1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * NET3 Protocol independent device support routines. 4 * 5 * Derived from the non IP parts of dev.c 1.0.19 6 * Authors: Ross Biro 7 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 8 * Mark Evans, <evansmp@uhura.aston.ac.uk> 9 * 10 * Additional Authors: 11 * Florian la Roche <rzsfl@rz.uni-sb.de> 12 * Alan Cox <gw4pts@gw4pts.ampr.org> 13 * David Hinds <dahinds@users.sourceforge.net> 14 * Alexey Kuznetsov <kuznet@ms2.inr.ac.ru> 15 * Adam Sulmicki <adam@cfar.umd.edu> 16 * Pekka Riikonen <priikone@poesidon.pspt.fi> 17 * 18 * Changes: 19 * D.J. Barrow : Fixed bug where dev->refcnt gets set 20 * to 2 if register_netdev gets called 21 * before net_dev_init & also removed a 22 * few lines of code in the process. 23 * Alan Cox : device private ioctl copies fields back. 24 * Alan Cox : Transmit queue code does relevant 25 * stunts to keep the queue safe. 26 * Alan Cox : Fixed double lock. 27 * Alan Cox : Fixed promisc NULL pointer trap 28 * ???????? : Support the full private ioctl range 29 * Alan Cox : Moved ioctl permission check into 30 * drivers 31 * Tim Kordas : SIOCADDMULTI/SIOCDELMULTI 32 * Alan Cox : 100 backlog just doesn't cut it when 33 * you start doing multicast video 8) 34 * Alan Cox : Rewrote net_bh and list manager. 35 * Alan Cox : Fix ETH_P_ALL echoback lengths. 36 * Alan Cox : Took out transmit every packet pass 37 * Saved a few bytes in the ioctl handler 38 * Alan Cox : Network driver sets packet type before 39 * calling netif_rx. Saves a function 40 * call a packet. 41 * Alan Cox : Hashed net_bh() 42 * Richard Kooijman: Timestamp fixes. 43 * Alan Cox : Wrong field in SIOCGIFDSTADDR 44 * Alan Cox : Device lock protection. 45 * Alan Cox : Fixed nasty side effect of device close 46 * changes. 47 * Rudi Cilibrasi : Pass the right thing to 48 * set_mac_address() 49 * Dave Miller : 32bit quantity for the device lock to 50 * make it work out on a Sparc. 51 * Bjorn Ekwall : Added KERNELD hack. 52 * Alan Cox : Cleaned up the backlog initialise. 53 * Craig Metz : SIOCGIFCONF fix if space for under 54 * 1 device. 55 * Thomas Bogendoerfer : Return ENODEV for dev_open, if there 56 * is no device open function. 57 * Andi Kleen : Fix error reporting for SIOCGIFCONF 58 * Michael Chastain : Fix signed/unsigned for SIOCGIFCONF 59 * Cyrus Durgin : Cleaned for KMOD 60 * Adam Sulmicki : Bug Fix : Network Device Unload 61 * A network device unload needs to purge 62 * the backlog queue. 63 * Paul Rusty Russell : SIOCSIFNAME 64 * Pekka Riikonen : Netdev boot-time settings code 65 * Andrew Morton : Make unregister_netdevice wait 66 * indefinitely on dev->refcnt 67 * J Hadi Salim : - Backlog queue sampling 68 * - netif_rx() feedback 69 */ 70 71 #include <linux/uaccess.h> 72 #include <linux/bitmap.h> 73 #include <linux/capability.h> 74 #include <linux/cpu.h> 75 #include <linux/types.h> 76 #include <linux/kernel.h> 77 #include <linux/hash.h> 78 #include <linux/slab.h> 79 #include <linux/sched.h> 80 #include <linux/sched/isolation.h> 81 #include <linux/sched/mm.h> 82 #include <linux/smpboot.h> 83 #include <linux/mutex.h> 84 #include <linux/rwsem.h> 85 #include <linux/string.h> 86 #include <linux/mm.h> 87 #include <linux/socket.h> 88 #include <linux/sockios.h> 89 #include <linux/errno.h> 90 #include <linux/interrupt.h> 91 #include <linux/if_ether.h> 92 #include <linux/netdevice.h> 93 #include <linux/etherdevice.h> 94 #include <linux/ethtool.h> 95 #include <linux/ethtool_netlink.h> 96 #include <linux/skbuff.h> 97 #include <linux/kthread.h> 98 #include <linux/bpf.h> 99 #include <linux/bpf_trace.h> 100 #include <net/net_namespace.h> 101 #include <net/sock.h> 102 #include <net/busy_poll.h> 103 #include <linux/rtnetlink.h> 104 #include <linux/stat.h> 105 #include <net/dsa.h> 106 #include <net/dst.h> 107 #include <net/dst_metadata.h> 108 #include <net/gro.h> 109 #include <net/netdev_queues.h> 110 #include <net/pkt_sched.h> 111 #include <net/pkt_cls.h> 112 #include <net/checksum.h> 113 #include <net/xfrm.h> 114 #include <net/tcx.h> 115 #include <linux/highmem.h> 116 #include <linux/init.h> 117 #include <linux/module.h> 118 #include <linux/netpoll.h> 119 #include <linux/rcupdate.h> 120 #include <linux/delay.h> 121 #include <net/iw_handler.h> 122 #include <asm/current.h> 123 #include <linux/audit.h> 124 #include <linux/dmaengine.h> 125 #include <linux/err.h> 126 #include <linux/ctype.h> 127 #include <linux/if_arp.h> 128 #include <linux/if_vlan.h> 129 #include <linux/ip.h> 130 #include <net/ip.h> 131 #include <net/mpls.h> 132 #include <linux/ipv6.h> 133 #include <linux/in.h> 134 #include <linux/jhash.h> 135 #include <linux/random.h> 136 #include <trace/events/napi.h> 137 #include <trace/events/net.h> 138 #include <trace/events/skb.h> 139 #include <trace/events/qdisc.h> 140 #include <trace/events/xdp.h> 141 #include <linux/inetdevice.h> 142 #include <linux/cpu_rmap.h> 143 #include <linux/static_key.h> 144 #include <linux/hashtable.h> 145 #include <linux/vmalloc.h> 146 #include <linux/if_macvlan.h> 147 #include <linux/errqueue.h> 148 #include <linux/hrtimer.h> 149 #include <linux/netfilter_netdev.h> 150 #include <linux/crash_dump.h> 151 #include <linux/sctp.h> 152 #include <net/udp_tunnel.h> 153 #include <linux/net_namespace.h> 154 #include <linux/indirect_call_wrapper.h> 155 #include <net/devlink.h> 156 #include <linux/pm_runtime.h> 157 #include <linux/prandom.h> 158 #include <linux/once_lite.h> 159 #include <net/netdev_lock.h> 160 #include <net/netdev_rx_queue.h> 161 #include <net/page_pool/types.h> 162 #include <net/page_pool/helpers.h> 163 #include <net/page_pool/memory_provider.h> 164 #include <net/rps.h> 165 #include <linux/phy_link_topology.h> 166 167 #include "dev.h" 168 #include "devmem.h" 169 #include "net-sysfs.h" 170 171 static DEFINE_SPINLOCK(ptype_lock); 172 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly; 173 174 static int netif_rx_internal(struct sk_buff *skb); 175 static int call_netdevice_notifiers_extack(unsigned long val, 176 struct net_device *dev, 177 struct netlink_ext_ack *extack); 178 179 static DEFINE_MUTEX(ifalias_mutex); 180 181 /* protects napi_hash addition/deletion and napi_gen_id */ 182 static DEFINE_SPINLOCK(napi_hash_lock); 183 184 static unsigned int napi_gen_id = NR_CPUS; 185 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8); 186 187 static inline void dev_base_seq_inc(struct net *net) 188 { 189 unsigned int val = net->dev_base_seq + 1; 190 191 WRITE_ONCE(net->dev_base_seq, val ?: 1); 192 } 193 194 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name) 195 { 196 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ)); 197 198 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)]; 199 } 200 201 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex) 202 { 203 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)]; 204 } 205 206 #ifndef CONFIG_PREEMPT_RT 207 208 static DEFINE_STATIC_KEY_FALSE(use_backlog_threads_key); 209 210 static int __init setup_backlog_napi_threads(char *arg) 211 { 212 static_branch_enable(&use_backlog_threads_key); 213 return 0; 214 } 215 early_param("thread_backlog_napi", setup_backlog_napi_threads); 216 217 static bool use_backlog_threads(void) 218 { 219 return static_branch_unlikely(&use_backlog_threads_key); 220 } 221 222 #else 223 224 static bool use_backlog_threads(void) 225 { 226 return true; 227 } 228 229 #endif 230 231 static inline void backlog_lock_irq_save(struct softnet_data *sd, 232 unsigned long *flags) 233 { 234 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 235 spin_lock_irqsave(&sd->input_pkt_queue.lock, *flags); 236 else 237 local_irq_save(*flags); 238 } 239 240 static inline void backlog_lock_irq_disable(struct softnet_data *sd) 241 { 242 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 243 spin_lock_irq(&sd->input_pkt_queue.lock); 244 else 245 local_irq_disable(); 246 } 247 248 static inline void backlog_unlock_irq_restore(struct softnet_data *sd, 249 unsigned long flags) 250 { 251 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 252 spin_unlock(&sd->input_pkt_queue.lock); 253 local_irq_restore(flags); 254 } 255 256 static inline void backlog_unlock_irq_enable(struct softnet_data *sd) 257 { 258 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 259 spin_unlock_irq(&sd->input_pkt_queue.lock); 260 else 261 local_irq_enable(); 262 } 263 264 static struct netdev_name_node *netdev_name_node_alloc(struct net_device *dev, 265 const char *name) 266 { 267 struct netdev_name_node *name_node; 268 269 name_node = kmalloc(sizeof(*name_node), GFP_KERNEL); 270 if (!name_node) 271 return NULL; 272 INIT_HLIST_NODE(&name_node->hlist); 273 name_node->dev = dev; 274 name_node->name = name; 275 return name_node; 276 } 277 278 static struct netdev_name_node * 279 netdev_name_node_head_alloc(struct net_device *dev) 280 { 281 struct netdev_name_node *name_node; 282 283 name_node = netdev_name_node_alloc(dev, dev->name); 284 if (!name_node) 285 return NULL; 286 INIT_LIST_HEAD(&name_node->list); 287 return name_node; 288 } 289 290 static void netdev_name_node_free(struct netdev_name_node *name_node) 291 { 292 kfree(name_node); 293 } 294 295 static void netdev_name_node_add(struct net *net, 296 struct netdev_name_node *name_node) 297 { 298 hlist_add_head_rcu(&name_node->hlist, 299 dev_name_hash(net, name_node->name)); 300 } 301 302 static void netdev_name_node_del(struct netdev_name_node *name_node) 303 { 304 hlist_del_rcu(&name_node->hlist); 305 } 306 307 static struct netdev_name_node *netdev_name_node_lookup(struct net *net, 308 const char *name) 309 { 310 struct hlist_head *head = dev_name_hash(net, name); 311 struct netdev_name_node *name_node; 312 313 hlist_for_each_entry(name_node, head, hlist) 314 if (!strcmp(name_node->name, name)) 315 return name_node; 316 return NULL; 317 } 318 319 static struct netdev_name_node *netdev_name_node_lookup_rcu(struct net *net, 320 const char *name) 321 { 322 struct hlist_head *head = dev_name_hash(net, name); 323 struct netdev_name_node *name_node; 324 325 hlist_for_each_entry_rcu(name_node, head, hlist) 326 if (!strcmp(name_node->name, name)) 327 return name_node; 328 return NULL; 329 } 330 331 bool netdev_name_in_use(struct net *net, const char *name) 332 { 333 return netdev_name_node_lookup(net, name); 334 } 335 EXPORT_SYMBOL(netdev_name_in_use); 336 337 int netdev_name_node_alt_create(struct net_device *dev, const char *name) 338 { 339 struct netdev_name_node *name_node; 340 struct net *net = dev_net(dev); 341 342 name_node = netdev_name_node_lookup(net, name); 343 if (name_node) 344 return -EEXIST; 345 name_node = netdev_name_node_alloc(dev, name); 346 if (!name_node) 347 return -ENOMEM; 348 netdev_name_node_add(net, name_node); 349 /* The node that holds dev->name acts as a head of per-device list. */ 350 list_add_tail_rcu(&name_node->list, &dev->name_node->list); 351 352 return 0; 353 } 354 355 static void netdev_name_node_alt_free(struct rcu_head *head) 356 { 357 struct netdev_name_node *name_node = 358 container_of(head, struct netdev_name_node, rcu); 359 360 kfree(name_node->name); 361 netdev_name_node_free(name_node); 362 } 363 364 static void __netdev_name_node_alt_destroy(struct netdev_name_node *name_node) 365 { 366 netdev_name_node_del(name_node); 367 list_del(&name_node->list); 368 call_rcu(&name_node->rcu, netdev_name_node_alt_free); 369 } 370 371 int netdev_name_node_alt_destroy(struct net_device *dev, const char *name) 372 { 373 struct netdev_name_node *name_node; 374 struct net *net = dev_net(dev); 375 376 name_node = netdev_name_node_lookup(net, name); 377 if (!name_node) 378 return -ENOENT; 379 /* lookup might have found our primary name or a name belonging 380 * to another device. 381 */ 382 if (name_node == dev->name_node || name_node->dev != dev) 383 return -EINVAL; 384 385 __netdev_name_node_alt_destroy(name_node); 386 return 0; 387 } 388 389 static void netdev_name_node_alt_flush(struct net_device *dev) 390 { 391 struct netdev_name_node *name_node, *tmp; 392 393 list_for_each_entry_safe(name_node, tmp, &dev->name_node->list, list) { 394 list_del(&name_node->list); 395 netdev_name_node_alt_free(&name_node->rcu); 396 } 397 } 398 399 /* Device list insertion */ 400 static void list_netdevice(struct net_device *dev) 401 { 402 struct netdev_name_node *name_node; 403 struct net *net = dev_net(dev); 404 405 ASSERT_RTNL(); 406 407 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head); 408 netdev_name_node_add(net, dev->name_node); 409 hlist_add_head_rcu(&dev->index_hlist, 410 dev_index_hash(net, dev->ifindex)); 411 412 netdev_for_each_altname(dev, name_node) 413 netdev_name_node_add(net, name_node); 414 415 /* We reserved the ifindex, this can't fail */ 416 WARN_ON(xa_store(&net->dev_by_index, dev->ifindex, dev, GFP_KERNEL)); 417 418 dev_base_seq_inc(net); 419 } 420 421 /* Device list removal 422 * caller must respect a RCU grace period before freeing/reusing dev 423 */ 424 static void unlist_netdevice(struct net_device *dev) 425 { 426 struct netdev_name_node *name_node; 427 struct net *net = dev_net(dev); 428 429 ASSERT_RTNL(); 430 431 xa_erase(&net->dev_by_index, dev->ifindex); 432 433 netdev_for_each_altname(dev, name_node) 434 netdev_name_node_del(name_node); 435 436 /* Unlink dev from the device chain */ 437 list_del_rcu(&dev->dev_list); 438 netdev_name_node_del(dev->name_node); 439 hlist_del_rcu(&dev->index_hlist); 440 441 dev_base_seq_inc(dev_net(dev)); 442 } 443 444 /* 445 * Our notifier list 446 */ 447 448 static RAW_NOTIFIER_HEAD(netdev_chain); 449 450 /* 451 * Device drivers call our routines to queue packets here. We empty the 452 * queue in the local softnet handler. 453 */ 454 455 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data) = { 456 .process_queue_bh_lock = INIT_LOCAL_LOCK(process_queue_bh_lock), 457 }; 458 EXPORT_PER_CPU_SYMBOL(softnet_data); 459 460 /* Page_pool has a lockless array/stack to alloc/recycle pages. 461 * PP consumers must pay attention to run APIs in the appropriate context 462 * (e.g. NAPI context). 463 */ 464 DEFINE_PER_CPU(struct page_pool_bh, system_page_pool) = { 465 .bh_lock = INIT_LOCAL_LOCK(bh_lock), 466 }; 467 468 #ifdef CONFIG_LOCKDEP 469 /* 470 * register_netdevice() inits txq->_xmit_lock and sets lockdep class 471 * according to dev->type 472 */ 473 static const unsigned short netdev_lock_type[] = { 474 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25, 475 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET, 476 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM, 477 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP, 478 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD, 479 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25, 480 ARPHRD_CAN, ARPHRD_MCTP, 481 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP, 482 ARPHRD_RAWHDLC, ARPHRD_RAWIP, 483 ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD, 484 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI, 485 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE, 486 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET, 487 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL, 488 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM, 489 ARPHRD_IEEE80211_RADIOTAP, 490 ARPHRD_IEEE802154, ARPHRD_IEEE802154_MONITOR, 491 ARPHRD_PHONET, ARPHRD_PHONET_PIPE, 492 ARPHRD_CAIF, ARPHRD_IP6GRE, ARPHRD_NETLINK, ARPHRD_6LOWPAN, 493 ARPHRD_VSOCKMON, 494 ARPHRD_VOID, ARPHRD_NONE}; 495 496 static const char *const netdev_lock_name[] = { 497 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25", 498 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET", 499 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM", 500 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP", 501 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD", 502 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25", 503 "_xmit_CAN", "_xmit_MCTP", 504 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP", 505 "_xmit_RAWHDLC", "_xmit_RAWIP", 506 "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD", 507 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI", 508 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE", 509 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET", 510 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL", 511 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM", 512 "_xmit_IEEE80211_RADIOTAP", 513 "_xmit_IEEE802154", "_xmit_IEEE802154_MONITOR", 514 "_xmit_PHONET", "_xmit_PHONET_PIPE", 515 "_xmit_CAIF", "_xmit_IP6GRE", "_xmit_NETLINK", "_xmit_6LOWPAN", 516 "_xmit_VSOCKMON", 517 "_xmit_VOID", "_xmit_NONE"}; 518 519 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)]; 520 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)]; 521 522 static inline unsigned short netdev_lock_pos(unsigned short dev_type) 523 { 524 int i; 525 526 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++) 527 if (netdev_lock_type[i] == dev_type) 528 return i; 529 /* the last key is used by default */ 530 WARN_ONCE(1, "netdev_lock_pos() could not find dev_type=%u\n", dev_type); 531 return ARRAY_SIZE(netdev_lock_type) - 1; 532 } 533 534 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 535 unsigned short dev_type) 536 { 537 int i; 538 539 i = netdev_lock_pos(dev_type); 540 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i], 541 netdev_lock_name[i]); 542 } 543 544 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 545 { 546 int i; 547 548 i = netdev_lock_pos(dev->type); 549 lockdep_set_class_and_name(&dev->addr_list_lock, 550 &netdev_addr_lock_key[i], 551 netdev_lock_name[i]); 552 } 553 #else 554 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 555 unsigned short dev_type) 556 { 557 } 558 559 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 560 { 561 } 562 #endif 563 564 /******************************************************************************* 565 * 566 * Protocol management and registration routines 567 * 568 *******************************************************************************/ 569 570 571 /* 572 * Add a protocol ID to the list. Now that the input handler is 573 * smarter we can dispense with all the messy stuff that used to be 574 * here. 575 * 576 * BEWARE!!! Protocol handlers, mangling input packets, 577 * MUST BE last in hash buckets and checking protocol handlers 578 * MUST start from promiscuous ptype_all chain in net_bh. 579 * It is true now, do not change it. 580 * Explanation follows: if protocol handler, mangling packet, will 581 * be the first on list, it is not able to sense, that packet 582 * is cloned and should be copied-on-write, so that it will 583 * change it and subsequent readers will get broken packet. 584 * --ANK (980803) 585 */ 586 587 static inline struct list_head *ptype_head(const struct packet_type *pt) 588 { 589 if (pt->type == htons(ETH_P_ALL)) { 590 if (!pt->af_packet_net && !pt->dev) 591 return NULL; 592 593 return pt->dev ? &pt->dev->ptype_all : 594 &pt->af_packet_net->ptype_all; 595 } 596 597 if (pt->dev) 598 return &pt->dev->ptype_specific; 599 600 return pt->af_packet_net ? &pt->af_packet_net->ptype_specific : 601 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK]; 602 } 603 604 /** 605 * dev_add_pack - add packet handler 606 * @pt: packet type declaration 607 * 608 * Add a protocol handler to the networking stack. The passed &packet_type 609 * is linked into kernel lists and may not be freed until it has been 610 * removed from the kernel lists. 611 * 612 * This call does not sleep therefore it can not 613 * guarantee all CPU's that are in middle of receiving packets 614 * will see the new packet type (until the next received packet). 615 */ 616 617 void dev_add_pack(struct packet_type *pt) 618 { 619 struct list_head *head = ptype_head(pt); 620 621 if (WARN_ON_ONCE(!head)) 622 return; 623 624 spin_lock(&ptype_lock); 625 list_add_rcu(&pt->list, head); 626 spin_unlock(&ptype_lock); 627 } 628 EXPORT_SYMBOL(dev_add_pack); 629 630 /** 631 * __dev_remove_pack - remove packet handler 632 * @pt: packet type declaration 633 * 634 * Remove a protocol handler that was previously added to the kernel 635 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 636 * from the kernel lists and can be freed or reused once this function 637 * returns. 638 * 639 * The packet type might still be in use by receivers 640 * and must not be freed until after all the CPU's have gone 641 * through a quiescent state. 642 */ 643 void __dev_remove_pack(struct packet_type *pt) 644 { 645 struct list_head *head = ptype_head(pt); 646 struct packet_type *pt1; 647 648 if (!head) 649 return; 650 651 spin_lock(&ptype_lock); 652 653 list_for_each_entry(pt1, head, list) { 654 if (pt == pt1) { 655 list_del_rcu(&pt->list); 656 goto out; 657 } 658 } 659 660 pr_warn("dev_remove_pack: %p not found\n", pt); 661 out: 662 spin_unlock(&ptype_lock); 663 } 664 EXPORT_SYMBOL(__dev_remove_pack); 665 666 /** 667 * dev_remove_pack - remove packet handler 668 * @pt: packet type declaration 669 * 670 * Remove a protocol handler that was previously added to the kernel 671 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 672 * from the kernel lists and can be freed or reused once this function 673 * returns. 674 * 675 * This call sleeps to guarantee that no CPU is looking at the packet 676 * type after return. 677 */ 678 void dev_remove_pack(struct packet_type *pt) 679 { 680 __dev_remove_pack(pt); 681 682 synchronize_net(); 683 } 684 EXPORT_SYMBOL(dev_remove_pack); 685 686 687 /******************************************************************************* 688 * 689 * Device Interface Subroutines 690 * 691 *******************************************************************************/ 692 693 /** 694 * dev_get_iflink - get 'iflink' value of a interface 695 * @dev: targeted interface 696 * 697 * Indicates the ifindex the interface is linked to. 698 * Physical interfaces have the same 'ifindex' and 'iflink' values. 699 */ 700 701 int dev_get_iflink(const struct net_device *dev) 702 { 703 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink) 704 return dev->netdev_ops->ndo_get_iflink(dev); 705 706 return READ_ONCE(dev->ifindex); 707 } 708 EXPORT_SYMBOL(dev_get_iflink); 709 710 /** 711 * dev_fill_metadata_dst - Retrieve tunnel egress information. 712 * @dev: targeted interface 713 * @skb: The packet. 714 * 715 * For better visibility of tunnel traffic OVS needs to retrieve 716 * egress tunnel information for a packet. Following API allows 717 * user to get this info. 718 */ 719 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb) 720 { 721 struct ip_tunnel_info *info; 722 723 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst) 724 return -EINVAL; 725 726 info = skb_tunnel_info_unclone(skb); 727 if (!info) 728 return -ENOMEM; 729 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX))) 730 return -EINVAL; 731 732 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb); 733 } 734 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst); 735 736 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack) 737 { 738 int k = stack->num_paths++; 739 740 if (WARN_ON_ONCE(k >= NET_DEVICE_PATH_STACK_MAX)) 741 return NULL; 742 743 return &stack->path[k]; 744 } 745 746 int dev_fill_forward_path(const struct net_device *dev, const u8 *daddr, 747 struct net_device_path_stack *stack) 748 { 749 const struct net_device *last_dev; 750 struct net_device_path_ctx ctx = { 751 .dev = dev, 752 }; 753 struct net_device_path *path; 754 int ret = 0; 755 756 memcpy(ctx.daddr, daddr, sizeof(ctx.daddr)); 757 stack->num_paths = 0; 758 while (ctx.dev && ctx.dev->netdev_ops->ndo_fill_forward_path) { 759 last_dev = ctx.dev; 760 path = dev_fwd_path(stack); 761 if (!path) 762 return -1; 763 764 memset(path, 0, sizeof(struct net_device_path)); 765 ret = ctx.dev->netdev_ops->ndo_fill_forward_path(&ctx, path); 766 if (ret < 0) 767 return -1; 768 769 if (WARN_ON_ONCE(last_dev == ctx.dev)) 770 return -1; 771 } 772 773 if (!ctx.dev) 774 return ret; 775 776 path = dev_fwd_path(stack); 777 if (!path) 778 return -1; 779 path->type = DEV_PATH_ETHERNET; 780 path->dev = ctx.dev; 781 782 return ret; 783 } 784 EXPORT_SYMBOL_GPL(dev_fill_forward_path); 785 786 /* must be called under rcu_read_lock(), as we dont take a reference */ 787 static struct napi_struct *napi_by_id(unsigned int napi_id) 788 { 789 unsigned int hash = napi_id % HASH_SIZE(napi_hash); 790 struct napi_struct *napi; 791 792 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node) 793 if (napi->napi_id == napi_id) 794 return napi; 795 796 return NULL; 797 } 798 799 /* must be called under rcu_read_lock(), as we dont take a reference */ 800 static struct napi_struct * 801 netdev_napi_by_id(struct net *net, unsigned int napi_id) 802 { 803 struct napi_struct *napi; 804 805 napi = napi_by_id(napi_id); 806 if (!napi) 807 return NULL; 808 809 if (WARN_ON_ONCE(!napi->dev)) 810 return NULL; 811 if (!net_eq(net, dev_net(napi->dev))) 812 return NULL; 813 814 return napi; 815 } 816 817 /** 818 * netdev_napi_by_id_lock() - find a device by NAPI ID and lock it 819 * @net: the applicable net namespace 820 * @napi_id: ID of a NAPI of a target device 821 * 822 * Find a NAPI instance with @napi_id. Lock its device. 823 * The device must be in %NETREG_REGISTERED state for lookup to succeed. 824 * netdev_unlock() must be called to release it. 825 * 826 * Return: pointer to NAPI, its device with lock held, NULL if not found. 827 */ 828 struct napi_struct * 829 netdev_napi_by_id_lock(struct net *net, unsigned int napi_id) 830 { 831 struct napi_struct *napi; 832 struct net_device *dev; 833 834 rcu_read_lock(); 835 napi = netdev_napi_by_id(net, napi_id); 836 if (!napi || READ_ONCE(napi->dev->reg_state) != NETREG_REGISTERED) { 837 rcu_read_unlock(); 838 return NULL; 839 } 840 841 dev = napi->dev; 842 dev_hold(dev); 843 rcu_read_unlock(); 844 845 dev = __netdev_put_lock(dev, net); 846 if (!dev) 847 return NULL; 848 849 rcu_read_lock(); 850 napi = netdev_napi_by_id(net, napi_id); 851 if (napi && napi->dev != dev) 852 napi = NULL; 853 rcu_read_unlock(); 854 855 if (!napi) 856 netdev_unlock(dev); 857 return napi; 858 } 859 860 /** 861 * __dev_get_by_name - find a device by its name 862 * @net: the applicable net namespace 863 * @name: name to find 864 * 865 * Find an interface by name. Must be called under RTNL semaphore. 866 * If the name is found a pointer to the device is returned. 867 * If the name is not found then %NULL is returned. The 868 * reference counters are not incremented so the caller must be 869 * careful with locks. 870 */ 871 872 struct net_device *__dev_get_by_name(struct net *net, const char *name) 873 { 874 struct netdev_name_node *node_name; 875 876 node_name = netdev_name_node_lookup(net, name); 877 return node_name ? node_name->dev : NULL; 878 } 879 EXPORT_SYMBOL(__dev_get_by_name); 880 881 /** 882 * dev_get_by_name_rcu - find a device by its name 883 * @net: the applicable net namespace 884 * @name: name to find 885 * 886 * Find an interface by name. 887 * If the name is found a pointer to the device is returned. 888 * If the name is not found then %NULL is returned. 889 * The reference counters are not incremented so the caller must be 890 * careful with locks. The caller must hold RCU lock. 891 */ 892 893 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name) 894 { 895 struct netdev_name_node *node_name; 896 897 node_name = netdev_name_node_lookup_rcu(net, name); 898 return node_name ? node_name->dev : NULL; 899 } 900 EXPORT_SYMBOL(dev_get_by_name_rcu); 901 902 /* Deprecated for new users, call netdev_get_by_name() instead */ 903 struct net_device *dev_get_by_name(struct net *net, const char *name) 904 { 905 struct net_device *dev; 906 907 rcu_read_lock(); 908 dev = dev_get_by_name_rcu(net, name); 909 dev_hold(dev); 910 rcu_read_unlock(); 911 return dev; 912 } 913 EXPORT_SYMBOL(dev_get_by_name); 914 915 /** 916 * netdev_get_by_name() - find a device by its name 917 * @net: the applicable net namespace 918 * @name: name to find 919 * @tracker: tracking object for the acquired reference 920 * @gfp: allocation flags for the tracker 921 * 922 * Find an interface by name. This can be called from any 923 * context and does its own locking. The returned handle has 924 * the usage count incremented and the caller must use netdev_put() to 925 * release it when it is no longer needed. %NULL is returned if no 926 * matching device is found. 927 */ 928 struct net_device *netdev_get_by_name(struct net *net, const char *name, 929 netdevice_tracker *tracker, gfp_t gfp) 930 { 931 struct net_device *dev; 932 933 dev = dev_get_by_name(net, name); 934 if (dev) 935 netdev_tracker_alloc(dev, tracker, gfp); 936 return dev; 937 } 938 EXPORT_SYMBOL(netdev_get_by_name); 939 940 /** 941 * __dev_get_by_index - find a device by its ifindex 942 * @net: the applicable net namespace 943 * @ifindex: index of device 944 * 945 * Search for an interface by index. Returns %NULL if the device 946 * is not found or a pointer to the device. The device has not 947 * had its reference counter increased so the caller must be careful 948 * about locking. The caller must hold the RTNL semaphore. 949 */ 950 951 struct net_device *__dev_get_by_index(struct net *net, int ifindex) 952 { 953 struct net_device *dev; 954 struct hlist_head *head = dev_index_hash(net, ifindex); 955 956 hlist_for_each_entry(dev, head, index_hlist) 957 if (dev->ifindex == ifindex) 958 return dev; 959 960 return NULL; 961 } 962 EXPORT_SYMBOL(__dev_get_by_index); 963 964 /** 965 * dev_get_by_index_rcu - find a device by its ifindex 966 * @net: the applicable net namespace 967 * @ifindex: index of device 968 * 969 * Search for an interface by index. Returns %NULL if the device 970 * is not found or a pointer to the device. The device has not 971 * had its reference counter increased so the caller must be careful 972 * about locking. The caller must hold RCU lock. 973 */ 974 975 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex) 976 { 977 struct net_device *dev; 978 struct hlist_head *head = dev_index_hash(net, ifindex); 979 980 hlist_for_each_entry_rcu(dev, head, index_hlist) 981 if (dev->ifindex == ifindex) 982 return dev; 983 984 return NULL; 985 } 986 EXPORT_SYMBOL(dev_get_by_index_rcu); 987 988 /* Deprecated for new users, call netdev_get_by_index() instead */ 989 struct net_device *dev_get_by_index(struct net *net, int ifindex) 990 { 991 struct net_device *dev; 992 993 rcu_read_lock(); 994 dev = dev_get_by_index_rcu(net, ifindex); 995 dev_hold(dev); 996 rcu_read_unlock(); 997 return dev; 998 } 999 EXPORT_SYMBOL(dev_get_by_index); 1000 1001 /** 1002 * netdev_get_by_index() - find a device by its ifindex 1003 * @net: the applicable net namespace 1004 * @ifindex: index of device 1005 * @tracker: tracking object for the acquired reference 1006 * @gfp: allocation flags for the tracker 1007 * 1008 * Search for an interface by index. Returns NULL if the device 1009 * is not found or a pointer to the device. The device returned has 1010 * had a reference added and the pointer is safe until the user calls 1011 * netdev_put() to indicate they have finished with it. 1012 */ 1013 struct net_device *netdev_get_by_index(struct net *net, int ifindex, 1014 netdevice_tracker *tracker, gfp_t gfp) 1015 { 1016 struct net_device *dev; 1017 1018 dev = dev_get_by_index(net, ifindex); 1019 if (dev) 1020 netdev_tracker_alloc(dev, tracker, gfp); 1021 return dev; 1022 } 1023 EXPORT_SYMBOL(netdev_get_by_index); 1024 1025 /** 1026 * dev_get_by_napi_id - find a device by napi_id 1027 * @napi_id: ID of the NAPI struct 1028 * 1029 * Search for an interface by NAPI ID. Returns %NULL if the device 1030 * is not found or a pointer to the device. The device has not had 1031 * its reference counter increased so the caller must be careful 1032 * about locking. The caller must hold RCU lock. 1033 */ 1034 struct net_device *dev_get_by_napi_id(unsigned int napi_id) 1035 { 1036 struct napi_struct *napi; 1037 1038 WARN_ON_ONCE(!rcu_read_lock_held()); 1039 1040 if (!napi_id_valid(napi_id)) 1041 return NULL; 1042 1043 napi = napi_by_id(napi_id); 1044 1045 return napi ? napi->dev : NULL; 1046 } 1047 1048 /* Release the held reference on the net_device, and if the net_device 1049 * is still registered try to lock the instance lock. If device is being 1050 * unregistered NULL will be returned (but the reference has been released, 1051 * either way!) 1052 * 1053 * This helper is intended for locking net_device after it has been looked up 1054 * using a lockless lookup helper. Lock prevents the instance from going away. 1055 */ 1056 struct net_device *__netdev_put_lock(struct net_device *dev, struct net *net) 1057 { 1058 netdev_lock(dev); 1059 if (dev->reg_state > NETREG_REGISTERED || 1060 dev->moving_ns || !net_eq(dev_net(dev), net)) { 1061 netdev_unlock(dev); 1062 dev_put(dev); 1063 return NULL; 1064 } 1065 dev_put(dev); 1066 return dev; 1067 } 1068 1069 static struct net_device * 1070 __netdev_put_lock_ops_compat(struct net_device *dev, struct net *net) 1071 { 1072 netdev_lock_ops_compat(dev); 1073 if (dev->reg_state > NETREG_REGISTERED || 1074 dev->moving_ns || !net_eq(dev_net(dev), net)) { 1075 netdev_unlock_ops_compat(dev); 1076 dev_put(dev); 1077 return NULL; 1078 } 1079 dev_put(dev); 1080 return dev; 1081 } 1082 1083 /** 1084 * netdev_get_by_index_lock() - find a device by its ifindex 1085 * @net: the applicable net namespace 1086 * @ifindex: index of device 1087 * 1088 * Search for an interface by index. If a valid device 1089 * with @ifindex is found it will be returned with netdev->lock held. 1090 * netdev_unlock() must be called to release it. 1091 * 1092 * Return: pointer to a device with lock held, NULL if not found. 1093 */ 1094 struct net_device *netdev_get_by_index_lock(struct net *net, int ifindex) 1095 { 1096 struct net_device *dev; 1097 1098 dev = dev_get_by_index(net, ifindex); 1099 if (!dev) 1100 return NULL; 1101 1102 return __netdev_put_lock(dev, net); 1103 } 1104 1105 struct net_device * 1106 netdev_get_by_index_lock_ops_compat(struct net *net, int ifindex) 1107 { 1108 struct net_device *dev; 1109 1110 dev = dev_get_by_index(net, ifindex); 1111 if (!dev) 1112 return NULL; 1113 1114 return __netdev_put_lock_ops_compat(dev, net); 1115 } 1116 1117 struct net_device * 1118 netdev_xa_find_lock(struct net *net, struct net_device *dev, 1119 unsigned long *index) 1120 { 1121 if (dev) 1122 netdev_unlock(dev); 1123 1124 do { 1125 rcu_read_lock(); 1126 dev = xa_find(&net->dev_by_index, index, ULONG_MAX, XA_PRESENT); 1127 if (!dev) { 1128 rcu_read_unlock(); 1129 return NULL; 1130 } 1131 dev_hold(dev); 1132 rcu_read_unlock(); 1133 1134 dev = __netdev_put_lock(dev, net); 1135 if (dev) 1136 return dev; 1137 1138 (*index)++; 1139 } while (true); 1140 } 1141 1142 struct net_device * 1143 netdev_xa_find_lock_ops_compat(struct net *net, struct net_device *dev, 1144 unsigned long *index) 1145 { 1146 if (dev) 1147 netdev_unlock_ops_compat(dev); 1148 1149 do { 1150 rcu_read_lock(); 1151 dev = xa_find(&net->dev_by_index, index, ULONG_MAX, XA_PRESENT); 1152 if (!dev) { 1153 rcu_read_unlock(); 1154 return NULL; 1155 } 1156 dev_hold(dev); 1157 rcu_read_unlock(); 1158 1159 dev = __netdev_put_lock_ops_compat(dev, net); 1160 if (dev) 1161 return dev; 1162 1163 (*index)++; 1164 } while (true); 1165 } 1166 1167 static DEFINE_SEQLOCK(netdev_rename_lock); 1168 1169 void netdev_copy_name(struct net_device *dev, char *name) 1170 { 1171 unsigned int seq; 1172 1173 do { 1174 seq = read_seqbegin(&netdev_rename_lock); 1175 strscpy(name, dev->name, IFNAMSIZ); 1176 } while (read_seqretry(&netdev_rename_lock, seq)); 1177 } 1178 EXPORT_IPV6_MOD_GPL(netdev_copy_name); 1179 1180 /** 1181 * netdev_get_name - get a netdevice name, knowing its ifindex. 1182 * @net: network namespace 1183 * @name: a pointer to the buffer where the name will be stored. 1184 * @ifindex: the ifindex of the interface to get the name from. 1185 */ 1186 int netdev_get_name(struct net *net, char *name, int ifindex) 1187 { 1188 struct net_device *dev; 1189 int ret; 1190 1191 rcu_read_lock(); 1192 1193 dev = dev_get_by_index_rcu(net, ifindex); 1194 if (!dev) { 1195 ret = -ENODEV; 1196 goto out; 1197 } 1198 1199 netdev_copy_name(dev, name); 1200 1201 ret = 0; 1202 out: 1203 rcu_read_unlock(); 1204 return ret; 1205 } 1206 1207 static bool dev_addr_cmp(struct net_device *dev, unsigned short type, 1208 const char *ha) 1209 { 1210 return dev->type == type && !memcmp(dev->dev_addr, ha, dev->addr_len); 1211 } 1212 1213 /** 1214 * dev_getbyhwaddr_rcu - find a device by its hardware address 1215 * @net: the applicable net namespace 1216 * @type: media type of device 1217 * @ha: hardware address 1218 * 1219 * Search for an interface by MAC address. Returns NULL if the device 1220 * is not found or a pointer to the device. 1221 * The caller must hold RCU. 1222 * The returned device has not had its ref count increased 1223 * and the caller must therefore be careful about locking 1224 * 1225 */ 1226 1227 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type, 1228 const char *ha) 1229 { 1230 struct net_device *dev; 1231 1232 for_each_netdev_rcu(net, dev) 1233 if (dev_addr_cmp(dev, type, ha)) 1234 return dev; 1235 1236 return NULL; 1237 } 1238 EXPORT_SYMBOL(dev_getbyhwaddr_rcu); 1239 1240 /** 1241 * dev_getbyhwaddr() - find a device by its hardware address 1242 * @net: the applicable net namespace 1243 * @type: media type of device 1244 * @ha: hardware address 1245 * 1246 * Similar to dev_getbyhwaddr_rcu(), but the owner needs to hold 1247 * rtnl_lock. 1248 * 1249 * Context: rtnl_lock() must be held. 1250 * Return: pointer to the net_device, or NULL if not found 1251 */ 1252 struct net_device *dev_getbyhwaddr(struct net *net, unsigned short type, 1253 const char *ha) 1254 { 1255 struct net_device *dev; 1256 1257 ASSERT_RTNL(); 1258 for_each_netdev(net, dev) 1259 if (dev_addr_cmp(dev, type, ha)) 1260 return dev; 1261 1262 return NULL; 1263 } 1264 EXPORT_SYMBOL(dev_getbyhwaddr); 1265 1266 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type) 1267 { 1268 struct net_device *dev, *ret = NULL; 1269 1270 rcu_read_lock(); 1271 for_each_netdev_rcu(net, dev) 1272 if (dev->type == type) { 1273 dev_hold(dev); 1274 ret = dev; 1275 break; 1276 } 1277 rcu_read_unlock(); 1278 return ret; 1279 } 1280 EXPORT_SYMBOL(dev_getfirstbyhwtype); 1281 1282 /** 1283 * netdev_get_by_flags_rcu - find any device with given flags 1284 * @net: the applicable net namespace 1285 * @tracker: tracking object for the acquired reference 1286 * @if_flags: IFF_* values 1287 * @mask: bitmask of bits in if_flags to check 1288 * 1289 * Search for any interface with the given flags. 1290 * 1291 * Context: rcu_read_lock() must be held. 1292 * Returns: NULL if a device is not found or a pointer to the device. 1293 */ 1294 struct net_device *netdev_get_by_flags_rcu(struct net *net, netdevice_tracker *tracker, 1295 unsigned short if_flags, unsigned short mask) 1296 { 1297 struct net_device *dev; 1298 1299 for_each_netdev_rcu(net, dev) { 1300 if (((READ_ONCE(dev->flags) ^ if_flags) & mask) == 0) { 1301 netdev_hold(dev, tracker, GFP_ATOMIC); 1302 return dev; 1303 } 1304 } 1305 1306 return NULL; 1307 } 1308 EXPORT_IPV6_MOD(netdev_get_by_flags_rcu); 1309 1310 /** 1311 * dev_valid_name - check if name is okay for network device 1312 * @name: name string 1313 * 1314 * Network device names need to be valid file names to 1315 * allow sysfs to work. We also disallow any kind of 1316 * whitespace. 1317 */ 1318 bool dev_valid_name(const char *name) 1319 { 1320 if (*name == '\0') 1321 return false; 1322 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ) 1323 return false; 1324 if (!strcmp(name, ".") || !strcmp(name, "..")) 1325 return false; 1326 1327 while (*name) { 1328 if (*name == '/' || *name == ':' || isspace(*name)) 1329 return false; 1330 name++; 1331 } 1332 return true; 1333 } 1334 EXPORT_SYMBOL(dev_valid_name); 1335 1336 /** 1337 * __dev_alloc_name - allocate a name for a device 1338 * @net: network namespace to allocate the device name in 1339 * @name: name format string 1340 * @res: result name string 1341 * 1342 * Passed a format string - eg "lt%d" it will try and find a suitable 1343 * id. It scans list of devices to build up a free map, then chooses 1344 * the first empty slot. The caller must hold the dev_base or rtnl lock 1345 * while allocating the name and adding the device in order to avoid 1346 * duplicates. 1347 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1348 * Returns the number of the unit assigned or a negative errno code. 1349 */ 1350 1351 static int __dev_alloc_name(struct net *net, const char *name, char *res) 1352 { 1353 int i = 0; 1354 const char *p; 1355 const int max_netdevices = 8*PAGE_SIZE; 1356 unsigned long *inuse; 1357 struct net_device *d; 1358 char buf[IFNAMSIZ]; 1359 1360 /* Verify the string as this thing may have come from the user. 1361 * There must be one "%d" and no other "%" characters. 1362 */ 1363 p = strchr(name, '%'); 1364 if (!p || p[1] != 'd' || strchr(p + 2, '%')) 1365 return -EINVAL; 1366 1367 /* Use one page as a bit array of possible slots */ 1368 inuse = bitmap_zalloc(max_netdevices, GFP_ATOMIC); 1369 if (!inuse) 1370 return -ENOMEM; 1371 1372 for_each_netdev(net, d) { 1373 struct netdev_name_node *name_node; 1374 1375 netdev_for_each_altname(d, name_node) { 1376 if (!sscanf(name_node->name, name, &i)) 1377 continue; 1378 if (i < 0 || i >= max_netdevices) 1379 continue; 1380 1381 /* avoid cases where sscanf is not exact inverse of printf */ 1382 snprintf(buf, IFNAMSIZ, name, i); 1383 if (!strncmp(buf, name_node->name, IFNAMSIZ)) 1384 __set_bit(i, inuse); 1385 } 1386 if (!sscanf(d->name, name, &i)) 1387 continue; 1388 if (i < 0 || i >= max_netdevices) 1389 continue; 1390 1391 /* avoid cases where sscanf is not exact inverse of printf */ 1392 snprintf(buf, IFNAMSIZ, name, i); 1393 if (!strncmp(buf, d->name, IFNAMSIZ)) 1394 __set_bit(i, inuse); 1395 } 1396 1397 i = find_first_zero_bit(inuse, max_netdevices); 1398 bitmap_free(inuse); 1399 if (i == max_netdevices) 1400 return -ENFILE; 1401 1402 /* 'res' and 'name' could overlap, use 'buf' as an intermediate buffer */ 1403 strscpy(buf, name, IFNAMSIZ); 1404 snprintf(res, IFNAMSIZ, buf, i); 1405 return i; 1406 } 1407 1408 /* Returns negative errno or allocated unit id (see __dev_alloc_name()) */ 1409 static int dev_prep_valid_name(struct net *net, struct net_device *dev, 1410 const char *want_name, char *out_name, 1411 int dup_errno) 1412 { 1413 if (!dev_valid_name(want_name)) 1414 return -EINVAL; 1415 1416 if (strchr(want_name, '%')) 1417 return __dev_alloc_name(net, want_name, out_name); 1418 1419 if (netdev_name_in_use(net, want_name)) 1420 return -dup_errno; 1421 if (out_name != want_name) 1422 strscpy(out_name, want_name, IFNAMSIZ); 1423 return 0; 1424 } 1425 1426 /** 1427 * dev_alloc_name - allocate a name for a device 1428 * @dev: device 1429 * @name: name format string 1430 * 1431 * Passed a format string - eg "lt%d" it will try and find a suitable 1432 * id. It scans list of devices to build up a free map, then chooses 1433 * the first empty slot. The caller must hold the dev_base or rtnl lock 1434 * while allocating the name and adding the device in order to avoid 1435 * duplicates. 1436 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1437 * Returns the number of the unit assigned or a negative errno code. 1438 */ 1439 1440 int dev_alloc_name(struct net_device *dev, const char *name) 1441 { 1442 return dev_prep_valid_name(dev_net(dev), dev, name, dev->name, ENFILE); 1443 } 1444 EXPORT_SYMBOL(dev_alloc_name); 1445 1446 static int dev_get_valid_name(struct net *net, struct net_device *dev, 1447 const char *name) 1448 { 1449 int ret; 1450 1451 ret = dev_prep_valid_name(net, dev, name, dev->name, EEXIST); 1452 return ret < 0 ? ret : 0; 1453 } 1454 1455 int netif_change_name(struct net_device *dev, const char *newname) 1456 { 1457 struct net *net = dev_net(dev); 1458 unsigned char old_assign_type; 1459 char oldname[IFNAMSIZ]; 1460 int err = 0; 1461 int ret; 1462 1463 ASSERT_RTNL_NET(net); 1464 1465 if (!strncmp(newname, dev->name, IFNAMSIZ)) 1466 return 0; 1467 1468 memcpy(oldname, dev->name, IFNAMSIZ); 1469 1470 write_seqlock_bh(&netdev_rename_lock); 1471 err = dev_get_valid_name(net, dev, newname); 1472 write_sequnlock_bh(&netdev_rename_lock); 1473 1474 if (err < 0) 1475 return err; 1476 1477 if (oldname[0] && !strchr(oldname, '%')) 1478 netdev_info(dev, "renamed from %s%s\n", oldname, 1479 dev->flags & IFF_UP ? " (while UP)" : ""); 1480 1481 old_assign_type = dev->name_assign_type; 1482 WRITE_ONCE(dev->name_assign_type, NET_NAME_RENAMED); 1483 1484 rollback: 1485 ret = device_rename(&dev->dev, dev->name); 1486 if (ret) { 1487 write_seqlock_bh(&netdev_rename_lock); 1488 memcpy(dev->name, oldname, IFNAMSIZ); 1489 write_sequnlock_bh(&netdev_rename_lock); 1490 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1491 return ret; 1492 } 1493 1494 netdev_adjacent_rename_links(dev, oldname); 1495 1496 netdev_name_node_del(dev->name_node); 1497 1498 synchronize_net(); 1499 1500 netdev_name_node_add(net, dev->name_node); 1501 1502 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev); 1503 ret = notifier_to_errno(ret); 1504 1505 if (ret) { 1506 /* err >= 0 after dev_alloc_name() or stores the first errno */ 1507 if (err >= 0) { 1508 err = ret; 1509 write_seqlock_bh(&netdev_rename_lock); 1510 memcpy(dev->name, oldname, IFNAMSIZ); 1511 write_sequnlock_bh(&netdev_rename_lock); 1512 memcpy(oldname, newname, IFNAMSIZ); 1513 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1514 old_assign_type = NET_NAME_RENAMED; 1515 goto rollback; 1516 } else { 1517 netdev_err(dev, "name change rollback failed: %d\n", 1518 ret); 1519 } 1520 } 1521 1522 return err; 1523 } 1524 1525 int netif_set_alias(struct net_device *dev, const char *alias, size_t len) 1526 { 1527 struct dev_ifalias *new_alias = NULL; 1528 1529 if (len >= IFALIASZ) 1530 return -EINVAL; 1531 1532 if (len) { 1533 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL); 1534 if (!new_alias) 1535 return -ENOMEM; 1536 1537 memcpy(new_alias->ifalias, alias, len); 1538 new_alias->ifalias[len] = 0; 1539 } 1540 1541 mutex_lock(&ifalias_mutex); 1542 new_alias = rcu_replace_pointer(dev->ifalias, new_alias, 1543 mutex_is_locked(&ifalias_mutex)); 1544 mutex_unlock(&ifalias_mutex); 1545 1546 if (new_alias) 1547 kfree_rcu(new_alias, rcuhead); 1548 1549 return len; 1550 } 1551 1552 /** 1553 * dev_get_alias - get ifalias of a device 1554 * @dev: device 1555 * @name: buffer to store name of ifalias 1556 * @len: size of buffer 1557 * 1558 * get ifalias for a device. Caller must make sure dev cannot go 1559 * away, e.g. rcu read lock or own a reference count to device. 1560 */ 1561 int dev_get_alias(const struct net_device *dev, char *name, size_t len) 1562 { 1563 const struct dev_ifalias *alias; 1564 int ret = 0; 1565 1566 rcu_read_lock(); 1567 alias = rcu_dereference(dev->ifalias); 1568 if (alias) 1569 ret = snprintf(name, len, "%s", alias->ifalias); 1570 rcu_read_unlock(); 1571 1572 return ret; 1573 } 1574 1575 /** 1576 * netdev_features_change - device changes features 1577 * @dev: device to cause notification 1578 * 1579 * Called to indicate a device has changed features. 1580 */ 1581 void netdev_features_change(struct net_device *dev) 1582 { 1583 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev); 1584 } 1585 EXPORT_SYMBOL(netdev_features_change); 1586 1587 void netif_state_change(struct net_device *dev) 1588 { 1589 netdev_ops_assert_locked_or_invisible(dev); 1590 1591 if (dev->flags & IFF_UP) { 1592 struct netdev_notifier_change_info change_info = { 1593 .info.dev = dev, 1594 }; 1595 1596 call_netdevice_notifiers_info(NETDEV_CHANGE, 1597 &change_info.info); 1598 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL, 0, NULL); 1599 } 1600 } 1601 1602 /** 1603 * __netdev_notify_peers - notify network peers about existence of @dev, 1604 * to be called when rtnl lock is already held. 1605 * @dev: network device 1606 * 1607 * Generate traffic such that interested network peers are aware of 1608 * @dev, such as by generating a gratuitous ARP. This may be used when 1609 * a device wants to inform the rest of the network about some sort of 1610 * reconfiguration such as a failover event or virtual machine 1611 * migration. 1612 */ 1613 void __netdev_notify_peers(struct net_device *dev) 1614 { 1615 ASSERT_RTNL(); 1616 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev); 1617 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev); 1618 } 1619 EXPORT_SYMBOL(__netdev_notify_peers); 1620 1621 /** 1622 * netdev_notify_peers - notify network peers about existence of @dev 1623 * @dev: network device 1624 * 1625 * Generate traffic such that interested network peers are aware of 1626 * @dev, such as by generating a gratuitous ARP. This may be used when 1627 * a device wants to inform the rest of the network about some sort of 1628 * reconfiguration such as a failover event or virtual machine 1629 * migration. 