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_ALIGNED(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 (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 /* 3779 * Heuristic to force contended enqueues to serialize on a 3780 * separate lock before trying to get qdisc main lock. 3781 * This permits qdisc->running owner to get the lock more 3782 * often and dequeue packets faster. 3783 * On PREEMPT_RT it is possible to preempt the qdisc owner during xmit 3784 * and then other tasks will only enqueue packets. The packets will be 3785 * sent after the qdisc owner is scheduled again. To prevent this 3786 * scenario the task always serialize on the lock. 3787 */ 3788 contended = qdisc_is_running(q) || IS_ENABLED(CONFIG_PREEMPT_RT); 3789 if (unlikely(contended)) 3790 spin_lock(&q->busylock); 3791 3792 spin_lock(root_lock); 3793 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) { 3794 __qdisc_drop(skb, &to_free); 3795 rc = NET_XMIT_DROP; 3796 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && 3797 qdisc_run_begin(q)) { 3798 /* 3799 * This is a work-conserving queue; there are no old skbs 3800 * waiting to be sent out; and the qdisc is not running - 3801 * xmit the skb directly. 3802 */ 3803 3804 qdisc_bstats_update(q, skb); 3805 3806 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) { 3807 if (unlikely(contended)) { 3808 spin_unlock(&q->busylock); 3809 contended = false; 3810 } 3811 __qdisc_run(q); 3812 } 3813 3814 qdisc_run_end(q); 3815 rc = NET_XMIT_SUCCESS; 3816 } else { 3817 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 3818 if (qdisc_run_begin(q)) { 3819 if (unlikely(contended)) { 3820 spin_unlock(&q->busylock); 3821 contended = false; 3822 } 3823 __qdisc_run(q); 3824 qdisc_run_end(q); 3825 } 3826 } 3827 spin_unlock(root_lock); 3828 if (unlikely(to_free)) 3829 kfree_skb_list_reason(to_free, 3830 tcf_get_drop_reason(to_free)); 3831 if (unlikely(contended)) 3832 spin_unlock(&q->busylock); 3833 return rc; 3834 } 3835 3836 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO) 3837 static void skb_update_prio(struct sk_buff *skb) 3838 { 3839 const struct netprio_map *map; 3840 const struct sock *sk; 3841 unsigned int prioidx; 3842 3843 if (skb->priority) 3844 return; 3845 map = rcu_dereference_bh(skb->dev->priomap); 3846 if (!map) 3847 return; 3848 sk = skb_to_full_sk(skb); 3849 if (!sk) 3850 return; 3851 3852 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data); 3853 3854 if (prioidx < map->priomap_len) 3855 skb->priority = map->priomap[prioidx]; 3856 } 3857 #else 3858 #define skb_update_prio(skb) 3859 #endif 3860 3861 /** 3862 * dev_loopback_xmit - loop back @skb 3863 * @net: network namespace this loopback is happening in 3864 * @sk: sk needed to be a netfilter okfn 3865 * @skb: buffer to transmit 3866 */ 3867 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb) 3868 { 3869 skb_reset_mac_header(skb); 3870 __skb_pull(skb, skb_network_offset(skb)); 3871 skb->pkt_type = PACKET_LOOPBACK; 3872 if (skb->ip_summed == CHECKSUM_NONE) 3873 skb->ip_summed = CHECKSUM_UNNECESSARY; 3874 DEBUG_NET_WARN_ON_ONCE(!skb_dst(skb)); 3875 skb_dst_force(skb); 3876 netif_rx(skb); 3877 return 0; 3878 } 3879 EXPORT_SYMBOL(dev_loopback_xmit); 3880 3881 #ifdef CONFIG_NET_EGRESS 3882 static struct netdev_queue * 3883 netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb) 3884 { 3885 int qm = skb_get_queue_mapping(skb); 3886 3887 return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm)); 3888 } 3889 3890 static bool netdev_xmit_txqueue_skipped(void) 3891 { 3892 return __this_cpu_read(softnet_data.xmit.skip_txqueue); 3893 } 3894 3895 void netdev_xmit_skip_txqueue(bool skip) 3896 { 3897 __this_cpu_write(softnet_data.xmit.skip_txqueue, skip); 3898 } 3899 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 3900 #endif /* CONFIG_NET_EGRESS */ 3901 3902 #ifdef CONFIG_NET_XGRESS 3903 static int tc_run(struct tcx_entry *entry, struct sk_buff *skb, 3904 enum skb_drop_reason *drop_reason) 3905 { 3906 int ret = TC_ACT_UNSPEC; 3907 #ifdef CONFIG_NET_CLS_ACT 3908 struct mini_Qdisc *miniq = rcu_dereference_bh(entry->miniq); 3909 struct tcf_result res; 3910 3911 if (!miniq) 3912 return ret; 3913 3914 tc_skb_cb(skb)->mru = 0; 3915 tc_skb_cb(skb)->post_ct = false; 3916 tcf_set_drop_reason(skb, *drop_reason); 3917 3918 mini_qdisc_bstats_cpu_update(miniq, skb); 3919 ret = tcf_classify(skb, miniq->block, miniq->filter_list, &res, false); 3920 /* Only tcf related quirks below. */ 3921 switch (ret) { 3922 case TC_ACT_SHOT: 3923 *drop_reason = tcf_get_drop_reason(skb); 3924 mini_qdisc_qstats_cpu_drop(miniq); 3925 break; 3926 case TC_ACT_OK: 3927 case TC_ACT_RECLASSIFY: 3928 skb->tc_index = TC_H_MIN(res.classid); 3929 break; 3930 } 3931 #endif /* CONFIG_NET_CLS_ACT */ 3932 return ret; 3933 } 3934 3935 static DEFINE_STATIC_KEY_FALSE(tcx_needed_key); 3936 3937 void tcx_inc(void) 3938 { 3939 static_branch_inc(&tcx_needed_key); 3940 } 3941 3942 void tcx_dec(void) 3943 { 3944 static_branch_dec(&tcx_needed_key); 3945 } 3946 3947 static __always_inline enum tcx_action_base 3948 tcx_run(const struct bpf_mprog_entry *entry, struct sk_buff *skb, 3949 const bool needs_mac) 3950 { 3951 const struct bpf_mprog_fp *fp; 3952 const struct bpf_prog *prog; 3953 int ret = TCX_NEXT; 3954 3955 if (needs_mac) 3956 __skb_push(skb, skb->mac_len); 3957 bpf_mprog_foreach_prog(entry, fp, prog) { 3958 bpf_compute_data_pointers(skb); 3959 ret = bpf_prog_run(prog, skb); 3960 if (ret != TCX_NEXT) 3961 break; 3962 } 3963 if (needs_mac) 3964 __skb_pull(skb, skb->mac_len); 3965 return tcx_action_code(skb, ret); 3966 } 3967 3968 static __always_inline struct sk_buff * 3969 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 3970 struct net_device *orig_dev, bool *another) 3971 { 3972 struct bpf_mprog_entry *entry = rcu_dereference_bh(skb->dev->tcx_ingress); 3973 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_INGRESS; 3974 int sch_ret; 3975 3976 if (!entry) 3977 return skb; 3978 if (*pt_prev) { 3979 *ret = deliver_skb(skb, *pt_prev, orig_dev); 3980 *pt_prev = NULL; 3981 } 3982 3983 qdisc_skb_cb(skb)->pkt_len = skb->len; 3984 tcx_set_ingress(skb, true); 3985 3986 if (static_branch_unlikely(&tcx_needed_key)) { 3987 sch_ret = tcx_run(entry, skb, true); 3988 if (sch_ret != TC_ACT_UNSPEC) 3989 goto ingress_verdict; 3990 } 3991 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 3992 ingress_verdict: 3993 switch (sch_ret) { 3994 case TC_ACT_REDIRECT: 3995 /* skb_mac_header check was done by BPF, so we can safely 3996 * push the L2 header back before redirecting to another 3997 * netdev. 3998 */ 3999 __skb_push(skb, skb->mac_len); 4000 if (skb_do_redirect(skb) == -EAGAIN) { 4001 __skb_pull(skb, skb->mac_len); 4002 *another = true; 4003 break; 4004 } 4005 *ret = NET_RX_SUCCESS; 4006 return NULL; 4007 case TC_ACT_SHOT: 4008 kfree_skb_reason(skb, drop_reason); 4009 *ret = NET_RX_DROP; 4010 return NULL; 4011 /* used by tc_run */ 4012 case TC_ACT_STOLEN: 4013 case TC_ACT_QUEUED: 4014 case TC_ACT_TRAP: 4015 consume_skb(skb); 4016 fallthrough; 4017 case TC_ACT_CONSUMED: 4018 *ret = NET_RX_SUCCESS; 4019 return NULL; 4020 } 4021 4022 return skb; 4023 } 4024 4025 static __always_inline struct sk_buff * 4026 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4027 { 4028 struct bpf_mprog_entry *entry = rcu_dereference_bh(dev->tcx_egress); 4029 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_EGRESS; 4030 int sch_ret; 4031 4032 if (!entry) 4033 return skb; 4034 4035 /* qdisc_skb_cb(skb)->pkt_len & tcx_set_ingress() was 4036 * already set by the caller. 4037 */ 4038 if (static_branch_unlikely(&tcx_needed_key)) { 4039 sch_ret = tcx_run(entry, skb, false); 4040 if (sch_ret != TC_ACT_UNSPEC) 4041 goto egress_verdict; 4042 } 4043 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4044 egress_verdict: 4045 switch (sch_ret) { 4046 case TC_ACT_REDIRECT: 4047 /* No need to push/pop skb's mac_header here on egress! */ 4048 skb_do_redirect(skb); 4049 *ret = NET_XMIT_SUCCESS; 4050 return NULL; 4051 case TC_ACT_SHOT: 4052 kfree_skb_reason(skb, drop_reason); 4053 *ret = NET_XMIT_DROP; 4054 return NULL; 4055 /* used by tc_run */ 4056 case TC_ACT_STOLEN: 4057 case TC_ACT_QUEUED: 4058 case TC_ACT_TRAP: 4059 consume_skb(skb); 4060 fallthrough; 4061 case TC_ACT_CONSUMED: 4062 *ret = NET_XMIT_SUCCESS; 4063 return NULL; 4064 } 4065 4066 return skb; 4067 } 4068 #else 4069 static __always_inline struct sk_buff * 4070 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4071 struct net_device *orig_dev, bool *another) 4072 { 4073 return skb; 4074 } 4075 4076 static __always_inline struct sk_buff * 4077 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4078 { 4079 return skb; 4080 } 4081 #endif /* CONFIG_NET_XGRESS */ 4082 4083 #ifdef CONFIG_XPS 4084 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb, 4085 struct xps_dev_maps *dev_maps, unsigned int tci) 4086 { 4087 int tc = netdev_get_prio_tc_map(dev, skb->priority); 4088 struct xps_map *map; 4089 int queue_index = -1; 4090 4091 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids) 4092 return queue_index; 4093 4094 tci *= dev_maps->num_tc; 4095 tci += tc; 4096 4097 map = rcu_dereference(dev_maps->attr_map[tci]); 4098 if (map) { 4099 if (map->len == 1) 4100 queue_index = map->queues[0]; 4101 else 4102 queue_index = map->queues[reciprocal_scale( 4103 skb_get_hash(skb), map->len)]; 4104 if (unlikely(queue_index >= dev->real_num_tx_queues)) 4105 queue_index = -1; 4106 } 4107 return queue_index; 4108 } 4109 #endif 4110 4111 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev, 4112 struct sk_buff *skb) 4113 { 4114 #ifdef CONFIG_XPS 4115 struct xps_dev_maps *dev_maps; 4116 struct sock *sk = skb->sk; 4117 int queue_index = -1; 4118 4119 if (!static_key_false(&xps_needed)) 4120 return -1; 4121 4122 rcu_read_lock(); 4123 if (!static_key_false(&xps_rxqs_needed)) 4124 goto get_cpus_map; 4125 4126 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]); 4127 if (dev_maps) { 4128 int tci = sk_rx_queue_get(sk); 4129 4130 if (tci >= 0) 4131 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4132 tci); 4133 } 4134 4135 get_cpus_map: 4136 if (queue_index < 0) { 4137 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]); 4138 if (dev_maps) { 4139 unsigned int tci = skb->sender_cpu - 1; 4140 4141 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4142 tci); 4143 } 4144 } 4145 rcu_read_unlock(); 4146 4147 return queue_index; 4148 #else 4149 return -1; 4150 #endif 4151 } 4152 4153 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb, 4154 struct net_device *sb_dev) 4155 { 4156 return 0; 4157 } 4158 EXPORT_SYMBOL(dev_pick_tx_zero); 4159 4160 u16 dev_pick_tx_cpu_id(struct net_device *dev, struct sk_buff *skb, 4161 struct net_device *sb_dev) 4162 { 4163 return (u16)raw_smp_processor_id() % dev->real_num_tx_queues; 4164 } 4165 EXPORT_SYMBOL(dev_pick_tx_cpu_id); 4166 4167 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb, 4168 struct net_device *sb_dev) 4169 { 4170 struct sock *sk = skb->sk; 4171 int queue_index = sk_tx_queue_get(sk); 4172 4173 sb_dev = sb_dev ? : dev; 4174 4175 if (queue_index < 0 || skb->ooo_okay || 4176 queue_index >= dev->real_num_tx_queues) { 4177 int new_index = get_xps_queue(dev, sb_dev, skb); 4178 4179 if (new_index < 0) 4180 new_index = skb_tx_hash(dev, sb_dev, skb); 4181 4182 if (queue_index != new_index && sk && 4183 sk_fullsock(sk) && 4184 rcu_access_pointer(sk->sk_dst_cache)) 4185 sk_tx_queue_set(sk, new_index); 4186 4187 queue_index = new_index; 4188 } 4189 4190 return queue_index; 4191 } 4192 EXPORT_SYMBOL(netdev_pick_tx); 4193 4194 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev, 4195 struct sk_buff *skb, 4196 struct net_device *sb_dev) 4197 { 4198 int queue_index = 0; 4199 4200 #ifdef CONFIG_XPS 4201 u32 sender_cpu = skb->sender_cpu - 1; 4202 4203 if (sender_cpu >= (u32)NR_CPUS) 4204 skb->sender_cpu = raw_smp_processor_id() + 1; 4205 #endif 4206 4207 if (dev->real_num_tx_queues != 1) { 4208 const struct net_device_ops *ops = dev->netdev_ops; 4209 4210 if (ops->ndo_select_queue) 4211 queue_index = ops->ndo_select_queue(dev, skb, sb_dev); 4212 else 4213 queue_index = netdev_pick_tx(dev, skb, sb_dev); 4214 4215 queue_index = netdev_cap_txqueue(dev, queue_index); 4216 } 4217 4218 skb_set_queue_mapping(skb, queue_index); 4219 return netdev_get_tx_queue(dev, queue_index); 4220 } 4221 4222 /** 4223 * __dev_queue_xmit() - transmit a buffer 4224 * @skb: buffer to transmit 4225 * @sb_dev: suboordinate device used for L2 forwarding offload 4226 * 4227 * Queue a buffer for transmission to a network device. The caller must 4228 * have set the device and priority and built the buffer before calling 4229 * this function. The function can be called from an interrupt. 4230 * 4231 * When calling this method, interrupts MUST be enabled. This is because 4232 * the BH enable code must have IRQs enabled so that it will not deadlock. 4233 * 4234 * Regardless of the return value, the skb is consumed, so it is currently 4235 * difficult to retry a send to this method. (You can bump the ref count 4236 * before sending to hold a reference for retry if you are careful.) 4237 * 4238 * Return: 4239 * * 0 - buffer successfully transmitted 4240 * * positive qdisc return code - NET_XMIT_DROP etc. 4241 * * negative errno - other errors 4242 */ 4243 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev) 4244 { 4245 struct net_device *dev = skb->dev; 4246 struct netdev_queue *txq = NULL; 4247 struct Qdisc *q; 4248 int rc = -ENOMEM; 4249 bool again = false; 4250 4251 skb_reset_mac_header(skb); 4252 skb_assert_len(skb); 4253 4254 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP)) 4255 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED); 4256 4257 /* Disable soft irqs for various locks below. Also 4258 * stops preemption for RCU. 4259 */ 4260 rcu_read_lock_bh(); 4261 4262 skb_update_prio(skb); 4263 4264 qdisc_pkt_len_init(skb); 4265 tcx_set_ingress(skb, false); 4266 #ifdef CONFIG_NET_EGRESS 4267 if (static_branch_unlikely(&egress_needed_key)) { 4268 if (nf_hook_egress_active()) { 4269 skb = nf_hook_egress(skb, &rc, dev); 4270 if (!skb) 4271 goto out; 4272 } 4273 4274 netdev_xmit_skip_txqueue(false); 4275 4276 nf_skip_egress(skb, true); 4277 skb = sch_handle_egress(skb, &rc, dev); 4278 if (!skb) 4279 goto out; 4280 nf_skip_egress(skb, false); 4281 4282 if (netdev_xmit_txqueue_skipped()) 4283 txq = netdev_tx_queue_mapping(dev, skb); 4284 } 4285 #endif 4286 /* If device/qdisc don't need skb->dst, release it right now while 4287 * its hot in this cpu cache. 4288 */ 4289 if (dev->priv_flags & IFF_XMIT_DST_RELEASE) 4290 skb_dst_drop(skb); 4291 else 4292 skb_dst_force(skb); 4293 4294 if (!txq) 4295 txq = netdev_core_pick_tx(dev, skb, sb_dev); 4296 4297 q = rcu_dereference_bh(txq->qdisc); 4298 4299 trace_net_dev_queue(skb); 4300 if (q->enqueue) { 4301 rc = __dev_xmit_skb(skb, q, dev, txq); 4302 goto out; 4303 } 4304 4305 /* The device has no queue. Common case for software devices: 4306 * loopback, all the sorts of tunnels... 4307 4308 * Really, it is unlikely that netif_tx_lock protection is necessary 4309 * here. (f.e. loopback and IP tunnels are clean ignoring statistics 4310 * counters.) 4311 * However, it is possible, that they rely on protection 4312 * made by us here. 4313 4314 * Check this and shot the lock. It is not prone from deadlocks. 4315 *Either shot noqueue qdisc, it is even simpler 8) 4316 */ 4317 if (dev->flags & IFF_UP) { 4318 int cpu = smp_processor_id(); /* ok because BHs are off */ 4319 4320 /* Other cpus might concurrently change txq->xmit_lock_owner 4321 * to -1 or to their cpu id, but not to our id. 4322 */ 4323 if (READ_ONCE(txq->xmit_lock_owner) != cpu) { 4324 if (dev_xmit_recursion()) 4325 goto recursion_alert; 4326 4327 skb = validate_xmit_skb(skb, dev, &again); 4328 if (!skb) 4329 goto out; 4330 4331 HARD_TX_LOCK(dev, txq, cpu); 4332 4333 if (!netif_xmit_stopped(txq)) { 4334 dev_xmit_recursion_inc(); 4335 skb = dev_hard_start_xmit(skb, dev, txq, &rc); 4336 dev_xmit_recursion_dec(); 4337 if (dev_xmit_complete(rc)) { 4338 HARD_TX_UNLOCK(dev, txq); 4339 goto out; 4340 } 4341 } 4342 HARD_TX_UNLOCK(dev, txq); 4343 net_crit_ratelimited("Virtual device %s asks to queue packet!\n", 4344 dev->name); 4345 } else { 4346 /* Recursion is detected! It is possible, 4347 * unfortunately 4348 */ 4349 recursion_alert: 4350 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n", 4351 dev->name); 4352 } 4353 } 4354 4355 rc = -ENETDOWN; 4356 rcu_read_unlock_bh(); 4357 4358 dev_core_stats_tx_dropped_inc(dev); 4359 kfree_skb_list(skb); 4360 return rc; 4361 out: 4362 rcu_read_unlock_bh(); 4363 return rc; 4364 } 4365 EXPORT_SYMBOL(__dev_queue_xmit); 4366 4367 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id) 4368 { 4369 struct net_device *dev = skb->dev; 4370 struct sk_buff *orig_skb = skb; 4371 struct netdev_queue *txq; 4372 int ret = NETDEV_TX_BUSY; 4373 bool again = false; 4374 4375 if (unlikely(!netif_running(dev) || 4376 !netif_carrier_ok(dev))) 4377 goto drop; 4378 4379 skb = validate_xmit_skb_list(skb, dev, &again); 4380 if (skb != orig_skb) 4381 goto drop; 4382 4383 skb_set_queue_mapping(skb, queue_id); 4384 txq = skb_get_tx_queue(dev, skb); 4385 4386 local_bh_disable(); 4387 4388 dev_xmit_recursion_inc(); 4389 HARD_TX_LOCK(dev, txq, smp_processor_id()); 4390 if (!netif_xmit_frozen_or_drv_stopped(txq)) 4391 ret = netdev_start_xmit(skb, dev, txq, false); 4392 HARD_TX_UNLOCK(dev, txq); 4393 dev_xmit_recursion_dec(); 4394 4395 local_bh_enable(); 4396 return ret; 4397 drop: 4398 dev_core_stats_tx_dropped_inc(dev); 4399 kfree_skb_list(skb); 4400 return NET_XMIT_DROP; 4401 } 4402 EXPORT_SYMBOL(__dev_direct_xmit); 4403 4404 /************************************************************************* 4405 * Receiver routines 4406 *************************************************************************/ 4407 4408 unsigned int sysctl_skb_defer_max __read_mostly = 64; 4409 int weight_p __read_mostly = 64; /* old backlog weight */ 4410 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */ 4411 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */ 4412 4413 /* Called with irq disabled */ 4414 static inline void ____napi_schedule(struct softnet_data *sd, 4415 struct napi_struct *napi) 4416 { 4417 struct task_struct *thread; 4418 4419 lockdep_assert_irqs_disabled(); 4420 4421 if (test_bit(NAPI_STATE_THREADED, &napi->state)) { 4422 /* Paired with smp_mb__before_atomic() in 4423 * napi_enable()/dev_set_threaded(). 4424 * Use READ_ONCE() to guarantee a complete 4425 * read on napi->thread. Only call 4426 * wake_up_process() when it's not NULL. 4427 */ 4428 thread = READ_ONCE(napi->thread); 4429 if (thread) { 4430 /* Avoid doing set_bit() if the thread is in 4431 * INTERRUPTIBLE state, cause napi_thread_wait() 4432 * makes sure to proceed with napi polling 4433 * if the thread is explicitly woken from here. 4434 */ 4435 if (READ_ONCE(thread->__state) != TASK_INTERRUPTIBLE) 4436 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 4437 wake_up_process(thread); 4438 return; 4439 } 4440 } 4441 4442 list_add_tail(&napi->poll_list, &sd->poll_list); 4443 WRITE_ONCE(napi->list_owner, smp_processor_id()); 4444 /* If not called from net_rx_action() 4445 * we have to raise NET_RX_SOFTIRQ. 4446 */ 4447 if (!sd->in_net_rx_action) 4448 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4449 } 4450 4451 #ifdef CONFIG_RPS 4452 4453 struct static_key_false rps_needed __read_mostly; 4454 EXPORT_SYMBOL(rps_needed); 4455 struct static_key_false rfs_needed __read_mostly; 4456 EXPORT_SYMBOL(rfs_needed); 4457 4458 static struct rps_dev_flow * 4459 set_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4460 struct rps_dev_flow *rflow, u16 next_cpu) 4461 { 4462 if (next_cpu < nr_cpu_ids) { 4463 #ifdef CONFIG_RFS_ACCEL 4464 struct netdev_rx_queue *rxqueue; 4465 struct rps_dev_flow_table *flow_table; 4466 struct rps_dev_flow *old_rflow; 4467 u32 flow_id; 4468 u16 rxq_index; 4469 int rc; 4470 4471 /* Should we steer this flow to a different hardware queue? */ 4472 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap || 4473 !(dev->features & NETIF_F_NTUPLE)) 4474 goto out; 4475 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu); 4476 if (rxq_index == skb_get_rx_queue(skb)) 4477 goto out; 4478 4479 rxqueue = dev->_rx + rxq_index; 4480 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4481 if (!flow_table) 4482 goto out; 4483 flow_id = skb_get_hash(skb) & flow_table->mask; 4484 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb, 4485 rxq_index, flow_id); 4486 if (rc < 0) 4487 goto out; 4488 old_rflow = rflow; 4489 rflow = &flow_table->flows[flow_id]; 4490 rflow->filter = rc; 4491 if (old_rflow->filter == rflow->filter) 4492 old_rflow->filter = RPS_NO_FILTER; 4493 out: 4494 #endif 4495 rflow->last_qtail = 4496 per_cpu(softnet_data, next_cpu).input_queue_head; 4497 } 4498 4499 rflow->cpu = next_cpu; 4500 return rflow; 4501 } 4502 4503 /* 4504 * get_rps_cpu is called from netif_receive_skb and returns the target 4505 * CPU from the RPS map of the receiving queue for a given skb. 4506 * rcu_read_lock must be held on entry. 4507 */ 4508 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4509 struct rps_dev_flow **rflowp) 4510 { 4511 const struct rps_sock_flow_table *sock_flow_table; 4512 struct netdev_rx_queue *rxqueue = dev->_rx; 4513 struct rps_dev_flow_table *flow_table; 4514 struct rps_map *map; 4515 int cpu = -1; 4516 u32 tcpu; 4517 u32 hash; 4518 4519 if (skb_rx_queue_recorded(skb)) { 4520 u16 index = skb_get_rx_queue(skb); 4521 4522 if (unlikely(index >= dev->real_num_rx_queues)) { 4523 WARN_ONCE(dev->real_num_rx_queues > 1, 4524 "%s received packet on queue %u, but number " 4525 "of RX queues is %u\n", 4526 dev->name, index, dev->real_num_rx_queues); 4527 goto done; 4528 } 4529 rxqueue += index; 4530 } 4531 4532 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */ 4533 4534 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4535 map = rcu_dereference(rxqueue->rps_map); 4536 if (!flow_table && !map) 4537 goto done; 4538 4539 skb_reset_network_header(skb); 4540 hash = skb_get_hash(skb); 4541 if (!hash) 4542 goto done; 4543 4544 sock_flow_table = rcu_dereference(net_hotdata.rps_sock_flow_table); 4545 if (flow_table && sock_flow_table) { 4546 struct rps_dev_flow *rflow; 4547 u32 next_cpu; 4548 u32 ident; 4549 4550 /* First check into global flow table if there is a match. 4551 * This READ_ONCE() pairs with WRITE_ONCE() from rps_record_sock_flow(). 4552 */ 4553 ident = READ_ONCE(sock_flow_table->ents[hash & sock_flow_table->mask]); 4554 if ((ident ^ hash) & ~net_hotdata.rps_cpu_mask) 4555 goto try_rps; 4556 4557 next_cpu = ident & net_hotdata.rps_cpu_mask; 4558 4559 /* OK, now we know there is a match, 4560 * we can look at the local (per receive queue) flow table 4561 */ 4562 rflow = &flow_table->flows[hash & flow_table->mask]; 4563 tcpu = rflow->cpu; 4564 4565 /* 4566 * If the desired CPU (where last recvmsg was done) is 4567 * different from current CPU (one in the rx-queue flow 4568 * table entry), switch if one of the following holds: 4569 * - Current CPU is unset (>= nr_cpu_ids). 