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(sizeof(*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 (WARN_ON_ONCE(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(struct_size_t(struct flush_backlogs, w, nr_cpu_ids), 6514 GFP_KERNEL); 6515 } 6516 6517 static struct flush_backlogs *flush_backlogs_fallback; 6518 static DEFINE_MUTEX(flush_backlogs_mutex); 6519 6520 static void flush_all_backlogs(void) 6521 { 6522 struct flush_backlogs *ptr = flush_backlogs_alloc(); 6523 unsigned int cpu; 6524 6525 if (!ptr) { 6526 mutex_lock(&flush_backlogs_mutex); 6527 ptr = flush_backlogs_fallback; 6528 } 6529 cpumask_clear(&ptr->flush_cpus); 6530 6531 cpus_read_lock(); 6532 6533 for_each_online_cpu(cpu) { 6534 if (flush_required(cpu)) { 6535 INIT_WORK(&ptr->w[cpu], flush_backlog); 6536 queue_work_on(cpu, system_highpri_wq, &ptr->w[cpu]); 6537 __cpumask_set_cpu(cpu, &ptr->flush_cpus); 6538 } 6539 } 6540 6541 /* we can have in flight packet[s] on the cpus we are not flushing, 6542 * synchronize_net() in unregister_netdevice_many() will take care of 6543 * them. 6544 */ 6545 for_each_cpu(cpu, &ptr->flush_cpus) 6546 flush_work(&ptr->w[cpu]); 6547 6548 cpus_read_unlock(); 6549 6550 if (ptr != flush_backlogs_fallback) 6551 kfree(ptr); 6552 else 6553 mutex_unlock(&flush_backlogs_mutex); 6554 } 6555 6556 static void net_rps_send_ipi(struct softnet_data *remsd) 6557 { 6558 #ifdef CONFIG_RPS 6559 while (remsd) { 6560 struct softnet_data *next = remsd->rps_ipi_next; 6561 6562 if (cpu_online(remsd->cpu)) 6563 smp_call_function_single_async(remsd->cpu, &remsd->csd); 6564 remsd = next; 6565 } 6566 #endif 6567 } 6568 6569 /* 6570 * net_rps_action_and_irq_enable sends any pending IPI's for rps. 6571 * Note: called with local irq disabled, but exits with local irq enabled. 6572 */ 6573 static void net_rps_action_and_irq_enable(struct softnet_data *sd) 6574 { 6575 #ifdef CONFIG_RPS 6576 struct softnet_data *remsd = sd->rps_ipi_list; 6577 6578 if (!use_backlog_threads() && remsd) { 6579 sd->rps_ipi_list = NULL; 6580 6581 local_irq_enable(); 6582 6583 /* Send pending IPI's to kick RPS processing on remote cpus. */ 6584 net_rps_send_ipi(remsd); 6585 } else 6586 #endif 6587 local_irq_enable(); 6588 } 6589 6590 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd) 6591 { 6592 #ifdef CONFIG_RPS 6593 return !use_backlog_threads() && sd->rps_ipi_list; 6594 #else 6595 return false; 6596 #endif 6597 } 6598 6599 static int process_backlog(struct napi_struct *napi, int quota) 6600 { 6601 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog); 6602 bool again = true; 6603 int work = 0; 6604 6605 /* Check if we have pending ipi, its better to send them now, 6606 * not waiting net_rx_action() end. 6607 */ 6608 if (sd_has_rps_ipi_waiting(sd)) { 6609 local_irq_disable(); 6610 net_rps_action_and_irq_enable(sd); 6611 } 6612 6613 napi->weight = READ_ONCE(net_hotdata.dev_rx_weight); 6614 while (again) { 6615 struct sk_buff *skb; 6616 6617 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6618 while ((skb = __skb_dequeue(&sd->process_queue))) { 6619 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6620 rcu_read_lock(); 6621 __netif_receive_skb(skb); 6622 rcu_read_unlock(); 6623 if (++work >= quota) { 6624 rps_input_queue_head_add(sd, work); 6625 return work; 6626 } 6627 6628 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6629 } 6630 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6631 6632 backlog_lock_irq_disable(sd); 6633 if (skb_queue_empty(&sd->input_pkt_queue)) { 6634 /* 6635 * Inline a custom version of __napi_complete(). 6636 * only current cpu owns and manipulates this napi, 6637 * and NAPI_STATE_SCHED is the only possible flag set 6638 * on backlog. 6639 * We can use a plain write instead of clear_bit(), 6640 * and we dont need an smp_mb() memory barrier. 6641 */ 6642 napi->state &= NAPIF_STATE_THREADED; 6643 again = false; 6644 } else { 6645 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6646 skb_queue_splice_tail_init(&sd->input_pkt_queue, 6647 &sd->process_queue); 6648 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6649 } 6650 backlog_unlock_irq_enable(sd); 6651 } 6652 6653 if (work) 6654 rps_input_queue_head_add(sd, work); 6655 return work; 6656 } 6657 6658 /** 6659 * __napi_schedule - schedule for receive 6660 * @n: entry to schedule 6661 * 6662 * The entry's receive function will be scheduled to run. 6663 * Consider using __napi_schedule_irqoff() if hard irqs are masked. 6664 */ 6665 void __napi_schedule(struct napi_struct *n) 6666 { 6667 unsigned long flags; 6668 6669 local_irq_save(flags); 6670 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6671 local_irq_restore(flags); 6672 } 6673 EXPORT_SYMBOL(__napi_schedule); 6674 6675 /** 6676 * napi_schedule_prep - check if napi can be scheduled 6677 * @n: napi context 6678 * 6679 * Test if NAPI routine is already running, and if not mark 6680 * it as running. This is used as a condition variable to 6681 * insure only one NAPI poll instance runs. We also make 6682 * sure there is no pending NAPI disable. 6683 */ 6684 bool napi_schedule_prep(struct napi_struct *n) 6685 { 6686 unsigned long new, val = READ_ONCE(n->state); 6687 6688 do { 6689 if (unlikely(val & NAPIF_STATE_DISABLE)) 6690 return false; 6691 new = val | NAPIF_STATE_SCHED; 6692 6693 /* Sets STATE_MISSED bit if STATE_SCHED was already set 6694 * This was suggested by Alexander Duyck, as compiler 6695 * emits better code than : 6696 * if (val & NAPIF_STATE_SCHED) 6697 * new |= NAPIF_STATE_MISSED; 6698 */ 6699 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED * 6700 NAPIF_STATE_MISSED; 6701 } while (!try_cmpxchg(&n->state, &val, new)); 6702 6703 return !(val & NAPIF_STATE_SCHED); 6704 } 6705 EXPORT_SYMBOL(napi_schedule_prep); 6706 6707 /** 6708 * __napi_schedule_irqoff - schedule for receive 6709 * @n: entry to schedule 6710 * 6711 * Variant of __napi_schedule() assuming hard irqs are masked. 6712 * 6713 * On PREEMPT_RT enabled kernels this maps to __napi_schedule() 6714 * because the interrupt disabled assumption might not be true 6715 * due to force-threaded interrupts and spinlock substitution. 6716 */ 6717 void __napi_schedule_irqoff(struct napi_struct *n) 6718 { 6719 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6720 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6721 else 6722 __napi_schedule(n); 6723 } 6724 EXPORT_SYMBOL(__napi_schedule_irqoff); 6725 6726 bool napi_complete_done(struct napi_struct *n, int work_done) 6727 { 6728 unsigned long flags, val, new, timeout = 0; 6729 bool ret = true; 6730 6731 /* 6732 * 1) Don't let napi dequeue from the cpu poll list 6733 * just in case its running on a different cpu. 6734 * 2) If we are busy polling, do nothing here, we have 6735 * the guarantee we will be called later. 6736 */ 6737 if (unlikely(n->state & (NAPIF_STATE_NPSVC | 6738 NAPIF_STATE_IN_BUSY_POLL))) 6739 return false; 6740 6741 if (work_done) { 6742 if (n->gro.bitmask) 6743 timeout = napi_get_gro_flush_timeout(n); 6744 n->defer_hard_irqs_count = napi_get_defer_hard_irqs(n); 6745 } 6746 if (n->defer_hard_irqs_count > 0) { 6747 n->defer_hard_irqs_count--; 6748 timeout = napi_get_gro_flush_timeout(n); 6749 if (timeout) 6750 ret = false; 6751 } 6752 6753 /* 6754 * When the NAPI instance uses a timeout and keeps postponing 6755 * it, we need to bound somehow the time packets are kept in 6756 * the GRO layer. 6757 */ 6758 gro_flush_normal(&n->gro, !!timeout); 6759 6760 if (unlikely(!list_empty(&n->poll_list))) { 6761 /* If n->poll_list is not empty, we need to mask irqs */ 6762 local_irq_save(flags); 6763 list_del_init(&n->poll_list); 6764 local_irq_restore(flags); 6765 } 6766 WRITE_ONCE(n->list_owner, -1); 6767 6768 val = READ_ONCE(n->state); 6769 do { 6770 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED)); 6771 6772 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED | 6773 NAPIF_STATE_SCHED_THREADED | 6774 NAPIF_STATE_PREFER_BUSY_POLL); 6775 6776 /* If STATE_MISSED was set, leave STATE_SCHED set, 6777 * because we will call napi->poll() one more time. 6778 * This C code was suggested by Alexander Duyck to help gcc. 6779 */ 6780 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED * 6781 NAPIF_STATE_SCHED; 6782 } while (!try_cmpxchg(&n->state, &val, new)); 6783 6784 if (unlikely(val & NAPIF_STATE_MISSED)) { 6785 __napi_schedule(n); 6786 return false; 6787 } 6788 6789 if (timeout) 6790 hrtimer_start(&n->timer, ns_to_ktime(timeout), 6791 HRTIMER_MODE_REL_PINNED); 6792 return ret; 6793 } 6794 EXPORT_SYMBOL(napi_complete_done); 6795 6796 static void skb_defer_free_flush(void) 6797 { 6798 struct llist_node *free_list; 6799 struct sk_buff *skb, *next; 6800 struct skb_defer_node *sdn; 6801 int node; 6802 6803 for_each_node(node) { 6804 sdn = this_cpu_ptr(net_hotdata.skb_defer_nodes) + node; 6805 6806 if (llist_empty(&sdn->defer_list)) 6807 continue; 6808 atomic_long_set(&sdn->defer_count, 0); 6809 free_list = llist_del_all(&sdn->defer_list); 6810 6811 llist_for_each_entry_safe(skb, next, free_list, ll_node) { 6812 prefetch(next); 6813 napi_consume_skb(skb, 1); 6814 } 6815 } 6816 } 6817 6818 #if defined(CONFIG_NET_RX_BUSY_POLL) 6819 6820 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule) 6821 { 6822 if (!skip_schedule) { 6823 gro_normal_list(&napi->gro); 6824 __napi_schedule(napi); 6825 return; 6826 } 6827 6828 /* Flush too old packets. If HZ < 1000, flush all packets */ 6829 gro_flush_normal(&napi->gro, HZ >= 1000); 6830 6831 clear_bit(NAPI_STATE_SCHED, &napi->state); 6832 } 6833 6834 enum { 6835 NAPI_F_PREFER_BUSY_POLL = 1, 6836 NAPI_F_END_ON_RESCHED = 2, 6837 }; 6838 6839 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, 6840 unsigned flags, u16 budget) 6841 { 6842 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6843 bool skip_schedule = false; 6844 unsigned long timeout; 6845 int rc; 6846 6847 /* Busy polling means there is a high chance device driver hard irq 6848 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was 6849 * set in napi_schedule_prep(). 6850 * Since we are about to call napi->poll() once more, we can safely 6851 * clear NAPI_STATE_MISSED. 6852 * 6853 * Note: x86 could use a single "lock and ..." instruction 6854 * to perform these two clear_bit() 6855 */ 6856 clear_bit(NAPI_STATE_MISSED, &napi->state); 6857 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state); 6858 6859 local_bh_disable(); 6860 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6861 6862 if (flags & NAPI_F_PREFER_BUSY_POLL) { 6863 napi->defer_hard_irqs_count = napi_get_defer_hard_irqs(napi); 6864 timeout = napi_get_gro_flush_timeout(napi); 6865 if (napi->defer_hard_irqs_count && timeout) { 6866 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED); 6867 skip_schedule = true; 6868 } 6869 } 6870 6871 /* All we really want here is to re-enable device interrupts. 6872 * Ideally, a new ndo_busy_poll_stop() could avoid another round. 6873 */ 6874 rc = napi->poll(napi, budget); 6875 /* We can't gro_normal_list() here, because napi->poll() might have 6876 * rearmed the napi (napi_complete_done()) in which case it could 6877 * already be running on another CPU. 6878 */ 6879 trace_napi_poll(napi, rc, budget); 6880 netpoll_poll_unlock(have_poll_lock); 6881 if (rc == budget) 6882 __busy_poll_stop(napi, skip_schedule); 6883 bpf_net_ctx_clear(bpf_net_ctx); 6884 local_bh_enable(); 6885 } 6886 6887 static void __napi_busy_loop(unsigned int napi_id, 6888 bool (*loop_end)(void *, unsigned long), 6889 void *loop_end_arg, unsigned flags, u16 budget) 6890 { 6891 unsigned long start_time = loop_end ? busy_loop_current_time() : 0; 6892 int (*napi_poll)(struct napi_struct *napi, int budget); 6893 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6894 void *have_poll_lock = NULL; 6895 struct napi_struct *napi; 6896 6897 WARN_ON_ONCE(!rcu_read_lock_held()); 6898 6899 restart: 6900 napi_poll = NULL; 6901 6902 napi = napi_by_id(napi_id); 6903 if (!napi) 6904 return; 6905 6906 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6907 preempt_disable(); 6908 for (;;) { 6909 int work = 0; 6910 6911 local_bh_disable(); 6912 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6913 if (!napi_poll) { 6914 unsigned long val = READ_ONCE(napi->state); 6915 6916 /* If multiple threads are competing for this napi, 6917 * we avoid dirtying napi->state as much as we can. 6918 */ 6919 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED | 6920 NAPIF_STATE_IN_BUSY_POLL)) { 6921 if (flags & NAPI_F_PREFER_BUSY_POLL) 6922 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6923 goto count; 6924 } 6925 if (cmpxchg(&napi->state, val, 6926 val | NAPIF_STATE_IN_BUSY_POLL | 6927 NAPIF_STATE_SCHED) != val) { 6928 if (flags & NAPI_F_PREFER_BUSY_POLL) 6929 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6930 goto count; 6931 } 6932 have_poll_lock = netpoll_poll_lock(napi); 6933 napi_poll = napi->poll; 6934 } 6935 work = napi_poll(napi, budget); 6936 trace_napi_poll(napi, work, budget); 6937 gro_normal_list(&napi->gro); 6938 count: 6939 if (work > 0) 6940 __NET_ADD_STATS(dev_net(napi->dev), 6941 LINUX_MIB_BUSYPOLLRXPACKETS, work); 6942 skb_defer_free_flush(); 6943 bpf_net_ctx_clear(bpf_net_ctx); 6944 local_bh_enable(); 6945 6946 if (!loop_end || loop_end(loop_end_arg, start_time)) 6947 break; 6948 6949 if (unlikely(need_resched())) { 6950 if (flags & NAPI_F_END_ON_RESCHED) 6951 break; 6952 if (napi_poll) 6953 busy_poll_stop(napi, have_poll_lock, flags, budget); 6954 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6955 preempt_enable(); 6956 rcu_read_unlock(); 6957 cond_resched(); 6958 rcu_read_lock(); 6959 if (loop_end(loop_end_arg, start_time)) 6960 return; 6961 goto restart; 6962 } 6963 cpu_relax(); 6964 } 6965 if (napi_poll) 6966 busy_poll_stop(napi, have_poll_lock, flags, budget); 6967 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6968 preempt_enable(); 6969 } 6970 6971 void napi_busy_loop_rcu(unsigned int napi_id, 6972 bool (*loop_end)(void *, unsigned long), 6973 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6974 { 6975 unsigned flags = NAPI_F_END_ON_RESCHED; 6976 6977 if (prefer_busy_poll) 6978 flags |= NAPI_F_PREFER_BUSY_POLL; 6979 6980 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6981 } 6982 6983 void napi_busy_loop(unsigned int napi_id, 6984 bool (*loop_end)(void *, unsigned long), 6985 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6986 { 6987 unsigned flags = prefer_busy_poll ? NAPI_F_PREFER_BUSY_POLL : 0; 6988 6989 rcu_read_lock(); 6990 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6991 rcu_read_unlock(); 6992 } 6993 EXPORT_SYMBOL(napi_busy_loop); 6994 6995 void napi_suspend_irqs(unsigned int napi_id) 6996 { 6997 struct napi_struct *napi; 6998 6999 rcu_read_lock(); 7000 napi = napi_by_id(napi_id); 7001 if (napi) { 7002 unsigned long timeout = napi_get_irq_suspend_timeout(napi); 7003 7004 if (timeout) 7005 hrtimer_start(&napi->timer, ns_to_ktime(timeout), 7006 HRTIMER_MODE_REL_PINNED); 7007 } 7008 rcu_read_unlock(); 7009 } 7010 7011 void napi_resume_irqs(unsigned int napi_id) 7012 { 7013 struct napi_struct *napi; 7014 7015 rcu_read_lock(); 7016 napi = napi_by_id(napi_id); 7017 if (napi) { 7018 /* If irq_suspend_timeout is set to 0 between the call to 7019 * napi_suspend_irqs and now, the original value still 7020 * determines the safety timeout as intended and napi_watchdog 7021 * will resume irq processing. 7022 */ 7023 if (napi_get_irq_suspend_timeout(napi)) { 7024 local_bh_disable(); 7025 napi_schedule(napi); 7026 local_bh_enable(); 7027 } 7028 } 7029 rcu_read_unlock(); 7030 } 7031 7032 #endif /* CONFIG_NET_RX_BUSY_POLL */ 7033 7034 static void __napi_hash_add_with_id(struct napi_struct *napi, 7035 unsigned int napi_id) 7036 { 7037 napi->gro.cached_napi_id = napi_id; 7038 7039 WRITE_ONCE(napi->napi_id, napi_id); 7040 hlist_add_head_rcu(&napi->napi_hash_node, 7041 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]); 7042 } 7043 7044 static void napi_hash_add_with_id(struct napi_struct *napi, 7045 unsigned int napi_id) 7046 { 7047 unsigned long flags; 7048 7049 spin_lock_irqsave(&napi_hash_lock, flags); 7050 WARN_ON_ONCE(napi_by_id(napi_id)); 7051 __napi_hash_add_with_id(napi, napi_id); 7052 spin_unlock_irqrestore(&napi_hash_lock, flags); 7053 } 7054 7055 static void napi_hash_add(struct napi_struct *napi) 7056 { 7057 unsigned long flags; 7058 7059 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state)) 7060 return; 7061 7062 spin_lock_irqsave(&napi_hash_lock, flags); 7063 7064 /* 0..NR_CPUS range is reserved for sender_cpu use */ 7065 do { 7066 if (unlikely(!napi_id_valid(++napi_gen_id))) 7067 napi_gen_id = MIN_NAPI_ID; 7068 } while (napi_by_id(napi_gen_id)); 7069 7070 __napi_hash_add_with_id(napi, napi_gen_id); 7071 7072 spin_unlock_irqrestore(&napi_hash_lock, flags); 7073 } 7074 7075 /* Warning : caller is responsible to make sure rcu grace period 7076 * is respected before freeing memory containing @napi 7077 */ 7078 static void napi_hash_del(struct napi_struct *napi) 7079 { 7080 unsigned long flags; 7081 7082 spin_lock_irqsave(&napi_hash_lock, flags); 7083 7084 hlist_del_init_rcu(&napi->napi_hash_node); 7085 7086 spin_unlock_irqrestore(&napi_hash_lock, flags); 7087 } 7088 7089 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer) 7090 { 7091 struct napi_struct *napi; 7092 7093 napi = container_of(timer, struct napi_struct, timer); 7094 7095 /* Note : we use a relaxed variant of napi_schedule_prep() not setting 7096 * NAPI_STATE_MISSED, since we do not react to a device IRQ. 7097 */ 7098 if (!napi_disable_pending(napi) && 7099 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) { 7100 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 7101 __napi_schedule_irqoff(napi); 7102 } 7103 7104 return HRTIMER_NORESTART; 7105 } 7106 7107 static void napi_stop_kthread(struct napi_struct *napi) 7108 { 7109 unsigned long val, new; 7110 7111 /* Wait until the napi STATE_THREADED is unset. */ 7112 while (true) { 7113 val = READ_ONCE(napi->state); 7114 7115 /* If napi kthread own this napi or the napi is idle, 7116 * STATE_THREADED can be unset here. 7117 */ 7118 if ((val & NAPIF_STATE_SCHED_THREADED) || 7119 !(val & NAPIF_STATE_SCHED)) { 7120 new = val & (~(NAPIF_STATE_THREADED | 7121 NAPIF_STATE_THREADED_BUSY_POLL)); 7122 } else { 7123 msleep(20); 7124 continue; 7125 } 7126 7127 if (try_cmpxchg(&napi->state, &val, new)) 7128 break; 7129 } 7130 7131 /* Once STATE_THREADED is unset, wait for SCHED_THREADED to be unset by 7132 * the kthread. 7133 */ 7134 while (true) { 7135 if (!test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) 7136 break; 7137 7138 msleep(20); 7139 } 7140 7141 kthread_stop(napi->thread); 7142 napi->thread = NULL; 7143 } 7144 7145 static void napi_set_threaded_state(struct napi_struct *napi, 7146 enum netdev_napi_threaded threaded_mode) 7147 { 7148 bool threaded = threaded_mode != NETDEV_NAPI_THREADED_DISABLED; 7149 bool busy_poll = threaded_mode == NETDEV_NAPI_THREADED_BUSY_POLL; 7150 7151 assign_bit(NAPI_STATE_THREADED, &napi->state, threaded); 7152 assign_bit(NAPI_STATE_THREADED_BUSY_POLL, &napi->state, busy_poll); 7153 } 7154 7155 int napi_set_threaded(struct napi_struct *napi, 7156 enum netdev_napi_threaded threaded) 7157 { 7158 if (threaded) { 7159 if (!napi->thread) { 7160 int err = napi_kthread_create(napi); 7161 7162 if (err) 7163 return err; 7164 } 7165 } 7166 7167 if (napi->config) 7168 napi->config->threaded = threaded; 7169 7170 /* Setting/unsetting threaded mode on a napi might not immediately 7171 * take effect, if the current napi instance is actively being 7172 * polled. In this case, the switch between threaded mode and 7173 * softirq mode will happen in the next round of napi_schedule(). 7174 * This should not cause hiccups/stalls to the live traffic. 7175 */ 7176 if (!threaded && napi->thread) { 7177 napi_stop_kthread(napi); 7178 } else { 7179 /* Make sure kthread is created before THREADED bit is set. */ 7180 smp_mb__before_atomic(); 7181 napi_set_threaded_state(napi, threaded); 7182 } 7183 7184 return 0; 7185 } 7186 7187 int netif_set_threaded(struct net_device *dev, 7188 enum netdev_napi_threaded threaded) 7189 { 7190 struct napi_struct *napi; 7191 int i, err = 0; 7192 7193 netdev_assert_locked_or_invisible(dev); 7194 7195 if (threaded) { 7196 list_for_each_entry(napi, &dev->napi_list, dev_list) { 7197 if (!napi->thread) { 7198 err = napi_kthread_create(napi); 7199 if (err) { 7200 threaded = NETDEV_NAPI_THREADED_DISABLED; 7201 break; 7202 } 7203 } 7204 } 7205 } 7206 7207 WRITE_ONCE(dev->threaded, threaded); 7208 7209 /* The error should not occur as the kthreads are already created. */ 7210 list_for_each_entry(napi, &dev->napi_list, dev_list) 7211 WARN_ON_ONCE(napi_set_threaded(napi, threaded)); 7212 7213 /* Override the config for all NAPIs even if currently not listed */ 7214 for (i = 0; i < dev->num_napi_configs; i++) 7215 dev->napi_config[i].threaded = threaded; 7216 7217 return err; 7218 } 7219 7220 /** 7221 * netif_threaded_enable() - enable threaded NAPIs 7222 * @dev: net_device instance 7223 * 7224 * Enable threaded mode for the NAPI instances of the device. This may be useful 7225 * for devices where multiple NAPI instances get scheduled by a single 7226 * interrupt. Threaded NAPI allows moving the NAPI processing to cores other 7227 * than the core where IRQ is mapped. 7228 * 7229 * This function should be called before @dev is registered. 7230 */ 7231 void netif_threaded_enable(struct net_device *dev) 7232 { 7233 WARN_ON_ONCE(netif_set_threaded(dev, NETDEV_NAPI_THREADED_ENABLED)); 7234 } 7235 EXPORT_SYMBOL(netif_threaded_enable); 7236 7237 /** 7238 * netif_queue_set_napi - Associate queue with the napi 7239 * @dev: device to which NAPI and queue belong 7240 * @queue_index: Index of queue 7241 * @type: queue type as RX or TX 7242 * @napi: NAPI context, pass NULL to clear previously set NAPI 7243 * 7244 * Set queue with its corresponding napi context. This should be done after 7245 * registering the NAPI handler for the queue-vector and the queues have been 7246 * mapped to the corresponding interrupt vector. 7247 */ 7248 void netif_queue_set_napi(struct net_device *dev, unsigned int queue_index, 7249 enum netdev_queue_type type, struct napi_struct *napi) 7250 { 7251 struct netdev_rx_queue *rxq; 7252 struct netdev_queue *txq; 7253 7254 if (WARN_ON_ONCE(napi && !napi->dev)) 7255 return; 7256 netdev_ops_assert_locked_or_invisible(dev); 7257 7258 switch (type) { 7259 case NETDEV_QUEUE_TYPE_RX: 7260 rxq = __netif_get_rx_queue(dev, queue_index); 7261 rxq->napi = napi; 7262 return; 7263 case NETDEV_QUEUE_TYPE_TX: 7264 txq = netdev_get_tx_queue(dev, queue_index); 7265 txq->napi = napi; 7266 return; 7267 default: 7268 return; 7269 } 7270 } 7271 EXPORT_SYMBOL(netif_queue_set_napi); 7272 7273 static void 7274 netif_napi_irq_notify(struct irq_affinity_notify *notify, 7275 const cpumask_t *mask) 7276 { 7277 struct napi_struct *napi = 7278 container_of(notify, struct napi_struct, notify); 7279 #ifdef CONFIG_RFS_ACCEL 7280 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 7281 int err; 7282 #endif 7283 7284 if (napi->config && napi->dev->irq_affinity_auto) 7285 cpumask_copy(&napi->config->affinity_mask, mask); 7286 7287 #ifdef CONFIG_RFS_ACCEL 7288 if (napi->dev->rx_cpu_rmap_auto) { 7289 err = cpu_rmap_update(rmap, napi->napi_rmap_idx, mask); 7290 if (err) 7291 netdev_warn(napi->dev, "RMAP update failed (%d)\n", 7292 err); 7293 } 7294 #endif 7295 } 7296 7297 #ifdef CONFIG_RFS_ACCEL 7298 static void netif_napi_affinity_release(struct kref *ref) 7299 { 7300 struct napi_struct *napi = 7301 container_of(ref, struct napi_struct, notify.kref); 7302 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 7303 7304 netdev_assert_locked(napi->dev); 7305 WARN_ON(test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, 7306 &napi->state)); 7307 7308 if (!napi->dev->rx_cpu_rmap_auto) 7309 return; 7310 rmap->obj[napi->napi_rmap_idx] = NULL; 7311 napi->napi_rmap_idx = -1; 7312 cpu_rmap_put(rmap); 7313 } 7314 7315 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 7316 { 7317 if (dev->rx_cpu_rmap_auto) 7318 return 0; 7319 7320 dev->rx_cpu_rmap = alloc_irq_cpu_rmap(num_irqs); 7321 if (!dev->rx_cpu_rmap) 7322 return -ENOMEM; 7323 7324 dev->rx_cpu_rmap_auto = true; 7325 return 0; 7326 } 7327 EXPORT_SYMBOL(netif_enable_cpu_rmap); 7328 7329 static void netif_del_cpu_rmap(struct net_device *dev) 7330 { 7331 struct cpu_rmap *rmap = dev->rx_cpu_rmap; 7332 7333 if (!dev->rx_cpu_rmap_auto) 7334 return; 7335 7336 /* Free the rmap */ 7337 cpu_rmap_put(rmap); 7338 dev->rx_cpu_rmap = NULL; 7339 dev->rx_cpu_rmap_auto = false; 7340 } 7341 7342 #else 7343 static void netif_napi_affinity_release(struct kref *ref) 7344 { 7345 } 7346 7347 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 7348 { 7349 return 0; 7350 } 7351 EXPORT_SYMBOL(netif_enable_cpu_rmap); 7352 7353 static void netif_del_cpu_rmap(struct net_device *dev) 7354 { 7355 } 7356 #endif 7357 7358 void netif_set_affinity_auto(struct net_device *dev) 7359 { 7360 unsigned int i, maxqs, numa; 7361 7362 maxqs = max(dev->num_tx_queues, dev->num_rx_queues); 7363 numa = dev_to_node(&dev->dev); 7364 7365 for (i = 0; i < maxqs; i++) 7366 cpumask_set_cpu(cpumask_local_spread(i, numa), 7367 &dev->napi_config[i].affinity_mask); 7368 7369 dev->irq_affinity_auto = true; 7370 } 7371 EXPORT_SYMBOL(netif_set_affinity_auto); 7372 7373 void netif_napi_set_irq_locked(struct napi_struct *napi, int irq) 7374 { 7375 int rc; 7376 7377 netdev_assert_locked_or_invisible(napi->dev); 7378 7379 if (napi->irq == irq) 7380 return; 7381 7382 /* Remove existing resources */ 7383 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7384 irq_set_affinity_notifier(napi->irq, NULL); 7385 7386 napi->irq = irq; 7387 if (irq < 0 || 7388 (!napi->dev->rx_cpu_rmap_auto && !napi->dev->irq_affinity_auto)) 7389 return; 7390 7391 /* Abort for buggy drivers */ 7392 if (napi->dev->irq_affinity_auto && WARN_ON_ONCE(!napi->config)) 7393 return; 7394 7395 #ifdef CONFIG_RFS_ACCEL 7396 if (napi->dev->rx_cpu_rmap_auto) { 7397 rc = cpu_rmap_add(napi->dev->rx_cpu_rmap, napi); 7398 if (rc < 0) 7399 return; 7400 7401 cpu_rmap_get(napi->dev->rx_cpu_rmap); 7402 napi->napi_rmap_idx = rc; 7403 } 7404 #endif 7405 7406 /* Use core IRQ notifier */ 7407 napi->notify.notify = netif_napi_irq_notify; 7408 napi->notify.release = netif_napi_affinity_release; 7409 rc = irq_set_affinity_notifier(irq, &napi->notify); 7410 if (rc) { 7411 netdev_warn(napi->dev, "Unable to set IRQ notifier (%d)\n", 7412 rc); 7413 goto put_rmap; 7414 } 7415 7416 set_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state); 7417 return; 7418 7419 put_rmap: 7420 #ifdef CONFIG_RFS_ACCEL 7421 if (napi->dev->rx_cpu_rmap_auto) { 7422 napi->dev->rx_cpu_rmap->obj[napi->napi_rmap_idx] = NULL; 7423 cpu_rmap_put(napi->dev->rx_cpu_rmap); 7424 napi->napi_rmap_idx = -1; 7425 } 7426 #endif 7427 napi->notify.notify = NULL; 7428 napi->notify.release = NULL; 7429 } 7430 EXPORT_SYMBOL(netif_napi_set_irq_locked); 7431 7432 static void napi_restore_config(struct napi_struct *n) 7433 { 7434 n->defer_hard_irqs = n->config->defer_hard_irqs; 7435 n->gro_flush_timeout = n->config->gro_flush_timeout; 7436 n->irq_suspend_timeout = n->config->irq_suspend_timeout; 7437 7438 if (n->dev->irq_affinity_auto && 7439 test_bit(NAPI_STATE_HAS_NOTIFIER, &n->state)) 7440 irq_set_affinity(n->irq, &n->config->affinity_mask); 7441 7442 /* a NAPI ID might be stored in the config, if so use it. if not, use 7443 * napi_hash_add to generate one for us. 7444 */ 7445 if (n->config->napi_id) { 7446 napi_hash_add_with_id(n, n->config->napi_id); 7447 } else { 7448 napi_hash_add(n); 7449 n->config->napi_id = n->napi_id; 7450 } 7451 7452 WARN_ON_ONCE(napi_set_threaded(n, n->config->threaded)); 7453 } 7454 7455 static void napi_save_config(struct napi_struct *n) 7456 { 7457 n->config->defer_hard_irqs = n->defer_hard_irqs; 7458 n->config->gro_flush_timeout = n->gro_flush_timeout; 7459 n->config->irq_suspend_timeout = n->irq_suspend_timeout; 7460 napi_hash_del(n); 7461 } 7462 7463 /* Netlink wants the NAPI list to be sorted by ID, if adding a NAPI which will 7464 * inherit an existing ID try to insert it at the right position. 