case NETDEV_PCPU_STAT_TSTATS: 11249 v = dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats); 11250 break; 11251 case NETDEV_PCPU_STAT_DSTATS: 11252 v = dev->dstats = netdev_alloc_pcpu_stats(struct pcpu_dstats); 11253 break; 11254 default: 11255 return -EINVAL; 11256 } 11257 11258 return v ? 0 : -ENOMEM; 11259 } 11260 11261 static void netdev_do_free_pcpu_stats(struct net_device *dev) 11262 { 11263 switch (dev->pcpu_stat_type) { 11264 case NETDEV_PCPU_STAT_NONE: 11265 return; 11266 case NETDEV_PCPU_STAT_LSTATS: 11267 free_percpu(dev->lstats); 11268 break; 11269 case NETDEV_PCPU_STAT_TSTATS: 11270 free_percpu(dev->tstats); 11271 break; 11272 case NETDEV_PCPU_STAT_DSTATS: 11273 free_percpu(dev->dstats); 11274 break; 11275 } 11276 } 11277 11278 static void netdev_free_phy_link_topology(struct net_device *dev) 11279 { 11280 struct phy_link_topology *topo = dev->link_topo; 11281 11282 if (IS_ENABLED(CONFIG_PHYLIB) && topo) { 11283 xa_destroy(&topo->phys); 11284 kfree(topo); 11285 dev->link_topo = NULL; 11286 } 11287 } 11288 11289 /** 11290 * register_netdevice() - register a network device 11291 * @dev: device to register 11292 * 11293 * Take a prepared network device structure and make it externally accessible. 11294 * A %NETDEV_REGISTER message is sent to the netdev notifier chain. 11295 * Callers must hold the rtnl lock - you may want register_netdev() 11296 * instead of this. 11297 */ 11298 int register_netdevice(struct net_device *dev) 11299 { 11300 int ret; 11301 struct net *net = dev_net(dev); 11302 11303 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE < 11304 NETDEV_FEATURE_COUNT); 11305 BUG_ON(dev_boot_phase); 11306 ASSERT_RTNL(); 11307 11308 might_sleep(); 11309 11310 /* When net_device's are persistent, this will be fatal. */ 11311 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED); 11312 BUG_ON(!net); 11313 11314 ret = ethtool_check_ops(dev->ethtool_ops); 11315 if (ret) 11316 return ret; 11317 11318 /* rss ctx ID 0 is reserved for the default context, start from 1 */ 11319 xa_init_flags(&dev->ethtool->rss_ctx, XA_FLAGS_ALLOC1); 11320 mutex_init(&dev->ethtool->rss_lock); 11321 11322 spin_lock_init(&dev->addr_list_lock); 11323 netdev_set_addr_lockdep_class(dev); 11324 11325 ret = dev_get_valid_name(net, dev, dev->name); 11326 if (ret < 0) 11327 goto out; 11328 11329 ret = -ENOMEM; 11330 dev->name_node = netdev_name_node_head_alloc(dev); 11331 if (!dev->name_node) 11332 goto out; 11333 11334 /* Init, if this function is available */ 11335 if (dev->netdev_ops->ndo_init) { 11336 ret = dev->netdev_ops->ndo_init(dev); 11337 if (ret) { 11338 if (ret > 0) 11339 ret = -EIO; 11340 goto err_free_name; 11341 } 11342 } 11343 11344 if (((dev->hw_features | dev->features) & 11345 NETIF_F_HW_VLAN_CTAG_FILTER) && 11346 (!dev->netdev_ops->ndo_vlan_rx_add_vid || 11347 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) { 11348 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n"); 11349 ret = -EINVAL; 11350 goto err_uninit; 11351 } 11352 11353 ret = netdev_do_alloc_pcpu_stats(dev); 11354 if (ret) 11355 goto err_uninit; 11356 11357 ret = dev_index_reserve(net, dev->ifindex); 11358 if (ret < 0) 11359 goto err_free_pcpu; 11360 dev->ifindex = ret; 11361 11362 /* Transfer changeable features to wanted_features and enable 11363 * software offloads (GSO and GRO). 11364 */ 11365 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF); 11366 dev->features |= NETIF_F_SOFT_FEATURES; 11367 11368 if (dev->udp_tunnel_nic_info) { 11369 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT; 11370 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT; 11371 } 11372 11373 dev->wanted_features = dev->features & dev->hw_features; 11374 11375 if (!(dev->flags & IFF_LOOPBACK)) 11376 dev->hw_features |= NETIF_F_NOCACHE_COPY; 11377 11378 /* If IPv4 TCP segmentation offload is supported we should also 11379 * allow the device to enable segmenting the frame with the option 11380 * of ignoring a static IP ID value. This doesn't enable the 11381 * feature itself but allows the user to enable it later. 11382 */ 11383 if (dev->hw_features & NETIF_F_TSO) 11384 dev->hw_features |= NETIF_F_TSO_MANGLEID; 11385 if (dev->vlan_features & NETIF_F_TSO) 11386 dev->vlan_features |= NETIF_F_TSO_MANGLEID; 11387 if (dev->mpls_features & NETIF_F_TSO) 11388 dev->mpls_features |= NETIF_F_TSO_MANGLEID; 11389 if (dev->hw_enc_features & NETIF_F_TSO) 11390 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID; 11391 11392 /* TSO_MANGLEID belongs in mangleid_features by definition */ 11393 dev->mangleid_features |= NETIF_F_TSO_MANGLEID; 11394 11395 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices. 11396 */ 11397 dev->vlan_features |= NETIF_F_HIGHDMA; 11398 11399 /* Make NETIF_F_SG inheritable to tunnel devices. 11400 */ 11401 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL; 11402 11403 /* Make NETIF_F_SG inheritable to MPLS. 11404 */ 11405 dev->mpls_features |= NETIF_F_SG; 11406 11407 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev); 11408 ret = notifier_to_errno(ret); 11409 if (ret) 11410 goto err_ifindex_release; 11411 11412 ret = netdev_register_kobject(dev); 11413 11414 netdev_lock(dev); 11415 WRITE_ONCE(dev->reg_state, ret ? NETREG_UNREGISTERED : NETREG_REGISTERED); 11416 netdev_unlock(dev); 11417 11418 if (ret) 11419 goto err_uninit_notify; 11420 11421 netdev_lock_ops(dev); 11422 __netdev_update_features(dev); 11423 netdev_unlock_ops(dev); 11424 11425 /* 11426 * Default initial state at registry is that the 11427 * device is present. 11428 */ 11429 11430 set_bit(__LINK_STATE_PRESENT, &dev->state); 11431 11432 linkwatch_init_dev(dev); 11433 11434 dev_init_scheduler(dev); 11435 11436 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL); 11437 list_netdevice(dev); 11438 11439 add_device_randomness(dev->dev_addr, dev->addr_len); 11440 11441 /* If the device has permanent device address, driver should 11442 * set dev_addr and also addr_assign_type should be set to 11443 * NET_ADDR_PERM (default value). 11444 */ 11445 if (dev->addr_assign_type == NET_ADDR_PERM) 11446 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len); 11447 11448 /* Notify protocols, that a new device appeared. */ 11449 netdev_lock_ops(dev); 11450 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev); 11451 netdev_unlock_ops(dev); 11452 ret = notifier_to_errno(ret); 11453 if (ret) { 11454 /* Expect explicit free_netdev() on failure */ 11455 dev->needs_free_netdev = false; 11456 unregister_netdevice_queue(dev, NULL); 11457 goto out; 11458 } 11459 /* 11460 * Prevent userspace races by waiting until the network 11461 * device is fully setup before sending notifications. 11462 */ 11463 if (!(dev->rtnl_link_ops && dev->rtnl_link_initializing)) 11464 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 11465 11466 out: 11467 return ret; 11468 11469 err_uninit_notify: 11470 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11471 err_ifindex_release: 11472 dev_index_release(net, dev->ifindex); 11473 err_free_pcpu: 11474 netdev_do_free_pcpu_stats(dev); 11475 err_uninit: 11476 if (dev->netdev_ops->ndo_uninit) 11477 dev->netdev_ops->ndo_uninit(dev); 11478 if (dev->priv_destructor) 11479 dev->priv_destructor(dev); 11480 err_free_name: 11481 netdev_name_node_free(dev->name_node); 11482 goto out; 11483 } 11484 EXPORT_SYMBOL(register_netdevice); 11485 11486 /* Initialize the core of a dummy net device. 11487 * The setup steps dummy netdevs need which normal netdevs get by going 11488 * through register_netdevice(). 11489 */ 11490 static void init_dummy_netdev(struct net_device *dev) 11491 { 11492 /* make sure we BUG if trying to hit standard 11493 * register/unregister code path 11494 */ 11495 dev->reg_state = NETREG_DUMMY; 11496 11497 /* a dummy interface is started by default */ 11498 set_bit(__LINK_STATE_PRESENT, &dev->state); 11499 set_bit(__LINK_STATE_START, &dev->state); 11500 11501 /* Note : We dont allocate pcpu_refcnt for dummy devices, 11502 * because users of this 'device' dont need to change 11503 * its refcount. 11504 */ 11505 } 11506 11507 /** 11508 * register_netdev - register a network device 11509 * @dev: device to register 11510 * 11511 * Take a completed network device structure and add it to the kernel 11512 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier 11513 * chain. 