1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * NET3 Protocol independent device support routines. 4 * 5 * Derived from the non IP parts of dev.c 1.0.19 6 * Authors: Ross Biro 7 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 8 * Mark Evans, <evansmp@uhura.aston.ac.uk> 9 * 10 * Additional Authors: 11 * Florian la Roche <rzsfl@rz.uni-sb.de> 12 * Alan Cox <gw4pts@gw4pts.ampr.org> 13 * David Hinds <dahinds@users.sourceforge.net> 14 * Alexey Kuznetsov <kuznet@ms2.inr.ac.ru> 15 * Adam Sulmicki <adam@cfar.umd.edu> 16 * Pekka Riikonen <priikone@poesidon.pspt.fi> 17 * 18 * Changes: 19 * D.J. Barrow : Fixed bug where dev->refcnt gets set 20 * to 2 if register_netdev gets called 21 * before net_dev_init & also removed a 22 * few lines of code in the process. 23 * Alan Cox : device private ioctl copies fields back. 24 * Alan Cox : Transmit queue code does relevant 25 * stunts to keep the queue safe. 26 * Alan Cox : Fixed double lock. 27 * Alan Cox : Fixed promisc NULL pointer trap 28 * ???????? : Support the full private ioctl range 29 * Alan Cox : Moved ioctl permission check into 30 * drivers 31 * Tim Kordas : SIOCADDMULTI/SIOCDELMULTI 32 * Alan Cox : 100 backlog just doesn't cut it when 33 * you start doing multicast video 8) 34 * Alan Cox : Rewrote net_bh and list manager. 35 * Alan Cox : Fix ETH_P_ALL echoback lengths. 36 * Alan Cox : Took out transmit every packet pass 37 * Saved a few bytes in the ioctl handler 38 * Alan Cox : Network driver sets packet type before 39 * calling netif_rx. Saves a function 40 * call a packet. 41 * Alan Cox : Hashed net_bh() 42 * Richard Kooijman: Timestamp fixes. 43 * Alan Cox : Wrong field in SIOCGIFDSTADDR 44 * Alan Cox : Device lock protection. 45 * Alan Cox : Fixed nasty side effect of device close 46 * changes. 47 * Rudi Cilibrasi : Pass the right thing to 48 * set_mac_address() 49 * Dave Miller : 32bit quantity for the device lock to 50 * make it work out on a Sparc. 51 * Bjorn Ekwall : Added KERNELD hack. 52 * Alan Cox : Cleaned up the backlog initialise. 53 * Craig Metz : SIOCGIFCONF fix if space for under 54 * 1 device. 55 * Thomas Bogendoerfer : Return ENODEV for dev_open, if there 56 * is no device open function. 57 * Andi Kleen : Fix error reporting for SIOCGIFCONF 58 * Michael Chastain : Fix signed/unsigned for SIOCGIFCONF 59 * Cyrus Durgin : Cleaned for KMOD 60 * Adam Sulmicki : Bug Fix : Network Device Unload 61 * A network device unload needs to purge 62 * the backlog queue. 63 * Paul Rusty Russell : SIOCSIFNAME 64 * Pekka Riikonen : Netdev boot-time settings code 65 * Andrew Morton : Make unregister_netdevice wait 66 * indefinitely on dev->refcnt 67 * J Hadi Salim : - Backlog queue sampling 68 * - netif_rx() feedback 69 */ 70 71 #include <linux/uaccess.h> 72 #include <linux/bitmap.h> 73 #include <linux/capability.h> 74 #include <linux/cpu.h> 75 #include <linux/types.h> 76 #include <linux/kernel.h> 77 #include <linux/hash.h> 78 #include <linux/slab.h> 79 #include <linux/sched.h> 80 #include <linux/sched/isolation.h> 81 #include <linux/sched/mm.h> 82 #include <linux/smpboot.h> 83 #include <linux/mutex.h> 84 #include <linux/rwsem.h> 85 #include <linux/string.h> 86 #include <linux/mm.h> 87 #include <linux/socket.h> 88 #include <linux/sockios.h> 89 #include <linux/errno.h> 90 #include <linux/interrupt.h> 91 #include <linux/if_ether.h> 92 #include <linux/netdevice.h> 93 #include <linux/etherdevice.h> 94 #include <linux/ethtool.h> 95 #include <linux/ethtool_netlink.h> 96 #include <linux/skbuff.h> 97 #include <linux/kthread.h> 98 #include <linux/bpf.h> 99 #include <linux/bpf_trace.h> 100 #include <net/net_namespace.h> 101 #include <net/sock.h> 102 #include <net/busy_poll.h> 103 #include <linux/rtnetlink.h> 104 #include <linux/stat.h> 105 #include <net/dsa.h> 106 #include <net/dst.h> 107 #include <net/dst_metadata.h> 108 #include <net/gro.h> 109 #include <net/netdev_queues.h> 110 #include <net/pkt_sched.h> 111 #include <net/pkt_cls.h> 112 #include <net/checksum.h> 113 #include <net/xfrm.h> 114 #include <net/tcx.h> 115 #include <linux/highmem.h> 116 #include <linux/init.h> 117 #include <linux/module.h> 118 #include <linux/netpoll.h> 119 #include <linux/rcupdate.h> 120 #include <linux/delay.h> 121 #include <net/iw_handler.h> 122 #include <asm/current.h> 123 #include <linux/audit.h> 124 #include <linux/dmaengine.h> 125 #include <linux/err.h> 126 #include <linux/ctype.h> 127 #include <linux/if_arp.h> 128 #include <linux/if_vlan.h> 129 #include <linux/ip.h> 130 #include <net/ip.h> 131 #include <net/mpls.h> 132 #include <linux/ipv6.h> 133 #include <linux/in.h> 134 #include <linux/jhash.h> 135 #include <linux/random.h> 136 #include <trace/events/napi.h> 137 #include <trace/events/net.h> 138 #include <trace/events/skb.h> 139 #include <trace/events/qdisc.h> 140 #include <trace/events/xdp.h> 141 #include <linux/inetdevice.h> 142 #include <linux/cpu_rmap.h> 143 #include <linux/static_key.h> 144 #include <linux/hashtable.h> 145 #include <linux/vmalloc.h> 146 #include <linux/if_macvlan.h> 147 #include <linux/errqueue.h> 148 #include <linux/hrtimer.h> 149 #include <linux/netfilter_netdev.h> 150 #include <linux/crash_dump.h> 151 #include <linux/sctp.h> 152 #include <net/udp_tunnel.h> 153 #include <linux/net_namespace.h> 154 #include <linux/indirect_call_wrapper.h> 155 #include <net/devlink.h> 156 #include <linux/pm_runtime.h> 157 #include <linux/prandom.h> 158 #include <linux/once_lite.h> 159 #include <net/netdev_lock.h> 160 #include <net/netdev_rx_queue.h> 161 #include <net/page_pool/types.h> 162 #include <net/page_pool/helpers.h> 163 #include <net/page_pool/memory_provider.h> 164 #include <net/rps.h> 165 #include <linux/phy_link_topology.h> 166 167 #include "dev.h" 168 #include "devmem.h" 169 #include "net-sysfs.h" 170 171 static DEFINE_SPINLOCK(ptype_lock); 172 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly; 173 174 static int netif_rx_internal(struct sk_buff *skb); 175 static int call_netdevice_notifiers_extack(unsigned long val, 176 struct net_device *dev, 177 struct netlink_ext_ack *extack); 178 179 static DEFINE_MUTEX(ifalias_mutex); 180 181 /* protects napi_hash addition/deletion and napi_gen_id */ 182 static DEFINE_SPINLOCK(napi_hash_lock); 183 184 static unsigned int napi_gen_id = NR_CPUS; 185 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8); 186 187 static inline void dev_base_seq_inc(struct net *net) 188 { 189 unsigned int val = net->dev_base_seq + 1; 190 191 WRITE_ONCE(net->dev_base_seq, val ?: 1); 192 } 193 194 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name) 195 { 196 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ)); 197 198 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)]; 199 } 200 201 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex) 202 { 203 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)]; 204 } 205 206 #ifndef CONFIG_PREEMPT_RT 207 208 static DEFINE_STATIC_KEY_FALSE(use_backlog_threads_key); 209 210 static int __init setup_backlog_napi_threads(char *arg) 211 { 212 static_branch_enable(&use_backlog_threads_key); 213 return 0; 214 } 215 early_param("thread_backlog_napi", setup_backlog_napi_threads); 216 217 static bool use_backlog_threads(void) 218 { 219 return static_branch_unlikely(&use_backlog_threads_key); 220 } 221 222 #else 223 224 static bool use_backlog_threads(void) 225 { 226 return true; 227 } 228 229 #endif 230 231 static inline void backlog_lock_irq_save(struct softnet_data *sd, 232 unsigned long *flags) 233 { 234 if (IS_ENABLED(CONFIG_PREEMPT_RT)) { 235 spin_lock_irqsave(&sd->input_pkt_queue.lock, *flags); 236 } else { 237 local_irq_save(*flags); 238 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 239 spin_lock(&sd->input_pkt_queue.lock); 240 } 241 } 242 243 static inline void backlog_lock_irq_disable(struct softnet_data *sd) 244 { 245 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 246 spin_lock_irq(&sd->input_pkt_queue.lock); 247 else 248 local_irq_disable(); 249 } 250 251 static inline void backlog_unlock_irq_restore(struct softnet_data *sd, 252 unsigned long flags) 253 { 254 if (IS_ENABLED(CONFIG_PREEMPT_RT)) { 255 spin_unlock_irqrestore(&sd->input_pkt_queue.lock, flags); 256 } else { 257 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 258 spin_unlock(&sd->input_pkt_queue.lock); 259 local_irq_restore(flags); 260 } 261 } 262 263 static inline void backlog_unlock_irq_enable(struct softnet_data *sd) 264 { 265 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 266 spin_unlock_irq(&sd->input_pkt_queue.lock); 267 else 268 local_irq_enable(); 269 } 270 271 static struct netdev_name_node *netdev_name_node_alloc(struct net_device *dev, 272 const char *name) 273 { 274 struct netdev_name_node *name_node; 275 276 name_node = kmalloc_obj(*name_node); 277 if (!name_node) 278 return NULL; 279 INIT_HLIST_NODE(&name_node->hlist); 280 name_node->dev = dev; 281 name_node->name = name; 282 return name_node; 283 } 284 285 static struct netdev_name_node * 286 netdev_name_node_head_alloc(struct net_device *dev) 287 { 288 struct netdev_name_node *name_node; 289 290 name_node = netdev_name_node_alloc(dev, dev->name); 291 if (!name_node) 292 return NULL; 293 INIT_LIST_HEAD(&name_node->list); 294 return name_node; 295 } 296 297 static void netdev_name_node_free(struct netdev_name_node *name_node) 298 { 299 kfree(name_node); 300 } 301 302 static void netdev_name_node_add(struct net *net, 303 struct netdev_name_node *name_node) 304 { 305 hlist_add_head_rcu(&name_node->hlist, 306 dev_name_hash(net, name_node->name)); 307 } 308 309 static void netdev_name_node_del(struct netdev_name_node *name_node) 310 { 311 hlist_del_rcu(&name_node->hlist); 312 } 313 314 static struct netdev_name_node *netdev_name_node_lookup(struct net *net, 315 const char *name) 316 { 317 struct hlist_head *head = dev_name_hash(net, name); 318 struct netdev_name_node *name_node; 319 320 hlist_for_each_entry(name_node, head, hlist) 321 if (!strcmp(name_node->name, name)) 322 return name_node; 323 return NULL; 324 } 325 326 static struct netdev_name_node *netdev_name_node_lookup_rcu(struct net *net, 327 const char *name) 328 { 329 struct hlist_head *head = dev_name_hash(net, name); 330 struct netdev_name_node *name_node; 331 332 hlist_for_each_entry_rcu(name_node, head, hlist) 333 if (!strcmp(name_node->name, name)) 334 return name_node; 335 return NULL; 336 } 337 338 bool netdev_name_in_use(struct net *net, const char *name) 339 { 340 return netdev_name_node_lookup(net, name); 341 } 342 EXPORT_SYMBOL(netdev_name_in_use); 343 344 int netdev_name_node_alt_create(struct net_device *dev, const char *name) 345 { 346 struct netdev_name_node *name_node; 347 struct net *net = dev_net(dev); 348 349 name_node = netdev_name_node_lookup(net, name); 350 if (name_node) 351 return -EEXIST; 352 name_node = netdev_name_node_alloc(dev, name); 353 if (!name_node) 354 return -ENOMEM; 355 netdev_name_node_add(net, name_node); 356 /* The node that holds dev->name acts as a head of per-device list. */ 357 list_add_tail_rcu(&name_node->list, &dev->name_node->list); 358 359 return 0; 360 } 361 362 static void netdev_name_node_alt_free(struct rcu_head *head) 363 { 364 struct netdev_name_node *name_node = 365 container_of(head, struct netdev_name_node, rcu); 366 367 kfree(name_node->name); 368 netdev_name_node_free(name_node); 369 } 370 371 static void __netdev_name_node_alt_destroy(struct netdev_name_node *name_node) 372 { 373 netdev_name_node_del(name_node); 374 list_del(&name_node->list); 375 call_rcu(&name_node->rcu, netdev_name_node_alt_free); 376 } 377 378 int netdev_name_node_alt_destroy(struct net_device *dev, const char *name) 379 { 380 struct netdev_name_node *name_node; 381 struct net *net = dev_net(dev); 382 383 name_node = netdev_name_node_lookup(net, name); 384 if (!name_node) 385 return -ENOENT; 386 /* lookup might have found our primary name or a name belonging 387 * to another device. 388 */ 389 if (name_node == dev->name_node || name_node->dev != dev) 390 return -EINVAL; 391 392 __netdev_name_node_alt_destroy(name_node); 393 return 0; 394 } 395 396 static void netdev_name_node_alt_flush(struct net_device *dev) 397 { 398 struct netdev_name_node *name_node, *tmp; 399 400 list_for_each_entry_safe(name_node, tmp, &dev->name_node->list, list) { 401 list_del(&name_node->list); 402 netdev_name_node_alt_free(&name_node->rcu); 403 } 404 } 405 406 /* Device list insertion */ 407 static void list_netdevice(struct net_device *dev) 408 { 409 struct netdev_name_node *name_node; 410 struct net *net = dev_net(dev); 411 412 ASSERT_RTNL(); 413 414 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head); 415 netdev_name_node_add(net, dev->name_node); 416 hlist_add_head_rcu(&dev->index_hlist, 417 dev_index_hash(net, dev->ifindex)); 418 419 netdev_for_each_altname(dev, name_node) 420 netdev_name_node_add(net, name_node); 421 422 /* We reserved the ifindex, this can't fail */ 423 WARN_ON(xa_store(&net->dev_by_index, dev->ifindex, dev, GFP_KERNEL)); 424 425 dev_base_seq_inc(net); 426 } 427 428 /* Device list removal 429 * caller must respect a RCU grace period before freeing/reusing dev 430 */ 431 static void unlist_netdevice(struct net_device *dev) 432 { 433 struct netdev_name_node *name_node; 434 struct net *net = dev_net(dev); 435 436 ASSERT_RTNL(); 437 438 xa_erase(&net->dev_by_index, dev->ifindex); 439 440 netdev_for_each_altname(dev, name_node) 441 netdev_name_node_del(name_node); 442 443 /* Unlink dev from the device chain */ 444 list_del_rcu(&dev->dev_list); 445 netdev_name_node_del(dev->name_node); 446 hlist_del_rcu(&dev->index_hlist); 447 448 dev_base_seq_inc(dev_net(dev)); 449 } 450 451 /* 452 * Our notifier list 453 */ 454 455 static RAW_NOTIFIER_HEAD(netdev_chain); 456 457 /* 458 * Device drivers call our routines to queue packets here. We empty the 459 * queue in the local softnet handler. 460 */ 461 462 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data) = { 463 .process_queue_bh_lock = INIT_LOCAL_LOCK(process_queue_bh_lock), 464 }; 465 EXPORT_PER_CPU_SYMBOL(softnet_data); 466 467 /* Page_pool has a lockless array/stack to alloc/recycle pages. 468 * PP consumers must pay attention to run APIs in the appropriate context 469 * (e.g. NAPI context). 470 */ 471 DEFINE_PER_CPU(struct page_pool_bh, system_page_pool) = { 472 .bh_lock = INIT_LOCAL_LOCK(bh_lock), 473 }; 474 475 #ifdef CONFIG_LOCKDEP 476 /* 477 * register_netdevice() inits txq->_xmit_lock and sets lockdep class 478 * according to dev->type 479 */ 480 static const unsigned short netdev_lock_type[] = { 481 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25, 482 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET, 483 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM, 484 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP, 485 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD, 486 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25, 487 ARPHRD_CAN, ARPHRD_MCTP, 488 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP, 489 ARPHRD_RAWHDLC, ARPHRD_RAWIP, 490 ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD, 491 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI, 492 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE, 493 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET, 494 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL, 495 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM, 496 ARPHRD_IEEE80211_RADIOTAP, 497 ARPHRD_IEEE802154, ARPHRD_IEEE802154_MONITOR, 498 ARPHRD_PHONET, ARPHRD_PHONET_PIPE, 499 ARPHRD_CAIF, ARPHRD_IP6GRE, ARPHRD_NETLINK, ARPHRD_6LOWPAN, 500 ARPHRD_VSOCKMON, 501 ARPHRD_VOID, ARPHRD_NONE}; 502 503 static const char *const netdev_lock_name[] = { 504 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25", 505 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET", 506 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM", 507 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP", 508 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD", 509 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25", 510 "_xmit_CAN", "_xmit_MCTP", 511 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP", 512 "_xmit_RAWHDLC", "_xmit_RAWIP", 513 "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD", 514 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI", 515 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE", 516 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET", 517 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL", 518 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM", 519 "_xmit_IEEE80211_RADIOTAP", 520 "_xmit_IEEE802154", "_xmit_IEEE802154_MONITOR", 521 "_xmit_PHONET", "_xmit_PHONET_PIPE", 522 "_xmit_CAIF", "_xmit_IP6GRE", "_xmit_NETLINK", "_xmit_6LOWPAN", 523 "_xmit_VSOCKMON", 524 "_xmit_VOID", "_xmit_NONE"}; 525 526 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)]; 527 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)]; 528 529 static inline unsigned short netdev_lock_pos(unsigned short dev_type) 530 { 531 int i; 532 533 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++) 534 if (netdev_lock_type[i] == dev_type) 535 return i; 536 /* the last key is used by default */ 537 WARN_ONCE(1, "netdev_lock_pos() could not find dev_type=%u\n", dev_type); 538 return ARRAY_SIZE(netdev_lock_type) - 1; 539 } 540 541 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 542 unsigned short dev_type) 543 { 544 int i; 545 546 i = netdev_lock_pos(dev_type); 547 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i], 548 netdev_lock_name[i]); 549 } 550 551 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 552 { 553 int i; 554 555 i = netdev_lock_pos(dev->type); 556 lockdep_set_class_and_name(&dev->addr_list_lock, 557 &netdev_addr_lock_key[i], 558 netdev_lock_name[i]); 559 } 560 #else 561 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 562 unsigned short dev_type) 563 { 564 } 565 566 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 567 { 568 } 569 #endif 570 571 /******************************************************************************* 572 * 573 * Protocol management and registration routines 574 * 575 *******************************************************************************/ 576 577 578 /* 579 * Add a protocol ID to the list. Now that the input handler is 580 * smarter we can dispense with all the messy stuff that used to be 581 * here. 582 * 583 * BEWARE!!! Protocol handlers, mangling input packets, 584 * MUST BE last in hash buckets and checking protocol handlers 585 * MUST start from promiscuous ptype_all chain in net_bh. 586 * It is true now, do not change it. 587 * Explanation follows: if protocol handler, mangling packet, will 588 * be the first on list, it is not able to sense, that packet 589 * is cloned and should be copied-on-write, so that it will 590 * change it and subsequent readers will get broken packet. 591 * --ANK (980803) 592 */ 593 594 static inline struct list_head *ptype_head(const struct packet_type *pt) 595 { 596 if (pt->type == htons(ETH_P_ALL)) { 597 if (!pt->af_packet_net && !pt->dev) 598 return NULL; 599 600 return pt->dev ? &pt->dev->ptype_all : 601 &pt->af_packet_net->ptype_all; 602 } 603 604 if (pt->dev) 605 return &pt->dev->ptype_specific; 606 607 return pt->af_packet_net ? &pt->af_packet_net->ptype_specific : 608 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK]; 609 } 610 611 /** 612 * dev_add_pack - add packet handler 613 * @pt: packet type declaration 614 * 615 * Add a protocol handler to the networking stack. The passed &packet_type 616 * is linked into kernel lists and may not be freed until it has been 617 * removed from the kernel lists. 618 * 619 * This call does not sleep therefore it can not 620 * guarantee all CPU's that are in middle of receiving packets 621 * will see the new packet type (until the next received packet). 622 */ 623 624 void dev_add_pack(struct packet_type *pt) 625 { 626 struct list_head *head = ptype_head(pt); 627 628 if (WARN_ON_ONCE(!head)) 629 return; 630 631 spin_lock(&ptype_lock); 632 list_add_rcu(&pt->list, head); 633 spin_unlock(&ptype_lock); 634 } 635 EXPORT_SYMBOL(dev_add_pack); 636 637 /** 638 * __dev_remove_pack - remove packet handler 639 * @pt: packet type declaration 640 * 641 * Remove a protocol handler that was previously added to the kernel 642 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 643 * from the kernel lists and can be freed or reused once this function 644 * returns. 645 * 646 * The packet type might still be in use by receivers 647 * and must not be freed until after all the CPU's have gone 648 * through a quiescent state. 649 */ 650 void __dev_remove_pack(struct packet_type *pt) 651 { 652 struct list_head *head = ptype_head(pt); 653 struct packet_type *pt1; 654 655 if (!head) 656 return; 657 658 spin_lock(&ptype_lock); 659 660 list_for_each_entry(pt1, head, list) { 661 if (pt == pt1) { 662 list_del_rcu(&pt->list); 663 goto out; 664 } 665 } 666 667 pr_warn("dev_remove_pack: %p not found\n", pt); 668 out: 669 spin_unlock(&ptype_lock); 670 } 671 EXPORT_SYMBOL(__dev_remove_pack); 672 673 /** 674 * dev_remove_pack - remove packet handler 675 * @pt: packet type declaration 676 * 677 * Remove a protocol handler that was previously added to the kernel 678 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 679 * from the kernel lists and can be freed or reused once this function 680 * returns. 681 * 682 * This call sleeps to guarantee that no CPU is looking at the packet 683 * type after return. 684 */ 685 void dev_remove_pack(struct packet_type *pt) 686 { 687 __dev_remove_pack(pt); 688 689 synchronize_net(); 690 } 691 EXPORT_SYMBOL(dev_remove_pack); 692 693 694 /******************************************************************************* 695 * 696 * Device Interface Subroutines 697 * 698 *******************************************************************************/ 699 700 /** 701 * dev_get_iflink - get 'iflink' value of a interface 702 * @dev: targeted interface 703 * 704 * Indicates the ifindex the interface is linked to. 705 * Physical interfaces have the same 'ifindex' and 'iflink' values. 706 */ 707 708 int dev_get_iflink(const struct net_device *dev) 709 { 710 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink) 711 return dev->netdev_ops->ndo_get_iflink(dev); 712 713 return READ_ONCE(dev->ifindex); 714 } 715 EXPORT_SYMBOL(dev_get_iflink); 716 717 /** 718 * dev_fill_metadata_dst - Retrieve tunnel egress information. 719 * @dev: targeted interface 720 * @skb: The packet. 721 * 722 * For better visibility of tunnel traffic OVS needs to retrieve 723 * egress tunnel information for a packet. Following API allows 724 * user to get this info. 725 */ 726 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb) 727 { 728 struct ip_tunnel_info *info; 729 730 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst) 731 return -EINVAL; 732 733 info = skb_tunnel_info_unclone(skb); 734 if (!info) 735 return -ENOMEM; 736 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX))) 737 return -EINVAL; 738 739 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb); 740 } 741 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst); 742 743 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack) 744 { 745 int k = stack->num_paths++; 746 747 if (k >= NET_DEVICE_PATH_STACK_MAX) 748 return NULL; 749 750 return &stack->path[k]; 751 } 752 753 int dev_fill_forward_path(const struct net_device *dev, const u8 *daddr, 754 struct net_device_path_stack *stack) 755 { 756 const struct net_device *last_dev; 757 struct net_device_path_ctx ctx = { 758 .dev = dev, 759 }; 760 struct net_device_path *path; 761 int ret = 0; 762 763 memcpy(ctx.daddr, daddr, sizeof(ctx.daddr)); 764 stack->num_paths = 0; 765 while (ctx.dev && ctx.dev->netdev_ops->ndo_fill_forward_path) { 766 last_dev = ctx.dev; 767 path = dev_fwd_path(stack); 768 if (!path) 769 return -1; 770 771 memset(path, 0, sizeof(struct net_device_path)); 772 ret = ctx.dev->netdev_ops->ndo_fill_forward_path(&ctx, path); 773 if (ret < 0) 774 return -1; 775 776 if (WARN_ON_ONCE(last_dev == ctx.dev)) 777 return -1; 778 } 779 780 if (!ctx.dev) 781 return ret; 782 783 path = dev_fwd_path(stack); 784 if (!path) 785 return -1; 786 path->type = DEV_PATH_ETHERNET; 787 path->dev = ctx.dev; 788 789 return ret; 790 } 791 EXPORT_SYMBOL_GPL(dev_fill_forward_path); 792 793 /* must be called under rcu_read_lock(), as we dont take a reference */ 794 static struct napi_struct *napi_by_id(unsigned int napi_id) 795 { 796 unsigned int hash = napi_id % HASH_SIZE(napi_hash); 797 struct napi_struct *napi; 798 799 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node) 800 if (napi->napi_id == napi_id) 801 return napi; 802 803 return NULL; 804 } 805 806 /* must be called under rcu_read_lock(), as we dont take a reference */ 807 static struct napi_struct * 808 netdev_napi_by_id(struct net *net, unsigned int napi_id) 809 { 810 struct napi_struct *napi; 811 812 napi = napi_by_id(napi_id); 813 if (!napi) 814 return NULL; 815 816 if (WARN_ON_ONCE(!napi->dev)) 817 return NULL; 818 if (!net_eq(net, dev_net(napi->dev))) 819 return NULL; 820 821 return napi; 822 } 823 824 /** 825 * netdev_napi_by_id_lock() - find a device by NAPI ID and lock it 826 * @net: the applicable net namespace 827 * @napi_id: ID of a NAPI of a target device 828 * 829 * Find a NAPI instance with @napi_id. Lock its device. 830 * The device must be in %NETREG_REGISTERED state for lookup to succeed. 831 * netdev_unlock() must be called to release it. 832 * 833 * Return: pointer to NAPI, its device with lock held, NULL if not found. 834 */ 835 struct napi_struct * 836 netdev_napi_by_id_lock(struct net *net, unsigned int napi_id) 837 { 838 struct napi_struct *napi; 839 struct net_device *dev; 840 841 rcu_read_lock(); 842 napi = netdev_napi_by_id(net, napi_id); 843 if (!napi || READ_ONCE(napi->dev->reg_state) != NETREG_REGISTERED) { 844 rcu_read_unlock(); 845 return NULL; 846 } 847 848 dev = napi->dev; 849 dev_hold(dev); 850 rcu_read_unlock(); 851 852 dev = __netdev_put_lock(dev, net); 853 if (!dev) 854 return NULL; 855 856 rcu_read_lock(); 857 napi = netdev_napi_by_id(net, napi_id); 858 if (napi && napi->dev != dev) 859 napi = NULL; 860 rcu_read_unlock(); 861 862 if (!napi) 863 netdev_unlock(dev); 864 return napi; 865 } 866 867 /** 868 * __dev_get_by_name - find a device by its name 869 * @net: the applicable net namespace 870 * @name: name to find 871 * 872 * Find an interface by name. Must be called under RTNL semaphore. 873 * If the name is found a pointer to the device is returned. 874 * If the name is not found then %NULL is returned. The 875 * reference counters are not incremented so the caller must be 876 * careful with locks. 877 */ 878 879 struct net_device *__dev_get_by_name(struct net *net, const char *name) 880 { 881 struct netdev_name_node *node_name; 882 883 node_name = netdev_name_node_lookup(net, name); 884 return node_name ? node_name->dev : NULL; 885 } 886 EXPORT_SYMBOL(__dev_get_by_name); 887 888 /** 889 * dev_get_by_name_rcu - find a device by its name 890 * @net: the applicable net namespace 891 * @name: name to find 892 * 893 * Find an interface by name. 894 * If the name is found a pointer to the device is returned. 895 * If the name is not found then %NULL is returned. 896 * The reference counters are not incremented so the caller must be 897 * careful with locks. The caller must hold RCU lock. 898 */ 899 900 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name) 901 { 902 struct netdev_name_node *node_name; 903 904 node_name = netdev_name_node_lookup_rcu(net, name); 905 return node_name ? node_name->dev : NULL; 906 } 907 EXPORT_SYMBOL(dev_get_by_name_rcu); 908 909 /* Deprecated for new users, call netdev_get_by_name() instead */ 910 struct net_device *dev_get_by_name(struct net *net, const char *name) 911 { 912 struct net_device *dev; 913 914 rcu_read_lock(); 915 dev = dev_get_by_name_rcu(net, name); 916 dev_hold(dev); 917 rcu_read_unlock(); 918 return dev; 919 } 920 EXPORT_SYMBOL(dev_get_by_name); 921 922 /** 923 * netdev_get_by_name() - find a device by its name 924 * @net: the applicable net namespace 925 * @name: name to find 926 * @tracker: tracking object for the acquired reference 927 * @gfp: allocation flags for the tracker 928 * 929 * Find an interface by name. This can be called from any 930 * context and does its own locking. The returned handle has 931 * the usage count incremented and the caller must use netdev_put() to 932 * release it when it is no longer needed. %NULL is returned if no 933 * matching device is found. 934 */ 935 struct net_device *netdev_get_by_name(struct net *net, const char *name, 936 netdevice_tracker *tracker, gfp_t gfp) 937 { 938 struct net_device *dev; 939 940 dev = dev_get_by_name(net, name); 941 if (dev) 942 netdev_tracker_alloc(dev, tracker, gfp); 943 return dev; 944 } 945 EXPORT_SYMBOL(netdev_get_by_name); 946 947 /** 948 * __dev_get_by_index - find a device by its ifindex 949 * @net: the applicable net namespace 950 * @ifindex: index of device 951 * 952 * Search for an interface by index. Returns %NULL if the device 953 * is not found or a pointer to the device. The device has not 954 * had its reference counter increased so the caller must be careful 955 * about locking. The caller must hold the RTNL semaphore. 956 */ 957 958 struct net_device *__dev_get_by_index(struct net *net, int ifindex) 959 { 960 struct net_device *dev; 961 struct hlist_head *head = dev_index_hash(net, ifindex); 962 963 hlist_for_each_entry(dev, head, index_hlist) 964 if (dev->ifindex == ifindex) 965 return dev; 966 967 return NULL; 968 } 969 EXPORT_SYMBOL(__dev_get_by_index); 970 971 /** 972 * dev_get_by_index_rcu - find a device by its ifindex 973 * @net: the applicable net namespace 974 * @ifindex: index of device 975 * 976 * Search for an interface by index. Returns %NULL if the device 977 * is not found or a pointer to the device. The device has not 978 * had its reference counter increased so the caller must be careful 979 * about locking. The caller must hold RCU lock. 980 */ 981 982 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex) 983 { 984 struct net_device *dev; 985 struct hlist_head *head = dev_index_hash(net, ifindex); 986 987 hlist_for_each_entry_rcu(dev, head, index_hlist) 988 if (dev->ifindex == ifindex) 989 return dev; 990 991 return NULL; 992 } 993 EXPORT_SYMBOL(dev_get_by_index_rcu); 994 995 /* Deprecated for new users, call netdev_get_by_index() instead */ 996 struct net_device *dev_get_by_index(struct net *net, int ifindex) 997 { 998 struct net_device *dev; 999 1000 rcu_read_lock(); 1001 dev = dev_get_by_index_rcu(net, ifindex); 1002 dev_hold(dev); 1003 rcu_read_unlock(); 1004 return dev; 1005 } 1006 EXPORT_SYMBOL(dev_get_by_index); 1007 1008 /** 1009 * netdev_get_by_index() - find a device by its ifindex 1010 * @net: the applicable net namespace 1011 * @ifindex: index of device 1012 * @tracker: tracking object for the acquired reference 1013 * @gfp: allocation flags for the tracker 1014 * 1015 * Search for an interface by index. Returns NULL if the device 1016 * is not found or a pointer to the device. The device returned has 1017 * had a reference added and the pointer is safe until the user calls 1018 * netdev_put() to indicate they have finished with it. 1019 */ 1020 struct net_device *netdev_get_by_index(struct net *net, int ifindex, 1021 netdevice_tracker *tracker, gfp_t gfp) 1022 { 1023 struct net_device *dev; 1024 1025 dev = dev_get_by_index(net, ifindex); 1026 if (dev) 1027 netdev_tracker_alloc(dev, tracker, gfp); 1028 return dev; 1029 } 1030 EXPORT_SYMBOL(netdev_get_by_index); 1031 1032 /** 1033 * dev_get_by_napi_id - find a device by napi_id 1034 * @napi_id: ID of the NAPI struct 1035 * 1036 * Search for an interface by NAPI ID. Returns %NULL if the device 1037 * is not found or a pointer to the device. The device has not had 1038 * its reference counter increased so the caller must be careful 1039 * about locking. The caller must hold RCU lock. 1040 */ 1041 struct net_device *dev_get_by_napi_id(unsigned int napi_id) 1042 { 1043 struct napi_struct *napi; 1044 1045 WARN_ON_ONCE(!rcu_read_lock_held()); 1046 1047 if (!napi_id_valid(napi_id)) 1048 return NULL; 1049 1050 napi = napi_by_id(napi_id); 1051 1052 return napi ? napi->dev : NULL; 1053 } 1054 1055 /* Release the held reference on the net_device, and if the net_device 1056 * is still registered try to lock the instance lock. If device is being 1057 * unregistered NULL will be returned (but the reference has been released, 1058 * either way!) 1059 * 1060 * This helper is intended for locking net_device after it has been looked up 1061 * using a lockless lookup helper. Lock prevents the instance from going away. 1062 */ 1063 struct net_device * 1064 netdev_put_lock(struct net_device *dev, struct net *net, 1065 netdevice_tracker *tracker) 1066 { 1067 netdev_lock(dev); 1068 if (dev->reg_state > NETREG_REGISTERED || 1069 dev->moving_ns || !net_eq(dev_net(dev), net)) { 1070 netdev_unlock(dev); 1071 netdev_put(dev, tracker); 1072 return NULL; 1073 } 1074 netdev_put(dev, tracker); 1075 return dev; 1076 } 1077 1078 static struct net_device * 1079 __netdev_put_lock_ops_compat(struct net_device *dev, struct net *net) 1080 { 1081 netdev_lock_ops_compat(dev); 1082 if (dev->reg_state > NETREG_REGISTERED || 1083 dev->moving_ns || !net_eq(dev_net(dev), net)) { 1084 netdev_unlock_ops_compat(dev); 1085 dev_put(dev); 1086 return NULL; 1087 } 1088 dev_put(dev); 1089 return dev; 1090 } 1091 1092 /** 1093 * netdev_get_by_index_lock() - find a device by its ifindex 1094 * @net: the applicable net namespace 1095 * @ifindex: index of device 1096 * 1097 * Search for an interface by index. If a valid device 1098 * with @ifindex is found it will be returned with netdev->lock held. 1099 * netdev_unlock() must be called to release it. 1100 * 1101 * Return: pointer to a device with lock held, NULL if not found. 1102 */ 1103 struct net_device *netdev_get_by_index_lock(struct net *net, int ifindex) 1104 { 1105 struct net_device *dev; 1106 1107 dev = dev_get_by_index(net, ifindex); 1108 if (!dev) 1109 return NULL; 1110 1111 return __netdev_put_lock(dev, net); 1112 } 1113 1114 struct net_device * 1115 netdev_get_by_index_lock_ops_compat(struct net *net, int ifindex) 1116 { 1117 struct net_device *dev; 1118 1119 dev = dev_get_by_index(net, ifindex); 1120 if (!dev) 1121 return NULL; 1122 1123 return __netdev_put_lock_ops_compat(dev, net); 1124 } 1125 1126 struct net_device * 1127 netdev_xa_find_lock(struct net *net, struct net_device *dev, 1128 unsigned long *index) 1129 { 1130 if (dev) 1131 netdev_unlock(dev); 1132 1133 do { 1134 rcu_read_lock(); 1135 dev = xa_find(&net->dev_by_index, index, ULONG_MAX, XA_PRESENT); 1136 if (!dev) { 1137 rcu_read_unlock(); 1138 return NULL; 1139 } 1140 dev_hold(dev); 1141 rcu_read_unlock(); 1142 1143 dev = __netdev_put_lock(dev, net); 1144 if (dev) 1145 return dev; 1146 1147 (*index)++; 1148 } while (true); 1149 } 1150 1151 struct net_device * 1152 netdev_xa_find_lock_ops_compat(struct net *net, struct net_device *dev, 1153 unsigned long *index) 1154 { 1155 if (dev) 1156 netdev_unlock_ops_compat(dev); 1157 1158 do { 1159 rcu_read_lock(); 1160 dev = xa_find(&net->dev_by_index, index, ULONG_MAX, XA_PRESENT); 1161 if (!dev) { 1162 rcu_read_unlock(); 1163 return NULL; 1164 } 1165 dev_hold(dev); 1166 rcu_read_unlock(); 1167 1168 dev = __netdev_put_lock_ops_compat(dev, net); 1169 if (dev) 1170 return dev; 1171 1172 (*index)++; 1173 } while (true); 1174 } 1175 1176 static DEFINE_SEQLOCK(netdev_rename_lock); 1177 1178 void netdev_copy_name(struct net_device *dev, char *name) 1179 { 1180 unsigned int seq; 1181 1182 do { 1183 seq = read_seqbegin(&netdev_rename_lock); 1184 strscpy(name, dev->name, IFNAMSIZ); 1185 } while (read_seqretry(&netdev_rename_lock, seq)); 1186 } 1187 1188 /** 1189 * netdev_get_name - get a netdevice name, knowing its ifindex. 1190 * @net: network namespace 1191 * @name: a pointer to the buffer where the name will be stored. 1192 * @ifindex: the ifindex of the interface to get the name from. 1193 */ 1194 int netdev_get_name(struct net *net, char *name, int ifindex) 1195 { 1196 struct net_device *dev; 1197 int ret; 1198 1199 rcu_read_lock(); 1200 1201 dev = dev_get_by_index_rcu(net, ifindex); 1202 if (!dev) { 1203 ret = -ENODEV; 1204 goto out; 1205 } 1206 1207 netdev_copy_name(dev, name); 1208 1209 ret = 0; 1210 out: 1211 rcu_read_unlock(); 1212 return ret; 1213 } 1214 1215 static bool dev_addr_cmp(struct net_device *dev, unsigned short type, 1216 const char *ha) 1217 { 1218 return dev->type == type && !memcmp(dev->dev_addr, ha, dev->addr_len); 1219 } 1220 1221 /** 1222 * dev_getbyhwaddr_rcu - find a device by its hardware address 1223 * @net: the applicable net namespace 1224 * @type: media type of device 1225 * @ha: hardware address 1226 * 1227 * Search for an interface by MAC address. Returns NULL if the device 1228 * is not found or a pointer to the device. 1229 * The caller must hold RCU. 1230 * The returned device has not had its ref count increased 1231 * and the caller must therefore be careful about locking 1232 * 1233 */ 1234 1235 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type, 1236 const char *ha) 1237 { 1238 struct net_device *dev; 1239 1240 for_each_netdev_rcu(net, dev) 1241 if (dev_addr_cmp(dev, type, ha)) 1242 return dev; 1243 1244 return NULL; 1245 } 1246 EXPORT_SYMBOL(dev_getbyhwaddr_rcu); 1247 1248 /** 1249 * dev_getbyhwaddr() - find a device by its hardware address 1250 * @net: the applicable net namespace 1251 * @type: media type of device 1252 * @ha: hardware address 1253 * 1254 * Similar to dev_getbyhwaddr_rcu(), but the owner needs to hold 1255 * rtnl_lock. 1256 * 1257 * Context: rtnl_lock() must be held. 1258 * Return: pointer to the net_device, or NULL if not found 1259 */ 1260 struct net_device *dev_getbyhwaddr(struct net *net, unsigned short type, 1261 const char *ha) 1262 { 1263 struct net_device *dev; 1264 1265 ASSERT_RTNL(); 1266 for_each_netdev(net, dev) 1267 if (dev_addr_cmp(dev, type, ha)) 1268 return dev; 1269 1270 return NULL; 1271 } 1272 EXPORT_SYMBOL(dev_getbyhwaddr); 1273 1274 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type) 1275 { 1276 struct net_device *dev, *ret = NULL; 1277 1278 rcu_read_lock(); 1279 for_each_netdev_rcu(net, dev) 1280 if (dev->type == type) { 1281 dev_hold(dev); 1282 ret = dev; 1283 break; 1284 } 1285 rcu_read_unlock(); 1286 return ret; 1287 } 1288 EXPORT_SYMBOL(dev_getfirstbyhwtype); 1289 1290 /** 1291 * netdev_get_by_flags_rcu - find any device with given flags 1292 * @net: the applicable net namespace 1293 * @tracker: tracking object for the acquired reference 1294 * @if_flags: IFF_* values 1295 * @mask: bitmask of bits in if_flags to check 1296 * 1297 * Search for any interface with the given flags. 1298 * 1299 * Context: rcu_read_lock() must be held. 1300 * Returns: NULL if a device is not found or a pointer to the device. 1301 */ 1302 struct net_device *netdev_get_by_flags_rcu(struct net *net, netdevice_tracker *tracker, 1303 unsigned short if_flags, unsigned short mask) 1304 { 1305 struct net_device *dev; 1306 1307 for_each_netdev_rcu(net, dev) { 1308 if (((READ_ONCE(dev->flags) ^ if_flags) & mask) == 0) { 1309 netdev_hold(dev, tracker, GFP_ATOMIC); 1310 return dev; 1311 } 1312 } 1313 1314 return NULL; 1315 } 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 EXPORT_SYMBOL(netif_open); 1735 1736 static void __dev_close_many(struct list_head *head) 1737 { 1738 struct net_device *dev; 1739 1740 ASSERT_RTNL(); 1741 might_sleep(); 1742 1743 list_for_each_entry(dev, head, close_list) { 1744 /* Temporarily disable netpoll until the interface is down */ 1745 netpoll_poll_disable(dev); 1746 1747 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev); 1748 1749 clear_bit(__LINK_STATE_START, &dev->state); 1750 1751 /* Synchronize to scheduled poll. We cannot touch poll list, it 1752 * can be even on different cpu. So just clear netif_running(). 1753 * 1754 * dev->stop() will invoke napi_disable() on all of it's 1755 * napi_struct instances on this device. 1756 */ 1757 smp_mb__after_atomic(); /* Commit netif_running(). */ 1758 } 1759 1760 dev_deactivate_many(head, true); 1761 1762 list_for_each_entry(dev, head, close_list) { 1763 const struct net_device_ops *ops = dev->netdev_ops; 1764 1765 /* 1766 * Call the device specific close. This cannot fail. 1767 * Only if device is UP 1768 * 1769 * We allow it to be called even after a DETACH hot-plug 1770 * event. 