7465 */ 7466 static void 7467 netif_napi_dev_list_add(struct net_device *dev, struct napi_struct *napi) 7468 { 7469 unsigned int new_id, pos_id; 7470 struct list_head *higher; 7471 struct napi_struct *pos; 7472 7473 new_id = UINT_MAX; 7474 if (napi->config && napi->config->napi_id) 7475 new_id = napi->config->napi_id; 7476 7477 higher = &dev->napi_list; 7478 list_for_each_entry(pos, &dev->napi_list, dev_list) { 7479 if (napi_id_valid(pos->napi_id)) 7480 pos_id = pos->napi_id; 7481 else if (pos->config) 7482 pos_id = pos->config->napi_id; 7483 else 7484 pos_id = UINT_MAX; 7485 7486 if (pos_id <= new_id) 7487 break; 7488 higher = &pos->dev_list; 7489 } 7490 list_add_rcu(&napi->dev_list, higher); /* adds after higher */ 7491 } 7492 7493 /* Double check that napi_get_frags() allocates skbs with 7494 * skb->head being backed by slab, not a page fragment. 7495 * This is to make sure bug fixed in 3226b158e67c 7496 * ("net: avoid 32 x truesize under-estimation for tiny skbs") 7497 * does not accidentally come back. 7498 */ 7499 static void napi_get_frags_check(struct napi_struct *napi) 7500 { 7501 struct sk_buff *skb; 7502 7503 local_bh_disable(); 7504 skb = napi_get_frags(napi); 7505 WARN_ON_ONCE(skb && skb->head_frag); 7506 napi_free_frags(napi); 7507 local_bh_enable(); 7508 } 7509 7510 void netif_napi_add_weight_locked(struct net_device *dev, 7511 struct napi_struct *napi, 7512 int (*poll)(struct napi_struct *, int), 7513 int weight) 7514 { 7515 netdev_assert_locked(dev); 7516 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state))) 7517 return; 7518 7519 INIT_LIST_HEAD(&napi->poll_list); 7520 INIT_HLIST_NODE(&napi->napi_hash_node); 7521 hrtimer_setup(&napi->timer, napi_watchdog, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); 7522 gro_init(&napi->gro); 7523 napi->skb = NULL; 7524 napi->poll = poll; 7525 if (weight > NAPI_POLL_WEIGHT) 7526 netdev_err_once(dev, "%s() called with weight %d\n", __func__, 7527 weight); 7528 napi->weight = weight; 7529 napi->dev = dev; 7530 #ifdef CONFIG_NETPOLL 7531 napi->poll_owner = -1; 7532 #endif 7533 napi->list_owner = -1; 7534 set_bit(NAPI_STATE_SCHED, &napi->state); 7535 set_bit(NAPI_STATE_NPSVC, &napi->state); 7536 netif_napi_dev_list_add(dev, napi); 7537 7538 /* default settings from sysfs are applied to all NAPIs. any per-NAPI 7539 * configuration will be loaded in napi_enable 7540 */ 7541 napi_set_defer_hard_irqs(napi, READ_ONCE(dev->napi_defer_hard_irqs)); 7542 napi_set_gro_flush_timeout(napi, READ_ONCE(dev->gro_flush_timeout)); 7543 7544 napi_get_frags_check(napi); 7545 /* Create kthread for this napi if dev->threaded is set. 7546 * Clear dev->threaded if kthread creation failed so that 7547 * threaded mode will not be enabled in napi_enable(). 7548 */ 7549 if (napi_get_threaded_config(dev, napi)) 7550 if (napi_kthread_create(napi)) 7551 dev->threaded = NETDEV_NAPI_THREADED_DISABLED; 7552 netif_napi_set_irq_locked(napi, -1); 7553 } 7554 EXPORT_SYMBOL(netif_napi_add_weight_locked); 7555 7556 void napi_disable_locked(struct napi_struct *n) 7557 { 7558 unsigned long val, new; 7559 7560 might_sleep(); 7561 netdev_assert_locked(n->dev); 7562 7563 set_bit(NAPI_STATE_DISABLE, &n->state); 7564 7565 val = READ_ONCE(n->state); 7566 do { 7567 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) { 7568 usleep_range(20, 200); 7569 val = READ_ONCE(n->state); 7570 } 7571 7572 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC; 7573 new &= ~(NAPIF_STATE_THREADED | 7574 NAPIF_STATE_THREADED_BUSY_POLL | 7575 NAPIF_STATE_PREFER_BUSY_POLL); 7576 } while (!try_cmpxchg(&n->state, &val, new)); 7577 7578 hrtimer_cancel(&n->timer); 7579 7580 if (n->config) 7581 napi_save_config(n); 7582 else 7583 napi_hash_del(n); 7584 7585 clear_bit(NAPI_STATE_DISABLE, &n->state); 7586 } 7587 EXPORT_SYMBOL(napi_disable_locked); 7588 7589 /** 7590 * napi_disable() - prevent NAPI from scheduling 7591 * @n: NAPI context 7592 * 7593 * Stop NAPI from being scheduled on this context. 7594 * Waits till any outstanding processing completes. 7595 * Takes netdev_lock() for associated net_device. 7596 */ 7597 void napi_disable(struct napi_struct *n) 7598 { 7599 netdev_lock(n->dev); 7600 napi_disable_locked(n); 7601 netdev_unlock(n->dev); 7602 } 7603 EXPORT_SYMBOL(napi_disable); 7604 7605 void napi_enable_locked(struct napi_struct *n) 7606 { 7607 unsigned long new, val = READ_ONCE(n->state); 7608 7609 if (n->config) 7610 napi_restore_config(n); 7611 else 7612 napi_hash_add(n); 7613 7614 do { 7615 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val)); 7616 7617 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC); 7618 if (n->dev->threaded && n->thread) 7619 new |= NAPIF_STATE_THREADED; 7620 } while (!try_cmpxchg(&n->state, &val, new)); 7621 } 7622 EXPORT_SYMBOL(napi_enable_locked); 7623 7624 /** 7625 * napi_enable() - enable NAPI scheduling 7626 * @n: NAPI context 7627 * 7628 * Enable scheduling of a NAPI instance. 7629 * Must be paired with napi_disable(). 7630 * Takes netdev_lock() for associated net_device. 7631 */ 7632 void napi_enable(struct napi_struct *n) 7633 { 7634 netdev_lock(n->dev); 7635 napi_enable_locked(n); 7636 netdev_unlock(n->dev); 7637 } 7638 EXPORT_SYMBOL(napi_enable); 7639 7640 /* Must be called in process context */ 7641 void __netif_napi_del_locked(struct napi_struct *napi) 7642 { 7643 netdev_assert_locked(napi->dev); 7644 7645 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state)) 7646 return; 7647 7648 /* Make sure NAPI is disabled (or was never enabled). */ 7649 WARN_ON(!test_bit(NAPI_STATE_SCHED, &napi->state)); 7650 7651 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7652 irq_set_affinity_notifier(napi->irq, NULL); 7653 7654 if (napi->config) { 7655 napi->index = -1; 7656 napi->config = NULL; 7657 } 7658 7659 list_del_rcu(&napi->dev_list); 7660 napi_free_frags(napi); 7661 7662 gro_cleanup(&napi->gro); 7663 7664 if (napi->thread) { 7665 kthread_stop(napi->thread); 7666 napi->thread = NULL; 7667 } 7668 } 7669 EXPORT_SYMBOL(__netif_napi_del_locked); 7670 7671 static int __napi_poll(struct napi_struct *n, bool *repoll) 7672 { 7673 int work, weight; 7674 7675 weight = n->weight; 7676 7677 /* This NAPI_STATE_SCHED test is for avoiding a race 7678 * with netpoll's poll_napi(). Only the entity which 7679 * obtains the lock and sees NAPI_STATE_SCHED set will 7680 * actually make the ->poll() call. Therefore we avoid 7681 * accidentally calling ->poll() when NAPI is not scheduled. 7682 */ 7683 work = 0; 7684 if (napi_is_scheduled(n)) { 7685 work = n->poll(n, weight); 7686 trace_napi_poll(n, work, weight); 7687 7688 xdp_do_check_flushed(n); 7689 } 7690 7691 if (unlikely(work > weight)) 7692 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n", 7693 n->poll, work, weight); 7694 7695 if (likely(work < weight)) 7696 return work; 7697 7698 /* Drivers must not modify the NAPI state if they 7699 * consume the entire weight. In such cases this code 7700 * still "owns" the NAPI instance and therefore can 7701 * move the instance around on the list at-will. 7702 */ 7703 if (unlikely(napi_disable_pending(n))) { 7704 napi_complete(n); 7705 return work; 7706 } 7707 7708 /* The NAPI context has more processing work, but busy-polling 7709 * is preferred. Exit early. 7710 */ 7711 if (napi_prefer_busy_poll(n)) { 7712 if (napi_complete_done(n, work)) { 7713 /* If timeout is not set, we need to make sure 7714 * that the NAPI is re-scheduled. 7715 */ 7716 napi_schedule(n); 7717 } 7718 return work; 7719 } 7720 7721 /* Flush too old packets. If HZ < 1000, flush all packets */ 7722 gro_flush_normal(&n->gro, HZ >= 1000); 7723 7724 /* Some drivers may have called napi_schedule 7725 * prior to exhausting their budget. 7726 */ 7727 if (unlikely(!list_empty(&n->poll_list))) { 7728 pr_warn_once("%s: Budget exhausted after napi rescheduled\n", 7729 n->dev ? n->dev->name : "backlog"); 7730 return work; 7731 } 7732 7733 *repoll = true; 7734 7735 return work; 7736 } 7737 7738 static int napi_poll(struct napi_struct *n, struct list_head *repoll) 7739 { 7740 bool do_repoll = false; 7741 void *have; 7742 int work; 7743 7744 list_del_init(&n->poll_list); 7745 7746 have = netpoll_poll_lock(n); 7747 7748 work = __napi_poll(n, &do_repoll); 7749 7750 if (do_repoll) { 7751 #if defined(CONFIG_DEBUG_NET) 7752 if (unlikely(!napi_is_scheduled(n))) 7753 pr_crit("repoll requested for device %s %ps but napi is not scheduled.\n", 7754 n->dev->name, n->poll); 7755 #endif 7756 list_add_tail(&n->poll_list, repoll); 7757 } 7758 netpoll_poll_unlock(have); 7759 7760 return work; 7761 } 7762 7763 static int napi_thread_wait(struct napi_struct *napi) 7764 { 7765 set_current_state(TASK_INTERRUPTIBLE); 7766 7767 while (!kthread_should_stop()) { 7768 /* Testing SCHED_THREADED bit here to make sure the current 7769 * kthread owns this napi and could poll on this napi. 7770 * Testing SCHED bit is not enough because SCHED bit might be 7771 * set by some other busy poll thread or by napi_disable(). 7772 */ 7773 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) { 7774 WARN_ON(!list_empty(&napi->poll_list)); 7775 __set_current_state(TASK_RUNNING); 7776 return 0; 7777 } 7778 7779 schedule(); 7780 set_current_state(TASK_INTERRUPTIBLE); 7781 } 7782 __set_current_state(TASK_RUNNING); 7783 7784 return -1; 7785 } 7786 7787 static void napi_threaded_poll_loop(struct napi_struct *napi, bool busy_poll) 7788 { 7789 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7790 struct softnet_data *sd; 7791 unsigned long last_qs = jiffies; 7792 7793 for (;;) { 7794 bool repoll = false; 7795 void *have; 7796 7797 local_bh_disable(); 7798 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7799 7800 sd = this_cpu_ptr(&softnet_data); 7801 sd->in_napi_threaded_poll = true; 7802 7803 have = netpoll_poll_lock(napi); 7804 __napi_poll(napi, &repoll); 7805 netpoll_poll_unlock(have); 7806 7807 sd->in_napi_threaded_poll = false; 7808 barrier(); 7809 7810 if (sd_has_rps_ipi_waiting(sd)) { 7811 local_irq_disable(); 7812 net_rps_action_and_irq_enable(sd); 7813 } 7814 skb_defer_free_flush(); 7815 bpf_net_ctx_clear(bpf_net_ctx); 7816 7817 /* When busy poll is enabled, the old packets are not flushed in 7818 * napi_complete_done. So flush them here. 7819 */ 7820 if (busy_poll) 7821 gro_flush_normal(&napi->gro, HZ >= 1000); 7822 local_bh_enable(); 7823 7824 /* Call cond_resched here to avoid watchdog warnings. */ 7825 if (repoll || busy_poll) { 7826 rcu_softirq_qs_periodic(last_qs); 7827 cond_resched(); 7828 } 7829 7830 if (!repoll) 7831 break; 7832 } 7833 } 7834 7835 static int napi_threaded_poll(void *data) 7836 { 7837 struct napi_struct *napi = data; 7838 bool want_busy_poll; 7839 bool in_busy_poll; 7840 unsigned long val; 7841 7842 while (!napi_thread_wait(napi)) { 7843 val = READ_ONCE(napi->state); 7844 7845 want_busy_poll = val & NAPIF_STATE_THREADED_BUSY_POLL; 7846 in_busy_poll = val & NAPIF_STATE_IN_BUSY_POLL; 7847 7848 if (unlikely(val & NAPIF_STATE_DISABLE)) 7849 want_busy_poll = false; 7850 7851 if (want_busy_poll != in_busy_poll) 7852 assign_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state, 7853 want_busy_poll); 7854 7855 napi_threaded_poll_loop(napi, want_busy_poll); 7856 } 7857 7858 return 0; 7859 } 7860 7861 static __latent_entropy void net_rx_action(void) 7862 { 7863 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 7864 unsigned long time_limit = jiffies + 7865 usecs_to_jiffies(READ_ONCE(net_hotdata.netdev_budget_usecs)); 7866 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7867 int budget = READ_ONCE(net_hotdata.netdev_budget); 7868 LIST_HEAD(list); 7869 LIST_HEAD(repoll); 7870 7871 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7872 start: 7873 sd->in_net_rx_action = true; 7874 local_irq_disable(); 7875 list_splice_init(&sd->poll_list, &list); 7876 local_irq_enable(); 7877 7878 for (;;) { 7879 struct napi_struct *n; 7880 7881 skb_defer_free_flush(); 7882 7883 if (list_empty(&list)) { 7884 if (list_empty(&repoll)) { 7885 sd->in_net_rx_action = false; 7886 barrier(); 7887 /* We need to check if ____napi_schedule() 7888 * had refilled poll_list while 7889 * sd->in_net_rx_action was true. 7890 */ 7891 if (!list_empty(&sd->poll_list)) 7892 goto start; 7893 if (!sd_has_rps_ipi_waiting(sd)) 7894 goto end; 7895 } 7896 break; 7897 } 7898 7899 n = list_first_entry(&list, struct napi_struct, poll_list); 7900 budget -= napi_poll(n, &repoll); 7901 7902 /* If softirq window is exhausted then punt. 7903 * Allow this to run for 2 jiffies since which will allow 7904 * an average latency of 1.5/HZ. 7905 */ 7906 if (unlikely(budget <= 0 || 7907 time_after_eq(jiffies, time_limit))) { 7908 /* Pairs with READ_ONCE() in softnet_seq_show() */ 7909 WRITE_ONCE(sd->time_squeeze, sd->time_squeeze + 1); 7910 break; 7911 } 7912 } 7913 7914 local_irq_disable(); 7915 7916 list_splice_tail_init(&sd->poll_list, &list); 7917 list_splice_tail(&repoll, &list); 7918 list_splice(&list, &sd->poll_list); 7919 if (!list_empty(&sd->poll_list)) 7920 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 7921 else 7922 sd->in_net_rx_action = false; 7923 7924 net_rps_action_and_irq_enable(sd); 7925 end: 7926 bpf_net_ctx_clear(bpf_net_ctx); 7927 } 7928 7929 struct netdev_adjacent { 7930 struct net_device *dev; 7931 netdevice_tracker dev_tracker; 7932 7933 /* upper master flag, there can only be one master device per list */ 7934 bool master; 7935 7936 /* lookup ignore flag */ 7937 bool ignore; 7938 7939 /* counter for the number of times this device was added to us */ 7940 u16 ref_nr; 7941 7942 /* private field for the users */ 7943 void *private; 7944 7945 struct list_head list; 7946 struct rcu_head rcu; 7947 }; 7948 7949 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev, 7950 struct list_head *adj_list) 7951 { 7952 struct netdev_adjacent *adj; 7953 7954 list_for_each_entry(adj, adj_list, list) { 7955 if (adj->dev == adj_dev) 7956 return adj; 7957 } 7958 return NULL; 7959 } 7960 7961 static int ____netdev_has_upper_dev(struct net_device *upper_dev, 7962 struct netdev_nested_priv *priv) 7963 { 7964 struct net_device *dev = (struct net_device *)priv->data; 7965 7966 return upper_dev == dev; 7967 } 7968 7969 /** 7970 * netdev_has_upper_dev - Check if device is linked to an upper device 7971 * @dev: device 7972 * @upper_dev: upper device to check 7973 * 7974 * Find out if a device is linked to specified upper device and return true 7975 * in case it is. Note that this checks only immediate upper device, 7976 * not through a complete stack of devices. The caller must hold the RTNL lock. 7977 */ 7978 bool netdev_has_upper_dev(struct net_device *dev, 7979 struct net_device *upper_dev) 7980 { 7981 struct netdev_nested_priv priv = { 7982 .data = (void *)upper_dev, 7983 }; 7984 7985 ASSERT_RTNL(); 7986 7987 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7988 &priv); 7989 } 7990 EXPORT_SYMBOL(netdev_has_upper_dev); 7991 7992 /** 7993 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device 7994 * @dev: device 7995 * @upper_dev: upper device to check 7996 * 7997 * Find out if a device is linked to specified upper device and return true 7998 * in case it is. Note that this checks the entire upper device chain. 7999 * The caller must hold rcu lock. 8000 */ 8001 8002 bool netdev_has_upper_dev_all_rcu(struct net_device *dev, 8003 struct net_device *upper_dev) 8004 { 8005 struct netdev_nested_priv priv = { 8006 .data = (void *)upper_dev, 8007 }; 8008 8009 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 8010 &priv); 8011 } 8012 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu); 8013 8014 /** 8015 * netdev_has_any_upper_dev - Check if device is linked to some device 8016 * @dev: device 8017 * 8018 * Find out if a device is linked to an upper device and return true in case 8019 * it is. The caller must hold the RTNL lock. 8020 */ 8021 bool netdev_has_any_upper_dev(struct net_device *dev) 8022 { 8023 ASSERT_RTNL(); 8024 8025 return !list_empty(&dev->adj_list.upper); 8026 } 8027 EXPORT_SYMBOL(netdev_has_any_upper_dev); 8028 8029 /** 8030 * netdev_master_upper_dev_get - Get master upper device 8031 * @dev: device 8032 * 8033 * Find a master upper device and return pointer to it or NULL in case 8034 * it's not there. The caller must hold the RTNL lock. 8035 */ 8036 struct net_device *netdev_master_upper_dev_get(struct net_device *dev) 8037 { 8038 struct netdev_adjacent *upper; 8039 8040 ASSERT_RTNL(); 8041 8042 if (list_empty(&dev->adj_list.upper)) 8043 return NULL; 8044 8045 upper = list_first_entry(&dev->adj_list.upper, 8046 struct netdev_adjacent, list); 8047 if (likely(upper->master)) 8048 return upper->dev; 8049 return NULL; 8050 } 8051 EXPORT_SYMBOL(netdev_master_upper_dev_get); 8052 8053 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev) 8054 { 8055 struct netdev_adjacent *upper; 8056 8057 ASSERT_RTNL(); 8058 8059 if (list_empty(&dev->adj_list.upper)) 8060 return NULL; 8061 8062 upper = list_first_entry(&dev->adj_list.upper, 8063 struct netdev_adjacent, list); 8064 if (likely(upper->master) && !upper->ignore) 8065 return upper->dev; 8066 return NULL; 8067 } 8068 8069 /** 8070 * netdev_has_any_lower_dev - Check if device is linked to some device 8071 * @dev: device 8072 * 8073 * Find out if a device is linked to a lower device and return true in case 8074 * it is. The caller must hold the RTNL lock. 8075 */ 8076 static bool netdev_has_any_lower_dev(struct net_device *dev) 8077 { 8078 ASSERT_RTNL(); 8079 8080 return !list_empty(&dev->adj_list.lower); 8081 } 8082 8083 void *netdev_adjacent_get_private(struct list_head *adj_list) 8084 { 8085 struct netdev_adjacent *adj; 8086 8087 adj = list_entry(adj_list, struct netdev_adjacent, list); 8088 8089 return adj->private; 8090 } 8091 EXPORT_SYMBOL(netdev_adjacent_get_private); 8092 8093 /** 8094 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list 8095 * @dev: device 8096 * @iter: list_head ** of the current position 8097 * 8098 * Gets the next device from the dev's upper list, starting from iter 8099 * position. The caller must hold RCU read lock. 8100 */ 8101 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev, 8102 struct list_head **iter) 8103 { 8104 struct netdev_adjacent *upper; 8105 8106 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 8107 8108 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8109 8110 if (&upper->list == &dev->adj_list.upper) 8111 return NULL; 8112 8113 *iter = &upper->list; 8114 8115 return upper->dev; 8116 } 8117 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu); 8118 8119 static struct net_device *__netdev_next_upper_dev(struct net_device *dev, 8120 struct list_head **iter, 8121 bool *ignore) 8122 { 8123 struct netdev_adjacent *upper; 8124 8125 upper = list_entry((*iter)->next, struct netdev_adjacent, list); 8126 8127 if (&upper->list == &dev->adj_list.upper) 8128 return NULL; 8129 8130 *iter = &upper->list; 8131 *ignore = upper->ignore; 8132 8133 return upper->dev; 8134 } 8135 8136 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev, 8137 struct list_head **iter) 8138 { 8139 struct netdev_adjacent *upper; 8140 8141 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 8142 8143 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8144 8145 if (&upper->list == &dev->adj_list.upper) 8146 return NULL; 8147 8148 *iter = &upper->list; 8149 8150 return upper->dev; 8151 } 8152 8153 static int __netdev_walk_all_upper_dev(struct net_device *dev, 8154 int (*fn)(struct net_device *dev, 8155 struct netdev_nested_priv *priv), 8156 struct netdev_nested_priv *priv) 8157 { 8158 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8159 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8160 int ret, cur = 0; 8161 bool ignore; 8162 8163 now = dev; 8164 iter = &dev->adj_list.upper; 8165 8166 while (1) { 8167 if (now != dev) { 8168 ret = fn(now, priv); 8169 if (ret) 8170 return ret; 8171 } 8172 8173 next = NULL; 8174 while (1) { 8175 udev = __netdev_next_upper_dev(now, &iter, &ignore); 8176 if (!udev) 8177 break; 8178 if (ignore) 8179 continue; 8180 8181 next = udev; 8182 niter = &udev->adj_list.upper; 8183 dev_stack[cur] = now; 8184 iter_stack[cur++] = iter; 8185 break; 8186 } 8187 8188 if (!next) { 8189 if (!cur) 8190 return 0; 8191 next = dev_stack[--cur]; 8192 niter = iter_stack[cur]; 8193 } 8194 8195 now = next; 8196 iter = niter; 8197 } 8198 8199 return 0; 8200 } 8201 8202 int netdev_walk_all_upper_dev_rcu(struct net_device *dev, 8203 int (*fn)(struct net_device *dev, 8204 struct netdev_nested_priv *priv), 8205 struct netdev_nested_priv *priv) 8206 { 8207 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8208 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8209 int ret, cur = 0; 8210 8211 now = dev; 8212 iter = &dev->adj_list.upper; 8213 8214 while (1) { 8215 if (now != dev) { 8216 ret = fn(now, priv); 8217 if (ret) 8218 return ret; 8219 } 8220 8221 next = NULL; 8222 while (1) { 8223 udev = netdev_next_upper_dev_rcu(now, &iter); 8224 if (!udev) 8225 break; 8226 8227 next = udev; 8228 niter = &udev->adj_list.upper; 8229 dev_stack[cur] = now; 8230 iter_stack[cur++] = iter; 8231 break; 8232 } 8233 8234 if (!next) { 8235 if (!cur) 8236 return 0; 8237 next = dev_stack[--cur]; 8238 niter = iter_stack[cur]; 8239 } 8240 8241 now = next; 8242 iter = niter; 8243 } 8244 8245 return 0; 8246 } 8247 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu); 8248 8249 static bool __netdev_has_upper_dev(struct net_device *dev, 8250 struct net_device *upper_dev) 8251 { 8252 struct netdev_nested_priv priv = { 8253 .flags = 0, 8254 .data = (void *)upper_dev, 8255 }; 8256 8257 ASSERT_RTNL(); 8258 8259 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev, 8260 &priv); 8261 } 8262 8263 /** 8264 * netdev_lower_get_next_private - Get the next ->private from the 8265 * lower neighbour list 8266 * @dev: device 8267 * @iter: list_head ** of the current position 8268 * 8269 * Gets the next netdev_adjacent->private from the dev's lower neighbour 8270 * list, starting from iter position. The caller must hold either hold the 8271 * RTNL lock or its own locking that guarantees that the neighbour lower 8272 * list will remain unchanged. 