1630 */ 1631 void netdev_notify_peers(struct net_device *dev) 1632 { 1633 rtnl_lock(); 1634 __netdev_notify_peers(dev); 1635 rtnl_unlock(); 1636 } 1637 EXPORT_SYMBOL(netdev_notify_peers); 1638 1639 static int napi_threaded_poll(void *data); 1640 1641 static int napi_kthread_create(struct napi_struct *n) 1642 { 1643 int err = 0; 1644 1645 /* Create and wake up the kthread once to put it in 1646 * TASK_INTERRUPTIBLE mode to avoid the blocked task 1647 * warning and work with loadavg. 1648 */ 1649 n->thread = kthread_run(napi_threaded_poll, n, "napi/%s-%d", 1650 n->dev->name, n->napi_id); 1651 if (IS_ERR(n->thread)) { 1652 err = PTR_ERR(n->thread); 1653 pr_err("kthread_run failed with err %d\n", err); 1654 n->thread = NULL; 1655 } 1656 1657 return err; 1658 } 1659 1660 static int __dev_open(struct net_device *dev, struct netlink_ext_ack *extack) 1661 { 1662 const struct net_device_ops *ops = dev->netdev_ops; 1663 int ret; 1664 1665 ASSERT_RTNL(); 1666 dev_addr_check(dev); 1667 1668 if (!netif_device_present(dev)) { 1669 /* may be detached because parent is runtime-suspended */ 1670 if (dev->dev.parent) 1671 pm_runtime_resume(dev->dev.parent); 1672 if (!netif_device_present(dev)) 1673 return -ENODEV; 1674 } 1675 1676 /* Block netpoll from trying to do any rx path servicing. 1677 * If we don't do this there is a chance ndo_poll_controller 1678 * or ndo_poll may be running while we open the device 1679 */ 1680 netpoll_poll_disable(dev); 1681 1682 ret = call_netdevice_notifiers_extack(NETDEV_PRE_UP, dev, extack); 1683 ret = notifier_to_errno(ret); 1684 if (ret) 1685 return ret; 1686 1687 set_bit(__LINK_STATE_START, &dev->state); 1688 1689 netdev_ops_assert_locked(dev); 1690 1691 if (ops->ndo_validate_addr) 1692 ret = ops->ndo_validate_addr(dev); 1693 1694 if (!ret && ops->ndo_open) 1695 ret = ops->ndo_open(dev); 1696 1697 netpoll_poll_enable(dev); 1698 1699 if (ret) 1700 clear_bit(__LINK_STATE_START, &dev->state); 1701 else { 1702 netif_set_up(dev, true); 1703 dev_set_rx_mode(dev); 1704 dev_activate(dev); 1705 add_device_randomness(dev->dev_addr, dev->addr_len); 1706 } 1707 1708 return ret; 1709 } 1710 1711 int netif_open(struct net_device *dev, struct netlink_ext_ack *extack) 1712 { 1713 int ret; 1714 1715 if (dev->flags & IFF_UP) 1716 return 0; 1717 1718 ret = __dev_open(dev, extack); 1719 if (ret < 0) 1720 return ret; 1721 1722 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1723 call_netdevice_notifiers(NETDEV_UP, dev); 1724 1725 return ret; 1726 } 1727 1728 static void __dev_close_many(struct list_head *head) 1729 { 1730 struct net_device *dev; 1731 1732 ASSERT_RTNL(); 1733 might_sleep(); 1734 1735 list_for_each_entry(dev, head, close_list) { 1736 /* Temporarily disable netpoll until the interface is down */ 1737 netpoll_poll_disable(dev); 1738 1739 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev); 1740 1741 clear_bit(__LINK_STATE_START, &dev->state); 1742 1743 /* Synchronize to scheduled poll. We cannot touch poll list, it 1744 * can be even on different cpu. So just clear netif_running(). 1745 * 1746 * dev->stop() will invoke napi_disable() on all of it's 1747 * napi_struct instances on this device. 1748 */ 1749 smp_mb__after_atomic(); /* Commit netif_running(). */ 1750 } 1751 1752 dev_deactivate_many(head); 1753 1754 list_for_each_entry(dev, head, close_list) { 1755 const struct net_device_ops *ops = dev->netdev_ops; 1756 1757 /* 1758 * Call the device specific close. This cannot fail. 1759 * Only if device is UP 1760 * 1761 * We allow it to be called even after a DETACH hot-plug 1762 * event. 1763 */ 1764 1765 netdev_ops_assert_locked(dev); 1766 1767 if (ops->ndo_stop) 1768 ops->ndo_stop(dev); 1769 1770 netif_set_up(dev, false); 1771 netpoll_poll_enable(dev); 1772 } 1773 } 1774 1775 static void __dev_close(struct net_device *dev) 1776 { 1777 LIST_HEAD(single); 1778 1779 list_add(&dev->close_list, &single); 1780 __dev_close_many(&single); 1781 list_del(&single); 1782 } 1783 1784 void netif_close_many(struct list_head *head, bool unlink) 1785 { 1786 struct net_device *dev, *tmp; 1787 1788 /* Remove the devices that don't need to be closed */ 1789 list_for_each_entry_safe(dev, tmp, head, close_list) 1790 if (!(dev->flags & IFF_UP)) 1791 list_del_init(&dev->close_list); 1792 1793 __dev_close_many(head); 1794 1795 list_for_each_entry_safe(dev, tmp, head, close_list) { 1796 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1797 call_netdevice_notifiers(NETDEV_DOWN, dev); 1798 if (unlink) 1799 list_del_init(&dev->close_list); 1800 } 1801 } 1802 EXPORT_SYMBOL_NS_GPL(netif_close_many, "NETDEV_INTERNAL"); 1803 1804 void netif_close(struct net_device *dev) 1805 { 1806 if (dev->flags & IFF_UP) { 1807 LIST_HEAD(single); 1808 1809 list_add(&dev->close_list, &single); 1810 netif_close_many(&single, true); 1811 list_del(&single); 1812 } 1813 } 1814 EXPORT_SYMBOL(netif_close); 1815 1816 void netif_disable_lro(struct net_device *dev) 1817 { 1818 struct net_device *lower_dev; 1819 struct list_head *iter; 1820 1821 dev->wanted_features &= ~NETIF_F_LRO; 1822 netdev_update_features(dev); 1823 1824 if (unlikely(dev->features & NETIF_F_LRO)) 1825 netdev_WARN(dev, "failed to disable LRO!\n"); 1826 1827 netdev_for_each_lower_dev(dev, lower_dev, iter) { 1828 netdev_lock_ops(lower_dev); 1829 netif_disable_lro(lower_dev); 1830 netdev_unlock_ops(lower_dev); 1831 } 1832 } 1833 EXPORT_IPV6_MOD(netif_disable_lro); 1834 1835 /** 1836 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device 1837 * @dev: device 1838 * 1839 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be 1840 * called under RTNL. This is needed if Generic XDP is installed on 1841 * the device. 1842 */ 1843 static void dev_disable_gro_hw(struct net_device *dev) 1844 { 1845 dev->wanted_features &= ~NETIF_F_GRO_HW; 1846 netdev_update_features(dev); 1847 1848 if (unlikely(dev->features & NETIF_F_GRO_HW)) 1849 netdev_WARN(dev, "failed to disable GRO_HW!\n"); 1850 } 1851 1852 const char *netdev_cmd_to_name(enum netdev_cmd cmd) 1853 { 1854 #define N(val) \ 1855 case NETDEV_##val: \ 1856 return "NETDEV_" __stringify(val); 1857 switch (cmd) { 1858 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER) 1859 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE) 1860 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE) 1861 N(POST_INIT) N(PRE_UNINIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) 1862 N(CHANGEUPPER) N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) 1863 N(BONDING_INFO) N(PRECHANGEUPPER) N(CHANGELOWERSTATE) 1864 N(UDP_TUNNEL_PUSH_INFO) N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN) 1865 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO) 1866 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO) 1867 N(PRE_CHANGEADDR) N(OFFLOAD_XSTATS_ENABLE) N(OFFLOAD_XSTATS_DISABLE) 1868 N(OFFLOAD_XSTATS_REPORT_USED) N(OFFLOAD_XSTATS_REPORT_DELTA) 1869 N(XDP_FEAT_CHANGE) 1870 } 1871 #undef N 1872 return "UNKNOWN_NETDEV_EVENT"; 1873 } 1874 EXPORT_SYMBOL_GPL(netdev_cmd_to_name); 1875 1876 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val, 1877 struct net_device *dev) 1878 { 1879 struct netdev_notifier_info info = { 1880 .dev = dev, 1881 }; 1882 1883 return nb->notifier_call(nb, val, &info); 1884 } 1885 1886 static int call_netdevice_register_notifiers(struct notifier_block *nb, 1887 struct net_device *dev) 1888 { 1889 int err; 1890 1891 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev); 1892 err = notifier_to_errno(err); 1893 if (err) 1894 return err; 1895 1896 if (!(dev->flags & IFF_UP)) 1897 return 0; 1898 1899 call_netdevice_notifier(nb, NETDEV_UP, dev); 1900 return 0; 1901 } 1902 1903 static void call_netdevice_unregister_notifiers(struct notifier_block *nb, 1904 struct net_device *dev) 1905 { 1906 if (dev->flags & IFF_UP) { 1907 call_netdevice_notifier(nb, NETDEV_GOING_DOWN, 1908 dev); 1909 call_netdevice_notifier(nb, NETDEV_DOWN, dev); 1910 } 1911 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev); 1912 } 1913 1914 static int call_netdevice_register_net_notifiers(struct notifier_block *nb, 1915 struct net *net) 1916 { 1917 struct net_device *dev; 1918 int err; 1919 1920 for_each_netdev(net, dev) { 1921 netdev_lock_ops(dev); 1922 err = call_netdevice_register_notifiers(nb, dev); 1923 netdev_unlock_ops(dev); 1924 if (err) 1925 goto rollback; 1926 } 1927 return 0; 1928 1929 rollback: 1930 for_each_netdev_continue_reverse(net, dev) 1931 call_netdevice_unregister_notifiers(nb, dev); 1932 return err; 1933 } 1934 1935 static void call_netdevice_unregister_net_notifiers(struct notifier_block *nb, 1936 struct net *net) 1937 { 1938 struct net_device *dev; 1939 1940 for_each_netdev(net, dev) 1941 call_netdevice_unregister_notifiers(nb, dev); 1942 } 1943 1944 static int dev_boot_phase = 1; 1945 1946 /** 1947 * register_netdevice_notifier - register a network notifier block 1948 * @nb: notifier 1949 * 1950 * Register a notifier to be called when network device events occur. 1951 * The notifier passed is linked into the kernel structures and must 1952 * not be reused until it has been unregistered. A negative errno code 1953 * is returned on a failure. 1954 * 1955 * When registered all registration and up events are replayed 1956 * to the new notifier to allow device to have a race free 1957 * view of the network device list. 1958 */ 1959 1960 int register_netdevice_notifier(struct notifier_block *nb) 1961 { 1962 struct net *net; 1963 int err; 1964 1965 /* Close race with setup_net() and cleanup_net() */ 1966 down_write(&pernet_ops_rwsem); 1967 1968 /* When RTNL is removed, we need protection for netdev_chain. */ 1969 rtnl_lock(); 1970 1971 err = raw_notifier_chain_register(&netdev_chain, nb); 1972 if (err) 1973 goto unlock; 1974 if (dev_boot_phase) 1975 goto unlock; 1976 for_each_net(net) { 1977 __rtnl_net_lock(net); 1978 err = call_netdevice_register_net_notifiers(nb, net); 1979 __rtnl_net_unlock(net); 1980 if (err) 1981 goto rollback; 1982 } 1983 1984 unlock: 1985 rtnl_unlock(); 1986 up_write(&pernet_ops_rwsem); 1987 return err; 1988 1989 rollback: 1990 for_each_net_continue_reverse(net) { 1991 __rtnl_net_lock(net); 1992 call_netdevice_unregister_net_notifiers(nb, net); 1993 __rtnl_net_unlock(net); 1994 } 1995 1996 raw_notifier_chain_unregister(&netdev_chain, nb); 1997 goto unlock; 1998 } 1999 EXPORT_SYMBOL(register_netdevice_notifier); 2000 2001 /** 2002 * unregister_netdevice_notifier - unregister a network notifier block 2003 * @nb: notifier 2004 * 2005 * Unregister a notifier previously registered by 2006 * register_netdevice_notifier(). The notifier is unlinked into the 2007 * kernel structures and may then be reused. A negative errno code 2008 * is returned on a failure. 2009 * 2010 * After unregistering unregister and down device events are synthesized 2011 * for all devices on the device list to the removed notifier to remove 2012 * the need for special case cleanup code. 2013 */ 2014 2015 int unregister_netdevice_notifier(struct notifier_block *nb) 2016 { 2017 struct net *net; 2018 int err; 2019 2020 /* Close race with setup_net() and cleanup_net() */ 2021 down_write(&pernet_ops_rwsem); 2022 rtnl_lock(); 2023 err = raw_notifier_chain_unregister(&netdev_chain, nb); 2024 if (err) 2025 goto unlock; 2026 2027 for_each_net(net) { 2028 __rtnl_net_lock(net); 2029 call_netdevice_unregister_net_notifiers(nb, net); 2030 __rtnl_net_unlock(net); 2031 } 2032 2033 unlock: 2034 rtnl_unlock(); 2035 up_write(&pernet_ops_rwsem); 2036 return err; 2037 } 2038 EXPORT_SYMBOL(unregister_netdevice_notifier); 2039 2040 static int __register_netdevice_notifier_net(struct net *net, 2041 struct notifier_block *nb, 2042 bool ignore_call_fail) 2043 { 2044 int err; 2045 2046 err = raw_notifier_chain_register(&net->netdev_chain, nb); 2047 if (err) 2048 return err; 2049 if (dev_boot_phase) 2050 return 0; 2051 2052 err = call_netdevice_register_net_notifiers(nb, net); 2053 if (err && !ignore_call_fail) 2054 goto chain_unregister; 2055 2056 return 0; 2057 2058 chain_unregister: 2059 raw_notifier_chain_unregister(&net->netdev_chain, nb); 2060 return err; 2061 } 2062 2063 static int __unregister_netdevice_notifier_net(struct net *net, 2064 struct notifier_block *nb) 2065 { 2066 int err; 2067 2068 err = raw_notifier_chain_unregister(&net->netdev_chain, nb); 2069 if (err) 2070 return err; 2071 2072 call_netdevice_unregister_net_notifiers(nb, net); 2073 return 0; 2074 } 2075 2076 /** 2077 * register_netdevice_notifier_net - register a per-netns network notifier block 2078 * @net: network namespace 2079 * @nb: notifier 2080 * 2081 * Register a notifier to be called when network device events occur. 2082 * The notifier passed is linked into the kernel structures and must 2083 * not be reused until it has been unregistered. A negative errno code 2084 * is returned on a failure. 2085 * 2086 * When registered all registration and up events are replayed 2087 * to the new notifier to allow device to have a race free 2088 * view of the network device list. 2089 */ 2090 2091 int register_netdevice_notifier_net(struct net *net, struct notifier_block *nb) 2092 { 2093 int err; 2094 2095 rtnl_net_lock(net); 2096 err = __register_netdevice_notifier_net(net, nb, false); 2097 rtnl_net_unlock(net); 2098 2099 return err; 2100 } 2101 EXPORT_SYMBOL(register_netdevice_notifier_net); 2102 2103 /** 2104 * unregister_netdevice_notifier_net - unregister a per-netns 2105 * network notifier block 2106 * @net: network namespace 2107 * @nb: notifier 2108 * 2109 * Unregister a notifier previously registered by 2110 * register_netdevice_notifier_net(). The notifier is unlinked from the 2111 * kernel structures and may then be reused. A negative errno code 2112 * is returned on a failure. 2113 * 2114 * After unregistering unregister and down device events are synthesized 2115 * for all devices on the device list to the removed notifier to remove 2116 * the need for special case cleanup code. 2117 */ 2118 2119 int unregister_netdevice_notifier_net(struct net *net, 2120 struct notifier_block *nb) 2121 { 2122 int err; 2123 2124 rtnl_net_lock(net); 2125 err = __unregister_netdevice_notifier_net(net, nb); 2126 rtnl_net_unlock(net); 2127 2128 return err; 2129 } 2130 EXPORT_SYMBOL(unregister_netdevice_notifier_net); 2131 2132 static void __move_netdevice_notifier_net(struct net *src_net, 2133 struct net *dst_net, 2134 struct notifier_block *nb) 2135 { 2136 __unregister_netdevice_notifier_net(src_net, nb); 2137 __register_netdevice_notifier_net(dst_net, nb, true); 2138 } 2139 2140 static void rtnl_net_dev_lock(struct net_device *dev) 2141 { 2142 bool again; 2143 2144 do { 2145 struct net *net; 2146 2147 again = false; 2148 2149 /* netns might be being dismantled. */ 2150 rcu_read_lock(); 2151 net = dev_net_rcu(dev); 2152 net_passive_inc(net); 2153 rcu_read_unlock(); 2154 2155 rtnl_net_lock(net); 2156 2157 #ifdef CONFIG_NET_NS 2158 /* dev might have been moved to another netns. */ 2159 if (!net_eq(net, rcu_access_pointer(dev->nd_net.net))) { 2160 rtnl_net_unlock(net); 2161 net_passive_dec(net); 2162 again = true; 2163 } 2164 #endif 2165 } while (again); 2166 } 2167 2168 static void rtnl_net_dev_unlock(struct net_device *dev) 2169 { 2170 struct net *net = dev_net(dev); 2171 2172 rtnl_net_unlock(net); 2173 net_passive_dec(net); 2174 } 2175 2176 int register_netdevice_notifier_dev_net(struct net_device *dev, 2177 struct notifier_block *nb, 2178 struct netdev_net_notifier *nn) 2179 { 2180 int err; 2181 2182 rtnl_net_dev_lock(dev); 2183 err = __register_netdevice_notifier_net(dev_net(dev), nb, false); 2184 if (!err) { 2185 nn->nb = nb; 2186 list_add(&nn->list, &dev->net_notifier_list); 2187 } 2188 rtnl_net_dev_unlock(dev); 2189 2190 return err; 2191 } 2192 EXPORT_SYMBOL(register_netdevice_notifier_dev_net); 2193 2194 int unregister_netdevice_notifier_dev_net(struct net_device *dev, 2195 struct notifier_block *nb, 2196 struct netdev_net_notifier *nn) 2197 { 2198 int err; 2199 2200 rtnl_net_dev_lock(dev); 2201 list_del(&nn->list); 2202 err = __unregister_netdevice_notifier_net(dev_net(dev), nb); 2203 rtnl_net_dev_unlock(dev); 2204 2205 return err; 2206 } 2207 EXPORT_SYMBOL(unregister_netdevice_notifier_dev_net); 2208 2209 static void move_netdevice_notifiers_dev_net(struct net_device *dev, 2210 struct net *net) 2211 { 2212 struct netdev_net_notifier *nn; 2213 2214 list_for_each_entry(nn, &dev->net_notifier_list, list) 2215 __move_netdevice_notifier_net(dev_net(dev), net, nn->nb); 2216 } 2217 2218 /** 2219 * call_netdevice_notifiers_info - call all network notifier blocks 2220 * @val: value passed unmodified to notifier function 2221 * @info: notifier information data 2222 * 2223 * Call all network notifier blocks. Parameters and return value 2224 * are as for raw_notifier_call_chain(). 2225 */ 2226 2227 int call_netdevice_notifiers_info(unsigned long val, 2228 struct netdev_notifier_info *info) 2229 { 2230 struct net *net = dev_net(info->dev); 2231 int ret; 2232 2233 ASSERT_RTNL(); 2234 2235 /* Run per-netns notifier block chain first, then run the global one. 2236 * Hopefully, one day, the global one is going to be removed after 2237 * all notifier block registrators get converted to be per-netns. 2238 */ 2239 ret = raw_notifier_call_chain(&net->netdev_chain, val, info); 2240 if (ret & NOTIFY_STOP_MASK) 2241 return ret; 2242 return raw_notifier_call_chain(&netdev_chain, val, info); 2243 } 2244 2245 /** 2246 * call_netdevice_notifiers_info_robust - call per-netns notifier blocks 2247 * for and rollback on error 2248 * @val_up: value passed unmodified to notifier function 2249 * @val_down: value passed unmodified to the notifier function when 2250 * recovering from an error on @val_up 2251 * @info: notifier information data 2252 * 2253 * Call all per-netns network notifier blocks, but not notifier blocks on 2254 * the global notifier chain. Parameters and return value are as for 2255 * raw_notifier_call_chain_robust(). 2256 */ 2257 2258 static int 2259 call_netdevice_notifiers_info_robust(unsigned long val_up, 2260 unsigned long val_down, 2261 struct netdev_notifier_info *info) 2262 { 2263 struct net *net = dev_net(info->dev); 2264 2265 ASSERT_RTNL(); 2266 2267 return raw_notifier_call_chain_robust(&net->netdev_chain, 2268 val_up, val_down, info); 2269 } 2270 2271 static int call_netdevice_notifiers_extack(unsigned long val, 2272 struct net_device *dev, 2273 struct netlink_ext_ack *extack) 2274 { 2275 struct netdev_notifier_info info = { 2276 .dev = dev, 2277 .extack = extack, 2278 }; 2279 2280 return call_netdevice_notifiers_info(val, &info); 2281 } 2282 2283 /** 2284 * call_netdevice_notifiers - call all network notifier blocks 2285 * @val: value passed unmodified to notifier function 2286 * @dev: net_device pointer passed unmodified to notifier function 2287 * 2288 * Call all network notifier blocks. Parameters and return value 2289 * are as for raw_notifier_call_chain(). 2290 */ 2291 2292 int call_netdevice_notifiers(unsigned long val, struct net_device *dev) 2293 { 2294 return call_netdevice_notifiers_extack(val, dev, NULL); 2295 } 2296 EXPORT_SYMBOL(call_netdevice_notifiers); 2297 2298 /** 2299 * call_netdevice_notifiers_mtu - call all network notifier blocks 2300 * @val: value passed unmodified to notifier function 2301 * @dev: net_device pointer passed unmodified to notifier function 2302 * @arg: additional u32 argument passed to the notifier function 2303 * 2304 * Call all network notifier blocks. Parameters and return value 2305 * are as for raw_notifier_call_chain(). 2306 */ 2307 static int call_netdevice_notifiers_mtu(unsigned long val, 2308 struct net_device *dev, u32 arg) 2309 { 2310 struct netdev_notifier_info_ext info = { 2311 .info.dev = dev, 2312 .ext.mtu = arg, 2313 }; 2314 2315 BUILD_BUG_ON(offsetof(struct netdev_notifier_info_ext, info) != 0); 2316 2317 return call_netdevice_notifiers_info(val, &info.info); 2318 } 2319 2320 #ifdef CONFIG_NET_INGRESS 2321 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key); 2322 2323 void net_inc_ingress_queue(void) 2324 { 2325 static_branch_inc(&ingress_needed_key); 2326 } 2327 EXPORT_SYMBOL_GPL(net_inc_ingress_queue); 2328 2329 void net_dec_ingress_queue(void) 2330 { 2331 static_branch_dec(&ingress_needed_key); 2332 } 2333 EXPORT_SYMBOL_GPL(net_dec_ingress_queue); 2334 #endif 2335 2336 #ifdef CONFIG_NET_EGRESS 2337 static DEFINE_STATIC_KEY_FALSE(egress_needed_key); 2338 2339 void net_inc_egress_queue(void) 2340 { 2341 static_branch_inc(&egress_needed_key); 2342 } 2343 EXPORT_SYMBOL_GPL(net_inc_egress_queue); 2344 2345 void net_dec_egress_queue(void) 2346 { 2347 static_branch_dec(&egress_needed_key); 2348 } 2349 EXPORT_SYMBOL_GPL(net_dec_egress_queue); 2350 #endif 2351 2352 #ifdef CONFIG_NET_CLS_ACT 2353 DEFINE_STATIC_KEY_FALSE(tcf_sw_enabled_key); 2354 EXPORT_SYMBOL(tcf_sw_enabled_key); 2355 #endif 2356 2357 DEFINE_STATIC_KEY_FALSE(netstamp_needed_key); 2358 EXPORT_SYMBOL(netstamp_needed_key); 2359 #ifdef CONFIG_JUMP_LABEL 2360 static atomic_t netstamp_needed_deferred; 2361 static atomic_t netstamp_wanted; 2362 static void netstamp_clear(struct work_struct *work) 2363 { 2364 int deferred = atomic_xchg(&netstamp_needed_deferred, 0); 2365 int wanted; 2366 2367 wanted = atomic_add_return(deferred, &netstamp_wanted); 2368 if (wanted > 0) 2369 static_branch_enable(&netstamp_needed_key); 2370 else 2371 static_branch_disable(&netstamp_needed_key); 2372 } 2373 static DECLARE_WORK(netstamp_work, netstamp_clear); 2374 #endif 2375 2376 void net_enable_timestamp(void) 2377 { 2378 #ifdef CONFIG_JUMP_LABEL 2379 int wanted = atomic_read(&netstamp_wanted); 2380 2381 while (wanted > 0) { 2382 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted + 1)) 2383 return; 2384 } 2385 atomic_inc(&netstamp_needed_deferred); 2386 schedule_work(&netstamp_work); 2387 #else 2388 static_branch_inc(&netstamp_needed_key); 2389 #endif 2390 } 2391 EXPORT_SYMBOL(net_enable_timestamp); 2392 2393 void net_disable_timestamp(void) 2394 { 2395 #ifdef CONFIG_JUMP_LABEL 2396 int wanted = atomic_read(&netstamp_wanted); 2397 2398 while (wanted > 1) { 2399 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted - 1)) 2400 return; 2401 } 2402 atomic_dec(&netstamp_needed_deferred); 2403 schedule_work(&netstamp_work); 2404 #else 2405 static_branch_dec(&netstamp_needed_key); 2406 #endif 2407 } 2408 EXPORT_SYMBOL(net_disable_timestamp); 2409 2410 static inline void net_timestamp_set(struct sk_buff *skb) 2411 { 2412 skb->tstamp = 0; 2413 skb->tstamp_type = SKB_CLOCK_REALTIME; 2414 if (static_branch_unlikely(&netstamp_needed_key)) 2415 skb->tstamp = ktime_get_real(); 2416 } 2417 2418 #define net_timestamp_check(COND, SKB) \ 2419 if (static_branch_unlikely(&netstamp_needed_key)) { \ 2420 if ((COND) && !(SKB)->tstamp) \ 2421 (SKB)->tstamp = ktime_get_real(); \ 2422 } \ 2423 2424 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb) 2425 { 2426 return __is_skb_forwardable(dev, skb, true); 2427 } 2428 EXPORT_SYMBOL_GPL(is_skb_forwardable); 2429 2430 static int __dev_forward_skb2(struct net_device *dev, struct sk_buff *skb, 2431 bool check_mtu) 2432 { 2433 int ret = ____dev_forward_skb(dev, skb, check_mtu); 2434 2435 if (likely(!ret)) { 2436 skb->protocol = eth_type_trans(skb, dev); 2437 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN); 2438 } 2439 2440 return ret; 2441 } 2442 2443 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2444 { 2445 return __dev_forward_skb2(dev, skb, true); 2446 } 2447 EXPORT_SYMBOL_GPL(__dev_forward_skb); 2448 2449 /** 2450 * dev_forward_skb - loopback an skb to another netif 2451 * 2452 * @dev: destination network device 2453 * @skb: buffer to forward 2454 * 2455 * return values: 2456 * NET_RX_SUCCESS (no congestion) 2457 * NET_RX_DROP (packet was dropped, but freed) 2458 * 2459 * dev_forward_skb can be used for injecting an skb from the 2460 * start_xmit function of one device into the receive queue 2461 * of another device. 2462 * 2463 * The receiving device may be in another namespace, so 2464 * we have to clear all information in the skb that could 2465 * impact namespace isolation. 2466 */ 2467 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2468 { 2469 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb); 2470 } 2471 EXPORT_SYMBOL_GPL(dev_forward_skb); 2472 2473 int dev_forward_skb_nomtu(struct net_device *dev, struct sk_buff *skb) 2474 { 2475 return __dev_forward_skb2(dev, skb, false) ?: netif_rx_internal(skb); 2476 } 2477 2478 static int deliver_skb(struct sk_buff *skb, 2479 struct packet_type *pt_prev, 2480 struct net_device *orig_dev) 2481 { 2482 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 2483 return -ENOMEM; 2484 refcount_inc(&skb->users); 2485 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 2486 } 2487 2488 static inline void deliver_ptype_list_skb(struct sk_buff *skb, 2489 struct packet_type **pt, 2490 struct net_device *orig_dev, 2491 __be16 type, 2492 struct list_head *ptype_list) 2493 { 2494 struct packet_type *ptype, *pt_prev = *pt; 2495 2496 list_for_each_entry_rcu(ptype, ptype_list, list) { 2497 if (ptype->type != type) 2498 continue; 2499 if (unlikely(pt_prev)) 2500 deliver_skb(skb, pt_prev, orig_dev); 2501 pt_prev = ptype; 2502 } 2503 *pt = pt_prev; 2504 } 2505 2506 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb) 2507 { 2508 if (!ptype->af_packet_priv || !skb->sk) 2509 return false; 2510 2511 if (ptype->id_match) 2512 return ptype->id_match(ptype, skb->sk); 2513 else if ((struct sock *)ptype->af_packet_priv == skb->sk) 2514 return true; 2515 2516 return false; 2517 } 2518 2519 /** 2520 * dev_nit_active_rcu - return true if any network interface taps are in use 2521 * 2522 * The caller must hold the RCU lock 2523 * 2524 * @dev: network device to check for the presence of taps 2525 */ 2526 bool dev_nit_active_rcu(const struct net_device *dev) 2527 { 2528 /* Callers may hold either RCU or RCU BH lock */ 2529 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 2530 2531 return !list_empty(&dev_net(dev)->ptype_all) || 2532 !list_empty(&dev->ptype_all); 2533 } 2534 EXPORT_SYMBOL_GPL(dev_nit_active_rcu); 2535 2536 /* 2537 * Support routine. Sends outgoing frames to any network 2538 * taps currently in use. 2539 */ 2540 2541 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev) 2542 { 2543 struct packet_type *ptype, *pt_prev = NULL; 2544 struct list_head *ptype_list; 2545 struct sk_buff *skb2 = NULL; 2546 2547 rcu_read_lock(); 2548 ptype_list = &dev_net_rcu(dev)->ptype_all; 2549 again: 2550 list_for_each_entry_rcu(ptype, ptype_list, list) { 2551 if (READ_ONCE(ptype->ignore_outgoing)) 2552 continue; 2553 2554 /* Never send packets back to the socket 2555 * they originated from - MvS (miquels@drinkel.ow.org) 2556 */ 2557 if (skb_loop_sk(ptype, skb)) 2558 continue; 2559 2560 if (unlikely(pt_prev)) { 2561 deliver_skb(skb2, pt_prev, skb->dev); 2562 pt_prev = ptype; 2563 continue; 2564 } 2565 2566 /* need to clone skb, done only once */ 2567 skb2 = skb_clone(skb, GFP_ATOMIC); 2568 if (!skb2) 2569 goto out_unlock; 2570 2571 net_timestamp_set(skb2); 2572 2573 /* skb->nh should be correctly 2574 * set by sender, so that the second statement is 2575 * just protection against buggy protocols. 2576 */ 2577 skb_reset_mac_header(skb2); 2578 2579 if (skb_network_header(skb2) < skb2->data || 2580 skb_network_header(skb2) > skb_tail_pointer(skb2)) { 2581 net_crit_ratelimited("protocol %04x is buggy, dev %s\n", 2582 ntohs(skb2->protocol), 2583 dev->name); 2584 skb_reset_network_header(skb2); 2585 } 2586 2587 skb2->transport_header = skb2->network_header; 2588 skb2->pkt_type = PACKET_OUTGOING; 2589 pt_prev = ptype; 2590 } 2591 2592 if (ptype_list != &dev->ptype_all) { 2593 ptype_list = &dev->ptype_all; 2594 goto again; 2595 } 2596 out_unlock: 2597 if (pt_prev) { 2598 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC)) 2599 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev); 2600 else 2601 kfree_skb(skb2); 2602 } 2603 rcu_read_unlock(); 2604 } 2605 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit); 2606 2607 /** 2608 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change 2609 * @dev: Network device 2610 * @txq: number of queues available 2611 * 2612 * If real_num_tx_queues is changed the tc mappings may no longer be 2613 * valid. To resolve this verify the tc mapping remains valid and if 2614 * not NULL the mapping. With no priorities mapping to this 2615 * offset/count pair it will no longer be used. In the worst case TC0 2616 * is invalid nothing can be done so disable priority mappings. If is 2617 * expected that drivers will fix this mapping if they can before 2618 * calling netif_set_real_num_tx_queues. 2619 */ 2620 static void netif_setup_tc(struct net_device *dev, unsigned int txq) 2621 { 2622 int i; 2623 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2624 2625 /* If TC0 is invalidated disable TC mapping */ 2626 if (tc->offset + tc->count > txq) { 2627 netdev_warn(dev, "Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n"); 2628 dev->num_tc = 0; 2629 return; 2630 } 2631 2632 /* Invalidated prio to tc mappings set to TC0 */ 2633 for (i = 1; i < TC_BITMASK + 1; i++) { 2634 int q = netdev_get_prio_tc_map(dev, i); 2635 2636 tc = &dev->tc_to_txq[q]; 2637 if (tc->offset + tc->count > txq) { 2638 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", 2639 i, q); 2640 netdev_set_prio_tc_map(dev, i, 0); 2641 } 2642 } 2643 } 2644 2645 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq) 2646 { 2647 if (dev->num_tc) { 2648 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2649 int i; 2650 2651 /* walk through the TCs and see if it falls into any of them */ 2652 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) { 2653 if ((txq - tc->offset) < tc->count) 2654 return i; 2655 } 2656 2657 /* didn't find it, just return -1 to indicate no match */ 2658 return -1; 2659 } 2660 2661 return 0; 2662 } 2663 EXPORT_SYMBOL(netdev_txq_to_tc); 2664 2665 #ifdef CONFIG_XPS 2666 static struct static_key xps_needed __read_mostly; 2667 static struct static_key xps_rxqs_needed __read_mostly; 2668 static DEFINE_MUTEX(xps_map_mutex); 2669 #define xmap_dereference(P) \ 2670 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex)) 2671 2672 static bool remove_xps_queue(struct xps_dev_maps *dev_maps, 2673 struct xps_dev_maps *old_maps, int tci, u16 index) 2674 { 2675 struct xps_map *map = NULL; 2676 int pos; 2677 2678 map = xmap_dereference(dev_maps->attr_map[tci]); 2679 if (!map) 2680 return false; 2681 2682 for (pos = map->len; pos--;) { 2683 if (map->queues[pos] != index) 2684 continue; 2685 2686 if (map->len > 1) { 2687 map->queues[pos] = map->queues[--map->len]; 2688 break; 2689 } 2690 2691 if (old_maps) 2692 RCU_INIT_POINTER(old_maps->attr_map[tci], NULL); 2693 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2694 kfree_rcu(map, rcu); 2695 return false; 2696 } 2697 2698 return true; 2699 } 2700 2701 static bool remove_xps_queue_cpu(struct net_device *dev, 2702 struct xps_dev_maps *dev_maps, 2703 int cpu, u16 offset, u16 count) 2704 { 2705 int num_tc = dev_maps->num_tc; 2706 bool active = false; 2707 int tci; 2708 2709 for (tci = cpu * num_tc; num_tc--; tci++) { 2710 int i, j; 2711 2712 for (i = count, j = offset; i--; j++) { 2713 if (!remove_xps_queue(dev_maps, NULL, tci, j)) 2714 break; 2715 } 2716 2717 active |= i < 0; 2718 } 2719 2720 return active; 2721 } 2722 2723 static void reset_xps_maps(struct net_device *dev, 2724 struct xps_dev_maps *dev_maps, 2725 enum xps_map_type type) 2726 { 2727 static_key_slow_dec_cpuslocked(&xps_needed); 2728 if (type == XPS_RXQS) 2729 static_key_slow_dec_cpuslocked(&xps_rxqs_needed); 2730 2731 RCU_INIT_POINTER(dev->xps_maps[type], NULL); 2732 2733 kfree_rcu(dev_maps, rcu); 2734 } 2735 2736 static void clean_xps_maps(struct net_device *dev, enum xps_map_type type, 2737 u16 offset, u16 count) 2738 { 2739 struct xps_dev_maps *dev_maps; 2740 bool active = false; 2741 int i, j; 2742 2743 dev_maps = xmap_dereference(dev->xps_maps[type]); 2744 if (!dev_maps) 2745 return; 2746 2747 for (j = 0; j < dev_maps->nr_ids; j++) 2748 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, count); 2749 if (!active) 2750 reset_xps_maps(dev, dev_maps, type); 2751 2752 if (type == XPS_CPUS) { 2753 for (i = offset + (count - 1); count--; i--) 2754 netdev_queue_numa_node_write( 2755 netdev_get_tx_queue(dev, i), NUMA_NO_NODE); 2756 } 2757 } 2758 2759 static void netif_reset_xps_queues(struct net_device *dev, u16 offset, 2760 u16 count) 2761 { 2762 if (!static_key_false(&xps_needed)) 2763 return; 2764 2765 cpus_read_lock(); 2766 mutex_lock(&xps_map_mutex); 2767 2768 if (static_key_false(&xps_rxqs_needed)) 2769 clean_xps_maps(dev, XPS_RXQS, offset, count); 2770 2771 clean_xps_maps(dev, XPS_CPUS, offset, count); 2772 2773 mutex_unlock(&xps_map_mutex); 2774 cpus_read_unlock(); 2775 } 2776 2777 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index) 2778 { 2779 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index); 2780 } 2781 2782 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index, 2783 u16 index, bool is_rxqs_map) 2784 { 2785 struct xps_map *new_map; 2786 int alloc_len = XPS_MIN_MAP_ALLOC; 2787 int i, pos; 2788 2789 for (pos = 0; map && pos < map->len; pos++) { 2790 if (map->queues[pos] != index) 2791 continue; 2792 return map; 2793 } 2794 2795 /* Need to add tx-queue to this CPU's/rx-queue's existing map */ 2796 if (map) { 2797 if (pos < map->alloc_len) 2798 return map; 2799 2800 alloc_len = map->alloc_len * 2; 2801 } 2802 2803 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's 2804 * map 2805 */ 2806 if (is_rxqs_map) 2807 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL); 2808 else 2809 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL, 2810 cpu_to_node(attr_index)); 2811 if (!new_map) 2812 return NULL; 2813 2814 for (i = 0; i < pos; i++) 2815 new_map->queues[i] = map->queues[i]; 2816 new_map->alloc_len = alloc_len; 2817 new_map->len = pos; 2818 2819 return new_map; 2820 } 2821 2822 /* Copy xps maps at a given index */ 2823 static void xps_copy_dev_maps(struct xps_dev_maps *dev_maps, 2824 struct xps_dev_maps *new_dev_maps, int index, 2825 int tc, bool skip_tc) 2826 { 2827 int i, tci = index * dev_maps->num_tc; 2828 struct xps_map *map; 2829 2830 /* copy maps belonging to foreign traffic classes */ 2831 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2832 if (i == tc && skip_tc) 2833 continue; 2834 2835 /* fill in the new device map from the old device map */ 2836 map = xmap_dereference(dev_maps->attr_map[tci]); 2837 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2838 } 2839 } 2840 2841 /* Must be called under cpus_read_lock */ 2842 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask, 2843 u16 index, enum xps_map_type type) 2844 { 2845 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL, *old_dev_maps = NULL; 2846 const unsigned long *online_mask = NULL; 2847 bool active = false, copy = false; 2848 int i, j, tci, numa_node_id = -2; 2849 int maps_sz, num_tc = 1, tc = 0; 2850 struct xps_map *map, *new_map; 2851 unsigned int nr_ids; 2852 2853 WARN_ON_ONCE(index >= dev->num_tx_queues); 2854 2855 if (dev->num_tc) { 2856 /* Do not allow XPS on subordinate device directly */ 2857 num_tc = dev->num_tc; 2858 if (num_tc < 0) 2859 return -EINVAL; 2860 2861 /* If queue belongs to subordinate dev use its map */ 2862 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev; 2863 2864 tc = netdev_txq_to_tc(dev, index); 2865 if (tc < 0) 2866 return -EINVAL; 2867 } 2868 2869 mutex_lock(&xps_map_mutex); 2870 2871 dev_maps = xmap_dereference(dev->xps_maps[type]); 2872 if (type == XPS_RXQS) { 2873 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues); 2874 nr_ids = dev->num_rx_queues; 2875 } else { 2876 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc); 2877 if (num_possible_cpus() > 1) 2878 online_mask = cpumask_bits(cpu_online_mask); 2879 nr_ids = nr_cpu_ids; 2880 } 2881 2882 if (maps_sz < L1_CACHE_BYTES) 2883 maps_sz = L1_CACHE_BYTES; 2884 2885 /* The old dev_maps could be larger or smaller than the one we're 2886 * setting up now, as dev->num_tc or nr_ids could have been updated in 2887 * between. We could try to be smart, but let's be safe instead and only 2888 * copy foreign traffic classes if the two map sizes match. 2889 */ 2890 if (dev_maps && 2891 dev_maps->num_tc == num_tc && dev_maps->nr_ids == nr_ids) 2892 copy = true; 2893 2894 /* allocate memory for queue storage */ 2895 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids), 2896 j < nr_ids;) { 2897 if (!new_dev_maps) { 2898 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL); 2899 if (!new_dev_maps) { 2900 mutex_unlock(&xps_map_mutex); 2901 return -ENOMEM; 2902 } 2903 2904 new_dev_maps->nr_ids = nr_ids; 2905 new_dev_maps->num_tc = num_tc; 2906 } 2907 2908 tci = j * num_tc + tc; 2909 map = copy ? xmap_dereference(dev_maps->attr_map[tci]) : NULL; 2910 2911 map = expand_xps_map(map, j, index, type == XPS_RXQS); 2912 if (!map) 2913 goto error; 2914 2915 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2916 } 2917 2918 if (!new_dev_maps) 2919 goto out_no_new_maps; 2920 2921 if (!dev_maps) { 2922 /* Increment static keys at most once per type */ 2923 static_key_slow_inc_cpuslocked(&xps_needed); 2924 if (type == XPS_RXQS) 2925 static_key_slow_inc_cpuslocked(&xps_rxqs_needed); 2926 } 2927 2928 for (j = 0; j < nr_ids; j++) { 2929 bool skip_tc = false; 2930 2931 tci = j * num_tc + tc; 2932 if (netif_attr_test_mask(j, mask, nr_ids) && 2933 netif_attr_test_online(j, online_mask, nr_ids)) { 2934 /* add tx-queue to CPU/rx-queue maps */ 2935 int pos = 0; 2936 2937 skip_tc = true; 2938 2939 map = xmap_dereference(new_dev_maps->attr_map[tci]); 2940 while ((pos < map->len) && (map->queues[pos] != index)) 2941 pos++; 2942 2943 if (pos == map->len) 2944 map->queues[map->len++] = index; 2945 #ifdef CONFIG_NUMA 2946 if (type == XPS_CPUS) { 2947 if (numa_node_id == -2) 2948 numa_node_id = cpu_to_node(j); 2949 else if (numa_node_id != cpu_to_node(j)) 2950 numa_node_id = -1; 2951 } 2952 #endif 2953 } 2954 2955 if (copy) 2956 xps_copy_dev_maps(dev_maps, new_dev_maps, j, tc, 2957 skip_tc); 2958 } 2959 2960 rcu_assign_pointer(dev->xps_maps[type], new_dev_maps); 2961 2962 /* Cleanup old maps */ 2963 if (!dev_maps) 2964 goto out_no_old_maps; 2965 2966 for (j = 0; j < dev_maps->nr_ids; j++) { 2967 for (i = num_tc, tci = j * dev_maps->num_tc; i--; tci++) { 2968 map = xmap_dereference(dev_maps->attr_map[tci]); 2969 if (!map) 2970 continue; 2971 2972 if (copy) { 2973 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2974 if (map == new_map) 2975 continue; 2976 } 2977 2978 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2979 kfree_rcu(map, rcu); 2980 } 2981 } 2982 2983 old_dev_maps = dev_maps; 2984 2985 out_no_old_maps: 2986 dev_maps = new_dev_maps; 2987 active = true; 2988 2989 out_no_new_maps: 2990 if (type == XPS_CPUS) 2991 /* update Tx queue numa node */ 2992 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index), 2993 (numa_node_id >= 0) ? 2994 numa_node_id : NUMA_NO_NODE); 2995 2996 if (!dev_maps) 2997 goto out_no_maps; 2998 2999 /* removes tx-queue from unused CPUs/rx-queues */ 3000 for (j = 0; j < dev_maps->nr_ids; j++) { 3001 tci = j * dev_maps->num_tc; 3002 3003 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 3004 if (i == tc && 3005 netif_attr_test_mask(j, mask, dev_maps->nr_ids) && 3006 netif_attr_test_online(j, online_mask, dev_maps->nr_ids)) 3007 continue; 3008 3009 active |= remove_xps_queue(dev_maps, 3010 copy ? old_dev_maps : NULL, 3011 tci, index); 3012 } 3013 } 3014 3015 if (old_dev_maps) 3016 kfree_rcu(old_dev_maps, rcu); 3017 3018 /* free map if not active */ 3019 if (!active) 3020 reset_xps_maps(dev, dev_maps, type); 3021 3022 out_no_maps: 3023 mutex_unlock(&xps_map_mutex); 3024 3025 return 0; 3026 error: 3027 /* remove any maps that we added */ 3028 for (j = 0; j < nr_ids; j++) { 3029 for (i = num_tc, tci = j * num_tc; i--; tci++) { 3030 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 3031 map = copy ? 3032 xmap_dereference(dev_maps->attr_map[tci]) : 3033 NULL; 3034 if (new_map && new_map != map) 3035 kfree(new_map); 3036 } 3037 } 3038 3039 mutex_unlock(&xps_map_mutex); 3040 3041 kfree(new_dev_maps); 3042 return -ENOMEM; 3043 } 3044 EXPORT_SYMBOL_GPL(__netif_set_xps_queue); 3045 3046 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, 3047 u16 index) 3048 { 3049 int ret; 3050 3051 cpus_read_lock(); 3052 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, XPS_CPUS); 3053 cpus_read_unlock(); 3054 3055 return ret; 3056 } 3057 EXPORT_SYMBOL(netif_set_xps_queue); 3058 3059 #endif 3060 static void netdev_unbind_all_sb_channels(struct net_device *dev) 3061 { 3062 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 3063 3064 /* Unbind any subordinate channels */ 3065 while (txq-- != &dev->_tx[0]) { 3066 if (txq->sb_dev) 3067 netdev_unbind_sb_channel(dev, txq->sb_dev); 3068 } 3069 } 3070 3071 void netdev_reset_tc(struct net_device *dev) 3072 { 3073 #ifdef CONFIG_XPS 3074 netif_reset_xps_queues_gt(dev, 0); 3075 #endif 3076 netdev_unbind_all_sb_channels(dev); 3077 3078 /* Reset TC configuration of device */ 3079 dev->num_tc = 0; 3080 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq)); 3081 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map)); 3082 } 3083 EXPORT_SYMBOL(netdev_reset_tc); 3084 3085 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset) 3086 { 3087 if (tc >= dev->num_tc) 3088 return -EINVAL; 3089 3090 #ifdef CONFIG_XPS 3091 netif_reset_xps_queues(dev, offset, count); 3092 #endif 3093 dev->tc_to_txq[tc].count = count; 3094 dev->tc_to_txq[tc].offset = offset; 3095 return 0; 3096 } 3097 EXPORT_SYMBOL(netdev_set_tc_queue); 3098 3099 int netdev_set_num_tc(struct net_device *dev, u8 num_tc) 3100 { 3101 if (num_tc > TC_MAX_QUEUE) 3102 return -EINVAL; 3103 3104 #ifdef CONFIG_XPS 3105 netif_reset_xps_queues_gt(dev, 0); 3106 #endif 3107 netdev_unbind_all_sb_channels(dev); 3108 3109 dev->num_tc = num_tc; 3110 return 0; 3111 } 3112 EXPORT_SYMBOL(netdev_set_num_tc); 3113 3114 void netdev_unbind_sb_channel(struct net_device *dev, 3115 struct net_device *sb_dev) 3116 { 3117 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 3118 3119 #ifdef CONFIG_XPS 3120 netif_reset_xps_queues_gt(sb_dev, 0); 3121 #endif 3122 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq)); 3123 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map)); 3124 3125 while (txq-- != &dev->_tx[0]) { 3126 if (txq->sb_dev == sb_dev) 3127 txq->sb_dev = NULL; 3128 } 3129 } 3130 EXPORT_SYMBOL(netdev_unbind_sb_channel); 3131 3132 int netdev_bind_sb_channel_queue(struct net_device *dev, 3133 struct net_device *sb_dev, 3134 u8 tc, u16 count, u16 offset) 3135 { 3136 /* Make certain the sb_dev and dev are already configured */ 3137 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc) 3138 return -EINVAL; 3139 3140 /* We cannot hand out queues we don't have */ 3141 if ((offset + count) > dev->real_num_tx_queues) 3142 return -EINVAL; 3143 3144 /* Record the mapping */ 3145 sb_dev->tc_to_txq[tc].count = count; 3146 sb_dev->tc_to_txq[tc].offset = offset; 3147 3148 /* Provide a way for Tx queue to find the tc_to_txq map or 3149 * XPS map for itself. 3150 */ 3151 while (count--) 3152 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev; 3153 3154 return 0; 3155 } 3156 EXPORT_SYMBOL(netdev_bind_sb_channel_queue); 3157 3158 int netdev_set_sb_channel(struct net_device *dev, u16 channel) 3159 { 3160 /* Do not use a multiqueue device to represent a subordinate channel */ 3161 if (netif_is_multiqueue(dev)) 3162 return -ENODEV; 3163 3164 /* We allow channels 1 - 32767 to be used for subordinate channels. 3165 * Channel 0 is meant to be "native" mode and used only to represent 3166 * the main root device. We allow writing 0 to reset the device back 3167 * to normal mode after being used as a subordinate channel. 3168 */ 3169 if (channel > S16_MAX) 3170 return -EINVAL; 3171 3172 dev->num_tc = -channel; 3173 3174 return 0; 3175 } 3176 EXPORT_SYMBOL(netdev_set_sb_channel); 3177 3178 /* 3179 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues 3180 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed. 3181 */ 3182 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq) 3183 { 3184 bool disabling; 3185 int rc; 3186 3187 disabling = txq < dev->real_num_tx_queues; 3188 3189 if (txq < 1 || txq > dev->num_tx_queues) 3190 return -EINVAL; 3191 3192 if (dev->reg_state == NETREG_REGISTERED || 3193 dev->reg_state == NETREG_UNREGISTERING) { 3194 netdev_ops_assert_locked(dev); 3195 3196 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues, 3197 txq); 3198 if (rc) 3199 return rc; 3200 3201 if (dev->num_tc) 3202 netif_setup_tc(dev, txq); 3203 3204 net_shaper_set_real_num_tx_queues(dev, txq); 3205 3206 dev_qdisc_change_real_num_tx(dev, txq); 3207 3208 dev->real_num_tx_queues = txq; 3209 3210 if (disabling) { 3211 synchronize_net(); 3212 qdisc_reset_all_tx_gt(dev, txq); 3213 #ifdef CONFIG_XPS 3214 netif_reset_xps_queues_gt(dev, txq); 3215 #endif 3216 } 3217 } else { 3218 dev->real_num_tx_queues = txq; 3219 } 3220 3221 return 0; 3222 } 3223 EXPORT_SYMBOL(netif_set_real_num_tx_queues); 3224 3225 /** 3226 * netif_set_real_num_rx_queues - set actual number of RX queues used 3227 * @dev: Network device 3228 * @rxq: Actual number of RX queues 3229 * 3230 * This must be called either with the rtnl_lock held or before 3231 * registration of the net device. Returns 0 on success, or a 3232 * negative error code. If called before registration, it always 3233 * succeeds. 3234 */ 3235 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq) 3236 { 3237 int rc; 3238 3239 if (rxq < 1 || rxq > dev->num_rx_queues) 3240 return -EINVAL; 3241 3242 if (dev->reg_state == NETREG_REGISTERED) { 3243 netdev_ops_assert_locked(dev); 3244 3245 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues, 3246 rxq); 3247 if (rc) 3248 return rc; 3249 } 3250 3251 dev->real_num_rx_queues = rxq; 3252 return 0; 3253 } 3254 EXPORT_SYMBOL(netif_set_real_num_rx_queues); 3255 3256 /** 3257 * netif_set_real_num_queues - set actual number of RX and TX queues used 3258 * @dev: Network device 3259 * @txq: Actual number of TX queues 3260 * @rxq: Actual number of RX queues 3261 * 3262 * Set the real number of both TX and RX queues. 3263 * Does nothing if the number of queues is already correct. 3264 */ 3265 int netif_set_real_num_queues(struct net_device *dev, 3266 unsigned int txq, unsigned int rxq) 3267 { 3268 unsigned int old_rxq = dev->real_num_rx_queues; 3269 int err; 3270 3271 if (txq < 1 || txq > dev->num_tx_queues || 3272 rxq < 1 || rxq > dev->num_rx_queues) 3273 return -EINVAL; 3274 3275 /* Start from increases, so the error path only does decreases - 3276 * decreases can't fail. 3277 */ 3278 if (rxq > dev->real_num_rx_queues) { 3279 err = netif_set_real_num_rx_queues(dev, rxq); 3280 if (err) 3281 return err; 3282 } 3283 if (txq > dev->real_num_tx_queues) { 3284 err = netif_set_real_num_tx_queues(dev, txq); 3285 if (err) 3286 goto undo_rx; 3287 } 3288 if (rxq < dev->real_num_rx_queues) 3289 WARN_ON(netif_set_real_num_rx_queues(dev, rxq)); 3290 if (txq < dev->real_num_tx_queues) 3291 WARN_ON(netif_set_real_num_tx_queues(dev, txq)); 3292 3293 return 0; 3294 undo_rx: 3295 WARN_ON(netif_set_real_num_rx_queues(dev, old_rxq)); 3296 return err; 3297 } 3298 EXPORT_SYMBOL(netif_set_real_num_queues); 3299 3300 /** 3301 * netif_set_tso_max_size() - set the max size of TSO frames supported 3302 * @dev: netdev to update 3303 * @size: max skb->len of a TSO frame 3304 * 3305 * Set the limit on the size of TSO super-frames the device can handle. 3306 * Unless explicitly set the stack will assume the value of 3307 * %GSO_LEGACY_MAX_SIZE. 3308 */ 3309 void netif_set_tso_max_size(struct net_device *dev, unsigned int size) 3310 { 3311 dev->tso_max_size = min(GSO_MAX_SIZE, size); 3312 if (size < READ_ONCE(dev->gso_max_size)) 3313 netif_set_gso_max_size(dev, size); 3314 if (size < READ_ONCE(dev->gso_ipv4_max_size)) 3315 netif_set_gso_ipv4_max_size(dev, size); 3316 } 3317 EXPORT_SYMBOL(netif_set_tso_max_size); 3318 3319 /** 3320 * netif_set_tso_max_segs() - set the max number of segs supported for TSO 3321 * @dev: netdev to update 3322 * @segs: max number of TCP segments 3323 * 3324 * Set the limit on the number of TCP segments the device can generate from 3325 * a single TSO super-frame. 