1771 */ 1772 1773 netdev_ops_assert_locked(dev); 1774 1775 if (ops->ndo_stop) 1776 ops->ndo_stop(dev); 1777 1778 netif_set_up(dev, false); 1779 netpoll_poll_enable(dev); 1780 } 1781 } 1782 1783 static void __dev_close(struct net_device *dev) 1784 { 1785 LIST_HEAD(single); 1786 1787 list_add(&dev->close_list, &single); 1788 __dev_close_many(&single); 1789 list_del(&single); 1790 } 1791 1792 void netif_close_many(struct list_head *head, bool unlink) 1793 { 1794 struct net_device *dev, *tmp; 1795 1796 /* Remove the devices that don't need to be closed */ 1797 list_for_each_entry_safe(dev, tmp, head, close_list) 1798 if (!(dev->flags & IFF_UP)) 1799 list_del_init(&dev->close_list); 1800 1801 __dev_close_many(head); 1802 1803 list_for_each_entry_safe(dev, tmp, head, close_list) { 1804 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1805 call_netdevice_notifiers(NETDEV_DOWN, dev); 1806 if (unlink) 1807 list_del_init(&dev->close_list); 1808 } 1809 } 1810 EXPORT_SYMBOL_NS_GPL(netif_close_many, "NETDEV_INTERNAL"); 1811 1812 void netif_close(struct net_device *dev) 1813 { 1814 if (dev->flags & IFF_UP) { 1815 LIST_HEAD(single); 1816 1817 list_add(&dev->close_list, &single); 1818 netif_close_many(&single, true); 1819 list_del(&single); 1820 } 1821 } 1822 EXPORT_SYMBOL(netif_close); 1823 1824 void netif_disable_lro(struct net_device *dev) 1825 { 1826 struct net_device *lower_dev; 1827 struct list_head *iter; 1828 1829 dev->wanted_features &= ~NETIF_F_LRO; 1830 netdev_update_features(dev); 1831 1832 if (unlikely(dev->features & NETIF_F_LRO)) 1833 netdev_WARN(dev, "failed to disable LRO!\n"); 1834 1835 netdev_for_each_lower_dev(dev, lower_dev, iter) { 1836 netdev_lock_ops(lower_dev); 1837 netif_disable_lro(lower_dev); 1838 netdev_unlock_ops(lower_dev); 1839 } 1840 } 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 bool skb_gso_has_extension_hdr(const struct sk_buff *skb) 3773 { 3774 if (!skb->encapsulation) 3775 return ((skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6 || 3776 (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4 && 3777 vlan_get_protocol(skb) == htons(ETH_P_IPV6))) && 3778 skb_transport_header_was_set(skb) && 3779 skb_network_header_len(skb) != sizeof(struct ipv6hdr)); 3780 else 3781 return (!skb_inner_network_header_was_set(skb) || 3782 ((skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6 || 3783 (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4 && 3784 inner_ip_hdr(skb)->version == 6)) && 3785 skb_inner_network_header_len(skb) != sizeof(struct ipv6hdr))); 3786 } 3787 3788 static netdev_features_t gso_features_check(const struct sk_buff *skb, 3789 struct net_device *dev, 3790 netdev_features_t features) 3791 { 3792 u16 gso_segs = skb_shinfo(skb)->gso_segs; 3793 3794 if (gso_segs > READ_ONCE(dev->gso_max_segs)) 3795 return features & ~NETIF_F_GSO_MASK; 3796 3797 if (unlikely(skb->len >= netif_get_gso_max_size(dev, skb))) 3798 return features & ~NETIF_F_GSO_MASK; 3799 3800 if (!skb_shinfo(skb)->gso_type) { 3801 skb_warn_bad_offload(skb); 3802 return features & ~NETIF_F_GSO_MASK; 3803 } 3804 3805 /* Support for GSO partial features requires software 3806 * intervention before we can actually process the packets 3807 * so we need to strip support for any partial features now 3808 * and we can pull them back in after we have partially 3809 * segmented the frame. 3810 */ 3811 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)) 3812 features &= ~dev->gso_partial_features; 3813 3814 /* Make sure to clear the IPv4 ID mangling feature if the IPv4 header 3815 * has the potential to be fragmented so that TSO does not generate 3816 * segments with the same ID. For encapsulated packets, the ID mangling 3817 * feature is guaranteed not to use the same ID for the outer IPv4 3818 * headers of the generated segments if the headers have the potential 3819 * to be fragmented, so there is no need to clear the IPv4 ID mangling 3820 * feature (see the section about NETIF_F_TSO_MANGLEID in 3821 * segmentation-offloads.rst). 3822 */ 3823 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) { 3824 const struct iphdr *iph; 3825 struct iphdr _iph; 3826 int nhoff = skb->encapsulation ? 3827 skb_inner_network_offset(skb) : 3828 skb_network_offset(skb); 3829 3830 iph = skb_header_pointer(skb, nhoff, sizeof(_iph), &_iph); 3831 3832 if (!iph || !(iph->frag_off & htons(IP_DF))) 3833 features &= ~dev->mangleid_features; 3834 } 3835 3836 /* NETIF_F_IPV6_CSUM does not support IPv6 extension headers, 3837 * so neither does TSO that depends on it. 3838 */ 3839 if (features & NETIF_F_IPV6_CSUM && 3840 skb_gso_has_extension_hdr(skb)) 3841 features &= ~(NETIF_F_IPV6_CSUM | NETIF_F_TSO6 | NETIF_F_GSO_UDP_L4); 3842 3843 return features; 3844 } 3845 3846 netdev_features_t netif_skb_features(struct sk_buff *skb) 3847 { 3848 struct net_device *dev = skb->dev; 3849 netdev_features_t features = dev->features; 3850 3851 if (skb_is_gso(skb)) 3852 features = gso_features_check(skb, dev, features); 3853 3854 /* If encapsulation offload request, verify we are testing 3855 * hardware encapsulation features instead of standard 3856 * features for the netdev 3857 */ 3858 if (skb->encapsulation) 3859 features &= dev->hw_enc_features; 3860 3861 if (skb_vlan_tagged(skb)) 3862 features = netdev_intersect_features(features, 3863 dev->vlan_features | 3864 NETIF_F_HW_VLAN_CTAG_TX | 3865 NETIF_F_HW_VLAN_STAG_TX); 3866 3867 if (dev->netdev_ops->ndo_features_check) 3868 features &= dev->netdev_ops->ndo_features_check(skb, dev, 3869 features); 3870 else 3871 features &= dflt_features_check(skb, dev, features); 3872 3873 return harmonize_features(skb, features); 3874 } 3875 EXPORT_SYMBOL(netif_skb_features); 3876 3877 static int xmit_one(struct sk_buff *skb, struct net_device *dev, 3878 struct netdev_queue *txq, bool more) 3879 { 3880 unsigned int len; 3881 int rc; 3882 3883 if (dev_nit_active_rcu(dev)) 3884 dev_queue_xmit_nit(skb, dev); 3885 3886 len = skb->len; 3887 trace_net_dev_start_xmit(skb, dev); 3888 rc = netdev_start_xmit(skb, dev, txq, more); 3889 trace_net_dev_xmit(skb, rc, dev, len); 3890 3891 return rc; 3892 } 3893 3894 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev, 3895 struct netdev_queue *txq, int *ret) 3896 { 3897 struct sk_buff *skb = first; 3898 int rc = NETDEV_TX_OK; 3899 3900 while (skb) { 3901 struct sk_buff *next = skb->next; 3902 3903 skb_mark_not_on_list(skb); 3904 rc = xmit_one(skb, dev, txq, next != NULL); 3905 if (unlikely(!dev_xmit_complete(rc))) { 3906 skb->next = next; 3907 goto out; 3908 } 3909 3910 skb = next; 3911 if (netif_tx_queue_stopped(txq) && skb) { 3912 rc = NETDEV_TX_BUSY; 3913 break; 3914 } 3915 } 3916 3917 out: 3918 *ret = rc; 3919 return skb; 3920 } 3921 3922 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb, 3923 netdev_features_t features) 3924 { 3925 if (skb_vlan_tag_present(skb) && 3926 !vlan_hw_offload_capable(features, skb->vlan_proto)) 3927 skb = __vlan_hwaccel_push_inside(skb); 3928 return skb; 3929 } 3930 3931 int skb_csum_hwoffload_help(struct sk_buff *skb, 3932 const netdev_features_t features) 3933 { 3934 if (unlikely(skb_csum_is_sctp(skb))) 3935 return !!(features & NETIF_F_SCTP_CRC) ? 0 : 3936 skb_crc32c_csum_help(skb); 3937 3938 if (features & NETIF_F_HW_CSUM) 3939 return 0; 3940 3941 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) { 3942 if (vlan_get_protocol(skb) == htons(ETH_P_IPV6) && 3943 skb_network_header_len(skb) != sizeof(struct ipv6hdr)) 3944 goto sw_checksum; 3945 3946 switch (skb->csum_offset) { 3947 case offsetof(struct tcphdr, check): 3948 case offsetof(struct udphdr, check): 3949 return 0; 3950 } 3951 } 3952 3953 sw_checksum: 3954 return skb_checksum_help(skb); 3955 } 3956 EXPORT_SYMBOL(skb_csum_hwoffload_help); 3957 3958 /* Checks if this SKB belongs to an HW offloaded socket 3959 * and whether any SW fallbacks are required based on dev. 3960 * Check decrypted mark in case skb_orphan() cleared socket. 3961 */ 3962 static struct sk_buff *sk_validate_xmit_skb(struct sk_buff *skb, 3963 struct net_device *dev) 3964 { 3965 #ifdef CONFIG_SOCK_VALIDATE_XMIT 3966 struct sk_buff *(*sk_validate)(struct sock *sk, struct net_device *dev, 3967 struct sk_buff *skb); 3968 struct sock *sk = skb->sk; 3969 3970 sk_validate = NULL; 3971 if (sk) { 3972 if (sk_fullsock(sk)) 3973 sk_validate = sk->sk_validate_xmit_skb; 3974 else if (sk_is_inet(sk) && sk->sk_state == TCP_TIME_WAIT) 3975 sk_validate = inet_twsk(sk)->tw_validate_xmit_skb; 3976 } 3977 3978 if (sk_validate) { 3979 skb = sk_validate(sk, dev, skb); 3980 } else if (unlikely(skb_is_decrypted(skb))) { 3981 pr_warn_ratelimited("unencrypted skb with no associated socket - dropping\n"); 3982 kfree_skb(skb); 3983 skb = NULL; 3984 } 3985 #endif 3986 3987 return skb; 3988 } 3989 3990 static struct sk_buff *validate_xmit_unreadable_skb(struct sk_buff *skb, 3991 struct net_device *dev) 3992 { 3993 struct skb_shared_info *shinfo; 3994 struct net_iov *niov; 3995 3996 if (likely(skb_frags_readable(skb))) 3997 goto out; 3998 3999 if (!dev->netmem_tx) 4000 goto out_free; 4001 4002 shinfo = skb_shinfo(skb); 4003 4004 if (shinfo->nr_frags > 0) { 4005 niov = netmem_to_net_iov(skb_frag_netmem(&shinfo->frags[0])); 4006 if (net_is_devmem_iov(niov) && 4007 READ_ONCE(net_devmem_iov_binding(niov)->dev) != dev) 4008 goto out_free; 4009 } 4010 4011 out: 4012 return skb; 4013 4014 out_free: 4015 kfree_skb(skb); 4016 return NULL; 4017 } 4018 4019 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again) 4020 { 4021 netdev_features_t features; 4022 4023 skb = validate_xmit_unreadable_skb(skb, dev); 4024 if (unlikely(!skb)) 4025 goto out_null; 4026 4027 features = netif_skb_features(skb); 4028 skb = validate_xmit_vlan(skb, features); 4029 if (unlikely(!skb)) 4030 goto out_null; 4031 4032 skb = sk_validate_xmit_skb(skb, dev); 4033 if (unlikely(!skb)) 4034 goto out_null; 4035 4036 if (netif_needs_gso(skb, features)) { 4037 struct sk_buff *segs; 4038 4039 segs = skb_gso_segment(skb, features); 4040 if (IS_ERR(segs)) { 4041 goto out_kfree_skb; 4042 } else if (segs) { 4043 consume_skb(skb); 4044 skb = segs; 4045 } 4046 } else { 4047 if (skb_needs_linearize(skb, features) && 4048 __skb_linearize(skb)) 4049 goto out_kfree_skb; 4050 4051 /* If packet is not checksummed and device does not 4052 * support checksumming for this protocol, complete 4053 * checksumming here. 4054 */ 4055 if (skb->ip_summed == CHECKSUM_PARTIAL) { 4056 if (skb->encapsulation) 4057 skb_set_inner_transport_header(skb, 4058 skb_checksum_start_offset(skb)); 4059 else 4060 skb_set_transport_header(skb, 4061 skb_checksum_start_offset(skb)); 4062 if (skb_csum_hwoffload_help(skb, features)) 4063 goto out_kfree_skb; 4064 } 4065 } 4066 4067 skb = validate_xmit_xfrm(skb, features, again); 4068 4069 return skb; 4070 4071 out_kfree_skb: 4072 kfree_skb(skb); 4073 out_null: 4074 dev_core_stats_tx_dropped_inc(dev); 4075 return NULL; 4076 } 4077 4078 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again) 4079 { 4080 struct sk_buff *next, *head = NULL, *tail; 4081 4082 for (; skb != NULL; skb = next) { 4083 next = skb->next; 4084 skb_mark_not_on_list(skb); 4085 4086 /* in case skb won't be segmented, point to itself */ 4087 skb->prev = skb; 4088 4089 skb = validate_xmit_skb(skb, dev, again); 4090 if (!skb) 4091 continue; 4092 4093 if (!head) 4094 head = skb; 4095 else 4096 tail->next = skb; 4097 /* If skb was segmented, skb->prev points to 4098 * the last segment. If not, it still contains skb. 4099 */ 4100 tail = skb->prev; 4101 } 4102 return head; 4103 } 4104 EXPORT_SYMBOL_GPL(validate_xmit_skb_list); 4105 4106 static enum skb_drop_reason qdisc_pkt_len_segs_init(struct sk_buff *skb) 4107 { 4108 struct skb_shared_info *shinfo = skb_shinfo(skb); 4109 unsigned int hdr_len, tlen; 4110 u16 gso_segs; 4111 4112 qdisc_skb_cb(skb)->pkt_len = skb->len; 4113 if (!shinfo->gso_size) { 4114 qdisc_skb_cb(skb)->pkt_segs = 1; 4115 return SKB_NOT_DROPPED_YET; 4116 } 4117 4118 qdisc_skb_cb(skb)->pkt_segs = gso_segs = shinfo->gso_segs; 4119 4120 /* To get more precise estimation of bytes sent on wire, 4121 * we add to pkt_len the headers size of all segments 4122 */ 4123 4124 /* mac layer + network layer */ 4125 if (!skb->encapsulation) { 4126 if (unlikely(!skb_transport_header_was_set(skb))) 4127 return SKB_NOT_DROPPED_YET; 4128 hdr_len = skb_transport_offset(skb); 4129 } else { 4130 hdr_len = skb_inner_transport_offset(skb); 4131 } 4132 /* + transport layer */ 4133 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) { 4134 const struct tcphdr *th; 4135 4136 if (!pskb_may_pull(skb, hdr_len + sizeof(struct tcphdr))) 4137 return SKB_DROP_REASON_SKB_BAD_GSO; 4138 4139 th = (const struct tcphdr *)(skb->data + hdr_len); 4140 tlen = __tcp_hdrlen(th); 4141 if (tlen < sizeof(*th)) 4142 return SKB_DROP_REASON_SKB_BAD_GSO; 4143 hdr_len += tlen; 4144 if (!pskb_may_pull(skb, hdr_len)) 4145 return SKB_DROP_REASON_SKB_BAD_GSO; 4146 } else if (shinfo->gso_type & SKB_GSO_UDP_L4) { 4147 if (!pskb_may_pull(skb, hdr_len + sizeof(struct udphdr))) 4148 return SKB_DROP_REASON_SKB_BAD_GSO; 4149 hdr_len += sizeof(struct udphdr); 4150 } 4151 4152 /* prior pskb_may_pull() might have changed skb->head. */ 4153 shinfo = skb_shinfo(skb); 4154 if (unlikely(shinfo->gso_type & SKB_GSO_DODGY)) { 4155 int payload = skb->len - hdr_len; 4156 4157 /* Malicious packet. */ 4158 if (payload <= 0) 4159 return SKB_DROP_REASON_SKB_BAD_GSO; 4160 gso_segs = DIV_ROUND_UP(payload, shinfo->gso_size); 4161 shinfo->gso_segs = gso_segs; 4162 qdisc_skb_cb(skb)->pkt_segs = gso_segs; 4163 } 4164 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len; 4165 return SKB_NOT_DROPPED_YET; 4166 } 4167 4168 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q, 4169 struct sk_buff **to_free, 4170 struct netdev_queue *txq) 4171 { 4172 int rc; 4173 4174 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK; 4175 if (rc == NET_XMIT_SUCCESS) 4176 trace_qdisc_enqueue(q, txq, skb); 4177 return rc; 4178 } 4179 4180 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q, 4181 struct net_device *dev, 4182 struct netdev_queue *txq) 4183 { 4184 struct sk_buff *next, *to_free = NULL, *to_free2 = NULL; 4185 spinlock_t *root_lock = qdisc_lock(q); 4186 struct llist_node *ll_list, *first_n; 4187 unsigned long defer_count = 0; 4188 int rc; 4189 4190 qdisc_calculate_pkt_len(skb, q); 4191 4192 tcf_set_qdisc_drop_reason(skb, QDISC_DROP_GENERIC); 4193 4194 if (q->flags & TCQ_F_NOLOCK) { 4195 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) && 4196 qdisc_run_begin(q)) { 4197 /* Retest nolock_qdisc_is_empty() within the protection 4198 * of q->seqlock to protect from racing with requeuing. 4199 */ 4200 if (unlikely(!nolock_qdisc_is_empty(q))) { 4201 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4202 __qdisc_run(q); 4203 to_free2 = qdisc_run_end(q); 4204 4205 goto free_skbs; 4206 } 4207 4208 qdisc_bstats_cpu_update(q, skb); 4209 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) && 4210 !nolock_qdisc_is_empty(q)) 4211 __qdisc_run(q); 4212 4213 to_free2 = qdisc_run_end(q); 4214 rc = NET_XMIT_SUCCESS; 4215 goto free_skbs; 4216 } 4217 4218 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4219 to_free2 = qdisc_run(q); 4220 goto free_skbs; 4221 } 4222 4223 /* Open code llist_add(&skb->ll_node, &q->defer_list) + queue limit. 4224 * In the try_cmpxchg() loop, we want to increment q->defer_count 4225 * at most once to limit the number of skbs in defer_list. 4226 * We perform the defer_count increment only if the list is not empty, 4227 * because some arches have slow atomic_long_inc_return(). 4228 */ 4229 first_n = READ_ONCE(q->defer_list.first); 4230 do { 4231 if (first_n && !defer_count) { 4232 defer_count = atomic_long_inc_return(&q->defer_count); 4233 if (unlikely(defer_count > READ_ONCE(net_hotdata.qdisc_max_burst))) { 4234 kfree_skb_reason(skb, SKB_DROP_REASON_QDISC_BURST_DROP); 4235 return NET_XMIT_DROP; 4236 } 4237 } 4238 skb->ll_node.next = first_n; 4239 } while (!try_cmpxchg(&q->defer_list.first, &first_n, &skb->ll_node)); 4240 4241 /* If defer_list was not empty, we know the cpu which queued 4242 * the first skb will process the whole list for us. 4243 */ 4244 if (first_n) 4245 return NET_XMIT_SUCCESS; 4246 4247 spin_lock(root_lock); 4248 4249 ll_list = llist_del_all(&q->defer_list); 4250 /* There is a small race because we clear defer_count not atomically 4251 * with the prior llist_del_all(). This means defer_list could grow 4252 * over qdisc_max_burst. 4253 */ 4254 atomic_long_set(&q->defer_count, 0); 4255 4256 ll_list = llist_reverse_order(ll_list); 4257 4258 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) { 4259 llist_for_each_entry_safe(skb, next, ll_list, ll_node) 4260 __qdisc_drop(skb, &to_free); 4261 rc = NET_XMIT_DROP; 4262 goto unlock; 4263 } 4264 if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && 4265 !llist_next(ll_list) && qdisc_run_begin(q)) { 4266 /* 4267 * This is a work-conserving queue; there are no old skbs 4268 * waiting to be sent out; and the qdisc is not running - 4269 * xmit the skb directly. 4270 */ 4271 4272 DEBUG_NET_WARN_ON_ONCE(skb != llist_entry(ll_list, 4273 struct sk_buff, 4274 ll_node)); 4275 qdisc_bstats_update(q, skb); 4276 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) 4277 __qdisc_run(q); 4278 to_free2 = qdisc_run_end(q); 4279 rc = NET_XMIT_SUCCESS; 4280 } else { 4281 int count = 0; 4282 4283 llist_for_each_entry_safe(skb, next, ll_list, ll_node) { 4284 if (next) { 4285 prefetch(next); 4286 prefetch(&next->priority); 4287 skb_mark_not_on_list(skb); 4288 } 4289 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4290 count++; 4291 } 4292 to_free2 = qdisc_run(q); 4293 if (count != 1) 4294 rc = NET_XMIT_SUCCESS; 4295 } 4296 unlock: 4297 spin_unlock(root_lock); 4298 4299 free_skbs: 4300 tcf_kfree_skb_list(to_free, q, txq, dev); 4301 tcf_kfree_skb_list(to_free2, q, txq, dev); 4302 return rc; 4303 } 4304 4305 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO) 4306 static void skb_update_prio(struct sk_buff *skb) 4307 { 4308 const struct netprio_map *map; 4309 const struct sock *sk; 4310 unsigned int prioidx; 4311 4312 if (skb->priority) 4313 return; 4314 map = rcu_dereference_bh(skb->dev->priomap); 4315 if (!map) 4316 return; 4317 sk = skb_to_full_sk(skb); 4318 if (!sk) 4319 return; 4320 4321 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data); 4322 4323 if (prioidx < map->priomap_len) 4324 skb->priority = map->priomap[prioidx]; 4325 } 4326 #else 4327 #define skb_update_prio(skb) 4328 #endif 4329 4330 /** 4331 * dev_loopback_xmit - loop back @skb 4332 * @net: network namespace this loopback is happening in 4333 * @sk: sk needed to be a netfilter okfn 4334 * @skb: buffer to transmit 4335 */ 4336 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb) 4337 { 4338 skb_reset_mac_header(skb); 4339 __skb_pull(skb, skb_network_offset(skb)); 4340 skb->pkt_type = PACKET_LOOPBACK; 4341 if (skb->ip_summed == CHECKSUM_NONE) 4342 skb->ip_summed = CHECKSUM_UNNECESSARY; 4343 DEBUG_NET_WARN_ON_ONCE(!skb_dst(skb)); 4344 skb_dst_force(skb); 4345 netif_rx(skb); 4346 return 0; 4347 } 4348 EXPORT_SYMBOL(dev_loopback_xmit); 4349 4350 #ifdef CONFIG_NET_EGRESS 4351 static struct netdev_queue * 4352 netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb) 4353 { 4354 int qm = skb_get_queue_mapping(skb); 4355 4356 return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm)); 4357 } 4358 4359 #ifndef CONFIG_PREEMPT_RT 4360 static bool netdev_xmit_txqueue_skipped(void) 4361 { 4362 return __this_cpu_read(softnet_data.xmit.skip_txqueue); 4363 } 4364 4365 void netdev_xmit_skip_txqueue(bool skip) 4366 { 4367 __this_cpu_write(softnet_data.xmit.skip_txqueue, skip); 4368 } 4369 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4370 4371 #else 4372 static bool netdev_xmit_txqueue_skipped(void) 4373 { 4374 return current->net_xmit.skip_txqueue; 4375 } 4376 4377 void netdev_xmit_skip_txqueue(bool skip) 4378 { 4379 current->net_xmit.skip_txqueue = skip; 4380 } 4381 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4382 #endif 4383 #endif /* CONFIG_NET_EGRESS */ 4384 4385 #ifdef CONFIG_NET_XGRESS 4386 static int tc_run(struct tcx_entry *entry, struct sk_buff *skb, 4387 enum skb_drop_reason *drop_reason) 4388 { 4389 int ret = TC_ACT_UNSPEC; 4390 #ifdef CONFIG_NET_CLS_ACT 4391 struct mini_Qdisc *miniq = rcu_dereference_bh(entry->miniq); 4392 struct tcf_result res; 4393 4394 if (!miniq) 4395 return ret; 4396 4397 /* Global bypass */ 4398 if (!static_branch_likely(&tcf_sw_enabled_key)) 4399 return ret; 4400 4401 /* Block-wise bypass */ 4402 if (tcf_block_bypass_sw(miniq->block)) 4403 return ret; 4404 4405 tc_skb_cb(skb)->mru = 0; 4406 qdisc_skb_cb(skb)->post_ct = false; 4407 tcf_set_drop_reason(skb, *drop_reason); 4408 4409 mini_qdisc_bstats_cpu_update(miniq, skb); 4410 ret = tcf_classify(skb, miniq->block, miniq->filter_list, &res, false); 4411 /* Only tcf related quirks below. */ 4412 switch (ret) { 4413 case TC_ACT_SHOT: 4414 *drop_reason = tcf_get_drop_reason(skb); 4415 mini_qdisc_qstats_cpu_drop(miniq); 4416 break; 4417 case TC_ACT_OK: 4418 case TC_ACT_RECLASSIFY: 4419 skb->tc_index = TC_H_MIN(res.classid); 4420 break; 4421 } 4422 #endif /* CONFIG_NET_CLS_ACT */ 4423 return ret; 4424 } 4425 4426 static DEFINE_STATIC_KEY_FALSE(tcx_needed_key); 4427 4428 void tcx_inc(void) 4429 { 4430 static_branch_inc(&tcx_needed_key); 4431 } 4432 4433 void tcx_dec(void) 4434 { 4435 static_branch_dec(&tcx_needed_key); 4436 } 4437 4438 static __always_inline enum tcx_action_base 4439 tcx_run(const struct bpf_mprog_entry *entry, struct sk_buff *skb, 4440 const bool needs_mac) 4441 { 4442 const struct bpf_mprog_fp *fp; 4443 const struct bpf_prog *prog; 4444 int ret = TCX_NEXT; 4445 4446 if (needs_mac) 4447 __skb_push(skb, skb->mac_len); 4448 bpf_mprog_foreach_prog(entry, fp, prog) { 4449 bpf_compute_data_pointers(skb); 4450 ret = bpf_prog_run(prog, skb); 4451 if (ret != TCX_NEXT) 4452 break; 4453 } 4454 if (needs_mac) 4455 __skb_pull(skb, skb->mac_len); 4456 return tcx_action_code(skb, ret); 4457 } 4458 4459 static __always_inline struct sk_buff * 4460 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4461 struct net_device *orig_dev, bool *another) 4462 { 4463 struct bpf_mprog_entry *entry = rcu_dereference_bh(skb->dev->tcx_ingress); 4464 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_INGRESS; 4465 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4466 int sch_ret; 4467 4468 if (!entry) 4469 return skb; 4470 4471 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4472 if (unlikely(*pt_prev)) { 4473 *ret = deliver_skb(skb, *pt_prev, orig_dev); 4474 *pt_prev = NULL; 4475 } 4476 4477 qdisc_pkt_len_segs_init(skb); 4478 tcx_set_ingress(skb, true); 4479 4480 if (static_branch_unlikely(&tcx_needed_key)) { 4481 sch_ret = tcx_run(entry, skb, true); 4482 if (sch_ret != TC_ACT_UNSPEC) 4483 goto ingress_verdict; 4484 } 4485 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4486 ingress_verdict: 4487 switch (sch_ret) { 4488 case TC_ACT_REDIRECT: 4489 /* skb_mac_header check was done by BPF, so we can safely 4490 * push the L2 header back before redirecting to another 4491 * netdev. 4492 */ 4493 __skb_push(skb, skb->mac_len); 4494 if (skb_do_redirect(skb) == -EAGAIN) { 4495 __skb_pull(skb, skb->mac_len); 4496 *another = true; 4497 break; 4498 } 4499 *ret = NET_RX_SUCCESS; 4500 bpf_net_ctx_clear(bpf_net_ctx); 4501 return NULL; 4502 case TC_ACT_SHOT: 4503 kfree_skb_reason(skb, drop_reason); 4504 *ret = NET_RX_DROP; 4505 bpf_net_ctx_clear(bpf_net_ctx); 4506 return NULL; 4507 /* used by tc_run */ 4508 case TC_ACT_STOLEN: 4509 case TC_ACT_QUEUED: 4510 case TC_ACT_TRAP: 4511 consume_skb(skb); 4512 fallthrough; 4513 case TC_ACT_CONSUMED: 4514 *ret = NET_RX_SUCCESS; 4515 bpf_net_ctx_clear(bpf_net_ctx); 4516 return NULL; 4517 } 4518 bpf_net_ctx_clear(bpf_net_ctx); 4519 4520 return skb; 4521 } 4522 4523 static __always_inline struct sk_buff * 4524 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4525 { 4526 struct bpf_mprog_entry *entry = rcu_dereference_bh(dev->tcx_egress); 4527 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_EGRESS; 4528 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4529 int sch_ret; 4530 4531 if (!entry) 4532 return skb; 4533 4534 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4535 4536 /* qdisc_skb_cb(skb)->pkt_len & tcx_set_ingress() was 4537 * already set by the caller. 4538 */ 4539 if (static_branch_unlikely(&tcx_needed_key)) { 4540 sch_ret = tcx_run(entry, skb, false); 4541 if (sch_ret != TC_ACT_UNSPEC) 4542 goto egress_verdict; 4543 } 4544 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4545 egress_verdict: 4546 switch (sch_ret) { 4547 case TC_ACT_REDIRECT: 4548 /* No need to push/pop skb's mac_header here on egress! */ 4549 skb_do_redirect(skb); 4550 *ret = NET_XMIT_SUCCESS; 4551 bpf_net_ctx_clear(bpf_net_ctx); 4552 return NULL; 4553 case TC_ACT_SHOT: 4554 kfree_skb_reason(skb, drop_reason); 4555 *ret = NET_XMIT_DROP; 4556 bpf_net_ctx_clear(bpf_net_ctx); 4557 return NULL; 4558 /* used by tc_run */ 4559 case TC_ACT_STOLEN: 4560 case TC_ACT_QUEUED: 4561 case TC_ACT_TRAP: 4562 consume_skb(skb); 4563 fallthrough; 4564 case TC_ACT_CONSUMED: 4565 *ret = NET_XMIT_SUCCESS; 4566 bpf_net_ctx_clear(bpf_net_ctx); 4567 return NULL; 4568 } 4569 bpf_net_ctx_clear(bpf_net_ctx); 4570 4571 return skb; 4572 } 4573 #else 4574 static __always_inline struct sk_buff * 4575 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4576 struct net_device *orig_dev, bool *another) 4577 { 4578 return skb; 4579 } 4580 4581 static __always_inline struct sk_buff * 4582 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4583 { 4584 return skb; 4585 } 4586 #endif /* CONFIG_NET_XGRESS */ 4587 4588 #ifdef CONFIG_XPS 4589 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb, 4590 struct xps_dev_maps *dev_maps, unsigned int tci) 4591 { 4592 int tc = netdev_get_prio_tc_map(dev, skb->priority); 4593 struct xps_map *map; 4594 int queue_index = -1; 4595 4596 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids) 4597 return queue_index; 4598 4599 tci *= dev_maps->num_tc; 4600 tci += tc; 4601 4602 map = rcu_dereference(dev_maps->attr_map[tci]); 4603 if (map) { 4604 if (map->len == 1) 4605 queue_index = map->queues[0]; 4606 else 4607 queue_index = map->queues[reciprocal_scale( 4608 skb_get_hash(skb), map->len)]; 4609 if (unlikely(queue_index >= dev->real_num_tx_queues)) 4610 queue_index = -1; 4611 } 4612 return queue_index; 4613 } 4614 #endif 4615 4616 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev, 4617 struct sk_buff *skb) 4618 { 4619 #ifdef CONFIG_XPS 4620 struct xps_dev_maps *dev_maps; 4621 struct sock *sk = skb->sk; 4622 int queue_index = -1; 4623 4624 if (!static_key_false(&xps_needed)) 4625 return -1; 4626 4627 rcu_read_lock(); 4628 if (!static_key_false(&xps_rxqs_needed)) 4629 goto get_cpus_map; 4630 4631 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]); 4632 if (dev_maps) { 4633 int tci = sk_rx_queue_get(sk); 4634 4635 if (tci >= 0) 4636 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4637 tci); 4638 } 4639 4640 get_cpus_map: 4641 if (queue_index < 0) { 4642 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]); 4643 if (dev_maps) { 4644 unsigned int tci = skb->sender_cpu - 1; 4645 4646 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4647 tci); 4648 } 4649 } 4650 rcu_read_unlock(); 4651 4652 return queue_index; 4653 #else 4654 return -1; 4655 #endif 4656 } 4657 4658 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb, 4659 struct net_device *sb_dev) 4660 { 4661 return 0; 4662 } 4663 EXPORT_SYMBOL(dev_pick_tx_zero); 4664 4665 int sk_tx_queue_get(const struct sock *sk) 4666 { 4667 int resel, val; 4668 4669 if (!sk) 4670 return -1; 4671 /* Paired with WRITE_ONCE() in sk_tx_queue_clear() 4672 * and sk_tx_queue_set(). 4673 */ 4674 val = READ_ONCE(sk->sk_tx_queue_mapping); 4675 4676 if (val == NO_QUEUE_MAPPING) 4677 return -1; 4678 4679 if (!sk_fullsock(sk)) 4680 return val; 4681 4682 resel = READ_ONCE(sock_net(sk)->core.sysctl_txq_reselection); 4683 if (resel && time_is_before_jiffies( 4684 READ_ONCE(sk->sk_tx_queue_mapping_jiffies) + resel)) 4685 return -1; 4686 4687 return val; 4688 } 4689 EXPORT_SYMBOL(sk_tx_queue_get); 4690 4691 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb, 4692 struct net_device *sb_dev) 4693 { 4694 struct sock *sk = skb->sk; 4695 int queue_index = sk_tx_queue_get(sk); 4696 4697 sb_dev = sb_dev ? : dev; 4698 4699 if (queue_index < 0 || skb->ooo_okay || 4700 queue_index >= dev->real_num_tx_queues) { 4701 int new_index = get_xps_queue(dev, sb_dev, skb); 4702 4703 if (new_index < 0) 4704 new_index = skb_tx_hash(dev, sb_dev, skb); 4705 4706 if (sk && sk_fullsock(sk) && 4707 rcu_access_pointer(sk->sk_dst_cache)) 4708 sk_tx_queue_set(sk, new_index); 4709 4710 queue_index = new_index; 4711 } 4712 4713 return queue_index; 4714 } 4715 EXPORT_SYMBOL(netdev_pick_tx); 4716 4717 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev, 4718 struct sk_buff *skb, 4719 struct net_device *sb_dev) 4720 { 4721 int queue_index = 0; 4722 4723 #ifdef CONFIG_XPS 4724 u32 sender_cpu = skb->sender_cpu - 1; 4725 4726 if (sender_cpu >= (u32)NR_CPUS) 4727 skb->sender_cpu = raw_smp_processor_id() + 1; 4728 #endif 4729 4730 if (dev->real_num_tx_queues != 1) { 4731 const struct net_device_ops *ops = dev->netdev_ops; 4732 4733 if (ops->ndo_select_queue) 4734 queue_index = ops->ndo_select_queue(dev, skb, sb_dev); 4735 else 4736 queue_index = netdev_pick_tx(dev, skb, sb_dev); 4737 4738 queue_index = netdev_cap_txqueue(dev, queue_index); 4739 } 4740 4741 skb_set_queue_mapping(skb, queue_index); 4742 return netdev_get_tx_queue(dev, queue_index); 4743 } 4744 4745 /** 4746 * __dev_queue_xmit() - transmit a buffer 4747 * @skb: buffer to transmit 4748 * @sb_dev: suboordinate device used for L2 forwarding offload 4749 * 4750 * Queue a buffer for transmission to a network device. The caller must 4751 * have set the device and priority and built the buffer before calling 4752 * this function. The function can be called from an interrupt. 4753 * 4754 * When calling this method, interrupts MUST be enabled. This is because 4755 * the BH enable code must have IRQs enabled so that it will not deadlock. 4756 * 4757 * Regardless of the return value, the skb is consumed, so it is currently 4758 * difficult to retry a send to this method. (You can bump the ref count 4759 * before sending to hold a reference for retry if you are careful.) 4760 * 4761 * Return: 4762 * * 0 - buffer successfully transmitted 4763 * * positive qdisc return code - NET_XMIT_DROP etc. 4764 * * negative errno - other errors 4765 */ 4766 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev) 4767 { 4768 struct net_device *dev = skb->dev; 4769 struct netdev_queue *txq = NULL; 4770 enum skb_drop_reason reason; 4771 int cpu, rc = -ENOMEM; 4772 bool again = false; 4773 struct Qdisc *q; 4774 4775 skb_reset_mac_header(skb); 4776 skb_assert_len(skb); 4777 4778 if (unlikely(skb_shinfo(skb)->tx_flags & 4779 (SKBTX_SCHED_TSTAMP | SKBTX_BPF))) 4780 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED); 4781 4782 reason = qdisc_pkt_len_segs_init(skb); 4783 if (unlikely(reason)) { 4784 dev_core_stats_tx_dropped_inc(dev); 4785 kfree_skb_reason(skb, reason); 4786 return -EINVAL; 4787 } 4788 /* Disable soft irqs for various locks below. Also 4789 * stops preemption for RCU. 4790 */ 4791 rcu_read_lock_bh(); 4792 4793 skb_update_prio(skb); 4794 4795 tcx_set_ingress(skb, false); 4796 #ifdef CONFIG_NET_EGRESS 4797 if (static_branch_unlikely(&egress_needed_key)) { 4798 if (nf_hook_egress_active()) { 4799 skb = nf_hook_egress(skb, &rc, dev); 4800 if (!skb) 4801 goto out; 4802 } 4803 4804 netdev_xmit_skip_txqueue(false); 4805 4806 nf_skip_egress(skb, true); 4807 skb = sch_handle_egress(skb, &rc, dev); 4808 if (!skb) 4809 goto out; 4810 nf_skip_egress(skb, false); 4811 4812 if (netdev_xmit_txqueue_skipped()) 4813 txq = netdev_tx_queue_mapping(dev, skb); 4814 } 4815 #endif 4816 /* If device/qdisc don't need skb->dst, release it right now while 4817 * its hot in this cpu cache. 4818 */ 4819 if (dev->priv_flags & IFF_XMIT_DST_RELEASE) 4820 skb_dst_drop(skb); 4821 else 4822 skb_dst_force(skb); 4823 4824 if (!txq) 4825 txq = netdev_core_pick_tx(dev, skb, sb_dev); 4826 4827 q = rcu_dereference_bh(txq->qdisc); 4828 4829 trace_net_dev_queue(skb); 4830 if (q->enqueue) { 4831 rc = __dev_xmit_skb(skb, q, dev, txq); 4832 goto out; 4833 } 4834 4835 /* The device has no queue. Common case for software devices: 4836 * loopback, all the sorts of tunnels... 4837 4838 * Really, it is unlikely that netif_tx_lock protection is necessary 4839 * here. (f.e. loopback and IP tunnels are clean ignoring statistics 4840 * counters.) 4841 * However, it is possible, that they rely on protection 4842 * made by us here. 4843 4844 * Check this and shot the lock. It is not prone from deadlocks. 4845 *Either shot noqueue qdisc, it is even simpler 8) 4846 */ 4847 if (unlikely(!(dev->flags & IFF_UP))) { 4848 reason = SKB_DROP_REASON_DEV_READY; 4849 goto drop; 4850 } 4851 4852 cpu = smp_processor_id(); /* ok because BHs are off */ 4853 4854 if (likely(!netif_tx_owned(txq, cpu))) { 4855 bool is_list = false; 4856 4857 if (dev_xmit_recursion()) 4858 goto recursion_alert; 4859 4860 skb = validate_xmit_skb(skb, dev, &again); 4861 if (!skb) 4862 goto out; 4863 4864 HARD_TX_LOCK(dev, txq, cpu); 4865 4866 if (!netif_xmit_stopped(txq)) { 4867 is_list = !!skb->next; 4868 4869 dev_xmit_recursion_inc(); 4870 skb = dev_hard_start_xmit(skb, dev, txq, &rc); 4871 dev_xmit_recursion_dec(); 4872 4873 /* GSO segments a single SKB into a list of frames. 4874 * TCP expects error to mean none of the data was sent. 4875 */ 4876 if (is_list) 4877 rc = NETDEV_TX_OK; 4878 } 4879 HARD_TX_UNLOCK(dev, txq); 4880 if (!skb) /* xmit completed */ 4881 goto out; 4882 4883 net_crit_ratelimited("Virtual device %s asks to queue packet!\n", 4884 dev->name); 4885 /* NETDEV_TX_BUSY or queue was stopped */ 4886 if (!is_list) 4887 rc = -ENETDOWN; 4888 } else { 4889 /* Recursion is detected! It is possible unfortunately. */ 4890 recursion_alert: 4891 net_crit_ratelimited("Dead loop on virtual device %s (net %llu), fix it urgently!\n", 4892 dev->name, dev_net(dev)->net_cookie); 4893 4894 rc = -ENETDOWN; 4895 } 4896 4897 reason = SKB_DROP_REASON_RECURSION_LIMIT; 4898 drop: 4899 rcu_read_unlock_bh(); 4900 4901 dev_core_stats_tx_dropped_inc(dev); 4902 kfree_skb_list_reason(skb, reason); 4903 return rc; 4904 out: 4905 rcu_read_unlock_bh(); 4906 return rc; 4907 } 4908 EXPORT_SYMBOL(__dev_queue_xmit); 4909 4910 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id) 4911 { 4912 struct net_device *dev = skb->dev; 4913 struct sk_buff *orig_skb = skb; 4914 struct netdev_queue *txq; 4915 int ret = NETDEV_TX_BUSY; 4916 bool again = false; 4917 4918 if (unlikely(!netif_running(dev) || 4919 !netif_carrier_ok(dev))) 4920 goto drop; 4921 4922 skb = validate_xmit_skb_list(skb, dev, &again); 4923 if (skb != orig_skb) 4924 goto drop; 4925 4926 skb_set_queue_mapping(skb, queue_id); 4927 txq = skb_get_tx_queue(dev, skb); 4928 4929 local_bh_disable(); 4930 4931 dev_xmit_recursion_inc(); 4932 HARD_TX_LOCK(dev, txq, smp_processor_id()); 4933 if (!netif_xmit_frozen_or_drv_stopped(txq)) 4934 ret = netdev_start_xmit(skb, dev, txq, false); 4935 HARD_TX_UNLOCK(dev, txq); 4936 dev_xmit_recursion_dec(); 4937 4938 local_bh_enable(); 4939 return ret; 4940 drop: 4941 dev_core_stats_tx_dropped_inc(dev); 4942 kfree_skb_list(skb); 4943 return NET_XMIT_DROP; 4944 } 4945 EXPORT_SYMBOL(__dev_direct_xmit); 4946 4947 /************************************************************************* 4948 * Receiver routines 4949 *************************************************************************/ 4950 static DEFINE_PER_CPU(struct task_struct *, backlog_napi); 4951 4952 int weight_p __read_mostly = 64; /* old backlog weight */ 4953 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */ 4954 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */ 4955 4956 /* Called with irq disabled */ 4957 static inline void ____napi_schedule(struct softnet_data *sd, 4958 struct napi_struct *napi) 4959 { 4960 struct task_struct *thread; 4961 4962 lockdep_assert_irqs_disabled(); 4963 4964 if (test_bit(NAPI_STATE_THREADED, &napi->state)) { 4965 /* Paired with smp_mb__before_atomic() in 4966 * napi_enable()/netif_set_threaded(). 4967 * Use READ_ONCE() to guarantee a complete 4968 * read on napi->thread. Only call 4969 * wake_up_process() when it's not NULL. 4970 */ 4971 thread = READ_ONCE(napi->thread); 4972 if (thread) { 4973 if (use_backlog_threads() && thread == raw_cpu_read(backlog_napi)) 4974 goto use_local_napi; 4975 4976 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 4977 wake_up_process(thread); 4978 return; 4979 } 4980 } 4981 4982 use_local_napi: 4983 DEBUG_NET_WARN_ON_ONCE(!list_empty(&napi->poll_list)); 4984 list_add_tail(&napi->poll_list, &sd->poll_list); 4985 WRITE_ONCE(napi->list_owner, smp_processor_id()); 4986 /* If not called from net_rx_action() 4987 * we have to raise NET_RX_SOFTIRQ. 4988 */ 4989 if (!sd->in_net_rx_action) 4990 raise_softirq_irqoff(NET_RX_SOFTIRQ); 4991 } 4992 4993 #ifdef CONFIG_RPS 4994 4995 struct static_key_false rps_needed __read_mostly; 4996 EXPORT_SYMBOL(rps_needed); 4997 struct static_key_false rfs_needed __read_mostly; 4998 EXPORT_SYMBOL(rfs_needed); 4999 5000 static u32 rfs_slot(u32 hash, rps_tag_ptr tag_ptr) 5001 { 5002 return hash_32(hash, rps_tag_to_log(tag_ptr)); 5003 } 5004 5005 #ifdef CONFIG_RFS_ACCEL 5006 /** 5007 * rps_flow_is_active - check whether the flow is recently active. 5008 * @rflow: Specific flow to check activity. 5009 * @log: ilog2(hashsize). 5010 * @cpu: CPU saved in @rflow. 5011 * 5012 * If the CPU has processed many packets since the flow's last activity 5013 * (beyond 10 times the table size), the flow is considered stale. 5014 * 5015 * Return: true if flow was recently active. 5016 */ 5017 static bool rps_flow_is_active(struct rps_dev_flow *rflow, 5018 u8 log, 5019 unsigned int cpu) 5020 { 5021 unsigned int flow_last_active; 5022 unsigned int sd_input_head; 5023 5024 if (cpu >= nr_cpu_ids) 5025 return false; 5026 5027 sd_input_head = READ_ONCE(per_cpu(softnet_data, cpu).input_queue_head); 5028 flow_last_active = READ_ONCE(rflow->last_qtail); 5029 5030 return (int)(sd_input_head - flow_last_active) < 5031 (int)(10 << log); 5032 } 5033 #endif 5034 5035 static struct rps_dev_flow * 5036 set_rps_cpu(struct net_device *dev, struct sk_buff *skb, 5037 struct rps_dev_flow *rflow, u16 next_cpu, u32 hash) 5038 { 5039 if (next_cpu < nr_cpu_ids) { 5040 u32 head; 5041 #ifdef CONFIG_RFS_ACCEL 5042 struct netdev_rx_queue *rxqueue; 5043 struct rps_dev_flow *flow_table; 5044 struct rps_dev_flow *old_rflow; 5045 struct rps_dev_flow *tmp_rflow; 5046 rps_tag_ptr q_tag_ptr; 5047 unsigned int tmp_cpu; 5048 u16 rxq_index; 5049 u32 flow_id; 5050 int rc; 5051 5052 /* Should we steer this flow to a different hardware queue? */ 5053 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap || 5054 !(dev->features & NETIF_F_NTUPLE)) 5055 goto out; 5056 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu); 5057 if (rxq_index == skb_get_rx_queue(skb)) 5058 goto out; 5059 5060 rxqueue = dev->_rx + rxq_index; 5061 q_tag_ptr = READ_ONCE(rxqueue->rps_flow_table); 5062 if (!q_tag_ptr) 5063 goto out; 5064 5065 flow_id = rfs_slot(hash, q_tag_ptr); 5066 flow_table = rps_tag_to_table(q_tag_ptr); 5067 tmp_rflow = flow_table + flow_id; 5068 tmp_cpu = READ_ONCE(tmp_rflow->cpu); 5069 5070 if (READ_ONCE(tmp_rflow->filter) != RPS_NO_FILTER) { 5071 if (rps_flow_is_active(tmp_rflow, 5072 rps_tag_to_log(q_tag_ptr), 5073 tmp_cpu)) { 5074 if (hash != READ_ONCE(tmp_rflow->hash) || 5075 next_cpu == tmp_cpu) 5076 goto out; 5077 } 5078 } 5079 5080 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb, 5081 rxq_index, flow_id); 5082 if (rc < 0) 5083 goto out; 5084 5085 old_rflow = rflow; 5086 rflow = tmp_rflow; 5087 WRITE_ONCE(rflow->filter, rc); 5088 WRITE_ONCE(rflow->hash, hash); 5089 5090 if (old_rflow->filter == rc) 5091 WRITE_ONCE(old_rflow->filter, RPS_NO_FILTER); 5092 out: 5093 #endif 5094 head = READ_ONCE(per_cpu(softnet_data, next_cpu).input_queue_head); 5095 rps_input_queue_tail_save(&rflow->last_qtail, head); 5096 } 5097 5098 WRITE_ONCE(rflow->cpu, next_cpu); 5099 return rflow; 5100 } 5101 5102 /* 5103 * get_rps_cpu is called from netif_receive_skb and returns the target 5104 * CPU from the RPS map of the receiving queue for a given skb. 5105 * rcu_read_lock must be held on entry. 