8273 */ 8274 void *netdev_lower_get_next_private(struct net_device *dev, 8275 struct list_head **iter) 8276 { 8277 struct netdev_adjacent *lower; 8278 8279 lower = list_entry(*iter, struct netdev_adjacent, list); 8280 8281 if (&lower->list == &dev->adj_list.lower) 8282 return NULL; 8283 8284 *iter = lower->list.next; 8285 8286 return lower->private; 8287 } 8288 EXPORT_SYMBOL(netdev_lower_get_next_private); 8289 8290 /** 8291 * netdev_lower_get_next_private_rcu - Get the next ->private from the 8292 * lower neighbour list, RCU 8293 * variant 8294 * @dev: device 8295 * @iter: list_head ** of the current position 8296 * 8297 * Gets the next netdev_adjacent->private from the dev's lower neighbour 8298 * list, starting from iter position. The caller must hold RCU read lock. 8299 */ 8300 void *netdev_lower_get_next_private_rcu(struct net_device *dev, 8301 struct list_head **iter) 8302 { 8303 struct netdev_adjacent *lower; 8304 8305 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 8306 8307 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8308 8309 if (&lower->list == &dev->adj_list.lower) 8310 return NULL; 8311 8312 *iter = &lower->list; 8313 8314 return lower->private; 8315 } 8316 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu); 8317 8318 /** 8319 * netdev_lower_get_next - Get the next device from the lower neighbour 8320 * list 8321 * @dev: device 8322 * @iter: list_head ** of the current position 8323 * 8324 * Gets the next netdev_adjacent from the dev's lower neighbour 8325 * list, starting from iter position. The caller must hold RTNL lock or 8326 * its own locking that guarantees that the neighbour lower 8327 * list will remain unchanged. 8328 */ 8329 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter) 8330 { 8331 struct netdev_adjacent *lower; 8332 8333 lower = list_entry(*iter, struct netdev_adjacent, list); 8334 8335 if (&lower->list == &dev->adj_list.lower) 8336 return NULL; 8337 8338 *iter = lower->list.next; 8339 8340 return lower->dev; 8341 } 8342 EXPORT_SYMBOL(netdev_lower_get_next); 8343 8344 static struct net_device *netdev_next_lower_dev(struct net_device *dev, 8345 struct list_head **iter) 8346 { 8347 struct netdev_adjacent *lower; 8348 8349 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 8350 8351 if (&lower->list == &dev->adj_list.lower) 8352 return NULL; 8353 8354 *iter = &lower->list; 8355 8356 return lower->dev; 8357 } 8358 8359 static struct net_device *__netdev_next_lower_dev(struct net_device *dev, 8360 struct list_head **iter, 8361 bool *ignore) 8362 { 8363 struct netdev_adjacent *lower; 8364 8365 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 8366 8367 if (&lower->list == &dev->adj_list.lower) 8368 return NULL; 8369 8370 *iter = &lower->list; 8371 *ignore = lower->ignore; 8372 8373 return lower->dev; 8374 } 8375 8376 int netdev_walk_all_lower_dev(struct net_device *dev, 8377 int (*fn)(struct net_device *dev, 8378 struct netdev_nested_priv *priv), 8379 struct netdev_nested_priv *priv) 8380 { 8381 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8382 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8383 int ret, cur = 0; 8384 8385 now = dev; 8386 iter = &dev->adj_list.lower; 8387 8388 while (1) { 8389 if (now != dev) { 8390 ret = fn(now, priv); 8391 if (ret) 8392 return ret; 8393 } 8394 8395 next = NULL; 8396 while (1) { 8397 ldev = netdev_next_lower_dev(now, &iter); 8398 if (!ldev) 8399 break; 8400 8401 next = ldev; 8402 niter = &ldev->adj_list.lower; 8403 dev_stack[cur] = now; 8404 iter_stack[cur++] = iter; 8405 break; 8406 } 8407 8408 if (!next) { 8409 if (!cur) 8410 return 0; 8411 next = dev_stack[--cur]; 8412 niter = iter_stack[cur]; 8413 } 8414 8415 now = next; 8416 iter = niter; 8417 } 8418 8419 return 0; 8420 } 8421 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev); 8422 8423 static int __netdev_walk_all_lower_dev(struct net_device *dev, 8424 int (*fn)(struct net_device *dev, 8425 struct netdev_nested_priv *priv), 8426 struct netdev_nested_priv *priv) 8427 { 8428 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8429 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8430 int ret, cur = 0; 8431 bool ignore; 8432 8433 now = dev; 8434 iter = &dev->adj_list.lower; 8435 8436 while (1) { 8437 if (now != dev) { 8438 ret = fn(now, priv); 8439 if (ret) 8440 return ret; 8441 } 8442 8443 next = NULL; 8444 while (1) { 8445 ldev = __netdev_next_lower_dev(now, &iter, &ignore); 8446 if (!ldev) 8447 break; 8448 if (ignore) 8449 continue; 8450 8451 next = ldev; 8452 niter = &ldev->adj_list.lower; 8453 dev_stack[cur] = now; 8454 iter_stack[cur++] = iter; 8455 break; 8456 } 8457 8458 if (!next) { 8459 if (!cur) 8460 return 0; 8461 next = dev_stack[--cur]; 8462 niter = iter_stack[cur]; 8463 } 8464 8465 now = next; 8466 iter = niter; 8467 } 8468 8469 return 0; 8470 } 8471 8472 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev, 8473 struct list_head **iter) 8474 { 8475 struct netdev_adjacent *lower; 8476 8477 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8478 if (&lower->list == &dev->adj_list.lower) 8479 return NULL; 8480 8481 *iter = &lower->list; 8482 8483 return lower->dev; 8484 } 8485 EXPORT_SYMBOL(netdev_next_lower_dev_rcu); 8486 8487 static u8 __netdev_upper_depth(struct net_device *dev) 8488 { 8489 struct net_device *udev; 8490 struct list_head *iter; 8491 u8 max_depth = 0; 8492 bool ignore; 8493 8494 for (iter = &dev->adj_list.upper, 8495 udev = __netdev_next_upper_dev(dev, &iter, &ignore); 8496 udev; 8497 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) { 8498 if (ignore) 8499 continue; 8500 if (max_depth < udev->upper_level) 8501 max_depth = udev->upper_level; 8502 } 8503 8504 return max_depth; 8505 } 8506 8507 static u8 __netdev_lower_depth(struct net_device *dev) 8508 { 8509 struct net_device *ldev; 8510 struct list_head *iter; 8511 u8 max_depth = 0; 8512 bool ignore; 8513 8514 for (iter = &dev->adj_list.lower, 8515 ldev = __netdev_next_lower_dev(dev, &iter, &ignore); 8516 ldev; 8517 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) { 8518 if (ignore) 8519 continue; 8520 if (max_depth < ldev->lower_level) 8521 max_depth = ldev->lower_level; 8522 } 8523 8524 return max_depth; 8525 } 8526 8527 static int __netdev_update_upper_level(struct net_device *dev, 8528 struct netdev_nested_priv *__unused) 8529 { 8530 dev->upper_level = __netdev_upper_depth(dev) + 1; 8531 return 0; 8532 } 8533 8534 #ifdef CONFIG_LOCKDEP 8535 static LIST_HEAD(net_unlink_list); 8536 8537 static void net_unlink_todo(struct net_device *dev) 8538 { 8539 if (list_empty(&dev->unlink_list)) 8540 list_add_tail(&dev->unlink_list, &net_unlink_list); 8541 } 8542 #endif 8543 8544 static int __netdev_update_lower_level(struct net_device *dev, 8545 struct netdev_nested_priv *priv) 8546 { 8547 dev->lower_level = __netdev_lower_depth(dev) + 1; 8548 8549 #ifdef CONFIG_LOCKDEP 8550 if (!priv) 8551 return 0; 8552 8553 if (priv->flags & NESTED_SYNC_IMM) 8554 dev->nested_level = dev->lower_level - 1; 8555 if (priv->flags & NESTED_SYNC_TODO) 8556 net_unlink_todo(dev); 8557 #endif 8558 return 0; 8559 } 8560 8561 int netdev_walk_all_lower_dev_rcu(struct net_device *dev, 8562 int (*fn)(struct net_device *dev, 8563 struct netdev_nested_priv *priv), 8564 struct netdev_nested_priv *priv) 8565 { 8566 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8567 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8568 int ret, cur = 0; 8569 8570 now = dev; 8571 iter = &dev->adj_list.lower; 8572 8573 while (1) { 8574 if (now != dev) { 8575 ret = fn(now, priv); 8576 if (ret) 8577 return ret; 8578 } 8579 8580 next = NULL; 8581 while (1) { 8582 ldev = netdev_next_lower_dev_rcu(now, &iter); 8583 if (!ldev) 8584 break; 8585 8586 next = ldev; 8587 niter = &ldev->adj_list.lower; 8588 dev_stack[cur] = now; 8589 iter_stack[cur++] = iter; 8590 break; 8591 } 8592 8593 if (!next) { 8594 if (!cur) 8595 return 0; 8596 next = dev_stack[--cur]; 8597 niter = iter_stack[cur]; 8598 } 8599 8600 now = next; 8601 iter = niter; 8602 } 8603 8604 return 0; 8605 } 8606 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu); 8607 8608 /** 8609 * netdev_lower_get_first_private_rcu - Get the first ->private from the 8610 * lower neighbour list, RCU 8611 * variant 8612 * @dev: device 8613 * 8614 * Gets the first netdev_adjacent->private from the dev's lower neighbour 8615 * list. The caller must hold RCU read lock. 8616 */ 8617 void *netdev_lower_get_first_private_rcu(struct net_device *dev) 8618 { 8619 struct netdev_adjacent *lower; 8620 8621 lower = list_first_or_null_rcu(&dev->adj_list.lower, 8622 struct netdev_adjacent, list); 8623 if (lower) 8624 return lower->private; 8625 return NULL; 8626 } 8627 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu); 8628 8629 /** 8630 * netdev_master_upper_dev_get_rcu - Get master upper device 8631 * @dev: device 8632 * 8633 * Find a master upper device and return pointer to it or NULL in case 8634 * it's not there. The caller must hold the RCU read lock. 8635 */ 8636 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev) 8637 { 8638 struct netdev_adjacent *upper; 8639 8640 upper = list_first_or_null_rcu(&dev->adj_list.upper, 8641 struct netdev_adjacent, list); 8642 if (upper && likely(upper->master)) 8643 return upper->dev; 8644 return NULL; 8645 } 8646 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu); 8647 8648 static int netdev_adjacent_sysfs_add(struct net_device *dev, 8649 struct net_device *adj_dev, 8650 struct list_head *dev_list) 8651 { 8652 char linkname[IFNAMSIZ+7]; 8653 8654 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8655 "upper_%s" : "lower_%s", adj_dev->name); 8656 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj), 8657 linkname); 8658 } 8659 static void netdev_adjacent_sysfs_del(struct net_device *dev, 8660 char *name, 8661 struct list_head *dev_list) 8662 { 8663 char linkname[IFNAMSIZ+7]; 8664 8665 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8666 "upper_%s" : "lower_%s", name); 8667 sysfs_remove_link(&(dev->dev.kobj), linkname); 8668 } 8669 8670 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev, 8671 struct net_device *adj_dev, 8672 struct list_head *dev_list) 8673 { 8674 return (dev_list == &dev->adj_list.upper || 8675 dev_list == &dev->adj_list.lower) && 8676 net_eq(dev_net(dev), dev_net(adj_dev)); 8677 } 8678 8679 static int __netdev_adjacent_dev_insert(struct net_device *dev, 8680 struct net_device *adj_dev, 8681 struct list_head *dev_list, 8682 void *private, bool master) 8683 { 8684 struct netdev_adjacent *adj; 8685 int ret; 8686 8687 adj = __netdev_find_adj(adj_dev, dev_list); 8688 8689 if (adj) { 8690 adj->ref_nr += 1; 8691 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n", 8692 dev->name, adj_dev->name, adj->ref_nr); 8693 8694 return 0; 8695 } 8696 8697 adj = kmalloc(sizeof(*adj), GFP_KERNEL); 8698 if (!adj) 8699 return -ENOMEM; 8700 8701 adj->dev = adj_dev; 8702 adj->master = master; 8703 adj->ref_nr = 1; 8704 adj->private = private; 8705 adj->ignore = false; 8706 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL); 8707 8708 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n", 8709 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name); 8710 8711 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) { 8712 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list); 8713 if (ret) 8714 goto free_adj; 8715 } 8716 8717 /* Ensure that master link is always the first item in list. */ 8718 if (master) { 8719 ret = sysfs_create_link(&(dev->dev.kobj), 8720 &(adj_dev->dev.kobj), "master"); 8721 if (ret) 8722 goto remove_symlinks; 8723 8724 list_add_rcu(&adj->list, dev_list); 8725 } else { 8726 list_add_tail_rcu(&adj->list, dev_list); 8727 } 8728 8729 return 0; 8730 8731 remove_symlinks: 8732 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8733 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8734 free_adj: 8735 netdev_put(adj_dev, &adj->dev_tracker); 8736 kfree(adj); 8737 8738 return ret; 8739 } 8740 8741 static void __netdev_adjacent_dev_remove(struct net_device *dev, 8742 struct net_device *adj_dev, 8743 u16 ref_nr, 8744 struct list_head *dev_list) 8745 { 8746 struct netdev_adjacent *adj; 8747 8748 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n", 8749 dev->name, adj_dev->name, ref_nr); 8750 8751 adj = __netdev_find_adj(adj_dev, dev_list); 8752 8753 if (!adj) { 8754 pr_err("Adjacency does not exist for device %s from %s\n", 8755 dev->name, adj_dev->name); 8756 WARN_ON(1); 8757 return; 8758 } 8759 8760 if (adj->ref_nr > ref_nr) { 8761 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n", 8762 dev->name, adj_dev->name, ref_nr, 8763 adj->ref_nr - ref_nr); 8764 adj->ref_nr -= ref_nr; 8765 return; 8766 } 8767 8768 if (adj->master) 8769 sysfs_remove_link(&(dev->dev.kobj), "master"); 8770 8771 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8772 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8773 8774 list_del_rcu(&adj->list); 8775 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n", 8776 adj_dev->name, dev->name, adj_dev->name); 8777 netdev_put(adj_dev, &adj->dev_tracker); 8778 kfree_rcu(adj, rcu); 8779 } 8780 8781 static int __netdev_adjacent_dev_link_lists(struct net_device *dev, 8782 struct net_device *upper_dev, 8783 struct list_head *up_list, 8784 struct list_head *down_list, 8785 void *private, bool master) 8786 { 8787 int ret; 8788 8789 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list, 8790 private, master); 8791 if (ret) 8792 return ret; 8793 8794 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list, 8795 private, false); 8796 if (ret) { 8797 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list); 8798 return ret; 8799 } 8800 8801 return 0; 8802 } 8803 8804 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev, 8805 struct net_device *upper_dev, 8806 u16 ref_nr, 8807 struct list_head *up_list, 8808 struct list_head *down_list) 8809 { 8810 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list); 8811 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list); 8812 } 8813 8814 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev, 8815 struct net_device *upper_dev, 8816 void *private, bool master) 8817 { 8818 return __netdev_adjacent_dev_link_lists(dev, upper_dev, 8819 &dev->adj_list.upper, 8820 &upper_dev->adj_list.lower, 8821 private, master); 8822 } 8823 8824 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev, 8825 struct net_device *upper_dev) 8826 { 8827 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1, 8828 &dev->adj_list.upper, 8829 &upper_dev->adj_list.lower); 8830 } 8831 8832 static int __netdev_upper_dev_link(struct net_device *dev, 8833 struct net_device *upper_dev, bool master, 8834 void *upper_priv, void *upper_info, 8835 struct netdev_nested_priv *priv, 8836 struct netlink_ext_ack *extack) 8837 { 8838 struct netdev_notifier_changeupper_info changeupper_info = { 8839 .info = { 8840 .dev = dev, 8841 .extack = extack, 8842 }, 8843 .upper_dev = upper_dev, 8844 .master = master, 8845 .linking = true, 8846 .upper_info = upper_info, 8847 }; 8848 struct net_device *master_dev; 8849 int ret = 0; 8850 8851 ASSERT_RTNL(); 8852 8853 if (dev == upper_dev) 8854 return -EBUSY; 8855 8856 /* To prevent loops, check if dev is not upper device to upper_dev. */ 8857 if (__netdev_has_upper_dev(upper_dev, dev)) 8858 return -EBUSY; 8859 8860 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV) 8861 return -EMLINK; 8862 8863 if (!master) { 8864 if (__netdev_has_upper_dev(dev, upper_dev)) 8865 return -EEXIST; 8866 } else { 8867 master_dev = __netdev_master_upper_dev_get(dev); 8868 if (master_dev) 8869 return master_dev == upper_dev ? -EEXIST : -EBUSY; 8870 } 8871 8872 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8873 &changeupper_info.info); 8874 ret = notifier_to_errno(ret); 8875 if (ret) 8876 return ret; 8877 8878 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv, 8879 master); 8880 if (ret) 8881 return ret; 8882 8883 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8884 &changeupper_info.info); 8885 ret = notifier_to_errno(ret); 8886 if (ret) 8887 goto rollback; 8888 8889 __netdev_update_upper_level(dev, NULL); 8890 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8891 8892 __netdev_update_lower_level(upper_dev, priv); 8893 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8894 priv); 8895 8896 return 0; 8897 8898 rollback: 8899 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8900 8901 return ret; 8902 } 8903 8904 /** 8905 * netdev_upper_dev_link - Add a link to the upper device 8906 * @dev: device 8907 * @upper_dev: new upper device 8908 * @extack: netlink extended ack 8909 * 8910 * Adds a link to device which is upper to this one. The caller must hold 8911 * the RTNL lock. On a failure a negative errno code is returned. 8912 * On success the reference counts are adjusted and the function 8913 * returns zero. 8914 */ 8915 int netdev_upper_dev_link(struct net_device *dev, 8916 struct net_device *upper_dev, 8917 struct netlink_ext_ack *extack) 8918 { 8919 struct netdev_nested_priv priv = { 8920 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8921 .data = NULL, 8922 }; 8923 8924 return __netdev_upper_dev_link(dev, upper_dev, false, 8925 NULL, NULL, &priv, extack); 8926 } 8927 EXPORT_SYMBOL(netdev_upper_dev_link); 8928 8929 /** 8930 * netdev_master_upper_dev_link - Add a master link to the upper device 8931 * @dev: device 8932 * @upper_dev: new upper device 8933 * @upper_priv: upper device private 8934 * @upper_info: upper info to be passed down via notifier 8935 * @extack: netlink extended ack 8936 * 8937 * Adds a link to device which is upper to this one. In this case, only 8938 * one master upper device can be linked, although other non-master devices 8939 * might be linked as well. The caller must hold the RTNL lock. 8940 * On a failure a negative errno code is returned. On success the reference 8941 * counts are adjusted and the function returns zero. 8942 */ 8943 int netdev_master_upper_dev_link(struct net_device *dev, 8944 struct net_device *upper_dev, 8945 void *upper_priv, void *upper_info, 8946 struct netlink_ext_ack *extack) 8947 { 8948 struct netdev_nested_priv priv = { 8949 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8950 .data = NULL, 8951 }; 8952 8953 return __netdev_upper_dev_link(dev, upper_dev, true, 8954 upper_priv, upper_info, &priv, extack); 8955 } 8956 EXPORT_SYMBOL(netdev_master_upper_dev_link); 8957 8958 static void __netdev_upper_dev_unlink(struct net_device *dev, 8959 struct net_device *upper_dev, 8960 struct netdev_nested_priv *priv) 8961 { 8962 struct netdev_notifier_changeupper_info changeupper_info = { 8963 .info = { 8964 .dev = dev, 8965 }, 8966 .upper_dev = upper_dev, 8967 .linking = false, 8968 }; 8969 8970 ASSERT_RTNL(); 8971 8972 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev; 8973 8974 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8975 &changeupper_info.info); 8976 8977 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8978 8979 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8980 &changeupper_info.info); 8981 8982 __netdev_update_upper_level(dev, NULL); 8983 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8984 8985 __netdev_update_lower_level(upper_dev, priv); 8986 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8987 priv); 8988 } 8989 8990 /** 8991 * netdev_upper_dev_unlink - Removes a link to upper device 8992 * @dev: device 8993 * @upper_dev: new upper device 8994 * 8995 * Removes a link to device which is upper to this one. The caller must hold 8996 * the RTNL lock. 8997 */ 8998 void netdev_upper_dev_unlink(struct net_device *dev, 8999 struct net_device *upper_dev) 9000 { 9001 struct netdev_nested_priv priv = { 9002 .flags = NESTED_SYNC_TODO, 9003 .data = NULL, 9004 }; 9005 9006 __netdev_upper_dev_unlink(dev, upper_dev, &priv); 9007 } 9008 EXPORT_SYMBOL(netdev_upper_dev_unlink); 9009 9010 static void __netdev_adjacent_dev_set(struct net_device *upper_dev, 9011 struct net_device *lower_dev, 9012 bool val) 9013 { 9014 struct netdev_adjacent *adj; 9015 9016 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower); 9017 if (adj) 9018 adj->ignore = val; 9019 9020 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper); 9021 if (adj) 9022 adj->ignore = val; 9023 } 9024 9025 static void netdev_adjacent_dev_disable(struct net_device *upper_dev, 9026 struct net_device *lower_dev) 9027 { 9028 __netdev_adjacent_dev_set(upper_dev, lower_dev, true); 9029 } 9030 9031 static void netdev_adjacent_dev_enable(struct net_device *upper_dev, 9032 struct net_device *lower_dev) 9033 { 9034 __netdev_adjacent_dev_set(upper_dev, lower_dev, false); 9035 } 9036 9037 int netdev_adjacent_change_prepare(struct net_device *old_dev, 9038 struct net_device *new_dev, 9039 struct net_device *dev, 9040 struct netlink_ext_ack *extack) 9041 { 9042 struct netdev_nested_priv priv = { 9043 .flags = 0, 9044 .data = NULL, 9045 }; 9046 int err; 9047 9048 if (!new_dev) 9049 return 0; 9050 9051 if (old_dev && new_dev != old_dev) 9052 netdev_adjacent_dev_disable(dev, old_dev); 9053 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv, 9054 extack); 9055 if (err) { 9056 if (old_dev && new_dev != old_dev) 9057 netdev_adjacent_dev_enable(dev, old_dev); 9058 return err; 9059 } 9060 9061 return 0; 9062 } 9063 EXPORT_SYMBOL(netdev_adjacent_change_prepare); 9064 9065 void netdev_adjacent_change_commit(struct net_device *old_dev, 9066 struct net_device *new_dev, 9067 struct net_device *dev) 9068 { 9069 struct netdev_nested_priv priv = { 9070 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 9071 .data = NULL, 9072 }; 9073 9074 if (!new_dev || !old_dev) 9075 return; 9076 9077 if (new_dev == old_dev) 9078 return; 9079 9080 netdev_adjacent_dev_enable(dev, old_dev); 9081 __netdev_upper_dev_unlink(old_dev, dev, &priv); 9082 } 9083 EXPORT_SYMBOL(netdev_adjacent_change_commit); 9084 9085 void netdev_adjacent_change_abort(struct net_device *old_dev, 9086 struct net_device *new_dev, 9087 struct net_device *dev) 9088 { 9089 struct netdev_nested_priv priv = { 9090 .flags = 0, 9091 .data = NULL, 9092 }; 9093 9094 if (!new_dev) 9095 return; 9096 9097 if (old_dev && new_dev != old_dev) 9098 netdev_adjacent_dev_enable(dev, old_dev); 9099 9100 __netdev_upper_dev_unlink(new_dev, dev, &priv); 9101 } 9102 EXPORT_SYMBOL(netdev_adjacent_change_abort); 9103 9104 /** 9105 * netdev_bonding_info_change - Dispatch event about slave change 9106 * @dev: device 9107 * @bonding_info: info to dispatch 9108 * 9109 * Send NETDEV_BONDING_INFO to netdev notifiers with info. 9110 * The caller must hold the RTNL lock. 9111 */ 9112 void netdev_bonding_info_change(struct net_device *dev, 9113 struct netdev_bonding_info *bonding_info) 9114 { 9115 struct netdev_notifier_bonding_info info = { 9116 .info.dev = dev, 9117 }; 9118 9119 memcpy(&info.bonding_info, bonding_info, 9120 sizeof(struct netdev_bonding_info)); 9121 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, 9122 &info.info); 9123 } 9124 EXPORT_SYMBOL(netdev_bonding_info_change); 9125 9126 static int netdev_offload_xstats_enable_l3(struct net_device *dev, 9127 struct netlink_ext_ack *extack) 9128 { 9129 struct netdev_notifier_offload_xstats_info info = { 9130 .info.dev = dev, 9131 .info.extack = extack, 9132 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 9133 }; 9134 int err; 9135 int rc; 9136 9137 dev->offload_xstats_l3 = kzalloc(sizeof(*dev->offload_xstats_l3), 9138 GFP_KERNEL); 9139 if (!dev->offload_xstats_l3) 9140 return -ENOMEM; 9141 9142 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE, 9143 NETDEV_OFFLOAD_XSTATS_DISABLE, 9144 &info.info); 9145 err = notifier_to_errno(rc); 9146 if (err) 9147 goto free_stats; 9148 9149 return 0; 9150 9151 free_stats: 9152 kfree(dev->offload_xstats_l3); 9153 dev->offload_xstats_l3 = NULL; 9154 return err; 9155 } 9156 9157 int netdev_offload_xstats_enable(struct net_device *dev, 9158 enum netdev_offload_xstats_type type, 9159 struct netlink_ext_ack *extack) 9160 { 9161 ASSERT_RTNL(); 9162 9163 if (netdev_offload_xstats_enabled(dev, type)) 9164 return -EALREADY; 9165 9166 switch (type) { 9167 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 9168 return netdev_offload_xstats_enable_l3(dev, extack); 9169 } 9170 9171 WARN_ON(1); 9172 return -EINVAL; 9173 } 9174 EXPORT_SYMBOL(netdev_offload_xstats_enable); 9175 9176 static void netdev_offload_xstats_disable_l3(struct net_device *dev) 9177 { 9178 struct netdev_notifier_offload_xstats_info info = { 9179 .info.dev = dev, 9180 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 9181 }; 9182 9183 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE, 9184 &info.info); 9185 kfree(dev->offload_xstats_l3); 9186 dev->offload_xstats_l3 = NULL; 9187 } 9188 9189 int netdev_offload_xstats_disable(struct net_device *dev, 9190 enum netdev_offload_xstats_type type) 9191 { 9192 ASSERT_RTNL(); 9193 9194 if (!netdev_offload_xstats_enabled(dev, type)) 9195 return -EALREADY; 9196 9197 switch (type) { 9198 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 9199 netdev_offload_xstats_disable_l3(dev); 9200 return 0; 9201 } 9202 9203 WARN_ON(1); 9204 return -EINVAL; 9205 } 9206 EXPORT_SYMBOL(netdev_offload_xstats_disable); 9207 9208 static void netdev_offload_xstats_disable_all(struct