4570 * - Current CPU is offline. 4571 * - The current CPU's queue tail has advanced beyond the 4572 * last packet that was enqueued using this table entry. 4573 * This guarantees that all previous packets for the flow 4574 * have been dequeued, thus preserving in order delivery. 4575 */ 4576 if (unlikely(tcpu != next_cpu) && 4577 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) || 4578 ((int)(per_cpu(softnet_data, tcpu).input_queue_head - 4579 rflow->last_qtail)) >= 0)) { 4580 tcpu = next_cpu; 4581 rflow = set_rps_cpu(dev, skb, rflow, next_cpu); 4582 } 4583 4584 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) { 4585 *rflowp = rflow; 4586 cpu = tcpu; 4587 goto done; 4588 } 4589 } 4590 4591 try_rps: 4592 4593 if (map) { 4594 tcpu = map->cpus[reciprocal_scale(hash, map->len)]; 4595 if (cpu_online(tcpu)) { 4596 cpu = tcpu; 4597 goto done; 4598 } 4599 } 4600 4601 done: 4602 return cpu; 4603 } 4604 4605 #ifdef CONFIG_RFS_ACCEL 4606 4607 /** 4608 * rps_may_expire_flow - check whether an RFS hardware filter may be removed 4609 * @dev: Device on which the filter was set 4610 * @rxq_index: RX queue index 4611 * @flow_id: Flow ID passed to ndo_rx_flow_steer() 4612 * @filter_id: Filter ID returned by ndo_rx_flow_steer() 4613 * 4614 * Drivers that implement ndo_rx_flow_steer() should periodically call 4615 * this function for each installed filter and remove the filters for 4616 * which it returns %true. 4617 */ 4618 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index, 4619 u32 flow_id, u16 filter_id) 4620 { 4621 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index; 4622 struct rps_dev_flow_table *flow_table; 4623 struct rps_dev_flow *rflow; 4624 bool expire = true; 4625 unsigned int cpu; 4626 4627 rcu_read_lock(); 4628 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4629 if (flow_table && flow_id <= flow_table->mask) { 4630 rflow = &flow_table->flows[flow_id]; 4631 cpu = READ_ONCE(rflow->cpu); 4632 if (rflow->filter == filter_id && cpu < nr_cpu_ids && 4633 ((int)(per_cpu(softnet_data, cpu).input_queue_head - 4634 rflow->last_qtail) < 4635 (int)(10 * flow_table->mask))) 4636 expire = false; 4637 } 4638 rcu_read_unlock(); 4639 return expire; 4640 } 4641 EXPORT_SYMBOL(rps_may_expire_flow); 4642 4643 #endif /* CONFIG_RFS_ACCEL */ 4644 4645 /* Called from hardirq (IPI) context */ 4646 static void rps_trigger_softirq(void *data) 4647 { 4648 struct softnet_data *sd = data; 4649 4650 ____napi_schedule(sd, &sd->backlog); 4651 sd->received_rps++; 4652 } 4653 4654 #endif /* CONFIG_RPS */ 4655 4656 /* Called from hardirq (IPI) context */ 4657 static void trigger_rx_softirq(void *data) 4658 { 4659 struct softnet_data *sd = data; 4660 4661 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4662 smp_store_release(&sd->defer_ipi_scheduled, 0); 4663 } 4664 4665 /* 4666 * After we queued a packet into sd->input_pkt_queue, 4667 * we need to make sure this queue is serviced soon. 4668 * 4669 * - If this is another cpu queue, link it to our rps_ipi_list, 4670 * and make sure we will process rps_ipi_list from net_rx_action(). 4671 * 4672 * - If this is our own queue, NAPI schedule our backlog. 4673 * Note that this also raises NET_RX_SOFTIRQ. 4674 */ 4675 static void napi_schedule_rps(struct softnet_data *sd) 4676 { 4677 struct softnet_data *mysd = this_cpu_ptr(&softnet_data); 4678 4679 #ifdef CONFIG_RPS 4680 if (sd != mysd) { 4681 sd->rps_ipi_next = mysd->rps_ipi_list; 4682 mysd->rps_ipi_list = sd; 4683 4684 /* If not called from net_rx_action() or napi_threaded_poll() 4685 * we have to raise NET_RX_SOFTIRQ. 4686 */ 4687 if (!mysd->in_net_rx_action && !mysd->in_napi_threaded_poll) 4688 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4689 return; 4690 } 4691 #endif /* CONFIG_RPS */ 4692 __napi_schedule_irqoff(&mysd->backlog); 4693 } 4694 4695 #ifdef CONFIG_NET_FLOW_LIMIT 4696 int netdev_flow_limit_table_len __read_mostly = (1 << 12); 4697 #endif 4698 4699 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen) 4700 { 4701 #ifdef CONFIG_NET_FLOW_LIMIT 4702 struct sd_flow_limit *fl; 4703 struct softnet_data *sd; 4704 unsigned int old_flow, new_flow; 4705 4706 if (qlen < (READ_ONCE(net_hotdata.max_backlog) >> 1)) 4707 return false; 4708 4709 sd = this_cpu_ptr(&softnet_data); 4710 4711 rcu_read_lock(); 4712 fl = rcu_dereference(sd->flow_limit); 4713 if (fl) { 4714 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1); 4715 old_flow = fl->history[fl->history_head]; 4716 fl->history[fl->history_head] = new_flow; 4717 4718 fl->history_head++; 4719 fl->history_head &= FLOW_LIMIT_HISTORY - 1; 4720 4721 if (likely(fl->buckets[old_flow])) 4722 fl->buckets[old_flow]--; 4723 4724 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) { 4725 fl->count++; 4726 rcu_read_unlock(); 4727 return true; 4728 } 4729 } 4730 rcu_read_unlock(); 4731 #endif 4732 return false; 4733 } 4734 4735 /* 4736 * enqueue_to_backlog is called to queue an skb to a per CPU backlog 4737 * queue (may be a remote CPU queue). 4738 */ 4739 static int enqueue_to_backlog(struct sk_buff *skb, int cpu, 4740 unsigned int *qtail) 4741 { 4742 enum skb_drop_reason reason; 4743 struct softnet_data *sd; 4744 unsigned long flags; 4745 unsigned int qlen; 4746 4747 reason = SKB_DROP_REASON_NOT_SPECIFIED; 4748 sd = &per_cpu(softnet_data, cpu); 4749 4750 rps_lock_irqsave(sd, &flags); 4751 if (!netif_running(skb->dev)) 4752 goto drop; 4753 qlen = skb_queue_len(&sd->input_pkt_queue); 4754 if (qlen <= READ_ONCE(net_hotdata.max_backlog) && 4755 !skb_flow_limit(skb, qlen)) { 4756 if (qlen) { 4757 enqueue: 4758 __skb_queue_tail(&sd->input_pkt_queue, skb); 4759 input_queue_tail_incr_save(sd, qtail); 4760 rps_unlock_irq_restore(sd, &flags); 4761 return NET_RX_SUCCESS; 4762 } 4763 4764 /* Schedule NAPI for backlog device 4765 * We can use non atomic operation since we own the queue lock 4766 */ 4767 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) 4768 napi_schedule_rps(sd); 4769 goto enqueue; 4770 } 4771 reason = SKB_DROP_REASON_CPU_BACKLOG; 4772 4773 drop: 4774 sd->dropped++; 4775 rps_unlock_irq_restore(sd, &flags); 4776 4777 dev_core_stats_rx_dropped_inc(skb->dev); 4778 kfree_skb_reason(skb, reason); 4779 return NET_RX_DROP; 4780 } 4781 4782 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb) 4783 { 4784 struct net_device *dev = skb->dev; 4785 struct netdev_rx_queue *rxqueue; 4786 4787 rxqueue = dev->_rx; 4788 4789 if (skb_rx_queue_recorded(skb)) { 4790 u16 index = skb_get_rx_queue(skb); 4791 4792 if (unlikely(index >= dev->real_num_rx_queues)) { 4793 WARN_ONCE(dev->real_num_rx_queues > 1, 4794 "%s received packet on queue %u, but number " 4795 "of RX queues is %u\n", 4796 dev->name, index, dev->real_num_rx_queues); 4797 4798 return rxqueue; /* Return first rxqueue */ 4799 } 4800 rxqueue += index; 4801 } 4802 return rxqueue; 4803 } 4804 4805 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, 4806 struct bpf_prog *xdp_prog) 4807 { 4808 void *orig_data, *orig_data_end, *hard_start; 4809 struct netdev_rx_queue *rxqueue; 4810 bool orig_bcast, orig_host; 4811 u32 mac_len, frame_sz; 4812 __be16 orig_eth_type; 4813 struct ethhdr *eth; 4814 u32 metalen, act; 4815 int off; 4816 4817 /* The XDP program wants to see the packet starting at the MAC 4818 * header. 4819 */ 4820 mac_len = skb->data - skb_mac_header(skb); 4821 hard_start = skb->data - skb_headroom(skb); 4822 4823 /* SKB "head" area always have tailroom for skb_shared_info */ 4824 frame_sz = (void *)skb_end_pointer(skb) - hard_start; 4825 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 4826 4827 rxqueue = netif_get_rxqueue(skb); 4828 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq); 4829 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len, 4830 skb_headlen(skb) + mac_len, true); 4831 if (skb_is_nonlinear(skb)) { 4832 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 4833 xdp_buff_set_frags_flag(xdp); 4834 } else { 4835 xdp_buff_clear_frags_flag(xdp); 4836 } 4837 4838 orig_data_end = xdp->data_end; 4839 orig_data = xdp->data; 4840 eth = (struct ethhdr *)xdp->data; 4841 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr); 4842 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest); 4843 orig_eth_type = eth->h_proto; 4844 4845 act = bpf_prog_run_xdp(xdp_prog, xdp); 4846 4847 /* check if bpf_xdp_adjust_head was used */ 4848 off = xdp->data - orig_data; 4849 if (off) { 4850 if (off > 0) 4851 __skb_pull(skb, off); 4852 else if (off < 0) 4853 __skb_push(skb, -off); 4854 4855 skb->mac_header += off; 4856 skb_reset_network_header(skb); 4857 } 4858 4859 /* check if bpf_xdp_adjust_tail was used */ 4860 off = xdp->data_end - orig_data_end; 4861 if (off != 0) { 4862 skb_set_tail_pointer(skb, xdp->data_end - xdp->data); 4863 skb->len += off; /* positive on grow, negative on shrink */ 4864 } 4865 4866 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 4867 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here. 4868 */ 4869 if (xdp_buff_has_frags(xdp)) 4870 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 4871 else 4872 skb->data_len = 0; 4873 4874 /* check if XDP changed eth hdr such SKB needs update */ 4875 eth = (struct ethhdr *)xdp->data; 4876 if ((orig_eth_type != eth->h_proto) || 4877 (orig_host != ether_addr_equal_64bits(eth->h_dest, 4878 skb->dev->dev_addr)) || 4879 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) { 4880 __skb_push(skb, ETH_HLEN); 4881 skb->pkt_type = PACKET_HOST; 4882 skb->protocol = eth_type_trans(skb, skb->dev); 4883 } 4884 4885 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull 4886 * before calling us again on redirect path. We do not call do_redirect 4887 * as we leave that up to the caller. 4888 * 4889 * Caller is responsible for managing lifetime of skb (i.e. calling 4890 * kfree_skb in response to actions it cannot handle/XDP_DROP). 4891 */ 4892 switch (act) { 4893 case XDP_REDIRECT: 4894 case XDP_TX: 4895 __skb_push(skb, mac_len); 4896 break; 4897 case XDP_PASS: 4898 metalen = xdp->data - xdp->data_meta; 4899 if (metalen) 4900 skb_metadata_set(skb, metalen); 4901 break; 4902 } 4903 4904 return act; 4905 } 4906 4907 static int 4908 netif_skb_check_for_xdp(struct sk_buff **pskb, struct bpf_prog *prog) 4909 { 4910 struct sk_buff *skb = *pskb; 4911 int err, hroom, troom; 4912 4913 if (!skb_cow_data_for_xdp(this_cpu_read(system_page_pool), pskb, prog)) 4914 return 0; 4915 4916 /* In case we have to go down the path and also linearize, 4917 * then lets do the pskb_expand_head() work just once here. 4918 */ 4919 hroom = XDP_PACKET_HEADROOM - skb_headroom(skb); 4920 troom = skb->tail + skb->data_len - skb->end; 4921 err = pskb_expand_head(skb, 4922 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0, 4923 troom > 0 ? troom + 128 : 0, GFP_ATOMIC); 4924 if (err) 4925 return err; 4926 4927 return skb_linearize(skb); 4928 } 4929 4930 static u32 netif_receive_generic_xdp(struct sk_buff **pskb, 4931 struct xdp_buff *xdp, 4932 struct bpf_prog *xdp_prog) 4933 { 4934 struct sk_buff *skb = *pskb; 4935 u32 mac_len, act = XDP_DROP; 4936 4937 /* Reinjected packets coming from act_mirred or similar should 4938 * not get XDP generic processing. 4939 */ 4940 if (skb_is_redirected(skb)) 4941 return XDP_PASS; 4942 4943 /* XDP packets must have sufficient headroom of XDP_PACKET_HEADROOM 4944 * bytes. This is the guarantee that also native XDP provides, 4945 * thus we need to do it here as well. 4946 */ 4947 mac_len = skb->data - skb_mac_header(skb); 4948 __skb_push(skb, mac_len); 4949 4950 if (skb_cloned(skb) || skb_is_nonlinear(skb) || 4951 skb_headroom(skb) < XDP_PACKET_HEADROOM) { 4952 if (netif_skb_check_for_xdp(pskb, xdp_prog)) 4953 goto do_drop; 4954 } 4955 4956 __skb_pull(*pskb, mac_len); 4957 4958 act = bpf_prog_run_generic_xdp(*pskb, xdp, xdp_prog); 4959 switch (act) { 4960 case XDP_REDIRECT: 4961 case XDP_TX: 4962 case XDP_PASS: 4963 break; 4964 default: 4965 bpf_warn_invalid_xdp_action((*pskb)->dev, xdp_prog, act); 4966 fallthrough; 4967 case XDP_ABORTED: 4968 trace_xdp_exception((*pskb)->dev, xdp_prog, act); 4969 fallthrough; 4970 case XDP_DROP: 4971 do_drop: 4972 kfree_skb(*pskb); 4973 break; 4974 } 4975 4976 return act; 4977 } 4978 4979 /* When doing generic XDP we have to bypass the qdisc layer and the 4980 * network taps in order to match in-driver-XDP behavior. This also means 4981 * that XDP packets are able to starve other packets going through a qdisc, 4982 * and DDOS attacks will be more effective. In-driver-XDP use dedicated TX 4983 * queues, so they do not have this starvation issue. 4984 */ 4985 void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog) 4986 { 4987 struct net_device *dev = skb->dev; 4988 struct netdev_queue *txq; 4989 bool free_skb = true; 4990 int cpu, rc; 4991 4992 txq = netdev_core_pick_tx(dev, skb, NULL); 4993 cpu = smp_processor_id(); 4994 HARD_TX_LOCK(dev, txq, cpu); 4995 if (!netif_xmit_frozen_or_drv_stopped(txq)) { 4996 rc = netdev_start_xmit(skb, dev, txq, 0); 4997 if (dev_xmit_complete(rc)) 4998 free_skb = false; 4999 } 5000 HARD_TX_UNLOCK(dev, txq); 5001 if (free_skb) { 5002 trace_xdp_exception(dev, xdp_prog, XDP_TX); 5003 dev_core_stats_tx_dropped_inc(dev); 5004 kfree_skb(skb); 5005 } 5006 } 5007 5008 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key); 5009 5010 int do_xdp_generic(struct bpf_prog *xdp_prog, struct sk_buff **pskb) 5011 { 5012 if (xdp_prog) { 5013 struct xdp_buff xdp; 5014 u32 act; 5015 int err; 5016 5017 act = netif_receive_generic_xdp(pskb, &xdp, xdp_prog); 5018 if (act != XDP_PASS) { 5019 switch (act) { 5020 case XDP_REDIRECT: 5021 err = xdp_do_generic_redirect((*pskb)->dev, *pskb, 5022 &xdp, xdp_prog); 5023 if (err) 5024 goto out_redir; 5025 break; 5026 case XDP_TX: 5027 generic_xdp_tx(*pskb, xdp_prog); 5028 break; 5029 } 5030 return XDP_DROP; 5031 } 5032 } 5033 return XDP_PASS; 5034 out_redir: 5035 kfree_skb_reason(*pskb, SKB_DROP_REASON_XDP); 5036 return XDP_DROP; 5037 } 5038 EXPORT_SYMBOL_GPL(do_xdp_generic); 5039 5040 static int netif_rx_internal(struct sk_buff *skb) 5041 { 5042 int ret; 5043 5044 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5045 5046 trace_netif_rx(skb); 5047 5048 #ifdef CONFIG_RPS 5049 if (static_branch_unlikely(&rps_needed)) { 5050 struct rps_dev_flow voidflow, *rflow = &voidflow; 5051 int cpu; 5052 5053 rcu_read_lock(); 5054 5055 cpu = get_rps_cpu(skb->dev, skb, &rflow); 5056 if (cpu < 0) 5057 cpu = smp_processor_id(); 5058 5059 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5060 5061 rcu_read_unlock(); 5062 } else 5063 #endif 5064 { 5065 unsigned int qtail; 5066 5067 ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail); 5068 } 5069 return ret; 5070 } 5071 5072 /** 5073 * __netif_rx - Slightly optimized version of netif_rx 5074 * @skb: buffer to post 5075 * 5076 * This behaves as netif_rx except that it does not disable bottom halves. 5077 * As a result this function may only be invoked from the interrupt context 5078 * (either hard or soft interrupt). 5079 */ 5080 int __netif_rx(struct sk_buff *skb) 5081 { 5082 int ret; 5083 5084 lockdep_assert_once(hardirq_count() | softirq_count()); 5085 5086 trace_netif_rx_entry(skb); 5087 ret = netif_rx_internal(skb); 5088 trace_netif_rx_exit(ret); 5089 return ret; 5090 } 5091 EXPORT_SYMBOL(__netif_rx); 5092 5093 /** 5094 * netif_rx - post buffer to the network code 5095 * @skb: buffer to post 5096 * 5097 * This function receives a packet from a device driver and queues it for 5098 * the upper (protocol) levels to process via the backlog NAPI device. It 5099 * always succeeds. The buffer may be dropped during processing for 5100 * congestion control or by the protocol layers. 5101 * The network buffer is passed via the backlog NAPI device. Modern NIC 5102 * driver should use NAPI and GRO. 5103 * This function can used from interrupt and from process context. The 5104 * caller from process context must not disable interrupts before invoking 5105 * this function. 5106 * 5107 * return values: 5108 * NET_RX_SUCCESS (no congestion) 5109 * NET_RX_DROP (packet was dropped) 5110 * 5111 */ 5112 int netif_rx(struct sk_buff *skb) 5113 { 5114 bool need_bh_off = !(hardirq_count() | softirq_count()); 5115 int ret; 5116 5117 if (need_bh_off) 5118 local_bh_disable(); 5119 trace_netif_rx_entry(skb); 5120 ret = netif_rx_internal(skb); 5121 trace_netif_rx_exit(ret); 5122 if (need_bh_off) 5123 local_bh_enable(); 5124 return ret; 5125 } 5126 EXPORT_SYMBOL(netif_rx); 5127 5128 static __latent_entropy void net_tx_action(struct softirq_action *h) 5129 { 5130 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 5131 5132 if (sd->completion_queue) { 5133 struct sk_buff *clist; 5134 5135 local_irq_disable(); 5136 clist = sd->completion_queue; 5137 sd->completion_queue = NULL; 5138 local_irq_enable(); 5139 5140 while (clist) { 5141 struct sk_buff *skb = clist; 5142 5143 clist = clist->next; 5144 5145 WARN_ON(refcount_read(&skb->users)); 5146 if (likely(get_kfree_skb_cb(skb)->reason == SKB_CONSUMED)) 5147 trace_consume_skb(skb, net_tx_action); 5148 else 5149 trace_kfree_skb(skb, net_tx_action, 5150 get_kfree_skb_cb(skb)->reason); 5151 5152 if (skb->fclone != SKB_FCLONE_UNAVAILABLE) 5153 __kfree_skb(skb); 5154 else 5155 __napi_kfree_skb(skb, 5156 get_kfree_skb_cb(skb)->reason); 5157 } 5158 } 5159 5160 if (sd->output_queue) { 5161 struct Qdisc *head; 5162 5163 local_irq_disable(); 5164 head = sd->output_queue; 5165 sd->output_queue = NULL; 5166 sd->output_queue_tailp = &sd->output_queue; 5167 local_irq_enable(); 5168 5169 rcu_read_lock(); 5170 5171 while (head) { 5172 struct Qdisc *q = head; 5173 spinlock_t *root_lock = NULL; 5174 5175 head = head->next_sched; 5176 5177 /* We need to make sure head->next_sched is read 5178 * before clearing __QDISC_STATE_SCHED 5179 */ 5180 smp_mb__before_atomic(); 5181 5182 if (!(q->flags & TCQ_F_NOLOCK)) { 5183 root_lock = qdisc_lock(q); 5184 spin_lock(root_lock); 5185 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, 5186 &q->state))) { 5187 /* There is a synchronize_net() between 5188 * STATE_DEACTIVATED flag being set and 5189 * qdisc_reset()/some_qdisc_is_busy() in 5190 * dev_deactivate(), so we can safely bail out 5191 * early here to avoid data race between 5192 * qdisc_deactivate() and some_qdisc_is_busy() 5193 * for lockless qdisc. 5194 */ 5195 clear_bit(__QDISC_STATE_SCHED, &q->state); 5196 continue; 5197 } 5198 5199 clear_bit(__QDISC_STATE_SCHED, &q->state); 5200 qdisc_run(q); 5201 if (root_lock) 5202 spin_unlock(root_lock); 5203 } 5204 5205 rcu_read_unlock(); 5206 } 5207 5208 xfrm_dev_backlog(sd); 5209 } 5210 5211 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE) 5212 /* This hook is defined here for ATM LANE */ 5213 int (*br_fdb_test_addr_hook)(struct net_device *dev, 5214 unsigned char *addr) __read_mostly; 5215 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook); 5216 #endif 5217 5218 /** 5219 * netdev_is_rx_handler_busy - check if receive handler is registered 5220 * @dev: device to check 5221 * 5222 * Check if a receive handler is already registered for a given device. 5223 * Return true if there one. 5224 * 5225 * The caller must hold the rtnl_mutex. 5226 */ 5227 bool netdev_is_rx_handler_busy(struct net_device *dev) 5228 { 5229 ASSERT_RTNL(); 5230 return dev && rtnl_dereference(dev->rx_handler); 5231 } 5232 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy); 5233 5234 /** 5235 * netdev_rx_handler_register - register receive handler 5236 * @dev: device to register a handler for 5237 * @rx_handler: receive handler to register 5238 * @rx_handler_data: data pointer that is used by rx handler 5239 * 5240 * Register a receive handler for a device. This handler will then be 5241 * called from __netif_receive_skb. A negative errno code is returned 5242 * on a failure. 5243 * 5244 * The caller must hold the rtnl_mutex. 5245 * 5246 * For a general description of rx_handler, see enum rx_handler_result. 5247 */ 5248 int netdev_rx_handler_register(struct net_device *dev, 5249 rx_handler_func_t *rx_handler, 5250 void *rx_handler_data) 5251 { 5252 if (netdev_is_rx_handler_busy(dev)) 5253 return -EBUSY; 5254 5255 if (dev->priv_flags & IFF_NO_RX_HANDLER) 5256 return -EINVAL; 5257 5258 /* Note: rx_handler_data must be set before rx_handler */ 5259 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data); 5260 rcu_assign_pointer(dev->rx_handler, rx_handler); 5261 5262 return 0; 5263 } 5264 EXPORT_SYMBOL_GPL(netdev_rx_handler_register); 5265 5266 /** 5267 * netdev_rx_handler_unregister - unregister receive handler 5268 * @dev: device to unregister a handler from 5269 * 5270 * Unregister a receive handler from a device. 5271 * 5272 * The caller must hold the rtnl_mutex. 