5106 */ 5107 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb, 5108 struct rps_dev_flow **rflowp) 5109 { 5110 struct netdev_rx_queue *rxqueue = dev->_rx; 5111 rps_tag_ptr global_tag_ptr, q_tag_ptr; 5112 struct rps_map *map; 5113 int cpu = -1; 5114 u32 tcpu; 5115 u32 hash; 5116 5117 if (skb_rx_queue_recorded(skb)) { 5118 u16 index = skb_get_rx_queue(skb); 5119 5120 if (unlikely(index >= dev->real_num_rx_queues)) { 5121 WARN_ONCE(dev->real_num_rx_queues > 1, 5122 "%s received packet on queue %u, but number " 5123 "of RX queues is %u\n", 5124 dev->name, index, dev->real_num_rx_queues); 5125 goto done; 5126 } 5127 rxqueue += index; 5128 } 5129 5130 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */ 5131 5132 q_tag_ptr = READ_ONCE(rxqueue->rps_flow_table); 5133 map = rcu_dereference(rxqueue->rps_map); 5134 if (!q_tag_ptr && !map) 5135 goto done; 5136 5137 skb_reset_network_header(skb); 5138 hash = skb_get_hash(skb); 5139 if (!hash) 5140 goto done; 5141 5142 global_tag_ptr = READ_ONCE(net_hotdata.rps_sock_flow_table); 5143 if (q_tag_ptr && global_tag_ptr) { 5144 struct rps_sock_flow_table *sock_flow_table; 5145 struct rps_dev_flow *flow_table; 5146 struct rps_dev_flow *rflow; 5147 u32 next_cpu; 5148 u32 flow_id; 5149 u32 ident; 5150 5151 /* First check into global flow table if there is a match. 5152 * This READ_ONCE() pairs with WRITE_ONCE() from rps_record_sock_flow(). 5153 */ 5154 flow_id = hash & rps_tag_to_mask(global_tag_ptr); 5155 sock_flow_table = rps_tag_to_table(global_tag_ptr); 5156 ident = READ_ONCE(sock_flow_table[flow_id].ent); 5157 if ((ident ^ hash) & ~net_hotdata.rps_cpu_mask) 5158 goto try_rps; 5159 5160 next_cpu = ident & net_hotdata.rps_cpu_mask; 5161 5162 /* OK, now we know there is a match, 5163 * we can look at the local (per receive queue) flow table 5164 */ 5165 flow_id = rfs_slot(hash, q_tag_ptr); 5166 flow_table = rps_tag_to_table(q_tag_ptr); 5167 rflow = flow_table + flow_id; 5168 tcpu = rflow->cpu; 5169 5170 /* 5171 * If the desired CPU (where last recvmsg was done) is 5172 * different from current CPU (one in the rx-queue flow 5173 * table entry), switch if one of the following holds: 5174 * - Current CPU is unset (>= nr_cpu_ids). 5175 * - Current CPU is offline. 5176 * - The current CPU's queue tail has advanced beyond the 5177 * last packet that was enqueued using this table entry. 5178 * This guarantees that all previous packets for the flow 5179 * have been dequeued, thus preserving in order delivery. 5180 */ 5181 if (unlikely(tcpu != next_cpu) && 5182 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) || 5183 ((int)(READ_ONCE(per_cpu(softnet_data, tcpu).input_queue_head) - 5184 rflow->last_qtail)) >= 0)) { 5185 tcpu = next_cpu; 5186 rflow = set_rps_cpu(dev, skb, rflow, next_cpu, hash); 5187 } 5188 5189 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) { 5190 *rflowp = rflow; 5191 cpu = tcpu; 5192 goto done; 5193 } 5194 } 5195 5196 try_rps: 5197 5198 if (map) { 5199 tcpu = map->cpus[reciprocal_scale(hash, map->len)]; 5200 if (cpu_online(tcpu)) { 5201 cpu = tcpu; 5202 goto done; 5203 } 5204 } 5205 5206 done: 5207 return cpu; 5208 } 5209 5210 #ifdef CONFIG_RFS_ACCEL 5211 5212 /** 5213 * rps_may_expire_flow - check whether an RFS hardware filter may be removed 5214 * @dev: Device on which the filter was set 5215 * @rxq_index: RX queue index 5216 * @flow_id: Flow ID passed to ndo_rx_flow_steer() 5217 * @filter_id: Filter ID returned by ndo_rx_flow_steer() 5218 * 5219 * Drivers that implement ndo_rx_flow_steer() should periodically call 5220 * this function for each installed filter and remove the filters for 5221 * which it returns %true. 5222 */ 5223 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index, 5224 u32 flow_id, u16 filter_id) 5225 { 5226 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index; 5227 struct rps_dev_flow *flow_table; 5228 struct rps_dev_flow *rflow; 5229 rps_tag_ptr q_tag_ptr; 5230 bool expire = true; 5231 u8 log; 5232 5233 rcu_read_lock(); 5234 q_tag_ptr = READ_ONCE(rxqueue->rps_flow_table); 5235 log = rps_tag_to_log(q_tag_ptr); 5236 if (q_tag_ptr && flow_id < (1UL << log)) { 5237 unsigned int cpu; 5238 5239 flow_table = rps_tag_to_table(q_tag_ptr); 5240 rflow = flow_table + flow_id; 5241 cpu = READ_ONCE(rflow->cpu); 5242 if (READ_ONCE(rflow->filter) == filter_id && 5243 rps_flow_is_active(rflow, log, cpu)) 5244 expire = false; 5245 } 5246 rcu_read_unlock(); 5247 return expire; 5248 } 5249 EXPORT_SYMBOL(rps_may_expire_flow); 5250 5251 #endif /* CONFIG_RFS_ACCEL */ 5252 5253 /* Called from hardirq (IPI) context */ 5254 static void rps_trigger_softirq(void *data) 5255 { 5256 struct softnet_data *sd = data; 5257 5258 ____napi_schedule(sd, &sd->backlog); 5259 /* Pairs with READ_ONCE() in softnet_seq_show() */ 5260 WRITE_ONCE(sd->received_rps, sd->received_rps + 1); 5261 } 5262 5263 #endif /* CONFIG_RPS */ 5264 5265 /* Called from hardirq (IPI) context */ 5266 static void trigger_rx_softirq(void *data) 5267 { 5268 struct softnet_data *sd = data; 5269 5270 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 5271 smp_store_release(&sd->defer_ipi_scheduled, 0); 5272 } 5273 5274 /* 5275 * After we queued a packet into sd->input_pkt_queue, 5276 * we need to make sure this queue is serviced soon. 5277 * 5278 * - If this is another cpu queue, link it to our rps_ipi_list, 5279 * and make sure we will process rps_ipi_list from net_rx_action(). 5280 * 5281 * - If this is our own queue, NAPI schedule our backlog. 5282 * Note that this also raises NET_RX_SOFTIRQ. 5283 */ 5284 static void napi_schedule_rps(struct softnet_data *sd) 5285 { 5286 struct softnet_data *mysd = this_cpu_ptr(&softnet_data); 5287 5288 #ifdef CONFIG_RPS 5289 if (sd != mysd) { 5290 if (use_backlog_threads()) { 5291 __napi_schedule_irqoff(&sd->backlog); 5292 return; 5293 } 5294 5295 sd->rps_ipi_next = mysd->rps_ipi_list; 5296 mysd->rps_ipi_list = sd; 5297 5298 /* If not called from net_rx_action() or napi_threaded_poll() 5299 * we have to raise NET_RX_SOFTIRQ. 5300 */ 5301 if (!mysd->in_net_rx_action && !mysd->in_napi_threaded_poll) 5302 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 5303 return; 5304 } 5305 #endif /* CONFIG_RPS */ 5306 __napi_schedule_irqoff(&mysd->backlog); 5307 } 5308 5309 void kick_defer_list_purge(unsigned int cpu) 5310 { 5311 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 5312 unsigned long flags; 5313 5314 if (use_backlog_threads()) { 5315 backlog_lock_irq_save(sd, &flags); 5316 5317 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) 5318 __napi_schedule_irqoff(&sd->backlog); 5319 5320 backlog_unlock_irq_restore(sd, flags); 5321 5322 } else if (!cmpxchg(&sd->defer_ipi_scheduled, 0, 1)) { 5323 smp_call_function_single_async(cpu, &sd->defer_csd); 5324 } 5325 } 5326 5327 #ifdef CONFIG_NET_FLOW_LIMIT 5328 int netdev_flow_limit_table_len __read_mostly = (1 << 12); 5329 #endif 5330 5331 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen, 5332 int max_backlog) 5333 { 5334 #ifdef CONFIG_NET_FLOW_LIMIT 5335 unsigned int old_flow, new_flow; 5336 const struct softnet_data *sd; 5337 struct sd_flow_limit *fl; 5338 5339 if (likely(qlen < (max_backlog >> 1))) 5340 return false; 5341 5342 sd = this_cpu_ptr(&softnet_data); 5343 5344 rcu_read_lock(); 5345 fl = rcu_dereference(sd->flow_limit); 5346 if (fl) { 5347 new_flow = hash_32(skb_get_hash(skb), fl->log_buckets); 5348 old_flow = fl->history[fl->history_head]; 5349 fl->history[fl->history_head] = new_flow; 5350 5351 fl->history_head++; 5352 fl->history_head &= FLOW_LIMIT_HISTORY - 1; 5353 5354 if (likely(fl->buckets[old_flow])) 5355 fl->buckets[old_flow]--; 5356 5357 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) { 5358 /* Pairs with READ_ONCE() in softnet_seq_show() */ 5359 WRITE_ONCE(fl->count, fl->count + 1); 5360 rcu_read_unlock(); 5361 return true; 5362 } 5363 } 5364 rcu_read_unlock(); 5365 #endif 5366 return false; 5367 } 5368 5369 /* 5370 * enqueue_to_backlog is called to queue an skb to a per CPU backlog 5371 * queue (may be a remote CPU queue). 5372 */ 5373 static int enqueue_to_backlog(struct sk_buff *skb, int cpu, 5374 unsigned int *qtail) 5375 { 5376 enum skb_drop_reason reason; 5377 struct softnet_data *sd; 5378 unsigned long flags; 5379 unsigned int qlen; 5380 int max_backlog; 5381 u32 tail; 5382 5383 reason = SKB_DROP_REASON_DEV_READY; 5384 if (unlikely(!netif_running(skb->dev))) 5385 goto bad_dev; 5386 5387 sd = &per_cpu(softnet_data, cpu); 5388 5389 qlen = skb_queue_len_lockless(&sd->input_pkt_queue); 5390 max_backlog = READ_ONCE(net_hotdata.max_backlog); 5391 if (unlikely(qlen > max_backlog) || 5392 skb_flow_limit(skb, qlen, max_backlog)) 5393 goto cpu_backlog_drop; 5394 backlog_lock_irq_save(sd, &flags); 5395 qlen = skb_queue_len(&sd->input_pkt_queue); 5396 if (likely(qlen <= max_backlog)) { 5397 if (!qlen) { 5398 /* Schedule NAPI for backlog device. We can use 5399 * non atomic operation as we own the queue lock. 5400 */ 5401 if (!__test_and_set_bit(NAPI_STATE_SCHED, 5402 &sd->backlog.state)) 5403 napi_schedule_rps(sd); 5404 } 5405 __skb_queue_tail(&sd->input_pkt_queue, skb); 5406 tail = rps_input_queue_tail_incr(sd); 5407 backlog_unlock_irq_restore(sd, flags); 5408 5409 /* save the tail outside of the critical section */ 5410 rps_input_queue_tail_save(qtail, tail); 5411 return NET_RX_SUCCESS; 5412 } 5413 5414 backlog_unlock_irq_restore(sd, flags); 5415 5416 cpu_backlog_drop: 5417 reason = SKB_DROP_REASON_CPU_BACKLOG; 5418 numa_drop_add(&sd->drop_counters, 1); 5419 bad_dev: 5420 dev_core_stats_rx_dropped_inc(skb->dev); 5421 kfree_skb_reason(skb, reason); 5422 return NET_RX_DROP; 5423 } 5424 5425 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb) 5426 { 5427 struct net_device *dev = skb->dev; 5428 struct netdev_rx_queue *rxqueue; 5429 5430 rxqueue = dev->_rx; 5431 5432 if (skb_rx_queue_recorded(skb)) { 5433 u16 index = skb_get_rx_queue(skb); 5434 5435 if (unlikely(index >= dev->real_num_rx_queues)) { 5436 WARN_ONCE(dev->real_num_rx_queues > 1, 5437 "%s received packet on queue %u, but number " 5438 "of RX queues is %u\n", 5439 dev->name, index, dev->real_num_rx_queues); 5440 5441 return rxqueue; /* Return first rxqueue */ 5442 } 5443 rxqueue += index; 5444 } 5445 return rxqueue; 5446 } 5447 5448 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, 5449 const struct bpf_prog *xdp_prog) 5450 { 5451 void *orig_data, *orig_data_end, *hard_start; 5452 struct netdev_rx_queue *rxqueue; 5453 bool orig_bcast, orig_host; 5454 u32 mac_len, frame_sz; 5455 __be16 orig_eth_type; 5456 struct ethhdr *eth; 5457 u32 metalen, act; 5458 int off; 5459 5460 /* The XDP program wants to see the packet starting at the MAC 5461 * header. 5462 */ 5463 mac_len = skb->data - skb_mac_header(skb); 5464 hard_start = skb->data - skb_headroom(skb); 5465 5466 /* SKB "head" area always have tailroom for skb_shared_info */ 5467 frame_sz = (void *)skb_end_pointer(skb) - hard_start; 5468 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 5469 5470 rxqueue = netif_get_rxqueue(skb); 5471 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq); 5472 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len, 5473 skb_headlen(skb) + mac_len, true); 5474 if (skb_is_nonlinear(skb)) { 5475 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 5476 xdp_buff_set_frags_flag(xdp); 5477 } else { 5478 xdp_buff_clear_frags_flag(xdp); 5479 } 5480 5481 orig_data_end = xdp->data_end; 5482 orig_data = xdp->data; 5483 eth = (struct ethhdr *)xdp->data; 5484 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr); 5485 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest); 5486 orig_eth_type = eth->h_proto; 5487 5488 act = bpf_prog_run_xdp(xdp_prog, xdp); 5489 5490 /* check if bpf_xdp_adjust_head was used */ 5491 off = xdp->data - orig_data; 5492 if (off) { 5493 if (off > 0) 5494 __skb_pull(skb, off); 5495 else if (off < 0) 5496 __skb_push(skb, -off); 5497 5498 skb->mac_header += off; 5499 skb_reset_network_header(skb); 5500 } 5501 5502 /* check if bpf_xdp_adjust_tail was used */ 5503 off = xdp->data_end - orig_data_end; 5504 if (off != 0) { 5505 skb_set_tail_pointer(skb, xdp->data_end - xdp->data); 5506 skb->len += off; /* positive on grow, negative on shrink */ 5507 } 5508 5509 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 5510 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here. 5511 */ 5512 if (xdp_buff_has_frags(xdp)) 5513 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 5514 else 5515 skb->data_len = 0; 5516 5517 /* check if XDP changed eth hdr such SKB needs update */ 5518 eth = (struct ethhdr *)xdp->data; 5519 if ((orig_eth_type != eth->h_proto) || 5520 (orig_host != ether_addr_equal_64bits(eth->h_dest, 5521 skb->dev->dev_addr)) || 5522 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) { 5523 __skb_push(skb, ETH_HLEN); 5524 skb->pkt_type = PACKET_HOST; 5525 skb->protocol = eth_type_trans(skb, skb->dev); 5526 } 5527 5528 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull 5529 * before calling us again on redirect path. We do not call do_redirect 5530 * as we leave that up to the caller. 5531 * 5532 * Caller is responsible for managing lifetime of skb (i.e. calling 5533 * kfree_skb in response to actions it cannot handle/XDP_DROP). 5534 */ 5535 switch (act) { 5536 case XDP_REDIRECT: 5537 case XDP_TX: 5538 __skb_push(skb, mac_len); 5539 break; 5540 case XDP_PASS: 5541 metalen = xdp->data - xdp->data_meta; 5542 if (metalen) 5543 skb_metadata_set(skb, metalen); 5544 break; 5545 } 5546 5547 return act; 5548 } 5549 5550 static int 5551 netif_skb_check_for_xdp(struct sk_buff **pskb, const struct bpf_prog *prog) 5552 { 5553 struct sk_buff *skb = *pskb; 5554 int err, hroom, troom; 5555 5556 local_lock_nested_bh(&system_page_pool.bh_lock); 5557 err = skb_cow_data_for_xdp(this_cpu_read(system_page_pool.pool), pskb, prog); 5558 local_unlock_nested_bh(&system_page_pool.bh_lock); 5559 if (!err) 5560 return 0; 5561 5562 /* In case we have to go down the path and also linearize, 5563 * then lets do the pskb_expand_head() work just once here. 5564 */ 5565 hroom = XDP_PACKET_HEADROOM - skb_headroom(skb); 5566 troom = skb->tail + skb->data_len - skb->end; 5567 err = pskb_expand_head(skb, 5568 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0, 5569 troom > 0 ? troom + 128 : 0, GFP_ATOMIC); 5570 if (err) 5571 return err; 5572 5573 return skb_linearize(skb); 5574 } 5575 5576 static u32 netif_receive_generic_xdp(struct sk_buff **pskb, 5577 struct xdp_buff *xdp, 5578 const struct bpf_prog *xdp_prog) 5579 { 5580 struct sk_buff *skb = *pskb; 5581 u32 mac_len, act = XDP_DROP; 5582 5583 /* Reinjected packets coming from act_mirred or similar should 5584 * not get XDP generic processing. 5585 */ 5586 if (skb_is_redirected(skb)) 5587 return XDP_PASS; 5588 5589 /* XDP packets must have sufficient headroom of XDP_PACKET_HEADROOM 5590 * bytes. This is the guarantee that also native XDP provides, 5591 * thus we need to do it here as well. 5592 */ 5593 mac_len = skb->data - skb_mac_header(skb); 5594 __skb_push(skb, mac_len); 5595 5596 if (skb_cloned(skb) || skb_is_nonlinear(skb) || 5597 skb_headroom(skb) < XDP_PACKET_HEADROOM) { 5598 if (netif_skb_check_for_xdp(pskb, xdp_prog)) 5599 goto do_drop; 5600 } 5601 5602 __skb_pull(*pskb, mac_len); 5603 5604 act = bpf_prog_run_generic_xdp(*pskb, xdp, xdp_prog); 5605 switch (act) { 5606 case XDP_REDIRECT: 5607 case XDP_TX: 5608 case XDP_PASS: 5609 break; 5610 default: 5611 bpf_warn_invalid_xdp_action((*pskb)->dev, xdp_prog, act); 5612 fallthrough; 5613 case XDP_ABORTED: 5614 trace_xdp_exception((*pskb)->dev, xdp_prog, act); 5615 fallthrough; 5616 case XDP_DROP: 5617 do_drop: 5618 kfree_skb(*pskb); 5619 break; 5620 } 5621 5622 return act; 5623 } 5624 5625 /* When doing generic XDP we have to bypass the qdisc layer and the 5626 * network taps in order to match in-driver-XDP behavior. This also means 5627 * that XDP packets are able to starve other packets going through a qdisc, 5628 * and DDOS attacks will be more effective. In-driver-XDP use dedicated TX 5629 * queues, so they do not have this starvation issue. 5630 */ 5631 void generic_xdp_tx(struct sk_buff *skb, const struct bpf_prog *xdp_prog) 5632 { 5633 struct net_device *dev = skb->dev; 5634 struct netdev_queue *txq; 5635 bool free_skb = true; 5636 int cpu, rc; 5637 5638 txq = netdev_core_pick_tx(dev, skb, NULL); 5639 cpu = smp_processor_id(); 5640 HARD_TX_LOCK(dev, txq, cpu); 5641 if (!netif_xmit_frozen_or_drv_stopped(txq)) { 5642 rc = netdev_start_xmit(skb, dev, txq, 0); 5643 if (dev_xmit_complete(rc)) 5644 free_skb = false; 5645 } 5646 HARD_TX_UNLOCK(dev, txq); 5647 if (free_skb) { 5648 trace_xdp_exception(dev, xdp_prog, XDP_TX); 5649 dev_core_stats_tx_dropped_inc(dev); 5650 kfree_skb(skb); 5651 } 5652 } 5653 5654 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key); 5655 5656 int do_xdp_generic(const struct bpf_prog *xdp_prog, struct sk_buff **pskb) 5657 { 5658 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 5659 5660 if (xdp_prog) { 5661 struct xdp_buff xdp; 5662 u32 act; 5663 int err; 5664 5665 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 5666 act = netif_receive_generic_xdp(pskb, &xdp, xdp_prog); 5667 if (act != XDP_PASS) { 5668 switch (act) { 5669 case XDP_REDIRECT: 5670 err = xdp_do_generic_redirect((*pskb)->dev, *pskb, 5671 &xdp, xdp_prog); 5672 if (err) 5673 goto out_redir; 5674 break; 5675 case XDP_TX: 5676 generic_xdp_tx(*pskb, xdp_prog); 5677 break; 5678 } 5679 bpf_net_ctx_clear(bpf_net_ctx); 5680 return XDP_DROP; 5681 } 5682 bpf_net_ctx_clear(bpf_net_ctx); 5683 } 5684 return XDP_PASS; 5685 out_redir: 5686 bpf_net_ctx_clear(bpf_net_ctx); 5687 kfree_skb_reason(*pskb, SKB_DROP_REASON_XDP); 5688 return XDP_DROP; 5689 } 5690 EXPORT_SYMBOL_GPL(do_xdp_generic); 5691 5692 static int netif_rx_internal(struct sk_buff *skb) 5693 { 5694 int ret; 5695 5696 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5697 5698 trace_netif_rx(skb); 5699 5700 #ifdef CONFIG_RPS 5701 if (static_branch_unlikely(&rps_needed)) { 5702 struct rps_dev_flow voidflow, *rflow = &voidflow; 5703 int cpu; 5704 5705 rcu_read_lock(); 5706 5707 cpu = get_rps_cpu(skb->dev, skb, &rflow); 5708 if (cpu < 0) 5709 cpu = smp_processor_id(); 5710 5711 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5712 5713 rcu_read_unlock(); 5714 } else 5715 #endif 5716 { 5717 unsigned int qtail; 5718 5719 ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail); 5720 } 5721 return ret; 5722 } 5723 5724 /** 5725 * __netif_rx - Slightly optimized version of netif_rx 5726 * @skb: buffer to post 5727 * 5728 * This behaves as netif_rx except that it does not disable bottom halves. 5729 * As a result this function may only be invoked from the interrupt context 5730 * (either hard or soft interrupt). 5731 */ 5732 int __netif_rx(struct sk_buff *skb) 5733 { 5734 int ret; 5735 5736 lockdep_assert_once(hardirq_count() | softirq_count()); 5737 5738 trace_netif_rx_entry(skb); 5739 ret = netif_rx_internal(skb); 5740 trace_netif_rx_exit(ret); 5741 return ret; 5742 } 5743 EXPORT_SYMBOL(__netif_rx); 5744 5745 /** 5746 * netif_rx - post buffer to the network code 5747 * @skb: buffer to post 5748 * 5749 * This function receives a packet from a device driver and queues it for 5750 * the upper (protocol) levels to process via the backlog NAPI device. It 5751 * always succeeds. The buffer may be dropped during processing for 5752 * congestion control or by the protocol layers. 5753 * The network buffer is passed via the backlog NAPI device. Modern NIC 5754 * driver should use NAPI and GRO. 5755 * This function can used from interrupt and from process context. The 5756 * caller from process context must not disable interrupts before invoking 5757 * this function. 5758 * 5759 * return values: 5760 * NET_RX_SUCCESS (no congestion) 5761 * NET_RX_DROP (packet was dropped) 5762 * 5763 */ 5764 int netif_rx(struct sk_buff *skb) 5765 { 5766 bool need_bh_off = !(hardirq_count() | softirq_count()); 5767 int ret; 5768 5769 if (need_bh_off) 5770 local_bh_disable(); 5771 trace_netif_rx_entry(skb); 5772 ret = netif_rx_internal(skb); 5773 trace_netif_rx_exit(ret); 5774 if (need_bh_off) 5775 local_bh_enable(); 5776 return ret; 5777 } 5778 EXPORT_SYMBOL(netif_rx); 5779 5780 static __latent_entropy void net_tx_action(void) 5781 { 5782 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 5783 5784 if (sd->completion_queue) { 5785 struct sk_buff *clist; 5786 5787 local_irq_disable(); 5788 clist = sd->completion_queue; 5789 sd->completion_queue = NULL; 5790 local_irq_enable(); 5791 5792 while (clist) { 5793 struct sk_buff *skb = clist; 5794 5795 clist = clist->next; 5796 5797 WARN_ON(refcount_read(&skb->users)); 5798 if (likely(get_kfree_skb_cb(skb)->reason == SKB_CONSUMED)) 5799 trace_consume_skb(skb, net_tx_action); 5800 else 5801 trace_kfree_skb(skb, net_tx_action, 5802 get_kfree_skb_cb(skb)->reason, NULL); 5803 5804 if (skb->fclone != SKB_FCLONE_UNAVAILABLE) 5805 __kfree_skb(skb); 5806 else 5807 __napi_kfree_skb(skb, 5808 get_kfree_skb_cb(skb)->reason); 5809 } 5810 } 5811 5812 if (sd->output_queue) { 5813 struct Qdisc *head; 5814 5815 local_irq_disable(); 5816 head = sd->output_queue; 5817 sd->output_queue = NULL; 5818 sd->output_queue_tailp = &sd->output_queue; 5819 local_irq_enable(); 5820 5821 rcu_read_lock(); 5822 5823 while (head) { 5824 spinlock_t *root_lock = NULL; 5825 struct sk_buff *to_free; 5826 struct Qdisc *q = head; 5827 5828 head = head->next_sched; 5829 5830 /* We need to make sure head->next_sched is read 5831 * before clearing __QDISC_STATE_SCHED 5832 */ 5833 smp_mb__before_atomic(); 5834 5835 if (!(q->flags & TCQ_F_NOLOCK)) { 5836 root_lock = qdisc_lock(q); 5837 spin_lock(root_lock); 5838 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, 5839 &q->state))) { 5840 /* There is a synchronize_net() between 5841 * STATE_DEACTIVATED flag being set and 5842 * qdisc_reset()/some_qdisc_is_busy() in 5843 * dev_deactivate(), so we can safely bail out 5844 * early here to avoid data race between 5845 * qdisc_deactivate() and some_qdisc_is_busy() 5846 * for lockless qdisc. 5847 */ 5848 clear_bit(__QDISC_STATE_SCHED, &q->state); 5849 continue; 5850 } 5851 5852 clear_bit(__QDISC_STATE_SCHED, &q->state); 5853 to_free = qdisc_run(q); 5854 if (root_lock) 5855 spin_unlock(root_lock); 5856 tcf_kfree_skb_list(to_free, q, NULL, qdisc_dev(q)); 5857 } 5858 5859 rcu_read_unlock(); 5860 } 5861 5862 xfrm_dev_backlog(sd); 5863 } 5864 5865 /** 5866 * netdev_is_rx_handler_busy - check if receive handler is registered 5867 * @dev: device to check 5868 * 5869 * Check if a receive handler is already registered for a given device. 5870 * Return true if there one. 5871 * 5872 * The caller must hold the rtnl_mutex. 5873 */ 5874 bool netdev_is_rx_handler_busy(struct net_device *dev) 5875 { 5876 ASSERT_RTNL(); 5877 return dev && rtnl_dereference(dev->rx_handler); 5878 } 5879 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy); 5880 5881 /** 5882 * netdev_rx_handler_register - register receive handler 5883 * @dev: device to register a handler for 5884 * @rx_handler: receive handler to register 5885 * @rx_handler_data: data pointer that is used by rx handler 5886 * 5887 * Register a receive handler for a device. This handler will then be 5888 * called from __netif_receive_skb. A negative errno code is returned 5889 * on a failure. 5890 * 5891 * The caller must hold the rtnl_mutex. 5892 * 5893 * For a general description of rx_handler, see enum rx_handler_result. 5894 */ 5895 int netdev_rx_handler_register(struct net_device *dev, 5896 rx_handler_func_t *rx_handler, 5897 void *rx_handler_data) 5898 { 5899 if (netdev_is_rx_handler_busy(dev)) 5900 return -EBUSY; 5901 5902 if (dev->priv_flags & IFF_NO_RX_HANDLER) 5903 return -EINVAL; 5904 5905 /* Note: rx_handler_data must be set before rx_handler */ 5906 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data); 5907 rcu_assign_pointer(dev->rx_handler, rx_handler); 5908 5909 return 0; 5910 } 5911 EXPORT_SYMBOL_GPL(netdev_rx_handler_register); 5912 5913 /** 5914 * netdev_rx_handler_unregister - unregister receive handler 5915 * @dev: device to unregister a handler from 5916 * 5917 * Unregister a receive handler from a device. 5918 * 5919 * The caller must hold the rtnl_mutex. 5920 */ 5921 void netdev_rx_handler_unregister(struct net_device *dev) 5922 { 5923 5924 ASSERT_RTNL(); 5925 RCU_INIT_POINTER(dev->rx_handler, NULL); 5926 /* a reader seeing a non NULL rx_handler in a rcu_read_lock() 5927 * section has a guarantee to see a non NULL rx_handler_data 5928 * as well. 5929 */ 5930 synchronize_net(); 5931 RCU_INIT_POINTER(dev->rx_handler_data, NULL); 5932 } 5933 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister); 5934 5935 /* 5936 * Limit the use of PFMEMALLOC reserves to those protocols that implement 5937 * the special handling of PFMEMALLOC skbs. 5938 */ 5939 static bool skb_pfmemalloc_protocol(struct sk_buff *skb) 5940 { 5941 switch (skb->protocol) { 5942 case htons(ETH_P_ARP): 5943 case htons(ETH_P_IP): 5944 case htons(ETH_P_IPV6): 5945 case htons(ETH_P_8021Q): 5946 case htons(ETH_P_8021AD): 5947 return true; 5948 default: 5949 return false; 5950 } 5951 } 5952 5953 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev, 5954 int *ret, struct net_device *orig_dev) 5955 { 5956 if (nf_hook_ingress_active(skb)) { 5957 int ingress_retval; 5958 5959 if (unlikely(*pt_prev)) { 5960 *ret = deliver_skb(skb, *pt_prev, orig_dev); 5961 *pt_prev = NULL; 5962 } 5963 5964 rcu_read_lock(); 5965 ingress_retval = nf_hook_ingress(skb); 5966 rcu_read_unlock(); 5967 return ingress_retval; 5968 } 5969 return 0; 5970 } 5971 5972 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc, 5973 struct packet_type **ppt_prev) 5974 { 5975 enum skb_drop_reason drop_reason = SKB_DROP_REASON_UNHANDLED_PROTO; 5976 struct packet_type *ptype, *pt_prev; 5977 rx_handler_func_t *rx_handler; 5978 struct sk_buff *skb = *pskb; 5979 struct net_device *orig_dev; 5980 bool deliver_exact = false; 5981 int ret = NET_RX_DROP; 5982 __be16 type; 5983 5984 net_timestamp_check(!READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5985 5986 trace_netif_receive_skb(skb); 5987 5988 orig_dev = skb->dev; 5989 5990 skb_reset_network_header(skb); 5991 #if !defined(CONFIG_DEBUG_NET) 5992 /* We plan to no longer reset the transport header here. 5993 * Give some time to fuzzers and dev build to catch bugs 5994 * in network stacks. 5995 */ 5996 if (!skb_transport_header_was_set(skb)) 5997 skb_reset_transport_header(skb); 5998 #endif 5999 skb_reset_mac_len(skb); 6000 6001 pt_prev = NULL; 6002 6003 another_round: 6004 skb->skb_iif = skb->dev->ifindex; 6005 6006 __this_cpu_inc(softnet_data.processed); 6007 6008 if (static_branch_unlikely(&generic_xdp_needed_key)) { 6009 int ret2; 6010 6011 migrate_disable(); 6012 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), 6013 &skb); 6014 migrate_enable(); 6015 6016 if (ret2 != XDP_PASS) { 6017 ret = NET_RX_DROP; 6018 goto out; 6019 } 6020 } 6021 6022 if (eth_type_vlan(skb->protocol)) { 6023 skb = skb_vlan_untag(skb); 6024 if (unlikely(!skb)) 6025 goto out; 6026 } 6027 6028 if (skb_skip_tc_classify(skb)) 6029 goto skip_classify; 6030 6031 if (pfmemalloc) 6032 goto skip_taps; 6033 6034 list_for_each_entry_rcu(ptype, &dev_net_rcu(skb->dev)->ptype_all, 6035 list) { 6036 if (unlikely(pt_prev)) 6037 ret = deliver_skb(skb, pt_prev, orig_dev); 6038 pt_prev = ptype; 6039 } 6040 6041 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) { 6042 if (unlikely(pt_prev)) 6043 ret = deliver_skb(skb, pt_prev, orig_dev); 6044 pt_prev = ptype; 6045 } 6046 6047 skip_taps: 6048 #ifdef CONFIG_NET_INGRESS 6049 if (static_branch_unlikely(&ingress_needed_key)) { 6050 bool another = false; 6051 6052 nf_skip_egress(skb, true); 6053 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev, 6054 &another); 6055 if (another) 6056 goto another_round; 6057 if (!skb) 6058 goto out; 6059 6060 nf_skip_egress(skb, false); 6061 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0) 6062 goto out; 6063 } 6064 #endif 6065 skb_reset_redirect(skb); 6066 skip_classify: 6067 if (pfmemalloc && !skb_pfmemalloc_protocol(skb)) { 6068 drop_reason = SKB_DROP_REASON_PFMEMALLOC; 6069 goto drop; 6070 } 6071 6072 if (skb_vlan_tag_present(skb)) { 6073 if (unlikely(pt_prev)) { 6074 ret = deliver_skb(skb, pt_prev, orig_dev); 6075 pt_prev = NULL; 6076 } 6077 if (vlan_do_receive(&skb)) 6078 goto another_round; 6079 else if (unlikely(!skb)) 6080 goto out; 6081 } 6082 6083 rx_handler = rcu_dereference(skb->dev->rx_handler); 6084 if (rx_handler) { 6085 if (unlikely(pt_prev)) { 6086 ret = deliver_skb(skb, pt_prev, orig_dev); 6087 pt_prev = NULL; 6088 } 6089 switch (rx_handler(&skb)) { 6090 case RX_HANDLER_CONSUMED: 6091 ret = NET_RX_SUCCESS; 6092 goto out; 6093 case RX_HANDLER_ANOTHER: 6094 goto another_round; 6095 case RX_HANDLER_EXACT: 6096 deliver_exact = true; 6097 break; 6098 case RX_HANDLER_PASS: 6099 break; 6100 default: 6101 BUG(); 6102 } 6103 } 6104 6105 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) { 6106 check_vlan_id: 6107 if (skb_vlan_tag_get_id(skb)) { 6108 /* Vlan id is non 0 and vlan_do_receive() above couldn't 6109 * find vlan device. 6110 */ 6111 skb->pkt_type = PACKET_OTHERHOST; 6112 } else if (eth_type_vlan(skb->protocol)) { 6113 /* Outer header is 802.1P with vlan 0, inner header is 6114 * 802.1Q or 802.1AD and vlan_do_receive() above could 6115 * not find vlan dev for vlan id 0. 6116 */ 6117 __vlan_hwaccel_clear_tag(skb); 6118 skb = skb_vlan_untag(skb); 6119 if (unlikely(!skb)) 6120 goto out; 6121 if (vlan_do_receive(&skb)) 6122 /* After stripping off 802.1P header with vlan 0 6123 * vlan dev is found for inner header. 6124 */ 6125 goto another_round; 6126 else if (unlikely(!skb)) 6127 goto out; 6128 else 6129 /* We have stripped outer 802.1P vlan 0 header. 6130 * But could not find vlan dev. 6131 * check again for vlan id to set OTHERHOST. 6132 */ 6133 goto check_vlan_id; 6134 } 6135 /* Note: we might in the future use prio bits 6136 * and set skb->priority like in vlan_do_receive() 6137 * For the time being, just ignore Priority Code Point 6138 */ 6139 __vlan_hwaccel_clear_tag(skb); 6140 } 6141 6142 type = skb->protocol; 6143 6144 /* deliver only exact match when indicated */ 6145 if (likely(!deliver_exact)) { 6146 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6147 &ptype_base[ntohs(type) & 6148 PTYPE_HASH_MASK]); 6149 6150 /* orig_dev and skb->dev could belong to different netns; 6151 * Even in such case we need to traverse only the list 6152 * coming from skb->dev, as the ptype owner (packet socket) 6153 * will use dev_net(skb->dev) to do namespace filtering. 6154 */ 6155 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6156 &dev_net_rcu(skb->dev)->ptype_specific); 6157 } 6158 6159 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6160 &orig_dev->ptype_specific); 6161 6162 if (unlikely(skb->dev != orig_dev)) { 6163 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6164 &skb->dev->ptype_specific); 6165 } 6166 6167 if (pt_prev) { 6168 *ppt_prev = pt_prev; 6169 } else { 6170 drop: 6171 if (!deliver_exact) 6172 dev_core_stats_rx_dropped_inc(skb->dev); 6173 else 6174 dev_core_stats_rx_nohandler_inc(skb->dev); 6175 6176 kfree_skb_reason(skb, drop_reason); 6177 /* Jamal, now you will not able to escape explaining 6178 * me how you were going to use this. :-) 6179 */ 6180 ret = NET_RX_DROP; 6181 } 6182 6183 out: 6184 /* The invariant here is that if *ppt_prev is not NULL 6185 * then skb should also be non-NULL. 6186 * 6187 * Apparently *ppt_prev assignment above holds this invariant due to 6188 * skb dereferencing near it. 6189 */ 6190 *pskb = skb; 6191 return ret; 6192 } 6193 6194 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc) 6195 { 6196 struct net_device *orig_dev = skb->dev; 6197 struct packet_type *pt_prev = NULL; 6198 int ret; 6199 6200 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 6201 if (pt_prev) 6202 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb, 6203 skb->dev, pt_prev, orig_dev); 6204 return ret; 6205 } 6206 6207 /** 6208 * netif_receive_skb_core - special purpose version of netif_receive_skb 6209 * @skb: buffer to process 6210 * 6211 * More direct receive version of netif_receive_skb(). It should 6212 * only be used by callers that have a need to skip RPS and Generic XDP. 6213 * Caller must also take care of handling if ``(page_is_)pfmemalloc``. 6214 * 6215 * This function may only be called from softirq context and interrupts 6216 * should be enabled. 6217 * 6218 * Return values (usually ignored): 6219 * NET_RX_SUCCESS: no congestion 6220 * NET_RX_DROP: packet was dropped 6221 */ 6222 int netif_receive_skb_core(struct sk_buff *skb) 6223 { 6224 int ret; 6225 6226 rcu_read_lock(); 6227 ret = __netif_receive_skb_one_core(skb, false); 6228 rcu_read_unlock(); 6229 6230 return ret; 6231 } 6232 EXPORT_SYMBOL(netif_receive_skb_core); 6233 6234 static inline void __netif_receive_skb_list_ptype(struct list_head *head, 6235 struct packet_type *pt_prev, 6236 struct net_device *orig_dev) 6237 { 6238 struct sk_buff *skb, *next; 6239 6240 if (!pt_prev) 6241 return; 6242 if (list_empty(head)) 6243 return; 6244 if (pt_prev->list_func != NULL) 6245 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv, 6246 ip_list_rcv, head, pt_prev, orig_dev); 6247 else 6248 list_for_each_entry_safe(skb, next, head, list) { 6249 skb_list_del_init(skb); 6250 pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 6251 } 6252 } 6253 6254 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc) 6255 { 6256 /* Fast-path assumptions: 6257 * - There is no RX handler. 6258 * - Only one packet_type matches. 6259 * If either of these fails, we will end up doing some per-packet 6260 * processing in-line, then handling the 'last ptype' for the whole 6261 * sublist. This can't cause out-of-order delivery to any single ptype, 6262 * because the 'last ptype' must be constant across the sublist, and all 6263 * other ptypes are handled per-packet. 6264 */ 6265 /* Current (common) ptype of sublist */ 6266 struct packet_type *pt_curr = NULL; 6267 /* Current (common) orig_dev of sublist */ 6268 struct net_device *od_curr = NULL; 6269 struct sk_buff *skb, *next; 6270 LIST_HEAD(sublist); 6271 6272 list_for_each_entry_safe(skb, next, head, list) { 6273 struct net_device *orig_dev = skb->dev; 6274 struct packet_type *pt_prev = NULL; 6275 6276 skb_list_del_init(skb); 6277 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 6278 if (!pt_prev) 6279 continue; 6280 if (pt_curr != pt_prev || od_curr != orig_dev) { 6281 /* dispatch old sublist */ 6282 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 6283 /* start new sublist */ 6284 INIT_LIST_HEAD(&sublist); 6285 pt_curr = pt_prev; 6286 od_curr = orig_dev; 6287 } 6288 list_add_tail(&skb->list, &sublist); 6289 } 6290 6291 /* dispatch final sublist */ 6292 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 6293 } 6294 6295 static int __netif_receive_skb(struct sk_buff *skb) 6296 { 6297 int ret; 6298 6299 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) { 6300 unsigned int noreclaim_flag; 6301 6302 /* 6303 * PFMEMALLOC skbs are special, they should 6304 * - be delivered to SOCK_MEMALLOC sockets only 6305 * - stay away from userspace 6306 * - have bounded memory usage 6307 * 6308 * Use PF_MEMALLOC as this saves us from propagating the allocation 6309 * context down to all allocation sites. 6310 */ 6311 noreclaim_flag = memalloc_noreclaim_save(); 6312 ret = __netif_receive_skb_one_core(skb, true); 6313 memalloc_noreclaim_restore(noreclaim_flag); 6314 } else 6315 ret = __netif_receive_skb_one_core(skb, false); 6316 6317 return ret; 6318 } 6319 6320 static void __netif_receive_skb_list(struct list_head *head) 6321 { 6322 unsigned long noreclaim_flag = 0; 6323 struct sk_buff *skb, *next; 6324 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */ 6325 6326 list_for_each_entry_safe(skb, next, head, list) { 6327 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) { 6328 struct list_head sublist; 6329 6330 /* Handle the previous sublist */ 6331 list_cut_before(&sublist, head, &skb->list); 6332 if (!list_empty(&sublist)) 6333 __netif_receive_skb_list_core(&sublist, pfmemalloc); 6334 pfmemalloc = !pfmemalloc; 6335 /* See comments in __netif_receive_skb */ 6336 if (pfmemalloc) 6337 noreclaim_flag = memalloc_noreclaim_save(); 6338 else 6339 memalloc_noreclaim_restore(noreclaim_flag); 6340 } 6341 } 6342 /* Handle the remaining sublist */ 6343 if (!list_empty(head)) 6344 __netif_receive_skb_list_core(head, pfmemalloc); 6345 /* Restore pflags */ 6346 if (pfmemalloc) 6347 memalloc_noreclaim_restore(noreclaim_flag); 6348 } 6349 6350 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp) 6351 { 6352 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog); 6353 struct bpf_prog *new = xdp->prog; 6354 int ret = 0; 6355 6356 switch (xdp->command) { 6357 case XDP_SETUP_PROG: 6358 rcu_assign_pointer(dev->xdp_prog, new); 6359 if (old) 6360 bpf_prog_put(old); 6361 6362 if (old && !new) { 6363 static_branch_dec(&generic_xdp_needed_key); 6364 } else if (new && !old) { 6365 static_branch_inc(&generic_xdp_needed_key); 6366 netif_disable_lro(dev); 6367 dev_disable_gro_hw(dev); 6368 } 6369 break; 6370 6371 default: 6372 ret = -EINVAL; 6373 break; 6374 } 6375 6376 return ret; 6377 } 6378 6379 static int netif_receive_skb_internal(struct sk_buff *skb) 6380 { 6381 int ret; 6382 6383 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 6384 6385 if (skb_defer_rx_timestamp(skb)) 6386 return NET_RX_SUCCESS; 6387 6388 rcu_read_lock(); 6389 #ifdef CONFIG_RPS 6390 if (static_branch_unlikely(&rps_needed)) { 6391 struct rps_dev_flow voidflow, *rflow = &voidflow; 6392 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6393 6394 if (cpu >= 0) { 6395 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6396 rcu_read_unlock(); 6397 return ret; 6398 } 6399 } 6400 #endif 6401 ret = __netif_receive_skb(skb); 6402 rcu_read_unlock(); 6403 return ret; 6404 } 6405 6406 void netif_receive_skb_list_internal(struct list_head *head) 6407 { 6408 struct sk_buff *skb, *next; 6409 LIST_HEAD(sublist); 6410 6411 list_for_each_entry_safe(skb, next, head, list) { 6412 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), 6413 skb); 6414 skb_list_del_init(skb); 6415 if (!skb_defer_rx_timestamp(skb)) 6416 list_add_tail(&skb->list, &sublist); 6417 } 6418 list_splice_init(&sublist, head); 6419 6420 rcu_read_lock(); 6421 #ifdef CONFIG_RPS 6422 if (static_branch_unlikely(&rps_needed)) { 6423 list_for_each_entry_safe(skb, next, head, list) { 6424 struct rps_dev_flow voidflow, *rflow = &voidflow; 6425 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6426 6427 if (cpu >= 0) { 6428 /* Will be handled, remove from list */ 6429 skb_list_del_init(skb); 6430 enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6431 } 6432 } 6433 } 6434 #endif 6435 __netif_receive_skb_list(head); 6436 rcu_read_unlock(); 6437 } 6438 6439 /** 6440 * netif_receive_skb - process receive buffer from network 6441 * @skb: buffer to process 6442 * 6443 * netif_receive_skb() is the main receive data processing function. 6444 * It always succeeds. The buffer may be dropped during processing 6445 * for congestion control or by the protocol layers. 6446 * 6447 * This function may only be called from softirq context and interrupts 6448 * should be enabled. 6449 * 6450 * Return values (usually ignored): 6451 * NET_RX_SUCCESS: no congestion 6452 * NET_RX_DROP: packet was dropped 6453 */ 6454 int netif_receive_skb(struct sk_buff *skb) 6455 { 6456 int ret; 6457 6458 trace_netif_receive_skb_entry(skb); 6459 6460 ret = netif_receive_skb_internal(skb); 6461 trace_netif_receive_skb_exit(ret); 6462 6463 return ret; 6464 } 6465 EXPORT_SYMBOL(netif_receive_skb); 6466 6467 /** 6468 * netif_receive_skb_list - process many receive buffers from network 6469 * @head: list of skbs to process. 6470 * 6471 * Since return value of netif_receive_skb() is normally ignored, and 6472 * wouldn't be meaningful for a list, this function returns void. 6473 * 6474 * This function may only be called from softirq context and interrupts 6475 * should be enabled. 