0 is returned on success. A negative errno code is returned 11514 * on a failure to set up the device, or if the name is a duplicate. 11515 * 11516 * This is a wrapper around register_netdevice that takes the rtnl semaphore 11517 * and expands the device name if you passed a format string to 11518 * alloc_netdev. 11519 */ 11520 int register_netdev(struct net_device *dev) 11521 { 11522 struct net *net = dev_net(dev); 11523 int err; 11524 11525 if (rtnl_net_lock_killable(net)) 11526 return -EINTR; 11527 11528 err = register_netdevice(dev); 11529 11530 rtnl_net_unlock(net); 11531 11532 return err; 11533 } 11534 EXPORT_SYMBOL(register_netdev); 11535 11536 int netdev_refcnt_read(const struct net_device *dev) 11537 { 11538 #ifdef CONFIG_PCPU_DEV_REFCNT 11539 int i, refcnt = 0; 11540 11541 for_each_possible_cpu(i) 11542 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i); 11543 return refcnt; 11544 #else 11545 return refcount_read(&dev->dev_refcnt); 11546 #endif 11547 } 11548 EXPORT_SYMBOL(netdev_refcnt_read); 11549 11550 int netdev_unregister_timeout_secs __read_mostly = 10; 11551 11552 #define WAIT_REFS_MIN_MSECS 1 11553 #define WAIT_REFS_MAX_MSECS 250 11554 /** 11555 * netdev_wait_allrefs_any - wait until all references are gone. 11556 * @list: list of net_devices to wait on 11557 * 11558 * This is called when unregistering network devices. 11559 * 11560 * Any protocol or device that holds a reference should register 11561 * for netdevice notification, and cleanup and put back the 11562 * reference if they receive an UNREGISTER event. 11563 * We can get stuck here if buggy protocols don't correctly 11564 * call dev_put. 11565 */ 11566 static struct net_device *netdev_wait_allrefs_any(struct list_head *list) 11567 { 11568 unsigned long rebroadcast_time, warning_time; 11569 struct net_device *dev; 11570 int wait = 0; 11571 11572 rebroadcast_time = warning_time = jiffies; 11573 11574 list_for_each_entry(dev, list, todo_list) 11575 if (netdev_refcnt_read(dev) == 1) 11576 return dev; 11577 11578 while (true) { 11579 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) { 11580 rtnl_lock(); 11581 11582 /* Rebroadcast unregister notification */ 11583 list_for_each_entry(dev, list, todo_list) 11584 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11585 11586 __rtnl_unlock(); 11587 rcu_barrier(); 11588 rtnl_lock(); 11589 11590 list_for_each_entry(dev, list, todo_list) 11591 if (test_bit(__LINK_STATE_LINKWATCH_PENDING, 11592 &dev->state)) { 11593 /* We must not have linkwatch events 11594 * pending on unregister. If this 11595 * happens, we simply run the queue 11596 * unscheduled, resulting in a noop 11597 * for this device. 11598 */ 11599 linkwatch_run_queue(); 11600 break; 11601 } 11602 11603 __rtnl_unlock(); 11604 11605 rebroadcast_time = jiffies; 11606 } 11607 11608 rcu_barrier(); 11609 11610 if (!wait) { 11611 wait = WAIT_REFS_MIN_MSECS; 11612 } else { 11613 msleep(wait); 11614 wait = min(wait << 1, WAIT_REFS_MAX_MSECS); 11615 } 11616 11617 list_for_each_entry(dev, list, todo_list) 11618 if (netdev_refcnt_read(dev) == 1) 11619 return dev; 11620 11621 if (time_after(jiffies, warning_time + 11622 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) { 11623 list_for_each_entry(dev, list, todo_list) { 11624 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n", 11625 dev->name, netdev_refcnt_read(dev)); 11626 ref_tracker_dir_print(&dev->refcnt_tracker, 10); 11627 } 11628 11629 warning_time = jiffies; 11630 } 11631 } 11632 } 11633 11634 /* The sequence is: 11635 * 11636 * rtnl_lock(); 11637 * ... 11638 * register_netdevice(x1); 11639 * register_netdevice(x2); 11640 * ... 11641 * unregister_netdevice(y1); 11642 * unregister_netdevice(y2); 11643 * ... 11644 * rtnl_unlock(); 11645 * free_netdev(y1); 11646 * free_netdev(y2); 11647 * 11648 * We are invoked by rtnl_unlock(). 11649 * This allows us to deal with problems: 11650 * 1) We can delete sysfs objects which invoke hotplug 11651 * without deadlocking with linkwatch via keventd. 11652 * 2) Since we run with the RTNL semaphore not held, we can sleep 11653 * safely in order to wait for the netdev refcnt to drop to zero. 11654 * 11655 * We must not return until all unregister events added during 11656 * the interval the lock was held have been completed. 11657 */ 11658 void netdev_run_todo(void) 11659 { 11660 struct net_device *dev, *tmp; 11661 struct list_head list; 11662 int cnt; 11663 #ifdef CONFIG_LOCKDEP 11664 struct list_head unlink_list; 11665 11666 list_replace_init(&net_unlink_list, &unlink_list); 11667 11668 while (!list_empty(&unlink_list)) { 11669 dev = list_first_entry(&unlink_list, struct net_device, 11670 unlink_list); 11671 list_del_init(&dev->unlink_list); 11672 dev->nested_level = dev->lower_level - 1; 11673 } 11674 #endif 11675 11676 /* Snapshot list, allow later requests */ 11677 list_replace_init(&net_todo_list, &list); 11678 11679 __rtnl_unlock(); 11680 11681 /* Wait for rcu callbacks to finish before next phase */ 11682 if (!list_empty(&list)) 11683 rcu_barrier(); 11684 11685 list_for_each_entry_safe(dev, tmp, &list, todo_list) { 11686 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) { 11687 netdev_WARN(dev, "run_todo but not unregistering\n"); 11688 list_del(&dev->todo_list); 11689 continue; 11690 } 11691 11692 netdev_lock(dev); 11693 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERED); 11694 netdev_unlock(dev); 11695 linkwatch_sync_dev(dev); 11696 } 11697 11698 cnt = 0; 11699 while (!list_empty(&list)) { 11700 dev = netdev_wait_allrefs_any(&list); 11701 list_del(&dev->todo_list); 11702 11703 /* paranoia */ 11704 BUG_ON(netdev_refcnt_read(dev) != 1); 11705 BUG_ON(!list_empty(&dev->ptype_all)); 11706 BUG_ON(!list_empty(&dev->ptype_specific)); 11707 WARN_ON(rcu_access_pointer(dev->ip_ptr)); 11708 WARN_ON(rcu_access_pointer(dev->ip6_ptr)); 11709 11710 netdev_do_free_pcpu_stats(dev); 11711 if (dev->priv_destructor) 11712 dev->priv_destructor(dev); 11713 if (dev->needs_free_netdev) 11714 free_netdev(dev); 11715 11716 cnt++; 11717 11718 /* Free network device */ 11719 kobject_put(&dev->dev.kobj); 11720 } 11721 if (cnt && atomic_sub_and_test(cnt, &dev_unreg_count)) 11722 wake_up(&netdev_unregistering_wq); 11723 } 11724 11725 /* Collate per-cpu network dstats statistics 11726 * 11727 * Read per-cpu network statistics from dev->dstats and populate the related 11728 * fields in @s. 11729 */ 11730 static void dev_fetch_dstats(struct rtnl_link_stats64 *s, 11731 const struct pcpu_dstats __percpu *dstats) 11732 { 11733 int cpu; 11734 11735 for_each_possible_cpu(cpu) { 11736 u64 rx_packets, rx_bytes, rx_drops; 11737 u64 tx_packets, tx_bytes, tx_drops; 11738 const struct pcpu_dstats *stats; 11739 unsigned int start; 11740 11741 stats = per_cpu_ptr(dstats, cpu); 11742 do { 11743 start = u64_stats_fetch_begin(&stats->syncp); 11744 rx_packets = u64_stats_read(&stats->rx_packets); 11745 rx_bytes = u64_stats_read(&stats->rx_bytes); 11746 rx_drops = u64_stats_read(&stats->rx_drops); 11747 tx_packets = u64_stats_read(&stats->tx_packets); 11748 tx_bytes = u64_stats_read(&stats->tx_bytes); 11749 tx_drops = u64_stats_read(&stats->tx_drops); 11750 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11751 11752 s->rx_packets += rx_packets; 11753 s->rx_bytes += rx_bytes; 11754 s->rx_dropped += rx_drops; 11755 s->tx_packets += tx_packets; 11756 s->tx_bytes += tx_bytes; 11757 s->tx_dropped += tx_drops; 11758 } 11759 } 11760 11761 /* ndo_get_stats64 implementation for dtstats-based accounting. 11762 * 11763 * Populate @s from dev->stats and dev->dstats. This is used internally by the 11764 * core for NETDEV_PCPU_STAT_DSTAT-type stats collection. 11765 */ 11766 static void dev_get_dstats64(const struct net_device *dev, 11767 struct rtnl_link_stats64 *s) 11768 { 11769 netdev_stats_to_stats64(s, &dev->stats); 11770 dev_fetch_dstats(s, dev->dstats); 11771 } 11772 11773 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has 11774 * all the same fields in the same order as net_device_stats, with only 11775 * the type differing, but rtnl_link_stats64 may have additional fields 11776 * at the end for newer counters. 