3326 * Unless explicitly set the stack will assume the value of %GSO_MAX_SEGS. 3327 */ 3328 void netif_set_tso_max_segs(struct net_device *dev, unsigned int segs) 3329 { 3330 dev->tso_max_segs = segs; 3331 if (segs < READ_ONCE(dev->gso_max_segs)) 3332 netif_set_gso_max_segs(dev, segs); 3333 } 3334 EXPORT_SYMBOL(netif_set_tso_max_segs); 3335 3336 /** 3337 * netif_inherit_tso_max() - copy all TSO limits from a lower device to an upper 3338 * @to: netdev to update 3339 * @from: netdev from which to copy the limits 3340 */ 3341 void netif_inherit_tso_max(struct net_device *to, const struct net_device *from) 3342 { 3343 netif_set_tso_max_size(to, from->tso_max_size); 3344 netif_set_tso_max_segs(to, from->tso_max_segs); 3345 } 3346 EXPORT_SYMBOL(netif_inherit_tso_max); 3347 3348 /** 3349 * netif_get_num_default_rss_queues - default number of RSS queues 3350 * 3351 * Default value is the number of physical cores if there are only 1 or 2, or 3352 * divided by 2 if there are more. 3353 */ 3354 int netif_get_num_default_rss_queues(void) 3355 { 3356 cpumask_var_t cpus; 3357 int cpu, count = 0; 3358 3359 if (unlikely(is_kdump_kernel() || !zalloc_cpumask_var(&cpus, GFP_KERNEL))) 3360 return 1; 3361 3362 cpumask_copy(cpus, cpu_online_mask); 3363 for_each_cpu(cpu, cpus) { 3364 ++count; 3365 cpumask_andnot(cpus, cpus, topology_sibling_cpumask(cpu)); 3366 } 3367 free_cpumask_var(cpus); 3368 3369 return count > 2 ? DIV_ROUND_UP(count, 2) : count; 3370 } 3371 EXPORT_SYMBOL(netif_get_num_default_rss_queues); 3372 3373 static void __netif_reschedule(struct Qdisc *q) 3374 { 3375 struct softnet_data *sd; 3376 unsigned long flags; 3377 3378 local_irq_save(flags); 3379 sd = this_cpu_ptr(&softnet_data); 3380 q->next_sched = NULL; 3381 *sd->output_queue_tailp = q; 3382 sd->output_queue_tailp = &q->next_sched; 3383 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3384 local_irq_restore(flags); 3385 } 3386 3387 void __netif_schedule(struct Qdisc *q) 3388 { 3389 /* If q->defer_list is not empty, at least one thread is 3390 * in __dev_xmit_skb() before llist_del_all(&q->defer_list). 3391 * This thread will attempt to run the queue. 3392 */ 3393 if (!llist_empty(&q->defer_list)) 3394 return; 3395 3396 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state)) 3397 __netif_reschedule(q); 3398 } 3399 EXPORT_SYMBOL(__netif_schedule); 3400 3401 struct dev_kfree_skb_cb { 3402 enum skb_drop_reason reason; 3403 }; 3404 3405 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb) 3406 { 3407 return (struct dev_kfree_skb_cb *)skb->cb; 3408 } 3409 3410 void netif_schedule_queue(struct netdev_queue *txq) 3411 { 3412 rcu_read_lock(); 3413 if (!netif_xmit_stopped(txq)) { 3414 struct Qdisc *q = rcu_dereference(txq->qdisc); 3415 3416 __netif_schedule(q); 3417 } 3418 rcu_read_unlock(); 3419 } 3420 EXPORT_SYMBOL(netif_schedule_queue); 3421 3422 void netif_tx_wake_queue(struct netdev_queue *dev_queue) 3423 { 3424 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) { 3425 struct Qdisc *q; 3426 3427 rcu_read_lock(); 3428 q = rcu_dereference(dev_queue->qdisc); 3429 __netif_schedule(q); 3430 rcu_read_unlock(); 3431 } 3432 } 3433 EXPORT_SYMBOL(netif_tx_wake_queue); 3434 3435 void dev_kfree_skb_irq_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3436 { 3437 unsigned long flags; 3438 3439 if (unlikely(!skb)) 3440 return; 3441 3442 if (likely(refcount_read(&skb->users) == 1)) { 3443 smp_rmb(); 3444 refcount_set(&skb->users, 0); 3445 } else if (likely(!refcount_dec_and_test(&skb->users))) { 3446 return; 3447 } 3448 get_kfree_skb_cb(skb)->reason = reason; 3449 local_irq_save(flags); 3450 skb->next = __this_cpu_read(softnet_data.completion_queue); 3451 __this_cpu_write(softnet_data.completion_queue, skb); 3452 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3453 local_irq_restore(flags); 3454 } 3455 EXPORT_SYMBOL(dev_kfree_skb_irq_reason); 3456 3457 void dev_kfree_skb_any_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3458 { 3459 if (in_hardirq() || irqs_disabled()) 3460 dev_kfree_skb_irq_reason(skb, reason); 3461 else 3462 kfree_skb_reason(skb, reason); 3463 } 3464 EXPORT_SYMBOL(dev_kfree_skb_any_reason); 3465 3466 3467 /** 3468 * netif_device_detach - mark device as removed 3469 * @dev: network device 3470 * 3471 * Mark device as removed from system and therefore no longer available. 3472 */ 3473 void netif_device_detach(struct net_device *dev) 3474 { 3475 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) && 3476 netif_running(dev)) { 3477 netif_tx_stop_all_queues(dev); 3478 } 3479 } 3480 EXPORT_SYMBOL(netif_device_detach); 3481 3482 /** 3483 * netif_device_attach - mark device as attached 3484 * @dev: network device 3485 * 3486 * Mark device as attached from system and restart if needed. 3487 */ 3488 void netif_device_attach(struct net_device *dev) 3489 { 3490 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) && 3491 netif_running(dev)) { 3492 netif_tx_wake_all_queues(dev); 3493 netdev_watchdog_up(dev); 3494 } 3495 } 3496 EXPORT_SYMBOL(netif_device_attach); 3497 3498 /* 3499 * Returns a Tx hash based on the given packet descriptor a Tx queues' number 3500 * to be used as a distribution range. 3501 */ 3502 static u16 skb_tx_hash(const struct net_device *dev, 3503 const struct net_device *sb_dev, 3504 struct sk_buff *skb) 3505 { 3506 u32 hash; 3507 u16 qoffset = 0; 3508 u16 qcount = dev->real_num_tx_queues; 3509 3510 if (dev->num_tc) { 3511 u8 tc = netdev_get_prio_tc_map(dev, skb->priority); 3512 3513 qoffset = sb_dev->tc_to_txq[tc].offset; 3514 qcount = sb_dev->tc_to_txq[tc].count; 3515 if (unlikely(!qcount)) { 3516 net_warn_ratelimited("%s: invalid qcount, qoffset %u for tc %u\n", 3517 sb_dev->name, qoffset, tc); 3518 qoffset = 0; 3519 qcount = dev->real_num_tx_queues; 3520 } 3521 } 3522 3523 if (skb_rx_queue_recorded(skb)) { 3524 DEBUG_NET_WARN_ON_ONCE(qcount == 0); 3525 hash = skb_get_rx_queue(skb); 3526 if (hash >= qoffset) 3527 hash -= qoffset; 3528 while (unlikely(hash >= qcount)) 3529 hash -= qcount; 3530 return hash + qoffset; 3531 } 3532 3533 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset; 3534 } 3535 3536 void skb_warn_bad_offload(const struct sk_buff *skb) 3537 { 3538 static const netdev_features_t null_features; 3539 struct net_device *dev = skb->dev; 3540 const char *name = ""; 3541 3542 if (!net_ratelimit()) 3543 return; 3544 3545 if (dev) { 3546 if (dev->dev.parent) 3547 name = dev_driver_string(dev->dev.parent); 3548 else 3549 name = netdev_name(dev); 3550 } 3551 skb_dump(KERN_WARNING, skb, false); 3552 WARN(1, "%s: caps=(%pNF, %pNF)\n", 3553 name, dev ? &dev->features : &null_features, 3554 skb->sk ? &skb->sk->sk_route_caps : &null_features); 3555 } 3556 3557 /* 3558 * Invalidate hardware checksum when packet is to be mangled, and 3559 * complete checksum manually on outgoing path. 3560 */ 3561 int skb_checksum_help(struct sk_buff *skb) 3562 { 3563 __wsum csum; 3564 int ret = 0, offset; 3565 3566 if (skb->ip_summed == CHECKSUM_COMPLETE) 3567 goto out_set_summed; 3568 3569 if (unlikely(skb_is_gso(skb))) { 3570 skb_warn_bad_offload(skb); 3571 return -EINVAL; 3572 } 3573 3574 if (!skb_frags_readable(skb)) { 3575 return -EFAULT; 3576 } 3577 3578 /* Before computing a checksum, we should make sure no frag could 3579 * be modified by an external entity : checksum could be wrong. 3580 */ 3581 if (skb_has_shared_frag(skb)) { 3582 ret = __skb_linearize(skb); 3583 if (ret) 3584 goto out; 3585 } 3586 3587 offset = skb_checksum_start_offset(skb); 3588 ret = -EINVAL; 3589 if (unlikely(offset >= skb_headlen(skb))) { 3590 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3591 WARN_ONCE(true, "offset (%d) >= skb_headlen() (%u)\n", 3592 offset, skb_headlen(skb)); 3593 goto out; 3594 } 3595 csum = skb_checksum(skb, offset, skb->len - offset, 0); 3596 3597 offset += skb->csum_offset; 3598 if (unlikely(offset + sizeof(__sum16) > skb_headlen(skb))) { 3599 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3600 WARN_ONCE(true, "offset+2 (%zu) > skb_headlen() (%u)\n", 3601 offset + sizeof(__sum16), skb_headlen(skb)); 3602 goto out; 3603 } 3604 ret = skb_ensure_writable(skb, offset + sizeof(__sum16)); 3605 if (ret) 3606 goto out; 3607 3608 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0; 3609 out_set_summed: 3610 skb->ip_summed = CHECKSUM_NONE; 3611 out: 3612 return ret; 3613 } 3614 EXPORT_SYMBOL(skb_checksum_help); 3615 3616 #ifdef CONFIG_NET_CRC32C 3617 int skb_crc32c_csum_help(struct sk_buff *skb) 3618 { 3619 u32 crc; 3620 int ret = 0, offset, start; 3621 3622 if (skb->ip_summed != CHECKSUM_PARTIAL) 3623 goto out; 3624 3625 if (unlikely(skb_is_gso(skb))) 3626 goto out; 3627 3628 /* Before computing a checksum, we should make sure no frag could 3629 * be modified by an external entity : checksum could be wrong. 3630 */ 3631 if (unlikely(skb_has_shared_frag(skb))) { 3632 ret = __skb_linearize(skb); 3633 if (ret) 3634 goto out; 3635 } 3636 start = skb_checksum_start_offset(skb); 3637 offset = start + offsetof(struct sctphdr, checksum); 3638 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) { 3639 ret = -EINVAL; 3640 goto out; 3641 } 3642 3643 ret = skb_ensure_writable(skb, offset + sizeof(__le32)); 3644 if (ret) 3645 goto out; 3646 3647 crc = ~skb_crc32c(skb, start, skb->len - start, ~0); 3648 *(__le32 *)(skb->data + offset) = cpu_to_le32(crc); 3649 skb_reset_csum_not_inet(skb); 3650 out: 3651 return ret; 3652 } 3653 EXPORT_SYMBOL(skb_crc32c_csum_help); 3654 #endif /* CONFIG_NET_CRC32C */ 3655 3656 __be16 skb_network_protocol(struct sk_buff *skb, int *depth) 3657 { 3658 __be16 type = skb->protocol; 3659 3660 /* Tunnel gso handlers can set protocol to ethernet. */ 3661 if (type == htons(ETH_P_TEB)) { 3662 struct ethhdr *eth; 3663 3664 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr)))) 3665 return 0; 3666 3667 eth = (struct ethhdr *)skb->data; 3668 type = eth->h_proto; 3669 } 3670 3671 return vlan_get_protocol_and_depth(skb, type, depth); 3672 } 3673 3674 3675 /* Take action when hardware reception checksum errors are detected. */ 3676 #ifdef CONFIG_BUG 3677 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3678 { 3679 netdev_err(dev, "hw csum failure\n"); 3680 skb_dump(KERN_ERR, skb, true); 3681 dump_stack(); 3682 } 3683 3684 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3685 { 3686 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb); 3687 } 3688 EXPORT_SYMBOL(netdev_rx_csum_fault); 3689 #endif 3690 3691 /* XXX: check that highmem exists at all on the given machine. */ 3692 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb) 3693 { 3694 #ifdef CONFIG_HIGHMEM 3695 int i; 3696 3697 if (!(dev->features & NETIF_F_HIGHDMA)) { 3698 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 3699 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 3700 struct page *page = skb_frag_page(frag); 3701 3702 if (page && PageHighMem(page)) 3703 return 1; 3704 } 3705 } 3706 #endif 3707 return 0; 3708 } 3709 3710 /* If MPLS offload request, verify we are testing hardware MPLS features 3711 * instead of standard features for the netdev. 3712 */ 3713 #if IS_ENABLED(CONFIG_NET_MPLS_GSO) 3714 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3715 netdev_features_t features, 3716 __be16 type) 3717 { 3718 if (eth_p_mpls(type)) 3719 features &= skb->dev->mpls_features; 3720 3721 return features; 3722 } 3723 #else 3724 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3725 netdev_features_t features, 3726 __be16 type) 3727 { 3728 return features; 3729 } 3730 #endif 3731 3732 static netdev_features_t harmonize_features(struct sk_buff *skb, 3733 netdev_features_t features) 3734 { 3735 __be16 type; 3736 3737 type = skb_network_protocol(skb, NULL); 3738 features = net_mpls_features(skb, features, type); 3739 3740 if (skb->ip_summed != CHECKSUM_NONE && 3741 !can_checksum_protocol(features, type)) { 3742 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3743 } 3744 if (illegal_highdma(skb->dev, skb)) 3745 features &= ~NETIF_F_SG; 3746 3747 return features; 3748 } 3749 3750 netdev_features_t passthru_features_check(struct sk_buff *skb, 3751 struct net_device *dev, 3752 netdev_features_t features) 3753 { 3754 return features; 3755 } 3756 EXPORT_SYMBOL(passthru_features_check); 3757 3758 static netdev_features_t dflt_features_check(struct sk_buff *skb, 3759 struct net_device *dev, 3760 netdev_features_t features) 3761 { 3762 return vlan_features_check(skb, features); 3763 } 3764 3765 static netdev_features_t gso_features_check(const struct sk_buff *skb, 3766 struct net_device *dev, 3767 netdev_features_t features) 3768 { 3769 u16 gso_segs = skb_shinfo(skb)->gso_segs; 3770 3771 if (gso_segs > READ_ONCE(dev->gso_max_segs)) 3772 return features & ~NETIF_F_GSO_MASK; 3773 3774 if (unlikely(skb->len >= netif_get_gso_max_size(dev, skb))) 3775 return features & ~NETIF_F_GSO_MASK; 3776 3777 if (!skb_shinfo(skb)->gso_type) { 3778 skb_warn_bad_offload(skb); 3779 return features & ~NETIF_F_GSO_MASK; 3780 } 3781 3782 /* Support for GSO partial features requires software 3783 * intervention before we can actually process the packets 3784 * so we need to strip support for any partial features now 3785 * and we can pull them back in after we have partially 3786 * segmented the frame. 3787 */ 3788 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)) 3789 features &= ~dev->gso_partial_features; 3790 3791 /* Make sure to clear the IPv4 ID mangling feature if the IPv4 header 3792 * has the potential to be fragmented so that TSO does not generate 3793 * segments with the same ID. For encapsulated packets, the ID mangling 3794 * feature is guaranteed not to use the same ID for the outer IPv4 3795 * headers of the generated segments if the headers have the potential 3796 * to be fragmented, so there is no need to clear the IPv4 ID mangling 3797 * feature (see the section about NETIF_F_TSO_MANGLEID in 3798 * segmentation-offloads.rst). 3799 */ 3800 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) { 3801 struct iphdr *iph = skb->encapsulation ? 3802 inner_ip_hdr(skb) : ip_hdr(skb); 3803 3804 if (!(iph->frag_off & htons(IP_DF))) 3805 features &= ~dev->mangleid_features; 3806 } 3807 3808 /* NETIF_F_IPV6_CSUM does not support IPv6 extension headers, 3809 * so neither does TSO that depends on it. 3810 */ 3811 if (features & NETIF_F_IPV6_CSUM && 3812 (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6 || 3813 (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4 && 3814 vlan_get_protocol(skb) == htons(ETH_P_IPV6))) && 3815 skb_transport_header_was_set(skb) && 3816 skb_network_header_len(skb) != sizeof(struct ipv6hdr)) 3817 features &= ~(NETIF_F_IPV6_CSUM | NETIF_F_TSO6 | NETIF_F_GSO_UDP_L4); 3818 3819 return features; 3820 } 3821 3822 netdev_features_t netif_skb_features(struct sk_buff *skb) 3823 { 3824 struct net_device *dev = skb->dev; 3825 netdev_features_t features = dev->features; 3826 3827 if (skb_is_gso(skb)) 3828 features = gso_features_check(skb, dev, features); 3829 3830 /* If encapsulation offload request, verify we are testing 3831 * hardware encapsulation features instead of standard 3832 * features for the netdev 3833 */ 3834 if (skb->encapsulation) 3835 features &= dev->hw_enc_features; 3836 3837 if (skb_vlan_tagged(skb)) 3838 features = netdev_intersect_features(features, 3839 dev->vlan_features | 3840 NETIF_F_HW_VLAN_CTAG_TX | 3841 NETIF_F_HW_VLAN_STAG_TX); 3842 3843 if (dev->netdev_ops->ndo_features_check) 3844 features &= dev->netdev_ops->ndo_features_check(skb, dev, 3845 features); 3846 else 3847 features &= dflt_features_check(skb, dev, features); 3848 3849 return harmonize_features(skb, features); 3850 } 3851 EXPORT_SYMBOL(netif_skb_features); 3852 3853 static int xmit_one(struct sk_buff *skb, struct net_device *dev, 3854 struct netdev_queue *txq, bool more) 3855 { 3856 unsigned int len; 3857 int rc; 3858 3859 if (dev_nit_active_rcu(dev)) 3860 dev_queue_xmit_nit(skb, dev); 3861 3862 len = skb->len; 3863 trace_net_dev_start_xmit(skb, dev); 3864 rc = netdev_start_xmit(skb, dev, txq, more); 3865 trace_net_dev_xmit(skb, rc, dev, len); 3866 3867 return rc; 3868 } 3869 3870 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev, 3871 struct netdev_queue *txq, int *ret) 3872 { 3873 struct sk_buff *skb = first; 3874 int rc = NETDEV_TX_OK; 3875 3876 while (skb) { 3877 struct sk_buff *next = skb->next; 3878 3879 skb_mark_not_on_list(skb); 3880 rc = xmit_one(skb, dev, txq, next != NULL); 3881 if (unlikely(!dev_xmit_complete(rc))) { 3882 skb->next = next; 3883 goto out; 3884 } 3885 3886 skb = next; 3887 if (netif_tx_queue_stopped(txq) && skb) { 3888 rc = NETDEV_TX_BUSY; 3889 break; 3890 } 3891 } 3892 3893 out: 3894 *ret = rc; 3895 return skb; 3896 } 3897 3898 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb, 3899 netdev_features_t features) 3900 { 3901 if (skb_vlan_tag_present(skb) && 3902 !vlan_hw_offload_capable(features, skb->vlan_proto)) 3903 skb = __vlan_hwaccel_push_inside(skb); 3904 return skb; 3905 } 3906 3907 int skb_csum_hwoffload_help(struct sk_buff *skb, 3908 const netdev_features_t features) 3909 { 3910 if (unlikely(skb_csum_is_sctp(skb))) 3911 return !!(features & NETIF_F_SCTP_CRC) ? 0 : 3912 skb_crc32c_csum_help(skb); 3913 3914 if (features & NETIF_F_HW_CSUM) 3915 return 0; 3916 3917 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) { 3918 if (vlan_get_protocol(skb) == htons(ETH_P_IPV6) && 3919 skb_network_header_len(skb) != sizeof(struct ipv6hdr)) 3920 goto sw_checksum; 3921 3922 switch (skb->csum_offset) { 3923 case offsetof(struct tcphdr, check): 3924 case offsetof(struct udphdr, check): 3925 return 0; 3926 } 3927 } 3928 3929 sw_checksum: 3930 return skb_checksum_help(skb); 3931 } 3932 EXPORT_SYMBOL(skb_csum_hwoffload_help); 3933 3934 /* Checks if this SKB belongs to an HW offloaded socket 3935 * and whether any SW fallbacks are required based on dev. 3936 * Check decrypted mark in case skb_orphan() cleared socket. 3937 */ 3938 static struct sk_buff *sk_validate_xmit_skb(struct sk_buff *skb, 3939 struct net_device *dev) 3940 { 3941 #ifdef CONFIG_SOCK_VALIDATE_XMIT 3942 struct sk_buff *(*sk_validate)(struct sock *sk, struct net_device *dev, 3943 struct sk_buff *skb); 3944 struct sock *sk = skb->sk; 3945 3946 sk_validate = NULL; 3947 if (sk) { 3948 if (sk_fullsock(sk)) 3949 sk_validate = sk->sk_validate_xmit_skb; 3950 else if (sk_is_inet(sk) && sk->sk_state == TCP_TIME_WAIT) 3951 sk_validate = inet_twsk(sk)->tw_validate_xmit_skb; 3952 } 3953 3954 if (sk_validate) { 3955 skb = sk_validate(sk, dev, skb); 3956 } else if (unlikely(skb_is_decrypted(skb))) { 3957 pr_warn_ratelimited("unencrypted skb with no associated socket - dropping\n"); 3958 kfree_skb(skb); 3959 skb = NULL; 3960 } 3961 #endif 3962 3963 return skb; 3964 } 3965 3966 static struct sk_buff *validate_xmit_unreadable_skb(struct sk_buff *skb, 3967 struct net_device *dev) 3968 { 3969 struct skb_shared_info *shinfo; 3970 struct net_iov *niov; 3971 3972 if (likely(skb_frags_readable(skb))) 3973 goto out; 3974 3975 if (!dev->netmem_tx) 3976 goto out_free; 3977 3978 shinfo = skb_shinfo(skb); 3979 3980 if (shinfo->nr_frags > 0) { 3981 niov = netmem_to_net_iov(skb_frag_netmem(&shinfo->frags[0])); 3982 if (net_is_devmem_iov(niov) && 3983 net_devmem_iov_binding(niov)->dev != dev) 3984 goto out_free; 3985 } 3986 3987 out: 3988 return skb; 3989 3990 out_free: 3991 kfree_skb(skb); 3992 return NULL; 3993 } 3994 3995 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again) 3996 { 3997 netdev_features_t features; 3998 3999 skb = validate_xmit_unreadable_skb(skb, dev); 4000 if (unlikely(!skb)) 4001 goto out_null; 4002 4003 features = netif_skb_features(skb); 4004 skb = validate_xmit_vlan(skb, features); 4005 if (unlikely(!skb)) 4006 goto out_null; 4007 4008 skb = sk_validate_xmit_skb(skb, dev); 4009 if (unlikely(!skb)) 4010 goto out_null; 4011 4012 if (netif_needs_gso(skb, features)) { 4013 struct sk_buff *segs; 4014 4015 segs = skb_gso_segment(skb, features); 4016 if (IS_ERR(segs)) { 4017 goto out_kfree_skb; 4018 } else if (segs) { 4019 consume_skb(skb); 4020 skb = segs; 4021 } 4022 } else { 4023 if (skb_needs_linearize(skb, features) && 4024 __skb_linearize(skb)) 4025 goto out_kfree_skb; 4026 4027 /* If packet is not checksummed and device does not 4028 * support checksumming for this protocol, complete 4029 * checksumming here. 4030 */ 4031 if (skb->ip_summed == CHECKSUM_PARTIAL) { 4032 if (skb->encapsulation) 4033 skb_set_inner_transport_header(skb, 4034 skb_checksum_start_offset(skb)); 4035 else 4036 skb_set_transport_header(skb, 4037 skb_checksum_start_offset(skb)); 4038 if (skb_csum_hwoffload_help(skb, features)) 4039 goto out_kfree_skb; 4040 } 4041 } 4042 4043 skb = validate_xmit_xfrm(skb, features, again); 4044 4045 return skb; 4046 4047 out_kfree_skb: 4048 kfree_skb(skb); 4049 out_null: 4050 dev_core_stats_tx_dropped_inc(dev); 4051 return NULL; 4052 } 4053 4054 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again) 4055 { 4056 struct sk_buff *next, *head = NULL, *tail; 4057 4058 for (; skb != NULL; skb = next) { 4059 next = skb->next; 4060 skb_mark_not_on_list(skb); 4061 4062 /* in case skb won't be segmented, point to itself */ 4063 skb->prev = skb; 4064 4065 skb = validate_xmit_skb(skb, dev, again); 4066 if (!skb) 4067 continue; 4068 4069 if (!head) 4070 head = skb; 4071 else 4072 tail->next = skb; 4073 /* If skb was segmented, skb->prev points to 4074 * the last segment. If not, it still contains skb. 4075 */ 4076 tail = skb->prev; 4077 } 4078 return head; 4079 } 4080 EXPORT_SYMBOL_GPL(validate_xmit_skb_list); 4081 4082 static void qdisc_pkt_len_segs_init(struct sk_buff *skb) 4083 { 4084 struct skb_shared_info *shinfo = skb_shinfo(skb); 4085 u16 gso_segs; 4086 4087 qdisc_skb_cb(skb)->pkt_len = skb->len; 4088 if (!shinfo->gso_size) { 4089 qdisc_skb_cb(skb)->pkt_segs = 1; 4090 return; 4091 } 4092 4093 qdisc_skb_cb(skb)->pkt_segs = gso_segs = shinfo->gso_segs; 4094 4095 /* To get more precise estimation of bytes sent on wire, 4096 * we add to pkt_len the headers size of all segments 4097 */ 4098 if (skb_transport_header_was_set(skb)) { 4099 unsigned int hdr_len; 4100 4101 /* mac layer + network layer */ 4102 if (!skb->encapsulation) 4103 hdr_len = skb_transport_offset(skb); 4104 else 4105 hdr_len = skb_inner_transport_offset(skb); 4106 4107 /* + transport layer */ 4108 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) { 4109 const struct tcphdr *th; 4110 struct tcphdr _tcphdr; 4111 4112 th = skb_header_pointer(skb, hdr_len, 4113 sizeof(_tcphdr), &_tcphdr); 4114 if (likely(th)) 4115 hdr_len += __tcp_hdrlen(th); 4116 } else if (shinfo->gso_type & SKB_GSO_UDP_L4) { 4117 struct udphdr _udphdr; 4118 4119 if (skb_header_pointer(skb, hdr_len, 4120 sizeof(_udphdr), &_udphdr)) 4121 hdr_len += sizeof(struct udphdr); 4122 } 4123 4124 if (unlikely(shinfo->gso_type & SKB_GSO_DODGY)) { 4125 int payload = skb->len - hdr_len; 4126 4127 /* Malicious packet. */ 4128 if (payload <= 0) 4129 return; 4130 gso_segs = DIV_ROUND_UP(payload, shinfo->gso_size); 4131 shinfo->gso_segs = gso_segs; 4132 qdisc_skb_cb(skb)->pkt_segs = gso_segs; 4133 } 4134 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len; 4135 } 4136 } 4137 4138 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q, 4139 struct sk_buff **to_free, 4140 struct netdev_queue *txq) 4141 { 4142 int rc; 4143 4144 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK; 4145 if (rc == NET_XMIT_SUCCESS) 4146 trace_qdisc_enqueue(q, txq, skb); 4147 return rc; 4148 } 4149 4150 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q, 4151 struct net_device *dev, 4152 struct netdev_queue *txq) 4153 { 4154 struct sk_buff *next, *to_free = NULL, *to_free2 = NULL; 4155 spinlock_t *root_lock = qdisc_lock(q); 4156 struct llist_node *ll_list, *first_n; 4157 unsigned long defer_count = 0; 4158 int rc; 4159 4160 qdisc_calculate_pkt_len(skb, q); 4161 4162 tcf_set_drop_reason(skb, SKB_DROP_REASON_QDISC_DROP); 4163 4164 if (q->flags & TCQ_F_NOLOCK) { 4165 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) && 4166 qdisc_run_begin(q)) { 4167 /* Retest nolock_qdisc_is_empty() within the protection 4168 * of q->seqlock to protect from racing with requeuing. 4169 */ 4170 if (unlikely(!nolock_qdisc_is_empty(q))) { 4171 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4172 __qdisc_run(q); 4173 to_free2 = qdisc_run_end(q); 4174 4175 goto free_skbs; 4176 } 4177 4178 qdisc_bstats_cpu_update(q, skb); 4179 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) && 4180 !nolock_qdisc_is_empty(q)) 4181 __qdisc_run(q); 4182 4183 to_free2 = qdisc_run_end(q); 4184 rc = NET_XMIT_SUCCESS; 4185 goto free_skbs; 4186 } 4187 4188 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4189 to_free2 = qdisc_run(q); 4190 goto free_skbs; 4191 } 4192 4193 /* Open code llist_add(&skb->ll_node, &q->defer_list) + queue limit. 4194 * In the try_cmpxchg() loop, we want to increment q->defer_count 4195 * at most once to limit the number of skbs in defer_list. 4196 * We perform the defer_count increment only if the list is not empty, 4197 * because some arches have slow atomic_long_inc_return(). 4198 */ 4199 first_n = READ_ONCE(q->defer_list.first); 4200 do { 4201 if (first_n && !defer_count) { 4202 defer_count = atomic_long_inc_return(&q->defer_count); 4203 if (unlikely(defer_count > READ_ONCE(net_hotdata.qdisc_max_burst))) { 4204 kfree_skb_reason(skb, SKB_DROP_REASON_QDISC_BURST_DROP); 4205 return NET_XMIT_DROP; 4206 } 4207 } 4208 skb->ll_node.next = first_n; 4209 } while (!try_cmpxchg(&q->defer_list.first, &first_n, &skb->ll_node)); 4210 4211 /* If defer_list was not empty, we know the cpu which queued 4212 * the first skb will process the whole list for us. 4213 */ 4214 if (first_n) 4215 return NET_XMIT_SUCCESS; 4216 4217 spin_lock(root_lock); 4218 4219 ll_list = llist_del_all(&q->defer_list); 4220 /* There is a small race because we clear defer_count not atomically 4221 * with the prior llist_del_all(). This means defer_list could grow 4222 * over qdisc_max_burst. 4223 */ 4224 atomic_long_set(&q->defer_count, 0); 4225 4226 ll_list = llist_reverse_order(ll_list); 4227 4228 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) { 4229 llist_for_each_entry_safe(skb, next, ll_list, ll_node) 4230 __qdisc_drop(skb, &to_free); 4231 rc = NET_XMIT_DROP; 4232 goto unlock; 4233 } 4234 if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && 4235 !llist_next(ll_list) && qdisc_run_begin(q)) { 4236 /* 4237 * This is a work-conserving queue; there are no old skbs 4238 * waiting to be sent out; and the qdisc is not running - 4239 * xmit the skb directly. 4240 */ 4241 4242 DEBUG_NET_WARN_ON_ONCE(skb != llist_entry(ll_list, 4243 struct sk_buff, 4244 ll_node)); 4245 qdisc_bstats_update(q, skb); 4246 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) 4247 __qdisc_run(q); 4248 to_free2 = qdisc_run_end(q); 4249 rc = NET_XMIT_SUCCESS; 4250 } else { 4251 int count = 0; 4252 4253 llist_for_each_entry_safe(skb, next, ll_list, ll_node) { 4254 if (next) { 4255 prefetch(next); 4256 prefetch(&next->priority); 4257 skb_mark_not_on_list(skb); 4258 } 4259 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4260 count++; 4261 } 4262 to_free2 = qdisc_run(q); 4263 if (count != 1) 4264 rc = NET_XMIT_SUCCESS; 4265 } 4266 unlock: 4267 spin_unlock(root_lock); 4268 4269 free_skbs: 4270 tcf_kfree_skb_list(to_free); 4271 tcf_kfree_skb_list(to_free2); 4272 return rc; 4273 } 4274 4275 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO) 4276 static void skb_update_prio(struct sk_buff *skb) 4277 { 4278 const struct netprio_map *map; 4279 const struct sock *sk; 4280 unsigned int prioidx; 4281 4282 if (skb->priority) 4283 return; 4284 map = rcu_dereference_bh(skb->dev->priomap); 4285 if (!map) 4286 return; 4287 sk = skb_to_full_sk(skb); 4288 if (!sk) 4289 return; 4290 4291 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data); 4292 4293 if (prioidx < map->priomap_len) 4294 skb->priority = map->priomap[prioidx]; 4295 } 4296 #else 4297 #define skb_update_prio(skb) 4298 #endif 4299 4300 /** 4301 * dev_loopback_xmit - loop back @skb 4302 * @net: network namespace this loopback is happening in 4303 * @sk: sk needed to be a netfilter okfn 4304 * @skb: buffer to transmit 4305 */ 4306 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb) 4307 { 4308 skb_reset_mac_header(skb); 4309 __skb_pull(skb, skb_network_offset(skb)); 4310 skb->pkt_type = PACKET_LOOPBACK; 4311 if (skb->ip_summed == CHECKSUM_NONE) 4312 skb->ip_summed = CHECKSUM_UNNECESSARY; 4313 DEBUG_NET_WARN_ON_ONCE(!skb_dst(skb)); 4314 skb_dst_force(skb); 4315 netif_rx(skb); 4316 return 0; 4317 } 4318 EXPORT_SYMBOL(dev_loopback_xmit); 4319 4320 #ifdef CONFIG_NET_EGRESS 4321 static struct netdev_queue * 4322 netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb) 4323 { 4324 int qm = skb_get_queue_mapping(skb); 4325 4326 return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm)); 4327 } 4328 4329 #ifndef CONFIG_PREEMPT_RT 4330 static bool netdev_xmit_txqueue_skipped(void) 4331 { 4332 return __this_cpu_read(softnet_data.xmit.skip_txqueue); 4333 } 4334 4335 void netdev_xmit_skip_txqueue(bool skip) 4336 { 4337 __this_cpu_write(softnet_data.xmit.skip_txqueue, skip); 4338 } 4339 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4340 4341 #else 4342 static bool netdev_xmit_txqueue_skipped(void) 4343 { 4344 return current->net_xmit.skip_txqueue; 4345 } 4346 4347 void netdev_xmit_skip_txqueue(bool skip) 4348 { 4349 current->net_xmit.skip_txqueue = skip; 4350 } 4351 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4352 #endif 4353 #endif /* CONFIG_NET_EGRESS */ 4354 4355 #ifdef CONFIG_NET_XGRESS 4356 static int tc_run(struct tcx_entry *entry, struct sk_buff *skb, 4357 enum skb_drop_reason *drop_reason) 4358 { 4359 int ret = TC_ACT_UNSPEC; 4360 #ifdef CONFIG_NET_CLS_ACT 4361 struct mini_Qdisc *miniq = rcu_dereference_bh(entry->miniq); 4362 struct tcf_result res; 4363 4364 if (!miniq) 4365 return ret; 4366 4367 /* Global bypass */ 4368 if (!static_branch_likely(&tcf_sw_enabled_key)) 4369 return ret; 4370 4371 /* Block-wise bypass */ 4372 if (tcf_block_bypass_sw(miniq->block)) 4373 return ret; 4374 4375 tc_skb_cb(skb)->mru = 0; 4376 qdisc_skb_cb(skb)->post_ct = false; 4377 tcf_set_drop_reason(skb, *drop_reason); 4378 4379 mini_qdisc_bstats_cpu_update(miniq, skb); 4380 ret = tcf_classify(skb, miniq->block, miniq->filter_list, &res, false); 4381 /* Only tcf related quirks below. */ 4382 switch (ret) { 4383 case TC_ACT_SHOT: 4384 *drop_reason = tcf_get_drop_reason(skb); 4385 mini_qdisc_qstats_cpu_drop(miniq); 4386 break; 4387 case TC_ACT_OK: 4388 case TC_ACT_RECLASSIFY: 4389 skb->tc_index = TC_H_MIN(res.classid); 4390 break; 4391 } 4392 #endif /* CONFIG_NET_CLS_ACT */ 4393 return ret; 4394 } 4395 4396 static DEFINE_STATIC_KEY_FALSE(tcx_needed_key); 4397 4398 void tcx_inc(void) 4399 { 4400 static_branch_inc(&tcx_needed_key); 4401 } 4402 4403 void tcx_dec(void) 4404 { 4405 static_branch_dec(&tcx_needed_key); 4406 } 4407 4408 static __always_inline enum tcx_action_base 4409 tcx_run(const struct bpf_mprog_entry *entry, struct sk_buff *skb, 4410 const bool needs_mac) 4411 { 4412 const struct bpf_mprog_fp *fp; 4413 const struct bpf_prog *prog; 4414 int ret = TCX_NEXT; 4415 4416 if (needs_mac) 4417 __skb_push(skb, skb->mac_len); 4418 bpf_mprog_foreach_prog(entry, fp, prog) { 4419 bpf_compute_data_pointers(skb); 4420 ret = bpf_prog_run(prog, skb); 4421 if (ret != TCX_NEXT) 4422 break; 4423 } 4424 if (needs_mac) 4425 __skb_pull(skb, skb->mac_len); 4426 return tcx_action_code(skb, ret); 4427 } 4428 4429 static __always_inline struct sk_buff * 4430 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4431 struct net_device *orig_dev, bool *another) 4432 { 4433 struct bpf_mprog_entry *entry = rcu_dereference_bh(skb->dev->tcx_ingress); 4434 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_INGRESS; 4435 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4436 int sch_ret; 4437 4438 if (!entry) 4439 return skb; 4440 4441 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4442 if (unlikely(*pt_prev)) { 4443 *ret = deliver_skb(skb, *pt_prev, orig_dev); 4444 *pt_prev = NULL; 4445 } 4446 4447 qdisc_pkt_len_segs_init(skb); 4448 tcx_set_ingress(skb, true); 4449 4450 if (static_branch_unlikely(&tcx_needed_key)) { 4451 sch_ret = tcx_run(entry, skb, true); 4452 if (sch_ret != TC_ACT_UNSPEC) 4453 goto ingress_verdict; 4454 } 4455 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4456 ingress_verdict: 4457 switch (sch_ret) { 4458 case TC_ACT_REDIRECT: 4459 /* skb_mac_header check was done by BPF, so we can safely 4460 * push the L2 header back before redirecting to another 4461 * netdev. 4462 */ 4463 __skb_push(skb, skb->mac_len); 4464 if (skb_do_redirect(skb) == -EAGAIN) { 4465 __skb_pull(skb, skb->mac_len); 4466 *another = true; 4467 break; 4468 } 4469 *ret = NET_RX_SUCCESS; 4470 bpf_net_ctx_clear(bpf_net_ctx); 4471 return NULL; 4472 case TC_ACT_SHOT: 4473 kfree_skb_reason(skb, drop_reason); 4474 *ret = NET_RX_DROP; 4475 bpf_net_ctx_clear(bpf_net_ctx); 4476 return NULL; 4477 /* used by tc_run */ 4478 case TC_ACT_STOLEN: 4479 case TC_ACT_QUEUED: 4480 case TC_ACT_TRAP: 4481 consume_skb(skb); 4482 fallthrough; 4483 case TC_ACT_CONSUMED: 4484 *ret = NET_RX_SUCCESS; 4485 bpf_net_ctx_clear(bpf_net_ctx); 4486 return NULL; 4487 } 4488 bpf_net_ctx_clear(bpf_net_ctx); 4489 4490 return skb; 4491 } 4492 4493 static __always_inline struct sk_buff * 4494 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4495 { 4496 struct bpf_mprog_entry *entry = rcu_dereference_bh(dev->tcx_egress); 4497 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_EGRESS; 4498 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4499 int sch_ret; 4500 4501 if (!entry) 4502 return skb; 4503 4504 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4505 4506 /* qdisc_skb_cb(skb)->pkt_len & tcx_set_ingress() was 4507 * already set by the caller. 4508 */ 4509 if (static_branch_unlikely(&tcx_needed_key)) { 4510 sch_ret = tcx_run(entry, skb, false); 4511 if (sch_ret != TC_ACT_UNSPEC) 4512 goto egress_verdict; 4513 } 4514 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4515 egress_verdict: 4516 switch (sch_ret) { 4517 case TC_ACT_REDIRECT: 4518 /* No need to push/pop skb's mac_header here on egress! */ 4519 skb_do_redirect(skb); 4520 *ret = NET_XMIT_SUCCESS; 4521 bpf_net_ctx_clear(bpf_net_ctx); 4522 return NULL; 4523 case TC_ACT_SHOT: 4524 kfree_skb_reason(skb, drop_reason); 4525 *ret = NET_XMIT_DROP; 4526 bpf_net_ctx_clear(bpf_net_ctx); 4527 return NULL; 4528 /* used by tc_run */ 4529 case TC_ACT_STOLEN: 4530 case TC_ACT_QUEUED: 4531 case TC_ACT_TRAP: 4532 consume_skb(skb); 4533 fallthrough; 4534 case TC_ACT_CONSUMED: 4535 *ret = NET_XMIT_SUCCESS; 4536 bpf_net_ctx_clear(bpf_net_ctx); 4537 return NULL; 4538 } 4539 bpf_net_ctx_clear(bpf_net_ctx); 4540 4541 return skb; 4542 } 4543 #else 4544 static __always_inline struct sk_buff * 4545 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4546 struct net_device *orig_dev, bool *another) 4547 { 4548 return skb; 4549 } 4550 4551 static __always_inline struct sk_buff * 4552 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4553 { 4554 return skb; 4555 } 4556 #endif /* CONFIG_NET_XGRESS */ 4557 4558 #ifdef CONFIG_XPS 4559 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb, 4560 struct xps_dev_maps *dev_maps, unsigned int tci) 4561 { 4562 int tc = netdev_get_prio_tc_map(dev, skb->priority); 4563 struct xps_map *map; 4564 int queue_index = -1; 4565 4566 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids) 4567 return queue_index; 4568 4569 tci *= dev_maps->num_tc; 4570 tci += tc; 4571 4572 map = rcu_dereference(dev_maps->attr_map[tci]); 4573 if (map) { 4574 if (map->len == 1) 4575 queue_index = map->queues[0]; 4576 else 4577 queue_index = map->queues[reciprocal_scale( 4578 skb_get_hash(skb), map->len)]; 4579 if (unlikely(queue_index >= dev->real_num_tx_queues)) 4580 queue_index = -1; 4581 } 4582 return queue_index; 4583 } 4584 #endif 4585 4586 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev, 4587 struct sk_buff *skb) 4588 { 4589 #ifdef CONFIG_XPS 4590 struct xps_dev_maps *dev_maps; 4591 struct sock *sk = skb->sk; 4592 int queue_index = -1; 4593 4594 if (!static_key_false(&xps_needed)) 4595 return -1; 4596 4597 rcu_read_lock(); 4598 if (!static_key_false(&xps_rxqs_needed)) 4599 goto get_cpus_map; 4600 4601 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]); 4602 if (dev_maps) { 4603 int tci = sk_rx_queue_get(sk); 4604 4605 if (tci >= 0) 4606 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4607 tci); 4608 } 4609 4610 get_cpus_map: 4611 if (queue_index < 0) { 4612 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]); 4613 if (dev_maps) { 4614 unsigned int tci = skb->sender_cpu - 1; 4615 4616 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4617 tci); 4618 } 4619 } 4620 rcu_read_unlock(); 4621 4622 return queue_index; 4623 #else 4624 return -1; 4625 #endif 4626 } 4627 4628 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb, 4629 struct net_device *sb_dev) 4630 { 4631 return 0; 4632 } 4633 EXPORT_SYMBOL(dev_pick_tx_zero); 4634 4635 int sk_tx_queue_get(const struct sock *sk) 4636 { 4637 int resel, val; 4638 4639 if (!sk) 4640 return -1; 4641 /* Paired with WRITE_ONCE() in sk_tx_queue_clear() 4642 * and sk_tx_queue_set(). 4643 */ 4644 val = READ_ONCE(sk->sk_tx_queue_mapping); 4645 4646 if (val == NO_QUEUE_MAPPING) 4647 return -1; 4648 4649 if (!sk_fullsock(sk)) 4650 return val; 4651 4652 resel = READ_ONCE(sock_net(sk)->core.sysctl_txq_reselection); 4653 if (resel && time_is_before_jiffies( 4654 READ_ONCE(sk->sk_tx_queue_mapping_jiffies) + resel)) 4655 return -1; 4656 4657 return val; 4658 } 4659 EXPORT_SYMBOL(sk_tx_queue_get); 4660 4661 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb, 4662 struct net_device *sb_dev) 4663 { 4664 struct sock *sk = skb->sk; 4665 int queue_index = sk_tx_queue_get(sk); 4666 4667 sb_dev = sb_dev ? : dev; 4668 4669 if (queue_index < 0 || skb->ooo_okay || 4670 queue_index >= dev->real_num_tx_queues) { 4671 int new_index = get_xps_queue(dev, sb_dev, skb); 4672 4673 if (new_index < 0) 4674 new_index = skb_tx_hash(dev, sb_dev, skb); 4675 4676 if (sk && sk_fullsock(sk) && 4677 rcu_access_pointer(sk->sk_dst_cache)) 4678 sk_tx_queue_set(sk, new_index); 4679 4680 queue_index = new_index; 4681 } 4682 4683 return queue_index; 4684 } 4685 EXPORT_SYMBOL(netdev_pick_tx); 4686 4687 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev, 4688 struct sk_buff *skb, 4689 struct net_device *sb_dev) 4690 { 4691 int queue_index = 0; 4692 4693 #ifdef CONFIG_XPS 4694 u32 sender_cpu = skb->sender_cpu - 1; 4695 4696 if (sender_cpu >= (u32)NR_CPUS) 4697 skb->sender_cpu = raw_smp_processor_id() + 1; 4698 #endif 4699 4700 if (dev->real_num_tx_queues != 1) { 4701 const struct net_device_ops *ops = dev->netdev_ops; 4702 4703 if (ops->ndo_select_queue) 4704 queue_index = ops->ndo_select_queue(dev, skb, sb_dev); 4705 else 4706 queue_index = netdev_pick_tx(dev, skb, sb_dev); 4707 4708 queue_index = netdev_cap_txqueue(dev, queue_index); 4709 } 4710 4711 skb_set_queue_mapping(skb, queue_index); 4712 return netdev_get_tx_queue(dev, queue_index); 4713 } 4714 4715 /** 4716 * __dev_queue_xmit() - transmit a buffer 4717 * @skb: buffer to transmit 4718 * @sb_dev: suboordinate device used for L2 forwarding offload 4719 * 4720 * Queue a buffer for transmission to a network device. The caller must 4721 * have set the device and priority and built the buffer before calling 4722 * this function. The function can be called from an interrupt. 4723 * 4724 * When calling this method, interrupts MUST be enabled. This is because 4725 * the BH enable code must have IRQs enabled so that it will not deadlock. 4726 * 4727 * Regardless of the return value, the skb is consumed, so it is currently 4728 * difficult to retry a send to this method. (You can bump the ref count 4729 * before sending to hold a reference for retry if you are careful.) 4730 * 4731 * Return: 4732 * * 0 - buffer successfully transmitted 4733 * * positive qdisc return code - NET_XMIT_DROP etc. 4734 * * negative errno - other errors 4735 */ 4736 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev) 4737 { 4738 struct net_device *dev = skb->dev; 4739 struct netdev_queue *txq = NULL; 4740 struct Qdisc *q; 4741 int rc = -ENOMEM; 4742 bool again = false; 4743 4744 skb_reset_mac_header(skb); 4745 skb_assert_len(skb); 4746 4747 if (unlikely(skb_shinfo(skb)->tx_flags & 4748 (SKBTX_SCHED_TSTAMP | SKBTX_BPF))) 4749 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED); 4750 4751 /* Disable soft irqs for various locks below. Also 4752 * stops preemption for RCU. 4753 */ 4754 rcu_read_lock_bh(); 4755 4756 skb_update_prio(skb); 4757 4758 qdisc_pkt_len_segs_init(skb); 4759 tcx_set_ingress(skb, false); 4760 #ifdef CONFIG_NET_EGRESS 4761 if (static_branch_unlikely(&egress_needed_key)) { 4762 if (nf_hook_egress_active()) { 4763 skb = nf_hook_egress(skb, &rc, dev); 4764 if (!skb) 4765 goto out; 4766 } 4767 4768 netdev_xmit_skip_txqueue(false); 4769 4770 nf_skip_egress(skb, true); 4771 skb = sch_handle_egress(skb, &rc, dev); 4772 if (!skb) 4773 goto out; 4774 nf_skip_egress(skb, false); 4775 4776 if (netdev_xmit_txqueue_skipped()) 4777 txq = netdev_tx_queue_mapping(dev, skb); 4778 } 4779 #endif 4780 /* If device/qdisc don't need skb->dst, release it right now while 4781 * its hot in this cpu cache. 4782 */ 4783 if (dev->priv_flags & IFF_XMIT_DST_RELEASE) 4784 skb_dst_drop(skb); 4785 else 4786 skb_dst_force(skb); 4787 4788 if (!txq) 4789 txq = netdev_core_pick_tx(dev, skb, sb_dev); 4790 4791 q = rcu_dereference_bh(txq->qdisc); 4792 4793 trace_net_dev_queue(skb); 4794 if (q->enqueue) { 4795 rc = __dev_xmit_skb(skb, q, dev, txq); 4796 goto out; 4797 } 4798 4799 /* The device has no queue. Common case for software devices: 4800 * loopback, all the sorts of tunnels... 4801 4802 * Really, it is unlikely that netif_tx_lock protection is necessary 4803 * here. (f.e. loopback and IP tunnels are clean ignoring statistics 4804 * counters.) 4805 * However, it is possible, that they rely on protection 4806 * made by us here. 4807 4808 * Check this and shot the lock. It is not prone from deadlocks. 4809 *Either shot noqueue qdisc, it is even simpler 8) 4810 */ 4811 if (dev->flags & IFF_UP) { 4812 int cpu = smp_processor_id(); /* ok because BHs are off */ 4813 4814 /* Other cpus might concurrently change txq->xmit_lock_owner 4815 * to -1 or to their cpu id, but not to our id. 4816 */ 4817 if (READ_ONCE(txq->xmit_lock_owner) != cpu) { 4818 if (dev_xmit_recursion()) 4819 goto recursion_alert; 4820 4821 skb = validate_xmit_skb(skb, dev, &again); 4822 if (!skb) 4823 goto out; 4824 4825 HARD_TX_LOCK(dev, txq, cpu); 4826 4827 if (!netif_xmit_stopped(txq)) { 4828 dev_xmit_recursion_inc(); 4829 skb = dev_hard_start_xmit(skb, dev, txq, &rc); 4830 dev_xmit_recursion_dec(); 4831 if (dev_xmit_complete(rc)) { 4832 HARD_TX_UNLOCK(dev, txq); 4833 goto out; 4834 } 4835 } 4836 HARD_TX_UNLOCK(dev, txq); 4837 net_crit_ratelimited("Virtual device %s asks to queue packet!\n", 4838 dev->name); 4839 } else { 4840 /* Recursion is detected! It is possible, 4841 * unfortunately 4842 */ 4843 recursion_alert: 4844 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n", 4845 dev->name); 4846 } 4847 } 4848 4849 rc = -ENETDOWN; 4850 rcu_read_unlock_bh(); 4851 4852 dev_core_stats_tx_dropped_inc(dev); 4853 kfree_skb_list(skb); 4854 return rc; 4855 out: 4856 rcu_read_unlock_bh(); 4857 return rc; 4858 } 4859 EXPORT_SYMBOL(__dev_queue_xmit); 4860 4861 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id) 4862 { 4863 struct net_device *dev = skb->dev; 4864 struct sk_buff *orig_skb = skb; 4865 struct netdev_queue *txq; 4866 int ret = NETDEV_TX_BUSY; 4867 bool again = false; 4868 4869 if (unlikely(!netif_running(dev) || 4870 !netif_carrier_ok(dev))) 4871 goto drop; 4872 4873 skb = validate_xmit_skb_list(skb, dev, &again); 4874 if (skb != orig_skb) 4875 goto drop; 4876 4877 skb_set_queue_mapping(skb, queue_id); 4878 txq = skb_get_tx_queue(dev, skb); 4879 4880 local_bh_disable(); 4881 4882 dev_xmit_recursion_inc(); 4883 HARD_TX_LOCK(dev, txq, smp_processor_id()); 4884 if (!netif_xmit_frozen_or_drv_stopped(txq)) 4885 ret = netdev_start_xmit(skb, dev, txq, false); 4886 HARD_TX_UNLOCK(dev, txq); 4887 dev_xmit_recursion_dec(); 4888 4889 local_bh_enable(); 4890 return ret; 4891 drop: 4892 dev_core_stats_tx_dropped_inc(dev); 4893 kfree_skb_list(skb); 4894 return NET_XMIT_DROP; 4895 } 4896 EXPORT_SYMBOL(__dev_direct_xmit); 4897 4898 /************************************************************************* 4899 * Receiver routines 4900 *************************************************************************/ 4901 static DEFINE_PER_CPU(struct task_struct *, backlog_napi); 4902 4903 int weight_p __read_mostly = 64; /* old backlog weight */ 4904 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */ 4905 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */ 4906 4907 /* Called with irq disabled */ 4908 static inline void ____napi_schedule(struct softnet_data *sd, 4909 struct napi_struct *napi) 4910 { 4911 struct task_struct *thread; 4912 4913 lockdep_assert_irqs_disabled(); 4914 4915 if (test_bit(NAPI_STATE_THREADED, &napi->state)) { 4916 /* Paired with smp_mb__before_atomic() in 4917 * napi_enable()/netif_set_threaded(). 4918 * Use READ_ONCE() to guarantee a complete 4919 * read on napi->thread. Only call 4920 * wake_up_process() when it's not NULL. 4921 */ 4922 thread = READ_ONCE(napi->thread); 4923 if (thread) { 4924 if (use_backlog_threads() && thread == raw_cpu_read(backlog_napi)) 4925 goto use_local_napi; 4926 4927 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 4928 wake_up_process(thread); 4929 return; 4930 } 4931 } 4932 4933 use_local_napi: 4934 DEBUG_NET_WARN_ON_ONCE(!list_empty(&napi->poll_list)); 4935 list_add_tail(&napi->poll_list, &sd->poll_list); 4936 WRITE_ONCE(napi->list_owner, smp_processor_id()); 4937 /* If not called from net_rx_action() 4938 * we have to raise NET_RX_SOFTIRQ. 4939 */ 4940 if (!sd->in_net_rx_action) 4941 raise_softirq_irqoff(NET_RX_SOFTIRQ); 4942 } 4943 4944 #ifdef CONFIG_RPS 4945 4946 struct static_key_false rps_needed __read_mostly; 4947 EXPORT_SYMBOL(rps_needed); 4948 struct static_key_false rfs_needed __read_mostly; 4949 EXPORT_SYMBOL(rfs_needed); 4950 4951 static u32 rfs_slot(u32 hash, const struct rps_dev_flow_table *flow_table) 4952 { 4953 return hash_32(hash, flow_table->log); 4954 } 4955 4956 #ifdef CONFIG_RFS_ACCEL 4957 /** 4958 * rps_flow_is_active - check whether the flow is recently active. 4959 * @rflow: Specific flow to check activity. 4960 * @flow_table: per-queue flowtable that @rflow belongs to. 4961 * @cpu: CPU saved in @rflow. 4962 * 4963 * If the CPU has processed many packets since the flow's last activity 4964 * (beyond 10 times the table size), the flow is considered stale. 4965 * 4966 * Return: true if flow was recently active. 4967 */ 4968 static bool rps_flow_is_active(struct rps_dev_flow *rflow, 4969 struct rps_dev_flow_table *flow_table, 4970 unsigned int cpu) 4971 { 4972 unsigned int flow_last_active; 4973 unsigned int sd_input_head; 4974 4975 if (cpu >= nr_cpu_ids) 4976 return false; 4977 4978 sd_input_head = READ_ONCE(per_cpu(softnet_data, cpu).input_queue_head); 4979 flow_last_active = READ_ONCE(rflow->last_qtail); 4980 4981 return (int)(sd_input_head - flow_last_active) < 4982 (int)(10 << flow_table->log); 4983 } 4984 #endif 4985 4986 static struct rps_dev_flow * 4987 set_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4988 struct rps_dev_flow *rflow, u16 next_cpu, u32 hash, 4989 u32 flow_id) 4990 { 4991 if (next_cpu < nr_cpu_ids) { 4992 u32 head; 4993 #ifdef CONFIG_RFS_ACCEL 4994 struct netdev_rx_queue *rxqueue; 4995 struct rps_dev_flow_table *flow_table; 4996 struct rps_dev_flow *old_rflow; 4997 struct rps_dev_flow *tmp_rflow; 4998 unsigned int tmp_cpu; 4999 u16 rxq_index; 5000 int rc; 5001 5002 /* Should we steer this flow to a different hardware queue? */ 5003 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap || 5004 !(dev->features & NETIF_F_NTUPLE)) 5005 goto out; 5006 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu); 5007 if (rxq_index == skb_get_rx_queue(skb)) 5008 goto out; 5009 5010 rxqueue = dev->_rx + rxq_index; 5011 flow_table = rcu_dereference(rxqueue->rps_flow_table); 5012 if (!flow_table) 5013 goto out; 5014 5015 tmp_rflow = &flow_table->flows[flow_id]; 5016 tmp_cpu = READ_ONCE(tmp_rflow->cpu); 5017 5018 if (READ_ONCE(tmp_rflow->filter) != RPS_NO_FILTER) { 5019 if (rps_flow_is_active(tmp_rflow, flow_table, 5020 tmp_cpu)) { 5021 if (hash != READ_ONCE(tmp_rflow->hash) || 5022 next_cpu == tmp_cpu) 5023 goto out; 5024 } 5025 } 5026 5027 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb, 5028 rxq_index, flow_id); 5029 if (rc < 0) 5030 goto out; 5031 5032 old_rflow = rflow; 5033 rflow = tmp_rflow; 5034 WRITE_ONCE(rflow->filter, rc); 5035 WRITE_ONCE(rflow->hash, hash); 5036 5037 if (old_rflow->filter == rc) 5038 WRITE_ONCE(old_rflow->filter, RPS_NO_FILTER); 5039 out: 5040 #endif 5041 head = READ_ONCE(per_cpu(softnet_data, next_cpu).input_queue_head); 5042 rps_input_queue_tail_save(&rflow->last_qtail, head); 5043 } 5044 5045 WRITE_ONCE(rflow->cpu, next_cpu); 5046 return rflow; 5047 } 5048 5049 /* 5050 * get_rps_cpu is called from netif_receive_skb and returns the target 5051 * CPU from the RPS map of the receiving queue for a given skb. 5052 * rcu_read_lock must be held on entry. 