6476 */ 6477 void netif_receive_skb_list(struct list_head *head) 6478 { 6479 struct sk_buff *skb; 6480 6481 if (list_empty(head)) 6482 return; 6483 if (trace_netif_receive_skb_list_entry_enabled()) { 6484 list_for_each_entry(skb, head, list) 6485 trace_netif_receive_skb_list_entry(skb); 6486 } 6487 netif_receive_skb_list_internal(head); 6488 trace_netif_receive_skb_list_exit(0); 6489 } 6490 EXPORT_SYMBOL(netif_receive_skb_list); 6491 6492 /* Network device is going away, flush any packets still pending */ 6493 static void flush_backlog(struct work_struct *work) 6494 { 6495 struct sk_buff *skb, *tmp; 6496 struct sk_buff_head list; 6497 struct softnet_data *sd; 6498 6499 __skb_queue_head_init(&list); 6500 local_bh_disable(); 6501 sd = this_cpu_ptr(&softnet_data); 6502 6503 backlog_lock_irq_disable(sd); 6504 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) { 6505 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6506 __skb_unlink(skb, &sd->input_pkt_queue); 6507 __skb_queue_tail(&list, skb); 6508 rps_input_queue_head_incr(sd); 6509 } 6510 } 6511 backlog_unlock_irq_enable(sd); 6512 6513 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6514 skb_queue_walk_safe(&sd->process_queue, skb, tmp) { 6515 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6516 __skb_unlink(skb, &sd->process_queue); 6517 __skb_queue_tail(&list, skb); 6518 rps_input_queue_head_incr(sd); 6519 } 6520 } 6521 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6522 local_bh_enable(); 6523 6524 __skb_queue_purge_reason(&list, SKB_DROP_REASON_DEV_READY); 6525 } 6526 6527 static bool flush_required(int cpu) 6528 { 6529 #if IS_ENABLED(CONFIG_RPS) 6530 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 6531 bool do_flush; 6532 6533 backlog_lock_irq_disable(sd); 6534 6535 /* as insertion into process_queue happens with the rps lock held, 6536 * process_queue access may race only with dequeue 6537 */ 6538 do_flush = !skb_queue_empty(&sd->input_pkt_queue) || 6539 !skb_queue_empty_lockless(&sd->process_queue); 6540 backlog_unlock_irq_enable(sd); 6541 6542 return do_flush; 6543 #endif 6544 /* without RPS we can't safely check input_pkt_queue: during a 6545 * concurrent remote skb_queue_splice() we can detect as empty both 6546 * input_pkt_queue and process_queue even if the latter could end-up 6547 * containing a lot of packets. 6548 */ 6549 return true; 6550 } 6551 6552 struct flush_backlogs { 6553 cpumask_t flush_cpus; 6554 struct work_struct w[]; 6555 }; 6556 6557 static struct flush_backlogs *flush_backlogs_alloc(void) 6558 { 6559 return kmalloc_flex(struct flush_backlogs, w, nr_cpu_ids); 6560 } 6561 6562 static struct flush_backlogs *flush_backlogs_fallback; 6563 static DEFINE_MUTEX(flush_backlogs_mutex); 6564 6565 static void flush_all_backlogs(void) 6566 { 6567 struct flush_backlogs *ptr = flush_backlogs_alloc(); 6568 unsigned int cpu; 6569 6570 if (!ptr) { 6571 mutex_lock(&flush_backlogs_mutex); 6572 ptr = flush_backlogs_fallback; 6573 } 6574 cpumask_clear(&ptr->flush_cpus); 6575 6576 cpus_read_lock(); 6577 6578 for_each_online_cpu(cpu) { 6579 if (flush_required(cpu)) { 6580 INIT_WORK(&ptr->w[cpu], flush_backlog); 6581 queue_work_on(cpu, system_highpri_wq, &ptr->w[cpu]); 6582 __cpumask_set_cpu(cpu, &ptr->flush_cpus); 6583 } 6584 } 6585 6586 /* we can have in flight packet[s] on the cpus we are not flushing, 6587 * synchronize_net() in unregister_netdevice_many() will take care of 6588 * them. 6589 */ 6590 for_each_cpu(cpu, &ptr->flush_cpus) 6591 flush_work(&ptr->w[cpu]); 6592 6593 cpus_read_unlock(); 6594 6595 if (ptr != flush_backlogs_fallback) 6596 kfree(ptr); 6597 else 6598 mutex_unlock(&flush_backlogs_mutex); 6599 } 6600 6601 static void net_rps_send_ipi(struct softnet_data *remsd) 6602 { 6603 #ifdef CONFIG_RPS 6604 while (remsd) { 6605 struct softnet_data *next = remsd->rps_ipi_next; 6606 6607 if (cpu_online(remsd->cpu)) 6608 smp_call_function_single_async(remsd->cpu, &remsd->csd); 6609 remsd = next; 6610 } 6611 #endif 6612 } 6613 6614 /* 6615 * net_rps_action_and_irq_enable sends any pending IPI's for rps. 6616 * Note: called with local irq disabled, but exits with local irq enabled. 6617 */ 6618 static void net_rps_action_and_irq_enable(struct softnet_data *sd) 6619 { 6620 #ifdef CONFIG_RPS 6621 struct softnet_data *remsd = sd->rps_ipi_list; 6622 6623 if (!use_backlog_threads() && remsd) { 6624 sd->rps_ipi_list = NULL; 6625 6626 local_irq_enable(); 6627 6628 /* Send pending IPI's to kick RPS processing on remote cpus. */ 6629 net_rps_send_ipi(remsd); 6630 } else 6631 #endif 6632 local_irq_enable(); 6633 } 6634 6635 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd) 6636 { 6637 #ifdef CONFIG_RPS 6638 return !use_backlog_threads() && sd->rps_ipi_list; 6639 #else 6640 return false; 6641 #endif 6642 } 6643 6644 static int process_backlog(struct napi_struct *napi, int quota) 6645 { 6646 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog); 6647 bool again = true; 6648 int work = 0; 6649 6650 /* Check if we have pending ipi, its better to send them now, 6651 * not waiting net_rx_action() end. 6652 */ 6653 if (sd_has_rps_ipi_waiting(sd)) { 6654 local_irq_disable(); 6655 net_rps_action_and_irq_enable(sd); 6656 } 6657 6658 napi->weight = READ_ONCE(net_hotdata.dev_rx_weight); 6659 while (again) { 6660 struct sk_buff *skb; 6661 6662 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6663 while ((skb = __skb_dequeue(&sd->process_queue))) { 6664 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6665 rcu_read_lock(); 6666 __netif_receive_skb(skb); 6667 rcu_read_unlock(); 6668 if (++work >= quota) { 6669 rps_input_queue_head_add(sd, work); 6670 return work; 6671 } 6672 6673 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6674 } 6675 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6676 6677 backlog_lock_irq_disable(sd); 6678 if (skb_queue_empty(&sd->input_pkt_queue)) { 6679 /* 6680 * Inline a custom version of __napi_complete(). 6681 * only current cpu owns and manipulates this napi, 6682 * and NAPI_STATE_SCHED is the only possible flag set 6683 * on backlog. 6684 * We can use a plain write instead of clear_bit(), 6685 * and we dont need an smp_mb() memory barrier. 6686 */ 6687 napi->state &= NAPIF_STATE_THREADED; 6688 again = false; 6689 } else { 6690 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6691 skb_queue_splice_tail_init(&sd->input_pkt_queue, 6692 &sd->process_queue); 6693 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6694 } 6695 backlog_unlock_irq_enable(sd); 6696 } 6697 6698 if (work) 6699 rps_input_queue_head_add(sd, work); 6700 return work; 6701 } 6702 6703 /** 6704 * __napi_schedule - schedule for receive 6705 * @n: entry to schedule 6706 * 6707 * The entry's receive function will be scheduled to run. 6708 * Consider using __napi_schedule_irqoff() if hard irqs are masked. 6709 */ 6710 void __napi_schedule(struct napi_struct *n) 6711 { 6712 unsigned long flags; 6713 6714 local_irq_save(flags); 6715 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6716 local_irq_restore(flags); 6717 } 6718 EXPORT_SYMBOL(__napi_schedule); 6719 6720 /** 6721 * napi_schedule_prep - check if napi can be scheduled 6722 * @n: napi context 6723 * 6724 * Test if NAPI routine is already running, and if not mark 6725 * it as running. This is used as a condition variable to 6726 * insure only one NAPI poll instance runs. We also make 6727 * sure there is no pending NAPI disable. 6728 */ 6729 bool napi_schedule_prep(struct napi_struct *n) 6730 { 6731 unsigned long new, val = READ_ONCE(n->state); 6732 6733 do { 6734 if (unlikely(val & NAPIF_STATE_DISABLE)) 6735 return false; 6736 new = val | NAPIF_STATE_SCHED; 6737 6738 /* Sets STATE_MISSED bit if STATE_SCHED was already set 6739 * This was suggested by Alexander Duyck, as compiler 6740 * emits better code than : 6741 * if (val & NAPIF_STATE_SCHED) 6742 * new |= NAPIF_STATE_MISSED; 6743 */ 6744 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED * 6745 NAPIF_STATE_MISSED; 6746 } while (!try_cmpxchg(&n->state, &val, new)); 6747 6748 return !(val & NAPIF_STATE_SCHED); 6749 } 6750 EXPORT_SYMBOL(napi_schedule_prep); 6751 6752 /** 6753 * __napi_schedule_irqoff - schedule for receive 6754 * @n: entry to schedule 6755 * 6756 * Variant of __napi_schedule() assuming hard irqs are masked. 6757 * 6758 * On PREEMPT_RT enabled kernels this maps to __napi_schedule() 6759 * because the interrupt disabled assumption might not be true 6760 * due to force-threaded interrupts and spinlock substitution. 6761 */ 6762 void __napi_schedule_irqoff(struct napi_struct *n) 6763 { 6764 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6765 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6766 else 6767 __napi_schedule(n); 6768 } 6769 EXPORT_SYMBOL(__napi_schedule_irqoff); 6770 6771 bool napi_complete_done(struct napi_struct *n, int work_done) 6772 { 6773 unsigned long flags, val, new, timeout = 0; 6774 bool ret = true; 6775 6776 /* 6777 * 1) Don't let napi dequeue from the cpu poll list 6778 * just in case its running on a different cpu. 6779 * 2) If we are busy polling, do nothing here, we have 6780 * the guarantee we will be called later. 6781 */ 6782 if (unlikely(n->state & (NAPIF_STATE_NPSVC | 6783 NAPIF_STATE_IN_BUSY_POLL))) 6784 return false; 6785 6786 if (work_done) { 6787 if (n->gro.bitmask) 6788 timeout = napi_get_gro_flush_timeout(n); 6789 n->defer_hard_irqs_count = napi_get_defer_hard_irqs(n); 6790 } 6791 if (n->defer_hard_irqs_count > 0) { 6792 n->defer_hard_irqs_count--; 6793 timeout = napi_get_gro_flush_timeout(n); 6794 if (timeout) 6795 ret = false; 6796 } 6797 6798 /* 6799 * When the NAPI instance uses a timeout and keeps postponing 6800 * it, we need to bound somehow the time packets are kept in 6801 * the GRO layer. 6802 */ 6803 gro_flush_normal(&n->gro, !!timeout); 6804 6805 if (unlikely(!list_empty(&n->poll_list))) { 6806 /* If n->poll_list is not empty, we need to mask irqs */ 6807 local_irq_save(flags); 6808 list_del_init(&n->poll_list); 6809 local_irq_restore(flags); 6810 } 6811 WRITE_ONCE(n->list_owner, -1); 6812 6813 val = READ_ONCE(n->state); 6814 do { 6815 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED)); 6816 6817 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED | 6818 NAPIF_STATE_SCHED_THREADED | 6819 NAPIF_STATE_PREFER_BUSY_POLL); 6820 6821 /* If STATE_MISSED was set, leave STATE_SCHED set, 6822 * because we will call napi->poll() one more time. 6823 * This C code was suggested by Alexander Duyck to help gcc. 6824 */ 6825 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED * 6826 NAPIF_STATE_SCHED; 6827 } while (!try_cmpxchg(&n->state, &val, new)); 6828 6829 if (unlikely(val & NAPIF_STATE_MISSED)) { 6830 __napi_schedule(n); 6831 return false; 6832 } 6833 6834 if (timeout) 6835 hrtimer_start(&n->timer, ns_to_ktime(timeout), 6836 HRTIMER_MODE_REL_PINNED); 6837 return ret; 6838 } 6839 EXPORT_SYMBOL(napi_complete_done); 6840 6841 static void skb_defer_free_flush(void) 6842 { 6843 struct llist_node *free_list; 6844 struct sk_buff *skb, *next; 6845 struct skb_defer_node *sdn; 6846 int node; 6847 6848 for_each_node(node) { 6849 sdn = this_cpu_ptr(net_hotdata.skb_defer_nodes) + node; 6850 6851 if (llist_empty(&sdn->defer_list)) 6852 continue; 6853 atomic_long_set(&sdn->defer_count, 0); 6854 free_list = llist_del_all(&sdn->defer_list); 6855 6856 llist_for_each_entry_safe(skb, next, free_list, ll_node) { 6857 prefetch(next); 6858 napi_consume_skb(skb, 1); 6859 } 6860 } 6861 } 6862 6863 #if defined(CONFIG_NET_RX_BUSY_POLL) 6864 6865 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule) 6866 { 6867 if (!skip_schedule) { 6868 gro_normal_list(&napi->gro); 6869 __napi_schedule(napi); 6870 return; 6871 } 6872 6873 /* Flush too old packets. If HZ < 1000, flush all packets */ 6874 gro_flush_normal(&napi->gro, HZ >= 1000); 6875 6876 clear_bit(NAPI_STATE_SCHED, &napi->state); 6877 } 6878 6879 enum { 6880 NAPI_F_PREFER_BUSY_POLL = 1, 6881 NAPI_F_END_ON_RESCHED = 2, 6882 }; 6883 6884 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, 6885 unsigned flags, u16 budget) 6886 { 6887 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6888 bool skip_schedule = false; 6889 unsigned long timeout; 6890 int rc; 6891 6892 /* Busy polling means there is a high chance device driver hard irq 6893 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was 6894 * set in napi_schedule_prep(). 6895 * Since we are about to call napi->poll() once more, we can safely 6896 * clear NAPI_STATE_MISSED. 6897 * 6898 * Note: x86 could use a single "lock and ..." instruction 6899 * to perform these two clear_bit() 6900 */ 6901 clear_bit(NAPI_STATE_MISSED, &napi->state); 6902 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state); 6903 6904 local_bh_disable(); 6905 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6906 6907 if (flags & NAPI_F_PREFER_BUSY_POLL) { 6908 napi->defer_hard_irqs_count = napi_get_defer_hard_irqs(napi); 6909 timeout = napi_get_gro_flush_timeout(napi); 6910 if (napi->defer_hard_irqs_count && timeout) { 6911 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED); 6912 skip_schedule = true; 6913 } 6914 } 6915 6916 /* All we really want here is to re-enable device interrupts. 6917 * Ideally, a new ndo_busy_poll_stop() could avoid another round. 6918 */ 6919 rc = napi->poll(napi, budget); 6920 /* We can't gro_normal_list() here, because napi->poll() might have 6921 * rearmed the napi (napi_complete_done()) in which case it could 6922 * already be running on another CPU. 6923 */ 6924 trace_napi_poll(napi, rc, budget); 6925 netpoll_poll_unlock(have_poll_lock); 6926 if (rc == budget) 6927 __busy_poll_stop(napi, skip_schedule); 6928 bpf_net_ctx_clear(bpf_net_ctx); 6929 local_bh_enable(); 6930 } 6931 6932 static void __napi_busy_loop(unsigned int napi_id, 6933 bool (*loop_end)(void *, unsigned long), 6934 void *loop_end_arg, unsigned flags, u16 budget) 6935 { 6936 unsigned long start_time = loop_end ? busy_loop_current_time() : 0; 6937 int (*napi_poll)(struct napi_struct *napi, int budget); 6938 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6939 void *have_poll_lock = NULL; 6940 struct napi_struct *napi; 6941 6942 WARN_ON_ONCE(!rcu_read_lock_held()); 6943 6944 restart: 6945 napi_poll = NULL; 6946 6947 napi = napi_by_id(napi_id); 6948 if (!napi) 6949 return; 6950 6951 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6952 preempt_disable(); 6953 for (;;) { 6954 int work = 0; 6955 6956 local_bh_disable(); 6957 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6958 if (!napi_poll) { 6959 unsigned long val = READ_ONCE(napi->state); 6960 6961 /* If multiple threads are competing for this napi, 6962 * we avoid dirtying napi->state as much as we can. 6963 */ 6964 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED | 6965 NAPIF_STATE_IN_BUSY_POLL)) { 6966 if (flags & NAPI_F_PREFER_BUSY_POLL) 6967 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6968 goto count; 6969 } 6970 if (cmpxchg(&napi->state, val, 6971 val | NAPIF_STATE_IN_BUSY_POLL | 6972 NAPIF_STATE_SCHED) != val) { 6973 if (flags & NAPI_F_PREFER_BUSY_POLL) 6974 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6975 goto count; 6976 } 6977 have_poll_lock = netpoll_poll_lock(napi); 6978 napi_poll = napi->poll; 6979 } 6980 work = napi_poll(napi, budget); 6981 trace_napi_poll(napi, work, budget); 6982 gro_normal_list(&napi->gro); 6983 count: 6984 if (work > 0) 6985 __NET_ADD_STATS(dev_net(napi->dev), 6986 LINUX_MIB_BUSYPOLLRXPACKETS, work); 6987 skb_defer_free_flush(); 6988 bpf_net_ctx_clear(bpf_net_ctx); 6989 local_bh_enable(); 6990 6991 if (!loop_end || loop_end(loop_end_arg, start_time)) 6992 break; 6993 6994 if (unlikely(need_resched())) { 6995 if (flags & NAPI_F_END_ON_RESCHED) 6996 break; 6997 if (napi_poll) 6998 busy_poll_stop(napi, have_poll_lock, flags, budget); 6999 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 7000 preempt_enable(); 7001 rcu_read_unlock(); 7002 cond_resched(); 7003 rcu_read_lock(); 7004 if (loop_end(loop_end_arg, start_time)) 7005 return; 7006 goto restart; 7007 } 7008 cpu_relax(); 7009 } 7010 if (napi_poll) 7011 busy_poll_stop(napi, have_poll_lock, flags, budget); 7012 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 7013 preempt_enable(); 7014 } 7015 7016 void napi_busy_loop_rcu(unsigned int napi_id, 7017 bool (*loop_end)(void *, unsigned long), 7018 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 7019 { 7020 unsigned flags = NAPI_F_END_ON_RESCHED; 7021 7022 if (prefer_busy_poll) 7023 flags |= NAPI_F_PREFER_BUSY_POLL; 7024 7025 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 7026 } 7027 7028 void napi_busy_loop(unsigned int napi_id, 7029 bool (*loop_end)(void *, unsigned long), 7030 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 7031 { 7032 unsigned flags = prefer_busy_poll ? NAPI_F_PREFER_BUSY_POLL : 0; 7033 7034 rcu_read_lock(); 7035 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 7036 rcu_read_unlock(); 7037 } 7038 EXPORT_SYMBOL(napi_busy_loop); 7039 7040 void napi_suspend_irqs(unsigned int napi_id) 7041 { 7042 struct napi_struct *napi; 7043 7044 rcu_read_lock(); 7045 napi = napi_by_id(napi_id); 7046 if (napi) { 7047 unsigned long timeout = napi_get_irq_suspend_timeout(napi); 7048 7049 if (timeout) 7050 hrtimer_start(&napi->timer, ns_to_ktime(timeout), 7051 HRTIMER_MODE_REL_PINNED); 7052 } 7053 rcu_read_unlock(); 7054 } 7055 7056 void napi_resume_irqs(unsigned int napi_id) 7057 { 7058 struct napi_struct *napi; 7059 7060 rcu_read_lock(); 7061 napi = napi_by_id(napi_id); 7062 if (napi) { 7063 /* If irq_suspend_timeout is set to 0 between the call to 7064 * napi_suspend_irqs and now, the original value still 7065 * determines the safety timeout as intended and napi_watchdog 7066 * will resume irq processing. 7067 */ 7068 if (napi_get_irq_suspend_timeout(napi)) { 7069 local_bh_disable(); 7070 napi_schedule(napi); 7071 local_bh_enable(); 7072 } 7073 } 7074 rcu_read_unlock(); 7075 } 7076 7077 #endif /* CONFIG_NET_RX_BUSY_POLL */ 7078 7079 static void __napi_hash_add_with_id(struct napi_struct *napi, 7080 unsigned int napi_id) 7081 { 7082 napi->gro.cached_napi_id = napi_id; 7083 7084 WRITE_ONCE(napi->napi_id, napi_id); 7085 hlist_add_head_rcu(&napi->napi_hash_node, 7086 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]); 7087 } 7088 7089 static void napi_hash_add_with_id(struct napi_struct *napi, 7090 unsigned int napi_id) 7091 { 7092 unsigned long flags; 7093 7094 spin_lock_irqsave(&napi_hash_lock, flags); 7095 WARN_ON_ONCE(napi_by_id(napi_id)); 7096 __napi_hash_add_with_id(napi, napi_id); 7097 spin_unlock_irqrestore(&napi_hash_lock, flags); 7098 } 7099 7100 static void napi_hash_add(struct napi_struct *napi) 7101 { 7102 unsigned long flags; 7103 7104 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state)) 7105 return; 7106 7107 spin_lock_irqsave(&napi_hash_lock, flags); 7108 7109 /* 0..NR_CPUS range is reserved for sender_cpu use */ 7110 do { 7111 if (unlikely(!napi_id_valid(++napi_gen_id))) 7112 napi_gen_id = MIN_NAPI_ID; 7113 } while (napi_by_id(napi_gen_id)); 7114 7115 __napi_hash_add_with_id(napi, napi_gen_id); 7116 7117 spin_unlock_irqrestore(&napi_hash_lock, flags); 7118 } 7119 7120 /* Warning : caller is responsible to make sure rcu grace period 7121 * is respected before freeing memory containing @napi 7122 */ 7123 static void napi_hash_del(struct napi_struct *napi) 7124 { 7125 unsigned long flags; 7126 7127 spin_lock_irqsave(&napi_hash_lock, flags); 7128 7129 hlist_del_init_rcu(&napi->napi_hash_node); 7130 7131 spin_unlock_irqrestore(&napi_hash_lock, flags); 7132 } 7133 7134 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer) 7135 { 7136 struct napi_struct *napi; 7137 7138 napi = container_of(timer, struct napi_struct, timer); 7139 7140 /* Note : we use a relaxed variant of napi_schedule_prep() not setting 7141 * NAPI_STATE_MISSED, since we do not react to a device IRQ. 7142 */ 7143 if (!napi_disable_pending(napi) && 7144 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) { 7145 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 7146 __napi_schedule_irqoff(napi); 7147 } 7148 7149 return HRTIMER_NORESTART; 7150 } 7151 7152 static void napi_stop_kthread(struct napi_struct *napi) 7153 { 7154 unsigned long val, new; 7155 7156 /* Wait until the napi STATE_THREADED is unset. */ 7157 while (true) { 7158 val = READ_ONCE(napi->state); 7159 7160 /* If napi kthread own this napi or the napi is idle, 7161 * STATE_THREADED can be unset here. 7162 */ 7163 if ((val & NAPIF_STATE_SCHED_THREADED) || 7164 !(val & NAPIF_STATE_SCHED)) { 7165 new = val & (~(NAPIF_STATE_THREADED | 7166 NAPIF_STATE_THREADED_BUSY_POLL)); 7167 } else { 7168 msleep(20); 7169 continue; 7170 } 7171 7172 if (try_cmpxchg(&napi->state, &val, new)) 7173 break; 7174 } 7175 7176 /* Once STATE_THREADED is unset, wait for SCHED_THREADED to be unset by 7177 * the kthread. 7178 */ 7179 while (true) { 7180 if (!test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) 7181 break; 7182 7183 msleep(20); 7184 } 7185 7186 kthread_stop(napi->thread); 7187 napi->thread = NULL; 7188 } 7189 7190 static void napi_set_threaded_state(struct napi_struct *napi, 7191 enum netdev_napi_threaded threaded_mode) 7192 { 7193 bool threaded = threaded_mode != NETDEV_NAPI_THREADED_DISABLED; 7194 bool busy_poll = threaded_mode == NETDEV_NAPI_THREADED_BUSY_POLL; 7195 7196 assign_bit(NAPI_STATE_THREADED, &napi->state, threaded); 7197 assign_bit(NAPI_STATE_THREADED_BUSY_POLL, &napi->state, busy_poll); 7198 } 7199 7200 int napi_set_threaded(struct napi_struct *napi, 7201 enum netdev_napi_threaded threaded) 7202 { 7203 if (threaded) { 7204 if (!napi->thread) { 7205 int err = napi_kthread_create(napi); 7206 7207 if (err) 7208 return err; 7209 } 7210 } 7211 7212 if (napi->config) 7213 napi->config->threaded = threaded; 7214 7215 /* Setting/unsetting threaded mode on a napi might not immediately 7216 * take effect, if the current napi instance is actively being 7217 * polled. In this case, the switch between threaded mode and 7218 * softirq mode will happen in the next round of napi_schedule(). 7219 * This should not cause hiccups/stalls to the live traffic. 7220 */ 7221 if (!threaded && napi->thread) { 7222 napi_stop_kthread(napi); 7223 } else { 7224 /* Make sure kthread is created before THREADED bit is set. */ 7225 smp_mb__before_atomic(); 7226 napi_set_threaded_state(napi, threaded); 7227 } 7228 7229 return 0; 7230 } 7231 7232 int netif_set_threaded(struct net_device *dev, 7233 enum netdev_napi_threaded threaded) 7234 { 7235 struct napi_struct *napi; 7236 int i, err = 0; 7237 7238 netdev_assert_locked_or_invisible(dev); 7239 7240 if (threaded) { 7241 list_for_each_entry(napi, &dev->napi_list, dev_list) { 7242 if (!napi->thread) { 7243 err = napi_kthread_create(napi); 7244 if (err) { 7245 threaded = NETDEV_NAPI_THREADED_DISABLED; 7246 break; 7247 } 7248 } 7249 } 7250 } 7251 7252 WRITE_ONCE(dev->threaded, threaded); 7253 7254 /* The error should not occur as the kthreads are already created. */ 7255 list_for_each_entry(napi, &dev->napi_list, dev_list) 7256 WARN_ON_ONCE(napi_set_threaded(napi, threaded)); 7257 7258 /* Override the config for all NAPIs even if currently not listed */ 7259 for (i = 0; i < dev->num_napi_configs; i++) 7260 dev->napi_config[i].threaded = threaded; 7261 7262 return err; 7263 } 7264 7265 /** 7266 * netif_threaded_enable() - enable threaded NAPIs 7267 * @dev: net_device instance 7268 * 7269 * Enable threaded mode for the NAPI instances of the device. This may be useful 7270 * for devices where multiple NAPI instances get scheduled by a single 7271 * interrupt. Threaded NAPI allows moving the NAPI processing to cores other 7272 * than the core where IRQ is mapped. 7273 * 7274 * This function should be called before @dev is registered. 7275 */ 7276 void netif_threaded_enable(struct net_device *dev) 7277 { 7278 WARN_ON_ONCE(netif_set_threaded(dev, NETDEV_NAPI_THREADED_ENABLED)); 7279 } 7280 EXPORT_SYMBOL(netif_threaded_enable); 7281 7282 /** 7283 * netif_queue_set_napi - Associate queue with the napi 7284 * @dev: device to which NAPI and queue belong 7285 * @queue_index: Index of queue 7286 * @type: queue type as RX or TX 7287 * @napi: NAPI context, pass NULL to clear previously set NAPI 7288 * 7289 * Set queue with its corresponding napi context. This should be done after 7290 * registering the NAPI handler for the queue-vector and the queues have been 7291 * mapped to the corresponding interrupt vector. 7292 */ 7293 void netif_queue_set_napi(struct net_device *dev, unsigned int queue_index, 7294 enum netdev_queue_type type, struct napi_struct *napi) 7295 { 7296 struct netdev_rx_queue *rxq; 7297 struct netdev_queue *txq; 7298 7299 if (WARN_ON_ONCE(napi && !napi->dev)) 7300 return; 7301 netdev_ops_assert_locked_or_invisible(dev); 7302 7303 switch (type) { 7304 case NETDEV_QUEUE_TYPE_RX: 7305 rxq = __netif_get_rx_queue(dev, queue_index); 7306 rxq->napi = napi; 7307 return; 7308 case NETDEV_QUEUE_TYPE_TX: 7309 txq = netdev_get_tx_queue(dev, queue_index); 7310 txq->napi = napi; 7311 return; 7312 default: 7313 return; 7314 } 7315 } 7316 EXPORT_SYMBOL(netif_queue_set_napi); 7317 7318 static void 7319 netif_napi_irq_notify(struct irq_affinity_notify *notify, 7320 const cpumask_t *mask) 7321 { 7322 struct napi_struct *napi = 7323 container_of(notify, struct napi_struct, notify); 7324 #ifdef CONFIG_RFS_ACCEL 7325 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 7326 int err; 7327 #endif 7328 7329 if (napi->config && napi->dev->irq_affinity_auto) 7330 cpumask_copy(&napi->config->affinity_mask, mask); 7331 7332 #ifdef CONFIG_RFS_ACCEL 7333 if (napi->dev->rx_cpu_rmap_auto) { 7334 err = cpu_rmap_update(rmap, napi->napi_rmap_idx, mask); 7335 if (err) 7336 netdev_warn(napi->dev, "RMAP update failed (%d)\n", 7337 err); 7338 } 7339 #endif 7340 } 7341 7342 #ifdef CONFIG_RFS_ACCEL 7343 static void netif_napi_affinity_release(struct kref *ref) 7344 { 7345 struct napi_struct *napi = 7346 container_of(ref, struct napi_struct, notify.kref); 7347 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 7348 7349 netdev_assert_locked(napi->dev); 7350 WARN_ON(test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, 7351 &napi->state)); 7352 7353 if (!napi->dev->rx_cpu_rmap_auto) 7354 return; 7355 rmap->obj[napi->napi_rmap_idx] = NULL; 7356 napi->napi_rmap_idx = -1; 7357 cpu_rmap_put(rmap); 7358 } 7359 7360 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 7361 { 7362 if (dev->rx_cpu_rmap_auto) 7363 return 0; 7364 7365 dev->rx_cpu_rmap = alloc_irq_cpu_rmap(num_irqs); 7366 if (!dev->rx_cpu_rmap) 7367 return -ENOMEM; 7368 7369 dev->rx_cpu_rmap_auto = true; 7370 return 0; 7371 } 7372 EXPORT_SYMBOL(netif_enable_cpu_rmap); 7373 7374 static void netif_del_cpu_rmap(struct net_device *dev) 7375 { 7376 struct cpu_rmap *rmap = dev->rx_cpu_rmap; 7377 7378 if (!dev->rx_cpu_rmap_auto) 7379 return; 7380 7381 /* Free the rmap */ 7382 cpu_rmap_put(rmap); 7383 dev->rx_cpu_rmap = NULL; 7384 dev->rx_cpu_rmap_auto = false; 7385 } 7386 7387 #else 7388 static void netif_napi_affinity_release(struct kref *ref) 7389 { 7390 } 7391 7392 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 7393 { 7394 return 0; 7395 } 7396 EXPORT_SYMBOL(netif_enable_cpu_rmap); 7397 7398 static void netif_del_cpu_rmap(struct net_device *dev) 7399 { 7400 } 7401 #endif 7402 7403 void netif_set_affinity_auto(struct net_device *dev) 7404 { 7405 unsigned int i, maxqs, numa; 7406 7407 maxqs = max(dev->num_tx_queues, dev->num_rx_queues); 7408 numa = dev_to_node(&dev->dev); 7409 7410 for (i = 0; i < maxqs; i++) 7411 cpumask_set_cpu(cpumask_local_spread(i, numa), 7412 &dev->napi_config[i].affinity_mask); 7413 7414 dev->irq_affinity_auto = true; 7415 } 7416 EXPORT_SYMBOL(netif_set_affinity_auto); 7417 7418 void netif_napi_set_irq_locked(struct napi_struct *napi, int irq) 7419 { 7420 int rc; 7421 7422 netdev_assert_locked_or_invisible(napi->dev); 7423 7424 if (napi->irq == irq) 7425 return; 7426 7427 /* Remove existing resources */ 7428 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7429 irq_set_affinity_notifier(napi->irq, NULL); 7430 7431 napi->irq = irq; 7432 if (irq < 0 || 7433 (!napi->dev->rx_cpu_rmap_auto && !napi->dev->irq_affinity_auto)) 7434 return; 7435 7436 /* Abort for buggy drivers */ 7437 if (napi->dev->irq_affinity_auto && WARN_ON_ONCE(!napi->config)) 7438 return; 7439 7440 #ifdef CONFIG_RFS_ACCEL 7441 if (napi->dev->rx_cpu_rmap_auto) { 7442 rc = cpu_rmap_add(napi->dev->rx_cpu_rmap, napi); 7443 if (rc < 0) 7444 return; 7445 7446 cpu_rmap_get(napi->dev->rx_cpu_rmap); 7447 napi->napi_rmap_idx = rc; 7448 } 7449 #endif 7450 7451 /* Use core IRQ notifier */ 7452 napi->notify.notify = netif_napi_irq_notify; 7453 napi->notify.release = netif_napi_affinity_release; 7454 rc = irq_set_affinity_notifier(irq, &napi->notify); 7455 if (rc) { 7456 netdev_warn(napi->dev, "Unable to set IRQ notifier (%d)\n", 7457 rc); 7458 goto put_rmap; 7459 } 7460 7461 set_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state); 7462 return; 7463 7464 put_rmap: 7465 #ifdef CONFIG_RFS_ACCEL 7466 if (napi->dev->rx_cpu_rmap_auto) { 7467 napi->dev->rx_cpu_rmap->obj[napi->napi_rmap_idx] = NULL; 7468 cpu_rmap_put(napi->dev->rx_cpu_rmap); 7469 napi->napi_rmap_idx = -1; 7470 } 7471 #endif 7472 napi->notify.notify = NULL; 7473 napi->notify.release = NULL; 7474 } 7475 EXPORT_SYMBOL(netif_napi_set_irq_locked); 7476 7477 static void napi_restore_config(struct napi_struct *n) 7478 { 7479 n->defer_hard_irqs = n->config->defer_hard_irqs; 7480 n->gro_flush_timeout = n->config->gro_flush_timeout; 7481 n->irq_suspend_timeout = n->config->irq_suspend_timeout; 7482 7483 if (n->dev->irq_affinity_auto && 7484 test_bit(NAPI_STATE_HAS_NOTIFIER, &n->state)) 7485 irq_set_affinity(n->irq, &n->config->affinity_mask); 7486 7487 /* a NAPI ID might be stored in the config, if so use it. if not, use 7488 * napi_hash_add to generate one for us. 7489 */ 7490 if (n->config->napi_id) { 7491 napi_hash_add_with_id(n, n->config->napi_id); 7492 } else { 7493 napi_hash_add(n); 7494 n->config->napi_id = n->napi_id; 7495 } 7496 7497 WARN_ON_ONCE(napi_set_threaded(n, n->config->threaded)); 7498 } 7499 7500 static void napi_save_config(struct napi_struct *n) 7501 { 7502 n->config->defer_hard_irqs = n->defer_hard_irqs; 7503 n->config->gro_flush_timeout = n->gro_flush_timeout; 7504 n->config->irq_suspend_timeout = n->irq_suspend_timeout; 7505 napi_hash_del(n); 7506 } 7507 7508 /* Netlink wants the NAPI list to be sorted by ID, if adding a NAPI which will 7509 * inherit an existing ID try to insert it at the right position. 7510 */ 7511 static void 7512 netif_napi_dev_list_add(struct net_device *dev, struct napi_struct *napi) 7513 { 7514 unsigned int new_id, pos_id; 7515 struct list_head *higher; 7516 struct napi_struct *pos; 7517 7518 new_id = UINT_MAX; 7519 if (napi->config && napi->config->napi_id) 7520 new_id = napi->config->napi_id; 7521 7522 higher = &dev->napi_list; 7523 list_for_each_entry(pos, &dev->napi_list, dev_list) { 7524 if (napi_id_valid(pos->napi_id)) 7525 pos_id = pos->napi_id; 7526 else if (pos->config) 7527 pos_id = pos->config->napi_id; 7528 else 7529 pos_id = UINT_MAX; 7530 7531 if (pos_id <= new_id) 7532 break; 7533 higher = &pos->dev_list; 7534 } 7535 list_add_rcu(&napi->dev_list, higher); /* adds after higher */ 7536 } 7537 7538 /* Double check that napi_get_frags() allocates skbs with 7539 * skb->head being backed by slab, not a page fragment. 7540 * This is to make sure bug fixed in 3226b158e67c 7541 * ("net: avoid 32 x truesize under-estimation for tiny skbs") 7542 * does not accidentally come back. 7543 */ 7544 static void napi_get_frags_check(struct napi_struct *napi) 7545 { 7546 struct sk_buff *skb; 7547 7548 local_bh_disable(); 7549 skb = napi_get_frags(napi); 7550 WARN_ON_ONCE(skb && skb->head_frag); 7551 napi_free_frags(napi); 7552 local_bh_enable(); 7553 } 7554 7555 void netif_napi_add_weight_locked(struct net_device *dev, 7556 struct napi_struct *napi, 7557 int (*poll)(struct napi_struct *, int), 7558 int weight) 7559 { 7560 netdev_assert_locked(dev); 7561 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state))) 7562 return; 7563 7564 INIT_LIST_HEAD(&napi->poll_list); 7565 INIT_HLIST_NODE(&napi->napi_hash_node); 7566 hrtimer_setup(&napi->timer, napi_watchdog, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); 7567 gro_init(&napi->gro); 7568 napi->skb = NULL; 7569 napi->poll = poll; 7570 if (weight > NAPI_POLL_WEIGHT) 7571 netdev_err_once(dev, "%s() called with weight %d\n", __func__, 7572 weight); 7573 napi->weight = weight; 7574 napi->dev = dev; 7575 #ifdef CONFIG_NETPOLL 7576 napi->poll_owner = -1; 7577 #endif 7578 napi->list_owner = -1; 7579 set_bit(NAPI_STATE_SCHED, &napi->state); 7580 set_bit(NAPI_STATE_NPSVC, &napi->state); 7581 netif_napi_dev_list_add(dev, napi); 7582 7583 /* default settings from sysfs are applied to all NAPIs. any per-NAPI 7584 * configuration will be loaded in napi_enable 7585 */ 7586 napi_set_defer_hard_irqs(napi, READ_ONCE(dev->napi_defer_hard_irqs)); 7587 napi_set_gro_flush_timeout(napi, READ_ONCE(dev->gro_flush_timeout)); 7588 7589 napi_get_frags_check(napi); 7590 /* Create kthread for this napi if dev->threaded is set. 7591 * Clear dev->threaded if kthread creation failed so that 7592 * threaded mode will not be enabled in napi_enable(). 7593 */ 7594 if (napi_get_threaded_config(dev, napi)) 7595 if (napi_kthread_create(napi)) 7596 dev->threaded = NETDEV_NAPI_THREADED_DISABLED; 7597 netif_napi_set_irq_locked(napi, -1); 7598 } 7599 EXPORT_SYMBOL(netif_napi_add_weight_locked); 7600 7601 void napi_disable_locked(struct napi_struct *n) 7602 { 7603 unsigned long val, new; 7604 7605 might_sleep(); 7606 netdev_assert_locked(n->dev); 7607 7608 set_bit(NAPI_STATE_DISABLE, &n->state); 7609 7610 val = READ_ONCE(n->state); 7611 do { 7612 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) { 7613 usleep_range(20, 200); 7614 val = READ_ONCE(n->state); 7615 } 7616 7617 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC; 7618 new &= ~(NAPIF_STATE_THREADED | 7619 NAPIF_STATE_THREADED_BUSY_POLL | 7620 NAPIF_STATE_PREFER_BUSY_POLL); 7621 } while (!try_cmpxchg(&n->state, &val, new)); 7622 7623 hrtimer_cancel(&n->timer); 7624 7625 if (n->config) 7626 napi_save_config(n); 7627 else 7628 napi_hash_del(n); 7629 7630 clear_bit(NAPI_STATE_DISABLE, &n->state); 7631 } 7632 EXPORT_SYMBOL(napi_disable_locked); 7633 7634 /** 7635 * napi_disable() - prevent NAPI from scheduling 7636 * @n: NAPI context 7637 * 7638 * Stop NAPI from being scheduled on this context. 7639 * Waits till any outstanding processing completes. 7640 * Takes netdev_lock() for associated net_device. 7641 */ 7642 void napi_disable(struct napi_struct *n) 7643 { 7644 netdev_lock(n->dev); 7645 napi_disable_locked(n); 7646 netdev_unlock(n->dev); 7647 } 7648 EXPORT_SYMBOL(napi_disable); 7649 7650 void napi_enable_locked(struct napi_struct *n) 7651 { 7652 unsigned long new, val = READ_ONCE(n->state); 7653 7654 if (n->config) 7655 napi_restore_config(n); 7656 else 7657 napi_hash_add(n); 7658 7659 do { 7660 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val)); 7661 7662 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC); 7663 if (n->dev->threaded && n->thread) 7664 new |= NAPIF_STATE_THREADED; 7665 } while (!try_cmpxchg(&n->state, &val, new)); 7666 } 7667 EXPORT_SYMBOL(napi_enable_locked); 7668 7669 /** 7670 * napi_enable() - enable NAPI scheduling 7671 * @n: NAPI context 7672 * 7673 * Enable scheduling of a NAPI instance. 7674 * Must be paired with napi_disable(). 7675 * Takes netdev_lock() for associated net_device. 7676 */ 7677 void napi_enable(struct napi_struct *n) 7678 { 7679 netdev_lock(n->dev); 7680 napi_enable_locked(n); 7681 netdev_unlock(n->dev); 7682 } 7683 EXPORT_SYMBOL(napi_enable); 7684 7685 /* Must be called in process context */ 7686 void __netif_napi_del_locked(struct napi_struct *napi) 7687 { 7688 netdev_assert_locked(napi->dev); 7689 7690 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state)) 7691 return; 7692 7693 /* Make sure NAPI is disabled (or was never enabled). */ 7694 WARN_ON(!test_bit(NAPI_STATE_SCHED, &napi->state)); 7695 7696 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7697 irq_set_affinity_notifier(napi->irq, NULL); 7698 7699 if (napi->config) { 7700 napi->index = -1; 7701 napi->config = NULL; 7702 } 7703 7704 list_del_rcu(&napi->dev_list); 7705 napi_free_frags(napi); 7706 7707 gro_cleanup(&napi->gro); 7708 7709 if (napi->thread) { 7710 kthread_stop(napi->thread); 7711 napi->thread = NULL; 7712 } 7713 } 7714 EXPORT_SYMBOL(__netif_napi_del_locked); 7715 7716 static int __napi_poll(struct napi_struct *n, bool *repoll) 7717 { 7718 int work, weight; 7719 7720 weight = n->weight; 7721 7722 /* This NAPI_STATE_SCHED test is for avoiding a race 7723 * with netpoll's poll_napi(). Only the entity which 7724 * obtains the lock and sees NAPI_STATE_SCHED set will 7725 * actually make the ->poll() call. Therefore we avoid 7726 * accidentally calling ->poll() when NAPI is not scheduled. 7727 */ 7728 work = 0; 7729 if (napi_is_scheduled(n)) { 7730 work = n->poll(n, weight); 7731 trace_napi_poll(n, work, weight); 7732 7733 xdp_do_check_flushed(n); 7734 } 7735 7736 if (unlikely(work > weight)) 7737 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n", 7738 n->poll, work, weight); 7739 7740 if (likely(work < weight)) 7741 return work; 7742 7743 /* Drivers must not modify the NAPI state if they 7744 * consume the entire weight. In such cases this code 7745 * still "owns" the NAPI instance and therefore can 7746 * move the instance around on the list at-will. 7747 */ 7748 if (unlikely(napi_disable_pending(n))) { 7749 napi_complete(n); 7750 return work; 7751 } 7752 7753 /* The NAPI context has more processing work, but busy-polling 7754 * is preferred. Exit early. 7755 */ 7756 if (napi_prefer_busy_poll(n)) { 7757 if (napi_complete_done(n, work)) { 7758 /* If timeout is not set, we need to make sure 7759 * that the NAPI is re-scheduled. 7760 */ 7761 napi_schedule(n); 7762 } 7763 return work; 7764 } 7765 7766 /* Flush too old packets. If HZ < 1000, flush all packets */ 7767 gro_flush_normal(&n->gro, HZ >= 1000); 7768 7769 /* Some drivers may have called napi_schedule 7770 * prior to exhausting their budget. 7771 */ 7772 if (unlikely(!list_empty(&n->poll_list))) { 7773 pr_warn_once("%s: Budget exhausted after napi rescheduled\n", 7774 n->dev ? n->dev->name : "backlog"); 7775 return work; 7776 } 7777 7778 *repoll = true; 7779 7780 return work; 7781 } 7782 7783 static int napi_poll(struct napi_struct *n, struct list_head *repoll) 7784 { 7785 bool do_repoll = false; 7786 void *have; 7787 int work; 7788 7789 list_del_init(&n->poll_list); 7790 7791 have = netpoll_poll_lock(n); 7792 7793 work = __napi_poll(n, &do_repoll); 7794 7795 if (do_repoll) { 7796 #if defined(CONFIG_DEBUG_NET) 7797 if (unlikely(!napi_is_scheduled(n))) 7798 pr_crit("repoll requested for device %s %ps but napi is not scheduled.\n", 7799 n->dev->name, n->poll); 7800 #endif 7801 list_add_tail(&n->poll_list, repoll); 7802 } 7803 netpoll_poll_unlock(have); 7804 7805 return work; 7806 } 7807 7808 static int napi_thread_wait(struct napi_struct *napi) 7809 { 7810 set_current_state(TASK_INTERRUPTIBLE); 7811 7812 while (!kthread_should_stop()) { 7813 /* Testing SCHED_THREADED bit here to make sure the current 7814 * kthread owns this napi and could poll on this napi. 7815 * Testing SCHED bit is not enough because SCHED bit might be 7816 * set by some other busy poll thread or by napi_disable(). 