net_device *dev) 9209 { 9210 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3); 9211 } 9212 9213 static struct rtnl_hw_stats64 * 9214 netdev_offload_xstats_get_ptr(const struct net_device *dev, 9215 enum netdev_offload_xstats_type type) 9216 { 9217 switch (type) { 9218 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 9219 return dev->offload_xstats_l3; 9220 } 9221 9222 WARN_ON(1); 9223 return NULL; 9224 } 9225 9226 bool netdev_offload_xstats_enabled(const struct net_device *dev, 9227 enum netdev_offload_xstats_type type) 9228 { 9229 ASSERT_RTNL(); 9230 9231 return netdev_offload_xstats_get_ptr(dev, type); 9232 } 9233 EXPORT_SYMBOL(netdev_offload_xstats_enabled); 9234 9235 struct netdev_notifier_offload_xstats_ru { 9236 bool used; 9237 }; 9238 9239 struct netdev_notifier_offload_xstats_rd { 9240 struct rtnl_hw_stats64 stats; 9241 bool used; 9242 }; 9243 9244 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest, 9245 const struct rtnl_hw_stats64 *src) 9246 { 9247 dest->rx_packets += src->rx_packets; 9248 dest->tx_packets += src->tx_packets; 9249 dest->rx_bytes += src->rx_bytes; 9250 dest->tx_bytes += src->tx_bytes; 9251 dest->rx_errors += src->rx_errors; 9252 dest->tx_errors += src->tx_errors; 9253 dest->rx_dropped += src->rx_dropped; 9254 dest->tx_dropped += src->tx_dropped; 9255 dest->multicast += src->multicast; 9256 } 9257 9258 static int netdev_offload_xstats_get_used(struct net_device *dev, 9259 enum netdev_offload_xstats_type type, 9260 bool *p_used, 9261 struct netlink_ext_ack *extack) 9262 { 9263 struct netdev_notifier_offload_xstats_ru report_used = {}; 9264 struct netdev_notifier_offload_xstats_info info = { 9265 .info.dev = dev, 9266 .info.extack = extack, 9267 .type = type, 9268 .report_used = &report_used, 9269 }; 9270 int rc; 9271 9272 WARN_ON(!netdev_offload_xstats_enabled(dev, type)); 9273 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED, 9274 &info.info); 9275 *p_used = report_used.used; 9276 return notifier_to_errno(rc); 9277 } 9278 9279 static int netdev_offload_xstats_get_stats(struct net_device *dev, 9280 enum netdev_offload_xstats_type type, 9281 struct rtnl_hw_stats64 *p_stats, 9282 bool *p_used, 9283 struct netlink_ext_ack *extack) 9284 { 9285 struct netdev_notifier_offload_xstats_rd report_delta = {}; 9286 struct netdev_notifier_offload_xstats_info info = { 9287 .info.dev = dev, 9288 .info.extack = extack, 9289 .type = type, 9290 .report_delta = &report_delta, 9291 }; 9292 struct rtnl_hw_stats64 *stats; 9293 int rc; 9294 9295 stats = netdev_offload_xstats_get_ptr(dev, type); 9296 if (WARN_ON(!stats)) 9297 return -EINVAL; 9298 9299 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA, 9300 &info.info); 9301 9302 /* Cache whatever we got, even if there was an error, otherwise the 9303 * successful stats retrievals would get lost. 9304 */ 9305 netdev_hw_stats64_add(stats, &report_delta.stats); 9306 9307 if (p_stats) 9308 *p_stats = *stats; 9309 *p_used = report_delta.used; 9310 9311 return notifier_to_errno(rc); 9312 } 9313 9314 int netdev_offload_xstats_get(struct net_device *dev, 9315 enum netdev_offload_xstats_type type, 9316 struct rtnl_hw_stats64 *p_stats, bool *p_used, 9317 struct netlink_ext_ack *extack) 9318 { 9319 ASSERT_RTNL(); 9320 9321 if (p_stats) 9322 return netdev_offload_xstats_get_stats(dev, type, p_stats, 9323 p_used, extack); 9324 else 9325 return netdev_offload_xstats_get_used(dev, type, p_used, 9326 extack); 9327 } 9328 EXPORT_SYMBOL(netdev_offload_xstats_get); 9329 9330 void 9331 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta, 9332 const struct rtnl_hw_stats64 *stats) 9333 { 9334 report_delta->used = true; 9335 netdev_hw_stats64_add(&report_delta->stats, stats); 9336 } 9337 EXPORT_SYMBOL(netdev_offload_xstats_report_delta); 9338 9339 void 9340 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used) 9341 { 9342 report_used->used = true; 9343 } 9344 EXPORT_SYMBOL(netdev_offload_xstats_report_used); 9345 9346 void netdev_offload_xstats_push_delta(struct net_device *dev, 9347 enum netdev_offload_xstats_type type, 9348 const struct rtnl_hw_stats64 *p_stats) 9349 { 9350 struct rtnl_hw_stats64 *stats; 9351 9352 ASSERT_RTNL(); 9353 9354 stats = netdev_offload_xstats_get_ptr(dev, type); 9355 if (WARN_ON(!stats)) 9356 return; 9357 9358 netdev_hw_stats64_add(stats, p_stats); 9359 } 9360 EXPORT_SYMBOL(netdev_offload_xstats_push_delta); 9361 9362 /** 9363 * netdev_get_xmit_slave - Get the xmit slave of master device 9364 * @dev: device 9365 * @skb: The packet 9366 * @all_slaves: assume all the slaves are active 9367 * 9368 * The reference counters are not incremented so the caller must be 9369 * careful with locks. The caller must hold RCU lock. 9370 * %NULL is returned if no slave is found. 9371 */ 9372 9373 struct net_device *netdev_get_xmit_slave(struct net_device *dev, 9374 struct sk_buff *skb, 9375 bool all_slaves) 9376 { 9377 const struct net_device_ops *ops = dev->netdev_ops; 9378 9379 if (!ops->ndo_get_xmit_slave) 9380 return NULL; 9381 return ops->ndo_get_xmit_slave(dev, skb, all_slaves); 9382 } 9383 EXPORT_SYMBOL(netdev_get_xmit_slave); 9384 9385 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev, 9386 struct sock *sk) 9387 { 9388 const struct net_device_ops *ops = dev->netdev_ops; 9389 9390 if (!ops->ndo_sk_get_lower_dev) 9391 return NULL; 9392 return ops->ndo_sk_get_lower_dev(dev, sk); 9393 } 9394 9395 /** 9396 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket 9397 * @dev: device 9398 * @sk: the socket 9399 * 9400 * %NULL is returned if no lower device is found. 9401 */ 9402 9403 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev, 9404 struct sock *sk) 9405 { 9406 struct net_device *lower; 9407 9408 lower = netdev_sk_get_lower_dev(dev, sk); 9409 while (lower) { 9410 dev = lower; 9411 lower = netdev_sk_get_lower_dev(dev, sk); 9412 } 9413 9414 return dev; 9415 } 9416 EXPORT_SYMBOL(netdev_sk_get_lowest_dev); 9417 9418 static void netdev_adjacent_add_links(struct net_device *dev) 9419 { 9420 struct netdev_adjacent *iter; 9421 9422 struct net *net = dev_net(dev); 9423 9424 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9425 if (!net_eq(net, dev_net(iter->dev))) 9426 continue; 9427 netdev_adjacent_sysfs_add(iter->dev, dev, 9428 &iter->dev->adj_list.lower); 9429 netdev_adjacent_sysfs_add(dev, iter->dev, 9430 &dev->adj_list.upper); 9431 } 9432 9433 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9434 if (!net_eq(net, dev_net(iter->dev))) 9435 continue; 9436 netdev_adjacent_sysfs_add(iter->dev, dev, 9437 &iter->dev->adj_list.upper); 9438 netdev_adjacent_sysfs_add(dev, iter->dev, 9439 &dev->adj_list.lower); 9440 } 9441 } 9442 9443 static void netdev_adjacent_del_links(struct net_device *dev) 9444 { 9445 struct netdev_adjacent *iter; 9446 9447 struct net *net = dev_net(dev); 9448 9449 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9450 if (!net_eq(net, dev_net(iter->dev))) 9451 continue; 9452 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9453 &iter->dev->adj_list.lower); 9454 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9455 &dev->adj_list.upper); 9456 } 9457 9458 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9459 if (!net_eq(net, dev_net(iter->dev))) 9460 continue; 9461 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9462 &iter->dev->adj_list.upper); 9463 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9464 &dev->adj_list.lower); 9465 } 9466 } 9467 9468 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname) 9469 { 9470 struct netdev_adjacent *iter; 9471 9472 struct net *net = dev_net(dev); 9473 9474 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9475 if (!net_eq(net, dev_net(iter->dev))) 9476 continue; 9477 netdev_adjacent_sysfs_del(iter->dev, oldname, 9478 &iter->dev->adj_list.lower); 9479 netdev_adjacent_sysfs_add(iter->dev, dev, 9480 &iter->dev->adj_list.lower); 9481 } 9482 9483 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9484 if (!net_eq(net, dev_net(iter->dev))) 9485 continue; 9486 netdev_adjacent_sysfs_del(iter->dev, oldname, 9487 &iter->dev->adj_list.upper); 9488 netdev_adjacent_sysfs_add(iter->dev, dev, 9489 &iter->dev->adj_list.upper); 9490 } 9491 } 9492 9493 void *netdev_lower_dev_get_private(struct net_device *dev, 9494 struct net_device *lower_dev) 9495 { 9496 struct netdev_adjacent *lower; 9497 9498 if (!lower_dev) 9499 return NULL; 9500 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower); 9501 if (!lower) 9502 return NULL; 9503 9504 return lower->private; 9505 } 9506 EXPORT_SYMBOL(netdev_lower_dev_get_private); 9507 9508 9509 /** 9510 * netdev_lower_state_changed - Dispatch event about lower device state change 9511 * @lower_dev: device 9512 * @lower_state_info: state to dispatch 9513 * 9514 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info. 9515 * The caller must hold the RTNL lock. 9516 */ 9517 void netdev_lower_state_changed(struct net_device *lower_dev, 9518 void *lower_state_info) 9519 { 9520 struct netdev_notifier_changelowerstate_info changelowerstate_info = { 9521 .info.dev = lower_dev, 9522 }; 9523 9524 ASSERT_RTNL(); 9525 changelowerstate_info.lower_state_info = lower_state_info; 9526 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, 9527 &changelowerstate_info.info); 9528 } 9529 EXPORT_SYMBOL(netdev_lower_state_changed); 9530 9531 static void dev_change_rx_flags(struct net_device *dev, int flags) 9532 { 9533 const struct net_device_ops *ops = dev->netdev_ops; 9534 9535 if (ops->ndo_change_rx_flags) 9536 ops->ndo_change_rx_flags(dev, flags); 9537 } 9538 9539 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify) 9540 { 9541 unsigned int old_flags = dev->flags; 9542 unsigned int promiscuity, flags; 9543 kuid_t uid; 9544 kgid_t gid; 9545 9546 ASSERT_RTNL(); 9547 9548 promiscuity = dev->promiscuity + inc; 9549 if (promiscuity == 0) { 9550 /* 9551 * Avoid overflow. 9552 * If inc causes overflow, untouch promisc and return error. 9553 */ 9554 if (unlikely(inc > 0)) { 9555 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n"); 9556 return -EOVERFLOW; 9557 } 9558 flags = old_flags & ~IFF_PROMISC; 9559 } else { 9560 flags = old_flags | IFF_PROMISC; 9561 } 9562 WRITE_ONCE(dev->promiscuity, promiscuity); 9563 if (flags != old_flags) { 9564 WRITE_ONCE(dev->flags, flags); 9565 netdev_info(dev, "%s promiscuous mode\n", 9566 dev->flags & IFF_PROMISC ? "entered" : "left"); 9567 if (audit_enabled) { 9568 current_uid_gid(&uid, &gid); 9569 audit_log(audit_context(), GFP_ATOMIC, 9570 AUDIT_ANOM_PROMISCUOUS, 9571 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u", 9572 dev->name, (dev->flags & IFF_PROMISC), 9573 (old_flags & IFF_PROMISC), 9574 from_kuid(&init_user_ns, audit_get_loginuid(current)), 9575 from_kuid(&init_user_ns, uid), 9576 from_kgid(&init_user_ns, gid), 9577 audit_get_sessionid(current)); 9578 } 9579 9580 dev_change_rx_flags(dev, IFF_PROMISC); 9581 } 9582 if (notify) { 9583 /* The ops lock is only required to ensure consistent locking 9584 * for `NETDEV_CHANGE` notifiers. This function is sometimes 9585 * called without the lock, even for devices that are ops 9586 * locked, such as in `dev_uc_sync_multiple` when using 9587 * bonding or teaming. 9588 */ 9589 netdev_ops_assert_locked(dev); 9590 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL); 9591 } 9592 return 0; 9593 } 9594 9595 int netif_set_promiscuity(struct net_device *dev, int inc) 9596 { 9597 unsigned int old_flags = dev->flags; 9598 int err; 9599 9600 err = __dev_set_promiscuity(dev, inc, true); 9601 if (err < 0) 9602 return err; 9603 if (dev->flags != old_flags) 9604 dev_set_rx_mode(dev); 9605 return err; 9606 } 9607 9608 int netif_set_allmulti(struct net_device *dev, int inc, bool notify) 9609 { 9610 unsigned int old_flags = dev->flags, old_gflags = dev->gflags; 9611 unsigned int allmulti, flags; 9612 9613 ASSERT_RTNL(); 9614 9615 allmulti = dev->allmulti + inc; 9616 if (allmulti == 0) { 9617 /* 9618 * Avoid overflow. 9619 * If inc causes overflow, untouch allmulti and return error. 9620 */ 9621 if (unlikely(inc > 0)) { 9622 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n"); 9623 return -EOVERFLOW; 9624 } 9625 flags = old_flags & ~IFF_ALLMULTI; 9626 } else { 9627 flags = old_flags | IFF_ALLMULTI; 9628 } 9629 WRITE_ONCE(dev->allmulti, allmulti); 9630 if (flags != old_flags) { 9631 WRITE_ONCE(dev->flags, flags); 9632 netdev_info(dev, "%s allmulticast mode\n", 9633 dev->flags & IFF_ALLMULTI ? "entered" : "left"); 9634 dev_change_rx_flags(dev, IFF_ALLMULTI); 9635 dev_set_rx_mode(dev); 9636 if (notify) 9637 __dev_notify_flags(dev, old_flags, 9638 dev->gflags ^ old_gflags, 0, NULL); 9639 } 9640 return 0; 9641 } 9642 9643 /* 9644 * Upload unicast and multicast address lists to device and 9645 * configure RX filtering. When the device doesn't support unicast 9646 * filtering it is put in promiscuous mode while unicast addresses 9647 * are present. 9648 */ 9649 void __dev_set_rx_mode(struct net_device *dev) 9650 { 9651 const struct net_device_ops *ops = dev->netdev_ops; 9652 9653 /* dev_open will call this function so the list will stay sane. */ 9654 if (!(dev->flags&IFF_UP)) 9655 return; 9656 9657 if (!netif_device_present(dev)) 9658 return; 9659 9660 if (!(dev->priv_flags & IFF_UNICAST_FLT)) { 9661 /* Unicast addresses changes may only happen under the rtnl, 9662 * therefore calling __dev_set_promiscuity here is safe. 9663 */ 9664 if (!netdev_uc_empty(dev) && !dev->uc_promisc) { 9665 __dev_set_promiscuity(dev, 1, false); 9666 dev->uc_promisc = true; 9667 } else if (netdev_uc_empty(dev) && dev->uc_promisc) { 9668 __dev_set_promiscuity(dev, -1, false); 9669 dev->uc_promisc = false; 9670 } 9671 } 9672 9673 if (ops->ndo_set_rx_mode) 9674 ops->ndo_set_rx_mode(dev); 9675 } 9676 9677 void dev_set_rx_mode(struct net_device *dev) 9678 { 9679 netif_addr_lock_bh(dev); 9680 __dev_set_rx_mode(dev); 9681 netif_addr_unlock_bh(dev); 9682 } 9683 9684 /** 9685 * netif_get_flags() - get flags reported to userspace 9686 * @dev: device 9687 * 9688 * Get the combination of flag bits exported through APIs to userspace. 9689 */ 9690 unsigned int netif_get_flags(const struct net_device *dev) 9691 { 9692 unsigned int flags; 9693 9694 flags = (READ_ONCE(dev->flags) & ~(IFF_PROMISC | 9695 IFF_ALLMULTI | 9696 IFF_RUNNING | 9697 IFF_LOWER_UP | 9698 IFF_DORMANT)) | 9699 (READ_ONCE(dev->gflags) & (IFF_PROMISC | 9700 IFF_ALLMULTI)); 9701 9702 if (netif_running(dev)) { 9703 if (netif_oper_up(dev)) 9704 flags |= IFF_RUNNING; 9705 if (netif_carrier_ok(dev)) 9706 flags |= IFF_LOWER_UP; 9707 if (netif_dormant(dev)) 9708 flags |= IFF_DORMANT; 9709 } 9710 9711 return flags; 9712 } 9713 EXPORT_SYMBOL(netif_get_flags); 9714 9715 int __dev_change_flags(struct net_device *dev, unsigned int flags, 9716 struct netlink_ext_ack *extack) 9717 { 9718 unsigned int old_flags = dev->flags; 9719 int ret; 9720 9721 ASSERT_RTNL(); 9722 9723 /* 9724 * Set the flags on our device. 9725 */ 9726 9727 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP | 9728 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL | 9729 IFF_AUTOMEDIA)) | 9730 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC | 9731 IFF_ALLMULTI)); 9732 9733 /* 9734 * Load in the correct multicast list now the flags have changed. 9735 */ 9736 9737 if ((old_flags ^ flags) & IFF_MULTICAST) 9738 dev_change_rx_flags(dev, IFF_MULTICAST); 9739 9740 dev_set_rx_mode(dev); 9741 9742 /* 9743 * Have we downed the interface. We handle IFF_UP ourselves 9744 * according to user attempts to set it, rather than blindly 9745 * setting it. 9746 */ 9747 9748 ret = 0; 9749 if ((old_flags ^ flags) & IFF_UP) { 9750 if (old_flags & IFF_UP) 9751 __dev_close(dev); 9752 else 9753 ret = __dev_open(dev, extack); 9754 } 9755 9756 if ((flags ^ dev->gflags) & IFF_PROMISC) { 9757 int inc = (flags & IFF_PROMISC) ? 1 : -1; 9758 old_flags = dev->flags; 9759 9760 dev->gflags ^= IFF_PROMISC; 9761 9762 if (__dev_set_promiscuity(dev, inc, false) >= 0) 9763 if (dev->flags != old_flags) 9764 dev_set_rx_mode(dev); 9765 } 9766 9767 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI 9768 * is important. Some (broken) drivers set IFF_PROMISC, when 9769 * IFF_ALLMULTI is requested not asking us and not reporting. 9770 */ 9771 if ((flags ^ dev->gflags) & IFF_ALLMULTI) { 9772 int inc = (flags & IFF_ALLMULTI) ? 1 : -1; 9773 9774 dev->gflags ^= IFF_ALLMULTI; 9775 netif_set_allmulti(dev, inc, false); 9776 } 9777 9778 return ret; 9779 } 9780 9781 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags, 9782 unsigned int gchanges, u32 portid, 9783 const struct nlmsghdr *nlh) 9784 { 9785 unsigned int changes = dev->flags ^ old_flags; 9786 9787 if (gchanges) 9788 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh); 9789 9790 if (changes & IFF_UP) { 9791 if (dev->flags & IFF_UP) 9792 call_netdevice_notifiers(NETDEV_UP, dev); 9793 else 9794 call_netdevice_notifiers(NETDEV_DOWN, dev); 9795 } 9796 9797 if (dev->flags & IFF_UP && 9798 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) { 9799 struct netdev_notifier_change_info change_info = { 9800 .info = { 9801 .dev = dev, 9802 }, 9803 .flags_changed = changes, 9804 }; 9805 9806 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info); 9807 } 9808 } 9809 9810 int netif_change_flags(struct net_device *dev, unsigned int flags, 9811 struct netlink_ext_ack *extack) 9812 { 9813 int ret; 9814 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags; 9815 9816 ret = __dev_change_flags(dev, flags, extack); 9817 if (ret < 0) 9818 return ret; 9819 9820 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags); 9821 __dev_notify_flags(dev, old_flags, changes, 0, NULL); 9822 return ret; 9823 } 9824 9825 int __netif_set_mtu(struct net_device *dev, int new_mtu) 9826 { 9827 const struct net_device_ops *ops = dev->netdev_ops; 9828 9829 if (ops->ndo_change_mtu) 9830 return ops->ndo_change_mtu(dev, new_mtu); 9831 9832 /* Pairs with all the lockless reads of dev->mtu in the stack */ 9833 WRITE_ONCE(dev->mtu, new_mtu); 9834 return 0; 9835 } 9836 EXPORT_SYMBOL_NS_GPL(__netif_set_mtu, "NETDEV_INTERNAL"); 9837 9838 int dev_validate_mtu(struct net_device *dev, int new_mtu, 9839 struct netlink_ext_ack *extack) 9840 { 9841 /* MTU must be positive, and in range */ 9842 if (new_mtu < 0 || new_mtu < dev->min_mtu) { 9843 NL_SET_ERR_MSG(extack, "mtu less than device minimum"); 9844 return -EINVAL; 9845 } 9846 9847 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) { 9848 NL_SET_ERR_MSG(extack, "mtu greater than device maximum"); 9849 return -EINVAL; 9850 } 9851 return 0; 9852 } 9853 9854 /** 9855 * netif_set_mtu_ext() - Change maximum transfer unit 9856 * @dev: device 9857 * @new_mtu: new transfer unit 9858 * @extack: netlink extended ack 9859 * 9860 * Change the maximum transfer size of the network device. 9861 * 9862 * Return: 0 on success, -errno on failure. 9863 */ 9864 int netif_set_mtu_ext(struct net_device *dev, int new_mtu, 9865 struct netlink_ext_ack *extack) 9866 { 9867 int err, orig_mtu; 9868 9869 netdev_ops_assert_locked(dev); 9870 9871 if (new_mtu == dev->mtu) 9872 return 0; 9873 9874 err = dev_validate_mtu(dev, new_mtu, extack); 9875 if (err) 9876 return err; 9877 9878 if (!netif_device_present(dev)) 9879 return -ENODEV; 9880 9881 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev); 9882 err = notifier_to_errno(err); 9883 if (err) 9884 return err; 9885 9886 orig_mtu = dev->mtu; 9887 err = __netif_set_mtu(dev, new_mtu); 9888 9889 if (!err) { 9890 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9891 orig_mtu); 9892 err = notifier_to_errno(err); 9893 if (err) { 9894 /* setting mtu back and notifying everyone again, 9895 * so that they have a chance to revert changes. 9896 */ 9897 __netif_set_mtu(dev, orig_mtu); 9898 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9899 new_mtu); 9900 } 9901 } 9902 return err; 9903 } 9904 9905 int netif_set_mtu(struct net_device *dev, int new_mtu) 9906 { 9907 struct netlink_ext_ack extack; 9908 int err; 9909 9910 memset(&extack, 0, sizeof(extack)); 9911 err = netif_set_mtu_ext(dev, new_mtu, &extack); 9912 if (err && extack._msg) 9913 net_err_ratelimited("%s: %s\n", dev->name, extack._msg); 9914 return err; 9915 } 9916 EXPORT_SYMBOL(netif_set_mtu); 9917 9918 int netif_change_tx_queue_len(struct net_device *dev, unsigned long new_len) 9919 { 9920 unsigned int orig_len = dev->tx_queue_len; 9921 int res; 9922 9923 if (new_len != (unsigned int)new_len) 9924 return -ERANGE; 9925 9926 if (new_len != orig_len) { 9927 WRITE_ONCE(dev->tx_queue_len, new_len); 9928 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev); 9929 res = notifier_to_errno(res); 9930 if (res) 9931 goto err_rollback; 9932 res = dev_qdisc_change_tx_queue_len(dev); 9933 if (res) 9934 goto err_rollback; 9935 } 9936 9937 return 0; 9938 9939 err_rollback: 9940 netdev_err(dev, "refused to change device tx_queue_len\n"); 9941 WRITE_ONCE(dev->tx_queue_len, orig_len); 9942 return res; 9943 } 9944 9945 void netif_set_group(struct net_device *dev, int new_group) 9946 { 9947 dev->group = new_group; 9948 } 9949 9950 /** 9951 * netif_pre_changeaddr_notify() - Call NETDEV_PRE_CHANGEADDR. 9952 * @dev: device 9953 * @addr: new address 9954 * @extack: netlink extended ack 9955 * 9956 * Return: 0 on success, -errno on failure. 9957 */ 9958 int netif_pre_changeaddr_notify(struct net_device *dev, const char *addr, 9959 struct netlink_ext_ack *extack) 9960 { 9961 struct netdev_notifier_pre_changeaddr_info info = { 9962 .info.dev = dev, 9963 .info.extack = extack, 9964 .dev_addr = addr, 9965 }; 9966 int rc; 9967 9968 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info); 9969 return notifier_to_errno(rc); 9970 } 9971 EXPORT_SYMBOL_NS_GPL(netif_pre_changeaddr_notify, "NETDEV_INTERNAL"); 9972 9973 int netif_set_mac_address(struct net_device *dev, struct sockaddr_storage *ss, 9974 struct netlink_ext_ack *extack) 9975 { 9976 const struct net_device_ops *ops = dev->netdev_ops; 9977 int err; 9978 9979 if (!ops->ndo_set_mac_address) 9980 return -EOPNOTSUPP; 9981 if (ss->ss_family != dev->type) 9982 return -EINVAL; 9983 if (!netif_device_present(dev)) 9984 return -ENODEV; 9985 err = netif_pre_changeaddr_notify(dev, ss->__data, extack); 9986 if (err) 9987 return err; 9988 if (memcmp(dev->dev_addr, ss->__data, dev->addr_len)) { 9989 err = ops->ndo_set_mac_address(dev, ss); 9990 if (err) 9991 return err; 9992 } 9993 dev->addr_assign_type = NET_ADDR_SET; 9994 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev); 9995 add_device_randomness(dev->dev_addr, dev->addr_len); 9996 return 0; 9997 } 9998 9999 DECLARE_RWSEM(dev_addr_sem); 10000 10001 /* "sa" is a true struct sockaddr with limited "sa_data" member. */ 10002 int netif_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name) 10003 { 10004 size_t size = sizeof(sa->sa_data); 10005 struct net_device *dev; 10006 int ret = 0; 10007 10008 down_read(&dev_addr_sem); 10009 rcu_read_lock(); 10010 10011 dev = dev_get_by_name_rcu(net, dev_name); 10012 if (!dev) { 10013 ret = -ENODEV; 10014 goto unlock; 10015 } 10016 if (!dev->addr_len) 10017 memset(sa->sa_data, 0, size); 10018 else 10019 memcpy(sa->sa_data, dev->dev_addr, 10020 min_t(size_t, size, dev->addr_len)); 10021 sa->sa_family = dev->type; 10022 10023 unlock: 10024 rcu_read_unlock(); 10025 up_read(&dev_addr_sem); 10026 return ret; 10027 } 10028 EXPORT_SYMBOL_NS_GPL(netif_get_mac_address, "NETDEV_INTERNAL"); 10029 10030 int netif_change_carrier(struct net_device *dev, bool new_carrier) 10031 { 10032 const struct net_device_ops *ops = dev->netdev_ops; 10033 10034 if (!ops->ndo_change_carrier) 10035 return -EOPNOTSUPP; 10036 if (!netif_device_present(dev)) 10037 return -ENODEV; 10038 return ops->ndo_change_carrier(dev, new_carrier); 10039 } 10040 10041 /** 10042 * dev_get_phys_port_id - Get device physical port ID 10043 * @dev: device 10044 * @ppid: port ID 10045 * 10046 * Get device physical port ID 10047 */ 10048 int dev_get_phys_port_id(struct net_device *dev, 10049 struct netdev_phys_item_id *ppid) 10050 { 10051 const struct net_device_ops *ops = dev->netdev_ops; 10052 10053 if (!ops->ndo_get_phys_port_id) 10054 return -EOPNOTSUPP; 10055 return ops->ndo_get_phys_port_id(dev, ppid); 10056 } 10057 10058 /** 10059 * dev_get_phys_port_name - Get device physical port name 10060 * @dev: device 10061 * @name: port name 10062 * @len: limit of bytes to copy to name 10063 * 10064 * Get device physical port name 10065 */ 10066 int dev_get_phys_port_name(struct net_device *dev, 10067 char *name, size_t len) 10068 { 10069 const struct net_device_ops *ops = dev->netdev_ops; 10070 int err; 10071 10072 if (ops->ndo_get_phys_port_name) { 10073 err = ops->ndo_get_phys_port_name(dev, name, len); 10074 if (err != -EOPNOTSUPP) 10075 return err; 10076 } 10077 return devlink_compat_phys_port_name_get(dev, name, len); 10078 } 10079 10080 /** 10081 * netif_get_port_parent_id() - Get the device's port parent identifier 10082 * @dev: network device 10083 * @ppid: pointer to a storage for the port's parent identifier 10084 * @recurse: allow/disallow recursion to lower devices 10085 * 10086 * Get the devices's port parent identifier. 10087 * 10088 * Return: 0 on success, -errno on failure. 