5053 */ 5054 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb, 5055 struct rps_dev_flow **rflowp) 5056 { 5057 const struct rps_sock_flow_table *sock_flow_table; 5058 struct netdev_rx_queue *rxqueue = dev->_rx; 5059 struct rps_dev_flow_table *flow_table; 5060 struct rps_map *map; 5061 int cpu = -1; 5062 u32 flow_id; 5063 u32 tcpu; 5064 u32 hash; 5065 5066 if (skb_rx_queue_recorded(skb)) { 5067 u16 index = skb_get_rx_queue(skb); 5068 5069 if (unlikely(index >= dev->real_num_rx_queues)) { 5070 WARN_ONCE(dev->real_num_rx_queues > 1, 5071 "%s received packet on queue %u, but number " 5072 "of RX queues is %u\n", 5073 dev->name, index, dev->real_num_rx_queues); 5074 goto done; 5075 } 5076 rxqueue += index; 5077 } 5078 5079 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */ 5080 5081 flow_table = rcu_dereference(rxqueue->rps_flow_table); 5082 map = rcu_dereference(rxqueue->rps_map); 5083 if (!flow_table && !map) 5084 goto done; 5085 5086 skb_reset_network_header(skb); 5087 hash = skb_get_hash(skb); 5088 if (!hash) 5089 goto done; 5090 5091 sock_flow_table = rcu_dereference(net_hotdata.rps_sock_flow_table); 5092 if (flow_table && sock_flow_table) { 5093 struct rps_dev_flow *rflow; 5094 u32 next_cpu; 5095 u32 ident; 5096 5097 /* First check into global flow table if there is a match. 5098 * This READ_ONCE() pairs with WRITE_ONCE() from rps_record_sock_flow(). 5099 */ 5100 ident = READ_ONCE(sock_flow_table->ents[hash & sock_flow_table->mask]); 5101 if ((ident ^ hash) & ~net_hotdata.rps_cpu_mask) 5102 goto try_rps; 5103 5104 next_cpu = ident & net_hotdata.rps_cpu_mask; 5105 5106 /* OK, now we know there is a match, 5107 * we can look at the local (per receive queue) flow table 5108 */ 5109 flow_id = rfs_slot(hash, flow_table); 5110 rflow = &flow_table->flows[flow_id]; 5111 tcpu = rflow->cpu; 5112 5113 /* 5114 * If the desired CPU (where last recvmsg was done) is 5115 * different from current CPU (one in the rx-queue flow 5116 * table entry), switch if one of the following holds: 5117 * - Current CPU is unset (>= nr_cpu_ids). 5118 * - Current CPU is offline. 5119 * - The current CPU's queue tail has advanced beyond the 5120 * last packet that was enqueued using this table entry. 5121 * This guarantees that all previous packets for the flow 5122 * have been dequeued, thus preserving in order delivery. 5123 */ 5124 if (unlikely(tcpu != next_cpu) && 5125 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) || 5126 ((int)(READ_ONCE(per_cpu(softnet_data, tcpu).input_queue_head) - 5127 rflow->last_qtail)) >= 0)) { 5128 tcpu = next_cpu; 5129 rflow = set_rps_cpu(dev, skb, rflow, next_cpu, hash, 5130 flow_id); 5131 } 5132 5133 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) { 5134 *rflowp = rflow; 5135 cpu = tcpu; 5136 goto done; 5137 } 5138 } 5139 5140 try_rps: 5141 5142 if (map) { 5143 tcpu = map->cpus[reciprocal_scale(hash, map->len)]; 5144 if (cpu_online(tcpu)) { 5145 cpu = tcpu; 5146 goto done; 5147 } 5148 } 5149 5150 done: 5151 return cpu; 5152 } 5153 5154 #ifdef CONFIG_RFS_ACCEL 5155 5156 /** 5157 * rps_may_expire_flow - check whether an RFS hardware filter may be removed 5158 * @dev: Device on which the filter was set 5159 * @rxq_index: RX queue index 5160 * @flow_id: Flow ID passed to ndo_rx_flow_steer() 5161 * @filter_id: Filter ID returned by ndo_rx_flow_steer() 5162 * 5163 * Drivers that implement ndo_rx_flow_steer() should periodically call 5164 * this function for each installed filter and remove the filters for 5165 * which it returns %true. 5166 */ 5167 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index, 5168 u32 flow_id, u16 filter_id) 5169 { 5170 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index; 5171 struct rps_dev_flow_table *flow_table; 5172 struct rps_dev_flow *rflow; 5173 bool expire = true; 5174 5175 rcu_read_lock(); 5176 flow_table = rcu_dereference(rxqueue->rps_flow_table); 5177 if (flow_table && flow_id < (1UL << flow_table->log)) { 5178 unsigned int cpu; 5179 5180 rflow = &flow_table->flows[flow_id]; 5181 cpu = READ_ONCE(rflow->cpu); 5182 if (READ_ONCE(rflow->filter) == filter_id && 5183 rps_flow_is_active(rflow, flow_table, cpu)) 5184 expire = false; 5185 } 5186 rcu_read_unlock(); 5187 return expire; 5188 } 5189 EXPORT_SYMBOL(rps_may_expire_flow); 5190 5191 #endif /* CONFIG_RFS_ACCEL */ 5192 5193 /* Called from hardirq (IPI) context */ 5194 static void rps_trigger_softirq(void *data) 5195 { 5196 struct softnet_data *sd = data; 5197 5198 ____napi_schedule(sd, &sd->backlog); 5199 /* Pairs with READ_ONCE() in softnet_seq_show() */ 5200 WRITE_ONCE(sd->received_rps, sd->received_rps + 1); 5201 } 5202 5203 #endif /* CONFIG_RPS */ 5204 5205 /* Called from hardirq (IPI) context */ 5206 static void trigger_rx_softirq(void *data) 5207 { 5208 struct softnet_data *sd = data; 5209 5210 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 5211 smp_store_release(&sd->defer_ipi_scheduled, 0); 5212 } 5213 5214 /* 5215 * After we queued a packet into sd->input_pkt_queue, 5216 * we need to make sure this queue is serviced soon. 5217 * 5218 * - If this is another cpu queue, link it to our rps_ipi_list, 5219 * and make sure we will process rps_ipi_list from net_rx_action(). 5220 * 5221 * - If this is our own queue, NAPI schedule our backlog. 5222 * Note that this also raises NET_RX_SOFTIRQ. 5223 */ 5224 static void napi_schedule_rps(struct softnet_data *sd) 5225 { 5226 struct softnet_data *mysd = this_cpu_ptr(&softnet_data); 5227 5228 #ifdef CONFIG_RPS 5229 if (sd != mysd) { 5230 if (use_backlog_threads()) { 5231 __napi_schedule_irqoff(&sd->backlog); 5232 return; 5233 } 5234 5235 sd->rps_ipi_next = mysd->rps_ipi_list; 5236 mysd->rps_ipi_list = sd; 5237 5238 /* If not called from net_rx_action() or napi_threaded_poll() 5239 * we have to raise NET_RX_SOFTIRQ. 5240 */ 5241 if (!mysd->in_net_rx_action && !mysd->in_napi_threaded_poll) 5242 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 5243 return; 5244 } 5245 #endif /* CONFIG_RPS */ 5246 __napi_schedule_irqoff(&mysd->backlog); 5247 } 5248 5249 void kick_defer_list_purge(unsigned int cpu) 5250 { 5251 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 5252 unsigned long flags; 5253 5254 if (use_backlog_threads()) { 5255 backlog_lock_irq_save(sd, &flags); 5256 5257 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) 5258 __napi_schedule_irqoff(&sd->backlog); 5259 5260 backlog_unlock_irq_restore(sd, flags); 5261 5262 } else if (!cmpxchg(&sd->defer_ipi_scheduled, 0, 1)) { 5263 smp_call_function_single_async(cpu, &sd->defer_csd); 5264 } 5265 } 5266 5267 #ifdef CONFIG_NET_FLOW_LIMIT 5268 int netdev_flow_limit_table_len __read_mostly = (1 << 12); 5269 #endif 5270 5271 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen, 5272 int max_backlog) 5273 { 5274 #ifdef CONFIG_NET_FLOW_LIMIT 5275 unsigned int old_flow, new_flow; 5276 const struct softnet_data *sd; 5277 struct sd_flow_limit *fl; 5278 5279 if (likely(qlen < (max_backlog >> 1))) 5280 return false; 5281 5282 sd = this_cpu_ptr(&softnet_data); 5283 5284 rcu_read_lock(); 5285 fl = rcu_dereference(sd->flow_limit); 5286 if (fl) { 5287 new_flow = hash_32(skb_get_hash(skb), fl->log_buckets); 5288 old_flow = fl->history[fl->history_head]; 5289 fl->history[fl->history_head] = new_flow; 5290 5291 fl->history_head++; 5292 fl->history_head &= FLOW_LIMIT_HISTORY - 1; 5293 5294 if (likely(fl->buckets[old_flow])) 5295 fl->buckets[old_flow]--; 5296 5297 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) { 5298 /* Pairs with READ_ONCE() in softnet_seq_show() */ 5299 WRITE_ONCE(fl->count, fl->count + 1); 5300 rcu_read_unlock(); 5301 return true; 5302 } 5303 } 5304 rcu_read_unlock(); 5305 #endif 5306 return false; 5307 } 5308 5309 /* 5310 * enqueue_to_backlog is called to queue an skb to a per CPU backlog 5311 * queue (may be a remote CPU queue). 5312 */ 5313 static int enqueue_to_backlog(struct sk_buff *skb, int cpu, 5314 unsigned int *qtail) 5315 { 5316 enum skb_drop_reason reason; 5317 struct softnet_data *sd; 5318 unsigned long flags; 5319 unsigned int qlen; 5320 int max_backlog; 5321 u32 tail; 5322 5323 reason = SKB_DROP_REASON_DEV_READY; 5324 if (unlikely(!netif_running(skb->dev))) 5325 goto bad_dev; 5326 5327 sd = &per_cpu(softnet_data, cpu); 5328 5329 qlen = skb_queue_len_lockless(&sd->input_pkt_queue); 5330 max_backlog = READ_ONCE(net_hotdata.max_backlog); 5331 if (unlikely(qlen > max_backlog) || 5332 skb_flow_limit(skb, qlen, max_backlog)) 5333 goto cpu_backlog_drop; 5334 backlog_lock_irq_save(sd, &flags); 5335 qlen = skb_queue_len(&sd->input_pkt_queue); 5336 if (likely(qlen <= max_backlog)) { 5337 if (!qlen) { 5338 /* Schedule NAPI for backlog device. We can use 5339 * non atomic operation as we own the queue lock. 5340 */ 5341 if (!__test_and_set_bit(NAPI_STATE_SCHED, 5342 &sd->backlog.state)) 5343 napi_schedule_rps(sd); 5344 } 5345 __skb_queue_tail(&sd->input_pkt_queue, skb); 5346 tail = rps_input_queue_tail_incr(sd); 5347 backlog_unlock_irq_restore(sd, flags); 5348 5349 /* save the tail outside of the critical section */ 5350 rps_input_queue_tail_save(qtail, tail); 5351 return NET_RX_SUCCESS; 5352 } 5353 5354 backlog_unlock_irq_restore(sd, flags); 5355 5356 cpu_backlog_drop: 5357 reason = SKB_DROP_REASON_CPU_BACKLOG; 5358 numa_drop_add(&sd->drop_counters, 1); 5359 bad_dev: 5360 dev_core_stats_rx_dropped_inc(skb->dev); 5361 kfree_skb_reason(skb, reason); 5362 return NET_RX_DROP; 5363 } 5364 5365 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb) 5366 { 5367 struct net_device *dev = skb->dev; 5368 struct netdev_rx_queue *rxqueue; 5369 5370 rxqueue = dev->_rx; 5371 5372 if (skb_rx_queue_recorded(skb)) { 5373 u16 index = skb_get_rx_queue(skb); 5374 5375 if (unlikely(index >= dev->real_num_rx_queues)) { 5376 WARN_ONCE(dev->real_num_rx_queues > 1, 5377 "%s received packet on queue %u, but number " 5378 "of RX queues is %u\n", 5379 dev->name, index, dev->real_num_rx_queues); 5380 5381 return rxqueue; /* Return first rxqueue */ 5382 } 5383 rxqueue += index; 5384 } 5385 return rxqueue; 5386 } 5387 5388 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, 5389 const struct bpf_prog *xdp_prog) 5390 { 5391 void *orig_data, *orig_data_end, *hard_start; 5392 struct netdev_rx_queue *rxqueue; 5393 bool orig_bcast, orig_host; 5394 u32 mac_len, frame_sz; 5395 __be16 orig_eth_type; 5396 struct ethhdr *eth; 5397 u32 metalen, act; 5398 int off; 5399 5400 /* The XDP program wants to see the packet starting at the MAC 5401 * header. 5402 */ 5403 mac_len = skb->data - skb_mac_header(skb); 5404 hard_start = skb->data - skb_headroom(skb); 5405 5406 /* SKB "head" area always have tailroom for skb_shared_info */ 5407 frame_sz = (void *)skb_end_pointer(skb) - hard_start; 5408 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 5409 5410 rxqueue = netif_get_rxqueue(skb); 5411 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq); 5412 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len, 5413 skb_headlen(skb) + mac_len, true); 5414 if (skb_is_nonlinear(skb)) { 5415 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 5416 xdp_buff_set_frags_flag(xdp); 5417 } else { 5418 xdp_buff_clear_frags_flag(xdp); 5419 } 5420 5421 orig_data_end = xdp->data_end; 5422 orig_data = xdp->data; 5423 eth = (struct ethhdr *)xdp->data; 5424 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr); 5425 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest); 5426 orig_eth_type = eth->h_proto; 5427 5428 act = bpf_prog_run_xdp(xdp_prog, xdp); 5429 5430 /* check if bpf_xdp_adjust_head was used */ 5431 off = xdp->data - orig_data; 5432 if (off) { 5433 if (off > 0) 5434 __skb_pull(skb, off); 5435 else if (off < 0) 5436 __skb_push(skb, -off); 5437 5438 skb->mac_header += off; 5439 skb_reset_network_header(skb); 5440 } 5441 5442 /* check if bpf_xdp_adjust_tail was used */ 5443 off = xdp->data_end - orig_data_end; 5444 if (off != 0) { 5445 skb_set_tail_pointer(skb, xdp->data_end - xdp->data); 5446 skb->len += off; /* positive on grow, negative on shrink */ 5447 } 5448 5449 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 5450 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here. 5451 */ 5452 if (xdp_buff_has_frags(xdp)) 5453 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 5454 else 5455 skb->data_len = 0; 5456 5457 /* check if XDP changed eth hdr such SKB needs update */ 5458 eth = (struct ethhdr *)xdp->data; 5459 if ((orig_eth_type != eth->h_proto) || 5460 (orig_host != ether_addr_equal_64bits(eth->h_dest, 5461 skb->dev->dev_addr)) || 5462 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) { 5463 __skb_push(skb, ETH_HLEN); 5464 skb->pkt_type = PACKET_HOST; 5465 skb->protocol = eth_type_trans(skb, skb->dev); 5466 } 5467 5468 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull 5469 * before calling us again on redirect path. We do not call do_redirect 5470 * as we leave that up to the caller. 5471 * 5472 * Caller is responsible for managing lifetime of skb (i.e. calling 5473 * kfree_skb in response to actions it cannot handle/XDP_DROP). 5474 */ 5475 switch (act) { 5476 case XDP_REDIRECT: 5477 case XDP_TX: 5478 __skb_push(skb, mac_len); 5479 break; 5480 case XDP_PASS: 5481 metalen = xdp->data - xdp->data_meta; 5482 if (metalen) 5483 skb_metadata_set(skb, metalen); 5484 break; 5485 } 5486 5487 return act; 5488 } 5489 5490 static int 5491 netif_skb_check_for_xdp(struct sk_buff **pskb, const struct bpf_prog *prog) 5492 { 5493 struct sk_buff *skb = *pskb; 5494 int err, hroom, troom; 5495 5496 local_lock_nested_bh(&system_page_pool.bh_lock); 5497 err = skb_cow_data_for_xdp(this_cpu_read(system_page_pool.pool), pskb, prog); 5498 local_unlock_nested_bh(&system_page_pool.bh_lock); 5499 if (!err) 5500 return 0; 5501 5502 /* In case we have to go down the path and also linearize, 5503 * then lets do the pskb_expand_head() work just once here. 5504 */ 5505 hroom = XDP_PACKET_HEADROOM - skb_headroom(skb); 5506 troom = skb->tail + skb->data_len - skb->end; 5507 err = pskb_expand_head(skb, 5508 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0, 5509 troom > 0 ? troom + 128 : 0, GFP_ATOMIC); 5510 if (err) 5511 return err; 5512 5513 return skb_linearize(skb); 5514 } 5515 5516 static u32 netif_receive_generic_xdp(struct sk_buff **pskb, 5517 struct xdp_buff *xdp, 5518 const struct bpf_prog *xdp_prog) 5519 { 5520 struct sk_buff *skb = *pskb; 5521 u32 mac_len, act = XDP_DROP; 5522 5523 /* Reinjected packets coming from act_mirred or similar should 5524 * not get XDP generic processing. 5525 */ 5526 if (skb_is_redirected(skb)) 5527 return XDP_PASS; 5528 5529 /* XDP packets must have sufficient headroom of XDP_PACKET_HEADROOM 5530 * bytes. This is the guarantee that also native XDP provides, 5531 * thus we need to do it here as well. 5532 */ 5533 mac_len = skb->data - skb_mac_header(skb); 5534 __skb_push(skb, mac_len); 5535 5536 if (skb_cloned(skb) || skb_is_nonlinear(skb) || 5537 skb_headroom(skb) < XDP_PACKET_HEADROOM) { 5538 if (netif_skb_check_for_xdp(pskb, xdp_prog)) 5539 goto do_drop; 5540 } 5541 5542 __skb_pull(*pskb, mac_len); 5543 5544 act = bpf_prog_run_generic_xdp(*pskb, xdp, xdp_prog); 5545 switch (act) { 5546 case XDP_REDIRECT: 5547 case XDP_TX: 5548 case XDP_PASS: 5549 break; 5550 default: 5551 bpf_warn_invalid_xdp_action((*pskb)->dev, xdp_prog, act); 5552 fallthrough; 5553 case XDP_ABORTED: 5554 trace_xdp_exception((*pskb)->dev, xdp_prog, act); 5555 fallthrough; 5556 case XDP_DROP: 5557 do_drop: 5558 kfree_skb(*pskb); 5559 break; 5560 } 5561 5562 return act; 5563 } 5564 5565 /* When doing generic XDP we have to bypass the qdisc layer and the 5566 * network taps in order to match in-driver-XDP behavior. This also means 5567 * that XDP packets are able to starve other packets going through a qdisc, 5568 * and DDOS attacks will be more effective. In-driver-XDP use dedicated TX 5569 * queues, so they do not have this starvation issue. 5570 */ 5571 void generic_xdp_tx(struct sk_buff *skb, const struct bpf_prog *xdp_prog) 5572 { 5573 struct net_device *dev = skb->dev; 5574 struct netdev_queue *txq; 5575 bool free_skb = true; 5576 int cpu, rc; 5577 5578 txq = netdev_core_pick_tx(dev, skb, NULL); 5579 cpu = smp_processor_id(); 5580 HARD_TX_LOCK(dev, txq, cpu); 5581 if (!netif_xmit_frozen_or_drv_stopped(txq)) { 5582 rc = netdev_start_xmit(skb, dev, txq, 0); 5583 if (dev_xmit_complete(rc)) 5584 free_skb = false; 5585 } 5586 HARD_TX_UNLOCK(dev, txq); 5587 if (free_skb) { 5588 trace_xdp_exception(dev, xdp_prog, XDP_TX); 5589 dev_core_stats_tx_dropped_inc(dev); 5590 kfree_skb(skb); 5591 } 5592 } 5593 5594 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key); 5595 5596 int do_xdp_generic(const struct bpf_prog *xdp_prog, struct sk_buff **pskb) 5597 { 5598 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 5599 5600 if (xdp_prog) { 5601 struct xdp_buff xdp; 5602 u32 act; 5603 int err; 5604 5605 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 5606 act = netif_receive_generic_xdp(pskb, &xdp, xdp_prog); 5607 if (act != XDP_PASS) { 5608 switch (act) { 5609 case XDP_REDIRECT: 5610 err = xdp_do_generic_redirect((*pskb)->dev, *pskb, 5611 &xdp, xdp_prog); 5612 if (err) 5613 goto out_redir; 5614 break; 5615 case XDP_TX: 5616 generic_xdp_tx(*pskb, xdp_prog); 5617 break; 5618 } 5619 bpf_net_ctx_clear(bpf_net_ctx); 5620 return XDP_DROP; 5621 } 5622 bpf_net_ctx_clear(bpf_net_ctx); 5623 } 5624 return XDP_PASS; 5625 out_redir: 5626 bpf_net_ctx_clear(bpf_net_ctx); 5627 kfree_skb_reason(*pskb, SKB_DROP_REASON_XDP); 5628 return XDP_DROP; 5629 } 5630 EXPORT_SYMBOL_GPL(do_xdp_generic); 5631 5632 static int netif_rx_internal(struct sk_buff *skb) 5633 { 5634 int ret; 5635 5636 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5637 5638 trace_netif_rx(skb); 5639 5640 #ifdef CONFIG_RPS 5641 if (static_branch_unlikely(&rps_needed)) { 5642 struct rps_dev_flow voidflow, *rflow = &voidflow; 5643 int cpu; 5644 5645 rcu_read_lock(); 5646 5647 cpu = get_rps_cpu(skb->dev, skb, &rflow); 5648 if (cpu < 0) 5649 cpu = smp_processor_id(); 5650 5651 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5652 5653 rcu_read_unlock(); 5654 } else 5655 #endif 5656 { 5657 unsigned int qtail; 5658 5659 ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail); 5660 } 5661 return ret; 5662 } 5663 5664 /** 5665 * __netif_rx - Slightly optimized version of netif_rx 5666 * @skb: buffer to post 5667 * 5668 * This behaves as netif_rx except that it does not disable bottom halves. 5669 * As a result this function may only be invoked from the interrupt context 5670 * (either hard or soft interrupt). 5671 */ 5672 int __netif_rx(struct sk_buff *skb) 5673 { 5674 int ret; 5675 5676 lockdep_assert_once(hardirq_count() | softirq_count()); 5677 5678 trace_netif_rx_entry(skb); 5679 ret = netif_rx_internal(skb); 5680 trace_netif_rx_exit(ret); 5681 return ret; 5682 } 5683 EXPORT_SYMBOL(__netif_rx); 5684 5685 /** 5686 * netif_rx - post buffer to the network code 5687 * @skb: buffer to post 5688 * 5689 * This function receives a packet from a device driver and queues it for 5690 * the upper (protocol) levels to process via the backlog NAPI device. It 5691 * always succeeds. The buffer may be dropped during processing for 5692 * congestion control or by the protocol layers. 5693 * The network buffer is passed via the backlog NAPI device. Modern NIC 5694 * driver should use NAPI and GRO. 5695 * This function can used from interrupt and from process context. The 5696 * caller from process context must not disable interrupts before invoking 5697 * this function. 5698 * 5699 * return values: 5700 * NET_RX_SUCCESS (no congestion) 5701 * NET_RX_DROP (packet was dropped) 5702 * 5703 */ 5704 int netif_rx(struct sk_buff *skb) 5705 { 5706 bool need_bh_off = !(hardirq_count() | softirq_count()); 5707 int ret; 5708 5709 if (need_bh_off) 5710 local_bh_disable(); 5711 trace_netif_rx_entry(skb); 5712 ret = netif_rx_internal(skb); 5713 trace_netif_rx_exit(ret); 5714 if (need_bh_off) 5715 local_bh_enable(); 5716 return ret; 5717 } 5718 EXPORT_SYMBOL(netif_rx); 5719 5720 static __latent_entropy void net_tx_action(void) 5721 { 5722 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 5723 5724 if (sd->completion_queue) { 5725 struct sk_buff *clist; 5726 5727 local_irq_disable(); 5728 clist = sd->completion_queue; 5729 sd->completion_queue = NULL; 5730 local_irq_enable(); 5731 5732 while (clist) { 5733 struct sk_buff *skb = clist; 5734 5735 clist = clist->next; 5736 5737 WARN_ON(refcount_read(&skb->users)); 5738 if (likely(get_kfree_skb_cb(skb)->reason == SKB_CONSUMED)) 5739 trace_consume_skb(skb, net_tx_action); 5740 else 5741 trace_kfree_skb(skb, net_tx_action, 5742 get_kfree_skb_cb(skb)->reason, NULL); 5743 5744 if (skb->fclone != SKB_FCLONE_UNAVAILABLE) 5745 __kfree_skb(skb); 5746 else 5747 __napi_kfree_skb(skb, 5748 get_kfree_skb_cb(skb)->reason); 5749 } 5750 } 5751 5752 if (sd->output_queue) { 5753 struct Qdisc *head; 5754 5755 local_irq_disable(); 5756 head = sd->output_queue; 5757 sd->output_queue = NULL; 5758 sd->output_queue_tailp = &sd->output_queue; 5759 local_irq_enable(); 5760 5761 rcu_read_lock(); 5762 5763 while (head) { 5764 spinlock_t *root_lock = NULL; 5765 struct sk_buff *to_free; 5766 struct Qdisc *q = head; 5767 5768 head = head->next_sched; 5769 5770 /* We need to make sure head->next_sched is read 5771 * before clearing __QDISC_STATE_SCHED 5772 */ 5773 smp_mb__before_atomic(); 5774 5775 if (!(q->flags & TCQ_F_NOLOCK)) { 5776 root_lock = qdisc_lock(q); 5777 spin_lock(root_lock); 5778 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, 5779 &q->state))) { 5780 /* There is a synchronize_net() between 5781 * STATE_DEACTIVATED flag being set and 5782 * qdisc_reset()/some_qdisc_is_busy() in 5783 * dev_deactivate(), so we can safely bail out 5784 * early here to avoid data race between 5785 * qdisc_deactivate() and some_qdisc_is_busy() 5786 * for lockless qdisc. 5787 */ 5788 clear_bit(__QDISC_STATE_SCHED, &q->state); 5789 continue; 5790 } 5791 5792 clear_bit(__QDISC_STATE_SCHED, &q->state); 5793 to_free = qdisc_run(q); 5794 if (root_lock) 5795 spin_unlock(root_lock); 5796 tcf_kfree_skb_list(to_free); 5797 } 5798 5799 rcu_read_unlock(); 5800 } 5801 5802 xfrm_dev_backlog(sd); 5803 } 5804 5805 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE) 5806 /* This hook is defined here for ATM LANE */ 5807 int (*br_fdb_test_addr_hook)(struct net_device *dev, 5808 unsigned char *addr) __read_mostly; 5809 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook); 5810 #endif 5811 5812 /** 5813 * netdev_is_rx_handler_busy - check if receive handler is registered 5814 * @dev: device to check 5815 * 5816 * Check if a receive handler is already registered for a given device. 5817 * Return true if there one. 5818 * 5819 * The caller must hold the rtnl_mutex. 5820 */ 5821 bool netdev_is_rx_handler_busy(struct net_device *dev) 5822 { 5823 ASSERT_RTNL(); 5824 return dev && rtnl_dereference(dev->rx_handler); 5825 } 5826 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy); 5827 5828 /** 5829 * netdev_rx_handler_register - register receive handler 5830 * @dev: device to register a handler for 5831 * @rx_handler: receive handler to register 5832 * @rx_handler_data: data pointer that is used by rx handler 5833 * 5834 * Register a receive handler for a device. This handler will then be 5835 * called from __netif_receive_skb. A negative errno code is returned 5836 * on a failure. 5837 * 5838 * The caller must hold the rtnl_mutex. 5839 * 5840 * For a general description of rx_handler, see enum rx_handler_result. 5841 */ 5842 int netdev_rx_handler_register(struct net_device *dev, 5843 rx_handler_func_t *rx_handler, 5844 void *rx_handler_data) 5845 { 5846 if (netdev_is_rx_handler_busy(dev)) 5847 return -EBUSY; 5848 5849 if (dev->priv_flags & IFF_NO_RX_HANDLER) 5850 return -EINVAL; 5851 5852 /* Note: rx_handler_data must be set before rx_handler */ 5853 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data); 5854 rcu_assign_pointer(dev->rx_handler, rx_handler); 5855 5856 return 0; 5857 } 5858 EXPORT_SYMBOL_GPL(netdev_rx_handler_register); 5859 5860 /** 5861 * netdev_rx_handler_unregister - unregister receive handler 5862 * @dev: device to unregister a handler from 5863 * 5864 * Unregister a receive handler from a device. 5865 * 5866 * The caller must hold the rtnl_mutex. 5867 */ 5868 void netdev_rx_handler_unregister(struct net_device *dev) 5869 { 5870 5871 ASSERT_RTNL(); 5872 RCU_INIT_POINTER(dev->rx_handler, NULL); 5873 /* a reader seeing a non NULL rx_handler in a rcu_read_lock() 5874 * section has a guarantee to see a non NULL rx_handler_data 5875 * as well. 5876 */ 5877 synchronize_net(); 5878 RCU_INIT_POINTER(dev->rx_handler_data, NULL); 5879 } 5880 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister); 5881 5882 /* 5883 * Limit the use of PFMEMALLOC reserves to those protocols that implement 5884 * the special handling of PFMEMALLOC skbs. 5885 */ 5886 static bool skb_pfmemalloc_protocol(struct sk_buff *skb) 5887 { 5888 switch (skb->protocol) { 5889 case htons(ETH_P_ARP): 5890 case htons(ETH_P_IP): 5891 case htons(ETH_P_IPV6): 5892 case htons(ETH_P_8021Q): 5893 case htons(ETH_P_8021AD): 5894 return true; 5895 default: 5896 return false; 5897 } 5898 } 5899 5900 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev, 5901 int *ret, struct net_device *orig_dev) 5902 { 5903 if (nf_hook_ingress_active(skb)) { 5904 int ingress_retval; 5905 5906 if (unlikely(*pt_prev)) { 5907 *ret = deliver_skb(skb, *pt_prev, orig_dev); 5908 *pt_prev = NULL; 5909 } 5910 5911 rcu_read_lock(); 5912 ingress_retval = nf_hook_ingress(skb); 5913 rcu_read_unlock(); 5914 return ingress_retval; 5915 } 5916 return 0; 5917 } 5918 5919 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc, 5920 struct packet_type **ppt_prev) 5921 { 5922 enum skb_drop_reason drop_reason = SKB_DROP_REASON_UNHANDLED_PROTO; 5923 struct packet_type *ptype, *pt_prev; 5924 rx_handler_func_t *rx_handler; 5925 struct sk_buff *skb = *pskb; 5926 struct net_device *orig_dev; 5927 bool deliver_exact = false; 5928 int ret = NET_RX_DROP; 5929 __be16 type; 5930 5931 net_timestamp_check(!READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5932 5933 trace_netif_receive_skb(skb); 5934 5935 orig_dev = skb->dev; 5936 5937 skb_reset_network_header(skb); 5938 #if !defined(CONFIG_DEBUG_NET) 5939 /* We plan to no longer reset the transport header here. 5940 * Give some time to fuzzers and dev build to catch bugs 5941 * in network stacks. 5942 */ 5943 if (!skb_transport_header_was_set(skb)) 5944 skb_reset_transport_header(skb); 5945 #endif 5946 skb_reset_mac_len(skb); 5947 5948 pt_prev = NULL; 5949 5950 another_round: 5951 skb->skb_iif = skb->dev->ifindex; 5952 5953 __this_cpu_inc(softnet_data.processed); 5954 5955 if (static_branch_unlikely(&generic_xdp_needed_key)) { 5956 int ret2; 5957 5958 migrate_disable(); 5959 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), 5960 &skb); 5961 migrate_enable(); 5962 5963 if (ret2 != XDP_PASS) { 5964 ret = NET_RX_DROP; 5965 goto out; 5966 } 5967 } 5968 5969 if (eth_type_vlan(skb->protocol)) { 5970 skb = skb_vlan_untag(skb); 5971 if (unlikely(!skb)) 5972 goto out; 5973 } 5974 5975 if (skb_skip_tc_classify(skb)) 5976 goto skip_classify; 5977 5978 if (pfmemalloc) 5979 goto skip_taps; 5980 5981 list_for_each_entry_rcu(ptype, &dev_net_rcu(skb->dev)->ptype_all, 5982 list) { 5983 if (unlikely(pt_prev)) 5984 ret = deliver_skb(skb, pt_prev, orig_dev); 5985 pt_prev = ptype; 5986 } 5987 5988 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) { 5989 if (unlikely(pt_prev)) 5990 ret = deliver_skb(skb, pt_prev, orig_dev); 5991 pt_prev = ptype; 5992 } 5993 5994 skip_taps: 5995 #ifdef CONFIG_NET_INGRESS 5996 if (static_branch_unlikely(&ingress_needed_key)) { 5997 bool another = false; 5998 5999 nf_skip_egress(skb, true); 6000 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev, 6001 &another); 6002 if (another) 6003 goto another_round; 6004 if (!skb) 6005 goto out; 6006 6007 nf_skip_egress(skb, false); 6008 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0) 6009 goto out; 6010 } 6011 #endif 6012 skb_reset_redirect(skb); 6013 skip_classify: 6014 if (pfmemalloc && !skb_pfmemalloc_protocol(skb)) { 6015 drop_reason = SKB_DROP_REASON_PFMEMALLOC; 6016 goto drop; 6017 } 6018 6019 if (skb_vlan_tag_present(skb)) { 6020 if (unlikely(pt_prev)) { 6021 ret = deliver_skb(skb, pt_prev, orig_dev); 6022 pt_prev = NULL; 6023 } 6024 if (vlan_do_receive(&skb)) 6025 goto another_round; 6026 else if (unlikely(!skb)) 6027 goto out; 6028 } 6029 6030 rx_handler = rcu_dereference(skb->dev->rx_handler); 6031 if (rx_handler) { 6032 if (unlikely(pt_prev)) { 6033 ret = deliver_skb(skb, pt_prev, orig_dev); 6034 pt_prev = NULL; 6035 } 6036 switch (rx_handler(&skb)) { 6037 case RX_HANDLER_CONSUMED: 6038 ret = NET_RX_SUCCESS; 6039 goto out; 6040 case RX_HANDLER_ANOTHER: 6041 goto another_round; 6042 case RX_HANDLER_EXACT: 6043 deliver_exact = true; 6044 break; 6045 case RX_HANDLER_PASS: 6046 break; 6047 default: 6048 BUG(); 6049 } 6050 } 6051 6052 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) { 6053 check_vlan_id: 6054 if (skb_vlan_tag_get_id(skb)) { 6055 /* Vlan id is non 0 and vlan_do_receive() above couldn't 6056 * find vlan device. 6057 */ 6058 skb->pkt_type = PACKET_OTHERHOST; 6059 } else if (eth_type_vlan(skb->protocol)) { 6060 /* Outer header is 802.1P with vlan 0, inner header is 6061 * 802.1Q or 802.1AD and vlan_do_receive() above could 6062 * not find vlan dev for vlan id 0. 6063 */ 6064 __vlan_hwaccel_clear_tag(skb); 6065 skb = skb_vlan_untag(skb); 6066 if (unlikely(!skb)) 6067 goto out; 6068 if (vlan_do_receive(&skb)) 6069 /* After stripping off 802.1P header with vlan 0 6070 * vlan dev is found for inner header. 6071 */ 6072 goto another_round; 6073 else if (unlikely(!skb)) 6074 goto out; 6075 else 6076 /* We have stripped outer 802.1P vlan 0 header. 6077 * But could not find vlan dev. 6078 * check again for vlan id to set OTHERHOST. 6079 */ 6080 goto check_vlan_id; 6081 } 6082 /* Note: we might in the future use prio bits 6083 * and set skb->priority like in vlan_do_receive() 6084 * For the time being, just ignore Priority Code Point 6085 */ 6086 __vlan_hwaccel_clear_tag(skb); 6087 } 6088 6089 type = skb->protocol; 6090 6091 /* deliver only exact match when indicated */ 6092 if (likely(!deliver_exact)) { 6093 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6094 &ptype_base[ntohs(type) & 6095 PTYPE_HASH_MASK]); 6096 6097 /* orig_dev and skb->dev could belong to different netns; 6098 * Even in such case we need to traverse only the list 6099 * coming from skb->dev, as the ptype owner (packet socket) 6100 * will use dev_net(skb->dev) to do namespace filtering. 6101 */ 6102 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6103 &dev_net_rcu(skb->dev)->ptype_specific); 6104 } 6105 6106 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6107 &orig_dev->ptype_specific); 6108 6109 if (unlikely(skb->dev != orig_dev)) { 6110 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6111 &skb->dev->ptype_specific); 6112 } 6113 6114 if (pt_prev) { 6115 *ppt_prev = pt_prev; 6116 } else { 6117 drop: 6118 if (!deliver_exact) 6119 dev_core_stats_rx_dropped_inc(skb->dev); 6120 else 6121 dev_core_stats_rx_nohandler_inc(skb->dev); 6122 6123 kfree_skb_reason(skb, drop_reason); 6124 /* Jamal, now you will not able to escape explaining 6125 * me how you were going to use this. :-) 6126 */ 6127 ret = NET_RX_DROP; 6128 } 6129 6130 out: 6131 /* The invariant here is that if *ppt_prev is not NULL 6132 * then skb should also be non-NULL. 6133 * 6134 * Apparently *ppt_prev assignment above holds this invariant due to 6135 * skb dereferencing near it. 6136 */ 6137 *pskb = skb; 6138 return ret; 6139 } 6140 6141 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc) 6142 { 6143 struct net_device *orig_dev = skb->dev; 6144 struct packet_type *pt_prev = NULL; 6145 int ret; 6146 6147 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 6148 if (pt_prev) 6149 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb, 6150 skb->dev, pt_prev, orig_dev); 6151 return ret; 6152 } 6153 6154 /** 6155 * netif_receive_skb_core - special purpose version of netif_receive_skb 6156 * @skb: buffer to process 6157 * 6158 * More direct receive version of netif_receive_skb(). It should 6159 * only be used by callers that have a need to skip RPS and Generic XDP. 6160 * Caller must also take care of handling if ``(page_is_)pfmemalloc``. 6161 * 6162 * This function may only be called from softirq context and interrupts 6163 * should be enabled. 6164 * 6165 * Return values (usually ignored): 6166 * NET_RX_SUCCESS: no congestion 6167 * NET_RX_DROP: packet was dropped 6168 */ 6169 int netif_receive_skb_core(struct sk_buff *skb) 6170 { 6171 int ret; 6172 6173 rcu_read_lock(); 6174 ret = __netif_receive_skb_one_core(skb, false); 6175 rcu_read_unlock(); 6176 6177 return ret; 6178 } 6179 EXPORT_SYMBOL(netif_receive_skb_core); 6180 6181 static inline void __netif_receive_skb_list_ptype(struct list_head *head, 6182 struct packet_type *pt_prev, 6183 struct net_device *orig_dev) 6184 { 6185 struct sk_buff *skb, *next; 6186 6187 if (!pt_prev) 6188 return; 6189 if (list_empty(head)) 6190 return; 6191 if (pt_prev->list_func != NULL) 6192 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv, 6193 ip_list_rcv, head, pt_prev, orig_dev); 6194 else 6195 list_for_each_entry_safe(skb, next, head, list) { 6196 skb_list_del_init(skb); 6197 pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 6198 } 6199 } 6200 6201 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc) 6202 { 6203 /* Fast-path assumptions: 6204 * - There is no RX handler. 6205 * - Only one packet_type matches. 6206 * If either of these fails, we will end up doing some per-packet 6207 * processing in-line, then handling the 'last ptype' for the whole 6208 * sublist. This can't cause out-of-order delivery to any single ptype, 6209 * because the 'last ptype' must be constant across the sublist, and all 6210 * other ptypes are handled per-packet. 6211 */ 6212 /* Current (common) ptype of sublist */ 6213 struct packet_type *pt_curr = NULL; 6214 /* Current (common) orig_dev of sublist */ 6215 struct net_device *od_curr = NULL; 6216 struct sk_buff *skb, *next; 6217 LIST_HEAD(sublist); 6218 6219 list_for_each_entry_safe(skb, next, head, list) { 6220 struct net_device *orig_dev = skb->dev; 6221 struct packet_type *pt_prev = NULL; 6222 6223 skb_list_del_init(skb); 6224 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 6225 if (!pt_prev) 6226 continue; 6227 if (pt_curr != pt_prev || od_curr != orig_dev) { 6228 /* dispatch old sublist */ 6229 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 6230 /* start new sublist */ 6231 INIT_LIST_HEAD(&sublist); 6232 pt_curr = pt_prev; 6233 od_curr = orig_dev; 6234 } 6235 list_add_tail(&skb->list, &sublist); 6236 } 6237 6238 /* dispatch final sublist */ 6239 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 6240 } 6241 6242 static int __netif_receive_skb(struct sk_buff *skb) 6243 { 6244 int ret; 6245 6246 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) { 6247 unsigned int noreclaim_flag; 6248 6249 /* 6250 * PFMEMALLOC skbs are special, they should 6251 * - be delivered to SOCK_MEMALLOC sockets only 6252 * - stay away from userspace 6253 * - have bounded memory usage 6254 * 6255 * Use PF_MEMALLOC as this saves us from propagating the allocation 6256 * context down to all allocation sites. 6257 */ 6258 noreclaim_flag = memalloc_noreclaim_save(); 6259 ret = __netif_receive_skb_one_core(skb, true); 6260 memalloc_noreclaim_restore(noreclaim_flag); 6261 } else 6262 ret = __netif_receive_skb_one_core(skb, false); 6263 6264 return ret; 6265 } 6266 6267 static void __netif_receive_skb_list(struct list_head *head) 6268 { 6269 unsigned long noreclaim_flag = 0; 6270 struct sk_buff *skb, *next; 6271 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */ 6272 6273 list_for_each_entry_safe(skb, next, head, list) { 6274 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) { 6275 struct list_head sublist; 6276 6277 /* Handle the previous sublist */ 6278 list_cut_before(&sublist, head, &skb->list); 6279 if (!list_empty(&sublist)) 6280 __netif_receive_skb_list_core(&sublist, pfmemalloc); 6281 pfmemalloc = !pfmemalloc; 6282 /* See comments in __netif_receive_skb */ 6283 if (pfmemalloc) 6284 noreclaim_flag = memalloc_noreclaim_save(); 6285 else 6286 memalloc_noreclaim_restore(noreclaim_flag); 6287 } 6288 } 6289 /* Handle the remaining sublist */ 6290 if (!list_empty(head)) 6291 __netif_receive_skb_list_core(head, pfmemalloc); 6292 /* Restore pflags */ 6293 if (pfmemalloc) 6294 memalloc_noreclaim_restore(noreclaim_flag); 6295 } 6296 6297 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp) 6298 { 6299 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog); 6300 struct bpf_prog *new = xdp->prog; 6301 int ret = 0; 6302 6303 switch (xdp->command) { 6304 case XDP_SETUP_PROG: 6305 rcu_assign_pointer(dev->xdp_prog, new); 6306 if (old) 6307 bpf_prog_put(old); 6308 6309 if (old && !new) { 6310 static_branch_dec(&generic_xdp_needed_key); 6311 } else if (new && !old) { 6312 static_branch_inc(&generic_xdp_needed_key); 6313 netif_disable_lro(dev); 6314 dev_disable_gro_hw(dev); 6315 } 6316 break; 6317 6318 default: 6319 ret = -EINVAL; 6320 break; 6321 } 6322 6323 return ret; 6324 } 6325 6326 static int netif_receive_skb_internal(struct sk_buff *skb) 6327 { 6328 int ret; 6329 6330 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 6331 6332 if (skb_defer_rx_timestamp(skb)) 6333 return NET_RX_SUCCESS; 6334 6335 rcu_read_lock(); 6336 #ifdef CONFIG_RPS 6337 if (static_branch_unlikely(&rps_needed)) { 6338 struct rps_dev_flow voidflow, *rflow = &voidflow; 6339 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6340 6341 if (cpu >= 0) { 6342 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6343 rcu_read_unlock(); 6344 return ret; 6345 } 6346 } 6347 #endif 6348 ret = __netif_receive_skb(skb); 6349 rcu_read_unlock(); 6350 return ret; 6351 } 6352 6353 void netif_receive_skb_list_internal(struct list_head *head) 6354 { 6355 struct sk_buff *skb, *next; 6356 LIST_HEAD(sublist); 6357 6358 list_for_each_entry_safe(skb, next, head, list) { 6359 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), 6360 skb); 6361 skb_list_del_init(skb); 6362 if (!skb_defer_rx_timestamp(skb)) 6363 list_add_tail(&skb->list, &sublist); 6364 } 6365 list_splice_init(&sublist, head); 6366 6367 rcu_read_lock(); 6368 #ifdef CONFIG_RPS 6369 if (static_branch_unlikely(&rps_needed)) { 6370 list_for_each_entry_safe(skb, next, head, list) { 6371 struct rps_dev_flow voidflow, *rflow = &voidflow; 6372 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6373 6374 if (cpu >= 0) { 6375 /* Will be handled, remove from list */ 6376 skb_list_del_init(skb); 6377 enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6378 } 6379 } 6380 } 6381 #endif 6382 __netif_receive_skb_list(head); 6383 rcu_read_unlock(); 6384 } 6385 6386 /** 6387 * netif_receive_skb - process receive buffer from network 6388 * @skb: buffer to process 6389 * 6390 * netif_receive_skb() is the main receive data processing function. 6391 * It always succeeds. The buffer may be dropped during processing 6392 * for congestion control or by the protocol layers. 6393 * 6394 * This function may only be called from softirq context and interrupts 6395 * should be enabled. 6396 * 6397 * Return values (usually ignored): 6398 * NET_RX_SUCCESS: no congestion 6399 * NET_RX_DROP: packet was dropped 6400 */ 6401 int netif_receive_skb(struct sk_buff *skb) 6402 { 6403 int ret; 6404 6405 trace_netif_receive_skb_entry(skb); 6406 6407 ret = netif_receive_skb_internal(skb); 6408 trace_netif_receive_skb_exit(ret); 6409 6410 return ret; 6411 } 6412 EXPORT_SYMBOL(netif_receive_skb); 6413 6414 /** 6415 * netif_receive_skb_list - process many receive buffers from network 6416 * @head: list of skbs to process. 6417 * 6418 * Since return value of netif_receive_skb() is normally ignored, and 6419 * wouldn't be meaningful for a list, this function returns void. 6420 * 6421 * This function may only be called from softirq context and interrupts 6422 * should be enabled. 6423 */ 6424 void netif_receive_skb_list(struct list_head *head) 6425 { 6426 struct sk_buff *skb; 6427 6428 if (list_empty(head)) 6429 return; 6430 if (trace_netif_receive_skb_list_entry_enabled()) { 6431 list_for_each_entry(skb, head, list) 6432 trace_netif_receive_skb_list_entry(skb); 6433 } 6434 netif_receive_skb_list_internal(head); 6435 trace_netif_receive_skb_list_exit(0); 6436 } 6437 EXPORT_SYMBOL(netif_receive_skb_list); 6438 6439 /* Network device is going away, flush any packets still pending */ 6440 static void flush_backlog(struct work_struct *work) 6441 { 6442 struct sk_buff *skb, *tmp; 6443 struct sk_buff_head list; 6444 struct softnet_data *sd; 6445 6446 __skb_queue_head_init(&list); 6447 local_bh_disable(); 6448 sd = this_cpu_ptr(&softnet_data); 6449 6450 backlog_lock_irq_disable(sd); 6451 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) { 6452 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6453 __skb_unlink(skb, &sd->input_pkt_queue); 6454 __skb_queue_tail(&list, skb); 6455 rps_input_queue_head_incr(sd); 6456 } 6457 } 6458 backlog_unlock_irq_enable(sd); 6459 6460 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6461 skb_queue_walk_safe(&sd->process_queue, skb, tmp) { 6462 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6463 __skb_unlink(skb, &sd->process_queue); 6464 __skb_queue_tail(&list, skb); 6465 rps_input_queue_head_incr(sd); 6466 } 6467 } 6468 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6469 local_bh_enable(); 6470 6471 __skb_queue_purge_reason(&list, SKB_DROP_REASON_DEV_READY); 6472 } 6473 6474 static bool flush_required(int cpu) 6475 { 6476 #if IS_ENABLED(CONFIG_RPS) 6477 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 6478 bool do_flush; 6479 6480 backlog_lock_irq_disable(sd); 6481 6482 /* as insertion into process_queue happens with the rps lock held, 6483 * process_queue access may race only with dequeue 6484 */ 6485 do_flush = !skb_queue_empty(&sd->input_pkt_queue) || 6486 !skb_queue_empty_lockless(&sd->process_queue); 6487 backlog_unlock_irq_enable(sd); 6488 6489 return do_flush; 6490 #endif 6491 /* without RPS we can't safely check input_pkt_queue: during a 6492 * concurrent remote skb_queue_splice() we can detect as empty both 6493 * input_pkt_queue and process_queue even if the latter could end-up 6494 * containing a lot of packets. 6495 */ 6496 return true; 6497 } 6498 6499 struct flush_backlogs { 6500 cpumask_t flush_cpus; 6501 struct work_struct w[]; 6502 }; 6503 6504 static struct flush_backlogs *flush_backlogs_alloc(void) 6505 { 6506 return kmalloc(struct_size_t(struct flush_backlogs, w, nr_cpu_ids), 6507 GFP_KERNEL); 6508 } 6509 6510 static struct flush_backlogs *flush_backlogs_fallback; 6511 static DEFINE_MUTEX(flush_backlogs_mutex); 6512 6513 static void flush_all_backlogs(void) 6514 { 6515 struct flush_backlogs *ptr = flush_backlogs_alloc(); 6516 unsigned int cpu; 6517 6518 if (!ptr) { 6519 mutex_lock(&flush_backlogs_mutex); 6520 ptr = flush_backlogs_fallback; 6521 } 6522 cpumask_clear(&ptr->flush_cpus); 6523 6524 cpus_read_lock(); 6525 6526 for_each_online_cpu(cpu) { 6527 if (flush_required(cpu)) { 6528 INIT_WORK(&ptr->w[cpu], flush_backlog); 6529 queue_work_on(cpu, system_highpri_wq, &ptr->w[cpu]); 6530 __cpumask_set_cpu(cpu, &ptr->flush_cpus); 6531 } 6532 } 6533 6534 /* we can have in flight packet[s] on the cpus we are not flushing, 6535 * synchronize_net() in unregister_netdevice_many() will take care of 6536 * them. 6537 */ 6538 for_each_cpu(cpu, &ptr->flush_cpus) 6539 flush_work(&ptr->w[cpu]); 6540 6541 cpus_read_unlock(); 6542 6543 if (ptr != flush_backlogs_fallback) 6544 kfree(ptr); 6545 else 6546 mutex_unlock(&flush_backlogs_mutex); 6547 } 6548 6549 static void net_rps_send_ipi(struct softnet_data *remsd) 6550 { 6551 #ifdef CONFIG_RPS 6552 while (remsd) { 6553 struct softnet_data *next = remsd->rps_ipi_next; 6554 6555 if (cpu_online(remsd->cpu)) 6556 smp_call_function_single_async(remsd->cpu, &remsd->csd); 6557 remsd = next; 6558 } 6559 #endif 6560 } 6561 6562 /* 6563 * net_rps_action_and_irq_enable sends any pending IPI's for rps. 6564 * Note: called with local irq disabled, but exits with local irq enabled. 6565 */ 6566 static void net_rps_action_and_irq_enable(struct softnet_data *sd) 6567 { 6568 #ifdef CONFIG_RPS 6569 struct softnet_data *remsd = sd->rps_ipi_list; 6570 6571 if (!use_backlog_threads() && remsd) { 6572 sd->rps_ipi_list = NULL; 6573 6574 local_irq_enable(); 6575 6576 /* Send pending IPI's to kick RPS processing on remote cpus. */ 6577 net_rps_send_ipi(remsd); 6578 } else 6579 #endif 6580 local_irq_enable(); 6581 } 6582 6583 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd) 6584 { 6585 #ifdef CONFIG_RPS 6586 return !use_backlog_threads() && sd->rps_ipi_list; 6587 #else 6588 return false; 6589 #endif 6590 } 6591 6592 static int process_backlog(struct napi_struct *napi, int quota) 6593 { 6594 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog); 6595 bool again = true; 6596 int work = 0; 6597 6598 /* Check if we have pending ipi, its better to send them now, 6599 * not waiting net_rx_action() end. 6600 */ 6601 if (sd_has_rps_ipi_waiting(sd)) { 6602 local_irq_disable(); 6603 net_rps_action_and_irq_enable(sd); 6604 } 6605 6606 napi->weight = READ_ONCE(net_hotdata.dev_rx_weight); 6607 while (again) { 6608 struct sk_buff *skb; 6609 6610 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6611 while ((skb = __skb_dequeue(&sd->process_queue))) { 6612 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6613 rcu_read_lock(); 6614 __netif_receive_skb(skb); 6615 rcu_read_unlock(); 6616 if (++work >= quota) { 6617 rps_input_queue_head_add(sd, work); 6618 return work; 6619 } 6620 6621 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6622 } 6623 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6624 6625 backlog_lock_irq_disable(sd); 6626 if (skb_queue_empty(&sd->input_pkt_queue)) { 6627 /* 6628 * Inline a custom version of __napi_complete(). 6629 * only current cpu owns and manipulates this napi, 6630 * and NAPI_STATE_SCHED is the only possible flag set 6631 * on backlog. 6632 * We can use a plain write instead of clear_bit(), 6633 * and we dont need an smp_mb() memory barrier. 6634 */ 6635 napi->state &= NAPIF_STATE_THREADED; 6636 again = false; 6637 } else { 6638 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6639 skb_queue_splice_tail_init(&sd->input_pkt_queue, 6640 &sd->process_queue); 6641 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6642 } 6643 backlog_unlock_irq_enable(sd); 6644 } 6645 6646 if (work) 6647 rps_input_queue_head_add(sd, work); 6648 return work; 6649 } 6650 6651 /** 6652 * __napi_schedule - schedule for receive 6653 * @n: entry to schedule 6654 * 6655 * The entry's receive function will be scheduled to run. 6656 * Consider using __napi_schedule_irqoff() if hard irqs are masked. 6657 */ 6658 void __napi_schedule(struct napi_struct *n) 6659 { 6660 unsigned long flags; 6661 6662 local_irq_save(flags); 6663 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6664 local_irq_restore(flags); 6665 } 6666 EXPORT_SYMBOL(__napi_schedule); 6667 6668 /** 6669 * napi_schedule_prep - check if napi can be scheduled 6670 * @n: napi context 6671 * 6672 * Test if NAPI routine is already running, and if not mark 6673 * it as running. This is used as a condition variable to 6674 * insure only one NAPI poll instance runs. We also make 6675 * sure there is no pending NAPI disable. 6676 */ 6677 bool napi_schedule_prep(struct napi_struct *n) 6678 { 6679 unsigned long new, val = READ_ONCE(n->state); 6680 6681 do { 6682 if (unlikely(val & NAPIF_STATE_DISABLE)) 6683 return false; 6684 new = val | NAPIF_STATE_SCHED; 6685 6686 /* Sets STATE_MISSED bit if STATE_SCHED was already set 6687 * This was suggested by Alexander Duyck, as compiler 6688 * emits better code than : 6689 * if (val & NAPIF_STATE_SCHED) 6690 * new |= NAPIF_STATE_MISSED; 6691 */ 6692 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED * 6693 NAPIF_STATE_MISSED; 6694 } while (!try_cmpxchg(&n->state, &val, new)); 6695 6696 return !(val & NAPIF_STATE_SCHED); 6697 } 6698 EXPORT_SYMBOL(napi_schedule_prep); 6699 6700 /** 6701 * __napi_schedule_irqoff - schedule for receive 6702 * @n: entry to schedule 6703 * 6704 * Variant of __napi_schedule() assuming hard irqs are masked. 6705 * 6706 * On PREEMPT_RT enabled kernels this maps to __napi_schedule() 6707 * because the interrupt disabled assumption might not be true 6708 * due to force-threaded interrupts and spinlock substitution. 