7817 */ 7818 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) { 7819 WARN_ON(!list_empty(&napi->poll_list)); 7820 __set_current_state(TASK_RUNNING); 7821 return 0; 7822 } 7823 7824 schedule(); 7825 set_current_state(TASK_INTERRUPTIBLE); 7826 } 7827 __set_current_state(TASK_RUNNING); 7828 7829 return -1; 7830 } 7831 7832 static void napi_threaded_poll_loop(struct napi_struct *napi, 7833 unsigned long *busy_poll_last_qs) 7834 { 7835 unsigned long last_qs = busy_poll_last_qs ? *busy_poll_last_qs : jiffies; 7836 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7837 struct softnet_data *sd; 7838 7839 for (;;) { 7840 bool repoll = false; 7841 void *have; 7842 7843 local_bh_disable(); 7844 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7845 7846 sd = this_cpu_ptr(&softnet_data); 7847 sd->in_napi_threaded_poll = true; 7848 7849 have = netpoll_poll_lock(napi); 7850 __napi_poll(napi, &repoll); 7851 netpoll_poll_unlock(have); 7852 7853 sd->in_napi_threaded_poll = false; 7854 barrier(); 7855 7856 if (sd_has_rps_ipi_waiting(sd)) { 7857 local_irq_disable(); 7858 net_rps_action_and_irq_enable(sd); 7859 } 7860 skb_defer_free_flush(); 7861 bpf_net_ctx_clear(bpf_net_ctx); 7862 7863 /* When busy poll is enabled, the old packets are not flushed in 7864 * napi_complete_done. So flush them here. 7865 */ 7866 if (busy_poll_last_qs) 7867 gro_flush_normal(&napi->gro, HZ >= 1000); 7868 local_bh_enable(); 7869 7870 /* Call cond_resched here to avoid watchdog warnings. */ 7871 if (repoll || busy_poll_last_qs) { 7872 rcu_softirq_qs_periodic(last_qs); 7873 cond_resched(); 7874 } 7875 7876 if (!repoll) 7877 break; 7878 } 7879 7880 if (busy_poll_last_qs) 7881 *busy_poll_last_qs = last_qs; 7882 } 7883 7884 static int napi_threaded_poll(void *data) 7885 { 7886 struct napi_struct *napi = data; 7887 unsigned long last_qs = jiffies; 7888 bool want_busy_poll; 7889 bool in_busy_poll; 7890 unsigned long val; 7891 7892 while (!napi_thread_wait(napi)) { 7893 val = READ_ONCE(napi->state); 7894 7895 want_busy_poll = val & NAPIF_STATE_THREADED_BUSY_POLL; 7896 in_busy_poll = val & NAPIF_STATE_IN_BUSY_POLL; 7897 7898 if (unlikely(val & NAPIF_STATE_DISABLE)) 7899 want_busy_poll = false; 7900 7901 if (want_busy_poll != in_busy_poll) 7902 assign_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state, 7903 want_busy_poll); 7904 7905 napi_threaded_poll_loop(napi, want_busy_poll ? &last_qs : NULL); 7906 } 7907 7908 return 0; 7909 } 7910 7911 static __latent_entropy void net_rx_action(void) 7912 { 7913 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 7914 unsigned long time_limit = jiffies + 7915 usecs_to_jiffies(READ_ONCE(net_hotdata.netdev_budget_usecs)); 7916 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7917 int budget = READ_ONCE(net_hotdata.netdev_budget); 7918 LIST_HEAD(list); 7919 LIST_HEAD(repoll); 7920 7921 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7922 start: 7923 sd->in_net_rx_action = true; 7924 local_irq_disable(); 7925 list_splice_init(&sd->poll_list, &list); 7926 local_irq_enable(); 7927 7928 for (;;) { 7929 struct napi_struct *n; 7930 7931 skb_defer_free_flush(); 7932 7933 if (list_empty(&list)) { 7934 if (list_empty(&repoll)) { 7935 sd->in_net_rx_action = false; 7936 barrier(); 7937 /* We need to check if ____napi_schedule() 7938 * had refilled poll_list while 7939 * sd->in_net_rx_action was true. 7940 */ 7941 if (!list_empty(&sd->poll_list)) 7942 goto start; 7943 if (!sd_has_rps_ipi_waiting(sd)) 7944 goto end; 7945 } 7946 break; 7947 } 7948 7949 n = list_first_entry(&list, struct napi_struct, poll_list); 7950 budget -= napi_poll(n, &repoll); 7951 7952 /* If softirq window is exhausted then punt. 7953 * Allow this to run for 2 jiffies since which will allow 7954 * an average latency of 1.5/HZ. 7955 */ 7956 if (unlikely(budget <= 0 || 7957 time_after_eq(jiffies, time_limit))) { 7958 /* Pairs with READ_ONCE() in softnet_seq_show() */ 7959 WRITE_ONCE(sd->time_squeeze, sd->time_squeeze + 1); 7960 break; 7961 } 7962 } 7963 7964 local_irq_disable(); 7965 7966 list_splice_tail_init(&sd->poll_list, &list); 7967 list_splice_tail(&repoll, &list); 7968 list_splice(&list, &sd->poll_list); 7969 if (!list_empty(&sd->poll_list)) 7970 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 7971 else 7972 sd->in_net_rx_action = false; 7973 7974 net_rps_action_and_irq_enable(sd); 7975 end: 7976 bpf_net_ctx_clear(bpf_net_ctx); 7977 } 7978 7979 struct netdev_adjacent { 7980 struct net_device *dev; 7981 netdevice_tracker dev_tracker; 7982 7983 /* upper master flag, there can only be one master device per list */ 7984 bool master; 7985 7986 /* lookup ignore flag */ 7987 bool ignore; 7988 7989 /* counter for the number of times this device was added to us */ 7990 u16 ref_nr; 7991 7992 /* private field for the users */ 7993 void *private; 7994 7995 struct list_head list; 7996 struct rcu_head rcu; 7997 }; 7998 7999 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev, 8000 struct list_head *adj_list) 8001 { 8002 struct netdev_adjacent *adj; 8003 8004 list_for_each_entry(adj, adj_list, list) { 8005 if (adj->dev == adj_dev) 8006 return adj; 8007 } 8008 return NULL; 8009 } 8010 8011 static int ____netdev_has_upper_dev(struct net_device *upper_dev, 8012 struct netdev_nested_priv *priv) 8013 { 8014 struct net_device *dev = (struct net_device *)priv->data; 8015 8016 return upper_dev == dev; 8017 } 8018 8019 /** 8020 * netdev_has_upper_dev - Check if device is linked to an upper device 8021 * @dev: device 8022 * @upper_dev: upper device to check 8023 * 8024 * Find out if a device is linked to specified upper device and return true 8025 * in case it is. Note that this checks only immediate upper device, 8026 * not through a complete stack of devices. The caller must hold the RTNL lock. 8027 */ 8028 bool netdev_has_upper_dev(struct net_device *dev, 8029 struct net_device *upper_dev) 8030 { 8031 struct netdev_nested_priv priv = { 8032 .data = (void *)upper_dev, 8033 }; 8034 8035 ASSERT_RTNL(); 8036 8037 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 8038 &priv); 8039 } 8040 EXPORT_SYMBOL(netdev_has_upper_dev); 8041 8042 /** 8043 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device 8044 * @dev: device 8045 * @upper_dev: upper device to check 8046 * 8047 * Find out if a device is linked to specified upper device and return true 8048 * in case it is. Note that this checks the entire upper device chain. 8049 * The caller must hold rcu lock. 8050 */ 8051 8052 bool netdev_has_upper_dev_all_rcu(struct net_device *dev, 8053 struct net_device *upper_dev) 8054 { 8055 struct netdev_nested_priv priv = { 8056 .data = (void *)upper_dev, 8057 }; 8058 8059 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 8060 &priv); 8061 } 8062 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu); 8063 8064 /** 8065 * netdev_has_any_upper_dev - Check if device is linked to some device 8066 * @dev: device 8067 * 8068 * Find out if a device is linked to an upper device and return true in case 8069 * it is. The caller must hold the RTNL lock. 8070 */ 8071 bool netdev_has_any_upper_dev(struct net_device *dev) 8072 { 8073 ASSERT_RTNL(); 8074 8075 return !list_empty(&dev->adj_list.upper); 8076 } 8077 EXPORT_SYMBOL(netdev_has_any_upper_dev); 8078 8079 /** 8080 * netdev_master_upper_dev_get - Get master upper device 8081 * @dev: device 8082 * 8083 * Find a master upper device and return pointer to it or NULL in case 8084 * it's not there. The caller must hold the RTNL lock. 8085 */ 8086 struct net_device *netdev_master_upper_dev_get(struct net_device *dev) 8087 { 8088 struct netdev_adjacent *upper; 8089 8090 ASSERT_RTNL(); 8091 8092 if (list_empty(&dev->adj_list.upper)) 8093 return NULL; 8094 8095 upper = list_first_entry(&dev->adj_list.upper, 8096 struct netdev_adjacent, list); 8097 if (likely(upper->master)) 8098 return upper->dev; 8099 return NULL; 8100 } 8101 EXPORT_SYMBOL(netdev_master_upper_dev_get); 8102 8103 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev) 8104 { 8105 struct netdev_adjacent *upper; 8106 8107 ASSERT_RTNL(); 8108 8109 if (list_empty(&dev->adj_list.upper)) 8110 return NULL; 8111 8112 upper = list_first_entry(&dev->adj_list.upper, 8113 struct netdev_adjacent, list); 8114 if (likely(upper->master) && !upper->ignore) 8115 return upper->dev; 8116 return NULL; 8117 } 8118 8119 /** 8120 * netdev_has_any_lower_dev - Check if device is linked to some device 8121 * @dev: device 8122 * 8123 * Find out if a device is linked to a lower device and return true in case 8124 * it is. The caller must hold the RTNL lock. 8125 */ 8126 static bool netdev_has_any_lower_dev(struct net_device *dev) 8127 { 8128 ASSERT_RTNL(); 8129 8130 return !list_empty(&dev->adj_list.lower); 8131 } 8132 8133 void *netdev_adjacent_get_private(struct list_head *adj_list) 8134 { 8135 struct netdev_adjacent *adj; 8136 8137 adj = list_entry(adj_list, struct netdev_adjacent, list); 8138 8139 return adj->private; 8140 } 8141 EXPORT_SYMBOL(netdev_adjacent_get_private); 8142 8143 /** 8144 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list 8145 * @dev: device 8146 * @iter: list_head ** of the current position 8147 * 8148 * Gets the next device from the dev's upper list, starting from iter 8149 * position. The caller must hold RCU read lock. 8150 */ 8151 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev, 8152 struct list_head **iter) 8153 { 8154 struct netdev_adjacent *upper; 8155 8156 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held() && 8157 !lockdep_rtnl_is_held()); 8158 8159 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8160 8161 if (&upper->list == &dev->adj_list.upper) 8162 return NULL; 8163 8164 *iter = &upper->list; 8165 8166 return upper->dev; 8167 } 8168 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu); 8169 8170 static struct net_device *__netdev_next_upper_dev(struct net_device *dev, 8171 struct list_head **iter, 8172 bool *ignore) 8173 { 8174 struct netdev_adjacent *upper; 8175 8176 upper = list_entry((*iter)->next, struct netdev_adjacent, list); 8177 8178 if (&upper->list == &dev->adj_list.upper) 8179 return NULL; 8180 8181 *iter = &upper->list; 8182 *ignore = upper->ignore; 8183 8184 return upper->dev; 8185 } 8186 8187 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev, 8188 struct list_head **iter) 8189 { 8190 struct netdev_adjacent *upper; 8191 8192 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 8193 8194 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8195 8196 if (&upper->list == &dev->adj_list.upper) 8197 return NULL; 8198 8199 *iter = &upper->list; 8200 8201 return upper->dev; 8202 } 8203 8204 static int __netdev_walk_all_upper_dev(struct net_device *dev, 8205 int (*fn)(struct net_device *dev, 8206 struct netdev_nested_priv *priv), 8207 struct netdev_nested_priv *priv) 8208 { 8209 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8210 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8211 int ret, cur = 0; 8212 bool ignore; 8213 8214 now = dev; 8215 iter = &dev->adj_list.upper; 8216 8217 while (1) { 8218 if (now != dev) { 8219 ret = fn(now, priv); 8220 if (ret) 8221 return ret; 8222 } 8223 8224 next = NULL; 8225 while (1) { 8226 udev = __netdev_next_upper_dev(now, &iter, &ignore); 8227 if (!udev) 8228 break; 8229 if (ignore) 8230 continue; 8231 8232 next = udev; 8233 niter = &udev->adj_list.upper; 8234 dev_stack[cur] = now; 8235 iter_stack[cur++] = iter; 8236 break; 8237 } 8238 8239 if (!next) { 8240 if (!cur) 8241 return 0; 8242 next = dev_stack[--cur]; 8243 niter = iter_stack[cur]; 8244 } 8245 8246 now = next; 8247 iter = niter; 8248 } 8249 8250 return 0; 8251 } 8252 8253 int netdev_walk_all_upper_dev_rcu(struct net_device *dev, 8254 int (*fn)(struct net_device *dev, 8255 struct netdev_nested_priv *priv), 8256 struct netdev_nested_priv *priv) 8257 { 8258 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8259 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8260 int ret, cur = 0; 8261 8262 now = dev; 8263 iter = &dev->adj_list.upper; 8264 8265 while (1) { 8266 if (now != dev) { 8267 ret = fn(now, priv); 8268 if (ret) 8269 return ret; 8270 } 8271 8272 next = NULL; 8273 while (1) { 8274 udev = netdev_next_upper_dev_rcu(now, &iter); 8275 if (!udev) 8276 break; 8277 8278 next = udev; 8279 niter = &udev->adj_list.upper; 8280 dev_stack[cur] = now; 8281 iter_stack[cur++] = iter; 8282 break; 8283 } 8284 8285 if (!next) { 8286 if (!cur) 8287 return 0; 8288 next = dev_stack[--cur]; 8289 niter = iter_stack[cur]; 8290 } 8291 8292 now = next; 8293 iter = niter; 8294 } 8295 8296 return 0; 8297 } 8298 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu); 8299 8300 static bool __netdev_has_upper_dev(struct net_device *dev, 8301 struct net_device *upper_dev) 8302 { 8303 struct netdev_nested_priv priv = { 8304 .flags = 0, 8305 .data = (void *)upper_dev, 8306 }; 8307 8308 ASSERT_RTNL(); 8309 8310 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev, 8311 &priv); 8312 } 8313 8314 /** 8315 * netdev_lower_get_next_private - Get the next ->private from the 8316 * lower neighbour list 8317 * @dev: device 8318 * @iter: list_head ** of the current position 8319 * 8320 * Gets the next netdev_adjacent->private from the dev's lower neighbour 8321 * list, starting from iter position. The caller must hold either hold the 8322 * RTNL lock or its own locking that guarantees that the neighbour lower 8323 * list will remain unchanged. 8324 */ 8325 void *netdev_lower_get_next_private(struct net_device *dev, 8326 struct list_head **iter) 8327 { 8328 struct netdev_adjacent *lower; 8329 8330 lower = list_entry(*iter, struct netdev_adjacent, list); 8331 8332 if (&lower->list == &dev->adj_list.lower) 8333 return NULL; 8334 8335 *iter = lower->list.next; 8336 8337 return lower->private; 8338 } 8339 EXPORT_SYMBOL(netdev_lower_get_next_private); 8340 8341 /** 8342 * netdev_lower_get_next_private_rcu - Get the next ->private from the 8343 * lower neighbour list, RCU 8344 * variant 8345 * @dev: device 8346 * @iter: list_head ** of the current position 8347 * 8348 * Gets the next netdev_adjacent->private from the dev's lower neighbour 8349 * list, starting from iter position. The caller must hold RCU read lock. 8350 */ 8351 void *netdev_lower_get_next_private_rcu(struct net_device *dev, 8352 struct list_head **iter) 8353 { 8354 struct netdev_adjacent *lower; 8355 8356 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 8357 8358 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8359 8360 if (&lower->list == &dev->adj_list.lower) 8361 return NULL; 8362 8363 *iter = &lower->list; 8364 8365 return lower->private; 8366 } 8367 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu); 8368 8369 /** 8370 * netdev_lower_get_next - Get the next device from the lower neighbour 8371 * list 8372 * @dev: device 8373 * @iter: list_head ** of the current position 8374 * 8375 * Gets the next netdev_adjacent from the dev's lower neighbour 8376 * list, starting from iter position. The caller must hold RTNL lock or 8377 * its own locking that guarantees that the neighbour lower 8378 * list will remain unchanged. 8379 */ 8380 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter) 8381 { 8382 struct netdev_adjacent *lower; 8383 8384 lower = list_entry(*iter, struct netdev_adjacent, list); 8385 8386 if (&lower->list == &dev->adj_list.lower) 8387 return NULL; 8388 8389 *iter = lower->list.next; 8390 8391 return lower->dev; 8392 } 8393 EXPORT_SYMBOL(netdev_lower_get_next); 8394 8395 static struct net_device *netdev_next_lower_dev(struct net_device *dev, 8396 struct list_head **iter) 8397 { 8398 struct netdev_adjacent *lower; 8399 8400 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 8401 8402 if (&lower->list == &dev->adj_list.lower) 8403 return NULL; 8404 8405 *iter = &lower->list; 8406 8407 return lower->dev; 8408 } 8409 8410 static struct net_device *__netdev_next_lower_dev(struct net_device *dev, 8411 struct list_head **iter, 8412 bool *ignore) 8413 { 8414 struct netdev_adjacent *lower; 8415 8416 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 8417 8418 if (&lower->list == &dev->adj_list.lower) 8419 return NULL; 8420 8421 *iter = &lower->list; 8422 *ignore = lower->ignore; 8423 8424 return lower->dev; 8425 } 8426 8427 int netdev_walk_all_lower_dev(struct net_device *dev, 8428 int (*fn)(struct net_device *dev, 8429 struct netdev_nested_priv *priv), 8430 struct netdev_nested_priv *priv) 8431 { 8432 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8433 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8434 int ret, cur = 0; 8435 8436 now = dev; 8437 iter = &dev->adj_list.lower; 8438 8439 while (1) { 8440 if (now != dev) { 8441 ret = fn(now, priv); 8442 if (ret) 8443 return ret; 8444 } 8445 8446 next = NULL; 8447 while (1) { 8448 ldev = netdev_next_lower_dev(now, &iter); 8449 if (!ldev) 8450 break; 8451 8452 next = ldev; 8453 niter = &ldev->adj_list.lower; 8454 dev_stack[cur] = now; 8455 iter_stack[cur++] = iter; 8456 break; 8457 } 8458 8459 if (!next) { 8460 if (!cur) 8461 return 0; 8462 next = dev_stack[--cur]; 8463 niter = iter_stack[cur]; 8464 } 8465 8466 now = next; 8467 iter = niter; 8468 } 8469 8470 return 0; 8471 } 8472 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev); 8473 8474 static int __netdev_walk_all_lower_dev(struct net_device *dev, 8475 int (*fn)(struct net_device *dev, 8476 struct netdev_nested_priv *priv), 8477 struct netdev_nested_priv *priv) 8478 { 8479 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8480 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8481 int ret, cur = 0; 8482 bool ignore; 8483 8484 now = dev; 8485 iter = &dev->adj_list.lower; 8486 8487 while (1) { 8488 if (now != dev) { 8489 ret = fn(now, priv); 8490 if (ret) 8491 return ret; 8492 } 8493 8494 next = NULL; 8495 while (1) { 8496 ldev = __netdev_next_lower_dev(now, &iter, &ignore); 8497 if (!ldev) 8498 break; 8499 if (ignore) 8500 continue; 8501 8502 next = ldev; 8503 niter = &ldev->adj_list.lower; 8504 dev_stack[cur] = now; 8505 iter_stack[cur++] = iter; 8506 break; 8507 } 8508 8509 if (!next) { 8510 if (!cur) 8511 return 0; 8512 next = dev_stack[--cur]; 8513 niter = iter_stack[cur]; 8514 } 8515 8516 now = next; 8517 iter = niter; 8518 } 8519 8520 return 0; 8521 } 8522 8523 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev, 8524 struct list_head **iter) 8525 { 8526 struct netdev_adjacent *lower; 8527 8528 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8529 if (&lower->list == &dev->adj_list.lower) 8530 return NULL; 8531 8532 *iter = &lower->list; 8533 8534 return lower->dev; 8535 } 8536 EXPORT_SYMBOL(netdev_next_lower_dev_rcu); 8537 8538 static u8 __netdev_upper_depth(struct net_device *dev) 8539 { 8540 struct net_device *udev; 8541 struct list_head *iter; 8542 u8 max_depth = 0; 8543 bool ignore; 8544 8545 for (iter = &dev->adj_list.upper, 8546 udev = __netdev_next_upper_dev(dev, &iter, &ignore); 8547 udev; 8548 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) { 8549 if (ignore) 8550 continue; 8551 if (max_depth < udev->upper_level) 8552 max_depth = udev->upper_level; 8553 } 8554 8555 return max_depth; 8556 } 8557 8558 static u8 __netdev_lower_depth(struct net_device *dev) 8559 { 8560 struct net_device *ldev; 8561 struct list_head *iter; 8562 u8 max_depth = 0; 8563 bool ignore; 8564 8565 for (iter = &dev->adj_list.lower, 8566 ldev = __netdev_next_lower_dev(dev, &iter, &ignore); 8567 ldev; 8568 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) { 8569 if (ignore) 8570 continue; 8571 if (max_depth < ldev->lower_level) 8572 max_depth = ldev->lower_level; 8573 } 8574 8575 return max_depth; 8576 } 8577 8578 static int __netdev_update_upper_level(struct net_device *dev, 8579 struct netdev_nested_priv *__unused) 8580 { 8581 dev->upper_level = __netdev_upper_depth(dev) + 1; 8582 return 0; 8583 } 8584 8585 #ifdef CONFIG_LOCKDEP 8586 static LIST_HEAD(net_unlink_list); 8587 8588 static void net_unlink_todo(struct net_device *dev) 8589 { 8590 if (list_empty(&dev->unlink_list)) 8591 list_add_tail(&dev->unlink_list, &net_unlink_list); 8592 } 8593 #endif 8594 8595 static int __netdev_update_lower_level(struct net_device *dev, 8596 struct netdev_nested_priv *priv) 8597 { 8598 dev->lower_level = __netdev_lower_depth(dev) + 1; 8599 8600 #ifdef CONFIG_LOCKDEP 8601 if (!priv) 8602 return 0; 8603 8604 if (priv->flags & NESTED_SYNC_IMM) 8605 dev->nested_level = dev->lower_level - 1; 8606 if (priv->flags & NESTED_SYNC_TODO) 8607 net_unlink_todo(dev); 8608 #endif 8609 return 0; 8610 } 8611 8612 int netdev_walk_all_lower_dev_rcu(struct net_device *dev, 8613 int (*fn)(struct net_device *dev, 8614 struct netdev_nested_priv *priv), 8615 struct netdev_nested_priv *priv) 8616 { 8617 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8618 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8619 int ret, cur = 0; 8620 8621 now = dev; 8622 iter = &dev->adj_list.lower; 8623 8624 while (1) { 8625 if (now != dev) { 8626 ret = fn(now, priv); 8627 if (ret) 8628 return ret; 8629 } 8630 8631 next = NULL; 8632 while (1) { 8633 ldev = netdev_next_lower_dev_rcu(now, &iter); 8634 if (!ldev) 8635 break; 8636 8637 next = ldev; 8638 niter = &ldev->adj_list.lower; 8639 dev_stack[cur] = now; 8640 iter_stack[cur++] = iter; 8641 break; 8642 } 8643 8644 if (!next) { 8645 if (!cur) 8646 return 0; 8647 next = dev_stack[--cur]; 8648 niter = iter_stack[cur]; 8649 } 8650 8651 now = next; 8652 iter = niter; 8653 } 8654 8655 return 0; 8656 } 8657 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu); 8658 8659 /** 8660 * netdev_lower_get_first_private_rcu - Get the first ->private from the 8661 * lower neighbour list, RCU 8662 * variant 8663 * @dev: device 8664 * 8665 * Gets the first netdev_adjacent->private from the dev's lower neighbour 8666 * list. The caller must hold RCU read lock. 8667 */ 8668 void *netdev_lower_get_first_private_rcu(struct net_device *dev) 8669 { 8670 struct netdev_adjacent *lower; 8671 8672 lower = list_first_or_null_rcu(&dev->adj_list.lower, 8673 struct netdev_adjacent, list); 8674 if (lower) 8675 return lower->private; 8676 return NULL; 8677 } 8678 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu); 8679 8680 /** 8681 * netdev_master_upper_dev_get_rcu - Get master upper device 8682 * @dev: device 8683 * 8684 * Find a master upper device and return pointer to it or NULL in case 8685 * it's not there. The caller must hold the RCU read lock. 8686 */ 8687 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev) 8688 { 8689 struct netdev_adjacent *upper; 8690 8691 upper = list_first_or_null_rcu(&dev->adj_list.upper, 8692 struct netdev_adjacent, list); 8693 if (upper && likely(upper->master)) 8694 return upper->dev; 8695 return NULL; 8696 } 8697 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu); 8698 8699 static int netdev_adjacent_sysfs_add(struct net_device *dev, 8700 struct net_device *adj_dev, 8701 struct list_head *dev_list) 8702 { 8703 char linkname[IFNAMSIZ+7]; 8704 8705 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8706 "upper_%s" : "lower_%s", adj_dev->name); 8707 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj), 8708 linkname); 8709 } 8710 static void netdev_adjacent_sysfs_del(struct net_device *dev, 8711 char *name, 8712 struct list_head *dev_list) 8713 { 8714 char linkname[IFNAMSIZ+7]; 8715 8716 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8717 "upper_%s" : "lower_%s", name); 8718 sysfs_remove_link(&(dev->dev.kobj), linkname); 8719 } 8720 8721 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev, 8722 struct net_device *adj_dev, 8723 struct list_head *dev_list) 8724 { 8725 return (dev_list == &dev->adj_list.upper || 8726 dev_list == &dev->adj_list.lower) && 8727 net_eq(dev_net(dev), dev_net(adj_dev)); 8728 } 8729 8730 static int __netdev_adjacent_dev_insert(struct net_device *dev, 8731 struct net_device *adj_dev, 8732 struct list_head *dev_list, 8733 void *private, bool master) 8734 { 8735 struct netdev_adjacent *adj; 8736 int ret; 8737 8738 adj = __netdev_find_adj(adj_dev, dev_list); 8739 8740 if (adj) { 8741 adj->ref_nr += 1; 8742 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n", 8743 dev->name, adj_dev->name, adj->ref_nr); 8744 8745 return 0; 8746 } 8747 8748 adj = kmalloc_obj(*adj); 8749 if (!adj) 8750 return -ENOMEM; 8751 8752 adj->dev = adj_dev; 8753 adj->master = master; 8754 adj->ref_nr = 1; 8755 adj->private = private; 8756 adj->ignore = false; 8757 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL); 8758 8759 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n", 8760 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name); 8761 8762 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) { 8763 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list); 8764 if (ret) 8765 goto free_adj; 8766 } 8767 8768 /* Ensure that master link is always the first item in list. */ 8769 if (master) { 8770 ret = sysfs_create_link(&(dev->dev.kobj), 8771 &(adj_dev->dev.kobj), "master"); 8772 if (ret) 8773 goto remove_symlinks; 8774 8775 list_add_rcu(&adj->list, dev_list); 8776 } else { 8777 list_add_tail_rcu(&adj->list, dev_list); 8778 } 8779 8780 return 0; 8781 8782 remove_symlinks: 8783 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8784 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8785 free_adj: 8786 netdev_put(adj_dev, &adj->dev_tracker); 8787 kfree(adj); 8788 8789 return ret; 8790 } 8791 8792 static void __netdev_adjacent_dev_remove(struct net_device *dev, 8793 struct net_device *adj_dev, 8794 u16 ref_nr, 8795 struct list_head *dev_list) 8796 { 8797 struct netdev_adjacent *adj; 8798 8799 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n", 8800 dev->name, adj_dev->name, ref_nr); 8801 8802 adj = __netdev_find_adj(adj_dev, dev_list); 8803 8804 if (!adj) { 8805 pr_err("Adjacency does not exist for device %s from %s\n", 8806 dev->name, adj_dev->name); 8807 WARN_ON(1); 8808 return; 8809 } 8810 8811 if (adj->ref_nr > ref_nr) { 8812 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n", 8813 dev->name, adj_dev->name, ref_nr, 8814 adj->ref_nr - ref_nr); 8815 adj->ref_nr -= ref_nr; 8816 return; 8817 } 8818 8819 if (adj->master) 8820 sysfs_remove_link(&(dev->dev.kobj), "master"); 8821 8822 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8823 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8824 8825 list_del_rcu(&adj->list); 8826 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n", 8827 adj_dev->name, dev->name, adj_dev->name); 8828 netdev_put(adj_dev, &adj->dev_tracker); 8829 kfree_rcu(adj, rcu); 8830 } 8831 8832 static int __netdev_adjacent_dev_link_lists(struct net_device *dev, 8833 struct net_device *upper_dev, 8834 struct list_head *up_list, 8835 struct list_head *down_list, 8836 void *private, bool master) 8837 { 8838 int ret; 8839 8840 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list, 8841 private, master); 8842 if (ret) 8843 return ret; 8844 8845 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list, 8846 private, false); 8847 if (ret) { 8848 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list); 8849 return ret; 8850 } 8851 8852 return 0; 8853 } 8854 8855 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev, 8856 struct net_device *upper_dev, 8857 u16 ref_nr, 8858 struct list_head *up_list, 8859 struct list_head *down_list) 8860 { 8861 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list); 8862 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list); 8863 } 8864 8865 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev, 8866 struct net_device *upper_dev, 8867 void *private, bool master) 8868 { 8869 return __netdev_adjacent_dev_link_lists(dev, upper_dev, 8870 &dev->adj_list.upper, 8871 &upper_dev->adj_list.lower, 8872 private, master); 8873 } 8874 8875 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev, 8876 struct net_device *upper_dev) 8877 { 8878 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1, 8879 &dev->adj_list.upper, 8880 &upper_dev->adj_list.lower); 8881 } 8882 8883 static int __netdev_upper_dev_link(struct net_device *dev, 8884 struct net_device *upper_dev, bool master, 8885 void *upper_priv, void *upper_info, 8886 struct netdev_nested_priv *priv, 8887 struct netlink_ext_ack *extack) 8888 { 8889 struct netdev_notifier_changeupper_info changeupper_info = { 8890 .info = { 8891 .dev = dev, 8892 .extack = extack, 8893 }, 8894 .upper_dev = upper_dev, 8895 .master = master, 8896 .linking = true, 8897 .upper_info = upper_info, 8898 }; 8899 struct net_device *master_dev; 8900 int ret = 0; 8901 8902 ASSERT_RTNL(); 8903 8904 if (dev == upper_dev) 8905 return -EBUSY; 8906 8907 /* To prevent loops, check if dev is not upper device to upper_dev. */ 8908 if (__netdev_has_upper_dev(upper_dev, dev)) 8909 return -EBUSY; 8910 8911 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV) 8912 return -EMLINK; 8913 8914 if (!master) { 8915 if (__netdev_has_upper_dev(dev, upper_dev)) 8916 return -EEXIST; 8917 } else { 8918 master_dev = __netdev_master_upper_dev_get(dev); 8919 if (master_dev) 8920 return master_dev == upper_dev ? -EEXIST : -EBUSY; 8921 } 8922 8923 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8924 &changeupper_info.info); 8925 ret = notifier_to_errno(ret); 8926 if (ret) 8927 return ret; 8928 8929 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv, 8930 master); 8931 if (ret) 8932 return ret; 8933 8934 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8935 &changeupper_info.info); 8936 ret = notifier_to_errno(ret); 8937 if (ret) 8938 goto rollback; 8939 8940 __netdev_update_upper_level(dev, NULL); 8941 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8942 8943 __netdev_update_lower_level(upper_dev, priv); 8944 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8945 priv); 8946 8947 return 0; 8948 8949 rollback: 8950 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8951 8952 return ret; 8953 } 8954 8955 /** 8956 * netdev_upper_dev_link - Add a link to the upper device 8957 * @dev: device 8958 * @upper_dev: new upper device 8959 * @extack: netlink extended ack 8960 * 8961 * Adds a link to device which is upper to this one. The caller must hold 8962 * the RTNL lock. On a failure a negative errno code is returned. 8963 * On success the reference counts are adjusted and the function 8964 * returns zero. 8965 */ 8966 int netdev_upper_dev_link(struct net_device *dev, 8967 struct net_device *upper_dev, 8968 struct netlink_ext_ack *extack) 8969 { 8970 struct netdev_nested_priv priv = { 8971 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8972 .data = NULL, 8973 }; 8974 8975 return __netdev_upper_dev_link(dev, upper_dev, false, 8976 NULL, NULL, &priv, extack); 8977 } 8978 EXPORT_SYMBOL(netdev_upper_dev_link); 8979 8980 /** 8981 * netdev_master_upper_dev_link - Add a master link to the upper device 8982 * @dev: device 8983 * @upper_dev: new upper device 8984 * @upper_priv: upper device private 8985 * @upper_info: upper info to be passed down via notifier 8986 * @extack: netlink extended ack 8987 * 8988 * Adds a link to device which is upper to this one. In this case, only 8989 * one master upper device can be linked, although other non-master devices 8990 * might be linked as well. The caller must hold the RTNL lock. 8991 * On a failure a negative errno code is returned. On success the reference 8992 * counts are adjusted and the function returns zero. 8993 */ 8994 int netdev_master_upper_dev_link(struct net_device *dev, 8995 struct net_device *upper_dev, 8996 void *upper_priv, void *upper_info, 8997 struct netlink_ext_ack *extack) 8998 { 8999 struct netdev_nested_priv priv = { 9000 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 9001 .data = NULL, 9002 }; 9003 9004 return __netdev_upper_dev_link(dev, upper_dev, true, 9005 upper_priv, upper_info, &priv, extack); 9006 } 9007 EXPORT_SYMBOL(netdev_master_upper_dev_link); 9008 9009 static void __netdev_upper_dev_unlink(struct net_device *dev, 9010 struct net_device *upper_dev, 9011 struct netdev_nested_priv *priv) 9012 { 9013 struct netdev_notifier_changeupper_info changeupper_info = { 9014 .info = { 9015 .dev = dev, 9016 }, 9017 .upper_dev = upper_dev, 9018 .linking = false, 9019 }; 9020 9021 ASSERT_RTNL(); 9022 9023 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev; 9024 9025 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 9026 &changeupper_info.info); 9027 9028 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 9029 9030 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 9031 &changeupper_info.info); 9032 9033 __netdev_update_upper_level(dev, NULL); 9034 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 9035 9036 __netdev_update_lower_level(upper_dev, priv); 9037 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 9038 priv); 9039 } 9040 9041 /** 9042 * netdev_upper_dev_unlink - Removes a link to upper device 9043 * @dev: device 9044 * @upper_dev: new upper device 9045 * 9046 * Removes a link to device which is upper to this one. The caller must hold 9047 * the RTNL lock. 9048 */ 9049 void netdev_upper_dev_unlink(struct net_device *dev, 9050 struct net_device *upper_dev) 9051 { 9052 struct netdev_nested_priv priv = { 9053 .flags = NESTED_SYNC_TODO, 9054 .data = NULL, 9055 }; 9056 9057 __netdev_upper_dev_unlink(dev, upper_dev, &priv); 9058 } 9059 EXPORT_SYMBOL(netdev_upper_dev_unlink); 9060 9061 static void __netdev_adjacent_dev_set(struct net_device *upper_dev, 9062 struct net_device *lower_dev, 9063 bool val) 9064 { 9065 struct netdev_adjacent *adj; 9066 9067 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower); 9068 if (adj) 9069 adj->ignore = val; 9070 9071 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper); 9072 if (adj) 9073 adj->ignore = val; 9074 } 9075 9076 static void netdev_adjacent_dev_disable(struct net_device *upper_dev, 9077 struct net_device *lower_dev) 9078 { 9079 __netdev_adjacent_dev_set(upper_dev, lower_dev, true); 9080 } 9081 9082 static void netdev_adjacent_dev_enable(struct net_device *upper_dev, 9083 struct net_device *lower_dev) 9084 { 9085 __netdev_adjacent_dev_set(upper_dev, lower_dev, false); 9086 } 9087 9088 int netdev_adjacent_change_prepare(struct net_device *old_dev, 9089 struct net_device *new_dev, 9090 struct net_device *dev, 9091 struct netlink_ext_ack *extack) 9092 { 9093 struct netdev_nested_priv priv = { 9094 .flags = 0, 9095 .data = NULL, 9096 }; 9097 int err; 9098 9099 if (!new_dev) 9100 return 0; 9101 9102 if (old_dev && new_dev != old_dev) 9103 netdev_adjacent_dev_disable(dev, old_dev); 9104 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv, 9105 extack); 9106 if (err) { 9107 if (old_dev && new_dev != old_dev) 9108 netdev_adjacent_dev_enable(dev, old_dev); 9109 return err; 9110 } 9111 9112 return 0; 9113 } 9114 EXPORT_SYMBOL(netdev_adjacent_change_prepare); 9115 9116 void netdev_adjacent_change_commit(struct net_device *old_dev, 9117 struct net_device *new_dev, 9118 struct net_device *dev) 9119 { 9120 struct netdev_nested_priv priv = { 9121 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 9122 .data = NULL, 9123 }; 9124 9125 if (!new_dev || !old_dev) 9126 return; 9127 9128 if (new_dev == old_dev) 9129 return; 9130 9131 netdev_adjacent_dev_enable(dev, old_dev); 9132 __netdev_upper_dev_unlink(old_dev, dev, &priv); 9133 } 9134 EXPORT_SYMBOL(netdev_adjacent_change_commit); 9135 9136 void netdev_adjacent_change_abort(struct net_device *old_dev, 9137 struct net_device *new_dev, 9138 struct net_device *dev) 9139 { 9140 struct netdev_nested_priv priv = { 9141 .flags = 0, 9142 .data = NULL, 9143 }; 9144 9145 if (!new_dev) 9146 return; 9147 9148 if (old_dev && new_dev != old_dev) 9149 netdev_adjacent_dev_enable(dev, old_dev); 9150 9151 __netdev_upper_dev_unlink(new_dev, dev, &priv); 9152 } 9153 EXPORT_SYMBOL(netdev_adjacent_change_abort); 9154 9155 /** 9156 * netdev_bonding_info_change - Dispatch event about slave change 9157 * @dev: device 9158 * @bonding_info: info to dispatch 9159 * 9160 * Send NETDEV_BONDING_INFO to netdev notifiers with info. 9161 * The caller must hold the RTNL lock. 9162 */ 9163 void netdev_bonding_info_change(struct net_device *dev, 9164 struct netdev_bonding_info *bonding_info) 9165 { 9166 struct netdev_notifier_bonding_info info = { 9167 .info.dev = dev, 9168 }; 9169 9170 memcpy(&info.bonding_info, bonding_info, 9171 sizeof(struct netdev_bonding_info)); 9172 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, 9173 &info.info); 9174 } 9175 EXPORT_SYMBOL(netdev_bonding_info_change); 9176 9177 static int netdev_offload_xstats_enable_l3(struct net_device *dev, 9178 struct netlink_ext_ack *extack) 9179 { 9180 struct netdev_notifier_offload_xstats_info info = { 9181 .info.dev = dev, 9182 .info.extack = extack, 9183 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 9184 }; 9185 int err; 9186 int rc; 9187 9188 dev->offload_xstats_l3 = kzalloc_obj(*dev->offload_xstats_l3); 9189 if (!dev->offload_xstats_l3) 9190 return -ENOMEM; 9191 9192 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE, 9193 NETDEV_OFFLOAD_XSTATS_DISABLE, 9194 &info.info); 9195 err = notifier_to_errno(rc); 9196 if (err) 9197 goto free_stats; 9198 9199 return 0; 9200 9201 free_stats: 9202 kfree(dev->offload_xstats_l3); 9203 dev->offload_xstats_l3 = NULL; 9204 return err; 9205 } 9206 9207 int netdev_offload_xstats_enable(struct net_device *dev, 9208 enum netdev_offload_xstats_type type, 9209 struct netlink_ext_ack *extack) 9210 { 9211 ASSERT_RTNL(); 9212 9213 if (netdev_offload_xstats_enabled(dev, type)) 9214 return -EALREADY; 9215 9216 switch (type) { 9217 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 9218 return netdev_offload_xstats_enable_l3(dev, extack); 9219 } 9220 9221 WARN_ON(1); 9222 return -EINVAL; 9223 } 9224 EXPORT_SYMBOL(netdev_offload_xstats_enable); 9225 9226 static void netdev_offload_xstats_disable_l3(struct net_device *dev) 9227 { 9228 struct netdev_notifier_offload_xstats_info info = { 9229 .info.dev = dev, 9230 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 9231 }; 9232 9233 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE, 9234 &info.info); 9235 kfree(dev->offload_xstats_l3); 9236 dev->offload_xstats_l3 = NULL; 9237 } 9238 9239 int netdev_offload_xstats_disable(struct net_device *dev, 9240 enum netdev_offload_xstats_type type) 9241 { 9242 ASSERT_RTNL(); 9243 9244 if (!netdev_offload_xstats_enabled(dev, type)) 9245 return -EALREADY; 9246 9247 switch (type) { 9248 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 9249 netdev_offload_xstats_disable_l3(dev); 9250 return 0; 9251 } 9252 9253 WARN_ON(1); 9254 return -EINVAL; 9255 } 9256 EXPORT_SYMBOL(netdev_offload_xstats_disable); 9257 9258 static void netdev_offload_xstats_disable_all(struct net_device *dev) 9259 { 9260 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3); 9261 } 9262 9263 static struct rtnl_hw_stats64 * 9264 netdev_offload_xstats_get_ptr(const struct net_device *dev, 9265 enum netdev_offload_xstats_type type) 9266 { 9267 switch (type) { 9268 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 9269 return dev->offload_xstats_l3; 9270 } 9271 9272 WARN_ON(1); 9273 return NULL; 9274 } 9275 9276 bool netdev_offload_xstats_enabled(const struct net_device *dev, 9277 enum netdev_offload_xstats_type type) 9278 { 9279 ASSERT_RTNL(); 9280 9281 return netdev_offload_xstats_get_ptr(dev, type); 9282 } 9283 EXPORT_SYMBOL(netdev_offload_xstats_enabled); 9284 9285 struct netdev_notifier_offload_xstats_ru { 9286 bool used; 9287 }; 9288 9289 struct netdev_notifier_offload_xstats_rd { 9290 struct rtnl_hw_stats64 stats; 9291 bool used; 9292 }; 9293 9294 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest, 9295 const struct rtnl_hw_stats64 *src) 9296 { 9297 dest->rx_packets += src->rx_packets; 9298 dest->tx_packets += src->tx_packets; 9299 dest->rx_bytes += src->rx_bytes; 9300 dest->tx_bytes += src->tx_bytes; 9301 dest->rx_errors += src->rx_errors; 9302 dest->tx_errors += src->tx_errors; 9303 dest->rx_dropped += src->rx_dropped; 9304 dest->tx_dropped += src->tx_dropped; 9305 dest->multicast += src->multicast; 9306 } 9307 9308 static int netdev_offload_xstats_get_used(struct net_device *dev, 9309 enum netdev_offload_xstats_type type, 9310 bool *p_used, 9311 struct netlink_ext_ack *extack) 9312 { 9313 struct netdev_notifier_offload_xstats_ru report_used = {}; 9314 struct netdev_notifier_offload_xstats_info info = { 9315 .info.dev = dev, 9316 .info.extack = extack, 9317 .type = type, 9318 .report_used = &report_used, 9319 }; 9320 int rc; 9321 9322 WARN_ON(!netdev_offload_xstats_enabled(dev, type)); 9323 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED, 9324 &info.info); 9325 *p_used = report_used.used; 9326 return notifier_to_errno(rc); 9327 } 9328 9329 static int netdev_offload_xstats_get_stats(struct net_device *dev, 9330 enum netdev_offload_xstats_type type, 9331 struct rtnl_hw_stats64 *p_stats, 9332 bool *p_used, 9333 struct netlink_ext_ack *extack) 9334 { 9335 struct netdev_notifier_offload_xstats_rd report_delta = {}; 9336 struct netdev_notifier_offload_xstats_info info = { 9337 .info.dev = dev, 9338 .info.extack = extack, 9339 .type = type, 9340 .report_delta = &report_delta, 9341 }; 9342 struct rtnl_hw_stats64 *stats; 9343 int rc; 9344 9345 stats = netdev_offload_xstats_get_ptr(dev, type); 9346 if (WARN_ON(!stats)) 9347 return -EINVAL; 9348 9349 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA, 9350 &info.info); 9351 9352 /* Cache whatever we got, even if there was an error, otherwise the 9353 * successful stats retrievals would get lost. 9354 */ 9355 netdev_hw_stats64_add(stats, &report_delta.stats); 9356 9357 if (p_stats) 9358 *p_stats = *stats; 9359 *p_used = report_delta.used; 9360 9361 return notifier_to_errno(rc); 9362 } 9363 9364 int netdev_offload_xstats_get(struct net_device *dev, 9365 enum netdev_offload_xstats_type type, 9366 struct rtnl_hw_stats64 *p_stats, bool *p_used, 9367 struct netlink_ext_ack *extack) 9368 { 9369 ASSERT_RTNL(); 9370 9371 if (p_stats) 9372 return netdev_offload_xstats_get_stats(dev, type, p_stats, 9373 p_used, extack); 9374 else 9375 return netdev_offload_xstats_get_used(dev, type, p_used, 9376 extack); 9377 } 9378 EXPORT_SYMBOL(netdev_offload_xstats_get); 9379 9380 void 9381 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta, 9382 const struct rtnl_hw_stats64 *stats) 9383 { 9384 report_delta->used = true; 9385 netdev_hw_stats64_add(&report_delta->stats, stats); 9386 } 9387 EXPORT_SYMBOL(netdev_offload_xstats_report_delta); 9388 9389 void 9390 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used) 9391 { 9392 report_used->used = true; 9393 } 9394 EXPORT_SYMBOL(netdev_offload_xstats_report_used); 9395 9396 void netdev_offload_xstats_push_delta(struct net_device *dev, 9397 enum netdev_offload_xstats_type type, 9398 const struct rtnl_hw_stats64 *p_stats) 9399 { 9400 struct rtnl_hw_stats64 *stats; 9401 9402 ASSERT_RTNL(); 9403 9404 stats = netdev_offload_xstats_get_ptr(dev, type); 9405 if (WARN_ON(!stats)) 9406 return; 9407 9408 netdev_hw_stats64_add(stats, p_stats); 9409 } 9410 EXPORT_SYMBOL(netdev_offload_xstats_push_delta); 9411 9412 /** 9413 * netdev_get_xmit_slave - Get the xmit slave of master device 9414 * @dev: device 9415 * @skb: The packet 9416 * @all_slaves: assume all the slaves are active 9417 * 9418 * The reference counters are not incremented so the caller must be 9419 * careful with locks. The caller must hold RCU lock. 9420 * %NULL is returned if no slave is found. 