1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * NET3 Protocol independent device support routines. 4 * 5 * Derived from the non IP parts of dev.c 1.0.19 6 * Authors: Ross Biro 7 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 8 * Mark Evans, <evansmp@uhura.aston.ac.uk> 9 * 10 * Additional Authors: 11 * Florian la Roche <rzsfl@rz.uni-sb.de> 12 * Alan Cox <gw4pts@gw4pts.ampr.org> 13 * David Hinds <dahinds@users.sourceforge.net> 14 * Alexey Kuznetsov <kuznet@ms2.inr.ac.ru> 15 * Adam Sulmicki <adam@cfar.umd.edu> 16 * Pekka Riikonen <priikone@poesidon.pspt.fi> 17 * 18 * Changes: 19 * D.J. Barrow : Fixed bug where dev->refcnt gets set 20 * to 2 if register_netdev gets called 21 * before net_dev_init & also removed a 22 * few lines of code in the process. 23 * Alan Cox : device private ioctl copies fields back. 24 * Alan Cox : Transmit queue code does relevant 25 * stunts to keep the queue safe. 26 * Alan Cox : Fixed double lock. 27 * Alan Cox : Fixed promisc NULL pointer trap 28 * ???????? : Support the full private ioctl range 29 * Alan Cox : Moved ioctl permission check into 30 * drivers 31 * Tim Kordas : SIOCADDMULTI/SIOCDELMULTI 32 * Alan Cox : 100 backlog just doesn't cut it when 33 * you start doing multicast video 8) 34 * Alan Cox : Rewrote net_bh and list manager. 35 * Alan Cox : Fix ETH_P_ALL echoback lengths. 36 * Alan Cox : Took out transmit every packet pass 37 * Saved a few bytes in the ioctl handler 38 * Alan Cox : Network driver sets packet type before 39 * calling netif_rx. Saves a function 40 * call a packet. 41 * Alan Cox : Hashed net_bh() 42 * Richard Kooijman: Timestamp fixes. 43 * Alan Cox : Wrong field in SIOCGIFDSTADDR 44 * Alan Cox : Device lock protection. 45 * Alan Cox : Fixed nasty side effect of device close 46 * changes. 47 * Rudi Cilibrasi : Pass the right thing to 48 * set_mac_address() 49 * Dave Miller : 32bit quantity for the device lock to 50 * make it work out on a Sparc. 51 * Bjorn Ekwall : Added KERNELD hack. 52 * Alan Cox : Cleaned up the backlog initialise. 53 * Craig Metz : SIOCGIFCONF fix if space for under 54 * 1 device. 55 * Thomas Bogendoerfer : Return ENODEV for dev_open, if there 56 * is no device open function. 57 * Andi Kleen : Fix error reporting for SIOCGIFCONF 58 * Michael Chastain : Fix signed/unsigned for SIOCGIFCONF 59 * Cyrus Durgin : Cleaned for KMOD 60 * Adam Sulmicki : Bug Fix : Network Device Unload 61 * A network device unload needs to purge 62 * the backlog queue. 63 * Paul Rusty Russell : SIOCSIFNAME 64 * Pekka Riikonen : Netdev boot-time settings code 65 * Andrew Morton : Make unregister_netdevice wait 66 * indefinitely on dev->refcnt 67 * J Hadi Salim : - Backlog queue sampling 68 * - netif_rx() feedback 69 */ 70 71 #include <linux/uaccess.h> 72 #include <linux/bitmap.h> 73 #include <linux/capability.h> 74 #include <linux/cpu.h> 75 #include <linux/types.h> 76 #include <linux/kernel.h> 77 #include <linux/hash.h> 78 #include <linux/slab.h> 79 #include <linux/sched.h> 80 #include <linux/sched/isolation.h> 81 #include <linux/sched/mm.h> 82 #include <linux/smpboot.h> 83 #include <linux/mutex.h> 84 #include <linux/rwsem.h> 85 #include <linux/string.h> 86 #include <linux/mm.h> 87 #include <linux/socket.h> 88 #include <linux/sockios.h> 89 #include <linux/errno.h> 90 #include <linux/interrupt.h> 91 #include <linux/if_ether.h> 92 #include <linux/netdevice.h> 93 #include <linux/etherdevice.h> 94 #include <linux/ethtool.h> 95 #include <linux/ethtool_netlink.h> 96 #include <linux/skbuff.h> 97 #include <linux/kthread.h> 98 #include <linux/bpf.h> 99 #include <linux/bpf_trace.h> 100 #include <net/net_namespace.h> 101 #include <net/sock.h> 102 #include <net/busy_poll.h> 103 #include <linux/rtnetlink.h> 104 #include <linux/stat.h> 105 #include <net/dsa.h> 106 #include <net/dst.h> 107 #include <net/dst_metadata.h> 108 #include <net/gro.h> 109 #include <net/netdev_queues.h> 110 #include <net/pkt_sched.h> 111 #include <net/pkt_cls.h> 112 #include <net/checksum.h> 113 #include <net/xfrm.h> 114 #include <net/tcx.h> 115 #include <linux/highmem.h> 116 #include <linux/init.h> 117 #include <linux/module.h> 118 #include <linux/netpoll.h> 119 #include <linux/rcupdate.h> 120 #include <linux/delay.h> 121 #include <net/iw_handler.h> 122 #include <asm/current.h> 123 #include <linux/audit.h> 124 #include <linux/dmaengine.h> 125 #include <linux/err.h> 126 #include <linux/ctype.h> 127 #include <linux/if_arp.h> 128 #include <linux/if_vlan.h> 129 #include <linux/ip.h> 130 #include <net/ip.h> 131 #include <net/mpls.h> 132 #include <linux/ipv6.h> 133 #include <linux/in.h> 134 #include <linux/jhash.h> 135 #include <linux/random.h> 136 #include <trace/events/napi.h> 137 #include <trace/events/net.h> 138 #include <trace/events/skb.h> 139 #include <trace/events/qdisc.h> 140 #include <trace/events/xdp.h> 141 #include <linux/inetdevice.h> 142 #include <linux/cpu_rmap.h> 143 #include <linux/static_key.h> 144 #include <linux/hashtable.h> 145 #include <linux/vmalloc.h> 146 #include <linux/if_macvlan.h> 147 #include <linux/errqueue.h> 148 #include <linux/hrtimer.h> 149 #include <linux/netfilter_netdev.h> 150 #include <linux/crash_dump.h> 151 #include <linux/sctp.h> 152 #include <net/udp_tunnel.h> 153 #include <linux/net_namespace.h> 154 #include <linux/indirect_call_wrapper.h> 155 #include <net/devlink.h> 156 #include <linux/pm_runtime.h> 157 #include <linux/prandom.h> 158 #include <linux/once_lite.h> 159 #include <net/netdev_lock.h> 160 #include <net/netdev_rx_queue.h> 161 #include <net/page_pool/types.h> 162 #include <net/page_pool/helpers.h> 163 #include <net/page_pool/memory_provider.h> 164 #include <net/rps.h> 165 #include <linux/phy_link_topology.h> 166 167 #include "dev.h" 168 #include "devmem.h" 169 #include "net-sysfs.h" 170 171 static DEFINE_SPINLOCK(ptype_lock); 172 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly; 173 174 static int netif_rx_internal(struct sk_buff *skb); 175 static int call_netdevice_notifiers_extack(unsigned long val, 176 struct net_device *dev, 177 struct netlink_ext_ack *extack); 178 179 static DEFINE_MUTEX(ifalias_mutex); 180 181 /* protects napi_hash addition/deletion and napi_gen_id */ 182 static DEFINE_SPINLOCK(napi_hash_lock); 183 184 static unsigned int napi_gen_id = NR_CPUS; 185 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8); 186 187 static inline void dev_base_seq_inc(struct net *net) 188 { 189 unsigned int val = net->dev_base_seq + 1; 190 191 WRITE_ONCE(net->dev_base_seq, val ?: 1); 192 } 193 194 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name) 195 { 196 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ)); 197 198 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)]; 199 } 200 201 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex) 202 { 203 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)]; 204 } 205 206 #ifndef CONFIG_PREEMPT_RT 207 208 static DEFINE_STATIC_KEY_FALSE(use_backlog_threads_key); 209 210 static int __init setup_backlog_napi_threads(char *arg) 211 { 212 static_branch_enable(&use_backlog_threads_key); 213 return 0; 214 } 215 early_param("thread_backlog_napi", setup_backlog_napi_threads); 216 217 static bool use_backlog_threads(void) 218 { 219 return static_branch_unlikely(&use_backlog_threads_key); 220 } 221 222 #else 223 224 static bool use_backlog_threads(void) 225 { 226 return true; 227 } 228 229 #endif 230 231 static inline void backlog_lock_irq_save(struct softnet_data *sd, 232 unsigned long *flags) 233 { 234 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 235 spin_lock_irqsave(&sd->input_pkt_queue.lock, *flags); 236 else 237 local_irq_save(*flags); 238 } 239 240 static inline void backlog_lock_irq_disable(struct softnet_data *sd) 241 { 242 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 243 spin_lock_irq(&sd->input_pkt_queue.lock); 244 else 245 local_irq_disable(); 246 } 247 248 static inline void backlog_unlock_irq_restore(struct softnet_data *sd, 249 unsigned long *flags) 250 { 251 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 252 spin_unlock_irqrestore(&sd->input_pkt_queue.lock, *flags); 253 else 254 local_irq_restore(*flags); 255 } 256 257 static inline void backlog_unlock_irq_enable(struct softnet_data *sd) 258 { 259 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 260 spin_unlock_irq(&sd->input_pkt_queue.lock); 261 else 262 local_irq_enable(); 263 } 264 265 static struct netdev_name_node *netdev_name_node_alloc(struct net_device *dev, 266 const char *name) 267 { 268 struct netdev_name_node *name_node; 269 270 name_node = kmalloc(sizeof(*name_node), GFP_KERNEL); 271 if (!name_node) 272 return NULL; 273 INIT_HLIST_NODE(&name_node->hlist); 274 name_node->dev = dev; 275 name_node->name = name; 276 return name_node; 277 } 278 279 static struct netdev_name_node * 280 netdev_name_node_head_alloc(struct net_device *dev) 281 { 282 struct netdev_name_node *name_node; 283 284 name_node = netdev_name_node_alloc(dev, dev->name); 285 if (!name_node) 286 return NULL; 287 INIT_LIST_HEAD(&name_node->list); 288 return name_node; 289 } 290 291 static void netdev_name_node_free(struct netdev_name_node *name_node) 292 { 293 kfree(name_node); 294 } 295 296 static void netdev_name_node_add(struct net *net, 297 struct netdev_name_node *name_node) 298 { 299 hlist_add_head_rcu(&name_node->hlist, 300 dev_name_hash(net, name_node->name)); 301 } 302 303 static void netdev_name_node_del(struct netdev_name_node *name_node) 304 { 305 hlist_del_rcu(&name_node->hlist); 306 } 307 308 static struct netdev_name_node *netdev_name_node_lookup(struct net *net, 309 const char *name) 310 { 311 struct hlist_head *head = dev_name_hash(net, name); 312 struct netdev_name_node *name_node; 313 314 hlist_for_each_entry(name_node, head, hlist) 315 if (!strcmp(name_node->name, name)) 316 return name_node; 317 return NULL; 318 } 319 320 static struct netdev_name_node *netdev_name_node_lookup_rcu(struct net *net, 321 const char *name) 322 { 323 struct hlist_head *head = dev_name_hash(net, name); 324 struct netdev_name_node *name_node; 325 326 hlist_for_each_entry_rcu(name_node, head, hlist) 327 if (!strcmp(name_node->name, name)) 328 return name_node; 329 return NULL; 330 } 331 332 bool netdev_name_in_use(struct net *net, const char *name) 333 { 334 return netdev_name_node_lookup(net, name); 335 } 336 EXPORT_SYMBOL(netdev_name_in_use); 337 338 int netdev_name_node_alt_create(struct net_device *dev, const char *name) 339 { 340 struct netdev_name_node *name_node; 341 struct net *net = dev_net(dev); 342 343 name_node = netdev_name_node_lookup(net, name); 344 if (name_node) 345 return -EEXIST; 346 name_node = netdev_name_node_alloc(dev, name); 347 if (!name_node) 348 return -ENOMEM; 349 netdev_name_node_add(net, name_node); 350 /* The node that holds dev->name acts as a head of per-device list. */ 351 list_add_tail_rcu(&name_node->list, &dev->name_node->list); 352 353 return 0; 354 } 355 356 static void netdev_name_node_alt_free(struct rcu_head *head) 357 { 358 struct netdev_name_node *name_node = 359 container_of(head, struct netdev_name_node, rcu); 360 361 kfree(name_node->name); 362 netdev_name_node_free(name_node); 363 } 364 365 static void __netdev_name_node_alt_destroy(struct netdev_name_node *name_node) 366 { 367 netdev_name_node_del(name_node); 368 list_del(&name_node->list); 369 call_rcu(&name_node->rcu, netdev_name_node_alt_free); 370 } 371 372 int netdev_name_node_alt_destroy(struct net_device *dev, const char *name) 373 { 374 struct netdev_name_node *name_node; 375 struct net *net = dev_net(dev); 376 377 name_node = netdev_name_node_lookup(net, name); 378 if (!name_node) 379 return -ENOENT; 380 /* lookup might have found our primary name or a name belonging 381 * to another device. 382 */ 383 if (name_node == dev->name_node || name_node->dev != dev) 384 return -EINVAL; 385 386 __netdev_name_node_alt_destroy(name_node); 387 return 0; 388 } 389 390 static void netdev_name_node_alt_flush(struct net_device *dev) 391 { 392 struct netdev_name_node *name_node, *tmp; 393 394 list_for_each_entry_safe(name_node, tmp, &dev->name_node->list, list) { 395 list_del(&name_node->list); 396 netdev_name_node_alt_free(&name_node->rcu); 397 } 398 } 399 400 /* Device list insertion */ 401 static void list_netdevice(struct net_device *dev) 402 { 403 struct netdev_name_node *name_node; 404 struct net *net = dev_net(dev); 405 406 ASSERT_RTNL(); 407 408 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head); 409 netdev_name_node_add(net, dev->name_node); 410 hlist_add_head_rcu(&dev->index_hlist, 411 dev_index_hash(net, dev->ifindex)); 412 413 netdev_for_each_altname(dev, name_node) 414 netdev_name_node_add(net, name_node); 415 416 /* We reserved the ifindex, this can't fail */ 417 WARN_ON(xa_store(&net->dev_by_index, dev->ifindex, dev, GFP_KERNEL)); 418 419 dev_base_seq_inc(net); 420 } 421 422 /* Device list removal 423 * caller must respect a RCU grace period before freeing/reusing dev 424 */ 425 static void unlist_netdevice(struct net_device *dev) 426 { 427 struct netdev_name_node *name_node; 428 struct net *net = dev_net(dev); 429 430 ASSERT_RTNL(); 431 432 xa_erase(&net->dev_by_index, dev->ifindex); 433 434 netdev_for_each_altname(dev, name_node) 435 netdev_name_node_del(name_node); 436 437 /* Unlink dev from the device chain */ 438 list_del_rcu(&dev->dev_list); 439 netdev_name_node_del(dev->name_node); 440 hlist_del_rcu(&dev->index_hlist); 441 442 dev_base_seq_inc(dev_net(dev)); 443 } 444 445 /* 446 * Our notifier list 447 */ 448 449 static RAW_NOTIFIER_HEAD(netdev_chain); 450 451 /* 452 * Device drivers call our routines to queue packets here. We empty the 453 * queue in the local softnet handler. 454 */ 455 456 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data) = { 457 .process_queue_bh_lock = INIT_LOCAL_LOCK(process_queue_bh_lock), 458 }; 459 EXPORT_PER_CPU_SYMBOL(softnet_data); 460 461 /* Page_pool has a lockless array/stack to alloc/recycle pages. 462 * PP consumers must pay attention to run APIs in the appropriate context 463 * (e.g. NAPI context). 464 */ 465 DEFINE_PER_CPU(struct page_pool *, system_page_pool); 466 467 #ifdef CONFIG_LOCKDEP 468 /* 469 * register_netdevice() inits txq->_xmit_lock and sets lockdep class 470 * according to dev->type 471 */ 472 static const unsigned short netdev_lock_type[] = { 473 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25, 474 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET, 475 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM, 476 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP, 477 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD, 478 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25, 479 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP, 480 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD, 481 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI, 482 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE, 483 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET, 484 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL, 485 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM, 486 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE, 487 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE}; 488 489 static const char *const netdev_lock_name[] = { 490 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25", 491 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET", 492 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM", 493 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP", 494 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD", 495 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25", 496 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP", 497 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD", 498 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI", 499 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE", 500 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET", 501 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL", 502 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM", 503 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE", 504 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"}; 505 506 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)]; 507 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)]; 508 509 static inline unsigned short netdev_lock_pos(unsigned short dev_type) 510 { 511 int i; 512 513 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++) 514 if (netdev_lock_type[i] == dev_type) 515 return i; 516 /* the last key is used by default */ 517 return ARRAY_SIZE(netdev_lock_type) - 1; 518 } 519 520 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 521 unsigned short dev_type) 522 { 523 int i; 524 525 i = netdev_lock_pos(dev_type); 526 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i], 527 netdev_lock_name[i]); 528 } 529 530 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 531 { 532 int i; 533 534 i = netdev_lock_pos(dev->type); 535 lockdep_set_class_and_name(&dev->addr_list_lock, 536 &netdev_addr_lock_key[i], 537 netdev_lock_name[i]); 538 } 539 #else 540 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 541 unsigned short dev_type) 542 { 543 } 544 545 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 546 { 547 } 548 #endif 549 550 /******************************************************************************* 551 * 552 * Protocol management and registration routines 553 * 554 *******************************************************************************/ 555 556 557 /* 558 * Add a protocol ID to the list. Now that the input handler is 559 * smarter we can dispense with all the messy stuff that used to be 560 * here. 561 * 562 * BEWARE!!! Protocol handlers, mangling input packets, 563 * MUST BE last in hash buckets and checking protocol handlers 564 * MUST start from promiscuous ptype_all chain in net_bh. 565 * It is true now, do not change it. 566 * Explanation follows: if protocol handler, mangling packet, will 567 * be the first on list, it is not able to sense, that packet 568 * is cloned and should be copied-on-write, so that it will 569 * change it and subsequent readers will get broken packet. 570 * --ANK (980803) 571 */ 572 573 static inline struct list_head *ptype_head(const struct packet_type *pt) 574 { 575 if (pt->type == htons(ETH_P_ALL)) 576 return pt->dev ? &pt->dev->ptype_all : &net_hotdata.ptype_all; 577 else 578 return pt->dev ? &pt->dev->ptype_specific : 579 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK]; 580 } 581 582 /** 583 * dev_add_pack - add packet handler 584 * @pt: packet type declaration 585 * 586 * Add a protocol handler to the networking stack. The passed &packet_type 587 * is linked into kernel lists and may not be freed until it has been 588 * removed from the kernel lists. 589 * 590 * This call does not sleep therefore it can not 591 * guarantee all CPU's that are in middle of receiving packets 592 * will see the new packet type (until the next received packet). 593 */ 594 595 void dev_add_pack(struct packet_type *pt) 596 { 597 struct list_head *head = ptype_head(pt); 598 599 spin_lock(&ptype_lock); 600 list_add_rcu(&pt->list, head); 601 spin_unlock(&ptype_lock); 602 } 603 EXPORT_SYMBOL(dev_add_pack); 604 605 /** 606 * __dev_remove_pack - remove packet handler 607 * @pt: packet type declaration 608 * 609 * Remove a protocol handler that was previously added to the kernel 610 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 611 * from the kernel lists and can be freed or reused once this function 612 * returns. 613 * 614 * The packet type might still be in use by receivers 615 * and must not be freed until after all the CPU's have gone 616 * through a quiescent state. 617 */ 618 void __dev_remove_pack(struct packet_type *pt) 619 { 620 struct list_head *head = ptype_head(pt); 621 struct packet_type *pt1; 622 623 spin_lock(&ptype_lock); 624 625 list_for_each_entry(pt1, head, list) { 626 if (pt == pt1) { 627 list_del_rcu(&pt->list); 628 goto out; 629 } 630 } 631 632 pr_warn("dev_remove_pack: %p not found\n", pt); 633 out: 634 spin_unlock(&ptype_lock); 635 } 636 EXPORT_SYMBOL(__dev_remove_pack); 637 638 /** 639 * dev_remove_pack - remove packet handler 640 * @pt: packet type declaration 641 * 642 * Remove a protocol handler that was previously added to the kernel 643 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 644 * from the kernel lists and can be freed or reused once this function 645 * returns. 646 * 647 * This call sleeps to guarantee that no CPU is looking at the packet 648 * type after return. 649 */ 650 void dev_remove_pack(struct packet_type *pt) 651 { 652 __dev_remove_pack(pt); 653 654 synchronize_net(); 655 } 656 EXPORT_SYMBOL(dev_remove_pack); 657 658 659 /******************************************************************************* 660 * 661 * Device Interface Subroutines 662 * 663 *******************************************************************************/ 664 665 /** 666 * dev_get_iflink - get 'iflink' value of a interface 667 * @dev: targeted interface 668 * 669 * Indicates the ifindex the interface is linked to. 670 * Physical interfaces have the same 'ifindex' and 'iflink' values. 671 */ 672 673 int dev_get_iflink(const struct net_device *dev) 674 { 675 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink) 676 return dev->netdev_ops->ndo_get_iflink(dev); 677 678 return READ_ONCE(dev->ifindex); 679 } 680 EXPORT_SYMBOL(dev_get_iflink); 681 682 /** 683 * dev_fill_metadata_dst - Retrieve tunnel egress information. 684 * @dev: targeted interface 685 * @skb: The packet. 686 * 687 * For better visibility of tunnel traffic OVS needs to retrieve 688 * egress tunnel information for a packet. Following API allows 689 * user to get this info. 690 */ 691 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb) 692 { 693 struct ip_tunnel_info *info; 694 695 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst) 696 return -EINVAL; 697 698 info = skb_tunnel_info_unclone(skb); 699 if (!info) 700 return -ENOMEM; 701 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX))) 702 return -EINVAL; 703 704 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb); 705 } 706 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst); 707 708 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack) 709 { 710 int k = stack->num_paths++; 711 712 if (WARN_ON_ONCE(k >= NET_DEVICE_PATH_STACK_MAX)) 713 return NULL; 714 715 return &stack->path[k]; 716 } 717 718 int dev_fill_forward_path(const struct net_device *dev, const u8 *daddr, 719 struct net_device_path_stack *stack) 720 { 721 const struct net_device *last_dev; 722 struct net_device_path_ctx ctx = { 723 .dev = dev, 724 }; 725 struct net_device_path *path; 726 int ret = 0; 727 728 memcpy(ctx.daddr, daddr, sizeof(ctx.daddr)); 729 stack->num_paths = 0; 730 while (ctx.dev && ctx.dev->netdev_ops->ndo_fill_forward_path) { 731 last_dev = ctx.dev; 732 path = dev_fwd_path(stack); 733 if (!path) 734 return -1; 735 736 memset(path, 0, sizeof(struct net_device_path)); 737 ret = ctx.dev->netdev_ops->ndo_fill_forward_path(&ctx, path); 738 if (ret < 0) 739 return -1; 740 741 if (WARN_ON_ONCE(last_dev == ctx.dev)) 742 return -1; 743 } 744 745 if (!ctx.dev) 746 return ret; 747 748 path = dev_fwd_path(stack); 749 if (!path) 750 return -1; 751 path->type = DEV_PATH_ETHERNET; 752 path->dev = ctx.dev; 753 754 return ret; 755 } 756 EXPORT_SYMBOL_GPL(dev_fill_forward_path); 757 758 /* must be called under rcu_read_lock(), as we dont take a reference */ 759 static struct napi_struct *napi_by_id(unsigned int napi_id) 760 { 761 unsigned int hash = napi_id % HASH_SIZE(napi_hash); 762 struct napi_struct *napi; 763 764 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node) 765 if (napi->napi_id == napi_id) 766 return napi; 767 768 return NULL; 769 } 770 771 /* must be called under rcu_read_lock(), as we dont take a reference */ 772 static struct napi_struct * 773 netdev_napi_by_id(struct net *net, unsigned int napi_id) 774 { 775 struct napi_struct *napi; 776 777 napi = napi_by_id(napi_id); 778 if (!napi) 779 return NULL; 780 781 if (WARN_ON_ONCE(!napi->dev)) 782 return NULL; 783 if (!net_eq(net, dev_net(napi->dev))) 784 return NULL; 785 786 return napi; 787 } 788 789 /** 790 * netdev_napi_by_id_lock() - find a device by NAPI ID and lock it 791 * @net: the applicable net namespace 792 * @napi_id: ID of a NAPI of a target device 793 * 794 * Find a NAPI instance with @napi_id. Lock its device. 795 * The device must be in %NETREG_REGISTERED state for lookup to succeed. 796 * netdev_unlock() must be called to release it. 797 * 798 * Return: pointer to NAPI, its device with lock held, NULL if not found. 799 */ 800 struct napi_struct * 801 netdev_napi_by_id_lock(struct net *net, unsigned int napi_id) 802 { 803 struct napi_struct *napi; 804 struct net_device *dev; 805 806 rcu_read_lock(); 807 napi = netdev_napi_by_id(net, napi_id); 808 if (!napi || READ_ONCE(napi->dev->reg_state) != NETREG_REGISTERED) { 809 rcu_read_unlock(); 810 return NULL; 811 } 812 813 dev = napi->dev; 814 dev_hold(dev); 815 rcu_read_unlock(); 816 817 dev = __netdev_put_lock(dev); 818 if (!dev) 819 return NULL; 820 821 rcu_read_lock(); 822 napi = netdev_napi_by_id(net, napi_id); 823 if (napi && napi->dev != dev) 824 napi = NULL; 825 rcu_read_unlock(); 826 827 if (!napi) 828 netdev_unlock(dev); 829 return napi; 830 } 831 832 /** 833 * __dev_get_by_name - find a device by its name 834 * @net: the applicable net namespace 835 * @name: name to find 836 * 837 * Find an interface by name. Must be called under RTNL semaphore. 838 * If the name is found a pointer to the device is returned. 839 * If the name is not found then %NULL is returned. The 840 * reference counters are not incremented so the caller must be 841 * careful with locks. 842 */ 843 844 struct net_device *__dev_get_by_name(struct net *net, const char *name) 845 { 846 struct netdev_name_node *node_name; 847 848 node_name = netdev_name_node_lookup(net, name); 849 return node_name ? node_name->dev : NULL; 850 } 851 EXPORT_SYMBOL(__dev_get_by_name); 852 853 /** 854 * dev_get_by_name_rcu - find a device by its name 855 * @net: the applicable net namespace 856 * @name: name to find 857 * 858 * Find an interface by name. 859 * If the name is found a pointer to the device is returned. 860 * If the name is not found then %NULL is returned. 861 * The reference counters are not incremented so the caller must be 862 * careful with locks. The caller must hold RCU lock. 863 */ 864 865 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name) 866 { 867 struct netdev_name_node *node_name; 868 869 node_name = netdev_name_node_lookup_rcu(net, name); 870 return node_name ? node_name->dev : NULL; 871 } 872 EXPORT_SYMBOL(dev_get_by_name_rcu); 873 874 /* Deprecated for new users, call netdev_get_by_name() instead */ 875 struct net_device *dev_get_by_name(struct net *net, const char *name) 876 { 877 struct net_device *dev; 878 879 rcu_read_lock(); 880 dev = dev_get_by_name_rcu(net, name); 881 dev_hold(dev); 882 rcu_read_unlock(); 883 return dev; 884 } 885 EXPORT_SYMBOL(dev_get_by_name); 886 887 /** 888 * netdev_get_by_name() - find a device by its name 889 * @net: the applicable net namespace 890 * @name: name to find 891 * @tracker: tracking object for the acquired reference 892 * @gfp: allocation flags for the tracker 893 * 894 * Find an interface by name. This can be called from any 895 * context and does its own locking. The returned handle has 896 * the usage count incremented and the caller must use netdev_put() to 897 * release it when it is no longer needed. %NULL is returned if no 898 * matching device is found. 899 */ 900 struct net_device *netdev_get_by_name(struct net *net, const char *name, 901 netdevice_tracker *tracker, gfp_t gfp) 902 { 903 struct net_device *dev; 904 905 dev = dev_get_by_name(net, name); 906 if (dev) 907 netdev_tracker_alloc(dev, tracker, gfp); 908 return dev; 909 } 910 EXPORT_SYMBOL(netdev_get_by_name); 911 912 /** 913 * __dev_get_by_index - find a device by its ifindex 914 * @net: the applicable net namespace 915 * @ifindex: index of device 916 * 917 * Search for an interface by index. Returns %NULL if the device 918 * is not found or a pointer to the device. The device has not 919 * had its reference counter increased so the caller must be careful 920 * about locking. The caller must hold the RTNL semaphore. 921 */ 922 923 struct net_device *__dev_get_by_index(struct net *net, int ifindex) 924 { 925 struct net_device *dev; 926 struct hlist_head *head = dev_index_hash(net, ifindex); 927 928 hlist_for_each_entry(dev, head, index_hlist) 929 if (dev->ifindex == ifindex) 930 return dev; 931 932 return NULL; 933 } 934 EXPORT_SYMBOL(__dev_get_by_index); 935 936 /** 937 * dev_get_by_index_rcu - find a device by its ifindex 938 * @net: the applicable net namespace 939 * @ifindex: index of device 940 * 941 * Search for an interface by index. Returns %NULL if the device 942 * is not found or a pointer to the device. The device has not 943 * had its reference counter increased so the caller must be careful 944 * about locking. The caller must hold RCU lock. 945 */ 946 947 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex) 948 { 949 struct net_device *dev; 950 struct hlist_head *head = dev_index_hash(net, ifindex); 951 952 hlist_for_each_entry_rcu(dev, head, index_hlist) 953 if (dev->ifindex == ifindex) 954 return dev; 955 956 return NULL; 957 } 958 EXPORT_SYMBOL(dev_get_by_index_rcu); 959 960 /* Deprecated for new users, call netdev_get_by_index() instead */ 961 struct net_device *dev_get_by_index(struct net *net, int ifindex) 962 { 963 struct net_device *dev; 964 965 rcu_read_lock(); 966 dev = dev_get_by_index_rcu(net, ifindex); 967 dev_hold(dev); 968 rcu_read_unlock(); 969 return dev; 970 } 971 EXPORT_SYMBOL(dev_get_by_index); 972 973 /** 974 * netdev_get_by_index() - find a device by its ifindex 975 * @net: the applicable net namespace 976 * @ifindex: index of device 977 * @tracker: tracking object for the acquired reference 978 * @gfp: allocation flags for the tracker 979 * 980 * Search for an interface by index. Returns NULL if the device 981 * is not found or a pointer to the device. The device returned has 982 * had a reference added and the pointer is safe until the user calls 983 * netdev_put() to indicate they have finished with it. 984 */ 985 struct net_device *netdev_get_by_index(struct net *net, int ifindex, 986 netdevice_tracker *tracker, gfp_t gfp) 987 { 988 struct net_device *dev; 989 990 dev = dev_get_by_index(net, ifindex); 991 if (dev) 992 netdev_tracker_alloc(dev, tracker, gfp); 993 return dev; 994 } 995 EXPORT_SYMBOL(netdev_get_by_index); 996 997 /** 998 * dev_get_by_napi_id - find a device by napi_id 999 * @napi_id: ID of the NAPI struct 1000 * 1001 * Search for an interface by NAPI ID. Returns %NULL if the device 1002 * is not found or a pointer to the device. The device has not had 1003 * its reference counter increased so the caller must be careful 1004 * about locking. The caller must hold RCU lock. 1005 */ 1006 struct net_device *dev_get_by_napi_id(unsigned int napi_id) 1007 { 1008 struct napi_struct *napi; 1009 1010 WARN_ON_ONCE(!rcu_read_lock_held()); 1011 1012 if (!napi_id_valid(napi_id)) 1013 return NULL; 1014 1015 napi = napi_by_id(napi_id); 1016 1017 return napi ? napi->dev : NULL; 1018 } 1019 1020 /* Release the held reference on the net_device, and if the net_device 1021 * is still registered try to lock the instance lock. If device is being 1022 * unregistered NULL will be returned (but the reference has been released, 1023 * either way!) 1024 * 1025 * This helper is intended for locking net_device after it has been looked up 1026 * using a lockless lookup helper. Lock prevents the instance from going away. 1027 */ 1028 struct net_device *__netdev_put_lock(struct net_device *dev) 1029 { 1030 netdev_lock(dev); 1031 if (dev->reg_state > NETREG_REGISTERED) { 1032 netdev_unlock(dev); 1033 dev_put(dev); 1034 return NULL; 1035 } 1036 dev_put(dev); 1037 return dev; 1038 } 1039 1040 /** 1041 * netdev_get_by_index_lock() - find a device by its ifindex 1042 * @net: the applicable net namespace 1043 * @ifindex: index of device 1044 * 1045 * Search for an interface by index. If a valid device 1046 * with @ifindex is found it will be returned with netdev->lock held. 1047 * netdev_unlock() must be called to release it. 1048 * 1049 * Return: pointer to a device with lock held, NULL if not found. 1050 */ 1051 struct net_device *netdev_get_by_index_lock(struct net *net, int ifindex) 1052 { 1053 struct net_device *dev; 1054 1055 dev = dev_get_by_index(net, ifindex); 1056 if (!dev) 1057 return NULL; 1058 1059 return __netdev_put_lock(dev); 1060 } 1061 1062 struct net_device * 1063 netdev_xa_find_lock(struct net *net, struct net_device *dev, 1064 unsigned long *index) 1065 { 1066 if (dev) 1067 netdev_unlock(dev); 1068 1069 do { 1070 rcu_read_lock(); 1071 dev = xa_find(&net->dev_by_index, index, ULONG_MAX, XA_PRESENT); 1072 if (!dev) { 1073 rcu_read_unlock(); 1074 return NULL; 1075 } 1076 dev_hold(dev); 1077 rcu_read_unlock(); 1078 1079 dev = __netdev_put_lock(dev); 1080 if (dev) 1081 return dev; 1082 1083 (*index)++; 1084 } while (true); 1085 } 1086 1087 static DEFINE_SEQLOCK(netdev_rename_lock); 1088 1089 void netdev_copy_name(struct net_device *dev, char *name) 1090 { 1091 unsigned int seq; 1092 1093 do { 1094 seq = read_seqbegin(&netdev_rename_lock); 1095 strscpy(name, dev->name, IFNAMSIZ); 1096 } while (read_seqretry(&netdev_rename_lock, seq)); 1097 } 1098 1099 /** 1100 * netdev_get_name - get a netdevice name, knowing its ifindex. 1101 * @net: network namespace 1102 * @name: a pointer to the buffer where the name will be stored. 1103 * @ifindex: the ifindex of the interface to get the name from. 1104 */ 1105 int netdev_get_name(struct net *net, char *name, int ifindex) 1106 { 1107 struct net_device *dev; 1108 int ret; 1109 1110 rcu_read_lock(); 1111 1112 dev = dev_get_by_index_rcu(net, ifindex); 1113 if (!dev) { 1114 ret = -ENODEV; 1115 goto out; 1116 } 1117 1118 netdev_copy_name(dev, name); 1119 1120 ret = 0; 1121 out: 1122 rcu_read_unlock(); 1123 return ret; 1124 } 1125 1126 static bool dev_addr_cmp(struct net_device *dev, unsigned short type, 1127 const char *ha) 1128 { 1129 return dev->type == type && !memcmp(dev->dev_addr, ha, dev->addr_len); 1130 } 1131 1132 /** 1133 * dev_getbyhwaddr_rcu - find a device by its hardware address 1134 * @net: the applicable net namespace 1135 * @type: media type of device 1136 * @ha: hardware address 1137 * 1138 * Search for an interface by MAC address. Returns NULL if the device 1139 * is not found or a pointer to the device. 1140 * The caller must hold RCU. 1141 * The returned device has not had its ref count increased 1142 * and the caller must therefore be careful about locking 1143 * 1144 */ 1145 1146 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type, 1147 const char *ha) 1148 { 1149 struct net_device *dev; 1150 1151 for_each_netdev_rcu(net, dev) 1152 if (dev_addr_cmp(dev, type, ha)) 1153 return dev; 1154 1155 return NULL; 1156 } 1157 EXPORT_SYMBOL(dev_getbyhwaddr_rcu); 1158 1159 /** 1160 * dev_getbyhwaddr() - find a device by its hardware address 1161 * @net: the applicable net namespace 1162 * @type: media type of device 1163 * @ha: hardware address 1164 * 1165 * Similar to dev_getbyhwaddr_rcu(), but the owner needs to hold 1166 * rtnl_lock. 1167 * 1168 * Context: rtnl_lock() must be held. 1169 * Return: pointer to the net_device, or NULL if not found 1170 */ 1171 struct net_device *dev_getbyhwaddr(struct net *net, unsigned short type, 1172 const char *ha) 1173 { 1174 struct net_device *dev; 1175 1176 ASSERT_RTNL(); 1177 for_each_netdev(net, dev) 1178 if (dev_addr_cmp(dev, type, ha)) 1179 return dev; 1180 1181 return NULL; 1182 } 1183 EXPORT_SYMBOL(dev_getbyhwaddr); 1184 1185 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type) 1186 { 1187 struct net_device *dev, *ret = NULL; 1188 1189 rcu_read_lock(); 1190 for_each_netdev_rcu(net, dev) 1191 if (dev->type == type) { 1192 dev_hold(dev); 1193 ret = dev; 1194 break; 1195 } 1196 rcu_read_unlock(); 1197 return ret; 1198 } 1199 EXPORT_SYMBOL(dev_getfirstbyhwtype); 1200 1201 /** 1202 * __dev_get_by_flags - find any device with given flags 1203 * @net: the applicable net namespace 1204 * @if_flags: IFF_* values 1205 * @mask: bitmask of bits in if_flags to check 1206 * 1207 * Search for any interface with the given flags. Returns NULL if a device 1208 * is not found or a pointer to the device. Must be called inside 1209 * rtnl_lock(), and result refcount is unchanged. 1210 */ 1211 1212 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags, 1213 unsigned short mask) 1214 { 1215 struct net_device *dev, *ret; 1216 1217 ASSERT_RTNL(); 1218 1219 ret = NULL; 1220 for_each_netdev(net, dev) { 1221 if (((dev->flags ^ if_flags) & mask) == 0) { 1222 ret = dev; 1223 break; 1224 } 1225 } 1226 return ret; 1227 } 1228 EXPORT_SYMBOL(__dev_get_by_flags); 1229 1230 /** 1231 * dev_valid_name - check if name is okay for network device 1232 * @name: name string 1233 * 1234 * Network device names need to be valid file names to 1235 * allow sysfs to work. We also disallow any kind of 1236 * whitespace. 1237 */ 1238 bool dev_valid_name(const char *name) 1239 { 1240 if (*name == '\0') 1241 return false; 1242 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ) 1243 return false; 1244 if (!strcmp(name, ".") || !strcmp(name, "..")) 1245 return false; 1246 1247 while (*name) { 1248 if (*name == '/' || *name == ':' || isspace(*name)) 1249 return false; 1250 name++; 1251 } 1252 return true; 1253 } 1254 EXPORT_SYMBOL(dev_valid_name); 1255 1256 /** 1257 * __dev_alloc_name - allocate a name for a device 1258 * @net: network namespace to allocate the device name in 1259 * @name: name format string 1260 * @res: result name string 1261 * 1262 * Passed a format string - eg "lt%d" it will try and find a suitable 1263 * id. It scans list of devices to build up a free map, then chooses 1264 * the first empty slot. The caller must hold the dev_base or rtnl lock 1265 * while allocating the name and adding the device in order to avoid 1266 * duplicates. 1267 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1268 * Returns the number of the unit assigned or a negative errno code. 1269 */ 1270 1271 static int __dev_alloc_name(struct net *net, const char *name, char *res) 1272 { 1273 int i = 0; 1274 const char *p; 1275 const int max_netdevices = 8*PAGE_SIZE; 1276 unsigned long *inuse; 1277 struct net_device *d; 1278 char buf[IFNAMSIZ]; 1279 1280 /* Verify the string as this thing may have come from the user. 1281 * There must be one "%d" and no other "%" characters. 1282 */ 1283 p = strchr(name, '%'); 1284 if (!p || p[1] != 'd' || strchr(p + 2, '%')) 1285 return -EINVAL; 1286 1287 /* Use one page as a bit array of possible slots */ 1288 inuse = bitmap_zalloc(max_netdevices, GFP_ATOMIC); 1289 if (!inuse) 1290 return -ENOMEM; 1291 1292 for_each_netdev(net, d) { 1293 struct netdev_name_node *name_node; 1294 1295 netdev_for_each_altname(d, name_node) { 1296 if (!sscanf(name_node->name, name, &i)) 1297 continue; 1298 if (i < 0 || i >= max_netdevices) 1299 continue; 1300 1301 /* avoid cases where sscanf is not exact inverse of printf */ 1302 snprintf(buf, IFNAMSIZ, name, i); 1303 if (!strncmp(buf, name_node->name, IFNAMSIZ)) 1304 __set_bit(i, inuse); 1305 } 1306 if (!sscanf(d->name, name, &i)) 1307 continue; 1308 if (i < 0 || i >= max_netdevices) 1309 continue; 1310 1311 /* avoid cases where sscanf is not exact inverse of printf */ 1312 snprintf(buf, IFNAMSIZ, name, i); 1313 if (!strncmp(buf, d->name, IFNAMSIZ)) 1314 __set_bit(i, inuse); 1315 } 1316 1317 i = find_first_zero_bit(inuse, max_netdevices); 1318 bitmap_free(inuse); 1319 if (i == max_netdevices) 1320 return -ENFILE; 1321 1322 /* 'res' and 'name' could overlap, use 'buf' as an intermediate buffer */ 1323 strscpy(buf, name, IFNAMSIZ); 1324 snprintf(res, IFNAMSIZ, buf, i); 1325 return i; 1326 } 1327 1328 /* Returns negative errno or allocated unit id (see __dev_alloc_name()) */ 1329 static int dev_prep_valid_name(struct net *net, struct net_device *dev, 1330 const char *want_name, char *out_name, 1331 int dup_errno) 1332 { 1333 if (!dev_valid_name(want_name)) 1334 return -EINVAL; 1335 1336 if (strchr(want_name, '%')) 1337 return __dev_alloc_name(net, want_name, out_name); 1338 1339 if (netdev_name_in_use(net, want_name)) 1340 return -dup_errno; 1341 if (out_name != want_name) 1342 strscpy(out_name, want_name, IFNAMSIZ); 1343 return 0; 1344 } 1345 1346 /** 1347 * dev_alloc_name - allocate a name for a device 1348 * @dev: device 1349 * @name: name format string 1350 * 1351 * Passed a format string - eg "lt%d" it will try and find a suitable 1352 * id. It scans list of devices to build up a free map, then chooses 1353 * the first empty slot. The caller must hold the dev_base or rtnl lock 1354 * while allocating the name and adding the device in order to avoid 1355 * duplicates. 1356 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1357 * Returns the number of the unit assigned or a negative errno code. 1358 */ 1359 1360 int dev_alloc_name(struct net_device *dev, const char *name) 1361 { 1362 return dev_prep_valid_name(dev_net(dev), dev, name, dev->name, ENFILE); 1363 } 1364 EXPORT_SYMBOL(dev_alloc_name); 1365 1366 static int dev_get_valid_name(struct net *net, struct net_device *dev, 1367 const char *name) 1368 { 1369 int ret; 1370 1371 ret = dev_prep_valid_name(net, dev, name, dev->name, EEXIST); 1372 return ret < 0 ? ret : 0; 1373 } 1374 1375 int netif_change_name(struct net_device *dev, const char *newname) 1376 { 1377 struct net *net = dev_net(dev); 1378 unsigned char old_assign_type; 1379 char oldname[IFNAMSIZ]; 1380 int err = 0; 1381 int ret; 1382 1383 ASSERT_RTNL_NET(net); 1384 1385 if (!strncmp(newname, dev->name, IFNAMSIZ)) 1386 return 0; 1387 1388 memcpy(oldname, dev->name, IFNAMSIZ); 1389 1390 write_seqlock_bh(&netdev_rename_lock); 1391 err = dev_get_valid_name(net, dev, newname); 1392 write_sequnlock_bh(&netdev_rename_lock); 1393 1394 if (err < 0) 1395 return err; 1396 1397 if (oldname[0] && !strchr(oldname, '%')) 1398 netdev_info(dev, "renamed from %s%s\n", oldname, 1399 dev->flags & IFF_UP ? " (while UP)" : ""); 1400 1401 old_assign_type = dev->name_assign_type; 1402 WRITE_ONCE(dev->name_assign_type, NET_NAME_RENAMED); 1403 1404 rollback: 1405 ret = device_rename(&dev->dev, dev->name); 1406 if (ret) { 1407 write_seqlock_bh(&netdev_rename_lock); 1408 memcpy(dev->name, oldname, IFNAMSIZ); 1409 write_sequnlock_bh(&netdev_rename_lock); 1410 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1411 return ret; 1412 } 1413 1414 netdev_adjacent_rename_links(dev, oldname); 1415 1416 netdev_name_node_del(dev->name_node); 1417 1418 synchronize_net(); 1419 1420 netdev_name_node_add(net, dev->name_node); 1421 1422 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev); 1423 ret = notifier_to_errno(ret); 1424 1425 if (ret) { 1426 /* err >= 0 after dev_alloc_name() or stores the first errno */ 1427 if (err >= 0) { 1428 err = ret; 1429 write_seqlock_bh(&netdev_rename_lock); 1430 memcpy(dev->name, oldname, IFNAMSIZ); 1431 write_sequnlock_bh(&netdev_rename_lock); 1432 memcpy(oldname, newname, IFNAMSIZ); 1433 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1434 old_assign_type = NET_NAME_RENAMED; 1435 goto rollback; 1436 } else { 1437 netdev_err(dev, "name change rollback failed: %d\n", 1438 ret); 1439 } 1440 } 1441 1442 return err; 1443 } 1444 1445 int netif_set_alias(struct net_device *dev, const char *alias, size_t len) 1446 { 1447 struct dev_ifalias *new_alias = NULL; 1448 1449 if (len >= IFALIASZ) 1450 return -EINVAL; 1451 1452 if (len) { 1453 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL); 1454 if (!new_alias) 1455 return -ENOMEM; 1456 1457 memcpy(new_alias->ifalias, alias, len); 1458 new_alias->ifalias[len] = 0; 1459 } 1460 1461 mutex_lock(&ifalias_mutex); 1462 new_alias = rcu_replace_pointer(dev->ifalias, new_alias, 1463 mutex_is_locked(&ifalias_mutex)); 1464 mutex_unlock(&ifalias_mutex); 1465 1466 if (new_alias) 1467 kfree_rcu(new_alias, rcuhead); 1468 1469 return len; 1470 } 1471 1472 /** 1473 * dev_get_alias - get ifalias of a device 1474 * @dev: device 1475 * @name: buffer to store name of ifalias 1476 * @len: size of buffer 1477 * 1478 * get ifalias for a device. Caller must make sure dev cannot go 1479 * away, e.g. rcu read lock or own a reference count to device. 1480 */ 1481 int dev_get_alias(const struct net_device *dev, char *name, size_t len) 1482 { 1483 const struct dev_ifalias *alias; 1484 int ret = 0; 1485 1486 rcu_read_lock(); 1487 alias = rcu_dereference(dev->ifalias); 1488 if (alias) 1489 ret = snprintf(name, len, "%s", alias->ifalias); 1490 rcu_read_unlock(); 1491 1492 return ret; 1493 } 1494 1495 /** 1496 * netdev_features_change - device changes features 1497 * @dev: device to cause notification 1498 * 1499 * Called to indicate a device has changed features. 1500 */ 1501 void netdev_features_change(struct net_device *dev) 1502 { 1503 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev); 1504 } 1505 EXPORT_SYMBOL(netdev_features_change); 1506 1507 /** 1508 * netdev_state_change - device changes state 1509 * @dev: device to cause notification 1510 * 1511 * Called to indicate a device has changed state. This function calls 1512 * the notifier chains for netdev_chain and sends a NEWLINK message 1513 * to the routing socket. 1514 */ 1515 void netdev_state_change(struct net_device *dev) 1516 { 1517 if (dev->flags & IFF_UP) { 1518 struct netdev_notifier_change_info change_info = { 1519 .info.dev = dev, 1520 }; 1521 1522 call_netdevice_notifiers_info(NETDEV_CHANGE, 1523 &change_info.info); 1524 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL, 0, NULL); 1525 } 1526 } 1527 EXPORT_SYMBOL(netdev_state_change); 1528 1529 /** 1530 * __netdev_notify_peers - notify network peers about existence of @dev, 1531 * to be called when rtnl lock is already held. 1532 * @dev: network device 1533 * 1534 * Generate traffic such that interested network peers are aware of 1535 * @dev, such as by generating a gratuitous ARP. This may be used when 1536 * a device wants to inform the rest of the network about some sort of 1537 * reconfiguration such as a failover event or virtual machine 1538 * migration. 