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