5273 */ 5274 void netdev_rx_handler_unregister(struct net_device *dev) 5275 { 5276 5277 ASSERT_RTNL(); 5278 RCU_INIT_POINTER(dev->rx_handler, NULL); 5279 /* a reader seeing a non NULL rx_handler in a rcu_read_lock() 5280 * section has a guarantee to see a non NULL rx_handler_data 5281 * as well. 5282 */ 5283 synchronize_net(); 5284 RCU_INIT_POINTER(dev->rx_handler_data, NULL); 5285 } 5286 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister); 5287 5288 /* 5289 * Limit the use of PFMEMALLOC reserves to those protocols that implement 5290 * the special handling of PFMEMALLOC skbs. 5291 */ 5292 static bool skb_pfmemalloc_protocol(struct sk_buff *skb) 5293 { 5294 switch (skb->protocol) { 5295 case htons(ETH_P_ARP): 5296 case htons(ETH_P_IP): 5297 case htons(ETH_P_IPV6): 5298 case htons(ETH_P_8021Q): 5299 case htons(ETH_P_8021AD): 5300 return true; 5301 default: 5302 return false; 5303 } 5304 } 5305 5306 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev, 5307 int *ret, struct net_device *orig_dev) 5308 { 5309 if (nf_hook_ingress_active(skb)) { 5310 int ingress_retval; 5311 5312 if (*pt_prev) { 5313 *ret = deliver_skb(skb, *pt_prev, orig_dev); 5314 *pt_prev = NULL; 5315 } 5316 5317 rcu_read_lock(); 5318 ingress_retval = nf_hook_ingress(skb); 5319 rcu_read_unlock(); 5320 return ingress_retval; 5321 } 5322 return 0; 5323 } 5324 5325 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc, 5326 struct packet_type **ppt_prev) 5327 { 5328 struct packet_type *ptype, *pt_prev; 5329 rx_handler_func_t *rx_handler; 5330 struct sk_buff *skb = *pskb; 5331 struct net_device *orig_dev; 5332 bool deliver_exact = false; 5333 int ret = NET_RX_DROP; 5334 __be16 type; 5335 5336 net_timestamp_check(!READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5337 5338 trace_netif_receive_skb(skb); 5339 5340 orig_dev = skb->dev; 5341 5342 skb_reset_network_header(skb); 5343 if (!skb_transport_header_was_set(skb)) 5344 skb_reset_transport_header(skb); 5345 skb_reset_mac_len(skb); 5346 5347 pt_prev = NULL; 5348 5349 another_round: 5350 skb->skb_iif = skb->dev->ifindex; 5351 5352 __this_cpu_inc(softnet_data.processed); 5353 5354 if (static_branch_unlikely(&generic_xdp_needed_key)) { 5355 int ret2; 5356 5357 migrate_disable(); 5358 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), 5359 &skb); 5360 migrate_enable(); 5361 5362 if (ret2 != XDP_PASS) { 5363 ret = NET_RX_DROP; 5364 goto out; 5365 } 5366 } 5367 5368 if (eth_type_vlan(skb->protocol)) { 5369 skb = skb_vlan_untag(skb); 5370 if (unlikely(!skb)) 5371 goto out; 5372 } 5373 5374 if (skb_skip_tc_classify(skb)) 5375 goto skip_classify; 5376 5377 if (pfmemalloc) 5378 goto skip_taps; 5379 5380 list_for_each_entry_rcu(ptype, &net_hotdata.ptype_all, list) { 5381 if (pt_prev) 5382 ret = deliver_skb(skb, pt_prev, orig_dev); 5383 pt_prev = ptype; 5384 } 5385 5386 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) { 5387 if (pt_prev) 5388 ret = deliver_skb(skb, pt_prev, orig_dev); 5389 pt_prev = ptype; 5390 } 5391 5392 skip_taps: 5393 #ifdef CONFIG_NET_INGRESS 5394 if (static_branch_unlikely(&ingress_needed_key)) { 5395 bool another = false; 5396 5397 nf_skip_egress(skb, true); 5398 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev, 5399 &another); 5400 if (another) 5401 goto another_round; 5402 if (!skb) 5403 goto out; 5404 5405 nf_skip_egress(skb, false); 5406 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0) 5407 goto out; 5408 } 5409 #endif 5410 skb_reset_redirect(skb); 5411 skip_classify: 5412 if (pfmemalloc && !skb_pfmemalloc_protocol(skb)) 5413 goto drop; 5414 5415 if (skb_vlan_tag_present(skb)) { 5416 if (pt_prev) { 5417 ret = deliver_skb(skb, pt_prev, orig_dev); 5418 pt_prev = NULL; 5419 } 5420 if (vlan_do_receive(&skb)) 5421 goto another_round; 5422 else if (unlikely(!skb)) 5423 goto out; 5424 } 5425 5426 rx_handler = rcu_dereference(skb->dev->rx_handler); 5427 if (rx_handler) { 5428 if (pt_prev) { 5429 ret = deliver_skb(skb, pt_prev, orig_dev); 5430 pt_prev = NULL; 5431 } 5432 switch (rx_handler(&skb)) { 5433 case RX_HANDLER_CONSUMED: 5434 ret = NET_RX_SUCCESS; 5435 goto out; 5436 case RX_HANDLER_ANOTHER: 5437 goto another_round; 5438 case RX_HANDLER_EXACT: 5439 deliver_exact = true; 5440 break; 5441 case RX_HANDLER_PASS: 5442 break; 5443 default: 5444 BUG(); 5445 } 5446 } 5447 5448 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) { 5449 check_vlan_id: 5450 if (skb_vlan_tag_get_id(skb)) { 5451 /* Vlan id is non 0 and vlan_do_receive() above couldn't 5452 * find vlan device. 5453 */ 5454 skb->pkt_type = PACKET_OTHERHOST; 5455 } else if (eth_type_vlan(skb->protocol)) { 5456 /* Outer header is 802.1P with vlan 0, inner header is 5457 * 802.1Q or 802.1AD and vlan_do_receive() above could 5458 * not find vlan dev for vlan id 0. 5459 */ 5460 __vlan_hwaccel_clear_tag(skb); 5461 skb = skb_vlan_untag(skb); 5462 if (unlikely(!skb)) 5463 goto out; 5464 if (vlan_do_receive(&skb)) 5465 /* After stripping off 802.1P header with vlan 0 5466 * vlan dev is found for inner header. 5467 */ 5468 goto another_round; 5469 else if (unlikely(!skb)) 5470 goto out; 5471 else 5472 /* We have stripped outer 802.1P vlan 0 header. 5473 * But could not find vlan dev. 5474 * check again for vlan id to set OTHERHOST. 5475 */ 5476 goto check_vlan_id; 5477 } 5478 /* Note: we might in the future use prio bits 5479 * and set skb->priority like in vlan_do_receive() 5480 * For the time being, just ignore Priority Code Point 5481 */ 5482 __vlan_hwaccel_clear_tag(skb); 5483 } 5484 5485 type = skb->protocol; 5486 5487 /* deliver only exact match when indicated */ 5488 if (likely(!deliver_exact)) { 5489 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5490 &ptype_base[ntohs(type) & 5491 PTYPE_HASH_MASK]); 5492 } 5493 5494 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5495 &orig_dev->ptype_specific); 5496 5497 if (unlikely(skb->dev != orig_dev)) { 5498 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5499 &skb->dev->ptype_specific); 5500 } 5501 5502 if (pt_prev) { 5503 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 5504 goto drop; 5505 *ppt_prev = pt_prev; 5506 } else { 5507 drop: 5508 if (!deliver_exact) 5509 dev_core_stats_rx_dropped_inc(skb->dev); 5510 else 5511 dev_core_stats_rx_nohandler_inc(skb->dev); 5512 kfree_skb_reason(skb, SKB_DROP_REASON_UNHANDLED_PROTO); 5513 /* Jamal, now you will not able to escape explaining 5514 * me how you were going to use this. :-) 5515 */ 5516 ret = NET_RX_DROP; 5517 } 5518 5519 out: 5520 /* The invariant here is that if *ppt_prev is not NULL 5521 * then skb should also be non-NULL. 5522 * 5523 * Apparently *ppt_prev assignment above holds this invariant due to 5524 * skb dereferencing near it. 5525 */ 5526 *pskb = skb; 5527 return ret; 5528 } 5529 5530 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc) 5531 { 5532 struct net_device *orig_dev = skb->dev; 5533 struct packet_type *pt_prev = NULL; 5534 int ret; 5535 5536 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5537 if (pt_prev) 5538 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb, 5539 skb->dev, pt_prev, orig_dev); 5540 return ret; 5541 } 5542 5543 /** 5544 * netif_receive_skb_core - special purpose version of netif_receive_skb 5545 * @skb: buffer to process 5546 * 5547 * More direct receive version of netif_receive_skb(). It should 5548 * only be used by callers that have a need to skip RPS and Generic XDP. 5549 * Caller must also take care of handling if ``(page_is_)pfmemalloc``. 5550 * 5551 * This function may only be called from softirq context and interrupts 5552 * should be enabled. 5553 * 5554 * Return values (usually ignored): 5555 * NET_RX_SUCCESS: no congestion 5556 * NET_RX_DROP: packet was dropped 5557 */ 5558 int netif_receive_skb_core(struct sk_buff *skb) 5559 { 5560 int ret; 5561 5562 rcu_read_lock(); 5563 ret = __netif_receive_skb_one_core(skb, false); 5564 rcu_read_unlock(); 5565 5566 return ret; 5567 } 5568 EXPORT_SYMBOL(netif_receive_skb_core); 5569 5570 static inline void __netif_receive_skb_list_ptype(struct list_head *head, 5571 struct packet_type *pt_prev, 5572 struct net_device *orig_dev) 5573 { 5574 struct sk_buff *skb, *next; 5575 5576 if (!pt_prev) 5577 return; 5578 if (list_empty(head)) 5579 return; 5580 if (pt_prev->list_func != NULL) 5581 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv, 5582 ip_list_rcv, head, pt_prev, orig_dev); 5583 else 5584 list_for_each_entry_safe(skb, next, head, list) { 5585 skb_list_del_init(skb); 5586 pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 5587 } 5588 } 5589 5590 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc) 5591 { 5592 /* Fast-path assumptions: 5593 * - There is no RX handler. 5594 * - Only one packet_type matches. 5595 * If either of these fails, we will end up doing some per-packet 5596 * processing in-line, then handling the 'last ptype' for the whole 5597 * sublist. This can't cause out-of-order delivery to any single ptype, 5598 * because the 'last ptype' must be constant across the sublist, and all 5599 * other ptypes are handled per-packet. 5600 */ 5601 /* Current (common) ptype of sublist */ 5602 struct packet_type *pt_curr = NULL; 5603 /* Current (common) orig_dev of sublist */ 5604 struct net_device *od_curr = NULL; 5605 struct list_head sublist; 5606 struct sk_buff *skb, *next; 5607 5608 INIT_LIST_HEAD(&sublist); 5609 list_for_each_entry_safe(skb, next, head, list) { 5610 struct net_device *orig_dev = skb->dev; 5611 struct packet_type *pt_prev = NULL; 5612 5613 skb_list_del_init(skb); 5614 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5615 if (!pt_prev) 5616 continue; 5617 if (pt_curr != pt_prev || od_curr != orig_dev) { 5618 /* dispatch old sublist */ 5619 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5620 /* start new sublist */ 5621 INIT_LIST_HEAD(&sublist); 5622 pt_curr = pt_prev; 5623 od_curr = orig_dev; 5624 } 5625 list_add_tail(&skb->list, &sublist); 5626 } 5627 5628 /* dispatch final sublist */ 5629 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5630 } 5631 5632 static int __netif_receive_skb(struct sk_buff *skb) 5633 { 5634 int ret; 5635 5636 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) { 5637 unsigned int noreclaim_flag; 5638 5639 /* 5640 * PFMEMALLOC skbs are special, they should 5641 * - be delivered to SOCK_MEMALLOC sockets only 5642 * - stay away from userspace 5643 * - have bounded memory usage 5644 * 5645 * Use PF_MEMALLOC as this saves us from propagating the allocation 5646 * context down to all allocation sites. 5647 */ 5648 noreclaim_flag = memalloc_noreclaim_save(); 5649 ret = __netif_receive_skb_one_core(skb, true); 5650 memalloc_noreclaim_restore(noreclaim_flag); 5651 } else 5652 ret = __netif_receive_skb_one_core(skb, false); 5653 5654 return ret; 5655 } 5656 5657 static void __netif_receive_skb_list(struct list_head *head) 5658 { 5659 unsigned long noreclaim_flag = 0; 5660 struct sk_buff *skb, *next; 5661 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */ 5662 5663 list_for_each_entry_safe(skb, next, head, list) { 5664 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) { 5665 struct list_head sublist; 5666 5667 /* Handle the previous sublist */ 5668 list_cut_before(&sublist, head, &skb->list); 5669 if (!list_empty(&sublist)) 5670 __netif_receive_skb_list_core(&sublist, pfmemalloc); 5671 pfmemalloc = !pfmemalloc; 5672 /* See comments in __netif_receive_skb */ 5673 if (pfmemalloc) 5674 noreclaim_flag = memalloc_noreclaim_save(); 5675 else 5676 memalloc_noreclaim_restore(noreclaim_flag); 5677 } 5678 } 5679 /* Handle the remaining sublist */ 5680 if (!list_empty(head)) 5681 __netif_receive_skb_list_core(head, pfmemalloc); 5682 /* Restore pflags */ 5683 if (pfmemalloc) 5684 memalloc_noreclaim_restore(noreclaim_flag); 5685 } 5686 5687 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp) 5688 { 5689 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog); 5690 struct bpf_prog *new = xdp->prog; 5691 int ret = 0; 5692 5693 switch (xdp->command) { 5694 case XDP_SETUP_PROG: 5695 rcu_assign_pointer(dev->xdp_prog, new); 5696 if (old) 5697 bpf_prog_put(old); 5698 5699 if (old && !new) { 5700 static_branch_dec(&generic_xdp_needed_key); 5701 } else if (new && !old) { 5702 static_branch_inc(&generic_xdp_needed_key); 5703 dev_disable_lro(dev); 5704 dev_disable_gro_hw(dev); 5705 } 5706 break; 5707 5708 default: 5709 ret = -EINVAL; 5710 break; 5711 } 5712 5713 return ret; 5714 } 5715 5716 static int netif_receive_skb_internal(struct sk_buff *skb) 5717 { 5718 int ret; 5719 5720 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5721 5722 if (skb_defer_rx_timestamp(skb)) 5723 return NET_RX_SUCCESS; 5724 5725 rcu_read_lock(); 5726 #ifdef CONFIG_RPS 5727 if (static_branch_unlikely(&rps_needed)) { 5728 struct rps_dev_flow voidflow, *rflow = &voidflow; 5729 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 5730 5731 if (cpu >= 0) { 5732 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5733 rcu_read_unlock(); 5734 return ret; 5735 } 5736 } 5737 #endif 5738 ret = __netif_receive_skb(skb); 5739 rcu_read_unlock(); 5740 return ret; 5741 } 5742 5743 void netif_receive_skb_list_internal(struct list_head *head) 5744 { 5745 struct sk_buff *skb, *next; 5746 struct list_head sublist; 5747 5748 INIT_LIST_HEAD(&sublist); 5749 list_for_each_entry_safe(skb, next, head, list) { 5750 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), 5751 skb); 5752 skb_list_del_init(skb); 5753 if (!skb_defer_rx_timestamp(skb)) 5754 list_add_tail(&skb->list, &sublist); 5755 } 5756 list_splice_init(&sublist, head); 5757 5758 rcu_read_lock(); 5759 #ifdef CONFIG_RPS 5760 if (static_branch_unlikely(&rps_needed)) { 5761 list_for_each_entry_safe(skb, next, head, list) { 5762 struct rps_dev_flow voidflow, *rflow = &voidflow; 5763 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 5764 5765 if (cpu >= 0) { 5766 /* Will be handled, remove from list */ 5767 skb_list_del_init(skb); 5768 enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5769 } 5770 } 5771 } 5772 #endif 5773 __netif_receive_skb_list(head); 5774 rcu_read_unlock(); 5775 } 5776 5777 /** 5778 * netif_receive_skb - process receive buffer from network 5779 * @skb: buffer to process 5780 * 5781 * netif_receive_skb() is the main receive data processing function. 5782 * It always succeeds. The buffer may be dropped during processing 5783 * for congestion control or by the protocol layers. 5784 * 5785 * This function may only be called from softirq context and interrupts 5786 * should be enabled. 5787 * 5788 * Return values (usually ignored): 5789 * NET_RX_SUCCESS: no congestion 5790 * NET_RX_DROP: packet was dropped 5791 */ 5792 int netif_receive_skb(struct sk_buff *skb) 5793 { 5794 int ret; 5795 5796 trace_netif_receive_skb_entry(skb); 5797 5798 ret = netif_receive_skb_internal(skb); 5799 trace_netif_receive_skb_exit(ret); 5800 5801 return ret; 5802 } 5803 EXPORT_SYMBOL(netif_receive_skb); 5804 5805 /** 5806 * netif_receive_skb_list - process many receive buffers from network 5807 * @head: list of skbs to process. 5808 * 5809 * Since return value of netif_receive_skb() is normally ignored, and 5810 * wouldn't be meaningful for a list, this function returns void. 5811 * 5812 * This function may only be called from softirq context and interrupts 5813 * should be enabled. 5814 */ 5815 void netif_receive_skb_list(struct list_head *head) 5816 { 5817 struct sk_buff *skb; 5818 5819 if (list_empty(head)) 5820 return; 5821 if (trace_netif_receive_skb_list_entry_enabled()) { 5822 list_for_each_entry(skb, head, list) 5823 trace_netif_receive_skb_list_entry(skb); 5824 } 5825 netif_receive_skb_list_internal(head); 5826 trace_netif_receive_skb_list_exit(0); 5827 } 5828 EXPORT_SYMBOL(netif_receive_skb_list); 5829 5830 static DEFINE_PER_CPU(struct work_struct, flush_works); 5831 5832 /* Network device is going away, flush any packets still pending */ 5833 static void flush_backlog(struct work_struct *work) 5834 { 5835 struct sk_buff *skb, *tmp; 5836 struct softnet_data *sd; 5837 5838 local_bh_disable(); 5839 sd = this_cpu_ptr(&softnet_data); 5840 5841 rps_lock_irq_disable(sd); 5842 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) { 5843 if (skb->dev->reg_state == NETREG_UNREGISTERING) { 5844 __skb_unlink(skb, &sd->input_pkt_queue); 5845 dev_kfree_skb_irq(skb); 5846 input_queue_head_incr(sd); 5847 } 5848 } 5849 rps_unlock_irq_enable(sd); 5850 5851 skb_queue_walk_safe(&sd->process_queue, skb, tmp) { 5852 if (skb->dev->reg_state == NETREG_UNREGISTERING) { 5853 __skb_unlink(skb, &sd->process_queue); 5854 kfree_skb(skb); 5855 input_queue_head_incr(sd); 5856 } 5857 } 5858 local_bh_enable(); 5859 } 5860 5861 static bool flush_required(int cpu) 5862 { 5863 #if IS_ENABLED(CONFIG_RPS) 5864 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 5865 bool do_flush; 5866 5867 rps_lock_irq_disable(sd); 5868 5869 /* as insertion into process_queue happens with the rps lock held, 5870 * process_queue access may race only with dequeue 5871 */ 5872 do_flush = !skb_queue_empty(&sd->input_pkt_queue) || 5873 !skb_queue_empty_lockless(&sd->process_queue); 5874 rps_unlock_irq_enable(sd); 5875 5876 return do_flush; 5877 #endif 5878 /* without RPS we can't safely check input_pkt_queue: during a 5879 * concurrent remote skb_queue_splice() we can detect as empty both 5880 * input_pkt_queue and process_queue even if the latter could end-up 5881 * containing a lot of packets. 5882 */ 5883 return true; 5884 } 5885 5886 static void flush_all_backlogs(void) 5887 { 5888 static cpumask_t flush_cpus; 5889 unsigned int cpu; 5890 5891 /* since we are under rtnl lock protection we can use static data 5892 * for the cpumask and avoid allocating on stack the possibly 5893 * large mask 5894 */ 5895 ASSERT_RTNL(); 5896 5897 cpus_read_lock(); 5898 5899 cpumask_clear(&flush_cpus); 5900 for_each_online_cpu(cpu) { 5901 if (flush_required(cpu)) { 5902 queue_work_on(cpu, system_highpri_wq, 5903 per_cpu_ptr(&flush_works, cpu)); 5904 cpumask_set_cpu(cpu, &flush_cpus); 5905 } 5906 } 5907 5908 /* we can have in flight packet[s] on the cpus we are not flushing, 5909 * synchronize_net() in unregister_netdevice_many() will take care of 5910 * them 5911 */ 5912 for_each_cpu(cpu, &flush_cpus) 5913 flush_work(per_cpu_ptr(&flush_works, cpu)); 5914 5915 cpus_read_unlock(); 5916 } 5917 5918 static void net_rps_send_ipi(struct softnet_data *remsd) 5919 { 5920 #ifdef CONFIG_RPS 5921 while (remsd) { 5922 struct softnet_data *next = remsd->rps_ipi_next; 5923 5924 if (cpu_online(remsd->cpu)) 5925 smp_call_function_single_async(remsd->cpu, &remsd->csd); 5926 remsd = next; 5927 } 5928 #endif 5929 } 5930 5931 /* 5932 * net_rps_action_and_irq_enable sends any pending IPI's for rps. 5933 * Note: called with local irq disabled, but exits with local irq enabled. 5934 */ 5935 static void net_rps_action_and_irq_enable(struct softnet_data *sd) 5936 { 5937 #ifdef CONFIG_RPS 5938 struct softnet_data *remsd = sd->rps_ipi_list; 5939 5940 if (remsd) { 5941 sd->rps_ipi_list = NULL; 5942 5943 local_irq_enable(); 5944 5945 /* Send pending IPI's to kick RPS processing on remote cpus. */ 5946 net_rps_send_ipi(remsd); 5947 } else 5948 #endif 5949 local_irq_enable(); 5950 } 5951 5952 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd) 5953 { 5954 #ifdef CONFIG_RPS 5955 return sd->rps_ipi_list != NULL; 5956 #else 5957 return false; 5958 #endif 5959 } 5960 5961 static int process_backlog(struct napi_struct *napi, int quota) 5962 { 5963 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog); 5964 bool again = true; 5965 int work = 0; 5966 5967 /* Check if we have pending ipi, its better to send them now, 5968 * not waiting net_rx_action() end. 5969 */ 5970 if (sd_has_rps_ipi_waiting(sd)) { 5971 local_irq_disable(); 5972 net_rps_action_and_irq_enable(sd); 5973 } 5974 5975 napi->weight = READ_ONCE(net_hotdata.dev_rx_weight); 5976 while (again) { 5977 struct sk_buff *skb; 5978 5979 while ((skb = __skb_dequeue(&sd->process_queue))) { 5980 rcu_read_lock(); 5981 __netif_receive_skb(skb); 5982 rcu_read_unlock(); 5983 input_queue_head_incr(sd); 5984 if (++work >= quota) 5985 return work; 5986 5987 } 5988 5989 rps_lock_irq_disable(sd); 5990 if (skb_queue_empty(&sd->input_pkt_queue)) { 5991 /* 5992 * Inline a custom version of __napi_complete(). 5993 * only current cpu owns and manipulates this napi, 5994 * and NAPI_STATE_SCHED is the only possible flag set 5995 * on backlog. 5996 * We can use a plain write instead of clear_bit(), 5997 * and we dont need an smp_mb() memory barrier. 5998 */ 5999 napi->state = 0; 6000 again = false; 6001 } else { 6002 skb_queue_splice_tail_init(&sd->input_pkt_queue, 6003 &sd->process_queue); 6004 } 6005 rps_unlock_irq_enable(sd); 6006 } 6007 6008 return work; 6009 } 6010 6011 /** 6012 * __napi_schedule - schedule for receive 6013 * @n: entry to schedule 6014 * 6015 * The entry's receive function will be scheduled to run. 6016 * Consider using __napi_schedule_irqoff() if hard irqs are masked. 6017 */ 6018 void __napi_schedule(struct napi_struct *n) 6019 { 6020 unsigned long flags; 6021 6022 local_irq_save(flags); 6023 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6024 local_irq_restore(flags); 6025 } 6026 EXPORT_SYMBOL(__napi_schedule); 6027 6028 /** 6029 * napi_schedule_prep - check if napi can be scheduled 6030 * @n: napi context 6031 * 6032 * Test if NAPI routine is already running, and if not mark 6033 * it as running. This is used as a condition variable to 6034 * insure only one NAPI poll instance runs. We also make 6035 * sure there is no pending NAPI disable. 6036 */ 6037 bool napi_schedule_prep(struct napi_struct *n) 6038 { 6039 unsigned long new, val = READ_ONCE(n->state); 6040 6041 do { 6042 if (unlikely(val & NAPIF_STATE_DISABLE)) 6043 return false; 6044 new = val | NAPIF_STATE_SCHED; 6045 6046 /* Sets STATE_MISSED bit if STATE_SCHED was already set 6047 * This was suggested by Alexander Duyck, as compiler 6048 * emits better code than : 6049 * if (val & NAPIF_STATE_SCHED) 6050 * new |= NAPIF_STATE_MISSED; 6051 */ 6052 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED * 6053 NAPIF_STATE_MISSED; 6054 } while (!try_cmpxchg(&n->state, &val, new)); 6055 6056 return !(val & NAPIF_STATE_SCHED); 6057 } 6058 EXPORT_SYMBOL(napi_schedule_prep); 6059 6060 /** 6061 * __napi_schedule_irqoff - schedule for receive 6062 * @n: entry to schedule 6063 * 6064 * Variant of __napi_schedule() assuming hard irqs are masked. 6065 * 6066 * On PREEMPT_RT enabled kernels this maps to __napi_schedule() 6067 * because the interrupt disabled assumption might not be true 6068 * due to force-threaded interrupts and spinlock substitution. 6069 */ 6070 void __napi_schedule_irqoff(struct napi_struct *n) 6071 { 6072 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6073 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6074 else 6075 __napi_schedule(n); 6076 } 6077 EXPORT_SYMBOL(__napi_schedule_irqoff); 6078 6079 bool napi_complete_done(struct napi_struct *n, int work_done) 6080 { 6081 unsigned long flags, val, new, timeout = 0; 6082 bool ret = true; 6083 6084 /* 6085 * 1) Don't let napi dequeue from the cpu poll list 6086 * just in case its running on a different cpu. 6087 * 2) If we are busy polling, do nothing here, we have 6088 * the guarantee we will be called later. 6089 */ 6090 if (unlikely(n->state & (NAPIF_STATE_NPSVC | 6091 NAPIF_STATE_IN_BUSY_POLL))) 6092 return false; 6093 6094 if (work_done) { 6095 if (n->gro_bitmask) 6096 timeout = READ_ONCE(n->dev->gro_flush_timeout); 6097 n->defer_hard_irqs_count = READ_ONCE(n->dev->napi_defer_hard_irqs); 6098 } 6099 if (n->defer_hard_irqs_count > 0) { 6100 n->defer_hard_irqs_count--; 6101 timeout = READ_ONCE(n->dev->gro_flush_timeout); 6102 if (timeout) 6103 ret = false; 6104 } 6105 if (n->gro_bitmask) { 6106 /* When the NAPI instance uses a timeout and keeps postponing 6107 * it, we need to bound somehow the time packets are kept in 6108 * the GRO layer 6109 */ 6110 napi_gro_flush(n, !!timeout); 6111 } 6112 6113 gro_normal_list(n); 6114 6115 if (unlikely(!list_empty(&n->poll_list))) { 6116 /* If n->poll_list is not empty, we need to mask irqs */ 6117 local_irq_save(flags); 6118 list_del_init(&n->poll_list); 6119 local_irq_restore(flags); 6120 } 6121 WRITE_ONCE(n->list_owner, -1); 6122 6123 val = READ_ONCE(n->state); 6124 do { 6125 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED)); 6126 6127 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED | 6128 NAPIF_STATE_SCHED_THREADED | 6129 NAPIF_STATE_PREFER_BUSY_POLL); 6130 6131 /* If STATE_MISSED was set, leave STATE_SCHED set, 6132 * because we will call napi->poll() one more time. 6133 * This C code was suggested by Alexander Duyck to help gcc. 6134 */ 6135 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED * 6136 NAPIF_STATE_SCHED; 6137 } while (!try_cmpxchg(&n->state, &val, new)); 6138 6139 if (unlikely(val & NAPIF_STATE_MISSED)) { 6140 __napi_schedule(n); 6141 return false; 6142 } 6143 6144 if (timeout) 6145 hrtimer_start(&n->timer, ns_to_ktime(timeout), 6146 HRTIMER_MODE_REL_PINNED); 6147 return ret; 6148 } 6149 EXPORT_SYMBOL(napi_complete_done); 6150 6151 /* must be called under rcu_read_lock(), as we dont take a reference */ 6152 struct napi_struct *napi_by_id(unsigned int napi_id) 6153 { 6154 unsigned int hash = napi_id % HASH_SIZE(napi_hash); 6155 struct napi_struct *napi; 6156 6157 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node) 6158 if (napi->napi_id == napi_id) 6159 return napi; 6160 6161 return NULL; 6162 } 6163 6164 static void skb_defer_free_flush(struct softnet_data *sd) 6165 { 6166 struct sk_buff *skb, *next; 6167 6168 /* Paired with WRITE_ONCE() in skb_attempt_defer_free() */ 6169 if (!READ_ONCE(sd->defer_list)) 6170 return; 6171 6172 spin_lock(&sd->defer_lock); 6173 skb = sd->defer_list; 6174 sd->defer_list = NULL; 6175 sd->defer_count = 0; 6176 spin_unlock(&sd->defer_lock); 6177 6178 while (skb != NULL) { 6179 next = skb->next; 6180 napi_consume_skb(skb, 1); 6181 skb = next; 6182 } 6183 } 6184 6185 #if defined(CONFIG_NET_RX_BUSY_POLL) 6186 6187 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule) 6188 { 6189 if (!skip_schedule) { 6190 gro_normal_list(napi); 6191 __napi_schedule(napi); 6192 return; 6193 } 6194 6195 if (napi->gro_bitmask) { 6196 /* flush too old packets 6197 * If HZ < 1000, flush all packets. 