11777 */ 11778 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64, 11779 const struct net_device_stats *netdev_stats) 11780 { 11781 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t); 11782 const atomic_long_t *src = (atomic_long_t *)netdev_stats; 11783 u64 *dst = (u64 *)stats64; 11784 11785 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64)); 11786 for (i = 0; i < n; i++) 11787 dst[i] = (unsigned long)atomic_long_read(&src[i]); 11788 /* zero out counters that only exist in rtnl_link_stats64 */ 11789 memset((char *)stats64 + n * sizeof(u64), 0, 11790 sizeof(*stats64) - n * sizeof(u64)); 11791 } 11792 EXPORT_SYMBOL(netdev_stats_to_stats64); 11793 11794 static __cold struct net_device_core_stats __percpu *netdev_core_stats_alloc( 11795 struct net_device *dev) 11796 { 11797 struct net_device_core_stats __percpu *p; 11798 11799 p = alloc_percpu_gfp(struct net_device_core_stats, 11800 GFP_ATOMIC | __GFP_NOWARN); 11801 11802 if (p && cmpxchg(&dev->core_stats, NULL, p)) 11803 free_percpu(p); 11804 11805 /* This READ_ONCE() pairs with the cmpxchg() above */ 11806 return READ_ONCE(dev->core_stats); 11807 } 11808 11809 noinline void netdev_core_stats_inc(struct net_device *dev, u32 offset) 11810 { 11811 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11812 struct net_device_core_stats __percpu *p = READ_ONCE(dev->core_stats); 11813 unsigned long __percpu *field; 11814 11815 if (unlikely(!p)) { 11816 p = netdev_core_stats_alloc(dev); 11817 if (!p) 11818 return; 11819 } 11820 11821 field = (unsigned long __percpu *)((void __percpu *)p + offset); 11822 this_cpu_inc(*field); 11823 } 11824 EXPORT_SYMBOL_GPL(netdev_core_stats_inc); 11825 11826 /** 11827 * dev_get_stats - get network device statistics 11828 * @dev: device to get statistics from 11829 * @storage: place to store stats 11830 * 11831 * Get network statistics from device. Return @storage. 11832 * The device driver may provide its own method by setting 11833 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats; 11834 * otherwise the internal statistics structure is used. 11835 */ 11836 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev, 11837 struct rtnl_link_stats64 *storage) 11838 { 11839 const struct net_device_ops *ops = dev->netdev_ops; 11840 const struct net_device_core_stats __percpu *p; 11841 11842 /* 11843 * IPv{4,6} and udp tunnels share common stat helpers and use 11844 * different stat type (NETDEV_PCPU_STAT_TSTATS vs 11845 * NETDEV_PCPU_STAT_DSTATS). Ensure the accounting is consistent. 11846 */ 11847 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_bytes) != 11848 offsetof(struct pcpu_dstats, rx_bytes)); 11849 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_packets) != 11850 offsetof(struct pcpu_dstats, rx_packets)); 11851 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_bytes) != 11852 offsetof(struct pcpu_dstats, tx_bytes)); 11853 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_packets) != 11854 offsetof(struct pcpu_dstats, tx_packets)); 11855 11856 if (ops->ndo_get_stats64) { 11857 memset(storage, 0, sizeof(*storage)); 11858 ops->ndo_get_stats64(dev, storage); 11859 } else if (ops->ndo_get_stats) { 11860 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev)); 11861 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_TSTATS) { 11862 dev_get_tstats64(dev, storage); 11863 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_DSTATS) { 11864 dev_get_dstats64(dev, storage); 11865 } else { 11866 netdev_stats_to_stats64(storage, &dev->stats); 11867 } 11868 11869 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11870 p = READ_ONCE(dev->core_stats); 11871 if (p) { 11872 const struct net_device_core_stats *core_stats; 11873 int i; 11874 11875 for_each_possible_cpu(i) { 11876 core_stats = per_cpu_ptr(p, i); 11877 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped); 11878 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped); 11879 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler); 11880 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped); 11881 } 11882 } 11883 return storage; 11884 } 11885 EXPORT_SYMBOL(dev_get_stats); 11886 11887 /** 11888 * dev_fetch_sw_netstats - get per-cpu network device statistics 11889 * @s: place to store stats 11890 * @netstats: per-cpu network stats to read from 11891 * 11892 * Read per-cpu network statistics and populate the related fields in @s. 11893 */ 11894 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s, 11895 const struct pcpu_sw_netstats __percpu *netstats) 11896 { 11897 int cpu; 11898 11899 for_each_possible_cpu(cpu) { 11900 u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 11901 const struct pcpu_sw_netstats *stats; 11902 unsigned int start; 11903 11904 stats = per_cpu_ptr(netstats, cpu); 11905 do { 11906 start = u64_stats_fetch_begin(&stats->syncp); 11907 rx_packets = u64_stats_read(&stats->rx_packets); 11908 rx_bytes = u64_stats_read(&stats->rx_bytes); 11909 tx_packets = u64_stats_read(&stats->tx_packets); 11910 tx_bytes = u64_stats_read(&stats->tx_bytes); 11911 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11912 11913 s->rx_packets += rx_packets; 11914 s->rx_bytes += rx_bytes; 11915 s->tx_packets += tx_packets; 11916 s->tx_bytes += tx_bytes; 11917 } 11918 } 11919 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats); 11920 11921 /** 11922 * dev_get_tstats64 - ndo_get_stats64 implementation 11923 * @dev: device to get statistics from 11924 * @s: place to store stats 11925 * 11926 * Populate @s from dev->stats and dev->tstats. Can be used as 11927 * ndo_get_stats64() callback. 11928 */ 11929 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s) 11930 { 11931 netdev_stats_to_stats64(s, &dev->stats); 11932 dev_fetch_sw_netstats(s, dev->tstats); 11933 } 11934 EXPORT_SYMBOL_GPL(dev_get_tstats64); 11935 11936 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev) 11937 { 11938 struct netdev_queue *queue = dev_ingress_queue(dev); 11939 11940 #ifdef CONFIG_NET_CLS_ACT 11941 if (queue) 11942 return queue; 11943 queue = kzalloc_obj(*queue, GFP_KERNEL); 11944 if (!queue) 11945 return NULL; 11946 netdev_init_one_queue(dev, queue, NULL); 11947 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc); 11948 RCU_INIT_POINTER(queue->qdisc_sleeping, &noop_qdisc); 11949 rcu_assign_pointer(dev->ingress_queue, queue); 11950 #endif 11951 return queue; 11952 } 11953 11954 static const struct ethtool_ops default_ethtool_ops; 11955 11956 void netdev_set_default_ethtool_ops(struct net_device *dev, 11957 const struct ethtool_ops *ops) 11958 { 11959 if (dev->ethtool_ops == &default_ethtool_ops) 11960 dev->ethtool_ops = ops; 11961 } 11962 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops); 11963 11964 /** 11965 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default 11966 * @dev: netdev to enable the IRQ coalescing on 11967 * 11968 * Sets a conservative default for SW IRQ coalescing. Users can use 11969 * sysfs attributes to override the default values. 11970 */ 11971 void netdev_sw_irq_coalesce_default_on(struct net_device *dev) 11972 { 11973 WARN_ON(dev->reg_state == NETREG_REGISTERED); 11974 11975 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) { 11976 netdev_set_gro_flush_timeout(dev, 20000); 11977 netdev_set_defer_hard_irqs(dev, 1); 11978 } 11979 } 11980 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on); 11981 11982 /** 11983 * alloc_netdev_mqs - allocate network device 11984 * @sizeof_priv: size of private data to allocate space for 11985 * @name: device name format string 11986 * @name_assign_type: origin of device name 11987 * @setup: callback to initialize device 11988 * @txqs: the number of TX subqueues to allocate 11989 * @rxqs: the number of RX subqueues to allocate 11990 * 11991 * Allocates a struct net_device with private data area for driver use 11992 * and performs basic initialization. Also allocates subqueue structs 11993 * for each queue on the device. 