1539 */ 1540 void __netdev_notify_peers(struct net_device *dev) 1541 { 1542 ASSERT_RTNL(); 1543 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev); 1544 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev); 1545 } 1546 EXPORT_SYMBOL(__netdev_notify_peers); 1547 1548 /** 1549 * netdev_notify_peers - notify network peers about existence of @dev 1550 * @dev: network device 1551 * 1552 * Generate traffic such that interested network peers are aware of 1553 * @dev, such as by generating a gratuitous ARP. This may be used when 1554 * a device wants to inform the rest of the network about some sort of 1555 * reconfiguration such as a failover event or virtual machine 1556 * migration. 1557 */ 1558 void netdev_notify_peers(struct net_device *dev) 1559 { 1560 rtnl_lock(); 1561 __netdev_notify_peers(dev); 1562 rtnl_unlock(); 1563 } 1564 EXPORT_SYMBOL(netdev_notify_peers); 1565 1566 static int napi_threaded_poll(void *data); 1567 1568 static int napi_kthread_create(struct napi_struct *n) 1569 { 1570 int err = 0; 1571 1572 /* Create and wake up the kthread once to put it in 1573 * TASK_INTERRUPTIBLE mode to avoid the blocked task 1574 * warning and work with loadavg. 1575 */ 1576 n->thread = kthread_run(napi_threaded_poll, n, "napi/%s-%d", 1577 n->dev->name, n->napi_id); 1578 if (IS_ERR(n->thread)) { 1579 err = PTR_ERR(n->thread); 1580 pr_err("kthread_run failed with err %d\n", err); 1581 n->thread = NULL; 1582 } 1583 1584 return err; 1585 } 1586 1587 static int __dev_open(struct net_device *dev, struct netlink_ext_ack *extack) 1588 { 1589 const struct net_device_ops *ops = dev->netdev_ops; 1590 int ret; 1591 1592 ASSERT_RTNL(); 1593 dev_addr_check(dev); 1594 1595 if (!netif_device_present(dev)) { 1596 /* may be detached because parent is runtime-suspended */ 1597 if (dev->dev.parent) 1598 pm_runtime_resume(dev->dev.parent); 1599 if (!netif_device_present(dev)) 1600 return -ENODEV; 1601 } 1602 1603 /* Block netpoll from trying to do any rx path servicing. 1604 * If we don't do this there is a chance ndo_poll_controller 1605 * or ndo_poll may be running while we open the device 1606 */ 1607 netpoll_poll_disable(dev); 1608 1609 ret = call_netdevice_notifiers_extack(NETDEV_PRE_UP, dev, extack); 1610 ret = notifier_to_errno(ret); 1611 if (ret) 1612 return ret; 1613 1614 set_bit(__LINK_STATE_START, &dev->state); 1615 1616 netdev_ops_assert_locked(dev); 1617 1618 if (ops->ndo_validate_addr) 1619 ret = ops->ndo_validate_addr(dev); 1620 1621 if (!ret && ops->ndo_open) 1622 ret = ops->ndo_open(dev); 1623 1624 netpoll_poll_enable(dev); 1625 1626 if (ret) 1627 clear_bit(__LINK_STATE_START, &dev->state); 1628 else { 1629 netif_set_up(dev, true); 1630 dev_set_rx_mode(dev); 1631 dev_activate(dev); 1632 add_device_randomness(dev->dev_addr, dev->addr_len); 1633 } 1634 1635 return ret; 1636 } 1637 1638 int netif_open(struct net_device *dev, struct netlink_ext_ack *extack) 1639 { 1640 int ret; 1641 1642 if (dev->flags & IFF_UP) 1643 return 0; 1644 1645 ret = __dev_open(dev, extack); 1646 if (ret < 0) 1647 return ret; 1648 1649 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1650 call_netdevice_notifiers(NETDEV_UP, dev); 1651 1652 return ret; 1653 } 1654 1655 static void __dev_close_many(struct list_head *head) 1656 { 1657 struct net_device *dev; 1658 1659 ASSERT_RTNL(); 1660 might_sleep(); 1661 1662 list_for_each_entry(dev, head, close_list) { 1663 /* Temporarily disable netpoll until the interface is down */ 1664 netpoll_poll_disable(dev); 1665 1666 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev); 1667 1668 clear_bit(__LINK_STATE_START, &dev->state); 1669 1670 /* Synchronize to scheduled poll. We cannot touch poll list, it 1671 * can be even on different cpu. So just clear netif_running(). 1672 * 1673 * dev->stop() will invoke napi_disable() on all of it's 1674 * napi_struct instances on this device. 1675 */ 1676 smp_mb__after_atomic(); /* Commit netif_running(). */ 1677 } 1678 1679 dev_deactivate_many(head); 1680 1681 list_for_each_entry(dev, head, close_list) { 1682 const struct net_device_ops *ops = dev->netdev_ops; 1683 1684 /* 1685 * Call the device specific close. This cannot fail. 1686 * Only if device is UP 1687 * 1688 * We allow it to be called even after a DETACH hot-plug 1689 * event. 1690 */ 1691 1692 netdev_ops_assert_locked(dev); 1693 1694 if (ops->ndo_stop) 1695 ops->ndo_stop(dev); 1696 1697 netif_set_up(dev, false); 1698 netpoll_poll_enable(dev); 1699 } 1700 } 1701 1702 static void __dev_close(struct net_device *dev) 1703 { 1704 LIST_HEAD(single); 1705 1706 list_add(&dev->close_list, &single); 1707 __dev_close_many(&single); 1708 list_del(&single); 1709 } 1710 1711 void dev_close_many(struct list_head *head, bool unlink) 1712 { 1713 struct net_device *dev, *tmp; 1714 1715 /* Remove the devices that don't need to be closed */ 1716 list_for_each_entry_safe(dev, tmp, head, close_list) 1717 if (!(dev->flags & IFF_UP)) 1718 list_del_init(&dev->close_list); 1719 1720 __dev_close_many(head); 1721 1722 list_for_each_entry_safe(dev, tmp, head, close_list) { 1723 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1724 call_netdevice_notifiers(NETDEV_DOWN, dev); 1725 if (unlink) 1726 list_del_init(&dev->close_list); 1727 } 1728 } 1729 EXPORT_SYMBOL(dev_close_many); 1730 1731 void netif_close(struct net_device *dev) 1732 { 1733 if (dev->flags & IFF_UP) { 1734 LIST_HEAD(single); 1735 1736 list_add(&dev->close_list, &single); 1737 dev_close_many(&single, true); 1738 list_del(&single); 1739 } 1740 } 1741 EXPORT_SYMBOL(netif_close); 1742 1743 void netif_disable_lro(struct net_device *dev) 1744 { 1745 struct net_device *lower_dev; 1746 struct list_head *iter; 1747 1748 dev->wanted_features &= ~NETIF_F_LRO; 1749 netdev_update_features(dev); 1750 1751 if (unlikely(dev->features & NETIF_F_LRO)) 1752 netdev_WARN(dev, "failed to disable LRO!\n"); 1753 1754 netdev_for_each_lower_dev(dev, lower_dev, iter) { 1755 netdev_lock_ops(lower_dev); 1756 netif_disable_lro(lower_dev); 1757 netdev_unlock_ops(lower_dev); 1758 } 1759 } 1760 1761 /** 1762 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device 1763 * @dev: device 1764 * 1765 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be 1766 * called under RTNL. This is needed if Generic XDP is installed on 1767 * the device. 1768 */ 1769 static void dev_disable_gro_hw(struct net_device *dev) 1770 { 1771 dev->wanted_features &= ~NETIF_F_GRO_HW; 1772 netdev_update_features(dev); 1773 1774 if (unlikely(dev->features & NETIF_F_GRO_HW)) 1775 netdev_WARN(dev, "failed to disable GRO_HW!\n"); 1776 } 1777 1778 const char *netdev_cmd_to_name(enum netdev_cmd cmd) 1779 { 1780 #define N(val) \ 1781 case NETDEV_##val: \ 1782 return "NETDEV_" __stringify(val); 1783 switch (cmd) { 1784 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER) 1785 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE) 1786 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE) 1787 N(POST_INIT) N(PRE_UNINIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) 1788 N(CHANGEUPPER) N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) 1789 N(BONDING_INFO) N(PRECHANGEUPPER) N(CHANGELOWERSTATE) 1790 N(UDP_TUNNEL_PUSH_INFO) N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN) 1791 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO) 1792 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO) 1793 N(PRE_CHANGEADDR) N(OFFLOAD_XSTATS_ENABLE) N(OFFLOAD_XSTATS_DISABLE) 1794 N(OFFLOAD_XSTATS_REPORT_USED) N(OFFLOAD_XSTATS_REPORT_DELTA) 1795 N(XDP_FEAT_CHANGE) 1796 } 1797 #undef N 1798 return "UNKNOWN_NETDEV_EVENT"; 1799 } 1800 EXPORT_SYMBOL_GPL(netdev_cmd_to_name); 1801 1802 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val, 1803 struct net_device *dev) 1804 { 1805 struct netdev_notifier_info info = { 1806 .dev = dev, 1807 }; 1808 1809 return nb->notifier_call(nb, val, &info); 1810 } 1811 1812 static int call_netdevice_register_notifiers(struct notifier_block *nb, 1813 struct net_device *dev) 1814 { 1815 int err; 1816 1817 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev); 1818 err = notifier_to_errno(err); 1819 if (err) 1820 return err; 1821 1822 if (!(dev->flags & IFF_UP)) 1823 return 0; 1824 1825 call_netdevice_notifier(nb, NETDEV_UP, dev); 1826 return 0; 1827 } 1828 1829 static void call_netdevice_unregister_notifiers(struct notifier_block *nb, 1830 struct net_device *dev) 1831 { 1832 if (dev->flags & IFF_UP) { 1833 call_netdevice_notifier(nb, NETDEV_GOING_DOWN, 1834 dev); 1835 call_netdevice_notifier(nb, NETDEV_DOWN, dev); 1836 } 1837 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev); 1838 } 1839 1840 static int call_netdevice_register_net_notifiers(struct notifier_block *nb, 1841 struct net *net) 1842 { 1843 struct net_device *dev; 1844 int err; 1845 1846 for_each_netdev(net, dev) { 1847 err = call_netdevice_register_notifiers(nb, dev); 1848 if (err) 1849 goto rollback; 1850 } 1851 return 0; 1852 1853 rollback: 1854 for_each_netdev_continue_reverse(net, dev) 1855 call_netdevice_unregister_notifiers(nb, dev); 1856 return err; 1857 } 1858 1859 static void call_netdevice_unregister_net_notifiers(struct notifier_block *nb, 1860 struct net *net) 1861 { 1862 struct net_device *dev; 1863 1864 for_each_netdev(net, dev) 1865 call_netdevice_unregister_notifiers(nb, dev); 1866 } 1867 1868 static int dev_boot_phase = 1; 1869 1870 /** 1871 * register_netdevice_notifier - register a network notifier block 1872 * @nb: notifier 1873 * 1874 * Register a notifier to be called when network device events occur. 1875 * The notifier passed is linked into the kernel structures and must 1876 * not be reused until it has been unregistered. A negative errno code 1877 * is returned on a failure. 1878 * 1879 * When registered all registration and up events are replayed 1880 * to the new notifier to allow device to have a race free 1881 * view of the network device list. 1882 */ 1883 1884 int register_netdevice_notifier(struct notifier_block *nb) 1885 { 1886 struct net *net; 1887 int err; 1888 1889 /* Close race with setup_net() and cleanup_net() */ 1890 down_write(&pernet_ops_rwsem); 1891 1892 /* When RTNL is removed, we need protection for netdev_chain. */ 1893 rtnl_lock(); 1894 1895 err = raw_notifier_chain_register(&netdev_chain, nb); 1896 if (err) 1897 goto unlock; 1898 if (dev_boot_phase) 1899 goto unlock; 1900 for_each_net(net) { 1901 __rtnl_net_lock(net); 1902 err = call_netdevice_register_net_notifiers(nb, net); 1903 __rtnl_net_unlock(net); 1904 if (err) 1905 goto rollback; 1906 } 1907 1908 unlock: 1909 rtnl_unlock(); 1910 up_write(&pernet_ops_rwsem); 1911 return err; 1912 1913 rollback: 1914 for_each_net_continue_reverse(net) { 1915 __rtnl_net_lock(net); 1916 call_netdevice_unregister_net_notifiers(nb, net); 1917 __rtnl_net_unlock(net); 1918 } 1919 1920 raw_notifier_chain_unregister(&netdev_chain, nb); 1921 goto unlock; 1922 } 1923 EXPORT_SYMBOL(register_netdevice_notifier); 1924 1925 /** 1926 * unregister_netdevice_notifier - unregister a network notifier block 1927 * @nb: notifier 1928 * 1929 * Unregister a notifier previously registered by 1930 * register_netdevice_notifier(). The notifier is unlinked into the 1931 * kernel structures and may then be reused. A negative errno code 1932 * is returned on a failure. 1933 * 1934 * After unregistering unregister and down device events are synthesized 1935 * for all devices on the device list to the removed notifier to remove 1936 * the need for special case cleanup code. 1937 */ 1938 1939 int unregister_netdevice_notifier(struct notifier_block *nb) 1940 { 1941 struct net *net; 1942 int err; 1943 1944 /* Close race with setup_net() and cleanup_net() */ 1945 down_write(&pernet_ops_rwsem); 1946 rtnl_lock(); 1947 err = raw_notifier_chain_unregister(&netdev_chain, nb); 1948 if (err) 1949 goto unlock; 1950 1951 for_each_net(net) { 1952 __rtnl_net_lock(net); 1953 call_netdevice_unregister_net_notifiers(nb, net); 1954 __rtnl_net_unlock(net); 1955 } 1956 1957 unlock: 1958 rtnl_unlock(); 1959 up_write(&pernet_ops_rwsem); 1960 return err; 1961 } 1962 EXPORT_SYMBOL(unregister_netdevice_notifier); 1963 1964 static int __register_netdevice_notifier_net(struct net *net, 1965 struct notifier_block *nb, 1966 bool ignore_call_fail) 1967 { 1968 int err; 1969 1970 err = raw_notifier_chain_register(&net->netdev_chain, nb); 1971 if (err) 1972 return err; 1973 if (dev_boot_phase) 1974 return 0; 1975 1976 err = call_netdevice_register_net_notifiers(nb, net); 1977 if (err && !ignore_call_fail) 1978 goto chain_unregister; 1979 1980 return 0; 1981 1982 chain_unregister: 1983 raw_notifier_chain_unregister(&net->netdev_chain, nb); 1984 return err; 1985 } 1986 1987 static int __unregister_netdevice_notifier_net(struct net *net, 1988 struct notifier_block *nb) 1989 { 1990 int err; 1991 1992 err = raw_notifier_chain_unregister(&net->netdev_chain, nb); 1993 if (err) 1994 return err; 1995 1996 call_netdevice_unregister_net_notifiers(nb, net); 1997 return 0; 1998 } 1999 2000 /** 2001 * register_netdevice_notifier_net - register a per-netns network notifier block 2002 * @net: network namespace 2003 * @nb: notifier 2004 * 2005 * Register a notifier to be called when network device events occur. 2006 * The notifier passed is linked into the kernel structures and must 2007 * not be reused until it has been unregistered. A negative errno code 2008 * is returned on a failure. 2009 * 2010 * When registered all registration and up events are replayed 2011 * to the new notifier to allow device to have a race free 2012 * view of the network device list. 2013 */ 2014 2015 int register_netdevice_notifier_net(struct net *net, struct notifier_block *nb) 2016 { 2017 int err; 2018 2019 rtnl_net_lock(net); 2020 err = __register_netdevice_notifier_net(net, nb, false); 2021 rtnl_net_unlock(net); 2022 2023 return err; 2024 } 2025 EXPORT_SYMBOL(register_netdevice_notifier_net); 2026 2027 /** 2028 * unregister_netdevice_notifier_net - unregister a per-netns 2029 * network notifier block 2030 * @net: network namespace 2031 * @nb: notifier 2032 * 2033 * Unregister a notifier previously registered by 2034 * register_netdevice_notifier_net(). The notifier is unlinked from the 2035 * kernel structures and may then be reused. A negative errno code 2036 * is returned on a failure. 2037 * 2038 * After unregistering unregister and down device events are synthesized 2039 * for all devices on the device list to the removed notifier to remove 2040 * the need for special case cleanup code. 2041 */ 2042 2043 int unregister_netdevice_notifier_net(struct net *net, 2044 struct notifier_block *nb) 2045 { 2046 int err; 2047 2048 rtnl_net_lock(net); 2049 err = __unregister_netdevice_notifier_net(net, nb); 2050 rtnl_net_unlock(net); 2051 2052 return err; 2053 } 2054 EXPORT_SYMBOL(unregister_netdevice_notifier_net); 2055 2056 static void __move_netdevice_notifier_net(struct net *src_net, 2057 struct net *dst_net, 2058 struct notifier_block *nb) 2059 { 2060 __unregister_netdevice_notifier_net(src_net, nb); 2061 __register_netdevice_notifier_net(dst_net, nb, true); 2062 } 2063 2064 static void rtnl_net_dev_lock(struct net_device *dev) 2065 { 2066 bool again; 2067 2068 do { 2069 struct net *net; 2070 2071 again = false; 2072 2073 /* netns might be being dismantled. */ 2074 rcu_read_lock(); 2075 net = dev_net_rcu(dev); 2076 net_passive_inc(net); 2077 rcu_read_unlock(); 2078 2079 rtnl_net_lock(net); 2080 2081 #ifdef CONFIG_NET_NS 2082 /* dev might have been moved to another netns. */ 2083 if (!net_eq(net, rcu_access_pointer(dev->nd_net.net))) { 2084 rtnl_net_unlock(net); 2085 net_passive_dec(net); 2086 again = true; 2087 } 2088 #endif 2089 } while (again); 2090 } 2091 2092 static void rtnl_net_dev_unlock(struct net_device *dev) 2093 { 2094 struct net *net = dev_net(dev); 2095 2096 rtnl_net_unlock(net); 2097 net_passive_dec(net); 2098 } 2099 2100 int register_netdevice_notifier_dev_net(struct net_device *dev, 2101 struct notifier_block *nb, 2102 struct netdev_net_notifier *nn) 2103 { 2104 int err; 2105 2106 rtnl_net_dev_lock(dev); 2107 err = __register_netdevice_notifier_net(dev_net(dev), nb, false); 2108 if (!err) { 2109 nn->nb = nb; 2110 list_add(&nn->list, &dev->net_notifier_list); 2111 } 2112 rtnl_net_dev_unlock(dev); 2113 2114 return err; 2115 } 2116 EXPORT_SYMBOL(register_netdevice_notifier_dev_net); 2117 2118 int unregister_netdevice_notifier_dev_net(struct net_device *dev, 2119 struct notifier_block *nb, 2120 struct netdev_net_notifier *nn) 2121 { 2122 int err; 2123 2124 rtnl_net_dev_lock(dev); 2125 list_del(&nn->list); 2126 err = __unregister_netdevice_notifier_net(dev_net(dev), nb); 2127 rtnl_net_dev_unlock(dev); 2128 2129 return err; 2130 } 2131 EXPORT_SYMBOL(unregister_netdevice_notifier_dev_net); 2132 2133 static void move_netdevice_notifiers_dev_net(struct net_device *dev, 2134 struct net *net) 2135 { 2136 struct netdev_net_notifier *nn; 2137 2138 list_for_each_entry(nn, &dev->net_notifier_list, list) 2139 __move_netdevice_notifier_net(dev_net(dev), net, nn->nb); 2140 } 2141 2142 /** 2143 * call_netdevice_notifiers_info - call all network notifier blocks 2144 * @val: value passed unmodified to notifier function 2145 * @info: notifier information data 2146 * 2147 * Call all network notifier blocks. Parameters and return value 2148 * are as for raw_notifier_call_chain(). 2149 */ 2150 2151 int call_netdevice_notifiers_info(unsigned long val, 2152 struct netdev_notifier_info *info) 2153 { 2154 struct net *net = dev_net(info->dev); 2155 int ret; 2156 2157 ASSERT_RTNL(); 2158 2159 /* Run per-netns notifier block chain first, then run the global one. 2160 * Hopefully, one day, the global one is going to be removed after 2161 * all notifier block registrators get converted to be per-netns. 2162 */ 2163 ret = raw_notifier_call_chain(&net->netdev_chain, val, info); 2164 if (ret & NOTIFY_STOP_MASK) 2165 return ret; 2166 return raw_notifier_call_chain(&netdev_chain, val, info); 2167 } 2168 2169 /** 2170 * call_netdevice_notifiers_info_robust - call per-netns notifier blocks 2171 * for and rollback on error 2172 * @val_up: value passed unmodified to notifier function 2173 * @val_down: value passed unmodified to the notifier function when 2174 * recovering from an error on @val_up 2175 * @info: notifier information data 2176 * 2177 * Call all per-netns network notifier blocks, but not notifier blocks on 2178 * the global notifier chain. Parameters and return value are as for 2179 * raw_notifier_call_chain_robust(). 2180 */ 2181 2182 static int 2183 call_netdevice_notifiers_info_robust(unsigned long val_up, 2184 unsigned long val_down, 2185 struct netdev_notifier_info *info) 2186 { 2187 struct net *net = dev_net(info->dev); 2188 2189 ASSERT_RTNL(); 2190 2191 return raw_notifier_call_chain_robust(&net->netdev_chain, 2192 val_up, val_down, info); 2193 } 2194 2195 static int call_netdevice_notifiers_extack(unsigned long val, 2196 struct net_device *dev, 2197 struct netlink_ext_ack *extack) 2198 { 2199 struct netdev_notifier_info info = { 2200 .dev = dev, 2201 .extack = extack, 2202 }; 2203 2204 return call_netdevice_notifiers_info(val, &info); 2205 } 2206 2207 /** 2208 * call_netdevice_notifiers - call all network notifier blocks 2209 * @val: value passed unmodified to notifier function 2210 * @dev: net_device pointer passed unmodified to notifier function 2211 * 2212 * Call all network notifier blocks. Parameters and return value 2213 * are as for raw_notifier_call_chain(). 2214 */ 2215 2216 int call_netdevice_notifiers(unsigned long val, struct net_device *dev) 2217 { 2218 return call_netdevice_notifiers_extack(val, dev, NULL); 2219 } 2220 EXPORT_SYMBOL(call_netdevice_notifiers); 2221 2222 /** 2223 * call_netdevice_notifiers_mtu - call all network notifier blocks 2224 * @val: value passed unmodified to notifier function 2225 * @dev: net_device pointer passed unmodified to notifier function 2226 * @arg: additional u32 argument passed to the notifier function 2227 * 2228 * Call all network notifier blocks. Parameters and return value 2229 * are as for raw_notifier_call_chain(). 2230 */ 2231 static int call_netdevice_notifiers_mtu(unsigned long val, 2232 struct net_device *dev, u32 arg) 2233 { 2234 struct netdev_notifier_info_ext info = { 2235 .info.dev = dev, 2236 .ext.mtu = arg, 2237 }; 2238 2239 BUILD_BUG_ON(offsetof(struct netdev_notifier_info_ext, info) != 0); 2240 2241 return call_netdevice_notifiers_info(val, &info.info); 2242 } 2243 2244 #ifdef CONFIG_NET_INGRESS 2245 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key); 2246 2247 void net_inc_ingress_queue(void) 2248 { 2249 static_branch_inc(&ingress_needed_key); 2250 } 2251 EXPORT_SYMBOL_GPL(net_inc_ingress_queue); 2252 2253 void net_dec_ingress_queue(void) 2254 { 2255 static_branch_dec(&ingress_needed_key); 2256 } 2257 EXPORT_SYMBOL_GPL(net_dec_ingress_queue); 2258 #endif 2259 2260 #ifdef CONFIG_NET_EGRESS 2261 static DEFINE_STATIC_KEY_FALSE(egress_needed_key); 2262 2263 void net_inc_egress_queue(void) 2264 { 2265 static_branch_inc(&egress_needed_key); 2266 } 2267 EXPORT_SYMBOL_GPL(net_inc_egress_queue); 2268 2269 void net_dec_egress_queue(void) 2270 { 2271 static_branch_dec(&egress_needed_key); 2272 } 2273 EXPORT_SYMBOL_GPL(net_dec_egress_queue); 2274 #endif 2275 2276 #ifdef CONFIG_NET_CLS_ACT 2277 DEFINE_STATIC_KEY_FALSE(tcf_sw_enabled_key); 2278 EXPORT_SYMBOL(tcf_sw_enabled_key); 2279 #endif 2280 2281 DEFINE_STATIC_KEY_FALSE(netstamp_needed_key); 2282 EXPORT_SYMBOL(netstamp_needed_key); 2283 #ifdef CONFIG_JUMP_LABEL 2284 static atomic_t netstamp_needed_deferred; 2285 static atomic_t netstamp_wanted; 2286 static void netstamp_clear(struct work_struct *work) 2287 { 2288 int deferred = atomic_xchg(&netstamp_needed_deferred, 0); 2289 int wanted; 2290 2291 wanted = atomic_add_return(deferred, &netstamp_wanted); 2292 if (wanted > 0) 2293 static_branch_enable(&netstamp_needed_key); 2294 else 2295 static_branch_disable(&netstamp_needed_key); 2296 } 2297 static DECLARE_WORK(netstamp_work, netstamp_clear); 2298 #endif 2299 2300 void net_enable_timestamp(void) 2301 { 2302 #ifdef CONFIG_JUMP_LABEL 2303 int wanted = atomic_read(&netstamp_wanted); 2304 2305 while (wanted > 0) { 2306 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted + 1)) 2307 return; 2308 } 2309 atomic_inc(&netstamp_needed_deferred); 2310 schedule_work(&netstamp_work); 2311 #else 2312 static_branch_inc(&netstamp_needed_key); 2313 #endif 2314 } 2315 EXPORT_SYMBOL(net_enable_timestamp); 2316 2317 void net_disable_timestamp(void) 2318 { 2319 #ifdef CONFIG_JUMP_LABEL 2320 int wanted = atomic_read(&netstamp_wanted); 2321 2322 while (wanted > 1) { 2323 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted - 1)) 2324 return; 2325 } 2326 atomic_dec(&netstamp_needed_deferred); 2327 schedule_work(&netstamp_work); 2328 #else 2329 static_branch_dec(&netstamp_needed_key); 2330 #endif 2331 } 2332 EXPORT_SYMBOL(net_disable_timestamp); 2333 2334 static inline void net_timestamp_set(struct sk_buff *skb) 2335 { 2336 skb->tstamp = 0; 2337 skb->tstamp_type = SKB_CLOCK_REALTIME; 2338 if (static_branch_unlikely(&netstamp_needed_key)) 2339 skb->tstamp = ktime_get_real(); 2340 } 2341 2342 #define net_timestamp_check(COND, SKB) \ 2343 if (static_branch_unlikely(&netstamp_needed_key)) { \ 2344 if ((COND) && !(SKB)->tstamp) \ 2345 (SKB)->tstamp = ktime_get_real(); \ 2346 } \ 2347 2348 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb) 2349 { 2350 return __is_skb_forwardable(dev, skb, true); 2351 } 2352 EXPORT_SYMBOL_GPL(is_skb_forwardable); 2353 2354 static int __dev_forward_skb2(struct net_device *dev, struct sk_buff *skb, 2355 bool check_mtu) 2356 { 2357 int ret = ____dev_forward_skb(dev, skb, check_mtu); 2358 2359 if (likely(!ret)) { 2360 skb->protocol = eth_type_trans(skb, dev); 2361 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN); 2362 } 2363 2364 return ret; 2365 } 2366 2367 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2368 { 2369 return __dev_forward_skb2(dev, skb, true); 2370 } 2371 EXPORT_SYMBOL_GPL(__dev_forward_skb); 2372 2373 /** 2374 * dev_forward_skb - loopback an skb to another netif 2375 * 2376 * @dev: destination network device 2377 * @skb: buffer to forward 2378 * 2379 * return values: 2380 * NET_RX_SUCCESS (no congestion) 2381 * NET_RX_DROP (packet was dropped, but freed) 2382 * 2383 * dev_forward_skb can be used for injecting an skb from the 2384 * start_xmit function of one device into the receive queue 2385 * of another device. 2386 * 2387 * The receiving device may be in another namespace, so 2388 * we have to clear all information in the skb that could 2389 * impact namespace isolation. 2390 */ 2391 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2392 { 2393 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb); 2394 } 2395 EXPORT_SYMBOL_GPL(dev_forward_skb); 2396 2397 int dev_forward_skb_nomtu(struct net_device *dev, struct sk_buff *skb) 2398 { 2399 return __dev_forward_skb2(dev, skb, false) ?: netif_rx_internal(skb); 2400 } 2401 2402 static inline int deliver_skb(struct sk_buff *skb, 2403 struct packet_type *pt_prev, 2404 struct net_device *orig_dev) 2405 { 2406 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 2407 return -ENOMEM; 2408 refcount_inc(&skb->users); 2409 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 2410 } 2411 2412 static inline void deliver_ptype_list_skb(struct sk_buff *skb, 2413 struct packet_type **pt, 2414 struct net_device *orig_dev, 2415 __be16 type, 2416 struct list_head *ptype_list) 2417 { 2418 struct packet_type *ptype, *pt_prev = *pt; 2419 2420 list_for_each_entry_rcu(ptype, ptype_list, list) { 2421 if (ptype->type != type) 2422 continue; 2423 if (pt_prev) 2424 deliver_skb(skb, pt_prev, orig_dev); 2425 pt_prev = ptype; 2426 } 2427 *pt = pt_prev; 2428 } 2429 2430 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb) 2431 { 2432 if (!ptype->af_packet_priv || !skb->sk) 2433 return false; 2434 2435 if (ptype->id_match) 2436 return ptype->id_match(ptype, skb->sk); 2437 else if ((struct sock *)ptype->af_packet_priv == skb->sk) 2438 return true; 2439 2440 return false; 2441 } 2442 2443 /** 2444 * dev_nit_active - return true if any network interface taps are in use 2445 * 2446 * @dev: network device to check for the presence of taps 2447 */ 2448 bool dev_nit_active(struct net_device *dev) 2449 { 2450 return !list_empty(&net_hotdata.ptype_all) || 2451 !list_empty(&dev->ptype_all); 2452 } 2453 EXPORT_SYMBOL_GPL(dev_nit_active); 2454 2455 /* 2456 * Support routine. Sends outgoing frames to any network 2457 * taps currently in use. 2458 */ 2459 2460 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev) 2461 { 2462 struct list_head *ptype_list = &net_hotdata.ptype_all; 2463 struct packet_type *ptype, *pt_prev = NULL; 2464 struct sk_buff *skb2 = NULL; 2465 2466 rcu_read_lock(); 2467 again: 2468 list_for_each_entry_rcu(ptype, ptype_list, list) { 2469 if (READ_ONCE(ptype->ignore_outgoing)) 2470 continue; 2471 2472 /* Never send packets back to the socket 2473 * they originated from - MvS (miquels@drinkel.ow.org) 2474 */ 2475 if (skb_loop_sk(ptype, skb)) 2476 continue; 2477 2478 if (pt_prev) { 2479 deliver_skb(skb2, pt_prev, skb->dev); 2480 pt_prev = ptype; 2481 continue; 2482 } 2483 2484 /* need to clone skb, done only once */ 2485 skb2 = skb_clone(skb, GFP_ATOMIC); 2486 if (!skb2) 2487 goto out_unlock; 2488 2489 net_timestamp_set(skb2); 2490 2491 /* skb->nh should be correctly 2492 * set by sender, so that the second statement is 2493 * just protection against buggy protocols. 2494 */ 2495 skb_reset_mac_header(skb2); 2496 2497 if (skb_network_header(skb2) < skb2->data || 2498 skb_network_header(skb2) > skb_tail_pointer(skb2)) { 2499 net_crit_ratelimited("protocol %04x is buggy, dev %s\n", 2500 ntohs(skb2->protocol), 2501 dev->name); 2502 skb_reset_network_header(skb2); 2503 } 2504 2505 skb2->transport_header = skb2->network_header; 2506 skb2->pkt_type = PACKET_OUTGOING; 2507 pt_prev = ptype; 2508 } 2509 2510 if (ptype_list == &net_hotdata.ptype_all) { 2511 ptype_list = &dev->ptype_all; 2512 goto again; 2513 } 2514 out_unlock: 2515 if (pt_prev) { 2516 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC)) 2517 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev); 2518 else 2519 kfree_skb(skb2); 2520 } 2521 rcu_read_unlock(); 2522 } 2523 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit); 2524 2525 /** 2526 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change 2527 * @dev: Network device 2528 * @txq: number of queues available 2529 * 2530 * If real_num_tx_queues is changed the tc mappings may no longer be 2531 * valid. To resolve this verify the tc mapping remains valid and if 2532 * not NULL the mapping. With no priorities mapping to this 2533 * offset/count pair it will no longer be used. In the worst case TC0 2534 * is invalid nothing can be done so disable priority mappings. If is 2535 * expected that drivers will fix this mapping if they can before 2536 * calling netif_set_real_num_tx_queues. 2537 */ 2538 static void netif_setup_tc(struct net_device *dev, unsigned int txq) 2539 { 2540 int i; 2541 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2542 2543 /* If TC0 is invalidated disable TC mapping */ 2544 if (tc->offset + tc->count > txq) { 2545 netdev_warn(dev, "Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n"); 2546 dev->num_tc = 0; 2547 return; 2548 } 2549 2550 /* Invalidated prio to tc mappings set to TC0 */ 2551 for (i = 1; i < TC_BITMASK + 1; i++) { 2552 int q = netdev_get_prio_tc_map(dev, i); 2553 2554 tc = &dev->tc_to_txq[q]; 2555 if (tc->offset + tc->count > txq) { 2556 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", 2557 i, q); 2558 netdev_set_prio_tc_map(dev, i, 0); 2559 } 2560 } 2561 } 2562 2563 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq) 2564 { 2565 if (dev->num_tc) { 2566 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2567 int i; 2568 2569 /* walk through the TCs and see if it falls into any of them */ 2570 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) { 2571 if ((txq - tc->offset) < tc->count) 2572 return i; 2573 } 2574 2575 /* didn't find it, just return -1 to indicate no match */ 2576 return -1; 2577 } 2578 2579 return 0; 2580 } 2581 EXPORT_SYMBOL(netdev_txq_to_tc); 2582 2583 #ifdef CONFIG_XPS 2584 static struct static_key xps_needed __read_mostly; 2585 static struct static_key xps_rxqs_needed __read_mostly; 2586 static DEFINE_MUTEX(xps_map_mutex); 2587 #define xmap_dereference(P) \ 2588 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex)) 2589 2590 static bool remove_xps_queue(struct xps_dev_maps *dev_maps, 2591 struct xps_dev_maps *old_maps, int tci, u16 index) 2592 { 2593 struct xps_map *map = NULL; 2594 int pos; 2595 2596 map = xmap_dereference(dev_maps->attr_map[tci]); 2597 if (!map) 2598 return false; 2599 2600 for (pos = map->len; pos--;) { 2601 if (map->queues[pos] != index) 2602 continue; 2603 2604 if (map->len > 1) { 2605 map->queues[pos] = map->queues[--map->len]; 2606 break; 2607 } 2608 2609 if (old_maps) 2610 RCU_INIT_POINTER(old_maps->attr_map[tci], NULL); 2611 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2612 kfree_rcu(map, rcu); 2613 return false; 2614 } 2615 2616 return true; 2617 } 2618 2619 static bool remove_xps_queue_cpu(struct net_device *dev, 2620 struct xps_dev_maps *dev_maps, 2621 int cpu, u16 offset, u16 count) 2622 { 2623 int num_tc = dev_maps->num_tc; 2624 bool active = false; 2625 int tci; 2626 2627 for (tci = cpu * num_tc; num_tc--; tci++) { 2628 int i, j; 2629 2630 for (i = count, j = offset; i--; j++) { 2631 if (!remove_xps_queue(dev_maps, NULL, tci, j)) 2632 break; 2633 } 2634 2635 active |= i < 0; 2636 } 2637 2638 return active; 2639 } 2640 2641 static void reset_xps_maps(struct net_device *dev, 2642 struct xps_dev_maps *dev_maps, 2643 enum xps_map_type type) 2644 { 2645 static_key_slow_dec_cpuslocked(&xps_needed); 2646 if (type == XPS_RXQS) 2647 static_key_slow_dec_cpuslocked(&xps_rxqs_needed); 2648 2649 RCU_INIT_POINTER(dev->xps_maps[type], NULL); 2650 2651 kfree_rcu(dev_maps, rcu); 2652 } 2653 2654 static void clean_xps_maps(struct net_device *dev, enum xps_map_type type, 2655 u16 offset, u16 count) 2656 { 2657 struct xps_dev_maps *dev_maps; 2658 bool active = false; 2659 int i, j; 2660 2661 dev_maps = xmap_dereference(dev->xps_maps[type]); 2662 if (!dev_maps) 2663 return; 2664 2665 for (j = 0; j < dev_maps->nr_ids; j++) 2666 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, count); 2667 if (!active) 2668 reset_xps_maps(dev, dev_maps, type); 2669 2670 if (type == XPS_CPUS) { 2671 for (i = offset + (count - 1); count--; i--) 2672 netdev_queue_numa_node_write( 2673 netdev_get_tx_queue(dev, i), NUMA_NO_NODE); 2674 } 2675 } 2676 2677 static void netif_reset_xps_queues(struct net_device *dev, u16 offset, 2678 u16 count) 2679 { 2680 if (!static_key_false(&xps_needed)) 2681 return; 2682 2683 cpus_read_lock(); 2684 mutex_lock(&xps_map_mutex); 2685 2686 if (static_key_false(&xps_rxqs_needed)) 2687 clean_xps_maps(dev, XPS_RXQS, offset, count); 2688 2689 clean_xps_maps(dev, XPS_CPUS, offset, count); 2690 2691 mutex_unlock(&xps_map_mutex); 2692 cpus_read_unlock(); 2693 } 2694 2695 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index) 2696 { 2697 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index); 2698 } 2699 2700 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index, 2701 u16 index, bool is_rxqs_map) 2702 { 2703 struct xps_map *new_map; 2704 int alloc_len = XPS_MIN_MAP_ALLOC; 2705 int i, pos; 2706 2707 for (pos = 0; map && pos < map->len; pos++) { 2708 if (map->queues[pos] != index) 2709 continue; 2710 return map; 2711 } 2712 2713 /* Need to add tx-queue to this CPU's/rx-queue's existing map */ 2714 if (map) { 2715 if (pos < map->alloc_len) 2716 return map; 2717 2718 alloc_len = map->alloc_len * 2; 2719 } 2720 2721 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's 2722 * map 2723 */ 2724 if (is_rxqs_map) 2725 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL); 2726 else 2727 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL, 2728 cpu_to_node(attr_index)); 2729 if (!new_map) 2730 return NULL; 2731 2732 for (i = 0; i < pos; i++) 2733 new_map->queues[i] = map->queues[i]; 2734 new_map->alloc_len = alloc_len; 2735 new_map->len = pos; 2736 2737 return new_map; 2738 } 2739 2740 /* Copy xps maps at a given index */ 2741 static void xps_copy_dev_maps(struct xps_dev_maps *dev_maps, 2742 struct xps_dev_maps *new_dev_maps, int index, 2743 int tc, bool skip_tc) 2744 { 2745 int i, tci = index * dev_maps->num_tc; 2746 struct xps_map *map; 2747 2748 /* copy maps belonging to foreign traffic classes */ 2749 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2750 if (i == tc && skip_tc) 2751 continue; 2752 2753 /* fill in the new device map from the old device map */ 2754 map = xmap_dereference(dev_maps->attr_map[tci]); 2755 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2756 } 2757 } 2758 2759 /* Must be called under cpus_read_lock */ 2760 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask, 2761 u16 index, enum xps_map_type type) 2762 { 2763 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL, *old_dev_maps = NULL; 2764 const unsigned long *online_mask = NULL; 2765 bool active = false, copy = false; 2766 int i, j, tci, numa_node_id = -2; 2767 int maps_sz, num_tc = 1, tc = 0; 2768 struct xps_map *map, *new_map; 2769 unsigned int nr_ids; 2770 2771 WARN_ON_ONCE(index >= dev->num_tx_queues); 2772 2773 if (dev->num_tc) { 2774 /* Do not allow XPS on subordinate device directly */ 2775 num_tc = dev->num_tc; 2776 if (num_tc < 0) 2777 return -EINVAL; 2778 2779 /* If queue belongs to subordinate dev use its map */ 2780 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev; 2781 2782 tc = netdev_txq_to_tc(dev, index); 2783 if (tc < 0) 2784 return -EINVAL; 2785 } 2786 2787 mutex_lock(&xps_map_mutex); 2788 2789 dev_maps = xmap_dereference(dev->xps_maps[type]); 2790 if (type == XPS_RXQS) { 2791 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues); 2792 nr_ids = dev->num_rx_queues; 2793 } else { 2794 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc); 2795 if (num_possible_cpus() > 1) 2796 online_mask = cpumask_bits(cpu_online_mask); 2797 nr_ids = nr_cpu_ids; 2798 } 2799 2800 if (maps_sz < L1_CACHE_BYTES) 2801 maps_sz = L1_CACHE_BYTES; 2802 2803 /* The old dev_maps could be larger or smaller than the one we're 2804 * setting up now, as dev->num_tc or nr_ids could have been updated in 2805 * between. We could try to be smart, but let's be safe instead and only 2806 * copy foreign traffic classes if the two map sizes match. 2807 */ 2808 if (dev_maps && 2809 dev_maps->num_tc == num_tc && dev_maps->nr_ids == nr_ids) 2810 copy = true; 2811 2812 /* allocate memory for queue storage */ 2813 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids), 2814 j < nr_ids;) { 2815 if (!new_dev_maps) { 2816 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL); 2817 if (!new_dev_maps) { 2818 mutex_unlock(&xps_map_mutex); 2819 return -ENOMEM; 2820 } 2821 2822 new_dev_maps->nr_ids = nr_ids; 2823 new_dev_maps->num_tc = num_tc; 2824 } 2825 2826 tci = j * num_tc + tc; 2827 map = copy ? xmap_dereference(dev_maps->attr_map[tci]) : NULL; 2828 2829 map = expand_xps_map(map, j, index, type == XPS_RXQS); 2830 if (!map) 2831 goto error; 2832 2833 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2834 } 2835 2836 if (!new_dev_maps) 2837 goto out_no_new_maps; 2838 2839 if (!dev_maps) { 2840 /* Increment static keys at most once per type */ 2841 static_key_slow_inc_cpuslocked(&xps_needed); 2842 if (type == XPS_RXQS) 2843 static_key_slow_inc_cpuslocked(&xps_rxqs_needed); 2844 } 2845 2846 for (j = 0; j < nr_ids; j++) { 2847 bool skip_tc = false; 2848 2849 tci = j * num_tc + tc; 2850 if (netif_attr_test_mask(j, mask, nr_ids) && 2851 netif_attr_test_online(j, online_mask, nr_ids)) { 2852 /* add tx-queue to CPU/rx-queue maps */ 2853 int pos = 0; 2854 2855 skip_tc = true; 2856 2857 map = xmap_dereference(new_dev_maps->attr_map[tci]); 2858 while ((pos < map->len) && (map->queues[pos] != index)) 2859 pos++; 2860 2861 if (pos == map->len) 2862 map->queues[map->len++] = index; 2863 #ifdef CONFIG_NUMA 2864 if (type == XPS_CPUS) { 2865 if (numa_node_id == -2) 2866 numa_node_id = cpu_to_node(j); 2867 else if (numa_node_id != cpu_to_node(j)) 2868 numa_node_id = -1; 2869 } 2870 #endif 2871 } 2872 2873 if (copy) 2874 xps_copy_dev_maps(dev_maps, new_dev_maps, j, tc, 2875 skip_tc); 2876 } 2877 2878 rcu_assign_pointer(dev->xps_maps[type], new_dev_maps); 2879 2880 /* Cleanup old maps */ 2881 if (!dev_maps) 2882 goto out_no_old_maps; 2883 2884 for (j = 0; j < dev_maps->nr_ids; j++) { 2885 for (i = num_tc, tci = j * dev_maps->num_tc; i--; tci++) { 2886 map = xmap_dereference(dev_maps->attr_map[tci]); 2887 if (!map) 2888 continue; 2889 2890 if (copy) { 2891 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2892 if (map == new_map) 2893 continue; 2894 } 2895 2896 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2897 kfree_rcu(map, rcu); 2898 } 2899 } 2900 2901 old_dev_maps = dev_maps; 2902 2903 out_no_old_maps: 2904 dev_maps = new_dev_maps; 2905 active = true; 2906 2907 out_no_new_maps: 2908 if (type == XPS_CPUS) 2909 /* update Tx queue numa node */ 2910 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index), 2911 (numa_node_id >= 0) ? 2912 numa_node_id : NUMA_NO_NODE); 2913 2914 if (!dev_maps) 2915 goto out_no_maps; 2916 2917 /* removes tx-queue from unused CPUs/rx-queues */ 2918 for (j = 0; j < dev_maps->nr_ids; j++) { 2919 tci = j * dev_maps->num_tc; 2920 2921 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2922 if (i == tc && 2923 netif_attr_test_mask(j, mask, dev_maps->nr_ids) && 2924 netif_attr_test_online(j, online_mask, dev_maps->nr_ids)) 2925 continue; 2926 2927 active |= remove_xps_queue(dev_maps, 2928 copy ? old_dev_maps : NULL, 2929 tci, index); 2930 } 2931 } 2932 2933 if (old_dev_maps) 2934 kfree_rcu(old_dev_maps, rcu); 2935 2936 /* free map if not active */ 2937 if (!active) 2938 reset_xps_maps(dev, dev_maps, type); 2939 2940 out_no_maps: 2941 mutex_unlock(&xps_map_mutex); 2942 2943 return 0; 2944 error: 2945 /* remove any maps that we added */ 2946 for (j = 0; j < nr_ids; j++) { 2947 for (i = num_tc, tci = j * num_tc; i--; tci++) { 2948 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2949 map = copy ? 2950 xmap_dereference(dev_maps->attr_map[tci]) : 2951 NULL; 2952 if (new_map && new_map != map) 2953 kfree(new_map); 2954 } 2955 } 2956 2957 mutex_unlock(&xps_map_mutex); 2958 2959 kfree(new_dev_maps); 2960 return -ENOMEM; 2961 } 2962 EXPORT_SYMBOL_GPL(__netif_set_xps_queue); 2963 2964 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, 2965 u16 index) 2966 { 2967 int ret; 2968 2969 cpus_read_lock(); 2970 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, XPS_CPUS); 2971 cpus_read_unlock(); 2972 2973 return ret; 2974 } 2975 EXPORT_SYMBOL(netif_set_xps_queue); 2976 2977 #endif 2978 static void netdev_unbind_all_sb_channels(struct net_device *dev) 2979 { 2980 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 2981 2982 /* Unbind any subordinate channels */ 2983 while (txq-- != &dev->_tx[0]) { 2984 if (txq->sb_dev) 2985 netdev_unbind_sb_channel(dev, txq->sb_dev); 2986 } 2987 } 2988 2989 void netdev_reset_tc(struct net_device *dev) 2990 { 2991 #ifdef CONFIG_XPS 2992 netif_reset_xps_queues_gt(dev, 0); 2993 #endif 2994 netdev_unbind_all_sb_channels(dev); 2995 2996 /* Reset TC configuration of device */ 2997 dev->num_tc = 0; 2998 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq)); 2999 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map)); 3000 } 3001 EXPORT_SYMBOL(netdev_reset_tc); 3002 3003 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset) 3004 { 3005 if (tc >= dev->num_tc) 3006 return -EINVAL; 3007 3008 #ifdef CONFIG_XPS 3009 netif_reset_xps_queues(dev, offset, count); 3010 #endif 3011 dev->tc_to_txq[tc].count = count; 3012 dev->tc_to_txq[tc].offset = offset; 3013 return 0; 3014 } 3015 EXPORT_SYMBOL(netdev_set_tc_queue); 3016 3017 int netdev_set_num_tc(struct net_device *dev, u8 num_tc) 3018 { 3019 if (num_tc > TC_MAX_QUEUE) 3020 return -EINVAL; 3021 3022 #ifdef CONFIG_XPS 3023 netif_reset_xps_queues_gt(dev, 0); 3024 #endif 3025 netdev_unbind_all_sb_channels(dev); 3026 3027 dev->num_tc = num_tc; 3028 return 0; 3029 } 3030 EXPORT_SYMBOL(netdev_set_num_tc); 3031 3032 void netdev_unbind_sb_channel(struct net_device *dev, 3033 struct net_device *sb_dev) 3034 { 3035 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 3036 3037 #ifdef CONFIG_XPS 3038 netif_reset_xps_queues_gt(sb_dev, 0); 3039 #endif 3040 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq)); 3041 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map)); 3042 3043 while (txq-- != &dev->_tx[0]) { 3044 if (txq->sb_dev == sb_dev) 3045 txq->sb_dev = NULL; 3046 } 3047 } 3048 EXPORT_SYMBOL(netdev_unbind_sb_channel); 3049 3050 int netdev_bind_sb_channel_queue(struct net_device *dev, 3051 struct net_device *sb_dev, 3052 u8 tc, u16 count, u16 offset) 3053 { 3054 /* Make certain the sb_dev and dev are already configured */ 3055 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc) 3056 return -EINVAL; 3057 3058 /* We cannot hand out queues we don't have */ 3059 if ((offset + count) > dev->real_num_tx_queues) 3060 return -EINVAL; 3061 3062 /* Record the mapping */ 3063 sb_dev->tc_to_txq[tc].count = count; 3064 sb_dev->tc_to_txq[tc].offset = offset; 3065 3066 /* Provide a way for Tx queue to find the tc_to_txq map or 3067 * XPS map for itself. 3068 */ 3069 while (count--) 3070 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev; 3071 3072 return 0; 3073 } 3074 EXPORT_SYMBOL(netdev_bind_sb_channel_queue); 3075 3076 int netdev_set_sb_channel(struct net_device *dev, u16 channel) 3077 { 3078 /* Do not use a multiqueue device to represent a subordinate channel */ 3079 if (netif_is_multiqueue(dev)) 3080 return -ENODEV; 3081 3082 /* We allow channels 1 - 32767 to be used for subordinate channels. 3083 * Channel 0 is meant to be "native" mode and used only to represent 3084 * the main root device. We allow writing 0 to reset the device back 3085 * to normal mode after being used as a subordinate channel. 3086 */ 3087 if (channel > S16_MAX) 3088 return -EINVAL; 3089 3090 dev->num_tc = -channel; 3091 3092 return 0; 3093 } 3094 EXPORT_SYMBOL(netdev_set_sb_channel); 3095 3096 /* 3097 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues 3098 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed. 3099 */ 3100 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq) 3101 { 3102 bool disabling; 3103 int rc; 3104 3105 disabling = txq < dev->real_num_tx_queues; 3106 3107 if (txq < 1 || txq > dev->num_tx_queues) 3108 return -EINVAL; 3109 3110 if (dev->reg_state == NETREG_REGISTERED || 3111 dev->reg_state == NETREG_UNREGISTERING) { 3112 ASSERT_RTNL(); 3113 3114 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues, 3115 txq); 3116 if (rc) 3117 return rc; 3118 3119 if (dev->num_tc) 3120 netif_setup_tc(dev, txq); 3121 3122 net_shaper_set_real_num_tx_queues(dev, txq); 3123 3124 dev_qdisc_change_real_num_tx(dev, txq); 3125 3126 dev->real_num_tx_queues = txq; 3127 3128 if (disabling) { 3129 synchronize_net(); 3130 qdisc_reset_all_tx_gt(dev, txq); 3131 #ifdef CONFIG_XPS 3132 netif_reset_xps_queues_gt(dev, txq); 3133 #endif 3134 } 3135 } else { 3136 dev->real_num_tx_queues = txq; 3137 } 3138 3139 return 0; 3140 } 3141 EXPORT_SYMBOL(netif_set_real_num_tx_queues); 3142 3143 #ifdef CONFIG_SYSFS 3144 /** 3145 * netif_set_real_num_rx_queues - set actual number of RX queues used 3146 * @dev: Network device 3147 * @rxq: Actual number of RX queues 3148 * 3149 * This must be called either with the rtnl_lock held or before 3150 * registration of the net device. Returns 0 on success, or a 3151 * negative error code. If called before registration, it always 3152 * succeeds. 3153 */ 3154 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq) 3155 { 3156 int rc; 3157 3158 if (rxq < 1 || rxq > dev->num_rx_queues) 3159 return -EINVAL; 3160 3161 if (dev->reg_state == NETREG_REGISTERED) { 3162 ASSERT_RTNL(); 3163 3164 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues, 3165 rxq); 3166 if (rc) 3167 return rc; 3168 } 3169 3170 dev->real_num_rx_queues = rxq; 3171 return 0; 3172 } 3173 EXPORT_SYMBOL(netif_set_real_num_rx_queues); 3174 #endif 3175 3176 /** 3177 * netif_set_real_num_queues - set actual number of RX and TX queues used 3178 * @dev: Network device 3179 * @txq: Actual number of TX queues 3180 * @rxq: Actual number of RX queues 3181 * 3182 * Set the real number of both TX and RX queues. 