9421 */ 9422 9423 struct net_device *netdev_get_xmit_slave(struct net_device *dev, 9424 struct sk_buff *skb, 9425 bool all_slaves) 9426 { 9427 const struct net_device_ops *ops = dev->netdev_ops; 9428 9429 if (!ops->ndo_get_xmit_slave) 9430 return NULL; 9431 return ops->ndo_get_xmit_slave(dev, skb, all_slaves); 9432 } 9433 EXPORT_SYMBOL(netdev_get_xmit_slave); 9434 9435 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev, 9436 struct sock *sk) 9437 { 9438 const struct net_device_ops *ops = dev->netdev_ops; 9439 9440 if (!ops->ndo_sk_get_lower_dev) 9441 return NULL; 9442 return ops->ndo_sk_get_lower_dev(dev, sk); 9443 } 9444 9445 /** 9446 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket 9447 * @dev: device 9448 * @sk: the socket 9449 * 9450 * %NULL is returned if no lower device is found. 9451 */ 9452 9453 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev, 9454 struct sock *sk) 9455 { 9456 struct net_device *lower; 9457 9458 lower = netdev_sk_get_lower_dev(dev, sk); 9459 while (lower) { 9460 dev = lower; 9461 lower = netdev_sk_get_lower_dev(dev, sk); 9462 } 9463 9464 return dev; 9465 } 9466 EXPORT_SYMBOL(netdev_sk_get_lowest_dev); 9467 9468 static void netdev_adjacent_add_links(struct net_device *dev) 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_add(iter->dev, dev, 9478 &iter->dev->adj_list.lower); 9479 netdev_adjacent_sysfs_add(dev, iter->dev, 9480 &dev->adj_list.upper); 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_add(iter->dev, dev, 9487 &iter->dev->adj_list.upper); 9488 netdev_adjacent_sysfs_add(dev, iter->dev, 9489 &dev->adj_list.lower); 9490 } 9491 } 9492 9493 static void netdev_adjacent_del_links(struct net_device *dev) 9494 { 9495 struct netdev_adjacent *iter; 9496 9497 struct net *net = dev_net(dev); 9498 9499 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9500 if (!net_eq(net, dev_net(iter->dev))) 9501 continue; 9502 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9503 &iter->dev->adj_list.lower); 9504 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9505 &dev->adj_list.upper); 9506 } 9507 9508 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9509 if (!net_eq(net, dev_net(iter->dev))) 9510 continue; 9511 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9512 &iter->dev->adj_list.upper); 9513 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9514 &dev->adj_list.lower); 9515 } 9516 } 9517 9518 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname) 9519 { 9520 struct netdev_adjacent *iter; 9521 9522 struct net *net = dev_net(dev); 9523 9524 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9525 if (!net_eq(net, dev_net(iter->dev))) 9526 continue; 9527 netdev_adjacent_sysfs_del(iter->dev, oldname, 9528 &iter->dev->adj_list.lower); 9529 netdev_adjacent_sysfs_add(iter->dev, dev, 9530 &iter->dev->adj_list.lower); 9531 } 9532 9533 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9534 if (!net_eq(net, dev_net(iter->dev))) 9535 continue; 9536 netdev_adjacent_sysfs_del(iter->dev, oldname, 9537 &iter->dev->adj_list.upper); 9538 netdev_adjacent_sysfs_add(iter->dev, dev, 9539 &iter->dev->adj_list.upper); 9540 } 9541 } 9542 9543 void *netdev_lower_dev_get_private(struct net_device *dev, 9544 struct net_device *lower_dev) 9545 { 9546 struct netdev_adjacent *lower; 9547 9548 if (!lower_dev) 9549 return NULL; 9550 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower); 9551 if (!lower) 9552 return NULL; 9553 9554 return lower->private; 9555 } 9556 EXPORT_SYMBOL(netdev_lower_dev_get_private); 9557 9558 9559 /** 9560 * netdev_lower_state_changed - Dispatch event about lower device state change 9561 * @lower_dev: device 9562 * @lower_state_info: state to dispatch 9563 * 9564 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info. 9565 * The caller must hold the RTNL lock. 9566 */ 9567 void netdev_lower_state_changed(struct net_device *lower_dev, 9568 void *lower_state_info) 9569 { 9570 struct netdev_notifier_changelowerstate_info changelowerstate_info = { 9571 .info.dev = lower_dev, 9572 }; 9573 9574 ASSERT_RTNL(); 9575 changelowerstate_info.lower_state_info = lower_state_info; 9576 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, 9577 &changelowerstate_info.info); 9578 } 9579 EXPORT_SYMBOL(netdev_lower_state_changed); 9580 9581 static void dev_change_rx_flags(struct net_device *dev, int flags) 9582 { 9583 const struct net_device_ops *ops = dev->netdev_ops; 9584 9585 if (ops->ndo_change_rx_flags) 9586 ops->ndo_change_rx_flags(dev, flags); 9587 } 9588 9589 int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify) 9590 { 9591 unsigned int old_flags = dev->flags; 9592 unsigned int promiscuity, flags; 9593 kuid_t uid; 9594 kgid_t gid; 9595 9596 netdev_ops_assert_locked(dev); 9597 9598 promiscuity = dev->promiscuity + inc; 9599 if (promiscuity == 0) { 9600 /* 9601 * Avoid overflow. 9602 * If inc causes overflow, untouch promisc and return error. 9603 */ 9604 if (unlikely(inc > 0)) { 9605 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n"); 9606 return -EOVERFLOW; 9607 } 9608 flags = old_flags & ~IFF_PROMISC; 9609 } else { 9610 flags = old_flags | IFF_PROMISC; 9611 } 9612 WRITE_ONCE(dev->promiscuity, promiscuity); 9613 if (flags != old_flags) { 9614 WRITE_ONCE(dev->flags, flags); 9615 netdev_info(dev, "%s promiscuous mode\n", 9616 dev->flags & IFF_PROMISC ? "entered" : "left"); 9617 if (audit_enabled) { 9618 current_uid_gid(&uid, &gid); 9619 audit_log(audit_context(), GFP_ATOMIC, 9620 AUDIT_ANOM_PROMISCUOUS, 9621 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u", 9622 dev->name, (dev->flags & IFF_PROMISC), 9623 (old_flags & IFF_PROMISC), 9624 from_kuid(&init_user_ns, audit_get_loginuid(current)), 9625 from_kuid(&init_user_ns, uid), 9626 from_kgid(&init_user_ns, gid), 9627 audit_get_sessionid(current)); 9628 } 9629 9630 dev_change_rx_flags(dev, IFF_PROMISC); 9631 } 9632 if (notify) 9633 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL); 9634 return 0; 9635 } 9636 9637 int netif_set_promiscuity(struct net_device *dev, int inc) 9638 { 9639 unsigned int old_flags = dev->flags; 9640 int err; 9641 9642 err = __dev_set_promiscuity(dev, inc, true); 9643 if (err < 0) 9644 return err; 9645 if (dev->flags != old_flags) 9646 dev_set_rx_mode(dev); 9647 return err; 9648 } 9649 9650 int netif_set_allmulti(struct net_device *dev, int inc, bool notify) 9651 { 9652 unsigned int old_flags = dev->flags, old_gflags = dev->gflags; 9653 unsigned int allmulti, flags; 9654 9655 netdev_ops_assert_locked(dev); 9656 9657 allmulti = dev->allmulti + inc; 9658 if (allmulti == 0) { 9659 /* 9660 * Avoid overflow. 9661 * If inc causes overflow, untouch allmulti and return error. 9662 */ 9663 if (unlikely(inc > 0)) { 9664 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n"); 9665 return -EOVERFLOW; 9666 } 9667 flags = old_flags & ~IFF_ALLMULTI; 9668 } else { 9669 flags = old_flags | IFF_ALLMULTI; 9670 } 9671 WRITE_ONCE(dev->allmulti, allmulti); 9672 if (flags != old_flags) { 9673 WRITE_ONCE(dev->flags, flags); 9674 netdev_info(dev, "%s allmulticast mode\n", 9675 dev->flags & IFF_ALLMULTI ? "entered" : "left"); 9676 dev_change_rx_flags(dev, IFF_ALLMULTI); 9677 dev_set_rx_mode(dev); 9678 if (notify) 9679 __dev_notify_flags(dev, old_flags, 9680 dev->gflags ^ old_gflags, 0, NULL); 9681 } 9682 return 0; 9683 } 9684 9685 9686 /** 9687 * netif_get_flags() - get flags reported to userspace 9688 * @dev: device 9689 * 9690 * Get the combination of flag bits exported through APIs to userspace. 9691 */ 9692 unsigned int netif_get_flags(const struct net_device *dev) 9693 { 9694 unsigned int flags; 9695 9696 flags = (READ_ONCE(dev->flags) & ~(IFF_PROMISC | 9697 IFF_ALLMULTI | 9698 IFF_RUNNING | 9699 IFF_LOWER_UP | 9700 IFF_DORMANT)) | 9701 (READ_ONCE(dev->gflags) & (IFF_PROMISC | 9702 IFF_ALLMULTI)); 9703 9704 if (netif_running(dev)) { 9705 if (netif_oper_up(dev)) 9706 flags |= IFF_RUNNING; 9707 if (netif_carrier_ok(dev)) 9708 flags |= IFF_LOWER_UP; 9709 if (netif_dormant(dev)) 9710 flags |= IFF_DORMANT; 9711 } 9712 9713 return flags; 9714 } 9715 EXPORT_SYMBOL(netif_get_flags); 9716 9717 int __dev_change_flags(struct net_device *dev, unsigned int flags, 9718 struct netlink_ext_ack *extack) 9719 { 9720 unsigned int old_flags = dev->flags; 9721 int ret; 9722 9723 netdev_ops_assert_locked(dev); 9724 9725 /* 9726 * Set the flags on our device. 9727 */ 9728 9729 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP | 9730 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL | 9731 IFF_AUTOMEDIA)) | 9732 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC | 9733 IFF_ALLMULTI)); 9734 9735 /* 9736 * Load in the correct multicast list now the flags have changed. 9737 */ 9738 9739 if ((old_flags ^ flags) & IFF_MULTICAST) 9740 dev_change_rx_flags(dev, IFF_MULTICAST); 9741 9742 dev_set_rx_mode(dev); 9743 9744 /* 9745 * Have we downed the interface. We handle IFF_UP ourselves 9746 * according to user attempts to set it, rather than blindly 9747 * setting it. 9748 */ 9749 9750 ret = 0; 9751 if ((old_flags ^ flags) & IFF_UP) { 9752 if (old_flags & IFF_UP) 9753 __dev_close(dev); 9754 else 9755 ret = __dev_open(dev, extack); 9756 } 9757 9758 if ((flags ^ dev->gflags) & IFF_PROMISC) { 9759 int inc = (flags & IFF_PROMISC) ? 1 : -1; 9760 old_flags = dev->flags; 9761 9762 dev->gflags ^= IFF_PROMISC; 9763 9764 if (__dev_set_promiscuity(dev, inc, false) >= 0) 9765 if (dev->flags != old_flags) 9766 dev_set_rx_mode(dev); 9767 } 9768 9769 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI 9770 * is important. Some (broken) drivers set IFF_PROMISC, when 9771 * IFF_ALLMULTI is requested not asking us and not reporting. 9772 */ 9773 if ((flags ^ dev->gflags) & IFF_ALLMULTI) { 9774 int inc = (flags & IFF_ALLMULTI) ? 1 : -1; 9775 9776 dev->gflags ^= IFF_ALLMULTI; 9777 netif_set_allmulti(dev, inc, false); 9778 } 9779 9780 return ret; 9781 } 9782 9783 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags, 9784 unsigned int gchanges, u32 portid, 9785 const struct nlmsghdr *nlh) 9786 { 9787 unsigned int changes = dev->flags ^ old_flags; 9788 9789 if (gchanges) 9790 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh); 9791 9792 if (changes & IFF_UP) { 9793 if (dev->flags & IFF_UP) 9794 call_netdevice_notifiers(NETDEV_UP, dev); 9795 else 9796 call_netdevice_notifiers(NETDEV_DOWN, dev); 9797 } 9798 9799 if (dev->flags & IFF_UP && 9800 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) { 9801 struct netdev_notifier_change_info change_info = { 9802 .info = { 9803 .dev = dev, 9804 }, 9805 .flags_changed = changes, 9806 }; 9807 9808 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info); 9809 } 9810 } 9811 9812 int netif_change_flags(struct net_device *dev, unsigned int flags, 9813 struct netlink_ext_ack *extack) 9814 { 9815 int ret; 9816 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags; 9817 9818 ret = __dev_change_flags(dev, flags, extack); 9819 if (ret < 0) 9820 return ret; 9821 9822 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags); 9823 __dev_notify_flags(dev, old_flags, changes, 0, NULL); 9824 return ret; 9825 } 9826 9827 int __netif_set_mtu(struct net_device *dev, int new_mtu) 9828 { 9829 const struct net_device_ops *ops = dev->netdev_ops; 9830 9831 if (ops->ndo_change_mtu) 9832 return ops->ndo_change_mtu(dev, new_mtu); 9833 9834 /* Pairs with all the lockless reads of dev->mtu in the stack */ 9835 WRITE_ONCE(dev->mtu, new_mtu); 9836 return 0; 9837 } 9838 EXPORT_SYMBOL_NS_GPL(__netif_set_mtu, "NETDEV_INTERNAL"); 9839 9840 int dev_validate_mtu(struct net_device *dev, int new_mtu, 9841 struct netlink_ext_ack *extack) 9842 { 9843 /* MTU must be positive, and in range */ 9844 if (new_mtu < 0 || new_mtu < dev->min_mtu) { 9845 NL_SET_ERR_MSG(extack, "mtu less than device minimum"); 9846 return -EINVAL; 9847 } 9848 9849 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) { 9850 NL_SET_ERR_MSG(extack, "mtu greater than device maximum"); 9851 return -EINVAL; 9852 } 9853 return 0; 9854 } 9855 9856 /** 9857 * netif_set_mtu_ext() - Change maximum transfer unit 9858 * @dev: device 9859 * @new_mtu: new transfer unit 9860 * @extack: netlink extended ack 9861 * 9862 * Change the maximum transfer size of the network device. 9863 * 9864 * Return: 0 on success, -errno on failure. 9865 */ 9866 int netif_set_mtu_ext(struct net_device *dev, int new_mtu, 9867 struct netlink_ext_ack *extack) 9868 { 9869 int err, orig_mtu; 9870 9871 netdev_ops_assert_locked(dev); 9872 9873 if (new_mtu == dev->mtu) 9874 return 0; 9875 9876 err = dev_validate_mtu(dev, new_mtu, extack); 9877 if (err) 9878 return err; 9879 9880 if (!netif_device_present(dev)) 9881 return -ENODEV; 9882 9883 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev); 9884 err = notifier_to_errno(err); 9885 if (err) 9886 return err; 9887 9888 orig_mtu = dev->mtu; 9889 err = __netif_set_mtu(dev, new_mtu); 9890 9891 if (!err) { 9892 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9893 orig_mtu); 9894 err = notifier_to_errno(err); 9895 if (err) { 9896 /* setting mtu back and notifying everyone again, 9897 * so that they have a chance to revert changes. 9898 */ 9899 __netif_set_mtu(dev, orig_mtu); 9900 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9901 new_mtu); 9902 } 9903 } 9904 return err; 9905 } 9906 9907 int netif_set_mtu(struct net_device *dev, int new_mtu) 9908 { 9909 struct netlink_ext_ack extack; 9910 int err; 9911 9912 memset(&extack, 0, sizeof(extack)); 9913 err = netif_set_mtu_ext(dev, new_mtu, &extack); 9914 if (err && extack._msg) 9915 net_err_ratelimited("%s: %s\n", dev->name, extack._msg); 9916 return err; 9917 } 9918 EXPORT_SYMBOL(netif_set_mtu); 9919 9920 int netif_change_tx_queue_len(struct net_device *dev, unsigned long new_len) 9921 { 9922 unsigned int orig_len = dev->tx_queue_len; 9923 int res; 9924 9925 if (new_len != (unsigned int)new_len) 9926 return -ERANGE; 9927 9928 if (new_len != orig_len) { 9929 WRITE_ONCE(dev->tx_queue_len, new_len); 9930 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev); 9931 res = notifier_to_errno(res); 9932 if (res) 9933 goto err_rollback; 9934 res = dev_qdisc_change_tx_queue_len(dev); 9935 if (res) 9936 goto err_rollback; 9937 } 9938 9939 return 0; 9940 9941 err_rollback: 9942 netdev_err(dev, "refused to change device tx_queue_len\n"); 9943 WRITE_ONCE(dev->tx_queue_len, orig_len); 9944 return res; 9945 } 9946 9947 void netif_set_group(struct net_device *dev, int new_group) 9948 { 9949 dev->group = new_group; 9950 } 9951 9952 /** 9953 * netif_pre_changeaddr_notify() - Call NETDEV_PRE_CHANGEADDR. 9954 * @dev: device 9955 * @addr: new address 9956 * @extack: netlink extended ack 9957 * 9958 * Return: 0 on success, -errno on failure. 9959 */ 9960 int netif_pre_changeaddr_notify(struct net_device *dev, const char *addr, 9961 struct netlink_ext_ack *extack) 9962 { 9963 struct netdev_notifier_pre_changeaddr_info info = { 9964 .info.dev = dev, 9965 .info.extack = extack, 9966 .dev_addr = addr, 9967 }; 9968 int rc; 9969 9970 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info); 9971 return notifier_to_errno(rc); 9972 } 9973 EXPORT_SYMBOL_NS_GPL(netif_pre_changeaddr_notify, "NETDEV_INTERNAL"); 9974 9975 int netif_set_mac_address(struct net_device *dev, struct sockaddr_storage *ss, 9976 struct netlink_ext_ack *extack) 9977 { 9978 const struct net_device_ops *ops = dev->netdev_ops; 9979 int err; 9980 9981 if (!ops->ndo_set_mac_address) 9982 return -EOPNOTSUPP; 9983 if (ss->ss_family != dev->type) 9984 return -EINVAL; 9985 if (!netif_device_present(dev)) 9986 return -ENODEV; 9987 err = netif_pre_changeaddr_notify(dev, ss->__data, extack); 9988 if (err) 9989 return err; 9990 if (memcmp(dev->dev_addr, ss->__data, dev->addr_len)) { 9991 err = ops->ndo_set_mac_address(dev, ss); 9992 if (err) 9993 return err; 9994 } 9995 dev->addr_assign_type = NET_ADDR_SET; 9996 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev); 9997 add_device_randomness(dev->dev_addr, dev->addr_len); 9998 return 0; 9999 } 10000 10001 DECLARE_RWSEM(dev_addr_sem); 10002 10003 /* "sa" is a true struct sockaddr with limited "sa_data" member. */ 10004 int netif_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name) 10005 { 10006 size_t size = sizeof(sa->sa_data); 10007 struct net_device *dev; 10008 int ret = 0; 10009 10010 down_read(&dev_addr_sem); 10011 rcu_read_lock(); 10012 10013 dev = dev_get_by_name_rcu(net, dev_name); 10014 if (!dev) { 10015 ret = -ENODEV; 10016 goto unlock; 10017 } 10018 if (!dev->addr_len) 10019 memset(sa->sa_data, 0, size); 10020 else 10021 memcpy(sa->sa_data, dev->dev_addr, 10022 min_t(size_t, size, dev->addr_len)); 10023 sa->sa_family = dev->type; 10024 10025 unlock: 10026 rcu_read_unlock(); 10027 up_read(&dev_addr_sem); 10028 return ret; 10029 } 10030 EXPORT_SYMBOL_NS_GPL(netif_get_mac_address, "NETDEV_INTERNAL"); 10031 10032 int netif_change_carrier(struct net_device *dev, bool new_carrier) 10033 { 10034 const struct net_device_ops *ops = dev->netdev_ops; 10035 10036 if (!ops->ndo_change_carrier) 10037 return -EOPNOTSUPP; 10038 if (!netif_device_present(dev)) 10039 return -ENODEV; 10040 return ops->ndo_change_carrier(dev, new_carrier); 10041 } 10042 10043 /** 10044 * dev_get_phys_port_id - Get device physical port ID 10045 * @dev: device 10046 * @ppid: port ID 10047 * 10048 * Get device physical port ID 10049 */ 10050 int dev_get_phys_port_id(struct net_device *dev, 10051 struct netdev_phys_item_id *ppid) 10052 { 10053 const struct net_device_ops *ops = dev->netdev_ops; 10054 10055 if (!ops->ndo_get_phys_port_id) 10056 return -EOPNOTSUPP; 10057 return ops->ndo_get_phys_port_id(dev, ppid); 10058 } 10059 10060 /** 10061 * dev_get_phys_port_name - Get device physical port name 10062 * @dev: device 10063 * @name: port name 10064 * @len: limit of bytes to copy to name 10065 * 10066 * Get device physical port name 10067 */ 10068 int dev_get_phys_port_name(struct net_device *dev, 10069 char *name, size_t len) 10070 { 10071 const struct net_device_ops *ops = dev->netdev_ops; 10072 int err; 10073 10074 if (ops->ndo_get_phys_port_name) { 10075 err = ops->ndo_get_phys_port_name(dev, name, len); 10076 if (err != -EOPNOTSUPP) 10077 return err; 10078 } 10079 return devlink_compat_phys_port_name_get(dev, name, len); 10080 } 10081 10082 /** 10083 * netif_get_port_parent_id() - Get the device's port parent identifier 10084 * @dev: network device 10085 * @ppid: pointer to a storage for the port's parent identifier 10086 * @recurse: allow/disallow recursion to lower devices 10087 * 10088 * Get the devices's port parent identifier. 10089 * 10090 * Return: 0 on success, -errno on failure. 10091 */ 10092 int netif_get_port_parent_id(struct net_device *dev, 10093 struct netdev_phys_item_id *ppid, bool recurse) 10094 { 10095 const struct net_device_ops *ops = dev->netdev_ops; 10096 struct netdev_phys_item_id first = { }; 10097 struct net_device *lower_dev; 10098 struct list_head *iter; 10099 int err; 10100 10101 if (ops->ndo_get_port_parent_id) { 10102 err = ops->ndo_get_port_parent_id(dev, ppid); 10103 if (err != -EOPNOTSUPP) 10104 return err; 10105 } 10106 10107 err = devlink_compat_switch_id_get(dev, ppid); 10108 if (!recurse || err != -EOPNOTSUPP) 10109 return err; 10110 10111 netdev_for_each_lower_dev(dev, lower_dev, iter) { 10112 err = netif_get_port_parent_id(lower_dev, ppid, true); 10113 if (err) 10114 break; 10115 if (!first.id_len) 10116 first = *ppid; 10117 else if (memcmp(&first, ppid, sizeof(*ppid))) 10118 return -EOPNOTSUPP; 10119 } 10120 10121 return err; 10122 } 10123 EXPORT_SYMBOL(netif_get_port_parent_id); 10124 10125 /** 10126 * netdev_port_same_parent_id - Indicate if two network devices have 10127 * the same port parent identifier 10128 * @a: first network device 10129 * @b: second network device 10130 */ 10131 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b) 10132 { 10133 struct netdev_phys_item_id a_id = { }; 10134 struct netdev_phys_item_id b_id = { }; 10135 10136 if (netif_get_port_parent_id(a, &a_id, true) || 10137 netif_get_port_parent_id(b, &b_id, true)) 10138 return false; 10139 10140 return netdev_phys_item_id_same(&a_id, &b_id); 10141 } 10142 EXPORT_SYMBOL(netdev_port_same_parent_id); 10143 10144 int netif_change_proto_down(struct net_device *dev, bool proto_down) 10145 { 10146 if (!dev->change_proto_down) 10147 return -EOPNOTSUPP; 10148 if (!netif_device_present(dev)) 10149 return -ENODEV; 10150 if (proto_down) 10151 netif_carrier_off(dev); 10152 else 10153 netif_carrier_on(dev); 10154 WRITE_ONCE(dev->proto_down, proto_down); 10155 return 0; 10156 } 10157 10158 /** 10159 * netdev_change_proto_down_reason_locked - proto down reason 10160 * 10161 * @dev: device 10162 * @mask: proto down mask 10163 * @value: proto down value 10164 */ 10165 void netdev_change_proto_down_reason_locked(struct net_device *dev, 10166 unsigned long mask, u32 value) 10167 { 10168 u32 proto_down_reason; 10169 int b; 10170 10171 if (!mask) { 10172 proto_down_reason = value; 10173 } else { 10174 proto_down_reason = dev->proto_down_reason; 10175 for_each_set_bit(b, &mask, 32) { 10176 if (value & (1 << b)) 10177 proto_down_reason |= BIT(b); 10178 else 10179 proto_down_reason &= ~BIT(b); 10180 } 10181 } 10182 WRITE_ONCE(dev->proto_down_reason, proto_down_reason); 10183 } 10184 10185 struct bpf_xdp_link { 10186 struct bpf_link link; 10187 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */ 10188 int flags; 10189 }; 10190 10191 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags) 10192 { 10193 if (flags & XDP_FLAGS_HW_MODE) 10194 return XDP_MODE_HW; 10195 if (flags & XDP_FLAGS_DRV_MODE) 10196 return XDP_MODE_DRV; 10197 if (flags & XDP_FLAGS_SKB_MODE) 10198 return XDP_MODE_SKB; 10199 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB; 10200 } 10201 10202 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode) 10203 { 10204 switch (mode) { 10205 case XDP_MODE_SKB: 10206 return generic_xdp_install; 10207 case XDP_MODE_DRV: 10208 case XDP_MODE_HW: 10209 return dev->netdev_ops->ndo_bpf; 10210 default: 10211 return NULL; 10212 } 10213 } 10214 10215 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev, 10216 enum bpf_xdp_mode mode) 10217 { 10218 return dev->xdp_state[mode].link; 10219 } 10220 10221 static struct bpf_prog *dev_xdp_prog(struct net_device *dev, 10222 enum bpf_xdp_mode mode) 10223 { 10224 struct bpf_xdp_link *link = dev_xdp_link(dev, mode); 10225 10226 if (link) 10227 return link->link.prog; 10228 return dev->xdp_state[mode].prog; 10229 } 10230 10231 u8 dev_xdp_prog_count(struct net_device *dev) 10232 { 10233 u8 count = 0; 10234 int i; 10235 10236 for (i = 0; i < __MAX_XDP_MODE; i++) 10237 if (dev->xdp_state[i].prog || dev->xdp_state[i].link) 10238 count++; 10239 return count; 10240 } 10241 EXPORT_SYMBOL_GPL(dev_xdp_prog_count); 10242 10243 u8 dev_xdp_sb_prog_count(struct net_device *dev) 10244 { 10245 u8 count = 0; 10246 int i; 10247 10248 for (i = 0; i < __MAX_XDP_MODE; i++) 10249 if (dev->xdp_state[i].prog && 10250 !dev->xdp_state[i].prog->aux->xdp_has_frags) 10251 count++; 10252 return count; 10253 } 10254 10255 int netif_xdp_propagate(struct net_device *dev, struct netdev_bpf *bpf) 10256 { 10257 if (!dev->netdev_ops->ndo_bpf) 10258 return -EOPNOTSUPP; 10259 10260 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 10261 bpf->command == XDP_SETUP_PROG && 10262 bpf->prog && !bpf->prog->aux->xdp_has_frags) { 10263 NL_SET_ERR_MSG(bpf->extack, 10264 "unable to propagate XDP to device using tcp-data-split"); 10265 return -EBUSY; 10266 } 10267 10268 if (dev_get_min_mp_channel_count(dev)) { 10269 NL_SET_ERR_MSG(bpf->extack, "unable to propagate XDP to device using memory provider"); 10270 return -EBUSY; 10271 } 10272 10273 return dev->netdev_ops->ndo_bpf(dev, bpf); 10274 } 10275 EXPORT_SYMBOL_GPL(netif_xdp_propagate); 10276 10277 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode) 10278 { 10279 struct bpf_prog *prog = dev_xdp_prog(dev, mode); 10280 10281 return prog ? prog->aux->id : 0; 10282 } 10283 10284 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode, 10285 struct bpf_xdp_link *link) 10286 { 10287 dev->xdp_state[mode].link = link; 10288 dev->xdp_state[mode].prog = NULL; 10289 } 10290 10291 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode, 10292 struct bpf_prog *prog) 10293 { 10294 dev->xdp_state[mode].link = NULL; 10295 dev->xdp_state[mode].prog = prog; 10296 } 10297 10298 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode, 10299 bpf_op_t bpf_op, struct netlink_ext_ack *extack, 10300 u32 flags, struct bpf_prog *prog) 10301 { 10302 struct netdev_bpf xdp; 10303 int err; 10304 10305 netdev_ops_assert_locked(dev); 10306 10307 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 10308 prog && !prog->aux->xdp_has_frags) { 10309 NL_SET_ERR_MSG(extack, "unable to install XDP to device using tcp-data-split"); 10310 return -EBUSY; 10311 } 10312 10313 if (dev_get_min_mp_channel_count(dev)) { 10314 NL_SET_ERR_MSG(extack, "unable to install XDP to device using memory provider"); 10315 return -EBUSY; 10316 } 10317 10318 memset(&xdp, 0, sizeof(xdp)); 10319 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG; 10320 xdp.extack = extack; 10321 xdp.flags = flags; 10322 xdp.prog = prog; 10323 10324 /* Drivers assume refcnt is already incremented (i.e, prog pointer is 10325 * "moved" into driver), so they don't increment it on their own, but 10326 * they do decrement refcnt when program is detached or replaced. 10327 * Given net_device also owns link/prog, we need to bump refcnt here 10328 * to prevent drivers from underflowing it. 10329 */ 10330 if (prog) 10331 bpf_prog_inc(prog); 10332 err = bpf_op(dev, &xdp); 10333 if (err) { 10334 if (prog) 10335 bpf_prog_put(prog); 10336 return err; 10337 } 10338 10339 if (mode != XDP_MODE_HW) 10340 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog); 10341 10342 return 0; 10343 } 10344 10345 static void dev_xdp_uninstall(struct net_device *dev) 10346 { 10347 struct bpf_xdp_link *link; 10348 struct bpf_prog *prog; 10349 enum bpf_xdp_mode mode; 10350 bpf_op_t bpf_op; 10351 10352 ASSERT_RTNL(); 10353 10354 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) { 10355 prog = dev_xdp_prog(dev, mode); 10356 if (!prog) 10357 continue; 10358 10359 bpf_op = dev_xdp_bpf_op(dev, mode); 10360 if (!bpf_op) 10361 continue; 10362 10363 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 10364 10365 /* auto-detach link from net device */ 10366 link = dev_xdp_link(dev, mode); 10367 if (link) 10368 link->dev = NULL; 10369 else 10370 bpf_prog_put(prog); 10371 10372 dev_xdp_set_link(dev, mode, NULL); 10373 } 10374 } 10375 10376 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack, 10377 struct bpf_xdp_link *link, struct bpf_prog *new_prog, 10378 struct bpf_prog *old_prog, u32 flags) 10379 { 10380 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES); 10381 struct bpf_prog *cur_prog; 10382 struct net_device *upper; 10383 struct list_head *iter; 10384 enum bpf_xdp_mode mode; 10385 bpf_op_t bpf_op; 10386 int err; 10387 10388 ASSERT_RTNL(); 10389 10390 /* either link or prog attachment, never both */ 10391 if (link && (new_prog || old_prog)) 10392 return -EINVAL; 10393 /* link supports only XDP mode flags */ 10394 if (link && (flags & ~XDP_FLAGS_MODES)) { 10395 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment"); 10396 return -EINVAL; 10397 } 10398 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */ 10399 if (num_modes > 1) { 10400 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set"); 10401 return -EINVAL; 10402 } 10403 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */ 10404 if (!num_modes && dev_xdp_prog_count(dev) > 1) { 10405 NL_SET_ERR_MSG(extack, 10406 "More than one program loaded, unset mode is ambiguous"); 10407 return -EINVAL; 10408 } 10409 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */ 10410 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) { 10411 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified"); 10412 return -EINVAL; 10413 } 10414 10415 mode = dev_xdp_mode(dev, flags); 10416 /* can't replace attached link */ 10417 if (dev_xdp_link(dev, mode)) { 10418 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link"); 10419 return -EBUSY; 10420 } 10421 10422 /* don't allow if an upper device already has a program */ 10423 netdev_for_each_upper_dev_rcu(dev, upper, iter) { 10424 if (dev_xdp_prog_count(upper) > 0) { 10425 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program"); 10426 return -EEXIST; 10427 } 10428 } 10429 10430 cur_prog = dev_xdp_prog(dev, mode); 10431 /* can't replace attached prog with link */ 10432 if (link && cur_prog) { 10433 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link"); 10434 return -EBUSY; 10435 } 10436 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) { 10437 NL_SET_ERR_MSG(extack, "Active program does not match expected"); 10438 return -EEXIST; 10439 } 10440 10441 /* put effective new program into new_prog */ 10442 if (link) 10443 new_prog = link->link.prog; 10444 10445 if (new_prog) { 10446 bool offload = mode == XDP_MODE_HW; 10447 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB 10448 ? XDP_MODE_DRV : XDP_MODE_SKB; 10449 10450 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) { 10451 NL_SET_ERR_MSG(extack, "XDP program already attached"); 10452 return -EBUSY; 10453 } 10454 if (!offload && dev_xdp_prog(dev, other_mode)) { 10455 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time"); 10456 return -EEXIST; 10457 } 10458 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) { 10459 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported"); 10460 return -EINVAL; 10461 } 10462 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) { 10463 NL_SET_ERR_MSG(extack, "Program bound to different device"); 10464 return -EINVAL; 10465 } 10466 if (bpf_prog_is_dev_bound(new_prog->aux) && mode == XDP_MODE_SKB) { 10467 NL_SET_ERR_MSG(extack, "Can't attach device-bound programs in generic mode"); 10468 return -EINVAL; 10469 } 10470 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) { 10471 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device"); 10472 return -EINVAL; 10473 } 10474 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) { 10475 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device"); 10476 return -EINVAL; 10477 } 10478 } 10479 10480 /* don't call drivers if the effective program didn't change */ 10481 if (new_prog != cur_prog) { 10482 bpf_op = dev_xdp_bpf_op(dev, mode); 10483 if (!bpf_op) { 10484 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode"); 10485 return -EOPNOTSUPP; 10486 } 10487 10488 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog); 10489 if (err) 10490 return err; 10491 } 10492 10493 if (link) 10494 dev_xdp_set_link(dev, mode, link); 10495 else 10496 dev_xdp_set_prog(dev, mode, new_prog); 10497 if (cur_prog) 10498 bpf_prog_put(cur_prog); 10499 10500 return 0; 10501 } 10502 10503 static int dev_xdp_attach_link(struct net_device *dev, 10504 struct netlink_ext_ack *extack, 10505 struct bpf_xdp_link *link) 10506 { 10507 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags); 10508 } 10509 10510 static int dev_xdp_detach_link(struct net_device *dev, 10511 struct netlink_ext_ack *extack, 10512 struct bpf_xdp_link *link) 10513 { 10514 enum bpf_xdp_mode mode; 10515 bpf_op_t bpf_op; 10516 10517 ASSERT_RTNL(); 10518 10519 mode = dev_xdp_mode(dev, link->flags); 10520 if (dev_xdp_link(dev, mode) != link) 10521 return -EINVAL; 10522 10523 bpf_op = dev_xdp_bpf_op(dev, mode); 10524 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 10525 dev_xdp_set_link(dev, mode, NULL); 10526 return 0; 10527 } 10528 10529 static void bpf_xdp_link_release(struct bpf_link *link) 10530 { 10531 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10532 10533 rtnl_lock(); 10534 10535 /* if racing with net_device's tear down, xdp_link->dev might be 10536 * already NULL, in which case link was already auto-detached 10537 */ 10538 if (xdp_link->dev) { 10539 netdev_lock_ops(xdp_link->dev); 10540 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link)); 10541 netdev_unlock_ops(xdp_link->dev); 10542 xdp_link->dev = NULL; 10543 } 10544 10545 rtnl_unlock(); 10546 } 10547 10548 static int bpf_xdp_link_detach(struct bpf_link *link) 10549 { 10550 bpf_xdp_link_release(link); 10551 return 0; 10552 } 10553 10554 static void bpf_xdp_link_dealloc(struct bpf_link *link) 10555 { 10556 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10557 10558 kfree(xdp_link); 10559 } 10560 10561 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link, 10562 struct seq_file *seq) 10563 { 10564 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10565 u32 ifindex = 0; 10566 10567 rtnl_lock(); 10568 if (xdp_link->dev) 10569 ifindex = xdp_link->dev->ifindex; 10570 rtnl_unlock(); 10571 10572 seq_printf(seq, "ifindex:\t%u\n", ifindex); 10573 } 10574 10575 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link, 10576 struct bpf_link_info *info) 10577 { 10578 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10579 u32 ifindex = 0; 10580 10581 rtnl_lock(); 10582 if (xdp_link->dev) 10583 ifindex = xdp_link->dev->ifindex; 10584 rtnl_unlock(); 10585 10586 info->xdp.ifindex = ifindex; 10587 return 0; 10588 } 10589 10590 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog, 10591 struct bpf_prog *old_prog) 10592 { 10593 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10594 enum bpf_xdp_mode mode; 10595 bpf_op_t bpf_op; 10596 int err = 0; 10597 10598 rtnl_lock(); 10599 10600 /* link might have been auto-released already, so fail */ 10601 if (!xdp_link->dev) { 10602 err = -ENOLINK; 10603 goto out_unlock; 10604 } 10605 10606 if (old_prog && link->prog != old_prog) { 10607 err = -EPERM; 10608 goto out_unlock; 10609 } 10610 old_prog = link->prog; 10611 if (old_prog->type != new_prog->type || 10612 old_prog->expected_attach_type != new_prog->expected_attach_type) { 10613 err = -EINVAL; 10614 goto out_unlock; 10615 } 10616 10617 if (old_prog == new_prog) { 10618 /* no-op, don't disturb drivers */ 10619 bpf_prog_put(new_prog); 10620 goto out_unlock; 10621 } 10622 10623 netdev_lock_ops(xdp_link->dev); 10624 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags); 10625 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode); 10626 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL, 10627 xdp_link->flags, new_prog); 10628 netdev_unlock_ops(xdp_link->dev); 10629 if (err) 10630 goto out_unlock; 10631 10632 old_prog = xchg(&link->prog, new_prog); 10633 bpf_prog_put(old_prog); 10634 10635 out_unlock: 10636 rtnl_unlock(); 10637 return err; 10638 } 10639 10640 static const struct bpf_link_ops bpf_xdp_link_lops = { 10641 .release = bpf_xdp_link_release, 10642 .dealloc = bpf_xdp_link_dealloc, 10643 .detach = bpf_xdp_link_detach, 10644 .show_fdinfo = bpf_xdp_link_show_fdinfo, 10645 .fill_link_info = bpf_xdp_link_fill_link_info, 10646 .update_prog = bpf_xdp_link_update, 10647 }; 10648 10649 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog) 10650 { 10651 struct net *net = current->nsproxy->net_ns; 10652 struct bpf_link_primer link_primer; 10653 struct netlink_ext_ack extack = {}; 10654 struct bpf_xdp_link *link; 10655 struct net_device *dev; 10656 int err, fd; 10657 10658 rtnl_lock(); 10659 dev = dev_get_by_index(net, attr->link_create.target_ifindex); 10660 if (!dev) { 10661 rtnl_unlock(); 10662 return -EINVAL; 10663 } 10664 10665 link = kzalloc_obj(*link, GFP_USER); 10666 if (!link) { 10667 err = -ENOMEM; 10668 goto unlock; 10669 } 10670 10671 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog, 10672 attr->link_create.attach_type); 10673 link->dev = dev; 10674 link->flags = attr->link_create.flags; 10675 10676 err = bpf_link_prime(&link->link, &link_primer); 10677 if (err) { 10678 kfree(link); 10679 goto unlock; 10680 } 10681 10682 netdev_lock_ops(dev); 10683 err = dev_xdp_attach_link(dev, &extack, link); 10684 netdev_unlock_ops(dev); 10685 rtnl_unlock(); 10686 10687 if (err) { 10688 link->dev = NULL; 10689 bpf_link_cleanup(&link_primer); 10690 trace_bpf_xdp_link_attach_failed(extack._msg); 10691 goto out_put_dev; 10692 } 10693 10694 fd = bpf_link_settle(&link_primer); 10695 /* link itself doesn't hold dev's refcnt to not complicate shutdown */ 10696 dev_put(dev); 10697 return fd; 10698 10699 unlock: 10700 rtnl_unlock(); 10701 10702 out_put_dev: 10703 dev_put(dev); 10704 return err; 10705 } 10706 10707 /** 10708 * dev_change_xdp_fd - set or clear a bpf program for a device rx path 10709 * @dev: device 10710 * @extack: netlink extended ack 10711 * @fd: new program fd or negative value to clear 10712 * @expected_fd: old program fd that userspace expects to replace or clear 10713 * @flags: xdp-related flags 10714 * 10715 * Set or clear a bpf program for a device 10716 */ 10717 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, 10718 int fd, int expected_fd, u32 flags) 10719 { 10720 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags); 10721 struct bpf_prog *new_prog = NULL, *old_prog = NULL; 10722 int err; 10723 10724 ASSERT_RTNL(); 10725 10726 if (fd >= 0) { 10727 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP, 10728 mode != XDP_MODE_SKB); 10729 if (IS_ERR(new_prog)) 10730 return PTR_ERR(new_prog); 10731 } 10732 10733 if (expected_fd >= 0) { 10734 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP, 10735 mode != XDP_MODE_SKB); 10736 if (IS_ERR(old_prog)) { 10737 err = PTR_ERR(old_prog); 10738 old_prog = NULL; 10739 goto err_out; 10740 } 10741 } 10742 10743 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags); 10744 10745 err_out: 10746 if (err && new_prog) 10747 bpf_prog_put(new_prog); 10748 if (old_prog) 10749 bpf_prog_put(old_prog); 10750 return err; 10751 } 10752 10753 u32 dev_get_min_mp_channel_count(const struct net_device *dev) 10754 { 10755 int i; 10756 10757 netdev_ops_assert_locked(dev); 10758 10759 for (i = dev->real_num_rx_queues - 1; i >= 0; i--) 10760 if (dev->_rx[i].mp_params.mp_priv) 10761 /* The channel count is the idx plus 1. */ 10762 return i + 1; 10763 10764 return 0; 10765 } 10766 10767 /** 10768 * dev_index_reserve() - allocate an ifindex in a namespace 10769 * @net: the applicable net namespace 10770 * @ifindex: requested ifindex, pass %0 to get one allocated 10771 * 10772 * Allocate a ifindex for a new device. Caller must either use the ifindex 10773 * to store the device (via list_netdevice()) or call dev_index_release() 10774 * to give the index up. 10775 * 10776 * Return: a suitable unique value for a new device interface number or -errno. 