11994 */ 11995 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name, 11996 unsigned char name_assign_type, 11997 void (*setup)(struct net_device *), 11998 unsigned int txqs, unsigned int rxqs) 11999 { 12000 struct net_device *dev; 12001 size_t napi_config_sz; 12002 unsigned int maxqs; 12003 12004 BUG_ON(strlen(name) >= sizeof(dev->name)); 12005 12006 if (txqs < 1) { 12007 pr_err("alloc_netdev: Unable to allocate device with zero queues\n"); 12008 return NULL; 12009 } 12010 12011 if (rxqs < 1) { 12012 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n"); 12013 return NULL; 12014 } 12015 12016 maxqs = max(txqs, rxqs); 12017 12018 dev = kvzalloc_flex(*dev, priv, sizeof_priv, 12019 GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 12020 if (!dev) 12021 return NULL; 12022 12023 dev->priv_len = sizeof_priv; 12024 12025 ref_tracker_dir_init(&dev->refcnt_tracker, 128, "netdev"); 12026 #ifdef CONFIG_PCPU_DEV_REFCNT 12027 dev->pcpu_refcnt = alloc_percpu(int); 12028 if (!dev->pcpu_refcnt) 12029 goto free_dev; 12030 __dev_hold(dev); 12031 #else 12032 refcount_set(&dev->dev_refcnt, 1); 12033 #endif 12034 12035 if (dev_addr_init(dev)) 12036 goto free_pcpu; 12037 12038 dev_mc_init(dev); 12039 dev_uc_init(dev); 12040 12041 dev_net_set(dev, &init_net); 12042 12043 dev->gso_max_size = GSO_LEGACY_MAX_SIZE; 12044 dev->xdp_zc_max_segs = 1; 12045 dev->gso_max_segs = GSO_MAX_SEGS; 12046 dev->gro_max_size = GRO_LEGACY_MAX_SIZE; 12047 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE; 12048 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE; 12049 dev->tso_max_size = TSO_LEGACY_MAX_SIZE; 12050 dev->tso_max_segs = TSO_MAX_SEGS; 12051 dev->upper_level = 1; 12052 dev->lower_level = 1; 12053 #ifdef CONFIG_LOCKDEP 12054 dev->nested_level = 0; 12055 INIT_LIST_HEAD(&dev->unlink_list); 12056 #endif 12057 12058 INIT_LIST_HEAD(&dev->napi_list); 12059 INIT_LIST_HEAD(&dev->unreg_list); 12060 INIT_LIST_HEAD(&dev->close_list); 12061 INIT_LIST_HEAD(&dev->link_watch_list); 12062 INIT_LIST_HEAD(&dev->adj_list.upper); 12063 INIT_LIST_HEAD(&dev->adj_list.lower); 12064 INIT_LIST_HEAD(&dev->ptype_all); 12065 INIT_LIST_HEAD(&dev->ptype_specific); 12066 INIT_LIST_HEAD(&dev->net_notifier_list); 12067 #ifdef CONFIG_NET_SCHED 12068 hash_init(dev->qdisc_hash); 12069 #endif 12070 12071 mutex_init(&dev->lock); 12072 12073 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 12074 setup(dev); 12075 12076 if (!dev->tx_queue_len) { 12077 dev->priv_flags |= IFF_NO_QUEUE; 12078 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN; 12079 } 12080 12081 dev->num_tx_queues = txqs; 12082 dev->real_num_tx_queues = txqs; 12083 if (netif_alloc_netdev_queues(dev)) 12084 goto free_all; 12085 12086 dev->num_rx_queues = rxqs; 12087 dev->real_num_rx_queues = rxqs; 12088 if (netif_alloc_rx_queues(dev)) 12089 goto free_all; 12090 dev->ethtool = kzalloc_obj(*dev->ethtool, GFP_KERNEL_ACCOUNT); 12091 if (!dev->ethtool) 12092 goto free_all; 12093 12094 dev->cfg = kzalloc_obj(*dev->cfg, GFP_KERNEL_ACCOUNT); 12095 if (!dev->cfg) 12096 goto free_all; 12097 dev->cfg_pending = dev->cfg; 12098 12099 dev->num_napi_configs = maxqs; 12100 napi_config_sz = array_size(maxqs, sizeof(*dev->napi_config)); 12101 dev->napi_config = kvzalloc(napi_config_sz, GFP_KERNEL_ACCOUNT); 12102 if (!dev->napi_config) 12103 goto free_all; 12104 12105 strscpy(dev->name, name); 12106 dev->name_assign_type = name_assign_type; 12107 dev->group = INIT_NETDEV_GROUP; 12108 if (!dev->ethtool_ops) 12109 dev->ethtool_ops = &default_ethtool_ops; 12110 12111 nf_hook_netdev_init(dev); 12112 12113 return dev; 12114 12115 free_all: 12116 free_netdev(dev); 12117 return NULL; 12118 12119 free_pcpu: 12120 #ifdef CONFIG_PCPU_DEV_REFCNT 12121 free_percpu(dev->pcpu_refcnt); 12122 free_dev: 12123 #endif 12124 kvfree(dev); 12125 return NULL; 12126 } 12127 EXPORT_SYMBOL(alloc_netdev_mqs); 12128 12129 static void netdev_napi_exit(struct net_device *dev) 12130 { 12131 if (!list_empty(&dev->napi_list)) { 12132 struct napi_struct *p, *n; 12133 12134 netdev_lock(dev); 12135 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list) 12136 __netif_napi_del_locked(p); 12137 netdev_unlock(dev); 12138 12139 synchronize_net(); 12140 } 12141 12142 kvfree(dev->napi_config); 12143 } 12144 12145 /** 12146 * free_netdev - free network device 12147 * @dev: device 12148 * 12149 * This function does the last stage of destroying an allocated device 12150 * interface. The reference to the device object is released. If this 12151 * is the last reference then it will be freed.Must be called in process 12152 * context. 12153 */ 12154 void free_netdev(struct net_device *dev) 12155 { 12156 might_sleep(); 12157 12158 /* When called immediately after register_netdevice() failed the unwind 12159 * handling may still be dismantling the device. Handle that case by 12160 * deferring the free. 12161 */ 12162 if (dev->reg_state == NETREG_UNREGISTERING) { 12163 ASSERT_RTNL(); 12164 dev->needs_free_netdev = true; 12165 return; 12166 } 12167 12168 WARN_ON(dev->cfg != dev->cfg_pending); 12169 kfree(dev->cfg); 12170 kfree(dev->ethtool); 12171 netif_free_tx_queues(dev); 12172 netif_free_rx_queues(dev); 12173 12174 kfree(rcu_dereference_protected(dev->ingress_queue, 1)); 12175 12176 /* Flush device addresses */ 12177 dev_addr_flush(dev); 12178 12179 netdev_napi_exit(dev); 12180 12181 netif_del_cpu_rmap(dev); 12182 12183 ref_tracker_dir_exit(&dev->refcnt_tracker); 12184 #ifdef CONFIG_PCPU_DEV_REFCNT 12185 free_percpu(dev->pcpu_refcnt); 12186 dev->pcpu_refcnt = NULL; 12187 #endif 12188 free_percpu(dev->core_stats); 12189 dev->core_stats = NULL; 12190 free_percpu(dev->xdp_bulkq); 12191 dev->xdp_bulkq = NULL; 12192 12193 netdev_free_phy_link_topology(dev); 12194 12195 mutex_destroy(&dev->lock); 12196 12197 /* Compatibility with error handling in drivers */ 12198 if (dev->reg_state == NETREG_UNINITIALIZED || 12199 dev->reg_state == NETREG_DUMMY) { 12200 kvfree(dev); 12201 return; 12202 } 12203 12204 BUG_ON(dev->reg_state != NETREG_UNREGISTERED); 12205 WRITE_ONCE(dev->reg_state, NETREG_RELEASED); 12206 12207 /* will free via device release */ 12208 put_device(&dev->dev); 12209 } 12210 EXPORT_SYMBOL(free_netdev); 12211 12212 /** 12213 * alloc_netdev_dummy - Allocate and initialize a dummy net device. 12214 * @sizeof_priv: size of private data to allocate space for 12215 * 12216 * Return: the allocated net_device on success, NULL otherwise 12217 */ 12218 struct net_device *alloc_netdev_dummy(int sizeof_priv) 12219 { 12220 return alloc_netdev(sizeof_priv, "dummy#", NET_NAME_UNKNOWN, 12221 init_dummy_netdev); 12222 } 12223 EXPORT_SYMBOL_GPL(alloc_netdev_dummy); 12224 12225 /** 12226 * synchronize_net - Synchronize with packet receive processing 12227 * 12228 * Wait for packets currently being received to be done. 12229 * Does not block later packets from starting. 12230 */ 12231 void synchronize_net(void) 12232 { 12233 might_sleep(); 12234 if (from_cleanup_net() || rtnl_is_locked()) 12235 synchronize_rcu_expedited(); 12236 else 12237 synchronize_rcu(); 12238 } 12239 EXPORT_SYMBOL(synchronize_net); 12240 12241 static void netdev_rss_contexts_free(struct net_device *dev) 12242 { 12243 struct ethtool_rxfh_context *ctx; 12244 unsigned long context; 12245 12246 mutex_lock(&dev->ethtool->rss_lock); 12247 xa_for_each(&dev->ethtool->rss_ctx, context, ctx) { 12248 xa_erase(&dev->ethtool->rss_ctx, context); 12249 dev->ethtool_ops->remove_rxfh_context(dev, ctx, context, NULL); 12250 kfree(ctx); 12251 } 12252 xa_destroy(&dev->ethtool->rss_ctx); 12253 mutex_unlock(&dev->ethtool->rss_lock); 12254 } 12255 12256 /** 12257 * unregister_netdevice_queue - remove device from the kernel 12258 * @dev: device 12259 * @head: list 12260 * 12261 * This function shuts down a device interface and removes it 12262 * from the kernel tables. 12263 * If head not NULL, device is queued to be unregistered later. 