3183 * Does nothing if the number of queues is already correct. 3184 */ 3185 int netif_set_real_num_queues(struct net_device *dev, 3186 unsigned int txq, unsigned int rxq) 3187 { 3188 unsigned int old_rxq = dev->real_num_rx_queues; 3189 int err; 3190 3191 if (txq < 1 || txq > dev->num_tx_queues || 3192 rxq < 1 || rxq > dev->num_rx_queues) 3193 return -EINVAL; 3194 3195 /* Start from increases, so the error path only does decreases - 3196 * decreases can't fail. 3197 */ 3198 if (rxq > dev->real_num_rx_queues) { 3199 err = netif_set_real_num_rx_queues(dev, rxq); 3200 if (err) 3201 return err; 3202 } 3203 if (txq > dev->real_num_tx_queues) { 3204 err = netif_set_real_num_tx_queues(dev, txq); 3205 if (err) 3206 goto undo_rx; 3207 } 3208 if (rxq < dev->real_num_rx_queues) 3209 WARN_ON(netif_set_real_num_rx_queues(dev, rxq)); 3210 if (txq < dev->real_num_tx_queues) 3211 WARN_ON(netif_set_real_num_tx_queues(dev, txq)); 3212 3213 return 0; 3214 undo_rx: 3215 WARN_ON(netif_set_real_num_rx_queues(dev, old_rxq)); 3216 return err; 3217 } 3218 EXPORT_SYMBOL(netif_set_real_num_queues); 3219 3220 /** 3221 * netif_set_tso_max_size() - set the max size of TSO frames supported 3222 * @dev: netdev to update 3223 * @size: max skb->len of a TSO frame 3224 * 3225 * Set the limit on the size of TSO super-frames the device can handle. 3226 * Unless explicitly set the stack will assume the value of 3227 * %GSO_LEGACY_MAX_SIZE. 3228 */ 3229 void netif_set_tso_max_size(struct net_device *dev, unsigned int size) 3230 { 3231 dev->tso_max_size = min(GSO_MAX_SIZE, size); 3232 if (size < READ_ONCE(dev->gso_max_size)) 3233 netif_set_gso_max_size(dev, size); 3234 if (size < READ_ONCE(dev->gso_ipv4_max_size)) 3235 netif_set_gso_ipv4_max_size(dev, size); 3236 } 3237 EXPORT_SYMBOL(netif_set_tso_max_size); 3238 3239 /** 3240 * netif_set_tso_max_segs() - set the max number of segs supported for TSO 3241 * @dev: netdev to update 3242 * @segs: max number of TCP segments 3243 * 3244 * Set the limit on the number of TCP segments the device can generate from 3245 * a single TSO super-frame. 3246 * Unless explicitly set the stack will assume the value of %GSO_MAX_SEGS. 3247 */ 3248 void netif_set_tso_max_segs(struct net_device *dev, unsigned int segs) 3249 { 3250 dev->tso_max_segs = segs; 3251 if (segs < READ_ONCE(dev->gso_max_segs)) 3252 netif_set_gso_max_segs(dev, segs); 3253 } 3254 EXPORT_SYMBOL(netif_set_tso_max_segs); 3255 3256 /** 3257 * netif_inherit_tso_max() - copy all TSO limits from a lower device to an upper 3258 * @to: netdev to update 3259 * @from: netdev from which to copy the limits 3260 */ 3261 void netif_inherit_tso_max(struct net_device *to, const struct net_device *from) 3262 { 3263 netif_set_tso_max_size(to, from->tso_max_size); 3264 netif_set_tso_max_segs(to, from->tso_max_segs); 3265 } 3266 EXPORT_SYMBOL(netif_inherit_tso_max); 3267 3268 /** 3269 * netif_get_num_default_rss_queues - default number of RSS queues 3270 * 3271 * Default value is the number of physical cores if there are only 1 or 2, or 3272 * divided by 2 if there are more. 3273 */ 3274 int netif_get_num_default_rss_queues(void) 3275 { 3276 cpumask_var_t cpus; 3277 int cpu, count = 0; 3278 3279 if (unlikely(is_kdump_kernel() || !zalloc_cpumask_var(&cpus, GFP_KERNEL))) 3280 return 1; 3281 3282 cpumask_copy(cpus, cpu_online_mask); 3283 for_each_cpu(cpu, cpus) { 3284 ++count; 3285 cpumask_andnot(cpus, cpus, topology_sibling_cpumask(cpu)); 3286 } 3287 free_cpumask_var(cpus); 3288 3289 return count > 2 ? DIV_ROUND_UP(count, 2) : count; 3290 } 3291 EXPORT_SYMBOL(netif_get_num_default_rss_queues); 3292 3293 static void __netif_reschedule(struct Qdisc *q) 3294 { 3295 struct softnet_data *sd; 3296 unsigned long flags; 3297 3298 local_irq_save(flags); 3299 sd = this_cpu_ptr(&softnet_data); 3300 q->next_sched = NULL; 3301 *sd->output_queue_tailp = q; 3302 sd->output_queue_tailp = &q->next_sched; 3303 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3304 local_irq_restore(flags); 3305 } 3306 3307 void __netif_schedule(struct Qdisc *q) 3308 { 3309 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state)) 3310 __netif_reschedule(q); 3311 } 3312 EXPORT_SYMBOL(__netif_schedule); 3313 3314 struct dev_kfree_skb_cb { 3315 enum skb_drop_reason reason; 3316 }; 3317 3318 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb) 3319 { 3320 return (struct dev_kfree_skb_cb *)skb->cb; 3321 } 3322 3323 void netif_schedule_queue(struct netdev_queue *txq) 3324 { 3325 rcu_read_lock(); 3326 if (!netif_xmit_stopped(txq)) { 3327 struct Qdisc *q = rcu_dereference(txq->qdisc); 3328 3329 __netif_schedule(q); 3330 } 3331 rcu_read_unlock(); 3332 } 3333 EXPORT_SYMBOL(netif_schedule_queue); 3334 3335 void netif_tx_wake_queue(struct netdev_queue *dev_queue) 3336 { 3337 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) { 3338 struct Qdisc *q; 3339 3340 rcu_read_lock(); 3341 q = rcu_dereference(dev_queue->qdisc); 3342 __netif_schedule(q); 3343 rcu_read_unlock(); 3344 } 3345 } 3346 EXPORT_SYMBOL(netif_tx_wake_queue); 3347 3348 void dev_kfree_skb_irq_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3349 { 3350 unsigned long flags; 3351 3352 if (unlikely(!skb)) 3353 return; 3354 3355 if (likely(refcount_read(&skb->users) == 1)) { 3356 smp_rmb(); 3357 refcount_set(&skb->users, 0); 3358 } else if (likely(!refcount_dec_and_test(&skb->users))) { 3359 return; 3360 } 3361 get_kfree_skb_cb(skb)->reason = reason; 3362 local_irq_save(flags); 3363 skb->next = __this_cpu_read(softnet_data.completion_queue); 3364 __this_cpu_write(softnet_data.completion_queue, skb); 3365 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3366 local_irq_restore(flags); 3367 } 3368 EXPORT_SYMBOL(dev_kfree_skb_irq_reason); 3369 3370 void dev_kfree_skb_any_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3371 { 3372 if (in_hardirq() || irqs_disabled()) 3373 dev_kfree_skb_irq_reason(skb, reason); 3374 else 3375 kfree_skb_reason(skb, reason); 3376 } 3377 EXPORT_SYMBOL(dev_kfree_skb_any_reason); 3378 3379 3380 /** 3381 * netif_device_detach - mark device as removed 3382 * @dev: network device 3383 * 3384 * Mark device as removed from system and therefore no longer available. 3385 */ 3386 void netif_device_detach(struct net_device *dev) 3387 { 3388 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) && 3389 netif_running(dev)) { 3390 netif_tx_stop_all_queues(dev); 3391 } 3392 } 3393 EXPORT_SYMBOL(netif_device_detach); 3394 3395 /** 3396 * netif_device_attach - mark device as attached 3397 * @dev: network device 3398 * 3399 * Mark device as attached from system and restart if needed. 3400 */ 3401 void netif_device_attach(struct net_device *dev) 3402 { 3403 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) && 3404 netif_running(dev)) { 3405 netif_tx_wake_all_queues(dev); 3406 netdev_watchdog_up(dev); 3407 } 3408 } 3409 EXPORT_SYMBOL(netif_device_attach); 3410 3411 /* 3412 * Returns a Tx hash based on the given packet descriptor a Tx queues' number 3413 * to be used as a distribution range. 3414 */ 3415 static u16 skb_tx_hash(const struct net_device *dev, 3416 const struct net_device *sb_dev, 3417 struct sk_buff *skb) 3418 { 3419 u32 hash; 3420 u16 qoffset = 0; 3421 u16 qcount = dev->real_num_tx_queues; 3422 3423 if (dev->num_tc) { 3424 u8 tc = netdev_get_prio_tc_map(dev, skb->priority); 3425 3426 qoffset = sb_dev->tc_to_txq[tc].offset; 3427 qcount = sb_dev->tc_to_txq[tc].count; 3428 if (unlikely(!qcount)) { 3429 net_warn_ratelimited("%s: invalid qcount, qoffset %u for tc %u\n", 3430 sb_dev->name, qoffset, tc); 3431 qoffset = 0; 3432 qcount = dev->real_num_tx_queues; 3433 } 3434 } 3435 3436 if (skb_rx_queue_recorded(skb)) { 3437 DEBUG_NET_WARN_ON_ONCE(qcount == 0); 3438 hash = skb_get_rx_queue(skb); 3439 if (hash >= qoffset) 3440 hash -= qoffset; 3441 while (unlikely(hash >= qcount)) 3442 hash -= qcount; 3443 return hash + qoffset; 3444 } 3445 3446 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset; 3447 } 3448 3449 void skb_warn_bad_offload(const struct sk_buff *skb) 3450 { 3451 static const netdev_features_t null_features; 3452 struct net_device *dev = skb->dev; 3453 const char *name = ""; 3454 3455 if (!net_ratelimit()) 3456 return; 3457 3458 if (dev) { 3459 if (dev->dev.parent) 3460 name = dev_driver_string(dev->dev.parent); 3461 else 3462 name = netdev_name(dev); 3463 } 3464 skb_dump(KERN_WARNING, skb, false); 3465 WARN(1, "%s: caps=(%pNF, %pNF)\n", 3466 name, dev ? &dev->features : &null_features, 3467 skb->sk ? &skb->sk->sk_route_caps : &null_features); 3468 } 3469 3470 /* 3471 * Invalidate hardware checksum when packet is to be mangled, and 3472 * complete checksum manually on outgoing path. 3473 */ 3474 int skb_checksum_help(struct sk_buff *skb) 3475 { 3476 __wsum csum; 3477 int ret = 0, offset; 3478 3479 if (skb->ip_summed == CHECKSUM_COMPLETE) 3480 goto out_set_summed; 3481 3482 if (unlikely(skb_is_gso(skb))) { 3483 skb_warn_bad_offload(skb); 3484 return -EINVAL; 3485 } 3486 3487 if (!skb_frags_readable(skb)) { 3488 return -EFAULT; 3489 } 3490 3491 /* Before computing a checksum, we should make sure no frag could 3492 * be modified by an external entity : checksum could be wrong. 3493 */ 3494 if (skb_has_shared_frag(skb)) { 3495 ret = __skb_linearize(skb); 3496 if (ret) 3497 goto out; 3498 } 3499 3500 offset = skb_checksum_start_offset(skb); 3501 ret = -EINVAL; 3502 if (unlikely(offset >= skb_headlen(skb))) { 3503 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3504 WARN_ONCE(true, "offset (%d) >= skb_headlen() (%u)\n", 3505 offset, skb_headlen(skb)); 3506 goto out; 3507 } 3508 csum = skb_checksum(skb, offset, skb->len - offset, 0); 3509 3510 offset += skb->csum_offset; 3511 if (unlikely(offset + sizeof(__sum16) > skb_headlen(skb))) { 3512 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3513 WARN_ONCE(true, "offset+2 (%zu) > skb_headlen() (%u)\n", 3514 offset + sizeof(__sum16), skb_headlen(skb)); 3515 goto out; 3516 } 3517 ret = skb_ensure_writable(skb, offset + sizeof(__sum16)); 3518 if (ret) 3519 goto out; 3520 3521 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0; 3522 out_set_summed: 3523 skb->ip_summed = CHECKSUM_NONE; 3524 out: 3525 return ret; 3526 } 3527 EXPORT_SYMBOL(skb_checksum_help); 3528 3529 int skb_crc32c_csum_help(struct sk_buff *skb) 3530 { 3531 __le32 crc32c_csum; 3532 int ret = 0, offset, start; 3533 3534 if (skb->ip_summed != CHECKSUM_PARTIAL) 3535 goto out; 3536 3537 if (unlikely(skb_is_gso(skb))) 3538 goto out; 3539 3540 /* Before computing a checksum, we should make sure no frag could 3541 * be modified by an external entity : checksum could be wrong. 3542 */ 3543 if (unlikely(skb_has_shared_frag(skb))) { 3544 ret = __skb_linearize(skb); 3545 if (ret) 3546 goto out; 3547 } 3548 start = skb_checksum_start_offset(skb); 3549 offset = start + offsetof(struct sctphdr, checksum); 3550 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) { 3551 ret = -EINVAL; 3552 goto out; 3553 } 3554 3555 ret = skb_ensure_writable(skb, offset + sizeof(__le32)); 3556 if (ret) 3557 goto out; 3558 3559 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start, 3560 skb->len - start, ~(__u32)0, 3561 crc32c_csum_stub)); 3562 *(__le32 *)(skb->data + offset) = crc32c_csum; 3563 skb_reset_csum_not_inet(skb); 3564 out: 3565 return ret; 3566 } 3567 EXPORT_SYMBOL(skb_crc32c_csum_help); 3568 3569 __be16 skb_network_protocol(struct sk_buff *skb, int *depth) 3570 { 3571 __be16 type = skb->protocol; 3572 3573 /* Tunnel gso handlers can set protocol to ethernet. */ 3574 if (type == htons(ETH_P_TEB)) { 3575 struct ethhdr *eth; 3576 3577 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr)))) 3578 return 0; 3579 3580 eth = (struct ethhdr *)skb->data; 3581 type = eth->h_proto; 3582 } 3583 3584 return vlan_get_protocol_and_depth(skb, type, depth); 3585 } 3586 3587 3588 /* Take action when hardware reception checksum errors are detected. */ 3589 #ifdef CONFIG_BUG 3590 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3591 { 3592 netdev_err(dev, "hw csum failure\n"); 3593 skb_dump(KERN_ERR, skb, true); 3594 dump_stack(); 3595 } 3596 3597 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3598 { 3599 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb); 3600 } 3601 EXPORT_SYMBOL(netdev_rx_csum_fault); 3602 #endif 3603 3604 /* XXX: check that highmem exists at all on the given machine. */ 3605 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb) 3606 { 3607 #ifdef CONFIG_HIGHMEM 3608 int i; 3609 3610 if (!(dev->features & NETIF_F_HIGHDMA)) { 3611 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 3612 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 3613 struct page *page = skb_frag_page(frag); 3614 3615 if (page && PageHighMem(page)) 3616 return 1; 3617 } 3618 } 3619 #endif 3620 return 0; 3621 } 3622 3623 /* If MPLS offload request, verify we are testing hardware MPLS features 3624 * instead of standard features for the netdev. 3625 */ 3626 #if IS_ENABLED(CONFIG_NET_MPLS_GSO) 3627 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3628 netdev_features_t features, 3629 __be16 type) 3630 { 3631 if (eth_p_mpls(type)) 3632 features &= skb->dev->mpls_features; 3633 3634 return features; 3635 } 3636 #else 3637 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3638 netdev_features_t features, 3639 __be16 type) 3640 { 3641 return features; 3642 } 3643 #endif 3644 3645 static netdev_features_t harmonize_features(struct sk_buff *skb, 3646 netdev_features_t features) 3647 { 3648 __be16 type; 3649 3650 type = skb_network_protocol(skb, NULL); 3651 features = net_mpls_features(skb, features, type); 3652 3653 if (skb->ip_summed != CHECKSUM_NONE && 3654 !can_checksum_protocol(features, type)) { 3655 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3656 } 3657 if (illegal_highdma(skb->dev, skb)) 3658 features &= ~NETIF_F_SG; 3659 3660 return features; 3661 } 3662 3663 netdev_features_t passthru_features_check(struct sk_buff *skb, 3664 struct net_device *dev, 3665 netdev_features_t features) 3666 { 3667 return features; 3668 } 3669 EXPORT_SYMBOL(passthru_features_check); 3670 3671 static netdev_features_t dflt_features_check(struct sk_buff *skb, 3672 struct net_device *dev, 3673 netdev_features_t features) 3674 { 3675 return vlan_features_check(skb, features); 3676 } 3677 3678 static netdev_features_t gso_features_check(const struct sk_buff *skb, 3679 struct net_device *dev, 3680 netdev_features_t features) 3681 { 3682 u16 gso_segs = skb_shinfo(skb)->gso_segs; 3683 3684 if (gso_segs > READ_ONCE(dev->gso_max_segs)) 3685 return features & ~NETIF_F_GSO_MASK; 3686 3687 if (unlikely(skb->len >= netif_get_gso_max_size(dev, skb))) 3688 return features & ~NETIF_F_GSO_MASK; 3689 3690 if (!skb_shinfo(skb)->gso_type) { 3691 skb_warn_bad_offload(skb); 3692 return features & ~NETIF_F_GSO_MASK; 3693 } 3694 3695 /* Support for GSO partial features requires software 3696 * intervention before we can actually process the packets 3697 * so we need to strip support for any partial features now 3698 * and we can pull them back in after we have partially 3699 * segmented the frame. 3700 */ 3701 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)) 3702 features &= ~dev->gso_partial_features; 3703 3704 /* Make sure to clear the IPv4 ID mangling feature if the 3705 * IPv4 header has the potential to be fragmented. 3706 */ 3707 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) { 3708 struct iphdr *iph = skb->encapsulation ? 3709 inner_ip_hdr(skb) : ip_hdr(skb); 3710 3711 if (!(iph->frag_off & htons(IP_DF))) 3712 features &= ~NETIF_F_TSO_MANGLEID; 3713 } 3714 3715 return features; 3716 } 3717 3718 netdev_features_t netif_skb_features(struct sk_buff *skb) 3719 { 3720 struct net_device *dev = skb->dev; 3721 netdev_features_t features = dev->features; 3722 3723 if (skb_is_gso(skb)) 3724 features = gso_features_check(skb, dev, features); 3725 3726 /* If encapsulation offload request, verify we are testing 3727 * hardware encapsulation features instead of standard 3728 * features for the netdev 3729 */ 3730 if (skb->encapsulation) 3731 features &= dev->hw_enc_features; 3732 3733 if (skb_vlan_tagged(skb)) 3734 features = netdev_intersect_features(features, 3735 dev->vlan_features | 3736 NETIF_F_HW_VLAN_CTAG_TX | 3737 NETIF_F_HW_VLAN_STAG_TX); 3738 3739 if (dev->netdev_ops->ndo_features_check) 3740 features &= dev->netdev_ops->ndo_features_check(skb, dev, 3741 features); 3742 else 3743 features &= dflt_features_check(skb, dev, features); 3744 3745 return harmonize_features(skb, features); 3746 } 3747 EXPORT_SYMBOL(netif_skb_features); 3748 3749 static int xmit_one(struct sk_buff *skb, struct net_device *dev, 3750 struct netdev_queue *txq, bool more) 3751 { 3752 unsigned int len; 3753 int rc; 3754 3755 if (dev_nit_active(dev)) 3756 dev_queue_xmit_nit(skb, dev); 3757 3758 len = skb->len; 3759 trace_net_dev_start_xmit(skb, dev); 3760 rc = netdev_start_xmit(skb, dev, txq, more); 3761 trace_net_dev_xmit(skb, rc, dev, len); 3762 3763 return rc; 3764 } 3765 3766 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev, 3767 struct netdev_queue *txq, int *ret) 3768 { 3769 struct sk_buff *skb = first; 3770 int rc = NETDEV_TX_OK; 3771 3772 while (skb) { 3773 struct sk_buff *next = skb->next; 3774 3775 skb_mark_not_on_list(skb); 3776 rc = xmit_one(skb, dev, txq, next != NULL); 3777 if (unlikely(!dev_xmit_complete(rc))) { 3778 skb->next = next; 3779 goto out; 3780 } 3781 3782 skb = next; 3783 if (netif_tx_queue_stopped(txq) && skb) { 3784 rc = NETDEV_TX_BUSY; 3785 break; 3786 } 3787 } 3788 3789 out: 3790 *ret = rc; 3791 return skb; 3792 } 3793 3794 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb, 3795 netdev_features_t features) 3796 { 3797 if (skb_vlan_tag_present(skb) && 3798 !vlan_hw_offload_capable(features, skb->vlan_proto)) 3799 skb = __vlan_hwaccel_push_inside(skb); 3800 return skb; 3801 } 3802 3803 int skb_csum_hwoffload_help(struct sk_buff *skb, 3804 const netdev_features_t features) 3805 { 3806 if (unlikely(skb_csum_is_sctp(skb))) 3807 return !!(features & NETIF_F_SCTP_CRC) ? 0 : 3808 skb_crc32c_csum_help(skb); 3809 3810 if (features & NETIF_F_HW_CSUM) 3811 return 0; 3812 3813 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) { 3814 if (vlan_get_protocol(skb) == htons(ETH_P_IPV6) && 3815 skb_network_header_len(skb) != sizeof(struct ipv6hdr) && 3816 !ipv6_has_hopopt_jumbo(skb)) 3817 goto sw_checksum; 3818 3819 switch (skb->csum_offset) { 3820 case offsetof(struct tcphdr, check): 3821 case offsetof(struct udphdr, check): 3822 return 0; 3823 } 3824 } 3825 3826 sw_checksum: 3827 return skb_checksum_help(skb); 3828 } 3829 EXPORT_SYMBOL(skb_csum_hwoffload_help); 3830 3831 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again) 3832 { 3833 netdev_features_t features; 3834 3835 if (!skb_frags_readable(skb)) 3836 goto out_kfree_skb; 3837 3838 features = netif_skb_features(skb); 3839 skb = validate_xmit_vlan(skb, features); 3840 if (unlikely(!skb)) 3841 goto out_null; 3842 3843 skb = sk_validate_xmit_skb(skb, dev); 3844 if (unlikely(!skb)) 3845 goto out_null; 3846 3847 if (netif_needs_gso(skb, features)) { 3848 struct sk_buff *segs; 3849 3850 segs = skb_gso_segment(skb, features); 3851 if (IS_ERR(segs)) { 3852 goto out_kfree_skb; 3853 } else if (segs) { 3854 consume_skb(skb); 3855 skb = segs; 3856 } 3857 } else { 3858 if (skb_needs_linearize(skb, features) && 3859 __skb_linearize(skb)) 3860 goto out_kfree_skb; 3861 3862 /* If packet is not checksummed and device does not 3863 * support checksumming for this protocol, complete 3864 * checksumming here. 3865 */ 3866 if (skb->ip_summed == CHECKSUM_PARTIAL) { 3867 if (skb->encapsulation) 3868 skb_set_inner_transport_header(skb, 3869 skb_checksum_start_offset(skb)); 3870 else 3871 skb_set_transport_header(skb, 3872 skb_checksum_start_offset(skb)); 3873 if (skb_csum_hwoffload_help(skb, features)) 3874 goto out_kfree_skb; 3875 } 3876 } 3877 3878 skb = validate_xmit_xfrm(skb, features, again); 3879 3880 return skb; 3881 3882 out_kfree_skb: 3883 kfree_skb(skb); 3884 out_null: 3885 dev_core_stats_tx_dropped_inc(dev); 3886 return NULL; 3887 } 3888 3889 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again) 3890 { 3891 struct sk_buff *next, *head = NULL, *tail; 3892 3893 for (; skb != NULL; skb = next) { 3894 next = skb->next; 3895 skb_mark_not_on_list(skb); 3896 3897 /* in case skb won't be segmented, point to itself */ 3898 skb->prev = skb; 3899 3900 skb = validate_xmit_skb(skb, dev, again); 3901 if (!skb) 3902 continue; 3903 3904 if (!head) 3905 head = skb; 3906 else 3907 tail->next = skb; 3908 /* If skb was segmented, skb->prev points to 3909 * the last segment. If not, it still contains skb. 3910 */ 3911 tail = skb->prev; 3912 } 3913 return head; 3914 } 3915 EXPORT_SYMBOL_GPL(validate_xmit_skb_list); 3916 3917 static void qdisc_pkt_len_init(struct sk_buff *skb) 3918 { 3919 const struct skb_shared_info *shinfo = skb_shinfo(skb); 3920 3921 qdisc_skb_cb(skb)->pkt_len = skb->len; 3922 3923 /* To get more precise estimation of bytes sent on wire, 3924 * we add to pkt_len the headers size of all segments 3925 */ 3926 if (shinfo->gso_size && skb_transport_header_was_set(skb)) { 3927 u16 gso_segs = shinfo->gso_segs; 3928 unsigned int hdr_len; 3929 3930 /* mac layer + network layer */ 3931 hdr_len = skb_transport_offset(skb); 3932 3933 /* + transport layer */ 3934 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) { 3935 const struct tcphdr *th; 3936 struct tcphdr _tcphdr; 3937 3938 th = skb_header_pointer(skb, hdr_len, 3939 sizeof(_tcphdr), &_tcphdr); 3940 if (likely(th)) 3941 hdr_len += __tcp_hdrlen(th); 3942 } else if (shinfo->gso_type & SKB_GSO_UDP_L4) { 3943 struct udphdr _udphdr; 3944 3945 if (skb_header_pointer(skb, hdr_len, 3946 sizeof(_udphdr), &_udphdr)) 3947 hdr_len += sizeof(struct udphdr); 3948 } 3949 3950 if (unlikely(shinfo->gso_type & SKB_GSO_DODGY)) { 3951 int payload = skb->len - hdr_len; 3952 3953 /* Malicious packet. */ 3954 if (payload <= 0) 3955 return; 3956 gso_segs = DIV_ROUND_UP(payload, shinfo->gso_size); 3957 } 3958 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len; 3959 } 3960 } 3961 3962 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q, 3963 struct sk_buff **to_free, 3964 struct netdev_queue *txq) 3965 { 3966 int rc; 3967 3968 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK; 3969 if (rc == NET_XMIT_SUCCESS) 3970 trace_qdisc_enqueue(q, txq, skb); 3971 return rc; 3972 } 3973 3974 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q, 3975 struct net_device *dev, 3976 struct netdev_queue *txq) 3977 { 3978 spinlock_t *root_lock = qdisc_lock(q); 3979 struct sk_buff *to_free = NULL; 3980 bool contended; 3981 int rc; 3982 3983 qdisc_calculate_pkt_len(skb, q); 3984 3985 tcf_set_drop_reason(skb, SKB_DROP_REASON_QDISC_DROP); 3986 3987 if (q->flags & TCQ_F_NOLOCK) { 3988 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) && 3989 qdisc_run_begin(q)) { 3990 /* Retest nolock_qdisc_is_empty() within the protection 3991 * of q->seqlock to protect from racing with requeuing. 3992 */ 3993 if (unlikely(!nolock_qdisc_is_empty(q))) { 3994 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 3995 __qdisc_run(q); 3996 qdisc_run_end(q); 3997 3998 goto no_lock_out; 3999 } 4000 4001 qdisc_bstats_cpu_update(q, skb); 4002 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) && 4003 !nolock_qdisc_is_empty(q)) 4004 __qdisc_run(q); 4005 4006 qdisc_run_end(q); 4007 return NET_XMIT_SUCCESS; 4008 } 4009 4010 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4011 qdisc_run(q); 4012 4013 no_lock_out: 4014 if (unlikely(to_free)) 4015 kfree_skb_list_reason(to_free, 4016 tcf_get_drop_reason(to_free)); 4017 return rc; 4018 } 4019 4020 if (unlikely(READ_ONCE(q->owner) == smp_processor_id())) { 4021 kfree_skb_reason(skb, SKB_DROP_REASON_TC_RECLASSIFY_LOOP); 4022 return NET_XMIT_DROP; 4023 } 4024 /* 4025 * Heuristic to force contended enqueues to serialize on a 4026 * separate lock before trying to get qdisc main lock. 4027 * This permits qdisc->running owner to get the lock more 4028 * often and dequeue packets faster. 4029 * On PREEMPT_RT it is possible to preempt the qdisc owner during xmit 4030 * and then other tasks will only enqueue packets. The packets will be 4031 * sent after the qdisc owner is scheduled again. To prevent this 4032 * scenario the task always serialize on the lock. 4033 */ 4034 contended = qdisc_is_running(q) || IS_ENABLED(CONFIG_PREEMPT_RT); 4035 if (unlikely(contended)) 4036 spin_lock(&q->busylock); 4037 4038 spin_lock(root_lock); 4039 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) { 4040 __qdisc_drop(skb, &to_free); 4041 rc = NET_XMIT_DROP; 4042 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && 4043 qdisc_run_begin(q)) { 4044 /* 4045 * This is a work-conserving queue; there are no old skbs 4046 * waiting to be sent out; and the qdisc is not running - 4047 * xmit the skb directly. 4048 */ 4049 4050 qdisc_bstats_update(q, skb); 4051 4052 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) { 4053 if (unlikely(contended)) { 4054 spin_unlock(&q->busylock); 4055 contended = false; 4056 } 4057 __qdisc_run(q); 4058 } 4059 4060 qdisc_run_end(q); 4061 rc = NET_XMIT_SUCCESS; 4062 } else { 4063 WRITE_ONCE(q->owner, smp_processor_id()); 4064 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4065 WRITE_ONCE(q->owner, -1); 4066 if (qdisc_run_begin(q)) { 4067 if (unlikely(contended)) { 4068 spin_unlock(&q->busylock); 4069 contended = false; 4070 } 4071 __qdisc_run(q); 4072 qdisc_run_end(q); 4073 } 4074 } 4075 spin_unlock(root_lock); 4076 if (unlikely(to_free)) 4077 kfree_skb_list_reason(to_free, 4078 tcf_get_drop_reason(to_free)); 4079 if (unlikely(contended)) 4080 spin_unlock(&q->busylock); 4081 return rc; 4082 } 4083 4084 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO) 4085 static void skb_update_prio(struct sk_buff *skb) 4086 { 4087 const struct netprio_map *map; 4088 const struct sock *sk; 4089 unsigned int prioidx; 4090 4091 if (skb->priority) 4092 return; 4093 map = rcu_dereference_bh(skb->dev->priomap); 4094 if (!map) 4095 return; 4096 sk = skb_to_full_sk(skb); 4097 if (!sk) 4098 return; 4099 4100 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data); 4101 4102 if (prioidx < map->priomap_len) 4103 skb->priority = map->priomap[prioidx]; 4104 } 4105 #else 4106 #define skb_update_prio(skb) 4107 #endif 4108 4109 /** 4110 * dev_loopback_xmit - loop back @skb 4111 * @net: network namespace this loopback is happening in 4112 * @sk: sk needed to be a netfilter okfn 4113 * @skb: buffer to transmit 4114 */ 4115 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb) 4116 { 4117 skb_reset_mac_header(skb); 4118 __skb_pull(skb, skb_network_offset(skb)); 4119 skb->pkt_type = PACKET_LOOPBACK; 4120 if (skb->ip_summed == CHECKSUM_NONE) 4121 skb->ip_summed = CHECKSUM_UNNECESSARY; 4122 DEBUG_NET_WARN_ON_ONCE(!skb_dst(skb)); 4123 skb_dst_force(skb); 4124 netif_rx(skb); 4125 return 0; 4126 } 4127 EXPORT_SYMBOL(dev_loopback_xmit); 4128 4129 #ifdef CONFIG_NET_EGRESS 4130 static struct netdev_queue * 4131 netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb) 4132 { 4133 int qm = skb_get_queue_mapping(skb); 4134 4135 return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm)); 4136 } 4137 4138 #ifndef CONFIG_PREEMPT_RT 4139 static bool netdev_xmit_txqueue_skipped(void) 4140 { 4141 return __this_cpu_read(softnet_data.xmit.skip_txqueue); 4142 } 4143 4144 void netdev_xmit_skip_txqueue(bool skip) 4145 { 4146 __this_cpu_write(softnet_data.xmit.skip_txqueue, skip); 4147 } 4148 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4149 4150 #else 4151 static bool netdev_xmit_txqueue_skipped(void) 4152 { 4153 return current->net_xmit.skip_txqueue; 4154 } 4155 4156 void netdev_xmit_skip_txqueue(bool skip) 4157 { 4158 current->net_xmit.skip_txqueue = skip; 4159 } 4160 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4161 #endif 4162 #endif /* CONFIG_NET_EGRESS */ 4163 4164 #ifdef CONFIG_NET_XGRESS 4165 static int tc_run(struct tcx_entry *entry, struct sk_buff *skb, 4166 enum skb_drop_reason *drop_reason) 4167 { 4168 int ret = TC_ACT_UNSPEC; 4169 #ifdef CONFIG_NET_CLS_ACT 4170 struct mini_Qdisc *miniq = rcu_dereference_bh(entry->miniq); 4171 struct tcf_result res; 4172 4173 if (!miniq) 4174 return ret; 4175 4176 /* Global bypass */ 4177 if (!static_branch_likely(&tcf_sw_enabled_key)) 4178 return ret; 4179 4180 /* Block-wise bypass */ 4181 if (tcf_block_bypass_sw(miniq->block)) 4182 return ret; 4183 4184 tc_skb_cb(skb)->mru = 0; 4185 tc_skb_cb(skb)->post_ct = false; 4186 tcf_set_drop_reason(skb, *drop_reason); 4187 4188 mini_qdisc_bstats_cpu_update(miniq, skb); 4189 ret = tcf_classify(skb, miniq->block, miniq->filter_list, &res, false); 4190 /* Only tcf related quirks below. */ 4191 switch (ret) { 4192 case TC_ACT_SHOT: 4193 *drop_reason = tcf_get_drop_reason(skb); 4194 mini_qdisc_qstats_cpu_drop(miniq); 4195 break; 4196 case TC_ACT_OK: 4197 case TC_ACT_RECLASSIFY: 4198 skb->tc_index = TC_H_MIN(res.classid); 4199 break; 4200 } 4201 #endif /* CONFIG_NET_CLS_ACT */ 4202 return ret; 4203 } 4204 4205 static DEFINE_STATIC_KEY_FALSE(tcx_needed_key); 4206 4207 void tcx_inc(void) 4208 { 4209 static_branch_inc(&tcx_needed_key); 4210 } 4211 4212 void tcx_dec(void) 4213 { 4214 static_branch_dec(&tcx_needed_key); 4215 } 4216 4217 static __always_inline enum tcx_action_base 4218 tcx_run(const struct bpf_mprog_entry *entry, struct sk_buff *skb, 4219 const bool needs_mac) 4220 { 4221 const struct bpf_mprog_fp *fp; 4222 const struct bpf_prog *prog; 4223 int ret = TCX_NEXT; 4224 4225 if (needs_mac) 4226 __skb_push(skb, skb->mac_len); 4227 bpf_mprog_foreach_prog(entry, fp, prog) { 4228 bpf_compute_data_pointers(skb); 4229 ret = bpf_prog_run(prog, skb); 4230 if (ret != TCX_NEXT) 4231 break; 4232 } 4233 if (needs_mac) 4234 __skb_pull(skb, skb->mac_len); 4235 return tcx_action_code(skb, ret); 4236 } 4237 4238 static __always_inline struct sk_buff * 4239 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4240 struct net_device *orig_dev, bool *another) 4241 { 4242 struct bpf_mprog_entry *entry = rcu_dereference_bh(skb->dev->tcx_ingress); 4243 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_INGRESS; 4244 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4245 int sch_ret; 4246 4247 if (!entry) 4248 return skb; 4249 4250 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4251 if (*pt_prev) { 4252 *ret = deliver_skb(skb, *pt_prev, orig_dev); 4253 *pt_prev = NULL; 4254 } 4255 4256 qdisc_skb_cb(skb)->pkt_len = skb->len; 4257 tcx_set_ingress(skb, true); 4258 4259 if (static_branch_unlikely(&tcx_needed_key)) { 4260 sch_ret = tcx_run(entry, skb, true); 4261 if (sch_ret != TC_ACT_UNSPEC) 4262 goto ingress_verdict; 4263 } 4264 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4265 ingress_verdict: 4266 switch (sch_ret) { 4267 case TC_ACT_REDIRECT: 4268 /* skb_mac_header check was done by BPF, so we can safely 4269 * push the L2 header back before redirecting to another 4270 * netdev. 4271 */ 4272 __skb_push(skb, skb->mac_len); 4273 if (skb_do_redirect(skb) == -EAGAIN) { 4274 __skb_pull(skb, skb->mac_len); 4275 *another = true; 4276 break; 4277 } 4278 *ret = NET_RX_SUCCESS; 4279 bpf_net_ctx_clear(bpf_net_ctx); 4280 return NULL; 4281 case TC_ACT_SHOT: 4282 kfree_skb_reason(skb, drop_reason); 4283 *ret = NET_RX_DROP; 4284 bpf_net_ctx_clear(bpf_net_ctx); 4285 return NULL; 4286 /* used by tc_run */ 4287 case TC_ACT_STOLEN: 4288 case TC_ACT_QUEUED: 4289 case TC_ACT_TRAP: 4290 consume_skb(skb); 4291 fallthrough; 4292 case TC_ACT_CONSUMED: 4293 *ret = NET_RX_SUCCESS; 4294 bpf_net_ctx_clear(bpf_net_ctx); 4295 return NULL; 4296 } 4297 bpf_net_ctx_clear(bpf_net_ctx); 4298 4299 return skb; 4300 } 4301 4302 static __always_inline struct sk_buff * 4303 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4304 { 4305 struct bpf_mprog_entry *entry = rcu_dereference_bh(dev->tcx_egress); 4306 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_EGRESS; 4307 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4308 int sch_ret; 4309 4310 if (!entry) 4311 return skb; 4312 4313 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4314 4315 /* qdisc_skb_cb(skb)->pkt_len & tcx_set_ingress() was 4316 * already set by the caller. 4317 */ 4318 if (static_branch_unlikely(&tcx_needed_key)) { 4319 sch_ret = tcx_run(entry, skb, false); 4320 if (sch_ret != TC_ACT_UNSPEC) 4321 goto egress_verdict; 4322 } 4323 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4324 egress_verdict: 4325 switch (sch_ret) { 4326 case TC_ACT_REDIRECT: 4327 /* No need to push/pop skb's mac_header here on egress! */ 4328 skb_do_redirect(skb); 4329 *ret = NET_XMIT_SUCCESS; 4330 bpf_net_ctx_clear(bpf_net_ctx); 4331 return NULL; 4332 case TC_ACT_SHOT: 4333 kfree_skb_reason(skb, drop_reason); 4334 *ret = NET_XMIT_DROP; 4335 bpf_net_ctx_clear(bpf_net_ctx); 4336 return NULL; 4337 /* used by tc_run */ 4338 case TC_ACT_STOLEN: 4339 case TC_ACT_QUEUED: 4340 case TC_ACT_TRAP: 4341 consume_skb(skb); 4342 fallthrough; 4343 case TC_ACT_CONSUMED: 4344 *ret = NET_XMIT_SUCCESS; 4345 bpf_net_ctx_clear(bpf_net_ctx); 4346 return NULL; 4347 } 4348 bpf_net_ctx_clear(bpf_net_ctx); 4349 4350 return skb; 4351 } 4352 #else 4353 static __always_inline struct sk_buff * 4354 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4355 struct net_device *orig_dev, bool *another) 4356 { 4357 return skb; 4358 } 4359 4360 static __always_inline struct sk_buff * 4361 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4362 { 4363 return skb; 4364 } 4365 #endif /* CONFIG_NET_XGRESS */ 4366 4367 #ifdef CONFIG_XPS 4368 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb, 4369 struct xps_dev_maps *dev_maps, unsigned int tci) 4370 { 4371 int tc = netdev_get_prio_tc_map(dev, skb->priority); 4372 struct xps_map *map; 4373 int queue_index = -1; 4374 4375 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids) 4376 return queue_index; 4377 4378 tci *= dev_maps->num_tc; 4379 tci += tc; 4380 4381 map = rcu_dereference(dev_maps->attr_map[tci]); 4382 if (map) { 4383 if (map->len == 1) 4384 queue_index = map->queues[0]; 4385 else 4386 queue_index = map->queues[reciprocal_scale( 4387 skb_get_hash(skb), map->len)]; 4388 if (unlikely(queue_index >= dev->real_num_tx_queues)) 4389 queue_index = -1; 4390 } 4391 return queue_index; 4392 } 4393 #endif 4394 4395 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev, 4396 struct sk_buff *skb) 4397 { 4398 #ifdef CONFIG_XPS 4399 struct xps_dev_maps *dev_maps; 4400 struct sock *sk = skb->sk; 4401 int queue_index = -1; 4402 4403 if (!static_key_false(&xps_needed)) 4404 return -1; 4405 4406 rcu_read_lock(); 4407 if (!static_key_false(&xps_rxqs_needed)) 4408 goto get_cpus_map; 4409 4410 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]); 4411 if (dev_maps) { 4412 int tci = sk_rx_queue_get(sk); 4413 4414 if (tci >= 0) 4415 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4416 tci); 4417 } 4418 4419 get_cpus_map: 4420 if (queue_index < 0) { 4421 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]); 4422 if (dev_maps) { 4423 unsigned int tci = skb->sender_cpu - 1; 4424 4425 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4426 tci); 4427 } 4428 } 4429 rcu_read_unlock(); 4430 4431 return queue_index; 4432 #else 4433 return -1; 4434 #endif 4435 } 4436 4437 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb, 4438 struct net_device *sb_dev) 4439 { 4440 return 0; 4441 } 4442 EXPORT_SYMBOL(dev_pick_tx_zero); 4443 4444 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb, 4445 struct net_device *sb_dev) 4446 { 4447 struct sock *sk = skb->sk; 4448 int queue_index = sk_tx_queue_get(sk); 4449 4450 sb_dev = sb_dev ? : dev; 4451 4452 if (queue_index < 0 || skb->ooo_okay || 4453 queue_index >= dev->real_num_tx_queues) { 4454 int new_index = get_xps_queue(dev, sb_dev, skb); 4455 4456 if (new_index < 0) 4457 new_index = skb_tx_hash(dev, sb_dev, skb); 4458 4459 if (queue_index != new_index && sk && 4460 sk_fullsock(sk) && 4461 rcu_access_pointer(sk->sk_dst_cache)) 4462 sk_tx_queue_set(sk, new_index); 4463 4464 queue_index = new_index; 4465 } 4466 4467 return queue_index; 4468 } 4469 EXPORT_SYMBOL(netdev_pick_tx); 4470 4471 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev, 4472 struct sk_buff *skb, 4473 struct net_device *sb_dev) 4474 { 4475 int queue_index = 0; 4476 4477 #ifdef CONFIG_XPS 4478 u32 sender_cpu = skb->sender_cpu - 1; 4479 4480 if (sender_cpu >= (u32)NR_CPUS) 4481 skb->sender_cpu = raw_smp_processor_id() + 1; 4482 #endif 4483 4484 if (dev->real_num_tx_queues != 1) { 4485 const struct net_device_ops *ops = dev->netdev_ops; 4486 4487 if (ops->ndo_select_queue) 4488 queue_index = ops->ndo_select_queue(dev, skb, sb_dev); 4489 else 4490 queue_index = netdev_pick_tx(dev, skb, sb_dev); 4491 4492 queue_index = netdev_cap_txqueue(dev, queue_index); 4493 } 4494 4495 skb_set_queue_mapping(skb, queue_index); 4496 return netdev_get_tx_queue(dev, queue_index); 4497 } 4498 4499 /** 4500 * __dev_queue_xmit() - transmit a buffer 4501 * @skb: buffer to transmit 4502 * @sb_dev: suboordinate device used for L2 forwarding offload 4503 * 4504 * Queue a buffer for transmission to a network device. The caller must 4505 * have set the device and priority and built the buffer before calling 4506 * this function. The function can be called from an interrupt. 4507 * 4508 * When calling this method, interrupts MUST be enabled. This is because 4509 * the BH enable code must have IRQs enabled so that it will not deadlock. 4510 * 4511 * Regardless of the return value, the skb is consumed, so it is currently 4512 * difficult to retry a send to this method. (You can bump the ref count 4513 * before sending to hold a reference for retry if you are careful.) 4514 * 4515 * Return: 4516 * * 0 - buffer successfully transmitted 4517 * * positive qdisc return code - NET_XMIT_DROP etc. 4518 * * negative errno - other errors 4519 */ 4520 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev) 4521 { 4522 struct net_device *dev = skb->dev; 4523 struct netdev_queue *txq = NULL; 4524 struct Qdisc *q; 4525 int rc = -ENOMEM; 4526 bool again = false; 4527 4528 skb_reset_mac_header(skb); 4529 skb_assert_len(skb); 4530 4531 if (unlikely(skb_shinfo(skb)->tx_flags & 4532 (SKBTX_SCHED_TSTAMP | SKBTX_BPF))) 4533 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED); 4534 4535 /* Disable soft irqs for various locks below. Also 4536 * stops preemption for RCU. 4537 */ 4538 rcu_read_lock_bh(); 4539 4540 skb_update_prio(skb); 4541 4542 qdisc_pkt_len_init(skb); 4543 tcx_set_ingress(skb, false); 4544 #ifdef CONFIG_NET_EGRESS 4545 if (static_branch_unlikely(&egress_needed_key)) { 4546 if (nf_hook_egress_active()) { 4547 skb = nf_hook_egress(skb, &rc, dev); 4548 if (!skb) 4549 goto out; 4550 } 4551 4552 netdev_xmit_skip_txqueue(false); 4553 4554 nf_skip_egress(skb, true); 4555 skb = sch_handle_egress(skb, &rc, dev); 4556 if (!skb) 4557 goto out; 4558 nf_skip_egress(skb, false); 4559 4560 if (netdev_xmit_txqueue_skipped()) 4561 txq = netdev_tx_queue_mapping(dev, skb); 4562 } 4563 #endif 4564 /* If device/qdisc don't need skb->dst, release it right now while 4565 * its hot in this cpu cache. 4566 */ 4567 if (dev->priv_flags & IFF_XMIT_DST_RELEASE) 4568 skb_dst_drop(skb); 4569 else 4570 skb_dst_force(skb); 4571 4572 if (!txq) 4573 txq = netdev_core_pick_tx(dev, skb, sb_dev); 4574 4575 q = rcu_dereference_bh(txq->qdisc); 4576 4577 trace_net_dev_queue(skb); 4578 if (q->enqueue) { 4579 rc = __dev_xmit_skb(skb, q, dev, txq); 4580 goto out; 4581 } 4582 4583 /* The device has no queue. Common case for software devices: 4584 * loopback, all the sorts of tunnels... 4585 4586 * Really, it is unlikely that netif_tx_lock protection is necessary 4587 * here. (f.e. loopback and IP tunnels are clean ignoring statistics 4588 * counters.) 4589 * However, it is possible, that they rely on protection 4590 * made by us here. 4591 4592 * Check this and shot the lock. It is not prone from deadlocks. 4593 *Either shot noqueue qdisc, it is even simpler 8) 4594 */ 4595 if (dev->flags & IFF_UP) { 4596 int cpu = smp_processor_id(); /* ok because BHs are off */ 4597 4598 /* Other cpus might concurrently change txq->xmit_lock_owner 4599 * to -1 or to their cpu id, but not to our id. 4600 */ 4601 if (READ_ONCE(txq->xmit_lock_owner) != cpu) { 4602 if (dev_xmit_recursion()) 4603 goto recursion_alert; 4604 4605 skb = validate_xmit_skb(skb, dev, &again); 4606 if (!skb) 4607 goto out; 4608 4609 HARD_TX_LOCK(dev, txq, cpu); 4610 4611 if (!netif_xmit_stopped(txq)) { 4612 dev_xmit_recursion_inc(); 4613 skb = dev_hard_start_xmit(skb, dev, txq, &rc); 4614 dev_xmit_recursion_dec(); 4615 if (dev_xmit_complete(rc)) { 4616 HARD_TX_UNLOCK(dev, txq); 4617 goto out; 4618 } 4619 } 4620 HARD_TX_UNLOCK(dev, txq); 4621 net_crit_ratelimited("Virtual device %s asks to queue packet!\n", 4622 dev->name); 4623 } else { 4624 /* Recursion is detected! It is possible, 4625 * unfortunately 4626 */ 4627 recursion_alert: 4628 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n", 4629 dev->name); 4630 } 4631 } 4632 4633 rc = -ENETDOWN; 4634 rcu_read_unlock_bh(); 4635 4636 dev_core_stats_tx_dropped_inc(dev); 4637 kfree_skb_list(skb); 4638 return rc; 4639 out: 4640 rcu_read_unlock_bh(); 4641 return rc; 4642 } 4643 EXPORT_SYMBOL(__dev_queue_xmit); 4644 4645 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id) 4646 { 4647 struct net_device *dev = skb->dev; 4648 struct sk_buff *orig_skb = skb; 4649 struct netdev_queue *txq; 4650 int ret = NETDEV_TX_BUSY; 4651 bool again = false; 4652 4653 if (unlikely(!netif_running(dev) || 4654 !netif_carrier_ok(dev))) 4655 goto drop; 4656 4657 skb = validate_xmit_skb_list(skb, dev, &again); 4658 if (skb != orig_skb) 4659 goto drop; 4660 4661 skb_set_queue_mapping(skb, queue_id); 4662 txq = skb_get_tx_queue(dev, skb); 4663 4664 local_bh_disable(); 4665 4666 dev_xmit_recursion_inc(); 4667 HARD_TX_LOCK(dev, txq, smp_processor_id()); 4668 if (!netif_xmit_frozen_or_drv_stopped(txq)) 4669 ret = netdev_start_xmit(skb, dev, txq, false); 4670 HARD_TX_UNLOCK(dev, txq); 4671 dev_xmit_recursion_dec(); 4672 4673 local_bh_enable(); 4674 return ret; 4675 drop: 4676 dev_core_stats_tx_dropped_inc(dev); 4677 kfree_skb_list(skb); 4678 return NET_XMIT_DROP; 4679 } 4680 EXPORT_SYMBOL(__dev_direct_xmit); 4681 4682 /************************************************************************* 4683 * Receiver routines 4684 *************************************************************************/ 4685 static DEFINE_PER_CPU(struct task_struct *, backlog_napi); 4686 4687 int weight_p __read_mostly = 64; /* old backlog weight */ 4688 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */ 4689 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */ 4690 4691 /* Called with irq disabled */ 4692 static inline void ____napi_schedule(struct softnet_data *sd, 4693 struct napi_struct *napi) 4694 { 4695 struct task_struct *thread; 4696 4697 lockdep_assert_irqs_disabled(); 4698 4699 if (test_bit(NAPI_STATE_THREADED, &napi->state)) { 4700 /* Paired with smp_mb__before_atomic() in 4701 * napi_enable()/dev_set_threaded(). 4702 * Use READ_ONCE() to guarantee a complete 4703 * read on napi->thread. Only call 4704 * wake_up_process() when it's not NULL. 4705 */ 4706 thread = READ_ONCE(napi->thread); 4707 if (thread) { 4708 if (use_backlog_threads() && thread == raw_cpu_read(backlog_napi)) 4709 goto use_local_napi; 4710 4711 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 4712 wake_up_process(thread); 4713 return; 4714 } 4715 } 4716 4717 use_local_napi: 4718 list_add_tail(&napi->poll_list, &sd->poll_list); 4719 WRITE_ONCE(napi->list_owner, smp_processor_id()); 4720 /* If not called from net_rx_action() 4721 * we have to raise NET_RX_SOFTIRQ. 4722 */ 4723 if (!sd->in_net_rx_action) 4724 raise_softirq_irqoff(NET_RX_SOFTIRQ); 4725 } 4726 4727 #ifdef CONFIG_RPS 4728 4729 struct static_key_false rps_needed __read_mostly; 4730 EXPORT_SYMBOL(rps_needed); 4731 struct static_key_false rfs_needed __read_mostly; 4732 EXPORT_SYMBOL(rfs_needed); 4733 4734 static struct rps_dev_flow * 4735 set_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4736 struct rps_dev_flow *rflow, u16 next_cpu) 4737 { 4738 if (next_cpu < nr_cpu_ids) { 4739 u32 head; 4740 #ifdef CONFIG_RFS_ACCEL 4741 struct netdev_rx_queue *rxqueue; 4742 struct rps_dev_flow_table *flow_table; 4743 struct rps_dev_flow *old_rflow; 4744 u16 rxq_index; 4745 u32 flow_id; 4746 int rc; 4747 4748 /* Should we steer this flow to a different hardware queue? */ 4749 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap || 4750 !(dev->features & NETIF_F_NTUPLE)) 4751 goto out; 4752 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu); 4753 if (rxq_index == skb_get_rx_queue(skb)) 4754 goto out; 4755 4756 rxqueue = dev->_rx + rxq_index; 4757 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4758 if (!flow_table) 4759 goto out; 4760 flow_id = skb_get_hash(skb) & flow_table->mask; 4761 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb, 4762 rxq_index, flow_id); 4763 if (rc < 0) 4764 goto out; 4765 old_rflow = rflow; 4766 rflow = &flow_table->flows[flow_id]; 4767 WRITE_ONCE(rflow->filter, rc); 4768 if (old_rflow->filter == rc) 4769 WRITE_ONCE(old_rflow->filter, RPS_NO_FILTER); 4770 out: 4771 #endif 4772 head = READ_ONCE(per_cpu(softnet_data, next_cpu).input_queue_head); 4773 rps_input_queue_tail_save(&rflow->last_qtail, head); 4774 } 4775 4776 WRITE_ONCE(rflow->cpu, next_cpu); 4777 return rflow; 4778 } 4779 4780 /* 4781 * get_rps_cpu is called from netif_receive_skb and returns the target 4782 * CPU from the RPS map of the receiving queue for a given skb. 4783 * rcu_read_lock must be held on entry. 4784 */ 4785 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4786 struct rps_dev_flow **rflowp) 4787 { 4788 const struct rps_sock_flow_table *sock_flow_table; 4789 struct netdev_rx_queue *rxqueue = dev->_rx; 4790 struct rps_dev_flow_table *flow_table; 4791 struct rps_map *map; 4792 int cpu = -1; 4793 u32 tcpu; 4794 u32 hash; 4795 4796 if (skb_rx_queue_recorded(skb)) { 4797 u16 index = skb_get_rx_queue(skb); 4798 4799 if (unlikely(index >= dev->real_num_rx_queues)) { 4800 WARN_ONCE(dev->real_num_rx_queues > 1, 4801 "%s received packet on queue %u, but number " 4802 "of RX queues is %u\n", 4803 dev->name, index, dev->real_num_rx_queues); 4804 goto done; 4805 } 4806 rxqueue += index; 4807 } 4808 4809 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */ 4810 4811 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4812 map = rcu_dereference(rxqueue->rps_map); 4813 if (!flow_table && !map) 4814 goto done; 4815 4816 skb_reset_network_header(skb); 4817 hash = skb_get_hash(skb); 4818 if (!hash) 4819 goto done; 4820 4821 sock_flow_table = rcu_dereference(net_hotdata.rps_sock_flow_table); 4822 if (flow_table && sock_flow_table) { 4823 struct rps_dev_flow *rflow; 4824 u32 next_cpu; 4825 u32 ident; 4826 4827 /* First check into global flow table if there is a match. 