6198 */ 6199 napi_gro_flush(napi, HZ >= 1000); 6200 } 6201 6202 gro_normal_list(napi); 6203 clear_bit(NAPI_STATE_SCHED, &napi->state); 6204 } 6205 6206 enum { 6207 NAPI_F_PREFER_BUSY_POLL = 1, 6208 NAPI_F_END_ON_RESCHED = 2, 6209 }; 6210 6211 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, 6212 unsigned flags, u16 budget) 6213 { 6214 bool skip_schedule = false; 6215 unsigned long timeout; 6216 int rc; 6217 6218 /* Busy polling means there is a high chance device driver hard irq 6219 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was 6220 * set in napi_schedule_prep(). 6221 * Since we are about to call napi->poll() once more, we can safely 6222 * clear NAPI_STATE_MISSED. 6223 * 6224 * Note: x86 could use a single "lock and ..." instruction 6225 * to perform these two clear_bit() 6226 */ 6227 clear_bit(NAPI_STATE_MISSED, &napi->state); 6228 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state); 6229 6230 local_bh_disable(); 6231 6232 if (flags & NAPI_F_PREFER_BUSY_POLL) { 6233 napi->defer_hard_irqs_count = READ_ONCE(napi->dev->napi_defer_hard_irqs); 6234 timeout = READ_ONCE(napi->dev->gro_flush_timeout); 6235 if (napi->defer_hard_irqs_count && timeout) { 6236 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED); 6237 skip_schedule = true; 6238 } 6239 } 6240 6241 /* All we really want here is to re-enable device interrupts. 6242 * Ideally, a new ndo_busy_poll_stop() could avoid another round. 6243 */ 6244 rc = napi->poll(napi, budget); 6245 /* We can't gro_normal_list() here, because napi->poll() might have 6246 * rearmed the napi (napi_complete_done()) in which case it could 6247 * already be running on another CPU. 6248 */ 6249 trace_napi_poll(napi, rc, budget); 6250 netpoll_poll_unlock(have_poll_lock); 6251 if (rc == budget) 6252 __busy_poll_stop(napi, skip_schedule); 6253 local_bh_enable(); 6254 } 6255 6256 static void __napi_busy_loop(unsigned int napi_id, 6257 bool (*loop_end)(void *, unsigned long), 6258 void *loop_end_arg, unsigned flags, u16 budget) 6259 { 6260 unsigned long start_time = loop_end ? busy_loop_current_time() : 0; 6261 int (*napi_poll)(struct napi_struct *napi, int budget); 6262 void *have_poll_lock = NULL; 6263 struct napi_struct *napi; 6264 6265 WARN_ON_ONCE(!rcu_read_lock_held()); 6266 6267 restart: 6268 napi_poll = NULL; 6269 6270 napi = napi_by_id(napi_id); 6271 if (!napi) 6272 return; 6273 6274 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6275 preempt_disable(); 6276 for (;;) { 6277 int work = 0; 6278 6279 local_bh_disable(); 6280 if (!napi_poll) { 6281 unsigned long val = READ_ONCE(napi->state); 6282 6283 /* If multiple threads are competing for this napi, 6284 * we avoid dirtying napi->state as much as we can. 6285 */ 6286 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED | 6287 NAPIF_STATE_IN_BUSY_POLL)) { 6288 if (flags & NAPI_F_PREFER_BUSY_POLL) 6289 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6290 goto count; 6291 } 6292 if (cmpxchg(&napi->state, val, 6293 val | NAPIF_STATE_IN_BUSY_POLL | 6294 NAPIF_STATE_SCHED) != val) { 6295 if (flags & NAPI_F_PREFER_BUSY_POLL) 6296 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6297 goto count; 6298 } 6299 have_poll_lock = netpoll_poll_lock(napi); 6300 napi_poll = napi->poll; 6301 } 6302 work = napi_poll(napi, budget); 6303 trace_napi_poll(napi, work, budget); 6304 gro_normal_list(napi); 6305 count: 6306 if (work > 0) 6307 __NET_ADD_STATS(dev_net(napi->dev), 6308 LINUX_MIB_BUSYPOLLRXPACKETS, work); 6309 skb_defer_free_flush(this_cpu_ptr(&softnet_data)); 6310 local_bh_enable(); 6311 6312 if (!loop_end || loop_end(loop_end_arg, start_time)) 6313 break; 6314 6315 if (unlikely(need_resched())) { 6316 if (flags & NAPI_F_END_ON_RESCHED) 6317 break; 6318 if (napi_poll) 6319 busy_poll_stop(napi, have_poll_lock, flags, budget); 6320 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6321 preempt_enable(); 6322 rcu_read_unlock(); 6323 cond_resched(); 6324 rcu_read_lock(); 6325 if (loop_end(loop_end_arg, start_time)) 6326 return; 6327 goto restart; 6328 } 6329 cpu_relax(); 6330 } 6331 if (napi_poll) 6332 busy_poll_stop(napi, have_poll_lock, flags, budget); 6333 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6334 preempt_enable(); 6335 } 6336 6337 void napi_busy_loop_rcu(unsigned int napi_id, 6338 bool (*loop_end)(void *, unsigned long), 6339 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6340 { 6341 unsigned flags = NAPI_F_END_ON_RESCHED; 6342 6343 if (prefer_busy_poll) 6344 flags |= NAPI_F_PREFER_BUSY_POLL; 6345 6346 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6347 } 6348 6349 void napi_busy_loop(unsigned int napi_id, 6350 bool (*loop_end)(void *, unsigned long), 6351 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6352 { 6353 unsigned flags = prefer_busy_poll ? NAPI_F_PREFER_BUSY_POLL : 0; 6354 6355 rcu_read_lock(); 6356 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6357 rcu_read_unlock(); 6358 } 6359 EXPORT_SYMBOL(napi_busy_loop); 6360 6361 #endif /* CONFIG_NET_RX_BUSY_POLL */ 6362 6363 static void napi_hash_add(struct napi_struct *napi) 6364 { 6365 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state)) 6366 return; 6367 6368 spin_lock(&napi_hash_lock); 6369 6370 /* 0..NR_CPUS range is reserved for sender_cpu use */ 6371 do { 6372 if (unlikely(++napi_gen_id < MIN_NAPI_ID)) 6373 napi_gen_id = MIN_NAPI_ID; 6374 } while (napi_by_id(napi_gen_id)); 6375 napi->napi_id = napi_gen_id; 6376 6377 hlist_add_head_rcu(&napi->napi_hash_node, 6378 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]); 6379 6380 spin_unlock(&napi_hash_lock); 6381 } 6382 6383 /* Warning : caller is responsible to make sure rcu grace period 6384 * is respected before freeing memory containing @napi 6385 */ 6386 static void napi_hash_del(struct napi_struct *napi) 6387 { 6388 spin_lock(&napi_hash_lock); 6389 6390 hlist_del_init_rcu(&napi->napi_hash_node); 6391 6392 spin_unlock(&napi_hash_lock); 6393 } 6394 6395 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer) 6396 { 6397 struct napi_struct *napi; 6398 6399 napi = container_of(timer, struct napi_struct, timer); 6400 6401 /* Note : we use a relaxed variant of napi_schedule_prep() not setting 6402 * NAPI_STATE_MISSED, since we do not react to a device IRQ. 6403 */ 6404 if (!napi_disable_pending(napi) && 6405 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) { 6406 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6407 __napi_schedule_irqoff(napi); 6408 } 6409 6410 return HRTIMER_NORESTART; 6411 } 6412 6413 static void init_gro_hash(struct napi_struct *napi) 6414 { 6415 int i; 6416 6417 for (i = 0; i < GRO_HASH_BUCKETS; i++) { 6418 INIT_LIST_HEAD(&napi->gro_hash[i].list); 6419 napi->gro_hash[i].count = 0; 6420 } 6421 napi->gro_bitmask = 0; 6422 } 6423 6424 int dev_set_threaded(struct net_device *dev, bool threaded) 6425 { 6426 struct napi_struct *napi; 6427 int err = 0; 6428 6429 if (dev->threaded == threaded) 6430 return 0; 6431 6432 if (threaded) { 6433 list_for_each_entry(napi, &dev->napi_list, dev_list) { 6434 if (!napi->thread) { 6435 err = napi_kthread_create(napi); 6436 if (err) { 6437 threaded = false; 6438 break; 6439 } 6440 } 6441 } 6442 } 6443 6444 dev->threaded = threaded; 6445 6446 /* Make sure kthread is created before THREADED bit 6447 * is set. 6448 */ 6449 smp_mb__before_atomic(); 6450 6451 /* Setting/unsetting threaded mode on a napi might not immediately 6452 * take effect, if the current napi instance is actively being 6453 * polled. In this case, the switch between threaded mode and 6454 * softirq mode will happen in the next round of napi_schedule(). 6455 * This should not cause hiccups/stalls to the live traffic. 6456 */ 6457 list_for_each_entry(napi, &dev->napi_list, dev_list) 6458 assign_bit(NAPI_STATE_THREADED, &napi->state, threaded); 6459 6460 return err; 6461 } 6462 EXPORT_SYMBOL(dev_set_threaded); 6463 6464 /** 6465 * netif_queue_set_napi - Associate queue with the napi 6466 * @dev: device to which NAPI and queue belong 6467 * @queue_index: Index of queue 6468 * @type: queue type as RX or TX 6469 * @napi: NAPI context, pass NULL to clear previously set NAPI 6470 * 6471 * Set queue with its corresponding napi context. This should be done after 6472 * registering the NAPI handler for the queue-vector and the queues have been 6473 * mapped to the corresponding interrupt vector. 6474 */ 6475 void netif_queue_set_napi(struct net_device *dev, unsigned int queue_index, 6476 enum netdev_queue_type type, struct napi_struct *napi) 6477 { 6478 struct netdev_rx_queue *rxq; 6479 struct netdev_queue *txq; 6480 6481 if (WARN_ON_ONCE(napi && !napi->dev)) 6482 return; 6483 if (dev->reg_state >= NETREG_REGISTERED) 6484 ASSERT_RTNL(); 6485 6486 switch (type) { 6487 case NETDEV_QUEUE_TYPE_RX: 6488 rxq = __netif_get_rx_queue(dev, queue_index); 6489 rxq->napi = napi; 6490 return; 6491 case NETDEV_QUEUE_TYPE_TX: 6492 txq = netdev_get_tx_queue(dev, queue_index); 6493 txq->napi = napi; 6494 return; 6495 default: 6496 return; 6497 } 6498 } 6499 EXPORT_SYMBOL(netif_queue_set_napi); 6500 6501 void netif_napi_add_weight(struct net_device *dev, struct napi_struct *napi, 6502 int (*poll)(struct napi_struct *, int), int weight) 6503 { 6504 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state))) 6505 return; 6506 6507 INIT_LIST_HEAD(&napi->poll_list); 6508 INIT_HLIST_NODE(&napi->napi_hash_node); 6509 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); 6510 napi->timer.function = napi_watchdog; 6511 init_gro_hash(napi); 6512 napi->skb = NULL; 6513 INIT_LIST_HEAD(&napi->rx_list); 6514 napi->rx_count = 0; 6515 napi->poll = poll; 6516 if (weight > NAPI_POLL_WEIGHT) 6517 netdev_err_once(dev, "%s() called with weight %d\n", __func__, 6518 weight); 6519 napi->weight = weight; 6520 napi->dev = dev; 6521 #ifdef CONFIG_NETPOLL 6522 napi->poll_owner = -1; 6523 #endif 6524 napi->list_owner = -1; 6525 set_bit(NAPI_STATE_SCHED, &napi->state); 6526 set_bit(NAPI_STATE_NPSVC, &napi->state); 6527 list_add_rcu(&napi->dev_list, &dev->napi_list); 6528 napi_hash_add(napi); 6529 napi_get_frags_check(napi); 6530 /* Create kthread for this napi if dev->threaded is set. 6531 * Clear dev->threaded if kthread creation failed so that 6532 * threaded mode will not be enabled in napi_enable(). 6533 */ 6534 if (dev->threaded && napi_kthread_create(napi)) 6535 dev->threaded = 0; 6536 netif_napi_set_irq(napi, -1); 6537 } 6538 EXPORT_SYMBOL(netif_napi_add_weight); 6539 6540 void napi_disable(struct napi_struct *n) 6541 { 6542 unsigned long val, new; 6543 6544 might_sleep(); 6545 set_bit(NAPI_STATE_DISABLE, &n->state); 6546 6547 val = READ_ONCE(n->state); 6548 do { 6549 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) { 6550 usleep_range(20, 200); 6551 val = READ_ONCE(n->state); 6552 } 6553 6554 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC; 6555 new &= ~(NAPIF_STATE_THREADED | NAPIF_STATE_PREFER_BUSY_POLL); 6556 } while (!try_cmpxchg(&n->state, &val, new)); 6557 6558 hrtimer_cancel(&n->timer); 6559 6560 clear_bit(NAPI_STATE_DISABLE, &n->state); 6561 } 6562 EXPORT_SYMBOL(napi_disable); 6563 6564 /** 6565 * napi_enable - enable NAPI scheduling 6566 * @n: NAPI context 6567 * 6568 * Resume NAPI from being scheduled on this context. 6569 * Must be paired with napi_disable. 6570 */ 6571 void napi_enable(struct napi_struct *n) 6572 { 6573 unsigned long new, val = READ_ONCE(n->state); 6574 6575 do { 6576 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val)); 6577 6578 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC); 6579 if (n->dev->threaded && n->thread) 6580 new |= NAPIF_STATE_THREADED; 6581 } while (!try_cmpxchg(&n->state, &val, new)); 6582 } 6583 EXPORT_SYMBOL(napi_enable); 6584 6585 static void flush_gro_hash(struct napi_struct *napi) 6586 { 6587 int i; 6588 6589 for (i = 0; i < GRO_HASH_BUCKETS; i++) { 6590 struct sk_buff *skb, *n; 6591 6592 list_for_each_entry_safe(skb, n, &napi->gro_hash[i].list, list) 6593 kfree_skb(skb); 6594 napi->gro_hash[i].count = 0; 6595 } 6596 } 6597 6598 /* Must be called in process context */ 6599 void __netif_napi_del(struct napi_struct *napi) 6600 { 6601 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state)) 6602 return; 6603 6604 napi_hash_del(napi); 6605 list_del_rcu(&napi->dev_list); 6606 napi_free_frags(napi); 6607 6608 flush_gro_hash(napi); 6609 napi->gro_bitmask = 0; 6610 6611 if (napi->thread) { 6612 kthread_stop(napi->thread); 6613 napi->thread = NULL; 6614 } 6615 } 6616 EXPORT_SYMBOL(__netif_napi_del); 6617 6618 static int __napi_poll(struct napi_struct *n, bool *repoll) 6619 { 6620 int work, weight; 6621 6622 weight = n->weight; 6623 6624 /* This NAPI_STATE_SCHED test is for avoiding a race 6625 * with netpoll's poll_napi(). Only the entity which 6626 * obtains the lock and sees NAPI_STATE_SCHED set will 6627 * actually make the ->poll() call. Therefore we avoid 6628 * accidentally calling ->poll() when NAPI is not scheduled. 6629 */ 6630 work = 0; 6631 if (napi_is_scheduled(n)) { 6632 work = n->poll(n, weight); 6633 trace_napi_poll(n, work, weight); 6634 6635 xdp_do_check_flushed(n); 6636 } 6637 6638 if (unlikely(work > weight)) 6639 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n", 6640 n->poll, work, weight); 6641 6642 if (likely(work < weight)) 6643 return work; 6644 6645 /* Drivers must not modify the NAPI state if they 6646 * consume the entire weight. In such cases this code 6647 * still "owns" the NAPI instance and therefore can 6648 * move the instance around on the list at-will. 6649 */ 6650 if (unlikely(napi_disable_pending(n))) { 6651 napi_complete(n); 6652 return work; 6653 } 6654 6655 /* The NAPI context has more processing work, but busy-polling 6656 * is preferred. Exit early. 6657 */ 6658 if (napi_prefer_busy_poll(n)) { 6659 if (napi_complete_done(n, work)) { 6660 /* If timeout is not set, we need to make sure 6661 * that the NAPI is re-scheduled. 6662 */ 6663 napi_schedule(n); 6664 } 6665 return work; 6666 } 6667 6668 if (n->gro_bitmask) { 6669 /* flush too old packets 6670 * If HZ < 1000, flush all packets. 6671 */ 6672 napi_gro_flush(n, HZ >= 1000); 6673 } 6674 6675 gro_normal_list(n); 6676 6677 /* Some drivers may have called napi_schedule 6678 * prior to exhausting their budget. 6679 */ 6680 if (unlikely(!list_empty(&n->poll_list))) { 6681 pr_warn_once("%s: Budget exhausted after napi rescheduled\n", 6682 n->dev ? n->dev->name : "backlog"); 6683 return work; 6684 } 6685 6686 *repoll = true; 6687 6688 return work; 6689 } 6690 6691 static int napi_poll(struct napi_struct *n, struct list_head *repoll) 6692 { 6693 bool do_repoll = false; 6694 void *have; 6695 int work; 6696 6697 list_del_init(&n->poll_list); 6698 6699 have = netpoll_poll_lock(n); 6700 6701 work = __napi_poll(n, &do_repoll); 6702 6703 if (do_repoll) 6704 list_add_tail(&n->poll_list, repoll); 6705 6706 netpoll_poll_unlock(have); 6707 6708 return work; 6709 } 6710 6711 static int napi_thread_wait(struct napi_struct *napi) 6712 { 6713 bool woken = false; 6714 6715 set_current_state(TASK_INTERRUPTIBLE); 6716 6717 while (!kthread_should_stop()) { 6718 /* Testing SCHED_THREADED bit here to make sure the current 6719 * kthread owns this napi and could poll on this napi. 6720 * Testing SCHED bit is not enough because SCHED bit might be 6721 * set by some other busy poll thread or by napi_disable(). 6722 */ 6723 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state) || woken) { 6724 WARN_ON(!list_empty(&napi->poll_list)); 6725 __set_current_state(TASK_RUNNING); 6726 return 0; 6727 } 6728 6729 schedule(); 6730 /* woken being true indicates this thread owns this napi. */ 6731 woken = true; 6732 set_current_state(TASK_INTERRUPTIBLE); 6733 } 6734 __set_current_state(TASK_RUNNING); 6735 6736 return -1; 6737 } 6738 6739 static int napi_threaded_poll(void *data) 6740 { 6741 struct napi_struct *napi = data; 6742 struct softnet_data *sd; 6743 void *have; 6744 6745 while (!napi_thread_wait(napi)) { 6746 for (;;) { 6747 bool repoll = false; 6748 6749 local_bh_disable(); 6750 sd = this_cpu_ptr(&softnet_data); 6751 sd->in_napi_threaded_poll = true; 6752 6753 have = netpoll_poll_lock(napi); 6754 __napi_poll(napi, &repoll); 6755 netpoll_poll_unlock(have); 6756 6757 sd->in_napi_threaded_poll = false; 6758 barrier(); 6759 6760 if (sd_has_rps_ipi_waiting(sd)) { 6761 local_irq_disable(); 6762 net_rps_action_and_irq_enable(sd); 6763 } 6764 skb_defer_free_flush(sd); 6765 local_bh_enable(); 6766 6767 if (!repoll) 6768 break; 6769 6770 cond_resched(); 6771 } 6772 } 6773 return 0; 6774 } 6775 6776 static __latent_entropy void net_rx_action(struct softirq_action *h) 6777 { 6778 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 6779 unsigned long time_limit = jiffies + 6780 usecs_to_jiffies(READ_ONCE(net_hotdata.netdev_budget_usecs)); 6781 int budget = READ_ONCE(net_hotdata.netdev_budget); 6782 LIST_HEAD(list); 6783 LIST_HEAD(repoll); 6784 6785 start: 6786 sd->in_net_rx_action = true; 6787 local_irq_disable(); 6788 list_splice_init(&sd->poll_list, &list); 6789 local_irq_enable(); 6790 6791 for (;;) { 6792 struct napi_struct *n; 6793 6794 skb_defer_free_flush(sd); 6795 6796 if (list_empty(&list)) { 6797 if (list_empty(&repoll)) { 6798 sd->in_net_rx_action = false; 6799 barrier(); 6800 /* We need to check if ____napi_schedule() 6801 * had refilled poll_list while 6802 * sd->in_net_rx_action was true. 6803 */ 6804 if (!list_empty(&sd->poll_list)) 6805 goto start; 6806 if (!sd_has_rps_ipi_waiting(sd)) 6807 goto end; 6808 } 6809 break; 6810 } 6811 6812 n = list_first_entry(&list, struct napi_struct, poll_list); 6813 budget -= napi_poll(n, &repoll); 6814 6815 /* If softirq window is exhausted then punt. 6816 * Allow this to run for 2 jiffies since which will allow 6817 * an average latency of 1.5/HZ. 6818 */ 6819 if (unlikely(budget <= 0 || 6820 time_after_eq(jiffies, time_limit))) { 6821 sd->time_squeeze++; 6822 break; 6823 } 6824 } 6825 6826 local_irq_disable(); 6827 6828 list_splice_tail_init(&sd->poll_list, &list); 6829 list_splice_tail(&repoll, &list); 6830 list_splice(&list, &sd->poll_list); 6831 if (!list_empty(&sd->poll_list)) 6832 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 6833 else 6834 sd->in_net_rx_action = false; 6835 6836 net_rps_action_and_irq_enable(sd); 6837 end:; 6838 } 6839 6840 struct netdev_adjacent { 6841 struct net_device *dev; 6842 netdevice_tracker dev_tracker; 6843 6844 /* upper master flag, there can only be one master device per list */ 6845 bool master; 6846 6847 /* lookup ignore flag */ 6848 bool ignore; 6849 6850 /* counter for the number of times this device was added to us */ 6851 u16 ref_nr; 6852 6853 /* private field for the users */ 6854 void *private; 6855 6856 struct list_head list; 6857 struct rcu_head rcu; 6858 }; 6859 6860 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev, 6861 struct list_head *adj_list) 6862 { 6863 struct netdev_adjacent *adj; 6864 6865 list_for_each_entry(adj, adj_list, list) { 6866 if (adj->dev == adj_dev) 6867 return adj; 6868 } 6869 return NULL; 6870 } 6871 6872 static int ____netdev_has_upper_dev(struct net_device *upper_dev, 6873 struct netdev_nested_priv *priv) 6874 { 6875 struct net_device *dev = (struct net_device *)priv->data; 6876 6877 return upper_dev == dev; 6878 } 6879 6880 /** 6881 * netdev_has_upper_dev - Check if device is linked to an upper device 6882 * @dev: device 6883 * @upper_dev: upper device to check 6884 * 6885 * Find out if a device is linked to specified upper device and return true 6886 * in case it is. Note that this checks only immediate upper device, 6887 * not through a complete stack of devices. The caller must hold the RTNL lock. 6888 */ 6889 bool netdev_has_upper_dev(struct net_device *dev, 6890 struct net_device *upper_dev) 6891 { 6892 struct netdev_nested_priv priv = { 6893 .data = (void *)upper_dev, 6894 }; 6895 6896 ASSERT_RTNL(); 6897 6898 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 6899 &priv); 6900 } 6901 EXPORT_SYMBOL(netdev_has_upper_dev); 6902 6903 /** 6904 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device 6905 * @dev: device 6906 * @upper_dev: upper device to check 6907 * 6908 * Find out if a device is linked to specified upper device and return true 6909 * in case it is. Note that this checks the entire upper device chain. 