10777 */ 10778 static int dev_index_reserve(struct net *net, u32 ifindex) 10779 { 10780 int err; 10781 10782 if (ifindex > INT_MAX) { 10783 DEBUG_NET_WARN_ON_ONCE(1); 10784 return -EINVAL; 10785 } 10786 10787 if (!ifindex) 10788 err = xa_alloc_cyclic(&net->dev_by_index, &ifindex, NULL, 10789 xa_limit_31b, &net->ifindex, GFP_KERNEL); 10790 else 10791 err = xa_insert(&net->dev_by_index, ifindex, NULL, GFP_KERNEL); 10792 if (err < 0) 10793 return err; 10794 10795 return ifindex; 10796 } 10797 10798 static void dev_index_release(struct net *net, int ifindex) 10799 { 10800 /* Expect only unused indexes, unlist_netdevice() removes the used */ 10801 WARN_ON(xa_erase(&net->dev_by_index, ifindex)); 10802 } 10803 10804 static bool from_cleanup_net(void) 10805 { 10806 #ifdef CONFIG_NET_NS 10807 return current == READ_ONCE(cleanup_net_task); 10808 #else 10809 return false; 10810 #endif 10811 } 10812 10813 /* Delayed registration/unregisteration */ 10814 LIST_HEAD(net_todo_list); 10815 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq); 10816 atomic_t dev_unreg_count = ATOMIC_INIT(0); 10817 10818 static void net_set_todo(struct net_device *dev) 10819 { 10820 list_add_tail(&dev->todo_list, &net_todo_list); 10821 } 10822 10823 static netdev_features_t netdev_sync_upper_features(struct net_device *lower, 10824 struct net_device *upper, netdev_features_t features) 10825 { 10826 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10827 netdev_features_t feature; 10828 int feature_bit; 10829 10830 for_each_netdev_feature(upper_disables, feature_bit) { 10831 feature = __NETIF_F_BIT(feature_bit); 10832 if (!(upper->wanted_features & feature) 10833 && (features & feature)) { 10834 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n", 10835 &feature, upper->name); 10836 features &= ~feature; 10837 } 10838 } 10839 10840 return features; 10841 } 10842 10843 static void netdev_sync_lower_features(struct net_device *upper, 10844 struct net_device *lower, netdev_features_t features) 10845 { 10846 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10847 netdev_features_t feature; 10848 int feature_bit; 10849 10850 for_each_netdev_feature(upper_disables, feature_bit) { 10851 feature = __NETIF_F_BIT(feature_bit); 10852 if (!(features & feature) && (lower->features & feature)) { 10853 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n", 10854 &feature, lower->name); 10855 netdev_lock_ops(lower); 10856 lower->wanted_features &= ~feature; 10857 __netdev_update_features(lower); 10858 10859 if (unlikely(lower->features & feature)) 10860 netdev_WARN(upper, "failed to disable %pNF on %s!\n", 10861 &feature, lower->name); 10862 else 10863 netdev_features_change(lower); 10864 netdev_unlock_ops(lower); 10865 } 10866 } 10867 } 10868 10869 static bool netdev_has_ip_or_hw_csum(netdev_features_t features) 10870 { 10871 netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 10872 bool ip_csum = (features & ip_csum_mask) == ip_csum_mask; 10873 bool hw_csum = features & NETIF_F_HW_CSUM; 10874 10875 return ip_csum || hw_csum; 10876 } 10877 10878 static netdev_features_t netdev_fix_features(struct net_device *dev, 10879 netdev_features_t features) 10880 { 10881 /* Fix illegal checksum combinations */ 10882 if ((features & NETIF_F_HW_CSUM) && 10883 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) { 10884 netdev_warn(dev, "mixed HW and IP checksum settings.\n"); 10885 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM); 10886 } 10887 10888 /* TSO requires that SG is present as well. */ 10889 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) { 10890 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n"); 10891 features &= ~NETIF_F_ALL_TSO; 10892 } 10893 10894 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) && 10895 !(features & NETIF_F_IP_CSUM)) { 10896 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n"); 10897 features &= ~NETIF_F_TSO; 10898 features &= ~NETIF_F_TSO_ECN; 10899 } 10900 10901 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) && 10902 !(features & NETIF_F_IPV6_CSUM)) { 10903 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n"); 10904 features &= ~NETIF_F_TSO6; 10905 } 10906 10907 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */ 10908 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO)) 10909 features &= ~NETIF_F_TSO_MANGLEID; 10910 10911 /* TSO ECN requires that TSO is present as well. */ 10912 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN) 10913 features &= ~NETIF_F_TSO_ECN; 10914 10915 /* Software GSO depends on SG. */ 10916 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) { 10917 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n"); 10918 features &= ~NETIF_F_GSO; 10919 } 10920 10921 /* GSO partial features require GSO partial be set */ 10922 if ((features & dev->gso_partial_features) && 10923 !(features & NETIF_F_GSO_PARTIAL)) { 10924 netdev_dbg(dev, 10925 "Dropping partially supported GSO features since no GSO partial.\n"); 10926 features &= ~dev->gso_partial_features; 10927 } 10928 10929 if (!(features & NETIF_F_RXCSUM)) { 10930 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet 10931 * successfully merged by hardware must also have the 10932 * checksum verified by hardware. If the user does not 10933 * want to enable RXCSUM, logically, we should disable GRO_HW. 10934 */ 10935 if (features & NETIF_F_GRO_HW) { 10936 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n"); 10937 features &= ~NETIF_F_GRO_HW; 10938 } 10939 } 10940 10941 /* LRO/HW-GRO features cannot be combined with RX-FCS */ 10942 if (features & NETIF_F_RXFCS) { 10943 if (features & NETIF_F_LRO) { 10944 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n"); 10945 features &= ~NETIF_F_LRO; 10946 } 10947 10948 if (features & NETIF_F_GRO_HW) { 10949 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n"); 10950 features &= ~NETIF_F_GRO_HW; 10951 } 10952 } 10953 10954 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) { 10955 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n"); 10956 features &= ~NETIF_F_LRO; 10957 } 10958 10959 if ((features & NETIF_F_HW_TLS_TX) && !netdev_has_ip_or_hw_csum(features)) { 10960 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n"); 10961 features &= ~NETIF_F_HW_TLS_TX; 10962 } 10963 10964 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) { 10965 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n"); 10966 features &= ~NETIF_F_HW_TLS_RX; 10967 } 10968 10969 if ((features & NETIF_F_GSO_UDP_L4) && !netdev_has_ip_or_hw_csum(features)) { 10970 netdev_dbg(dev, "Dropping USO feature since no CSUM feature.\n"); 10971 features &= ~NETIF_F_GSO_UDP_L4; 10972 } 10973 10974 return features; 10975 } 10976 10977 int __netdev_update_features(struct net_device *dev) 10978 { 10979 struct net_device *upper, *lower; 10980 netdev_features_t features; 10981 struct list_head *iter; 10982 int err = -1; 10983 10984 ASSERT_RTNL(); 10985 netdev_ops_assert_locked(dev); 10986 10987 features = netdev_get_wanted_features(dev); 10988 10989 if (dev->netdev_ops->ndo_fix_features) 10990 features = dev->netdev_ops->ndo_fix_features(dev, features); 10991 10992 /* driver might be less strict about feature dependencies */ 10993 features = netdev_fix_features(dev, features); 10994 10995 /* some features can't be enabled if they're off on an upper device */ 10996 netdev_for_each_upper_dev_rcu(dev, upper, iter) 10997 features = netdev_sync_upper_features(dev, upper, features); 10998 10999 if (dev->features == features) 11000 goto sync_lower; 11001 11002 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n", 11003 &dev->features, &features); 11004 11005 if (dev->netdev_ops->ndo_set_features) 11006 err = dev->netdev_ops->ndo_set_features(dev, features); 11007 else 11008 err = 0; 11009 11010 if (unlikely(err < 0)) { 11011 netdev_err(dev, 11012 "set_features() failed (%d); wanted %pNF, left %pNF\n", 11013 err, &features, &dev->features); 11014 /* return non-0 since some features might have changed and 11015 * it's better to fire a spurious notification than miss it 11016 */ 11017 return -1; 11018 } 11019 11020 sync_lower: 11021 /* some features must be disabled on lower devices when disabled 11022 * on an upper device (think: bonding master or bridge) 11023 */ 11024 netdev_for_each_lower_dev(dev, lower, iter) 11025 netdev_sync_lower_features(dev, lower, features); 11026 11027 if (!err) { 11028 netdev_features_t diff = features ^ dev->features; 11029 11030 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) { 11031 /* udp_tunnel_{get,drop}_rx_info both need 11032 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the 11033 * device, or they won't do anything. 11034 * Thus we need to update dev->features 11035 * *before* calling udp_tunnel_get_rx_info, 11036 * but *after* calling udp_tunnel_drop_rx_info. 11037 */ 11038 udp_tunnel_nic_lock(dev); 11039 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) { 11040 dev->features = features; 11041 udp_tunnel_get_rx_info(dev); 11042 } else { 11043 udp_tunnel_drop_rx_info(dev); 11044 } 11045 udp_tunnel_nic_unlock(dev); 11046 } 11047 11048 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) { 11049 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) { 11050 dev->features = features; 11051 err |= vlan_get_rx_ctag_filter_info(dev); 11052 } else { 11053 vlan_drop_rx_ctag_filter_info(dev); 11054 } 11055 } 11056 11057 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) { 11058 if (features & NETIF_F_HW_VLAN_STAG_FILTER) { 11059 dev->features = features; 11060 err |= vlan_get_rx_stag_filter_info(dev); 11061 } else { 11062 vlan_drop_rx_stag_filter_info(dev); 11063 } 11064 } 11065 11066 dev->features = features; 11067 } 11068 11069 return err < 0 ? 0 : 1; 11070 } 11071 11072 /** 11073 * netdev_update_features - recalculate device features 11074 * @dev: the device to check 11075 * 11076 * Recalculate dev->features set and send notifications if it 11077 * has changed. Should be called after driver or hardware dependent 11078 * conditions might have changed that influence the features. 11079 */ 11080 void netdev_update_features(struct net_device *dev) 11081 { 11082 if (__netdev_update_features(dev)) 11083 netdev_features_change(dev); 11084 } 11085 EXPORT_SYMBOL(netdev_update_features); 11086 11087 /** 11088 * netdev_change_features - recalculate device features 11089 * @dev: the device to check 11090 * 11091 * Recalculate dev->features set and send notifications even 11092 * if they have not changed. Should be called instead of 11093 * netdev_update_features() if also dev->vlan_features might 11094 * have changed to allow the changes to be propagated to stacked 11095 * VLAN devices. 11096 */ 11097 void netdev_change_features(struct net_device *dev) 11098 { 11099 __netdev_update_features(dev); 11100 netdev_features_change(dev); 11101 } 11102 EXPORT_SYMBOL(netdev_change_features); 11103 11104 /** 11105 * netif_stacked_transfer_operstate - transfer operstate 11106 * @rootdev: the root or lower level device to transfer state from 11107 * @dev: the device to transfer operstate to 11108 * 11109 * Transfer operational state from root to device. This is normally 11110 * called when a stacking relationship exists between the root 11111 * device and the device(a leaf device). 11112 */ 11113 void netif_stacked_transfer_operstate(const struct net_device *rootdev, 11114 struct net_device *dev) 11115 { 11116 if (rootdev->operstate == IF_OPER_DORMANT) 11117 netif_dormant_on(dev); 11118 else 11119 netif_dormant_off(dev); 11120 11121 if (rootdev->operstate == IF_OPER_TESTING) 11122 netif_testing_on(dev); 11123 else 11124 netif_testing_off(dev); 11125 11126 if (netif_carrier_ok(rootdev)) 11127 netif_carrier_on(dev); 11128 else 11129 netif_carrier_off(dev); 11130 } 11131 EXPORT_SYMBOL(netif_stacked_transfer_operstate); 11132 11133 static int netif_alloc_rx_queues(struct net_device *dev) 11134 { 11135 unsigned int i, count = dev->num_rx_queues; 11136 struct netdev_rx_queue *rx; 11137 size_t sz = count * sizeof(*rx); 11138 int err = 0; 11139 11140 BUG_ON(count < 1); 11141 11142 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11143 if (!rx) 11144 return -ENOMEM; 11145 11146 dev->_rx = rx; 11147 11148 for (i = 0; i < count; i++) { 11149 rx[i].dev = dev; 11150 11151 /* XDP RX-queue setup */ 11152 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0); 11153 if (err < 0) 11154 goto err_rxq_info; 11155 } 11156 return 0; 11157 11158 err_rxq_info: 11159 /* Rollback successful reg's and free other resources */ 11160 while (i--) 11161 xdp_rxq_info_unreg(&rx[i].xdp_rxq); 11162 kvfree(dev->_rx); 11163 dev->_rx = NULL; 11164 return err; 11165 } 11166 11167 static void netif_free_rx_queues(struct net_device *dev) 11168 { 11169 unsigned int i, count = dev->num_rx_queues; 11170 11171 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */ 11172 if (!dev->_rx) 11173 return; 11174 11175 for (i = 0; i < count; i++) 11176 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq); 11177 11178 kvfree(dev->_rx); 11179 } 11180 11181 static void netdev_init_one_queue(struct net_device *dev, 11182 struct netdev_queue *queue, void *_unused) 11183 { 11184 /* Initialize queue lock */ 11185 spin_lock_init(&queue->_xmit_lock); 11186 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type); 11187 queue->xmit_lock_owner = -1; 11188 netdev_queue_numa_node_write(queue, NUMA_NO_NODE); 11189 queue->dev = dev; 11190 #ifdef CONFIG_BQL 11191 dql_init(&queue->dql, HZ); 11192 #endif 11193 } 11194 11195 static void netif_free_tx_queues(struct net_device *dev) 11196 { 11197 kvfree(dev->_tx); 11198 } 11199 11200 static int netif_alloc_netdev_queues(struct net_device *dev) 11201 { 11202 unsigned int count = dev->num_tx_queues; 11203 struct netdev_queue *tx; 11204 size_t sz = count * sizeof(*tx); 11205 11206 if (count < 1 || count > 0xffff) 11207 return -EINVAL; 11208 11209 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11210 if (!tx) 11211 return -ENOMEM; 11212 11213 dev->_tx = tx; 11214 11215 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL); 11216 spin_lock_init(&dev->tx_global_lock); 11217 11218 return 0; 11219 } 11220 11221 void netif_tx_stop_all_queues(struct net_device *dev) 11222 { 11223 unsigned int i; 11224 11225 for (i = 0; i < dev->num_tx_queues; i++) { 11226 struct netdev_queue *txq = netdev_get_tx_queue(dev, i); 11227 11228 netif_tx_stop_queue(txq); 11229 } 11230 } 11231 EXPORT_SYMBOL(netif_tx_stop_all_queues); 11232 11233 static int netdev_do_alloc_pcpu_stats(struct net_device *dev) 11234 { 11235 void __percpu *v; 11236 11237 /* Drivers implementing ndo_get_peer_dev must support tstat 11238 * accounting, so that skb_do_redirect() can bump the dev's 11239 * RX stats upon network namespace switch. 11240 */ 11241 if (dev->netdev_ops->ndo_get_peer_dev && 11242 dev->pcpu_stat_type != NETDEV_PCPU_STAT_TSTATS) 11243 return -EOPNOTSUPP; 11244 11245 switch (dev->pcpu_stat_type) { 11246 case NETDEV_PCPU_STAT_NONE: 11247 return 0; 11248 case NETDEV_PCPU_STAT_LSTATS: 11249 v = dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats); 11250 break; 11251 case NETDEV_PCPU_STAT_TSTATS: 11252 v = dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats); 11253 break; 11254 case NETDEV_PCPU_STAT_DSTATS: 11255 v = dev->dstats = netdev_alloc_pcpu_stats(struct pcpu_dstats); 11256 break; 11257 default: 11258 return -EINVAL; 11259 } 11260 11261 return v ? 0 : -ENOMEM; 11262 } 11263 11264 static void netdev_do_free_pcpu_stats(struct net_device *dev) 11265 { 11266 switch (dev->pcpu_stat_type) { 11267 case NETDEV_PCPU_STAT_NONE: 11268 return; 11269 case NETDEV_PCPU_STAT_LSTATS: 11270 free_percpu(dev->lstats); 11271 break; 11272 case NETDEV_PCPU_STAT_TSTATS: 11273 free_percpu(dev->tstats); 11274 break; 11275 case NETDEV_PCPU_STAT_DSTATS: 11276 free_percpu(dev->dstats); 11277 break; 11278 } 11279 } 11280 11281 static void netdev_free_phy_link_topology(struct net_device *dev) 11282 { 11283 struct phy_link_topology *topo = dev->link_topo; 11284 11285 if (IS_ENABLED(CONFIG_PHYLIB) && topo) { 11286 xa_destroy(&topo->phys); 11287 kfree(topo); 11288 dev->link_topo = NULL; 11289 } 11290 } 11291 11292 /** 11293 * register_netdevice() - register a network device 11294 * @dev: device to register 11295 * 11296 * Take a prepared network device structure and make it externally accessible. 11297 * A %NETDEV_REGISTER message is sent to the netdev notifier chain. 11298 * Callers must hold the rtnl lock - you may want register_netdev() 11299 * instead of this. 11300 */ 11301 int register_netdevice(struct net_device *dev) 11302 { 11303 int ret; 11304 struct net *net = dev_net(dev); 11305 11306 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE < 11307 NETDEV_FEATURE_COUNT); 11308 BUG_ON(dev_boot_phase); 11309 ASSERT_RTNL(); 11310 11311 might_sleep(); 11312 11313 /* When net_device's are persistent, this will be fatal. */ 11314 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED); 11315 BUG_ON(!net); 11316 11317 ret = ethtool_check_ops(dev->ethtool_ops); 11318 if (ret) 11319 return ret; 11320 11321 /* rss ctx ID 0 is reserved for the default context, start from 1 */ 11322 xa_init_flags(&dev->ethtool->rss_ctx, XA_FLAGS_ALLOC1); 11323 mutex_init(&dev->ethtool->rss_lock); 11324 11325 spin_lock_init(&dev->addr_list_lock); 11326 netdev_set_addr_lockdep_class(dev); 11327 11328 ret = dev_get_valid_name(net, dev, dev->name); 11329 if (ret < 0) 11330 goto out; 11331 11332 ret = -ENOMEM; 11333 dev->name_node = netdev_name_node_head_alloc(dev); 11334 if (!dev->name_node) 11335 goto out; 11336 11337 /* Init, if this function is available */ 11338 if (dev->netdev_ops->ndo_init) { 11339 ret = dev->netdev_ops->ndo_init(dev); 11340 if (ret) { 11341 if (ret > 0) 11342 ret = -EIO; 11343 goto err_free_name; 11344 } 11345 } 11346 11347 if (((dev->hw_features | dev->features) & 11348 NETIF_F_HW_VLAN_CTAG_FILTER) && 11349 (!dev->netdev_ops->ndo_vlan_rx_add_vid || 11350 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) { 11351 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n"); 11352 ret = -EINVAL; 11353 goto err_uninit; 11354 } 11355 11356 if (netdev_need_ops_lock(dev) && 11357 dev->netdev_ops->ndo_set_rx_mode && 11358 !dev->netdev_ops->ndo_set_rx_mode_async) 11359 netdev_WARN(dev, "ops-locked drivers should use ndo_set_rx_mode_async\n"); 11360 11361 ret = netdev_do_alloc_pcpu_stats(dev); 11362 if (ret) 11363 goto err_uninit; 11364 11365 ret = dev_index_reserve(net, dev->ifindex); 11366 if (ret < 0) 11367 goto err_free_pcpu; 11368 dev->ifindex = ret; 11369 11370 /* Transfer changeable features to wanted_features and enable 11371 * software offloads (GSO and GRO). 11372 */ 11373 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF); 11374 dev->features |= NETIF_F_SOFT_FEATURES; 11375 11376 if (dev->udp_tunnel_nic_info) { 11377 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT; 11378 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT; 11379 } 11380 11381 dev->wanted_features = dev->features & dev->hw_features; 11382 11383 if (!(dev->flags & IFF_LOOPBACK)) 11384 dev->hw_features |= NETIF_F_NOCACHE_COPY; 11385 11386 /* If IPv4 TCP segmentation offload is supported we should also 11387 * allow the device to enable segmenting the frame with the option 11388 * of ignoring a static IP ID value. This doesn't enable the 11389 * feature itself but allows the user to enable it later. 11390 */ 11391 if (dev->hw_features & NETIF_F_TSO) 11392 dev->hw_features |= NETIF_F_TSO_MANGLEID; 11393 if (dev->vlan_features & NETIF_F_TSO) 11394 dev->vlan_features |= NETIF_F_TSO_MANGLEID; 11395 if (dev->mpls_features & NETIF_F_TSO) 11396 dev->mpls_features |= NETIF_F_TSO_MANGLEID; 11397 if (dev->hw_enc_features & NETIF_F_TSO) 11398 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID; 11399 11400 /* TSO_MANGLEID belongs in mangleid_features by definition */ 11401 dev->mangleid_features |= NETIF_F_TSO_MANGLEID; 11402 11403 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices. 11404 */ 11405 dev->vlan_features |= NETIF_F_HIGHDMA; 11406 11407 /* Make NETIF_F_SG inheritable to tunnel devices. 11408 */ 11409 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL; 11410 11411 /* Make NETIF_F_SG inheritable to MPLS. 11412 */ 11413 dev->mpls_features |= NETIF_F_SG; 11414 11415 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev); 11416 ret = notifier_to_errno(ret); 11417 if (ret) 11418 goto err_ifindex_release; 11419 11420 ret = netdev_register_kobject(dev); 11421 11422 netdev_lock(dev); 11423 WRITE_ONCE(dev->reg_state, ret ? NETREG_UNREGISTERED : NETREG_REGISTERED); 11424 netdev_unlock(dev); 11425 11426 if (ret) 11427 goto err_uninit_notify; 11428 11429 netdev_lock_ops(dev); 11430 __netdev_update_features(dev); 11431 netdev_unlock_ops(dev); 11432 11433 /* 11434 * Default initial state at registry is that the 11435 * device is present. 11436 */ 11437 11438 set_bit(__LINK_STATE_PRESENT, &dev->state); 11439 11440 linkwatch_init_dev(dev); 11441 11442 dev_init_scheduler(dev); 11443 11444 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL); 11445 list_netdevice(dev); 11446 11447 add_device_randomness(dev->dev_addr, dev->addr_len); 11448 11449 /* If the device has permanent device address, driver should 11450 * set dev_addr and also addr_assign_type should be set to 11451 * NET_ADDR_PERM (default value). 11452 */ 11453 if (dev->addr_assign_type == NET_ADDR_PERM) 11454 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len); 11455 11456 /* Notify protocols, that a new device appeared. */ 11457 netdev_lock_ops(dev); 11458 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev); 11459 netdev_unlock_ops(dev); 11460 ret = notifier_to_errno(ret); 11461 if (ret) { 11462 /* Expect explicit free_netdev() on failure */ 11463 dev->needs_free_netdev = false; 11464 unregister_netdevice_queue(dev, NULL); 11465 goto out; 11466 } 11467 /* 11468 * Prevent userspace races by waiting until the network 11469 * device is fully setup before sending notifications. 11470 */ 11471 if (!(dev->rtnl_link_ops && dev->rtnl_link_initializing)) 11472 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 11473 11474 out: 11475 return ret; 11476 11477 err_uninit_notify: 11478 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11479 err_ifindex_release: 11480 dev_index_release(net, dev->ifindex); 11481 err_free_pcpu: 11482 netdev_do_free_pcpu_stats(dev); 11483 err_uninit: 11484 if (dev->netdev_ops->ndo_uninit) 11485 dev->netdev_ops->ndo_uninit(dev); 11486 if (dev->priv_destructor) 11487 dev->priv_destructor(dev); 11488 err_free_name: 11489 netdev_name_node_free(dev->name_node); 11490 goto out; 11491 } 11492 EXPORT_SYMBOL(register_netdevice); 11493 11494 /* Initialize the core of a dummy net device. 11495 * The setup steps dummy netdevs need which normal netdevs get by going 11496 * through register_netdevice(). 11497 */ 11498 static void init_dummy_netdev(struct net_device *dev) 11499 { 11500 /* make sure we BUG if trying to hit standard 11501 * register/unregister code path 11502 */ 11503 dev->reg_state = NETREG_DUMMY; 11504 11505 /* a dummy interface is started by default */ 11506 set_bit(__LINK_STATE_PRESENT, &dev->state); 11507 set_bit(__LINK_STATE_START, &dev->state); 11508 11509 /* Note : We dont allocate pcpu_refcnt for dummy devices, 11510 * because users of this 'device' dont need to change 11511 * its refcount. 11512 */ 11513 } 11514 11515 /** 11516 * register_netdev - register a network device 11517 * @dev: device to register 11518 * 11519 * Take a completed network device structure and add it to the kernel 11520 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier 11521 * chain. 0 is returned on success. A negative errno code is returned 11522 * on a failure to set up the device, or if the name is a duplicate. 11523 * 11524 * This is a wrapper around register_netdevice that takes the rtnl semaphore 11525 * and expands the device name if you passed a format string to 11526 * alloc_netdev. 11527 */ 11528 int register_netdev(struct net_device *dev) 11529 { 11530 struct net *net = dev_net(dev); 11531 int err; 11532 11533 if (rtnl_net_lock_killable(net)) 11534 return -EINTR; 11535 11536 err = register_netdevice(dev); 11537 11538 rtnl_net_unlock(net); 11539 11540 return err; 11541 } 11542 EXPORT_SYMBOL(register_netdev); 11543 11544 int netdev_refcnt_read(const struct net_device *dev) 11545 { 11546 #ifdef CONFIG_PCPU_DEV_REFCNT 11547 int i, refcnt = 0; 11548 11549 for_each_possible_cpu(i) 11550 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i); 11551 return refcnt; 11552 #else 11553 return refcount_read(&dev->dev_refcnt); 11554 #endif 11555 } 11556 EXPORT_SYMBOL(netdev_refcnt_read); 11557 11558 int netdev_unregister_timeout_secs __read_mostly = 10; 11559 11560 #define WAIT_REFS_MIN_MSECS 1 11561 #define WAIT_REFS_MAX_MSECS 250 11562 /** 11563 * netdev_wait_allrefs_any - wait until all references are gone. 11564 * @list: list of net_devices to wait on 11565 * 11566 * This is called when unregistering network devices. 11567 * 11568 * Any protocol or device that holds a reference should register 11569 * for netdevice notification, and cleanup and put back the 11570 * reference if they receive an UNREGISTER event. 11571 * We can get stuck here if buggy protocols don't correctly 11572 * call dev_put. 11573 */ 11574 static struct net_device *netdev_wait_allrefs_any(struct list_head *list) 11575 { 11576 unsigned long rebroadcast_time, warning_time; 11577 struct net_device *dev; 11578 int wait = 0; 11579 11580 rebroadcast_time = warning_time = jiffies; 11581 11582 list_for_each_entry(dev, list, todo_list) 11583 if (netdev_refcnt_read(dev) == 1) 11584 return dev; 11585 11586 while (true) { 11587 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) { 11588 rtnl_lock(); 11589 11590 /* Rebroadcast unregister notification */ 11591 list_for_each_entry(dev, list, todo_list) 11592 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11593 11594 __rtnl_unlock(); 11595 rcu_barrier(); 11596 rtnl_lock(); 11597 11598 list_for_each_entry(dev, list, todo_list) 11599 if (test_bit(__LINK_STATE_LINKWATCH_PENDING, 11600 &dev->state)) { 11601 /* We must not have linkwatch events 11602 * pending on unregister. If this 11603 * happens, we simply run the queue 11604 * unscheduled, resulting in a noop 11605 * for this device. 11606 */ 11607 linkwatch_run_queue(); 11608 break; 11609 } 11610 11611 __rtnl_unlock(); 11612 11613 rebroadcast_time = jiffies; 11614 } 11615 11616 rcu_barrier(); 11617 11618 if (!wait) { 11619 wait = WAIT_REFS_MIN_MSECS; 11620 } else { 11621 msleep(wait); 11622 wait = min(wait << 1, WAIT_REFS_MAX_MSECS); 11623 } 11624 11625 list_for_each_entry(dev, list, todo_list) 11626 if (netdev_refcnt_read(dev) == 1) 11627 return dev; 11628 11629 if (time_after(jiffies, warning_time + 11630 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) { 11631 list_for_each_entry(dev, list, todo_list) { 11632 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n", 11633 dev->name, netdev_refcnt_read(dev)); 11634 ref_tracker_dir_print(&dev->refcnt_tracker, 10); 11635 } 11636 11637 warning_time = jiffies; 11638 } 11639 } 11640 } 11641 11642 /* The sequence is: 11643 * 11644 * rtnl_lock(); 11645 * ... 11646 * register_netdevice(x1); 11647 * register_netdevice(x2); 11648 * ... 11649 * unregister_netdevice(y1); 11650 * unregister_netdevice(y2); 11651 * ... 11652 * rtnl_unlock(); 11653 * free_netdev(y1); 11654 * free_netdev(y2); 11655 * 11656 * We are invoked by rtnl_unlock(). 11657 * This allows us to deal with problems: 11658 * 1) We can delete sysfs objects which invoke hotplug 11659 * without deadlocking with linkwatch via keventd. 11660 * 2) Since we run with the RTNL semaphore not held, we can sleep 11661 * safely in order to wait for the netdev refcnt to drop to zero. 11662 * 11663 * We must not return until all unregister events added during 11664 * the interval the lock was held have been completed. 11665 */ 11666 void netdev_run_todo(void) 11667 { 11668 struct net_device *dev, *tmp; 11669 struct list_head list; 11670 int cnt; 11671 #ifdef CONFIG_LOCKDEP 11672 struct list_head unlink_list; 11673 11674 list_replace_init(&net_unlink_list, &unlink_list); 11675 11676 while (!list_empty(&unlink_list)) { 11677 dev = list_first_entry(&unlink_list, struct net_device, 11678 unlink_list); 11679 list_del_init(&dev->unlink_list); 11680 dev->nested_level = dev->lower_level - 1; 11681 } 11682 #endif 11683 11684 /* Snapshot list, allow later requests */ 11685 list_replace_init(&net_todo_list, &list); 11686 11687 __rtnl_unlock(); 11688 11689 /* Wait for rcu callbacks to finish before next phase */ 11690 if (!list_empty(&list)) 11691 rcu_barrier(); 11692 11693 list_for_each_entry_safe(dev, tmp, &list, todo_list) { 11694 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) { 11695 netdev_WARN(dev, "run_todo but not unregistering\n"); 11696 list_del(&dev->todo_list); 11697 continue; 11698 } 11699 11700 netdev_lock(dev); 11701 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERED); 11702 netdev_unlock(dev); 11703 linkwatch_sync_dev(dev); 11704 } 11705 11706 cnt = 0; 11707 while (!list_empty(&list)) { 11708 dev = netdev_wait_allrefs_any(&list); 11709 list_del(&dev->todo_list); 11710 11711 /* paranoia */ 11712 BUG_ON(netdev_refcnt_read(dev) != 1); 11713 BUG_ON(!list_empty(&dev->ptype_all)); 11714 BUG_ON(!list_empty(&dev->ptype_specific)); 11715 WARN_ON(rcu_access_pointer(dev->ip_ptr)); 11716 WARN_ON(rcu_access_pointer(dev->ip6_ptr)); 11717 11718 netdev_do_free_pcpu_stats(dev); 11719 if (dev->priv_destructor) 11720 dev->priv_destructor(dev); 11721 if (dev->needs_free_netdev) 11722 free_netdev(dev); 11723 11724 cnt++; 11725 11726 /* Free network device */ 11727 kobject_put(&dev->dev.kobj); 11728 } 11729 if (cnt && atomic_sub_and_test(cnt, &dev_unreg_count)) 11730 wake_up(&netdev_unregistering_wq); 11731 } 11732 11733 /* Collate per-cpu network dstats statistics 11734 * 11735 * Read per-cpu network statistics from dev->dstats and populate the related 11736 * fields in @s. 11737 */ 11738 static void dev_fetch_dstats(struct rtnl_link_stats64 *s, 11739 const struct pcpu_dstats __percpu *dstats) 11740 { 11741 int cpu; 11742 11743 for_each_possible_cpu(cpu) { 11744 u64 rx_packets, rx_bytes, rx_drops; 11745 u64 tx_packets, tx_bytes, tx_drops; 11746 const struct pcpu_dstats *stats; 11747 unsigned int start; 11748 11749 stats = per_cpu_ptr(dstats, cpu); 11750 do { 11751 start = u64_stats_fetch_begin(&stats->syncp); 11752 rx_packets = u64_stats_read(&stats->rx_packets); 11753 rx_bytes = u64_stats_read(&stats->rx_bytes); 11754 rx_drops = u64_stats_read(&stats->rx_drops); 11755 tx_packets = u64_stats_read(&stats->tx_packets); 11756 tx_bytes = u64_stats_read(&stats->tx_bytes); 11757 tx_drops = u64_stats_read(&stats->tx_drops); 11758 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11759 11760 s->rx_packets += rx_packets; 11761 s->rx_bytes += rx_bytes; 11762 s->rx_dropped += rx_drops; 11763 s->tx_packets += tx_packets; 11764 s->tx_bytes += tx_bytes; 11765 s->tx_dropped += tx_drops; 11766 } 11767 } 11768 11769 /* ndo_get_stats64 implementation for dtstats-based accounting. 11770 * 11771 * Populate @s from dev->stats and dev->dstats. This is used internally by the 11772 * core for NETDEV_PCPU_STAT_DSTAT-type stats collection. 11773 */ 11774 static void dev_get_dstats64(const struct net_device *dev, 11775 struct rtnl_link_stats64 *s) 11776 { 11777 netdev_stats_to_stats64(s, &dev->stats); 11778 dev_fetch_dstats(s, dev->dstats); 11779 } 11780 11781 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has 11782 * all the same fields in the same order as net_device_stats, with only 11783 * the type differing, but rtnl_link_stats64 may have additional fields 11784 * at the end for newer counters. 11785 */ 11786 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64, 11787 const struct net_device_stats *netdev_stats) 11788 { 11789 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t); 11790 const atomic_long_t *src = (atomic_long_t *)netdev_stats; 11791 u64 *dst = (u64 *)stats64; 11792 11793 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64)); 11794 for (i = 0; i < n; i++) 11795 dst[i] = (unsigned long)atomic_long_read(&src[i]); 11796 /* zero out counters that only exist in rtnl_link_stats64 */ 11797 memset((char *)stats64 + n * sizeof(u64), 0, 11798 sizeof(*stats64) - n * sizeof(u64)); 11799 } 11800 EXPORT_SYMBOL(netdev_stats_to_stats64); 11801 11802 static __cold struct net_device_core_stats __percpu *netdev_core_stats_alloc( 11803 struct net_device *dev) 11804 { 11805 struct net_device_core_stats __percpu *p; 11806 11807 p = alloc_percpu_gfp(struct net_device_core_stats, 11808 GFP_ATOMIC | __GFP_NOWARN); 11809 11810 if (p && cmpxchg(&dev->core_stats, NULL, p)) 11811 free_percpu(p); 11812 11813 /* This READ_ONCE() pairs with the cmpxchg() above */ 11814 return READ_ONCE(dev->core_stats); 11815 } 11816 11817 noinline void netdev_core_stats_inc(struct net_device *dev, u32 offset) 11818 { 11819 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11820 struct net_device_core_stats __percpu *p = READ_ONCE(dev->core_stats); 11821 unsigned long __percpu *field; 11822 11823 if (unlikely(!p)) { 11824 p = netdev_core_stats_alloc(dev); 11825 if (!p) 11826 return; 11827 } 11828 11829 field = (unsigned long __percpu *)((void __percpu *)p + offset); 11830 this_cpu_inc(*field); 11831 } 11832 EXPORT_SYMBOL_GPL(netdev_core_stats_inc); 11833 11834 /** 11835 * dev_get_stats - get network device statistics 11836 * @dev: device to get statistics from 11837 * @storage: place to store stats 11838 * 11839 * Get network statistics from device. Return @storage. 11840 * The device driver may provide its own method by setting 11841 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats; 11842 * otherwise the internal statistics structure is used. 11843 */ 11844 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev, 11845 struct rtnl_link_stats64 *storage) 11846 { 11847 const struct net_device_ops *ops = dev->netdev_ops; 11848 const struct net_device_core_stats __percpu *p; 11849 11850 /* 11851 * IPv{4,6} and udp tunnels share common stat helpers and use 11852 * different stat type (NETDEV_PCPU_STAT_TSTATS vs 11853 * NETDEV_PCPU_STAT_DSTATS). Ensure the accounting is consistent. 11854 */ 11855 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_bytes) != 11856 offsetof(struct pcpu_dstats, rx_bytes)); 11857 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_packets) != 11858 offsetof(struct pcpu_dstats, rx_packets)); 11859 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_bytes) != 11860 offsetof(struct pcpu_dstats, tx_bytes)); 11861 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_packets) != 11862 offsetof(struct pcpu_dstats, tx_packets)); 11863 11864 if (ops->ndo_get_stats64) { 11865 memset(storage, 0, sizeof(*storage)); 11866 ops->ndo_get_stats64(dev, storage); 11867 } else if (ops->ndo_get_stats) { 11868 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev)); 11869 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_TSTATS) { 11870 dev_get_tstats64(dev, storage); 11871 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_DSTATS) { 11872 dev_get_dstats64(dev, storage); 11873 } else { 11874 netdev_stats_to_stats64(storage, &dev->stats); 11875 } 11876 11877 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11878 p = READ_ONCE(dev->core_stats); 11879 if (p) { 11880 const struct net_device_core_stats *core_stats; 11881 int i; 11882 11883 for_each_possible_cpu(i) { 11884 core_stats = per_cpu_ptr(p, i); 11885 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped); 11886 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped); 11887 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler); 11888 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped); 11889 } 11890 } 11891 return storage; 11892 } 11893 EXPORT_SYMBOL(dev_get_stats); 11894 11895 /** 11896 * dev_fetch_sw_netstats - get per-cpu network device statistics 11897 * @s: place to store stats 11898 * @netstats: per-cpu network stats to read from 11899 * 11900 * Read per-cpu network statistics and populate the related fields in @s. 11901 */ 11902 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s, 11903 const struct pcpu_sw_netstats __percpu *netstats) 11904 { 11905 int cpu; 11906 11907 for_each_possible_cpu(cpu) { 11908 u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 11909 const struct pcpu_sw_netstats *stats; 11910 unsigned int start; 11911 11912 stats = per_cpu_ptr(netstats, cpu); 11913 do { 11914 start = u64_stats_fetch_begin(&stats->syncp); 11915 rx_packets = u64_stats_read(&stats->rx_packets); 11916 rx_bytes = u64_stats_read(&stats->rx_bytes); 11917 tx_packets = u64_stats_read(&stats->tx_packets); 11918 tx_bytes = u64_stats_read(&stats->tx_bytes); 11919 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11920 11921 s->rx_packets += rx_packets; 11922 s->rx_bytes += rx_bytes; 11923 s->tx_packets += tx_packets; 11924 s->tx_bytes += tx_bytes; 11925 } 11926 } 11927 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats); 11928 11929 /** 11930 * dev_get_tstats64 - ndo_get_stats64 implementation 11931 * @dev: device to get statistics from 11932 * @s: place to store stats 11933 * 11934 * Populate @s from dev->stats and dev->tstats. Can be used as 11935 * ndo_get_stats64() callback. 11936 */ 11937 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s) 11938 { 11939 netdev_stats_to_stats64(s, &dev->stats); 11940 dev_fetch_sw_netstats(s, dev->tstats); 11941 } 11942 EXPORT_SYMBOL_GPL(dev_get_tstats64); 11943 11944 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev) 11945 { 11946 struct netdev_queue *queue = dev_ingress_queue(dev); 11947 11948 #ifdef CONFIG_NET_CLS_ACT 11949 if (queue) 11950 return queue; 11951 queue = kzalloc_obj(*queue); 11952 if (!queue) 11953 return NULL; 11954 netdev_init_one_queue(dev, queue, NULL); 11955 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc); 11956 RCU_INIT_POINTER(queue->qdisc_sleeping, &noop_qdisc); 11957 rcu_assign_pointer(dev->ingress_queue, queue); 11958 #endif 11959 return queue; 11960 } 11961 11962 static const struct ethtool_ops default_ethtool_ops; 11963 11964 void netdev_set_default_ethtool_ops(struct net_device *dev, 11965 const struct ethtool_ops *ops) 11966 { 11967 if (dev->ethtool_ops == &default_ethtool_ops) 11968 dev->ethtool_ops = ops; 11969 } 11970 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops); 11971 11972 /** 11973 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default 11974 * @dev: netdev to enable the IRQ coalescing on 11975 * 11976 * Sets a conservative default for SW IRQ coalescing. Users can use 11977 * sysfs attributes to override the default values. 11978 */ 11979 void netdev_sw_irq_coalesce_default_on(struct net_device *dev) 11980 { 11981 WARN_ON(dev->reg_state == NETREG_REGISTERED); 11982 11983 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) { 11984 netdev_set_gro_flush_timeout(dev, 20000); 11985 netdev_set_defer_hard_irqs(dev, 1); 11986 } 11987 } 11988 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on); 11989 11990 /** 11991 * alloc_netdev_mqs - allocate network device 11992 * @sizeof_priv: size of private data to allocate space for 11993 * @name: device name format string 11994 * @name_assign_type: origin of device name 11995 * @setup: callback to initialize device 11996 * @txqs: the number of TX subqueues to allocate 11997 * @rxqs: the number of RX subqueues to allocate 11998 * 11999 * Allocates a struct net_device with private data area for driver use 12000 * and performs basic initialization. Also allocates subqueue structs 12001 * for each queue on the device. 