4828 * This READ_ONCE() pairs with WRITE_ONCE() from rps_record_sock_flow(). 4829 */ 4830 ident = READ_ONCE(sock_flow_table->ents[hash & sock_flow_table->mask]); 4831 if ((ident ^ hash) & ~net_hotdata.rps_cpu_mask) 4832 goto try_rps; 4833 4834 next_cpu = ident & net_hotdata.rps_cpu_mask; 4835 4836 /* OK, now we know there is a match, 4837 * we can look at the local (per receive queue) flow table 4838 */ 4839 rflow = &flow_table->flows[hash & flow_table->mask]; 4840 tcpu = rflow->cpu; 4841 4842 /* 4843 * If the desired CPU (where last recvmsg was done) is 4844 * different from current CPU (one in the rx-queue flow 4845 * table entry), switch if one of the following holds: 4846 * - Current CPU is unset (>= nr_cpu_ids). 4847 * - Current CPU is offline. 4848 * - The current CPU's queue tail has advanced beyond the 4849 * last packet that was enqueued using this table entry. 4850 * This guarantees that all previous packets for the flow 4851 * have been dequeued, thus preserving in order delivery. 4852 */ 4853 if (unlikely(tcpu != next_cpu) && 4854 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) || 4855 ((int)(READ_ONCE(per_cpu(softnet_data, tcpu).input_queue_head) - 4856 rflow->last_qtail)) >= 0)) { 4857 tcpu = next_cpu; 4858 rflow = set_rps_cpu(dev, skb, rflow, next_cpu); 4859 } 4860 4861 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) { 4862 *rflowp = rflow; 4863 cpu = tcpu; 4864 goto done; 4865 } 4866 } 4867 4868 try_rps: 4869 4870 if (map) { 4871 tcpu = map->cpus[reciprocal_scale(hash, map->len)]; 4872 if (cpu_online(tcpu)) { 4873 cpu = tcpu; 4874 goto done; 4875 } 4876 } 4877 4878 done: 4879 return cpu; 4880 } 4881 4882 #ifdef CONFIG_RFS_ACCEL 4883 4884 /** 4885 * rps_may_expire_flow - check whether an RFS hardware filter may be removed 4886 * @dev: Device on which the filter was set 4887 * @rxq_index: RX queue index 4888 * @flow_id: Flow ID passed to ndo_rx_flow_steer() 4889 * @filter_id: Filter ID returned by ndo_rx_flow_steer() 4890 * 4891 * Drivers that implement ndo_rx_flow_steer() should periodically call 4892 * this function for each installed filter and remove the filters for 4893 * which it returns %true. 4894 */ 4895 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index, 4896 u32 flow_id, u16 filter_id) 4897 { 4898 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index; 4899 struct rps_dev_flow_table *flow_table; 4900 struct rps_dev_flow *rflow; 4901 bool expire = true; 4902 unsigned int cpu; 4903 4904 rcu_read_lock(); 4905 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4906 if (flow_table && flow_id <= flow_table->mask) { 4907 rflow = &flow_table->flows[flow_id]; 4908 cpu = READ_ONCE(rflow->cpu); 4909 if (READ_ONCE(rflow->filter) == filter_id && cpu < nr_cpu_ids && 4910 ((int)(READ_ONCE(per_cpu(softnet_data, cpu).input_queue_head) - 4911 READ_ONCE(rflow->last_qtail)) < 4912 (int)(10 * flow_table->mask))) 4913 expire = false; 4914 } 4915 rcu_read_unlock(); 4916 return expire; 4917 } 4918 EXPORT_SYMBOL(rps_may_expire_flow); 4919 4920 #endif /* CONFIG_RFS_ACCEL */ 4921 4922 /* Called from hardirq (IPI) context */ 4923 static void rps_trigger_softirq(void *data) 4924 { 4925 struct softnet_data *sd = data; 4926 4927 ____napi_schedule(sd, &sd->backlog); 4928 sd->received_rps++; 4929 } 4930 4931 #endif /* CONFIG_RPS */ 4932 4933 /* Called from hardirq (IPI) context */ 4934 static void trigger_rx_softirq(void *data) 4935 { 4936 struct softnet_data *sd = data; 4937 4938 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4939 smp_store_release(&sd->defer_ipi_scheduled, 0); 4940 } 4941 4942 /* 4943 * After we queued a packet into sd->input_pkt_queue, 4944 * we need to make sure this queue is serviced soon. 4945 * 4946 * - If this is another cpu queue, link it to our rps_ipi_list, 4947 * and make sure we will process rps_ipi_list from net_rx_action(). 4948 * 4949 * - If this is our own queue, NAPI schedule our backlog. 4950 * Note that this also raises NET_RX_SOFTIRQ. 4951 */ 4952 static void napi_schedule_rps(struct softnet_data *sd) 4953 { 4954 struct softnet_data *mysd = this_cpu_ptr(&softnet_data); 4955 4956 #ifdef CONFIG_RPS 4957 if (sd != mysd) { 4958 if (use_backlog_threads()) { 4959 __napi_schedule_irqoff(&sd->backlog); 4960 return; 4961 } 4962 4963 sd->rps_ipi_next = mysd->rps_ipi_list; 4964 mysd->rps_ipi_list = sd; 4965 4966 /* If not called from net_rx_action() or napi_threaded_poll() 4967 * we have to raise NET_RX_SOFTIRQ. 4968 */ 4969 if (!mysd->in_net_rx_action && !mysd->in_napi_threaded_poll) 4970 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4971 return; 4972 } 4973 #endif /* CONFIG_RPS */ 4974 __napi_schedule_irqoff(&mysd->backlog); 4975 } 4976 4977 void kick_defer_list_purge(struct softnet_data *sd, unsigned int cpu) 4978 { 4979 unsigned long flags; 4980 4981 if (use_backlog_threads()) { 4982 backlog_lock_irq_save(sd, &flags); 4983 4984 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) 4985 __napi_schedule_irqoff(&sd->backlog); 4986 4987 backlog_unlock_irq_restore(sd, &flags); 4988 4989 } else if (!cmpxchg(&sd->defer_ipi_scheduled, 0, 1)) { 4990 smp_call_function_single_async(cpu, &sd->defer_csd); 4991 } 4992 } 4993 4994 #ifdef CONFIG_NET_FLOW_LIMIT 4995 int netdev_flow_limit_table_len __read_mostly = (1 << 12); 4996 #endif 4997 4998 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen) 4999 { 5000 #ifdef CONFIG_NET_FLOW_LIMIT 5001 struct sd_flow_limit *fl; 5002 struct softnet_data *sd; 5003 unsigned int old_flow, new_flow; 5004 5005 if (qlen < (READ_ONCE(net_hotdata.max_backlog) >> 1)) 5006 return false; 5007 5008 sd = this_cpu_ptr(&softnet_data); 5009 5010 rcu_read_lock(); 5011 fl = rcu_dereference(sd->flow_limit); 5012 if (fl) { 5013 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1); 5014 old_flow = fl->history[fl->history_head]; 5015 fl->history[fl->history_head] = new_flow; 5016 5017 fl->history_head++; 5018 fl->history_head &= FLOW_LIMIT_HISTORY - 1; 5019 5020 if (likely(fl->buckets[old_flow])) 5021 fl->buckets[old_flow]--; 5022 5023 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) { 5024 fl->count++; 5025 rcu_read_unlock(); 5026 return true; 5027 } 5028 } 5029 rcu_read_unlock(); 5030 #endif 5031 return false; 5032 } 5033 5034 /* 5035 * enqueue_to_backlog is called to queue an skb to a per CPU backlog 5036 * queue (may be a remote CPU queue). 5037 */ 5038 static int enqueue_to_backlog(struct sk_buff *skb, int cpu, 5039 unsigned int *qtail) 5040 { 5041 enum skb_drop_reason reason; 5042 struct softnet_data *sd; 5043 unsigned long flags; 5044 unsigned int qlen; 5045 int max_backlog; 5046 u32 tail; 5047 5048 reason = SKB_DROP_REASON_DEV_READY; 5049 if (!netif_running(skb->dev)) 5050 goto bad_dev; 5051 5052 reason = SKB_DROP_REASON_CPU_BACKLOG; 5053 sd = &per_cpu(softnet_data, cpu); 5054 5055 qlen = skb_queue_len_lockless(&sd->input_pkt_queue); 5056 max_backlog = READ_ONCE(net_hotdata.max_backlog); 5057 if (unlikely(qlen > max_backlog)) 5058 goto cpu_backlog_drop; 5059 backlog_lock_irq_save(sd, &flags); 5060 qlen = skb_queue_len(&sd->input_pkt_queue); 5061 if (qlen <= max_backlog && !skb_flow_limit(skb, qlen)) { 5062 if (!qlen) { 5063 /* Schedule NAPI for backlog device. We can use 5064 * non atomic operation as we own the queue lock. 5065 */ 5066 if (!__test_and_set_bit(NAPI_STATE_SCHED, 5067 &sd->backlog.state)) 5068 napi_schedule_rps(sd); 5069 } 5070 __skb_queue_tail(&sd->input_pkt_queue, skb); 5071 tail = rps_input_queue_tail_incr(sd); 5072 backlog_unlock_irq_restore(sd, &flags); 5073 5074 /* save the tail outside of the critical section */ 5075 rps_input_queue_tail_save(qtail, tail); 5076 return NET_RX_SUCCESS; 5077 } 5078 5079 backlog_unlock_irq_restore(sd, &flags); 5080 5081 cpu_backlog_drop: 5082 atomic_inc(&sd->dropped); 5083 bad_dev: 5084 dev_core_stats_rx_dropped_inc(skb->dev); 5085 kfree_skb_reason(skb, reason); 5086 return NET_RX_DROP; 5087 } 5088 5089 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb) 5090 { 5091 struct net_device *dev = skb->dev; 5092 struct netdev_rx_queue *rxqueue; 5093 5094 rxqueue = dev->_rx; 5095 5096 if (skb_rx_queue_recorded(skb)) { 5097 u16 index = skb_get_rx_queue(skb); 5098 5099 if (unlikely(index >= dev->real_num_rx_queues)) { 5100 WARN_ONCE(dev->real_num_rx_queues > 1, 5101 "%s received packet on queue %u, but number " 5102 "of RX queues is %u\n", 5103 dev->name, index, dev->real_num_rx_queues); 5104 5105 return rxqueue; /* Return first rxqueue */ 5106 } 5107 rxqueue += index; 5108 } 5109 return rxqueue; 5110 } 5111 5112 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, 5113 const struct bpf_prog *xdp_prog) 5114 { 5115 void *orig_data, *orig_data_end, *hard_start; 5116 struct netdev_rx_queue *rxqueue; 5117 bool orig_bcast, orig_host; 5118 u32 mac_len, frame_sz; 5119 __be16 orig_eth_type; 5120 struct ethhdr *eth; 5121 u32 metalen, act; 5122 int off; 5123 5124 /* The XDP program wants to see the packet starting at the MAC 5125 * header. 5126 */ 5127 mac_len = skb->data - skb_mac_header(skb); 5128 hard_start = skb->data - skb_headroom(skb); 5129 5130 /* SKB "head" area always have tailroom for skb_shared_info */ 5131 frame_sz = (void *)skb_end_pointer(skb) - hard_start; 5132 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 5133 5134 rxqueue = netif_get_rxqueue(skb); 5135 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq); 5136 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len, 5137 skb_headlen(skb) + mac_len, true); 5138 if (skb_is_nonlinear(skb)) { 5139 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 5140 xdp_buff_set_frags_flag(xdp); 5141 } else { 5142 xdp_buff_clear_frags_flag(xdp); 5143 } 5144 5145 orig_data_end = xdp->data_end; 5146 orig_data = xdp->data; 5147 eth = (struct ethhdr *)xdp->data; 5148 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr); 5149 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest); 5150 orig_eth_type = eth->h_proto; 5151 5152 act = bpf_prog_run_xdp(xdp_prog, xdp); 5153 5154 /* check if bpf_xdp_adjust_head was used */ 5155 off = xdp->data - orig_data; 5156 if (off) { 5157 if (off > 0) 5158 __skb_pull(skb, off); 5159 else if (off < 0) 5160 __skb_push(skb, -off); 5161 5162 skb->mac_header += off; 5163 skb_reset_network_header(skb); 5164 } 5165 5166 /* check if bpf_xdp_adjust_tail was used */ 5167 off = xdp->data_end - orig_data_end; 5168 if (off != 0) { 5169 skb_set_tail_pointer(skb, xdp->data_end - xdp->data); 5170 skb->len += off; /* positive on grow, negative on shrink */ 5171 } 5172 5173 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 5174 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here. 5175 */ 5176 if (xdp_buff_has_frags(xdp)) 5177 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 5178 else 5179 skb->data_len = 0; 5180 5181 /* check if XDP changed eth hdr such SKB needs update */ 5182 eth = (struct ethhdr *)xdp->data; 5183 if ((orig_eth_type != eth->h_proto) || 5184 (orig_host != ether_addr_equal_64bits(eth->h_dest, 5185 skb->dev->dev_addr)) || 5186 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) { 5187 __skb_push(skb, ETH_HLEN); 5188 skb->pkt_type = PACKET_HOST; 5189 skb->protocol = eth_type_trans(skb, skb->dev); 5190 } 5191 5192 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull 5193 * before calling us again on redirect path. We do not call do_redirect 5194 * as we leave that up to the caller. 5195 * 5196 * Caller is responsible for managing lifetime of skb (i.e. calling 5197 * kfree_skb in response to actions it cannot handle/XDP_DROP). 5198 */ 5199 switch (act) { 5200 case XDP_REDIRECT: 5201 case XDP_TX: 5202 __skb_push(skb, mac_len); 5203 break; 5204 case XDP_PASS: 5205 metalen = xdp->data - xdp->data_meta; 5206 if (metalen) 5207 skb_metadata_set(skb, metalen); 5208 break; 5209 } 5210 5211 return act; 5212 } 5213 5214 static int 5215 netif_skb_check_for_xdp(struct sk_buff **pskb, const struct bpf_prog *prog) 5216 { 5217 struct sk_buff *skb = *pskb; 5218 int err, hroom, troom; 5219 5220 if (!skb_cow_data_for_xdp(this_cpu_read(system_page_pool), pskb, prog)) 5221 return 0; 5222 5223 /* In case we have to go down the path and also linearize, 5224 * then lets do the pskb_expand_head() work just once here. 5225 */ 5226 hroom = XDP_PACKET_HEADROOM - skb_headroom(skb); 5227 troom = skb->tail + skb->data_len - skb->end; 5228 err = pskb_expand_head(skb, 5229 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0, 5230 troom > 0 ? troom + 128 : 0, GFP_ATOMIC); 5231 if (err) 5232 return err; 5233 5234 return skb_linearize(skb); 5235 } 5236 5237 static u32 netif_receive_generic_xdp(struct sk_buff **pskb, 5238 struct xdp_buff *xdp, 5239 const struct bpf_prog *xdp_prog) 5240 { 5241 struct sk_buff *skb = *pskb; 5242 u32 mac_len, act = XDP_DROP; 5243 5244 /* Reinjected packets coming from act_mirred or similar should 5245 * not get XDP generic processing. 5246 */ 5247 if (skb_is_redirected(skb)) 5248 return XDP_PASS; 5249 5250 /* XDP packets must have sufficient headroom of XDP_PACKET_HEADROOM 5251 * bytes. This is the guarantee that also native XDP provides, 5252 * thus we need to do it here as well. 5253 */ 5254 mac_len = skb->data - skb_mac_header(skb); 5255 __skb_push(skb, mac_len); 5256 5257 if (skb_cloned(skb) || skb_is_nonlinear(skb) || 5258 skb_headroom(skb) < XDP_PACKET_HEADROOM) { 5259 if (netif_skb_check_for_xdp(pskb, xdp_prog)) 5260 goto do_drop; 5261 } 5262 5263 __skb_pull(*pskb, mac_len); 5264 5265 act = bpf_prog_run_generic_xdp(*pskb, xdp, xdp_prog); 5266 switch (act) { 5267 case XDP_REDIRECT: 5268 case XDP_TX: 5269 case XDP_PASS: 5270 break; 5271 default: 5272 bpf_warn_invalid_xdp_action((*pskb)->dev, xdp_prog, act); 5273 fallthrough; 5274 case XDP_ABORTED: 5275 trace_xdp_exception((*pskb)->dev, xdp_prog, act); 5276 fallthrough; 5277 case XDP_DROP: 5278 do_drop: 5279 kfree_skb(*pskb); 5280 break; 5281 } 5282 5283 return act; 5284 } 5285 5286 /* When doing generic XDP we have to bypass the qdisc layer and the 5287 * network taps in order to match in-driver-XDP behavior. This also means 5288 * that XDP packets are able to starve other packets going through a qdisc, 5289 * and DDOS attacks will be more effective. In-driver-XDP use dedicated TX 5290 * queues, so they do not have this starvation issue. 5291 */ 5292 void generic_xdp_tx(struct sk_buff *skb, const struct bpf_prog *xdp_prog) 5293 { 5294 struct net_device *dev = skb->dev; 5295 struct netdev_queue *txq; 5296 bool free_skb = true; 5297 int cpu, rc; 5298 5299 txq = netdev_core_pick_tx(dev, skb, NULL); 5300 cpu = smp_processor_id(); 5301 HARD_TX_LOCK(dev, txq, cpu); 5302 if (!netif_xmit_frozen_or_drv_stopped(txq)) { 5303 rc = netdev_start_xmit(skb, dev, txq, 0); 5304 if (dev_xmit_complete(rc)) 5305 free_skb = false; 5306 } 5307 HARD_TX_UNLOCK(dev, txq); 5308 if (free_skb) { 5309 trace_xdp_exception(dev, xdp_prog, XDP_TX); 5310 dev_core_stats_tx_dropped_inc(dev); 5311 kfree_skb(skb); 5312 } 5313 } 5314 5315 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key); 5316 5317 int do_xdp_generic(const struct bpf_prog *xdp_prog, struct sk_buff **pskb) 5318 { 5319 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 5320 5321 if (xdp_prog) { 5322 struct xdp_buff xdp; 5323 u32 act; 5324 int err; 5325 5326 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 5327 act = netif_receive_generic_xdp(pskb, &xdp, xdp_prog); 5328 if (act != XDP_PASS) { 5329 switch (act) { 5330 case XDP_REDIRECT: 5331 err = xdp_do_generic_redirect((*pskb)->dev, *pskb, 5332 &xdp, xdp_prog); 5333 if (err) 5334 goto out_redir; 5335 break; 5336 case XDP_TX: 5337 generic_xdp_tx(*pskb, xdp_prog); 5338 break; 5339 } 5340 bpf_net_ctx_clear(bpf_net_ctx); 5341 return XDP_DROP; 5342 } 5343 bpf_net_ctx_clear(bpf_net_ctx); 5344 } 5345 return XDP_PASS; 5346 out_redir: 5347 bpf_net_ctx_clear(bpf_net_ctx); 5348 kfree_skb_reason(*pskb, SKB_DROP_REASON_XDP); 5349 return XDP_DROP; 5350 } 5351 EXPORT_SYMBOL_GPL(do_xdp_generic); 5352 5353 static int netif_rx_internal(struct sk_buff *skb) 5354 { 5355 int ret; 5356 5357 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5358 5359 trace_netif_rx(skb); 5360 5361 #ifdef CONFIG_RPS 5362 if (static_branch_unlikely(&rps_needed)) { 5363 struct rps_dev_flow voidflow, *rflow = &voidflow; 5364 int cpu; 5365 5366 rcu_read_lock(); 5367 5368 cpu = get_rps_cpu(skb->dev, skb, &rflow); 5369 if (cpu < 0) 5370 cpu = smp_processor_id(); 5371 5372 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5373 5374 rcu_read_unlock(); 5375 } else 5376 #endif 5377 { 5378 unsigned int qtail; 5379 5380 ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail); 5381 } 5382 return ret; 5383 } 5384 5385 /** 5386 * __netif_rx - Slightly optimized version of netif_rx 5387 * @skb: buffer to post 5388 * 5389 * This behaves as netif_rx except that it does not disable bottom halves. 5390 * As a result this function may only be invoked from the interrupt context 5391 * (either hard or soft interrupt). 5392 */ 5393 int __netif_rx(struct sk_buff *skb) 5394 { 5395 int ret; 5396 5397 lockdep_assert_once(hardirq_count() | softirq_count()); 5398 5399 trace_netif_rx_entry(skb); 5400 ret = netif_rx_internal(skb); 5401 trace_netif_rx_exit(ret); 5402 return ret; 5403 } 5404 EXPORT_SYMBOL(__netif_rx); 5405 5406 /** 5407 * netif_rx - post buffer to the network code 5408 * @skb: buffer to post 5409 * 5410 * This function receives a packet from a device driver and queues it for 5411 * the upper (protocol) levels to process via the backlog NAPI device. It 5412 * always succeeds. The buffer may be dropped during processing for 5413 * congestion control or by the protocol layers. 5414 * The network buffer is passed via the backlog NAPI device. Modern NIC 5415 * driver should use NAPI and GRO. 5416 * This function can used from interrupt and from process context. The 5417 * caller from process context must not disable interrupts before invoking 5418 * this function. 5419 * 5420 * return values: 5421 * NET_RX_SUCCESS (no congestion) 5422 * NET_RX_DROP (packet was dropped) 5423 * 5424 */ 5425 int netif_rx(struct sk_buff *skb) 5426 { 5427 bool need_bh_off = !(hardirq_count() | softirq_count()); 5428 int ret; 5429 5430 if (need_bh_off) 5431 local_bh_disable(); 5432 trace_netif_rx_entry(skb); 5433 ret = netif_rx_internal(skb); 5434 trace_netif_rx_exit(ret); 5435 if (need_bh_off) 5436 local_bh_enable(); 5437 return ret; 5438 } 5439 EXPORT_SYMBOL(netif_rx); 5440 5441 static __latent_entropy void net_tx_action(void) 5442 { 5443 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 5444 5445 if (sd->completion_queue) { 5446 struct sk_buff *clist; 5447 5448 local_irq_disable(); 5449 clist = sd->completion_queue; 5450 sd->completion_queue = NULL; 5451 local_irq_enable(); 5452 5453 while (clist) { 5454 struct sk_buff *skb = clist; 5455 5456 clist = clist->next; 5457 5458 WARN_ON(refcount_read(&skb->users)); 5459 if (likely(get_kfree_skb_cb(skb)->reason == SKB_CONSUMED)) 5460 trace_consume_skb(skb, net_tx_action); 5461 else 5462 trace_kfree_skb(skb, net_tx_action, 5463 get_kfree_skb_cb(skb)->reason, NULL); 5464 5465 if (skb->fclone != SKB_FCLONE_UNAVAILABLE) 5466 __kfree_skb(skb); 5467 else 5468 __napi_kfree_skb(skb, 5469 get_kfree_skb_cb(skb)->reason); 5470 } 5471 } 5472 5473 if (sd->output_queue) { 5474 struct Qdisc *head; 5475 5476 local_irq_disable(); 5477 head = sd->output_queue; 5478 sd->output_queue = NULL; 5479 sd->output_queue_tailp = &sd->output_queue; 5480 local_irq_enable(); 5481 5482 rcu_read_lock(); 5483 5484 while (head) { 5485 struct Qdisc *q = head; 5486 spinlock_t *root_lock = NULL; 5487 5488 head = head->next_sched; 5489 5490 /* We need to make sure head->next_sched is read 5491 * before clearing __QDISC_STATE_SCHED 5492 */ 5493 smp_mb__before_atomic(); 5494 5495 if (!(q->flags & TCQ_F_NOLOCK)) { 5496 root_lock = qdisc_lock(q); 5497 spin_lock(root_lock); 5498 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, 5499 &q->state))) { 5500 /* There is a synchronize_net() between 5501 * STATE_DEACTIVATED flag being set and 5502 * qdisc_reset()/some_qdisc_is_busy() in 5503 * dev_deactivate(), so we can safely bail out 5504 * early here to avoid data race between 5505 * qdisc_deactivate() and some_qdisc_is_busy() 5506 * for lockless qdisc. 5507 */ 5508 clear_bit(__QDISC_STATE_SCHED, &q->state); 5509 continue; 5510 } 5511 5512 clear_bit(__QDISC_STATE_SCHED, &q->state); 5513 qdisc_run(q); 5514 if (root_lock) 5515 spin_unlock(root_lock); 5516 } 5517 5518 rcu_read_unlock(); 5519 } 5520 5521 xfrm_dev_backlog(sd); 5522 } 5523 5524 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE) 5525 /* This hook is defined here for ATM LANE */ 5526 int (*br_fdb_test_addr_hook)(struct net_device *dev, 5527 unsigned char *addr) __read_mostly; 5528 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook); 5529 #endif 5530 5531 /** 5532 * netdev_is_rx_handler_busy - check if receive handler is registered 5533 * @dev: device to check 5534 * 5535 * Check if a receive handler is already registered for a given device. 5536 * Return true if there one. 5537 * 5538 * The caller must hold the rtnl_mutex. 5539 */ 5540 bool netdev_is_rx_handler_busy(struct net_device *dev) 5541 { 5542 ASSERT_RTNL(); 5543 return dev && rtnl_dereference(dev->rx_handler); 5544 } 5545 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy); 5546 5547 /** 5548 * netdev_rx_handler_register - register receive handler 5549 * @dev: device to register a handler for 5550 * @rx_handler: receive handler to register 5551 * @rx_handler_data: data pointer that is used by rx handler 5552 * 5553 * Register a receive handler for a device. This handler will then be 5554 * called from __netif_receive_skb. A negative errno code is returned 5555 * on a failure. 5556 * 5557 * The caller must hold the rtnl_mutex. 5558 * 5559 * For a general description of rx_handler, see enum rx_handler_result. 5560 */ 5561 int netdev_rx_handler_register(struct net_device *dev, 5562 rx_handler_func_t *rx_handler, 5563 void *rx_handler_data) 5564 { 5565 if (netdev_is_rx_handler_busy(dev)) 5566 return -EBUSY; 5567 5568 if (dev->priv_flags & IFF_NO_RX_HANDLER) 5569 return -EINVAL; 5570 5571 /* Note: rx_handler_data must be set before rx_handler */ 5572 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data); 5573 rcu_assign_pointer(dev->rx_handler, rx_handler); 5574 5575 return 0; 5576 } 5577 EXPORT_SYMBOL_GPL(netdev_rx_handler_register); 5578 5579 /** 5580 * netdev_rx_handler_unregister - unregister receive handler 5581 * @dev: device to unregister a handler from 5582 * 5583 * Unregister a receive handler from a device. 5584 * 5585 * The caller must hold the rtnl_mutex. 5586 */ 5587 void netdev_rx_handler_unregister(struct net_device *dev) 5588 { 5589 5590 ASSERT_RTNL(); 5591 RCU_INIT_POINTER(dev->rx_handler, NULL); 5592 /* a reader seeing a non NULL rx_handler in a rcu_read_lock() 5593 * section has a guarantee to see a non NULL rx_handler_data 5594 * as well. 5595 */ 5596 synchronize_net(); 5597 RCU_INIT_POINTER(dev->rx_handler_data, NULL); 5598 } 5599 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister); 5600 5601 /* 5602 * Limit the use of PFMEMALLOC reserves to those protocols that implement 5603 * the special handling of PFMEMALLOC skbs. 5604 */ 5605 static bool skb_pfmemalloc_protocol(struct sk_buff *skb) 5606 { 5607 switch (skb->protocol) { 5608 case htons(ETH_P_ARP): 5609 case htons(ETH_P_IP): 5610 case htons(ETH_P_IPV6): 5611 case htons(ETH_P_8021Q): 5612 case htons(ETH_P_8021AD): 5613 return true; 5614 default: 5615 return false; 5616 } 5617 } 5618 5619 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev, 5620 int *ret, struct net_device *orig_dev) 5621 { 5622 if (nf_hook_ingress_active(skb)) { 5623 int ingress_retval; 5624 5625 if (*pt_prev) { 5626 *ret = deliver_skb(skb, *pt_prev, orig_dev); 5627 *pt_prev = NULL; 5628 } 5629 5630 rcu_read_lock(); 5631 ingress_retval = nf_hook_ingress(skb); 5632 rcu_read_unlock(); 5633 return ingress_retval; 5634 } 5635 return 0; 5636 } 5637 5638 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc, 5639 struct packet_type **ppt_prev) 5640 { 5641 struct packet_type *ptype, *pt_prev; 5642 rx_handler_func_t *rx_handler; 5643 struct sk_buff *skb = *pskb; 5644 struct net_device *orig_dev; 5645 bool deliver_exact = false; 5646 int ret = NET_RX_DROP; 5647 __be16 type; 5648 5649 net_timestamp_check(!READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5650 5651 trace_netif_receive_skb(skb); 5652 5653 orig_dev = skb->dev; 5654 5655 skb_reset_network_header(skb); 5656 #if !defined(CONFIG_DEBUG_NET) 5657 /* We plan to no longer reset the transport header here. 5658 * Give some time to fuzzers and dev build to catch bugs 5659 * in network stacks. 5660 */ 5661 if (!skb_transport_header_was_set(skb)) 5662 skb_reset_transport_header(skb); 5663 #endif 5664 skb_reset_mac_len(skb); 5665 5666 pt_prev = NULL; 5667 5668 another_round: 5669 skb->skb_iif = skb->dev->ifindex; 5670 5671 __this_cpu_inc(softnet_data.processed); 5672 5673 if (static_branch_unlikely(&generic_xdp_needed_key)) { 5674 int ret2; 5675 5676 migrate_disable(); 5677 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), 5678 &skb); 5679 migrate_enable(); 5680 5681 if (ret2 != XDP_PASS) { 5682 ret = NET_RX_DROP; 5683 goto out; 5684 } 5685 } 5686 5687 if (eth_type_vlan(skb->protocol)) { 5688 skb = skb_vlan_untag(skb); 5689 if (unlikely(!skb)) 5690 goto out; 5691 } 5692 5693 if (skb_skip_tc_classify(skb)) 5694 goto skip_classify; 5695 5696 if (pfmemalloc) 5697 goto skip_taps; 5698 5699 list_for_each_entry_rcu(ptype, &net_hotdata.ptype_all, list) { 5700 if (pt_prev) 5701 ret = deliver_skb(skb, pt_prev, orig_dev); 5702 pt_prev = ptype; 5703 } 5704 5705 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) { 5706 if (pt_prev) 5707 ret = deliver_skb(skb, pt_prev, orig_dev); 5708 pt_prev = ptype; 5709 } 5710 5711 skip_taps: 5712 #ifdef CONFIG_NET_INGRESS 5713 if (static_branch_unlikely(&ingress_needed_key)) { 5714 bool another = false; 5715 5716 nf_skip_egress(skb, true); 5717 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev, 5718 &another); 5719 if (another) 5720 goto another_round; 5721 if (!skb) 5722 goto out; 5723 5724 nf_skip_egress(skb, false); 5725 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0) 5726 goto out; 5727 } 5728 #endif 5729 skb_reset_redirect(skb); 5730 skip_classify: 5731 if (pfmemalloc && !skb_pfmemalloc_protocol(skb)) 5732 goto drop; 5733 5734 if (skb_vlan_tag_present(skb)) { 5735 if (pt_prev) { 5736 ret = deliver_skb(skb, pt_prev, orig_dev); 5737 pt_prev = NULL; 5738 } 5739 if (vlan_do_receive(&skb)) 5740 goto another_round; 5741 else if (unlikely(!skb)) 5742 goto out; 5743 } 5744 5745 rx_handler = rcu_dereference(skb->dev->rx_handler); 5746 if (rx_handler) { 5747 if (pt_prev) { 5748 ret = deliver_skb(skb, pt_prev, orig_dev); 5749 pt_prev = NULL; 5750 } 5751 switch (rx_handler(&skb)) { 5752 case RX_HANDLER_CONSUMED: 5753 ret = NET_RX_SUCCESS; 5754 goto out; 5755 case RX_HANDLER_ANOTHER: 5756 goto another_round; 5757 case RX_HANDLER_EXACT: 5758 deliver_exact = true; 5759 break; 5760 case RX_HANDLER_PASS: 5761 break; 5762 default: 5763 BUG(); 5764 } 5765 } 5766 5767 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) { 5768 check_vlan_id: 5769 if (skb_vlan_tag_get_id(skb)) { 5770 /* Vlan id is non 0 and vlan_do_receive() above couldn't 5771 * find vlan device. 5772 */ 5773 skb->pkt_type = PACKET_OTHERHOST; 5774 } else if (eth_type_vlan(skb->protocol)) { 5775 /* Outer header is 802.1P with vlan 0, inner header is 5776 * 802.1Q or 802.1AD and vlan_do_receive() above could 5777 * not find vlan dev for vlan id 0. 5778 */ 5779 __vlan_hwaccel_clear_tag(skb); 5780 skb = skb_vlan_untag(skb); 5781 if (unlikely(!skb)) 5782 goto out; 5783 if (vlan_do_receive(&skb)) 5784 /* After stripping off 802.1P header with vlan 0 5785 * vlan dev is found for inner header. 5786 */ 5787 goto another_round; 5788 else if (unlikely(!skb)) 5789 goto out; 5790 else 5791 /* We have stripped outer 802.1P vlan 0 header. 5792 * But could not find vlan dev. 5793 * check again for vlan id to set OTHERHOST. 5794 */ 5795 goto check_vlan_id; 5796 } 5797 /* Note: we might in the future use prio bits 5798 * and set skb->priority like in vlan_do_receive() 5799 * For the time being, just ignore Priority Code Point 5800 */ 5801 __vlan_hwaccel_clear_tag(skb); 5802 } 5803 5804 type = skb->protocol; 5805 5806 /* deliver only exact match when indicated */ 5807 if (likely(!deliver_exact)) { 5808 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5809 &ptype_base[ntohs(type) & 5810 PTYPE_HASH_MASK]); 5811 } 5812 5813 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5814 &orig_dev->ptype_specific); 5815 5816 if (unlikely(skb->dev != orig_dev)) { 5817 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5818 &skb->dev->ptype_specific); 5819 } 5820 5821 if (pt_prev) { 5822 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 5823 goto drop; 5824 *ppt_prev = pt_prev; 5825 } else { 5826 drop: 5827 if (!deliver_exact) 5828 dev_core_stats_rx_dropped_inc(skb->dev); 5829 else 5830 dev_core_stats_rx_nohandler_inc(skb->dev); 5831 kfree_skb_reason(skb, SKB_DROP_REASON_UNHANDLED_PROTO); 5832 /* Jamal, now you will not able to escape explaining 5833 * me how you were going to use this. :-) 5834 */ 5835 ret = NET_RX_DROP; 5836 } 5837 5838 out: 5839 /* The invariant here is that if *ppt_prev is not NULL 5840 * then skb should also be non-NULL. 5841 * 5842 * Apparently *ppt_prev assignment above holds this invariant due to 5843 * skb dereferencing near it. 5844 */ 5845 *pskb = skb; 5846 return ret; 5847 } 5848 5849 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc) 5850 { 5851 struct net_device *orig_dev = skb->dev; 5852 struct packet_type *pt_prev = NULL; 5853 int ret; 5854 5855 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5856 if (pt_prev) 5857 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb, 5858 skb->dev, pt_prev, orig_dev); 5859 return ret; 5860 } 5861 5862 /** 5863 * netif_receive_skb_core - special purpose version of netif_receive_skb 5864 * @skb: buffer to process 5865 * 5866 * More direct receive version of netif_receive_skb(). It should 5867 * only be used by callers that have a need to skip RPS and Generic XDP. 5868 * Caller must also take care of handling if ``(page_is_)pfmemalloc``. 5869 * 5870 * This function may only be called from softirq context and interrupts 5871 * should be enabled. 5872 * 5873 * Return values (usually ignored): 5874 * NET_RX_SUCCESS: no congestion 5875 * NET_RX_DROP: packet was dropped 5876 */ 5877 int netif_receive_skb_core(struct sk_buff *skb) 5878 { 5879 int ret; 5880 5881 rcu_read_lock(); 5882 ret = __netif_receive_skb_one_core(skb, false); 5883 rcu_read_unlock(); 5884 5885 return ret; 5886 } 5887 EXPORT_SYMBOL(netif_receive_skb_core); 5888 5889 static inline void __netif_receive_skb_list_ptype(struct list_head *head, 5890 struct packet_type *pt_prev, 5891 struct net_device *orig_dev) 5892 { 5893 struct sk_buff *skb, *next; 5894 5895 if (!pt_prev) 5896 return; 5897 if (list_empty(head)) 5898 return; 5899 if (pt_prev->list_func != NULL) 5900 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv, 5901 ip_list_rcv, head, pt_prev, orig_dev); 5902 else 5903 list_for_each_entry_safe(skb, next, head, list) { 5904 skb_list_del_init(skb); 5905 pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 5906 } 5907 } 5908 5909 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc) 5910 { 5911 /* Fast-path assumptions: 5912 * - There is no RX handler. 5913 * - Only one packet_type matches. 5914 * If either of these fails, we will end up doing some per-packet 5915 * processing in-line, then handling the 'last ptype' for the whole 5916 * sublist. This can't cause out-of-order delivery to any single ptype, 5917 * because the 'last ptype' must be constant across the sublist, and all 5918 * other ptypes are handled per-packet. 5919 */ 5920 /* Current (common) ptype of sublist */ 5921 struct packet_type *pt_curr = NULL; 5922 /* Current (common) orig_dev of sublist */ 5923 struct net_device *od_curr = NULL; 5924 struct sk_buff *skb, *next; 5925 LIST_HEAD(sublist); 5926 5927 list_for_each_entry_safe(skb, next, head, list) { 5928 struct net_device *orig_dev = skb->dev; 5929 struct packet_type *pt_prev = NULL; 5930 5931 skb_list_del_init(skb); 5932 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5933 if (!pt_prev) 5934 continue; 5935 if (pt_curr != pt_prev || od_curr != orig_dev) { 5936 /* dispatch old sublist */ 5937 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5938 /* start new sublist */ 5939 INIT_LIST_HEAD(&sublist); 5940 pt_curr = pt_prev; 5941 od_curr = orig_dev; 5942 } 5943 list_add_tail(&skb->list, &sublist); 5944 } 5945 5946 /* dispatch final sublist */ 5947 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5948 } 5949 5950 static int __netif_receive_skb(struct sk_buff *skb) 5951 { 5952 int ret; 5953 5954 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) { 5955 unsigned int noreclaim_flag; 5956 5957 /* 5958 * PFMEMALLOC skbs are special, they should 5959 * - be delivered to SOCK_MEMALLOC sockets only 5960 * - stay away from userspace 5961 * - have bounded memory usage 5962 * 5963 * Use PF_MEMALLOC as this saves us from propagating the allocation 5964 * context down to all allocation sites. 5965 */ 5966 noreclaim_flag = memalloc_noreclaim_save(); 5967 ret = __netif_receive_skb_one_core(skb, true); 5968 memalloc_noreclaim_restore(noreclaim_flag); 5969 } else 5970 ret = __netif_receive_skb_one_core(skb, false); 5971 5972 return ret; 5973 } 5974 5975 static void __netif_receive_skb_list(struct list_head *head) 5976 { 5977 unsigned long noreclaim_flag = 0; 5978 struct sk_buff *skb, *next; 5979 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */ 5980 5981 list_for_each_entry_safe(skb, next, head, list) { 5982 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) { 5983 struct list_head sublist; 5984 5985 /* Handle the previous sublist */ 5986 list_cut_before(&sublist, head, &skb->list); 5987 if (!list_empty(&sublist)) 5988 __netif_receive_skb_list_core(&sublist, pfmemalloc); 5989 pfmemalloc = !pfmemalloc; 5990 /* See comments in __netif_receive_skb */ 5991 if (pfmemalloc) 5992 noreclaim_flag = memalloc_noreclaim_save(); 5993 else 5994 memalloc_noreclaim_restore(noreclaim_flag); 5995 } 5996 } 5997 /* Handle the remaining sublist */ 5998 if (!list_empty(head)) 5999 __netif_receive_skb_list_core(head, pfmemalloc); 6000 /* Restore pflags */ 6001 if (pfmemalloc) 6002 memalloc_noreclaim_restore(noreclaim_flag); 6003 } 6004 6005 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp) 6006 { 6007 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog); 6008 struct bpf_prog *new = xdp->prog; 6009 int ret = 0; 6010 6011 switch (xdp->command) { 6012 case XDP_SETUP_PROG: 6013 rcu_assign_pointer(dev->xdp_prog, new); 6014 if (old) 6015 bpf_prog_put(old); 6016 6017 if (old && !new) { 6018 static_branch_dec(&generic_xdp_needed_key); 6019 } else if (new && !old) { 6020 static_branch_inc(&generic_xdp_needed_key); 6021 netif_disable_lro(dev); 6022 dev_disable_gro_hw(dev); 6023 } 6024 break; 6025 6026 default: 6027 ret = -EINVAL; 6028 break; 6029 } 6030 6031 return ret; 6032 } 6033 6034 static int netif_receive_skb_internal(struct sk_buff *skb) 6035 { 6036 int ret; 6037 6038 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 6039 6040 if (skb_defer_rx_timestamp(skb)) 6041 return NET_RX_SUCCESS; 6042 6043 rcu_read_lock(); 6044 #ifdef CONFIG_RPS 6045 if (static_branch_unlikely(&rps_needed)) { 6046 struct rps_dev_flow voidflow, *rflow = &voidflow; 6047 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6048 6049 if (cpu >= 0) { 6050 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6051 rcu_read_unlock(); 6052 return ret; 6053 } 6054 } 6055 #endif 6056 ret = __netif_receive_skb(skb); 6057 rcu_read_unlock(); 6058 return ret; 6059 } 6060 6061 void netif_receive_skb_list_internal(struct list_head *head) 6062 { 6063 struct sk_buff *skb, *next; 6064 LIST_HEAD(sublist); 6065 6066 list_for_each_entry_safe(skb, next, head, list) { 6067 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), 6068 skb); 6069 skb_list_del_init(skb); 6070 if (!skb_defer_rx_timestamp(skb)) 6071 list_add_tail(&skb->list, &sublist); 6072 } 6073 list_splice_init(&sublist, head); 6074 6075 rcu_read_lock(); 6076 #ifdef CONFIG_RPS 6077 if (static_branch_unlikely(&rps_needed)) { 6078 list_for_each_entry_safe(skb, next, head, list) { 6079 struct rps_dev_flow voidflow, *rflow = &voidflow; 6080 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6081 6082 if (cpu >= 0) { 6083 /* Will be handled, remove from list */ 6084 skb_list_del_init(skb); 6085 enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6086 } 6087 } 6088 } 6089 #endif 6090 __netif_receive_skb_list(head); 6091 rcu_read_unlock(); 6092 } 6093 6094 /** 6095 * netif_receive_skb - process receive buffer from network 6096 * @skb: buffer to process 6097 * 6098 * netif_receive_skb() is the main receive data processing function. 6099 * It always succeeds. The buffer may be dropped during processing 6100 * for congestion control or by the protocol layers. 6101 * 6102 * This function may only be called from softirq context and interrupts 6103 * should be enabled. 6104 * 6105 * Return values (usually ignored): 6106 * NET_RX_SUCCESS: no congestion 6107 * NET_RX_DROP: packet was dropped 6108 */ 6109 int netif_receive_skb(struct sk_buff *skb) 6110 { 6111 int ret; 6112 6113 trace_netif_receive_skb_entry(skb); 6114 6115 ret = netif_receive_skb_internal(skb); 6116 trace_netif_receive_skb_exit(ret); 6117 6118 return ret; 6119 } 6120 EXPORT_SYMBOL(netif_receive_skb); 6121 6122 /** 6123 * netif_receive_skb_list - process many receive buffers from network 6124 * @head: list of skbs to process. 6125 * 6126 * Since return value of netif_receive_skb() is normally ignored, and 6127 * wouldn't be meaningful for a list, this function returns void. 6128 * 6129 * This function may only be called from softirq context and interrupts 6130 * should be enabled. 6131 */ 6132 void netif_receive_skb_list(struct list_head *head) 6133 { 6134 struct sk_buff *skb; 6135 6136 if (list_empty(head)) 6137 return; 6138 if (trace_netif_receive_skb_list_entry_enabled()) { 6139 list_for_each_entry(skb, head, list) 6140 trace_netif_receive_skb_list_entry(skb); 6141 } 6142 netif_receive_skb_list_internal(head); 6143 trace_netif_receive_skb_list_exit(0); 6144 } 6145 EXPORT_SYMBOL(netif_receive_skb_list); 6146 6147 /* Network device is going away, flush any packets still pending */ 6148 static void flush_backlog(struct work_struct *work) 6149 { 6150 struct sk_buff *skb, *tmp; 6151 struct sk_buff_head list; 6152 struct softnet_data *sd; 6153 6154 __skb_queue_head_init(&list); 6155 local_bh_disable(); 6156 sd = this_cpu_ptr(&softnet_data); 6157 6158 backlog_lock_irq_disable(sd); 6159 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) { 6160 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6161 __skb_unlink(skb, &sd->input_pkt_queue); 6162 __skb_queue_tail(&list, skb); 6163 rps_input_queue_head_incr(sd); 6164 } 6165 } 6166 backlog_unlock_irq_enable(sd); 6167 6168 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6169 skb_queue_walk_safe(&sd->process_queue, skb, tmp) { 6170 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6171 __skb_unlink(skb, &sd->process_queue); 6172 __skb_queue_tail(&list, skb); 6173 rps_input_queue_head_incr(sd); 6174 } 6175 } 6176 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6177 local_bh_enable(); 6178 6179 __skb_queue_purge_reason(&list, SKB_DROP_REASON_DEV_READY); 6180 } 6181 6182 static bool flush_required(int cpu) 6183 { 6184 #if IS_ENABLED(CONFIG_RPS) 6185 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 6186 bool do_flush; 6187 6188 backlog_lock_irq_disable(sd); 6189 6190 /* as insertion into process_queue happens with the rps lock held, 6191 * process_queue access may race only with dequeue 6192 */ 6193 do_flush = !skb_queue_empty(&sd->input_pkt_queue) || 6194 !skb_queue_empty_lockless(&sd->process_queue); 6195 backlog_unlock_irq_enable(sd); 6196 6197 return do_flush; 6198 #endif 6199 /* without RPS we can't safely check input_pkt_queue: during a 6200 * concurrent remote skb_queue_splice() we can detect as empty both 6201 * input_pkt_queue and process_queue even if the latter could end-up 6202 * containing a lot of packets. 6203 */ 6204 return true; 6205 } 6206 6207 struct flush_backlogs { 6208 cpumask_t flush_cpus; 6209 struct work_struct w[]; 6210 }; 6211 6212 static struct flush_backlogs *flush_backlogs_alloc(void) 6213 { 6214 return kmalloc(struct_size_t(struct flush_backlogs, w, nr_cpu_ids), 6215 GFP_KERNEL); 6216 } 6217 6218 static struct flush_backlogs *flush_backlogs_fallback; 6219 static DEFINE_MUTEX(flush_backlogs_mutex); 6220 6221 static void flush_all_backlogs(void) 6222 { 6223 struct flush_backlogs *ptr = flush_backlogs_alloc(); 6224 unsigned int cpu; 6225 6226 if (!ptr) { 6227 mutex_lock(&flush_backlogs_mutex); 6228 ptr = flush_backlogs_fallback; 6229 } 6230 cpumask_clear(&ptr->flush_cpus); 6231 6232 cpus_read_lock(); 6233 6234 for_each_online_cpu(cpu) { 6235 if (flush_required(cpu)) { 6236 INIT_WORK(&ptr->w[cpu], flush_backlog); 6237 queue_work_on(cpu, system_highpri_wq, &ptr->w[cpu]); 6238 __cpumask_set_cpu(cpu, &ptr->flush_cpus); 6239 } 6240 } 6241 6242 /* we can have in flight packet[s] on the cpus we are not flushing, 6243 * synchronize_net() in unregister_netdevice_many() will take care of 6244 * them. 6245 */ 6246 for_each_cpu(cpu, &ptr->flush_cpus) 6247 flush_work(&ptr->w[cpu]); 6248 6249 cpus_read_unlock(); 6250 6251 if (ptr != flush_backlogs_fallback) 6252 kfree(ptr); 6253 else 6254 mutex_unlock(&flush_backlogs_mutex); 6255 } 6256 6257 static void net_rps_send_ipi(struct softnet_data *remsd) 6258 { 6259 #ifdef CONFIG_RPS 6260 while (remsd) { 6261 struct softnet_data *next = remsd->rps_ipi_next; 6262 6263 if (cpu_online(remsd->cpu)) 6264 smp_call_function_single_async(remsd->cpu, &remsd->csd); 6265 remsd = next; 6266 } 6267 #endif 6268 } 6269 6270 /* 6271 * net_rps_action_and_irq_enable sends any pending IPI's for rps. 6272 * Note: called with local irq disabled, but exits with local irq enabled. 6273 */ 6274 static void net_rps_action_and_irq_enable(struct softnet_data *sd) 6275 { 6276 #ifdef CONFIG_RPS 6277 struct softnet_data *remsd = sd->rps_ipi_list; 6278 6279 if (!use_backlog_threads() && remsd) { 6280 sd->rps_ipi_list = NULL; 6281 6282 local_irq_enable(); 6283 6284 /* Send pending IPI's to kick RPS processing on remote cpus. */ 6285 net_rps_send_ipi(remsd); 6286 } else 6287 #endif 6288 local_irq_enable(); 6289 } 6290 6291 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd) 6292 { 6293 #ifdef CONFIG_RPS 6294 return !use_backlog_threads() && sd->rps_ipi_list; 6295 #else 6296 return false; 6297 #endif 6298 } 6299 6300 static int process_backlog(struct napi_struct *napi, int quota) 6301 { 6302 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog); 6303 bool again = true; 6304 int work = 0; 6305 6306 /* Check if we have pending ipi, its better to send them now, 6307 * not waiting net_rx_action() end. 6308 */ 6309 if (sd_has_rps_ipi_waiting(sd)) { 6310 local_irq_disable(); 6311 net_rps_action_and_irq_enable(sd); 6312 } 6313 6314 napi->weight = READ_ONCE(net_hotdata.dev_rx_weight); 6315 while (again) { 6316 struct sk_buff *skb; 6317 6318 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6319 while ((skb = __skb_dequeue(&sd->process_queue))) { 6320 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6321 rcu_read_lock(); 6322 __netif_receive_skb(skb); 6323 rcu_read_unlock(); 6324 if (++work >= quota) { 6325 rps_input_queue_head_add(sd, work); 6326 return work; 6327 } 6328 6329 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6330 } 6331 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6332 6333 backlog_lock_irq_disable(sd); 6334 if (skb_queue_empty(&sd->input_pkt_queue)) { 6335 /* 6336 * Inline a custom version of __napi_complete(). 6337 * only current cpu owns and manipulates this napi, 6338 * and NAPI_STATE_SCHED is the only possible flag set 6339 * on backlog. 6340 * We can use a plain write instead of clear_bit(), 6341 * and we dont need an smp_mb() memory barrier. 6342 */ 6343 napi->state &= NAPIF_STATE_THREADED; 6344 again = false; 6345 } else { 6346 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6347 skb_queue_splice_tail_init(&sd->input_pkt_queue, 6348 &sd->process_queue); 6349 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6350 } 6351 backlog_unlock_irq_enable(sd); 6352 } 6353 6354 if (work) 6355 rps_input_queue_head_add(sd, work); 6356 return work; 6357 } 6358 6359 /** 6360 * __napi_schedule - schedule for receive 6361 * @n: entry to schedule 6362 * 6363 * The entry's receive function will be scheduled to run. 6364 * Consider using __napi_schedule_irqoff() if hard irqs are masked. 6365 */ 6366 void __napi_schedule(struct napi_struct *n) 6367 { 6368 unsigned long flags; 6369 6370 local_irq_save(flags); 6371 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6372 local_irq_restore(flags); 6373 } 6374 EXPORT_SYMBOL(__napi_schedule); 6375 6376 /** 6377 * napi_schedule_prep - check if napi can be scheduled 6378 * @n: napi context 6379 * 6380 * Test if NAPI routine is already running, and if not mark 6381 * it as running. This is used as a condition variable to 6382 * insure only one NAPI poll instance runs. We also make 6383 * sure there is no pending NAPI disable. 6384 */ 6385 bool napi_schedule_prep(struct napi_struct *n) 6386 { 6387 unsigned long new, val = READ_ONCE(n->state); 6388 6389 do { 6390 if (unlikely(val & NAPIF_STATE_DISABLE)) 6391 return false; 6392 new = val | NAPIF_STATE_SCHED; 6393 6394 /* Sets STATE_MISSED bit if STATE_SCHED was already set 6395 * This was suggested by Alexander Duyck, as compiler 6396 * emits better code than : 6397 * if (val & NAPIF_STATE_SCHED) 6398 * new |= NAPIF_STATE_MISSED; 6399 */ 6400 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED * 6401 NAPIF_STATE_MISSED; 6402 } while (!try_cmpxchg(&n->state, &val, new)); 6403 6404 return !