6910 * The caller must hold rcu lock. 6911 */ 6912 6913 bool netdev_has_upper_dev_all_rcu(struct net_device *dev, 6914 struct net_device *upper_dev) 6915 { 6916 struct netdev_nested_priv priv = { 6917 .data = (void *)upper_dev, 6918 }; 6919 6920 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 6921 &priv); 6922 } 6923 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu); 6924 6925 /** 6926 * netdev_has_any_upper_dev - Check if device is linked to some device 6927 * @dev: device 6928 * 6929 * Find out if a device is linked to an upper device and return true in case 6930 * it is. The caller must hold the RTNL lock. 6931 */ 6932 bool netdev_has_any_upper_dev(struct net_device *dev) 6933 { 6934 ASSERT_RTNL(); 6935 6936 return !list_empty(&dev->adj_list.upper); 6937 } 6938 EXPORT_SYMBOL(netdev_has_any_upper_dev); 6939 6940 /** 6941 * netdev_master_upper_dev_get - Get master upper device 6942 * @dev: device 6943 * 6944 * Find a master upper device and return pointer to it or NULL in case 6945 * it's not there. The caller must hold the RTNL lock. 6946 */ 6947 struct net_device *netdev_master_upper_dev_get(struct net_device *dev) 6948 { 6949 struct netdev_adjacent *upper; 6950 6951 ASSERT_RTNL(); 6952 6953 if (list_empty(&dev->adj_list.upper)) 6954 return NULL; 6955 6956 upper = list_first_entry(&dev->adj_list.upper, 6957 struct netdev_adjacent, list); 6958 if (likely(upper->master)) 6959 return upper->dev; 6960 return NULL; 6961 } 6962 EXPORT_SYMBOL(netdev_master_upper_dev_get); 6963 6964 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev) 6965 { 6966 struct netdev_adjacent *upper; 6967 6968 ASSERT_RTNL(); 6969 6970 if (list_empty(&dev->adj_list.upper)) 6971 return NULL; 6972 6973 upper = list_first_entry(&dev->adj_list.upper, 6974 struct netdev_adjacent, list); 6975 if (likely(upper->master) && !upper->ignore) 6976 return upper->dev; 6977 return NULL; 6978 } 6979 6980 /** 6981 * netdev_has_any_lower_dev - Check if device is linked to some device 6982 * @dev: device 6983 * 6984 * Find out if a device is linked to a lower device and return true in case 6985 * it is. The caller must hold the RTNL lock. 6986 */ 6987 static bool netdev_has_any_lower_dev(struct net_device *dev) 6988 { 6989 ASSERT_RTNL(); 6990 6991 return !list_empty(&dev->adj_list.lower); 6992 } 6993 6994 void *netdev_adjacent_get_private(struct list_head *adj_list) 6995 { 6996 struct netdev_adjacent *adj; 6997 6998 adj = list_entry(adj_list, struct netdev_adjacent, list); 6999 7000 return adj->private; 7001 } 7002 EXPORT_SYMBOL(netdev_adjacent_get_private); 7003 7004 /** 7005 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list 7006 * @dev: device 7007 * @iter: list_head ** of the current position 7008 * 7009 * Gets the next device from the dev's upper list, starting from iter 7010 * position. The caller must hold RCU read lock. 7011 */ 7012 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev, 7013 struct list_head **iter) 7014 { 7015 struct netdev_adjacent *upper; 7016 7017 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7018 7019 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7020 7021 if (&upper->list == &dev->adj_list.upper) 7022 return NULL; 7023 7024 *iter = &upper->list; 7025 7026 return upper->dev; 7027 } 7028 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu); 7029 7030 static struct net_device *__netdev_next_upper_dev(struct net_device *dev, 7031 struct list_head **iter, 7032 bool *ignore) 7033 { 7034 struct netdev_adjacent *upper; 7035 7036 upper = list_entry((*iter)->next, struct netdev_adjacent, list); 7037 7038 if (&upper->list == &dev->adj_list.upper) 7039 return NULL; 7040 7041 *iter = &upper->list; 7042 *ignore = upper->ignore; 7043 7044 return upper->dev; 7045 } 7046 7047 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev, 7048 struct list_head **iter) 7049 { 7050 struct netdev_adjacent *upper; 7051 7052 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7053 7054 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7055 7056 if (&upper->list == &dev->adj_list.upper) 7057 return NULL; 7058 7059 *iter = &upper->list; 7060 7061 return upper->dev; 7062 } 7063 7064 static int __netdev_walk_all_upper_dev(struct net_device *dev, 7065 int (*fn)(struct net_device *dev, 7066 struct netdev_nested_priv *priv), 7067 struct netdev_nested_priv *priv) 7068 { 7069 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7070 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7071 int ret, cur = 0; 7072 bool ignore; 7073 7074 now = dev; 7075 iter = &dev->adj_list.upper; 7076 7077 while (1) { 7078 if (now != dev) { 7079 ret = fn(now, priv); 7080 if (ret) 7081 return ret; 7082 } 7083 7084 next = NULL; 7085 while (1) { 7086 udev = __netdev_next_upper_dev(now, &iter, &ignore); 7087 if (!udev) 7088 break; 7089 if (ignore) 7090 continue; 7091 7092 next = udev; 7093 niter = &udev->adj_list.upper; 7094 dev_stack[cur] = now; 7095 iter_stack[cur++] = iter; 7096 break; 7097 } 7098 7099 if (!next) { 7100 if (!cur) 7101 return 0; 7102 next = dev_stack[--cur]; 7103 niter = iter_stack[cur]; 7104 } 7105 7106 now = next; 7107 iter = niter; 7108 } 7109 7110 return 0; 7111 } 7112 7113 int netdev_walk_all_upper_dev_rcu(struct net_device *dev, 7114 int (*fn)(struct net_device *dev, 7115 struct netdev_nested_priv *priv), 7116 struct netdev_nested_priv *priv) 7117 { 7118 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7119 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7120 int ret, cur = 0; 7121 7122 now = dev; 7123 iter = &dev->adj_list.upper; 7124 7125 while (1) { 7126 if (now != dev) { 7127 ret = fn(now, priv); 7128 if (ret) 7129 return ret; 7130 } 7131 7132 next = NULL; 7133 while (1) { 7134 udev = netdev_next_upper_dev_rcu(now, &iter); 7135 if (!udev) 7136 break; 7137 7138 next = udev; 7139 niter = &udev->adj_list.upper; 7140 dev_stack[cur] = now; 7141 iter_stack[cur++] = iter; 7142 break; 7143 } 7144 7145 if (!next) { 7146 if (!cur) 7147 return 0; 7148 next = dev_stack[--cur]; 7149 niter = iter_stack[cur]; 7150 } 7151 7152 now = next; 7153 iter = niter; 7154 } 7155 7156 return 0; 7157 } 7158 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu); 7159 7160 static bool __netdev_has_upper_dev(struct net_device *dev, 7161 struct net_device *upper_dev) 7162 { 7163 struct netdev_nested_priv priv = { 7164 .flags = 0, 7165 .data = (void *)upper_dev, 7166 }; 7167 7168 ASSERT_RTNL(); 7169 7170 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev, 7171 &priv); 7172 } 7173 7174 /** 7175 * netdev_lower_get_next_private - Get the next ->private from the 7176 * lower neighbour list 7177 * @dev: device 7178 * @iter: list_head ** of the current position 7179 * 7180 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7181 * list, starting from iter position. The caller must hold either hold the 7182 * RTNL lock or its own locking that guarantees that the neighbour lower 7183 * list will remain unchanged. 7184 */ 7185 void *netdev_lower_get_next_private(struct net_device *dev, 7186 struct list_head **iter) 7187 { 7188 struct netdev_adjacent *lower; 7189 7190 lower = list_entry(*iter, struct netdev_adjacent, list); 7191 7192 if (&lower->list == &dev->adj_list.lower) 7193 return NULL; 7194 7195 *iter = lower->list.next; 7196 7197 return lower->private; 7198 } 7199 EXPORT_SYMBOL(netdev_lower_get_next_private); 7200 7201 /** 7202 * netdev_lower_get_next_private_rcu - Get the next ->private from the 7203 * lower neighbour list, RCU 7204 * variant 7205 * @dev: device 7206 * @iter: list_head ** of the current position 7207 * 7208 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7209 * list, starting from iter position. The caller must hold RCU read lock. 7210 */ 7211 void *netdev_lower_get_next_private_rcu(struct net_device *dev, 7212 struct list_head **iter) 7213 { 7214 struct netdev_adjacent *lower; 7215 7216 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 7217 7218 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7219 7220 if (&lower->list == &dev->adj_list.lower) 7221 return NULL; 7222 7223 *iter = &lower->list; 7224 7225 return lower->private; 7226 } 7227 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu); 7228 7229 /** 7230 * netdev_lower_get_next - Get the next device from the lower neighbour 7231 * list 7232 * @dev: device 7233 * @iter: list_head ** of the current position 7234 * 7235 * Gets the next netdev_adjacent from the dev's lower neighbour 7236 * list, starting from iter position. The caller must hold RTNL lock or 7237 * its own locking that guarantees that the neighbour lower 7238 * list will remain unchanged. 7239 */ 7240 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter) 7241 { 7242 struct netdev_adjacent *lower; 7243 7244 lower = list_entry(*iter, struct netdev_adjacent, list); 7245 7246 if (&lower->list == &dev->adj_list.lower) 7247 return NULL; 7248 7249 *iter = lower->list.next; 7250 7251 return lower->dev; 7252 } 7253 EXPORT_SYMBOL(netdev_lower_get_next); 7254 7255 static struct net_device *netdev_next_lower_dev(struct net_device *dev, 7256 struct list_head **iter) 7257 { 7258 struct netdev_adjacent *lower; 7259 7260 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7261 7262 if (&lower->list == &dev->adj_list.lower) 7263 return NULL; 7264 7265 *iter = &lower->list; 7266 7267 return lower->dev; 7268 } 7269 7270 static struct net_device *__netdev_next_lower_dev(struct net_device *dev, 7271 struct list_head **iter, 7272 bool *ignore) 7273 { 7274 struct netdev_adjacent *lower; 7275 7276 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7277 7278 if (&lower->list == &dev->adj_list.lower) 7279 return NULL; 7280 7281 *iter = &lower->list; 7282 *ignore = lower->ignore; 7283 7284 return lower->dev; 7285 } 7286 7287 int netdev_walk_all_lower_dev(struct net_device *dev, 7288 int (*fn)(struct net_device *dev, 7289 struct netdev_nested_priv *priv), 7290 struct netdev_nested_priv *priv) 7291 { 7292 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7293 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7294 int ret, cur = 0; 7295 7296 now = dev; 7297 iter = &dev->adj_list.lower; 7298 7299 while (1) { 7300 if (now != dev) { 7301 ret = fn(now, priv); 7302 if (ret) 7303 return ret; 7304 } 7305 7306 next = NULL; 7307 while (1) { 7308 ldev = netdev_next_lower_dev(now, &iter); 7309 if (!ldev) 7310 break; 7311 7312 next = ldev; 7313 niter = &ldev->adj_list.lower; 7314 dev_stack[cur] = now; 7315 iter_stack[cur++] = iter; 7316 break; 7317 } 7318 7319 if (!next) { 7320 if (!cur) 7321 return 0; 7322 next = dev_stack[--cur]; 7323 niter = iter_stack[cur]; 7324 } 7325 7326 now = next; 7327 iter = niter; 7328 } 7329 7330 return 0; 7331 } 7332 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev); 7333 7334 static int __netdev_walk_all_lower_dev(struct net_device *dev, 7335 int (*fn)(struct net_device *dev, 7336 struct netdev_nested_priv *priv), 7337 struct netdev_nested_priv *priv) 7338 { 7339 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7340 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7341 int ret, cur = 0; 7342 bool ignore; 7343 7344 now = dev; 7345 iter = &dev->adj_list.lower; 7346 7347 while (1) { 7348 if (now != dev) { 7349 ret = fn(now, priv); 7350 if (ret) 7351 return ret; 7352 } 7353 7354 next = NULL; 7355 while (1) { 7356 ldev = __netdev_next_lower_dev(now, &iter, &ignore); 7357 if (!ldev) 7358 break; 7359 if (ignore) 7360 continue; 7361 7362 next = ldev; 7363 niter = &ldev->adj_list.lower; 7364 dev_stack[cur] = now; 7365 iter_stack[cur++] = iter; 7366 break; 7367 } 7368 7369 if (!next) { 7370 if (!cur) 7371 return 0; 7372 next = dev_stack[--cur]; 7373 niter = iter_stack[cur]; 7374 } 7375 7376 now = next; 7377 iter = niter; 7378 } 7379 7380 return 0; 7381 } 7382 7383 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev, 7384 struct list_head **iter) 7385 { 7386 struct netdev_adjacent *lower; 7387 7388 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7389 if (&lower->list == &dev->adj_list.lower) 7390 return NULL; 7391 7392 *iter = &lower->list; 7393 7394 return lower->dev; 7395 } 7396 EXPORT_SYMBOL(netdev_next_lower_dev_rcu); 7397 7398 static u8 __netdev_upper_depth(struct net_device *dev) 7399 { 7400 struct net_device *udev; 7401 struct list_head *iter; 7402 u8 max_depth = 0; 7403 bool ignore; 7404 7405 for (iter = &dev->adj_list.upper, 7406 udev = __netdev_next_upper_dev(dev, &iter, &ignore); 7407 udev; 7408 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) { 7409 if (ignore) 7410 continue; 7411 if (max_depth < udev->upper_level) 7412 max_depth = udev->upper_level; 7413 } 7414 7415 return max_depth; 7416 } 7417 7418 static u8 __netdev_lower_depth(struct net_device *dev) 7419 { 7420 struct net_device *ldev; 7421 struct list_head *iter; 7422 u8 max_depth = 0; 7423 bool ignore; 7424 7425 for (iter = &dev->adj_list.lower, 7426 ldev = __netdev_next_lower_dev(dev, &iter, &ignore); 7427 ldev; 7428 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) { 7429 if (ignore) 7430 continue; 7431 if (max_depth < ldev->lower_level) 7432 max_depth = ldev->lower_level; 7433 } 7434 7435 return max_depth; 7436 } 7437 7438 static int __netdev_update_upper_level(struct net_device *dev, 7439 struct netdev_nested_priv *__unused) 7440 { 7441 dev->upper_level = __netdev_upper_depth(dev) + 1; 7442 return 0; 7443 } 7444 7445 #ifdef CONFIG_LOCKDEP 7446 static LIST_HEAD(net_unlink_list); 7447 7448 static void net_unlink_todo(struct net_device *dev) 7449 { 7450 if (list_empty(&dev->unlink_list)) 7451 list_add_tail(&dev->unlink_list, &net_unlink_list); 7452 } 7453 #endif 7454 7455 static int __netdev_update_lower_level(struct net_device *dev, 7456 struct netdev_nested_priv *priv) 7457 { 7458 dev->lower_level = __netdev_lower_depth(dev) + 1; 7459 7460 #ifdef CONFIG_LOCKDEP 7461 if (!priv) 7462 return 0; 7463 7464 if (priv->flags & NESTED_SYNC_IMM) 7465 dev->nested_level = dev->lower_level - 1; 7466 if (priv->flags & NESTED_SYNC_TODO) 7467 net_unlink_todo(dev); 7468 #endif 7469 return 0; 7470 } 7471 7472 int netdev_walk_all_lower_dev_rcu(struct net_device *dev, 7473 int (*fn)(struct net_device *dev, 7474 struct netdev_nested_priv *priv), 7475 struct netdev_nested_priv *priv) 7476 { 7477 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7478 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7479 int ret, cur = 0; 7480 7481 now = dev; 7482 iter = &dev->adj_list.lower; 7483 7484 while (1) { 7485 if (now != dev) { 7486 ret = fn(now, priv); 7487 if (ret) 7488 return ret; 7489 } 7490 7491 next = NULL; 7492 while (1) { 7493 ldev = netdev_next_lower_dev_rcu(now, &iter); 7494 if (!ldev) 7495 break; 7496 7497 next = ldev; 7498 niter = &ldev->adj_list.lower; 7499 dev_stack[cur] = now; 7500 iter_stack[cur++] = iter; 7501 break; 7502 } 7503 7504 if (!next) { 7505 if (!cur) 7506 return 0; 7507 next = dev_stack[--cur]; 7508 niter = iter_stack[cur]; 7509 } 7510 7511 now = next; 7512 iter = niter; 7513 } 7514 7515 return 0; 7516 } 7517 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu); 7518 7519 /** 7520 * netdev_lower_get_first_private_rcu - Get the first ->private from the 7521 * lower neighbour list, RCU 7522 * variant 7523 * @dev: device 7524 * 7525 * Gets the first netdev_adjacent->private from the dev's lower neighbour 7526 * list. The caller must hold RCU read lock. 7527 */ 7528 void *netdev_lower_get_first_private_rcu(struct net_device *dev) 7529 { 7530 struct netdev_adjacent *lower; 7531 7532 lower = list_first_or_null_rcu(&dev->adj_list.lower, 7533 struct netdev_adjacent, list); 7534 if (lower) 7535 return lower->private; 7536 return NULL; 7537 } 7538 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu); 7539 7540 /** 7541 * netdev_master_upper_dev_get_rcu - Get master upper device 7542 * @dev: device 7543 * 7544 * Find a master upper device and return pointer to it or NULL in case 7545 * it's not there. The caller must hold the RCU read lock. 7546 */ 7547 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev) 7548 { 7549 struct netdev_adjacent *upper; 7550 7551 upper = list_first_or_null_rcu(&dev->adj_list.upper, 7552 struct netdev_adjacent, list); 7553 if (upper && likely(upper->master)) 7554 return upper->dev; 7555 return NULL; 7556 } 7557 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu); 7558 7559 static int netdev_adjacent_sysfs_add(struct net_device *dev, 7560 struct net_device *adj_dev, 7561 struct list_head *dev_list) 7562 { 7563 char linkname[IFNAMSIZ+7]; 7564 7565 sprintf(linkname, dev_list == &dev->adj_list.upper ? 7566 "upper_%s" : "lower_%s", adj_dev->name); 7567 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj), 7568 linkname); 7569 } 7570 static void netdev_adjacent_sysfs_del(struct net_device *dev, 7571 char *name, 7572 struct list_head *dev_list) 7573 { 7574 char linkname[IFNAMSIZ+7]; 7575 7576 sprintf(linkname, dev_list == &dev->adj_list.upper ? 7577 "upper_%s" : "lower_%s", name); 7578 sysfs_remove_link(&(dev->dev.kobj), linkname); 7579 } 7580 7581 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev, 7582 struct net_device *adj_dev, 7583 struct list_head *dev_list) 7584 { 7585 return (dev_list == &dev->adj_list.upper || 7586 dev_list == &dev->adj_list.lower) && 7587 net_eq(dev_net(dev), dev_net(adj_dev)); 7588 } 7589 7590 static int __netdev_adjacent_dev_insert(struct net_device *dev, 7591 struct net_device *adj_dev, 7592 struct list_head *dev_list, 7593 void *private, bool master) 7594 { 7595 struct netdev_adjacent *adj; 7596 int ret; 7597 7598 adj = __netdev_find_adj(adj_dev, dev_list); 7599 7600 if (adj) { 7601 adj->ref_nr += 1; 7602 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n", 7603 dev->name, adj_dev->name, adj->ref_nr); 7604 7605 return 0; 7606 } 7607 7608 adj = kmalloc(sizeof(*adj), GFP_KERNEL); 7609 if (!adj) 7610 return -ENOMEM; 7611 7612 adj->dev = adj_dev; 7613 adj->master = master; 7614 adj->ref_nr = 1; 7615 adj->private = private; 7616 adj->ignore = false; 7617 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL); 7618 7619 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n", 7620 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name); 7621 7622 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) { 7623 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list); 7624 if (ret) 7625 goto free_adj; 7626 } 7627 7628 /* Ensure that master link is always the first item in list. */ 7629 if (master) { 7630 ret = sysfs_create_link(&(dev->dev.kobj), 7631 &(adj_dev->dev.kobj), "master"); 7632 if (ret) 7633 goto remove_symlinks; 7634 7635 list_add_rcu(&adj->list, dev_list); 7636 } else { 7637 list_add_tail_rcu(&adj->list, dev_list); 7638 } 7639 7640 return 0; 7641 7642 remove_symlinks: 7643 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 7644 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 7645 free_adj: 7646 netdev_put(adj_dev, &adj->dev_tracker); 7647 kfree(adj); 7648 7649 return ret; 7650 } 7651 7652 static void __netdev_adjacent_dev_remove(struct net_device *dev, 7653 struct net_device *adj_dev, 7654 u16 ref_nr, 7655 struct list_head *dev_list) 7656 { 7657 struct netdev_adjacent *adj; 7658 7659 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n", 7660 dev->name, adj_dev->name, ref_nr); 7661 7662 adj = __netdev_find_adj(adj_dev, dev_list); 7663 7664 if (!adj) { 7665 pr_err("Adjacency does not exist for device %s from %s\n", 7666 dev->name, adj_dev->name); 7667 WARN_ON(1); 7668 return; 7669 } 7670 7671 if (adj->ref_nr > ref_nr) { 7672 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n", 7673 dev->name, adj_dev->name, ref_nr, 7674 adj->ref_nr - ref_nr); 7675 adj->ref_nr -= ref_nr; 7676 return; 7677 } 7678 7679 if (adj->master) 7680 sysfs_remove_link(&(dev->dev.kobj), "master"); 7681 7682 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 7683 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 7684 7685 list_del_rcu(&adj->list); 7686 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n", 7687 adj_dev->name, dev->name, adj_dev->name); 7688 netdev_put(adj_dev, &adj->dev_tracker); 7689 kfree_rcu(adj, rcu); 7690 } 7691 7692 static int __netdev_adjacent_dev_link_lists(struct net_device *dev, 7693 struct net_device *upper_dev, 7694 struct list_head *up_list, 7695 struct list_head *down_list, 7696 void *private, bool master) 7697 { 7698 int ret; 7699 7700 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list, 7701 private, master); 7702 if (ret) 7703 return ret; 7704 7705 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list, 7706 private, false); 7707 if (ret) { 7708 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list); 7709 return ret; 7710 } 7711 7712 return 0; 7713 } 7714 7715 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev, 7716 struct net_device *upper_dev, 7717 u16 ref_nr, 7718 struct list_head *up_list, 7719 struct list_head *down_list) 7720 { 7721 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list); 7722 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list); 7723 } 7724 7725 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev, 7726 struct net_device *upper_dev, 7727 void *private, bool master) 7728 { 7729 return __netdev_adjacent_dev_link_lists(dev, upper_dev, 7730 &dev->adj_list.upper, 7731 &upper_dev->adj_list.lower, 7732 private, master); 7733 } 7734 7735 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev, 7736 struct net_device *upper_dev) 7737 { 7738 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1, 7739 &dev->adj_list.upper, 7740 &upper_dev->adj_list.lower); 7741 } 7742 7743 static int __netdev_upper_dev_link(struct net_device *dev, 7744 struct net_device *upper_dev, bool master, 7745 void *upper_priv, void *upper_info, 7746 struct netdev_nested_priv *priv, 7747 struct netlink_ext_ack *extack) 7748 { 7749 struct netdev_notifier_changeupper_info changeupper_info = { 7750 .info = { 7751 .dev = dev, 7752 .extack = extack, 7753 }, 7754 .upper_dev = upper_dev, 7755 .master = master, 7756 .linking = true, 7757 .upper_info = upper_info, 7758 }; 7759 struct net_device *master_dev; 7760 int ret = 0; 7761 7762 ASSERT_RTNL(); 7763 7764 if (dev == upper_dev) 7765 return -EBUSY; 7766 7767 /* To prevent loops, check if dev is not upper device to upper_dev. */ 7768 if (__netdev_has_upper_dev(upper_dev, dev)) 7769 return -EBUSY; 7770 7771 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV) 7772 return -EMLINK; 7773 7774 if (!master) { 7775 if (__netdev_has_upper_dev(dev, upper_dev)) 7776 return -EEXIST; 7777 } else { 7778 master_dev = __netdev_master_upper_dev_get(dev); 7779 if (master_dev) 7780 return master_dev == upper_dev ? -EEXIST : -EBUSY; 7781 } 7782 7783 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 7784 &changeupper_info.info); 7785 ret = notifier_to_errno(ret); 7786 if (ret) 7787 return ret; 7788 7789 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv, 7790 master); 7791 if (ret) 7792 return ret; 7793 7794 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 7795 &changeupper_info.info); 7796 ret = notifier_to_errno(ret); 7797 if (ret) 7798 goto rollback; 7799 7800 __netdev_update_upper_level(dev, NULL); 7801 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 7802 7803 __netdev_update_lower_level(upper_dev, priv); 7804 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 7805 priv); 7806 7807 return 0; 7808 7809 rollback: 7810 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 7811 7812 return ret; 7813 } 7814 7815 /** 7816 * netdev_upper_dev_link - Add a link to the upper device 7817 * @dev: device 7818 * @upper_dev: new upper device 7819 * @extack: netlink extended ack 7820 * 7821 * Adds a link to device which is upper to this one. The caller must hold 7822 * the RTNL lock. On a failure a negative errno code is returned. 