10089 */ 10090 int netif_get_port_parent_id(struct net_device *dev, 10091 struct netdev_phys_item_id *ppid, bool recurse) 10092 { 10093 const struct net_device_ops *ops = dev->netdev_ops; 10094 struct netdev_phys_item_id first = { }; 10095 struct net_device *lower_dev; 10096 struct list_head *iter; 10097 int err; 10098 10099 if (ops->ndo_get_port_parent_id) { 10100 err = ops->ndo_get_port_parent_id(dev, ppid); 10101 if (err != -EOPNOTSUPP) 10102 return err; 10103 } 10104 10105 err = devlink_compat_switch_id_get(dev, ppid); 10106 if (!recurse || err != -EOPNOTSUPP) 10107 return err; 10108 10109 netdev_for_each_lower_dev(dev, lower_dev, iter) { 10110 err = netif_get_port_parent_id(lower_dev, ppid, true); 10111 if (err) 10112 break; 10113 if (!first.id_len) 10114 first = *ppid; 10115 else if (memcmp(&first, ppid, sizeof(*ppid))) 10116 return -EOPNOTSUPP; 10117 } 10118 10119 return err; 10120 } 10121 EXPORT_SYMBOL(netif_get_port_parent_id); 10122 10123 /** 10124 * netdev_port_same_parent_id - Indicate if two network devices have 10125 * the same port parent identifier 10126 * @a: first network device 10127 * @b: second network device 10128 */ 10129 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b) 10130 { 10131 struct netdev_phys_item_id a_id = { }; 10132 struct netdev_phys_item_id b_id = { }; 10133 10134 if (netif_get_port_parent_id(a, &a_id, true) || 10135 netif_get_port_parent_id(b, &b_id, true)) 10136 return false; 10137 10138 return netdev_phys_item_id_same(&a_id, &b_id); 10139 } 10140 EXPORT_SYMBOL(netdev_port_same_parent_id); 10141 10142 int netif_change_proto_down(struct net_device *dev, bool proto_down) 10143 { 10144 if (!dev->change_proto_down) 10145 return -EOPNOTSUPP; 10146 if (!netif_device_present(dev)) 10147 return -ENODEV; 10148 if (proto_down) 10149 netif_carrier_off(dev); 10150 else 10151 netif_carrier_on(dev); 10152 WRITE_ONCE(dev->proto_down, proto_down); 10153 return 0; 10154 } 10155 10156 /** 10157 * netdev_change_proto_down_reason_locked - proto down reason 10158 * 10159 * @dev: device 10160 * @mask: proto down mask 10161 * @value: proto down value 10162 */ 10163 void netdev_change_proto_down_reason_locked(struct net_device *dev, 10164 unsigned long mask, u32 value) 10165 { 10166 u32 proto_down_reason; 10167 int b; 10168 10169 if (!mask) { 10170 proto_down_reason = value; 10171 } else { 10172 proto_down_reason = dev->proto_down_reason; 10173 for_each_set_bit(b, &mask, 32) { 10174 if (value & (1 << b)) 10175 proto_down_reason |= BIT(b); 10176 else 10177 proto_down_reason &= ~BIT(b); 10178 } 10179 } 10180 WRITE_ONCE(dev->proto_down_reason, proto_down_reason); 10181 } 10182 10183 struct bpf_xdp_link { 10184 struct bpf_link link; 10185 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */ 10186 int flags; 10187 }; 10188 10189 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags) 10190 { 10191 if (flags & XDP_FLAGS_HW_MODE) 10192 return XDP_MODE_HW; 10193 if (flags & XDP_FLAGS_DRV_MODE) 10194 return XDP_MODE_DRV; 10195 if (flags & XDP_FLAGS_SKB_MODE) 10196 return XDP_MODE_SKB; 10197 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB; 10198 } 10199 10200 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode) 10201 { 10202 switch (mode) { 10203 case XDP_MODE_SKB: 10204 return generic_xdp_install; 10205 case XDP_MODE_DRV: 10206 case XDP_MODE_HW: 10207 return dev->netdev_ops->ndo_bpf; 10208 default: 10209 return NULL; 10210 } 10211 } 10212 10213 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev, 10214 enum bpf_xdp_mode mode) 10215 { 10216 return dev->xdp_state[mode].link; 10217 } 10218 10219 static struct bpf_prog *dev_xdp_prog(struct net_device *dev, 10220 enum bpf_xdp_mode mode) 10221 { 10222 struct bpf_xdp_link *link = dev_xdp_link(dev, mode); 10223 10224 if (link) 10225 return link->link.prog; 10226 return dev->xdp_state[mode].prog; 10227 } 10228 10229 u8 dev_xdp_prog_count(struct net_device *dev) 10230 { 10231 u8 count = 0; 10232 int i; 10233 10234 for (i = 0; i < __MAX_XDP_MODE; i++) 10235 if (dev->xdp_state[i].prog || dev->xdp_state[i].link) 10236 count++; 10237 return count; 10238 } 10239 EXPORT_SYMBOL_GPL(dev_xdp_prog_count); 10240 10241 u8 dev_xdp_sb_prog_count(struct net_device *dev) 10242 { 10243 u8 count = 0; 10244 int i; 10245 10246 for (i = 0; i < __MAX_XDP_MODE; i++) 10247 if (dev->xdp_state[i].prog && 10248 !dev->xdp_state[i].prog->aux->xdp_has_frags) 10249 count++; 10250 return count; 10251 } 10252 10253 int netif_xdp_propagate(struct net_device *dev, struct netdev_bpf *bpf) 10254 { 10255 if (!dev->netdev_ops->ndo_bpf) 10256 return -EOPNOTSUPP; 10257 10258 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 10259 bpf->command == XDP_SETUP_PROG && 10260 bpf->prog && !bpf->prog->aux->xdp_has_frags) { 10261 NL_SET_ERR_MSG(bpf->extack, 10262 "unable to propagate XDP to device using tcp-data-split"); 10263 return -EBUSY; 10264 } 10265 10266 if (dev_get_min_mp_channel_count(dev)) { 10267 NL_SET_ERR_MSG(bpf->extack, "unable to propagate XDP to device using memory provider"); 10268 return -EBUSY; 10269 } 10270 10271 return dev->netdev_ops->ndo_bpf(dev, bpf); 10272 } 10273 EXPORT_SYMBOL_GPL(netif_xdp_propagate); 10274 10275 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode) 10276 { 10277 struct bpf_prog *prog = dev_xdp_prog(dev, mode); 10278 10279 return prog ? prog->aux->id : 0; 10280 } 10281 10282 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode, 10283 struct bpf_xdp_link *link) 10284 { 10285 dev->xdp_state[mode].link = link; 10286 dev->xdp_state[mode].prog = NULL; 10287 } 10288 10289 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode, 10290 struct bpf_prog *prog) 10291 { 10292 dev->xdp_state[mode].link = NULL; 10293 dev->xdp_state[mode].prog = prog; 10294 } 10295 10296 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode, 10297 bpf_op_t bpf_op, struct netlink_ext_ack *extack, 10298 u32 flags, struct bpf_prog *prog) 10299 { 10300 struct netdev_bpf xdp; 10301 int err; 10302 10303 netdev_ops_assert_locked(dev); 10304 10305 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 10306 prog && !prog->aux->xdp_has_frags) { 10307 NL_SET_ERR_MSG(extack, "unable to install XDP to device using tcp-data-split"); 10308 return -EBUSY; 10309 } 10310 10311 if (dev_get_min_mp_channel_count(dev)) { 10312 NL_SET_ERR_MSG(extack, "unable to install XDP to device using memory provider"); 10313 return -EBUSY; 10314 } 10315 10316 memset(&xdp, 0, sizeof(xdp)); 10317 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG; 10318 xdp.extack = extack; 10319 xdp.flags = flags; 10320 xdp.prog = prog; 10321 10322 /* Drivers assume refcnt is already incremented (i.e, prog pointer is 10323 * "moved" into driver), so they don't increment it on their own, but 10324 * they do decrement refcnt when program is detached or replaced. 10325 * Given net_device also owns link/prog, we need to bump refcnt here 10326 * to prevent drivers from underflowing it. 10327 */ 10328 if (prog) 10329 bpf_prog_inc(prog); 10330 err = bpf_op(dev, &xdp); 10331 if (err) { 10332 if (prog) 10333 bpf_prog_put(prog); 10334 return err; 10335 } 10336 10337 if (mode != XDP_MODE_HW) 10338 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog); 10339 10340 return 0; 10341 } 10342 10343 static void dev_xdp_uninstall(struct net_device *dev) 10344 { 10345 struct bpf_xdp_link *link; 10346 struct bpf_prog *prog; 10347 enum bpf_xdp_mode mode; 10348 bpf_op_t bpf_op; 10349 10350 ASSERT_RTNL(); 10351 10352 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) { 10353 prog = dev_xdp_prog(dev, mode); 10354 if (!prog) 10355 continue; 10356 10357 bpf_op = dev_xdp_bpf_op(dev, mode); 10358 if (!bpf_op) 10359 continue; 10360 10361 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 10362 10363 /* auto-detach link from net device */ 10364 link = dev_xdp_link(dev, mode); 10365 if (link) 10366 link->dev = NULL; 10367 else 10368 bpf_prog_put(prog); 10369 10370 dev_xdp_set_link(dev, mode, NULL); 10371 } 10372 } 10373 10374 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack, 10375 struct bpf_xdp_link *link, struct bpf_prog *new_prog, 10376 struct bpf_prog *old_prog, u32 flags) 10377 { 10378 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES); 10379 struct bpf_prog *cur_prog; 10380 struct net_device *upper; 10381 struct list_head *iter; 10382 enum bpf_xdp_mode mode; 10383 bpf_op_t bpf_op; 10384 int err; 10385 10386 ASSERT_RTNL(); 10387 10388 /* either link or prog attachment, never both */ 10389 if (link && (new_prog || old_prog)) 10390 return -EINVAL; 10391 /* link supports only XDP mode flags */ 10392 if (link && (flags & ~XDP_FLAGS_MODES)) { 10393 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment"); 10394 return -EINVAL; 10395 } 10396 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */ 10397 if (num_modes > 1) { 10398 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set"); 10399 return -EINVAL; 10400 } 10401 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */ 10402 if (!num_modes && dev_xdp_prog_count(dev) > 1) { 10403 NL_SET_ERR_MSG(extack, 10404 "More than one program loaded, unset mode is ambiguous"); 10405 return -EINVAL; 10406 } 10407 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */ 10408 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) { 10409 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified"); 10410 return -EINVAL; 10411 } 10412 10413 mode = dev_xdp_mode(dev, flags); 10414 /* can't replace attached link */ 10415 if (dev_xdp_link(dev, mode)) { 10416 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link"); 10417 return -EBUSY; 10418 } 10419 10420 /* don't allow if an upper device already has a program */ 10421 netdev_for_each_upper_dev_rcu(dev, upper, iter) { 10422 if (dev_xdp_prog_count(upper) > 0) { 10423 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program"); 10424 return -EEXIST; 10425 } 10426 } 10427 10428 cur_prog = dev_xdp_prog(dev, mode); 10429 /* can't replace attached prog with link */ 10430 if (link && cur_prog) { 10431 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link"); 10432 return -EBUSY; 10433 } 10434 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) { 10435 NL_SET_ERR_MSG(extack, "Active program does not match expected"); 10436 return -EEXIST; 10437 } 10438 10439 /* put effective new program into new_prog */ 10440 if (link) 10441 new_prog = link->link.prog; 10442 10443 if (new_prog) { 10444 bool offload = mode == XDP_MODE_HW; 10445 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB 10446 ? XDP_MODE_DRV : XDP_MODE_SKB; 10447 10448 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) { 10449 NL_SET_ERR_MSG(extack, "XDP program already attached"); 10450 return -EBUSY; 10451 } 10452 if (!offload && dev_xdp_prog(dev, other_mode)) { 10453 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time"); 10454 return -EEXIST; 10455 } 10456 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) { 10457 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported"); 10458 return -EINVAL; 10459 } 10460 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) { 10461 NL_SET_ERR_MSG(extack, "Program bound to different device"); 10462 return -EINVAL; 10463 } 10464 if (bpf_prog_is_dev_bound(new_prog->aux) && mode == XDP_MODE_SKB) { 10465 NL_SET_ERR_MSG(extack, "Can't attach device-bound programs in generic mode"); 10466 return -EINVAL; 10467 } 10468 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) { 10469 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device"); 10470 return -EINVAL; 10471 } 10472 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) { 10473 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device"); 10474 return -EINVAL; 10475 } 10476 } 10477 10478 /* don't call drivers if the effective program didn't change */ 10479 if (new_prog != cur_prog) { 10480 bpf_op = dev_xdp_bpf_op(dev, mode); 10481 if (!bpf_op) { 10482 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode"); 10483 return -EOPNOTSUPP; 10484 } 10485 10486 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog); 10487 if (err) 10488 return err; 10489 } 10490 10491 if (link) 10492 dev_xdp_set_link(dev, mode, link); 10493 else 10494 dev_xdp_set_prog(dev, mode, new_prog); 10495 if (cur_prog) 10496 bpf_prog_put(cur_prog); 10497 10498 return 0; 10499 } 10500 10501 static int dev_xdp_attach_link(struct net_device *dev, 10502 struct netlink_ext_ack *extack, 10503 struct bpf_xdp_link *link) 10504 { 10505 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags); 10506 } 10507 10508 static int dev_xdp_detach_link(struct net_device *dev, 10509 struct netlink_ext_ack *extack, 10510 struct bpf_xdp_link *link) 10511 { 10512 enum bpf_xdp_mode mode; 10513 bpf_op_t bpf_op; 10514 10515 ASSERT_RTNL(); 10516 10517 mode = dev_xdp_mode(dev, link->flags); 10518 if (dev_xdp_link(dev, mode) != link) 10519 return -EINVAL; 10520 10521 bpf_op = dev_xdp_bpf_op(dev, mode); 10522 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 10523 dev_xdp_set_link(dev, mode, NULL); 10524 return 0; 10525 } 10526 10527 static void bpf_xdp_link_release(struct bpf_link *link) 10528 { 10529 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10530 10531 rtnl_lock(); 10532 10533 /* if racing with net_device's tear down, xdp_link->dev might be 10534 * already NULL, in which case link was already auto-detached 10535 */ 10536 if (xdp_link->dev) { 10537 netdev_lock_ops(xdp_link->dev); 10538 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link)); 10539 netdev_unlock_ops(xdp_link->dev); 10540 xdp_link->dev = NULL; 10541 } 10542 10543 rtnl_unlock(); 10544 } 10545 10546 static int bpf_xdp_link_detach(struct bpf_link *link) 10547 { 10548 bpf_xdp_link_release(link); 10549 return 0; 10550 } 10551 10552 static void bpf_xdp_link_dealloc(struct bpf_link *link) 10553 { 10554 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10555 10556 kfree(xdp_link); 10557 } 10558 10559 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link, 10560 struct seq_file *seq) 10561 { 10562 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10563 u32 ifindex = 0; 10564 10565 rtnl_lock(); 10566 if (xdp_link->dev) 10567 ifindex = xdp_link->dev->ifindex; 10568 rtnl_unlock(); 10569 10570 seq_printf(seq, "ifindex:\t%u\n", ifindex); 10571 } 10572 10573 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link, 10574 struct bpf_link_info *info) 10575 { 10576 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10577 u32 ifindex = 0; 10578 10579 rtnl_lock(); 10580 if (xdp_link->dev) 10581 ifindex = xdp_link->dev->ifindex; 10582 rtnl_unlock(); 10583 10584 info->xdp.ifindex = ifindex; 10585 return 0; 10586 } 10587 10588 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog, 10589 struct bpf_prog *old_prog) 10590 { 10591 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10592 enum bpf_xdp_mode mode; 10593 bpf_op_t bpf_op; 10594 int err = 0; 10595 10596 rtnl_lock(); 10597 10598 /* link might have been auto-released already, so fail */ 10599 if (!xdp_link->dev) { 10600 err = -ENOLINK; 10601 goto out_unlock; 10602 } 10603 10604 if (old_prog && link->prog != old_prog) { 10605 err = -EPERM; 10606 goto out_unlock; 10607 } 10608 old_prog = link->prog; 10609 if (old_prog->type != new_prog->type || 10610 old_prog->expected_attach_type != new_prog->expected_attach_type) { 10611 err = -EINVAL; 10612 goto out_unlock; 10613 } 10614 10615 if (old_prog == new_prog) { 10616 /* no-op, don't disturb drivers */ 10617 bpf_prog_put(new_prog); 10618 goto out_unlock; 10619 } 10620 10621 netdev_lock_ops(xdp_link->dev); 10622 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags); 10623 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode); 10624 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL, 10625 xdp_link->flags, new_prog); 10626 netdev_unlock_ops(xdp_link->dev); 10627 if (err) 10628 goto out_unlock; 10629 10630 old_prog = xchg(&link->prog, new_prog); 10631 bpf_prog_put(old_prog); 10632 10633 out_unlock: 10634 rtnl_unlock(); 10635 return err; 10636 } 10637 10638 static const struct bpf_link_ops bpf_xdp_link_lops = { 10639 .release = bpf_xdp_link_release, 10640 .dealloc = bpf_xdp_link_dealloc, 10641 .detach = bpf_xdp_link_detach, 10642 .show_fdinfo = bpf_xdp_link_show_fdinfo, 10643 .fill_link_info = bpf_xdp_link_fill_link_info, 10644 .update_prog = bpf_xdp_link_update, 10645 }; 10646 10647 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog) 10648 { 10649 struct net *net = current->nsproxy->net_ns; 10650 struct bpf_link_primer link_primer; 10651 struct netlink_ext_ack extack = {}; 10652 struct bpf_xdp_link *link; 10653 struct net_device *dev; 10654 int err, fd; 10655 10656 rtnl_lock(); 10657 dev = dev_get_by_index(net, attr->link_create.target_ifindex); 10658 if (!dev) { 10659 rtnl_unlock(); 10660 return -EINVAL; 10661 } 10662 10663 link = kzalloc(sizeof(*link), GFP_USER); 10664 if (!link) { 10665 err = -ENOMEM; 10666 goto unlock; 10667 } 10668 10669 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog, 10670 attr->link_create.attach_type); 10671 link->dev = dev; 10672 link->flags = attr->link_create.flags; 10673 10674 err = bpf_link_prime(&link->link, &link_primer); 10675 if (err) { 10676 kfree(link); 10677 goto unlock; 10678 } 10679 10680 netdev_lock_ops(dev); 10681 err = dev_xdp_attach_link(dev, &extack, link); 10682 netdev_unlock_ops(dev); 10683 rtnl_unlock(); 10684 10685 if (err) { 10686 link->dev = NULL; 10687 bpf_link_cleanup(&link_primer); 10688 trace_bpf_xdp_link_attach_failed(extack._msg); 10689 goto out_put_dev; 10690 } 10691 10692 fd = bpf_link_settle(&link_primer); 10693 /* link itself doesn't hold dev's refcnt to not complicate shutdown */ 10694 dev_put(dev); 10695 return fd; 10696 10697 unlock: 10698 rtnl_unlock(); 10699 10700 out_put_dev: 10701 dev_put(dev); 10702 return err; 10703 } 10704 10705 /** 10706 * dev_change_xdp_fd - set or clear a bpf program for a device rx path 10707 * @dev: device 10708 * @extack: netlink extended ack 10709 * @fd: new program fd or negative value to clear 10710 * @expected_fd: old program fd that userspace expects to replace or clear 10711 * @flags: xdp-related flags 10712 * 10713 * Set or clear a bpf program for a device 10714 */ 10715 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, 10716 int fd, int expected_fd, u32 flags) 10717 { 10718 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags); 10719 struct bpf_prog *new_prog = NULL, *old_prog = NULL; 10720 int err; 10721 10722 ASSERT_RTNL(); 10723 10724 if (fd >= 0) { 10725 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP, 10726 mode != XDP_MODE_SKB); 10727 if (IS_ERR(new_prog)) 10728 return PTR_ERR(new_prog); 10729 } 10730 10731 if (expected_fd >= 0) { 10732 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP, 10733 mode != XDP_MODE_SKB); 10734 if (IS_ERR(old_prog)) { 10735 err = PTR_ERR(old_prog); 10736 old_prog = NULL; 10737 goto err_out; 10738 } 10739 } 10740 10741 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags); 10742 10743 err_out: 10744 if (err && new_prog) 10745 bpf_prog_put(new_prog); 10746 if (old_prog) 10747 bpf_prog_put(old_prog); 10748 return err; 10749 } 10750 10751 u32 dev_get_min_mp_channel_count(const struct net_device *dev) 10752 { 10753 int i; 10754 10755 netdev_ops_assert_locked(dev); 10756 10757 for (i = dev->real_num_rx_queues - 1; i >= 0; i--) 10758 if (dev->_rx[i].mp_params.mp_priv) 10759 /* The channel count is the idx plus 1. */ 10760 return i + 1; 10761 10762 return 0; 10763 } 10764 10765 /** 10766 * dev_index_reserve() - allocate an ifindex in a namespace 10767 * @net: the applicable net namespace 10768 * @ifindex: requested ifindex, pass %0 to get one allocated 10769 * 10770 * Allocate a ifindex for a new device. Caller must either use the ifindex 10771 * to store the device (via list_netdevice()) or call dev_index_release() 10772 * to give the index up. 10773 * 10774 * Return: a suitable unique value for a new device interface number or -errno. 10775 */ 10776 static int dev_index_reserve(struct net *net, u32 ifindex) 10777 { 10778 int err; 10779 10780 if (ifindex > INT_MAX) { 10781 DEBUG_NET_WARN_ON_ONCE(1); 10782 return -EINVAL; 10783 } 10784 10785 if (!ifindex) 10786 err = xa_alloc_cyclic(&net->dev_by_index, &ifindex, NULL, 10787 xa_limit_31b, &net->ifindex, GFP_KERNEL); 10788 else 10789 err = xa_insert(&net->dev_by_index, ifindex, NULL, GFP_KERNEL); 10790 if (err < 0) 10791 return err; 10792 10793 return ifindex; 10794 } 10795 10796 static void dev_index_release(struct net *net, int ifindex) 10797 { 10798 /* Expect only unused indexes, unlist_netdevice() removes the used */ 10799 WARN_ON(xa_erase(&net->dev_by_index, ifindex)); 10800 } 10801 10802 static bool from_cleanup_net(void) 10803 { 10804 #ifdef CONFIG_NET_NS 10805 return current == READ_ONCE(cleanup_net_task); 10806 #else 10807 return false; 10808 #endif 10809 } 10810 10811 /* Delayed registration/unregisteration */ 10812 LIST_HEAD(net_todo_list); 10813 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq); 10814 atomic_t dev_unreg_count = ATOMIC_INIT(0); 10815 10816 static void net_set_todo(struct net_device *dev) 10817 { 10818 list_add_tail(&dev->todo_list, &net_todo_list); 10819 } 10820 10821 static netdev_features_t netdev_sync_upper_features(struct net_device *lower, 10822 struct net_device *upper, netdev_features_t features) 10823 { 10824 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10825 netdev_features_t feature; 10826 int feature_bit; 10827 10828 for_each_netdev_feature(upper_disables, feature_bit) { 10829 feature = __NETIF_F_BIT(feature_bit); 10830 if (!(upper->wanted_features & feature) 10831 && (features & feature)) { 10832 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n", 10833 &feature, upper->name); 10834 features &= ~feature; 10835 } 10836 } 10837 10838 return features; 10839 } 10840 10841 static void netdev_sync_lower_features(struct net_device *upper, 10842 struct net_device *lower, netdev_features_t features) 10843 { 10844 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10845 netdev_features_t feature; 10846 int feature_bit; 10847 10848 for_each_netdev_feature(upper_disables, feature_bit) { 10849 feature = __NETIF_F_BIT(feature_bit); 10850 if (!(features & feature) && (lower->features & feature)) { 10851 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n", 10852 &feature, lower->name); 10853 netdev_lock_ops(lower); 10854 lower->wanted_features &= ~feature; 10855 __netdev_update_features(lower); 10856 10857 if (unlikely(lower->features & feature)) 10858 netdev_WARN(upper, "failed to disable %pNF on %s!\n", 10859 &feature, lower->name); 10860 else 10861 netdev_features_change(lower); 10862 netdev_unlock_ops(lower); 10863 } 10864 } 10865 } 10866 10867 static bool netdev_has_ip_or_hw_csum(netdev_features_t features) 10868 { 10869 netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 10870 bool ip_csum = (features & ip_csum_mask) == ip_csum_mask; 10871 bool hw_csum = features & NETIF_F_HW_CSUM; 10872 10873 return ip_csum || hw_csum; 10874 } 10875 10876 static netdev_features_t netdev_fix_features(struct net_device *dev, 10877 netdev_features_t features) 10878 { 10879 /* Fix illegal checksum combinations */ 10880 if ((features & NETIF_F_HW_CSUM) && 10881 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) { 10882 netdev_warn(dev, "mixed HW and IP checksum settings.\n"); 10883 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM); 10884 } 10885 10886 /* TSO requires that SG is present as well. */ 10887 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) { 10888 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n"); 10889 features &= ~NETIF_F_ALL_TSO; 10890 } 10891 10892 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) && 10893 !(features & NETIF_F_IP_CSUM)) { 10894 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n"); 10895 features &= ~NETIF_F_TSO; 10896 features &= ~NETIF_F_TSO_ECN; 10897 } 10898 10899 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) && 10900 !(features & NETIF_F_IPV6_CSUM)) { 10901 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n"); 10902 features &= ~NETIF_F_TSO6; 10903 } 10904 10905 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */ 10906 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO)) 10907 features &= ~NETIF_F_TSO_MANGLEID; 10908 10909 /* TSO ECN requires that TSO is present as well. */ 10910 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN) 10911 features &= ~NETIF_F_TSO_ECN; 10912 10913 /* Software GSO depends on SG. */ 10914 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) { 10915 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n"); 10916 features &= ~NETIF_F_GSO; 10917 } 10918 10919 /* GSO partial features require GSO partial be set */ 10920 if ((features & dev->gso_partial_features) && 10921 !(features & NETIF_F_GSO_PARTIAL)) { 10922 netdev_dbg(dev, 10923 "Dropping partially supported GSO features since no GSO partial.\n"); 10924 features &= ~dev->gso_partial_features; 10925 } 10926 10927 if (!