(val & NAPIF_STATE_SCHED); 6405 } 6406 EXPORT_SYMBOL(napi_schedule_prep); 6407 6408 /** 6409 * __napi_schedule_irqoff - schedule for receive 6410 * @n: entry to schedule 6411 * 6412 * Variant of __napi_schedule() assuming hard irqs are masked. 6413 * 6414 * On PREEMPT_RT enabled kernels this maps to __napi_schedule() 6415 * because the interrupt disabled assumption might not be true 6416 * due to force-threaded interrupts and spinlock substitution. 6417 */ 6418 void __napi_schedule_irqoff(struct napi_struct *n) 6419 { 6420 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6421 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6422 else 6423 __napi_schedule(n); 6424 } 6425 EXPORT_SYMBOL(__napi_schedule_irqoff); 6426 6427 bool napi_complete_done(struct napi_struct *n, int work_done) 6428 { 6429 unsigned long flags, val, new, timeout = 0; 6430 bool ret = true; 6431 6432 /* 6433 * 1) Don't let napi dequeue from the cpu poll list 6434 * just in case its running on a different cpu. 6435 * 2) If we are busy polling, do nothing here, we have 6436 * the guarantee we will be called later. 6437 */ 6438 if (unlikely(n->state & (NAPIF_STATE_NPSVC | 6439 NAPIF_STATE_IN_BUSY_POLL))) 6440 return false; 6441 6442 if (work_done) { 6443 if (n->gro.bitmask) 6444 timeout = napi_get_gro_flush_timeout(n); 6445 n->defer_hard_irqs_count = napi_get_defer_hard_irqs(n); 6446 } 6447 if (n->defer_hard_irqs_count > 0) { 6448 n->defer_hard_irqs_count--; 6449 timeout = napi_get_gro_flush_timeout(n); 6450 if (timeout) 6451 ret = false; 6452 } 6453 6454 /* 6455 * When the NAPI instance uses a timeout and keeps postponing 6456 * it, we need to bound somehow the time packets are kept in 6457 * the GRO layer. 6458 */ 6459 gro_flush(&n->gro, !!timeout); 6460 gro_normal_list(&n->gro); 6461 6462 if (unlikely(!list_empty(&n->poll_list))) { 6463 /* If n->poll_list is not empty, we need to mask irqs */ 6464 local_irq_save(flags); 6465 list_del_init(&n->poll_list); 6466 local_irq_restore(flags); 6467 } 6468 WRITE_ONCE(n->list_owner, -1); 6469 6470 val = READ_ONCE(n->state); 6471 do { 6472 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED)); 6473 6474 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED | 6475 NAPIF_STATE_SCHED_THREADED | 6476 NAPIF_STATE_PREFER_BUSY_POLL); 6477 6478 /* If STATE_MISSED was set, leave STATE_SCHED set, 6479 * because we will call napi->poll() one more time. 6480 * This C code was suggested by Alexander Duyck to help gcc. 6481 */ 6482 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED * 6483 NAPIF_STATE_SCHED; 6484 } while (!try_cmpxchg(&n->state, &val, new)); 6485 6486 if (unlikely(val & NAPIF_STATE_MISSED)) { 6487 __napi_schedule(n); 6488 return false; 6489 } 6490 6491 if (timeout) 6492 hrtimer_start(&n->timer, ns_to_ktime(timeout), 6493 HRTIMER_MODE_REL_PINNED); 6494 return ret; 6495 } 6496 EXPORT_SYMBOL(napi_complete_done); 6497 6498 static void skb_defer_free_flush(struct softnet_data *sd) 6499 { 6500 struct sk_buff *skb, *next; 6501 6502 /* Paired with WRITE_ONCE() in skb_attempt_defer_free() */ 6503 if (!READ_ONCE(sd->defer_list)) 6504 return; 6505 6506 spin_lock(&sd->defer_lock); 6507 skb = sd->defer_list; 6508 sd->defer_list = NULL; 6509 sd->defer_count = 0; 6510 spin_unlock(&sd->defer_lock); 6511 6512 while (skb != NULL) { 6513 next = skb->next; 6514 napi_consume_skb(skb, 1); 6515 skb = next; 6516 } 6517 } 6518 6519 #if defined(CONFIG_NET_RX_BUSY_POLL) 6520 6521 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule) 6522 { 6523 if (!skip_schedule) { 6524 gro_normal_list(&napi->gro); 6525 __napi_schedule(napi); 6526 return; 6527 } 6528 6529 /* Flush too old packets. If HZ < 1000, flush all packets */ 6530 gro_flush(&napi->gro, HZ >= 1000); 6531 gro_normal_list(&napi->gro); 6532 6533 clear_bit(NAPI_STATE_SCHED, &napi->state); 6534 } 6535 6536 enum { 6537 NAPI_F_PREFER_BUSY_POLL = 1, 6538 NAPI_F_END_ON_RESCHED = 2, 6539 }; 6540 6541 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, 6542 unsigned flags, u16 budget) 6543 { 6544 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6545 bool skip_schedule = false; 6546 unsigned long timeout; 6547 int rc; 6548 6549 /* Busy polling means there is a high chance device driver hard irq 6550 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was 6551 * set in napi_schedule_prep(). 6552 * Since we are about to call napi->poll() once more, we can safely 6553 * clear NAPI_STATE_MISSED. 6554 * 6555 * Note: x86 could use a single "lock and ..." instruction 6556 * to perform these two clear_bit() 6557 */ 6558 clear_bit(NAPI_STATE_MISSED, &napi->state); 6559 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state); 6560 6561 local_bh_disable(); 6562 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6563 6564 if (flags & NAPI_F_PREFER_BUSY_POLL) { 6565 napi->defer_hard_irqs_count = napi_get_defer_hard_irqs(napi); 6566 timeout = napi_get_gro_flush_timeout(napi); 6567 if (napi->defer_hard_irqs_count && timeout) { 6568 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED); 6569 skip_schedule = true; 6570 } 6571 } 6572 6573 /* All we really want here is to re-enable device interrupts. 6574 * Ideally, a new ndo_busy_poll_stop() could avoid another round. 6575 */ 6576 rc = napi->poll(napi, budget); 6577 /* We can't gro_normal_list() here, because napi->poll() might have 6578 * rearmed the napi (napi_complete_done()) in which case it could 6579 * already be running on another CPU. 6580 */ 6581 trace_napi_poll(napi, rc, budget); 6582 netpoll_poll_unlock(have_poll_lock); 6583 if (rc == budget) 6584 __busy_poll_stop(napi, skip_schedule); 6585 bpf_net_ctx_clear(bpf_net_ctx); 6586 local_bh_enable(); 6587 } 6588 6589 static void __napi_busy_loop(unsigned int napi_id, 6590 bool (*loop_end)(void *, unsigned long), 6591 void *loop_end_arg, unsigned flags, u16 budget) 6592 { 6593 unsigned long start_time = loop_end ? busy_loop_current_time() : 0; 6594 int (*napi_poll)(struct napi_struct *napi, int budget); 6595 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6596 void *have_poll_lock = NULL; 6597 struct napi_struct *napi; 6598 6599 WARN_ON_ONCE(!rcu_read_lock_held()); 6600 6601 restart: 6602 napi_poll = NULL; 6603 6604 napi = napi_by_id(napi_id); 6605 if (!napi) 6606 return; 6607 6608 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6609 preempt_disable(); 6610 for (;;) { 6611 int work = 0; 6612 6613 local_bh_disable(); 6614 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6615 if (!napi_poll) { 6616 unsigned long val = READ_ONCE(napi->state); 6617 6618 /* If multiple threads are competing for this napi, 6619 * we avoid dirtying napi->state as much as we can. 6620 */ 6621 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED | 6622 NAPIF_STATE_IN_BUSY_POLL)) { 6623 if (flags & NAPI_F_PREFER_BUSY_POLL) 6624 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6625 goto count; 6626 } 6627 if (cmpxchg(&napi->state, val, 6628 val | NAPIF_STATE_IN_BUSY_POLL | 6629 NAPIF_STATE_SCHED) != val) { 6630 if (flags & NAPI_F_PREFER_BUSY_POLL) 6631 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6632 goto count; 6633 } 6634 have_poll_lock = netpoll_poll_lock(napi); 6635 napi_poll = napi->poll; 6636 } 6637 work = napi_poll(napi, budget); 6638 trace_napi_poll(napi, work, budget); 6639 gro_normal_list(&napi->gro); 6640 count: 6641 if (work > 0) 6642 __NET_ADD_STATS(dev_net(napi->dev), 6643 LINUX_MIB_BUSYPOLLRXPACKETS, work); 6644 skb_defer_free_flush(this_cpu_ptr(&softnet_data)); 6645 bpf_net_ctx_clear(bpf_net_ctx); 6646 local_bh_enable(); 6647 6648 if (!loop_end || loop_end(loop_end_arg, start_time)) 6649 break; 6650 6651 if (unlikely(need_resched())) { 6652 if (flags & NAPI_F_END_ON_RESCHED) 6653 break; 6654 if (napi_poll) 6655 busy_poll_stop(napi, have_poll_lock, flags, budget); 6656 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6657 preempt_enable(); 6658 rcu_read_unlock(); 6659 cond_resched(); 6660 rcu_read_lock(); 6661 if (loop_end(loop_end_arg, start_time)) 6662 return; 6663 goto restart; 6664 } 6665 cpu_relax(); 6666 } 6667 if (napi_poll) 6668 busy_poll_stop(napi, have_poll_lock, flags, budget); 6669 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6670 preempt_enable(); 6671 } 6672 6673 void napi_busy_loop_rcu(unsigned int napi_id, 6674 bool (*loop_end)(void *, unsigned long), 6675 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6676 { 6677 unsigned flags = NAPI_F_END_ON_RESCHED; 6678 6679 if (prefer_busy_poll) 6680 flags |= NAPI_F_PREFER_BUSY_POLL; 6681 6682 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6683 } 6684 6685 void napi_busy_loop(unsigned int napi_id, 6686 bool (*loop_end)(void *, unsigned long), 6687 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6688 { 6689 unsigned flags = prefer_busy_poll ? NAPI_F_PREFER_BUSY_POLL : 0; 6690 6691 rcu_read_lock(); 6692 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6693 rcu_read_unlock(); 6694 } 6695 EXPORT_SYMBOL(napi_busy_loop); 6696 6697 void napi_suspend_irqs(unsigned int napi_id) 6698 { 6699 struct napi_struct *napi; 6700 6701 rcu_read_lock(); 6702 napi = napi_by_id(napi_id); 6703 if (napi) { 6704 unsigned long timeout = napi_get_irq_suspend_timeout(napi); 6705 6706 if (timeout) 6707 hrtimer_start(&napi->timer, ns_to_ktime(timeout), 6708 HRTIMER_MODE_REL_PINNED); 6709 } 6710 rcu_read_unlock(); 6711 } 6712 6713 void napi_resume_irqs(unsigned int napi_id) 6714 { 6715 struct napi_struct *napi; 6716 6717 rcu_read_lock(); 6718 napi = napi_by_id(napi_id); 6719 if (napi) { 6720 /* If irq_suspend_timeout is set to 0 between the call to 6721 * napi_suspend_irqs and now, the original value still 6722 * determines the safety timeout as intended and napi_watchdog 6723 * will resume irq processing. 6724 */ 6725 if (napi_get_irq_suspend_timeout(napi)) { 6726 local_bh_disable(); 6727 napi_schedule(napi); 6728 local_bh_enable(); 6729 } 6730 } 6731 rcu_read_unlock(); 6732 } 6733 6734 #endif /* CONFIG_NET_RX_BUSY_POLL */ 6735 6736 static void __napi_hash_add_with_id(struct napi_struct *napi, 6737 unsigned int napi_id) 6738 { 6739 napi->gro.cached_napi_id = napi_id; 6740 6741 WRITE_ONCE(napi->napi_id, napi_id); 6742 hlist_add_head_rcu(&napi->napi_hash_node, 6743 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]); 6744 } 6745 6746 static void napi_hash_add_with_id(struct napi_struct *napi, 6747 unsigned int napi_id) 6748 { 6749 unsigned long flags; 6750 6751 spin_lock_irqsave(&napi_hash_lock, flags); 6752 WARN_ON_ONCE(napi_by_id(napi_id)); 6753 __napi_hash_add_with_id(napi, napi_id); 6754 spin_unlock_irqrestore(&napi_hash_lock, flags); 6755 } 6756 6757 static void napi_hash_add(struct napi_struct *napi) 6758 { 6759 unsigned long flags; 6760 6761 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state)) 6762 return; 6763 6764 spin_lock_irqsave(&napi_hash_lock, flags); 6765 6766 /* 0..NR_CPUS range is reserved for sender_cpu use */ 6767 do { 6768 if (unlikely(!napi_id_valid(++napi_gen_id))) 6769 napi_gen_id = MIN_NAPI_ID; 6770 } while (napi_by_id(napi_gen_id)); 6771 6772 __napi_hash_add_with_id(napi, napi_gen_id); 6773 6774 spin_unlock_irqrestore(&napi_hash_lock, flags); 6775 } 6776 6777 /* Warning : caller is responsible to make sure rcu grace period 6778 * is respected before freeing memory containing @napi 6779 */ 6780 static void napi_hash_del(struct napi_struct *napi) 6781 { 6782 unsigned long flags; 6783 6784 spin_lock_irqsave(&napi_hash_lock, flags); 6785 6786 hlist_del_init_rcu(&napi->napi_hash_node); 6787 6788 spin_unlock_irqrestore(&napi_hash_lock, flags); 6789 } 6790 6791 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer) 6792 { 6793 struct napi_struct *napi; 6794 6795 napi = container_of(timer, struct napi_struct, timer); 6796 6797 /* Note : we use a relaxed variant of napi_schedule_prep() not setting 6798 * NAPI_STATE_MISSED, since we do not react to a device IRQ. 6799 */ 6800 if (!napi_disable_pending(napi) && 6801 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) { 6802 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6803 __napi_schedule_irqoff(napi); 6804 } 6805 6806 return HRTIMER_NORESTART; 6807 } 6808 6809 int dev_set_threaded(struct net_device *dev, bool threaded) 6810 { 6811 struct napi_struct *napi; 6812 int err = 0; 6813 6814 netdev_assert_locked_or_invisible(dev); 6815 6816 if (dev->threaded == threaded) 6817 return 0; 6818 6819 if (threaded) { 6820 list_for_each_entry(napi, &dev->napi_list, dev_list) { 6821 if (!napi->thread) { 6822 err = napi_kthread_create(napi); 6823 if (err) { 6824 threaded = false; 6825 break; 6826 } 6827 } 6828 } 6829 } 6830 6831 WRITE_ONCE(dev->threaded, threaded); 6832 6833 /* Make sure kthread is created before THREADED bit 6834 * is set. 6835 */ 6836 smp_mb__before_atomic(); 6837 6838 /* Setting/unsetting threaded mode on a napi might not immediately 6839 * take effect, if the current napi instance is actively being 6840 * polled. In this case, the switch between threaded mode and 6841 * softirq mode will happen in the next round of napi_schedule(). 6842 * This should not cause hiccups/stalls to the live traffic. 6843 */ 6844 list_for_each_entry(napi, &dev->napi_list, dev_list) 6845 assign_bit(NAPI_STATE_THREADED, &napi->state, threaded); 6846 6847 return err; 6848 } 6849 EXPORT_SYMBOL(dev_set_threaded); 6850 6851 /** 6852 * netif_queue_set_napi - Associate queue with the napi 6853 * @dev: device to which NAPI and queue belong 6854 * @queue_index: Index of queue 6855 * @type: queue type as RX or TX 6856 * @napi: NAPI context, pass NULL to clear previously set NAPI 6857 * 6858 * Set queue with its corresponding napi context. This should be done after 6859 * registering the NAPI handler for the queue-vector and the queues have been 6860 * mapped to the corresponding interrupt vector. 6861 */ 6862 void netif_queue_set_napi(struct net_device *dev, unsigned int queue_index, 6863 enum netdev_queue_type type, struct napi_struct *napi) 6864 { 6865 struct netdev_rx_queue *rxq; 6866 struct netdev_queue *txq; 6867 6868 if (WARN_ON_ONCE(napi && !napi->dev)) 6869 return; 6870 if (dev->reg_state >= NETREG_REGISTERED) 6871 ASSERT_RTNL(); 6872 6873 switch (type) { 6874 case NETDEV_QUEUE_TYPE_RX: 6875 rxq = __netif_get_rx_queue(dev, queue_index); 6876 rxq->napi = napi; 6877 return; 6878 case NETDEV_QUEUE_TYPE_TX: 6879 txq = netdev_get_tx_queue(dev, queue_index); 6880 txq->napi = napi; 6881 return; 6882 default: 6883 return; 6884 } 6885 } 6886 EXPORT_SYMBOL(netif_queue_set_napi); 6887 6888 static void 6889 netif_napi_irq_notify(struct irq_affinity_notify *notify, 6890 const cpumask_t *mask) 6891 { 6892 struct napi_struct *napi = 6893 container_of(notify, struct napi_struct, notify); 6894 #ifdef CONFIG_RFS_ACCEL 6895 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 6896 int err; 6897 #endif 6898 6899 if (napi->config && napi->dev->irq_affinity_auto) 6900 cpumask_copy(&napi->config->affinity_mask, mask); 6901 6902 #ifdef CONFIG_RFS_ACCEL 6903 if (napi->dev->rx_cpu_rmap_auto) { 6904 err = cpu_rmap_update(rmap, napi->napi_rmap_idx, mask); 6905 if (err) 6906 netdev_warn(napi->dev, "RMAP update failed (%d)\n", 6907 err); 6908 } 6909 #endif 6910 } 6911 6912 #ifdef CONFIG_RFS_ACCEL 6913 static void netif_napi_affinity_release(struct kref *ref) 6914 { 6915 struct napi_struct *napi = 6916 container_of(ref, struct napi_struct, notify.kref); 6917 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 6918 6919 netdev_assert_locked(napi->dev); 6920 WARN_ON(test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, 6921 &napi->state)); 6922 6923 if (!napi->dev->rx_cpu_rmap_auto) 6924 return; 6925 rmap->obj[napi->napi_rmap_idx] = NULL; 6926 napi->napi_rmap_idx = -1; 6927 cpu_rmap_put(rmap); 6928 } 6929 6930 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 6931 { 6932 if (dev->rx_cpu_rmap_auto) 6933 return 0; 6934 6935 dev->rx_cpu_rmap = alloc_irq_cpu_rmap(num_irqs); 6936 if (!dev->rx_cpu_rmap) 6937 return -ENOMEM; 6938 6939 dev->rx_cpu_rmap_auto = true; 6940 return 0; 6941 } 6942 EXPORT_SYMBOL(netif_enable_cpu_rmap); 6943 6944 static void netif_del_cpu_rmap(struct net_device *dev) 6945 { 6946 struct cpu_rmap *rmap = dev->rx_cpu_rmap; 6947 6948 if (!dev->rx_cpu_rmap_auto) 6949 return; 6950 6951 /* Free the rmap */ 6952 cpu_rmap_put(rmap); 6953 dev->rx_cpu_rmap = NULL; 6954 dev->rx_cpu_rmap_auto = false; 6955 } 6956 6957 #else 6958 static void netif_napi_affinity_release(struct kref *ref) 6959 { 6960 } 6961 6962 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 6963 { 6964 return 0; 6965 } 6966 EXPORT_SYMBOL(netif_enable_cpu_rmap); 6967 6968 static void netif_del_cpu_rmap(struct net_device *dev) 6969 { 6970 } 6971 #endif 6972 6973 void netif_set_affinity_auto(struct net_device *dev) 6974 { 6975 unsigned int i, maxqs, numa; 6976 6977 maxqs = max(dev->num_tx_queues, dev->num_rx_queues); 6978 numa = dev_to_node(&dev->dev); 6979 6980 for (i = 0; i < maxqs; i++) 6981 cpumask_set_cpu(cpumask_local_spread(i, numa), 6982 &dev->napi_config[i].affinity_mask); 6983 6984 dev->irq_affinity_auto = true; 6985 } 6986 EXPORT_SYMBOL(netif_set_affinity_auto); 6987 6988 void netif_napi_set_irq_locked(struct napi_struct *napi, int irq) 6989 { 6990 int rc; 6991 6992 netdev_assert_locked_or_invisible(napi->dev); 6993 6994 if (napi->irq == irq) 6995 return; 6996 6997 /* Remove existing resources */ 6998 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 6999 irq_set_affinity_notifier(napi->irq, NULL); 7000 7001 napi->irq = irq; 7002 if (irq < 0 || 7003 (!napi->dev->rx_cpu_rmap_auto && !napi->dev->irq_affinity_auto)) 7004 return; 7005 7006 /* Abort for buggy drivers */ 7007 if (napi->dev->irq_affinity_auto && WARN_ON_ONCE(!napi->config)) 7008 return; 7009 7010 #ifdef CONFIG_RFS_ACCEL 7011 if (napi->dev->rx_cpu_rmap_auto) { 7012 rc = cpu_rmap_add(napi->dev->rx_cpu_rmap, napi); 7013 if (rc < 0) 7014 return; 7015 7016 cpu_rmap_get(napi->dev->rx_cpu_rmap); 7017 napi->napi_rmap_idx = rc; 7018 } 7019 #endif 7020 7021 /* Use core IRQ notifier */ 7022 napi->notify.notify = netif_napi_irq_notify; 7023 napi->notify.release = netif_napi_affinity_release; 7024 rc = irq_set_affinity_notifier(irq, &napi->notify); 7025 if (rc) { 7026 netdev_warn(napi->dev, "Unable to set IRQ notifier (%d)\n", 7027 rc); 7028 goto put_rmap; 7029 } 7030 7031 set_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state); 7032 return; 7033 7034 put_rmap: 7035 #ifdef CONFIG_RFS_ACCEL 7036 if (napi->dev->rx_cpu_rmap_auto) { 7037 napi->dev->rx_cpu_rmap->obj[napi->napi_rmap_idx] = NULL; 7038 cpu_rmap_put(napi->dev->rx_cpu_rmap); 7039 napi->napi_rmap_idx = -1; 7040 } 7041 #endif 7042 napi->notify.notify = NULL; 7043 napi->notify.release = NULL; 7044 } 7045 EXPORT_SYMBOL(netif_napi_set_irq_locked); 7046 7047 static void napi_restore_config(struct napi_struct *n) 7048 { 7049 n->defer_hard_irqs = n->config->defer_hard_irqs; 7050 n->gro_flush_timeout = n->config->gro_flush_timeout; 7051 n->irq_suspend_timeout = n->config->irq_suspend_timeout; 7052 7053 if (n->dev->irq_affinity_auto && 7054 test_bit(NAPI_STATE_HAS_NOTIFIER, &n->state)) 7055 irq_set_affinity(n->irq, &n->config->affinity_mask); 7056 7057 /* a NAPI ID might be stored in the config, if so use it. if not, use 7058 * napi_hash_add to generate one for us. 7059 */ 7060 if (n->config->napi_id) { 7061 napi_hash_add_with_id(n, n->config->napi_id); 7062 } else { 7063 napi_hash_add(n); 7064 n->config->napi_id = n->napi_id; 7065 } 7066 } 7067 7068 static void napi_save_config(struct napi_struct *n) 7069 { 7070 n->config->defer_hard_irqs = n->defer_hard_irqs; 7071 n->config->gro_flush_timeout = n->gro_flush_timeout; 7072 n->config->irq_suspend_timeout = n->irq_suspend_timeout; 7073 napi_hash_del(n); 7074 } 7075 7076 /* Netlink wants the NAPI list to be sorted by ID, if adding a NAPI which will 7077 * inherit an existing ID try to insert it at the right position. 7078 */ 7079 static void 7080 netif_napi_dev_list_add(struct net_device *dev, struct napi_struct *napi) 7081 { 7082 unsigned int new_id, pos_id; 7083 struct list_head *higher; 7084 struct napi_struct *pos; 7085 7086 new_id = UINT_MAX; 7087 if (napi->config && napi->config->napi_id) 7088 new_id = napi->config->napi_id; 7089 7090 higher = &dev->napi_list; 7091 list_for_each_entry(pos, &dev->napi_list, dev_list) { 7092 if (napi_id_valid(pos->napi_id)) 7093 pos_id = pos->napi_id; 7094 else if (pos->config) 7095 pos_id = pos->config->napi_id; 7096 else 7097 pos_id = UINT_MAX; 7098 7099 if (pos_id <= new_id) 7100 break; 7101 higher = &pos->dev_list; 7102 } 7103 list_add_rcu(&napi->dev_list, higher); /* adds after higher */ 7104 } 7105 7106 /* Double check that napi_get_frags() allocates skbs with 7107 * skb->head being backed by slab, not a page fragment. 7108 * This is to make sure bug fixed in 3226b158e67c 7109 * ("net: avoid 32 x truesize under-estimation for tiny skbs") 7110 * does not accidentally come back. 7111 */ 7112 static void napi_get_frags_check(struct napi_struct *napi) 7113 { 7114 struct sk_buff *skb; 7115 7116 local_bh_disable(); 7117 skb = napi_get_frags(napi); 7118 WARN_ON_ONCE(skb && skb->head_frag); 7119 napi_free_frags(napi); 7120 local_bh_enable(); 7121 } 7122 7123 void netif_napi_add_weight_locked(struct net_device *dev, 7124 struct napi_struct *napi, 7125 int (*poll)(struct napi_struct *, int), 7126 int weight) 7127 { 7128 netdev_assert_locked(dev); 7129 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state))) 7130 return; 7131 7132 INIT_LIST_HEAD(&napi->poll_list); 7133 INIT_HLIST_NODE(&napi->napi_hash_node); 7134 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); 7135 napi->timer.function = napi_watchdog; 7136 gro_init(&napi->gro); 7137 napi->skb = NULL; 7138 napi->poll = poll; 7139 if (weight > NAPI_POLL_WEIGHT) 7140 netdev_err_once(dev, "%s() called with weight %d\n", __func__, 7141 weight); 7142 napi->weight = weight; 7143 napi->dev = dev; 7144 #ifdef CONFIG_NETPOLL 7145 napi->poll_owner = -1; 7146 #endif 7147 napi->list_owner = -1; 7148 set_bit(NAPI_STATE_SCHED, &napi->state); 7149 set_bit(NAPI_STATE_NPSVC, &napi->state); 7150 netif_napi_dev_list_add(dev, napi); 7151 7152 /* default settings from sysfs are applied to all NAPIs. any per-NAPI 7153 * configuration will be loaded in napi_enable 7154 */ 7155 napi_set_defer_hard_irqs(napi, READ_ONCE(dev->napi_defer_hard_irqs)); 7156 napi_set_gro_flush_timeout(napi, READ_ONCE(dev->gro_flush_timeout)); 7157 7158 napi_get_frags_check(napi); 7159 /* Create kthread for this napi if dev->threaded is set. 7160 * Clear dev->threaded if kthread creation failed so that 7161 * threaded mode will not be enabled in napi_enable(). 7162 */ 7163 if (dev->threaded && napi_kthread_create(napi)) 7164 dev->threaded = false; 7165 netif_napi_set_irq_locked(napi, -1); 7166 } 7167 EXPORT_SYMBOL(netif_napi_add_weight_locked); 7168 7169 void napi_disable_locked(struct napi_struct *n) 7170 { 7171 unsigned long val, new; 7172 7173 might_sleep(); 7174 netdev_assert_locked(n->dev); 7175 7176 set_bit(NAPI_STATE_DISABLE, &n->state); 7177 7178 val = READ_ONCE(n->state); 7179 do { 7180 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) { 7181 usleep_range(20, 200); 7182 val = READ_ONCE(n->state); 7183 } 7184 7185 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC; 7186 new &= ~(NAPIF_STATE_THREADED | NAPIF_STATE_PREFER_BUSY_POLL); 7187 } while (!try_cmpxchg(&n->state, &val, new)); 7188 7189 hrtimer_cancel(&n->timer); 7190 7191 if (n->config) 7192 napi_save_config(n); 7193 else 7194 napi_hash_del(n); 7195 7196 clear_bit(NAPI_STATE_DISABLE, &n->state); 7197 } 7198 EXPORT_SYMBOL(napi_disable_locked); 7199 7200 /** 7201 * napi_disable() - prevent NAPI from scheduling 7202 * @n: NAPI context 7203 * 7204 * Stop NAPI from being scheduled on this context. 7205 * Waits till any outstanding processing completes. 7206 * Takes netdev_lock() for associated net_device. 7207 */ 7208 void napi_disable(struct napi_struct *n) 7209 { 7210 netdev_lock(n->dev); 7211 napi_disable_locked(n); 7212 netdev_unlock(n->dev); 7213 } 7214 EXPORT_SYMBOL(napi_disable); 7215 7216 void napi_enable_locked(struct napi_struct *n) 7217 { 7218 unsigned long new, val = READ_ONCE(n->state); 7219 7220 if (n->config) 7221 napi_restore_config(n); 7222 else 7223 napi_hash_add(n); 7224 7225 do { 7226 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val)); 7227 7228 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC); 7229 if (n->dev->threaded && n->thread) 7230 new |= NAPIF_STATE_THREADED; 7231 } while (!try_cmpxchg(&n->state, &val, new)); 7232 } 7233 EXPORT_SYMBOL(napi_enable_locked); 7234 7235 /** 7236 * napi_enable() - enable NAPI scheduling 7237 * @n: NAPI context 7238 * 7239 * Enable scheduling of a NAPI instance. 7240 * Must be paired with napi_disable(). 7241 * Takes netdev_lock() for associated net_device. 7242 */ 7243 void napi_enable(struct napi_struct *n) 7244 { 7245 netdev_lock(n->dev); 7246 napi_enable_locked(n); 7247 netdev_unlock(n->dev); 7248 } 7249 EXPORT_SYMBOL(napi_enable); 7250 7251 /* Must be called in process context */ 7252 void __netif_napi_del_locked(struct napi_struct *napi) 7253 { 7254 netdev_assert_locked(napi->dev); 7255 7256 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state)) 7257 return; 7258 7259 /* Make sure NAPI is disabled (or was never enabled). */ 7260 WARN_ON(!test_bit(NAPI_STATE_SCHED, &napi->state)); 7261 7262 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7263 irq_set_affinity_notifier(napi->irq, NULL); 7264 7265 if (napi->config) { 7266 napi->index = -1; 7267 napi->config = NULL; 7268 } 7269 7270 list_del_rcu(&napi->dev_list); 7271 napi_free_frags(napi); 7272 7273 gro_cleanup(&napi->gro); 7274 7275 if (napi->thread) { 7276 kthread_stop(napi->thread); 7277 napi->thread = NULL; 7278 } 7279 } 7280 EXPORT_SYMBOL(__netif_napi_del_locked); 7281 7282 static int __napi_poll(struct napi_struct *n, bool *repoll) 7283 { 7284 int work, weight; 7285 7286 weight = n->weight; 7287 7288 /* This NAPI_STATE_SCHED test is for avoiding a race 7289 * with netpoll's poll_napi(). Only the entity which 7290 * obtains the lock and sees NAPI_STATE_SCHED set will 7291 * actually make the ->poll() call. Therefore we avoid 7292 * accidentally calling ->poll() when NAPI is not scheduled. 7293 */ 7294 work = 0; 7295 if (napi_is_scheduled(n)) { 7296 work = n->poll(n, weight); 7297 trace_napi_poll(n, work, weight); 7298 7299 xdp_do_check_flushed(n); 7300 } 7301 7302 if (unlikely(work > weight)) 7303 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n", 7304 n->poll, work, weight); 7305 7306 if (likely(work < weight)) 7307 return work; 7308 7309 /* Drivers must not modify the NAPI state if they 7310 * consume the entire weight. In such cases this code 7311 * still "owns" the NAPI instance and therefore can 7312 * move the instance around on the list at-will. 7313 */ 7314 if (unlikely(napi_disable_pending(n))) { 7315 napi_complete(n); 7316 return work; 7317 } 7318 7319 /* The NAPI context has more processing work, but busy-polling 7320 * is preferred. Exit early. 7321 */ 7322 if (napi_prefer_busy_poll(n)) { 7323 if (napi_complete_done(n, work)) { 7324 /* If timeout is not set, we need to make sure 7325 * that the NAPI is re-scheduled. 7326 */ 7327 napi_schedule(n); 7328 } 7329 return work; 7330 } 7331 7332 /* Flush too old packets. If HZ < 1000, flush all packets */ 7333 gro_flush(&n->gro, HZ >= 1000); 7334 gro_normal_list(&n->gro); 7335 7336 /* Some drivers may have called napi_schedule 7337 * prior to exhausting their budget. 7338 */ 7339 if (unlikely(!list_empty(&n->poll_list))) { 7340 pr_warn_once("%s: Budget exhausted after napi rescheduled\n", 7341 n->dev ? n->dev->name : "backlog"); 7342 return work; 7343 } 7344 7345 *repoll = true; 7346 7347 return work; 7348 } 7349 7350 static int napi_poll(struct napi_struct *n, struct list_head *repoll) 7351 { 7352 bool do_repoll = false; 7353 void *have; 7354 int work; 7355 7356 list_del_init(&n->poll_list); 7357 7358 have = netpoll_poll_lock(n); 7359 7360 work = __napi_poll(n, &do_repoll); 7361 7362 if (do_repoll) 7363 list_add_tail(&n->poll_list, repoll); 7364 7365 netpoll_poll_unlock(have); 7366 7367 return work; 7368 } 7369 7370 static int napi_thread_wait(struct napi_struct *napi) 7371 { 7372 set_current_state(TASK_INTERRUPTIBLE); 7373 7374 while (!kthread_should_stop()) { 7375 /* Testing SCHED_THREADED bit here to make sure the current 7376 * kthread owns this napi and could poll on this napi. 7377 * Testing SCHED bit is not enough because SCHED bit might be 7378 * set by some other busy poll thread or by napi_disable(). 7379 */ 7380 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) { 7381 WARN_ON(!list_empty(&napi->poll_list)); 7382 __set_current_state(TASK_RUNNING); 7383 return 0; 7384 } 7385 7386 schedule(); 7387 set_current_state(TASK_INTERRUPTIBLE); 7388 } 7389 __set_current_state(TASK_RUNNING); 7390 7391 return -1; 7392 } 7393 7394 static void napi_threaded_poll_loop(struct napi_struct *napi) 7395 { 7396 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7397 struct softnet_data *sd; 7398 unsigned long last_qs = jiffies; 7399 7400 for (;;) { 7401 bool repoll = false; 7402 void *have; 7403 7404 local_bh_disable(); 7405 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7406 7407 sd = this_cpu_ptr(&softnet_data); 7408 sd->in_napi_threaded_poll = true; 7409 7410 have = netpoll_poll_lock(napi); 7411 __napi_poll(napi, &repoll); 7412 netpoll_poll_unlock(have); 7413 7414 sd->in_napi_threaded_poll = false; 7415 barrier(); 7416 7417 if (sd_has_rps_ipi_waiting(sd)) { 7418 local_irq_disable(); 7419 net_rps_action_and_irq_enable(sd); 7420 } 7421 skb_defer_free_flush(sd); 7422 bpf_net_ctx_clear(bpf_net_ctx); 7423 local_bh_enable(); 7424 7425 if (!repoll) 7426 break; 7427 7428 rcu_softirq_qs_periodic(last_qs); 7429 cond_resched(); 7430 } 7431 } 7432 7433 static int napi_threaded_poll(void *data) 7434 { 7435 struct napi_struct *napi = data; 7436 7437 while (!napi_thread_wait(napi)) 7438 napi_threaded_poll_loop(napi); 7439 7440 return 0; 7441 } 7442 7443 static __latent_entropy void net_rx_action(void) 7444 { 7445 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 7446 unsigned long time_limit = jiffies + 7447 usecs_to_jiffies(READ_ONCE(net_hotdata.netdev_budget_usecs)); 7448 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7449 int budget = READ_ONCE(net_hotdata.netdev_budget); 7450 LIST_HEAD(list); 7451 LIST_HEAD(repoll); 7452 7453 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7454 start: 7455 sd->in_net_rx_action = true; 7456 local_irq_disable(); 7457 list_splice_init(&sd->poll_list, &list); 7458 local_irq_enable(); 7459 7460 for (;;) { 7461 struct napi_struct *n; 7462 7463 skb_defer_free_flush(sd); 7464 7465 if (list_empty(&list)) { 7466 if (list_empty(&repoll)) { 7467 sd->in_net_rx_action = false; 7468 barrier(); 7469 /* We need to check if ____napi_schedule() 7470 * had refilled poll_list while 7471 * sd->in_net_rx_action was true. 7472 */ 7473 if (!list_empty(&sd->poll_list)) 7474 goto start; 7475 if (!sd_has_rps_ipi_waiting(sd)) 7476 goto end; 7477 } 7478 break; 7479 } 7480 7481 n = list_first_entry(&list, struct napi_struct, poll_list); 7482 budget -= napi_poll(n, &repoll); 7483 7484 /* If softirq window is exhausted then punt. 7485 * Allow this to run for 2 jiffies since which will allow 7486 * an average latency of 1.5/HZ. 7487 */ 7488 if (unlikely(budget <= 0 || 7489 time_after_eq(jiffies, time_limit))) { 7490 sd->time_squeeze++; 7491 break; 7492 } 7493 } 7494 7495 local_irq_disable(); 7496 7497 list_splice_tail_init(&sd->poll_list, &list); 7498 list_splice_tail(&repoll, &list); 7499 list_splice(&list, &sd->poll_list); 7500 if (!list_empty(&sd->poll_list)) 7501 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 7502 else 7503 sd->in_net_rx_action = false; 7504 7505 net_rps_action_and_irq_enable(sd); 7506 end: 7507 bpf_net_ctx_clear(bpf_net_ctx); 7508 } 7509 7510 struct netdev_adjacent { 7511 struct net_device *dev; 7512 netdevice_tracker dev_tracker; 7513 7514 /* upper master flag, there can only be one master device per list */ 7515 bool master; 7516 7517 /* lookup ignore flag */ 7518 bool ignore; 7519 7520 /* counter for the number of times this device was added to us */ 7521 u16 ref_nr; 7522 7523 /* private field for the users */ 7524 void *private; 7525 7526 struct list_head list; 7527 struct rcu_head rcu; 7528 }; 7529 7530 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev, 7531 struct list_head *adj_list) 7532 { 7533 struct netdev_adjacent *adj; 7534 7535 list_for_each_entry(adj, adj_list, list) { 7536 if (adj->dev == adj_dev) 7537 return adj; 7538 } 7539 return NULL; 7540 } 7541 7542 static int ____netdev_has_upper_dev(struct net_device *upper_dev, 7543 struct netdev_nested_priv *priv) 7544 { 7545 struct net_device *dev = (struct net_device *)priv->data; 7546 7547 return upper_dev == dev; 7548 } 7549 7550 /** 7551 * netdev_has_upper_dev - Check if device is linked to an upper device 7552 * @dev: device 7553 * @upper_dev: upper device to check 7554 * 7555 * Find out if a device is linked to specified upper device and return true 7556 * in case it is. Note that this checks only immediate upper device, 7557 * not through a complete stack of devices. The caller must hold the RTNL lock. 7558 */ 7559 bool netdev_has_upper_dev(struct net_device *dev, 7560 struct net_device *upper_dev) 7561 { 7562 struct netdev_nested_priv priv = { 7563 .data = (void *)upper_dev, 7564 }; 7565 7566 ASSERT_RTNL(); 7567 7568 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7569 &priv); 7570 } 7571 EXPORT_SYMBOL(netdev_has_upper_dev); 7572 7573 /** 7574 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device 7575 * @dev: device 7576 * @upper_dev: upper device to check 7577 * 7578 * Find out if a device is linked to specified upper device and return true 7579 * in case it is. Note that this checks the entire upper device chain. 7580 * The caller must hold rcu lock. 7581 */ 7582 7583 bool netdev_has_upper_dev_all_rcu(struct net_device *dev, 7584 struct net_device *upper_dev) 7585 { 7586 struct netdev_nested_priv priv = { 7587 .data = (void *)upper_dev, 7588 }; 7589 7590 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7591 &priv); 7592 } 7593 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu); 7594 7595 /** 7596 * netdev_has_any_upper_dev - Check if device is linked to some device 7597 * @dev: device 7598 * 7599 * Find out if a device is linked to an upper device and return true in case 7600 * it is. The caller must hold the RTNL lock. 7601 */ 7602 bool netdev_has_any_upper_dev(struct net_device *dev) 7603 { 7604 ASSERT_RTNL(); 7605 7606 return !list_empty(&dev->adj_list.upper); 7607 } 7608 EXPORT_SYMBOL(netdev_has_any_upper_dev); 7609 7610 /** 7611 * netdev_master_upper_dev_get - Get master upper device 7612 * @dev: device 7613 * 7614 * Find a master upper device and return pointer to it or NULL in case 7615 * it's not there. The caller must hold the RTNL lock. 7616 */ 7617 struct net_device *netdev_master_upper_dev_get(struct net_device *dev) 7618 { 7619 struct netdev_adjacent *upper; 7620 7621 ASSERT_RTNL(); 7622 7623 if (list_empty(&dev->adj_list.upper)) 7624 return NULL; 7625 7626 upper = list_first_entry(&dev->adj_list.upper, 7627 struct netdev_adjacent, list); 7628 if (likely(upper->master)) 7629 return upper->dev; 7630 return NULL; 7631 } 7632 EXPORT_SYMBOL(netdev_master_upper_dev_get); 7633 7634 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev) 7635 { 7636 struct netdev_adjacent *upper; 7637 7638 ASSERT_RTNL(); 7639 7640 if (list_empty(&dev->adj_list.upper)) 7641 return NULL; 7642 7643 upper = list_first_entry(&dev->adj_list.upper, 7644 struct netdev_adjacent, list); 7645 if (likely(upper->master) && !upper->ignore) 7646 return upper->dev; 7647 return NULL; 7648 } 7649 7650 /** 7651 * netdev_has_any_lower_dev - Check if device is linked to some device 7652 * @dev: device 7653 * 7654 * Find out if a device is linked to a lower device and return true in case 7655 * it is. The caller must hold the RTNL lock. 7656 */ 7657 static bool netdev_has_any_lower_dev(struct net_device *dev) 7658 { 7659 ASSERT_RTNL(); 7660 7661 return !list_empty(&dev->adj_list.lower); 7662 } 7663 7664 void *netdev_adjacent_get_private(struct list_head *adj_list) 7665 { 7666 struct netdev_adjacent *adj; 7667 7668 adj = list_entry(adj_list, struct netdev_adjacent, list); 7669 7670 return adj->private; 7671 } 7672 EXPORT_SYMBOL(netdev_adjacent_get_private); 7673 7674 /** 7675 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list 7676 * @dev: device 7677 * @iter: list_head ** of the current position 7678 * 7679 * Gets the next device from the dev's upper list, starting from iter 7680 * position. The caller must hold RCU read lock. 7681 */ 7682 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev, 7683 struct list_head **iter) 7684 { 7685 struct netdev_adjacent *upper; 7686 7687 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7688 7689 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7690 7691 if (&upper->list == &dev->adj_list.upper) 7692 return NULL; 7693 7694 *iter = &upper->list; 7695 7696 return upper->dev; 7697 } 7698 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu); 7699 7700 static struct net_device *__netdev_next_upper_dev(struct net_device *dev, 7701 struct list_head **iter, 7702 bool *ignore) 7703 { 7704 struct netdev_adjacent *upper; 7705 7706 upper = list_entry((*iter)->next, struct netdev_adjacent, list); 7707 7708 if (&upper->list == &dev->adj_list.upper) 7709 return NULL; 7710 7711 *iter = &upper->list; 7712 *ignore = upper->ignore; 7713 7714 return upper->dev; 7715 } 7716 7717 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev, 7718 struct list_head **iter) 7719 { 7720 struct netdev_adjacent *upper; 7721 7722 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7723 7724 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7725 7726 if (&upper->list == &dev->adj_list.upper) 7727 return NULL; 7728 7729 *iter = &upper->list; 7730 7731 return upper->dev; 7732 } 7733 7734 static int __netdev_walk_all_upper_dev(struct net_device *dev, 7735 int (*fn)(struct net_device *dev, 7736 struct netdev_nested_priv *priv), 7737 struct netdev_nested_priv *priv) 7738 { 7739 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7740 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7741 int ret, cur = 0; 7742 bool ignore; 7743 7744 now = dev; 7745 iter = &dev->adj_list.upper; 7746 7747 while (1) { 7748 if (now != dev) { 7749 ret = fn(now, priv); 7750 if (ret) 7751 return ret; 7752 } 7753 7754 next = NULL; 7755 while (1) { 7756 udev = __netdev_next_upper_dev(now, &iter, &ignore); 7757 if (!udev) 7758 break; 7759 if (ignore) 7760 continue; 7761 7762 next = udev; 7763 niter = &udev->adj_list.upper; 7764 dev_stack[cur] = now; 7765 iter_stack[cur++] = iter; 7766 break; 7767 } 7768 7769 if (!next) { 7770 if (!cur) 7771 return 0; 7772 next = dev_stack[--cur]; 7773 niter = iter_stack[cur]; 7774 } 7775 7776 now = next; 7777 iter = niter; 7778 } 7779 7780 return 0; 7781 } 7782 7783 int netdev_walk_all_upper_dev_rcu(struct net_device *dev, 7784 int (*fn)(struct net_device *dev, 7785 struct netdev_nested_priv *priv), 7786 struct netdev_nested_priv *priv) 7787 { 7788 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7789 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7790 int ret, cur = 0; 7791 7792 now = dev; 7793 iter = &dev->adj_list.upper; 7794 7795 while (1) { 7796 if (now != dev) { 7797 ret = fn(now, priv); 7798 if (ret) 7799 return ret; 7800 } 7801 7802 next = NULL; 7803 while (1) { 7804 udev = netdev_next_upper_dev_rcu(now, &iter); 7805 if (!udev) 7806 break; 7807 7808 next = udev; 7809 niter = &udev->adj_list.upper; 7810 dev_stack[cur] = now; 7811 iter_stack[cur++] = iter; 7812 break; 7813 } 7814 7815 if (!next) { 7816 if (!cur) 7817 return 0; 7818 next = dev_stack[--cur]; 7819 niter = iter_stack[cur]; 7820 } 7821 7822 now = next; 7823 iter = niter; 7824 } 7825 7826 return 0; 7827 } 7828 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu); 7829 7830 static bool __netdev_has_upper_dev(struct net_device *dev, 7831 struct net_device *upper_dev) 7832 { 7833 struct netdev_nested_priv priv = { 7834 .flags = 0, 7835 .data = (void *)upper_dev, 7836 }; 7837 7838 ASSERT_RTNL(); 7839 7840 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev, 7841 &priv); 7842 } 7843 7844 /** 7845 * netdev_lower_get_next_private - Get the next ->private from the 7846 * lower neighbour list 7847 * @dev: device 7848 * @iter: list_head ** of the current position 7849 * 7850 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7851 * list, starting from iter position. The caller must hold either hold the 7852 * RTNL lock or its own locking that guarantees that the neighbour lower 7853 * list will remain unchanged. 7854 */ 7855 void *netdev_lower_get_next_private(struct net_device *dev, 7856 struct list_head **iter) 7857 { 7858 struct netdev_adjacent *lower; 7859 7860 lower = list_entry(*iter, struct netdev_adjacent, list); 7861 7862 if (&lower->list == &dev->adj_list.lower) 7863 return NULL; 7864 7865 *iter = lower->list.next; 7866 7867 return lower->private; 7868 } 7869 EXPORT_SYMBOL(netdev_lower_get_next_private); 7870 7871 /** 7872 * netdev_lower_get_next_private_rcu - Get the next ->private from the 7873 * lower neighbour list, RCU 7874 * variant 7875 * @dev: device 7876 * @iter: list_head ** of the current position 7877 * 7878 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7879 * list, starting from iter position. The caller must hold RCU read lock. 7880 */ 7881 void *netdev_lower_get_next_private_rcu(struct net_device *dev, 7882 struct list_head **iter) 7883 { 7884 struct netdev_adjacent *lower; 7885 7886 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 7887 7888 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7889 7890 if (&lower->list == &dev->adj_list.lower) 7891 return NULL; 7892 7893 *iter = &lower->list; 7894 7895 return lower->private; 7896 } 7897 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu); 7898 7899 /** 7900 * netdev_lower_get_next - Get the next device from the lower neighbour 7901 * list 7902 * @dev: device 7903 * @iter: list_head ** of the current position 7904 * 7905 * Gets the next netdev_adjacent from the dev's lower neighbour 7906 * list, starting from iter position. The caller must hold RTNL lock or 7907 * its own locking that guarantees that the neighbour lower 7908 * list will remain unchanged. 