12002 */ 12003 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name, 12004 unsigned char name_assign_type, 12005 void (*setup)(struct net_device *), 12006 unsigned int txqs, unsigned int rxqs) 12007 { 12008 struct net_device *dev; 12009 size_t napi_config_sz; 12010 unsigned int maxqs; 12011 12012 BUG_ON(strlen(name) >= sizeof(dev->name)); 12013 12014 if (txqs < 1) { 12015 pr_err("alloc_netdev: Unable to allocate device with zero queues\n"); 12016 return NULL; 12017 } 12018 12019 if (rxqs < 1) { 12020 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n"); 12021 return NULL; 12022 } 12023 12024 maxqs = max(txqs, rxqs); 12025 12026 dev = kvzalloc_flex(*dev, priv, sizeof_priv, 12027 GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 12028 if (!dev) 12029 return NULL; 12030 12031 dev->priv_len = sizeof_priv; 12032 12033 ref_tracker_dir_init(&dev->refcnt_tracker, 128, "netdev"); 12034 #ifdef CONFIG_PCPU_DEV_REFCNT 12035 dev->pcpu_refcnt = alloc_percpu(int); 12036 if (!dev->pcpu_refcnt) 12037 goto free_dev; 12038 __dev_hold(dev); 12039 #else 12040 refcount_set(&dev->dev_refcnt, 1); 12041 #endif 12042 12043 if (dev_addr_init(dev)) 12044 goto free_pcpu; 12045 12046 dev_mc_init(dev); 12047 dev_uc_init(dev); 12048 12049 dev_net_set(dev, &init_net); 12050 12051 dev->gso_max_size = GSO_LEGACY_MAX_SIZE; 12052 dev->xdp_zc_max_segs = 1; 12053 dev->gso_max_segs = GSO_MAX_SEGS; 12054 dev->gro_max_size = GRO_LEGACY_MAX_SIZE; 12055 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE; 12056 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE; 12057 dev->tso_max_size = TSO_LEGACY_MAX_SIZE; 12058 dev->tso_max_segs = TSO_MAX_SEGS; 12059 dev->upper_level = 1; 12060 dev->lower_level = 1; 12061 #ifdef CONFIG_LOCKDEP 12062 dev->nested_level = 0; 12063 INIT_LIST_HEAD(&dev->unlink_list); 12064 #endif 12065 12066 INIT_LIST_HEAD(&dev->napi_list); 12067 INIT_LIST_HEAD(&dev->unreg_list); 12068 INIT_LIST_HEAD(&dev->close_list); 12069 INIT_LIST_HEAD(&dev->link_watch_list); 12070 INIT_LIST_HEAD(&dev->adj_list.upper); 12071 INIT_LIST_HEAD(&dev->adj_list.lower); 12072 INIT_LIST_HEAD(&dev->ptype_all); 12073 INIT_LIST_HEAD(&dev->ptype_specific); 12074 INIT_LIST_HEAD(&dev->net_notifier_list); 12075 #ifdef CONFIG_NET_SCHED 12076 hash_init(dev->qdisc_hash); 12077 #endif 12078 12079 mutex_init(&dev->lock); 12080 INIT_LIST_HEAD(&dev->rx_mode_node); 12081 __hw_addr_init(&dev->rx_mode_addr_cache); 12082 12083 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 12084 setup(dev); 12085 12086 if (!dev->tx_queue_len) { 12087 dev->priv_flags |= IFF_NO_QUEUE; 12088 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN; 12089 } 12090 12091 dev->num_tx_queues = txqs; 12092 dev->real_num_tx_queues = txqs; 12093 if (netif_alloc_netdev_queues(dev)) 12094 goto free_all; 12095 12096 dev->num_rx_queues = rxqs; 12097 dev->real_num_rx_queues = rxqs; 12098 if (netif_alloc_rx_queues(dev)) 12099 goto free_all; 12100 dev->ethtool = kzalloc_obj(*dev->ethtool, GFP_KERNEL_ACCOUNT); 12101 if (!dev->ethtool) 12102 goto free_all; 12103 12104 dev->cfg = kzalloc_obj(*dev->cfg, GFP_KERNEL_ACCOUNT); 12105 if (!dev->cfg) 12106 goto free_all; 12107 dev->cfg_pending = dev->cfg; 12108 12109 dev->num_napi_configs = maxqs; 12110 napi_config_sz = array_size(maxqs, sizeof(*dev->napi_config)); 12111 dev->napi_config = kvzalloc(napi_config_sz, GFP_KERNEL_ACCOUNT); 12112 if (!dev->napi_config) 12113 goto free_all; 12114 12115 strscpy(dev->name, name); 12116 dev->name_assign_type = name_assign_type; 12117 dev->group = INIT_NETDEV_GROUP; 12118 if (!dev->ethtool_ops) 12119 dev->ethtool_ops = &default_ethtool_ops; 12120 12121 nf_hook_netdev_init(dev); 12122 12123 return dev; 12124 12125 free_all: 12126 free_netdev(dev); 12127 return NULL; 12128 12129 free_pcpu: 12130 #ifdef CONFIG_PCPU_DEV_REFCNT 12131 free_percpu(dev->pcpu_refcnt); 12132 free_dev: 12133 #endif 12134 kvfree(dev); 12135 return NULL; 12136 } 12137 EXPORT_SYMBOL(alloc_netdev_mqs); 12138 12139 static void netdev_napi_exit(struct net_device *dev) 12140 { 12141 if (!list_empty(&dev->napi_list)) { 12142 struct napi_struct *p, *n; 12143 12144 netdev_lock(dev); 12145 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list) 12146 __netif_napi_del_locked(p); 12147 netdev_unlock(dev); 12148 12149 synchronize_net(); 12150 } 12151 12152 kvfree(dev->napi_config); 12153 } 12154 12155 /** 12156 * free_netdev - free network device 12157 * @dev: device 12158 * 12159 * This function does the last stage of destroying an allocated device 12160 * interface. The reference to the device object is released. If this 12161 * is the last reference then it will be freed.Must be called in process 12162 * context. 12163 */ 12164 void free_netdev(struct net_device *dev) 12165 { 12166 might_sleep(); 12167 12168 /* When called immediately after register_netdevice() failed the unwind 12169 * handling may still be dismantling the device. Handle that case by 12170 * deferring the free. 12171 */ 12172 if (dev->reg_state == NETREG_UNREGISTERING) { 12173 ASSERT_RTNL(); 12174 dev->needs_free_netdev = true; 12175 return; 12176 } 12177 12178 WARN_ON(dev->cfg != dev->cfg_pending); 12179 kfree(dev->cfg); 12180 kfree(dev->ethtool); 12181 netif_free_tx_queues(dev); 12182 netif_free_rx_queues(dev); 12183 12184 kfree(rcu_dereference_protected(dev->ingress_queue, 1)); 12185 12186 __hw_addr_flush(&dev->rx_mode_addr_cache); 12187 12188 /* Flush device addresses */ 12189 dev_addr_flush(dev); 12190 12191 netdev_napi_exit(dev); 12192 12193 netif_del_cpu_rmap(dev); 12194 12195 ref_tracker_dir_exit(&dev->refcnt_tracker); 12196 #ifdef CONFIG_PCPU_DEV_REFCNT 12197 free_percpu(dev->pcpu_refcnt); 12198 dev->pcpu_refcnt = NULL; 12199 #endif 12200 free_percpu(dev->core_stats); 12201 dev->core_stats = NULL; 12202 free_percpu(dev->xdp_bulkq); 12203 dev->xdp_bulkq = NULL; 12204 12205 netdev_free_phy_link_topology(dev); 12206 12207 mutex_destroy(&dev->lock); 12208 12209 /* Compatibility with error handling in drivers */ 12210 if (dev->reg_state == NETREG_UNINITIALIZED || 12211 dev->reg_state == NETREG_DUMMY) { 12212 kvfree(dev); 12213 return; 12214 } 12215 12216 BUG_ON(dev->reg_state != NETREG_UNREGISTERED); 12217 WRITE_ONCE(dev->reg_state, NETREG_RELEASED); 12218 12219 /* will free via device release */ 12220 put_device(&dev->dev); 12221 } 12222 EXPORT_SYMBOL(free_netdev); 12223 12224 /** 12225 * alloc_netdev_dummy - Allocate and initialize a dummy net device. 12226 * @sizeof_priv: size of private data to allocate space for 12227 * 12228 * Return: the allocated net_device on success, NULL otherwise 12229 */ 12230 struct net_device *alloc_netdev_dummy(int sizeof_priv) 12231 { 12232 return alloc_netdev(sizeof_priv, "dummy#", NET_NAME_UNKNOWN, 12233 init_dummy_netdev); 12234 } 12235 EXPORT_SYMBOL_GPL(alloc_netdev_dummy); 12236 12237 /** 12238 * synchronize_net - Synchronize with packet receive processing 12239 * 12240 * Wait for packets currently being received to be done. 12241 * Does not block later packets from starting. 12242 */ 12243 void synchronize_net(void) 12244 { 12245 might_sleep(); 12246 if (from_cleanup_net() || rtnl_is_locked()) 12247 synchronize_rcu_expedited(); 12248 else 12249 synchronize_rcu(); 12250 } 12251 EXPORT_SYMBOL(synchronize_net); 12252 12253 static void netdev_rss_contexts_free(struct net_device *dev) 12254 { 12255 struct ethtool_rxfh_context *ctx; 12256 unsigned long context; 12257 12258 mutex_lock(&dev->ethtool->rss_lock); 12259 xa_for_each(&dev->ethtool->rss_ctx, context, ctx) { 12260 xa_erase(&dev->ethtool->rss_ctx, context); 12261 dev->ethtool_ops->remove_rxfh_context(dev, ctx, context, NULL); 12262 kfree(ctx); 12263 } 12264 xa_destroy(&dev->ethtool->rss_ctx); 12265 mutex_unlock(&dev->ethtool->rss_lock); 12266 } 12267 12268 /** 12269 * unregister_netdevice_queue - remove device from the kernel 12270 * @dev: device 12271 * @head: list 12272 * 12273 * This function shuts down a device interface and removes it 12274 * from the kernel tables. 12275 * If head not NULL, device is queued to be unregistered later. 12276 * 12277 * Callers must hold the rtnl semaphore. You may want 12278 * unregister_netdev() instead of this. 12279 */ 12280 12281 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head) 12282 { 12283 ASSERT_RTNL(); 12284 12285 if (head) { 12286 list_move_tail(&dev->unreg_list, head); 12287 } else { 12288 LIST_HEAD(single); 12289 12290 list_add(&dev->unreg_list, &single); 12291 unregister_netdevice_many(&single); 12292 } 12293 } 12294 EXPORT_SYMBOL(unregister_netdevice_queue); 12295 12296 static void dev_memory_provider_uninstall(struct net_device *dev) 12297 { 12298 unsigned int i; 12299 12300 for (i = 0; i < dev->real_num_rx_queues; i++) { 12301 struct netdev_rx_queue *rxq = &dev->_rx[i]; 12302 12303 __netif_mp_uninstall_rxq(rxq, &rxq->mp_params); 12304 } 12305 } 12306 12307 /* devices must be UP and netdev_lock()'d */ 12308 static void netif_close_many_and_unlock(struct list_head *close_head) 12309 { 12310 struct net_device *dev, *tmp; 12311 12312 netif_close_many(close_head, false); 12313 12314 /* ... now unlock them */ 12315 list_for_each_entry_safe(dev, tmp, close_head, close_list) { 12316 netdev_unlock(dev); 12317 list_del_init(&dev->close_list); 12318 } 12319 } 12320 12321 static void netif_close_many_and_unlock_cond(struct list_head *close_head) 12322 { 12323 #ifdef CONFIG_LOCKDEP 12324 /* We can only track up to MAX_LOCK_DEPTH locks per task. 12325 * 12326 * Reserve half the available slots for additional locks possibly 12327 * taken by notifiers and (soft)irqs. 12328 */ 12329 unsigned int limit = MAX_LOCK_DEPTH / 2; 12330 12331 if (lockdep_depth(current) > limit) 12332 netif_close_many_and_unlock(close_head); 12333 #endif 12334 } 12335 12336 bool unregister_netdevice_queued(const struct net_device *dev) 12337 { 12338 ASSERT_RTNL(); 12339 return !list_empty(&dev->unreg_list); 12340 } 12341 12342 void unregister_netdevice_many_notify(struct list_head *head, 12343 u32 portid, const struct nlmsghdr *nlh) 12344 { 12345 struct net_device *dev, *tmp; 12346 LIST_HEAD(close_head); 12347 int cnt = 0; 12348 12349 BUG_ON(dev_boot_phase); 12350 ASSERT_RTNL(); 12351 12352 if (list_empty(head)) 12353 return; 12354 12355 list_for_each_entry_safe(dev, tmp, head, unreg_list) { 12356 /* Some devices call without registering 12357 * for initialization unwind. Remove those 12358 * devices and proceed with the remaining. 12359 */ 12360 if (dev->reg_state == NETREG_UNINITIALIZED) { 12361 pr_debug("unregister_netdevice: device %s/%p never was registered\n", 12362 dev->name, dev); 12363 12364 WARN_ON(1); 12365 list_del(&dev->unreg_list); 12366 continue; 12367 } 12368 dev->dismantle = true; 12369 BUG_ON(dev->reg_state != NETREG_REGISTERED); 12370 } 12371 12372 /* If device is running, close it first. Start with ops locked... */ 12373 list_for_each_entry(dev, head, unreg_list) { 12374 if (!(dev->flags & IFF_UP)) 12375 continue; 12376 if (netdev_need_ops_lock(dev)) { 12377 list_add_tail(&dev->close_list, &close_head); 12378 netdev_lock(dev); 12379 } 12380 netif_close_many_and_unlock_cond(&close_head); 12381 } 12382 netif_close_many_and_unlock(&close_head); 12383 /* ... now go over the rest. */ 12384 list_for_each_entry(dev, head, unreg_list) { 12385 if (!netdev_need_ops_lock(dev)) 12386 list_add_tail(&dev->close_list, &close_head); 12387 } 12388 netif_close_many(&close_head, true); 12389 12390 list_for_each_entry(dev, head, unreg_list) { 12391 /* And unlink it from device chain. */ 12392 unlist_netdevice(dev); 12393 netdev_lock(dev); 12394 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERING); 12395 netdev_unlock(dev); 12396 } 12397 flush_all_backlogs(); 12398 12399 synchronize_net(); 12400 12401 list_for_each_entry(dev, head, unreg_list) { 12402 struct sk_buff *skb = NULL; 12403 12404 /* Shutdown queueing discipline. */ 12405 netdev_lock_ops(dev); 12406 dev_shutdown(dev); 12407 dev_tcx_uninstall(dev); 12408 dev_xdp_uninstall(dev); 12409 dev_memory_provider_uninstall(dev); 12410 netdev_unlock_ops(dev); 12411 bpf_dev_bound_netdev_unregister(dev); 12412 12413 netdev_offload_xstats_disable_all(dev); 12414 12415 /* Notify protocols, that we are about to destroy 12416 * this device. They should clean all the things. 12417 */ 12418 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 12419 12420 if (!(dev->rtnl_link_ops && dev->rtnl_link_initializing)) 12421 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0, 12422 GFP_KERNEL, NULL, 0, 12423 portid, nlh); 12424 12425 /* 12426 * Flush the unicast and multicast chains 12427 */ 12428 dev_uc_flush(dev); 12429 dev_mc_flush(dev); 12430 12431 netdev_name_node_alt_flush(dev); 12432 netdev_name_node_free(dev->name_node); 12433 12434 netdev_rss_contexts_free(dev); 12435 12436 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 12437 12438 if (dev->netdev_ops->ndo_uninit) 12439 dev->netdev_ops->ndo_uninit(dev); 12440 12441 mutex_destroy(&dev->ethtool->rss_lock); 12442 12443 net_shaper_flush_netdev(dev); 12444 12445 if (skb) 12446 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh); 12447 12448 /* Notifier chain MUST detach us all upper devices. */ 12449 WARN_ON(netdev_has_any_upper_dev(dev)); 12450 WARN_ON(netdev_has_any_lower_dev(dev)); 12451 12452 /* Remove entries from kobject tree */ 12453 netdev_unregister_kobject(dev); 12454 #ifdef CONFIG_XPS 12455 /* Remove XPS queueing entries */ 12456 netif_reset_xps_queues_gt(dev, 0); 12457 #endif 12458 } 12459 12460 synchronize_net(); 12461 12462 list_for_each_entry(dev, head, unreg_list) { 12463 netdev_put(dev, &dev->dev_registered_tracker); 12464 net_set_todo(dev); 12465 cnt++; 12466 } 12467 atomic_add(cnt, &dev_unreg_count); 12468 12469 list_del(head); 12470 } 12471 12472 /** 12473 * unregister_netdevice_many - unregister many devices 12474 * @head: list of devices 12475 * 12476 * Note: As most callers use a stack allocated list_head, 12477 * we force a list_del() to make sure stack won't be corrupted later. 12478 */ 12479 void unregister_netdevice_many(struct list_head *head) 12480 { 12481 unregister_netdevice_many_notify(head, 0, NULL); 12482 } 12483 EXPORT_SYMBOL(unregister_netdevice_many); 12484 12485 /** 12486 * unregister_netdev - remove device from the kernel 12487 * @dev: device 12488 * 12489 * This function shuts down a device interface and removes it 12490 * from the kernel tables. 12491 * 12492 * This is just a wrapper for unregister_netdevice that takes 12493 * the rtnl semaphore. In general you want to use this and not 12494 * unregister_netdevice. 12495 */ 12496 void unregister_netdev(struct net_device *dev) 12497 { 12498 rtnl_net_dev_lock(dev); 12499 unregister_netdevice(dev); 12500 rtnl_net_dev_unlock(dev); 12501 } 12502 EXPORT_SYMBOL(unregister_netdev); 12503 12504 int __dev_change_net_namespace(struct net_device *dev, struct net *net, 12505 const char *pat, int new_ifindex, 12506 struct netlink_ext_ack *extack) 12507 { 12508 struct netdev_name_node *name_node; 12509 struct net *net_old = dev_net(dev); 12510 char new_name[IFNAMSIZ] = {}; 12511 int err, new_nsid; 12512 12513 ASSERT_RTNL(); 12514 12515 /* Don't allow namespace local devices to be moved. */ 12516 err = -EINVAL; 12517 if (dev->netns_immutable) { 12518 NL_SET_ERR_MSG(extack, "The interface netns is immutable"); 12519 goto out; 12520 } 12521 12522 /* Ensure the device has been registered */ 12523 if (dev->reg_state != NETREG_REGISTERED) { 12524 NL_SET_ERR_MSG(extack, "The interface isn't registered"); 12525 goto out; 12526 } 12527 12528 /* Get out if there is nothing todo */ 12529 err = 0; 12530 if (net_eq(net_old, net)) 12531 goto out; 12532 12533 /* Pick the destination device name, and ensure 12534 * we can use it in the destination network namespace. 12535 */ 12536 err = -EEXIST; 12537 if (netdev_name_in_use(net, dev->name)) { 12538 /* We get here if we can't use the current device name */ 12539 if (!pat) { 12540 NL_SET_ERR_MSG(extack, 12541 "An interface with the same name exists in the target netns"); 12542 goto out; 12543 } 12544 err = dev_prep_valid_name(net, dev, pat, new_name, EEXIST); 12545 if (err < 0) { 12546 NL_SET_ERR_MSG_FMT(extack, 12547 "Unable to use '%s' for the new interface name in the target netns", 12548 pat); 12549 goto out; 12550 } 12551 } 12552 /* Check that none of the altnames conflicts. */ 12553 err = -EEXIST; 12554 netdev_for_each_altname(dev, name_node) { 12555 if (netdev_name_in_use(net, name_node->name)) { 12556 NL_SET_ERR_MSG_FMT(extack, 12557 "An interface with the altname %s exists in the target netns", 12558 name_node->name); 12559 goto out; 12560 } 12561 } 12562 12563 /* Check that new_ifindex isn't used yet. */ 12564 if (new_ifindex) { 12565 err = dev_index_reserve(net, new_ifindex); 12566 if (err < 0) { 12567 NL_SET_ERR_MSG_FMT(extack, 12568 "The ifindex %d is not available in the target netns", 12569 new_ifindex); 12570 goto out; 12571 } 12572 } else { 12573 /* If there is an ifindex conflict assign a new one */ 12574 err = dev_index_reserve(net, dev->ifindex); 12575 if (err == -EBUSY) 12576 err = dev_index_reserve(net, 0); 12577 if (err < 0) { 12578 NL_SET_ERR_MSG(extack, 12579 "Unable to allocate a new ifindex in the target netns"); 12580 goto out; 12581 } 12582 new_ifindex = err; 12583 } 12584 12585 /* 12586 * And now a mini version of register_netdevice unregister_netdevice. 12587 */ 12588 12589 netdev_lock_ops(dev); 12590 /* If device is running close it first. */ 12591 netif_close(dev); 12592 /* And unlink it from device chain */ 12593 unlist_netdevice(dev); 12594 12595 if (!netdev_need_ops_lock(dev)) 12596 netdev_lock(dev); 12597 dev->moving_ns = true; 12598 netdev_unlock(dev); 12599 12600 synchronize_net(); 12601 12602 /* Shutdown queueing discipline. */ 12603 netdev_lock_ops(dev); 12604 dev_shutdown(dev); 12605 netdev_unlock_ops(dev); 12606 12607 /* Notify protocols, that we are about to destroy 12608 * this device. They should clean all the things. 12609 * 12610 * Note that dev->reg_state stays at NETREG_REGISTERED. 12611 * This is wanted because this way 8021q and macvlan know 12612 * the device is just moving and can keep their slaves up. 12613 */ 12614 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 12615 rcu_barrier(); 12616 12617 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL); 12618 12619 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid, 12620 new_ifindex); 12621 12622 /* 12623 * Flush the unicast and multicast chains 12624 */ 12625 dev_uc_flush(dev); 12626 dev_mc_flush(dev); 12627 12628 /* Send a netdev-removed uevent to the old namespace */ 12629 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE); 12630 netdev_adjacent_del_links(dev); 12631 12632 /* Move per-net netdevice notifiers that are following the netdevice */ 12633 move_netdevice_notifiers_dev_net(dev, net); 12634 12635 /* Actually switch the network namespace */ 12636 netdev_lock(dev); 12637 dev_net_set(dev, net); 12638 netdev_unlock(dev); 12639 dev->ifindex = new_ifindex; 12640 12641 if (new_name[0]) { 12642 /* Rename the netdev to prepared name */ 12643 write_seqlock_bh(&netdev_rename_lock); 12644 strscpy(dev->name, new_name, IFNAMSIZ); 12645 write_sequnlock_bh(&netdev_rename_lock); 12646 } 12647 12648 /* Fixup kobjects */ 12649 dev_set_uevent_suppress(&dev->dev, 1); 12650 err = device_rename(&dev->dev, dev->name); 12651 dev_set_uevent_suppress(&dev->dev, 0); 12652 WARN_ON(err); 12653 12654 /* Send a netdev-add uevent to the new namespace */ 12655 kobject_uevent(&dev->dev.kobj, KOBJ_ADD); 12656 netdev_adjacent_add_links(dev); 12657 12658 /* Adapt owner in case owning user namespace of target network 12659 * namespace is different from the original one. 12660 */ 12661 err = netdev_change_owner(dev, net_old, net); 12662 WARN_ON(err); 12663 12664 netdev_lock(dev); 12665 dev->moving_ns = false; 12666 if (!netdev_need_ops_lock(dev)) 12667 netdev_unlock(dev); 12668 12669 /* Add the device back in the hashes */ 12670 list_netdevice(dev); 12671 /* Notify protocols, that a new device appeared. */ 12672 call_netdevice_notifiers(NETDEV_REGISTER, dev); 12673 netdev_unlock_ops(dev); 12674 12675 /* 12676 * Prevent userspace races by waiting until the network 12677 * device is fully setup before sending notifications. 12678 */ 12679 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 12680 12681 synchronize_net(); 12682 err = 0; 12683 out: 12684 return err; 12685 } 12686 12687 static int dev_cpu_dead(unsigned int oldcpu) 12688 { 12689 struct sk_buff **list_skb; 12690 struct sk_buff *skb; 12691 unsigned int cpu; 12692 struct softnet_data *sd, *oldsd, *remsd = NULL; 12693 12694 local_irq_disable(); 12695 cpu = smp_processor_id(); 12696 sd = &per_cpu(softnet_data, cpu); 12697 oldsd = &per_cpu(softnet_data, oldcpu); 12698 12699 /* Find end of our completion_queue. */ 12700 list_skb = &sd->completion_queue; 12701 while (*list_skb) 12702 list_skb = &(*list_skb)->next; 12703 /* Append completion queue from offline CPU. */ 12704 *list_skb = oldsd->completion_queue; 12705 oldsd->completion_queue = NULL; 12706 12707 /* Append output queue from offline CPU. */ 12708 if (oldsd->output_queue) { 12709 *sd->output_queue_tailp = oldsd->output_queue; 12710 sd->output_queue_tailp = oldsd->output_queue_tailp; 12711 oldsd->output_queue = NULL; 12712 oldsd->output_queue_tailp = &oldsd->output_queue; 12713 } 12714 /* Append NAPI poll list from offline CPU, with one exception : 12715 * process_backlog() must be called by cpu owning percpu backlog. 12716 * We properly handle process_queue & input_pkt_queue later. 12717 */ 12718 while (!list_empty(&oldsd->poll_list)) { 12719 struct napi_struct *napi = list_first_entry(&oldsd->poll_list, 12720 struct napi_struct, 12721 poll_list); 12722 12723 list_del_init(&napi->poll_list); 12724 if (napi->poll == process_backlog) 12725 napi->state &= NAPIF_STATE_THREADED; 12726 else 12727 ____napi_schedule(sd, napi); 12728 } 12729 12730 raise_softirq_irqoff(NET_TX_SOFTIRQ); 12731 local_irq_enable(); 12732 12733 if (!use_backlog_threads()) { 12734 #ifdef CONFIG_RPS 12735 remsd = oldsd->rps_ipi_list; 12736 oldsd->rps_ipi_list = NULL; 12737 #endif 12738 /* send out pending IPI's on offline CPU */ 12739 net_rps_send_ipi(remsd); 12740 } 12741 12742 /* Process offline CPU's input_pkt_queue */ 12743 while ((skb = __skb_dequeue(&oldsd->process_queue))) { 12744 netif_rx(skb); 12745 rps_input_queue_head_incr(oldsd); 12746 } 12747 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) { 12748 netif_rx(skb); 12749 rps_input_queue_head_incr(oldsd); 12750 } 12751 12752 return 0; 12753 } 12754 12755 /** 12756 * netdev_increment_features - increment feature set by one 12757 * @all: current feature set 12758 * @one: new feature set 12759 * @mask: mask feature set 12760 * 12761 * Computes a new feature set after adding a device with feature set 12762 * @one to the master device with current feature set @all. Will not 12763 * enable anything that is off in @mask. Returns the new feature set. 12764 */ 12765 netdev_features_t netdev_increment_features(netdev_features_t all, 12766 netdev_features_t one, netdev_features_t mask) 12767 { 12768 if (mask & NETIF_F_HW_CSUM) 12769 mask |= NETIF_F_CSUM_MASK; 12770 mask |= NETIF_F_VLAN_CHALLENGED; 12771 12772 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask; 12773 all &= one | ~NETIF_F_ALL_FOR_ALL; 12774 12775 /* If one device supports hw checksumming, set for all. */ 12776 if (all & NETIF_F_HW_CSUM) 12777 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM); 12778 12779 return all; 12780 } 12781 EXPORT_SYMBOL(netdev_increment_features); 12782 12783 /** 12784 * netdev_compute_master_upper_features - compute feature from lowers 12785 * @dev: the upper device 12786 * @update_header: whether to update upper device's header_len/headroom/tailroom 12787 * 12788 * Recompute the upper device's feature based on all lower devices. 12789 */ 12790 void netdev_compute_master_upper_features(struct net_device *dev, bool update_header) 12791 { 12792 unsigned int dst_release_flag = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 12793 netdev_features_t gso_partial_features = MASTER_UPPER_DEV_GSO_PARTIAL_FEATURES; 12794 netdev_features_t xfrm_features = MASTER_UPPER_DEV_XFRM_FEATURES; 12795 netdev_features_t mpls_features = MASTER_UPPER_DEV_MPLS_FEATURES; 12796 netdev_features_t vlan_features = MASTER_UPPER_DEV_VLAN_FEATURES; 12797 netdev_features_t enc_features = MASTER_UPPER_DEV_ENC_FEATURES; 12798 unsigned short max_header_len = ETH_HLEN; 12799 unsigned int tso_max_size = TSO_MAX_SIZE; 12800 unsigned short max_headroom = 0; 12801 unsigned short max_tailroom = 0; 12802 u16 tso_max_segs = TSO_MAX_SEGS; 12803 struct net_device *lower_dev; 12804 struct list_head *iter; 12805 12806 mpls_features = netdev_base_features(mpls_features); 12807 vlan_features = netdev_base_features(vlan_features); 12808 enc_features = netdev_base_features(enc_features); 12809 12810 netdev_for_each_lower_dev(dev, lower_dev, iter) { 12811 gso_partial_features = netdev_increment_features(gso_partial_features, 12812 lower_dev->gso_partial_features, 12813 MASTER_UPPER_DEV_GSO_PARTIAL_FEATURES); 12814 12815 vlan_features = netdev_increment_features(vlan_features, 12816 lower_dev->vlan_features, 12817 MASTER_UPPER_DEV_VLAN_FEATURES); 12818 12819 enc_features = netdev_increment_features(enc_features, 12820 lower_dev->hw_enc_features, 12821 MASTER_UPPER_DEV_ENC_FEATURES); 12822 12823 if (IS_ENABLED(CONFIG_XFRM_OFFLOAD)) 12824 xfrm_features = netdev_increment_features(xfrm_features, 12825 lower_dev->hw_enc_features, 12826 MASTER_UPPER_DEV_XFRM_FEATURES); 12827 12828 mpls_features = netdev_increment_features(mpls_features, 12829 lower_dev->mpls_features, 12830 MASTER_UPPER_DEV_MPLS_FEATURES); 12831 12832 dst_release_flag &= lower_dev->priv_flags; 12833 12834 if (update_header) { 12835 max_header_len = max(max_header_len, lower_dev->hard_header_len); 12836 max_headroom = max(max_headroom, lower_dev->needed_headroom); 12837 max_tailroom = max(max_tailroom, lower_dev->needed_tailroom); 12838 } 12839 12840 tso_max_size = min(tso_max_size, lower_dev->tso_max_size); 12841 tso_max_segs = min(tso_max_segs, lower_dev->tso_max_segs); 12842 } 12843 12844 dev->gso_partial_features = gso_partial_features; 12845 dev->vlan_features = vlan_features; 12846 dev->hw_enc_features = enc_features | NETIF_F_GSO_ENCAP_ALL | 12847 NETIF_F_HW_VLAN_CTAG_TX | 12848 NETIF_F_HW_VLAN_STAG_TX; 12849 if (IS_ENABLED(CONFIG_XFRM_OFFLOAD)) 12850 dev->hw_enc_features |= xfrm_features; 12851 dev->mpls_features = mpls_features; 12852 12853 dev->priv_flags &= ~IFF_XMIT_DST_RELEASE; 12854 if ((dev->priv_flags & IFF_XMIT_DST_RELEASE_PERM) && 12855 dst_release_flag == (IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM)) 12856 dev->priv_flags |= IFF_XMIT_DST_RELEASE; 12857 12858 if (update_header) { 12859 dev->hard_header_len = max_header_len; 12860 dev->needed_headroom = max_headroom; 12861 dev->needed_tailroom = max_tailroom; 12862 } 12863 12864 netif_set_tso_max_segs(dev, tso_max_segs); 12865 netif_set_tso_max_size(dev, tso_max_size); 12866 12867 netdev_change_features(dev); 12868 } 12869 EXPORT_SYMBOL(netdev_compute_master_upper_features); 12870 12871 static struct hlist_head * __net_init netdev_create_hash(void) 12872 { 12873 int i; 12874 struct hlist_head *hash; 12875 12876 hash = kmalloc_objs(*hash, NETDEV_HASHENTRIES); 12877 if (hash != NULL) 12878 for (i = 0; i < NETDEV_HASHENTRIES; i++) 12879 INIT_HLIST_HEAD(&hash[i]); 12880 12881 return hash; 12882 } 12883 12884 /* Initialize per network namespace state */ 12885 static int __net_init netdev_init(struct net *net) 12886 { 12887 BUILD_BUG_ON(GRO_HASH_BUCKETS > 12888 BITS_PER_BYTE * sizeof_field(struct gro_node, bitmask)); 12889 12890 INIT_LIST_HEAD(&net->dev_base_head); 12891 12892 net->dev_name_head = netdev_create_hash(); 12893 if (net->dev_name_head == NULL) 12894 goto err_name; 12895 12896 net->dev_index_head = netdev_create_hash(); 12897 if (net->dev_index_head == NULL) 12898 goto err_idx; 12899 12900 xa_init_flags(&net->dev_by_index, XA_FLAGS_ALLOC1); 12901 12902 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain); 12903 12904 return 0; 12905 12906 err_idx: 12907 kfree(net->dev_name_head); 12908 err_name: 12909 return -ENOMEM; 12910 } 12911 12912 /** 12913 * netdev_drivername - network driver for the device 12914 * @dev: network device 12915 * 12916 * Determine network driver for device. 12917 */ 12918 const char *netdev_drivername(const struct net_device *dev) 12919 { 12920 const struct device_driver *driver; 12921 const struct device *parent; 12922 const char *empty = ""; 12923 12924 parent = dev->dev.parent; 12925 if (!parent) 12926 return empty; 12927 12928 driver = parent->driver; 12929 if (driver && driver->name) 12930 return driver->name; 12931 return empty; 12932 } 12933 12934 static void __netdev_printk(const char *level, const struct net_device *dev, 12935 struct va_format *vaf) 12936 { 12937 if (dev && dev->dev.parent) { 12938 dev_printk_emit(level[1] - '0', 12939 dev->dev.parent, 12940 "%s %s %s%s: %pV", 12941 dev_driver_string(dev->dev.parent), 12942 dev_name(dev->dev.parent), 12943 netdev_name(dev), netdev_reg_state(dev), 12944 vaf); 12945 } else if (dev) { 12946 printk("%s%s%s: %pV", 12947 level, netdev_name(dev), netdev_reg_state(dev), vaf); 12948 } else { 12949 printk("%s(NULL net_device): %pV", level, vaf); 12950 } 12951 } 12952 12953 void netdev_printk(const char *level, const struct net_device *dev, 12954 const char *format, ...) 12955 { 12956 struct va_format vaf; 12957 va_list args; 12958 12959 va_start(args, format); 12960 12961 vaf.fmt = format; 12962 vaf.va = &args; 12963 12964 __netdev_printk(level, dev, &vaf); 12965 12966 va_end(args); 12967 } 12968 EXPORT_SYMBOL(netdev_printk); 12969 12970 #define define_netdev_printk_level(func, level) \ 12971 void func(const struct net_device *dev, const char *fmt, ...) \ 12972 { \ 12973 struct va_format vaf; \ 12974 va_list args; \ 12975 \ 12976 va_start(args, fmt); \ 12977 \ 12978 vaf.fmt = fmt; \ 12979 vaf.va = &args; \ 12980 \ 12981 __netdev_printk(level, dev, &vaf); \ 12982 \ 12983 va_end(args); \ 12984 } \ 12985 EXPORT_SYMBOL(func); 12986 12987 define_netdev_printk_level(netdev_emerg, KERN_EMERG); 12988 define_netdev_printk_level(netdev_alert, KERN_ALERT); 12989 define_netdev_printk_level(netdev_crit, KERN_CRIT); 12990 define_netdev_printk_level(netdev_err, KERN_ERR); 12991 define_netdev_printk_level(netdev_warn, KERN_WARNING); 12992 define_netdev_printk_level(netdev_notice, KERN_NOTICE); 12993 define_netdev_printk_level(netdev_info, KERN_INFO); 12994 12995 static void __net_exit netdev_exit(struct net *net) 12996 { 12997 kfree(net->dev_name_head); 12998 kfree(net->dev_index_head); 12999 xa_destroy(&net->dev_by_index); 13000 if (net != &init_net) 13001 WARN_ON_ONCE(!list_empty(&net->dev_base_head)); 13002 } 13003 13004 static struct pernet_operations __net_initdata netdev_net_ops = { 13005 .init = netdev_init, 13006 .exit = netdev_exit, 13007 }; 13008 13009 static void __net_exit default_device_exit_net(struct net *net) 13010 { 13011 struct netdev_name_node *name_node, *tmp; 13012 struct net_device *dev, *aux; 13013 /* 13014 * Push all migratable network devices back to the 13015 * initial network namespace 13016 */ 13017 ASSERT_RTNL(); 13018 for_each_netdev_safe(net, dev, aux) { 13019 int err; 13020 char fb_name[IFNAMSIZ]; 13021 13022 /* Ignore unmoveable devices (i.e. loopback) */ 13023 if (dev->netns_immutable) 13024 continue; 13025 13026 /* Leave virtual devices for the generic cleanup */ 13027 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund) 13028 continue; 13029 13030 /* Push remaining network devices to init_net */ 13031 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex); 13032 if (netdev_name_in_use(&init_net, fb_name)) 13033 snprintf(fb_name, IFNAMSIZ, "dev%%d"); 13034 13035 netdev_for_each_altname_safe(dev, name_node, tmp) 13036 if (netdev_name_in_use(&init_net, name_node->name)) 13037 __netdev_name_node_alt_destroy(name_node); 13038 13039 err = dev_change_net_namespace(dev, &init_net, fb_name); 13040 if (err) { 13041 pr_emerg("%s: failed to move %s to init_net: %d\n", 13042 __func__, dev->name, err); 13043 BUG(); 13044 } 13045 } 13046 } 13047 13048 static void __net_exit default_device_exit_batch(struct list_head *net_list) 13049 { 13050 /* At exit all network devices most be removed from a network 13051 * namespace. Do this in the reverse order of registration. 13052 * Do this across as many network namespaces as possible to 13053 * improve batching efficiency. 13054 */ 13055 struct net_device *dev; 13056 struct net *net; 13057 LIST_HEAD(dev_kill_list); 13058 13059 rtnl_lock(); 13060 list_for_each_entry(net, net_list, exit_list) { 13061 default_device_exit_net(net); 13062 cond_resched(); 13063 } 13064 13065 list_for_each_entry(net, net_list, exit_list) { 13066 for_each_netdev_reverse(net, dev) { 13067 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink) 13068 dev->rtnl_link_ops->dellink(dev, &dev_kill_list); 13069 else 13070 unregister_netdevice_queue(dev, &dev_kill_list); 13071 } 13072 } 13073 unregister_netdevice_many(&dev_kill_list); 13074 rtnl_unlock(); 13075 } 13076 13077 static struct pernet_operations __net_initdata default_device_ops = { 13078 .exit_batch = default_device_exit_batch, 13079 }; 13080 13081 static void __init net_dev_struct_check(void) 13082 { 13083 /* TX read-mostly hotpath */ 13084 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, priv_flags_fast); 13085 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, netdev_ops); 13086 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, header_ops); 13087 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, _tx); 13088 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, real_num_tx_queues); 13089 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_size); 13090 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_ipv4_max_size); 13091 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_segs); 13092 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_partial_features); 13093 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, num_tc); 13094 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, mtu); 13095 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, needed_headroom); 13096 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tc_to_txq); 13097 #ifdef CONFIG_XPS 13098 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, xps_maps); 13099 #endif 13100 #ifdef CONFIG_NETFILTER_EGRESS 13101 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, nf_hooks_egress); 13102 #endif 13103 #ifdef CONFIG_NET_XGRESS 13104 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tcx_egress); 13105 #endif 13106 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_tx, 160); 13107 13108 /* TXRX read-mostly hotpath */ 13109 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, lstats); 13110 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, state); 13111 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, flags); 13112 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, hard_header_len); 13113 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, features); 13114 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, ip6_ptr); 13115 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_txrx, 46); 13116 13117 /* RX read-mostly hotpath */ 13118 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ptype_specific); 13119 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ifindex); 13120 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, real_num_rx_queues); 13121 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, _rx); 13122 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_max_size); 13123 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_ipv4_max_size); 13124 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler); 13125 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler_data); 13126 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, nd_net); 13127 #ifdef CONFIG_NETPOLL 13128 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, npinfo); 13129 #endif 13130 #ifdef CONFIG_NET_XGRESS 13131 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, tcx_ingress); 13132 #endif 13133 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_rx, 92); 13134 } 13135 13136 /* 13137 * Initialize the DEV module. At boot time this walks the device list and 13138 * unhooks any devices that fail to initialise (normally hardware not 13139 * present) and leaves us with a valid list of present and active devices. 13140 * 13141 */ 13142 13143 /* We allocate 256 pages for each CPU if PAGE_SHIFT is 12 */ 13144 #define SYSTEM_PERCPU_PAGE_POOL_SIZE ((1 << 20) / PAGE_SIZE) 13145 13146 static int net_page_pool_create(int cpuid) 13147 { 13148 #if IS_ENABLED(CONFIG_PAGE_POOL) 13149 struct page_pool_params page_pool_params = { 13150 .pool_size = SYSTEM_PERCPU_PAGE_POOL_SIZE, 13151 .flags = PP_FLAG_SYSTEM_POOL, 13152 .nid = cpu_to_mem(cpuid), 13153 }; 13154 struct page_pool *pp_ptr; 13155 int err; 13156 13157 pp_ptr = page_pool_create_percpu(&page_pool_params, cpuid); 13158 if (IS_ERR(pp_ptr)) 13159 return -ENOMEM; 13160 13161 err = xdp_reg_page_pool(pp_ptr); 13162 if (err) { 13163 page_pool_destroy(pp_ptr); 13164 return err; 13165 } 13166 13167 per_cpu(system_page_pool.pool, cpuid) = pp_ptr; 13168 #endif 13169 return 0; 13170 } 13171 13172 static int backlog_napi_should_run(unsigned int cpu) 13173 { 13174 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 13175 struct napi_struct *napi = &sd->backlog; 13176 13177 return test_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 13178 } 13179 13180 static void run_backlog_napi(unsigned int cpu) 13181 { 13182 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 13183 13184 napi_threaded_poll_loop(&sd->backlog, NULL); 13185 } 13186 13187 static void backlog_napi_setup(unsigned int cpu) 13188 { 13189 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 13190 struct napi_struct *napi = &sd->backlog; 13191 13192 napi->thread = this_cpu_read(backlog_napi); 13193 set_bit(NAPI_STATE_THREADED, &napi->state); 13194 } 13195 13196 static struct smp_hotplug_thread backlog_threads = { 13197 .store = &backlog_napi, 13198 .thread_should_run = backlog_napi_should_run, 13199 .thread_fn = run_backlog_napi, 13200 .thread_comm = "backlog_napi/%u", 13201 .setup = backlog_napi_setup, 13202 }; 13203 13204 /* 13205 * This is called single threaded during boot, so no need 13206 * to take the rtnl semaphore. 13207 */ 13208 static int __init net_dev_init(void) 13209 { 13210 int i, rc = -ENOMEM; 13211 13212 BUG_ON(!dev_boot_phase); 13213 13214 net_dev_struct_check(); 13215 13216 if (dev_proc_init()) 13217 goto out; 13218 13219 if (netdev_kobject_init()) 13220 goto out; 13221 13222 for (i = 0; i < PTYPE_HASH_SIZE; i++) 13223 INIT_LIST_HEAD(&ptype_base[i]); 13224 13225 if (register_pernet_subsys(&netdev_net_ops)) 13226 goto out; 13227 13228 /* 13229 * Initialise the packet receive queues. 13230 */ 13231 13232 flush_backlogs_fallback = flush_backlogs_alloc(); 13233 if (!flush_backlogs_fallback) 13234 goto out; 13235 13236 for_each_possible_cpu(i) { 13237 struct softnet_data *sd = &per_cpu(softnet_data, i); 13238 13239 skb_queue_head_init(&sd->input_pkt_queue); 13240 skb_queue_head_init(&sd->process_queue); 13241 #ifdef CONFIG_XFRM_OFFLOAD 13242 skb_queue_head_init(&sd->xfrm_backlog); 13243 #endif 13244 INIT_LIST_HEAD(&sd->poll_list); 13245 sd->output_queue_tailp = &sd->output_queue; 13246 #ifdef CONFIG_RPS 13247 INIT_CSD(&sd->csd, rps_trigger_softirq, sd); 13248 sd->cpu = i; 13249 #endif 13250 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd); 13251 13252 gro_init(&sd->backlog.gro); 13253 sd->backlog.poll = process_backlog; 13254 sd->backlog.weight = weight_p; 13255 INIT_LIST_HEAD(&sd->backlog.poll_list); 13256 13257 if (net_page_pool_create(i)) 13258 goto out; 13259 } 13260 net_hotdata.skb_defer_nodes = 13261 __alloc_percpu(sizeof(struct skb_defer_node) * nr_node_ids, 13262 __alignof__(struct skb_defer_node)); 13263 if (!net_hotdata.skb_defer_nodes) 13264 goto out; 13265 if (use_backlog_threads()) 13266 smpboot_register_percpu_thread(&backlog_threads); 13267 13268 dev_boot_phase = 0; 13269 13270 /* The loopback device is special if any other network devices 13271 * is present in a network namespace the loopback device must 13272 * be present. Since we now dynamically allocate and free the 13273 * loopback device ensure this invariant is maintained by 13274 * keeping the loopback device as the first device on the 13275 * list of network devices. Ensuring the loopback devices 13276 * is the first device that appears and the last network device 13277 * that disappears. 13278 */ 13279 if (register_pernet_device(&loopback_net_ops)) 13280 goto out; 13281 13282 if (register_pernet_device(&default_device_ops)) 13283 goto out; 13284 13285 open_softirq(NET_TX_SOFTIRQ, net_tx_action); 13286 open_softirq(NET_RX_SOFTIRQ, net_rx_action); 13287 13288 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead", 13289 NULL, dev_cpu_dead); 13290 WARN_ON(rc < 0); 13291 rc = 0; 13292 13293 /* avoid static key IPIs to isolated CPUs */ 13294 if (housekeeping_enabled(HK_TYPE_MISC)) 13295 net_enable_timestamp(); 13296 out: 13297 if (rc < 0) { 13298 for_each_possible_cpu(i) { 13299 struct page_pool *pp_ptr; 13300 13301 pp_ptr = per_cpu(system_page_pool.pool, i); 13302 if (!pp_ptr) 13303 continue; 13304 13305 xdp_unreg_page_pool(pp_ptr); 13306 page_pool_destroy(pp_ptr); 13307 per_cpu(system_page_pool.pool, i) = NULL; 13308 } 13309 } 13310 13311 return rc; 13312 } 13313 13314 subsys_initcall(net_dev_init); 13315