6709 */ 6710 void __napi_schedule_irqoff(struct napi_struct *n) 6711 { 6712 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6713 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6714 else 6715 __napi_schedule(n); 6716 } 6717 EXPORT_SYMBOL(__napi_schedule_irqoff); 6718 6719 bool napi_complete_done(struct napi_struct *n, int work_done) 6720 { 6721 unsigned long flags, val, new, timeout = 0; 6722 bool ret = true; 6723 6724 /* 6725 * 1) Don't let napi dequeue from the cpu poll list 6726 * just in case its running on a different cpu. 6727 * 2) If we are busy polling, do nothing here, we have 6728 * the guarantee we will be called later. 6729 */ 6730 if (unlikely(n->state & (NAPIF_STATE_NPSVC | 6731 NAPIF_STATE_IN_BUSY_POLL))) 6732 return false; 6733 6734 if (work_done) { 6735 if (n->gro.bitmask) 6736 timeout = napi_get_gro_flush_timeout(n); 6737 n->defer_hard_irqs_count = napi_get_defer_hard_irqs(n); 6738 } 6739 if (n->defer_hard_irqs_count > 0) { 6740 n->defer_hard_irqs_count--; 6741 timeout = napi_get_gro_flush_timeout(n); 6742 if (timeout) 6743 ret = false; 6744 } 6745 6746 /* 6747 * When the NAPI instance uses a timeout and keeps postponing 6748 * it, we need to bound somehow the time packets are kept in 6749 * the GRO layer. 6750 */ 6751 gro_flush_normal(&n->gro, !!timeout); 6752 6753 if (unlikely(!list_empty(&n->poll_list))) { 6754 /* If n->poll_list is not empty, we need to mask irqs */ 6755 local_irq_save(flags); 6756 list_del_init(&n->poll_list); 6757 local_irq_restore(flags); 6758 } 6759 WRITE_ONCE(n->list_owner, -1); 6760 6761 val = READ_ONCE(n->state); 6762 do { 6763 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED)); 6764 6765 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED | 6766 NAPIF_STATE_SCHED_THREADED | 6767 NAPIF_STATE_PREFER_BUSY_POLL); 6768 6769 /* If STATE_MISSED was set, leave STATE_SCHED set, 6770 * because we will call napi->poll() one more time. 6771 * This C code was suggested by Alexander Duyck to help gcc. 6772 */ 6773 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED * 6774 NAPIF_STATE_SCHED; 6775 } while (!try_cmpxchg(&n->state, &val, new)); 6776 6777 if (unlikely(val & NAPIF_STATE_MISSED)) { 6778 __napi_schedule(n); 6779 return false; 6780 } 6781 6782 if (timeout) 6783 hrtimer_start(&n->timer, ns_to_ktime(timeout), 6784 HRTIMER_MODE_REL_PINNED); 6785 return ret; 6786 } 6787 EXPORT_SYMBOL(napi_complete_done); 6788 6789 static void skb_defer_free_flush(void) 6790 { 6791 struct llist_node *free_list; 6792 struct sk_buff *skb, *next; 6793 struct skb_defer_node *sdn; 6794 int node; 6795 6796 for_each_node(node) { 6797 sdn = this_cpu_ptr(net_hotdata.skb_defer_nodes) + node; 6798 6799 if (llist_empty(&sdn->defer_list)) 6800 continue; 6801 atomic_long_set(&sdn->defer_count, 0); 6802 free_list = llist_del_all(&sdn->defer_list); 6803 6804 llist_for_each_entry_safe(skb, next, free_list, ll_node) { 6805 prefetch(next); 6806 napi_consume_skb(skb, 1); 6807 } 6808 } 6809 } 6810 6811 #if defined(CONFIG_NET_RX_BUSY_POLL) 6812 6813 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule) 6814 { 6815 if (!skip_schedule) { 6816 gro_normal_list(&napi->gro); 6817 __napi_schedule(napi); 6818 return; 6819 } 6820 6821 /* Flush too old packets. If HZ < 1000, flush all packets */ 6822 gro_flush_normal(&napi->gro, HZ >= 1000); 6823 6824 clear_bit(NAPI_STATE_SCHED, &napi->state); 6825 } 6826 6827 enum { 6828 NAPI_F_PREFER_BUSY_POLL = 1, 6829 NAPI_F_END_ON_RESCHED = 2, 6830 }; 6831 6832 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, 6833 unsigned flags, u16 budget) 6834 { 6835 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6836 bool skip_schedule = false; 6837 unsigned long timeout; 6838 int rc; 6839 6840 /* Busy polling means there is a high chance device driver hard irq 6841 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was 6842 * set in napi_schedule_prep(). 6843 * Since we are about to call napi->poll() once more, we can safely 6844 * clear NAPI_STATE_MISSED. 6845 * 6846 * Note: x86 could use a single "lock and ..." instruction 6847 * to perform these two clear_bit() 6848 */ 6849 clear_bit(NAPI_STATE_MISSED, &napi->state); 6850 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state); 6851 6852 local_bh_disable(); 6853 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6854 6855 if (flags & NAPI_F_PREFER_BUSY_POLL) { 6856 napi->defer_hard_irqs_count = napi_get_defer_hard_irqs(napi); 6857 timeout = napi_get_gro_flush_timeout(napi); 6858 if (napi->defer_hard_irqs_count && timeout) { 6859 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED); 6860 skip_schedule = true; 6861 } 6862 } 6863 6864 /* All we really want here is to re-enable device interrupts. 6865 * Ideally, a new ndo_busy_poll_stop() could avoid another round. 6866 */ 6867 rc = napi->poll(napi, budget); 6868 /* We can't gro_normal_list() here, because napi->poll() might have 6869 * rearmed the napi (napi_complete_done()) in which case it could 6870 * already be running on another CPU. 6871 */ 6872 trace_napi_poll(napi, rc, budget); 6873 netpoll_poll_unlock(have_poll_lock); 6874 if (rc == budget) 6875 __busy_poll_stop(napi, skip_schedule); 6876 bpf_net_ctx_clear(bpf_net_ctx); 6877 local_bh_enable(); 6878 } 6879 6880 static void __napi_busy_loop(unsigned int napi_id, 6881 bool (*loop_end)(void *, unsigned long), 6882 void *loop_end_arg, unsigned flags, u16 budget) 6883 { 6884 unsigned long start_time = loop_end ? busy_loop_current_time() : 0; 6885 int (*napi_poll)(struct napi_struct *napi, int budget); 6886 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6887 void *have_poll_lock = NULL; 6888 struct napi_struct *napi; 6889 6890 WARN_ON_ONCE(!rcu_read_lock_held()); 6891 6892 restart: 6893 napi_poll = NULL; 6894 6895 napi = napi_by_id(napi_id); 6896 if (!napi) 6897 return; 6898 6899 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6900 preempt_disable(); 6901 for (;;) { 6902 int work = 0; 6903 6904 local_bh_disable(); 6905 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6906 if (!napi_poll) { 6907 unsigned long val = READ_ONCE(napi->state); 6908 6909 /* If multiple threads are competing for this napi, 6910 * we avoid dirtying napi->state as much as we can. 6911 */ 6912 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED | 6913 NAPIF_STATE_IN_BUSY_POLL)) { 6914 if (flags & NAPI_F_PREFER_BUSY_POLL) 6915 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6916 goto count; 6917 } 6918 if (cmpxchg(&napi->state, val, 6919 val | NAPIF_STATE_IN_BUSY_POLL | 6920 NAPIF_STATE_SCHED) != val) { 6921 if (flags & NAPI_F_PREFER_BUSY_POLL) 6922 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6923 goto count; 6924 } 6925 have_poll_lock = netpoll_poll_lock(napi); 6926 napi_poll = napi->poll; 6927 } 6928 work = napi_poll(napi, budget); 6929 trace_napi_poll(napi, work, budget); 6930 gro_normal_list(&napi->gro); 6931 count: 6932 if (work > 0) 6933 __NET_ADD_STATS(dev_net(napi->dev), 6934 LINUX_MIB_BUSYPOLLRXPACKETS, work); 6935 skb_defer_free_flush(); 6936 bpf_net_ctx_clear(bpf_net_ctx); 6937 local_bh_enable(); 6938 6939 if (!loop_end || loop_end(loop_end_arg, start_time)) 6940 break; 6941 6942 if (unlikely(need_resched())) { 6943 if (flags & NAPI_F_END_ON_RESCHED) 6944 break; 6945 if (napi_poll) 6946 busy_poll_stop(napi, have_poll_lock, flags, budget); 6947 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6948 preempt_enable(); 6949 rcu_read_unlock(); 6950 cond_resched(); 6951 rcu_read_lock(); 6952 if (loop_end(loop_end_arg, start_time)) 6953 return; 6954 goto restart; 6955 } 6956 cpu_relax(); 6957 } 6958 if (napi_poll) 6959 busy_poll_stop(napi, have_poll_lock, flags, budget); 6960 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6961 preempt_enable(); 6962 } 6963 6964 void napi_busy_loop_rcu(unsigned int napi_id, 6965 bool (*loop_end)(void *, unsigned long), 6966 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6967 { 6968 unsigned flags = NAPI_F_END_ON_RESCHED; 6969 6970 if (prefer_busy_poll) 6971 flags |= NAPI_F_PREFER_BUSY_POLL; 6972 6973 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6974 } 6975 6976 void napi_busy_loop(unsigned int napi_id, 6977 bool (*loop_end)(void *, unsigned long), 6978 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6979 { 6980 unsigned flags = prefer_busy_poll ? NAPI_F_PREFER_BUSY_POLL : 0; 6981 6982 rcu_read_lock(); 6983 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6984 rcu_read_unlock(); 6985 } 6986 EXPORT_SYMBOL(napi_busy_loop); 6987 6988 void napi_suspend_irqs(unsigned int napi_id) 6989 { 6990 struct napi_struct *napi; 6991 6992 rcu_read_lock(); 6993 napi = napi_by_id(napi_id); 6994 if (napi) { 6995 unsigned long timeout = napi_get_irq_suspend_timeout(napi); 6996 6997 if (timeout) 6998 hrtimer_start(&napi->timer, ns_to_ktime(timeout), 6999 HRTIMER_MODE_REL_PINNED); 7000 } 7001 rcu_read_unlock(); 7002 } 7003 7004 void napi_resume_irqs(unsigned int napi_id) 7005 { 7006 struct napi_struct *napi; 7007 7008 rcu_read_lock(); 7009 napi = napi_by_id(napi_id); 7010 if (napi) { 7011 /* If irq_suspend_timeout is set to 0 between the call to 7012 * napi_suspend_irqs and now, the original value still 7013 * determines the safety timeout as intended and napi_watchdog 7014 * will resume irq processing. 7015 */ 7016 if (napi_get_irq_suspend_timeout(napi)) { 7017 local_bh_disable(); 7018 napi_schedule(napi); 7019 local_bh_enable(); 7020 } 7021 } 7022 rcu_read_unlock(); 7023 } 7024 7025 #endif /* CONFIG_NET_RX_BUSY_POLL */ 7026 7027 static void __napi_hash_add_with_id(struct napi_struct *napi, 7028 unsigned int napi_id) 7029 { 7030 napi->gro.cached_napi_id = napi_id; 7031 7032 WRITE_ONCE(napi->napi_id, napi_id); 7033 hlist_add_head_rcu(&napi->napi_hash_node, 7034 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]); 7035 } 7036 7037 static void napi_hash_add_with_id(struct napi_struct *napi, 7038 unsigned int napi_id) 7039 { 7040 unsigned long flags; 7041 7042 spin_lock_irqsave(&napi_hash_lock, flags); 7043 WARN_ON_ONCE(napi_by_id(napi_id)); 7044 __napi_hash_add_with_id(napi, napi_id); 7045 spin_unlock_irqrestore(&napi_hash_lock, flags); 7046 } 7047 7048 static void napi_hash_add(struct napi_struct *napi) 7049 { 7050 unsigned long flags; 7051 7052 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state)) 7053 return; 7054 7055 spin_lock_irqsave(&napi_hash_lock, flags); 7056 7057 /* 0..NR_CPUS range is reserved for sender_cpu use */ 7058 do { 7059 if (unlikely(!napi_id_valid(++napi_gen_id))) 7060 napi_gen_id = MIN_NAPI_ID; 7061 } while (napi_by_id(napi_gen_id)); 7062 7063 __napi_hash_add_with_id(napi, napi_gen_id); 7064 7065 spin_unlock_irqrestore(&napi_hash_lock, flags); 7066 } 7067 7068 /* Warning : caller is responsible to make sure rcu grace period 7069 * is respected before freeing memory containing @napi 7070 */ 7071 static void napi_hash_del(struct napi_struct *napi) 7072 { 7073 unsigned long flags; 7074 7075 spin_lock_irqsave(&napi_hash_lock, flags); 7076 7077 hlist_del_init_rcu(&napi->napi_hash_node); 7078 7079 spin_unlock_irqrestore(&napi_hash_lock, flags); 7080 } 7081 7082 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer) 7083 { 7084 struct napi_struct *napi; 7085 7086 napi = container_of(timer, struct napi_struct, timer); 7087 7088 /* Note : we use a relaxed variant of napi_schedule_prep() not setting 7089 * NAPI_STATE_MISSED, since we do not react to a device IRQ. 7090 */ 7091 if (!napi_disable_pending(napi) && 7092 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) { 7093 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 7094 __napi_schedule_irqoff(napi); 7095 } 7096 7097 return HRTIMER_NORESTART; 7098 } 7099 7100 static void napi_stop_kthread(struct napi_struct *napi) 7101 { 7102 unsigned long val, new; 7103 7104 /* Wait until the napi STATE_THREADED is unset. */ 7105 while (true) { 7106 val = READ_ONCE(napi->state); 7107 7108 /* If napi kthread own this napi or the napi is idle, 7109 * STATE_THREADED can be unset here. 7110 */ 7111 if ((val & NAPIF_STATE_SCHED_THREADED) || 7112 !(val & NAPIF_STATE_SCHED)) { 7113 new = val & (~(NAPIF_STATE_THREADED | 7114 NAPIF_STATE_THREADED_BUSY_POLL)); 7115 } else { 7116 msleep(20); 7117 continue; 7118 } 7119 7120 if (try_cmpxchg(&napi->state, &val, new)) 7121 break; 7122 } 7123 7124 /* Once STATE_THREADED is unset, wait for SCHED_THREADED to be unset by 7125 * the kthread. 7126 */ 7127 while (true) { 7128 if (!test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) 7129 break; 7130 7131 msleep(20); 7132 } 7133 7134 kthread_stop(napi->thread); 7135 napi->thread = NULL; 7136 } 7137 7138 static void napi_set_threaded_state(struct napi_struct *napi, 7139 enum netdev_napi_threaded threaded_mode) 7140 { 7141 bool threaded = threaded_mode != NETDEV_NAPI_THREADED_DISABLED; 7142 bool busy_poll = threaded_mode == NETDEV_NAPI_THREADED_BUSY_POLL; 7143 7144 assign_bit(NAPI_STATE_THREADED, &napi->state, threaded); 7145 assign_bit(NAPI_STATE_THREADED_BUSY_POLL, &napi->state, busy_poll); 7146 } 7147 7148 int napi_set_threaded(struct napi_struct *napi, 7149 enum netdev_napi_threaded threaded) 7150 { 7151 if (threaded) { 7152 if (!napi->thread) { 7153 int err = napi_kthread_create(napi); 7154 7155 if (err) 7156 return err; 7157 } 7158 } 7159 7160 if (napi->config) 7161 napi->config->threaded = threaded; 7162 7163 /* Setting/unsetting threaded mode on a napi might not immediately 7164 * take effect, if the current napi instance is actively being 7165 * polled. In this case, the switch between threaded mode and 7166 * softirq mode will happen in the next round of napi_schedule(). 7167 * This should not cause hiccups/stalls to the live traffic. 7168 */ 7169 if (!threaded && napi->thread) { 7170 napi_stop_kthread(napi); 7171 } else { 7172 /* Make sure kthread is created before THREADED bit is set. */ 7173 smp_mb__before_atomic(); 7174 napi_set_threaded_state(napi, threaded); 7175 } 7176 7177 return 0; 7178 } 7179 7180 int netif_set_threaded(struct net_device *dev, 7181 enum netdev_napi_threaded threaded) 7182 { 7183 struct napi_struct *napi; 7184 int i, err = 0; 7185 7186 netdev_assert_locked_or_invisible(dev); 7187 7188 if (threaded) { 7189 list_for_each_entry(napi, &dev->napi_list, dev_list) { 7190 if (!napi->thread) { 7191 err = napi_kthread_create(napi); 7192 if (err) { 7193 threaded = NETDEV_NAPI_THREADED_DISABLED; 7194 break; 7195 } 7196 } 7197 } 7198 } 7199 7200 WRITE_ONCE(dev->threaded, threaded); 7201 7202 /* The error should not occur as the kthreads are already created. */ 7203 list_for_each_entry(napi, &dev->napi_list, dev_list) 7204 WARN_ON_ONCE(napi_set_threaded(napi, threaded)); 7205 7206 /* Override the config for all NAPIs even if currently not listed */ 7207 for (i = 0; i < dev->num_napi_configs; i++) 7208 dev->napi_config[i].threaded = threaded; 7209 7210 return err; 7211 } 7212 7213 /** 7214 * netif_threaded_enable() - enable threaded NAPIs 7215 * @dev: net_device instance 7216 * 7217 * Enable threaded mode for the NAPI instances of the device. This may be useful 7218 * for devices where multiple NAPI instances get scheduled by a single 7219 * interrupt. Threaded NAPI allows moving the NAPI processing to cores other 7220 * than the core where IRQ is mapped. 7221 * 7222 * This function should be called before @dev is registered. 7223 */ 7224 void netif_threaded_enable(struct net_device *dev) 7225 { 7226 WARN_ON_ONCE(netif_set_threaded(dev, NETDEV_NAPI_THREADED_ENABLED)); 7227 } 7228 EXPORT_SYMBOL(netif_threaded_enable); 7229 7230 /** 7231 * netif_queue_set_napi - Associate queue with the napi 7232 * @dev: device to which NAPI and queue belong 7233 * @queue_index: Index of queue 7234 * @type: queue type as RX or TX 7235 * @napi: NAPI context, pass NULL to clear previously set NAPI 7236 * 7237 * Set queue with its corresponding napi context. This should be done after 7238 * registering the NAPI handler for the queue-vector and the queues have been 7239 * mapped to the corresponding interrupt vector. 7240 */ 7241 void netif_queue_set_napi(struct net_device *dev, unsigned int queue_index, 7242 enum netdev_queue_type type, struct napi_struct *napi) 7243 { 7244 struct netdev_rx_queue *rxq; 7245 struct netdev_queue *txq; 7246 7247 if (WARN_ON_ONCE(napi && !napi->dev)) 7248 return; 7249 netdev_ops_assert_locked_or_invisible(dev); 7250 7251 switch (type) { 7252 case NETDEV_QUEUE_TYPE_RX: 7253 rxq = __netif_get_rx_queue(dev, queue_index); 7254 rxq->napi = napi; 7255 return; 7256 case NETDEV_QUEUE_TYPE_TX: 7257 txq = netdev_get_tx_queue(dev, queue_index); 7258 txq->napi = napi; 7259 return; 7260 default: 7261 return; 7262 } 7263 } 7264 EXPORT_SYMBOL(netif_queue_set_napi); 7265 7266 static void 7267 netif_napi_irq_notify(struct irq_affinity_notify *notify, 7268 const cpumask_t *mask) 7269 { 7270 struct napi_struct *napi = 7271 container_of(notify, struct napi_struct, notify); 7272 #ifdef CONFIG_RFS_ACCEL 7273 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 7274 int err; 7275 #endif 7276 7277 if (napi->config && napi->dev->irq_affinity_auto) 7278 cpumask_copy(&napi->config->affinity_mask, mask); 7279 7280 #ifdef CONFIG_RFS_ACCEL 7281 if (napi->dev->rx_cpu_rmap_auto) { 7282 err = cpu_rmap_update(rmap, napi->napi_rmap_idx, mask); 7283 if (err) 7284 netdev_warn(napi->dev, "RMAP update failed (%d)\n", 7285 err); 7286 } 7287 #endif 7288 } 7289 7290 #ifdef CONFIG_RFS_ACCEL 7291 static void netif_napi_affinity_release(struct kref *ref) 7292 { 7293 struct napi_struct *napi = 7294 container_of(ref, struct napi_struct, notify.kref); 7295 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 7296 7297 netdev_assert_locked(napi->dev); 7298 WARN_ON(test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, 7299 &napi->state)); 7300 7301 if (!napi->dev->rx_cpu_rmap_auto) 7302 return; 7303 rmap->obj[napi->napi_rmap_idx] = NULL; 7304 napi->napi_rmap_idx = -1; 7305 cpu_rmap_put(rmap); 7306 } 7307 7308 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 7309 { 7310 if (dev->rx_cpu_rmap_auto) 7311 return 0; 7312 7313 dev->rx_cpu_rmap = alloc_irq_cpu_rmap(num_irqs); 7314 if (!dev->rx_cpu_rmap) 7315 return -ENOMEM; 7316 7317 dev->rx_cpu_rmap_auto = true; 7318 return 0; 7319 } 7320 EXPORT_SYMBOL(netif_enable_cpu_rmap); 7321 7322 static void netif_del_cpu_rmap(struct net_device *dev) 7323 { 7324 struct cpu_rmap *rmap = dev->rx_cpu_rmap; 7325 7326 if (!dev->rx_cpu_rmap_auto) 7327 return; 7328 7329 /* Free the rmap */ 7330 cpu_rmap_put(rmap); 7331 dev->rx_cpu_rmap = NULL; 7332 dev->rx_cpu_rmap_auto = false; 7333 } 7334 7335 #else 7336 static void netif_napi_affinity_release(struct kref *ref) 7337 { 7338 } 7339 7340 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 7341 { 7342 return 0; 7343 } 7344 EXPORT_SYMBOL(netif_enable_cpu_rmap); 7345 7346 static void netif_del_cpu_rmap(struct net_device *dev) 7347 { 7348 } 7349 #endif 7350 7351 void netif_set_affinity_auto(struct net_device *dev) 7352 { 7353 unsigned int i, maxqs, numa; 7354 7355 maxqs = max(dev->num_tx_queues, dev->num_rx_queues); 7356 numa = dev_to_node(&dev->dev); 7357 7358 for (i = 0; i < maxqs; i++) 7359 cpumask_set_cpu(cpumask_local_spread(i, numa), 7360 &dev->napi_config[i].affinity_mask); 7361 7362 dev->irq_affinity_auto = true; 7363 } 7364 EXPORT_SYMBOL(netif_set_affinity_auto); 7365 7366 void netif_napi_set_irq_locked(struct napi_struct *napi, int irq) 7367 { 7368 int rc; 7369 7370 netdev_assert_locked_or_invisible(napi->dev); 7371 7372 if (napi->irq == irq) 7373 return; 7374 7375 /* Remove existing resources */ 7376 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7377 irq_set_affinity_notifier(napi->irq, NULL); 7378 7379 napi->irq = irq; 7380 if (irq < 0 || 7381 (!napi->dev->rx_cpu_rmap_auto && !napi->dev->irq_affinity_auto)) 7382 return; 7383 7384 /* Abort for buggy drivers */ 7385 if (napi->dev->irq_affinity_auto && WARN_ON_ONCE(!napi->config)) 7386 return; 7387 7388 #ifdef CONFIG_RFS_ACCEL 7389 if (napi->dev->rx_cpu_rmap_auto) { 7390 rc = cpu_rmap_add(napi->dev->rx_cpu_rmap, napi); 7391 if (rc < 0) 7392 return; 7393 7394 cpu_rmap_get(napi->dev->rx_cpu_rmap); 7395 napi->napi_rmap_idx = rc; 7396 } 7397 #endif 7398 7399 /* Use core IRQ notifier */ 7400 napi->notify.notify = netif_napi_irq_notify; 7401 napi->notify.release = netif_napi_affinity_release; 7402 rc = irq_set_affinity_notifier(irq, &napi->notify); 7403 if (rc) { 7404 netdev_warn(napi->dev, "Unable to set IRQ notifier (%d)\n", 7405 rc); 7406 goto put_rmap; 7407 } 7408 7409 set_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state); 7410 return; 7411 7412 put_rmap: 7413 #ifdef CONFIG_RFS_ACCEL 7414 if (napi->dev->rx_cpu_rmap_auto) { 7415 napi->dev->rx_cpu_rmap->obj[napi->napi_rmap_idx] = NULL; 7416 cpu_rmap_put(napi->dev->rx_cpu_rmap); 7417 napi->napi_rmap_idx = -1; 7418 } 7419 #endif 7420 napi->notify.notify = NULL; 7421 napi->notify.release = NULL; 7422 } 7423 EXPORT_SYMBOL(netif_napi_set_irq_locked); 7424 7425 static void napi_restore_config(struct napi_struct *n) 7426 { 7427 n->defer_hard_irqs = n->config->defer_hard_irqs; 7428 n->gro_flush_timeout = n->config->gro_flush_timeout; 7429 n->irq_suspend_timeout = n->config->irq_suspend_timeout; 7430 7431 if (n->dev->irq_affinity_auto && 7432 test_bit(NAPI_STATE_HAS_NOTIFIER, &n->state)) 7433 irq_set_affinity(n->irq, &n->config->affinity_mask); 7434 7435 /* a NAPI ID might be stored in the config, if so use it. if not, use 7436 * napi_hash_add to generate one for us. 7437 */ 7438 if (n->config->napi_id) { 7439 napi_hash_add_with_id(n, n->config->napi_id); 7440 } else { 7441 napi_hash_add(n); 7442 n->config->napi_id = n->napi_id; 7443 } 7444 7445 WARN_ON_ONCE(napi_set_threaded(n, n->config->threaded)); 7446 } 7447 7448 static void napi_save_config(struct napi_struct *n) 7449 { 7450 n->config->defer_hard_irqs = n->defer_hard_irqs; 7451 n->config->gro_flush_timeout = n->gro_flush_timeout; 7452 n->config->irq_suspend_timeout = n->irq_suspend_timeout; 7453 napi_hash_del(n); 7454 } 7455 7456 /* Netlink wants the NAPI list to be sorted by ID, if adding a NAPI which will 7457 * inherit an existing ID try to insert it at the right position. 7458 */ 7459 static void 7460 netif_napi_dev_list_add(struct net_device *dev, struct napi_struct *napi) 7461 { 7462 unsigned int new_id, pos_id; 7463 struct list_head *higher; 7464 struct napi_struct *pos; 7465 7466 new_id = UINT_MAX; 7467 if (napi->config && napi->config->napi_id) 7468 new_id = napi->config->napi_id; 7469 7470 higher = &dev->napi_list; 7471 list_for_each_entry(pos, &dev->napi_list, dev_list) { 7472 if (napi_id_valid(pos->napi_id)) 7473 pos_id = pos->napi_id; 7474 else if (pos->config) 7475 pos_id = pos->config->napi_id; 7476 else 7477 pos_id = UINT_MAX; 7478 7479 if (pos_id <= new_id) 7480 break; 7481 higher = &pos->dev_list; 7482 } 7483 list_add_rcu(&napi->dev_list, higher); /* adds after higher */ 7484 } 7485 7486 /* Double check that napi_get_frags() allocates skbs with 7487 * skb->head being backed by slab, not a page fragment. 7488 * This is to make sure bug fixed in 3226b158e67c 7489 * ("net: avoid 32 x truesize under-estimation for tiny skbs") 7490 * does not accidentally come back. 7491 */ 7492 static void napi_get_frags_check(struct napi_struct *napi) 7493 { 7494 struct sk_buff *skb; 7495 7496 local_bh_disable(); 7497 skb = napi_get_frags(napi); 7498 WARN_ON_ONCE(skb && skb->head_frag); 7499 napi_free_frags(napi); 7500 local_bh_enable(); 7501 } 7502 7503 void netif_napi_add_weight_locked(struct net_device *dev, 7504 struct napi_struct *napi, 7505 int (*poll)(struct napi_struct *, int), 7506 int weight) 7507 { 7508 netdev_assert_locked(dev); 7509 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state))) 7510 return; 7511 7512 INIT_LIST_HEAD(&napi->poll_list); 7513 INIT_HLIST_NODE(&napi->napi_hash_node); 7514 hrtimer_setup(&napi->timer, napi_watchdog, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); 7515 gro_init(&napi->gro); 7516 napi->skb = NULL; 7517 napi->poll = poll; 7518 if (weight > NAPI_POLL_WEIGHT) 7519 netdev_err_once(dev, "%s() called with weight %d\n", __func__, 7520 weight); 7521 napi->weight = weight; 7522 napi->dev = dev; 7523 #ifdef CONFIG_NETPOLL 7524 napi->poll_owner = -1; 7525 #endif 7526 napi->list_owner = -1; 7527 set_bit(NAPI_STATE_SCHED, &napi->state); 7528 set_bit(NAPI_STATE_NPSVC, &napi->state); 7529 netif_napi_dev_list_add(dev, napi); 7530 7531 /* default settings from sysfs are applied to all NAPIs. any per-NAPI 7532 * configuration will be loaded in napi_enable 7533 */ 7534 napi_set_defer_hard_irqs(napi, READ_ONCE(dev->napi_defer_hard_irqs)); 7535 napi_set_gro_flush_timeout(napi, READ_ONCE(dev->gro_flush_timeout)); 7536 7537 napi_get_frags_check(napi); 7538 /* Create kthread for this napi if dev->threaded is set. 7539 * Clear dev->threaded if kthread creation failed so that 7540 * threaded mode will not be enabled in napi_enable(). 7541 */ 7542 if (napi_get_threaded_config(dev, napi)) 7543 if (napi_kthread_create(napi)) 7544 dev->threaded = NETDEV_NAPI_THREADED_DISABLED; 7545 netif_napi_set_irq_locked(napi, -1); 7546 } 7547 EXPORT_SYMBOL(netif_napi_add_weight_locked); 7548 7549 void napi_disable_locked(struct napi_struct *n) 7550 { 7551 unsigned long val, new; 7552 7553 might_sleep(); 7554 netdev_assert_locked(n->dev); 7555 7556 set_bit(NAPI_STATE_DISABLE, &n->state); 7557 7558 val = READ_ONCE(n->state); 7559 do { 7560 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) { 7561 usleep_range(20, 200); 7562 val = READ_ONCE(n->state); 7563 } 7564 7565 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC; 7566 new &= ~(NAPIF_STATE_THREADED | 7567 NAPIF_STATE_THREADED_BUSY_POLL | 7568 NAPIF_STATE_PREFER_BUSY_POLL); 7569 } while (!try_cmpxchg(&n->state, &val, new)); 7570 7571 hrtimer_cancel(&n->timer); 7572 7573 if (n->config) 7574 napi_save_config(n); 7575 else 7576 napi_hash_del(n); 7577 7578 clear_bit(NAPI_STATE_DISABLE, &n->state); 7579 } 7580 EXPORT_SYMBOL(napi_disable_locked); 7581 7582 /** 7583 * napi_disable() - prevent NAPI from scheduling 7584 * @n: NAPI context 7585 * 7586 * Stop NAPI from being scheduled on this context. 7587 * Waits till any outstanding processing completes. 7588 * Takes netdev_lock() for associated net_device. 7589 */ 7590 void napi_disable(struct napi_struct *n) 7591 { 7592 netdev_lock(n->dev); 7593 napi_disable_locked(n); 7594 netdev_unlock(n->dev); 7595 } 7596 EXPORT_SYMBOL(napi_disable); 7597 7598 void napi_enable_locked(struct napi_struct *n) 7599 { 7600 unsigned long new, val = READ_ONCE(n->state); 7601 7602 if (n->config) 7603 napi_restore_config(n); 7604 else 7605 napi_hash_add(n); 7606 7607 do { 7608 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val)); 7609 7610 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC); 7611 if (n->dev->threaded && n->thread) 7612 new |= NAPIF_STATE_THREADED; 7613 } while (!try_cmpxchg(&n->state, &val, new)); 7614 } 7615 EXPORT_SYMBOL(napi_enable_locked); 7616 7617 /** 7618 * napi_enable() - enable NAPI scheduling 7619 * @n: NAPI context 7620 * 7621 * Enable scheduling of a NAPI instance. 7622 * Must be paired with napi_disable(). 7623 * Takes netdev_lock() for associated net_device. 7624 */ 7625 void napi_enable(struct napi_struct *n) 7626 { 7627 netdev_lock(n->dev); 7628 napi_enable_locked(n); 7629 netdev_unlock(n->dev); 7630 } 7631 EXPORT_SYMBOL(napi_enable); 7632 7633 /* Must be called in process context */ 7634 void __netif_napi_del_locked(struct napi_struct *napi) 7635 { 7636 netdev_assert_locked(napi->dev); 7637 7638 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state)) 7639 return; 7640 7641 /* Make sure NAPI is disabled (or was never enabled). */ 7642 WARN_ON(!test_bit(NAPI_STATE_SCHED, &napi->state)); 7643 7644 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7645 irq_set_affinity_notifier(napi->irq, NULL); 7646 7647 if (napi->config) { 7648 napi->index = -1; 7649 napi->config = NULL; 7650 } 7651 7652 list_del_rcu(&napi->dev_list); 7653 napi_free_frags(napi); 7654 7655 gro_cleanup(&napi->gro); 7656 7657 if (napi->thread) { 7658 kthread_stop(napi->thread); 7659 napi->thread = NULL; 7660 } 7661 } 7662 EXPORT_SYMBOL(__netif_napi_del_locked); 7663 7664 static int __napi_poll(struct napi_struct *n, bool *repoll) 7665 { 7666 int work, weight; 7667 7668 weight = n->weight; 7669 7670 /* This NAPI_STATE_SCHED test is for avoiding a race 7671 * with netpoll's poll_napi(). Only the entity which 7672 * obtains the lock and sees NAPI_STATE_SCHED set will 7673 * actually make the ->poll() call. Therefore we avoid 7674 * accidentally calling ->poll() when NAPI is not scheduled. 7675 */ 7676 work = 0; 7677 if (napi_is_scheduled(n)) { 7678 work = n->poll(n, weight); 7679 trace_napi_poll(n, work, weight); 7680 7681 xdp_do_check_flushed(n); 7682 } 7683 7684 if (unlikely(work > weight)) 7685 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n", 7686 n->poll, work, weight); 7687 7688 if (likely(work < weight)) 7689 return work; 7690 7691 /* Drivers must not modify the NAPI state if they 7692 * consume the entire weight. In such cases this code 7693 * still "owns" the NAPI instance and therefore can 7694 * move the instance around on the list at-will. 7695 */ 7696 if (unlikely(napi_disable_pending(n))) { 7697 napi_complete(n); 7698 return work; 7699 } 7700 7701 /* The NAPI context has more processing work, but busy-polling 7702 * is preferred. Exit early. 7703 */ 7704 if (napi_prefer_busy_poll(n)) { 7705 if (napi_complete_done(n, work)) { 7706 /* If timeout is not set, we need to make sure 7707 * that the NAPI is re-scheduled. 7708 */ 7709 napi_schedule(n); 7710 } 7711 return work; 7712 } 7713 7714 /* Flush too old packets. If HZ < 1000, flush all packets */ 7715 gro_flush_normal(&n->gro, HZ >= 1000); 7716 7717 /* Some drivers may have called napi_schedule 7718 * prior to exhausting their budget. 7719 */ 7720 if (unlikely(!list_empty(&n->poll_list))) { 7721 pr_warn_once("%s: Budget exhausted after napi rescheduled\n", 7722 n->dev ? n->dev->name : "backlog"); 7723 return work; 7724 } 7725 7726 *repoll = true; 7727 7728 return work; 7729 } 7730 7731 static int napi_poll(struct napi_struct *n, struct list_head *repoll) 7732 { 7733 bool do_repoll = false; 7734 void *have; 7735 int work; 7736 7737 list_del_init(&n->poll_list); 7738 7739 have = netpoll_poll_lock(n); 7740 7741 work = __napi_poll(n, &do_repoll); 7742 7743 if (do_repoll) { 7744 #if defined(CONFIG_DEBUG_NET) 7745 if (unlikely(!napi_is_scheduled(n))) 7746 pr_crit("repoll requested for device %s %ps but napi is not scheduled.\n", 7747 n->dev->name, n->poll); 7748 #endif 7749 list_add_tail(&n->poll_list, repoll); 7750 } 7751 netpoll_poll_unlock(have); 7752 7753 return work; 7754 } 7755 7756 static int napi_thread_wait(struct napi_struct *napi) 7757 { 7758 set_current_state(TASK_INTERRUPTIBLE); 7759 7760 while (!kthread_should_stop()) { 7761 /* Testing SCHED_THREADED bit here to make sure the current 7762 * kthread owns this napi and could poll on this napi. 7763 * Testing SCHED bit is not enough because SCHED bit might be 7764 * set by some other busy poll thread or by napi_disable(). 7765 */ 7766 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) { 7767 WARN_ON(!list_empty(&napi->poll_list)); 7768 __set_current_state(TASK_RUNNING); 7769 return 0; 7770 } 7771 7772 schedule(); 7773 set_current_state(TASK_INTERRUPTIBLE); 7774 } 7775 __set_current_state(TASK_RUNNING); 7776 7777 return -1; 7778 } 7779 7780 static void napi_threaded_poll_loop(struct napi_struct *napi, bool busy_poll) 7781 { 7782 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7783 struct softnet_data *sd; 7784 unsigned long last_qs = jiffies; 7785 7786 for (;;) { 7787 bool repoll = false; 7788 void *have; 7789 7790 local_bh_disable(); 7791 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7792 7793 sd = this_cpu_ptr(&softnet_data); 7794 sd->in_napi_threaded_poll = true; 7795 7796 have = netpoll_poll_lock(napi); 7797 __napi_poll(napi, &repoll); 7798 netpoll_poll_unlock(have); 7799 7800 sd->in_napi_threaded_poll = false; 7801 barrier(); 7802 7803 if (sd_has_rps_ipi_waiting(sd)) { 7804 local_irq_disable(); 7805 net_rps_action_and_irq_enable(sd); 7806 } 7807 skb_defer_free_flush(); 7808 bpf_net_ctx_clear(bpf_net_ctx); 7809 7810 /* When busy poll is enabled, the old packets are not flushed in 7811 * napi_complete_done. So flush them here. 7812 */ 7813 if (busy_poll) 7814 gro_flush_normal(&napi->gro, HZ >= 1000); 7815 local_bh_enable(); 7816 7817 /* Call cond_resched here to avoid watchdog warnings. */ 7818 if (repoll || busy_poll) { 7819 rcu_softirq_qs_periodic(last_qs); 7820 cond_resched(); 7821 } 7822 7823 if (!repoll) 7824 break; 7825 } 7826 } 7827 7828 static int napi_threaded_poll(void *data) 7829 { 7830 struct napi_struct *napi = data; 7831 bool want_busy_poll; 7832 bool in_busy_poll; 7833 unsigned long val; 7834 7835 while (!napi_thread_wait(napi)) { 7836 val = READ_ONCE(napi->state); 7837 7838 want_busy_poll = val & NAPIF_STATE_THREADED_BUSY_POLL; 7839 in_busy_poll = val & NAPIF_STATE_IN_BUSY_POLL; 7840 7841 if (unlikely(val & NAPIF_STATE_DISABLE)) 7842 want_busy_poll = false; 7843 7844 if (want_busy_poll != in_busy_poll) 7845 assign_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state, 7846 want_busy_poll); 7847 7848 napi_threaded_poll_loop(napi, want_busy_poll); 7849 } 7850 7851 return 0; 7852 } 7853 7854 static __latent_entropy void net_rx_action(void) 7855 { 7856 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 7857 unsigned long time_limit = jiffies + 7858 usecs_to_jiffies(READ_ONCE(net_hotdata.netdev_budget_usecs)); 7859 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7860 int budget = READ_ONCE(net_hotdata.netdev_budget); 7861 LIST_HEAD(list); 7862 LIST_HEAD(repoll); 7863 7864 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7865 start: 7866 sd->in_net_rx_action = true; 7867 local_irq_disable(); 7868 list_splice_init(&sd->poll_list, &list); 7869 local_irq_enable(); 7870 7871 for (;;) { 7872 struct napi_struct *n; 7873 7874 skb_defer_free_flush(); 7875 7876 if (list_empty(&list)) { 7877 if (list_empty(&repoll)) { 7878 sd->in_net_rx_action = false; 7879 barrier(); 7880 /* We need to check if ____napi_schedule() 7881 * had refilled poll_list while 7882 * sd->in_net_rx_action was true. 7883 */ 7884 if (!list_empty(&sd->poll_list)) 7885 goto start; 7886 if (!sd_has_rps_ipi_waiting(sd)) 7887 goto end; 7888 } 7889 break; 7890 } 7891 7892 n = list_first_entry(&list, struct napi_struct, poll_list); 7893 budget -= napi_poll(n, &repoll); 7894 7895 /* If softirq window is exhausted then punt. 7896 * Allow this to run for 2 jiffies since which will allow 7897 * an average latency of 1.5/HZ. 7898 */ 7899 if (unlikely(budget <= 0 || 7900 time_after_eq(jiffies, time_limit))) { 7901 /* Pairs with READ_ONCE() in softnet_seq_show() */ 7902 WRITE_ONCE(sd->time_squeeze, sd->time_squeeze + 1); 7903 break; 7904 } 7905 } 7906 7907 local_irq_disable(); 7908 7909 list_splice_tail_init(&sd->poll_list, &list); 7910 list_splice_tail(&repoll, &list); 7911 list_splice(&list, &sd->poll_list); 7912 if (!list_empty(&sd->poll_list)) 7913 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 7914 else 7915 sd->in_net_rx_action = false; 7916 7917 net_rps_action_and_irq_enable(sd); 7918 end: 7919 bpf_net_ctx_clear(bpf_net_ctx); 7920 } 7921 7922 struct netdev_adjacent { 7923 struct net_device *dev; 7924 netdevice_tracker dev_tracker; 7925 7926 /* upper master flag, there can only be one master device per list */ 7927 bool master; 7928 7929 /* lookup ignore flag */ 7930 bool ignore; 7931 7932 /* counter for the number of times this device was added to us */ 7933 u16 ref_nr; 7934 7935 /* private field for the users */ 7936 void *private; 7937 7938 struct list_head list; 7939 struct rcu_head rcu; 7940 }; 7941 7942 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev, 7943 struct list_head *adj_list) 7944 { 7945 struct netdev_adjacent *adj; 7946 7947 list_for_each_entry(adj, adj_list, list) { 7948 if (adj->dev == adj_dev) 7949 return adj; 7950 } 7951 return NULL; 7952 } 7953 7954 static int ____netdev_has_upper_dev(struct net_device *upper_dev, 7955 struct netdev_nested_priv *priv) 7956 { 7957 struct net_device *dev = (struct net_device *)priv->data; 7958 7959 return upper_dev == dev; 7960 } 7961 7962 /** 7963 * netdev_has_upper_dev - Check if device is linked to an upper device 7964 * @dev: device 7965 * @upper_dev: upper device to check 7966 * 7967 * Find out if a device is linked to specified upper device and return true 7968 * in case it is. Note that this checks only immediate upper device, 7969 * not through a complete stack of devices. The caller must hold the RTNL lock. 7970 */ 7971 bool netdev_has_upper_dev(struct net_device *dev, 7972 struct net_device *upper_dev) 7973 { 7974 struct netdev_nested_priv priv = { 7975 .data = (void *)upper_dev, 7976 }; 7977 7978 ASSERT_RTNL(); 7979 7980 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7981 &priv); 7982 } 7983 EXPORT_SYMBOL(netdev_has_upper_dev); 7984 7985 /** 7986 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device 7987 * @dev: device 7988 * @upper_dev: upper device to check 7989 * 7990 * Find out if a device is linked to specified upper device and return true 7991 * in case it is. Note that this checks the entire upper device chain. 7992 * The caller must hold rcu lock. 7993 */ 7994 7995 bool netdev_has_upper_dev_all_rcu(struct net_device *dev, 7996 struct net_device *upper_dev) 7997 { 7998 struct netdev_nested_priv priv = { 7999 .data = (void *)upper_dev, 8000 }; 8001 8002 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 8003 &priv); 8004 } 8005 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu); 8006 8007 /** 8008 * netdev_has_any_upper_dev - Check if device is linked to some device 8009 * @dev: device 8010 * 8011 * Find out if a device is linked to an upper device and return true in case 8012 * it is. The caller must hold the RTNL lock. 8013 */ 8014 bool netdev_has_any_upper_dev(struct net_device *dev) 8015 { 8016 ASSERT_RTNL(); 8017 8018 return !list_empty(&dev->adj_list.upper); 8019 } 8020 EXPORT_SYMBOL(netdev_has_any_upper_dev); 8021 8022 /** 8023 * netdev_master_upper_dev_get - Get master upper device 8024 * @dev: device 8025 * 8026 * Find a master upper device and return pointer to it or NULL in case 8027 * it's not there. The caller must hold the RTNL lock. 8028 */ 8029 struct net_device *netdev_master_upper_dev_get(struct net_device *dev) 8030 { 8031 struct netdev_adjacent *upper; 8032 8033 ASSERT_RTNL(); 8034 8035 if (list_empty(&dev->adj_list.upper)) 8036 return NULL; 8037 8038 upper = list_first_entry(&dev->adj_list.upper, 8039 struct netdev_adjacent, list); 8040 if (likely(upper->master)) 8041 return upper->dev; 8042 return NULL; 8043 } 8044 EXPORT_SYMBOL(netdev_master_upper_dev_get); 8045 8046 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev) 8047 { 8048 struct netdev_adjacent *upper; 8049 8050 ASSERT_RTNL(); 8051 8052 if (list_empty(&dev->adj_list.upper)) 8053 return NULL; 8054 8055 upper = list_first_entry(&dev->adj_list.upper, 8056 struct netdev_adjacent, list); 8057 if (likely(upper->master) && !upper->ignore) 8058 return upper->dev; 8059 return NULL; 8060 } 8061 8062 /** 8063 * netdev_has_any_lower_dev - Check if device is linked to some device 8064 * @dev: device 8065 * 8066 * Find out if a device is linked to a lower device and return true in case 8067 * it is. The caller must hold the RTNL lock. 8068 */ 8069 static bool netdev_has_any_lower_dev(struct net_device *dev) 8070 { 8071 ASSERT_RTNL(); 8072 8073 return !list_empty(&dev->adj_list.lower); 8074 } 8075 8076 void *netdev_adjacent_get_private(struct list_head *adj_list) 8077 { 8078 struct netdev_adjacent *adj; 8079 8080 adj = list_entry(adj_list, struct netdev_adjacent, list); 8081 8082 return adj->private; 8083 } 8084 EXPORT_SYMBOL(netdev_adjacent_get_private); 8085 8086 /** 8087 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list 8088 * @dev: device 8089 * @iter: list_head ** of the current position 8090 * 8091 * Gets the next device from the dev's upper list, starting from iter 8092 * position. The caller must hold RCU read lock. 8093 */ 8094 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev, 8095 struct list_head **iter) 8096 { 8097 struct netdev_adjacent *upper; 8098 8099 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 8100 8101 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8102 8103 if (&upper->list == &dev->adj_list.upper) 8104 return NULL; 8105 8106 *iter = &upper->list; 8107 8108 return upper->dev; 8109 } 8110 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu); 8111 8112 static struct net_device *__netdev_next_upper_dev(struct net_device *dev, 8113 struct list_head **iter, 8114 bool *ignore) 8115 { 8116 struct netdev_adjacent *upper; 8117 8118 upper = list_entry((*iter)->next, struct netdev_adjacent, list); 8119 8120 if (&upper->list == &dev->adj_list.upper) 8121 return NULL; 8122 8123 *iter = &upper->list; 8124 *ignore = upper->ignore; 8125 8126 return upper->dev; 8127 } 8128 8129 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev, 8130 struct list_head **iter) 8131 { 8132 struct netdev_adjacent *upper; 8133 8134 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 8135 8136 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8137 8138 if (&upper->list == &dev->adj_list.upper) 8139 return NULL; 8140 8141 *iter = &upper->list; 8142 8143 return upper->dev; 8144 } 8145 8146 static int __netdev_walk_all_upper_dev(struct net_device *dev, 8147 int (*fn)(struct net_device *dev, 8148 struct netdev_nested_priv *priv), 8149 struct netdev_nested_priv *priv) 8150 { 8151 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8152 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8153 int ret, cur = 0; 8154 bool ignore; 8155 8156 now = dev; 8157 iter = &dev->adj_list.upper; 8158 8159 while (1) { 8160 if (now != dev) { 8161 ret = fn(now, priv); 8162 if (ret) 8163 return ret; 8164 } 8165 8166 next = NULL; 8167 while (1) { 8168 udev = __netdev_next_upper_dev(now, &iter, &ignore); 8169 if (!udev) 8170 break; 8171 if (ignore) 8172 continue; 8173 8174 next = udev; 8175 niter = &udev->adj_list.upper; 8176 dev_stack[cur] = now; 8177 iter_stack[cur++] = iter; 8178 break; 8179 } 8180 8181 if (!next) { 8182 if (!cur) 8183 return 0; 8184 next = dev_stack[--cur]; 8185 niter = iter_stack[cur]; 8186 } 8187 8188 now = next; 8189 iter = niter; 8190 } 8191 8192 return 0; 8193 } 8194 8195 int netdev_walk_all_upper_dev_rcu(struct net_device *dev, 8196 int (*fn)(struct net_device *dev, 8197 struct netdev_nested_priv *priv), 8198 struct netdev_nested_priv *priv) 8199 { 8200 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8201 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8202 int ret, cur = 0; 8203 8204 now = dev; 8205 iter = &dev->adj_list.upper; 8206 8207 while (1) { 8208 if (now != dev) { 8209 ret = fn(now, priv); 8210 if (ret) 8211 return ret; 8212 } 8213 8214 next = NULL; 8215 while (1) { 8216 udev = netdev_next_upper_dev_rcu(now, &iter); 8217 if (!udev) 8218 break; 8219 8220 next = udev; 8221 niter = &udev->adj_list.upper; 8222 dev_stack[cur] = now; 8223 iter_stack[cur++] = iter; 8224 break; 8225 } 8226 8227 if (!next) { 8228 if (!cur) 8229 return 0; 8230 next = dev_stack[--cur]; 8231 niter = iter_stack[cur]; 8232 } 8233 8234 now = next; 8235 iter = niter; 8236 } 8237 8238 return 0; 8239 } 8240 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu); 8241 8242 static bool __netdev_has_upper_dev(struct net_device *dev, 8243 struct net_device *upper_dev) 8244 { 8245 struct netdev_nested_priv priv = { 8246 .flags = 0, 8247 .data = (void *)upper_dev, 8248 }; 8249 8250 ASSERT_RTNL(); 8251 8252 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev, 8253 &priv); 8254 } 8255 8256 /** 8257 * netdev_lower_get_next_private - Get the next ->private from the 8258 * lower neighbour list 8259 * @dev: device 8260 * @iter: list_head ** of the current position 8261 * 8262 * Gets the next netdev_adjacent->private from the dev's lower neighbour 8263 * list, starting from iter position. The caller must hold either hold the 8264 * RTNL lock or its own locking that guarantees that the neighbour lower 8265 * list will remain unchanged. 8266 */ 8267 void *netdev_lower_get_next_private(struct net_device *dev, 8268 struct list_head **iter) 8269 { 8270 struct netdev_adjacent *lower; 8271 8272 lower = list_entry(*iter, struct netdev_adjacent, list); 8273 8274 if (&lower->list == &dev->adj_list.lower) 8275 return NULL; 8276 8277 *iter = lower->list.next; 8278 8279 return lower->private; 8280 } 8281 EXPORT_SYMBOL(netdev_lower_get_next_private); 8282 8283 /** 8284 * netdev_lower_get_next_private_rcu - Get the next ->private from the 8285 * lower neighbour list, RCU 8286 * variant 8287 * @dev: device 8288 * @iter: list_head ** of the current position 8289 * 8290 * Gets the next netdev_adjacent->private from the dev's lower neighbour 8291 * list, starting from iter position. The caller must hold RCU read lock. 8292 */ 8293 void *netdev_lower_get_next_private_rcu(struct net_device *dev, 8294 struct list_head **iter) 8295 { 8296 struct netdev_adjacent *lower; 8297 8298 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 8299 8300 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8301 8302 if (&lower->list == &dev->adj_list.lower) 8303 return NULL; 8304 8305 *iter = &lower->list; 8306 8307 return lower->private; 8308 } 8309 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu); 8310 8311 /** 8312 * netdev_lower_get_next - Get the next device from the lower neighbour 8313 * list 8314 * @dev: device 8315 * @iter: list_head ** of the current position 8316 * 8317 * Gets the next netdev_adjacent from the dev's lower neighbour 8318 * list, starting from iter position. The caller must hold RTNL lock or 8319 * its own locking that guarantees that the neighbour lower 8320 * list will remain unchanged. 