12264 * 12265 * Callers must hold the rtnl semaphore. You may want 12266 * unregister_netdev() instead of this. 12267 */ 12268 12269 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head) 12270 { 12271 ASSERT_RTNL(); 12272 12273 if (head) { 12274 list_move_tail(&dev->unreg_list, head); 12275 } else { 12276 LIST_HEAD(single); 12277 12278 list_add(&dev->unreg_list, &single); 12279 unregister_netdevice_many(&single); 12280 } 12281 } 12282 EXPORT_SYMBOL(unregister_netdevice_queue); 12283 12284 static void dev_memory_provider_uninstall(struct net_device *dev) 12285 { 12286 unsigned int i; 12287 12288 for (i = 0; i < dev->real_num_rx_queues; i++) { 12289 struct netdev_rx_queue *rxq = &dev->_rx[i]; 12290 struct pp_memory_provider_params *p = &rxq->mp_params; 12291 12292 if (p->mp_ops && p->mp_ops->uninstall) 12293 p->mp_ops->uninstall(rxq->mp_params.mp_priv, rxq); 12294 } 12295 } 12296 12297 /* devices must be UP and netdev_lock()'d */ 12298 static void netif_close_many_and_unlock(struct list_head *close_head) 12299 { 12300 struct net_device *dev, *tmp; 12301 12302 netif_close_many(close_head, false); 12303 12304 /* ... now unlock them */ 12305 list_for_each_entry_safe(dev, tmp, close_head, close_list) { 12306 netdev_unlock(dev); 12307 list_del_init(&dev->close_list); 12308 } 12309 } 12310 12311 static void netif_close_many_and_unlock_cond(struct list_head *close_head) 12312 { 12313 #ifdef CONFIG_LOCKDEP 12314 /* We can only track up to MAX_LOCK_DEPTH locks per task. 12315 * 12316 * Reserve half the available slots for additional locks possibly 12317 * taken by notifiers and (soft)irqs. 12318 */ 12319 unsigned int limit = MAX_LOCK_DEPTH / 2; 12320 12321 if (lockdep_depth(current) > limit) 12322 netif_close_many_and_unlock(close_head); 12323 #endif 12324 } 12325 12326 void unregister_netdevice_many_notify(struct list_head *head, 12327 u32 portid, const struct nlmsghdr *nlh) 12328 { 12329 struct net_device *dev, *tmp; 12330 LIST_HEAD(close_head); 12331 int cnt = 0; 12332 12333 BUG_ON(dev_boot_phase); 12334 ASSERT_RTNL(); 12335 12336 if (list_empty(head)) 12337 return; 12338 12339 list_for_each_entry_safe(dev, tmp, head, unreg_list) { 12340 /* Some devices call without registering 12341 * for initialization unwind. Remove those 12342 * devices and proceed with the remaining. 12343 */ 12344 if (dev->reg_state == NETREG_UNINITIALIZED) { 12345 pr_debug("unregister_netdevice: device %s/%p never was registered\n", 12346 dev->name, dev); 12347 12348 WARN_ON(1); 12349 list_del(&dev->unreg_list); 12350 continue; 12351 } 12352 dev->dismantle = true; 12353 BUG_ON(dev->reg_state != NETREG_REGISTERED); 12354 } 12355 12356 /* If device is running, close it first. Start with ops locked... */ 12357 list_for_each_entry(dev, head, unreg_list) { 12358 if (!(dev->flags & IFF_UP)) 12359 continue; 12360 if (netdev_need_ops_lock(dev)) { 12361 list_add_tail(&dev->close_list, &close_head); 12362 netdev_lock(dev); 12363 } 12364 netif_close_many_and_unlock_cond(&close_head); 12365 } 12366 netif_close_many_and_unlock(&close_head); 12367 /* ... now go over the rest. */ 12368 list_for_each_entry(dev, head, unreg_list) { 12369 if (!netdev_need_ops_lock(dev)) 12370 list_add_tail(&dev->close_list, &close_head); 12371 } 12372 netif_close_many(&close_head, true); 12373 12374 list_for_each_entry(dev, head, unreg_list) { 12375 /* And unlink it from device chain. */ 12376 unlist_netdevice(dev); 12377 netdev_lock(dev); 12378 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERING); 12379 netdev_unlock(dev); 12380 } 12381 flush_all_backlogs(); 12382 12383 synchronize_net(); 12384 12385 list_for_each_entry(dev, head, unreg_list) { 12386 struct sk_buff *skb = NULL; 12387 12388 /* Shutdown queueing discipline. */ 12389 netdev_lock_ops(dev); 12390 dev_shutdown(dev); 12391 dev_tcx_uninstall(dev); 12392 dev_xdp_uninstall(dev); 12393 dev_memory_provider_uninstall(dev); 12394 netdev_unlock_ops(dev); 12395 bpf_dev_bound_netdev_unregister(dev); 12396 12397 netdev_offload_xstats_disable_all(dev); 12398 12399 /* Notify protocols, that we are about to destroy 12400 * this device. They should clean all the things. 12401 */ 12402 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 12403 12404 if (!(dev->rtnl_link_ops && dev->rtnl_link_initializing)) 12405 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0, 12406 GFP_KERNEL, NULL, 0, 12407 portid, nlh); 12408 12409 /* 12410 * Flush the unicast and multicast chains 12411 */ 12412 dev_uc_flush(dev); 12413 dev_mc_flush(dev); 12414 12415 netdev_name_node_alt_flush(dev); 12416 netdev_name_node_free(dev->name_node); 12417 12418 netdev_rss_contexts_free(dev); 12419 12420 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 12421 12422 if (dev->netdev_ops->ndo_uninit) 12423 dev->netdev_ops->ndo_uninit(dev); 12424 12425 mutex_destroy(&dev->ethtool->rss_lock); 12426 12427 net_shaper_flush_netdev(dev); 12428 12429 if (skb) 12430 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh); 12431 12432 /* Notifier chain MUST detach us all upper devices. */ 12433 WARN_ON(netdev_has_any_upper_dev(dev)); 12434 WARN_ON(netdev_has_any_lower_dev(dev)); 12435 12436 /* Remove entries from kobject tree */ 12437 netdev_unregister_kobject(dev); 12438 #ifdef CONFIG_XPS 12439 /* Remove XPS queueing entries */ 12440 netif_reset_xps_queues_gt(dev, 0); 12441 #endif 12442 } 12443 12444 synchronize_net(); 12445 12446 list_for_each_entry(dev, head, unreg_list) { 12447 netdev_put(dev, &dev->dev_registered_tracker); 12448 net_set_todo(dev); 12449 cnt++; 12450 } 12451 atomic_add(cnt, &dev_unreg_count); 12452 12453 list_del(head); 12454 } 12455 12456 /** 12457 * unregister_netdevice_many - unregister many devices 12458 * @head: list of devices 12459 * 12460 * Note: As most callers use a stack allocated list_head, 12461 * we force a list_del() to make sure stack won't be corrupted later. 12462 */ 12463 void unregister_netdevice_many(struct list_head *head) 12464 { 12465 unregister_netdevice_many_notify(head, 0, NULL); 12466 } 12467 EXPORT_SYMBOL(unregister_netdevice_many); 12468 12469 /** 12470 * unregister_netdev - remove device from the kernel 12471 * @dev: device 12472 * 12473 * This function shuts down a device interface and removes it 12474 * from the kernel tables. 12475 * 12476 * This is just a wrapper for unregister_netdevice that takes 12477 * the rtnl semaphore. In general you want to use this and not 12478 * unregister_netdevice. 12479 */ 12480 void unregister_netdev(struct net_device *dev) 12481 { 12482 rtnl_net_dev_lock(dev); 12483 unregister_netdevice(dev); 12484 rtnl_net_dev_unlock(dev); 12485 } 12486 EXPORT_SYMBOL(unregister_netdev); 12487 12488 int __dev_change_net_namespace(struct net_device *dev, struct net *net, 12489 const char *pat, int new_ifindex, 12490 struct netlink_ext_ack *extack) 12491 { 12492 struct netdev_name_node *name_node; 12493 struct net *net_old = dev_net(dev); 12494 char new_name[IFNAMSIZ] = {}; 12495 int err, new_nsid; 12496 12497 ASSERT_RTNL(); 12498 12499 /* Don't allow namespace local devices to be moved. */ 12500 err = -EINVAL; 12501 if (dev->netns_immutable) { 12502 NL_SET_ERR_MSG(extack, "The interface netns is immutable"); 12503 goto out; 12504 } 12505 12506 /* Ensure the device has been registered */ 12507 if (dev->reg_state != NETREG_REGISTERED) { 12508 NL_SET_ERR_MSG(extack, "The interface isn't registered"); 12509 goto out; 12510 } 12511 12512 /* Get out if there is nothing todo */ 12513 err = 0; 12514 if (net_eq(net_old, net)) 12515 goto out; 12516 12517 /* Pick the destination device name, and ensure 12518 * we can use it in the destination network namespace. 12519 */ 12520 err = -EEXIST; 12521 if (netdev_name_in_use(net, dev->name)) { 12522 /* We get here if we can't use the current device name */ 12523 if (!pat) { 12524 NL_SET_ERR_MSG(extack, 12525 "An interface with the same name exists in the target netns"); 12526 goto out; 12527 } 12528 err = dev_prep_valid_name(net, dev, pat, new_name, EEXIST); 12529 if (err < 0) { 12530 NL_SET_ERR_MSG_FMT(extack, 12531 "Unable to use '%s' for the new interface name in the target netns", 12532 pat); 12533 goto out; 12534 } 12535 } 12536 /* Check that none of the altnames conflicts. */ 12537 err = -EEXIST; 12538 netdev_for_each_altname(dev, name_node) { 12539 if (netdev_name_in_use(net, name_node->name)) { 12540 NL_SET_ERR_MSG_FMT(extack, 12541 "An interface with the altname %s exists in the target netns", 12542 name_node->name); 12543 goto out; 12544 } 12545 } 12546 12547 /* Check that new_ifindex isn't used yet. */ 12548 if (new_ifindex) { 12549 err = dev_index_reserve(net, new_ifindex); 12550 if (err < 0) { 12551 NL_SET_ERR_MSG_FMT(extack, 12552 "The ifindex %d is not available in the target netns", 12553 new_ifindex); 12554 goto out; 12555 } 12556 } else { 12557 /* If there is an ifindex conflict assign a new one */ 12558 err = dev_index_reserve(net, dev->ifindex); 12559 if (err == -EBUSY) 12560 err = dev_index_reserve(net, 0); 12561 if (err < 0) { 12562 NL_SET_ERR_MSG(extack, 12563 "Unable to allocate a new ifindex in the target netns"); 12564 goto out; 12565 } 12566 new_ifindex = err; 12567 } 12568 12569 /* 12570 * And now a mini version of register_netdevice unregister_netdevice. 