7909 */ 7910 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter) 7911 { 7912 struct netdev_adjacent *lower; 7913 7914 lower = list_entry(*iter, struct netdev_adjacent, list); 7915 7916 if (&lower->list == &dev->adj_list.lower) 7917 return NULL; 7918 7919 *iter = lower->list.next; 7920 7921 return lower->dev; 7922 } 7923 EXPORT_SYMBOL(netdev_lower_get_next); 7924 7925 static struct net_device *netdev_next_lower_dev(struct net_device *dev, 7926 struct list_head **iter) 7927 { 7928 struct netdev_adjacent *lower; 7929 7930 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7931 7932 if (&lower->list == &dev->adj_list.lower) 7933 return NULL; 7934 7935 *iter = &lower->list; 7936 7937 return lower->dev; 7938 } 7939 7940 static struct net_device *__netdev_next_lower_dev(struct net_device *dev, 7941 struct list_head **iter, 7942 bool *ignore) 7943 { 7944 struct netdev_adjacent *lower; 7945 7946 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7947 7948 if (&lower->list == &dev->adj_list.lower) 7949 return NULL; 7950 7951 *iter = &lower->list; 7952 *ignore = lower->ignore; 7953 7954 return lower->dev; 7955 } 7956 7957 int netdev_walk_all_lower_dev(struct net_device *dev, 7958 int (*fn)(struct net_device *dev, 7959 struct netdev_nested_priv *priv), 7960 struct netdev_nested_priv *priv) 7961 { 7962 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7963 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7964 int ret, cur = 0; 7965 7966 now = dev; 7967 iter = &dev->adj_list.lower; 7968 7969 while (1) { 7970 if (now != dev) { 7971 ret = fn(now, priv); 7972 if (ret) 7973 return ret; 7974 } 7975 7976 next = NULL; 7977 while (1) { 7978 ldev = netdev_next_lower_dev(now, &iter); 7979 if (!ldev) 7980 break; 7981 7982 next = ldev; 7983 niter = &ldev->adj_list.lower; 7984 dev_stack[cur] = now; 7985 iter_stack[cur++] = iter; 7986 break; 7987 } 7988 7989 if (!next) { 7990 if (!cur) 7991 return 0; 7992 next = dev_stack[--cur]; 7993 niter = iter_stack[cur]; 7994 } 7995 7996 now = next; 7997 iter = niter; 7998 } 7999 8000 return 0; 8001 } 8002 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev); 8003 8004 static int __netdev_walk_all_lower_dev(struct net_device *dev, 8005 int (*fn)(struct net_device *dev, 8006 struct netdev_nested_priv *priv), 8007 struct netdev_nested_priv *priv) 8008 { 8009 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8010 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8011 int ret, cur = 0; 8012 bool ignore; 8013 8014 now = dev; 8015 iter = &dev->adj_list.lower; 8016 8017 while (1) { 8018 if (now != dev) { 8019 ret = fn(now, priv); 8020 if (ret) 8021 return ret; 8022 } 8023 8024 next = NULL; 8025 while (1) { 8026 ldev = __netdev_next_lower_dev(now, &iter, &ignore); 8027 if (!ldev) 8028 break; 8029 if (ignore) 8030 continue; 8031 8032 next = ldev; 8033 niter = &ldev->adj_list.lower; 8034 dev_stack[cur] = now; 8035 iter_stack[cur++] = iter; 8036 break; 8037 } 8038 8039 if (!next) { 8040 if (!cur) 8041 return 0; 8042 next = dev_stack[--cur]; 8043 niter = iter_stack[cur]; 8044 } 8045 8046 now = next; 8047 iter = niter; 8048 } 8049 8050 return 0; 8051 } 8052 8053 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev, 8054 struct list_head **iter) 8055 { 8056 struct netdev_adjacent *lower; 8057 8058 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8059 if (&lower->list == &dev->adj_list.lower) 8060 return NULL; 8061 8062 *iter = &lower->list; 8063 8064 return lower->dev; 8065 } 8066 EXPORT_SYMBOL(netdev_next_lower_dev_rcu); 8067 8068 static u8 __netdev_upper_depth(struct net_device *dev) 8069 { 8070 struct net_device *udev; 8071 struct list_head *iter; 8072 u8 max_depth = 0; 8073 bool ignore; 8074 8075 for (iter = &dev->adj_list.upper, 8076 udev = __netdev_next_upper_dev(dev, &iter, &ignore); 8077 udev; 8078 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) { 8079 if (ignore) 8080 continue; 8081 if (max_depth < udev->upper_level) 8082 max_depth = udev->upper_level; 8083 } 8084 8085 return max_depth; 8086 } 8087 8088 static u8 __netdev_lower_depth(struct net_device *dev) 8089 { 8090 struct net_device *ldev; 8091 struct list_head *iter; 8092 u8 max_depth = 0; 8093 bool ignore; 8094 8095 for (iter = &dev->adj_list.lower, 8096 ldev = __netdev_next_lower_dev(dev, &iter, &ignore); 8097 ldev; 8098 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) { 8099 if (ignore) 8100 continue; 8101 if (max_depth < ldev->lower_level) 8102 max_depth = ldev->lower_level; 8103 } 8104 8105 return max_depth; 8106 } 8107 8108 static int __netdev_update_upper_level(struct net_device *dev, 8109 struct netdev_nested_priv *__unused) 8110 { 8111 dev->upper_level = __netdev_upper_depth(dev) + 1; 8112 return 0; 8113 } 8114 8115 #ifdef CONFIG_LOCKDEP 8116 static LIST_HEAD(net_unlink_list); 8117 8118 static void net_unlink_todo(struct net_device *dev) 8119 { 8120 if (list_empty(&dev->unlink_list)) 8121 list_add_tail(&dev->unlink_list, &net_unlink_list); 8122 } 8123 #endif 8124 8125 static int __netdev_update_lower_level(struct net_device *dev, 8126 struct netdev_nested_priv *priv) 8127 { 8128 dev->lower_level = __netdev_lower_depth(dev) + 1; 8129 8130 #ifdef CONFIG_LOCKDEP 8131 if (!priv) 8132 return 0; 8133 8134 if (priv->flags & NESTED_SYNC_IMM) 8135 dev->nested_level = dev->lower_level - 1; 8136 if (priv->flags & NESTED_SYNC_TODO) 8137 net_unlink_todo(dev); 8138 #endif 8139 return 0; 8140 } 8141 8142 int netdev_walk_all_lower_dev_rcu(struct net_device *dev, 8143 int (*fn)(struct net_device *dev, 8144 struct netdev_nested_priv *priv), 8145 struct netdev_nested_priv *priv) 8146 { 8147 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8148 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8149 int ret, cur = 0; 8150 8151 now = dev; 8152 iter = &dev->adj_list.lower; 8153 8154 while (1) { 8155 if (now != dev) { 8156 ret = fn(now, priv); 8157 if (ret) 8158 return ret; 8159 } 8160 8161 next = NULL; 8162 while (1) { 8163 ldev = netdev_next_lower_dev_rcu(now, &iter); 8164 if (!ldev) 8165 break; 8166 8167 next = ldev; 8168 niter = &ldev->adj_list.lower; 8169 dev_stack[cur] = now; 8170 iter_stack[cur++] = iter; 8171 break; 8172 } 8173 8174 if (!next) { 8175 if (!cur) 8176 return 0; 8177 next = dev_stack[--cur]; 8178 niter = iter_stack[cur]; 8179 } 8180 8181 now = next; 8182 iter = niter; 8183 } 8184 8185 return 0; 8186 } 8187 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu); 8188 8189 /** 8190 * netdev_lower_get_first_private_rcu - Get the first ->private from the 8191 * lower neighbour list, RCU 8192 * variant 8193 * @dev: device 8194 * 8195 * Gets the first netdev_adjacent->private from the dev's lower neighbour 8196 * list. The caller must hold RCU read lock. 8197 */ 8198 void *netdev_lower_get_first_private_rcu(struct net_device *dev) 8199 { 8200 struct netdev_adjacent *lower; 8201 8202 lower = list_first_or_null_rcu(&dev->adj_list.lower, 8203 struct netdev_adjacent, list); 8204 if (lower) 8205 return lower->private; 8206 return NULL; 8207 } 8208 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu); 8209 8210 /** 8211 * netdev_master_upper_dev_get_rcu - Get master upper device 8212 * @dev: device 8213 * 8214 * Find a master upper device and return pointer to it or NULL in case 8215 * it's not there. The caller must hold the RCU read lock. 8216 */ 8217 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev) 8218 { 8219 struct netdev_adjacent *upper; 8220 8221 upper = list_first_or_null_rcu(&dev->adj_list.upper, 8222 struct netdev_adjacent, list); 8223 if (upper && likely(upper->master)) 8224 return upper->dev; 8225 return NULL; 8226 } 8227 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu); 8228 8229 static int netdev_adjacent_sysfs_add(struct net_device *dev, 8230 struct net_device *adj_dev, 8231 struct list_head *dev_list) 8232 { 8233 char linkname[IFNAMSIZ+7]; 8234 8235 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8236 "upper_%s" : "lower_%s", adj_dev->name); 8237 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj), 8238 linkname); 8239 } 8240 static void netdev_adjacent_sysfs_del(struct net_device *dev, 8241 char *name, 8242 struct list_head *dev_list) 8243 { 8244 char linkname[IFNAMSIZ+7]; 8245 8246 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8247 "upper_%s" : "lower_%s", name); 8248 sysfs_remove_link(&(dev->dev.kobj), linkname); 8249 } 8250 8251 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev, 8252 struct net_device *adj_dev, 8253 struct list_head *dev_list) 8254 { 8255 return (dev_list == &dev->adj_list.upper || 8256 dev_list == &dev->adj_list.lower) && 8257 net_eq(dev_net(dev), dev_net(adj_dev)); 8258 } 8259 8260 static int __netdev_adjacent_dev_insert(struct net_device *dev, 8261 struct net_device *adj_dev, 8262 struct list_head *dev_list, 8263 void *private, bool master) 8264 { 8265 struct netdev_adjacent *adj; 8266 int ret; 8267 8268 adj = __netdev_find_adj(adj_dev, dev_list); 8269 8270 if (adj) { 8271 adj->ref_nr += 1; 8272 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n", 8273 dev->name, adj_dev->name, adj->ref_nr); 8274 8275 return 0; 8276 } 8277 8278 adj = kmalloc(sizeof(*adj), GFP_KERNEL); 8279 if (!adj) 8280 return -ENOMEM; 8281 8282 adj->dev = adj_dev; 8283 adj->master = master; 8284 adj->ref_nr = 1; 8285 adj->private = private; 8286 adj->ignore = false; 8287 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL); 8288 8289 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n", 8290 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name); 8291 8292 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) { 8293 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list); 8294 if (ret) 8295 goto free_adj; 8296 } 8297 8298 /* Ensure that master link is always the first item in list. */ 8299 if (master) { 8300 ret = sysfs_create_link(&(dev->dev.kobj), 8301 &(adj_dev->dev.kobj), "master"); 8302 if (ret) 8303 goto remove_symlinks; 8304 8305 list_add_rcu(&adj->list, dev_list); 8306 } else { 8307 list_add_tail_rcu(&adj->list, dev_list); 8308 } 8309 8310 return 0; 8311 8312 remove_symlinks: 8313 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8314 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8315 free_adj: 8316 netdev_put(adj_dev, &adj->dev_tracker); 8317 kfree(adj); 8318 8319 return ret; 8320 } 8321 8322 static void __netdev_adjacent_dev_remove(struct net_device *dev, 8323 struct net_device *adj_dev, 8324 u16 ref_nr, 8325 struct list_head *dev_list) 8326 { 8327 struct netdev_adjacent *adj; 8328 8329 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n", 8330 dev->name, adj_dev->name, ref_nr); 8331 8332 adj = __netdev_find_adj(adj_dev, dev_list); 8333 8334 if (!adj) { 8335 pr_err("Adjacency does not exist for device %s from %s\n", 8336 dev->name, adj_dev->name); 8337 WARN_ON(1); 8338 return; 8339 } 8340 8341 if (adj->ref_nr > ref_nr) { 8342 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n", 8343 dev->name, adj_dev->name, ref_nr, 8344 adj->ref_nr - ref_nr); 8345 adj->ref_nr -= ref_nr; 8346 return; 8347 } 8348 8349 if (adj->master) 8350 sysfs_remove_link(&(dev->dev.kobj), "master"); 8351 8352 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8353 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8354 8355 list_del_rcu(&adj->list); 8356 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n", 8357 adj_dev->name, dev->name, adj_dev->name); 8358 netdev_put(adj_dev, &adj->dev_tracker); 8359 kfree_rcu(adj, rcu); 8360 } 8361 8362 static int __netdev_adjacent_dev_link_lists(struct net_device *dev, 8363 struct net_device *upper_dev, 8364 struct list_head *up_list, 8365 struct list_head *down_list, 8366 void *private, bool master) 8367 { 8368 int ret; 8369 8370 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list, 8371 private, master); 8372 if (ret) 8373 return ret; 8374 8375 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list, 8376 private, false); 8377 if (ret) { 8378 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list); 8379 return ret; 8380 } 8381 8382 return 0; 8383 } 8384 8385 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev, 8386 struct net_device *upper_dev, 8387 u16 ref_nr, 8388 struct list_head *up_list, 8389 struct list_head *down_list) 8390 { 8391 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list); 8392 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list); 8393 } 8394 8395 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev, 8396 struct net_device *upper_dev, 8397 void *private, bool master) 8398 { 8399 return __netdev_adjacent_dev_link_lists(dev, upper_dev, 8400 &dev->adj_list.upper, 8401 &upper_dev->adj_list.lower, 8402 private, master); 8403 } 8404 8405 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev, 8406 struct net_device *upper_dev) 8407 { 8408 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1, 8409 &dev->adj_list.upper, 8410 &upper_dev->adj_list.lower); 8411 } 8412 8413 static int __netdev_upper_dev_link(struct net_device *dev, 8414 struct net_device *upper_dev, bool master, 8415 void *upper_priv, void *upper_info, 8416 struct netdev_nested_priv *priv, 8417 struct netlink_ext_ack *extack) 8418 { 8419 struct netdev_notifier_changeupper_info changeupper_info = { 8420 .info = { 8421 .dev = dev, 8422 .extack = extack, 8423 }, 8424 .upper_dev = upper_dev, 8425 .master = master, 8426 .linking = true, 8427 .upper_info = upper_info, 8428 }; 8429 struct net_device *master_dev; 8430 int ret = 0; 8431 8432 ASSERT_RTNL(); 8433 8434 if (dev == upper_dev) 8435 return -EBUSY; 8436 8437 /* To prevent loops, check if dev is not upper device to upper_dev. */ 8438 if (__netdev_has_upper_dev(upper_dev, dev)) 8439 return -EBUSY; 8440 8441 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV) 8442 return -EMLINK; 8443 8444 if (!master) { 8445 if (__netdev_has_upper_dev(dev, upper_dev)) 8446 return -EEXIST; 8447 } else { 8448 master_dev = __netdev_master_upper_dev_get(dev); 8449 if (master_dev) 8450 return master_dev == upper_dev ? -EEXIST : -EBUSY; 8451 } 8452 8453 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8454 &changeupper_info.info); 8455 ret = notifier_to_errno(ret); 8456 if (ret) 8457 return ret; 8458 8459 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv, 8460 master); 8461 if (ret) 8462 return ret; 8463 8464 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8465 &changeupper_info.info); 8466 ret = notifier_to_errno(ret); 8467 if (ret) 8468 goto rollback; 8469 8470 __netdev_update_upper_level(dev, NULL); 8471 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8472 8473 __netdev_update_lower_level(upper_dev, priv); 8474 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8475 priv); 8476 8477 return 0; 8478 8479 rollback: 8480 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8481 8482 return ret; 8483 } 8484 8485 /** 8486 * netdev_upper_dev_link - Add a link to the upper device 8487 * @dev: device 8488 * @upper_dev: new upper device 8489 * @extack: netlink extended ack 8490 * 8491 * Adds a link to device which is upper to this one. The caller must hold 8492 * the RTNL lock. On a failure a negative errno code is returned. 8493 * On success the reference counts are adjusted and the function 8494 * returns zero. 8495 */ 8496 int netdev_upper_dev_link(struct net_device *dev, 8497 struct net_device *upper_dev, 8498 struct netlink_ext_ack *extack) 8499 { 8500 struct netdev_nested_priv priv = { 8501 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8502 .data = NULL, 8503 }; 8504 8505 return __netdev_upper_dev_link(dev, upper_dev, false, 8506 NULL, NULL, &priv, extack); 8507 } 8508 EXPORT_SYMBOL(netdev_upper_dev_link); 8509 8510 /** 8511 * netdev_master_upper_dev_link - Add a master link to the upper device 8512 * @dev: device 8513 * @upper_dev: new upper device 8514 * @upper_priv: upper device private 8515 * @upper_info: upper info to be passed down via notifier 8516 * @extack: netlink extended ack 8517 * 8518 * Adds a link to device which is upper to this one. In this case, only 8519 * one master upper device can be linked, although other non-master devices 8520 * might be linked as well. The caller must hold the RTNL lock. 8521 * On a failure a negative errno code is returned. On success the reference 8522 * counts are adjusted and the function returns zero. 8523 */ 8524 int netdev_master_upper_dev_link(struct net_device *dev, 8525 struct net_device *upper_dev, 8526 void *upper_priv, void *upper_info, 8527 struct netlink_ext_ack *extack) 8528 { 8529 struct netdev_nested_priv priv = { 8530 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8531 .data = NULL, 8532 }; 8533 8534 return __netdev_upper_dev_link(dev, upper_dev, true, 8535 upper_priv, upper_info, &priv, extack); 8536 } 8537 EXPORT_SYMBOL(netdev_master_upper_dev_link); 8538 8539 static void __netdev_upper_dev_unlink(struct net_device *dev, 8540 struct net_device *upper_dev, 8541 struct netdev_nested_priv *priv) 8542 { 8543 struct netdev_notifier_changeupper_info changeupper_info = { 8544 .info = { 8545 .dev = dev, 8546 }, 8547 .upper_dev = upper_dev, 8548 .linking = false, 8549 }; 8550 8551 ASSERT_RTNL(); 8552 8553 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev; 8554 8555 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8556 &changeupper_info.info); 8557 8558 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8559 8560 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8561 &changeupper_info.info); 8562 8563 __netdev_update_upper_level(dev, NULL); 8564 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8565 8566 __netdev_update_lower_level(upper_dev, priv); 8567 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8568 priv); 8569 } 8570 8571 /** 8572 * netdev_upper_dev_unlink - Removes a link to upper device 8573 * @dev: device 8574 * @upper_dev: new upper device 8575 * 8576 * Removes a link to device which is upper to this one. The caller must hold 8577 * the RTNL lock. 8578 */ 8579 void netdev_upper_dev_unlink(struct net_device *dev, 8580 struct net_device *upper_dev) 8581 { 8582 struct netdev_nested_priv priv = { 8583 .flags = NESTED_SYNC_TODO, 8584 .data = NULL, 8585 }; 8586 8587 __netdev_upper_dev_unlink(dev, upper_dev, &priv); 8588 } 8589 EXPORT_SYMBOL(netdev_upper_dev_unlink); 8590 8591 static void __netdev_adjacent_dev_set(struct net_device *upper_dev, 8592 struct net_device *lower_dev, 8593 bool val) 8594 { 8595 struct netdev_adjacent *adj; 8596 8597 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower); 8598 if (adj) 8599 adj->ignore = val; 8600 8601 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper); 8602 if (adj) 8603 adj->ignore = val; 8604 } 8605 8606 static void netdev_adjacent_dev_disable(struct net_device *upper_dev, 8607 struct net_device *lower_dev) 8608 { 8609 __netdev_adjacent_dev_set(upper_dev, lower_dev, true); 8610 } 8611 8612 static void netdev_adjacent_dev_enable(struct net_device *upper_dev, 8613 struct net_device *lower_dev) 8614 { 8615 __netdev_adjacent_dev_set(upper_dev, lower_dev, false); 8616 } 8617 8618 int netdev_adjacent_change_prepare(struct net_device *old_dev, 8619 struct net_device *new_dev, 8620 struct net_device *dev, 8621 struct netlink_ext_ack *extack) 8622 { 8623 struct netdev_nested_priv priv = { 8624 .flags = 0, 8625 .data = NULL, 8626 }; 8627 int err; 8628 8629 if (!new_dev) 8630 return 0; 8631 8632 if (old_dev && new_dev != old_dev) 8633 netdev_adjacent_dev_disable(dev, old_dev); 8634 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv, 8635 extack); 8636 if (err) { 8637 if (old_dev && new_dev != old_dev) 8638 netdev_adjacent_dev_enable(dev, old_dev); 8639 return err; 8640 } 8641 8642 return 0; 8643 } 8644 EXPORT_SYMBOL(netdev_adjacent_change_prepare); 8645 8646 void netdev_adjacent_change_commit(struct net_device *old_dev, 8647 struct net_device *new_dev, 8648 struct net_device *dev) 8649 { 8650 struct netdev_nested_priv priv = { 8651 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8652 .data = NULL, 8653 }; 8654 8655 if (!new_dev || !old_dev) 8656 return; 8657 8658 if (new_dev == old_dev) 8659 return; 8660 8661 netdev_adjacent_dev_enable(dev, old_dev); 8662 __netdev_upper_dev_unlink(old_dev, dev, &priv); 8663 } 8664 EXPORT_SYMBOL(netdev_adjacent_change_commit); 8665 8666 void netdev_adjacent_change_abort(struct net_device *old_dev, 8667 struct net_device *new_dev, 8668 struct net_device *dev) 8669 { 8670 struct netdev_nested_priv priv = { 8671 .flags = 0, 8672 .data = NULL, 8673 }; 8674 8675 if (!new_dev) 8676 return; 8677 8678 if (old_dev && new_dev != old_dev) 8679 netdev_adjacent_dev_enable(dev, old_dev); 8680 8681 __netdev_upper_dev_unlink(new_dev, dev, &priv); 8682 } 8683 EXPORT_SYMBOL(netdev_adjacent_change_abort); 8684 8685 /** 8686 * netdev_bonding_info_change - Dispatch event about slave change 8687 * @dev: device 8688 * @bonding_info: info to dispatch 8689 * 8690 * Send NETDEV_BONDING_INFO to netdev notifiers with info. 8691 * The caller must hold the RTNL lock. 8692 */ 8693 void netdev_bonding_info_change(struct net_device *dev, 8694 struct netdev_bonding_info *bonding_info) 8695 { 8696 struct netdev_notifier_bonding_info info = { 8697 .info.dev = dev, 8698 }; 8699 8700 memcpy(&info.bonding_info, bonding_info, 8701 sizeof(struct netdev_bonding_info)); 8702 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, 8703 &info.info); 8704 } 8705 EXPORT_SYMBOL(netdev_bonding_info_change); 8706 8707 static int netdev_offload_xstats_enable_l3(struct net_device *dev, 8708 struct netlink_ext_ack *extack) 8709 { 8710 struct netdev_notifier_offload_xstats_info info = { 8711 .info.dev = dev, 8712 .info.extack = extack, 8713 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8714 }; 8715 int err; 8716 int rc; 8717 8718 dev->offload_xstats_l3 = kzalloc(sizeof(*dev->offload_xstats_l3), 8719 GFP_KERNEL); 8720 if (!dev->offload_xstats_l3) 8721 return -ENOMEM; 8722 8723 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE, 8724 NETDEV_OFFLOAD_XSTATS_DISABLE, 8725 &info.info); 8726 err = notifier_to_errno(rc); 8727 if (err) 8728 goto free_stats; 8729 8730 return 0; 8731 8732 free_stats: 8733 kfree(dev->offload_xstats_l3); 8734 dev->offload_xstats_l3 = NULL; 8735 return err; 8736 } 8737 8738 int netdev_offload_xstats_enable(struct net_device *dev, 8739 enum netdev_offload_xstats_type type, 8740 struct netlink_ext_ack *extack) 8741 { 8742 ASSERT_RTNL(); 8743 8744 if (netdev_offload_xstats_enabled(dev, type)) 8745 return -EALREADY; 8746 8747 switch (type) { 8748 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8749 return netdev_offload_xstats_enable_l3(dev, extack); 8750 } 8751 8752 WARN_ON(1); 8753 return -EINVAL; 8754 } 8755 EXPORT_SYMBOL(netdev_offload_xstats_enable); 8756 8757 static void netdev_offload_xstats_disable_l3(struct net_device *dev) 8758 { 8759 struct netdev_notifier_offload_xstats_info info = { 8760 .info.dev = dev, 8761 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8762 }; 8763 8764 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE, 8765 &info.info); 8766 kfree(dev->offload_xstats_l3); 8767 dev->offload_xstats_l3 = NULL; 8768 } 8769 8770 int netdev_offload_xstats_disable(struct net_device *dev, 8771 enum netdev_offload_xstats_type type) 8772 { 8773 ASSERT_RTNL(); 8774 8775 if (!netdev_offload_xstats_enabled(dev, type)) 8776 return -EALREADY; 8777 8778 switch (type) { 8779 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8780 netdev_offload_xstats_disable_l3(dev); 8781 return 0; 8782 } 8783 8784 WARN_ON(1); 8785 return -EINVAL; 8786 } 8787 EXPORT_SYMBOL(netdev_offload_xstats_disable); 8788 8789 static void netdev_offload_xstats_disable_all(struct net_device *dev) 8790 { 8791 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3); 8792 } 8793 8794 static struct rtnl_hw_stats64 * 8795 netdev_offload_xstats_get_ptr(const struct net_device *dev, 8796 enum netdev_offload_xstats_type type) 8797 { 8798 switch (type) { 8799 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8800 return dev->offload_xstats_l3; 8801 } 8802 8803 WARN_ON(1); 8804 return NULL; 8805 } 8806 8807 bool netdev_offload_xstats_enabled(const struct net_device *dev, 8808 enum netdev_offload_xstats_type type) 8809 { 8810 ASSERT_RTNL(); 8811 8812 return netdev_offload_xstats_get_ptr(dev, type); 8813 } 8814 EXPORT_SYMBOL(netdev_offload_xstats_enabled); 8815 8816 struct netdev_notifier_offload_xstats_ru { 8817 bool used; 8818 }; 8819 8820 struct netdev_notifier_offload_xstats_rd { 8821 struct rtnl_hw_stats64 stats; 8822 bool used; 8823 }; 8824 8825 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest, 8826 const struct rtnl_hw_stats64 *src) 8827 { 8828 dest->rx_packets += src->rx_packets; 8829 dest->tx_packets += src->tx_packets; 8830 dest->rx_bytes += src->rx_bytes; 8831 dest->tx_bytes += src->tx_bytes; 8832 dest->rx_errors += src->rx_errors; 8833 dest->tx_errors += src->tx_errors; 8834 dest->rx_dropped += src->rx_dropped; 8835 dest->tx_dropped += src->tx_dropped; 8836 dest->multicast += src->multicast; 8837 } 8838 8839 static int netdev_offload_xstats_get_used(struct net_device *dev, 8840 enum netdev_offload_xstats_type type, 8841 bool *p_used, 8842 struct netlink_ext_ack *extack) 8843 { 8844 struct netdev_notifier_offload_xstats_ru report_used = {}; 8845 struct netdev_notifier_offload_xstats_info info = { 8846 .info.dev = dev, 8847 .info.extack = extack, 8848 .type = type, 8849 .report_used = &report_used, 8850 }; 8851 int rc; 8852 8853 WARN_ON(!netdev_offload_xstats_enabled(dev, type)); 8854 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED, 8855 &info.info); 8856 *p_used = report_used.used; 8857 return notifier_to_errno(rc); 8858 } 8859 8860 static int netdev_offload_xstats_get_stats(struct net_device *dev, 8861 enum netdev_offload_xstats_type type, 8862 struct rtnl_hw_stats64 *p_stats, 8863 bool *p_used, 8864 struct netlink_ext_ack *extack) 8865 { 8866 struct netdev_notifier_offload_xstats_rd report_delta = {}; 8867 struct netdev_notifier_offload_xstats_info info = { 8868 .info.dev = dev, 8869 .info.extack = extack, 8870 .type = type, 8871 .report_delta = &report_delta, 8872 }; 8873 struct rtnl_hw_stats64 *stats; 8874 int rc; 8875 8876 stats = netdev_offload_xstats_get_ptr(dev, type); 8877 if (WARN_ON(!stats)) 8878 return -EINVAL; 8879 8880 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA, 8881 &info.info); 8882 8883 /* Cache whatever we got, even if there was an error, otherwise the 8884 * successful stats retrievals would get lost. 8885 */ 8886 netdev_hw_stats64_add(stats, &report_delta.stats); 8887 8888 if (p_stats) 8889 *p_stats = *stats; 8890 *p_used = report_delta.used; 8891 8892 return notifier_to_errno(rc); 8893 } 8894 8895 int netdev_offload_xstats_get(struct net_device *dev, 8896 enum netdev_offload_xstats_type type, 8897 struct rtnl_hw_stats64 *p_stats, bool *p_used, 8898 struct netlink_ext_ack *extack) 8899 { 8900 ASSERT_RTNL(); 8901 8902 if (p_stats) 8903 return netdev_offload_xstats_get_stats(dev, type, p_stats, 8904 p_used, extack); 8905 else 8906 return netdev_offload_xstats_get_used(dev, type, p_used, 8907 extack); 8908 } 8909 EXPORT_SYMBOL(netdev_offload_xstats_get); 8910 8911 void 8912 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta, 8913 const struct rtnl_hw_stats64 *stats) 8914 { 8915 report_delta->used = true; 8916 netdev_hw_stats64_add(&report_delta->stats, stats); 8917 } 8918 EXPORT_SYMBOL(netdev_offload_xstats_report_delta); 8919 8920 void 8921 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used) 8922 { 8923 report_used->used = true; 8924 } 8925 EXPORT_SYMBOL(netdev_offload_xstats_report_used); 8926 8927 void netdev_offload_xstats_push_delta(struct net_device *dev, 8928 enum netdev_offload_xstats_type type, 8929 const struct rtnl_hw_stats64 *p_stats) 8930 { 8931 struct rtnl_hw_stats64 *stats; 8932 8933 ASSERT_RTNL(); 8934 8935 stats = netdev_offload_xstats_get_ptr(dev, type); 8936 if (WARN_ON(!stats)) 8937 return; 8938 8939 netdev_hw_stats64_add(stats, p_stats); 8940 } 8941 EXPORT_SYMBOL(netdev_offload_xstats_push_delta); 8942 8943 /** 8944 * netdev_get_xmit_slave - Get the xmit slave of master device 8945 * @dev: device 8946 * @skb: The packet 8947 * @all_slaves: assume all the slaves are active 8948 * 8949 * The reference counters are not incremented so the caller must be 8950 * careful with locks. The caller must hold RCU lock. 8951 * %NULL is returned if no slave is found. 8952 */ 8953 8954 struct net_device *netdev_get_xmit_slave(struct net_device *dev, 8955 struct sk_buff *skb, 8956 bool all_slaves) 8957 { 8958 const struct net_device_ops *ops = dev->netdev_ops; 8959 8960 if (!ops->ndo_get_xmit_slave) 8961 return NULL; 8962 return ops->ndo_get_xmit_slave(dev, skb, all_slaves); 8963 } 8964 EXPORT_SYMBOL(netdev_get_xmit_slave); 8965 8966 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev, 8967 struct sock *sk) 8968 { 8969 const struct net_device_ops *ops = dev->netdev_ops; 8970 8971 if (!ops->ndo_sk_get_lower_dev) 8972 return NULL; 8973 return ops->ndo_sk_get_lower_dev(dev, sk); 8974 } 8975 8976 /** 8977 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket 8978 * @dev: device 8979 * @sk: the socket 8980 * 8981 * %NULL is returned if no lower device is found. 8982 */ 8983 8984 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev, 8985 struct sock *sk) 8986 { 8987 struct net_device *lower; 8988 8989 lower = netdev_sk_get_lower_dev(dev, sk); 8990 while (lower) { 8991 dev = lower; 8992 lower = netdev_sk_get_lower_dev(dev, sk); 8993 } 8994 8995 return dev; 8996 } 8997 EXPORT_SYMBOL(netdev_sk_get_lowest_dev); 8998 8999 static void netdev_adjacent_add_links(struct net_device *dev) 9000 { 9001 struct netdev_adjacent *iter; 9002 9003 struct net *net = dev_net(dev); 9004 9005 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9006 if (!net_eq(net, dev_net(iter->dev))) 9007 continue; 9008 netdev_adjacent_sysfs_add(iter->dev, dev, 9009 &iter->dev->adj_list.lower); 9010 netdev_adjacent_sysfs_add(dev, iter->dev, 9011 &dev->adj_list.upper); 9012 } 9013 9014 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9015 if (!net_eq(net, dev_net(iter->dev))) 9016 continue; 9017 netdev_adjacent_sysfs_add(iter->dev, dev, 9018 &iter->dev->adj_list.upper); 9019 netdev_adjacent_sysfs_add(dev, iter->dev, 9020 &dev->adj_list.lower); 9021 } 9022 } 9023 9024 static void netdev_adjacent_del_links(struct net_device *dev) 9025 { 9026 struct netdev_adjacent *iter; 9027 9028 struct net *net = dev_net(dev); 9029 9030 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9031 if (!net_eq(net, dev_net(iter->dev))) 9032 continue; 9033 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9034 &iter->dev->adj_list.lower); 9035 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9036 &dev->adj_list.upper); 9037 } 9038 9039 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9040 if (!net_eq(net, dev_net(iter->dev))) 9041 continue; 9042 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9043 &iter->dev->adj_list.upper); 9044 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9045 &dev->adj_list.lower); 9046 } 9047 } 9048 9049 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname) 9050 { 9051 struct netdev_adjacent *iter; 9052 9053 struct net *net = dev_net(dev); 9054 9055 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9056 if (!net_eq(net, dev_net(iter->dev))) 9057 continue; 9058 netdev_adjacent_sysfs_del(iter->dev, oldname, 9059 &iter->dev->adj_list.lower); 9060 netdev_adjacent_sysfs_add(iter->dev, dev, 9061 &iter->dev->adj_list.lower); 9062 } 9063 9064 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9065 if (!net_eq(net, dev_net(iter->dev))) 9066 continue; 9067 netdev_adjacent_sysfs_del(iter->dev, oldname, 9068 &iter->dev->adj_list.upper); 9069 netdev_adjacent_sysfs_add(iter->dev, dev, 9070 &iter->dev->adj_list.upper); 9071 } 9072 } 9073 9074 void *netdev_lower_dev_get_private(struct net_device *dev, 9075 struct net_device *lower_dev) 9076 { 9077 struct netdev_adjacent *lower; 9078 9079 if (!lower_dev) 9080 return NULL; 9081 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower); 9082 if (!lower) 9083 return NULL; 9084 9085 return lower->private; 9086 } 9087 EXPORT_SYMBOL(netdev_lower_dev_get_private); 9088 9089 9090 /** 9091 * netdev_lower_state_changed - Dispatch event about lower device state change 9092 * @lower_dev: device 9093 * @lower_state_info: state to dispatch 9094 * 9095 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info. 9096 * The caller must hold the RTNL lock. 9097 */ 9098 void netdev_lower_state_changed(struct net_device *lower_dev, 9099 void *lower_state_info) 9100 { 9101 struct netdev_notifier_changelowerstate_info changelowerstate_info = { 9102 .info.dev = lower_dev, 9103 }; 9104 9105 ASSERT_RTNL(); 9106 changelowerstate_info.lower_state_info = lower_state_info; 9107 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, 9108 &changelowerstate_info.info); 9109 } 9110 EXPORT_SYMBOL(netdev_lower_state_changed); 9111 9112 static void dev_change_rx_flags(struct net_device *dev, int flags) 9113 { 9114 const struct net_device_ops *ops = dev->netdev_ops; 9115 9116 if (ops->ndo_change_rx_flags) 9117 ops->ndo_change_rx_flags(dev, flags); 9118 } 9119 9120 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify) 9121 { 9122 unsigned int old_flags = dev->flags; 9123 unsigned int promiscuity, flags; 9124 kuid_t uid; 9125 kgid_t gid; 9126 9127 ASSERT_RTNL(); 9128 9129 promiscuity = dev->promiscuity + inc; 9130 if (promiscuity == 0) { 9131 /* 9132 * Avoid overflow. 9133 * If inc causes overflow, untouch promisc and return error. 9134 */ 9135 if (unlikely(inc > 0)) { 9136 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n"); 9137 return -EOVERFLOW; 9138 } 9139 flags = old_flags & ~IFF_PROMISC; 9140 } else { 9141 flags = old_flags | IFF_PROMISC; 9142 } 9143 WRITE_ONCE(dev->promiscuity, promiscuity); 9144 if (flags != old_flags) { 9145 WRITE_ONCE(dev->flags, flags); 9146 netdev_info(dev, "%s promiscuous mode\n", 9147 dev->flags & IFF_PROMISC ? "entered" : "left"); 9148 if (audit_enabled) { 9149 current_uid_gid(&uid, &gid); 9150 audit_log(audit_context(), GFP_ATOMIC, 9151 AUDIT_ANOM_PROMISCUOUS, 9152 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u", 9153 dev->name, (dev->flags & IFF_PROMISC), 9154 (old_flags & IFF_PROMISC), 9155 from_kuid(&init_user_ns, audit_get_loginuid(current)), 9156 from_kuid(&init_user_ns, uid), 9157 from_kgid(&init_user_ns, gid), 9158 audit_get_sessionid(current)); 9159 } 9160 9161 dev_change_rx_flags(dev, IFF_PROMISC); 9162 } 9163 if (notify) 9164 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL); 9165 return 0; 9166 } 9167 9168 /** 9169 * dev_set_promiscuity - update promiscuity count on a device 9170 * @dev: device 9171 * @inc: modifier 9172 * 9173 * Add or remove promiscuity from a device. While the count in the device 9174 * remains above zero the interface remains promiscuous. Once it hits zero 9175 * the device reverts back to normal filtering operation. A negative inc 9176 * value is used to drop promiscuity on the device. 9177 * Return 0 if successful or a negative errno code on error. 9178 */ 9179 int dev_set_promiscuity(struct net_device *dev, int inc) 9180 { 9181 unsigned int old_flags = dev->flags; 9182 int err; 9183 9184 err = __dev_set_promiscuity(dev, inc, true); 9185 if (err < 0) 9186 return err; 9187 if (dev->flags != old_flags) 9188 dev_set_rx_mode(dev); 9189 return err; 9190 } 9191 EXPORT_SYMBOL(dev_set_promiscuity); 9192 9193 int netif_set_allmulti(struct net_device *dev, int inc, bool notify) 9194 { 9195 unsigned int old_flags = dev->flags, old_gflags = dev->gflags; 9196 unsigned int allmulti, flags; 9197 9198 ASSERT_RTNL(); 9199 9200 allmulti = dev->allmulti + inc; 9201 if (allmulti == 0) { 9202 /* 9203 * Avoid overflow. 9204 * If inc causes overflow, untouch allmulti and return error. 9205 */ 9206 if (unlikely(inc > 0)) { 9207 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n"); 9208 return -EOVERFLOW; 9209 } 9210 flags = old_flags & ~IFF_ALLMULTI; 9211 } else { 9212 flags = old_flags | IFF_ALLMULTI; 9213 } 9214 WRITE_ONCE(dev->allmulti, allmulti); 9215 if (flags != old_flags) { 9216 WRITE_ONCE(dev->flags, flags); 9217 netdev_info(dev, "%s allmulticast mode\n", 9218 dev->flags & IFF_ALLMULTI ? "entered" : "left"); 9219 dev_change_rx_flags(dev, IFF_ALLMULTI); 9220 dev_set_rx_mode(dev); 9221 if (notify) 9222 __dev_notify_flags(dev, old_flags, 9223 dev->gflags ^ old_gflags, 0, NULL); 9224 } 9225 return 0; 9226 } 9227 9228 /* 9229 * Upload unicast and multicast address lists to device and 9230 * configure RX filtering. When the device doesn't support unicast 9231 * filtering it is put in promiscuous mode while unicast addresses 9232 * are present. 9233 */ 9234 void __dev_set_rx_mode(struct net_device *dev) 9235 { 9236 const struct net_device_ops *ops = dev->netdev_ops; 9237 9238 /* dev_open will call this function so the list will stay sane. */ 9239 if (!(dev->flags&IFF_UP)) 9240 return; 9241 9242 if (!netif_device_present(dev)) 9243 return; 9244 9245 if (!(dev->priv_flags & IFF_UNICAST_FLT)) { 9246 /* Unicast addresses changes may only happen under the rtnl, 9247 * therefore calling __dev_set_promiscuity here is safe. 9248 */ 9249 if (!netdev_uc_empty(dev) && !dev->uc_promisc) { 9250 __dev_set_promiscuity(dev, 1, false); 9251 dev->uc_promisc = true; 9252 } else if (netdev_uc_empty(dev) && dev->uc_promisc) { 9253 __dev_set_promiscuity(dev, -1, false); 9254 dev->uc_promisc = false; 9255 } 9256 } 9257 9258 if (ops->ndo_set_rx_mode) 9259 ops->ndo_set_rx_mode(dev); 9260 } 9261 9262 void dev_set_rx_mode(struct net_device *dev) 9263 { 9264 netif_addr_lock_bh(dev); 9265 __dev_set_rx_mode(dev); 9266 netif_addr_unlock_bh(dev); 9267 } 9268 9269 /** 9270 * dev_get_flags - get flags reported to userspace 9271 * @dev: device 9272 * 9273 * Get the combination of flag bits exported through APIs to userspace. 9274 */ 9275 unsigned int dev_get_flags(const struct net_device *dev) 9276 { 9277 unsigned int flags; 9278 9279 flags = (READ_ONCE(dev->flags) & ~(IFF_PROMISC | 9280 IFF_ALLMULTI | 9281 IFF_RUNNING | 9282 IFF_LOWER_UP | 9283 IFF_DORMANT)) | 9284 (READ_ONCE(dev->gflags) & (IFF_PROMISC | 9285 IFF_ALLMULTI)); 9286 9287 if (netif_running(dev)) { 9288 if (netif_oper_up(dev)) 9289 flags |= IFF_RUNNING; 9290 if (netif_carrier_ok(dev)) 9291 flags |= IFF_LOWER_UP; 9292 if (netif_dormant(dev)) 9293 flags |= IFF_DORMANT; 9294 } 9295 9296 return flags; 9297 } 9298 EXPORT_SYMBOL(dev_get_flags); 9299 9300 int __dev_change_flags(struct net_device *dev, unsigned int flags, 9301 struct netlink_ext_ack *extack) 9302 { 9303 unsigned int old_flags = dev->flags; 9304 int ret; 9305 9306 ASSERT_RTNL(); 9307 9308 /* 9309 * Set the flags on our device. 9310 */ 9311 9312 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP | 9313 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL | 9314 IFF_AUTOMEDIA)) | 9315 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC | 9316 IFF_ALLMULTI)); 9317 9318 /* 9319 * Load in the correct multicast list now the flags have changed. 9320 */ 9321 9322 if ((old_flags ^ flags) & IFF_MULTICAST) 9323 dev_change_rx_flags(dev, IFF_MULTICAST); 9324 9325 dev_set_rx_mode(dev); 9326 9327 /* 9328 * Have we downed the interface. We handle IFF_UP ourselves 9329 * according to user attempts to set it, rather than blindly 9330 * setting it. 9331 */ 9332 9333 ret = 0; 9334 if ((old_flags ^ flags) & IFF_UP) { 9335 if (old_flags & IFF_UP) 9336 __dev_close(dev); 9337 else 9338 ret = __dev_open(dev, extack); 9339 } 9340 9341 if ((flags ^ dev->gflags) & IFF_PROMISC) { 9342 int inc = (flags & IFF_PROMISC) ? 1 : -1; 9343 old_flags = dev->flags; 9344 9345 dev->gflags ^= IFF_PROMISC; 9346 9347 if (__dev_set_promiscuity(dev, inc, false) >= 0) 9348 if (dev->flags != old_flags) 9349 dev_set_rx_mode(dev); 9350 } 9351 9352 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI 9353 * is important. Some (broken) drivers set IFF_PROMISC, when 9354 * IFF_ALLMULTI is requested not asking us and not reporting. 9355 */ 9356 if ((flags ^ dev->gflags) & IFF_ALLMULTI) { 9357 int inc = (flags & IFF_ALLMULTI) ? 1 : -1; 9358 9359 dev->gflags ^= IFF_ALLMULTI; 9360 netif_set_allmulti(dev, inc, false); 9361 } 9362 9363 return ret; 9364 } 9365 9366 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags, 9367 unsigned int gchanges, u32 portid, 9368 const struct nlmsghdr *nlh) 9369 { 9370 unsigned int changes = dev->flags ^ old_flags; 9371 9372 if (gchanges) 9373 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh); 9374 9375 if (changes & IFF_UP) { 9376 if (dev->flags & IFF_UP) 9377 call_netdevice_notifiers(NETDEV_UP, dev); 9378 else 9379 call_netdevice_notifiers(NETDEV_DOWN, dev); 9380 } 9381 9382 if (dev->flags & IFF_UP && 9383 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) { 9384 struct netdev_notifier_change_info change_info = { 9385 .info = { 9386 .dev = dev, 9387 }, 9388 .flags_changed = changes, 9389 }; 9390 9391 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info); 9392 } 9393 } 9394 9395 int netif_change_flags(struct net_device *dev, unsigned int flags, 9396 struct netlink_ext_ack *extack) 9397 { 9398 int ret; 9399 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags; 9400 9401 ret = __dev_change_flags(dev, flags, extack); 9402 if (ret < 0) 9403 return ret; 9404 9405 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags); 9406 __dev_notify_flags(dev, old_flags, changes, 0, NULL); 9407 return ret; 9408 } 9409 9410 int __dev_set_mtu(struct net_device *dev, int new_mtu) 9411 { 9412 const struct net_device_ops *ops = dev->netdev_ops; 9413 9414 if (ops->ndo_change_mtu) 9415 return ops->ndo_change_mtu(dev, new_mtu); 9416 9417 /* Pairs with all the lockless reads of dev->mtu in the stack */ 9418 WRITE_ONCE(dev->mtu, new_mtu); 9419 return 0; 9420 } 9421 EXPORT_SYMBOL(__dev_set_mtu); 9422 9423 int dev_validate_mtu(struct net_device *dev, int new_mtu, 9424 struct netlink_ext_ack *extack) 9425 { 9426 /* MTU must be positive, and in range */ 9427 if (new_mtu < 0 || new_mtu < dev->min_mtu) { 9428 NL_SET_ERR_MSG(extack, "mtu less than device minimum"); 9429 return -EINVAL; 9430 } 9431 9432 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) { 9433 NL_SET_ERR_MSG(extack, "mtu greater than device maximum"); 9434 return -EINVAL; 9435 } 9436 return 0; 9437 } 9438 9439 /** 9440 * netif_set_mtu_ext - Change maximum transfer unit 9441 * @dev: device 9442 * @new_mtu: new transfer unit 9443 * @extack: netlink extended ack 9444 * 9445 * Change the maximum transfer size of the network device. 9446 */ 9447 int netif_set_mtu_ext(struct net_device *dev, int new_mtu, 9448 struct netlink_ext_ack *extack) 9449 { 9450 int err, orig_mtu; 9451 9452 if (new_mtu == dev->mtu) 9453 return 0; 9454 9455 err = dev_validate_mtu(dev, new_mtu, extack); 9456 if (err) 9457 return err; 9458 9459 if (!netif_device_present(dev)) 9460 return -ENODEV; 9461 9462 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev); 9463 err = notifier_to_errno(err); 9464 if (err) 9465 return err; 9466 9467 orig_mtu = dev->mtu; 9468 err = __dev_set_mtu(dev, new_mtu); 9469 9470 if (!err) { 9471 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9472 orig_mtu); 9473 err = notifier_to_errno(err); 9474 if (err) { 9475 /* setting mtu back and notifying everyone again, 9476 * so that they have a chance to revert changes. 9477 */ 9478 __dev_set_mtu(dev, orig_mtu); 9479 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9480 new_mtu); 9481 } 9482 } 9483 return err; 9484 } 9485 9486 int netif_set_mtu(struct net_device *dev, int new_mtu) 9487 { 9488 struct netlink_ext_ack extack; 9489 int err; 9490 9491 memset(&extack, 0, sizeof(extack)); 9492 err = netif_set_mtu_ext(dev, new_mtu, &extack); 9493 if (err && extack._msg) 9494 net_err_ratelimited("%s: %s\n", dev->name, extack._msg); 9495 return err; 9496 } 9497 EXPORT_SYMBOL(netif_set_mtu); 9498 9499 int netif_change_tx_queue_len(struct net_device *dev, unsigned long new_len) 9500 { 9501 unsigned int orig_len = dev->tx_queue_len; 9502 int res; 9503 9504 if (new_len != (unsigned int)new_len) 9505 return -ERANGE; 9506 9507 if (new_len != orig_len) { 9508 WRITE_ONCE(dev->tx_queue_len, new_len); 9509 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev); 9510 res = notifier_to_errno(res); 9511 if (res) 9512 goto err_rollback; 9513 res = dev_qdisc_change_tx_queue_len(dev); 9514 if (res) 9515 goto err_rollback; 9516 } 9517 9518 return 0; 9519 9520 err_rollback: 9521 netdev_err(dev, "refused to change device tx_queue_len\n"); 9522 WRITE_ONCE(dev->tx_queue_len, orig_len); 9523 return res; 9524 } 9525 9526 void netif_set_group(struct net_device *dev, int new_group) 9527 { 9528 dev->group = new_group; 9529 } 9530 9531 /** 9532 * dev_pre_changeaddr_notify - Call NETDEV_PRE_CHANGEADDR. 9533 * @dev: device 9534 * @addr: new address 9535 * @extack: netlink extended ack 9536 */ 9537 int dev_pre_changeaddr_notify(struct net_device *dev, const char *addr, 9538 struct netlink_ext_ack *extack) 9539 { 9540 struct netdev_notifier_pre_changeaddr_info info = { 9541 .info.dev = dev, 9542 .info.extack = extack, 9543 .dev_addr = addr, 9544 }; 9545 int rc; 9546 9547 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info); 9548 return notifier_to_errno(rc); 9549 } 9550 EXPORT_SYMBOL(dev_pre_changeaddr_notify); 9551 9552 int netif_set_mac_address(struct net_device *dev, struct sockaddr *sa, 9553 struct netlink_ext_ack *extack) 9554 { 9555 const struct net_device_ops *ops = dev->netdev_ops; 9556 int err; 9557 9558 if (!ops->ndo_set_mac_address) 9559 return -EOPNOTSUPP; 9560 if (sa->sa_family != dev->type) 9561 return -EINVAL; 9562 if (!netif_device_present(dev)) 9563 return -ENODEV; 9564 err = dev_pre_changeaddr_notify(dev, sa->sa_data, extack); 9565 if (err) 9566 return err; 9567 if (memcmp(dev->dev_addr, sa->sa_data, dev->addr_len)) { 9568 err = ops->ndo_set_mac_address(dev, sa); 9569 if (err) 9570 return err; 9571 } 9572 dev->addr_assign_type = NET_ADDR_SET; 9573 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev); 9574 add_device_randomness(dev->dev_addr, dev->addr_len); 9575 return 0; 9576 } 9577 9578 DECLARE_RWSEM(dev_addr_sem); 9579 9580 int dev_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name) 9581 { 9582 size_t size = sizeof(sa->sa_data_min); 9583 struct net_device *dev; 9584 int ret = 0; 9585 9586 down_read(&dev_addr_sem); 9587 rcu_read_lock(); 9588 9589 dev = dev_get_by_name_rcu(net, dev_name); 9590 if (!dev) { 9591 ret = -ENODEV; 9592 goto unlock; 9593 } 9594 if (!dev->addr_len) 9595 memset(sa->sa_data, 0, size); 9596 else 9597 memcpy(sa->sa_data, dev->dev_addr, 9598 min_t(size_t, size, dev->addr_len)); 9599 sa->sa_family = dev->type; 9600 9601 unlock: 9602 rcu_read_unlock(); 9603 up_read(&dev_addr_sem); 9604 return ret; 9605 } 9606 EXPORT_SYMBOL(dev_get_mac_address); 9607 9608 int netif_change_carrier(struct net_device *dev, bool new_carrier) 9609 { 9610 const struct net_device_ops *ops = dev->netdev_ops; 9611 9612 if (!ops->ndo_change_carrier) 9613 return -EOPNOTSUPP; 9614 if (!netif_device_present(dev)) 9615 return -ENODEV; 9616 return ops->ndo_change_carrier(dev, new_carrier); 9617 } 9618 9619 /** 9620 * dev_get_phys_port_id - Get device physical port ID 9621 * @dev: device 9622 * @ppid: port ID 9623 * 9624 * Get device physical port ID 9625 */ 9626 int dev_get_phys_port_id(struct net_device *dev, 9627 struct netdev_phys_item_id *ppid) 9628 { 9629 const struct net_device_ops *ops = dev->netdev_ops; 9630 9631 if (!ops->ndo_get_phys_port_id) 9632 return -EOPNOTSUPP; 9633 return ops->ndo_get_phys_port_id(dev, ppid); 9634 } 9635 9636 /** 9637 * dev_get_phys_port_name - Get device physical port name 9638 * @dev: device 9639 * @name: port name 9640 * @len: limit of bytes to copy to name 9641 * 9642 * Get device physical port name 9643 */ 9644 int dev_get_phys_port_name(struct net_device *dev, 9645 char *name, size_t len) 9646 { 9647 const struct net_device_ops *ops = dev->netdev_ops; 9648 int err; 9649 9650 if (ops->ndo_get_phys_port_name) { 9651 err = ops->ndo_get_phys_port_name(dev, name, len); 9652 if (err != -EOPNOTSUPP) 9653 return err; 9654 } 9655 return devlink_compat_phys_port_name_get(dev, name, len); 9656 } 9657 9658 /** 9659 * dev_get_port_parent_id - Get the device's port parent identifier 9660 * @dev: network device 9661 * @ppid: pointer to a storage for the port's parent identifier 9662 * @recurse: allow/disallow recursion to lower devices 9663 * 9664 * Get the devices's port parent identifier 9665 */ 9666 int dev_get_port_parent_id(struct net_device *dev, 9667 struct netdev_phys_item_id *ppid, 9668 bool recurse) 9669 { 9670 const struct net_device_ops *ops = dev->netdev_ops; 9671 struct netdev_phys_item_id first = { }; 9672 struct net_device *lower_dev; 9673 struct list_head *iter; 9674 int err; 9675 9676 if (ops->ndo_get_port_parent_id) { 9677 err = ops->ndo_get_port_parent_id(dev, ppid); 9678 if (err != -EOPNOTSUPP) 9679 return err; 9680 } 9681 9682 err = devlink_compat_switch_id_get(dev, ppid); 9683 if (!recurse || err != -EOPNOTSUPP) 9684 return err; 9685 9686 netdev_for_each_lower_dev(dev, lower_dev, iter) { 9687 err = dev_get_port_parent_id(lower_dev, ppid, true); 9688 if (err) 9689 break; 9690 if (!first.id_len) 9691 first = *ppid; 9692 else if (memcmp(&first, ppid, sizeof(*ppid))) 9693 return -EOPNOTSUPP; 9694 } 9695 9696 return err; 9697 } 9698 EXPORT_SYMBOL(dev_get_port_parent_id); 9699 9700 /** 9701 * netdev_port_same_parent_id - Indicate if two network devices have 9702 * the same port parent identifier 9703 * @a: first network device 9704 * @b: second network device 9705 */ 9706 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b) 9707 { 9708 struct netdev_phys_item_id a_id = { }; 9709 struct netdev_phys_item_id b_id = { }; 9710 9711 if (dev_get_port_parent_id(a, &a_id, true) || 9712 dev_get_port_parent_id(b, &b_id, true)) 9713 return false; 9714 9715 return netdev_phys_item_id_same(&a_id, &b_id); 9716 } 9717 EXPORT_SYMBOL(netdev_port_same_parent_id); 9718 9719 int netif_change_proto_down(struct net_device *dev, bool proto_down) 9720 { 9721 if (!dev->change_proto_down) 9722 return -EOPNOTSUPP; 9723 if (!netif_device_present(dev)) 9724 return -ENODEV; 9725 if (proto_down) 9726 netif_carrier_off(dev); 9727 else 9728 netif_carrier_on(dev); 9729 WRITE_ONCE(dev->proto_down, proto_down); 9730 return 0; 9731 } 9732 9733 /** 9734 * netdev_change_proto_down_reason_locked - proto down reason 9735 * 9736 * @dev: device 9737 * @mask: proto down mask 9738 * @value: proto down value 9739 */ 9740 void netdev_change_proto_down_reason_locked(struct net_device *dev, 9741 unsigned long mask, u32 value) 9742 { 9743 u32 proto_down_reason; 9744 int b; 9745 9746 if (!mask) { 9747 proto_down_reason = value; 9748 } else { 9749 proto_down_reason = dev->proto_down_reason; 9750 for_each_set_bit(b, &mask, 32) { 9751 if (value & (1 << b)) 9752 proto_down_reason |= BIT(b); 9753 else 9754 proto_down_reason &= ~BIT(b); 9755 } 9756 } 9757 WRITE_ONCE(dev->proto_down_reason, proto_down_reason); 9758 } 9759 9760 struct bpf_xdp_link { 9761 struct bpf_link link; 9762 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */ 9763 int flags; 9764 }; 9765 9766 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags) 9767 { 9768 if (flags & XDP_FLAGS_HW_MODE) 9769 return XDP_MODE_HW; 9770 if (flags & XDP_FLAGS_DRV_MODE) 9771 return XDP_MODE_DRV; 9772 if (flags & XDP_FLAGS_SKB_MODE) 9773 return XDP_MODE_SKB; 9774 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB; 9775 } 9776 9777 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode) 9778 { 9779 switch (mode) { 9780 case XDP_MODE_SKB: 9781 return generic_xdp_install; 9782 case XDP_MODE_DRV: 9783 case XDP_MODE_HW: 9784 return dev->netdev_ops->ndo_bpf; 9785 default: 9786 return NULL; 9787 } 9788 } 9789 9790 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev, 9791 enum bpf_xdp_mode mode) 9792 { 9793 return dev->xdp_state[mode].link; 9794 } 9795 9796 static struct bpf_prog *dev_xdp_prog(struct net_device *dev, 9797 enum bpf_xdp_mode mode) 9798 { 9799 struct bpf_xdp_link *link = dev_xdp_link(dev, mode); 9800 9801 if (link) 9802 return link->link.prog; 9803 return dev->xdp_state[mode].prog; 9804 } 9805 9806 u8 dev_xdp_prog_count(struct net_device *dev) 9807 { 9808 u8 count = 0; 9809 int i; 9810 9811 for (i = 0; i < __MAX_XDP_MODE; i++) 9812 if (dev->xdp_state[i].prog || dev->xdp_state[i].link) 9813 count++; 9814 return count; 9815 } 9816 EXPORT_SYMBOL_GPL(dev_xdp_prog_count); 9817 9818 u8 dev_xdp_sb_prog_count(struct net_device *dev) 9819 { 9820 u8 count = 0; 9821 int i; 9822 9823 for (i = 0; i < __MAX_XDP_MODE; i++) 9824 if (dev->xdp_state[i].prog && 9825 !dev->xdp_state[i].prog->aux->xdp_has_frags) 9826 count++; 9827 return count; 9828 } 9829 9830 int netif_xdp_propagate(struct net_device *dev, struct netdev_bpf *bpf) 9831 { 9832 if (!dev->netdev_ops->ndo_bpf) 9833 return -EOPNOTSUPP; 9834 9835 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 9836 bpf->command == XDP_SETUP_PROG && 9837 bpf->prog && !bpf->prog->aux->xdp_has_frags) { 9838 NL_SET_ERR_MSG(bpf->extack, 9839 "unable to propagate XDP to device using tcp-data-split"); 9840 return -EBUSY; 9841 } 9842 9843 if (dev_get_min_mp_channel_count(dev)) { 9844 NL_SET_ERR_MSG(bpf->extack, "unable to propagate XDP to device using memory provider"); 9845 return -EBUSY; 9846 } 9847 9848 return dev->netdev_ops->ndo_bpf(dev, bpf); 9849 } 9850 9851 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode) 9852 { 9853 struct bpf_prog *prog = dev_xdp_prog(dev, mode); 9854 9855 return prog ? prog->aux->id : 0; 9856 } 9857 9858 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode, 9859 struct bpf_xdp_link *link) 9860 { 9861 dev->xdp_state[mode].link = link; 9862 dev->xdp_state[mode].prog = NULL; 9863 } 9864 9865 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode, 9866 struct bpf_prog *prog) 9867 { 9868 dev->xdp_state[mode].link = NULL; 9869 dev->xdp_state[mode].prog = prog; 9870 } 9871 9872 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode, 9873 bpf_op_t bpf_op, struct netlink_ext_ack *extack, 9874 u32 flags, struct bpf_prog *prog) 9875 { 9876 struct netdev_bpf xdp; 9877 int err; 9878 9879 netdev_ops_assert_locked(dev); 9880 9881 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 9882 prog && !prog->aux->xdp_has_frags) { 9883 NL_SET_ERR_MSG(extack, "unable to install XDP to device using tcp-data-split"); 9884 return -EBUSY; 9885 } 9886 9887 if (dev_get_min_mp_channel_count(dev)) { 9888 NL_SET_ERR_MSG(extack, "unable to install XDP to device using memory provider"); 9889 return -EBUSY; 9890 } 9891 9892 memset(&xdp, 0, sizeof(xdp)); 9893 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG; 9894 xdp.extack = extack; 9895 xdp.flags = flags; 9896 xdp.prog = prog; 9897 9898 /* Drivers assume refcnt is already incremented (i.e, prog pointer is 9899 * "moved" into driver), so they don't increment it on their own, but 9900 * they do decrement refcnt when program is detached or replaced. 