(features & NETIF_F_RXCSUM)) { 10928 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet 10929 * successfully merged by hardware must also have the 10930 * checksum verified by hardware. If the user does not 10931 * want to enable RXCSUM, logically, we should disable GRO_HW. 10932 */ 10933 if (features & NETIF_F_GRO_HW) { 10934 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n"); 10935 features &= ~NETIF_F_GRO_HW; 10936 } 10937 } 10938 10939 /* LRO/HW-GRO features cannot be combined with RX-FCS */ 10940 if (features & NETIF_F_RXFCS) { 10941 if (features & NETIF_F_LRO) { 10942 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n"); 10943 features &= ~NETIF_F_LRO; 10944 } 10945 10946 if (features & NETIF_F_GRO_HW) { 10947 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n"); 10948 features &= ~NETIF_F_GRO_HW; 10949 } 10950 } 10951 10952 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) { 10953 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n"); 10954 features &= ~NETIF_F_LRO; 10955 } 10956 10957 if ((features & NETIF_F_HW_TLS_TX) && !netdev_has_ip_or_hw_csum(features)) { 10958 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n"); 10959 features &= ~NETIF_F_HW_TLS_TX; 10960 } 10961 10962 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) { 10963 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n"); 10964 features &= ~NETIF_F_HW_TLS_RX; 10965 } 10966 10967 if ((features & NETIF_F_GSO_UDP_L4) && !netdev_has_ip_or_hw_csum(features)) { 10968 netdev_dbg(dev, "Dropping USO feature since no CSUM feature.\n"); 10969 features &= ~NETIF_F_GSO_UDP_L4; 10970 } 10971 10972 return features; 10973 } 10974 10975 int __netdev_update_features(struct net_device *dev) 10976 { 10977 struct net_device *upper, *lower; 10978 netdev_features_t features; 10979 struct list_head *iter; 10980 int err = -1; 10981 10982 ASSERT_RTNL(); 10983 netdev_ops_assert_locked(dev); 10984 10985 features = netdev_get_wanted_features(dev); 10986 10987 if (dev->netdev_ops->ndo_fix_features) 10988 features = dev->netdev_ops->ndo_fix_features(dev, features); 10989 10990 /* driver might be less strict about feature dependencies */ 10991 features = netdev_fix_features(dev, features); 10992 10993 /* some features can't be enabled if they're off on an upper device */ 10994 netdev_for_each_upper_dev_rcu(dev, upper, iter) 10995 features = netdev_sync_upper_features(dev, upper, features); 10996 10997 if (dev->features == features) 10998 goto sync_lower; 10999 11000 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n", 11001 &dev->features, &features); 11002 11003 if (dev->netdev_ops->ndo_set_features) 11004 err = dev->netdev_ops->ndo_set_features(dev, features); 11005 else 11006 err = 0; 11007 11008 if (unlikely(err < 0)) { 11009 netdev_err(dev, 11010 "set_features() failed (%d); wanted %pNF, left %pNF\n", 11011 err, &features, &dev->features); 11012 /* return non-0 since some features might have changed and 11013 * it's better to fire a spurious notification than miss it 11014 */ 11015 return -1; 11016 } 11017 11018 sync_lower: 11019 /* some features must be disabled on lower devices when disabled 11020 * on an upper device (think: bonding master or bridge) 11021 */ 11022 netdev_for_each_lower_dev(dev, lower, iter) 11023 netdev_sync_lower_features(dev, lower, features); 11024 11025 if (!err) { 11026 netdev_features_t diff = features ^ dev->features; 11027 11028 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) { 11029 /* udp_tunnel_{get,drop}_rx_info both need 11030 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the 11031 * device, or they won't do anything. 11032 * Thus we need to update dev->features 11033 * *before* calling udp_tunnel_get_rx_info, 11034 * but *after* calling udp_tunnel_drop_rx_info. 11035 */ 11036 udp_tunnel_nic_lock(dev); 11037 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) { 11038 dev->features = features; 11039 udp_tunnel_get_rx_info(dev); 11040 } else { 11041 udp_tunnel_drop_rx_info(dev); 11042 } 11043 udp_tunnel_nic_unlock(dev); 11044 } 11045 11046 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) { 11047 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) { 11048 dev->features = features; 11049 err |= vlan_get_rx_ctag_filter_info(dev); 11050 } else { 11051 vlan_drop_rx_ctag_filter_info(dev); 11052 } 11053 } 11054 11055 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) { 11056 if (features & NETIF_F_HW_VLAN_STAG_FILTER) { 11057 dev->features = features; 11058 err |= vlan_get_rx_stag_filter_info(dev); 11059 } else { 11060 vlan_drop_rx_stag_filter_info(dev); 11061 } 11062 } 11063 11064 dev->features = features; 11065 } 11066 11067 return err < 0 ? 0 : 1; 11068 } 11069 11070 /** 11071 * netdev_update_features - recalculate device features 11072 * @dev: the device to check 11073 * 11074 * Recalculate dev->features set and send notifications if it 11075 * has changed. Should be called after driver or hardware dependent 11076 * conditions might have changed that influence the features. 11077 */ 11078 void netdev_update_features(struct net_device *dev) 11079 { 11080 if (__netdev_update_features(dev)) 11081 netdev_features_change(dev); 11082 } 11083 EXPORT_SYMBOL(netdev_update_features); 11084 11085 /** 11086 * netdev_change_features - recalculate device features 11087 * @dev: the device to check 11088 * 11089 * Recalculate dev->features set and send notifications even 11090 * if they have not changed. Should be called instead of 11091 * netdev_update_features() if also dev->vlan_features might 11092 * have changed to allow the changes to be propagated to stacked 11093 * VLAN devices. 11094 */ 11095 void netdev_change_features(struct net_device *dev) 11096 { 11097 __netdev_update_features(dev); 11098 netdev_features_change(dev); 11099 } 11100 EXPORT_SYMBOL(netdev_change_features); 11101 11102 /** 11103 * netif_stacked_transfer_operstate - transfer operstate 11104 * @rootdev: the root or lower level device to transfer state from 11105 * @dev: the device to transfer operstate to 11106 * 11107 * Transfer operational state from root to device. This is normally 11108 * called when a stacking relationship exists between the root 11109 * device and the device(a leaf device). 11110 */ 11111 void netif_stacked_transfer_operstate(const struct net_device *rootdev, 11112 struct net_device *dev) 11113 { 11114 if (rootdev->operstate == IF_OPER_DORMANT) 11115 netif_dormant_on(dev); 11116 else 11117 netif_dormant_off(dev); 11118 11119 if (rootdev->operstate == IF_OPER_TESTING) 11120 netif_testing_on(dev); 11121 else 11122 netif_testing_off(dev); 11123 11124 if (netif_carrier_ok(rootdev)) 11125 netif_carrier_on(dev); 11126 else 11127 netif_carrier_off(dev); 11128 } 11129 EXPORT_SYMBOL(netif_stacked_transfer_operstate); 11130 11131 static int netif_alloc_rx_queues(struct net_device *dev) 11132 { 11133 unsigned int i, count = dev->num_rx_queues; 11134 struct netdev_rx_queue *rx; 11135 size_t sz = count * sizeof(*rx); 11136 int err = 0; 11137 11138 BUG_ON(count < 1); 11139 11140 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11141 if (!rx) 11142 return -ENOMEM; 11143 11144 dev->_rx = rx; 11145 11146 for (i = 0; i < count; i++) { 11147 rx[i].dev = dev; 11148 11149 /* XDP RX-queue setup */ 11150 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0); 11151 if (err < 0) 11152 goto err_rxq_info; 11153 } 11154 return 0; 11155 11156 err_rxq_info: 11157 /* Rollback successful reg's and free other resources */ 11158 while (i--) 11159 xdp_rxq_info_unreg(&rx[i].xdp_rxq); 11160 kvfree(dev->_rx); 11161 dev->_rx = NULL; 11162 return err; 11163 } 11164 11165 static void netif_free_rx_queues(struct net_device *dev) 11166 { 11167 unsigned int i, count = dev->num_rx_queues; 11168 11169 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */ 11170 if (!dev->_rx) 11171 return; 11172 11173 for (i = 0; i < count; i++) 11174 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq); 11175 11176 kvfree(dev->_rx); 11177 } 11178 11179 static void netdev_init_one_queue(struct net_device *dev, 11180 struct netdev_queue *queue, void *_unused) 11181 { 11182 /* Initialize queue lock */ 11183 spin_lock_init(&queue->_xmit_lock); 11184 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type); 11185 queue->xmit_lock_owner = -1; 11186 netdev_queue_numa_node_write(queue, NUMA_NO_NODE); 11187 queue->dev = dev; 11188 #ifdef CONFIG_BQL 11189 dql_init(&queue->dql, HZ); 11190 #endif 11191 } 11192 11193 static void netif_free_tx_queues(struct net_device *dev) 11194 { 11195 kvfree(dev->_tx); 11196 } 11197 11198 static int netif_alloc_netdev_queues(struct net_device *dev) 11199 { 11200 unsigned int count = dev->num_tx_queues; 11201 struct netdev_queue *tx; 11202 size_t sz = count * sizeof(*tx); 11203 11204 if (count < 1 || count > 0xffff) 11205 return -EINVAL; 11206 11207 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11208 if (!tx) 11209 return -ENOMEM; 11210 11211 dev->_tx = tx; 11212 11213 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL); 11214 spin_lock_init(&dev->tx_global_lock); 11215 11216 return 0; 11217 } 11218 11219 void netif_tx_stop_all_queues(struct net_device *dev) 11220 { 11221 unsigned int i; 11222 11223 for (i = 0; i < dev->num_tx_queues; i++) { 11224 struct netdev_queue *txq = netdev_get_tx_queue(dev, i); 11225 11226 netif_tx_stop_queue(txq); 11227 } 11228 } 11229 EXPORT_SYMBOL(netif_tx_stop_all_queues); 11230 11231 static int netdev_do_alloc_pcpu_stats(struct net_device *dev) 11232 { 11233 void __percpu *v; 11234 11235 /* Drivers implementing ndo_get_peer_dev must support tstat 11236 * accounting, so that skb_do_redirect() can bump the dev's 11237 * RX stats upon network namespace switch. 11238 */ 11239 if (dev->netdev_ops->ndo_get_peer_dev && 11240 dev->pcpu_stat_type != NETDEV_PCPU_STAT_TSTATS) 11241 return -EOPNOTSUPP; 11242 11243 switch (dev->pcpu_stat_type) { 11244 case NETDEV_PCPU_STAT_NONE: 11245 return 0; 11246 case NETDEV_PCPU_STAT_LSTATS: 11247 v = dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats); 11248 break; 11249 case NETDEV_PCPU_STAT_TSTATS: 11250 v = dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats); 11251 break; 11252 case NETDEV_PCPU_STAT_DSTATS: 11253 v = dev->dstats = netdev_alloc_pcpu_stats(struct pcpu_dstats); 11254 break; 11255 default: 11256 return -EINVAL; 11257 } 11258 11259 return v ? 0 : -ENOMEM; 11260 } 11261 11262 static void netdev_do_free_pcpu_stats(struct net_device *dev) 11263 { 11264 switch (dev->pcpu_stat_type) { 11265 case NETDEV_PCPU_STAT_NONE: 11266 return; 11267 case NETDEV_PCPU_STAT_LSTATS: 11268 free_percpu(dev->lstats); 11269 break; 11270 case NETDEV_PCPU_STAT_TSTATS: 11271 free_percpu(dev->tstats); 11272 break; 11273 case NETDEV_PCPU_STAT_DSTATS: 11274 free_percpu(dev->dstats); 11275 break; 11276 } 11277 } 11278 11279 static void netdev_free_phy_link_topology(struct net_device *dev) 11280 { 11281 struct phy_link_topology *topo = dev->link_topo; 11282 11283 if (IS_ENABLED(CONFIG_PHYLIB) && topo) { 11284 xa_destroy(&topo->phys); 11285 kfree(topo); 11286 dev->link_topo = NULL; 11287 } 11288 } 11289 11290 /** 11291 * register_netdevice() - register a network device 11292 * @dev: device to register 11293 * 11294 * Take a prepared network device structure and make it externally accessible. 11295 * A %NETDEV_REGISTER message is sent to the netdev notifier chain. 11296 * Callers must hold the rtnl lock - you may want register_netdev() 11297 * instead of this. 11298 */ 11299 int register_netdevice(struct net_device *dev) 11300 { 11301 int ret; 11302 struct net *net = dev_net(dev); 11303 11304 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE < 11305 NETDEV_FEATURE_COUNT); 11306 BUG_ON(dev_boot_phase); 11307 ASSERT_RTNL(); 11308 11309 might_sleep(); 11310 11311 /* When net_device's are persistent, this will be fatal. */ 11312 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED); 11313 BUG_ON(!net); 11314 11315 ret = ethtool_check_ops(dev->ethtool_ops); 11316 if (ret) 11317 return ret; 11318 11319 /* rss ctx ID 0 is reserved for the default context, start from 1 */ 11320 xa_init_flags(&dev->ethtool->rss_ctx, XA_FLAGS_ALLOC1); 11321 mutex_init(&dev->ethtool->rss_lock); 11322 11323 spin_lock_init(&dev->addr_list_lock); 11324 netdev_set_addr_lockdep_class(dev); 11325 11326 ret = dev_get_valid_name(net, dev, dev->name); 11327 if (ret < 0) 11328 goto out; 11329 11330 ret = -ENOMEM; 11331 dev->name_node = netdev_name_node_head_alloc(dev); 11332 if (!dev->name_node) 11333 goto out; 11334 11335 /* Init, if this function is available */ 11336 if (dev->netdev_ops->ndo_init) { 11337 ret = dev->netdev_ops->ndo_init(dev); 11338 if (ret) { 11339 if (ret > 0) 11340 ret = -EIO; 11341 goto err_free_name; 11342 } 11343 } 11344 11345 if (((dev->hw_features | dev->features) & 11346 NETIF_F_HW_VLAN_CTAG_FILTER) && 11347 (!dev->netdev_ops->ndo_vlan_rx_add_vid || 11348 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) { 11349 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n"); 11350 ret = -EINVAL; 11351 goto err_uninit; 11352 } 11353 11354 ret = netdev_do_alloc_pcpu_stats(dev); 11355 if (ret) 11356 goto err_uninit; 11357 11358 ret = dev_index_reserve(net, dev->ifindex); 11359 if (ret < 0) 11360 goto err_free_pcpu; 11361 dev->ifindex = ret; 11362 11363 /* Transfer changeable features to wanted_features and enable 11364 * software offloads (GSO and GRO). 11365 */ 11366 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF); 11367 dev->features |= NETIF_F_SOFT_FEATURES; 11368 11369 if (dev->udp_tunnel_nic_info) { 11370 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT; 11371 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT; 11372 } 11373 11374 dev->wanted_features = dev->features & dev->hw_features; 11375 11376 if (!(dev->flags & IFF_LOOPBACK)) 11377 dev->hw_features |= NETIF_F_NOCACHE_COPY; 11378 11379 /* If IPv4 TCP segmentation offload is supported we should also 11380 * allow the device to enable segmenting the frame with the option 11381 * of ignoring a static IP ID value. This doesn't enable the 11382 * feature itself but allows the user to enable it later. 11383 */ 11384 if (dev->hw_features & NETIF_F_TSO) 11385 dev->hw_features |= NETIF_F_TSO_MANGLEID; 11386 if (dev->vlan_features & NETIF_F_TSO) 11387 dev->vlan_features |= NETIF_F_TSO_MANGLEID; 11388 if (dev->mpls_features & NETIF_F_TSO) 11389 dev->mpls_features |= NETIF_F_TSO_MANGLEID; 11390 if (dev->hw_enc_features & NETIF_F_TSO) 11391 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID; 11392 11393 /* TSO_MANGLEID belongs in mangleid_features by definition */ 11394 dev->mangleid_features |= NETIF_F_TSO_MANGLEID; 11395 11396 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices. 11397 */ 11398 dev->vlan_features |= NETIF_F_HIGHDMA; 11399 11400 /* Make NETIF_F_SG inheritable to tunnel devices. 11401 */ 11402 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL; 11403 11404 /* Make NETIF_F_SG inheritable to MPLS. 11405 */ 11406 dev->mpls_features |= NETIF_F_SG; 11407 11408 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev); 11409 ret = notifier_to_errno(ret); 11410 if (ret) 11411 goto err_ifindex_release; 11412 11413 ret = netdev_register_kobject(dev); 11414 11415 netdev_lock(dev); 11416 WRITE_ONCE(dev->reg_state, ret ? NETREG_UNREGISTERED : NETREG_REGISTERED); 11417 netdev_unlock(dev); 11418 11419 if (ret) 11420 goto err_uninit_notify; 11421 11422 netdev_lock_ops(dev); 11423 __netdev_update_features(dev); 11424 netdev_unlock_ops(dev); 11425 11426 /* 11427 * Default initial state at registry is that the 11428 * device is present. 11429 */ 11430 11431 set_bit(__LINK_STATE_PRESENT, &dev->state); 11432 11433 linkwatch_init_dev(dev); 11434 11435 dev_init_scheduler(dev); 11436 11437 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL); 11438 list_netdevice(dev); 11439 11440 add_device_randomness(dev->dev_addr, dev->addr_len); 11441 11442 /* If the device has permanent device address, driver should 11443 * set dev_addr and also addr_assign_type should be set to 11444 * NET_ADDR_PERM (default value). 11445 */ 11446 if (dev->addr_assign_type == NET_ADDR_PERM) 11447 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len); 11448 11449 /* Notify protocols, that a new device appeared. */ 11450 netdev_lock_ops(dev); 11451 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev); 11452 netdev_unlock_ops(dev); 11453 ret = notifier_to_errno(ret); 11454 if (ret) { 11455 /* Expect explicit free_netdev() on failure */ 11456 dev->needs_free_netdev = false; 11457 unregister_netdevice_queue(dev, NULL); 11458 goto out; 11459 } 11460 /* 11461 * Prevent userspace races by waiting until the network 11462 * device is fully setup before sending notifications. 11463 */ 11464 if (!(dev->rtnl_link_ops && dev->rtnl_link_initializing)) 11465 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 11466 11467 out: 11468 return ret; 11469 11470 err_uninit_notify: 11471 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11472 err_ifindex_release: 11473 dev_index_release(net, dev->ifindex); 11474 err_free_pcpu: 11475 netdev_do_free_pcpu_stats(dev); 11476 err_uninit: 11477 if (dev->netdev_ops->ndo_uninit) 11478 dev->netdev_ops->ndo_uninit(dev); 11479 if (dev->priv_destructor) 11480 dev->priv_destructor(dev); 11481 err_free_name: 11482 netdev_name_node_free(dev->name_node); 11483 goto out; 11484 } 11485 EXPORT_SYMBOL(register_netdevice); 11486 11487 /* Initialize the core of a dummy net device. 11488 * The setup steps dummy netdevs need which normal netdevs get by going 11489 * through register_netdevice(). 11490 */ 11491 static void init_dummy_netdev(struct net_device *dev) 11492 { 11493 /* make sure we BUG if trying to hit standard 11494 * register/unregister code path 11495 */ 11496 dev->reg_state = NETREG_DUMMY; 11497 11498 /* a dummy interface is started by default */ 11499 set_bit(__LINK_STATE_PRESENT, &dev->state); 11500 set_bit(__LINK_STATE_START, &dev->state); 11501 11502 /* Note : We dont allocate pcpu_refcnt for dummy devices, 11503 * because users of this 'device' dont need to change 11504 * its refcount. 11505 */ 11506 } 11507 11508 /** 11509 * register_netdev - register a network device 11510 * @dev: device to register 11511 * 11512 * Take a completed network device structure and add it to the kernel 11513 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier 11514 * chain. 0 is returned on success. A negative errno code is returned 11515 * on a failure to set up the device, or if the name is a duplicate. 11516 * 11517 * This is a wrapper around register_netdevice that takes the rtnl semaphore 11518 * and expands the device name if you passed a format string to 11519 * alloc_netdev. 11520 */ 11521 int register_netdev(struct net_device *dev) 11522 { 11523 struct net *net = dev_net(dev); 11524 int err; 11525 11526 if (rtnl_net_lock_killable(net)) 11527 return -EINTR; 11528 11529 err = register_netdevice(dev); 11530 11531 rtnl_net_unlock(net); 11532 11533 return err; 11534 } 11535 EXPORT_SYMBOL(register_netdev); 11536 11537 int netdev_refcnt_read(const struct net_device *dev) 11538 { 11539 #ifdef CONFIG_PCPU_DEV_REFCNT 11540 int i, refcnt = 0; 11541 11542 for_each_possible_cpu(i) 11543 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i); 11544 return refcnt; 11545 #else 11546 return refcount_read(&dev->dev_refcnt); 11547 #endif 11548 } 11549 EXPORT_SYMBOL(netdev_refcnt_read); 11550 11551 int netdev_unregister_timeout_secs __read_mostly = 10; 11552 11553 #define WAIT_REFS_MIN_MSECS 1 11554 #define WAIT_REFS_MAX_MSECS 250 11555 /** 11556 * netdev_wait_allrefs_any - wait until all references are gone. 11557 * @list: list of net_devices to wait on 11558 * 11559 * This is called when unregistering network devices. 11560 * 11561 * Any protocol or device that holds a reference should register 11562 * for netdevice notification, and cleanup and put back the 11563 * reference if they receive an UNREGISTER event. 11564 * We can get stuck here if buggy protocols don't correctly 11565 * call dev_put. 11566 */ 11567 static struct net_device *netdev_wait_allrefs_any(struct list_head *list) 11568 { 11569 unsigned long rebroadcast_time, warning_time; 11570 struct net_device *dev; 11571 int wait = 0; 11572 11573 rebroadcast_time = warning_time = jiffies; 11574 11575 list_for_each_entry(dev, list, todo_list) 11576 if (netdev_refcnt_read(dev) == 1) 11577 return dev; 11578 11579 while (true) { 11580 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) { 11581 rtnl_lock(); 11582 11583 /* Rebroadcast unregister notification */ 11584 list_for_each_entry(dev, list, todo_list) 11585 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11586 11587 __rtnl_unlock(); 11588 rcu_barrier(); 11589 rtnl_lock(); 11590 11591 list_for_each_entry(dev, list, todo_list) 11592 if (test_bit(__LINK_STATE_LINKWATCH_PENDING, 11593 &dev->state)) { 11594 /* We must not have linkwatch events 11595 * pending on unregister. If this 11596 * happens, we simply run the queue 11597 * unscheduled, resulting in a noop 11598 * for this device. 11599 */ 11600 linkwatch_run_queue(); 11601 break; 11602 } 11603 11604 __rtnl_unlock(); 11605 11606 rebroadcast_time = jiffies; 11607 } 11608 11609 rcu_barrier(); 11610 11611 if (!wait) { 11612 wait = WAIT_REFS_MIN_MSECS; 11613 } else { 11614 msleep(wait); 11615 wait = min(wait << 1, WAIT_REFS_MAX_MSECS); 11616 } 11617 11618 list_for_each_entry(dev, list, todo_list) 11619 if (netdev_refcnt_read(dev) == 1) 11620 return dev; 11621 11622 if (time_after(jiffies, warning_time + 11623 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) { 11624 list_for_each_entry(dev, list, todo_list) { 11625 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n", 11626 dev->name, netdev_refcnt_read(dev)); 11627 ref_tracker_dir_print(&dev->refcnt_tracker, 10); 11628 } 11629 11630 warning_time = jiffies; 11631 } 11632 } 11633 } 11634 11635 /* The sequence is: 11636 * 11637 * rtnl_lock(); 11638 * ... 11639 * register_netdevice(x1); 11640 * register_netdevice(x2); 11641 * ... 11642 * unregister_netdevice(y1); 11643 * unregister_netdevice(y2); 11644 * ... 11645 * rtnl_unlock(); 11646 * free_netdev(y1); 11647 * free_netdev(y2); 11648 * 11649 * We are invoked by rtnl_unlock(). 11650 * This allows us to deal with problems: 11651 * 1) We can delete sysfs objects which invoke hotplug 11652 * without deadlocking with linkwatch via keventd. 11653 * 2) Since we run with the RTNL semaphore not held, we can sleep 11654 * safely in order to wait for the netdev refcnt to drop to zero. 11655 * 11656 * We must not return until all unregister events added during 11657 * the interval the lock was held have been completed. 11658 */ 11659 void netdev_run_todo(void) 11660 { 11661 struct net_device *dev, *tmp; 11662 struct list_head list; 11663 int cnt; 11664 #ifdef CONFIG_LOCKDEP 11665 struct list_head unlink_list; 11666 11667 list_replace_init(&net_unlink_list, &unlink_list); 11668 11669 while (!list_empty(&unlink_list)) { 11670 dev = list_first_entry(&unlink_list, struct net_device, 11671 unlink_list); 11672 list_del_init(&dev->unlink_list); 11673 dev->nested_level = dev->lower_level - 1; 11674 } 11675 #endif 11676 11677 /* Snapshot list, allow later requests */ 11678 list_replace_init(&net_todo_list, &list); 11679 11680 __rtnl_unlock(); 11681 11682 /* Wait for rcu callbacks to finish before next phase */ 11683 if (!list_empty(&list)) 11684 rcu_barrier(); 11685 11686 list_for_each_entry_safe(dev, tmp, &list, todo_list) { 11687 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) { 11688 netdev_WARN(dev, "run_todo but not unregistering\n"); 11689 list_del(&dev->todo_list); 11690 continue; 11691 } 11692 11693 netdev_lock(dev); 11694 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERED); 11695 netdev_unlock(dev); 11696 linkwatch_sync_dev(dev); 11697 } 11698 11699 cnt = 0; 11700 while (!list_empty(&list)) { 11701 dev = netdev_wait_allrefs_any(&list); 11702 list_del(&dev->todo_list); 11703 11704 /* paranoia */ 11705 BUG_ON(netdev_refcnt_read(dev) != 1); 11706 BUG_ON(!list_empty(&dev->ptype_all)); 11707 BUG_ON(!list_empty(&dev->ptype_specific)); 11708 WARN_ON(rcu_access_pointer(dev->ip_ptr)); 11709 WARN_ON(rcu_access_pointer(dev->ip6_ptr)); 11710 11711 netdev_do_free_pcpu_stats(dev); 11712 if (dev->priv_destructor) 11713 dev->priv_destructor(dev); 11714 if (dev->needs_free_netdev) 11715 free_netdev(dev); 11716 11717 cnt++; 11718 11719 /* Free network device */ 11720 kobject_put(&dev->dev.kobj); 11721 } 11722 if (cnt && atomic_sub_and_test(cnt, &dev_unreg_count)) 11723 wake_up(&netdev_unregistering_wq); 11724 } 11725 11726 /* Collate per-cpu network dstats statistics 11727 * 11728 * Read per-cpu network statistics from dev->dstats and populate the related 11729 * fields in @s. 11730 */ 11731 static void dev_fetch_dstats(struct rtnl_link_stats64 *s, 11732 const struct pcpu_dstats __percpu *dstats) 11733 { 11734 int cpu; 11735 11736 for_each_possible_cpu(cpu) { 11737 u64 rx_packets, rx_bytes, rx_drops; 11738 u64 tx_packets, tx_bytes, tx_drops; 11739 const struct pcpu_dstats *stats; 11740 unsigned int start; 11741 11742 stats = per_cpu_ptr(dstats, cpu); 11743 do { 11744 start = u64_stats_fetch_begin(&stats->syncp); 11745 rx_packets = u64_stats_read(&stats->rx_packets); 11746 rx_bytes = u64_stats_read(&stats->rx_bytes); 11747 rx_drops = u64_stats_read(&stats->rx_drops); 11748 tx_packets = u64_stats_read(&stats->tx_packets); 11749 tx_bytes = u64_stats_read(&stats->tx_bytes); 11750 tx_drops = u64_stats_read(&stats->tx_drops); 11751 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11752 11753 s->rx_packets += rx_packets; 11754 s->rx_bytes += rx_bytes; 11755 s->rx_dropped += rx_drops; 11756 s->tx_packets += tx_packets; 11757 s->tx_bytes += tx_bytes; 11758 s->tx_dropped += tx_drops; 11759 } 11760 } 11761 11762 /* ndo_get_stats64 implementation for dtstats-based accounting. 