12571 */ 12572 12573 netdev_lock_ops(dev); 12574 /* If device is running close it first. */ 12575 netif_close(dev); 12576 /* And unlink it from device chain */ 12577 unlist_netdevice(dev); 12578 12579 if (!netdev_need_ops_lock(dev)) 12580 netdev_lock(dev); 12581 dev->moving_ns = true; 12582 netdev_unlock(dev); 12583 12584 synchronize_net(); 12585 12586 /* Shutdown queueing discipline. */ 12587 netdev_lock_ops(dev); 12588 dev_shutdown(dev); 12589 netdev_unlock_ops(dev); 12590 12591 /* Notify protocols, that we are about to destroy 12592 * this device. They should clean all the things. 12593 * 12594 * Note that dev->reg_state stays at NETREG_REGISTERED. 12595 * This is wanted because this way 8021q and macvlan know 12596 * the device is just moving and can keep their slaves up. 12597 */ 12598 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 12599 rcu_barrier(); 12600 12601 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL); 12602 12603 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid, 12604 new_ifindex); 12605 12606 /* 12607 * Flush the unicast and multicast chains 12608 */ 12609 dev_uc_flush(dev); 12610 dev_mc_flush(dev); 12611 12612 /* Send a netdev-removed uevent to the old namespace */ 12613 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE); 12614 netdev_adjacent_del_links(dev); 12615 12616 /* Move per-net netdevice notifiers that are following the netdevice */ 12617 move_netdevice_notifiers_dev_net(dev, net); 12618 12619 /* Actually switch the network namespace */ 12620 netdev_lock(dev); 12621 dev_net_set(dev, net); 12622 netdev_unlock(dev); 12623 dev->ifindex = new_ifindex; 12624 12625 if (new_name[0]) { 12626 /* Rename the netdev to prepared name */ 12627 write_seqlock_bh(&netdev_rename_lock); 12628 strscpy(dev->name, new_name, IFNAMSIZ); 12629 write_sequnlock_bh(&netdev_rename_lock); 12630 } 12631 12632 /* Fixup kobjects */ 12633 dev_set_uevent_suppress(&dev->dev, 1); 12634 err = device_rename(&dev->dev, dev->name); 12635 dev_set_uevent_suppress(&dev->dev, 0); 12636 WARN_ON(err); 12637 12638 /* Send a netdev-add uevent to the new namespace */ 12639 kobject_uevent(&dev->dev.kobj, KOBJ_ADD); 12640 netdev_adjacent_add_links(dev); 12641 12642 /* Adapt owner in case owning user namespace of target network 12643 * namespace is different from the original one. 12644 */ 12645 err = netdev_change_owner(dev, net_old, net); 12646 WARN_ON(err); 12647 12648 netdev_lock(dev); 12649 dev->moving_ns = false; 12650 if (!netdev_need_ops_lock(dev)) 12651 netdev_unlock(dev); 12652 12653 /* Add the device back in the hashes */ 12654 list_netdevice(dev); 12655 /* Notify protocols, that a new device appeared. */ 12656 call_netdevice_notifiers(NETDEV_REGISTER, dev); 12657 netdev_unlock_ops(dev); 12658 12659 /* 12660 * Prevent userspace races by waiting until the network 12661 * device is fully setup before sending notifications. 12662 */ 12663 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 12664 12665 synchronize_net(); 12666 err = 0; 12667 out: 12668 return err; 12669 } 12670 12671 static int dev_cpu_dead(unsigned int oldcpu) 12672 { 12673 struct sk_buff **list_skb; 12674 struct sk_buff *skb; 12675 unsigned int cpu; 12676 struct softnet_data *sd, *oldsd, *remsd = NULL; 12677 12678 local_irq_disable(); 12679 cpu = smp_processor_id(); 12680 sd = &per_cpu(softnet_data, cpu); 12681 oldsd = &per_cpu(softnet_data, oldcpu); 12682 12683 /* Find end of our completion_queue. */ 12684 list_skb = &sd->completion_queue; 12685 while (*list_skb) 12686 list_skb = &(*list_skb)->next; 12687 /* Append completion queue from offline CPU. */ 12688 *list_skb = oldsd->completion_queue; 12689 oldsd->completion_queue = NULL; 12690 12691 /* Append output queue from offline CPU. */ 12692 if (oldsd->output_queue) { 12693 *sd->output_queue_tailp = oldsd->output_queue; 12694 sd->output_queue_tailp = oldsd->output_queue_tailp; 12695 oldsd->output_queue = NULL; 12696 oldsd->output_queue_tailp = &oldsd->output_queue; 12697 } 12698 /* Append NAPI poll list from offline CPU, with one exception : 12699 * process_backlog() must be called by cpu owning percpu backlog. 12700 * We properly handle process_queue & input_pkt_queue later. 12701 */ 12702 while (!list_empty(&oldsd->poll_list)) { 12703 struct napi_struct *napi = list_first_entry(&oldsd->poll_list, 12704 struct napi_struct, 12705 poll_list); 12706 12707 list_del_init(&napi->poll_list); 12708 if (napi->poll == process_backlog) 12709 napi->state &= NAPIF_STATE_THREADED; 12710 else 12711 ____napi_schedule(sd, napi); 12712 } 12713 12714 raise_softirq_irqoff(NET_TX_SOFTIRQ); 12715 local_irq_enable(); 12716 12717 if (!use_backlog_threads()) { 12718 #ifdef CONFIG_RPS 12719 remsd = oldsd->rps_ipi_list; 12720 oldsd->rps_ipi_list = NULL; 12721 #endif 12722 /* send out pending IPI's on offline CPU */ 12723 net_rps_send_ipi(remsd); 12724 } 12725 12726 /* Process offline CPU's input_pkt_queue */ 12727 while ((skb = __skb_dequeue(&oldsd->process_queue))) { 12728 netif_rx(skb); 12729 rps_input_queue_head_incr(oldsd); 12730 } 12731 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) { 12732 netif_rx(skb); 12733 rps_input_queue_head_incr(oldsd); 12734 } 12735 12736 return 0; 12737 } 12738 12739 /** 12740 * netdev_increment_features - increment feature set by one 12741 * @all: current feature set 12742 * @one: new feature set 12743 * @mask: mask feature set 12744 * 12745 * Computes a new feature set after adding a device with feature set 12746 * @one to the master device with current feature set @all. Will not 12747 * enable anything that is off in @mask. Returns the new feature set. 12748 */ 12749 netdev_features_t netdev_increment_features(netdev_features_t all, 12750 netdev_features_t one, netdev_features_t mask) 12751 { 12752 if (mask & NETIF_F_HW_CSUM) 12753 mask |= NETIF_F_CSUM_MASK; 12754 mask |= NETIF_F_VLAN_CHALLENGED; 12755 12756 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask; 12757 all &= one | ~NETIF_F_ALL_FOR_ALL; 12758 12759 /* If one device supports hw checksumming, set for all. */ 12760 if (all & NETIF_F_HW_CSUM) 12761 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM); 12762 12763 return all; 12764 } 12765 EXPORT_SYMBOL(netdev_increment_features); 12766 12767 /** 12768 * netdev_compute_master_upper_features - compute feature from lowers 12769 * @dev: the upper device 12770 * @update_header: whether to update upper device's header_len/headroom/tailroom 12771 * 12772 * Recompute the upper device's feature based on all lower devices. 12773 */ 12774 void netdev_compute_master_upper_features(struct net_device *dev, bool update_header) 12775 { 12776 unsigned int dst_release_flag = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 12777 netdev_features_t gso_partial_features = MASTER_UPPER_DEV_GSO_PARTIAL_FEATURES; 12778 netdev_features_t xfrm_features = MASTER_UPPER_DEV_XFRM_FEATURES; 12779 netdev_features_t mpls_features = MASTER_UPPER_DEV_MPLS_FEATURES; 12780 netdev_features_t vlan_features = MASTER_UPPER_DEV_VLAN_FEATURES; 12781 netdev_features_t enc_features = MASTER_UPPER_DEV_ENC_FEATURES; 12782 unsigned short max_header_len = ETH_HLEN; 12783 unsigned int tso_max_size = TSO_MAX_SIZE; 12784 unsigned short max_headroom = 0; 12785 unsigned short max_tailroom = 0; 12786 u16 tso_max_segs = TSO_MAX_SEGS; 12787 struct net_device *lower_dev; 12788 struct list_head *iter; 12789 12790 mpls_features = netdev_base_features(mpls_features); 12791 vlan_features = netdev_base_features(vlan_features); 12792 enc_features = netdev_base_features(enc_features); 12793 12794 netdev_for_each_lower_dev(dev, lower_dev, iter) { 12795 gso_partial_features = netdev_increment_features(gso_partial_features, 12796 lower_dev->gso_partial_features, 12797 MASTER_UPPER_DEV_GSO_PARTIAL_FEATURES); 12798 12799 vlan_features = netdev_increment_features(vlan_features, 12800 lower_dev->vlan_features, 12801 