7823 * On success the reference counts are adjusted and the function 7824 * returns zero. 7825 */ 7826 int netdev_upper_dev_link(struct net_device *dev, 7827 struct net_device *upper_dev, 7828 struct netlink_ext_ack *extack) 7829 { 7830 struct netdev_nested_priv priv = { 7831 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 7832 .data = NULL, 7833 }; 7834 7835 return __netdev_upper_dev_link(dev, upper_dev, false, 7836 NULL, NULL, &priv, extack); 7837 } 7838 EXPORT_SYMBOL(netdev_upper_dev_link); 7839 7840 /** 7841 * netdev_master_upper_dev_link - Add a master link to the upper device 7842 * @dev: device 7843 * @upper_dev: new upper device 7844 * @upper_priv: upper device private 7845 * @upper_info: upper info to be passed down via notifier 7846 * @extack: netlink extended ack 7847 * 7848 * Adds a link to device which is upper to this one. In this case, only 7849 * one master upper device can be linked, although other non-master devices 7850 * might be linked as well. The caller must hold the RTNL lock. 7851 * On a failure a negative errno code is returned. On success the reference 7852 * counts are adjusted and the function returns zero. 7853 */ 7854 int netdev_master_upper_dev_link(struct net_device *dev, 7855 struct net_device *upper_dev, 7856 void *upper_priv, void *upper_info, 7857 struct netlink_ext_ack *extack) 7858 { 7859 struct netdev_nested_priv priv = { 7860 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 7861 .data = NULL, 7862 }; 7863 7864 return __netdev_upper_dev_link(dev, upper_dev, true, 7865 upper_priv, upper_info, &priv, extack); 7866 } 7867 EXPORT_SYMBOL(netdev_master_upper_dev_link); 7868 7869 static void __netdev_upper_dev_unlink(struct net_device *dev, 7870 struct net_device *upper_dev, 7871 struct netdev_nested_priv *priv) 7872 { 7873 struct netdev_notifier_changeupper_info changeupper_info = { 7874 .info = { 7875 .dev = dev, 7876 }, 7877 .upper_dev = upper_dev, 7878 .linking = false, 7879 }; 7880 7881 ASSERT_RTNL(); 7882 7883 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev; 7884 7885 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 7886 &changeupper_info.info); 7887 7888 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 7889 7890 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 7891 &changeupper_info.info); 7892 7893 __netdev_update_upper_level(dev, NULL); 7894 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 7895 7896 __netdev_update_lower_level(upper_dev, priv); 7897 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 7898 priv); 7899 } 7900 7901 /** 7902 * netdev_upper_dev_unlink - Removes a link to upper device 7903 * @dev: device 7904 * @upper_dev: new upper device 7905 * 7906 * Removes a link to device which is upper to this one. The caller must hold 7907 * the RTNL lock. 7908 */ 7909 void netdev_upper_dev_unlink(struct net_device *dev, 7910 struct net_device *upper_dev) 7911 { 7912 struct netdev_nested_priv priv = { 7913 .flags = NESTED_SYNC_TODO, 7914 .data = NULL, 7915 }; 7916 7917 __netdev_upper_dev_unlink(dev, upper_dev, &priv); 7918 } 7919 EXPORT_SYMBOL(netdev_upper_dev_unlink); 7920 7921 static void __netdev_adjacent_dev_set(struct net_device *upper_dev, 7922 struct net_device *lower_dev, 7923 bool val) 7924 { 7925 struct netdev_adjacent *adj; 7926 7927 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower); 7928 if (adj) 7929 adj->ignore = val; 7930 7931 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper); 7932 if (adj) 7933 adj->ignore = val; 7934 } 7935 7936 static void netdev_adjacent_dev_disable(struct net_device *upper_dev, 7937 struct net_device *lower_dev) 7938 { 7939 __netdev_adjacent_dev_set(upper_dev, lower_dev, true); 7940 } 7941 7942 static void netdev_adjacent_dev_enable(struct net_device *upper_dev, 7943 struct net_device *lower_dev) 7944 { 7945 __netdev_adjacent_dev_set(upper_dev, lower_dev, false); 7946 } 7947 7948 int netdev_adjacent_change_prepare(struct net_device *old_dev, 7949 struct net_device *new_dev, 7950 struct net_device *dev, 7951 struct netlink_ext_ack *extack) 7952 { 7953 struct netdev_nested_priv priv = { 7954 .flags = 0, 7955 .data = NULL, 7956 }; 7957 int err; 7958 7959 if (!new_dev) 7960 return 0; 7961 7962 if (old_dev && new_dev != old_dev) 7963 netdev_adjacent_dev_disable(dev, old_dev); 7964 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv, 7965 extack); 7966 if (err) { 7967 if (old_dev && new_dev != old_dev) 7968 netdev_adjacent_dev_enable(dev, old_dev); 7969 return err; 7970 } 7971 7972 return 0; 7973 } 7974 EXPORT_SYMBOL(netdev_adjacent_change_prepare); 7975 7976 void netdev_adjacent_change_commit(struct net_device *old_dev, 7977 struct net_device *new_dev, 7978 struct net_device *dev) 7979 { 7980 struct netdev_nested_priv priv = { 7981 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 7982 .data = NULL, 7983 }; 7984 7985 if (!new_dev || !old_dev) 7986 return; 7987 7988 if (new_dev == old_dev) 7989 return; 7990 7991 netdev_adjacent_dev_enable(dev, old_dev); 7992 __netdev_upper_dev_unlink(old_dev, dev, &priv); 7993 } 7994 EXPORT_SYMBOL(netdev_adjacent_change_commit); 7995 7996 void netdev_adjacent_change_abort(struct net_device *old_dev, 7997 struct net_device *new_dev, 7998 struct net_device *dev) 7999 { 8000 struct netdev_nested_priv priv = { 8001 .flags = 0, 8002 .data = NULL, 8003 }; 8004 8005 if (!new_dev) 8006 return; 8007 8008 if (old_dev && new_dev != old_dev) 8009 netdev_adjacent_dev_enable(dev, old_dev); 8010 8011 __netdev_upper_dev_unlink(new_dev, dev, &priv); 8012 } 8013 EXPORT_SYMBOL(netdev_adjacent_change_abort); 8014 8015 /** 8016 * netdev_bonding_info_change - Dispatch event about slave change 8017 * @dev: device 8018 * @bonding_info: info to dispatch 8019 * 8020 * Send NETDEV_BONDING_INFO to netdev notifiers with info. 8021 * The caller must hold the RTNL lock. 8022 */ 8023 void netdev_bonding_info_change(struct net_device *dev, 8024 struct netdev_bonding_info *bonding_info) 8025 { 8026 struct netdev_notifier_bonding_info info = { 8027 .info.dev = dev, 8028 }; 8029 8030 memcpy(&info.bonding_info, bonding_info, 8031 sizeof(struct netdev_bonding_info)); 8032 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, 8033 &info.info); 8034 } 8035 EXPORT_SYMBOL(netdev_bonding_info_change); 8036 8037 static int netdev_offload_xstats_enable_l3(struct net_device *dev, 8038 struct netlink_ext_ack *extack) 8039 { 8040 struct netdev_notifier_offload_xstats_info info = { 8041 .info.dev = dev, 8042 .info.extack = extack, 8043 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8044 }; 8045 int err; 8046 int rc; 8047 8048 dev->offload_xstats_l3 = kzalloc(sizeof(*dev->offload_xstats_l3), 8049 GFP_KERNEL); 8050 if (!dev->offload_xstats_l3) 8051 return -ENOMEM; 8052 8053 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE, 8054 NETDEV_OFFLOAD_XSTATS_DISABLE, 8055 &info.info); 8056 err = notifier_to_errno(rc); 8057 if (err) 8058 goto free_stats; 8059 8060 return 0; 8061 8062 free_stats: 8063 kfree(dev->offload_xstats_l3); 8064 dev->offload_xstats_l3 = NULL; 8065 return err; 8066 } 8067 8068 int netdev_offload_xstats_enable(struct net_device *dev, 8069 enum netdev_offload_xstats_type type, 8070 struct netlink_ext_ack *extack) 8071 { 8072 ASSERT_RTNL(); 8073 8074 if (netdev_offload_xstats_enabled(dev, type)) 8075 return -EALREADY; 8076 8077 switch (type) { 8078 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8079 return netdev_offload_xstats_enable_l3(dev, extack); 8080 } 8081 8082 WARN_ON(1); 8083 return -EINVAL; 8084 } 8085 EXPORT_SYMBOL(netdev_offload_xstats_enable); 8086 8087 static void netdev_offload_xstats_disable_l3(struct net_device *dev) 8088 { 8089 struct netdev_notifier_offload_xstats_info info = { 8090 .info.dev = dev, 8091 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8092 }; 8093 8094 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE, 8095 &info.info); 8096 kfree(dev->offload_xstats_l3); 8097 dev->offload_xstats_l3 = NULL; 8098 } 8099 8100 int netdev_offload_xstats_disable(struct net_device *dev, 8101 enum netdev_offload_xstats_type type) 8102 { 8103 ASSERT_RTNL(); 8104 8105 if (!netdev_offload_xstats_enabled(dev, type)) 8106 return -EALREADY; 8107 8108 switch (type) { 8109 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8110 netdev_offload_xstats_disable_l3(dev); 8111 return 0; 8112 } 8113 8114 WARN_ON(1); 8115 return -EINVAL; 8116 } 8117 EXPORT_SYMBOL(netdev_offload_xstats_disable); 8118 8119 static void netdev_offload_xstats_disable_all(struct net_device *dev) 8120 { 8121 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3); 8122 } 8123 8124 static struct rtnl_hw_stats64 * 8125 netdev_offload_xstats_get_ptr(const struct net_device *dev, 8126 enum netdev_offload_xstats_type type) 8127 { 8128 switch (type) { 8129 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8130 return dev->offload_xstats_l3; 8131 } 8132 8133 WARN_ON(1); 8134 return NULL; 8135 } 8136 8137 bool netdev_offload_xstats_enabled(const struct net_device *dev, 8138 enum netdev_offload_xstats_type type) 8139 { 8140 ASSERT_RTNL(); 8141 8142 return netdev_offload_xstats_get_ptr(dev, type); 8143 } 8144 EXPORT_SYMBOL(netdev_offload_xstats_enabled); 8145 8146 struct netdev_notifier_offload_xstats_ru { 8147 bool used; 8148 }; 8149 8150 struct netdev_notifier_offload_xstats_rd { 8151 struct rtnl_hw_stats64 stats; 8152 bool used; 8153 }; 8154 8155 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest, 8156 const struct rtnl_hw_stats64 *src) 8157 { 8158 dest->rx_packets += src->rx_packets; 8159 dest->tx_packets += src->tx_packets; 8160 dest->rx_bytes += src->rx_bytes; 8161 dest->tx_bytes += src->tx_bytes; 8162 dest->rx_errors += src->rx_errors; 8163 dest->tx_errors += src->tx_errors; 8164 dest->rx_dropped += src->rx_dropped; 8165 dest->tx_dropped += src->tx_dropped; 8166 dest->multicast += src->multicast; 8167 } 8168 8169 static int netdev_offload_xstats_get_used(struct net_device *dev, 8170 enum netdev_offload_xstats_type type, 8171 bool *p_used, 8172 struct netlink_ext_ack *extack) 8173 { 8174 struct netdev_notifier_offload_xstats_ru report_used = {}; 8175 struct netdev_notifier_offload_xstats_info info = { 8176 .info.dev = dev, 8177 .info.extack = extack, 8178 .type = type, 8179 .report_used = &report_used, 8180 }; 8181 int rc; 8182 8183 WARN_ON(!netdev_offload_xstats_enabled(dev, type)); 8184 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED, 8185 &info.info); 8186 *p_used = report_used.used; 8187 return notifier_to_errno(rc); 8188 } 8189 8190 static int netdev_offload_xstats_get_stats(struct net_device *dev, 8191 enum netdev_offload_xstats_type type, 8192 struct rtnl_hw_stats64 *p_stats, 8193 bool *p_used, 8194 struct netlink_ext_ack *extack) 8195 { 8196 struct netdev_notifier_offload_xstats_rd report_delta = {}; 8197 struct netdev_notifier_offload_xstats_info info = { 8198 .info.dev = dev, 8199 .info.extack = extack, 8200 .type = type, 8201 .report_delta = &report_delta, 8202 }; 8203 struct rtnl_hw_stats64 *stats; 8204 int rc; 8205 8206 stats = netdev_offload_xstats_get_ptr(dev, type); 8207 if (WARN_ON(!stats)) 8208 return -EINVAL; 8209 8210 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA, 8211 &info.info); 8212 8213 /* Cache whatever we got, even if there was an error, otherwise the 8214 * successful stats retrievals would get lost. 8215 */ 8216 netdev_hw_stats64_add(stats, &report_delta.stats); 8217 8218 if (p_stats) 8219 *p_stats = *stats; 8220 *p_used = report_delta.used; 8221 8222 return notifier_to_errno(rc); 8223 } 8224 8225 int netdev_offload_xstats_get(struct net_device *dev, 8226 enum netdev_offload_xstats_type type, 8227 struct rtnl_hw_stats64 *p_stats, bool *p_used, 8228 struct netlink_ext_ack *extack) 8229 { 8230 ASSERT_RTNL(); 8231 8232 if (p_stats) 8233 return netdev_offload_xstats_get_stats(dev, type, p_stats, 8234 p_used, extack); 8235 else 8236 return netdev_offload_xstats_get_used(dev, type, p_used, 8237 extack); 8238 } 8239 EXPORT_SYMBOL(netdev_offload_xstats_get); 8240 8241 void 8242 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta, 8243 const struct rtnl_hw_stats64 *stats) 8244 { 8245 report_delta->used = true; 8246 netdev_hw_stats64_add(&report_delta->stats, stats); 8247 } 8248 EXPORT_SYMBOL(netdev_offload_xstats_report_delta); 8249 8250 void 8251 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used) 8252 { 8253 report_used->used = true; 8254 } 8255 EXPORT_SYMBOL(netdev_offload_xstats_report_used); 8256 8257 void netdev_offload_xstats_push_delta(struct net_device *dev, 8258 enum netdev_offload_xstats_type type, 8259 const struct rtnl_hw_stats64 *p_stats) 8260 { 8261 struct rtnl_hw_stats64 *stats; 8262 8263 ASSERT_RTNL(); 8264 8265 stats = netdev_offload_xstats_get_ptr(dev, type); 8266 if (WARN_ON(!stats)) 8267 return; 8268 8269 netdev_hw_stats64_add(stats, p_stats); 8270 } 8271 EXPORT_SYMBOL(netdev_offload_xstats_push_delta); 8272 8273 /** 8274 * netdev_get_xmit_slave - Get the xmit slave of master device 8275 * @dev: device 8276 * @skb: The packet 8277 * @all_slaves: assume all the slaves are active 8278 * 8279 * The reference counters are not incremented so the caller must be 8280 * careful with locks. The caller must hold RCU lock. 8281 * %NULL is returned if no slave is found. 8282 */ 8283 8284 struct net_device *netdev_get_xmit_slave(struct net_device *dev, 8285 struct sk_buff *skb, 8286 bool all_slaves) 8287 { 8288 const struct net_device_ops *ops = dev->netdev_ops; 8289 8290 if (!ops->ndo_get_xmit_slave) 8291 return NULL; 8292 return ops->ndo_get_xmit_slave(dev, skb, all_slaves); 8293 } 8294 EXPORT_SYMBOL(netdev_get_xmit_slave); 8295 8296 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev, 8297 struct sock *sk) 8298 { 8299 const struct net_device_ops *ops = dev->netdev_ops; 8300 8301 if (!ops->ndo_sk_get_lower_dev) 8302 return NULL; 8303 return ops->ndo_sk_get_lower_dev(dev, sk); 8304 } 8305 8306 /** 8307 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket 8308 * @dev: device 8309 * @sk: the socket 8310 * 8311 * %NULL is returned if no lower device is found. 8312 */ 8313 8314 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev, 8315 struct sock *sk) 8316 { 8317 struct net_device *lower; 8318 8319 lower = netdev_sk_get_lower_dev(dev, sk); 8320 while (lower) { 8321 dev = lower; 8322 lower = netdev_sk_get_lower_dev(dev, sk); 8323 } 8324 8325 return dev; 8326 } 8327 EXPORT_SYMBOL(netdev_sk_get_lowest_dev); 8328 8329 static void netdev_adjacent_add_links(struct net_device *dev) 8330 { 8331 struct netdev_adjacent *iter; 8332 8333 struct net *net = dev_net(dev); 8334 8335 list_for_each_entry(iter, &dev->adj_list.upper, list) { 8336 if (!net_eq(net, dev_net(iter->dev))) 8337 continue; 8338 netdev_adjacent_sysfs_add(iter->dev, dev, 8339 &iter->dev->adj_list.lower); 8340 netdev_adjacent_sysfs_add(dev, iter->dev, 8341 &dev->adj_list.upper); 8342 } 8343 8344 list_for_each_entry(iter, &dev->adj_list.lower, 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.upper); 8349 netdev_adjacent_sysfs_add(dev, iter->dev, 8350 &dev->adj_list.lower); 8351 } 8352 } 8353 8354 static void netdev_adjacent_del_links(struct net_device *dev) 8355 { 8356 struct netdev_adjacent *iter; 8357 8358 struct net *net = dev_net(dev); 8359 8360 list_for_each_entry(iter, &dev->adj_list.upper, list) { 8361 if (!net_eq(net, dev_net(iter->dev))) 8362 continue; 8363 netdev_adjacent_sysfs_del(iter->dev, dev->name, 8364 &iter->dev->adj_list.lower); 8365 netdev_adjacent_sysfs_del(dev, iter->dev->name, 8366 &dev->adj_list.upper); 8367 } 8368 8369 list_for_each_entry(iter, &dev->adj_list.lower, 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.upper); 8374 netdev_adjacent_sysfs_del(dev, iter->dev->name, 8375 &dev->adj_list.lower); 8376 } 8377 } 8378 8379 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname) 8380 { 8381 struct netdev_adjacent *iter; 8382 8383 struct net *net = dev_net(dev); 8384 8385 list_for_each_entry(iter, &dev->adj_list.upper, list) { 8386 if (!net_eq(net, dev_net(iter->dev))) 8387 continue; 8388 netdev_adjacent_sysfs_del(iter->dev, oldname, 8389 &iter->dev->adj_list.lower); 8390 netdev_adjacent_sysfs_add(iter->dev, dev, 8391 &iter->dev->adj_list.lower); 8392 } 8393 8394 list_for_each_entry(iter, &dev->adj_list.lower, 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.upper); 8399 netdev_adjacent_sysfs_add(iter->dev, dev, 8400 &iter->dev->adj_list.upper); 8401 } 8402 } 8403 8404 void *netdev_lower_dev_get_private(struct net_device *dev, 8405 struct net_device *lower_dev) 8406 { 8407 struct netdev_adjacent *lower; 8408 8409 if (!lower_dev) 8410 return NULL; 8411 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower); 8412 if (!lower) 8413 return NULL; 8414 8415 return lower->private; 8416 } 8417 EXPORT_SYMBOL(netdev_lower_dev_get_private); 8418 8419 8420 /** 8421 * netdev_lower_state_changed - Dispatch event about lower device state change 8422 * @lower_dev: device 8423 * @lower_state_info: state to dispatch 8424 * 8425 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info. 8426 * The caller must hold the RTNL lock. 8427 */ 8428 void netdev_lower_state_changed(struct net_device *lower_dev, 8429 void *lower_state_info) 8430 { 8431 struct netdev_notifier_changelowerstate_info changelowerstate_info = { 8432 .info.dev = lower_dev, 8433 }; 8434 8435 ASSERT_RTNL(); 8436 changelowerstate_info.lower_state_info = lower_state_info; 8437 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, 8438 &changelowerstate_info.info); 8439 } 8440 EXPORT_SYMBOL(netdev_lower_state_changed); 8441 8442 static void dev_change_rx_flags(struct net_device *dev, int flags) 8443 { 8444 const struct net_device_ops *ops = dev->netdev_ops; 8445 8446 if (ops->ndo_change_rx_flags) 8447 ops->ndo_change_rx_flags(dev, flags); 8448 } 8449 8450 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify) 8451 { 8452 unsigned int old_flags = dev->flags; 8453 kuid_t uid; 8454 kgid_t gid; 8455 8456 ASSERT_RTNL(); 8457 8458 dev->flags |= IFF_PROMISC; 8459 dev->promiscuity += inc; 8460 if (dev->promiscuity == 0) { 8461 /* 8462 * Avoid overflow. 8463 * If inc causes overflow, untouch promisc and return error. 8464 */ 8465 if (inc < 0) 8466 dev->flags &= ~IFF_PROMISC; 8467 else { 8468 dev->promiscuity -= inc; 8469 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n"); 8470 return -EOVERFLOW; 8471 } 8472 } 8473 if (dev->flags != old_flags) { 8474 netdev_info(dev, "%s promiscuous mode\n", 8475 dev->flags & IFF_PROMISC ? "entered" : "left"); 8476 if (audit_enabled) { 8477 current_uid_gid(&uid, &gid); 8478 audit_log(audit_context(), GFP_ATOMIC, 8479 AUDIT_ANOM_PROMISCUOUS, 8480 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u", 8481 dev->name, (dev->flags & IFF_PROMISC), 8482 (old_flags & IFF_PROMISC), 8483 from_kuid(&init_user_ns, audit_get_loginuid(current)), 8484 from_kuid(&init_user_ns, uid), 8485 from_kgid(&init_user_ns, gid), 8486 audit_get_sessionid(current)); 8487 } 8488 8489 dev_change_rx_flags(dev, IFF_PROMISC); 8490 } 8491 if (notify) 8492 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL); 8493 return 0; 8494 } 8495 8496 /** 8497 * dev_set_promiscuity - update promiscuity count on a device 8498 * @dev: device 8499 * @inc: modifier 8500 * 8501 * Add or remove promiscuity from a device. While the count in the device 8502 * remains above zero the interface remains promiscuous. Once it hits zero 8503 * the device reverts back to normal filtering operation. A negative inc 8504 * value is used to drop promiscuity on the device. 8505 * Return 0 if successful or a negative errno code on error. 8506 */ 8507 int dev_set_promiscuity(struct net_device *dev, int inc) 8508 { 8509 unsigned int old_flags = dev->flags; 8510 int err; 8511 8512 err = __dev_set_promiscuity(dev, inc, true); 8513 if (err < 0) 8514 return err; 8515 if (dev->flags != old_flags) 8516 dev_set_rx_mode(dev); 8517 return err; 8518 } 8519 EXPORT_SYMBOL(dev_set_promiscuity); 8520 8521 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify) 8522 { 8523 unsigned int old_flags = dev->flags, old_gflags = dev->gflags; 8524 8525 ASSERT_RTNL(); 8526 8527 dev->flags |= IFF_ALLMULTI; 8528 dev->allmulti += inc; 8529 if (dev->allmulti == 0) { 8530 /* 8531 * Avoid overflow. 8532 * If inc causes overflow, untouch allmulti and return error. 8533 */ 8534 if (inc < 0) 8535 dev->flags &= ~IFF_ALLMULTI; 8536 else { 8537 dev->allmulti -= inc; 8538 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n"); 8539 return -EOVERFLOW; 8540 } 8541 } 8542 if (dev->flags ^ old_flags) { 8543 netdev_info(dev, "%s allmulticast mode\n", 8544 dev->flags & IFF_ALLMULTI ? "entered" : "left"); 8545 dev_change_rx_flags(dev, IFF_ALLMULTI); 8546 dev_set_rx_mode(dev); 8547 if (notify) 8548 __dev_notify_flags(dev, old_flags, 8549 dev->gflags ^ old_gflags, 0, NULL); 8550 } 8551 return 0; 8552 } 8553 8554 /** 8555 * dev_set_allmulti - update allmulti count on a device 8556 * @dev: device 8557 * @inc: modifier 8558 * 8559 * Add or remove reception of all multicast frames to a device. While the 8560 * count in the device remains above zero the interface remains listening 8561 * to all interfaces. Once it hits zero the device reverts back to normal 8562 * filtering operation. A negative @inc value is used to drop the counter 8563 * when releasing a resource needing all multicasts. 8564 * Return 0 if successful or a negative errno code on error. 8565 */ 8566 8567 int dev_set_allmulti(struct net_device *dev, int inc) 8568 { 8569 return __dev_set_allmulti(dev, inc, true); 8570 } 8571 EXPORT_SYMBOL(dev_set_allmulti); 8572 8573 /* 8574 * Upload unicast and multicast address lists to device and 8575 * configure RX filtering. When the device doesn't support unicast 8576 * filtering it is put in promiscuous mode while unicast addresses 8577 * are present. 8578 */ 8579 void __dev_set_rx_mode(struct net_device *dev) 8580 { 8581 const struct net_device_ops *ops = dev->netdev_ops; 8582 8583 /* dev_open will call this function so the list will stay sane. */ 8584 if (!(dev->flags&IFF_UP)) 8585 return; 8586 8587 if (!netif_device_present(dev)) 8588 return; 8589 8590 if (!(dev->priv_flags & IFF_UNICAST_FLT)) { 8591 /* Unicast addresses changes may only happen under the rtnl, 8592 * therefore calling __dev_set_promiscuity here is safe. 8593 */ 8594 if (!netdev_uc_empty(dev) && !dev->uc_promisc) { 8595 __dev_set_promiscuity(dev, 1, false); 8596 dev->uc_promisc = true; 8597 } else if (netdev_uc_empty(dev) && dev->uc_promisc) { 8598 __dev_set_promiscuity(dev, -1, false); 8599 dev->uc_promisc = false; 8600 } 8601 } 8602 8603 if (ops->ndo_set_rx_mode) 8604 ops->ndo_set_rx_mode(dev); 8605 } 8606 8607 void dev_set_rx_mode(struct net_device *dev) 8608 { 8609 netif_addr_lock_bh(dev); 8610 __dev_set_rx_mode(dev); 8611 netif_addr_unlock_bh(dev); 8612 } 8613 8614 /** 8615 * dev_get_flags - get flags reported to userspace 8616 * @dev: device 8617 * 8618 * Get the combination of flag bits exported through APIs to userspace. 