8321 */ 8322 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter) 8323 { 8324 struct netdev_adjacent *lower; 8325 8326 lower = list_entry(*iter, struct netdev_adjacent, list); 8327 8328 if (&lower->list == &dev->adj_list.lower) 8329 return NULL; 8330 8331 *iter = lower->list.next; 8332 8333 return lower->dev; 8334 } 8335 EXPORT_SYMBOL(netdev_lower_get_next); 8336 8337 static struct net_device *netdev_next_lower_dev(struct net_device *dev, 8338 struct list_head **iter) 8339 { 8340 struct netdev_adjacent *lower; 8341 8342 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 8343 8344 if (&lower->list == &dev->adj_list.lower) 8345 return NULL; 8346 8347 *iter = &lower->list; 8348 8349 return lower->dev; 8350 } 8351 8352 static struct net_device *__netdev_next_lower_dev(struct net_device *dev, 8353 struct list_head **iter, 8354 bool *ignore) 8355 { 8356 struct netdev_adjacent *lower; 8357 8358 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 8359 8360 if (&lower->list == &dev->adj_list.lower) 8361 return NULL; 8362 8363 *iter = &lower->list; 8364 *ignore = lower->ignore; 8365 8366 return lower->dev; 8367 } 8368 8369 int netdev_walk_all_lower_dev(struct net_device *dev, 8370 int (*fn)(struct net_device *dev, 8371 struct netdev_nested_priv *priv), 8372 struct netdev_nested_priv *priv) 8373 { 8374 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8375 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8376 int ret, cur = 0; 8377 8378 now = dev; 8379 iter = &dev->adj_list.lower; 8380 8381 while (1) { 8382 if (now != dev) { 8383 ret = fn(now, priv); 8384 if (ret) 8385 return ret; 8386 } 8387 8388 next = NULL; 8389 while (1) { 8390 ldev = netdev_next_lower_dev(now, &iter); 8391 if (!ldev) 8392 break; 8393 8394 next = ldev; 8395 niter = &ldev->adj_list.lower; 8396 dev_stack[cur] = now; 8397 iter_stack[cur++] = iter; 8398 break; 8399 } 8400 8401 if (!next) { 8402 if (!cur) 8403 return 0; 8404 next = dev_stack[--cur]; 8405 niter = iter_stack[cur]; 8406 } 8407 8408 now = next; 8409 iter = niter; 8410 } 8411 8412 return 0; 8413 } 8414 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev); 8415 8416 static int __netdev_walk_all_lower_dev(struct net_device *dev, 8417 int (*fn)(struct net_device *dev, 8418 struct netdev_nested_priv *priv), 8419 struct netdev_nested_priv *priv) 8420 { 8421 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8422 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8423 int ret, cur = 0; 8424 bool ignore; 8425 8426 now = dev; 8427 iter = &dev->adj_list.lower; 8428 8429 while (1) { 8430 if (now != dev) { 8431 ret = fn(now, priv); 8432 if (ret) 8433 return ret; 8434 } 8435 8436 next = NULL; 8437 while (1) { 8438 ldev = __netdev_next_lower_dev(now, &iter, &ignore); 8439 if (!ldev) 8440 break; 8441 if (ignore) 8442 continue; 8443 8444 next = ldev; 8445 niter = &ldev->adj_list.lower; 8446 dev_stack[cur] = now; 8447 iter_stack[cur++] = iter; 8448 break; 8449 } 8450 8451 if (!next) { 8452 if (!cur) 8453 return 0; 8454 next = dev_stack[--cur]; 8455 niter = iter_stack[cur]; 8456 } 8457 8458 now = next; 8459 iter = niter; 8460 } 8461 8462 return 0; 8463 } 8464 8465 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev, 8466 struct list_head **iter) 8467 { 8468 struct netdev_adjacent *lower; 8469 8470 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8471 if (&lower->list == &dev->adj_list.lower) 8472 return NULL; 8473 8474 *iter = &lower->list; 8475 8476 return lower->dev; 8477 } 8478 EXPORT_SYMBOL(netdev_next_lower_dev_rcu); 8479 8480 static u8 __netdev_upper_depth(struct net_device *dev) 8481 { 8482 struct net_device *udev; 8483 struct list_head *iter; 8484 u8 max_depth = 0; 8485 bool ignore; 8486 8487 for (iter = &dev->adj_list.upper, 8488 udev = __netdev_next_upper_dev(dev, &iter, &ignore); 8489 udev; 8490 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) { 8491 if (ignore) 8492 continue; 8493 if (max_depth < udev->upper_level) 8494 max_depth = udev->upper_level; 8495 } 8496 8497 return max_depth; 8498 } 8499 8500 static u8 __netdev_lower_depth(struct net_device *dev) 8501 { 8502 struct net_device *ldev; 8503 struct list_head *iter; 8504 u8 max_depth = 0; 8505 bool ignore; 8506 8507 for (iter = &dev->adj_list.lower, 8508 ldev = __netdev_next_lower_dev(dev, &iter, &ignore); 8509 ldev; 8510 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) { 8511 if (ignore) 8512 continue; 8513 if (max_depth < ldev->lower_level) 8514 max_depth = ldev->lower_level; 8515 } 8516 8517 return max_depth; 8518 } 8519 8520 static int __netdev_update_upper_level(struct net_device *dev, 8521 struct netdev_nested_priv *__unused) 8522 { 8523 dev->upper_level = __netdev_upper_depth(dev) + 1; 8524 return 0; 8525 } 8526 8527 #ifdef CONFIG_LOCKDEP 8528 static LIST_HEAD(net_unlink_list); 8529 8530 static void net_unlink_todo(struct net_device *dev) 8531 { 8532 if (list_empty(&dev->unlink_list)) 8533 list_add_tail(&dev->unlink_list, &net_unlink_list); 8534 } 8535 #endif 8536 8537 static int __netdev_update_lower_level(struct net_device *dev, 8538 struct netdev_nested_priv *priv) 8539 { 8540 dev->lower_level = __netdev_lower_depth(dev) + 1; 8541 8542 #ifdef CONFIG_LOCKDEP 8543 if (!priv) 8544 return 0; 8545 8546 if (priv->flags & NESTED_SYNC_IMM) 8547 dev->nested_level = dev->lower_level - 1; 8548 if (priv->flags & NESTED_SYNC_TODO) 8549 net_unlink_todo(dev); 8550 #endif 8551 return 0; 8552 } 8553 8554 int netdev_walk_all_lower_dev_rcu(struct net_device *dev, 8555 int (*fn)(struct net_device *dev, 8556 struct netdev_nested_priv *priv), 8557 struct netdev_nested_priv *priv) 8558 { 8559 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8560 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8561 int ret, cur = 0; 8562 8563 now = dev; 8564 iter = &dev->adj_list.lower; 8565 8566 while (1) { 8567 if (now != dev) { 8568 ret = fn(now, priv); 8569 if (ret) 8570 return ret; 8571 } 8572 8573 next = NULL; 8574 while (1) { 8575 ldev = netdev_next_lower_dev_rcu(now, &iter); 8576 if (!ldev) 8577 break; 8578 8579 next = ldev; 8580 niter = &ldev->adj_list.lower; 8581 dev_stack[cur] = now; 8582 iter_stack[cur++] = iter; 8583 break; 8584 } 8585 8586 if (!next) { 8587 if (!cur) 8588 return 0; 8589 next = dev_stack[--cur]; 8590 niter = iter_stack[cur]; 8591 } 8592 8593 now = next; 8594 iter = niter; 8595 } 8596 8597 return 0; 8598 } 8599 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu); 8600 8601 /** 8602 * netdev_lower_get_first_private_rcu - Get the first ->private from the 8603 * lower neighbour list, RCU 8604 * variant 8605 * @dev: device 8606 * 8607 * Gets the first netdev_adjacent->private from the dev's lower neighbour 8608 * list. The caller must hold RCU read lock. 8609 */ 8610 void *netdev_lower_get_first_private_rcu(struct net_device *dev) 8611 { 8612 struct netdev_adjacent *lower; 8613 8614 lower = list_first_or_null_rcu(&dev->adj_list.lower, 8615 struct netdev_adjacent, list); 8616 if (lower) 8617 return lower->private; 8618 return NULL; 8619 } 8620 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu); 8621 8622 /** 8623 * netdev_master_upper_dev_get_rcu - Get master upper device 8624 * @dev: device 8625 * 8626 * Find a master upper device and return pointer to it or NULL in case 8627 * it's not there. The caller must hold the RCU read lock. 8628 */ 8629 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev) 8630 { 8631 struct netdev_adjacent *upper; 8632 8633 upper = list_first_or_null_rcu(&dev->adj_list.upper, 8634 struct netdev_adjacent, list); 8635 if (upper && likely(upper->master)) 8636 return upper->dev; 8637 return NULL; 8638 } 8639 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu); 8640 8641 static int netdev_adjacent_sysfs_add(struct net_device *dev, 8642 struct net_device *adj_dev, 8643 struct list_head *dev_list) 8644 { 8645 char linkname[IFNAMSIZ+7]; 8646 8647 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8648 "upper_%s" : "lower_%s", adj_dev->name); 8649 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj), 8650 linkname); 8651 } 8652 static void netdev_adjacent_sysfs_del(struct net_device *dev, 8653 char *name, 8654 struct list_head *dev_list) 8655 { 8656 char linkname[IFNAMSIZ+7]; 8657 8658 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8659 "upper_%s" : "lower_%s", name); 8660 sysfs_remove_link(&(dev->dev.kobj), linkname); 8661 } 8662 8663 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev, 8664 struct net_device *adj_dev, 8665 struct list_head *dev_list) 8666 { 8667 return (dev_list == &dev->adj_list.upper || 8668 dev_list == &dev->adj_list.lower) && 8669 net_eq(dev_net(dev), dev_net(adj_dev)); 8670 } 8671 8672 static int __netdev_adjacent_dev_insert(struct net_device *dev, 8673 struct net_device *adj_dev, 8674 struct list_head *dev_list, 8675 void *private, bool master) 8676 { 8677 struct netdev_adjacent *adj; 8678 int ret; 8679 8680 adj = __netdev_find_adj(adj_dev, dev_list); 8681 8682 if (adj) { 8683 adj->ref_nr += 1; 8684 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n", 8685 dev->name, adj_dev->name, adj->ref_nr); 8686 8687 return 0; 8688 } 8689 8690 adj = kmalloc(sizeof(*adj), GFP_KERNEL); 8691 if (!adj) 8692 return -ENOMEM; 8693 8694 adj->dev = adj_dev; 8695 adj->master = master; 8696 adj->ref_nr = 1; 8697 adj->private = private; 8698 adj->ignore = false; 8699 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL); 8700 8701 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n", 8702 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name); 8703 8704 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) { 8705 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list); 8706 if (ret) 8707 goto free_adj; 8708 } 8709 8710 /* Ensure that master link is always the first item in list. */ 8711 if (master) { 8712 ret = sysfs_create_link(&(dev->dev.kobj), 8713 &(adj_dev->dev.kobj), "master"); 8714 if (ret) 8715 goto remove_symlinks; 8716 8717 list_add_rcu(&adj->list, dev_list); 8718 } else { 8719 list_add_tail_rcu(&adj->list, dev_list); 8720 } 8721 8722 return 0; 8723 8724 remove_symlinks: 8725 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8726 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8727 free_adj: 8728 netdev_put(adj_dev, &adj->dev_tracker); 8729 kfree(adj); 8730 8731 return ret; 8732 } 8733 8734 static void __netdev_adjacent_dev_remove(struct net_device *dev, 8735 struct net_device *adj_dev, 8736 u16 ref_nr, 8737 struct list_head *dev_list) 8738 { 8739 struct netdev_adjacent *adj; 8740 8741 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n", 8742 dev->name, adj_dev->name, ref_nr); 8743 8744 adj = __netdev_find_adj(adj_dev, dev_list); 8745 8746 if (!adj) { 8747 pr_err("Adjacency does not exist for device %s from %s\n", 8748 dev->name, adj_dev->name); 8749 WARN_ON(1); 8750 return; 8751 } 8752 8753 if (adj->ref_nr > ref_nr) { 8754 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n", 8755 dev->name, adj_dev->name, ref_nr, 8756 adj->ref_nr - ref_nr); 8757 adj->ref_nr -= ref_nr; 8758 return; 8759 } 8760 8761 if (adj->master) 8762 sysfs_remove_link(&(dev->dev.kobj), "master"); 8763 8764 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8765 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8766 8767 list_del_rcu(&adj->list); 8768 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n", 8769 adj_dev->name, dev->name, adj_dev->name); 8770 netdev_put(adj_dev, &adj->dev_tracker); 8771 kfree_rcu(adj, rcu); 8772 } 8773 8774 static int __netdev_adjacent_dev_link_lists(struct net_device *dev, 8775 struct net_device *upper_dev, 8776 struct list_head *up_list, 8777 struct list_head *down_list, 8778 void *private, bool master) 8779 { 8780 int ret; 8781 8782 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list, 8783 private, master); 8784 if (ret) 8785 return ret; 8786 8787 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list, 8788 private, false); 8789 if (ret) { 8790 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list); 8791 return ret; 8792 } 8793 8794 return 0; 8795 } 8796 8797 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev, 8798 struct net_device *upper_dev, 8799 u16 ref_nr, 8800 struct list_head *up_list, 8801 struct list_head *down_list) 8802 { 8803 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list); 8804 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list); 8805 } 8806 8807 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev, 8808 struct net_device *upper_dev, 8809 void *private, bool master) 8810 { 8811 return __netdev_adjacent_dev_link_lists(dev, upper_dev, 8812 &dev->adj_list.upper, 8813 &upper_dev->adj_list.lower, 8814 private, master); 8815 } 8816 8817 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev, 8818 struct net_device *upper_dev) 8819 { 8820 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1, 8821 &dev->adj_list.upper, 8822 &upper_dev->adj_list.lower); 8823 } 8824 8825 static int __netdev_upper_dev_link(struct net_device *dev, 8826 struct net_device *upper_dev, bool master, 8827 void *upper_priv, void *upper_info, 8828 struct netdev_nested_priv *priv, 8829 struct netlink_ext_ack *extack) 8830 { 8831 struct netdev_notifier_changeupper_info changeupper_info = { 8832 .info = { 8833 .dev = dev, 8834 .extack = extack, 8835 }, 8836 .upper_dev = upper_dev, 8837 .master = master, 8838 .linking = true, 8839 .upper_info = upper_info, 8840 }; 8841 struct net_device *master_dev; 8842 int ret = 0; 8843 8844 ASSERT_RTNL(); 8845 8846 if (dev == upper_dev) 8847 return -EBUSY; 8848 8849 /* To prevent loops, check if dev is not upper device to upper_dev. */ 8850 if (__netdev_has_upper_dev(upper_dev, dev)) 8851 return -EBUSY; 8852 8853 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV) 8854 return -EMLINK; 8855 8856 if (!master) { 8857 if (__netdev_has_upper_dev(dev, upper_dev)) 8858 return -EEXIST; 8859 } else { 8860 master_dev = __netdev_master_upper_dev_get(dev); 8861 if (master_dev) 8862 return master_dev == upper_dev ? -EEXIST : -EBUSY; 8863 } 8864 8865 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8866 &changeupper_info.info); 8867 ret = notifier_to_errno(ret); 8868 if (ret) 8869 return ret; 8870 8871 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv, 8872 master); 8873 if (ret) 8874 return ret; 8875 8876 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8877 &changeupper_info.info); 8878 ret = notifier_to_errno(ret); 8879 if (ret) 8880 goto rollback; 8881 8882 __netdev_update_upper_level(dev, NULL); 8883 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8884 8885 __netdev_update_lower_level(upper_dev, priv); 8886 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8887 priv); 8888 8889 return 0; 8890 8891 rollback: 8892 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8893 8894 return ret; 8895 } 8896 8897 /** 8898 * netdev_upper_dev_link - Add a link to the upper device 8899 * @dev: device 8900 * @upper_dev: new upper device 8901 * @extack: netlink extended ack 8902 * 8903 * Adds a link to device which is upper to this one. The caller must hold 8904 * the RTNL lock. On a failure a negative errno code is returned. 8905 * On success the reference counts are adjusted and the function 8906 * returns zero. 8907 */ 8908 int netdev_upper_dev_link(struct net_device *dev, 8909 struct net_device *upper_dev, 8910 struct netlink_ext_ack *extack) 8911 { 8912 struct netdev_nested_priv priv = { 8913 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8914 .data = NULL, 8915 }; 8916 8917 return __netdev_upper_dev_link(dev, upper_dev, false, 8918 NULL, NULL, &priv, extack); 8919 } 8920 EXPORT_SYMBOL(netdev_upper_dev_link); 8921 8922 /** 8923 * netdev_master_upper_dev_link - Add a master link to the upper device 8924 * @dev: device 8925 * @upper_dev: new upper device 8926 * @upper_priv: upper device private 8927 * @upper_info: upper info to be passed down via notifier 8928 * @extack: netlink extended ack 8929 * 8930 * Adds a link to device which is upper to this one. In this case, only 8931 * one master upper device can be linked, although other non-master devices 8932 * might be linked as well. The caller must hold the RTNL lock. 8933 * On a failure a negative errno code is returned. On success the reference 8934 * counts are adjusted and the function returns zero. 8935 */ 8936 int netdev_master_upper_dev_link(struct net_device *dev, 8937 struct net_device *upper_dev, 8938 void *upper_priv, void *upper_info, 8939 struct netlink_ext_ack *extack) 8940 { 8941 struct netdev_nested_priv priv = { 8942 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8943 .data = NULL, 8944 }; 8945 8946 return __netdev_upper_dev_link(dev, upper_dev, true, 8947 upper_priv, upper_info, &priv, extack); 8948 } 8949 EXPORT_SYMBOL(netdev_master_upper_dev_link); 8950 8951 static void __netdev_upper_dev_unlink(struct net_device *dev, 8952 struct net_device *upper_dev, 8953 struct netdev_nested_priv *priv) 8954 { 8955 struct netdev_notifier_changeupper_info changeupper_info = { 8956 .info = { 8957 .dev = dev, 8958 }, 8959 .upper_dev = upper_dev, 8960 .linking = false, 8961 }; 8962 8963 ASSERT_RTNL(); 8964 8965 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev; 8966 8967 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8968 &changeupper_info.info); 8969 8970 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8971 8972 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8973 &changeupper_info.info); 8974 8975 __netdev_update_upper_level(dev, NULL); 8976 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8977 8978 __netdev_update_lower_level(upper_dev, priv); 8979 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8980 priv); 8981 } 8982 8983 /** 8984 * netdev_upper_dev_unlink - Removes a link to upper device 8985 * @dev: device 8986 * @upper_dev: new upper device 8987 * 8988 * Removes a link to device which is upper to this one. The caller must hold 8989 * the RTNL lock. 8990 */ 8991 void netdev_upper_dev_unlink(struct net_device *dev, 8992 struct net_device *upper_dev) 8993 { 8994 struct netdev_nested_priv priv = { 8995 .flags = NESTED_SYNC_TODO, 8996 .data = NULL, 8997 }; 8998 8999 __netdev_upper_dev_unlink(dev, upper_dev, &priv); 9000 } 9001 EXPORT_SYMBOL(netdev_upper_dev_unlink); 9002 9003 static void __netdev_adjacent_dev_set(struct net_device *upper_dev, 9004 struct net_device *lower_dev, 9005 bool val) 9006 { 9007 struct netdev_adjacent *adj; 9008 9009 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower); 9010 if (adj) 9011 adj->ignore = val; 9012 9013 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper); 9014 if (adj) 9015 adj->ignore = val; 9016 } 9017 9018 static void netdev_adjacent_dev_disable(struct net_device *upper_dev, 9019 struct net_device *lower_dev) 9020 { 9021 __netdev_adjacent_dev_set(upper_dev, lower_dev, true); 9022 } 9023 9024 static void netdev_adjacent_dev_enable(struct net_device *upper_dev, 9025 struct net_device *lower_dev) 9026 { 9027 __netdev_adjacent_dev_set(upper_dev, lower_dev, false); 9028 } 9029 9030 int netdev_adjacent_change_prepare(struct net_device *old_dev, 9031 struct net_device *new_dev, 9032 struct net_device *dev, 9033 struct netlink_ext_ack *extack) 9034 { 9035 struct netdev_nested_priv priv = { 9036 .flags = 0, 9037 .data = NULL, 9038 }; 9039 int err; 9040 9041 if (!new_dev) 9042 return 0; 9043 9044 if (old_dev && new_dev != old_dev) 9045 netdev_adjacent_dev_disable(dev, old_dev); 9046 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv, 9047 extack); 9048 if (err) { 9049 if (old_dev && new_dev != old_dev) 9050 netdev_adjacent_dev_enable(dev, old_dev); 9051 return err; 9052 } 9053 9054 return 0; 9055 } 9056 EXPORT_SYMBOL(netdev_adjacent_change_prepare); 9057 9058 void netdev_adjacent_change_commit(struct net_device *old_dev, 9059 struct net_device *new_dev, 9060 struct net_device *dev) 9061 { 9062 struct netdev_nested_priv priv = { 9063 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 9064 .data = NULL, 9065 }; 9066 9067 if (!new_dev || !old_dev) 9068 return; 9069 9070 if (new_dev == old_dev) 9071 return; 9072 9073 netdev_adjacent_dev_enable(dev, old_dev); 9074 __netdev_upper_dev_unlink(old_dev, dev, &priv); 9075 } 9076 EXPORT_SYMBOL(netdev_adjacent_change_commit); 9077 9078 void netdev_adjacent_change_abort(struct net_device *old_dev, 9079 struct net_device *new_dev, 9080 struct net_device *dev) 9081 { 9082 struct netdev_nested_priv priv = { 9083 .flags = 0, 9084 .data = NULL, 9085 }; 9086 9087 if (!new_dev) 9088 return; 9089 9090 if (old_dev && new_dev != old_dev) 9091 netdev_adjacent_dev_enable(dev, old_dev); 9092 9093 __netdev_upper_dev_unlink(new_dev, dev, &priv); 9094 } 9095 EXPORT_SYMBOL(netdev_adjacent_change_abort); 9096 9097 /** 9098 * netdev_bonding_info_change - Dispatch event about slave change 9099 * @dev: device 9100 * @bonding_info: info to dispatch 9101 * 9102 * Send NETDEV_BONDING_INFO to netdev notifiers with info. 9103 * The caller must hold the RTNL lock. 9104 */ 9105 void netdev_bonding_info_change(struct net_device *dev, 9106 struct netdev_bonding_info *bonding_info) 9107 { 9108 struct netdev_notifier_bonding_info info = { 9109 .info.dev = dev, 9110 }; 9111 9112 memcpy(&info.bonding_info, bonding_info, 9113 sizeof(struct netdev_bonding_info)); 9114 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, 9115 &info.info); 9116 } 9117 EXPORT_SYMBOL(netdev_bonding_info_change); 9118 9119 static int netdev_offload_xstats_enable_l3(struct net_device *dev, 9120 struct netlink_ext_ack *extack) 9121 { 9122 struct netdev_notifier_offload_xstats_info info = { 9123 .info.dev = dev, 9124 .info.extack = extack, 9125 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 9126 }; 9127 int err; 9128 int rc; 9129 9130 dev->offload_xstats_l3 = kzalloc(sizeof(*dev->offload_xstats_l3), 9131 GFP_KERNEL); 9132 if (!dev->offload_xstats_l3) 9133 return -ENOMEM; 9134 9135 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE, 9136 NETDEV_OFFLOAD_XSTATS_DISABLE, 9137 &info.info); 9138 err = notifier_to_errno(rc); 9139 if (err) 9140 goto free_stats; 9141 9142 return 0; 9143 9144 free_stats: 9145 kfree(dev->offload_xstats_l3); 9146 dev->offload_xstats_l3 = NULL; 9147 return err; 9148 } 9149 9150 int netdev_offload_xstats_enable(struct net_device *dev, 9151 enum netdev_offload_xstats_type type, 9152 struct netlink_ext_ack *extack) 9153 { 9154 ASSERT_RTNL(); 9155 9156 if (netdev_offload_xstats_enabled(dev, type)) 9157 return -EALREADY; 9158 9159 switch (type) { 9160 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 9161 return netdev_offload_xstats_enable_l3(dev, extack); 9162 } 9163 9164 WARN_ON(1); 9165 return -EINVAL; 9166 } 9167 EXPORT_SYMBOL(netdev_offload_xstats_enable); 9168 9169 static void netdev_offload_xstats_disable_l3(struct net_device *dev) 9170 { 9171 struct netdev_notifier_offload_xstats_info info = { 9172 .info.dev = dev, 9173 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 9174 }; 9175 9176 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE, 9177 &info.info); 9178 kfree(dev->offload_xstats_l3); 9179 dev->offload_xstats_l3 = NULL; 9180 } 9181 9182 int netdev_offload_xstats_disable(struct net_device *dev, 9183 enum netdev_offload_xstats_type type) 9184 { 9185 ASSERT_RTNL(); 9186 9187 if (!netdev_offload_xstats_enabled(dev, type)) 9188 return -EALREADY; 9189 9190 switch (type) { 9191 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 9192 netdev_offload_xstats_disable_l3(dev); 9193 return 0; 9194 } 9195 9196 WARN_ON(1); 9197 return -EINVAL; 9198 } 9199 EXPORT_SYMBOL(netdev_offload_xstats_disable); 9200 9201 static void netdev_offload_xstats_disable_all(struct net_device *dev) 9202 { 9203 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3); 9204 } 9205 9206 static struct rtnl_hw_stats64 * 9207 netdev_offload_xstats_get_ptr(const struct net_device *dev, 9208 enum netdev_offload_xstats_type type) 9209 { 9210 switch (type) { 9211 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 9212 return dev->offload_xstats_l3; 9213 } 9214 9215 WARN_ON(1); 9216 return NULL; 9217 } 9218 9219 bool netdev_offload_xstats_enabled(const struct net_device *dev, 9220 enum netdev_offload_xstats_type type) 9221 { 9222 ASSERT_RTNL(); 9223 9224 return netdev_offload_xstats_get_ptr(dev, type); 9225 } 9226 EXPORT_SYMBOL(netdev_offload_xstats_enabled); 9227 9228 struct netdev_notifier_offload_xstats_ru { 9229 bool used; 9230 }; 9231 9232 struct netdev_notifier_offload_xstats_rd { 9233 struct rtnl_hw_stats64 stats; 9234 bool used; 9235 }; 9236 9237 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest, 9238 const struct rtnl_hw_stats64 *src) 9239 { 9240 dest->rx_packets += src->rx_packets; 9241 dest->tx_packets += src->tx_packets; 9242 dest->rx_bytes += src->rx_bytes; 9243 dest->tx_bytes += src->tx_bytes; 9244 dest->rx_errors += src->rx_errors; 9245 dest->tx_errors += src->tx_errors; 9246 dest->rx_dropped += src->rx_dropped; 9247 dest->tx_dropped += src->tx_dropped; 9248 dest->multicast += src->multicast; 9249 } 9250 9251 static int netdev_offload_xstats_get_used(struct net_device *dev, 9252 enum netdev_offload_xstats_type type, 9253 bool *p_used, 9254 struct netlink_ext_ack *extack) 9255 { 9256 struct netdev_notifier_offload_xstats_ru report_used = {}; 9257 struct netdev_notifier_offload_xstats_info info = { 9258 .info.dev = dev, 9259 .info.extack = extack, 9260 .type = type, 9261 .report_used = &report_used, 9262 }; 9263 int rc; 9264 9265 WARN_ON(!netdev_offload_xstats_enabled(dev, type)); 9266 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED, 9267 &info.info); 9268 *p_used = report_used.used; 9269 return notifier_to_errno(rc); 9270 } 9271 9272 static int netdev_offload_xstats_get_stats(struct net_device *dev, 9273 enum netdev_offload_xstats_type type, 9274 struct rtnl_hw_stats64 *p_stats, 9275 bool *p_used, 9276 struct netlink_ext_ack *extack) 9277 { 9278 struct netdev_notifier_offload_xstats_rd report_delta = {}; 9279 struct netdev_notifier_offload_xstats_info info = { 9280 .info.dev = dev, 9281 .info.extack = extack, 9282 .type = type, 9283 .report_delta = &report_delta, 9284 }; 9285 struct rtnl_hw_stats64 *stats; 9286 int rc; 9287 9288 stats = netdev_offload_xstats_get_ptr(dev, type); 9289 if (WARN_ON(!stats)) 9290 return -EINVAL; 9291 9292 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA, 9293 &info.info); 9294 9295 /* Cache whatever we got, even if there was an error, otherwise the 9296 * successful stats retrievals would get lost. 9297 */ 9298 netdev_hw_stats64_add(stats, &report_delta.stats); 9299 9300 if (p_stats) 9301 *p_stats = *stats; 9302 *p_used = report_delta.used; 9303 9304 return notifier_to_errno(rc); 9305 } 9306 9307 int netdev_offload_xstats_get(struct net_device *dev, 9308 enum netdev_offload_xstats_type type, 9309 struct rtnl_hw_stats64 *p_stats, bool *p_used, 9310 struct netlink_ext_ack *extack) 9311 { 9312 ASSERT_RTNL(); 9313 9314 if (p_stats) 9315 return netdev_offload_xstats_get_stats(dev, type, p_stats, 9316 p_used, extack); 9317 else 9318 return netdev_offload_xstats_get_used(dev, type, p_used, 9319 extack); 9320 } 9321 EXPORT_SYMBOL(netdev_offload_xstats_get); 9322 9323 void 9324 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta, 9325 const struct rtnl_hw_stats64 *stats) 9326 { 9327 report_delta->used = true; 9328 netdev_hw_stats64_add(&report_delta->stats, stats); 9329 } 9330 EXPORT_SYMBOL(netdev_offload_xstats_report_delta); 9331 9332 void 9333 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used) 9334 { 9335 report_used->used = true; 9336 } 9337 EXPORT_SYMBOL(netdev_offload_xstats_report_used); 9338 9339 void netdev_offload_xstats_push_delta(struct net_device *dev, 9340 enum netdev_offload_xstats_type type, 9341 const struct rtnl_hw_stats64 *p_stats) 9342 { 9343 struct rtnl_hw_stats64 *stats; 9344 9345 ASSERT_RTNL(); 9346 9347 stats = netdev_offload_xstats_get_ptr(dev, type); 9348 if (WARN_ON(!stats)) 9349 return; 9350 9351 netdev_hw_stats64_add(stats, p_stats); 9352 } 9353 EXPORT_SYMBOL(netdev_offload_xstats_push_delta); 9354 9355 /** 9356 * netdev_get_xmit_slave - Get the xmit slave of master device 9357 * @dev: device 9358 * @skb: The packet 9359 * @all_slaves: assume all the slaves are active 9360 * 9361 * The reference counters are not incremented so the caller must be 9362 * careful with locks. The caller must hold RCU lock. 9363 * %NULL is returned if no slave is found. 9364 */ 9365 9366 struct net_device *netdev_get_xmit_slave(struct net_device *dev, 9367 struct sk_buff *skb, 9368 bool all_slaves) 9369 { 9370 const struct net_device_ops *ops = dev->netdev_ops; 9371 9372 if (!ops->ndo_get_xmit_slave) 9373 return NULL; 9374 return ops->ndo_get_xmit_slave(dev, skb, all_slaves); 9375 } 9376 EXPORT_SYMBOL(netdev_get_xmit_slave); 9377 9378 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev, 9379 struct sock *sk) 9380 { 9381 const struct net_device_ops *ops = dev->netdev_ops; 9382 9383 if (!ops->ndo_sk_get_lower_dev) 9384 return NULL; 9385 return ops->ndo_sk_get_lower_dev(dev, sk); 9386 } 9387 9388 /** 9389 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket 9390 * @dev: device 9391 * @sk: the socket 9392 * 9393 * %NULL is returned if no lower device is found. 9394 */ 9395 9396 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev, 9397 struct sock *sk) 9398 { 9399 struct net_device *lower; 9400 9401 lower = netdev_sk_get_lower_dev(dev, sk); 9402 while (lower) { 9403 dev = lower; 9404 lower = netdev_sk_get_lower_dev(dev, sk); 9405 } 9406 9407 return dev; 9408 } 9409 EXPORT_SYMBOL(netdev_sk_get_lowest_dev); 9410 9411 static void netdev_adjacent_add_links(struct net_device *dev) 9412 { 9413 struct netdev_adjacent *iter; 9414 9415 struct net *net = dev_net(dev); 9416 9417 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9418 if (!net_eq(net, dev_net(iter->dev))) 9419 continue; 9420 netdev_adjacent_sysfs_add(iter->dev, dev, 9421 &iter->dev->adj_list.lower); 9422 netdev_adjacent_sysfs_add(dev, iter->dev, 9423 &dev->adj_list.upper); 9424 } 9425 9426 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9427 if (!net_eq(net, dev_net(iter->dev))) 9428 continue; 9429 netdev_adjacent_sysfs_add(iter->dev, dev, 9430 &iter->dev->adj_list.upper); 9431 netdev_adjacent_sysfs_add(dev, iter->dev, 9432 &dev->adj_list.lower); 9433 } 9434 } 9435 9436 static void netdev_adjacent_del_links(struct net_device *dev) 9437 { 9438 struct netdev_adjacent *iter; 9439 9440 struct net *net = dev_net(dev); 9441 9442 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9443 if (!net_eq(net, dev_net(iter->dev))) 9444 continue; 9445 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9446 &iter->dev->adj_list.lower); 9447 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9448 &dev->adj_list.upper); 9449 } 9450 9451 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9452 if (!net_eq(net, dev_net(iter->dev))) 9453 continue; 9454 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9455 &iter->dev->adj_list.upper); 9456 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9457 &dev->adj_list.lower); 9458 } 9459 } 9460 9461 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname) 9462 { 9463 struct netdev_adjacent *iter; 9464 9465 struct net *net = dev_net(dev); 9466 9467 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9468 if (!net_eq(net, dev_net(iter->dev))) 9469 continue; 9470 netdev_adjacent_sysfs_del(iter->dev, oldname, 9471 &iter->dev->adj_list.lower); 9472 netdev_adjacent_sysfs_add(iter->dev, dev, 9473 &iter->dev->adj_list.lower); 9474 } 9475 9476 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9477 if (!net_eq(net, dev_net(iter->dev))) 9478 continue; 9479 netdev_adjacent_sysfs_del(iter->dev, oldname, 9480 &iter->dev->adj_list.upper); 9481 netdev_adjacent_sysfs_add(iter->dev, dev, 9482 &iter->dev->adj_list.upper); 9483 } 9484 } 9485 9486 void *netdev_lower_dev_get_private(struct net_device *dev, 9487 struct net_device *lower_dev) 9488 { 9489 struct netdev_adjacent *lower; 9490 9491 if (!lower_dev) 9492 return NULL; 9493 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower); 9494 if (!lower) 9495 return NULL; 9496 9497 return lower->private; 9498 } 9499 EXPORT_SYMBOL(netdev_lower_dev_get_private); 9500 9501 9502 /** 9503 * netdev_lower_state_changed - Dispatch event about lower device state change 9504 * @lower_dev: device 9505 * @lower_state_info: state to dispatch 9506 * 9507 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info. 9508 * The caller must hold the RTNL lock. 9509 */ 9510 void netdev_lower_state_changed(struct net_device *lower_dev, 9511 void *lower_state_info) 9512 { 9513 struct netdev_notifier_changelowerstate_info changelowerstate_info = { 9514 .info.dev = lower_dev, 9515 }; 9516 9517 ASSERT_RTNL(); 9518 changelowerstate_info.lower_state_info = lower_state_info; 9519 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, 9520 &changelowerstate_info.info); 9521 } 9522 EXPORT_SYMBOL(netdev_lower_state_changed); 9523 9524 static void dev_change_rx_flags(struct net_device *dev, int flags) 9525 { 9526 const struct net_device_ops *ops = dev->netdev_ops; 9527 9528 if (ops->ndo_change_rx_flags) 9529 ops->ndo_change_rx_flags(dev, flags); 9530 } 9531 9532 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify) 9533 { 9534 unsigned int old_flags = dev->flags; 9535 unsigned int promiscuity, flags; 9536 kuid_t uid; 9537 kgid_t gid; 9538 9539 ASSERT_RTNL(); 9540 9541 promiscuity = dev->promiscuity + inc; 9542 if (promiscuity == 0) { 9543 /* 9544 * Avoid overflow. 9545 * If inc causes overflow, untouch promisc and return error. 9546 */ 9547 if (unlikely(inc > 0)) { 9548 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n"); 9549 return -EOVERFLOW; 9550 } 9551 flags = old_flags & ~IFF_PROMISC; 9552 } else { 9553 flags = old_flags | IFF_PROMISC; 9554 } 9555 WRITE_ONCE(dev->promiscuity, promiscuity); 9556 if (flags != old_flags) { 9557 WRITE_ONCE(dev->flags, flags); 9558 netdev_info(dev, "%s promiscuous mode\n", 9559 dev->flags & IFF_PROMISC ? "entered" : "left"); 9560 if (audit_enabled) { 9561 current_uid_gid(&uid, &gid); 9562 audit_log(audit_context(), GFP_ATOMIC, 9563 AUDIT_ANOM_PROMISCUOUS, 9564 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u", 9565 dev->name, (dev->flags & IFF_PROMISC), 9566 (old_flags & IFF_PROMISC), 9567 from_kuid(&init_user_ns, audit_get_loginuid(current)), 9568 from_kuid(&init_user_ns, uid), 9569 from_kgid(&init_user_ns, gid), 9570 audit_get_sessionid(current)); 9571 } 9572 9573 dev_change_rx_flags(dev, IFF_PROMISC); 9574 } 9575 if (notify) { 9576 /* The ops lock is only required to ensure consistent locking 9577 * for `NETDEV_CHANGE` notifiers. This function is sometimes 9578 * called without the lock, even for devices that are ops 9579 * locked, such as in `dev_uc_sync_multiple` when using 9580 * bonding or teaming. 9581 */ 9582 netdev_ops_assert_locked(dev); 9583 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL); 9584 } 9585 return 0; 9586 } 9587 9588 int netif_set_promiscuity(struct net_device *dev, int inc) 9589 { 9590 unsigned int old_flags = dev->flags; 9591 int err; 9592 9593 err = __dev_set_promiscuity(dev, inc, true); 9594 if (err < 0) 9595 return err; 9596 if (dev->flags != old_flags) 9597 dev_set_rx_mode(dev); 9598 return err; 9599 } 9600 9601 int netif_set_allmulti(struct net_device *dev, int inc, bool notify) 9602 { 9603 unsigned int old_flags = dev->flags, old_gflags = dev->gflags; 9604 unsigned int allmulti, flags; 9605 9606 ASSERT_RTNL(); 9607 9608 allmulti = dev->allmulti + inc; 9609 if (allmulti == 0) { 9610 /* 9611 * Avoid overflow. 9612 * If inc causes overflow, untouch allmulti and return error. 9613 */ 9614 if (unlikely(inc > 0)) { 9615 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n"); 9616 return -EOVERFLOW; 9617 } 9618 flags = old_flags & ~IFF_ALLMULTI; 9619 } else { 9620 flags = old_flags | IFF_ALLMULTI; 9621 } 9622 WRITE_ONCE(dev->allmulti, allmulti); 9623 if (flags != old_flags) { 9624 WRITE_ONCE(dev->flags, flags); 9625 netdev_info(dev, "%s allmulticast mode\n", 9626 dev->flags & IFF_ALLMULTI ? "entered" : "left"); 9627 dev_change_rx_flags(dev, IFF_ALLMULTI); 9628 dev_set_rx_mode(dev); 9629 if (notify) 9630 __dev_notify_flags(dev, old_flags, 9631 dev->gflags ^ old_gflags, 0, NULL); 9632 } 9633 return 0; 9634 } 9635 9636 /* 9637 * Upload unicast and multicast address lists to device and 9638 * configure RX filtering. When the device doesn't support unicast 9639 * filtering it is put in promiscuous mode while unicast addresses 9640 * are present. 9641 */ 9642 void __dev_set_rx_mode(struct net_device *dev) 9643 { 9644 const struct net_device_ops *ops = dev->netdev_ops; 9645 9646 /* dev_open will call this function so the list will stay sane. */ 9647 if (!(dev->flags&IFF_UP)) 9648 return; 9649 9650 if (!netif_device_present(dev)) 9651 return; 9652 9653 if (!(dev->priv_flags & IFF_UNICAST_FLT)) { 9654 /* Unicast addresses changes may only happen under the rtnl, 9655 * therefore calling __dev_set_promiscuity here is safe. 9656 */ 9657 if (!netdev_uc_empty(dev) && !dev->uc_promisc) { 9658 __dev_set_promiscuity(dev, 1, false); 9659 dev->uc_promisc = true; 9660 } else if (netdev_uc_empty(dev) && dev->uc_promisc) { 9661 __dev_set_promiscuity(dev, -1, false); 9662 dev->uc_promisc = false; 9663 } 9664 } 9665 9666 if (ops->ndo_set_rx_mode) 9667 ops->ndo_set_rx_mode(dev); 9668 } 9669 9670 void dev_set_rx_mode(struct net_device *dev) 9671 { 9672 netif_addr_lock_bh(dev); 9673 __dev_set_rx_mode(dev); 9674 netif_addr_unlock_bh(dev); 9675 } 9676 9677 /** 9678 * netif_get_flags() - get flags reported to userspace 9679 * @dev: device 9680 * 9681 * Get the combination of flag bits exported through APIs to userspace. 9682 */ 9683 unsigned int netif_get_flags(const struct net_device *dev) 9684 { 9685 unsigned int flags; 9686 9687 flags = (READ_ONCE(dev->flags) & ~(IFF_PROMISC | 9688 IFF_ALLMULTI | 9689 IFF_RUNNING | 9690 IFF_LOWER_UP | 9691 IFF_DORMANT)) | 9692 (READ_ONCE(dev->gflags) & (IFF_PROMISC | 9693 IFF_ALLMULTI)); 9694 9695 if (netif_running(dev)) { 9696 if (netif_oper_up(dev)) 9697 flags |= IFF_RUNNING; 9698 if (netif_carrier_ok(dev)) 9699 flags |= IFF_LOWER_UP; 9700 if (netif_dormant(dev)) 9701 flags |= IFF_DORMANT; 9702 } 9703 9704 return flags; 9705 } 9706 EXPORT_SYMBOL(netif_get_flags); 9707 9708 int __dev_change_flags(struct net_device *dev, unsigned int flags, 9709 struct netlink_ext_ack *extack) 9710 { 9711 unsigned int old_flags = dev->flags; 9712 int ret; 9713 9714 ASSERT_RTNL(); 9715 9716 /* 9717 * Set the flags on our device. 9718 */ 9719 9720 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP | 9721 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL | 9722 IFF_AUTOMEDIA)) | 9723 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC | 9724 IFF_ALLMULTI)); 9725 9726 /* 9727 * Load in the correct multicast list now the flags have changed. 9728 */ 9729 9730 if ((old_flags ^ flags) & IFF_MULTICAST) 9731 dev_change_rx_flags(dev, IFF_MULTICAST); 9732 9733 dev_set_rx_mode(dev); 9734 9735 /* 9736 * Have we downed the interface. We handle IFF_UP ourselves 9737 * according to user attempts to set it, rather than blindly 9738 * setting it. 9739 */ 9740 9741 ret = 0; 9742 if ((old_flags ^ flags) & IFF_UP) { 9743 if (old_flags & IFF_UP) 9744 __dev_close(dev); 9745 else 9746 ret = __dev_open(dev, extack); 9747 } 9748 9749 if ((flags ^ dev->gflags) & IFF_PROMISC) { 9750 int inc = (flags & IFF_PROMISC) ? 1 : -1; 9751 old_flags = dev->flags; 9752 9753 dev->gflags ^= IFF_PROMISC; 9754 9755 if (__dev_set_promiscuity(dev, inc, false) >= 0) 9756 if (dev->flags != old_flags) 9757 dev_set_rx_mode(dev); 9758 } 9759 9760 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI 9761 * is important. Some (broken) drivers set IFF_PROMISC, when 9762 * IFF_ALLMULTI is requested not asking us and not reporting. 9763 */ 9764 if ((flags ^ dev->gflags) & IFF_ALLMULTI) { 9765 int inc = (flags & IFF_ALLMULTI) ? 1 : -1; 9766 9767 dev->gflags ^= IFF_ALLMULTI; 9768 netif_set_allmulti(dev, inc, false); 9769 } 9770 9771 return ret; 9772 } 9773 9774 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags, 9775 unsigned int gchanges, u32 portid, 9776 const struct nlmsghdr *nlh) 9777 { 9778 unsigned int changes = dev->flags ^ old_flags; 9779 9780 if (gchanges) 9781 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh); 9782 9783 if (changes & IFF_UP) { 9784 if (dev->flags & IFF_UP) 9785 call_netdevice_notifiers(NETDEV_UP, dev); 9786 else 9787 call_netdevice_notifiers(NETDEV_DOWN, dev); 9788 } 9789 9790 if (dev->flags & IFF_UP && 9791 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) { 9792 struct netdev_notifier_change_info change_info = { 9793 .info = { 9794 .dev = dev, 9795 }, 9796 .flags_changed = changes, 9797 }; 9798 9799 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info); 9800 } 9801 } 9802 9803 int netif_change_flags(struct net_device *dev, unsigned int flags, 9804 struct netlink_ext_ack *extack) 9805 { 9806 int ret; 9807 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags; 9808 9809 ret = __dev_change_flags(dev, flags, extack); 9810 if (ret < 0) 9811 return ret; 9812 9813 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags); 9814 __dev_notify_flags(dev, old_flags, changes, 0, NULL); 9815 return ret; 9816 } 9817 9818 int __netif_set_mtu(struct net_device *dev, int new_mtu) 9819 { 9820 const struct net_device_ops *ops = dev->netdev_ops; 9821 9822 if (ops->ndo_change_mtu) 9823 return ops->ndo_change_mtu(dev, new_mtu); 9824 9825 /* Pairs with all the lockless reads of dev->mtu in the stack */ 9826 WRITE_ONCE(dev->mtu, new_mtu); 9827 return 0; 9828 } 9829 EXPORT_SYMBOL_NS_GPL(__netif_set_mtu, "NETDEV_INTERNAL"); 9830 9831 int dev_validate_mtu(struct net_device *dev, int new_mtu, 9832 struct netlink_ext_ack *extack) 9833 { 9834 /* MTU must be positive, and in range */ 9835 if (new_mtu < 0 || new_mtu < dev->min_mtu) { 9836 NL_SET_ERR_MSG(extack, "mtu less than device minimum"); 9837 return -EINVAL; 9838 } 9839 9840 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) { 9841 NL_SET_ERR_MSG(extack, "mtu greater than device maximum"); 9842 return -EINVAL; 9843 } 9844 return 0; 9845 } 9846 9847 /** 9848 * netif_set_mtu_ext() - Change maximum transfer unit 9849 * @dev: device 9850 * @new_mtu: new transfer unit 9851 * @extack: netlink extended ack 9852 * 9853 * Change the maximum transfer size of the network device. 9854 * 9855 * Return: 0 on success, -errno on failure. 9856 */ 9857 int netif_set_mtu_ext(struct net_device *dev, int new_mtu, 9858 struct netlink_ext_ack *extack) 9859 { 9860 int err, orig_mtu; 9861 9862 netdev_ops_assert_locked(dev); 9863 9864 if (new_mtu == dev->mtu) 9865 return 0; 9866 9867 err = dev_validate_mtu(dev, new_mtu, extack); 9868 if (err) 9869 return err; 9870 9871 if (!netif_device_present(dev)) 9872 return -ENODEV; 9873 9874 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev); 9875 err = notifier_to_errno(err); 9876 if (err) 9877 return err; 9878 9879 orig_mtu = dev->mtu; 9880 err = __netif_set_mtu(dev, new_mtu); 9881 9882 if (!err) { 9883 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9884 orig_mtu); 9885 err = notifier_to_errno(err); 9886 if (err) { 9887 /* setting mtu back and notifying everyone again, 9888 * so that they have a chance to revert changes. 9889 */ 9890 __netif_set_mtu(dev, orig_mtu); 9891 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9892 new_mtu); 9893 } 9894 } 9895 return err; 9896 } 9897 9898 int netif_set_mtu(struct net_device *dev, int new_mtu) 9899 { 9900 struct netlink_ext_ack extack; 9901 int err; 9902 9903 memset(&extack, 0, sizeof(extack)); 9904 err = netif_set_mtu_ext(dev, new_mtu, &extack); 9905 if (err && extack._msg) 9906 net_err_ratelimited("%s: %s\n", dev->name, extack._msg); 9907 return err; 9908 } 9909 EXPORT_SYMBOL(netif_set_mtu); 9910 9911 int netif_change_tx_queue_len(struct net_device *dev, unsigned long new_len) 9912 { 9913 unsigned int orig_len = dev->tx_queue_len; 9914 int res; 9915 9916 if (new_len != (unsigned int)new_len) 9917 return -ERANGE; 9918 9919 if (new_len != orig_len) { 9920 WRITE_ONCE(dev->tx_queue_len, new_len); 9921 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev); 9922 res = notifier_to_errno(res); 9923 if (res) 9924 goto err_rollback; 9925 res = dev_qdisc_change_tx_queue_len(dev); 9926 if (res) 9927 goto err_rollback; 9928 } 9929 9930 return 0; 9931 9932 err_rollback: 9933 netdev_err(dev, "refused to change device tx_queue_len\n"); 9934 WRITE_ONCE(dev->tx_queue_len, orig_len); 9935 return res; 9936 } 9937 9938 void netif_set_group(struct net_device *dev, int new_group) 9939 { 9940 dev->group = new_group; 9941 } 9942 9943 /** 9944 * netif_pre_changeaddr_notify() - Call NETDEV_PRE_CHANGEADDR. 9945 * @dev: device 9946 * @addr: new address 9947 * @extack: netlink extended ack 9948 * 9949 * Return: 0 on success, -errno on failure. 9950 */ 9951 int netif_pre_changeaddr_notify(struct net_device *dev, const char *addr, 9952 struct netlink_ext_ack *extack) 9953 { 9954 struct netdev_notifier_pre_changeaddr_info info = { 9955 .info.dev = dev, 9956 .info.extack = extack, 9957 .dev_addr = addr, 9958 }; 9959 int rc; 9960 9961 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info); 9962 return notifier_to_errno(rc); 9963 } 9964 EXPORT_SYMBOL_NS_GPL(netif_pre_changeaddr_notify, "NETDEV_INTERNAL"); 9965 9966 int netif_set_mac_address(struct net_device *dev, struct sockaddr_storage *ss, 9967 struct netlink_ext_ack *extack) 9968 { 9969 const struct net_device_ops *ops = dev->netdev_ops; 9970 int err; 9971 9972 if (!ops->ndo_set_mac_address) 9973 return -EOPNOTSUPP; 9974 if (ss->ss_family != dev->type) 9975 return -EINVAL; 9976 if (!netif_device_present(dev)) 9977 return -ENODEV; 9978 err = netif_pre_changeaddr_notify(dev, ss->__data, extack); 9979 if (err) 9980 return err; 9981 if (memcmp(dev->dev_addr, ss->__data, dev->addr_len)) { 9982 err = ops->ndo_set_mac_address(dev, ss); 9983 if (err) 9984 return err; 9985 } 9986 dev->addr_assign_type = NET_ADDR_SET; 9987 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev); 9988 add_device_randomness(dev->dev_addr, dev->addr_len); 9989 return 0; 9990 } 9991 9992 DECLARE_RWSEM(dev_addr_sem); 9993 9994 /* "sa" is a true struct sockaddr with limited "sa_data" member. */ 9995 int netif_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name) 9996 { 9997 size_t size = sizeof(sa->sa_data); 9998 struct net_device *dev; 9999 int ret = 0; 10000 10001 down_read(&dev_addr_sem); 10002 rcu_read_lock(); 10003 10004 dev = dev_get_by_name_rcu(net, dev_name); 10005 if (!dev) { 10006 ret = -ENODEV; 10007 goto unlock; 10008 } 10009 if (!dev->addr_len) 10010 memset(sa->sa_data, 0, size); 10011 else 10012 memcpy(sa->sa_data, dev->dev_addr, 10013 min_t(size_t, size, dev->addr_len)); 10014 sa->sa_family = dev->type; 10015 10016 unlock: 10017 rcu_read_unlock(); 10018 up_read(&dev_addr_sem); 10019 return ret; 10020 } 10021 EXPORT_SYMBOL_NS_GPL(netif_get_mac_address, "NETDEV_INTERNAL"); 10022 10023 int netif_change_carrier(struct net_device *dev, bool new_carrier) 10024 { 10025 const struct net_device_ops *ops = dev->netdev_ops; 10026 10027 if (!ops->ndo_change_carrier) 10028 return -EOPNOTSUPP; 10029 if (!netif_device_present(dev)) 10030 return -ENODEV; 10031 return ops->ndo_change_carrier(dev, new_carrier); 10032 } 10033 10034 /** 10035 * dev_get_phys_port_id - Get device physical port ID 10036 * @dev: device 10037 * @ppid: port ID 10038 * 10039 * Get device physical port ID 10040 */ 10041 int dev_get_phys_port_id(struct net_device *dev, 10042 struct netdev_phys_item_id *ppid) 10043 { 10044 const struct net_device_ops *ops = dev->netdev_ops; 10045 10046 if (!ops->ndo_get_phys_port_id) 10047 return -EOPNOTSUPP; 10048 return ops->ndo_get_phys_port_id(dev, ppid); 10049 } 10050 10051 /** 10052 * dev_get_phys_port_name - Get device physical port name 10053 * @dev: device 10054 * @name: port name 10055 * @len: limit of bytes to copy to name 10056 * 10057 * Get device physical port name 10058 */ 10059 int dev_get_phys_port_name(struct net_device *dev, 10060 char *name, size_t len) 10061 { 10062 const struct net_device_ops *ops = dev->netdev_ops; 10063 int err; 10064 10065 if (ops->ndo_get_phys_port_name) { 10066 err = ops->ndo_get_phys_port_name(dev, name, len); 10067 if (err != -EOPNOTSUPP) 10068 return err; 10069 } 10070 return devlink_compat_phys_port_name_get(dev, name, len); 10071 } 10072 10073 /** 10074 * netif_get_port_parent_id() - Get the device's port parent identifier 10075 * @dev: network device 10076 * @ppid: pointer to a storage for the port's parent identifier 10077 * @recurse: allow/disallow recursion to lower devices 10078 * 10079 * Get the devices's port parent identifier. 10080 * 10081 * Return: 0 on success, -errno on failure. 