MASTER_UPPER_DEV_VLAN_FEATURES); 12802 12803 enc_features = netdev_increment_features(enc_features, 12804 lower_dev->hw_enc_features, 12805 MASTER_UPPER_DEV_ENC_FEATURES); 12806 12807 if (IS_ENABLED(CONFIG_XFRM_OFFLOAD)) 12808 xfrm_features = netdev_increment_features(xfrm_features, 12809 lower_dev->hw_enc_features, 12810 MASTER_UPPER_DEV_XFRM_FEATURES); 12811 12812 mpls_features = netdev_increment_features(mpls_features, 12813 lower_dev->mpls_features, 12814 MASTER_UPPER_DEV_MPLS_FEATURES); 12815 12816 dst_release_flag &= lower_dev->priv_flags; 12817 12818 if (update_header) { 12819 max_header_len = max(max_header_len, lower_dev->hard_header_len); 12820 max_headroom = max(max_headroom, lower_dev->needed_headroom); 12821 max_tailroom = max(max_tailroom, lower_dev->needed_tailroom); 12822 } 12823 12824 tso_max_size = min(tso_max_size, lower_dev->tso_max_size); 12825 tso_max_segs = min(tso_max_segs, lower_dev->tso_max_segs); 12826 } 12827 12828 dev->gso_partial_features = gso_partial_features; 12829 dev->vlan_features = vlan_features; 12830 dev->hw_enc_features = enc_features | NETIF_F_GSO_ENCAP_ALL | 12831 NETIF_F_HW_VLAN_CTAG_TX | 12832 NETIF_F_HW_VLAN_STAG_TX; 12833 if (IS_ENABLED(CONFIG_XFRM_OFFLOAD)) 12834 dev->hw_enc_features |= xfrm_features; 12835 dev->mpls_features = mpls_features; 12836 12837 dev->priv_flags &= ~IFF_XMIT_DST_RELEASE; 12838 if ((dev->priv_flags & IFF_XMIT_DST_RELEASE_PERM) && 12839 dst_release_flag == (IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM)) 12840 dev->priv_flags |= IFF_XMIT_DST_RELEASE; 12841 12842 if (update_header) { 12843 dev->hard_header_len = max_header_len; 12844 dev->needed_headroom = max_headroom; 12845 dev->needed_tailroom = max_tailroom; 12846 } 12847 12848 netif_set_tso_max_segs(dev, tso_max_segs); 12849 netif_set_tso_max_size(dev, tso_max_size); 12850 12851 netdev_change_features(dev); 12852 } 12853 EXPORT_SYMBOL(netdev_compute_master_upper_features); 12854 12855 static struct hlist_head * __net_init netdev_create_hash(void) 12856 { 12857 int i; 12858 struct hlist_head *hash; 12859 12860 hash = kmalloc_objs(*hash, NETDEV_HASHENTRIES, GFP_KERNEL); 12861 if (hash != NULL) 12862 for (i = 0; i < NETDEV_HASHENTRIES; i++) 12863 INIT_HLIST_HEAD(&hash[i]); 12864 12865 return hash; 12866 } 12867 12868 /* Initialize per network namespace state */ 12869 static int __net_init netdev_init(struct net *net) 12870 { 12871 BUILD_BUG_ON(GRO_HASH_BUCKETS > 12872 BITS_PER_BYTE * sizeof_field(struct gro_node, bitmask)); 12873 12874 INIT_LIST_HEAD(&net->dev_base_head); 12875 12876 net->dev_name_head = netdev_create_hash(); 12877 if (net->dev_name_head == NULL) 12878 goto err_name; 12879 12880 net->dev_index_head = netdev_create_hash(); 12881 if (net->dev_index_head == NULL) 12882 goto err_idx; 12883 12884 xa_init_flags(&net->dev_by_index, XA_FLAGS_ALLOC1); 12885 12886 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain); 12887 12888 return 0; 12889 12890 err_idx: 12891 kfree(net->dev_name_head); 12892 err_name: 12893 return -ENOMEM; 12894 } 12895 12896 /** 12897 * netdev_drivername - network driver for the device 12898 * @dev: network device 12899 * 12900 * Determine network driver for device. 12901 */ 12902 const char *netdev_drivername(const struct net_device *dev) 12903 { 12904 const struct device_driver *driver; 12905 const struct device *parent; 12906 const char *empty = ""; 12907 12908 parent = dev->dev.parent; 12909 if (!parent) 12910 return empty; 12911 12912 driver = parent->driver; 12913 if (driver && driver->name) 12914 return driver->name; 12915 return empty; 12916 } 12917 12918 static void __netdev_printk(const char *level, const struct net_device *dev, 12919 struct va_format *vaf) 12920 { 12921 if (dev && dev->dev.parent) { 12922 dev_printk_emit(level[1] - '0', 12923 dev->dev.parent, 12924 "%s %s %s%s: %pV", 12925 dev_driver_string(dev->dev.parent), 12926 dev_name(dev->dev.parent), 12927 netdev_name(dev), netdev_reg_state(dev), 12928 vaf); 12929 } else if (dev) { 12930 printk("%s%s%s: %pV", 12931 level, netdev_name(dev), netdev_reg_state(dev), vaf); 12932 } else { 12933 printk("%s(NULL net_device): %pV", level, vaf); 12934 } 12935 } 12936 12937 void netdev_printk(const char *level, const struct net_device *dev, 12938 const char *format, ...) 12939 { 12940 struct va_format vaf; 12941 va_list args; 12942 12943 va_start(args, format); 12944 12945 vaf.fmt = format; 12946 vaf.va = &args; 12947 12948 __netdev_printk(level, dev, &vaf); 12949 12950 va_end(args); 12951 } 12952 EXPORT_SYMBOL(netdev_printk); 12953 12954 #define define_netdev_printk_level(func, level) \ 12955 void func(const struct net_device *dev, const char *fmt, ...) \ 12956 { \ 12957 struct va_format vaf; \ 12958 va_list args; \ 12959 \ 12960 va_start(args, fmt); \ 12961 \ 12962 vaf.fmt = fmt; \ 12963 vaf.va = &args; \ 12964 \ 12965 __netdev_printk(level, dev, &vaf); \ 12966 \ 12967 va_end(args); \ 12968 } \ 12969 EXPORT_SYMBOL(func); 12970 12971 define_netdev_printk_level(netdev_emerg, KERN_EMERG); 12972 define_netdev_printk_level(netdev_alert, KERN_ALERT); 12973 define_netdev_printk_level(netdev_crit, KERN_CRIT); 12974 define_netdev_printk_level(netdev_err, KERN_ERR); 12975 define_netdev_printk_level(netdev_warn, KERN_WARNING); 12976 define_netdev_printk_level(netdev_notice, KERN_NOTICE); 12977 define_netdev_printk_level(netdev_info, KERN_INFO); 12978 12979 static void __net_exit netdev_exit(struct net *net) 12980 { 12981 kfree(net->dev_name_head); 12982 kfree(net->dev_index_head); 12983 xa_destroy(&net->dev_by_index); 12984 if (net != &init_net) 12985 WARN_ON_ONCE(!list_empty(&net->dev_base_head)); 12986 } 12987 12988 static struct pernet_operations __net_initdata netdev_net_ops = { 12989 .init = netdev_init, 12990 .exit = netdev_exit, 12991 }; 12992 12993 static void __net_exit default_device_exit_net(struct net *net) 12994 { 12995 struct netdev_name_node *name_node, *tmp; 12996 struct net_device *dev, *aux; 12997 /* 12998 * Push all migratable network devices back to the 12999 * initial network namespace 13000 */ 13001 ASSERT_RTNL(); 13002 for_each_netdev_safe(net, dev, aux) { 13003 int err; 13004 char fb_name[IFNAMSIZ]; 13005 13006 /* Ignore unmoveable devices (i.e. loopback) */ 13007 if (dev->netns_immutable) 13008 continue; 13009 13010 /* Leave virtual devices for the generic cleanup */ 13011 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund) 13012 continue; 13013 13014 /* Push remaining network devices to init_net */ 13015 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex); 13016 if (netdev_name_in_use(&init_net, fb_name)) 13017 snprintf(fb_name, IFNAMSIZ, "dev%%d"); 13018 13019 netdev_for_each_altname_safe(dev, name_node, tmp) 13020 if (netdev_name_in_use(&init_net, name_node->name)) 13021 __netdev_name_node_alt_destroy(name_node); 13022 13023 err = dev_change_net_namespace(dev, &init_net, fb_name); 13024 if (err) { 13025 pr_emerg("%s: failed to move %s to init_net: %d\n", 13026 __func__, dev->name, err); 13027 BUG(); 13028 } 13029 } 13030 } 13031 13032 static void __net_exit default_device_exit_batch(struct list_head *net_list) 13033 { 13034 /* At exit all network devices most be removed from a network 13035 * namespace. Do this in the reverse order of registration. 13036 * Do this across as many network namespaces as possible to 13037 * improve batching efficiency. 