11763 * 11764 * Populate @s from dev->stats and dev->dstats. This is used internally by the 11765 * core for NETDEV_PCPU_STAT_DSTAT-type stats collection. 11766 */ 11767 static void dev_get_dstats64(const struct net_device *dev, 11768 struct rtnl_link_stats64 *s) 11769 { 11770 netdev_stats_to_stats64(s, &dev->stats); 11771 dev_fetch_dstats(s, dev->dstats); 11772 } 11773 11774 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has 11775 * all the same fields in the same order as net_device_stats, with only 11776 * the type differing, but rtnl_link_stats64 may have additional fields 11777 * at the end for newer counters. 11778 */ 11779 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64, 11780 const struct net_device_stats *netdev_stats) 11781 { 11782 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t); 11783 const atomic_long_t *src = (atomic_long_t *)netdev_stats; 11784 u64 *dst = (u64 *)stats64; 11785 11786 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64)); 11787 for (i = 0; i < n; i++) 11788 dst[i] = (unsigned long)atomic_long_read(&src[i]); 11789 /* zero out counters that only exist in rtnl_link_stats64 */ 11790 memset((char *)stats64 + n * sizeof(u64), 0, 11791 sizeof(*stats64) - n * sizeof(u64)); 11792 } 11793 EXPORT_SYMBOL(netdev_stats_to_stats64); 11794 11795 static __cold struct net_device_core_stats __percpu *netdev_core_stats_alloc( 11796 struct net_device *dev) 11797 { 11798 struct net_device_core_stats __percpu *p; 11799 11800 p = alloc_percpu_gfp(struct net_device_core_stats, 11801 GFP_ATOMIC | __GFP_NOWARN); 11802 11803 if (p && cmpxchg(&dev->core_stats, NULL, p)) 11804 free_percpu(p); 11805 11806 /* This READ_ONCE() pairs with the cmpxchg() above */ 11807 return READ_ONCE(dev->core_stats); 11808 } 11809 11810 noinline void netdev_core_stats_inc(struct net_device *dev, u32 offset) 11811 { 11812 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11813 struct net_device_core_stats __percpu *p = READ_ONCE(dev->core_stats); 11814 unsigned long __percpu *field; 11815 11816 if (unlikely(!p)) { 11817 p = netdev_core_stats_alloc(dev); 11818 if (!p) 11819 return; 11820 } 11821 11822 field = (unsigned long __percpu *)((void __percpu *)p + offset); 11823 this_cpu_inc(*field); 11824 } 11825 EXPORT_SYMBOL_GPL(netdev_core_stats_inc); 11826 11827 /** 11828 * dev_get_stats - get network device statistics 11829 * @dev: device to get statistics from 11830 * @storage: place to store stats 11831 * 11832 * Get network statistics from device. Return @storage. 11833 * The device driver may provide its own method by setting 11834 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats; 11835 * otherwise the internal statistics structure is used. 11836 */ 11837 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev, 11838 struct rtnl_link_stats64 *storage) 11839 { 11840 const struct net_device_ops *ops = dev->netdev_ops; 11841 const struct net_device_core_stats __percpu *p; 11842 11843 /* 11844 * IPv{4,6} and udp tunnels share common stat helpers and use 11845 * different stat type (NETDEV_PCPU_STAT_TSTATS vs 11846 * NETDEV_PCPU_STAT_DSTATS). Ensure the accounting is consistent. 11847 */ 11848 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_bytes) != 11849 offsetof(struct pcpu_dstats, rx_bytes)); 11850 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_packets) != 11851 offsetof(struct pcpu_dstats, rx_packets)); 11852 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_bytes) != 11853 offsetof(struct pcpu_dstats, tx_bytes)); 11854 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_packets) != 11855 offsetof(struct pcpu_dstats, tx_packets)); 11856 11857 if (ops->ndo_get_stats64) { 11858 memset(storage, 0, sizeof(*storage)); 11859 ops->ndo_get_stats64(dev, storage); 11860 } else if (ops->ndo_get_stats) { 11861 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev)); 11862 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_TSTATS) { 11863 dev_get_tstats64(dev, storage); 11864 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_DSTATS) { 11865 dev_get_dstats64(dev, storage); 11866 } else { 11867 netdev_stats_to_stats64(storage, &dev->stats); 11868 } 11869 11870 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11871 p = READ_ONCE(dev->core_stats); 11872 if (p) { 11873 const struct net_device_core_stats *core_stats; 11874 int i; 11875 11876 for_each_possible_cpu(i) { 11877 core_stats = per_cpu_ptr(p, i); 11878 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped); 11879 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped); 11880 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler); 11881 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped); 11882 } 11883 } 11884 return storage; 11885 } 11886 EXPORT_SYMBOL(dev_get_stats); 11887 11888 /** 11889 * dev_fetch_sw_netstats - get per-cpu network device statistics 11890 * @s: place to store stats 11891 * @netstats: per-cpu network stats to read from 11892 * 11893 * Read per-cpu network statistics and populate the related fields in @s. 11894 */ 11895 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s, 11896 const struct pcpu_sw_netstats __percpu *netstats) 11897 { 11898 int cpu; 11899 11900 for_each_possible_cpu(cpu) { 11901 u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 11902 const struct pcpu_sw_netstats *stats; 11903 unsigned int start; 11904 11905 stats = per_cpu_ptr(netstats, cpu); 11906 do { 11907 start = u64_stats_fetch_begin(&stats->syncp); 11908 rx_packets = u64_stats_read(&stats->rx_packets); 11909 rx_bytes = u64_stats_read(&stats->rx_bytes); 11910 tx_packets = u64_stats_read(&stats->tx_packets); 11911 tx_bytes = u64_stats_read(&stats->tx_bytes); 11912 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11913 11914 s->rx_packets += rx_packets; 11915 s->rx_bytes += rx_bytes; 11916 s->tx_packets += tx_packets; 11917 s->tx_bytes += tx_bytes; 11918 } 11919 } 11920 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats); 11921 11922 /** 11923 * dev_get_tstats64 - ndo_get_stats64 implementation 11924 * @dev: device to get statistics from 11925 * @s: place to store stats 11926 * 11927 * Populate @s from dev->stats and dev->tstats. Can be used as 11928 * ndo_get_stats64() callback. 11929 */ 11930 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s) 11931 { 11932 netdev_stats_to_stats64(s, &dev->stats); 11933 dev_fetch_sw_netstats(s, dev->tstats); 11934 } 11935 EXPORT_SYMBOL_GPL(dev_get_tstats64); 11936 11937 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev) 11938 { 11939 struct netdev_queue *queue = dev_ingress_queue(dev); 11940 11941 #ifdef CONFIG_NET_CLS_ACT 11942 if (queue) 11943 return queue; 11944 queue = kzalloc(sizeof(*queue), GFP_KERNEL); 11945 if (!queue) 11946 return NULL; 11947 netdev_init_one_queue(dev, queue, NULL); 11948 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc); 11949 RCU_INIT_POINTER(queue->qdisc_sleeping, &noop_qdisc); 11950 rcu_assign_pointer(dev->ingress_queue, queue); 11951 #endif 11952 return queue; 11953 } 11954 11955 static const struct ethtool_ops default_ethtool_ops; 11956 11957 void netdev_set_default_ethtool_ops(struct net_device *dev, 11958 const struct ethtool_ops *ops) 11959 { 11960 if (dev->ethtool_ops == &default_ethtool_ops) 11961 dev->ethtool_ops = ops; 11962 } 11963 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops); 11964 11965 /** 11966 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default 11967 * @dev: netdev to enable the IRQ coalescing on 11968 * 11969 * Sets a conservative default for SW IRQ coalescing. Users can use 11970 * sysfs attributes to override the default values. 11971 */ 11972 void netdev_sw_irq_coalesce_default_on(struct net_device *dev) 11973 { 11974 WARN_ON(dev->reg_state == NETREG_REGISTERED); 11975 11976 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) { 11977 netdev_set_gro_flush_timeout(dev, 20000); 11978 netdev_set_defer_hard_irqs(dev, 1); 11979 } 11980 } 11981 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on); 11982 11983 /** 11984 * alloc_netdev_mqs - allocate network device 11985 * @sizeof_priv: size of private data to allocate space for 11986 * @name: device name format string 11987 * @name_assign_type: origin of device name 11988 * @setup: callback to initialize device 11989 * @txqs: the number of TX subqueues to allocate 11990 * @rxqs: the number of RX subqueues to allocate 11991 * 11992 * Allocates a struct net_device with private data area for driver use 11993 * and performs basic initialization. Also allocates subqueue structs 11994 * for each queue on the device. 11995 */ 11996 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name, 11997 unsigned char name_assign_type, 11998 void (*setup)(struct net_device *), 11999 unsigned int txqs, unsigned int rxqs) 12000 { 12001 struct net_device *dev; 12002 size_t napi_config_sz; 12003 unsigned int maxqs; 12004 12005 BUG_ON(strlen(name) >= sizeof(dev->name)); 12006 12007 if (txqs < 1) { 12008 pr_err("alloc_netdev: Unable to allocate device with zero queues\n"); 12009 return NULL; 12010 } 12011 12012 if (rxqs < 1) { 12013 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n"); 12014 return NULL; 12015 } 12016 12017 maxqs = max(txqs, rxqs); 12018 12019 dev = kvzalloc(struct_size(dev, priv, sizeof_priv), 12020 GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 12021 if (!dev) 12022 return NULL; 12023 12024 dev->priv_len = sizeof_priv; 12025 12026 ref_tracker_dir_init(&dev->refcnt_tracker, 128, "netdev"); 12027 #ifdef CONFIG_PCPU_DEV_REFCNT 12028 dev->pcpu_refcnt = alloc_percpu(int); 12029 if (!dev->pcpu_refcnt) 12030 goto free_dev; 12031 __dev_hold(dev); 12032 #else 12033 refcount_set(&dev->dev_refcnt, 1); 12034 #endif 12035 12036 if (dev_addr_init(dev)) 12037 goto free_pcpu; 12038 12039 dev_mc_init(dev); 12040 dev_uc_init(dev); 12041 12042 dev_net_set(dev, &init_net); 12043 12044 dev->gso_max_size = GSO_LEGACY_MAX_SIZE; 12045 dev->xdp_zc_max_segs = 1; 12046 dev->gso_max_segs = GSO_MAX_SEGS; 12047 dev->gro_max_size = GRO_LEGACY_MAX_SIZE; 12048 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE; 12049 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE; 12050 dev->tso_max_size = TSO_LEGACY_MAX_SIZE; 12051 dev->tso_max_segs = TSO_MAX_SEGS; 12052 dev->upper_level = 1; 12053 dev->lower_level = 1; 12054 #ifdef CONFIG_LOCKDEP 12055 dev->nested_level = 0; 12056 INIT_LIST_HEAD(&dev->unlink_list); 12057 #endif 12058 12059 INIT_LIST_HEAD(&dev->napi_list); 12060 INIT_LIST_HEAD(&dev->unreg_list); 12061 INIT_LIST_HEAD(&dev->close_list); 12062 INIT_LIST_HEAD(&dev->link_watch_list); 12063 INIT_LIST_HEAD(&dev->adj_list.upper); 12064 INIT_LIST_HEAD(&dev->adj_list.lower); 12065 INIT_LIST_HEAD(&dev->ptype_all); 12066 INIT_LIST_HEAD(&dev->ptype_specific); 12067 INIT_LIST_HEAD(&dev->net_notifier_list); 12068 #ifdef CONFIG_NET_SCHED 12069 hash_init(dev->qdisc_hash); 12070 #endif 12071 12072 mutex_init(&dev->lock); 12073 12074 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 12075 setup(dev); 12076 12077 if (!dev->tx_queue_len) { 12078 dev->priv_flags |= IFF_NO_QUEUE; 12079 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN; 12080 } 12081 12082 dev->num_tx_queues = txqs; 12083 dev->real_num_tx_queues = txqs; 12084 if (netif_alloc_netdev_queues(dev)) 12085 goto free_all; 12086 12087 dev->num_rx_queues = rxqs; 12088 dev->real_num_rx_queues = rxqs; 12089 if (netif_alloc_rx_queues(dev)) 12090 goto free_all; 12091 dev->ethtool = kzalloc(sizeof(*dev->ethtool), GFP_KERNEL_ACCOUNT); 12092 if (!dev->ethtool) 12093 goto free_all; 12094 12095 dev->cfg = kzalloc(sizeof(*dev->cfg), GFP_KERNEL_ACCOUNT); 12096 if (!dev->cfg) 12097 goto free_all; 12098 dev->cfg_pending = dev->cfg; 12099 12100 dev->num_napi_configs = maxqs; 12101 napi_config_sz = array_size(maxqs, sizeof(*dev->napi_config)); 12102 dev->napi_config = kvzalloc(napi_config_sz, GFP_KERNEL_ACCOUNT); 12103 if (!dev->napi_config) 12104 goto free_all; 12105 12106 strscpy(dev->name, name); 12107 dev->name_assign_type = name_assign_type; 12108 dev->group = INIT_NETDEV_GROUP; 12109 if (!dev->ethtool_ops) 12110 dev->ethtool_ops = &default_ethtool_ops; 12111 12112 nf_hook_netdev_init(dev); 12113 12114 return dev; 12115 12116 free_all: 12117 free_netdev(dev); 12118 return NULL; 12119 12120 free_pcpu: 12121 #ifdef CONFIG_PCPU_DEV_REFCNT 12122 free_percpu(dev->pcpu_refcnt); 12123 free_dev: 12124 #endif 12125 kvfree(dev); 12126 return NULL; 12127 } 12128 EXPORT_SYMBOL(alloc_netdev_mqs); 12129 12130 static void netdev_napi_exit(struct net_device *dev) 12131 { 12132 if (!list_empty(&dev->napi_list)) { 12133 struct napi_struct *p, *n; 12134 12135 netdev_lock(dev); 12136 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list) 12137 __netif_napi_del_locked(p); 12138 netdev_unlock(dev); 12139 12140 synchronize_net(); 12141 } 12142 12143 kvfree(dev->napi_config); 12144 } 12145 12146 /** 12147 * free_netdev - free network device 12148 * @dev: device 12149 * 12150 * This function does the last stage of destroying an allocated device 12151 * interface. The reference to the device object is released. If this 12152 * is the last reference then it will be freed.Must be called in process 12153 * context. 12154 */ 12155 void free_netdev(struct net_device *dev) 12156 { 12157 might_sleep(); 12158 12159 /* When called immediately after register_netdevice() failed the unwind 12160 * handling may still be dismantling the device. Handle that case by 12161 * deferring the free. 12162 */ 12163 if (dev->reg_state == NETREG_UNREGISTERING) { 12164 ASSERT_RTNL(); 12165 dev->needs_free_netdev = true; 12166 return; 12167 } 12168 12169 WARN_ON(dev->cfg != dev->cfg_pending); 12170 kfree(dev->cfg); 12171 kfree(dev->ethtool); 12172 netif_free_tx_queues(dev); 12173 netif_free_rx_queues(dev); 12174 12175 kfree(rcu_dereference_protected(dev->ingress_queue, 1)); 12176 12177 /* Flush device addresses */ 12178 dev_addr_flush(dev); 12179 12180 netdev_napi_exit(dev); 12181 12182 netif_del_cpu_rmap(dev); 12183 12184 ref_tracker_dir_exit(&dev->refcnt_tracker); 12185 #ifdef CONFIG_PCPU_DEV_REFCNT 12186 free_percpu(dev->pcpu_refcnt); 12187 dev->pcpu_refcnt = NULL; 12188 #endif 12189 free_percpu(dev->core_stats); 12190 dev->core_stats = NULL; 12191 free_percpu(dev->xdp_bulkq); 12192 dev->xdp_bulkq = NULL; 12193 12194 netdev_free_phy_link_topology(dev); 12195 12196 mutex_destroy(&dev->lock); 12197 12198 /* Compatibility with error handling in drivers */ 12199 if (dev->reg_state == NETREG_UNINITIALIZED || 12200 dev->reg_state == NETREG_DUMMY) { 12201 kvfree(dev); 12202 return; 12203 } 12204 12205 BUG_ON(dev->reg_state != NETREG_UNREGISTERED); 12206 WRITE_ONCE(dev->reg_state, NETREG_RELEASED); 12207 12208 /* will free via device release */ 12209 put_device(&dev->dev); 12210 } 12211 EXPORT_SYMBOL(free_netdev); 12212 12213 /** 12214 * alloc_netdev_dummy - Allocate and initialize a dummy net device. 12215 * @sizeof_priv: size of private data to allocate space for 12216 * 12217 * Return: the allocated net_device on success, NULL otherwise 12218 */ 12219 struct net_device *alloc_netdev_dummy(int sizeof_priv) 12220 { 12221 return alloc_netdev(sizeof_priv, "dummy#", NET_NAME_UNKNOWN, 12222 init_dummy_netdev); 12223 } 12224 EXPORT_SYMBOL_GPL(alloc_netdev_dummy); 12225 12226 /** 12227 * synchronize_net - Synchronize with packet receive processing 12228 * 12229 * Wait for packets currently being received to be done. 12230 * Does not block later packets from starting. 12231 */ 12232 void synchronize_net(void) 12233 { 12234 might_sleep(); 12235 if (from_cleanup_net() || rtnl_is_locked()) 12236 synchronize_rcu_expedited(); 12237 else 12238 synchronize_rcu(); 12239 } 12240 EXPORT_SYMBOL(synchronize_net); 12241 12242 static void netdev_rss_contexts_free(struct net_device *dev) 12243 { 12244 struct ethtool_rxfh_context *ctx; 12245 unsigned long context; 12246 12247 mutex_lock(&dev->ethtool->rss_lock); 12248 xa_for_each(&dev->ethtool->rss_ctx, context, ctx) { 12249 xa_erase(&dev->ethtool->rss_ctx, context); 12250 dev->ethtool_ops->remove_rxfh_context(dev, ctx, context, NULL); 12251 kfree(ctx); 12252 } 12253 xa_destroy(&dev->ethtool->rss_ctx); 12254 mutex_unlock(&dev->ethtool->rss_lock); 12255 } 12256 12257 /** 12258 * unregister_netdevice_queue - remove device from the kernel 12259 * @dev: device 12260 * @head: list 12261 * 12262 * This function shuts down a device interface and removes it 12263 * from the kernel tables. 12264 * If head not NULL, device is queued to be unregistered later. 12265 * 12266 * Callers must hold the rtnl semaphore. You may want 12267 * unregister_netdev() instead of this. 12268 */ 12269 12270 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head) 12271 { 12272 ASSERT_RTNL(); 12273 12274 if (head) { 12275 list_move_tail(&dev->unreg_list, head); 12276 } else { 12277 LIST_HEAD(single); 12278 12279 list_add(&dev->unreg_list, &single); 12280 unregister_netdevice_many(&single); 12281 } 12282 } 12283 EXPORT_SYMBOL(unregister_netdevice_queue); 12284 12285 static void dev_memory_provider_uninstall(struct net_device *dev) 12286 { 12287 unsigned int i; 12288 12289 for (i = 0; i < dev->real_num_rx_queues; i++) { 12290 struct netdev_rx_queue *rxq = &dev->_rx[i]; 12291 struct pp_memory_provider_params *p = &rxq->mp_params; 12292 12293 if (p->mp_ops && p->mp_ops->uninstall) 12294 p->mp_ops->uninstall(rxq->mp_params.mp_priv, rxq); 12295 } 12296 } 12297 12298 /* devices must be UP and netdev_lock()'d */ 12299 static void netif_close_many_and_unlock(struct list_head *close_head) 12300 { 12301 struct net_device *dev, *tmp; 12302 12303 netif_close_many(close_head, false); 12304 12305 /* ... now unlock them */ 12306 list_for_each_entry_safe(dev, tmp, close_head, close_list) { 12307 netdev_unlock(dev); 12308 list_del_init(&dev->close_list); 12309 } 12310 } 12311 12312 static void netif_close_many_and_unlock_cond(struct list_head *close_head) 12313 { 12314 #ifdef CONFIG_LOCKDEP 12315 /* We can only track up to MAX_LOCK_DEPTH locks per task. 12316 * 12317 * Reserve half the available slots for additional locks possibly 12318 * taken by notifiers and (soft)irqs. 12319 */ 12320 unsigned int limit = MAX_LOCK_DEPTH / 2; 12321 12322 if (lockdep_depth(current) > limit) 12323 netif_close_many_and_unlock(close_head); 12324 #endif 12325 } 12326 12327 void unregister_netdevice_many_notify(struct list_head *head, 12328 u32 portid, const struct nlmsghdr *nlh) 12329 { 12330 struct net_device *dev, *tmp; 12331 LIST_HEAD(close_head); 12332 int cnt = 0; 12333 12334 BUG_ON(dev_boot_phase); 12335 ASSERT_RTNL(); 12336 12337 if (list_empty(head)) 12338 return; 12339 12340 list_for_each_entry_safe(dev, tmp, head, unreg_list) { 12341 /* Some devices call without registering 12342 * for initialization unwind. Remove those 12343 * devices and proceed with the remaining. 12344 */ 12345 if (dev->reg_state == NETREG_UNINITIALIZED) { 12346 pr_debug("unregister_netdevice: device %s/%p never was registered\n", 12347 dev->name, dev); 12348 12349 WARN_ON(1); 12350 list_del(&dev->unreg_list); 12351 continue; 12352 } 12353 dev->dismantle = true; 12354 BUG_ON(dev->reg_state != NETREG_REGISTERED); 12355 } 12356 12357 /* If device is running, close it first. Start with ops locked... */ 12358 list_for_each_entry(dev, head, unreg_list) { 12359 if (!(dev->flags & IFF_UP)) 12360 continue; 12361 if (netdev_need_ops_lock(dev)) { 12362 list_add_tail(&dev->close_list, &close_head); 12363 netdev_lock(dev); 12364 } 12365 netif_close_many_and_unlock_cond(&close_head); 12366 } 12367 netif_close_many_and_unlock(&close_head); 12368 /* ... now go over the rest. */ 12369 list_for_each_entry(dev, head, unreg_list) { 12370 if (!netdev_need_ops_lock(dev)) 12371 list_add_tail(&dev->close_list, &close_head); 12372 } 12373 netif_close_many(&close_head, true); 12374 12375 list_for_each_entry(dev, head, unreg_list) { 12376 /* And unlink it from device chain. */ 12377 unlist_netdevice(dev); 12378 netdev_lock(dev); 12379 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERING); 12380 netdev_unlock(dev); 12381 } 12382 flush_all_backlogs(); 12383 12384 synchronize_net(); 12385 12386 list_for_each_entry(dev, head, unreg_list) { 12387 struct sk_buff *skb = NULL; 12388 12389 /* Shutdown queueing discipline. */ 12390 netdev_lock_ops(dev); 12391 dev_shutdown(dev); 12392 dev_tcx_uninstall(dev); 12393 dev_xdp_uninstall(dev); 12394 dev_memory_provider_uninstall(dev); 12395 netdev_unlock_ops(dev); 12396 bpf_dev_bound_netdev_unregister(dev); 12397 12398 netdev_offload_xstats_disable_all(dev); 12399 12400 /* Notify protocols, that we are about to destroy 12401 * this device. They should clean all the things. 12402 */ 12403 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 12404 12405 if (!(dev->rtnl_link_ops && dev->rtnl_link_initializing)) 12406 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0, 12407 GFP_KERNEL, NULL, 0, 12408 portid, nlh); 12409 12410 /* 12411 * Flush the unicast and multicast chains 12412 */ 12413 dev_uc_flush(dev); 12414 dev_mc_flush(dev); 12415 12416 netdev_name_node_alt_flush(dev); 12417 netdev_name_node_free(dev->name_node); 12418 12419 netdev_rss_contexts_free(dev); 12420 12421 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 12422 12423 if (dev->netdev_ops->ndo_uninit) 12424 dev->netdev_ops->ndo_uninit(dev); 12425 12426 mutex_destroy(&dev->ethtool->rss_lock); 12427 12428 net_shaper_flush_netdev(dev); 12429 12430 if (skb) 12431 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh); 12432 12433 /* Notifier chain MUST detach us all upper devices. */ 12434 WARN_ON(netdev_has_any_upper_dev(dev)); 12435 WARN_ON(netdev_has_any_lower_dev(dev)); 12436 12437 /* Remove entries from kobject tree */ 12438 netdev_unregister_kobject(dev); 12439 #ifdef CONFIG_XPS 12440 /* Remove XPS queueing entries */ 12441 netif_reset_xps_queues_gt(dev, 0); 12442 #endif 12443 } 12444 12445 synchronize_net(); 12446 12447 list_for_each_entry(dev, head, unreg_list) { 12448 netdev_put(dev, &dev->dev_registered_tracker); 12449 net_set_todo(dev); 12450 cnt++; 12451 } 12452 atomic_add(cnt, &dev_unreg_count); 12453 12454 list_del(head); 12455 } 12456 12457 /** 12458 * unregister_netdevice_many - unregister many devices 12459 * @head: list of devices 12460 * 12461 * Note: As most callers use a stack allocated list_head, 12462 * we force a list_del() to make sure stack won't be corrupted later. 12463 */ 12464 void unregister_netdevice_many(struct list_head *head) 12465 { 12466 unregister_netdevice_many_notify(head, 0, NULL); 12467 } 12468 EXPORT_SYMBOL(unregister_netdevice_many); 12469 12470 /** 12471 * unregister_netdev - remove device from the kernel 12472 * @dev: device 12473 * 12474 * This function shuts down a device interface and removes it 12475 * from the kernel tables. 12476 * 12477 * This is just a wrapper for unregister_netdevice that takes 12478 * the rtnl semaphore. In general you want to use this and not 12479 * unregister_netdevice. 12480 */ 12481 void unregister_netdev(struct net_device *dev) 12482 { 12483 rtnl_net_dev_lock(dev); 12484 unregister_netdevice(dev); 12485 rtnl_net_dev_unlock(dev); 12486 } 12487 EXPORT_SYMBOL(unregister_netdev); 12488 12489 int __dev_change_net_namespace(struct net_device *dev, struct net *net, 12490 const char *pat, int new_ifindex, 12491 struct netlink_ext_ack *extack) 12492 { 12493 struct netdev_name_node *name_node; 12494 struct net *net_old = dev_net(dev); 12495 char new_name[IFNAMSIZ] = {}; 12496 int err, new_nsid; 12497 12498 ASSERT_RTNL(); 12499 12500 /* Don't allow namespace local devices to be moved. */ 12501 err = -EINVAL; 12502 if (dev->netns_immutable) { 12503 NL_SET_ERR_MSG(extack, "The interface netns is immutable"); 12504 goto out; 12505 } 12506 12507 /* Ensure the device has been registered */ 12508 if (dev->reg_state != NETREG_REGISTERED) { 12509 NL_SET_ERR_MSG(extack, "The interface isn't registered"); 12510 goto out; 12511 } 12512 12513 /* Get out if there is nothing todo */ 12514 err = 0; 12515 if (net_eq(net_old, net)) 12516 goto out; 12517 12518 /* Pick the destination device name, and ensure 12519 * we can use it in the destination network namespace. 