9901 * Given net_device also owns link/prog, we need to bump refcnt here 9902 * to prevent drivers from underflowing it. 9903 */ 9904 if (prog) 9905 bpf_prog_inc(prog); 9906 err = bpf_op(dev, &xdp); 9907 if (err) { 9908 if (prog) 9909 bpf_prog_put(prog); 9910 return err; 9911 } 9912 9913 if (mode != XDP_MODE_HW) 9914 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog); 9915 9916 return 0; 9917 } 9918 9919 static void dev_xdp_uninstall(struct net_device *dev) 9920 { 9921 struct bpf_xdp_link *link; 9922 struct bpf_prog *prog; 9923 enum bpf_xdp_mode mode; 9924 bpf_op_t bpf_op; 9925 9926 ASSERT_RTNL(); 9927 9928 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) { 9929 prog = dev_xdp_prog(dev, mode); 9930 if (!prog) 9931 continue; 9932 9933 bpf_op = dev_xdp_bpf_op(dev, mode); 9934 if (!bpf_op) 9935 continue; 9936 9937 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 9938 9939 /* auto-detach link from net device */ 9940 link = dev_xdp_link(dev, mode); 9941 if (link) 9942 link->dev = NULL; 9943 else 9944 bpf_prog_put(prog); 9945 9946 dev_xdp_set_link(dev, mode, NULL); 9947 } 9948 } 9949 9950 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack, 9951 struct bpf_xdp_link *link, struct bpf_prog *new_prog, 9952 struct bpf_prog *old_prog, u32 flags) 9953 { 9954 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES); 9955 struct bpf_prog *cur_prog; 9956 struct net_device *upper; 9957 struct list_head *iter; 9958 enum bpf_xdp_mode mode; 9959 bpf_op_t bpf_op; 9960 int err; 9961 9962 ASSERT_RTNL(); 9963 9964 /* either link or prog attachment, never both */ 9965 if (link && (new_prog || old_prog)) 9966 return -EINVAL; 9967 /* link supports only XDP mode flags */ 9968 if (link && (flags & ~XDP_FLAGS_MODES)) { 9969 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment"); 9970 return -EINVAL; 9971 } 9972 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */ 9973 if (num_modes > 1) { 9974 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set"); 9975 return -EINVAL; 9976 } 9977 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */ 9978 if (!num_modes && dev_xdp_prog_count(dev) > 1) { 9979 NL_SET_ERR_MSG(extack, 9980 "More than one program loaded, unset mode is ambiguous"); 9981 return -EINVAL; 9982 } 9983 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */ 9984 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) { 9985 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified"); 9986 return -EINVAL; 9987 } 9988 9989 mode = dev_xdp_mode(dev, flags); 9990 /* can't replace attached link */ 9991 if (dev_xdp_link(dev, mode)) { 9992 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link"); 9993 return -EBUSY; 9994 } 9995 9996 /* don't allow if an upper device already has a program */ 9997 netdev_for_each_upper_dev_rcu(dev, upper, iter) { 9998 if (dev_xdp_prog_count(upper) > 0) { 9999 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program"); 10000 return -EEXIST; 10001 } 10002 } 10003 10004 cur_prog = dev_xdp_prog(dev, mode); 10005 /* can't replace attached prog with link */ 10006 if (link && cur_prog) { 10007 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link"); 10008 return -EBUSY; 10009 } 10010 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) { 10011 NL_SET_ERR_MSG(extack, "Active program does not match expected"); 10012 return -EEXIST; 10013 } 10014 10015 /* put effective new program into new_prog */ 10016 if (link) 10017 new_prog = link->link.prog; 10018 10019 if (new_prog) { 10020 bool offload = mode == XDP_MODE_HW; 10021 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB 10022 ? XDP_MODE_DRV : XDP_MODE_SKB; 10023 10024 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) { 10025 NL_SET_ERR_MSG(extack, "XDP program already attached"); 10026 return -EBUSY; 10027 } 10028 if (!offload && dev_xdp_prog(dev, other_mode)) { 10029 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time"); 10030 return -EEXIST; 10031 } 10032 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) { 10033 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported"); 10034 return -EINVAL; 10035 } 10036 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) { 10037 NL_SET_ERR_MSG(extack, "Program bound to different device"); 10038 return -EINVAL; 10039 } 10040 if (bpf_prog_is_dev_bound(new_prog->aux) && mode == XDP_MODE_SKB) { 10041 NL_SET_ERR_MSG(extack, "Can't attach device-bound programs in generic mode"); 10042 return -EINVAL; 10043 } 10044 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) { 10045 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device"); 10046 return -EINVAL; 10047 } 10048 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) { 10049 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device"); 10050 return -EINVAL; 10051 } 10052 } 10053 10054 /* don't call drivers if the effective program didn't change */ 10055 if (new_prog != cur_prog) { 10056 bpf_op = dev_xdp_bpf_op(dev, mode); 10057 if (!bpf_op) { 10058 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode"); 10059 return -EOPNOTSUPP; 10060 } 10061 10062 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog); 10063 if (err) 10064 return err; 10065 } 10066 10067 if (link) 10068 dev_xdp_set_link(dev, mode, link); 10069 else 10070 dev_xdp_set_prog(dev, mode, new_prog); 10071 if (cur_prog) 10072 bpf_prog_put(cur_prog); 10073 10074 return 0; 10075 } 10076 10077 static int dev_xdp_attach_link(struct net_device *dev, 10078 struct netlink_ext_ack *extack, 10079 struct bpf_xdp_link *link) 10080 { 10081 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags); 10082 } 10083 10084 static int dev_xdp_detach_link(struct net_device *dev, 10085 struct netlink_ext_ack *extack, 10086 struct bpf_xdp_link *link) 10087 { 10088 enum bpf_xdp_mode mode; 10089 bpf_op_t bpf_op; 10090 10091 ASSERT_RTNL(); 10092 10093 mode = dev_xdp_mode(dev, link->flags); 10094 if (dev_xdp_link(dev, mode) != link) 10095 return -EINVAL; 10096 10097 bpf_op = dev_xdp_bpf_op(dev, mode); 10098 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 10099 dev_xdp_set_link(dev, mode, NULL); 10100 return 0; 10101 } 10102 10103 static void bpf_xdp_link_release(struct bpf_link *link) 10104 { 10105 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10106 10107 rtnl_lock(); 10108 10109 /* if racing with net_device's tear down, xdp_link->dev might be 10110 * already NULL, in which case link was already auto-detached 10111 */ 10112 if (xdp_link->dev) { 10113 netdev_lock_ops(xdp_link->dev); 10114 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link)); 10115 netdev_unlock_ops(xdp_link->dev); 10116 xdp_link->dev = NULL; 10117 } 10118 10119 rtnl_unlock(); 10120 } 10121 10122 static int bpf_xdp_link_detach(struct bpf_link *link) 10123 { 10124 bpf_xdp_link_release(link); 10125 return 0; 10126 } 10127 10128 static void bpf_xdp_link_dealloc(struct bpf_link *link) 10129 { 10130 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10131 10132 kfree(xdp_link); 10133 } 10134 10135 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link, 10136 struct seq_file *seq) 10137 { 10138 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10139 u32 ifindex = 0; 10140 10141 rtnl_lock(); 10142 if (xdp_link->dev) 10143 ifindex = xdp_link->dev->ifindex; 10144 rtnl_unlock(); 10145 10146 seq_printf(seq, "ifindex:\t%u\n", ifindex); 10147 } 10148 10149 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link, 10150 struct bpf_link_info *info) 10151 { 10152 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10153 u32 ifindex = 0; 10154 10155 rtnl_lock(); 10156 if (xdp_link->dev) 10157 ifindex = xdp_link->dev->ifindex; 10158 rtnl_unlock(); 10159 10160 info->xdp.ifindex = ifindex; 10161 return 0; 10162 } 10163 10164 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog, 10165 struct bpf_prog *old_prog) 10166 { 10167 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10168 enum bpf_xdp_mode mode; 10169 bpf_op_t bpf_op; 10170 int err = 0; 10171 10172 rtnl_lock(); 10173 10174 /* link might have been auto-released already, so fail */ 10175 if (!xdp_link->dev) { 10176 err = -ENOLINK; 10177 goto out_unlock; 10178 } 10179 10180 if (old_prog && link->prog != old_prog) { 10181 err = -EPERM; 10182 goto out_unlock; 10183 } 10184 old_prog = link->prog; 10185 if (old_prog->type != new_prog->type || 10186 old_prog->expected_attach_type != new_prog->expected_attach_type) { 10187 err = -EINVAL; 10188 goto out_unlock; 10189 } 10190 10191 if (old_prog == new_prog) { 10192 /* no-op, don't disturb drivers */ 10193 bpf_prog_put(new_prog); 10194 goto out_unlock; 10195 } 10196 10197 netdev_lock_ops(xdp_link->dev); 10198 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags); 10199 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode); 10200 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL, 10201 xdp_link->flags, new_prog); 10202 netdev_unlock_ops(xdp_link->dev); 10203 if (err) 10204 goto out_unlock; 10205 10206 old_prog = xchg(&link->prog, new_prog); 10207 bpf_prog_put(old_prog); 10208 10209 out_unlock: 10210 rtnl_unlock(); 10211 return err; 10212 } 10213 10214 static const struct bpf_link_ops bpf_xdp_link_lops = { 10215 .release = bpf_xdp_link_release, 10216 .dealloc = bpf_xdp_link_dealloc, 10217 .detach = bpf_xdp_link_detach, 10218 .show_fdinfo = bpf_xdp_link_show_fdinfo, 10219 .fill_link_info = bpf_xdp_link_fill_link_info, 10220 .update_prog = bpf_xdp_link_update, 10221 }; 10222 10223 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog) 10224 { 10225 struct net *net = current->nsproxy->net_ns; 10226 struct bpf_link_primer link_primer; 10227 struct netlink_ext_ack extack = {}; 10228 struct bpf_xdp_link *link; 10229 struct net_device *dev; 10230 int err, fd; 10231 10232 rtnl_lock(); 10233 dev = dev_get_by_index(net, attr->link_create.target_ifindex); 10234 if (!dev) { 10235 rtnl_unlock(); 10236 return -EINVAL; 10237 } 10238 10239 link = kzalloc(sizeof(*link), GFP_USER); 10240 if (!link) { 10241 err = -ENOMEM; 10242 goto unlock; 10243 } 10244 10245 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog); 10246 link->dev = dev; 10247 link->flags = attr->link_create.flags; 10248 10249 err = bpf_link_prime(&link->link, &link_primer); 10250 if (err) { 10251 kfree(link); 10252 goto unlock; 10253 } 10254 10255 err = dev_xdp_attach_link(dev, &extack, link); 10256 rtnl_unlock(); 10257 10258 if (err) { 10259 link->dev = NULL; 10260 bpf_link_cleanup(&link_primer); 10261 trace_bpf_xdp_link_attach_failed(extack._msg); 10262 goto out_put_dev; 10263 } 10264 10265 fd = bpf_link_settle(&link_primer); 10266 /* link itself doesn't hold dev's refcnt to not complicate shutdown */ 10267 dev_put(dev); 10268 return fd; 10269 10270 unlock: 10271 rtnl_unlock(); 10272 10273 out_put_dev: 10274 dev_put(dev); 10275 return err; 10276 } 10277 10278 /** 10279 * dev_change_xdp_fd - set or clear a bpf program for a device rx path 10280 * @dev: device 10281 * @extack: netlink extended ack 10282 * @fd: new program fd or negative value to clear 10283 * @expected_fd: old program fd that userspace expects to replace or clear 10284 * @flags: xdp-related flags 10285 * 10286 * Set or clear a bpf program for a device 10287 */ 10288 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, 10289 int fd, int expected_fd, u32 flags) 10290 { 10291 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags); 10292 struct bpf_prog *new_prog = NULL, *old_prog = NULL; 10293 int err; 10294 10295 ASSERT_RTNL(); 10296 10297 if (fd >= 0) { 10298 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP, 10299 mode != XDP_MODE_SKB); 10300 if (IS_ERR(new_prog)) 10301 return PTR_ERR(new_prog); 10302 } 10303 10304 if (expected_fd >= 0) { 10305 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP, 10306 mode != XDP_MODE_SKB); 10307 if (IS_ERR(old_prog)) { 10308 err = PTR_ERR(old_prog); 10309 old_prog = NULL; 10310 goto err_out; 10311 } 10312 } 10313 10314 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags); 10315 10316 err_out: 10317 if (err && new_prog) 10318 bpf_prog_put(new_prog); 10319 if (old_prog) 10320 bpf_prog_put(old_prog); 10321 return err; 10322 } 10323 10324 u32 dev_get_min_mp_channel_count(const struct net_device *dev) 10325 { 10326 int i; 10327 10328 ASSERT_RTNL(); 10329 10330 for (i = dev->real_num_rx_queues - 1; i >= 0; i--) 10331 if (dev->_rx[i].mp_params.mp_priv) 10332 /* The channel count is the idx plus 1. */ 10333 return i + 1; 10334 10335 return 0; 10336 } 10337 10338 /** 10339 * dev_index_reserve() - allocate an ifindex in a namespace 10340 * @net: the applicable net namespace 10341 * @ifindex: requested ifindex, pass %0 to get one allocated 10342 * 10343 * Allocate a ifindex for a new device. Caller must either use the ifindex 10344 * to store the device (via list_netdevice()) or call dev_index_release() 10345 * to give the index up. 10346 * 10347 * Return: a suitable unique value for a new device interface number or -errno. 10348 */ 10349 static int dev_index_reserve(struct net *net, u32 ifindex) 10350 { 10351 int err; 10352 10353 if (ifindex > INT_MAX) { 10354 DEBUG_NET_WARN_ON_ONCE(1); 10355 return -EINVAL; 10356 } 10357 10358 if (!ifindex) 10359 err = xa_alloc_cyclic(&net->dev_by_index, &ifindex, NULL, 10360 xa_limit_31b, &net->ifindex, GFP_KERNEL); 10361 else 10362 err = xa_insert(&net->dev_by_index, ifindex, NULL, GFP_KERNEL); 10363 if (err < 0) 10364 return err; 10365 10366 return ifindex; 10367 } 10368 10369 static void dev_index_release(struct net *net, int ifindex) 10370 { 10371 /* Expect only unused indexes, unlist_netdevice() removes the used */ 10372 WARN_ON(xa_erase(&net->dev_by_index, ifindex)); 10373 } 10374 10375 static bool from_cleanup_net(void) 10376 { 10377 #ifdef CONFIG_NET_NS 10378 return current == cleanup_net_task; 10379 #else 10380 return false; 10381 #endif 10382 } 10383 10384 /* Delayed registration/unregisteration */ 10385 LIST_HEAD(net_todo_list); 10386 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq); 10387 atomic_t dev_unreg_count = ATOMIC_INIT(0); 10388 10389 static void net_set_todo(struct net_device *dev) 10390 { 10391 list_add_tail(&dev->todo_list, &net_todo_list); 10392 } 10393 10394 static netdev_features_t netdev_sync_upper_features(struct net_device *lower, 10395 struct net_device *upper, netdev_features_t features) 10396 { 10397 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10398 netdev_features_t feature; 10399 int feature_bit; 10400 10401 for_each_netdev_feature(upper_disables, feature_bit) { 10402 feature = __NETIF_F_BIT(feature_bit); 10403 if (!(upper->wanted_features & feature) 10404 && (features & feature)) { 10405 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n", 10406 &feature, upper->name); 10407 features &= ~feature; 10408 } 10409 } 10410 10411 return features; 10412 } 10413 10414 static void netdev_sync_lower_features(struct net_device *upper, 10415 struct net_device *lower, netdev_features_t features) 10416 { 10417 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10418 netdev_features_t feature; 10419 int feature_bit; 10420 10421 for_each_netdev_feature(upper_disables, feature_bit) { 10422 feature = __NETIF_F_BIT(feature_bit); 10423 if (!(features & feature) && (lower->features & feature)) { 10424 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n", 10425 &feature, lower->name); 10426 lower->wanted_features &= ~feature; 10427 __netdev_update_features(lower); 10428 10429 if (unlikely(lower->features & feature)) 10430 netdev_WARN(upper, "failed to disable %pNF on %s!\n", 10431 &feature, lower->name); 10432 else 10433 netdev_features_change(lower); 10434 } 10435 } 10436 } 10437 10438 static bool netdev_has_ip_or_hw_csum(netdev_features_t features) 10439 { 10440 netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 10441 bool ip_csum = (features & ip_csum_mask) == ip_csum_mask; 10442 bool hw_csum = features & NETIF_F_HW_CSUM; 10443 10444 return ip_csum || hw_csum; 10445 } 10446 10447 static netdev_features_t netdev_fix_features(struct net_device *dev, 10448 netdev_features_t features) 10449 { 10450 /* Fix illegal checksum combinations */ 10451 if ((features & NETIF_F_HW_CSUM) && 10452 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) { 10453 netdev_warn(dev, "mixed HW and IP checksum settings.\n"); 10454 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM); 10455 } 10456 10457 /* TSO requires that SG is present as well. */ 10458 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) { 10459 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n"); 10460 features &= ~NETIF_F_ALL_TSO; 10461 } 10462 10463 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) && 10464 !(features & NETIF_F_IP_CSUM)) { 10465 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n"); 10466 features &= ~NETIF_F_TSO; 10467 features &= ~NETIF_F_TSO_ECN; 10468 } 10469 10470 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) && 10471 !(features & NETIF_F_IPV6_CSUM)) { 10472 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n"); 10473 features &= ~NETIF_F_TSO6; 10474 } 10475 10476 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */ 10477 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO)) 10478 features &= ~NETIF_F_TSO_MANGLEID; 10479 10480 /* TSO ECN requires that TSO is present as well. */ 10481 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN) 10482 features &= ~NETIF_F_TSO_ECN; 10483 10484 /* Software GSO depends on SG. */ 10485 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) { 10486 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n"); 10487 features &= ~NETIF_F_GSO; 10488 } 10489 10490 /* GSO partial features require GSO partial be set */ 10491 if ((features & dev->gso_partial_features) && 10492 !(features & NETIF_F_GSO_PARTIAL)) { 10493 netdev_dbg(dev, 10494 "Dropping partially supported GSO features since no GSO partial.\n"); 10495 features &= ~dev->gso_partial_features; 10496 } 10497 10498 if (!(features & NETIF_F_RXCSUM)) { 10499 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet 10500 * successfully merged by hardware must also have the 10501 * checksum verified by hardware. If the user does not 10502 * want to enable RXCSUM, logically, we should disable GRO_HW. 10503 */ 10504 if (features & NETIF_F_GRO_HW) { 10505 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n"); 10506 features &= ~NETIF_F_GRO_HW; 10507 } 10508 } 10509 10510 /* LRO/HW-GRO features cannot be combined with RX-FCS */ 10511 if (features & NETIF_F_RXFCS) { 10512 if (features & NETIF_F_LRO) { 10513 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n"); 10514 features &= ~NETIF_F_LRO; 10515 } 10516 10517 if (features & NETIF_F_GRO_HW) { 10518 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n"); 10519 features &= ~NETIF_F_GRO_HW; 10520 } 10521 } 10522 10523 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) { 10524 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n"); 10525 features &= ~NETIF_F_LRO; 10526 } 10527 10528 if ((features & NETIF_F_HW_TLS_TX) && !netdev_has_ip_or_hw_csum(features)) { 10529 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n"); 10530 features &= ~NETIF_F_HW_TLS_TX; 10531 } 10532 10533 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) { 10534 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n"); 10535 features &= ~NETIF_F_HW_TLS_RX; 10536 } 10537 10538 if ((features & NETIF_F_GSO_UDP_L4) && !netdev_has_ip_or_hw_csum(features)) { 10539 netdev_dbg(dev, "Dropping USO feature since no CSUM feature.\n"); 10540 features &= ~NETIF_F_GSO_UDP_L4; 10541 } 10542 10543 return features; 10544 } 10545 10546 int __netdev_update_features(struct net_device *dev) 10547 { 10548 struct net_device *upper, *lower; 10549 netdev_features_t features; 10550 struct list_head *iter; 10551 int err = -1; 10552 10553 ASSERT_RTNL(); 10554 netdev_ops_assert_locked(dev); 10555 10556 features = netdev_get_wanted_features(dev); 10557 10558 if (dev->netdev_ops->ndo_fix_features) 10559 features = dev->netdev_ops->ndo_fix_features(dev, features); 10560 10561 /* driver might be less strict about feature dependencies */ 10562 features = netdev_fix_features(dev, features); 10563 10564 /* some features can't be enabled if they're off on an upper device */ 10565 netdev_for_each_upper_dev_rcu(dev, upper, iter) 10566 features = netdev_sync_upper_features(dev, upper, features); 10567 10568 if (dev->features == features) 10569 goto sync_lower; 10570 10571 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n", 10572 &dev->features, &features); 10573 10574 if (dev->netdev_ops->ndo_set_features) 10575 err = dev->netdev_ops->ndo_set_features(dev, features); 10576 else 10577 err = 0; 10578 10579 if (unlikely(err < 0)) { 10580 netdev_err(dev, 10581 "set_features() failed (%d); wanted %pNF, left %pNF\n", 10582 err, &features, &dev->features); 10583 /* return non-0 since some features might have changed and 10584 * it's better to fire a spurious notification than miss it 10585 */ 10586 return -1; 10587 } 10588 10589 sync_lower: 10590 /* some features must be disabled on lower devices when disabled 10591 * on an upper device (think: bonding master or bridge) 10592 */ 10593 netdev_for_each_lower_dev(dev, lower, iter) 10594 netdev_sync_lower_features(dev, lower, features); 10595 10596 if (!err) { 10597 netdev_features_t diff = features ^ dev->features; 10598 10599 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) { 10600 /* udp_tunnel_{get,drop}_rx_info both need 10601 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the 10602 * device, or they won't do anything. 10603 * Thus we need to update dev->features 10604 * *before* calling udp_tunnel_get_rx_info, 10605 * but *after* calling udp_tunnel_drop_rx_info. 10606 */ 10607 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) { 10608 dev->features = features; 10609 udp_tunnel_get_rx_info(dev); 10610 } else { 10611 udp_tunnel_drop_rx_info(dev); 10612 } 10613 } 10614 10615 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) { 10616 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) { 10617 dev->features = features; 10618 err |= vlan_get_rx_ctag_filter_info(dev); 10619 } else { 10620 vlan_drop_rx_ctag_filter_info(dev); 10621 } 10622 } 10623 10624 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) { 10625 if (features & NETIF_F_HW_VLAN_STAG_FILTER) { 10626 dev->features = features; 10627 err |= vlan_get_rx_stag_filter_info(dev); 10628 } else { 10629 vlan_drop_rx_stag_filter_info(dev); 10630 } 10631 } 10632 10633 dev->features = features; 10634 } 10635 10636 return err < 0 ? 0 : 1; 10637 } 10638 10639 /** 10640 * netdev_update_features - recalculate device features 10641 * @dev: the device to check 10642 * 10643 * Recalculate dev->features set and send notifications if it 10644 * has changed. Should be called after driver or hardware dependent 10645 * conditions might have changed that influence the features. 10646 */ 10647 void netdev_update_features(struct net_device *dev) 10648 { 10649 if (__netdev_update_features(dev)) 10650 netdev_features_change(dev); 10651 } 10652 EXPORT_SYMBOL(netdev_update_features); 10653 10654 /** 10655 * netdev_change_features - recalculate device features 10656 * @dev: the device to check 10657 * 10658 * Recalculate dev->features set and send notifications even 10659 * if they have not changed. Should be called instead of 10660 * netdev_update_features() if also dev->vlan_features might 10661 * have changed to allow the changes to be propagated to stacked 10662 * VLAN devices. 10663 */ 10664 void netdev_change_features(struct net_device *dev) 10665 { 10666 __netdev_update_features(dev); 10667 netdev_features_change(dev); 10668 } 10669 EXPORT_SYMBOL(netdev_change_features); 10670 10671 /** 10672 * netif_stacked_transfer_operstate - transfer operstate 10673 * @rootdev: the root or lower level device to transfer state from 10674 * @dev: the device to transfer operstate to 10675 * 10676 * Transfer operational state from root to device. This is normally 10677 * called when a stacking relationship exists between the root 10678 * device and the device(a leaf device). 10679 */ 10680 void netif_stacked_transfer_operstate(const struct net_device *rootdev, 10681 struct net_device *dev) 10682 { 10683 if (rootdev->operstate == IF_OPER_DORMANT) 10684 netif_dormant_on(dev); 10685 else 10686 netif_dormant_off(dev); 10687 10688 if (rootdev->operstate == IF_OPER_TESTING) 10689 netif_testing_on(dev); 10690 else 10691 netif_testing_off(dev); 10692 10693 if (netif_carrier_ok(rootdev)) 10694 netif_carrier_on(dev); 10695 else 10696 netif_carrier_off(dev); 10697 } 10698 EXPORT_SYMBOL(netif_stacked_transfer_operstate); 10699 10700 static int netif_alloc_rx_queues(struct net_device *dev) 10701 { 10702 unsigned int i, count = dev->num_rx_queues; 10703 struct netdev_rx_queue *rx; 10704 size_t sz = count * sizeof(*rx); 10705 int err = 0; 10706 10707 BUG_ON(count < 1); 10708 10709 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10710 if (!rx) 10711 return -ENOMEM; 10712 10713 dev->_rx = rx; 10714 10715 for (i = 0; i < count; i++) { 10716 rx[i].dev = dev; 10717 10718 /* XDP RX-queue setup */ 10719 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0); 10720 if (err < 0) 10721 goto err_rxq_info; 10722 } 10723 return 0; 10724 10725 err_rxq_info: 10726 /* Rollback successful reg's and free other resources */ 10727 while (i--) 10728 xdp_rxq_info_unreg(&rx[i].xdp_rxq); 10729 kvfree(dev->_rx); 10730 dev->_rx = NULL; 10731 return err; 10732 } 10733 10734 static void netif_free_rx_queues(struct net_device *dev) 10735 { 10736 unsigned int i, count = dev->num_rx_queues; 10737 10738 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */ 10739 if (!dev->_rx) 10740 return; 10741 10742 for (i = 0; i < count; i++) 10743 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq); 10744 10745 kvfree(dev->_rx); 10746 } 10747 10748 static void netdev_init_one_queue(struct net_device *dev, 10749 struct netdev_queue *queue, void *_unused) 10750 { 10751 /* Initialize queue lock */ 10752 spin_lock_init(&queue->_xmit_lock); 10753 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type); 10754 queue->xmit_lock_owner = -1; 10755 netdev_queue_numa_node_write(queue, NUMA_NO_NODE); 10756 queue->dev = dev; 10757 #ifdef CONFIG_BQL 10758 dql_init(&queue->dql, HZ); 10759 #endif 10760 } 10761 10762 static void netif_free_tx_queues(struct net_device *dev) 10763 { 10764 kvfree(dev->_tx); 10765 } 10766 10767 static int netif_alloc_netdev_queues(struct net_device *dev) 10768 { 10769 unsigned int count = dev->num_tx_queues; 10770 struct netdev_queue *tx; 10771 size_t sz = count * sizeof(*tx); 10772 10773 if (count < 1 || count > 0xffff) 10774 return -EINVAL; 10775 10776 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10777 if (!tx) 10778 return -ENOMEM; 10779 10780 dev->_tx = tx; 10781 10782 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL); 10783 spin_lock_init(&dev->tx_global_lock); 10784 10785 return 0; 10786 } 10787 10788 void netif_tx_stop_all_queues(struct net_device *dev) 10789 { 10790 unsigned int i; 10791 10792 for (i = 0; i < dev->num_tx_queues; i++) { 10793 struct netdev_queue *txq = netdev_get_tx_queue(dev, i); 10794 10795 netif_tx_stop_queue(txq); 10796 } 10797 } 10798 EXPORT_SYMBOL(netif_tx_stop_all_queues); 10799 10800 static int netdev_do_alloc_pcpu_stats(struct net_device *dev) 10801 { 10802 void __percpu *v; 10803 10804 /* Drivers implementing ndo_get_peer_dev must support tstat 10805 * accounting, so that skb_do_redirect() can bump the dev's 10806 * RX stats upon network namespace switch. 10807 */ 10808 if (dev->netdev_ops->ndo_get_peer_dev && 10809 dev->pcpu_stat_type != NETDEV_PCPU_STAT_TSTATS) 10810 return -EOPNOTSUPP; 10811 10812 switch (dev->pcpu_stat_type) { 10813 case NETDEV_PCPU_STAT_NONE: 10814 return 0; 10815 case NETDEV_PCPU_STAT_LSTATS: 10816 v = dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats); 10817 break; 10818 case NETDEV_PCPU_STAT_TSTATS: 10819 v = dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats); 10820 break; 10821 case NETDEV_PCPU_STAT_DSTATS: 10822 v = dev->dstats = netdev_alloc_pcpu_stats(struct pcpu_dstats); 10823 break; 10824 default: 10825 return -EINVAL; 10826 } 10827 10828 return v ? 0 : -ENOMEM; 10829 } 10830 10831 static void netdev_do_free_pcpu_stats(struct net_device *dev) 10832 { 10833 switch (dev->pcpu_stat_type) { 10834 case NETDEV_PCPU_STAT_NONE: 10835 return; 10836 case NETDEV_PCPU_STAT_LSTATS: 10837 free_percpu(dev->lstats); 10838 break; 10839 case NETDEV_PCPU_STAT_TSTATS: 10840 free_percpu(dev->tstats); 10841 break; 10842 case NETDEV_PCPU_STAT_DSTATS: 10843 free_percpu(dev->dstats); 10844 break; 10845 } 10846 } 10847 10848 static void netdev_free_phy_link_topology(struct net_device *dev) 10849 { 10850 struct phy_link_topology *topo = dev->link_topo; 10851 10852 if (IS_ENABLED(CONFIG_PHYLIB) && topo) { 10853 xa_destroy(&topo->phys); 10854 kfree(topo); 10855 dev->link_topo = NULL; 10856 } 10857 } 10858 10859 /** 10860 * register_netdevice() - register a network device 10861 * @dev: device to register 10862 * 10863 * Take a prepared network device structure and make it externally accessible. 10864 * A %NETDEV_REGISTER message is sent to the netdev notifier chain. 10865 * Callers must hold the rtnl lock - you may want register_netdev() 10866 * instead of this. 10867 */ 10868 int register_netdevice(struct net_device *dev) 10869 { 10870 int ret; 10871 struct net *net = dev_net(dev); 10872 10873 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE < 10874 NETDEV_FEATURE_COUNT); 10875 BUG_ON(dev_boot_phase); 10876 ASSERT_RTNL(); 10877 10878 might_sleep(); 10879 10880 /* When net_device's are persistent, this will be fatal. */ 10881 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED); 10882 BUG_ON(!net); 10883 10884 ret = ethtool_check_ops(dev->ethtool_ops); 10885 if (ret) 10886 return ret; 10887 10888 /* rss ctx ID 0 is reserved for the default context, start from 1 */ 10889 xa_init_flags(&dev->ethtool->rss_ctx, XA_FLAGS_ALLOC1); 10890 mutex_init(&dev->ethtool->rss_lock); 10891 10892 spin_lock_init(&dev->addr_list_lock); 10893 netdev_set_addr_lockdep_class(dev); 10894 10895 ret = dev_get_valid_name(net, dev, dev->name); 10896 if (ret < 0) 10897 goto out; 10898 10899 ret = -ENOMEM; 10900 dev->name_node = netdev_name_node_head_alloc(dev); 10901 if (!dev->name_node) 10902 goto out; 10903 10904 /* Init, if this function is available */ 10905 if (dev->netdev_ops->ndo_init) { 10906 ret = dev->netdev_ops->ndo_init(dev); 10907 if (ret) { 10908 if (ret > 0) 10909 ret = -EIO; 10910 goto err_free_name; 10911 } 10912 } 10913 10914 if (((dev->hw_features | dev->features) & 10915 NETIF_F_HW_VLAN_CTAG_FILTER) && 10916 (!dev->netdev_ops->ndo_vlan_rx_add_vid || 10917 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) { 10918 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n"); 10919 ret = -EINVAL; 10920 goto err_uninit; 10921 } 10922 10923 ret = netdev_do_alloc_pcpu_stats(dev); 10924 if (ret) 10925 goto err_uninit; 10926 10927 ret = dev_index_reserve(net, dev->ifindex); 10928 if (ret < 0) 10929 goto err_free_pcpu; 10930 dev->ifindex = ret; 10931 10932 /* Transfer changeable features to wanted_features and enable 10933 * software offloads (GSO and GRO). 10934 */ 10935 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF); 10936 dev->features |= NETIF_F_SOFT_FEATURES; 10937 10938 if (dev->udp_tunnel_nic_info) { 10939 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10940 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10941 } 10942 10943 dev->wanted_features = dev->features & dev->hw_features; 10944 10945 if (!(dev->flags & IFF_LOOPBACK)) 10946 dev->hw_features |= NETIF_F_NOCACHE_COPY; 10947 10948 /* If IPv4 TCP segmentation offload is supported we should also 10949 * allow the device to enable segmenting the frame with the option 10950 * of ignoring a static IP ID value. This doesn't enable the 10951 * feature itself but allows the user to enable it later. 10952 */ 10953 if (dev->hw_features & NETIF_F_TSO) 10954 dev->hw_features |= NETIF_F_TSO_MANGLEID; 10955 if (dev->vlan_features & NETIF_F_TSO) 10956 dev->vlan_features |= NETIF_F_TSO_MANGLEID; 10957 if (dev->mpls_features & NETIF_F_TSO) 10958 dev->mpls_features |= NETIF_F_TSO_MANGLEID; 10959 if (dev->hw_enc_features & NETIF_F_TSO) 10960 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID; 10961 10962 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices. 10963 */ 10964 dev->vlan_features |= NETIF_F_HIGHDMA; 10965 10966 /* Make NETIF_F_SG inheritable to tunnel devices. 10967 */ 10968 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL; 10969 10970 /* Make NETIF_F_SG inheritable to MPLS. 10971 */ 10972 dev->mpls_features |= NETIF_F_SG; 10973 10974 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev); 10975 ret = notifier_to_errno(ret); 10976 if (ret) 10977 goto err_ifindex_release; 10978 10979 ret = netdev_register_kobject(dev); 10980 10981 netdev_lock(dev); 10982 WRITE_ONCE(dev->reg_state, ret ? NETREG_UNREGISTERED : NETREG_REGISTERED); 10983 netdev_unlock(dev); 10984 10985 if (ret) 10986 goto err_uninit_notify; 10987 10988 netdev_lock_ops(dev); 10989 __netdev_update_features(dev); 10990 netdev_unlock_ops(dev); 10991 10992 /* 10993 * Default initial state at registry is that the 10994 * device is present. 10995 */ 10996 10997 set_bit(__LINK_STATE_PRESENT, &dev->state); 10998 10999 linkwatch_init_dev(dev); 11000 11001 dev_init_scheduler(dev); 11002 11003 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL); 11004 list_netdevice(dev); 11005 11006 add_device_randomness(dev->dev_addr, dev->addr_len); 11007 11008 /* If the device has permanent device address, driver should 11009 * set dev_addr and also addr_assign_type should be set to 11010 * NET_ADDR_PERM (default value). 11011 */ 11012 if (dev->addr_assign_type == NET_ADDR_PERM) 11013 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len); 11014 11015 /* Notify protocols, that a new device appeared. */ 11016 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev); 11017 ret = notifier_to_errno(ret); 11018 if (ret) { 11019 /* Expect explicit free_netdev() on failure */ 11020 dev->needs_free_netdev = false; 11021 unregister_netdevice_queue(dev, NULL); 11022 goto out; 11023 } 11024 /* 11025 * Prevent userspace races by waiting until the network 11026 * device is fully setup before sending notifications. 11027 */ 11028 if (!dev->rtnl_link_ops || 11029 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 11030 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 11031 11032 out: 11033 return ret; 11034 11035 err_uninit_notify: 11036 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11037 err_ifindex_release: 11038 dev_index_release(net, dev->ifindex); 11039 err_free_pcpu: 11040 netdev_do_free_pcpu_stats(dev); 11041 err_uninit: 11042 if (dev->netdev_ops->ndo_uninit) 11043 dev->netdev_ops->ndo_uninit(dev); 11044 if (dev->priv_destructor) 11045 dev->priv_destructor(dev); 11046 err_free_name: 11047 netdev_name_node_free(dev->name_node); 11048 goto out; 11049 } 11050 EXPORT_SYMBOL(register_netdevice); 11051 11052 /* Initialize the core of a dummy net device. 11053 * The setup steps dummy netdevs need which normal netdevs get by going 11054 * through register_netdevice(). 11055 */ 11056 static void init_dummy_netdev(struct net_device *dev) 11057 { 11058 /* make sure we BUG if trying to hit standard 11059 * register/unregister code path 11060 */ 11061 dev->reg_state = NETREG_DUMMY; 11062 11063 /* a dummy interface is started by default */ 11064 set_bit(__LINK_STATE_PRESENT, &dev->state); 11065 set_bit(__LINK_STATE_START, &dev->state); 11066 11067 /* Note : We dont allocate pcpu_refcnt for dummy devices, 11068 * because users of this 'device' dont need to change 11069 * its refcount. 11070 */ 11071 } 11072 11073 /** 11074 * register_netdev - register a network device 11075 * @dev: device to register 11076 * 11077 * Take a completed network device structure and add it to the kernel 11078 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier 11079 * chain. 0 is returned on success. A negative errno code is returned 11080 * on a failure to set up the device, or if the name is a duplicate. 11081 * 11082 * This is a wrapper around register_netdevice that takes the rtnl semaphore 11083 * and expands the device name if you passed a format string to 11084 * alloc_netdev. 11085 */ 11086 int register_netdev(struct net_device *dev) 11087 { 11088 struct net *net = dev_net(dev); 11089 int err; 11090 11091 if (rtnl_net_lock_killable(net)) 11092 return -EINTR; 11093 11094 err = register_netdevice(dev); 11095 11096 rtnl_net_unlock(net); 11097 11098 return err; 11099 } 11100 EXPORT_SYMBOL(register_netdev); 11101 11102 int netdev_refcnt_read(const struct net_device *dev) 11103 { 11104 #ifdef CONFIG_PCPU_DEV_REFCNT 11105 int i, refcnt = 0; 11106 11107 for_each_possible_cpu(i) 11108 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i); 11109 return refcnt; 11110 #else 11111 return refcount_read(&dev->dev_refcnt); 11112 #endif 11113 } 11114 EXPORT_SYMBOL(netdev_refcnt_read); 11115 11116 int netdev_unregister_timeout_secs __read_mostly = 10; 11117 11118 #define WAIT_REFS_MIN_MSECS 1 11119 #define WAIT_REFS_MAX_MSECS 250 11120 /** 11121 * netdev_wait_allrefs_any - wait until all references are gone. 11122 * @list: list of net_devices to wait on 11123 * 11124 * This is called when unregistering network devices. 11125 * 11126 * Any protocol or device that holds a reference should register 11127 * for netdevice notification, and cleanup and put back the 11128 * reference if they receive an UNREGISTER event. 11129 * We can get stuck here if buggy protocols don't correctly 11130 * call dev_put. 11131 */ 11132 static struct net_device *netdev_wait_allrefs_any(struct list_head *list) 11133 { 11134 unsigned long rebroadcast_time, warning_time; 11135 struct net_device *dev; 11136 int wait = 0; 11137 11138 rebroadcast_time = warning_time = jiffies; 11139 11140 list_for_each_entry(dev, list, todo_list) 11141 if (netdev_refcnt_read(dev) == 1) 11142 return dev; 11143 11144 while (true) { 11145 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) { 11146 rtnl_lock(); 11147 11148 /* Rebroadcast unregister notification */ 11149 list_for_each_entry(dev, list, todo_list) 11150 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11151 11152 __rtnl_unlock(); 11153 rcu_barrier(); 11154 rtnl_lock(); 11155 11156 list_for_each_entry(dev, list, todo_list) 11157 if (test_bit(__LINK_STATE_LINKWATCH_PENDING, 11158 &dev->state)) { 11159 /* We must not have linkwatch events 11160 * pending on unregister. If this 11161 * happens, we simply run the queue 11162 * unscheduled, resulting in a noop 11163 * for this device. 11164 */ 11165 linkwatch_run_queue(); 11166 break; 11167 } 11168 11169 __rtnl_unlock(); 11170 11171 rebroadcast_time = jiffies; 11172 } 11173 11174 rcu_barrier(); 11175 11176 if (!wait) { 11177 wait = WAIT_REFS_MIN_MSECS; 11178 } else { 11179 msleep(wait); 11180 wait = min(wait << 1, WAIT_REFS_MAX_MSECS); 11181 } 11182 11183 list_for_each_entry(dev, list, todo_list) 11184 if (netdev_refcnt_read(dev) == 1) 11185 return dev; 11186 11187 if (time_after(jiffies, warning_time + 11188 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) { 11189 list_for_each_entry(dev, list, todo_list) { 11190 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n", 11191 dev->name, netdev_refcnt_read(dev)); 11192 ref_tracker_dir_print(&dev->refcnt_tracker, 10); 11193 } 11194 11195 warning_time = jiffies; 11196 } 11197 } 11198 } 11199 11200 /* The sequence is: 11201 * 11202 * rtnl_lock(); 11203 * ... 11204 * register_netdevice(x1); 11205 * register_netdevice(x2); 11206 * ... 11207 * unregister_netdevice(y1); 11208 * unregister_netdevice(y2); 11209 * ... 11210 * rtnl_unlock(); 11211 * free_netdev(y1); 11212 * free_netdev(y2); 11213 * 11214 * We are invoked by rtnl_unlock(). 11215 * This allows us to deal with problems: 11216 * 1) We can delete sysfs objects which invoke hotplug 11217 * without deadlocking with linkwatch via keventd. 11218 * 2) Since we run with the RTNL semaphore not held, we can sleep 11219 * safely in order to wait for the netdev refcnt to drop to zero. 11220 * 11221 * We must not return until all unregister events added during 11222 * the interval the lock was held have been completed. 11223 */ 11224 void netdev_run_todo(void) 11225 { 11226 struct net_device *dev, *tmp; 11227 struct list_head list; 11228 int cnt; 11229 #ifdef CONFIG_LOCKDEP 11230 struct list_head unlink_list; 11231 11232 list_replace_init(&net_unlink_list, &unlink_list); 11233 11234 while (!list_empty(&unlink_list)) { 11235 dev = list_first_entry(&unlink_list, struct net_device, 11236 unlink_list); 11237 list_del_init(&dev->unlink_list); 11238 dev->nested_level = dev->lower_level - 1; 11239 } 11240 #endif 11241 11242 /* Snapshot list, allow later requests */ 11243 list_replace_init(&net_todo_list, &list); 11244 11245 __rtnl_unlock(); 11246 11247 /* Wait for rcu callbacks to finish before next phase */ 11248 if (!list_empty(&list)) 11249 rcu_barrier(); 11250 11251 list_for_each_entry_safe(dev, tmp, &list, todo_list) { 11252 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) { 11253 netdev_WARN(dev, "run_todo but not unregistering\n"); 11254 list_del(&dev->todo_list); 11255 continue; 11256 } 11257 11258 netdev_lock(dev); 11259 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERED); 11260 netdev_unlock(dev); 11261 linkwatch_sync_dev(dev); 11262 } 11263 11264 cnt = 0; 11265 while (!list_empty(&list)) { 11266 dev = netdev_wait_allrefs_any(&list); 11267 list_del(&dev->todo_list); 11268 11269 /* paranoia */ 11270 BUG_ON(netdev_refcnt_read(dev) != 1); 11271 BUG_ON(!list_empty(&dev->ptype_all)); 11272 BUG_ON(!list_empty(&dev->ptype_specific)); 11273 WARN_ON(rcu_access_pointer(dev->ip_ptr)); 11274 WARN_ON(rcu_access_pointer(dev->ip6_ptr)); 11275 11276 netdev_do_free_pcpu_stats(dev); 11277 if (dev->priv_destructor) 11278 dev->priv_destructor(dev); 11279 if (dev->needs_free_netdev) 11280 free_netdev(dev); 11281 11282 cnt++; 11283 11284 /* Free network device */ 11285 kobject_put(&dev->dev.kobj); 11286 } 11287 if (cnt && atomic_sub_and_test(cnt, &dev_unreg_count)) 11288 wake_up(&netdev_unregistering_wq); 11289 } 11290 11291 /* Collate per-cpu network dstats statistics 11292 * 11293 * Read per-cpu network statistics from dev->dstats and populate the related 11294 * fields in @s. 11295 */ 11296 static void dev_fetch_dstats(struct rtnl_link_stats64 *s, 11297 const struct pcpu_dstats __percpu *dstats) 11298 { 11299 int cpu; 11300 11301 for_each_possible_cpu(cpu) { 11302 u64 rx_packets, rx_bytes, rx_drops; 11303 u64 tx_packets, tx_bytes, tx_drops; 11304 const struct pcpu_dstats *stats; 11305 unsigned int start; 11306 11307 stats = per_cpu_ptr(dstats, cpu); 11308 do { 11309 start = u64_stats_fetch_begin(&stats->syncp); 11310 rx_packets = u64_stats_read(&stats->rx_packets); 11311 rx_bytes = u64_stats_read(&stats->rx_bytes); 11312 rx_drops = u64_stats_read(&stats->rx_drops); 11313 tx_packets = u64_stats_read(&stats->tx_packets); 11314 tx_bytes = u64_stats_read(&stats->tx_bytes); 11315 tx_drops = u64_stats_read(&stats->tx_drops); 11316 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11317 11318 s->rx_packets += rx_packets; 11319 s->rx_bytes += rx_bytes; 11320 s->rx_dropped += rx_drops; 11321 s->tx_packets += tx_packets; 11322 s->tx_bytes += tx_bytes; 11323 s->tx_dropped += tx_drops; 11324 } 11325 } 11326 11327 /* ndo_get_stats64 implementation for dtstats-based accounting. 