13038 */ 13039 struct net_device *dev; 13040 struct net *net; 13041 LIST_HEAD(dev_kill_list); 13042 13043 rtnl_lock(); 13044 list_for_each_entry(net, net_list, exit_list) { 13045 default_device_exit_net(net); 13046 cond_resched(); 13047 } 13048 13049 list_for_each_entry(net, net_list, exit_list) { 13050 for_each_netdev_reverse(net, dev) { 13051 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink) 13052 dev->rtnl_link_ops->dellink(dev, &dev_kill_list); 13053 else 13054 unregister_netdevice_queue(dev, &dev_kill_list); 13055 } 13056 } 13057 unregister_netdevice_many(&dev_kill_list); 13058 rtnl_unlock(); 13059 } 13060 13061 static struct pernet_operations __net_initdata default_device_ops = { 13062 .exit_batch = default_device_exit_batch, 13063 }; 13064 13065 static void __init net_dev_struct_check(void) 13066 { 13067 /* TX read-mostly hotpath */ 13068 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, priv_flags_fast); 13069 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, netdev_ops); 13070 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, header_ops); 13071 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, _tx); 13072 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, real_num_tx_queues); 13073 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_size); 13074 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_ipv4_max_size); 13075 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_segs); 13076 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_partial_features); 13077 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, num_tc); 13078 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, mtu); 13079 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, needed_headroom); 13080 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tc_to_txq); 13081 #ifdef CONFIG_XPS 13082 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, xps_maps); 13083 #endif 13084 #ifdef CONFIG_NETFILTER_EGRESS 13085 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, nf_hooks_egress); 13086 #endif 13087 #ifdef CONFIG_NET_XGRESS 13088 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tcx_egress); 13089 #endif 13090 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_tx, 160); 13091 13092 /* TXRX read-mostly hotpath */ 13093 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, lstats); 13094 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, state); 13095 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, flags); 13096 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, hard_header_len); 13097 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, features); 13098 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, ip6_ptr); 13099 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_txrx, 46); 13100 13101 /* RX read-mostly hotpath */ 13102 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ptype_specific); 13103 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ifindex); 13104 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, real_num_rx_queues); 13105 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, _rx); 13106 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_max_size); 13107 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_ipv4_max_size); 13108 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler); 13109 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler_data); 13110 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, nd_net); 13111 #ifdef CONFIG_NETPOLL 13112 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, npinfo); 13113 #endif 13114 #ifdef CONFIG_NET_XGRESS 13115 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, tcx_ingress); 13116 #endif 13117 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_rx, 92); 13118 } 13119 13120 /* 13121 * Initialize the DEV module. At boot time this walks the device list and 13122 * unhooks any devices that fail to initialise (normally hardware not 13123 * present) and leaves us with a valid list of present and active devices. 13124 * 13125 */ 13126 13127 /* We allocate 256 pages for each CPU if PAGE_SHIFT is 12 */ 13128 #define SYSTEM_PERCPU_PAGE_POOL_SIZE ((1 << 20) / PAGE_SIZE) 13129 13130 static int net_page_pool_create(int cpuid) 13131 { 13132 #if IS_ENABLED(CONFIG_PAGE_POOL) 13133 struct page_pool_params page_pool_params = { 13134 .pool_size = SYSTEM_PERCPU_PAGE_POOL_SIZE, 13135 .flags = PP_FLAG_SYSTEM_POOL, 13136 .nid = cpu_to_mem(cpuid), 13137 }; 13138 struct page_pool *pp_ptr; 13139 int err; 13140 13141 pp_ptr = page_pool_create_percpu(&page_pool_params, cpuid); 13142 if (IS_ERR(pp_ptr)) 13143 return -ENOMEM; 13144 13145 err = xdp_reg_page_pool(pp_ptr); 13146 if (err) { 13147 page_pool_destroy(pp_ptr); 13148 return err; 13149 } 13150 13151 per_cpu(system_page_pool.pool, cpuid) = pp_ptr; 13152 #endif 13153 return 0; 13154 } 13155 13156 static int backlog_napi_should_run(unsigned int cpu) 13157 { 13158 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 13159 struct napi_struct *napi = &sd->backlog; 13160 13161 return test_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 13162 } 13163 13164 static void run_backlog_napi(unsigned int cpu) 13165 { 13166 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 13167 13168 napi_threaded_poll_loop(&sd->backlog, false); 13169 } 13170 13171 static void backlog_napi_setup(unsigned int cpu) 13172 { 13173 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 13174 struct napi_struct *napi = &sd->backlog; 13175 13176 napi->thread = this_cpu_read(backlog_napi); 13177 set_bit(NAPI_STATE_THREADED, &napi->state); 13178 } 13179 13180 static struct smp_hotplug_thread backlog_threads = { 13181 .store = &backlog_napi, 13182 .thread_should_run = backlog_napi_should_run, 13183 .thread_fn = run_backlog_napi, 13184 .thread_comm = "backlog_napi/%u", 13185 .setup = backlog_napi_setup, 13186 }; 13187 13188 /* 13189 * This is called single threaded during boot, so no need 13190 * to take the rtnl semaphore. 13191 */ 13192 static int __init net_dev_init(void) 13193 { 13194 int i, rc = -ENOMEM; 13195 13196 BUG_ON(!dev_boot_phase); 13197 13198 net_dev_struct_check(); 13199 13200 if (dev_proc_init()) 13201 goto out; 13202 13203 if (netdev_kobject_init()) 13204 goto out; 13205 13206 for (i = 0; i < PTYPE_HASH_SIZE; i++) 13207 INIT_LIST_HEAD(&ptype_base[i]); 13208 13209 if (register_pernet_subsys(&netdev_net_ops)) 13210 goto out; 13211 13212 /* 13213 * Initialise the packet receive queues. 13214 */ 13215 13216 flush_backlogs_fallback = flush_backlogs_alloc(); 13217 if (!flush_backlogs_fallback) 13218 goto out; 13219 13220 for_each_possible_cpu(i) { 13221 struct softnet_data *sd = &per_cpu(softnet_data, i); 13222 13223 skb_queue_head_init(&sd->input_pkt_queue); 13224 skb_queue_head_init(&sd->process_queue); 13225 #ifdef CONFIG_XFRM_OFFLOAD 13226 skb_queue_head_init(&sd->xfrm_backlog); 13227 #endif 13228 INIT_LIST_HEAD(&sd->poll_list); 13229 sd->output_queue_tailp = &sd->output_queue; 13230 #ifdef CONFIG_RPS 13231 INIT_CSD(&sd->csd, rps_trigger_softirq, sd); 13232 sd->cpu = i; 13233 #endif 13234 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd); 13235 13236 gro_init(&sd->backlog.gro); 13237 sd->backlog.poll = process_backlog; 13238 sd->backlog.weight = weight_p; 13239 INIT_LIST_HEAD(&sd->backlog.poll_list); 13240 13241 if (net_page_pool_create(i)) 13242 goto out; 13243 } 13244 net_hotdata.skb_defer_nodes = 13245 __alloc_percpu(sizeof(struct skb_defer_node) * nr_node_ids, 13246 __alignof__(struct skb_defer_node)); 13247 if (!net_hotdata.skb_defer_nodes) 13248 goto out; 13249 if (use_backlog_threads()) 13250 smpboot_register_percpu_thread(&backlog_threads); 13251 13252 dev_boot_phase = 0; 13253 13254 /* The loopback device is special if any other network devices 13255 * is present in a network namespace the loopback device must 13256 * be present. Since we now dynamically allocate and free the 13257 * loopback device ensure this invariant is maintained by 13258 * keeping the loopback device as the first device on the 13259 * list of network devices. Ensuring the loopback devices 13260 * is the first device that appears and the last network device 13261 * that disappears. 13262 */ 13263 if (register_pernet_device(&loopback_net_ops)) 13264 goto out; 13265 13266 if (register_pernet_device(&default_device_ops)) 13267 goto out; 13268 13269 open_softirq(NET_TX_SOFTIRQ, net_tx_action); 13270 open_softirq(NET_RX_SOFTIRQ, net_rx_action); 13271 13272 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead", 13273 NULL, dev_cpu_dead); 13274 WARN_ON(rc < 0); 13275 rc = 0; 13276 13277 /* avoid static key IPIs to isolated CPUs */ 13278 if (housekeeping_enabled(HK_TYPE_MISC)) 13279 net_enable_timestamp(); 13280 out: 13281 if (rc < 0) { 13282 for_each_possible_cpu(i) { 13283 struct page_pool *pp_ptr; 13284 13285 pp_ptr = per_cpu(system_page_pool.pool, i); 13286 if (!pp_ptr) 13287 continue; 13288 13289 xdp_unreg_page_pool(pp_ptr); 13290 page_pool_destroy(pp_ptr); 13291 per_cpu(system_page_pool.pool, i) = NULL; 13292 } 13293 } 13294 13295 return rc; 13296 } 13297 13298 subsys_initcall(net_dev_init); 13299