12520 */ 12521 err = -EEXIST; 12522 if (netdev_name_in_use(net, dev->name)) { 12523 /* We get here if we can't use the current device name */ 12524 if (!pat) { 12525 NL_SET_ERR_MSG(extack, 12526 "An interface with the same name exists in the target netns"); 12527 goto out; 12528 } 12529 err = dev_prep_valid_name(net, dev, pat, new_name, EEXIST); 12530 if (err < 0) { 12531 NL_SET_ERR_MSG_FMT(extack, 12532 "Unable to use '%s' for the new interface name in the target netns", 12533 pat); 12534 goto out; 12535 } 12536 } 12537 /* Check that none of the altnames conflicts. */ 12538 err = -EEXIST; 12539 netdev_for_each_altname(dev, name_node) { 12540 if (netdev_name_in_use(net, name_node->name)) { 12541 NL_SET_ERR_MSG_FMT(extack, 12542 "An interface with the altname %s exists in the target netns", 12543 name_node->name); 12544 goto out; 12545 } 12546 } 12547 12548 /* Check that new_ifindex isn't used yet. */ 12549 if (new_ifindex) { 12550 err = dev_index_reserve(net, new_ifindex); 12551 if (err < 0) { 12552 NL_SET_ERR_MSG_FMT(extack, 12553 "The ifindex %d is not available in the target netns", 12554 new_ifindex); 12555 goto out; 12556 } 12557 } else { 12558 /* If there is an ifindex conflict assign a new one */ 12559 err = dev_index_reserve(net, dev->ifindex); 12560 if (err == -EBUSY) 12561 err = dev_index_reserve(net, 0); 12562 if (err < 0) { 12563 NL_SET_ERR_MSG(extack, 12564 "Unable to allocate a new ifindex in the target netns"); 12565 goto out; 12566 } 12567 new_ifindex = err; 12568 } 12569 12570 /* 12571 * And now a mini version of register_netdevice unregister_netdevice. 12572 */ 12573 12574 netdev_lock_ops(dev); 12575 /* If device is running close it first. */ 12576 netif_close(dev); 12577 /* And unlink it from device chain */ 12578 unlist_netdevice(dev); 12579 12580 if (!netdev_need_ops_lock(dev)) 12581 netdev_lock(dev); 12582 dev->moving_ns = true; 12583 netdev_unlock(dev); 12584 12585 synchronize_net(); 12586 12587 /* Shutdown queueing discipline. */ 12588 netdev_lock_ops(dev); 12589 dev_shutdown(dev); 12590 netdev_unlock_ops(dev); 12591 12592 /* Notify protocols, that we are about to destroy 12593 * this device. They should clean all the things. 12594 * 12595 * Note that dev->reg_state stays at NETREG_REGISTERED. 12596 * This is wanted because this way 8021q and macvlan know 12597 * the device is just moving and can keep their slaves up. 12598 */ 12599 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 12600 rcu_barrier(); 12601 12602 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL); 12603 12604 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid, 12605 new_ifindex); 12606 12607 /* 12608 * Flush the unicast and multicast chains 12609 */ 12610 dev_uc_flush(dev); 12611 dev_mc_flush(dev); 12612 12613 /* Send a netdev-removed uevent to the old namespace */ 12614 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE); 12615 netdev_adjacent_del_links(dev); 12616 12617 /* Move per-net netdevice notifiers that are following the netdevice */ 12618 move_netdevice_notifiers_dev_net(dev, net); 12619 12620 /* Actually switch the network namespace */ 12621 netdev_lock(dev); 12622 dev_net_set(dev, net); 12623 netdev_unlock(dev); 12624 dev->ifindex = new_ifindex; 12625 12626 if (new_name[0]) { 12627 /* Rename the netdev to prepared name */ 12628 write_seqlock_bh(&netdev_rename_lock); 12629 strscpy(dev->name, new_name, IFNAMSIZ); 12630 write_sequnlock_bh(&netdev_rename_lock); 12631 } 12632 12633 /* Fixup kobjects */ 12634 dev_set_uevent_suppress(&dev->dev, 1); 12635 err = device_rename(&dev->dev, dev->name); 12636 dev_set_uevent_suppress(&dev->dev, 0); 12637 WARN_ON(err); 12638 12639 /* Send a netdev-add uevent to the new namespace */ 12640 kobject_uevent(&dev->dev.kobj, KOBJ_ADD); 12641 netdev_adjacent_add_links(dev); 12642 12643 /* Adapt owner in case owning user namespace of target network 12644 * namespace is different from the original one. 12645 */ 12646 err = netdev_change_owner(dev, net_old, net); 12647 WARN_ON(err); 12648 12649 netdev_lock(dev); 12650 dev->moving_ns = false; 12651 if (!netdev_need_ops_lock(dev)) 12652 netdev_unlock(dev); 12653 12654 /* Add the device back in the hashes */ 12655 list_netdevice(dev); 12656 /* Notify protocols, that a new device appeared. */ 12657 call_netdevice_notifiers(NETDEV_REGISTER, dev); 12658 netdev_unlock_ops(dev); 12659 12660 /* 12661 * Prevent userspace races by waiting until the network 12662 * device is fully setup before sending notifications. 12663 */ 12664 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 12665 12666 synchronize_net(); 12667 err = 0; 12668 out: 12669 return err; 12670 } 12671 12672 static int dev_cpu_dead(unsigned int oldcpu) 12673 { 12674 struct sk_buff **list_skb; 12675 struct sk_buff *skb; 12676 unsigned int cpu; 12677 struct softnet_data *sd, *oldsd, *remsd = NULL; 12678 12679 local_irq_disable(); 12680 cpu = smp_processor_id(); 12681 sd = &per_cpu(softnet_data, cpu); 12682 oldsd = &per_cpu(softnet_data, oldcpu); 12683 12684 /* Find end of our completion_queue. */ 12685 list_skb = &sd->completion_queue; 12686 while (*list_skb) 12687 list_skb = &(*list_skb)->next; 12688 /* Append completion queue from offline CPU. */ 12689 *list_skb = oldsd->completion_queue; 12690 oldsd->completion_queue = NULL; 12691 12692 /* Append output queue from offline CPU. */ 12693 if (oldsd->output_queue) { 12694 *sd->output_queue_tailp = oldsd->output_queue; 12695 sd->output_queue_tailp = oldsd->output_queue_tailp; 12696 oldsd->output_queue = NULL; 12697 oldsd->output_queue_tailp = &oldsd->output_queue; 12698 } 12699 /* Append NAPI poll list from offline CPU, with one exception : 12700 * process_backlog() must be called by cpu owning percpu backlog. 12701 * We properly handle process_queue & input_pkt_queue later. 12702 */ 12703 while (!list_empty(&oldsd->poll_list)) { 12704 struct napi_struct *napi = list_first_entry(&oldsd->poll_list, 12705 struct napi_struct, 12706 poll_list); 12707 12708 list_del_init(&napi->poll_list); 12709 if (napi->poll == process_backlog) 12710 napi->state &= NAPIF_STATE_THREADED; 12711 else 12712 ____napi_schedule(sd, napi); 12713 } 12714 12715 raise_softirq_irqoff(NET_TX_SOFTIRQ); 12716 local_irq_enable(); 12717 12718 if (!use_backlog_threads()) { 12719 #ifdef CONFIG_RPS 12720 remsd = oldsd->rps_ipi_list; 12721 oldsd->rps_ipi_list = NULL; 12722 #endif 12723 /* send out pending IPI's on offline CPU */ 12724 net_rps_send_ipi(remsd); 12725 } 12726 12727 /* Process offline CPU's input_pkt_queue */ 12728 while ((skb = __skb_dequeue(&oldsd->process_queue))) { 12729 netif_rx(skb); 12730 rps_input_queue_head_incr(oldsd); 12731 } 12732 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) { 12733 netif_rx(skb); 12734 rps_input_queue_head_incr(oldsd); 12735 } 12736 12737 return 0; 12738 } 12739 12740 /** 12741 * netdev_increment_features - increment feature set by one 12742 * @all: current feature set 12743 * @one: new feature set 12744 * @mask: mask feature set 12745 * 12746 * Computes a new feature set after adding a device with feature set 12747 * @one to the master device with current feature set @all. Will not 12748 * enable anything that is off in @mask. Returns the new feature set. 12749 */ 12750 netdev_features_t netdev_increment_features(netdev_features_t all, 12751 netdev_features_t one, netdev_features_t mask) 12752 { 12753 if (mask & NETIF_F_HW_CSUM) 12754 mask |= NETIF_F_CSUM_MASK; 12755 mask |= NETIF_F_VLAN_CHALLENGED; 12756 12757 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask; 12758 all &= one | ~NETIF_F_ALL_FOR_ALL; 12759 12760 /* If one device supports hw checksumming, set for all. */ 12761 if (all & NETIF_F_HW_CSUM) 12762 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM); 12763 12764 return all; 12765 } 12766 EXPORT_SYMBOL(netdev_increment_features); 12767 12768 /** 12769 * netdev_compute_master_upper_features - compute feature from lowers 12770 * @dev: the upper device 12771 * @update_header: whether to update upper device's header_len/headroom/tailroom 12772 * 12773 * Recompute the upper device's feature based on all lower devices. 12774 */ 12775 void netdev_compute_master_upper_features(struct net_device *dev, bool update_header) 12776 { 12777 unsigned int dst_release_flag = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 12778 netdev_features_t gso_partial_features = MASTER_UPPER_DEV_GSO_PARTIAL_FEATURES; 12779 netdev_features_t xfrm_features = MASTER_UPPER_DEV_XFRM_FEATURES; 12780 netdev_features_t mpls_features = MASTER_UPPER_DEV_MPLS_FEATURES; 12781 netdev_features_t vlan_features = MASTER_UPPER_DEV_VLAN_FEATURES; 12782 netdev_features_t enc_features = MASTER_UPPER_DEV_ENC_FEATURES; 12783 unsigned short max_header_len = ETH_HLEN; 12784 unsigned int tso_max_size = TSO_MAX_SIZE; 12785 unsigned short max_headroom = 0; 12786 unsigned short max_tailroom = 0; 12787 u16 tso_max_segs = TSO_MAX_SEGS; 12788 struct net_device *lower_dev; 12789 struct list_head *iter; 12790 12791 mpls_features = netdev_base_features(mpls_features); 12792 vlan_features = netdev_base_features(vlan_features); 12793 enc_features = netdev_base_features(enc_features); 12794 12795 netdev_for_each_lower_dev(dev, lower_dev, iter) { 12796 gso_partial_features = netdev_increment_features(gso_partial_features, 12797 lower_dev->gso_partial_features, 12798 MASTER_UPPER_DEV_GSO_PARTIAL_FEATURES); 12799 12800 vlan_features = netdev_increment_features(vlan_features, 12801 lower_dev->vlan_features, 12802 MASTER_UPPER_DEV_VLAN_FEATURES); 12803 12804 enc_features = netdev_increment_features(enc_features, 12805 lower_dev->hw_enc_features, 12806 MASTER_UPPER_DEV_ENC_FEATURES); 12807 12808 if (IS_ENABLED(CONFIG_XFRM_OFFLOAD)) 12809 xfrm_features = netdev_increment_features(xfrm_features, 12810 lower_dev->hw_enc_features, 12811 MASTER_UPPER_DEV_XFRM_FEATURES); 12812 12813 mpls_features = netdev_increment_features(mpls_features, 12814 lower_dev->mpls_features, 12815 MASTER_UPPER_DEV_MPLS_FEATURES); 12816 12817 dst_release_flag &= lower_dev->priv_flags; 12818 12819 if (update_header) { 12820 max_header_len = max(max_header_len, lower_dev->hard_header_len); 12821 max_headroom = max(max_headroom, lower_dev->needed_headroom); 12822 max_tailroom = max(max_tailroom, lower_dev->needed_tailroom); 12823 } 12824 12825 tso_max_size = min(tso_max_size, lower_dev->tso_max_size); 12826 tso_max_segs = min(tso_max_segs, lower_dev->tso_max_segs); 12827 } 12828 12829 dev->gso_partial_features = gso_partial_features; 12830 dev->vlan_features = vlan_features; 12831 dev->hw_enc_features = enc_features | NETIF_F_GSO_ENCAP_ALL | 12832 NETIF_F_HW_VLAN_CTAG_TX | 12833 NETIF_F_HW_VLAN_STAG_TX; 12834 if (IS_ENABLED(CONFIG_XFRM_OFFLOAD)) 12835 dev->hw_enc_features |= xfrm_features; 12836 dev->mpls_features = mpls_features; 12837 12838 dev->priv_flags &= ~IFF_XMIT_DST_RELEASE; 12839 if ((dev->priv_flags & IFF_XMIT_DST_RELEASE_PERM) && 12840 dst_release_flag == (IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM)) 12841 dev->priv_flags |= IFF_XMIT_DST_RELEASE; 12842 12843 if (update_header) { 12844 dev->hard_header_len = max_header_len; 12845 dev->needed_headroom = max_headroom; 12846 dev->needed_tailroom = max_tailroom; 12847 } 12848 12849 netif_set_tso_max_segs(dev, tso_max_segs); 12850 netif_set_tso_max_size(dev, tso_max_size); 12851 12852 netdev_change_features(dev); 12853 } 12854 EXPORT_SYMBOL(netdev_compute_master_upper_features); 12855 12856 static struct hlist_head * __net_init netdev_create_hash(void) 12857 { 12858 int i; 12859 struct hlist_head *hash; 12860 12861 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL); 12862 if (hash != NULL) 12863 for (i = 0; i < NETDEV_HASHENTRIES; i++) 12864 INIT_HLIST_HEAD(&hash[i]); 12865 12866 return hash; 12867 } 12868 12869 /* Initialize per network namespace state */ 12870 static int __net_init netdev_init(struct net *net) 12871 { 12872 BUILD_BUG_ON(GRO_HASH_BUCKETS > 12873 BITS_PER_BYTE * sizeof_field(struct gro_node, bitmask)); 12874 12875 INIT_LIST_HEAD(&net->dev_base_head); 12876 12877 net->dev_name_head = netdev_create_hash(); 12878 if (net->dev_name_head == NULL) 12879 goto err_name; 12880 12881 net->dev_index_head = netdev_create_hash(); 12882 if (net->dev_index_head == NULL) 12883 goto err_idx; 12884 12885 xa_init_flags(&net->dev_by_index, XA_FLAGS_ALLOC1); 12886 12887 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain); 12888 12889 return 0; 12890 12891 err_idx: 12892 kfree(net->dev_name_head); 12893 err_name: 12894 return -ENOMEM; 12895 } 12896 12897 /** 12898 * netdev_drivername - network driver for the device 12899 * @dev: network device 12900 * 12901 * Determine network driver for device. 12902 */ 12903 const char *netdev_drivername(const struct net_device *dev) 12904 { 12905 const struct device_driver *driver; 12906 const struct device *parent; 12907 const char *empty = ""; 12908 12909 parent = dev->dev.parent; 12910 if (!parent) 12911 return empty; 12912 12913 driver = parent->driver; 12914 if (driver && driver->name) 12915 return driver->name; 12916 return empty; 12917 } 12918 12919 static void __netdev_printk(const char *level, const struct net_device *dev, 12920 struct va_format *vaf) 12921 { 12922 if (dev && dev->dev.parent) { 12923 dev_printk_emit(level[1] - '0', 12924 dev->dev.parent, 12925 "%s %s %s%s: %pV", 12926 dev_driver_string(dev->dev.parent), 12927 dev_name(dev->dev.parent), 12928 netdev_name(dev), netdev_reg_state(dev), 12929 vaf); 12930 } else if (dev) { 12931 printk("%s%s%s: %pV", 12932 level, netdev_name(dev), netdev_reg_state(dev), vaf); 12933 } else { 12934 printk("%s(NULL net_device): %pV", level, vaf); 12935 } 12936 } 12937 12938 void netdev_printk(const char *level, const struct net_device *dev, 12939 const char *format, ...) 12940 { 12941 struct va_format vaf; 12942 va_list args; 12943 12944 va_start(args, format); 12945 12946 vaf.fmt = format; 12947 vaf.va = &args; 12948 12949 __netdev_printk(level, dev, &vaf); 12950 12951 va_end(args); 12952 } 12953 EXPORT_SYMBOL(netdev_printk); 12954 12955 #define define_netdev_printk_level(func, level) \ 12956 void func(const struct net_device *dev, const char *fmt, ...) \ 12957 { \ 12958 struct va_format vaf; \ 12959 va_list args; \ 12960 \ 12961 va_start(args, fmt); \ 12962 \ 12963 vaf.fmt = fmt; \ 12964 vaf.va = &args; \ 12965 \ 12966 __netdev_printk(level, dev, &vaf); \ 12967 \ 12968 va_end(args); \ 12969 } \ 12970 EXPORT_SYMBOL(func); 12971 12972 define_netdev_printk_level(netdev_emerg, KERN_EMERG); 12973 define_netdev_printk_level(netdev_alert, KERN_ALERT); 12974 define_netdev_printk_level(netdev_crit, KERN_CRIT); 12975 define_netdev_printk_level(netdev_err, KERN_ERR); 12976 define_netdev_printk_level(netdev_warn, KERN_WARNING); 12977 define_netdev_printk_level(netdev_notice, KERN_NOTICE); 12978 define_netdev_printk_level(netdev_info, KERN_INFO); 12979 12980 static void __net_exit netdev_exit(struct net *net) 12981 { 12982 kfree(net->dev_name_head); 12983 kfree(net->dev_index_head); 12984 xa_destroy(&net->dev_by_index); 12985 if (net != &init_net) 12986 WARN_ON_ONCE(!list_empty(&net->dev_base_head)); 12987 } 12988 12989 static struct pernet_operations __net_initdata netdev_net_ops = { 12990 .init = netdev_init, 12991 .exit = netdev_exit, 12992 }; 12993 12994 static void __net_exit default_device_exit_net(struct net *net) 12995 { 12996 struct netdev_name_node *name_node, *tmp; 12997 struct net_device *dev, *aux; 12998 /* 12999 * Push all migratable network devices back to the 13000 * initial network namespace 13001 */ 13002 ASSERT_RTNL(); 13003 for_each_netdev_safe(net, dev, aux) { 13004 int err; 13005 char fb_name[IFNAMSIZ]; 13006 13007 /* Ignore unmoveable devices (i.e. loopback) */ 13008 if (dev->netns_immutable) 13009 continue; 13010 13011 /* Leave virtual devices for the generic cleanup */ 13012 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund) 13013 continue; 13014 13015 /* Push remaining network devices to init_net */ 13016 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex); 13017 if (netdev_name_in_use(&init_net, fb_name)) 13018 snprintf(fb_name, IFNAMSIZ, "dev%%d"); 13019 13020 netdev_for_each_altname_safe(dev, name_node, tmp) 13021 if (netdev_name_in_use(&init_net, name_node->name)) 13022 __netdev_name_node_alt_destroy(name_node); 13023 13024 err = dev_change_net_namespace(dev, &init_net, fb_name); 13025 if (err) { 13026 pr_emerg("%s: failed to move %s to init_net: %d\n", 13027 __func__, dev->name, err); 13028 BUG(); 13029 } 13030 } 13031 } 13032 13033 static void __net_exit default_device_exit_batch(struct list_head *net_list) 13034 { 13035 /* At exit all network devices most be removed from a network 13036 * namespace. Do this in the reverse order of registration. 13037 * Do this across as many network namespaces as possible to 13038 * improve batching efficiency. 13039 */ 13040 struct net_device *dev; 13041 struct net *net; 13042 LIST_HEAD(dev_kill_list); 13043 13044 rtnl_lock(); 13045 list_for_each_entry(net, net_list, exit_list) { 13046 default_device_exit_net(net); 13047 cond_resched(); 13048 } 13049 13050 list_for_each_entry(net, net_list, exit_list) { 13051 for_each_netdev_reverse(net, dev) { 13052 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink) 13053 dev->rtnl_link_ops->dellink(dev, &dev_kill_list); 13054 else 13055 unregister_netdevice_queue(dev, &dev_kill_list); 13056 } 13057 } 13058 unregister_netdevice_many(&dev_kill_list); 13059 rtnl_unlock(); 13060 } 13061 13062 static struct pernet_operations __net_initdata default_device_ops = { 13063 .exit_batch = default_device_exit_batch, 13064 }; 13065 13066 static void __init net_dev_struct_check(void) 13067 { 13068 /* TX read-mostly hotpath */ 13069 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, priv_flags_fast); 13070 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, netdev_ops); 13071 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, header_ops); 13072 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, _tx); 13073 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, real_num_tx_queues); 13074 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_size); 13075 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_ipv4_max_size); 13076 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_segs); 13077 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_partial_features); 13078 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, num_tc); 13079 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, mtu); 13080 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, needed_headroom); 13081 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tc_to_txq); 13082 #ifdef CONFIG_XPS 13083 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, xps_maps); 13084 #endif 13085 #ifdef CONFIG_NETFILTER_EGRESS 13086 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, nf_hooks_egress); 13087 #endif 13088 #ifdef CONFIG_NET_XGRESS 13089 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tcx_egress); 13090 #endif 13091 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_tx, 160); 13092 13093 /* TXRX read-mostly hotpath */ 13094 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, lstats); 13095 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, state); 13096 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, flags); 13097 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, hard_header_len); 13098 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, features); 13099 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, ip6_ptr); 13100 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_txrx, 46); 13101 13102 /* RX read-mostly hotpath */ 13103 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ptype_specific); 13104 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ifindex); 13105 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, real_num_rx_queues); 13106 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, _rx); 13107 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_max_size); 13108 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_ipv4_max_size); 13109 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler); 13110 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler_data); 13111 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, nd_net); 13112 #ifdef CONFIG_NETPOLL 13113 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, npinfo); 13114 #endif 13115 #ifdef CONFIG_NET_XGRESS 13116 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, tcx_ingress); 13117 #endif 13118 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_rx, 92); 13119 } 13120 13121 /* 13122 * Initialize the DEV module. At boot time this walks the device list and 13123 * unhooks any devices that fail to initialise (normally hardware not 13124 * present) and leaves us with a valid list of present and active devices. 13125 * 13126 */ 13127 13128 /* We allocate 256 pages for each CPU if PAGE_SHIFT is 12 */ 13129 #define SYSTEM_PERCPU_PAGE_POOL_SIZE ((1 << 20) / PAGE_SIZE) 13130 13131 static int net_page_pool_create(int cpuid) 13132 { 13133 #if IS_ENABLED(CONFIG_PAGE_POOL) 13134 struct page_pool_params page_pool_params = { 13135 .pool_size = SYSTEM_PERCPU_PAGE_POOL_SIZE, 13136 .flags = PP_FLAG_SYSTEM_POOL, 13137 .nid = cpu_to_mem(cpuid), 13138 }; 13139 struct page_pool *pp_ptr; 13140 int err; 13141 13142 pp_ptr = page_pool_create_percpu(&page_pool_params, cpuid); 13143 if (IS_ERR(pp_ptr)) 13144 return -ENOMEM; 13145 13146 err = xdp_reg_page_pool(pp_ptr); 13147 if (err) { 13148 page_pool_destroy(pp_ptr); 13149 return err; 13150 } 13151 13152 per_cpu(system_page_pool.pool, cpuid) = pp_ptr; 13153 #endif 13154 return 0; 13155 } 13156 13157 static int backlog_napi_should_run(unsigned int cpu) 13158 { 13159 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 13160 struct napi_struct *napi = &sd->backlog; 13161 13162 return test_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 13163 } 13164 13165 static void run_backlog_napi(unsigned int cpu) 13166 { 13167 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 13168 13169 napi_threaded_poll_loop(&sd->backlog, false); 13170 } 13171 13172 static void backlog_napi_setup(unsigned int cpu) 13173 { 13174 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 13175 struct napi_struct *napi = &sd->backlog; 13176 13177 napi->thread = this_cpu_read(backlog_napi); 13178 set_bit(NAPI_STATE_THREADED, &napi->state); 13179 } 13180 13181 static struct smp_hotplug_thread backlog_threads = { 13182 .store = &backlog_napi, 13183 .thread_should_run = backlog_napi_should_run, 13184 .thread_fn = run_backlog_napi, 13185 .thread_comm = "backlog_napi/%u", 13186 .setup = backlog_napi_setup, 13187 }; 13188 13189 /* 13190 * This is called single threaded during boot, so no need 13191 * to take the rtnl semaphore. 13192 */ 13193 static int __init net_dev_init(void) 13194 { 13195 int i, rc = -ENOMEM; 13196 13197 BUG_ON(!dev_boot_phase); 13198 13199 net_dev_struct_check(); 13200 13201 if (dev_proc_init()) 13202 goto out; 13203 13204 if (netdev_kobject_init()) 13205 goto out; 13206 13207 for (i = 0; i < PTYPE_HASH_SIZE; i++) 13208 INIT_LIST_HEAD(&ptype_base[i]); 13209 13210 if (register_pernet_subsys(&netdev_net_ops)) 13211 goto out; 13212 13213 /* 13214 * Initialise the packet receive queues. 13215 */ 13216 13217 flush_backlogs_fallback = flush_backlogs_alloc(); 13218 if (!flush_backlogs_fallback) 13219 goto out; 13220 13221 for_each_possible_cpu(i) { 13222 struct softnet_data *sd = &per_cpu(softnet_data, i); 13223 13224 skb_queue_head_init(&sd->input_pkt_queue); 13225 skb_queue_head_init(&sd->process_queue); 13226 #ifdef CONFIG_XFRM_OFFLOAD 13227 skb_queue_head_init(&sd->xfrm_backlog); 13228 #endif 13229 INIT_LIST_HEAD(&sd->poll_list); 13230 sd->output_queue_tailp = &sd->output_queue; 13231 #ifdef CONFIG_RPS 13232 INIT_CSD(&sd->csd, rps_trigger_softirq, sd); 13233 sd->cpu = i; 13234 #endif 13235 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd); 13236 13237 gro_init(&sd->backlog.gro); 13238 sd->backlog.poll = process_backlog; 13239 sd->backlog.weight = weight_p; 13240 INIT_LIST_HEAD(&sd->backlog.poll_list); 13241 13242 if (net_page_pool_create(i)) 13243 goto out; 13244 } 13245 net_hotdata.skb_defer_nodes = 13246 __alloc_percpu(sizeof(struct skb_defer_node) * nr_node_ids, 13247 __alignof__(struct skb_defer_node)); 13248 if (!net_hotdata.skb_defer_nodes) 13249 goto out; 13250 if (use_backlog_threads()) 13251 smpboot_register_percpu_thread(&backlog_threads); 13252 13253 dev_boot_phase = 0; 13254 13255 /* The loopback device is special if any other network devices 13256 * is present in a network namespace the loopback device must 13257 * be present. Since we now dynamically allocate and free the 13258 * loopback device ensure this invariant is maintained by 13259 * keeping the loopback device as the first device on the 13260 * list of network devices. Ensuring the loopback devices 13261 * is the first device that appears and the last network device 13262 * that disappears. 13263 */ 13264 if (register_pernet_device(&loopback_net_ops)) 13265 goto out; 13266 13267 if (register_pernet_device(&default_device_ops)) 13268 goto out; 13269 13270 open_softirq(NET_TX_SOFTIRQ, net_tx_action); 13271 open_softirq(NET_RX_SOFTIRQ, net_rx_action); 13272 13273 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead", 13274 NULL, dev_cpu_dead); 13275 WARN_ON(rc < 0); 13276 rc = 0; 13277 13278 /* avoid static key IPIs to isolated CPUs */ 13279 if (housekeeping_enabled(HK_TYPE_MISC)) 13280 net_enable_timestamp(); 13281 out: 13282 if (rc < 0) { 13283 for_each_possible_cpu(i) { 13284 struct page_pool *pp_ptr; 13285 13286 pp_ptr = per_cpu(system_page_pool.pool, i); 13287 if (!pp_ptr) 13288 continue; 13289 13290 xdp_unreg_page_pool(pp_ptr); 13291 page_pool_destroy(pp_ptr); 13292 per_cpu(system_page_pool.pool, i) = NULL; 13293 } 13294 } 13295 13296 return rc; 13297 } 13298 13299 subsys_initcall(net_dev_init); 13300