8619 */ 8620 unsigned int dev_get_flags(const struct net_device *dev) 8621 { 8622 unsigned int flags; 8623 8624 flags = (READ_ONCE(dev->flags) & ~(IFF_PROMISC | 8625 IFF_ALLMULTI | 8626 IFF_RUNNING | 8627 IFF_LOWER_UP | 8628 IFF_DORMANT)) | 8629 (READ_ONCE(dev->gflags) & (IFF_PROMISC | 8630 IFF_ALLMULTI)); 8631 8632 if (netif_running(dev)) { 8633 if (netif_oper_up(dev)) 8634 flags |= IFF_RUNNING; 8635 if (netif_carrier_ok(dev)) 8636 flags |= IFF_LOWER_UP; 8637 if (netif_dormant(dev)) 8638 flags |= IFF_DORMANT; 8639 } 8640 8641 return flags; 8642 } 8643 EXPORT_SYMBOL(dev_get_flags); 8644 8645 int __dev_change_flags(struct net_device *dev, unsigned int flags, 8646 struct netlink_ext_ack *extack) 8647 { 8648 unsigned int old_flags = dev->flags; 8649 int ret; 8650 8651 ASSERT_RTNL(); 8652 8653 /* 8654 * Set the flags on our device. 8655 */ 8656 8657 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP | 8658 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL | 8659 IFF_AUTOMEDIA)) | 8660 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC | 8661 IFF_ALLMULTI)); 8662 8663 /* 8664 * Load in the correct multicast list now the flags have changed. 8665 */ 8666 8667 if ((old_flags ^ flags) & IFF_MULTICAST) 8668 dev_change_rx_flags(dev, IFF_MULTICAST); 8669 8670 dev_set_rx_mode(dev); 8671 8672 /* 8673 * Have we downed the interface. We handle IFF_UP ourselves 8674 * according to user attempts to set it, rather than blindly 8675 * setting it. 8676 */ 8677 8678 ret = 0; 8679 if ((old_flags ^ flags) & IFF_UP) { 8680 if (old_flags & IFF_UP) 8681 __dev_close(dev); 8682 else 8683 ret = __dev_open(dev, extack); 8684 } 8685 8686 if ((flags ^ dev->gflags) & IFF_PROMISC) { 8687 int inc = (flags & IFF_PROMISC) ? 1 : -1; 8688 unsigned int old_flags = dev->flags; 8689 8690 dev->gflags ^= IFF_PROMISC; 8691 8692 if (__dev_set_promiscuity(dev, inc, false) >= 0) 8693 if (dev->flags != old_flags) 8694 dev_set_rx_mode(dev); 8695 } 8696 8697 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI 8698 * is important. Some (broken) drivers set IFF_PROMISC, when 8699 * IFF_ALLMULTI is requested not asking us and not reporting. 8700 */ 8701 if ((flags ^ dev->gflags) & IFF_ALLMULTI) { 8702 int inc = (flags & IFF_ALLMULTI) ? 1 : -1; 8703 8704 dev->gflags ^= IFF_ALLMULTI; 8705 __dev_set_allmulti(dev, inc, false); 8706 } 8707 8708 return ret; 8709 } 8710 8711 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags, 8712 unsigned int gchanges, u32 portid, 8713 const struct nlmsghdr *nlh) 8714 { 8715 unsigned int changes = dev->flags ^ old_flags; 8716 8717 if (gchanges) 8718 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh); 8719 8720 if (changes & IFF_UP) { 8721 if (dev->flags & IFF_UP) 8722 call_netdevice_notifiers(NETDEV_UP, dev); 8723 else 8724 call_netdevice_notifiers(NETDEV_DOWN, dev); 8725 } 8726 8727 if (dev->flags & IFF_UP && 8728 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) { 8729 struct netdev_notifier_change_info change_info = { 8730 .info = { 8731 .dev = dev, 8732 }, 8733 .flags_changed = changes, 8734 }; 8735 8736 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info); 8737 } 8738 } 8739 8740 /** 8741 * dev_change_flags - change device settings 8742 * @dev: device 8743 * @flags: device state flags 8744 * @extack: netlink extended ack 8745 * 8746 * Change settings on device based state flags. The flags are 8747 * in the userspace exported format. 8748 */ 8749 int dev_change_flags(struct net_device *dev, unsigned int flags, 8750 struct netlink_ext_ack *extack) 8751 { 8752 int ret; 8753 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags; 8754 8755 ret = __dev_change_flags(dev, flags, extack); 8756 if (ret < 0) 8757 return ret; 8758 8759 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags); 8760 __dev_notify_flags(dev, old_flags, changes, 0, NULL); 8761 return ret; 8762 } 8763 EXPORT_SYMBOL(dev_change_flags); 8764 8765 int __dev_set_mtu(struct net_device *dev, int new_mtu) 8766 { 8767 const struct net_device_ops *ops = dev->netdev_ops; 8768 8769 if (ops->ndo_change_mtu) 8770 return ops->ndo_change_mtu(dev, new_mtu); 8771 8772 /* Pairs with all the lockless reads of dev->mtu in the stack */ 8773 WRITE_ONCE(dev->mtu, new_mtu); 8774 return 0; 8775 } 8776 EXPORT_SYMBOL(__dev_set_mtu); 8777 8778 int dev_validate_mtu(struct net_device *dev, int new_mtu, 8779 struct netlink_ext_ack *extack) 8780 { 8781 /* MTU must be positive, and in range */ 8782 if (new_mtu < 0 || new_mtu < dev->min_mtu) { 8783 NL_SET_ERR_MSG(extack, "mtu less than device minimum"); 8784 return -EINVAL; 8785 } 8786 8787 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) { 8788 NL_SET_ERR_MSG(extack, "mtu greater than device maximum"); 8789 return -EINVAL; 8790 } 8791 return 0; 8792 } 8793 8794 /** 8795 * dev_set_mtu_ext - Change maximum transfer unit 8796 * @dev: device 8797 * @new_mtu: new transfer unit 8798 * @extack: netlink extended ack 8799 * 8800 * Change the maximum transfer size of the network device. 8801 */ 8802 int dev_set_mtu_ext(struct net_device *dev, int new_mtu, 8803 struct netlink_ext_ack *extack) 8804 { 8805 int err, orig_mtu; 8806 8807 if (new_mtu == dev->mtu) 8808 return 0; 8809 8810 err = dev_validate_mtu(dev, new_mtu, extack); 8811 if (err) 8812 return err; 8813 8814 if (!netif_device_present(dev)) 8815 return -ENODEV; 8816 8817 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev); 8818 err = notifier_to_errno(err); 8819 if (err) 8820 return err; 8821 8822 orig_mtu = dev->mtu; 8823 err = __dev_set_mtu(dev, new_mtu); 8824 8825 if (!err) { 8826 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 8827 orig_mtu); 8828 err = notifier_to_errno(err); 8829 if (err) { 8830 /* setting mtu back and notifying everyone again, 8831 * so that they have a chance to revert changes. 8832 */ 8833 __dev_set_mtu(dev, orig_mtu); 8834 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 8835 new_mtu); 8836 } 8837 } 8838 return err; 8839 } 8840 8841 int dev_set_mtu(struct net_device *dev, int new_mtu) 8842 { 8843 struct netlink_ext_ack extack; 8844 int err; 8845 8846 memset(&extack, 0, sizeof(extack)); 8847 err = dev_set_mtu_ext(dev, new_mtu, &extack); 8848 if (err && extack._msg) 8849 net_err_ratelimited("%s: %s\n", dev->name, extack._msg); 8850 return err; 8851 } 8852 EXPORT_SYMBOL(dev_set_mtu); 8853 8854 /** 8855 * dev_change_tx_queue_len - Change TX queue length of a netdevice 8856 * @dev: device 8857 * @new_len: new tx queue length 8858 */ 8859 int dev_change_tx_queue_len(struct net_device *dev, unsigned long new_len) 8860 { 8861 unsigned int orig_len = dev->tx_queue_len; 8862 int res; 8863 8864 if (new_len != (unsigned int)new_len) 8865 return -ERANGE; 8866 8867 if (new_len != orig_len) { 8868 dev->tx_queue_len = new_len; 8869 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev); 8870 res = notifier_to_errno(res); 8871 if (res) 8872 goto err_rollback; 8873 res = dev_qdisc_change_tx_queue_len(dev); 8874 if (res) 8875 goto err_rollback; 8876 } 8877 8878 return 0; 8879 8880 err_rollback: 8881 netdev_err(dev, "refused to change device tx_queue_len\n"); 8882 dev->tx_queue_len = orig_len; 8883 return res; 8884 } 8885 8886 /** 8887 * dev_set_group - Change group this device belongs to 8888 * @dev: device 8889 * @new_group: group this device should belong to 8890 */ 8891 void dev_set_group(struct net_device *dev, int new_group) 8892 { 8893 dev->group = new_group; 8894 } 8895 8896 /** 8897 * dev_pre_changeaddr_notify - Call NETDEV_PRE_CHANGEADDR. 8898 * @dev: device 8899 * @addr: new address 8900 * @extack: netlink extended ack 8901 */ 8902 int dev_pre_changeaddr_notify(struct net_device *dev, const char *addr, 8903 struct netlink_ext_ack *extack) 8904 { 8905 struct netdev_notifier_pre_changeaddr_info info = { 8906 .info.dev = dev, 8907 .info.extack = extack, 8908 .dev_addr = addr, 8909 }; 8910 int rc; 8911 8912 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info); 8913 return notifier_to_errno(rc); 8914 } 8915 EXPORT_SYMBOL(dev_pre_changeaddr_notify); 8916 8917 /** 8918 * dev_set_mac_address - Change Media Access Control Address 8919 * @dev: device 8920 * @sa: new address 8921 * @extack: netlink extended ack 8922 * 8923 * Change the hardware (MAC) address of the device 8924 */ 8925 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa, 8926 struct netlink_ext_ack *extack) 8927 { 8928 const struct net_device_ops *ops = dev->netdev_ops; 8929 int err; 8930 8931 if (!ops->ndo_set_mac_address) 8932 return -EOPNOTSUPP; 8933 if (sa->sa_family != dev->type) 8934 return -EINVAL; 8935 if (!netif_device_present(dev)) 8936 return -ENODEV; 8937 err = dev_pre_changeaddr_notify(dev, sa->sa_data, extack); 8938 if (err) 8939 return err; 8940 if (memcmp(dev->dev_addr, sa->sa_data, dev->addr_len)) { 8941 err = ops->ndo_set_mac_address(dev, sa); 8942 if (err) 8943 return err; 8944 } 8945 dev->addr_assign_type = NET_ADDR_SET; 8946 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev); 8947 add_device_randomness(dev->dev_addr, dev->addr_len); 8948 return 0; 8949 } 8950 EXPORT_SYMBOL(dev_set_mac_address); 8951 8952 DECLARE_RWSEM(dev_addr_sem); 8953 8954 int dev_set_mac_address_user(struct net_device *dev, struct sockaddr *sa, 8955 struct netlink_ext_ack *extack) 8956 { 8957 int ret; 8958 8959 down_write(&dev_addr_sem); 8960 ret = dev_set_mac_address(dev, sa, extack); 8961 up_write(&dev_addr_sem); 8962 return ret; 8963 } 8964 EXPORT_SYMBOL(dev_set_mac_address_user); 8965 8966 int dev_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name) 8967 { 8968 size_t size = sizeof(sa->sa_data_min); 8969 struct net_device *dev; 8970 int ret = 0; 8971 8972 down_read(&dev_addr_sem); 8973 rcu_read_lock(); 8974 8975 dev = dev_get_by_name_rcu(net, dev_name); 8976 if (!dev) { 8977 ret = -ENODEV; 8978 goto unlock; 8979 } 8980 if (!dev->addr_len) 8981 memset(sa->sa_data, 0, size); 8982 else 8983 memcpy(sa->sa_data, dev->dev_addr, 8984 min_t(size_t, size, dev->addr_len)); 8985 sa->sa_family = dev->type; 8986 8987 unlock: 8988 rcu_read_unlock(); 8989 up_read(&dev_addr_sem); 8990 return ret; 8991 } 8992 EXPORT_SYMBOL(dev_get_mac_address); 8993 8994 /** 8995 * dev_change_carrier - Change device carrier 8996 * @dev: device 8997 * @new_carrier: new value 8998 * 8999 * Change device carrier 9000 */ 9001 int dev_change_carrier(struct net_device *dev, bool new_carrier) 9002 { 9003 const struct net_device_ops *ops = dev->netdev_ops; 9004 9005 if (!ops->ndo_change_carrier) 9006 return -EOPNOTSUPP; 9007 if (!netif_device_present(dev)) 9008 return -ENODEV; 9009 return ops->ndo_change_carrier(dev, new_carrier); 9010 } 9011 9012 /** 9013 * dev_get_phys_port_id - Get device physical port ID 9014 * @dev: device 9015 * @ppid: port ID 9016 * 9017 * Get device physical port ID 9018 */ 9019 int dev_get_phys_port_id(struct net_device *dev, 9020 struct netdev_phys_item_id *ppid) 9021 { 9022 const struct net_device_ops *ops = dev->netdev_ops; 9023 9024 if (!ops->ndo_get_phys_port_id) 9025 return -EOPNOTSUPP; 9026 return ops->ndo_get_phys_port_id(dev, ppid); 9027 } 9028 9029 /** 9030 * dev_get_phys_port_name - Get device physical port name 9031 * @dev: device 9032 * @name: port name 9033 * @len: limit of bytes to copy to name 9034 * 9035 * Get device physical port name 9036 */ 9037 int dev_get_phys_port_name(struct net_device *dev, 9038 char *name, size_t len) 9039 { 9040 const struct net_device_ops *ops = dev->netdev_ops; 9041 int err; 9042 9043 if (ops->ndo_get_phys_port_name) { 9044 err = ops->ndo_get_phys_port_name(dev, name, len); 9045 if (err != -EOPNOTSUPP) 9046 return err; 9047 } 9048 return devlink_compat_phys_port_name_get(dev, name, len); 9049 } 9050 9051 /** 9052 * dev_get_port_parent_id - Get the device's port parent identifier 9053 * @dev: network device 9054 * @ppid: pointer to a storage for the port's parent identifier 9055 * @recurse: allow/disallow recursion to lower devices 9056 * 9057 * Get the devices's port parent identifier 9058 */ 9059 int dev_get_port_parent_id(struct net_device *dev, 9060 struct netdev_phys_item_id *ppid, 9061 bool recurse) 9062 { 9063 const struct net_device_ops *ops = dev->netdev_ops; 9064 struct netdev_phys_item_id first = { }; 9065 struct net_device *lower_dev; 9066 struct list_head *iter; 9067 int err; 9068 9069 if (ops->ndo_get_port_parent_id) { 9070 err = ops->ndo_get_port_parent_id(dev, ppid); 9071 if (err != -EOPNOTSUPP) 9072 return err; 9073 } 9074 9075 err = devlink_compat_switch_id_get(dev, ppid); 9076 if (!recurse || err != -EOPNOTSUPP) 9077 return err; 9078 9079 netdev_for_each_lower_dev(dev, lower_dev, iter) { 9080 err = dev_get_port_parent_id(lower_dev, ppid, true); 9081 if (err) 9082 break; 9083 if (!first.id_len) 9084 first = *ppid; 9085 else if (memcmp(&first, ppid, sizeof(*ppid))) 9086 return -EOPNOTSUPP; 9087 } 9088 9089 return err; 9090 } 9091 EXPORT_SYMBOL(dev_get_port_parent_id); 9092 9093 /** 9094 * netdev_port_same_parent_id - Indicate if two network devices have 9095 * the same port parent identifier 9096 * @a: first network device 9097 * @b: second network device 9098 */ 9099 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b) 9100 { 9101 struct netdev_phys_item_id a_id = { }; 9102 struct netdev_phys_item_id b_id = { }; 9103 9104 if (dev_get_port_parent_id(a, &a_id, true) || 9105 dev_get_port_parent_id(b, &b_id, true)) 9106 return false; 9107 9108 return netdev_phys_item_id_same(&a_id, &b_id); 9109 } 9110 EXPORT_SYMBOL(netdev_port_same_parent_id); 9111 9112 /** 9113 * dev_change_proto_down - set carrier according to proto_down. 9114 * 9115 * @dev: device 9116 * @proto_down: new value 9117 */ 9118 int dev_change_proto_down(struct net_device *dev, bool proto_down) 9119 { 9120 if (!(dev->priv_flags & IFF_CHANGE_PROTO_DOWN)) 9121 return -EOPNOTSUPP; 9122 if (!netif_device_present(dev)) 9123 return -ENODEV; 9124 if (proto_down) 9125 netif_carrier_off(dev); 9126 else 9127 netif_carrier_on(dev); 9128 dev->proto_down = proto_down; 9129 return 0; 9130 } 9131 9132 /** 9133 * dev_change_proto_down_reason - proto down reason 9134 * 9135 * @dev: device 9136 * @mask: proto down mask 9137 * @value: proto down value 9138 */ 9139 void dev_change_proto_down_reason(struct net_device *dev, unsigned long mask, 9140 u32 value) 9141 { 9142 int b; 9143 9144 if (!mask) { 9145 dev->proto_down_reason = value; 9146 } else { 9147 for_each_set_bit(b, &mask, 32) { 9148 if (value & (1 << b)) 9149 dev->proto_down_reason |= BIT(b); 9150 else 9151 dev->proto_down_reason &= ~BIT(b); 9152 } 9153 } 9154 } 9155 9156 struct bpf_xdp_link { 9157 struct bpf_link link; 9158 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */ 9159 int flags; 9160 }; 9161 9162 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags) 9163 { 9164 if (flags & XDP_FLAGS_HW_MODE) 9165 return XDP_MODE_HW; 9166 if (flags & XDP_FLAGS_DRV_MODE) 9167 return XDP_MODE_DRV; 9168 if (flags & XDP_FLAGS_SKB_MODE) 9169 return XDP_MODE_SKB; 9170 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB; 9171 } 9172 9173 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode) 9174 { 9175 switch (mode) { 9176 case XDP_MODE_SKB: 9177 return generic_xdp_install; 9178 case XDP_MODE_DRV: 9179 case XDP_MODE_HW: 9180 return dev->netdev_ops->ndo_bpf; 9181 default: 9182 return NULL; 9183 } 9184 } 9185 9186 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev, 9187 enum bpf_xdp_mode mode) 9188 { 9189 return dev->xdp_state[mode].link; 9190 } 9191 9192 static struct bpf_prog *dev_xdp_prog(struct net_device *dev, 9193 enum bpf_xdp_mode mode) 9194 { 9195 struct bpf_xdp_link *link = dev_xdp_link(dev, mode); 9196 9197 if (link) 9198 return link->link.prog; 9199 return dev->xdp_state[mode].prog; 9200 } 9201 9202 u8 dev_xdp_prog_count(struct net_device *dev) 9203 { 9204 u8 count = 0; 9205 int i; 9206 9207 for (i = 0; i < __MAX_XDP_MODE; i++) 9208 if (dev->xdp_state[i].prog || dev->xdp_state[i].link) 9209 count++; 9210 return count; 9211 } 9212 EXPORT_SYMBOL_GPL(dev_xdp_prog_count); 9213 9214 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode) 9215 { 9216 struct bpf_prog *prog = dev_xdp_prog(dev, mode); 9217 9218 return prog ? prog->aux->id : 0; 9219 } 9220 9221 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode, 9222 struct bpf_xdp_link *link) 9223 { 9224 dev->xdp_state[mode].link = link; 9225 dev->xdp_state[mode].prog = NULL; 9226 } 9227 9228 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode, 9229 struct bpf_prog *prog) 9230 { 9231 dev->xdp_state[mode].link = NULL; 9232 dev->xdp_state[mode].prog = prog; 9233 } 9234 9235 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode, 9236 bpf_op_t bpf_op, struct netlink_ext_ack *extack, 9237 u32 flags, struct bpf_prog *prog) 9238 { 9239 struct netdev_bpf xdp; 9240 int err; 9241 9242 memset(&xdp, 0, sizeof(xdp)); 9243 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG; 9244 xdp.extack = extack; 9245 xdp.flags = flags; 9246 xdp.prog = prog; 9247 9248 /* Drivers assume refcnt is already incremented (i.e, prog pointer is 9249 * "moved" into driver), so they don't increment it on their own, but 9250 * they do decrement refcnt when program is detached or replaced. 9251 * Given net_device also owns link/prog, we need to bump refcnt here 9252 * to prevent drivers from underflowing it. 9253 */ 9254 if (prog) 9255 bpf_prog_inc(prog); 9256 err = bpf_op(dev, &xdp); 9257 if (err) { 9258 if (prog) 9259 bpf_prog_put(prog); 9260 return err; 9261 } 9262 9263 if (mode != XDP_MODE_HW) 9264 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog); 9265 9266 return 0; 9267 } 9268 9269 static void dev_xdp_uninstall(struct net_device *dev) 9270 { 9271 struct bpf_xdp_link *link; 9272 struct bpf_prog *prog; 9273 enum bpf_xdp_mode mode; 9274 bpf_op_t bpf_op; 9275 9276 ASSERT_RTNL(); 9277 9278 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) { 9279 prog = dev_xdp_prog(dev, mode); 9280 if (!prog) 9281 continue; 9282 9283 bpf_op = dev_xdp_bpf_op(dev, mode); 9284 if (!bpf_op) 9285 continue; 9286 9287 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 9288 9289 /* auto-detach link from net device */ 9290 link = dev_xdp_link(dev, mode); 9291 if (link) 9292 link->dev = NULL; 9293 else 9294 bpf_prog_put(prog); 9295 9296 dev_xdp_set_link(dev, mode, NULL); 9297 } 9298 } 9299 9300 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack, 9301 struct bpf_xdp_link *link, struct bpf_prog *new_prog, 9302 struct bpf_prog *old_prog, u32 flags) 9303 { 9304 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES); 9305 struct bpf_prog *cur_prog; 9306 struct net_device *upper; 9307 struct list_head *iter; 9308 enum bpf_xdp_mode mode; 9309 bpf_op_t bpf_op; 9310 int err; 9311 9312 ASSERT_RTNL(); 9313 9314 /* either link or prog attachment, never both */ 9315 if (link && (new_prog || old_prog)) 9316 return -EINVAL; 9317 /* link supports only XDP mode flags */ 9318 if (link && (flags & ~XDP_FLAGS_MODES)) { 9319 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment"); 9320 return -EINVAL; 9321 } 9322 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */ 9323 if (num_modes > 1) { 9324 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set"); 9325 return -EINVAL; 9326 } 9327 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */ 9328 if (!num_modes && dev_xdp_prog_count(dev) > 1) { 9329 NL_SET_ERR_MSG(extack, 9330 "More than one program loaded, unset mode is ambiguous"); 9331 return -EINVAL; 9332 } 9333 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */ 9334 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) { 9335 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified"); 9336 return -EINVAL; 9337 } 9338 9339 mode = dev_xdp_mode(dev, flags); 9340 /* can't replace attached link */ 9341 if (dev_xdp_link(dev, mode)) { 9342 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link"); 9343 return -EBUSY; 9344 } 9345 9346 /* don't allow if an upper device already has a program */ 9347 netdev_for_each_upper_dev_rcu(dev, upper, iter) { 9348 if (dev_xdp_prog_count(upper) > 0) { 9349 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program"); 9350 return -EEXIST; 9351 } 9352 } 9353 9354 cur_prog = dev_xdp_prog(dev, mode); 9355 /* can't replace attached prog with link */ 9356 if (link && cur_prog) { 9357 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link"); 9358 return -EBUSY; 9359 } 9360 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) { 9361 NL_SET_ERR_MSG(extack, "Active program does not match expected"); 9362 return -EEXIST; 9363 } 9364 9365 /* put effective new program into new_prog */ 9366 if (link) 9367 new_prog = link->link.prog; 9368 9369 if (new_prog) { 9370 bool offload = mode == XDP_MODE_HW; 9371 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB 9372 ? XDP_MODE_DRV : XDP_MODE_SKB; 9373 9374 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) { 9375 NL_SET_ERR_MSG(extack, "XDP program already attached"); 9376 return -EBUSY; 9377 } 9378 if (!offload && dev_xdp_prog(dev, other_mode)) { 9379 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time"); 9380 return -EEXIST; 9381 } 9382 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) { 9383 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported"); 9384 return -EINVAL; 9385 } 9386 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) { 9387 NL_SET_ERR_MSG(extack, "Program bound to different device"); 9388 return -EINVAL; 9389 } 9390 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) { 9391 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device"); 9392 return -EINVAL; 9393 } 9394 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) { 9395 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device"); 9396 return -EINVAL; 9397 } 9398 } 9399 9400 /* don't call drivers if the effective program didn't change */ 9401 if (new_prog != cur_prog) { 9402 bpf_op = dev_xdp_bpf_op(dev, mode); 9403 if (!bpf_op) { 9404 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode"); 9405 return -EOPNOTSUPP; 9406 } 9407 9408 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog); 9409 if (err) 9410 return err; 9411 } 9412 9413 if (link) 9414 dev_xdp_set_link(dev, mode, link); 9415 else 9416 dev_xdp_set_prog(dev, mode, new_prog); 9417 if (cur_prog) 9418 bpf_prog_put(cur_prog); 9419 9420 return 0; 9421 } 9422 9423 static int dev_xdp_attach_link(struct net_device *dev, 9424 struct netlink_ext_ack *extack, 9425 struct bpf_xdp_link *link) 9426 { 9427 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags); 9428 } 9429 9430 static int dev_xdp_detach_link(struct net_device *dev, 9431 struct netlink_ext_ack *extack, 9432 struct bpf_xdp_link *link) 9433 { 9434 enum bpf_xdp_mode mode; 9435 bpf_op_t bpf_op; 9436 9437 ASSERT_RTNL(); 9438 9439 mode = dev_xdp_mode(dev, link->flags); 9440 if (dev_xdp_link(dev, mode) != link) 9441 return -EINVAL; 9442 9443 bpf_op = dev_xdp_bpf_op(dev, mode); 9444 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 9445 dev_xdp_set_link(dev, mode, NULL); 9446 return 0; 9447 } 9448 9449 static void bpf_xdp_link_release(struct bpf_link *link) 9450 { 9451 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9452 9453 rtnl_lock(); 9454 9455 /* if racing with net_device's tear down, xdp_link->dev might be 9456 * already NULL, in