11328 * 11329 * Populate @s from dev->stats and dev->dstats. This is used internally by the 11330 * core for NETDEV_PCPU_STAT_DSTAT-type stats collection. 11331 */ 11332 static void dev_get_dstats64(const struct net_device *dev, 11333 struct rtnl_link_stats64 *s) 11334 { 11335 netdev_stats_to_stats64(s, &dev->stats); 11336 dev_fetch_dstats(s, dev->dstats); 11337 } 11338 11339 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has 11340 * all the same fields in the same order as net_device_stats, with only 11341 * the type differing, but rtnl_link_stats64 may have additional fields 11342 * at the end for newer counters. 11343 */ 11344 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64, 11345 const struct net_device_stats *netdev_stats) 11346 { 11347 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t); 11348 const atomic_long_t *src = (atomic_long_t *)netdev_stats; 11349 u64 *dst = (u64 *)stats64; 11350 11351 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64)); 11352 for (i = 0; i < n; i++) 11353 dst[i] = (unsigned long)atomic_long_read(&src[i]); 11354 /* zero out counters that only exist in rtnl_link_stats64 */ 11355 memset((char *)stats64 + n * sizeof(u64), 0, 11356 sizeof(*stats64) - n * sizeof(u64)); 11357 } 11358 EXPORT_SYMBOL(netdev_stats_to_stats64); 11359 11360 static __cold struct net_device_core_stats __percpu *netdev_core_stats_alloc( 11361 struct net_device *dev) 11362 { 11363 struct net_device_core_stats __percpu *p; 11364 11365 p = alloc_percpu_gfp(struct net_device_core_stats, 11366 GFP_ATOMIC | __GFP_NOWARN); 11367 11368 if (p && cmpxchg(&dev->core_stats, NULL, p)) 11369 free_percpu(p); 11370 11371 /* This READ_ONCE() pairs with the cmpxchg() above */ 11372 return READ_ONCE(dev->core_stats); 11373 } 11374 11375 noinline void netdev_core_stats_inc(struct net_device *dev, u32 offset) 11376 { 11377 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11378 struct net_device_core_stats __percpu *p = READ_ONCE(dev->core_stats); 11379 unsigned long __percpu *field; 11380 11381 if (unlikely(!p)) { 11382 p = netdev_core_stats_alloc(dev); 11383 if (!p) 11384 return; 11385 } 11386 11387 field = (unsigned long __percpu *)((void __percpu *)p + offset); 11388 this_cpu_inc(*field); 11389 } 11390 EXPORT_SYMBOL_GPL(netdev_core_stats_inc); 11391 11392 /** 11393 * dev_get_stats - get network device statistics 11394 * @dev: device to get statistics from 11395 * @storage: place to store stats 11396 * 11397 * Get network statistics from device. Return @storage. 11398 * The device driver may provide its own method by setting 11399 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats; 11400 * otherwise the internal statistics structure is used. 11401 */ 11402 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev, 11403 struct rtnl_link_stats64 *storage) 11404 { 11405 const struct net_device_ops *ops = dev->netdev_ops; 11406 const struct net_device_core_stats __percpu *p; 11407 11408 /* 11409 * IPv{4,6} and udp tunnels share common stat helpers and use 11410 * different stat type (NETDEV_PCPU_STAT_TSTATS vs 11411 * NETDEV_PCPU_STAT_DSTATS). Ensure the accounting is consistent. 11412 */ 11413 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_bytes) != 11414 offsetof(struct pcpu_dstats, rx_bytes)); 11415 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_packets) != 11416 offsetof(struct pcpu_dstats, rx_packets)); 11417 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_bytes) != 11418 offsetof(struct pcpu_dstats, tx_bytes)); 11419 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_packets) != 11420 offsetof(struct pcpu_dstats, tx_packets)); 11421 11422 if (ops->ndo_get_stats64) { 11423 memset(storage, 0, sizeof(*storage)); 11424 ops->ndo_get_stats64(dev, storage); 11425 } else if (ops->ndo_get_stats) { 11426 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev)); 11427 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_TSTATS) { 11428 dev_get_tstats64(dev, storage); 11429 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_DSTATS) { 11430 dev_get_dstats64(dev, storage); 11431 } else { 11432 netdev_stats_to_stats64(storage, &dev->stats); 11433 } 11434 11435 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11436 p = READ_ONCE(dev->core_stats); 11437 if (p) { 11438 const struct net_device_core_stats *core_stats; 11439 int i; 11440 11441 for_each_possible_cpu(i) { 11442 core_stats = per_cpu_ptr(p, i); 11443 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped); 11444 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped); 11445 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler); 11446 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped); 11447 } 11448 } 11449 return storage; 11450 } 11451 EXPORT_SYMBOL(dev_get_stats); 11452 11453 /** 11454 * dev_fetch_sw_netstats - get per-cpu network device statistics 11455 * @s: place to store stats 11456 * @netstats: per-cpu network stats to read from 11457 * 11458 * Read per-cpu network statistics and populate the related fields in @s. 11459 */ 11460 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s, 11461 const struct pcpu_sw_netstats __percpu *netstats) 11462 { 11463 int cpu; 11464 11465 for_each_possible_cpu(cpu) { 11466 u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 11467 const struct pcpu_sw_netstats *stats; 11468 unsigned int start; 11469 11470 stats = per_cpu_ptr(netstats, cpu); 11471 do { 11472 start = u64_stats_fetch_begin(&stats->syncp); 11473 rx_packets = u64_stats_read(&stats->rx_packets); 11474 rx_bytes = u64_stats_read(&stats->rx_bytes); 11475 tx_packets = u64_stats_read(&stats->tx_packets); 11476 tx_bytes = u64_stats_read(&stats->tx_bytes); 11477 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11478 11479 s->rx_packets += rx_packets; 11480 s->rx_bytes += rx_bytes; 11481 s->tx_packets += tx_packets; 11482 s->tx_bytes += tx_bytes; 11483 } 11484 } 11485 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats); 11486 11487 /** 11488 * dev_get_tstats64 - ndo_get_stats64 implementation 11489 * @dev: device to get statistics from 11490 * @s: place to store stats 11491 * 11492 * Populate @s from dev->stats and dev->tstats. Can be used as 11493 * ndo_get_stats64() callback. 11494 */ 11495 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s) 11496 { 11497 netdev_stats_to_stats64(s, &dev->stats); 11498 dev_fetch_sw_netstats(s, dev->tstats); 11499 } 11500 EXPORT_SYMBOL_GPL(dev_get_tstats64); 11501 11502 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev) 11503 { 11504 struct netdev_queue *queue = dev_ingress_queue(dev); 11505 11506 #ifdef CONFIG_NET_CLS_ACT 11507 if (queue) 11508 return queue; 11509 queue = kzalloc(sizeof(*queue), GFP_KERNEL); 11510 if (!queue) 11511 return NULL; 11512 netdev_init_one_queue(dev, queue, NULL); 11513 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc); 11514 RCU_INIT_POINTER(queue->qdisc_sleeping, &noop_qdisc); 11515 rcu_assign_pointer(dev->ingress_queue, queue); 11516 #endif 11517 return queue; 11518 } 11519 11520 static const struct ethtool_ops default_ethtool_ops; 11521 11522 void netdev_set_default_ethtool_ops(struct net_device *dev, 11523 const struct ethtool_ops *ops) 11524 { 11525 if (dev->ethtool_ops == &default_ethtool_ops) 11526 dev->ethtool_ops = ops; 11527 } 11528 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops); 11529 11530 /** 11531 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default 11532 * @dev: netdev to enable the IRQ coalescing on 11533 * 11534 * Sets a conservative default for SW IRQ coalescing. Users can use 11535 * sysfs attributes to override the default values. 11536 */ 11537 void netdev_sw_irq_coalesce_default_on(struct net_device *dev) 11538 { 11539 WARN_ON(dev->reg_state == NETREG_REGISTERED); 11540 11541 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) { 11542 netdev_set_gro_flush_timeout(dev, 20000); 11543 netdev_set_defer_hard_irqs(dev, 1); 11544 } 11545 } 11546 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on); 11547 11548 /** 11549 * alloc_netdev_mqs - allocate network device 11550 * @sizeof_priv: size of private data to allocate space for 11551 * @name: device name format string 11552 * @name_assign_type: origin of device name 11553 * @setup: callback to initialize device 11554 * @txqs: the number of TX subqueues to allocate 11555 * @rxqs: the number of RX subqueues to allocate 11556 * 11557 * Allocates a struct net_device with private data area for driver use 11558 * and performs basic initialization. Also allocates subqueue structs 11559 * for each queue on the device. 11560 */ 11561 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name, 11562 unsigned char name_assign_type, 11563 void (*setup)(struct net_device *), 11564 unsigned int txqs, unsigned int rxqs) 11565 { 11566 struct net_device *dev; 11567 size_t napi_config_sz; 11568 unsigned int maxqs; 11569 11570 BUG_ON(strlen(name) >= sizeof(dev->name)); 11571 11572 if (txqs < 1) { 11573 pr_err("alloc_netdev: Unable to allocate device with zero queues\n"); 11574 return NULL; 11575 } 11576 11577 if (rxqs < 1) { 11578 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n"); 11579 return NULL; 11580 } 11581 11582 maxqs = max(txqs, rxqs); 11583 11584 dev = kvzalloc(struct_size(dev, priv, sizeof_priv), 11585 GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11586 if (!dev) 11587 return NULL; 11588 11589 dev->priv_len = sizeof_priv; 11590 11591 ref_tracker_dir_init(&dev->refcnt_tracker, 128, name); 11592 #ifdef CONFIG_PCPU_DEV_REFCNT 11593 dev->pcpu_refcnt = alloc_percpu(int); 11594 if (!dev->pcpu_refcnt) 11595 goto free_dev; 11596 __dev_hold(dev); 11597 #else 11598 refcount_set(&dev->dev_refcnt, 1); 11599 #endif 11600 11601 if (dev_addr_init(dev)) 11602 goto free_pcpu; 11603 11604 dev_mc_init(dev); 11605 dev_uc_init(dev); 11606 11607 dev_net_set(dev, &init_net); 11608 11609 dev->gso_max_size = GSO_LEGACY_MAX_SIZE; 11610 dev->xdp_zc_max_segs = 1; 11611 dev->gso_max_segs = GSO_MAX_SEGS; 11612 dev->gro_max_size = GRO_LEGACY_MAX_SIZE; 11613 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE; 11614 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE; 11615 dev->tso_max_size = TSO_LEGACY_MAX_SIZE; 11616 dev->tso_max_segs = TSO_MAX_SEGS; 11617 dev->upper_level = 1; 11618 dev->lower_level = 1; 11619 #ifdef CONFIG_LOCKDEP 11620 dev->nested_level = 0; 11621 INIT_LIST_HEAD(&dev->unlink_list); 11622 #endif 11623 11624 INIT_LIST_HEAD(&dev->napi_list); 11625 INIT_LIST_HEAD(&dev->unreg_list); 11626 INIT_LIST_HEAD(&dev->close_list); 11627 INIT_LIST_HEAD(&dev->link_watch_list); 11628 INIT_LIST_HEAD(&dev->adj_list.upper); 11629 INIT_LIST_HEAD(&dev->adj_list.lower); 11630 INIT_LIST_HEAD(&dev->ptype_all); 11631 INIT_LIST_HEAD(&dev->ptype_specific); 11632 INIT_LIST_HEAD(&dev->net_notifier_list); 11633 #ifdef CONFIG_NET_SCHED 11634 hash_init(dev->qdisc_hash); 11635 #endif 11636 11637 mutex_init(&dev->lock); 11638 11639 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 11640 setup(dev); 11641 11642 if (!dev->tx_queue_len) { 11643 dev->priv_flags |= IFF_NO_QUEUE; 11644 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN; 11645 } 11646 11647 dev->num_tx_queues = txqs; 11648 dev->real_num_tx_queues = txqs; 11649 if (netif_alloc_netdev_queues(dev)) 11650 goto free_all; 11651 11652 dev->num_rx_queues = rxqs; 11653 dev->real_num_rx_queues = rxqs; 11654 if (netif_alloc_rx_queues(dev)) 11655 goto free_all; 11656 dev->ethtool = kzalloc(sizeof(*dev->ethtool), GFP_KERNEL_ACCOUNT); 11657 if (!dev->ethtool) 11658 goto free_all; 11659 11660 dev->cfg = kzalloc(sizeof(*dev->cfg), GFP_KERNEL_ACCOUNT); 11661 if (!dev->cfg) 11662 goto free_all; 11663 dev->cfg_pending = dev->cfg; 11664 11665 napi_config_sz = array_size(maxqs, sizeof(*dev->napi_config)); 11666 dev->napi_config = kvzalloc(napi_config_sz, GFP_KERNEL_ACCOUNT); 11667 if (!dev->napi_config) 11668 goto free_all; 11669 11670 strscpy(dev->name, name); 11671 dev->name_assign_type = name_assign_type; 11672 dev->group = INIT_NETDEV_GROUP; 11673 if (!dev->ethtool_ops) 11674 dev->ethtool_ops = &default_ethtool_ops; 11675 11676 nf_hook_netdev_init(dev); 11677 11678 return dev; 11679 11680 free_all: 11681 free_netdev(dev); 11682 return NULL; 11683 11684 free_pcpu: 11685 #ifdef CONFIG_PCPU_DEV_REFCNT 11686 free_percpu(dev->pcpu_refcnt); 11687 free_dev: 11688 #endif 11689 kvfree(dev); 11690 return NULL; 11691 } 11692 EXPORT_SYMBOL(alloc_netdev_mqs); 11693 11694 static void netdev_napi_exit(struct net_device *dev) 11695 { 11696 if (!list_empty(&dev->napi_list)) { 11697 struct napi_struct *p, *n; 11698 11699 netdev_lock(dev); 11700 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list) 11701 __netif_napi_del_locked(p); 11702 netdev_unlock(dev); 11703 11704 synchronize_net(); 11705 } 11706 11707 kvfree(dev->napi_config); 11708 } 11709 11710 /** 11711 * free_netdev - free network device 11712 * @dev: device 11713 * 11714 * This function does the last stage of destroying an allocated device 11715 * interface. The reference to the device object is released. If this 11716 * is the last reference then it will be freed.Must be called in process 11717 * context. 11718 */ 11719 void free_netdev(struct net_device *dev) 11720 { 11721 might_sleep(); 11722 11723 /* When called immediately after register_netdevice() failed the unwind 11724 * handling may still be dismantling the device. Handle that case by 11725 * deferring the free. 11726 */ 11727 if (dev->reg_state == NETREG_UNREGISTERING) { 11728 ASSERT_RTNL(); 11729 dev->needs_free_netdev = true; 11730 return; 11731 } 11732 11733 WARN_ON(dev->cfg != dev->cfg_pending); 11734 kfree(dev->cfg); 11735 kfree(dev->ethtool); 11736 netif_free_tx_queues(dev); 11737 netif_free_rx_queues(dev); 11738 11739 kfree(rcu_dereference_protected(dev->ingress_queue, 1)); 11740 11741 /* Flush device addresses */ 11742 dev_addr_flush(dev); 11743 11744 netdev_napi_exit(dev); 11745 11746 netif_del_cpu_rmap(dev); 11747 11748 ref_tracker_dir_exit(&dev->refcnt_tracker); 11749 #ifdef CONFIG_PCPU_DEV_REFCNT 11750 free_percpu(dev->pcpu_refcnt); 11751 dev->pcpu_refcnt = NULL; 11752 #endif 11753 free_percpu(dev->core_stats); 11754 dev->core_stats = NULL; 11755 free_percpu(dev->xdp_bulkq); 11756 dev->xdp_bulkq = NULL; 11757 11758 netdev_free_phy_link_topology(dev); 11759 11760 mutex_destroy(&dev->lock); 11761 11762 /* Compatibility with error handling in drivers */ 11763 if (dev->reg_state == NETREG_UNINITIALIZED || 11764 dev->reg_state == NETREG_DUMMY) { 11765 kvfree(dev); 11766 return; 11767 } 11768 11769 BUG_ON(dev->reg_state != NETREG_UNREGISTERED); 11770 WRITE_ONCE(dev->reg_state, NETREG_RELEASED); 11771 11772 /* will free via device release */ 11773 put_device(&dev->dev); 11774 } 11775 EXPORT_SYMBOL(free_netdev); 11776 11777 /** 11778 * alloc_netdev_dummy - Allocate and initialize a dummy net device. 11779 * @sizeof_priv: size of private data to allocate space for 11780 * 11781 * Return: the allocated net_device on success, NULL otherwise 11782 */ 11783 struct net_device *alloc_netdev_dummy(int sizeof_priv) 11784 { 11785 return alloc_netdev(sizeof_priv, "dummy#", NET_NAME_UNKNOWN, 11786 init_dummy_netdev); 11787 } 11788 EXPORT_SYMBOL_GPL(alloc_netdev_dummy); 11789 11790 /** 11791 * synchronize_net - Synchronize with packet receive processing 11792 * 11793 * Wait for packets currently being received to be done. 11794 * Does not block later packets from starting. 11795 */ 11796 void synchronize_net(void) 11797 { 11798 might_sleep(); 11799 if (from_cleanup_net() || rtnl_is_locked()) 11800 synchronize_rcu_expedited(); 11801 else 11802 synchronize_rcu(); 11803 } 11804 EXPORT_SYMBOL(synchronize_net); 11805 11806 static void netdev_rss_contexts_free(struct net_device *dev) 11807 { 11808 struct ethtool_rxfh_context *ctx; 11809 unsigned long context; 11810 11811 mutex_lock(&dev->ethtool->rss_lock); 11812 xa_for_each(&dev->ethtool->rss_ctx, context, ctx) { 11813 struct ethtool_rxfh_param rxfh; 11814 11815 rxfh.indir = ethtool_rxfh_context_indir(ctx); 11816 rxfh.key = ethtool_rxfh_context_key(ctx); 11817 rxfh.hfunc = ctx->hfunc; 11818 rxfh.input_xfrm = ctx->input_xfrm; 11819 rxfh.rss_context = context; 11820 rxfh.rss_delete = true; 11821 11822 xa_erase(&dev->ethtool->rss_ctx, context); 11823 if (dev->ethtool_ops->create_rxfh_context) 11824 dev->ethtool_ops->remove_rxfh_context(dev, ctx, 11825 context, NULL); 11826 else 11827 dev->ethtool_ops->set_rxfh(dev, &rxfh, NULL); 11828 kfree(ctx); 11829 } 11830 xa_destroy(&dev->ethtool->rss_ctx); 11831 mutex_unlock(&dev->ethtool->rss_lock); 11832 } 11833 11834 /** 11835 * unregister_netdevice_queue - remove device from the kernel 11836 * @dev: device 11837 * @head: list 11838 * 11839 * This function shuts down a device interface and removes it 11840 * from the kernel tables. 11841 * If head not NULL, device is queued to be unregistered later. 11842 * 11843 * Callers must hold the rtnl semaphore. You may want 11844 * unregister_netdev() instead of this. 11845 */ 11846 11847 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head) 11848 { 11849 ASSERT_RTNL(); 11850 11851 if (head) { 11852 list_move_tail(&dev->unreg_list, head); 11853 } else { 11854 LIST_HEAD(single); 11855 11856 list_add(&dev->unreg_list, &single); 11857 unregister_netdevice_many(&single); 11858 } 11859 } 11860 EXPORT_SYMBOL(unregister_netdevice_queue); 11861 11862 static void dev_memory_provider_uninstall(struct net_device *dev) 11863 { 11864 unsigned int i; 11865 11866 for (i = 0; i < dev->real_num_rx_queues; i++) { 11867 struct netdev_rx_queue *rxq = &dev->_rx[i]; 11868 struct pp_memory_provider_params *p = &rxq->mp_params; 11869 11870 if (p->mp_ops && p->mp_ops->uninstall) 11871 p->mp_ops->uninstall(rxq->mp_params.mp_priv, rxq); 11872 } 11873 } 11874 11875 void unregister_netdevice_many_notify(struct list_head *head, 11876 u32 portid, const struct nlmsghdr *nlh) 11877 { 11878 struct net_device *dev, *tmp; 11879 LIST_HEAD(close_head); 11880 int cnt = 0; 11881 11882 BUG_ON(dev_boot_phase); 11883 ASSERT_RTNL(); 11884 11885 if (list_empty(head)) 11886 return; 11887 11888 list_for_each_entry_safe(dev, tmp, head, unreg_list) { 11889 /* Some devices call without registering 11890 * for initialization unwind. Remove those 11891 * devices and proceed with the remaining. 11892 */ 11893 if (dev->reg_state == NETREG_UNINITIALIZED) { 11894 pr_debug("unregister_netdevice: device %s/%p never was registered\n", 11895 dev->name, dev); 11896 11897 WARN_ON(1); 11898 list_del(&dev->unreg_list); 11899 continue; 11900 } 11901 dev->dismantle = true; 11902 BUG_ON(dev->reg_state != NETREG_REGISTERED); 11903 } 11904 11905 /* If device is running, close it first. */ 11906 list_for_each_entry(dev, head, unreg_list) { 11907 list_add_tail(&dev->close_list, &close_head); 11908 netdev_lock_ops(dev); 11909 } 11910 dev_close_many(&close_head, true); 11911 11912 list_for_each_entry(dev, head, unreg_list) { 11913 netdev_unlock_ops(dev); 11914 /* And unlink it from device chain. */ 11915 unlist_netdevice(dev); 11916 netdev_lock(dev); 11917 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERING); 11918 netdev_unlock(dev); 11919 } 11920 flush_all_backlogs(); 11921 11922 synchronize_net(); 11923 11924 list_for_each_entry(dev, head, unreg_list) { 11925 struct sk_buff *skb = NULL; 11926 11927 /* Shutdown queueing discipline. */ 11928 dev_shutdown(dev); 11929 dev_tcx_uninstall(dev); 11930 netdev_lock_ops(dev); 11931 dev_xdp_uninstall(dev); 11932 netdev_unlock_ops(dev); 11933 bpf_dev_bound_netdev_unregister(dev); 11934 dev_memory_provider_uninstall(dev); 11935 11936 netdev_offload_xstats_disable_all(dev); 11937 11938 /* Notify protocols, that we are about to destroy 11939 * this device. They should clean all the things. 11940 */ 11941 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11942 11943 if (!dev->rtnl_link_ops || 11944 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 11945 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0, 11946 GFP_KERNEL, NULL, 0, 11947 portid, nlh); 11948 11949 /* 11950 * Flush the unicast and multicast chains 11951 */ 11952 dev_uc_flush(dev); 11953 dev_mc_flush(dev); 11954 11955 netdev_name_node_alt_flush(dev); 11956 netdev_name_node_free(dev->name_node); 11957 11958 netdev_rss_contexts_free(dev); 11959 11960 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11961 11962 if (dev->netdev_ops->ndo_uninit) 11963 dev->netdev_ops->ndo_uninit(dev); 11964 11965 mutex_destroy(&dev->ethtool->rss_lock); 11966 11967 net_shaper_flush_netdev(dev); 11968 11969 if (skb) 11970 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh); 11971 11972 /* Notifier chain MUST detach us all upper devices. */ 11973 WARN_ON(netdev_has_any_upper_dev(dev)); 11974 WARN_ON(netdev_has_any_lower_dev(dev)); 11975 11976 /* Remove entries from kobject tree */ 11977 netdev_unregister_kobject(dev); 11978 #ifdef CONFIG_XPS 11979 /* Remove XPS queueing entries */ 11980 netif_reset_xps_queues_gt(dev, 0); 11981 #endif 11982 } 11983 11984 synchronize_net(); 11985 11986 list_for_each_entry(dev, head, unreg_list) { 11987 netdev_put(dev, &dev->dev_registered_tracker); 11988 net_set_todo(dev); 11989 cnt++; 11990 } 11991 atomic_add(cnt, &dev_unreg_count); 11992 11993 list_del(head); 11994 } 11995 11996 /** 11997 * unregister_netdevice_many - unregister many devices 11998 * @head: list of devices 11999 * 12000 * Note: As most callers use a stack allocated list_head, 12001 * we force a list_del() to make sure stack won't be corrupted later. 12002 */ 12003 void unregister_netdevice_many(struct list_head *head) 12004 { 12005 unregister_netdevice_many_notify(head, 0, NULL); 12006 } 12007 EXPORT_SYMBOL(unregister_netdevice_many); 12008 12009 /** 12010 * unregister_netdev - remove device from the kernel 12011 * @dev: device 12012 * 12013 * This function shuts down a device interface and removes it 12014 * from the kernel tables. 12015 * 12016 * This is just a wrapper for unregister_netdevice that takes 12017 * the rtnl semaphore. In general you want to use this and not 12018 * unregister_netdevice. 12019 */ 12020 void unregister_netdev(struct net_device *dev) 12021 { 12022 rtnl_net_dev_lock(dev); 12023 unregister_netdevice(dev); 12024 rtnl_net_dev_unlock(dev); 12025 } 12026 EXPORT_SYMBOL(unregister_netdev); 12027 12028 int netif_change_net_namespace(struct net_device *dev, struct net *net, 12029 const char *pat, int new_ifindex, 12030 struct netlink_ext_ack *extack) 12031 { 12032 struct netdev_name_node *name_node; 12033 struct net *net_old = dev_net(dev); 12034 char new_name[IFNAMSIZ] = {}; 12035 int err, new_nsid; 12036 12037 ASSERT_RTNL(); 12038 12039 /* Don't allow namespace local devices to be moved. */ 12040 err = -EINVAL; 12041 if (dev->netns_immutable) { 12042 NL_SET_ERR_MSG(extack, "The interface netns is immutable"); 12043 goto out; 12044 } 12045 12046 /* Ensure the device has been registered */ 12047 if (dev->reg_state != NETREG_REGISTERED) { 12048 NL_SET_ERR_MSG(extack, "The interface isn't registered"); 12049 goto out; 12050 } 12051 12052 /* Get out if there is nothing todo */ 12053 err = 0; 12054 if (net_eq(net_old, net)) 12055 goto out; 12056 12057 /* Pick the destination device name, and ensure 12058 * we can use it in the destination network namespace. 12059 */ 12060 err = -EEXIST; 12061 if (netdev_name_in_use(net, dev->name)) { 12062 /* We get here if we can't use the current device name */ 12063 if (!pat) { 12064 NL_SET_ERR_MSG(extack, 12065 "An interface with the same name exists in the target netns"); 12066 goto out; 12067 } 12068 err = dev_prep_valid_name(net, dev, pat, new_name, EEXIST); 12069 if (err < 0) { 12070 NL_SET_ERR_MSG_FMT(extack, 12071 "Unable to use '%s' for the new interface name in the target netns", 12072 pat); 12073 goto out; 12074 } 12075 } 12076 /* Check that none of the altnames conflicts. */ 12077 err = -EEXIST; 12078 netdev_for_each_altname(dev, name_node) { 12079 if (netdev_name_in_use(net, name_node->name)) { 12080 NL_SET_ERR_MSG_FMT(extack, 12081 "An interface with the altname %s exists in the target netns", 12082 name_node->name); 12083 goto out; 12084 } 12085 } 12086 12087 /* Check that new_ifindex isn't used yet. */ 12088 if (new_ifindex) { 12089 err = dev_index_reserve(net, new_ifindex); 12090 if (err < 0) { 12091 NL_SET_ERR_MSG_FMT(extack, 12092 "The ifindex %d is not available in the target netns", 12093 new_ifindex); 12094 goto out; 12095 } 12096 } else { 12097 /* If there is an ifindex conflict assign a new one */ 12098 err = dev_index_reserve(net, dev->ifindex); 12099 if (err == -EBUSY) 12100 err = dev_index_reserve(net, 0); 12101 if (err < 0) { 12102 NL_SET_ERR_MSG(extack, 12103 "Unable to allocate a new ifindex in the target netns"); 12104 goto out; 12105 } 12106 new_ifindex = err; 12107 } 12108 12109 /* 12110 * And now a mini version of register_netdevice unregister_netdevice. 12111 */ 12112 12113 /* If device is running close it first. */ 12114 netif_close(dev); 12115 12116 /* And unlink it from device chain */ 12117 unlist_netdevice(dev); 12118 12119 synchronize_net(); 12120 12121 /* Shutdown queueing discipline. */ 12122 dev_shutdown(dev); 12123 12124 /* Notify protocols, that we are about to destroy 12125 * this device. They should clean all the things. 12126 * 12127 * Note that dev->reg_state stays at NETREG_REGISTERED. 12128 * This is wanted because this way 8021q and macvlan know 12129 * the device is just moving and can keep their slaves up. 12130 */ 12131 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 12132 rcu_barrier(); 12133 12134 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL); 12135 12136 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid, 12137 new_ifindex); 12138 12139 /* 12140 * Flush the unicast and multicast chains 12141 */ 12142 dev_uc_flush(dev); 12143 dev_mc_flush(dev); 12144 12145 /* Send a netdev-removed uevent to the old namespace */ 12146 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE); 12147 netdev_adjacent_del_links(dev); 12148 12149 /* Move per-net netdevice notifiers that are following the netdevice */ 12150 move_netdevice_notifiers_dev_net(dev, net); 12151 12152 /* Actually switch the network namespace */ 12153 dev_net_set(dev, net); 12154 dev->ifindex = new_ifindex; 12155 12156 if (new_name[0]) { 12157 /* Rename the netdev to prepared name */ 12158 write_seqlock_bh(&netdev_rename_lock); 12159 strscpy(dev->name, new_name, IFNAMSIZ); 12160 write_sequnlock_bh(&netdev_rename_lock); 12161 } 12162 12163 /* Fixup kobjects */ 12164 dev_set_uevent_suppress(&dev->dev, 1); 12165 err = device_rename(&dev->dev, dev->name); 12166 dev_set_uevent_suppress(&dev->dev, 0); 12167 WARN_ON(err); 12168 12169 /* Send a netdev-add uevent to the new namespace */ 12170 kobject_uevent(&dev->dev.kobj, KOBJ_ADD); 12171 netdev_adjacent_add_links(dev); 12172 12173 /* Adapt owner in case owning user namespace of target network 12174 * namespace is different from the original one. 12175 */ 12176 err = netdev_change_owner(dev, net_old, net); 12177 WARN_ON(err); 12178 12179 /* Add the device back in the hashes */ 12180 list_netdevice(dev); 12181 12182 /* Notify protocols, that a new device appeared. */ 12183 call_netdevice_notifiers(NETDEV_REGISTER, dev); 12184 12185 /* 12186 * Prevent userspace races by waiting until the network 12187 * device is fully setup before sending notifications. 12188 */ 12189 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 12190 12191 synchronize_net(); 12192 err = 0; 12193 out: 12194 return err; 12195 } 12196 12197 static int dev_cpu_dead(unsigned int oldcpu) 12198 { 12199 struct sk_buff **list_skb; 12200 struct sk_buff *skb; 12201 unsigned int cpu; 12202 struct softnet_data *sd, *oldsd, *remsd = NULL; 12203 12204 local_irq_disable(); 12205 cpu = smp_processor_id(); 12206 sd = &per_cpu(softnet_data, cpu); 12207 oldsd = &per_cpu(softnet_data, oldcpu); 12208 12209 /* Find end of our completion_queue. */ 12210 list_skb = &sd->completion_queue; 12211 while (*list_skb) 12212 list_skb = &(*list_skb)->next; 12213 /* Append completion queue from offline CPU. */ 12214 *list_skb = oldsd->completion_queue; 12215 oldsd->completion_queue = NULL; 12216 12217 /* Append output queue from offline CPU. */ 12218 if (oldsd->output_queue) { 12219 *sd->output_queue_tailp = oldsd->output_queue; 12220 sd->output_queue_tailp = oldsd->output_queue_tailp; 12221 oldsd->output_queue = NULL; 12222 oldsd->output_queue_tailp = &oldsd->output_queue; 12223 } 12224 /* Append NAPI poll list from offline CPU, with one exception : 12225 * process_backlog() must be called by cpu owning percpu backlog. 12226 * We properly handle process_queue & input_pkt_queue later. 12227 */ 12228 while (!list_empty(&oldsd->poll_list)) { 12229 struct napi_struct *napi = list_first_entry(&oldsd->poll_list, 12230 struct napi_struct, 12231 poll_list); 12232 12233 list_del_init(&napi->poll_list); 12234 if (napi->poll == process_backlog) 12235 napi->state &= NAPIF_STATE_THREADED; 12236 else 12237 ____napi_schedule(sd, napi); 12238 } 12239 12240 raise_softirq_irqoff(NET_TX_SOFTIRQ); 12241 local_irq_enable(); 12242 12243 if (!use_backlog_threads()) { 12244 #ifdef CONFIG_RPS 12245 remsd = oldsd->rps_ipi_list; 12246 oldsd->rps_ipi_list = NULL; 12247 #endif 12248 /* send out pending IPI's on offline CPU */ 12249 net_rps_send_ipi(remsd); 12250 } 12251 12252 /* Process offline CPU's input_pkt_queue */ 12253 while ((skb = __skb_dequeue(&oldsd->process_queue))) { 12254 netif_rx(skb); 12255 rps_input_queue_head_incr(oldsd); 12256 } 12257 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) { 12258 netif_rx(skb); 12259 rps_input_queue_head_incr(oldsd); 12260 } 12261 12262 return 0; 12263 } 12264 12265 /** 12266 * netdev_increment_features - increment feature set by one 12267 * @all: current feature set 12268 * @one: new feature set 12269 * @mask: mask feature set 12270 * 12271 * Computes a new feature set after adding a device with feature set 12272 * @one to the master device with current feature set @all. Will not 12273 * enable anything that is off in @mask. Returns the new feature set. 12274 */ 12275 netdev_features_t netdev_increment_features(netdev_features_t all, 12276 netdev_features_t one, netdev_features_t mask) 12277 { 12278 if (mask & NETIF_F_HW_CSUM) 12279 mask |= NETIF_F_CSUM_MASK; 12280 mask |= NETIF_F_VLAN_CHALLENGED; 12281 12282 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask; 12283 all &= one | ~NETIF_F_ALL_FOR_ALL; 12284 12285 /* If one device supports hw checksumming, set for all. */ 12286 if (all & NETIF_F_HW_CSUM) 12287 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM); 12288 12289 return all; 12290 } 12291 EXPORT_SYMBOL(netdev_increment_features); 12292 12293 static struct hlist_head * __net_init netdev_create_hash(void) 12294 { 12295 int i; 12296 struct hlist_head *hash; 12297 12298 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL); 12299 if (hash != NULL) 12300 for (i = 0; i < NETDEV_HASHENTRIES; i++) 12301 INIT_HLIST_HEAD(&hash[i]); 12302 12303 return hash; 12304 } 12305 12306 /* Initialize per network namespace state */ 12307 static int __net_init netdev_init(struct net *net) 12308 { 12309 BUILD_BUG_ON(GRO_HASH_BUCKETS > 12310 BITS_PER_BYTE * sizeof_field(struct gro_node, bitmask)); 12311 12312 INIT_LIST_HEAD(&net->dev_base_head); 12313 12314 net->dev_name_head = netdev_create_hash(); 12315 if (net->dev_name_head == NULL) 12316 goto err_name; 12317 12318 net->dev_index_head = netdev_create_hash(); 12319 if (net->dev_index_head == NULL) 12320 goto err_idx; 12321 12322 xa_init_flags(&net->dev_by_index, XA_FLAGS_ALLOC1); 12323 12324 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain); 12325 12326 return 0; 12327 12328 err_idx: 12329 kfree(net->dev_name_head); 12330 err_name: 12331 return -ENOMEM; 12332 } 12333 12334 /** 12335 * netdev_drivername - network driver for the device 12336 * @dev: network device 12337 * 12338 * Determine network driver for device. 12339 */ 12340 const char *netdev_drivername(const struct net_device *dev) 12341 { 12342 const struct device_driver *driver; 12343 const struct device *parent; 12344 const char *empty = ""; 12345 12346 parent = dev->dev.parent; 12347 if (!parent) 12348 return empty; 12349 12350 driver = parent->driver; 12351 if (driver && driver->name) 12352 return driver->name; 12353 return empty; 12354 } 12355 12356 static void __netdev_printk(const char *level, const struct net_device *dev, 12357 struct va_format *vaf) 12358 { 12359 if (dev && dev->dev.parent) { 12360 dev_printk_emit(level[1] - '0', 12361 dev->dev.parent, 12362 "%s %s %s%s: %pV", 12363 dev_driver_string(dev->dev.parent), 12364 dev_name(dev->dev.parent), 12365 netdev_name(dev), netdev_reg_state(dev), 12366 vaf); 12367 } else if (dev) { 12368 printk("%s%s%s: %pV", 12369 level, netdev_name(dev), netdev_reg_state(dev), vaf); 12370 } else { 12371 printk("%s(NULL net_device): %pV", level, vaf); 12372 } 12373 } 12374 12375 void netdev_printk(const char *level, const struct net_device *dev, 12376 const char *format, ...) 12377 { 12378 struct va_format vaf; 12379 va_list args; 12380 12381 va_start(args, format); 12382 12383 vaf.fmt = format; 12384 vaf.va = &args; 12385 12386 __netdev_printk(level, dev, &vaf); 12387 12388 va_end(args); 12389 } 12390 EXPORT_SYMBOL(netdev_printk); 12391 12392 #define define_netdev_printk_level(func, level) \ 12393 void func(const struct net_device *dev, const char *fmt, ...) \ 12394 { \ 12395 struct va_format vaf; \ 12396 va_list args; \ 12397 \ 12398 va_start(args, fmt); \ 12399 \ 12400 vaf.fmt = fmt; \ 12401 vaf.va = &args; \ 12402 \ 12403 __netdev_printk(level, dev, &vaf); \ 12404 \ 12405 va_end(args); \ 12406 } \ 12407 EXPORT_SYMBOL(func); 12408 12409 define_netdev_printk_level(netdev_emerg, KERN_EMERG); 12410 define_netdev_printk_level(netdev_alert, KERN_ALERT); 12411 define_netdev_printk_level(netdev_crit, KERN_CRIT); 12412 define_netdev_printk_level(netdev_err, KERN_ERR); 12413 define_netdev_printk_level(netdev_warn, KERN_WARNING); 12414 define_netdev_printk_level(netdev_notice, KERN_NOTICE); 12415 define_netdev_printk_level(netdev_info, KERN_INFO); 12416 12417 static void __net_exit netdev_exit(struct net *net) 12418 { 12419 kfree(net->dev_name_head); 12420 kfree(net->dev_index_head); 12421 xa_destroy(&net->dev_by_index); 12422 if (net != &init_net) 12423 WARN_ON_ONCE(!list_empty(&net->dev_base_head)); 12424 } 12425 12426 static struct pernet_operations __net_initdata netdev_net_ops = { 12427 .init = netdev_init, 12428 .exit = netdev_exit, 12429 }; 12430 12431 static void __net_exit default_device_exit_net(struct net *net) 12432 { 12433 struct netdev_name_node *name_node, *tmp; 12434 struct net_device *dev, *aux; 12435 /* 12436 * Push all migratable network devices back to the 12437 * initial network namespace 12438 */ 12439 ASSERT_RTNL(); 12440 for_each_netdev_safe(net, dev, aux) { 12441 int err; 12442 char fb_name[IFNAMSIZ]; 12443 12444 /* Ignore unmoveable devices (i.e. loopback) */ 12445 if (dev->netns_immutable) 12446 continue; 12447 12448 /* Leave virtual devices for the generic cleanup */ 12449 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund) 12450 continue; 12451 12452 /* Push remaining network devices to init_net */ 12453 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex); 12454 if (netdev_name_in_use(&init_net, fb_name)) 12455 snprintf(fb_name, IFNAMSIZ, "dev%%d"); 12456 12457 netdev_for_each_altname_safe(dev, name_node, tmp) 12458 if (netdev_name_in_use(&init_net, name_node->name)) 12459 __netdev_name_node_alt_destroy(name_node); 12460 12461 err = dev_change_net_namespace(dev, &init_net, fb_name); 12462 if (err) { 12463 pr_emerg("%s: failed to move %s to init_net: %d\n", 12464 __func__, dev->name, err); 12465 BUG(); 12466 } 12467 } 12468 } 12469 12470 static void __net_exit default_device_exit_batch(struct list_head *net_list) 12471 { 12472 /* At exit all network devices most be removed from a network 12473 * namespace. Do this in the reverse order of registration. 12474 * Do this across as many network namespaces as possible to 12475 * improve batching efficiency. 12476 */ 12477 struct net_device *dev; 12478 struct net *net; 12479 LIST_HEAD(dev_kill_list); 12480 12481 rtnl_lock(); 12482 list_for_each_entry(net, net_list, exit_list) { 12483 default_device_exit_net(net); 12484 cond_resched(); 12485 } 12486 12487 list_for_each_entry(net, net_list, exit_list) { 12488 for_each_netdev_reverse(net, dev) { 12489 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink) 12490 dev->rtnl_link_ops->dellink(dev, &dev_kill_list); 12491 else 12492 unregister_netdevice_queue(dev, &dev_kill_list); 12493 } 12494 } 12495 unregister_netdevice_many(&dev_kill_list); 12496 rtnl_unlock(); 12497 } 12498 12499 static struct pernet_operations __net_initdata default_device_ops = { 12500 .exit_batch = default_device_exit_batch, 12501 }; 12502 12503 static void __init net_dev_struct_check(void) 12504 { 12505 /* TX read-mostly hotpath */ 12506 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, priv_flags_fast); 12507 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, netdev_ops); 12508 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, header_ops); 12509 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, _tx); 12510 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, real_num_tx_queues); 12511 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_size); 12512 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_ipv4_max_size); 12513 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_segs); 12514 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_partial_features); 12515 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, num_tc); 12516 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, mtu); 12517 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, needed_headroom); 12518 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tc_to_txq); 12519 #ifdef CONFIG_XPS 12520 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, xps_maps); 12521 #endif 12522 #ifdef CONFIG_NETFILTER_EGRESS 12523 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, nf_hooks_egress); 12524 #endif 12525 #ifdef CONFIG_NET_XGRESS 12526 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tcx_egress); 12527 #endif 12528 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_tx, 160); 12529 12530 /* TXRX read-mostly hotpath */ 12531 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, lstats); 12532 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, state); 12533 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, flags); 12534 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, hard_header_len); 12535 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, features); 12536 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, ip6_ptr); 12537 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_txrx, 46); 12538 12539 /* RX read-mostly hotpath */ 12540 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ptype_specific); 12541 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ifindex); 12542 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, real_num_rx_queues); 12543 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, _rx); 12544 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_max_size); 12545 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_ipv4_max_size); 12546 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler); 12547 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler_data); 12548 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, nd_net); 12549 #ifdef CONFIG_NETPOLL 12550 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, npinfo); 12551 #endif 12552 #ifdef CONFIG_NET_XGRESS 12553 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, tcx_ingress); 12554 #endif 12555 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_rx, 92); 12556 } 12557 12558 /* 12559 * Initialize the DEV module. At boot time this walks the device list and 12560 * unhooks any devices that fail to initialise (normally hardware not 12561 * present) and leaves us with a valid list of present and active devices. 12562 * 12563 */ 12564 12565 /* We allocate 256 pages for each CPU if PAGE_SHIFT is 12 */ 12566 #define SYSTEM_PERCPU_PAGE_POOL_SIZE ((1 << 20) / PAGE_SIZE) 12567 12568 static int net_page_pool_create(int cpuid) 12569 { 12570 #if IS_ENABLED(CONFIG_PAGE_POOL) 12571 struct page_pool_params page_pool_params = { 12572 .pool_size = SYSTEM_PERCPU_PAGE_POOL_SIZE, 12573 .flags = PP_FLAG_SYSTEM_POOL, 12574 .nid = cpu_to_mem(cpuid), 12575 }; 12576 struct page_pool *pp_ptr; 12577 int err; 12578 12579 pp_ptr = page_pool_create_percpu(&page_pool_params, cpuid); 12580 if (IS_ERR(pp_ptr)) 12581 return -ENOMEM; 12582 12583 err = xdp_reg_page_pool(pp_ptr); 12584 if (err) { 12585 page_pool_destroy(pp_ptr); 12586 return err; 12587 } 12588 12589 per_cpu(system_page_pool, cpuid) = pp_ptr; 12590 #endif 12591 return 0; 12592 } 12593 12594 static int backlog_napi_should_run(unsigned int cpu) 12595 { 12596 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12597 struct napi_struct *napi = &sd->backlog; 12598 12599 return test_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 12600 } 12601 12602 static void run_backlog_napi(unsigned int cpu) 12603 { 12604 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12605 12606 napi_threaded_poll_loop(&sd->backlog); 12607 } 12608 12609 static void backlog_napi_setup(unsigned int cpu) 12610 { 12611 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12612 struct napi_struct *napi = &sd->backlog; 12613 12614 napi->thread = this_cpu_read(backlog_napi); 12615 set_bit(NAPI_STATE_THREADED, &napi->state); 12616 } 12617 12618 static struct smp_hotplug_thread backlog_threads = { 12619 .store = &backlog_napi, 12620 .thread_should_run = backlog_napi_should_run, 12621 .thread_fn = run_backlog_napi, 12622 .thread_comm = "backlog_napi/%u", 12623 .setup = backlog_napi_setup, 12624 }; 12625 12626 /* 12627 * This is called single threaded during boot, so no need 12628 * to take the rtnl semaphore. 12629 */ 12630 static int __init net_dev_init(void) 12631 { 12632 int i, rc = -ENOMEM; 12633 12634 BUG_ON(!dev_boot_phase); 12635 12636 net_dev_struct_check(); 12637 12638 if (dev_proc_init()) 12639 goto out; 12640 12641 if (netdev_kobject_init()) 12642 goto out; 12643 12644 for (i = 0; i < PTYPE_HASH_SIZE; i++) 12645 INIT_LIST_HEAD(&ptype_base[i]); 12646 12647 if (register_pernet_subsys(&netdev_net_ops)) 12648 goto out; 12649 12650 /* 12651 * Initialise the packet receive queues. 12652 */ 12653 12654 flush_backlogs_fallback = flush_backlogs_alloc(); 12655 if (!flush_backlogs_fallback) 12656 goto out; 12657 12658 for_each_possible_cpu(i) { 12659 struct softnet_data *sd = &per_cpu(softnet_data, i); 12660 12661 skb_queue_head_init(&sd->input_pkt_queue); 12662 skb_queue_head_init(&sd->process_queue); 12663 #ifdef CONFIG_XFRM_OFFLOAD 12664 skb_queue_head_init(&sd->xfrm_backlog); 12665 #endif 12666 INIT_LIST_HEAD(&sd->poll_list); 12667 sd->output_queue_tailp = &sd->output_queue; 12668 #ifdef CONFIG_RPS 12669 INIT_CSD(&sd->csd, rps_trigger_softirq, sd); 12670 sd->cpu = i; 12671 #endif 12672 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd); 12673 spin_lock_init(&sd->defer_lock); 12674 12675 gro_init(&sd->backlog.gro); 12676 sd->backlog.poll = process_backlog; 12677 sd->backlog.weight = weight_p; 12678 INIT_LIST_HEAD(&sd->backlog.poll_list); 12679 12680 if (net_page_pool_create(i)) 12681 goto out; 12682 } 12683 if (use_backlog_threads()) 12684 smpboot_register_percpu_thread(&backlog_threads); 12685 12686 dev_boot_phase = 0; 12687 12688 /* The loopback device is special if any other network devices 12689 * is present in a network namespace the loopback device must 12690 * be present. Since we now dynamically allocate and free the 12691 * loopback device ensure this invariant is maintained by 12692 * keeping the loopback device as the first device on the 12693 * list of network devices. Ensuring the loopback devices 12694 * is the first device that appears and the last network device 12695 * that disappears. 12696 */ 12697 if (register_pernet_device(&loopback_net_ops)) 12698 goto out; 12699 12700 if (register_pernet_device(&default_device_ops)) 12701 goto out; 12702 12703 open_softirq(NET_TX_SOFTIRQ, net_tx_action); 12704 open_softirq(NET_RX_SOFTIRQ, net_rx_action); 12705 12706 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead", 12707 NULL, dev_cpu_dead); 12708 WARN_ON(rc < 0); 12709 rc = 0; 12710 12711 /* avoid static key IPIs to isolated CPUs */ 12712 if (housekeeping_enabled(HK_TYPE_MISC)) 12713 net_enable_timestamp(); 12714 out: 12715 if (rc < 0) { 12716 for_each_possible_cpu(i) { 12717 struct page_pool *pp_ptr; 12718 12719 pp_ptr = per_cpu(system_page_pool, i); 12720 if (!pp_ptr) 12721 continue; 12722 12723 xdp_unreg_page_pool(pp_ptr); 12724 page_pool_destroy(pp_ptr); 12725 per_cpu(system_page_pool, i) = NULL; 12726 } 12727 } 12728 12729 return rc; 12730 } 12731 12732 subsys_initcall(net_dev_init); 12733