10082 */ 10083 int netif_get_port_parent_id(struct net_device *dev, 10084 struct netdev_phys_item_id *ppid, bool recurse) 10085 { 10086 const struct net_device_ops *ops = dev->netdev_ops; 10087 struct netdev_phys_item_id first = { }; 10088 struct net_device *lower_dev; 10089 struct list_head *iter; 10090 int err; 10091 10092 if (ops->ndo_get_port_parent_id) { 10093 err = ops->ndo_get_port_parent_id(dev, ppid); 10094 if (err != -EOPNOTSUPP) 10095 return err; 10096 } 10097 10098 err = devlink_compat_switch_id_get(dev, ppid); 10099 if (!recurse || err != -EOPNOTSUPP) 10100 return err; 10101 10102 netdev_for_each_lower_dev(dev, lower_dev, iter) { 10103 err = netif_get_port_parent_id(lower_dev, ppid, true); 10104 if (err) 10105 break; 10106 if (!first.id_len) 10107 first = *ppid; 10108 else if (memcmp(&first, ppid, sizeof(*ppid))) 10109 return -EOPNOTSUPP; 10110 } 10111 10112 return err; 10113 } 10114 EXPORT_SYMBOL(netif_get_port_parent_id); 10115 10116 /** 10117 * netdev_port_same_parent_id - Indicate if two network devices have 10118 * the same port parent identifier 10119 * @a: first network device 10120 * @b: second network device 10121 */ 10122 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b) 10123 { 10124 struct netdev_phys_item_id a_id = { }; 10125 struct netdev_phys_item_id b_id = { }; 10126 10127 if (netif_get_port_parent_id(a, &a_id, true) || 10128 netif_get_port_parent_id(b, &b_id, true)) 10129 return false; 10130 10131 return netdev_phys_item_id_same(&a_id, &b_id); 10132 } 10133 EXPORT_SYMBOL(netdev_port_same_parent_id); 10134 10135 int netif_change_proto_down(struct net_device *dev, bool proto_down) 10136 { 10137 if (!dev->change_proto_down) 10138 return -EOPNOTSUPP; 10139 if (!netif_device_present(dev)) 10140 return -ENODEV; 10141 if (proto_down) 10142 netif_carrier_off(dev); 10143 else 10144 netif_carrier_on(dev); 10145 WRITE_ONCE(dev->proto_down, proto_down); 10146 return 0; 10147 } 10148 10149 /** 10150 * netdev_change_proto_down_reason_locked - proto down reason 10151 * 10152 * @dev: device 10153 * @mask: proto down mask 10154 * @value: proto down value 10155 */ 10156 void netdev_change_proto_down_reason_locked(struct net_device *dev, 10157 unsigned long mask, u32 value) 10158 { 10159 u32 proto_down_reason; 10160 int b; 10161 10162 if (!mask) { 10163 proto_down_reason = value; 10164 } else { 10165 proto_down_reason = dev->proto_down_reason; 10166 for_each_set_bit(b, &mask, 32) { 10167 if (value & (1 << b)) 10168 proto_down_reason |= BIT(b); 10169 else 10170 proto_down_reason &= ~BIT(b); 10171 } 10172 } 10173 WRITE_ONCE(dev->proto_down_reason, proto_down_reason); 10174 } 10175 10176 struct bpf_xdp_link { 10177 struct bpf_link link; 10178 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */ 10179 int flags; 10180 }; 10181 10182 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags) 10183 { 10184 if (flags & XDP_FLAGS_HW_MODE) 10185 return XDP_MODE_HW; 10186 if (flags & XDP_FLAGS_DRV_MODE) 10187 return XDP_MODE_DRV; 10188 if (flags & XDP_FLAGS_SKB_MODE) 10189 return XDP_MODE_SKB; 10190 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB; 10191 } 10192 10193 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode) 10194 { 10195 switch (mode) { 10196 case XDP_MODE_SKB: 10197 return generic_xdp_install; 10198 case XDP_MODE_DRV: 10199 case XDP_MODE_HW: 10200 return dev->netdev_ops->ndo_bpf; 10201 default: 10202 return NULL; 10203 } 10204 } 10205 10206 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev, 10207 enum bpf_xdp_mode mode) 10208 { 10209 return dev->xdp_state[mode].link; 10210 } 10211 10212 static struct bpf_prog *dev_xdp_prog(struct net_device *dev, 10213 enum bpf_xdp_mode mode) 10214 { 10215 struct bpf_xdp_link *link = dev_xdp_link(dev, mode); 10216 10217 if (link) 10218 return link->link.prog; 10219 return dev->xdp_state[mode].prog; 10220 } 10221 10222 u8 dev_xdp_prog_count(struct net_device *dev) 10223 { 10224 u8 count = 0; 10225 int i; 10226 10227 for (i = 0; i < __MAX_XDP_MODE; i++) 10228 if (dev->xdp_state[i].prog || dev->xdp_state[i].link) 10229 count++; 10230 return count; 10231 } 10232 EXPORT_SYMBOL_GPL(dev_xdp_prog_count); 10233 10234 u8 dev_xdp_sb_prog_count(struct net_device *dev) 10235 { 10236 u8 count = 0; 10237 int i; 10238 10239 for (i = 0; i < __MAX_XDP_MODE; i++) 10240 if (dev->xdp_state[i].prog && 10241 !dev->xdp_state[i].prog->aux->xdp_has_frags) 10242 count++; 10243 return count; 10244 } 10245 10246 int netif_xdp_propagate(struct net_device *dev, struct netdev_bpf *bpf) 10247 { 10248 if (!dev->netdev_ops->ndo_bpf) 10249 return -EOPNOTSUPP; 10250 10251 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 10252 bpf->command == XDP_SETUP_PROG && 10253 bpf->prog && !bpf->prog->aux->xdp_has_frags) { 10254 NL_SET_ERR_MSG(bpf->extack, 10255 "unable to propagate XDP to device using tcp-data-split"); 10256 return -EBUSY; 10257 } 10258 10259 if (dev_get_min_mp_channel_count(dev)) { 10260 NL_SET_ERR_MSG(bpf->extack, "unable to propagate XDP to device using memory provider"); 10261 return -EBUSY; 10262 } 10263 10264 return dev->netdev_ops->ndo_bpf(dev, bpf); 10265 } 10266 EXPORT_SYMBOL_GPL(netif_xdp_propagate); 10267 10268 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode) 10269 { 10270 struct bpf_prog *prog = dev_xdp_prog(dev, mode); 10271 10272 return prog ? prog->aux->id : 0; 10273 } 10274 10275 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode, 10276 struct bpf_xdp_link *link) 10277 { 10278 dev->xdp_state[mode].link = link; 10279 dev->xdp_state[mode].prog = NULL; 10280 } 10281 10282 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode, 10283 struct bpf_prog *prog) 10284 { 10285 dev->xdp_state[mode].link = NULL; 10286 dev->xdp_state[mode].prog = prog; 10287 } 10288 10289 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode, 10290 bpf_op_t bpf_op, struct netlink_ext_ack *extack, 10291 u32 flags, struct bpf_prog *prog) 10292 { 10293 struct netdev_bpf xdp; 10294 int err; 10295 10296 netdev_ops_assert_locked(dev); 10297 10298 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 10299 prog && !prog->aux->xdp_has_frags) { 10300 NL_SET_ERR_MSG(extack, "unable to install XDP to device using tcp-data-split"); 10301 return -EBUSY; 10302 } 10303 10304 if (dev_get_min_mp_channel_count(dev)) { 10305 NL_SET_ERR_MSG(extack, "unable to install XDP to device using memory provider"); 10306 return -EBUSY; 10307 } 10308 10309 memset(&xdp, 0, sizeof(xdp)); 10310 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG; 10311 xdp.extack = extack; 10312 xdp.flags = flags; 10313 xdp.prog = prog; 10314 10315 /* Drivers assume refcnt is already incremented (i.e, prog pointer is 10316 * "moved" into driver), so they don't increment it on their own, but 10317 * they do decrement refcnt when program is detached or replaced. 10318 * Given net_device also owns link/prog, we need to bump refcnt here 10319 * to prevent drivers from underflowing it. 10320 */ 10321 if (prog) 10322 bpf_prog_inc(prog); 10323 err = bpf_op(dev, &xdp); 10324 if (err) { 10325 if (prog) 10326 bpf_prog_put(prog); 10327 return err; 10328 } 10329 10330 if (mode != XDP_MODE_HW) 10331 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog); 10332 10333 return 0; 10334 } 10335 10336 static void dev_xdp_uninstall(struct net_device *dev) 10337 { 10338 struct bpf_xdp_link *link; 10339 struct bpf_prog *prog; 10340 enum bpf_xdp_mode mode; 10341 bpf_op_t bpf_op; 10342 10343 ASSERT_RTNL(); 10344 10345 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) { 10346 prog = dev_xdp_prog(dev, mode); 10347 if (!prog) 10348 continue; 10349 10350 bpf_op = dev_xdp_bpf_op(dev, mode); 10351 if (!bpf_op) 10352 continue; 10353 10354 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 10355 10356 /* auto-detach link from net device */ 10357 link = dev_xdp_link(dev, mode); 10358 if (link) 10359 link->dev = NULL; 10360 else 10361 bpf_prog_put(prog); 10362 10363 dev_xdp_set_link(dev, mode, NULL); 10364 } 10365 } 10366 10367 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack, 10368 struct bpf_xdp_link *link, struct bpf_prog *new_prog, 10369 struct bpf_prog *old_prog, u32 flags) 10370 { 10371 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES); 10372 struct bpf_prog *cur_prog; 10373 struct net_device *upper; 10374 struct list_head *iter; 10375 enum bpf_xdp_mode mode; 10376 bpf_op_t bpf_op; 10377 int err; 10378 10379 ASSERT_RTNL(); 10380 10381 /* either link or prog attachment, never both */ 10382 if (link && (new_prog || old_prog)) 10383 return -EINVAL; 10384 /* link supports only XDP mode flags */ 10385 if (link && (flags & ~XDP_FLAGS_MODES)) { 10386 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment"); 10387 return -EINVAL; 10388 } 10389 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */ 10390 if (num_modes > 1) { 10391 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set"); 10392 return -EINVAL; 10393 } 10394 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */ 10395 if (!num_modes && dev_xdp_prog_count(dev) > 1) { 10396 NL_SET_ERR_MSG(extack, 10397 "More than one program loaded, unset mode is ambiguous"); 10398 return -EINVAL; 10399 } 10400 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */ 10401 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) { 10402 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified"); 10403 return -EINVAL; 10404 } 10405 10406 mode = dev_xdp_mode(dev, flags); 10407 /* can't replace attached link */ 10408 if (dev_xdp_link(dev, mode)) { 10409 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link"); 10410 return -EBUSY; 10411 } 10412 10413 /* don't allow if an upper device already has a program */ 10414 netdev_for_each_upper_dev_rcu(dev, upper, iter) { 10415 if (dev_xdp_prog_count(upper) > 0) { 10416 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program"); 10417 return -EEXIST; 10418 } 10419 } 10420 10421 cur_prog = dev_xdp_prog(dev, mode); 10422 /* can't replace attached prog with link */ 10423 if (link && cur_prog) { 10424 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link"); 10425 return -EBUSY; 10426 } 10427 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) { 10428 NL_SET_ERR_MSG(extack, "Active program does not match expected"); 10429 return -EEXIST; 10430 } 10431 10432 /* put effective new program into new_prog */ 10433 if (link) 10434 new_prog = link->link.prog; 10435 10436 if (new_prog) { 10437 bool offload = mode == XDP_MODE_HW; 10438 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB 10439 ? XDP_MODE_DRV : XDP_MODE_SKB; 10440 10441 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) { 10442 NL_SET_ERR_MSG(extack, "XDP program already attached"); 10443 return -EBUSY; 10444 } 10445 if (!offload && dev_xdp_prog(dev, other_mode)) { 10446 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time"); 10447 return -EEXIST; 10448 } 10449 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) { 10450 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported"); 10451 return -EINVAL; 10452 } 10453 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) { 10454 NL_SET_ERR_MSG(extack, "Program bound to different device"); 10455 return -EINVAL; 10456 } 10457 if (bpf_prog_is_dev_bound(new_prog->aux) && mode == XDP_MODE_SKB) { 10458 NL_SET_ERR_MSG(extack, "Can't attach device-bound programs in generic mode"); 10459 return -EINVAL; 10460 } 10461 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) { 10462 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device"); 10463 return -EINVAL; 10464 } 10465 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) { 10466 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device"); 10467 return -EINVAL; 10468 } 10469 } 10470 10471 /* don't call drivers if the effective program didn't change */ 10472 if (new_prog != cur_prog) { 10473 bpf_op = dev_xdp_bpf_op(dev, mode); 10474 if (!bpf_op) { 10475 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode"); 10476 return -EOPNOTSUPP; 10477 } 10478 10479 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog); 10480 if (err) 10481 return err; 10482 } 10483 10484 if (link) 10485 dev_xdp_set_link(dev, mode, link); 10486 else 10487 dev_xdp_set_prog(dev, mode, new_prog); 10488 if (cur_prog) 10489 bpf_prog_put(cur_prog); 10490 10491 return 0; 10492 } 10493 10494 static int dev_xdp_attach_link(struct net_device *dev, 10495 struct netlink_ext_ack *extack, 10496 struct bpf_xdp_link *link) 10497 { 10498 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags); 10499 } 10500 10501 static int dev_xdp_detach_link(struct net_device *dev, 10502 struct netlink_ext_ack *extack, 10503 struct bpf_xdp_link *link) 10504 { 10505 enum bpf_xdp_mode mode; 10506 bpf_op_t bpf_op; 10507 10508 ASSERT_RTNL(); 10509 10510 mode = dev_xdp_mode(dev, link->flags); 10511 if (dev_xdp_link(dev, mode) != link) 10512 return -EINVAL; 10513 10514 bpf_op = dev_xdp_bpf_op(dev, mode); 10515 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 10516 dev_xdp_set_link(dev, mode, NULL); 10517 return 0; 10518 } 10519 10520 static void bpf_xdp_link_release(struct bpf_link *link) 10521 { 10522 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10523 10524 rtnl_lock(); 10525 10526 /* if racing with net_device's tear down, xdp_link->dev might be 10527 * already NULL, in which case link was already auto-detached 10528 */ 10529 if (xdp_link->dev) { 10530 netdev_lock_ops(xdp_link->dev); 10531 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link)); 10532 netdev_unlock_ops(xdp_link->dev); 10533 xdp_link->dev = NULL; 10534 } 10535 10536 rtnl_unlock(); 10537 } 10538 10539 static int bpf_xdp_link_detach(struct bpf_link *link) 10540 { 10541 bpf_xdp_link_release(link); 10542 return 0; 10543 } 10544 10545 static void bpf_xdp_link_dealloc(struct bpf_link *link) 10546 { 10547 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10548 10549 kfree(xdp_link); 10550 } 10551 10552 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link, 10553 struct seq_file *seq) 10554 { 10555 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10556 u32 ifindex = 0; 10557 10558 rtnl_lock(); 10559 if (xdp_link->dev) 10560 ifindex = xdp_link->dev->ifindex; 10561 rtnl_unlock(); 10562 10563 seq_printf(seq, "ifindex:\t%u\n", ifindex); 10564 } 10565 10566 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link, 10567 struct bpf_link_info *info) 10568 { 10569 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10570 u32 ifindex = 0; 10571 10572 rtnl_lock(); 10573 if (xdp_link->dev) 10574 ifindex = xdp_link->dev->ifindex; 10575 rtnl_unlock(); 10576 10577 info->xdp.ifindex = ifindex; 10578 return 0; 10579 } 10580 10581 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog, 10582 struct bpf_prog *old_prog) 10583 { 10584 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10585 enum bpf_xdp_mode mode; 10586 bpf_op_t bpf_op; 10587 int err = 0; 10588 10589 rtnl_lock(); 10590 10591 /* link might have been auto-released already, so fail */ 10592 if (!xdp_link->dev) { 10593 err = -ENOLINK; 10594 goto out_unlock; 10595 } 10596 10597 if (old_prog && link->prog != old_prog) { 10598 err = -EPERM; 10599 goto out_unlock; 10600 } 10601 old_prog = link->prog; 10602 if (old_prog->type != new_prog->type || 10603 old_prog->expected_attach_type != new_prog->expected_attach_type) { 10604 err = -EINVAL; 10605 goto out_unlock; 10606 } 10607 10608 if (old_prog == new_prog) { 10609 /* no-op, don't disturb drivers */ 10610 bpf_prog_put(new_prog); 10611 goto out_unlock; 10612 } 10613 10614 netdev_lock_ops(xdp_link->dev); 10615 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags); 10616 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode); 10617 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL, 10618 xdp_link->flags, new_prog); 10619 netdev_unlock_ops(xdp_link->dev); 10620 if (err) 10621 goto out_unlock; 10622 10623 old_prog = xchg(&link->prog, new_prog); 10624 bpf_prog_put(old_prog); 10625 10626 out_unlock: 10627 rtnl_unlock(); 10628 return err; 10629 } 10630 10631 static const struct bpf_link_ops bpf_xdp_link_lops = { 10632 .release = bpf_xdp_link_release, 10633 .dealloc = bpf_xdp_link_dealloc, 10634 .detach = bpf_xdp_link_detach, 10635 .show_fdinfo = bpf_xdp_link_show_fdinfo, 10636 .fill_link_info = bpf_xdp_link_fill_link_info, 10637 .update_prog = bpf_xdp_link_update, 10638 }; 10639 10640 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog) 10641 { 10642 struct net *net = current->nsproxy->net_ns; 10643 struct bpf_link_primer link_primer; 10644 struct netlink_ext_ack extack = {}; 10645 struct bpf_xdp_link *link; 10646 struct net_device *dev; 10647 int err, fd; 10648 10649 rtnl_lock(); 10650 dev = dev_get_by_index(net, attr->link_create.target_ifindex); 10651 if (!dev) { 10652 rtnl_unlock(); 10653 return -EINVAL; 10654 } 10655 10656 link = kzalloc(sizeof(*link), GFP_USER); 10657 if (!link) { 10658 err = -ENOMEM; 10659 goto unlock; 10660 } 10661 10662 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog, 10663 attr->link_create.attach_type); 10664 link->dev = dev; 10665 link->flags = attr->link_create.flags; 10666 10667 err = bpf_link_prime(&link->link, &link_primer); 10668 if (err) { 10669 kfree(link); 10670 goto unlock; 10671 } 10672 10673 netdev_lock_ops(dev); 10674 err = dev_xdp_attach_link(dev, &extack, link); 10675 netdev_unlock_ops(dev); 10676 rtnl_unlock(); 10677 10678 if (err) { 10679 link->dev = NULL; 10680 bpf_link_cleanup(&link_primer); 10681 trace_bpf_xdp_link_attach_failed(extack._msg); 10682 goto out_put_dev; 10683 } 10684 10685 fd = bpf_link_settle(&link_primer); 10686 /* link itself doesn't hold dev's refcnt to not complicate shutdown */ 10687 dev_put(dev); 10688 return fd; 10689 10690 unlock: 10691 rtnl_unlock(); 10692 10693 out_put_dev: 10694 dev_put(dev); 10695 return err; 10696 } 10697 10698 /** 10699 * dev_change_xdp_fd - set or clear a bpf program for a device rx path 10700 * @dev: device 10701 * @extack: netlink extended ack 10702 * @fd: new program fd or negative value to clear 10703 * @expected_fd: old program fd that userspace expects to replace or clear 10704 * @flags: xdp-related flags 10705 * 10706 * Set or clear a bpf program for a device 10707 */ 10708 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, 10709 int fd, int expected_fd, u32 flags) 10710 { 10711 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags); 10712 struct bpf_prog *new_prog = NULL, *old_prog = NULL; 10713 int err; 10714 10715 ASSERT_RTNL(); 10716 10717 if (fd >= 0) { 10718 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP, 10719 mode != XDP_MODE_SKB); 10720 if (IS_ERR(new_prog)) 10721 return PTR_ERR(new_prog); 10722 } 10723 10724 if (expected_fd >= 0) { 10725 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP, 10726 mode != XDP_MODE_SKB); 10727 if (IS_ERR(old_prog)) { 10728 err = PTR_ERR(old_prog); 10729 old_prog = NULL; 10730 goto err_out; 10731 } 10732 } 10733 10734 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags); 10735 10736 err_out: 10737 if (err && new_prog) 10738 bpf_prog_put(new_prog); 10739 if (old_prog) 10740 bpf_prog_put(old_prog); 10741 return err; 10742 } 10743 10744 u32 dev_get_min_mp_channel_count(const struct net_device *dev) 10745 { 10746 int i; 10747 10748 netdev_ops_assert_locked(dev); 10749 10750 for (i = dev->real_num_rx_queues - 1; i >= 0; i--) 10751 if (dev->_rx[i].mp_params.mp_priv) 10752 /* The channel count is the idx plus 1. */ 10753 return i + 1; 10754 10755 return 0; 10756 } 10757 10758 /** 10759 * dev_index_reserve() - allocate an ifindex in a namespace 10760 * @net: the applicable net namespace 10761 * @ifindex: requested ifindex, pass %0 to get one allocated 10762 * 10763 * Allocate a ifindex for a new device. Caller must either use the ifindex 10764 * to store the device (via list_netdevice()) or call dev_index_release() 10765 * to give the index up. 10766 * 10767 * Return: a suitable unique value for a new device interface number or -errno. 10768 */ 10769 static int dev_index_reserve(struct net *net, u32 ifindex) 10770 { 10771 int err; 10772 10773 if (ifindex > INT_MAX) { 10774 DEBUG_NET_WARN_ON_ONCE(1); 10775 return -EINVAL; 10776 } 10777 10778 if (!ifindex) 10779 err = xa_alloc_cyclic(&net->dev_by_index, &ifindex, NULL, 10780 xa_limit_31b, &net->ifindex, GFP_KERNEL); 10781 else 10782 err = xa_insert(&net->dev_by_index, ifindex, NULL, GFP_KERNEL); 10783 if (err < 0) 10784 return err; 10785 10786 return ifindex; 10787 } 10788 10789 static void dev_index_release(struct net *net, int ifindex) 10790 { 10791 /* Expect only unused indexes, unlist_netdevice() removes the used */ 10792 WARN_ON(xa_erase(&net->dev_by_index, ifindex)); 10793 } 10794 10795 static bool from_cleanup_net(void) 10796 { 10797 #ifdef CONFIG_NET_NS 10798 return current == READ_ONCE(cleanup_net_task); 10799 #else 10800 return false; 10801 #endif 10802 } 10803 10804 /* Delayed registration/unregisteration */ 10805 LIST_HEAD(net_todo_list); 10806 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq); 10807 atomic_t dev_unreg_count = ATOMIC_INIT(0); 10808 10809 static void net_set_todo(struct net_device *dev) 10810 { 10811 list_add_tail(&dev->todo_list, &net_todo_list); 10812 } 10813 10814 static netdev_features_t netdev_sync_upper_features(struct net_device *lower, 10815 struct net_device *upper, netdev_features_t features) 10816 { 10817 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10818 netdev_features_t feature; 10819 int feature_bit; 10820 10821 for_each_netdev_feature(upper_disables, feature_bit) { 10822 feature = __NETIF_F_BIT(feature_bit); 10823 if (!(upper->wanted_features & feature) 10824 && (features & feature)) { 10825 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n", 10826 &feature, upper->name); 10827 features &= ~feature; 10828 } 10829 } 10830 10831 return features; 10832 } 10833 10834 static void netdev_sync_lower_features(struct net_device *upper, 10835 struct net_device *lower, netdev_features_t features) 10836 { 10837 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10838 netdev_features_t feature; 10839 int feature_bit; 10840 10841 for_each_netdev_feature(upper_disables, feature_bit) { 10842 feature = __NETIF_F_BIT(feature_bit); 10843 if (!(features & feature) && (lower->features & feature)) { 10844 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n", 10845 &feature, lower->name); 10846 netdev_lock_ops(lower); 10847 lower->wanted_features &= ~feature; 10848 __netdev_update_features(lower); 10849 10850 if (unlikely(lower->features & feature)) 10851 netdev_WARN(upper, "failed to disable %pNF on %s!\n", 10852 &feature, lower->name); 10853 else 10854 netdev_features_change(lower); 10855 netdev_unlock_ops(lower); 10856 } 10857 } 10858 } 10859 10860 static bool netdev_has_ip_or_hw_csum(netdev_features_t features) 10861 { 10862 netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 10863 bool ip_csum = (features & ip_csum_mask) == ip_csum_mask; 10864 bool hw_csum = features & NETIF_F_HW_CSUM; 10865 10866 return ip_csum || hw_csum; 10867 } 10868 10869 static netdev_features_t netdev_fix_features(struct net_device *dev, 10870 netdev_features_t features) 10871 { 10872 /* Fix illegal checksum combinations */ 10873 if ((features & NETIF_F_HW_CSUM) && 10874 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) { 10875 netdev_warn(dev, "mixed HW and IP checksum settings.\n"); 10876 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM); 10877 } 10878 10879 /* TSO requires that SG is present as well. */ 10880 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) { 10881 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n"); 10882 features &= ~NETIF_F_ALL_TSO; 10883 } 10884 10885 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) && 10886 !(features & NETIF_F_IP_CSUM)) { 10887 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n"); 10888 features &= ~NETIF_F_TSO; 10889 features &= ~NETIF_F_TSO_ECN; 10890 } 10891 10892 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) && 10893 !(features & NETIF_F_IPV6_CSUM)) { 10894 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n"); 10895 features &= ~NETIF_F_TSO6; 10896 } 10897 10898 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */ 10899 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO)) 10900 features &= ~NETIF_F_TSO_MANGLEID; 10901 10902 /* TSO ECN requires that TSO is present as well. */ 10903 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN) 10904 features &= ~NETIF_F_TSO_ECN; 10905 10906 /* Software GSO depends on SG. */ 10907 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) { 10908 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n"); 10909 features &= ~NETIF_F_GSO; 10910 } 10911 10912 /* GSO partial features require GSO partial be set */ 10913 if ((features & dev->gso_partial_features) && 10914 !(features & NETIF_F_GSO_PARTIAL)) { 10915 netdev_dbg(dev, 10916 "Dropping partially supported GSO features since no GSO partial.\n"); 10917 features &= ~dev->gso_partial_features; 10918 } 10919 10920 if (!(features & NETIF_F_RXCSUM)) { 10921 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet 10922 * successfully merged by hardware must also have the 10923 * checksum verified by hardware. If the user does not 10924 * want to enable RXCSUM, logically, we should disable GRO_HW. 10925 */ 10926 if (features & NETIF_F_GRO_HW) { 10927 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n"); 10928 features &= ~NETIF_F_GRO_HW; 10929 } 10930 } 10931 10932 /* LRO/HW-GRO features cannot be combined with RX-FCS */ 10933 if (features & NETIF_F_RXFCS) { 10934 if (features & NETIF_F_LRO) { 10935 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n"); 10936 features &= ~NETIF_F_LRO; 10937 } 10938 10939 if (features & NETIF_F_GRO_HW) { 10940 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n"); 10941 features &= ~NETIF_F_GRO_HW; 10942 } 10943 } 10944 10945 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) { 10946 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n"); 10947 features &= ~NETIF_F_LRO; 10948 } 10949 10950 if ((features & NETIF_F_HW_TLS_TX) && !netdev_has_ip_or_hw_csum(features)) { 10951 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n"); 10952 features &= ~NETIF_F_HW_TLS_TX; 10953 } 10954 10955 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) { 10956 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n"); 10957 features &= ~NETIF_F_HW_TLS_RX; 10958 } 10959 10960 if ((features & NETIF_F_GSO_UDP_L4) && !netdev_has_ip_or_hw_csum(features)) { 10961 netdev_dbg(dev, "Dropping USO feature since no CSUM feature.\n"); 10962 features &= ~NETIF_F_GSO_UDP_L4; 10963 } 10964 10965 return features; 10966 } 10967 10968 int __netdev_update_features(struct net_device *dev) 10969 { 10970 struct net_device *upper, *lower; 10971 netdev_features_t features; 10972 struct list_head *iter; 10973 int err = -1; 10974 10975 ASSERT_RTNL(); 10976 netdev_ops_assert_locked(dev); 10977 10978 features = netdev_get_wanted_features(dev); 10979 10980 if (dev->netdev_ops->ndo_fix_features) 10981 features = dev->netdev_ops->ndo_fix_features(dev, features); 10982 10983 /* driver might be less strict about feature dependencies */ 10984 features = netdev_fix_features(dev, features); 10985 10986 /* some features can't be enabled if they're off on an upper device */ 10987 netdev_for_each_upper_dev_rcu(dev, upper, iter) 10988 features = netdev_sync_upper_features(dev, upper, features); 10989 10990 if (dev->features == features) 10991 goto sync_lower; 10992 10993 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n", 10994 &dev->features, &features); 10995 10996 if (dev->netdev_ops->ndo_set_features) 10997 err = dev->netdev_ops->ndo_set_features(dev, features); 10998 else 10999 err = 0; 11000 11001 if (unlikely(err < 0)) { 11002 netdev_err(dev, 11003 "set_features() failed (%d); wanted %pNF, left %pNF\n", 11004 err, &features, &dev->features); 11005 /* return non-0 since some features might have changed and 11006 * it's better to fire a spurious notification than miss it 11007 */ 11008 return -1; 11009 } 11010 11011 sync_lower: 11012 /* some features must be disabled on lower devices when disabled 11013 * on an upper device (think: bonding master or bridge) 11014 */ 11015 netdev_for_each_lower_dev(dev, lower, iter) 11016 netdev_sync_lower_features(dev, lower, features); 11017 11018 if (!err) { 11019 netdev_features_t diff = features ^ dev->features; 11020 11021 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) { 11022 /* udp_tunnel_{get,drop}_rx_info both need 11023 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the 11024 * device, or they won't do anything. 11025 * Thus we need to update dev->features 11026 * *before* calling udp_tunnel_get_rx_info, 11027 * but *after* calling udp_tunnel_drop_rx_info. 11028 */ 11029 udp_tunnel_nic_lock(dev); 11030 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) { 11031 dev->features = features; 11032 udp_tunnel_get_rx_info(dev); 11033 } else { 11034 udp_tunnel_drop_rx_info(dev); 11035 } 11036 udp_tunnel_nic_unlock(dev); 11037 } 11038 11039 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) { 11040 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) { 11041 dev->features = features; 11042 err |= vlan_get_rx_ctag_filter_info(dev); 11043 } else { 11044 vlan_drop_rx_ctag_filter_info(dev); 11045 } 11046 } 11047 11048 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) { 11049 if (features & NETIF_F_HW_VLAN_STAG_FILTER) { 11050 dev->features = features; 11051 err |= vlan_get_rx_stag_filter_info(dev); 11052 } else { 11053 vlan_drop_rx_stag_filter_info(dev); 11054 } 11055 } 11056 11057 dev->features = features; 11058 } 11059 11060 return err < 0 ? 0 : 1; 11061 } 11062 11063 /** 11064 * netdev_update_features - recalculate device features 11065 * @dev: the device to check 11066 * 11067 * Recalculate dev->features set and send notifications if it 11068 * has changed. Should be called after driver or hardware dependent 11069 * conditions might have changed that influence the features. 11070 */ 11071 void netdev_update_features(struct net_device *dev) 11072 { 11073 if (__netdev_update_features(dev)) 11074 netdev_features_change(dev); 11075 } 11076 EXPORT_SYMBOL(netdev_update_features); 11077 11078 /** 11079 * netdev_change_features - recalculate device features 11080 * @dev: the device to check 11081 * 11082 * Recalculate dev->features set and send notifications even 11083 * if they have not changed. Should be called instead of 11084 * netdev_update_features() if also dev->vlan_features might 11085 * have changed to allow the changes to be propagated to stacked 11086 * VLAN devices. 11087 */ 11088 void netdev_change_features(struct net_device *dev) 11089 { 11090 __netdev_update_features(dev); 11091 netdev_features_change(dev); 11092 } 11093 EXPORT_SYMBOL(netdev_change_features); 11094 11095 /** 11096 * netif_stacked_transfer_operstate - transfer operstate 11097 * @rootdev: the root or lower level device to transfer state from 11098 * @dev: the device to transfer operstate to 11099 * 11100 * Transfer operational state from root to device. This is normally 11101 * called when a stacking relationship exists between the root 11102 * device and the device(a leaf device). 11103 */ 11104 void netif_stacked_transfer_operstate(const struct net_device *rootdev, 11105 struct net_device *dev) 11106 { 11107 if (rootdev->operstate == IF_OPER_DORMANT) 11108 netif_dormant_on(dev); 11109 else 11110 netif_dormant_off(dev); 11111 11112 if (rootdev->operstate == IF_OPER_TESTING) 11113 netif_testing_on(dev); 11114 else 11115 netif_testing_off(dev); 11116 11117 if (netif_carrier_ok(rootdev)) 11118 netif_carrier_on(dev); 11119 else 11120 netif_carrier_off(dev); 11121 } 11122 EXPORT_SYMBOL(netif_stacked_transfer_operstate); 11123 11124 static int netif_alloc_rx_queues(struct net_device *dev) 11125 { 11126 unsigned int i, count = dev->num_rx_queues; 11127 struct netdev_rx_queue *rx; 11128 size_t sz = count * sizeof(*rx); 11129 int err = 0; 11130 11131 BUG_ON(count < 1); 11132 11133 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11134 if (!rx) 11135 return -ENOMEM; 11136 11137 dev->_rx = rx; 11138 11139 for (i = 0; i < count; i++) { 11140 rx[i].dev = dev; 11141 11142 /* XDP RX-queue setup */ 11143 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0); 11144 if (err < 0) 11145 goto err_rxq_info; 11146 } 11147 return 0; 11148 11149 err_rxq_info: 11150 /* Rollback successful reg's and free other resources */ 11151 while (i--) 11152 xdp_rxq_info_unreg(&rx[i].xdp_rxq); 11153 kvfree(dev->_rx); 11154 dev->_rx = NULL; 11155 return err; 11156 } 11157 11158 static void netif_free_rx_queues(struct net_device *dev) 11159 { 11160 unsigned int i, count = dev->num_rx_queues; 11161 11162 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */ 11163 if (!dev->_rx) 11164 return; 11165 11166 for (i = 0; i < count; i++) 11167 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq); 11168 11169 kvfree(dev->_rx); 11170 } 11171 11172 static void netdev_init_one_queue(struct net_device *dev, 11173 struct netdev_queue *queue, void *_unused) 11174 { 11175 /* Initialize queue lock */ 11176 spin_lock_init(&queue->_xmit_lock); 11177 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type); 11178 queue->xmit_lock_owner = -1; 11179 netdev_queue_numa_node_write(queue, NUMA_NO_NODE); 11180 queue->dev = dev; 11181 #ifdef CONFIG_BQL 11182 dql_init(&queue->dql, HZ); 11183 #endif 11184 } 11185 11186 static void netif_free_tx_queues(struct net_device *dev) 11187 { 11188 kvfree(dev->_tx); 11189 } 11190 11191 static int netif_alloc_netdev_queues(struct net_device *dev) 11192 { 11193 unsigned int count = dev->num_tx_queues; 11194 struct netdev_queue *tx; 11195 size_t sz = count * sizeof(*tx); 11196 11197 if (count < 1 || count > 0xffff) 11198 return -EINVAL; 11199 11200 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11201 if (!tx) 11202 return -ENOMEM; 11203 11204 dev->_tx = tx; 11205 11206 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL); 11207 spin_lock_init(&dev->tx_global_lock); 11208 11209 return 0; 11210 } 11211 11212 void netif_tx_stop_all_queues(struct net_device *dev) 11213 { 11214 unsigned int i; 11215 11216 for (i = 0; i < dev->num_tx_queues; i++) { 11217 struct netdev_queue *txq = netdev_get_tx_queue(dev, i); 11218 11219 netif_tx_stop_queue(txq); 11220 } 11221 } 11222 EXPORT_SYMBOL(netif_tx_stop_all_queues); 11223 11224 static int netdev_do_alloc_pcpu_stats(struct net_device *dev) 11225 { 11226 void __percpu *v; 11227 11228 /* Drivers implementing ndo_get_peer_dev must support tstat 11229 * accounting, so that skb_do_redirect() can bump the dev's 11230 * RX stats upon network namespace switch. 11231 */ 11232 if (dev->netdev_ops->ndo_get_peer_dev && 11233 dev->pcpu_stat_type != NETDEV_PCPU_STAT_TSTATS) 11234 return -EOPNOTSUPP; 11235 11236 switch (dev->pcpu_stat_type) { 11237 case NETDEV_PCPU_STAT_NONE: 11238 return 0; 11239 case NETDEV_PCPU_STAT_LSTATS: 11240 v = dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats); 11241 break; 11242 case NETDEV_PCPU_STAT_TSTATS: 11243 v = dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats); 11244 break; 11245 case NETDEV_PCPU_STAT_DSTATS: 11246 v = dev->dstats = netdev_alloc_pcpu_stats(struct pcpu_dstats); 11247 break; 11248 default: 11249 return -EINVAL; 11250 } 11251 11252 return v ? 0 : -ENOMEM; 11253 } 11254 11255 static void netdev_do_free_pcpu_stats(struct net_device *dev) 11256 { 11257 switch (dev->pcpu_stat_type) { 11258 case NETDEV_PCPU_STAT_NONE: 11259 return; 11260 case NETDEV_PCPU_STAT_LSTATS: 11261 free_percpu(dev->lstats); 11262 break; 11263 case NETDEV_PCPU_STAT_TSTATS: 11264 free_percpu(dev->tstats); 11265 break; 11266 case NETDEV_PCPU_STAT_DSTATS: 11267 free_percpu(dev->dstats); 11268 break; 11269 } 11270 } 11271 11272 static void netdev_free_phy_link_topology(struct net_device *dev) 11273 { 11274 struct phy_link_topology *topo = dev->link_topo; 11275 11276 if (IS_ENABLED(CONFIG_PHYLIB) && topo) { 11277 xa_destroy(&topo->phys); 11278 kfree(topo); 11279 dev->link_topo = NULL; 11280 } 11281 } 11282 11283 /** 11284 * register_netdevice() - register a network device 11285 * @dev: device to register 11286 * 11287 * Take a prepared network device structure and make it externally accessible. 11288 * A %NETDEV_REGISTER message is sent to the netdev notifier chain. 11289 * Callers must hold the rtnl lock - you may want register_netdev() 11290 * instead of this. 11291 */ 11292 int register_netdevice(struct net_device *dev) 11293 { 11294 int ret; 11295 struct net *net = dev_net(dev); 11296 11297 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE < 11298 NETDEV_FEATURE_COUNT); 11299 BUG_ON(dev_boot_phase); 11300 ASSERT_RTNL(); 11301 11302 might_sleep(); 11303 11304 /* When net_device's are persistent, this will be fatal. */ 11305 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED); 11306 BUG_ON(!net); 11307 11308 ret = ethtool_check_ops(dev->ethtool_ops); 11309 if (ret) 11310 return ret; 11311 11312 /* rss ctx ID 0 is reserved for the default context, start from 1 */ 11313 xa_init_flags(&dev->ethtool->rss_ctx, XA_FLAGS_ALLOC1); 11314 mutex_init(&dev->ethtool->rss_lock); 11315 11316 spin_lock_init(&dev->addr_list_lock); 11317 netdev_set_addr_lockdep_class(dev); 11318 11319 ret = dev_get_valid_name(net, dev, dev->name); 11320 if (ret < 0) 11321 goto out; 11322 11323 ret = -ENOMEM; 11324 dev->name_node = netdev_name_node_head_alloc(dev); 11325 if (!dev->name_node) 11326 goto out; 11327 11328 /* Init, if this function is available */ 11329 if (dev->netdev_ops->ndo_init) { 11330 ret = dev->netdev_ops->ndo_init(dev); 11331 if (ret) { 11332 if (ret > 0) 11333 ret = -EIO; 11334 goto err_free_name; 11335 } 11336 } 11337 11338 if (((dev->hw_features | dev->features) & 11339 NETIF_F_HW_VLAN_CTAG_FILTER) && 11340 (!dev->netdev_ops->ndo_vlan_rx_add_vid || 11341 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) { 11342 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n"); 11343 ret = -EINVAL; 11344 goto err_uninit; 11345 } 11346 11347 ret = netdev_do_alloc_pcpu_stats(dev); 11348 if (ret) 11349 goto err_uninit; 11350 11351 ret = dev_index_reserve(net, dev->ifindex); 11352 if (ret < 0) 11353 goto err_free_pcpu; 11354 dev->ifindex = ret; 11355 11356 /* Transfer changeable features to wanted_features and enable 11357 * software offloads (GSO and GRO). 11358 */ 11359 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF); 11360 dev->features |= NETIF_F_SOFT_FEATURES; 11361 11362 if (dev->udp_tunnel_nic_info) { 11363 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT; 11364 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT; 11365 } 11366 11367 dev->wanted_features = dev->features & dev->hw_features; 11368 11369 if (!(dev->flags & IFF_LOOPBACK)) 11370 dev->hw_features |= NETIF_F_NOCACHE_COPY; 11371 11372 /* If IPv4 TCP segmentation offload is supported we should also 11373 * allow the device to enable segmenting the frame with the option 11374 * of ignoring a static IP ID value. This doesn't enable the 11375 * feature itself but allows the user to enable it later. 11376 */ 11377 if (dev->hw_features & NETIF_F_TSO) 11378 dev->hw_features |= NETIF_F_TSO_MANGLEID; 11379 if (dev->vlan_features & NETIF_F_TSO) 11380 dev->vlan_features |= NETIF_F_TSO_MANGLEID; 11381 if (dev->mpls_features & NETIF_F_TSO) 11382 dev->mpls_features |= NETIF_F_TSO_MANGLEID; 11383 if (dev->hw_enc_features & NETIF_F_TSO) 11384 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID; 11385 11386 /* TSO_MANGLEID belongs in mangleid_features by definition */ 11387 dev->mangleid_features |= NETIF_F_TSO_MANGLEID; 11388 11389 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices. 11390 */ 11391 dev->vlan_features |= NETIF_F_HIGHDMA; 11392 11393 /* Make NETIF_F_SG inheritable to tunnel devices. 11394 */ 11395 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL; 11396 11397 /* Make NETIF_F_SG inheritable to MPLS. 11398 */ 11399 dev->mpls_features |= NETIF_F_SG; 11400 11401 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev); 11402 ret = notifier_to_errno(ret); 11403 if (ret) 11404 goto err_ifindex_release; 11405 11406 ret = netdev_register_kobject(dev); 11407 11408 netdev_lock(dev); 11409 WRITE_ONCE(dev->reg_state, ret ? NETREG_UNREGISTERED : NETREG_REGISTERED); 11410 netdev_unlock(dev); 11411 11412 if (ret) 11413 goto err_uninit_notify; 11414 11415 netdev_lock_ops(dev); 11416 __netdev_update_features(dev); 11417 netdev_unlock_ops(dev); 11418 11419 /* 11420 * Default initial state at registry is that the 11421 * device is present. 11422 */ 11423 11424 set_bit(__LINK_STATE_PRESENT, &dev->state); 11425 11426 linkwatch_init_dev(dev); 11427 11428 dev_init_scheduler(dev); 11429 11430 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL); 11431 list_netdevice(dev); 11432 11433 add_device_randomness(dev->dev_addr, dev->addr_len); 11434 11435 /* If the device has permanent device address, driver should 11436 * set dev_addr and also addr_assign_type should be set to 11437 * NET_ADDR_PERM (default value). 11438 */ 11439 if (dev->addr_assign_type == NET_ADDR_PERM) 11440 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len); 11441 11442 /* Notify protocols, that a new device appeared. */ 11443 netdev_lock_ops(dev); 11444 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev); 11445 netdev_unlock_ops(dev); 11446 ret = notifier_to_errno(ret); 11447 if (ret) { 11448 /* Expect explicit free_netdev() on failure */ 11449 dev->needs_free_netdev = false; 11450 unregister_netdevice_queue(dev, NULL); 11451 goto out; 11452 } 11453 /* 11454 * Prevent userspace races by waiting until the network 11455 * device is fully setup before sending notifications. 11456 */ 11457 if (!(dev->rtnl_link_ops && dev->rtnl_link_initializing)) 11458 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 11459 11460 out: 11461 return ret; 11462 11463 err_uninit_notify: 11464 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11465 err_ifindex_release: 11466 dev_index_release(net, dev->ifindex); 11467 err_free_pcpu: 11468 netdev_do_free_pcpu_stats(dev); 11469 err_uninit: 11470 if (dev->netdev_ops->ndo_uninit) 11471 dev->netdev_ops->ndo_uninit(dev); 11472 if (dev->priv_destructor) 11473 dev->priv_destructor(dev); 11474 err_free_name: 11475 netdev_name_node_free(dev->name_node); 11476 goto out; 11477 } 11478 EXPORT_SYMBOL(register_netdevice); 11479 11480 /* Initialize the core of a dummy net device. 11481 * The setup steps dummy netdevs need which normal netdevs get by going 11482 * through register_netdevice(). 11483 */ 11484 static void init_dummy_netdev(struct net_device *dev) 11485 { 11486 /* make sure we BUG if trying to hit standard 11487 * register/unregister code path 11488 */ 11489 dev->reg_state = NETREG_DUMMY; 11490 11491 /* a dummy interface is started by default */ 11492 set_bit(__LINK_STATE_PRESENT, &dev->state); 11493 set_bit(__LINK_STATE_START, &dev->state); 11494 11495 /* Note : We dont allocate pcpu_refcnt for dummy devices, 11496 * because users of this 'device' dont need to change 11497 * its refcount. 11498 */ 11499 } 11500 11501 /** 11502 * register_netdev - register a network device 11503 * @dev: device to register 11504 * 11505 * Take a completed network device structure and add it to the kernel 11506 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier 11507 * chain. 0 is returned on success. A negative errno code is returned 11508 * on a failure to set up the device, or if the name is a duplicate. 11509 * 11510 * This is a wrapper around register_netdevice that takes the rtnl semaphore 11511 * and expands the device name if you passed a format string to 11512 * alloc_netdev. 11513 */ 11514 int register_netdev(struct net_device *dev) 11515 { 11516 struct net *net = dev_net(dev); 11517 int err; 11518 11519 if (rtnl_net_lock_killable(net)) 11520 return -EINTR; 11521 11522 err = register_netdevice(dev); 11523 11524 rtnl_net_unlock(net); 11525 11526 return err; 11527 } 11528 EXPORT_SYMBOL(register_netdev); 11529 11530 int netdev_refcnt_read(const struct net_device *dev) 11531 { 11532 #ifdef CONFIG_PCPU_DEV_REFCNT 11533 int i, refcnt = 0; 11534 11535 for_each_possible_cpu(i) 11536 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i); 11537 return refcnt; 11538 #else 11539 return refcount_read(&dev->dev_refcnt); 11540 #endif 11541 } 11542 EXPORT_SYMBOL(netdev_refcnt_read); 11543 11544 int netdev_unregister_timeout_secs __read_mostly = 10; 11545 11546 #define WAIT_REFS_MIN_MSECS 1 11547 #define WAIT_REFS_MAX_MSECS 250 11548 /** 11549 * netdev_wait_allrefs_any - wait until all references are gone. 11550 * @list: list of net_devices to wait on 11551 * 11552 * This is called when unregistering network devices. 11553 * 11554 * Any protocol or device that holds a reference should register 11555 * for netdevice notification, and cleanup and put back the 11556 * reference if they receive an UNREGISTER event. 11557 * We can get stuck here if buggy protocols don't correctly 11558 * call dev_put. 11559 */ 11560 static struct net_device *netdev_wait_allrefs_any(struct list_head *list) 11561 { 11562 unsigned long rebroadcast_time, warning_time; 11563 struct net_device *dev; 11564 int wait = 0; 11565 11566 rebroadcast_time = warning_time = jiffies; 11567 11568 list_for_each_entry(dev, list, todo_list) 11569 if (netdev_refcnt_read(dev) == 1) 11570 return dev; 11571 11572 while (true) { 11573 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) { 11574 rtnl_lock(); 11575 11576 /* Rebroadcast unregister notification */ 11577 list_for_each_entry(dev, list, todo_list) 11578 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11579 11580 __rtnl_unlock(); 11581 rcu_barrier(); 11582 rtnl_lock(); 11583 11584 list_for_each_entry(dev, list, todo_list) 11585 if (test_bit(__LINK_STATE_LINKWATCH_PENDING, 11586 &dev->state)) { 11587 /* We must not have linkwatch events 11588 * pending on unregister. If this 11589 * happens, we simply run the queue 11590 * unscheduled, resulting in a noop 11591 * for this device. 11592 */ 11593 linkwatch_run_queue(); 11594 break; 11595 } 11596 11597 __rtnl_unlock(); 11598 11599 rebroadcast_time = jiffies; 11600 } 11601 11602 rcu_barrier(); 11603 11604 if (!wait) { 11605 wait = WAIT_REFS_MIN_MSECS; 11606 } else { 11607 msleep(wait); 11608 wait = min(wait << 1, WAIT_REFS_MAX_MSECS); 11609 } 11610 11611 list_for_each_entry(dev, list, todo_list) 11612 if (netdev_refcnt_read(dev) == 1) 11613 return dev; 11614 11615 if (time_after(jiffies, warning_time + 11616 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) { 11617 list_for_each_entry(dev, list, todo_list) { 11618 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n", 11619 dev->name, netdev_refcnt_read(dev)); 11620 ref_tracker_dir_print(&dev->refcnt_tracker, 10); 11621 } 11622 11623 warning_time = jiffies; 11624 } 11625 } 11626 } 11627 11628 /* The sequence is: 11629 * 11630 * rtnl_lock(); 11631 * ... 11632 * register_netdevice(x1); 11633 * register_netdevice(x2); 11634 * ... 11635 * unregister_netdevice(y1); 11636 * unregister_netdevice(y2); 11637 * ... 11638 * rtnl_unlock(); 11639 * free_netdev(y1); 11640 * free_netdev(y2); 11641 * 11642 * We are invoked by rtnl_unlock(). 11643 * This allows us to deal with problems: 11644 * 1) We can delete sysfs objects which invoke hotplug 11645 * without deadlocking with linkwatch via keventd. 11646 * 2) Since we run with the RTNL semaphore not held, we can sleep 11647 * safely in order to wait for the netdev refcnt to drop to zero. 11648 * 11649 * We must not return until all unregister events added during 11650 * the interval the lock was held have been completed. 11651 */ 11652 void netdev_run_todo(void) 11653 { 11654 struct net_device *dev, *tmp; 11655 struct list_head list; 11656 int cnt; 11657 #ifdef CONFIG_LOCKDEP 11658 struct list_head unlink_list; 11659 11660 list_replace_init(&net_unlink_list, &unlink_list); 11661 11662 while (!list_empty(&unlink_list)) { 11663 dev = list_first_entry(&unlink_list, struct net_device, 11664 unlink_list); 11665 list_del_init(&dev->unlink_list); 11666 dev->nested_level = dev->lower_level - 1; 11667 } 11668 #endif 11669 11670 /* Snapshot list, allow later requests */ 11671 list_replace_init(&net_todo_list, &list); 11672 11673 __rtnl_unlock(); 11674 11675 /* Wait for rcu callbacks to finish before next phase */ 11676 if (!list_empty(&list)) 11677 rcu_barrier(); 11678 11679 list_for_each_entry_safe(dev, tmp, &list, todo_list) { 11680 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) { 11681 netdev_WARN(dev, "run_todo but not unregistering\n"); 11682 list_del(&dev->todo_list); 11683 continue; 11684 } 11685 11686 netdev_lock(dev); 11687 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERED); 11688 netdev_unlock(dev); 11689 linkwatch_sync_dev(dev); 11690 } 11691 11692 cnt = 0; 11693 while (!list_empty(&list)) { 11694 dev = netdev_wait_allrefs_any(&list); 11695 list_del(&dev->todo_list); 11696 11697 /* paranoia */ 11698 BUG_ON(netdev_refcnt_read(dev) != 1); 11699 BUG_ON(!list_empty(&dev->ptype_all)); 11700 BUG_ON(!list_empty(&dev->ptype_specific)); 11701 WARN_ON(rcu_access_pointer(dev->ip_ptr)); 11702 WARN_ON(rcu_access_pointer(dev->ip6_ptr)); 11703 11704 netdev_do_free_pcpu_stats(dev); 11705 if (dev->priv_destructor) 11706 dev->priv_destructor(dev); 11707 if (dev->needs_free_netdev) 11708 free_netdev(dev); 11709 11710 cnt++; 11711 11712 /* Free network device */ 11713 kobject_put(&dev->dev.kobj); 11714 } 11715 if (cnt && atomic_sub_and_test(cnt, &dev_unreg_count)) 11716 wake_up(&netdev_unregistering_wq); 11717 } 11718 11719 /* Collate per-cpu network dstats statistics 11720 * 11721 * Read per-cpu network statistics from dev->dstats and populate the related 11722 * fields in @s. 11723 */ 11724 static void dev_fetch_dstats(struct rtnl_link_stats64 *s, 11725 const struct pcpu_dstats __percpu *dstats) 11726 { 11727 int cpu; 11728 11729 for_each_possible_cpu(cpu) { 11730 u64 rx_packets, rx_bytes, rx_drops; 11731 u64 tx_packets, tx_bytes, tx_drops; 11732 const struct pcpu_dstats *stats; 11733 unsigned int start; 11734 11735 stats = per_cpu_ptr(dstats, cpu); 11736 do { 11737 start = u64_stats_fetch_begin(&stats->syncp); 11738 rx_packets = u64_stats_read(&stats->rx_packets); 11739 rx_bytes = u64_stats_read(&stats->rx_bytes); 11740 rx_drops = u64_stats_read(&stats->rx_drops); 11741 tx_packets = u64_stats_read(&stats->tx_packets); 11742 tx_bytes = u64_stats_read(&stats->tx_bytes); 11743 tx_drops = u64_stats_read(&stats->tx_drops); 11744 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11745 11746 s->rx_packets += rx_packets; 11747 s->rx_bytes += rx_bytes; 11748 s->rx_dropped += rx_drops; 11749 s->tx_packets += tx_packets; 11750 s->tx_bytes += tx_bytes; 11751 s->tx_dropped += tx_drops; 11752 } 11753 } 11754 11755 /* ndo_get_stats64 implementation for dtstats-based accounting. 11756 * 11757 * Populate @s from dev->stats and dev->dstats. This is used internally by the 11758 * core for NETDEV_PCPU_STAT_DSTAT-type stats collection. 11759 */ 11760 static void dev_get_dstats64(const struct net_device *dev, 11761 struct rtnl_link_stats64 *s) 11762 { 11763 netdev_stats_to_stats64(s, &dev->stats); 11764 dev_fetch_dstats(s, dev->dstats); 11765 } 11766 11767 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has 11768 * all the same fields in the same order as net_device_stats, with only 11769 * the type differing, but rtnl_link_stats64 may have additional fields 11770 * at the end for newer counters. 11771 */ 11772 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64, 11773 const struct net_device_stats *netdev_stats) 11774 { 11775 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t); 11776 const atomic_long_t *src = (atomic_long_t *)netdev_stats; 11777 u64 *dst = (u64 *)stats64; 11778 11779 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64)); 11780 for (i = 0; i < n; i++) 11781 dst[i] = (unsigned long)atomic_long_read(&src[i]); 11782 /* zero out counters that only exist in rtnl_link_stats64 */ 11783 memset((char *)stats64 + n * sizeof(u64), 0, 11784 sizeof(*stats64) - n * sizeof(u64)); 11785 } 11786 EXPORT_SYMBOL(netdev_stats_to_stats64); 11787 11788 static __cold struct net_device_core_stats __percpu *netdev_core_stats_alloc( 11789 struct net_device *dev) 11790 { 11791 struct net_device_core_stats __percpu *p; 11792 11793 p = alloc_percpu_gfp(struct net_device_core_stats, 11794 GFP_ATOMIC | __GFP_NOWARN); 11795 11796 if (p && cmpxchg(&dev->core_stats, NULL, p)) 11797 free_percpu(p); 11798 11799 /* This READ_ONCE() pairs with the cmpxchg() above */ 11800 return READ_ONCE(dev->core_stats); 11801 } 11802 11803 noinline void netdev_core_stats_inc(struct net_device *dev, u32 offset) 11804 { 11805 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11806 struct net_device_core_stats __percpu *p = READ_ONCE(dev->core_stats); 11807 unsigned long __percpu *field; 11808 11809 if (unlikely(!p)) { 11810 p = netdev_core_stats_alloc(dev); 11811 if (!p) 11812 return; 11813 } 11814 11815 field = (unsigned long __percpu *)((void __percpu *)p + offset); 11816 this_cpu_inc(*field); 11817 } 11818 EXPORT_SYMBOL_GPL(netdev_core_stats_inc); 11819 11820 /** 11821 * dev_get_stats - get network device statistics 11822 * @dev: device to get statistics from 11823 * @storage: place to store stats 11824 * 11825 * Get network statistics from device. Return @storage. 11826 * The device driver may provide its own method by setting 11827 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats; 11828 * otherwise the internal statistics structure is used. 11829 */ 11830 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev, 11831 struct rtnl_link_stats64 *storage) 11832 { 11833 const struct net_device_ops *ops = dev->netdev_ops; 11834 const struct net_device_core_stats __percpu *p; 11835 11836 /* 11837 * IPv{4,6} and udp tunnels share common stat helpers and use 11838 * different stat type (NETDEV_PCPU_STAT_TSTATS vs 11839 * NETDEV_PCPU_STAT_DSTATS). Ensure the accounting is consistent. 11840 */ 11841 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_bytes) != 11842 offsetof(struct pcpu_dstats, rx_bytes)); 11843 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_packets) != 11844 offsetof(struct pcpu_dstats, rx_packets)); 11845 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_bytes) != 11846 offsetof(struct pcpu_dstats, tx_bytes)); 11847 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_packets) != 11848 offsetof(struct pcpu_dstats, tx_packets)); 11849 11850 if (ops->ndo_get_stats64) { 11851 memset(storage, 0, sizeof(*storage)); 11852 ops->ndo_get_stats64(dev, storage); 11853 } else if (ops->ndo_get_stats) { 11854 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev)); 11855 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_TSTATS) { 11856 dev_get_tstats64(dev, storage); 11857 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_DSTATS) { 11858 dev_get_dstats64(dev, storage); 11859 } else { 11860 netdev_stats_to_stats64(storage, &dev->stats); 11861 } 11862 11863 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11864 p = READ_ONCE(dev->core_stats); 11865 if (p) { 11866 const struct net_device_core_stats *core_stats; 11867 int i; 11868 11869 for_each_possible_cpu(i) { 11870 core_stats = per_cpu_ptr(p, i); 11871 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped); 11872 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped); 11873 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler); 11874 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped); 11875 } 11876 } 11877 return storage; 11878 } 11879 EXPORT_SYMBOL(dev_get_stats); 11880 11881 /** 11882 * dev_fetch_sw_netstats - get per-cpu network device statistics 11883 * @s: place to store stats 11884 * @netstats: per-cpu network stats to read from 11885 * 11886 * Read per-cpu network statistics and populate the related fields in @s. 11887 */ 11888 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s, 11889 const struct pcpu_sw_netstats __percpu *netstats) 11890 { 11891 int cpu; 11892 11893 for_each_possible_cpu(cpu) { 11894 u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 11895 const struct pcpu_sw_netstats *stats; 11896 unsigned int start; 11897 11898 stats = per_cpu_ptr(netstats, cpu); 11899 do { 11900 start = u64_stats_fetch_begin(&stats->syncp); 11901 rx_packets = u64_stats_read(&stats->rx_packets); 11902 rx_bytes = u64_stats_read(&stats->rx_bytes); 11903 tx_packets = u64_stats_read(&stats->tx_packets); 11904 tx_bytes = u64_stats_read(&stats->tx_bytes); 11905 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11906 11907 s->rx_packets += rx_packets; 11908 s->rx_bytes += rx_bytes; 11909 s->tx_packets += tx_packets; 11910 s->tx_bytes += tx_bytes; 11911 } 11912 } 11913 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats); 11914 11915 /** 11916 * dev_get_tstats64 - ndo_get_stats64 implementation 11917 * @dev: device to get statistics from 11918 * @s: place to store stats 11919 * 11920 * Populate @s from dev->stats and dev->tstats. Can be used as 11921 * ndo_get_stats64() callback. 11922 */ 11923 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s) 11924 { 11925 netdev_stats_to_stats64(s, &dev->stats); 11926 dev_fetch_sw_netstats(s, dev->tstats); 11927 } 11928 EXPORT_SYMBOL_GPL(dev_get_tstats64); 11929 11930 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev) 11931 { 11932 struct netdev_queue *queue = dev_ingress_queue(dev); 11933 11934 #ifdef CONFIG_NET_CLS_ACT 11935 if (queue) 11936 return queue; 11937 queue = kzalloc(sizeof(*queue), GFP_KERNEL); 11938 if (!queue) 11939 return NULL; 11940 netdev_init_one_queue(dev, queue, NULL); 11941 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc); 11942 RCU_INIT_POINTER(queue->qdisc_sleeping, &noop_qdisc); 11943 rcu_assign_pointer(dev->ingress_queue, queue); 11944 #endif 11945 return queue; 11946 } 11947 11948 static const struct ethtool_ops default_ethtool_ops; 11949 11950 void netdev_set_default_ethtool_ops(struct net_device *dev, 11951 const struct ethtool_ops *ops) 11952 { 11953 if (dev->ethtool_ops == &default_ethtool_ops) 11954 dev->ethtool_ops = ops; 11955 } 11956 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops); 11957 11958 /** 11959 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default 11960 * @dev: netdev to enable the IRQ coalescing on 11961 * 11962 * Sets a conservative default for SW IRQ coalescing. Users can use 11963 * sysfs attributes to override the default values. 11964 */ 11965 void netdev_sw_irq_coalesce_default_on(struct net_device *dev) 11966 { 11967 WARN_ON(dev->reg_state == NETREG_REGISTERED); 11968 11969 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) { 11970 netdev_set_gro_flush_timeout(dev, 20000); 11971 netdev_set_defer_hard_irqs(dev, 1); 11972 } 11973 } 11974 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on); 11975 11976 /** 11977 * alloc_netdev_mqs - allocate network device 11978 * @sizeof_priv: size of private data to allocate space for 11979 * @name: device name format string 11980 * @name_assign_type: origin of device name 11981 * @setup: callback to initialize device 11982 * @txqs: the number of TX subqueues to allocate 11983 * @rxqs: the number of RX subqueues to allocate 11984 * 11985 * Allocates a struct net_device with private data area for driver use 11986 * and performs basic initialization. Also allocates subqueue structs 11987 * for each queue on the device. 11988 */ 11989 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name, 11990 unsigned char name_assign_type, 11991 void (*setup)(struct net_device *), 11992 unsigned int txqs, unsigned int rxqs) 11993 { 11994 struct net_device *dev; 11995 size_t napi_config_sz; 11996 unsigned int maxqs; 11997 11998 BUG_ON(strlen(name) >= sizeof(dev->name)); 11999 12000 if (txqs < 1) { 12001 pr_err("alloc_netdev: Unable to allocate device with zero queues\n"); 12002 return NULL; 12003 } 12004 12005 if (rxqs < 1) { 12006 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n"); 12007 return NULL; 12008 } 12009 12010 maxqs = max(txqs, rxqs); 12011 12012 dev = kvzalloc(struct_size(dev, priv, sizeof_priv), 12013 GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 12014 if (!dev) 12015 return NULL; 12016 12017 dev->priv_len = sizeof_priv; 12018 12019 ref_tracker_dir_init(&dev->refcnt_tracker, 128, "netdev"); 12020 #ifdef CONFIG_PCPU_DEV_REFCNT 12021 dev->pcpu_refcnt = alloc_percpu(int); 12022 if (!dev->pcpu_refcnt) 12023 goto free_dev; 12024 __dev_hold(dev); 12025 #else 12026 refcount_set(&dev->dev_refcnt, 1); 12027 #endif 12028 12029 if (dev_addr_init(dev)) 12030 goto free_pcpu; 12031 12032 dev_mc_init(dev); 12033 dev_uc_init(dev); 12034 12035 dev_net_set(dev, &init_net); 12036 12037 dev->gso_max_size = GSO_LEGACY_MAX_SIZE; 12038 dev->xdp_zc_max_segs = 1; 12039 dev->gso_max_segs = GSO_MAX_SEGS; 12040 dev->gro_max_size = GRO_LEGACY_MAX_SIZE; 12041 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE; 12042 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE; 12043 dev->tso_max_size = TSO_LEGACY_MAX_SIZE; 12044 dev->tso_max_segs = TSO_MAX_SEGS; 12045 dev->upper_level = 1; 12046 dev->lower_level = 1; 12047 #ifdef CONFIG_LOCKDEP 12048 dev->nested_level = 0; 12049 INIT_LIST_HEAD(&dev->unlink_list); 12050 #endif 12051 12052 INIT_LIST_HEAD(&dev->napi_list); 12053 INIT_LIST_HEAD(&dev->unreg_list); 12054 INIT_LIST_HEAD(&dev->close_list); 12055 INIT_LIST_HEAD(&dev->link_watch_list); 12056 INIT_LIST_HEAD(&dev->adj_list.upper); 12057 INIT_LIST_HEAD(&dev->adj_list.lower); 12058 INIT_LIST_HEAD(&dev->ptype_all); 12059 INIT_LIST_HEAD(&dev->ptype_specific); 12060 INIT_LIST_HEAD(&dev->net_notifier_list); 12061 #ifdef CONFIG_NET_SCHED 12062 hash_init(dev->qdisc_hash); 12063 #endif 12064 12065 mutex_init(&dev->lock); 12066 12067 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 12068 setup(dev); 12069 12070 if (!dev->tx_queue_len) { 12071 dev->priv_flags |= IFF_NO_QUEUE; 12072 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN; 12073 } 12074 12075 dev->num_tx_queues = txqs; 12076 dev->real_num_tx_queues = txqs; 12077 if (netif_alloc_netdev_queues(dev)) 12078 goto free_all; 12079 12080 dev->num_rx_queues = rxqs; 12081 dev->real_num_rx_queues = rxqs; 12082 if (netif_alloc_rx_queues(dev)) 12083 goto free_all; 12084 dev->ethtool = kzalloc(sizeof(*dev->ethtool), GFP_KERNEL_ACCOUNT); 12085 if (!dev->ethtool) 12086 goto free_all; 12087 12088 dev->cfg = kzalloc(sizeof(*dev->cfg), GFP_KERNEL_ACCOUNT); 12089 if (!dev->cfg) 12090 goto free_all; 12091 dev->cfg_pending = dev->cfg; 12092 12093 dev->num_napi_configs = maxqs; 12094 napi_config_sz = array_size(maxqs, sizeof(*dev->napi_config)); 12095 dev->napi_config = kvzalloc(napi_config_sz, GFP_KERNEL_ACCOUNT); 12096 if (!dev->napi_config) 12097 goto free_all; 12098 12099 strscpy(dev->name, name); 12100 dev->name_assign_type = name_assign_type; 12101 dev->group = INIT_NETDEV_GROUP; 12102 if (!dev->ethtool_ops) 12103 dev->ethtool_ops = &default_ethtool_ops; 12104 12105 nf_hook_netdev_init(dev); 12106 12107 return dev; 12108 12109 free_all: 12110 free_netdev(dev); 12111 return NULL; 12112 12113 free_pcpu: 12114 #ifdef CONFIG_PCPU_DEV_REFCNT 12115 free_percpu(dev->pcpu_refcnt); 12116 free_dev: 12117 #endif 12118 kvfree(dev); 12119 return NULL; 12120 } 12121 EXPORT_SYMBOL(alloc_netdev_mqs); 12122 12123 static void netdev_napi_exit(struct net_device *dev) 12124 { 12125 if (!list_empty(&dev->napi_list)) { 12126 struct napi_struct *p, *n; 12127 12128 netdev_lock(dev); 12129 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list) 12130 __netif_napi_del_locked(p); 12131 netdev_unlock(dev); 12132 12133 synchronize_net(); 12134 } 12135 12136 kvfree(dev->napi_config); 12137 } 12138 12139 /** 12140 * free_netdev - free network device 12141 * @dev: device 12142 * 12143 * This function does the last stage of destroying an allocated device 12144 * interface. The reference to the device object is released. If this 12145 * is the last reference then it will be freed.Must be called in process 12146 * context. 12147 */ 12148 void free_netdev(struct net_device *dev) 12149 { 12150 might_sleep(); 12151 12152 /* When called immediately after register_netdevice() failed the unwind 12153 * handling may still be dismantling the device. Handle that case by 12154 * deferring the free. 12155 */ 12156 if (dev->reg_state == NETREG_UNREGISTERING) { 12157 ASSERT_RTNL(); 12158 dev->needs_free_netdev = true; 12159 return; 12160 } 12161 12162 WARN_ON(dev->cfg != dev->cfg_pending); 12163 kfree(dev->cfg); 12164 kfree(dev->ethtool); 12165 netif_free_tx_queues(dev); 12166 netif_free_rx_queues(dev); 12167 12168 kfree(rcu_dereference_protected(dev->ingress_queue, 1)); 12169 12170 /* Flush device addresses */ 12171 dev_addr_flush(dev); 12172 12173 netdev_napi_exit(dev); 12174 12175 netif_del_cpu_rmap(dev); 12176 12177 ref_tracker_dir_exit(&dev->refcnt_tracker); 12178 #ifdef CONFIG_PCPU_DEV_REFCNT 12179 free_percpu(dev->pcpu_refcnt); 12180 dev->pcpu_refcnt = NULL; 12181 #endif 12182 free_percpu(dev->core_stats); 12183 dev->core_stats = NULL; 12184 free_percpu(dev->xdp_bulkq); 12185 dev->xdp_bulkq = NULL; 12186 12187 netdev_free_phy_link_topology(dev); 12188 12189 mutex_destroy(&dev->lock); 12190 12191 /* Compatibility with error handling in drivers */ 12192 if (dev->reg_state == NETREG_UNINITIALIZED || 12193 dev->reg_state == NETREG_DUMMY) { 12194 kvfree(dev); 12195 return; 12196 } 12197 12198 BUG_ON(dev->reg_state != NETREG_UNREGISTERED); 12199 WRITE_ONCE(dev->reg_state, NETREG_RELEASED); 12200 12201 /* will free via device release */ 12202 put_device(&dev->dev); 12203 } 12204 EXPORT_SYMBOL(free_netdev); 12205 12206 /** 12207 * alloc_netdev_dummy - Allocate and initialize a dummy net device. 12208 * @sizeof_priv: size of private data to allocate space for 12209 * 12210 * Return: the allocated net_device on success, NULL otherwise 12211 */ 12212 struct net_device *alloc_netdev_dummy(int sizeof_priv) 12213 { 12214 return alloc_netdev(sizeof_priv, "dummy#", NET_NAME_UNKNOWN, 12215 init_dummy_netdev); 12216 } 12217 EXPORT_SYMBOL_GPL(alloc_netdev_dummy); 12218 12219 /** 12220 * synchronize_net - Synchronize with packet receive processing 12221 * 12222 * Wait for packets currently being received to be done. 12223 * Does not block later packets from starting. 12224 */ 12225 void synchronize_net(void) 12226 { 12227 might_sleep(); 12228 if (from_cleanup_net() || rtnl_is_locked()) 12229 synchronize_rcu_expedited(); 12230 else 12231 synchronize_rcu(); 12232 } 12233 EXPORT_SYMBOL(synchronize_net); 12234 12235 static void netdev_rss_contexts_free(struct net_device *dev) 12236 { 12237 struct ethtool_rxfh_context *ctx; 12238 unsigned long context; 12239 12240 mutex_lock(&dev->ethtool->rss_lock); 12241 xa_for_each(&dev->ethtool->rss_ctx, context, ctx) { 12242 xa_erase(&dev->ethtool->rss_ctx, context); 12243 dev->ethtool_ops->remove_rxfh_context(dev, ctx, context, NULL); 12244 kfree(ctx); 12245 } 12246 xa_destroy(&dev->ethtool->rss_ctx); 12247 mutex_unlock(&dev->ethtool->rss_lock); 12248 } 12249 12250 /** 12251 * unregister_netdevice_queue - remove device from the kernel 12252 * @dev: device 12253 * @head: list 12254 * 12255 * This function shuts down a device interface and removes it 12256 * from the kernel tables. 12257 * If head not NULL, device is queued to be unregistered later. 12258 * 12259 * Callers must hold the rtnl semaphore. You may want 12260 * unregister_netdev() instead of this. 12261 */ 12262 12263 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head) 12264 { 12265 ASSERT_RTNL(); 12266 12267 if (head) { 12268 list_move_tail(&dev->unreg_list, head); 12269 } else { 12270 LIST_HEAD(single); 12271 12272 list_add(&dev->unreg_list, &single); 12273 unregister_netdevice_many(&single); 12274 } 12275 } 12276 EXPORT_SYMBOL(unregister_netdevice_queue); 12277 12278 static void dev_memory_provider_uninstall(struct net_device *dev) 12279 { 12280 unsigned int i; 12281 12282 for (i = 0; i < dev->real_num_rx_queues; i++) { 12283 struct netdev_rx_queue *rxq = &dev->_rx[i]; 12284 struct pp_memory_provider_params *p = &rxq->mp_params; 12285 12286 if (p->mp_ops && p->mp_ops->uninstall) 12287 p->mp_ops->uninstall(rxq->mp_params.mp_priv, rxq); 12288 } 12289 } 12290 12291 /* devices must be UP and netdev_lock()'d */ 12292 static void netif_close_many_and_unlock(struct list_head *close_head) 12293 { 12294 struct net_device *dev, *tmp; 12295 12296 netif_close_many(close_head, false); 12297 12298 /* ... now unlock them */ 12299 list_for_each_entry_safe(dev, tmp, close_head, close_list) { 12300 netdev_unlock(dev); 12301 list_del_init(&dev->close_list); 12302 } 12303 } 12304 12305 static void netif_close_many_and_unlock_cond(struct list_head *close_head) 12306 { 12307 #ifdef CONFIG_LOCKDEP 12308 /* We can only track up to MAX_LOCK_DEPTH locks per task. 12309 * 12310 * Reserve half the available slots for additional locks possibly 12311 * taken by notifiers and (soft)irqs. 12312 */ 12313 unsigned int limit = MAX_LOCK_DEPTH / 2; 12314 12315 if (lockdep_depth(current) > limit) 12316 netif_close_many_and_unlock(close_head); 12317 #endif 12318 } 12319 12320 void unregister_netdevice_many_notify(struct list_head *head, 12321 u32 portid, const struct nlmsghdr *nlh) 12322 { 12323 struct net_device *dev, *tmp; 12324 LIST_HEAD(close_head); 12325 int cnt = 0; 12326 12327 BUG_ON(dev_boot_phase); 12328 ASSERT_RTNL(); 12329 12330 if (list_empty(head)) 12331 return; 12332 12333 list_for_each_entry_safe(dev, tmp, head, unreg_list) { 12334 /* Some devices call without registering 12335 * for initialization unwind. Remove those 12336 * devices and proceed with the remaining. 12337 */ 12338 if (dev->reg_state == NETREG_UNINITIALIZED) { 12339 pr_debug("unregister_netdevice: device %s/%p never was registered\n", 12340 dev->name, dev); 12341 12342 WARN_ON(1); 12343 list_del(&dev->unreg_list); 12344 continue; 12345 } 12346 dev->dismantle = true; 12347 BUG_ON(dev->reg_state != NETREG_REGISTERED); 12348 } 12349 12350 /* If device is running, close it first. Start with ops locked... */ 12351 list_for_each_entry(dev, head, unreg_list) { 12352 if (!(dev->flags & IFF_UP)) 12353 continue; 12354 if (netdev_need_ops_lock(dev)) { 12355 list_add_tail(&dev->close_list, &close_head); 12356 netdev_lock(dev); 12357 } 12358 netif_close_many_and_unlock_cond(&close_head); 12359 } 12360 netif_close_many_and_unlock(&close_head); 12361 /* ... now go over the rest. */ 12362 list_for_each_entry(dev, head, unreg_list) { 12363 if (!netdev_need_ops_lock(dev)) 12364 list_add_tail(&dev->close_list, &close_head); 12365 } 12366 netif_close_many(&close_head, true); 12367 12368 list_for_each_entry(dev, head, unreg_list) { 12369 /* And unlink it from device chain. */ 12370 unlist_netdevice(dev); 12371 netdev_lock(dev); 12372 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERING); 12373 netdev_unlock(dev); 12374 } 12375 flush_all_backlogs(); 12376 12377 synchronize_net(); 12378 12379 list_for_each_entry(dev, head, unreg_list) { 12380 struct sk_buff *skb = NULL; 12381 12382 /* Shutdown queueing discipline. */ 12383 netdev_lock_ops(dev); 12384 dev_shutdown(dev); 12385 dev_tcx_uninstall(dev); 12386 dev_xdp_uninstall(dev); 12387 dev_memory_provider_uninstall(dev); 12388 netdev_unlock_ops(dev); 12389 bpf_dev_bound_netdev_unregister(dev); 12390 12391 netdev_offload_xstats_disable_all(dev); 12392 12393 /* Notify protocols, that we are about to destroy 12394 * this device. They should clean all the things. 12395 */ 12396 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 12397 12398 if (!(dev->rtnl_link_ops && dev->rtnl_link_initializing)) 12399 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0, 12400 GFP_KERNEL, NULL, 0, 12401 portid, nlh); 12402 12403 /* 12404 * Flush the unicast and multicast chains 12405 */ 12406 dev_uc_flush(dev); 12407 dev_mc_flush(dev); 12408 12409 netdev_name_node_alt_flush(dev); 12410 netdev_name_node_free(dev->name_node); 12411 12412 netdev_rss_contexts_free(dev); 12413 12414 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 12415 12416 if (dev->netdev_ops->ndo_uninit) 12417 dev->netdev_ops->ndo_uninit(dev); 12418 12419 mutex_destroy(&dev->ethtool->rss_lock); 12420 12421 net_shaper_flush_netdev(dev); 12422 12423 if (skb) 12424 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh); 12425 12426 /* Notifier chain MUST detach us all upper devices. */ 12427 WARN_ON(netdev_has_any_upper_dev(dev)); 12428 WARN_ON(netdev_has_any_lower_dev(dev)); 12429 12430 /* Remove entries from kobject tree */ 12431 netdev_unregister_kobject(dev); 12432 #ifdef CONFIG_XPS 12433 /* Remove XPS queueing entries */ 12434 netif_reset_xps_queues_gt(dev, 0); 12435 #endif 12436 } 12437 12438 synchronize_net(); 12439 12440 list_for_each_entry(dev, head, unreg_list) { 12441 netdev_put(dev, &dev->dev_registered_tracker); 12442 net_set_todo(dev); 12443 cnt++; 12444 } 12445 atomic_add(cnt, &dev_unreg_count); 12446 12447 list_del(head); 12448 } 12449 12450 /** 12451 * unregister_netdevice_many - unregister many devices 12452 * @head: list of devices 12453 * 12454 * Note: As most callers use a stack allocated list_head, 12455 * we force a list_del() to make sure stack won't be corrupted later. 12456 */ 12457 void unregister_netdevice_many(struct list_head *head) 12458 { 12459 unregister_netdevice_many_notify(head, 0, NULL); 12460 } 12461 EXPORT_SYMBOL(unregister_netdevice_many); 12462 12463 /** 12464 * unregister_netdev - remove device from the kernel 12465 * @dev: device 12466 * 12467 * This function shuts down a device interface and removes it 12468 * from the kernel tables. 12469 * 12470 * This is just a wrapper for unregister_netdevice that takes 12471 * the rtnl semaphore. In general you want to use this and not 12472 * unregister_netdevice. 12473 */ 12474 void unregister_netdev(struct net_device *dev) 12475 { 12476 rtnl_net_dev_lock(dev); 12477 unregister_netdevice(dev); 12478 rtnl_net_dev_unlock(dev); 12479 } 12480 EXPORT_SYMBOL(unregister_netdev); 12481 12482 int __dev_change_net_namespace(struct net_device *dev, struct net *net, 12483 const char *pat, int new_ifindex, 12484 struct netlink_ext_ack *extack) 12485 { 12486 struct netdev_name_node *name_node; 12487 struct net *net_old = dev_net(dev); 12488 char new_name[IFNAMSIZ] = {}; 12489 int err, new_nsid; 12490 12491 ASSERT_RTNL(); 12492 12493 /* Don't allow namespace local devices to be moved. */ 12494 err = -EINVAL; 12495 if (dev->netns_immutable) { 12496 NL_SET_ERR_MSG(extack, "The interface netns is immutable"); 12497 goto out; 12498 } 12499 12500 /* Ensure the device has been registered */ 12501 if (dev->reg_state != NETREG_REGISTERED) { 12502 NL_SET_ERR_MSG(extack, "The interface isn't registered"); 12503 goto out; 12504 } 12505 12506 /* Get out if there is nothing todo */ 12507 err = 0; 12508 if (net_eq(net_old, net)) 12509 goto out; 12510 12511 /* Pick the destination device name, and ensure 12512 * we can use it in the destination network namespace. 12513 */ 12514 err = -EEXIST; 12515 if (netdev_name_in_use(net, dev->name)) { 12516 /* We get here if we can't use the current device name */ 12517 if (!pat) { 12518 NL_SET_ERR_MSG(extack, 12519 "An interface with the same name exists in the target netns"); 12520 goto out; 12521 } 12522 err = dev_prep_valid_name(net, dev, pat, new_name, EEXIST); 12523 if (err < 0) { 12524 NL_SET_ERR_MSG_FMT(extack, 12525 "Unable to use '%s' for the new interface name in the target netns", 12526 pat); 12527 goto out; 12528 } 12529 } 12530 /* Check that none of the altnames conflicts. */ 12531 err = -EEXIST; 12532 netdev_for_each_altname(dev, name_node) { 12533 if (netdev_name_in_use(net, name_node->name)) { 12534 NL_SET_ERR_MSG_FMT(extack, 12535 "An interface with the altname %s exists in the target netns", 12536 name_node->name); 12537 goto out; 12538 } 12539 } 12540 12541 /* Check that new_ifindex isn't used yet. */ 12542 if (new_ifindex) { 12543 err = dev_index_reserve(net, new_ifindex); 12544 if (err < 0) { 12545 NL_SET_ERR_MSG_FMT(extack, 12546 "The ifindex %d is not available in the target netns", 12547 new_ifindex); 12548 goto out; 12549 } 12550 } else { 12551 /* If there is an ifindex conflict assign a new one */ 12552 err = dev_index_reserve(net, dev->ifindex); 12553 if (err == -EBUSY) 12554 err = dev_index_reserve(net, 0); 12555 if (err < 0) { 12556 NL_SET_ERR_MSG(extack, 12557 "Unable to allocate a new ifindex in the target netns"); 12558 goto out; 12559 } 12560 new_ifindex = err; 12561 } 12562 12563 /* 12564 * And now a mini version of register_netdevice unregister_netdevice. 12565 */ 12566 12567 netdev_lock_ops(dev); 12568 /* If device is running close it first. */ 12569 netif_close(dev); 12570 /* And unlink it from device chain */ 12571 unlist_netdevice(dev); 12572 12573 if (!netdev_need_ops_lock(dev)) 12574 netdev_lock(dev); 12575 dev->moving_ns = true; 12576 netdev_unlock(dev); 12577 12578 synchronize_net(); 12579 12580 /* Shutdown queueing discipline. */ 12581 netdev_lock_ops(dev); 12582 dev_shutdown(dev); 12583 netdev_unlock_ops(dev); 12584 12585 /* Notify protocols, that we are about to destroy 12586 * this device. They should clean all the things. 12587 * 12588 * Note that dev->reg_state stays at NETREG_REGISTERED. 12589 * This is wanted because this way 8021q and macvlan know 12590 * the device is just moving and can keep their slaves up. 12591 */ 12592 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 12593 rcu_barrier(); 12594 12595 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL); 12596 12597 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid, 12598 new_ifindex); 12599 12600 /* 12601 * Flush the unicast and multicast chains 12602 */ 12603 dev_uc_flush(dev); 12604 dev_mc_flush(dev); 12605 12606 /* Send a netdev-removed uevent to the old namespace */ 12607 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE); 12608 netdev_adjacent_del_links(dev); 12609 12610 /* Move per-net netdevice notifiers that are following the netdevice */ 12611 move_netdevice_notifiers_dev_net(dev, net); 12612 12613 /* Actually switch the network namespace */ 12614 netdev_lock(dev); 12615 dev_net_set(dev, net); 12616 netdev_unlock(dev); 12617 dev->ifindex = new_ifindex; 12618 12619 if (new_name[0]) { 12620 /* Rename the netdev to prepared name */ 12621 write_seqlock_bh(&netdev_rename_lock); 12622 strscpy(dev->name, new_name, IFNAMSIZ); 12623 write_sequnlock_bh(&netdev_rename_lock); 12624 } 12625 12626 /* Fixup kobjects */ 12627 dev_set_uevent_suppress(&dev->dev, 1); 12628 err = device_rename(&dev->dev, dev->name); 12629 dev_set_uevent_suppress(&dev->dev, 0); 12630 WARN_ON(err); 12631 12632 /* Send a netdev-add uevent to the new namespace */ 12633 kobject_uevent(&dev->dev.kobj, KOBJ_ADD); 12634 netdev_adjacent_add_links(dev); 12635 12636 /* Adapt owner in case owning user namespace of target network 12637 * namespace is different from the original one. 12638 */ 12639 err = netdev_change_owner(dev, net_old, net); 12640 WARN_ON(err); 12641 12642 netdev_lock(dev); 12643 dev->moving_ns = false; 12644 if (!netdev_need_ops_lock(dev)) 12645 netdev_unlock(dev); 12646 12647 /* Add the device back in the hashes */ 12648 list_netdevice(dev); 12649 /* Notify protocols, that a new device appeared. */ 12650 call_netdevice_notifiers(NETDEV_REGISTER, dev); 12651 netdev_unlock_ops(dev); 12652 12653 /* 12654 * Prevent userspace races by waiting until the network 12655 * device is fully setup before sending notifications. 12656 */ 12657 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 12658 12659 synchronize_net(); 12660 err = 0; 12661 out: 12662 return err; 12663 } 12664 12665 static int dev_cpu_dead(unsigned int oldcpu) 12666 { 12667 struct sk_buff **list_skb; 12668 struct sk_buff *skb; 12669 unsigned int cpu; 12670 struct softnet_data *sd, *oldsd, *remsd = NULL; 12671 12672 local_irq_disable(); 12673 cpu = smp_processor_id(); 12674 sd = &per_cpu(softnet_data, cpu); 12675 oldsd = &per_cpu(softnet_data, oldcpu); 12676 12677 /* Find end of our completion_queue. */ 12678 list_skb = &sd->completion_queue; 12679 while (*list_skb) 12680 list_skb = &(*list_skb)->next; 12681 /* Append completion queue from offline CPU. */ 12682 *list_skb = oldsd->completion_queue; 12683 oldsd->completion_queue = NULL; 12684 12685 /* Append output queue from offline CPU. */ 12686 if (oldsd->output_queue) { 12687 *sd->output_queue_tailp = oldsd->output_queue; 12688 sd->output_queue_tailp = oldsd->output_queue_tailp; 12689 oldsd->output_queue = NULL; 12690 oldsd->output_queue_tailp = &oldsd->output_queue; 12691 } 12692 /* Append NAPI poll list from offline CPU, with one exception : 12693 * process_backlog() must be called by cpu owning percpu backlog. 12694 * We properly handle process_queue & input_pkt_queue later. 12695 */ 12696 while (!list_empty(&oldsd->poll_list)) { 12697 struct napi_struct *napi = list_first_entry(&oldsd->poll_list, 12698 struct napi_struct, 12699 poll_list); 12700 12701 list_del_init(&napi->poll_list); 12702 if (napi->poll == process_backlog) 12703 napi->state &= NAPIF_STATE_THREADED; 12704 else 12705 ____napi_schedule(sd, napi); 12706 } 12707 12708 raise_softirq_irqoff(NET_TX_SOFTIRQ); 12709 local_irq_enable(); 12710 12711 if (!use_backlog_threads()) { 12712 #ifdef CONFIG_RPS 12713 remsd = oldsd->rps_ipi_list; 12714 oldsd->rps_ipi_list = NULL; 12715 #endif 12716 /* send out pending IPI's on offline CPU */ 12717 net_rps_send_ipi(remsd); 12718 } 12719 12720 /* Process offline CPU's input_pkt_queue */ 12721 while ((skb = __skb_dequeue(&oldsd->process_queue))) { 12722 netif_rx(skb); 12723 rps_input_queue_head_incr(oldsd); 12724 } 12725 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) { 12726 netif_rx(skb); 12727 rps_input_queue_head_incr(oldsd); 12728 } 12729 12730 return 0; 12731 } 12732 12733 /** 12734 * netdev_increment_features - increment feature set by one 12735 * @all: current feature set 12736 * @one: new feature set 12737 * @mask: mask feature set 12738 * 12739 * Computes a new feature set after adding a device with feature set 12740 * @one to the master device with current feature set @all. Will not 12741 * enable anything that is off in @mask. Returns the new feature set. 12742 */ 12743 netdev_features_t netdev_increment_features(netdev_features_t all, 12744 netdev_features_t one, netdev_features_t mask) 12745 { 12746 if (mask & NETIF_F_HW_CSUM) 12747 mask |= NETIF_F_CSUM_MASK; 12748 mask |= NETIF_F_VLAN_CHALLENGED; 12749 12750 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask; 12751 all &= one | ~NETIF_F_ALL_FOR_ALL; 12752 12753 /* If one device supports hw checksumming, set for all. */ 12754 if (all & NETIF_F_HW_CSUM) 12755 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM); 12756 12757 return all; 12758 } 12759 EXPORT_SYMBOL(netdev_increment_features); 12760 12761 /** 12762 * netdev_compute_master_upper_features - compute feature from lowers 12763 * @dev: the upper device 12764 * @update_header: whether to update upper device's header_len/headroom/tailroom 12765 * 12766 * Recompute the upper device's feature based on all lower devices. 12767 */ 12768 void netdev_compute_master_upper_features(struct net_device *dev, bool update_header) 12769 { 12770 unsigned int dst_release_flag = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 12771 netdev_features_t gso_partial_features = MASTER_UPPER_DEV_GSO_PARTIAL_FEATURES; 12772 netdev_features_t xfrm_features = MASTER_UPPER_DEV_XFRM_FEATURES; 12773 netdev_features_t mpls_features = MASTER_UPPER_DEV_MPLS_FEATURES; 12774 netdev_features_t vlan_features = MASTER_UPPER_DEV_VLAN_FEATURES; 12775 netdev_features_t enc_features = MASTER_UPPER_DEV_ENC_FEATURES; 12776 unsigned short max_header_len = ETH_HLEN; 12777 unsigned int tso_max_size = TSO_MAX_SIZE; 12778 unsigned short max_headroom = 0; 12779 unsigned short max_tailroom = 0; 12780 u16 tso_max_segs = TSO_MAX_SEGS; 12781 struct net_device *lower_dev; 12782 struct list_head *iter; 12783 12784 mpls_features = netdev_base_features(mpls_features); 12785 vlan_features = netdev_base_features(vlan_features); 12786 enc_features = netdev_base_features(enc_features); 12787 12788 netdev_for_each_lower_dev(dev, lower_dev, iter) { 12789 gso_partial_features = netdev_increment_features(gso_partial_features, 12790 lower_dev->gso_partial_features, 12791 MASTER_UPPER_DEV_GSO_PARTIAL_FEATURES); 12792 12793 vlan_features = netdev_increment_features(vlan_features, 12794 lower_dev->vlan_features, 12795 MASTER_UPPER_DEV_VLAN_FEATURES); 12796 12797 enc_features = netdev_increment_features(enc_features, 12798 lower_dev->hw_enc_features, 12799 MASTER_UPPER_DEV_ENC_FEATURES); 12800 12801 if (IS_ENABLED(CONFIG_XFRM_OFFLOAD)) 12802 xfrm_features = netdev_increment_features(xfrm_features, 12803 lower_dev->hw_enc_features, 12804 MASTER_UPPER_DEV_XFRM_FEATURES); 12805 12806 mpls_features = netdev_increment_features(mpls_features, 12807 lower_dev->mpls_features, 12808 MASTER_UPPER_DEV_MPLS_FEATURES); 12809 12810 dst_release_flag &= lower_dev->priv_flags; 12811 12812 if (update_header) { 12813 max_header_len = max(max_header_len, lower_dev->hard_header_len); 12814 max_headroom = max(max_headroom, lower_dev->needed_headroom); 12815 max_tailroom = max(max_tailroom, lower_dev->needed_tailroom); 12816 } 12817 12818 tso_max_size = min(tso_max_size, lower_dev->tso_max_size); 12819 tso_max_segs = min(tso_max_segs, lower_dev->tso_max_segs); 12820 } 12821 12822 dev->gso_partial_features = gso_partial_features; 12823 dev->vlan_features = vlan_features; 12824 dev->hw_enc_features = enc_features | NETIF_F_GSO_ENCAP_ALL | 12825 NETIF_F_HW_VLAN_CTAG_TX | 12826 NETIF_F_HW_VLAN_STAG_TX; 12827 if (IS_ENABLED(CONFIG_XFRM_OFFLOAD)) 12828 dev->hw_enc_features |= xfrm_features; 12829 dev->mpls_features = mpls_features; 12830 12831 dev->priv_flags &= ~IFF_XMIT_DST_RELEASE; 12832 if ((dev->priv_flags & IFF_XMIT_DST_RELEASE_PERM) && 12833 dst_release_flag == (IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM)) 12834 dev->priv_flags |= IFF_XMIT_DST_RELEASE; 12835 12836 if (update_header) { 12837 dev->hard_header_len = max_header_len; 12838 dev->needed_headroom = max_headroom; 12839 dev->needed_tailroom = max_tailroom; 12840 } 12841 12842 netif_set_tso_max_segs(dev, tso_max_segs); 12843 netif_set_tso_max_size(dev, tso_max_size); 12844 12845 netdev_change_features(dev); 12846 } 12847 EXPORT_SYMBOL(netdev_compute_master_upper_features); 12848 12849 static struct hlist_head * __net_init netdev_create_hash(void) 12850 { 12851 int i; 12852 struct hlist_head *hash; 12853 12854 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL); 12855 if (hash != NULL) 12856 for (i = 0; i < NETDEV_HASHENTRIES; i++) 12857 INIT_HLIST_HEAD(&hash[i]); 12858 12859 return hash; 12860 } 12861 12862 /* Initialize per network namespace state */ 12863 static int __net_init netdev_init(struct net *net) 12864 { 12865 BUILD_BUG_ON(GRO_HASH_BUCKETS > 12866 BITS_PER_BYTE * sizeof_field(struct gro_node, bitmask)); 12867 12868 INIT_LIST_HEAD(&net->dev_base_head); 12869 12870 net->dev_name_head = netdev_create_hash(); 12871 if (net->dev_name_head == NULL) 12872 goto err_name; 12873 12874 net->dev_index_head = netdev_create_hash(); 12875 if (net->dev_index_head == NULL) 12876 goto err_idx; 12877 12878 xa_init_flags(&net->dev_by_index, XA_FLAGS_ALLOC1); 12879 12880 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain); 12881 12882 return 0; 12883 12884 err_idx: 12885 kfree(net->dev_name_head); 12886 err_name: 12887 return -ENOMEM; 12888 } 12889 12890 /** 12891 * netdev_drivername - network driver for the device 12892 * @dev: network device 12893 * 12894 * Determine network driver for device. 12895 */ 12896 const char *netdev_drivername(const struct net_device *dev) 12897 { 12898 const struct device_driver *driver; 12899 const struct device *parent; 12900 const char *empty = ""; 12901 12902 parent = dev->dev.parent; 12903 if (!parent) 12904 return empty; 12905 12906 driver = parent->driver; 12907 if (driver && driver->name) 12908 return driver->name; 12909 return empty; 12910 } 12911 12912 static void __netdev_printk(const char *level, const struct net_device *dev, 12913 struct va_format *vaf) 12914 { 12915 if (dev && dev->dev.parent) { 12916 dev_printk_emit(level[1] - '0', 12917 dev->dev.parent, 12918 "%s %s %s%s: %pV", 12919 dev_driver_string(dev->dev.parent), 12920 dev_name(dev->dev.parent), 12921 netdev_name(dev), netdev_reg_state(dev), 12922 vaf); 12923 } else if (dev) { 12924 printk("%s%s%s: %pV", 12925 level, netdev_name(dev), netdev_reg_state(dev), vaf); 12926 } else { 12927 printk("%s(NULL net_device): %pV", level, vaf); 12928 } 12929 } 12930 12931 void netdev_printk(const char *level, const struct net_device *dev, 12932 const char *format, ...) 12933 { 12934 struct va_format vaf; 12935 va_list args; 12936 12937 va_start(args, format); 12938 12939 vaf.fmt = format; 12940 vaf.va = &args; 12941 12942 __netdev_printk(level, dev, &vaf); 12943 12944 va_end(args); 12945 } 12946 EXPORT_SYMBOL(netdev_printk); 12947 12948 #define define_netdev_printk_level(func, level) \ 12949 void func(const struct net_device *dev, const char *fmt, ...) \ 12950 { \ 12951 struct va_format vaf; \ 12952 va_list args; \ 12953 \ 12954 va_start(args, fmt); \ 12955 \ 12956 vaf.fmt = fmt; \ 12957 vaf.va = &args; \ 12958 \ 12959 __netdev_printk(level, dev, &vaf); \ 12960 \ 12961 va_end(args); \ 12962 } \ 12963 EXPORT_SYMBOL(func); 12964 12965 define_netdev_printk_level(netdev_emerg, KERN_EMERG); 12966 define_netdev_printk_level(netdev_alert, KERN_ALERT); 12967 define_netdev_printk_level(netdev_crit, KERN_CRIT); 12968 define_netdev_printk_level(netdev_err, KERN_ERR); 12969 define_netdev_printk_level(netdev_warn, KERN_WARNING); 12970 define_netdev_printk_level(netdev_notice, KERN_NOTICE); 12971 define_netdev_printk_level(netdev_info, KERN_INFO); 12972 12973 static void __net_exit netdev_exit(struct net *net) 12974 { 12975 kfree(net->dev_name_head); 12976 kfree(net->dev_index_head); 12977 xa_destroy(&net->dev_by_index); 12978 if (net != &init_net) 12979 WARN_ON_ONCE(!list_empty(&net->dev_base_head)); 12980 } 12981 12982 static struct pernet_operations __net_initdata netdev_net_ops = { 12983 .init = netdev_init, 12984 .exit = netdev_exit, 12985 }; 12986 12987 static void __net_exit default_device_exit_net(struct net *net) 12988 { 12989 struct netdev_name_node *name_node, *tmp; 12990 struct net_device *dev, *aux; 12991 /* 12992 * Push all migratable network devices back to the 12993 * initial network namespace 12994 */ 12995 ASSERT_RTNL(); 12996 for_each_netdev_safe(net, dev, aux) { 12997 int err; 12998 char fb_name[IFNAMSIZ]; 12999 13000 /* Ignore unmoveable devices (i.e. loopback) */ 13001 if (dev->netns_immutable) 13002 continue; 13003 13004 /* Leave virtual devices for the generic cleanup */ 13005 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund) 13006 continue; 13007 13008 /* Push remaining network devices to init_net */ 13009 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex); 13010 if (netdev_name_in_use(&init_net, fb_name)) 13011 snprintf(fb_name, IFNAMSIZ, "dev%%d"); 13012 13013 netdev_for_each_altname_safe(dev, name_node, tmp) 13014 if (netdev_name_in_use(&init_net, name_node->name)) 13015 __netdev_name_node_alt_destroy(name_node); 13016 13017 err = dev_change_net_namespace(dev, &init_net, fb_name); 13018 if (err) { 13019 pr_emerg("%s: failed to move %s to init_net: %d\n", 13020 __func__, dev->name, err); 13021 BUG(); 13022 } 13023 } 13024 } 13025 13026 static void __net_exit default_device_exit_batch(struct list_head *net_list) 13027 { 13028 /* At exit all network devices most be removed from a network 13029 * namespace. Do this in the reverse order of registration. 13030 * Do this across as many network namespaces as possible to 13031 * improve batching efficiency. 13032 */ 13033 struct net_device *dev; 13034 struct net *net; 13035 LIST_HEAD(dev_kill_list); 13036 13037 rtnl_lock(); 13038 list_for_each_entry(net, net_list, exit_list) { 13039 default_device_exit_net(net); 13040 cond_resched(); 13041 } 13042 13043 list_for_each_entry(net, net_list, exit_list) { 13044 for_each_netdev_reverse(net, dev) { 13045 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink) 13046 dev->rtnl_link_ops->dellink(dev, &dev_kill_list); 13047 else 13048 unregister_netdevice_queue(dev, &dev_kill_list); 13049 } 13050 } 13051 unregister_netdevice_many(&dev_kill_list); 13052 rtnl_unlock(); 13053 } 13054 13055 static struct pernet_operations __net_initdata default_device_ops = { 13056 .exit_batch = default_device_exit_batch, 13057 }; 13058 13059 static void __init net_dev_struct_check(void) 13060 { 13061 /* TX read-mostly hotpath */ 13062 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, priv_flags_fast); 13063 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, netdev_ops); 13064 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, header_ops); 13065 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, _tx); 13066 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, real_num_tx_queues); 13067 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_size); 13068 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_ipv4_max_size); 13069 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_segs); 13070 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_partial_features); 13071 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, num_tc); 13072 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, mtu); 13073 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, needed_headroom); 13074 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tc_to_txq); 13075 #ifdef CONFIG_XPS 13076 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, xps_maps); 13077 #endif 13078 #ifdef CONFIG_NETFILTER_EGRESS 13079 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, nf_hooks_egress); 13080 #endif 13081 #ifdef CONFIG_NET_XGRESS 13082 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tcx_egress); 13083 #endif 13084 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_tx, 160); 13085 13086 /* TXRX read-mostly hotpath */ 13087 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, lstats); 13088 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, state); 13089 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, flags); 13090 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, hard_header_len); 13091 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, features); 13092 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, ip6_ptr); 13093 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_txrx, 46); 13094 13095 /* RX read-mostly hotpath */ 13096 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ptype_specific); 13097 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ifindex); 13098 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, real_num_rx_queues); 13099 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, _rx); 13100 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_max_size); 13101 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_ipv4_max_size); 13102 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler); 13103 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler_data); 13104 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, nd_net); 13105 #ifdef CONFIG_NETPOLL 13106 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, npinfo); 13107 #endif 13108 #ifdef CONFIG_NET_XGRESS 13109 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, tcx_ingress); 13110 #endif 13111 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_rx, 92); 13112 } 13113 13114 /* 13115 * Initialize the DEV module. At boot time this walks the device list and 13116 * unhooks any devices that fail to initialise (normally hardware not 13117 * present) and leaves us with a valid list of present and active devices. 13118 * 13119 */ 13120 13121 /* We allocate 256 pages for each CPU if PAGE_SHIFT is 12 */ 13122 #define SYSTEM_PERCPU_PAGE_POOL_SIZE ((1 << 20) / PAGE_SIZE) 13123 13124 static int net_page_pool_create(int cpuid) 13125 { 13126 #if IS_ENABLED(CONFIG_PAGE_POOL) 13127 struct page_pool_params page_pool_params = { 13128 .pool_size = SYSTEM_PERCPU_PAGE_POOL_SIZE, 13129 .flags = PP_FLAG_SYSTEM_POOL, 13130 .nid = cpu_to_mem(cpuid), 13131 }; 13132 struct page_pool *pp_ptr; 13133 int err; 13134 13135 pp_ptr = page_pool_create_percpu(&page_pool_params, cpuid); 13136 if (IS_ERR(pp_ptr)) 13137 return -ENOMEM; 13138 13139 err = xdp_reg_page_pool(pp_ptr); 13140 if (err) { 13141 page_pool_destroy(pp_ptr); 13142 return err; 13143 } 13144 13145 per_cpu(system_page_pool.pool, cpuid) = pp_ptr; 13146 #endif 13147 return 0; 13148 } 13149 13150 static int backlog_napi_should_run(unsigned int cpu) 13151 { 13152 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 13153 struct napi_struct *napi = &sd->backlog; 13154 13155 return test_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 13156 } 13157 13158 static void run_backlog_napi(unsigned int cpu) 13159 { 13160 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 13161 13162 napi_threaded_poll_loop(&sd->backlog, false); 13163 } 13164 13165 static void backlog_napi_setup(unsigned int cpu) 13166 { 13167 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 13168 struct napi_struct *napi = &sd->backlog; 13169 13170 napi->thread = this_cpu_read(backlog_napi); 13171 set_bit(NAPI_STATE_THREADED, &napi->state); 13172 } 13173 13174 static struct smp_hotplug_thread backlog_threads = { 13175 .store = &backlog_napi, 13176 .thread_should_run = backlog_napi_should_run, 13177 .thread_fn = run_backlog_napi, 13178 .thread_comm = "backlog_napi/%u", 13179 .setup = backlog_napi_setup, 13180 }; 13181 13182 /* 13183 * This is called single threaded during boot, so no need 13184 * to take the rtnl semaphore. 13185 */ 13186 static int __init net_dev_init(void) 13187 { 13188 int i, rc = -ENOMEM; 13189 13190 BUG_ON(!dev_boot_phase); 13191 13192 net_dev_struct_check(); 13193 13194 if (dev_proc_init()) 13195 goto out; 13196 13197 if (netdev_kobject_init()) 13198 goto out; 13199 13200 for (i = 0; i < PTYPE_HASH_SIZE; i++) 13201 INIT_LIST_HEAD(&ptype_base[i]); 13202 13203 if (register_pernet_subsys(&netdev_net_ops)) 13204 goto out; 13205 13206 /* 13207 * Initialise the packet receive queues. 13208 */ 13209 13210 flush_backlogs_fallback = flush_backlogs_alloc(); 13211 if (!flush_backlogs_fallback) 13212 goto out; 13213 13214 for_each_possible_cpu(i) { 13215 struct softnet_data *sd = &per_cpu(softnet_data, i); 13216 13217 skb_queue_head_init(&sd->input_pkt_queue); 13218 skb_queue_head_init(&sd->process_queue); 13219 #ifdef CONFIG_XFRM_OFFLOAD 13220 skb_queue_head_init(&sd->xfrm_backlog); 13221 #endif 13222 INIT_LIST_HEAD(&sd->poll_list); 13223 sd->output_queue_tailp = &sd->output_queue; 13224 #ifdef CONFIG_RPS 13225 INIT_CSD(&sd->csd, rps_trigger_softirq, sd); 13226 sd->cpu = i; 13227 #endif 13228 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd); 13229 13230 gro_init(&sd->backlog.gro); 13231 sd->backlog.poll = process_backlog; 13232 sd->backlog.weight = weight_p; 13233 INIT_LIST_HEAD(&sd->backlog.poll_list); 13234 13235 if (net_page_pool_create(i)) 13236 goto out; 13237 } 13238 net_hotdata.skb_defer_nodes = 13239 __alloc_percpu(sizeof(struct skb_defer_node) * nr_node_ids, 13240 __alignof__(struct skb_defer_node)); 13241 if (!net_hotdata.skb_defer_nodes) 13242 goto out; 13243 if (use_backlog_threads()) 13244 smpboot_register_percpu_thread(&backlog_threads); 13245 13246 dev_boot_phase = 0; 13247 13248 /* The loopback device is special if any other network devices 13249 * is present in a network namespace the loopback device must 13250 * be present. Since we now dynamically allocate and free the 13251 * loopback device ensure this invariant is maintained by 13252 * keeping the loopback device as the first device on the 13253 * list of network devices. Ensuring the loopback devices 13254 * is the first device that appears and the last network device 13255 * that disappears. 13256 */ 13257 if (register_pernet_device(&loopback_net_ops)) 13258 goto out; 13259 13260 if (register_pernet_device(&default_device_ops)) 13261 goto out; 13262 13263 open_softirq(NET_TX_SOFTIRQ, net_tx_action); 13264 open_softirq(NET_RX_SOFTIRQ, net_rx_action); 13265 13266 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead", 13267 NULL, dev_cpu_dead); 13268 WARN_ON(rc < 0); 13269 rc = 0; 13270 13271 /* avoid static key IPIs to isolated CPUs */ 13272 if (housekeeping_enabled(HK_TYPE_MISC)) 13273 net_enable_timestamp(); 13274 out: 13275 if (rc < 0) { 13276 for_each_possible_cpu(i) { 13277 struct page_pool *pp_ptr; 13278 13279 pp_ptr = per_cpu(system_page_pool.pool, i); 13280 if (!pp_ptr) 13281 continue; 13282 13283 xdp_unreg_page_pool(pp_ptr); 13284 page_pool_destroy(pp_ptr); 13285 per_cpu(system_page_pool.pool, i) = NULL; 13286 } 13287 } 13288 13289 return rc; 13290 } 13291 13292 subsys_initcall(net_dev_init); 13293