which case link was already auto-detached 9457 */ 9458 if (xdp_link->dev) { 9459 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link)); 9460 xdp_link->dev = NULL; 9461 } 9462 9463 rtnl_unlock(); 9464 } 9465 9466 static int bpf_xdp_link_detach(struct bpf_link *link) 9467 { 9468 bpf_xdp_link_release(link); 9469 return 0; 9470 } 9471 9472 static void bpf_xdp_link_dealloc(struct bpf_link *link) 9473 { 9474 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9475 9476 kfree(xdp_link); 9477 } 9478 9479 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link, 9480 struct seq_file *seq) 9481 { 9482 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9483 u32 ifindex = 0; 9484 9485 rtnl_lock(); 9486 if (xdp_link->dev) 9487 ifindex = xdp_link->dev->ifindex; 9488 rtnl_unlock(); 9489 9490 seq_printf(seq, "ifindex:\t%u\n", ifindex); 9491 } 9492 9493 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link, 9494 struct bpf_link_info *info) 9495 { 9496 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9497 u32 ifindex = 0; 9498 9499 rtnl_lock(); 9500 if (xdp_link->dev) 9501 ifindex = xdp_link->dev->ifindex; 9502 rtnl_unlock(); 9503 9504 info->xdp.ifindex = ifindex; 9505 return 0; 9506 } 9507 9508 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog, 9509 struct bpf_prog *old_prog) 9510 { 9511 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 9512 enum bpf_xdp_mode mode; 9513 bpf_op_t bpf_op; 9514 int err = 0; 9515 9516 rtnl_lock(); 9517 9518 /* link might have been auto-released already, so fail */ 9519 if (!xdp_link->dev) { 9520 err = -ENOLINK; 9521 goto out_unlock; 9522 } 9523 9524 if (old_prog && link->prog != old_prog) { 9525 err = -EPERM; 9526 goto out_unlock; 9527 } 9528 old_prog = link->prog; 9529 if (old_prog->type != new_prog->type || 9530 old_prog->expected_attach_type != new_prog->expected_attach_type) { 9531 err = -EINVAL; 9532 goto out_unlock; 9533 } 9534 9535 if (old_prog == new_prog) { 9536 /* no-op, don't disturb drivers */ 9537 bpf_prog_put(new_prog); 9538 goto out_unlock; 9539 } 9540 9541 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags); 9542 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode); 9543 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL, 9544 xdp_link->flags, new_prog); 9545 if (err) 9546 goto out_unlock; 9547 9548 old_prog = xchg(&link->prog, new_prog); 9549 bpf_prog_put(old_prog); 9550 9551 out_unlock: 9552 rtnl_unlock(); 9553 return err; 9554 } 9555 9556 static const struct bpf_link_ops bpf_xdp_link_lops = { 9557 .release = bpf_xdp_link_release, 9558 .dealloc = bpf_xdp_link_dealloc, 9559 .detach = bpf_xdp_link_detach, 9560 .show_fdinfo = bpf_xdp_link_show_fdinfo, 9561 .fill_link_info = bpf_xdp_link_fill_link_info, 9562 .update_prog = bpf_xdp_link_update, 9563 }; 9564 9565 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog) 9566 { 9567 struct net *net = current->nsproxy->net_ns; 9568 struct bpf_link_primer link_primer; 9569 struct netlink_ext_ack extack = {}; 9570 struct bpf_xdp_link *link; 9571 struct net_device *dev; 9572 int err, fd; 9573 9574 rtnl_lock(); 9575 dev = dev_get_by_index(net, attr->link_create.target_ifindex); 9576 if (!dev) { 9577 rtnl_unlock(); 9578 return -EINVAL; 9579 } 9580 9581 link = kzalloc(sizeof(*link), GFP_USER); 9582 if (!link) { 9583 err = -ENOMEM; 9584 goto unlock; 9585 } 9586 9587 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog); 9588 link->dev = dev; 9589 link->flags = attr->link_create.flags; 9590 9591 err = bpf_link_prime(&link->link, &link_primer); 9592 if (err) { 9593 kfree(link); 9594 goto unlock; 9595 } 9596 9597 err = dev_xdp_attach_link(dev, &extack, link); 9598 rtnl_unlock(); 9599 9600 if (err) { 9601 link->dev = NULL; 9602 bpf_link_cleanup(&link_primer); 9603 trace_bpf_xdp_link_attach_failed(extack._msg); 9604 goto out_put_dev; 9605 } 9606 9607 fd = bpf_link_settle(&link_primer); 9608 /* link itself doesn't hold dev's refcnt to not complicate shutdown */ 9609 dev_put(dev); 9610 return fd; 9611 9612 unlock: 9613 rtnl_unlock(); 9614 9615 out_put_dev: 9616 dev_put(dev); 9617 return err; 9618 } 9619 9620 /** 9621 * dev_change_xdp_fd - set or clear a bpf program for a device rx path 9622 * @dev: device 9623 * @extack: netlink extended ack 9624 * @fd: new program fd or negative value to clear 9625 * @expected_fd: old program fd that userspace expects to replace or clear 9626 * @flags: xdp-related flags 9627 * 9628 * Set or clear a bpf program for a device 9629 */ 9630 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, 9631 int fd, int expected_fd, u32 flags) 9632 { 9633 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags); 9634 struct bpf_prog *new_prog = NULL, *old_prog = NULL; 9635 int err; 9636 9637 ASSERT_RTNL(); 9638 9639 if (fd >= 0) { 9640 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP, 9641 mode != XDP_MODE_SKB); 9642 if (IS_ERR(new_prog)) 9643 return PTR_ERR(new_prog); 9644 } 9645 9646 if (expected_fd >= 0) { 9647 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP, 9648 mode != XDP_MODE_SKB); 9649 if (IS_ERR(old_prog)) { 9650 err = PTR_ERR(old_prog); 9651 old_prog = NULL; 9652 goto err_out; 9653 } 9654 } 9655 9656 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags); 9657 9658 err_out: 9659 if (err && new_prog) 9660 bpf_prog_put(new_prog); 9661 if (old_prog) 9662 bpf_prog_put(old_prog); 9663 return err; 9664 } 9665 9666 /** 9667 * dev_index_reserve() - allocate an ifindex in a namespace 9668 * @net: the applicable net namespace 9669 * @ifindex: requested ifindex, pass %0 to get one allocated 9670 * 9671 * Allocate a ifindex for a new device. Caller must either use the ifindex 9672 * to store the device (via list_netdevice()) or call dev_index_release() 9673 * to give the index up. 9674 * 9675 * Return: a suitable unique value for a new device interface number or -errno. 9676 */ 9677 static int dev_index_reserve(struct net *net, u32 ifindex) 9678 { 9679 int err; 9680 9681 if (ifindex > INT_MAX) { 9682 DEBUG_NET_WARN_ON_ONCE(1); 9683 return -EINVAL; 9684 } 9685 9686 if (!ifindex) 9687 err = xa_alloc_cyclic(&net->dev_by_index, &ifindex, NULL, 9688 xa_limit_31b, &net->ifindex, GFP_KERNEL); 9689 else 9690 err = xa_insert(&net->dev_by_index, ifindex, NULL, GFP_KERNEL); 9691 if (err < 0) 9692 return err; 9693 9694 return ifindex; 9695 } 9696 9697 static void dev_index_release(struct net *net, int ifindex) 9698 { 9699 /* Expect only unused indexes, unlist_netdevice() removes the used */ 9700 WARN_ON(xa_erase(&net->dev_by_index, ifindex)); 9701 } 9702 9703 /* Delayed registration/unregisteration */ 9704 LIST_HEAD(net_todo_list); 9705 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq); 9706 atomic_t dev_unreg_count = ATOMIC_INIT(0); 9707 9708 static void net_set_todo(struct net_device *dev) 9709 { 9710 list_add_tail(&dev->todo_list, &net_todo_list); 9711 } 9712 9713 static netdev_features_t netdev_sync_upper_features(struct net_device *lower, 9714 struct net_device *upper, netdev_features_t features) 9715 { 9716 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 9717 netdev_features_t feature; 9718 int feature_bit; 9719 9720 for_each_netdev_feature(upper_disables, feature_bit) { 9721 feature = __NETIF_F_BIT(feature_bit); 9722 if (!(upper->wanted_features & feature) 9723 && (features & feature)) { 9724 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n", 9725 &feature, upper->name); 9726 features &= ~feature; 9727 } 9728 } 9729 9730 return features; 9731 } 9732 9733 static void netdev_sync_lower_features(struct net_device *upper, 9734 struct net_device *lower, netdev_features_t features) 9735 { 9736 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 9737 netdev_features_t feature; 9738 int feature_bit; 9739 9740 for_each_netdev_feature(upper_disables, feature_bit) { 9741 feature = __NETIF_F_BIT(feature_bit); 9742 if (!(features & feature) && (lower->features & feature)) { 9743 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n", 9744 &feature, lower->name); 9745 lower->wanted_features &= ~feature; 9746 __netdev_update_features(lower); 9747 9748 if (unlikely(lower->features & feature)) 9749 netdev_WARN(upper, "failed to disable %pNF on %s!\n", 9750 &feature, lower->name); 9751 else 9752 netdev_features_change(lower); 9753 } 9754 } 9755 } 9756 9757 static netdev_features_t netdev_fix_features(struct net_device *dev, 9758 netdev_features_t features) 9759 { 9760 /* Fix illegal checksum combinations */ 9761 if ((features & NETIF_F_HW_CSUM) && 9762 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) { 9763 netdev_warn(dev, "mixed HW and IP checksum settings.\n"); 9764 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM); 9765 } 9766 9767 /* TSO requires that SG is present as well. */ 9768 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) { 9769 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n"); 9770 features &= ~NETIF_F_ALL_TSO; 9771 } 9772 9773 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) && 9774 !(features & NETIF_F_IP_CSUM)) { 9775 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n"); 9776 features &= ~NETIF_F_TSO; 9777 features &= ~NETIF_F_TSO_ECN; 9778 } 9779 9780 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) && 9781 !(features & NETIF_F_IPV6_CSUM)) { 9782 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n"); 9783 features &= ~NETIF_F_TSO6; 9784 } 9785 9786 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */ 9787 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO)) 9788 features &= ~NETIF_F_TSO_MANGLEID; 9789 9790 /* TSO ECN requires that TSO is present as well. */ 9791 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN) 9792 features &= ~NETIF_F_TSO_ECN; 9793 9794 /* Software GSO depends on SG. */ 9795 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) { 9796 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n"); 9797 features &= ~NETIF_F_GSO; 9798 } 9799 9800 /* GSO partial features require GSO partial be set */ 9801 if ((features & dev->gso_partial_features) && 9802 !(features & NETIF_F_GSO_PARTIAL)) { 9803 netdev_dbg(dev, 9804 "Dropping partially supported GSO features since no GSO partial.\n"); 9805 features &= ~dev->gso_partial_features; 9806 } 9807 9808 if (!(features & NETIF_F_RXCSUM)) { 9809 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet 9810 * successfully merged by hardware must also have the 9811 * checksum verified by hardware. If the user does not 9812 * want to enable RXCSUM, logically, we should disable GRO_HW. 9813 */ 9814 if (features & NETIF_F_GRO_HW) { 9815 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n"); 9816 features &= ~NETIF_F_GRO_HW; 9817 } 9818 } 9819 9820 /* LRO/HW-GRO features cannot be combined with RX-FCS */ 9821 if (features & NETIF_F_RXFCS) { 9822 if (features & NETIF_F_LRO) { 9823 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n"); 9824 features &= ~NETIF_F_LRO; 9825 } 9826 9827 if (features & NETIF_F_GRO_HW) { 9828 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n"); 9829 features &= ~NETIF_F_GRO_HW; 9830 } 9831 } 9832 9833 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) { 9834 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n"); 9835 features &= ~NETIF_F_LRO; 9836 } 9837 9838 if (features & NETIF_F_HW_TLS_TX) { 9839 bool ip_csum = (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) == 9840 (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM); 9841 bool hw_csum = features & NETIF_F_HW_CSUM; 9842 9843 if (!ip_csum && !hw_csum) { 9844 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n"); 9845 features &= ~NETIF_F_HW_TLS_TX; 9846 } 9847 } 9848 9849 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) { 9850 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n"); 9851 features &= ~NETIF_F_HW_TLS_RX; 9852 } 9853 9854 return features; 9855 } 9856 9857 int __netdev_update_features(struct net_device *dev) 9858 { 9859 struct net_device *upper, *lower; 9860 netdev_features_t features; 9861 struct list_head *iter; 9862 int err = -1; 9863 9864 ASSERT_RTNL(); 9865 9866 features = netdev_get_wanted_features(dev); 9867 9868 if (dev->netdev_ops->ndo_fix_features) 9869 features = dev->netdev_ops->ndo_fix_features(dev, features); 9870 9871 /* driver might be less strict about feature dependencies */ 9872 features = netdev_fix_features(dev, features); 9873 9874 /* some features can't be enabled if they're off on an upper device */ 9875 netdev_for_each_upper_dev_rcu(dev, upper, iter) 9876 features = netdev_sync_upper_features(dev, upper, features); 9877 9878 if (dev->features == features) 9879 goto sync_lower; 9880 9881 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n", 9882 &dev->features, &features); 9883 9884 if (dev->netdev_ops->ndo_set_features) 9885 err = dev->netdev_ops->ndo_set_features(dev, features); 9886 else 9887 err = 0; 9888 9889 if (unlikely(err < 0)) { 9890 netdev_err(dev, 9891 "set_features() failed (%d); wanted %pNF, left %pNF\n", 9892 err, &features, &dev->features); 9893 /* return non-0 since some features might have changed and 9894 * it's better to fire a spurious notification than miss it 9895 */ 9896 return -1; 9897 } 9898 9899 sync_lower: 9900 /* some features must be disabled on lower devices when disabled 9901 * on an upper device (think: bonding master or bridge) 9902 */ 9903 netdev_for_each_lower_dev(dev, lower, iter) 9904 netdev_sync_lower_features(dev, lower, features); 9905 9906 if (!err) { 9907 netdev_features_t diff = features ^ dev->features; 9908 9909 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) { 9910 /* udp_tunnel_{get,drop}_rx_info both need 9911 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the 9912 * device, or they won't do anything. 9913 * Thus we need to update dev->features 9914 * *before* calling udp_tunnel_get_rx_info, 9915 * but *after* calling udp_tunnel_drop_rx_info. 9916 */ 9917 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) { 9918 dev->features = features; 9919 udp_tunnel_get_rx_info(dev); 9920 } else { 9921 udp_tunnel_drop_rx_info(dev); 9922 } 9923 } 9924 9925 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) { 9926 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) { 9927 dev->features = features; 9928 err |= vlan_get_rx_ctag_filter_info(dev); 9929 } else { 9930 vlan_drop_rx_ctag_filter_info(dev); 9931 } 9932 } 9933 9934 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) { 9935 if (features & NETIF_F_HW_VLAN_STAG_FILTER) { 9936 dev->features = features; 9937 err |= vlan_get_rx_stag_filter_info(dev); 9938 } else { 9939 vlan_drop_rx_stag_filter_info(dev); 9940 } 9941 } 9942 9943 dev->features = features; 9944 } 9945 9946 return err < 0 ? 0 : 1; 9947 } 9948 9949 /** 9950 * netdev_update_features - recalculate device features 9951 * @dev: the device to check 9952 * 9953 * Recalculate dev->features set and send notifications if it 9954 * has changed. Should be called after driver or hardware dependent 9955 * conditions might have changed that influence the features. 9956 */ 9957 void netdev_update_features(struct net_device *dev) 9958 { 9959 if (__netdev_update_features(dev)) 9960 netdev_features_change(dev); 9961 } 9962 EXPORT_SYMBOL(netdev_update_features); 9963 9964 /** 9965 * netdev_change_features - recalculate device features 9966 * @dev: the device to check 9967 * 9968 * Recalculate dev->features set and send notifications even 9969 * if they have not changed. Should be called instead of 9970 * netdev_update_features() if also dev->vlan_features might 9971 * have changed to allow the changes to be propagated to stacked 9972 * VLAN devices. 9973 */ 9974 void netdev_change_features(struct net_device *dev) 9975 { 9976 __netdev_update_features(dev); 9977 netdev_features_change(dev); 9978 } 9979 EXPORT_SYMBOL(netdev_change_features); 9980 9981 /** 9982 * netif_stacked_transfer_operstate - transfer operstate 9983 * @rootdev: the root or lower level device to transfer state from 9984 * @dev: the device to transfer operstate to 9985 * 9986 * Transfer operational state from root to device. This is normally 9987 * called when a stacking relationship exists between the root 9988 * device and the device(a leaf device). 9989 */ 9990 void netif_stacked_transfer_operstate(const struct net_device *rootdev, 9991 struct net_device *dev) 9992 { 9993 if (rootdev->operstate == IF_OPER_DORMANT) 9994 netif_dormant_on(dev); 9995 else 9996 netif_dormant_off(dev); 9997 9998 if (rootdev->operstate == IF_OPER_TESTING) 9999 netif_testing_on(dev); 10000 else 10001 netif_testing_off(dev); 10002 10003 if (netif_carrier_ok(rootdev)) 10004 netif_carrier_on(dev); 10005 else 10006 netif_carrier_off(dev); 10007 } 10008 EXPORT_SYMBOL(netif_stacked_transfer_operstate); 10009 10010 static int netif_alloc_rx_queues(struct net_device *dev) 10011 { 10012 unsigned int i, count = dev->num_rx_queues; 10013 struct netdev_rx_queue *rx; 10014 size_t sz = count * sizeof(*rx); 10015 int err = 0; 10016 10017 BUG_ON(count < 1); 10018 10019 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10020 if (!rx) 10021 return -ENOMEM; 10022 10023 dev->_rx = rx; 10024 10025 for (i = 0; i < count; i++) { 10026 rx[i].dev = dev; 10027 10028 /* XDP RX-queue setup */ 10029 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0); 10030 if (err < 0) 10031 goto err_rxq_info; 10032 } 10033 return 0; 10034 10035 err_rxq_info: 10036 /* Rollback successful reg's and free other resources */ 10037 while (i--) 10038 xdp_rxq_info_unreg(&rx[i].xdp_rxq); 10039 kvfree(dev->_rx); 10040 dev->_rx = NULL; 10041 return err; 10042 } 10043 10044 static void netif_free_rx_queues(struct net_device *dev) 10045 { 10046 unsigned int i, count = dev->num_rx_queues; 10047 10048 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */ 10049 if (!dev->_rx) 10050 return; 10051 10052 for (i = 0; i < count; i++) 10053 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq); 10054 10055 kvfree(dev->_rx); 10056 } 10057 10058 static void netdev_init_one_queue(struct net_device *dev, 10059 struct netdev_queue *queue, void *_unused) 10060 { 10061 /* Initialize queue lock */ 10062 spin_lock_init(&queue->_xmit_lock); 10063 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type); 10064 queue->xmit_lock_owner = -1; 10065 netdev_queue_numa_node_write(queue, NUMA_NO_NODE); 10066 queue->dev = dev; 10067 #ifdef CONFIG_BQL 10068 dql_init(&queue->dql, HZ); 10069 #endif 10070 } 10071 10072 static void netif_free_tx_queues(struct net_device *dev) 10073 { 10074 kvfree(dev->_tx); 10075 } 10076 10077 static int netif_alloc_netdev_queues(struct net_device *dev) 10078 { 10079 unsigned int count = dev->num_tx_queues; 10080 struct netdev_queue *tx; 10081 size_t sz = count * sizeof(*tx); 10082 10083 if (count < 1 || count > 0xffff) 10084 return -EINVAL; 10085 10086 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10087 if (!tx) 10088 return -ENOMEM; 10089 10090 dev->_tx = tx; 10091 10092 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL); 10093 spin_lock_init(&dev->tx_global_lock); 10094 10095 return 0; 10096 } 10097 10098 void netif_tx_stop_all_queues(struct net_device *dev) 10099 { 10100 unsigned int i; 10101 10102 for (i = 0; i < dev->num_tx_queues; i++) { 10103 struct netdev_queue *txq = netdev_get_tx_queue(dev, i); 10104 10105 netif_tx_stop_queue(txq); 10106 } 10107 } 10108 EXPORT_SYMBOL(netif_tx_stop_all_queues); 10109 10110 static int netdev_do_alloc_pcpu_stats(struct net_device *dev) 10111 { 10112 void __percpu *v; 10113 10114 /* Drivers implementing ndo_get_peer_dev must support tstat 10115 * accounting, so that skb_do_redirect() can bump the dev's 10116 * RX stats upon network namespace switch. 10117 */ 10118 if (dev->netdev_ops->ndo_get_peer_dev && 10119 dev->pcpu_stat_type != NETDEV_PCPU_STAT_TSTATS) 10120 return -EOPNOTSUPP; 10121 10122 switch (dev->pcpu_stat_type) { 10123 case NETDEV_PCPU_STAT_NONE: 10124 return 0; 10125 case NETDEV_PCPU_STAT_LSTATS: 10126 v = dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats); 10127 break; 10128 case NETDEV_PCPU_STAT_TSTATS: 10129 v = dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats); 10130 break; 10131 case NETDEV_PCPU_STAT_DSTATS: 10132 v = dev->dstats = netdev_alloc_pcpu_stats(struct pcpu_dstats); 10133 break; 10134 default: 10135 return -EINVAL; 10136 } 10137 10138 return v ? 0 : -ENOMEM; 10139 } 10140 10141 static void netdev_do_free_pcpu_stats(struct net_device *dev) 10142 { 10143 switch (dev->pcpu_stat_type) { 10144 case NETDEV_PCPU_STAT_NONE: 10145 return; 10146 case NETDEV_PCPU_STAT_LSTATS: 10147 free_percpu(dev->lstats); 10148 break; 10149 case NETDEV_PCPU_STAT_TSTATS: 10150 free_percpu(dev->tstats); 10151 break; 10152 case NETDEV_PCPU_STAT_DSTATS: 10153 free_percpu(dev->dstats); 10154 break; 10155 } 10156 } 10157 10158 /** 10159 * register_netdevice() - register a network device 10160 * @dev: device to register 10161 * 10162 * Take a prepared network device structure and make it externally accessible. 10163 * A %NETDEV_REGISTER message is sent to the netdev notifier chain. 10164 * Callers must hold the rtnl lock - you may want register_netdev() 10165 * instead of this. 10166 */ 10167 int register_netdevice(struct net_device *dev) 10168 { 10169 int ret; 10170 struct net *net = dev_net(dev); 10171 10172 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE < 10173 NETDEV_FEATURE_COUNT); 10174 BUG_ON(dev_boot_phase); 10175 ASSERT_RTNL(); 10176 10177 might_sleep(); 10178 10179 /* When net_device's are persistent, this will be fatal. */ 10180 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED); 10181 BUG_ON(!net); 10182 10183 ret = ethtool_check_ops(dev->ethtool_ops); 10184 if (ret) 10185 return ret; 10186 10187 spin_lock_init(&dev->addr_list_lock); 10188 netdev_set_addr_lockdep_class(dev); 10189 10190 ret = dev_get_valid_name(net, dev, dev->name); 10191 if (ret < 0) 10192 goto out; 10193 10194 ret = -ENOMEM; 10195 dev->name_node = netdev_name_node_head_alloc(dev); 10196 if (!dev->name_node) 10197 goto out; 10198 10199 /* Init, if this function is available */ 10200 if (dev->netdev_ops->ndo_init) { 10201 ret = dev->netdev_ops->ndo_init(dev); 10202 if (ret) { 10203 if (ret > 0) 10204 ret = -EIO; 10205 goto err_free_name; 10206 } 10207 } 10208 10209 if (((dev->hw_features | dev->features) & 10210 NETIF_F_HW_VLAN_CTAG_FILTER) && 10211 (!dev->netdev_ops->ndo_vlan_rx_add_vid || 10212 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) { 10213 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n"); 10214 ret = -EINVAL; 10215 goto err_uninit; 10216 } 10217 10218 ret = netdev_do_alloc_pcpu_stats(dev); 10219 if (ret) 10220 goto err_uninit; 10221 10222 ret = dev_index_reserve(net, dev->ifindex); 10223 if (ret < 0) 10224 goto err_free_pcpu; 10225 dev->ifindex = ret; 10226 10227 /* Transfer changeable features to wanted_features and enable 10228 * software offloads (GSO and GRO). 10229 */ 10230 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF); 10231 dev->features |= NETIF_F_SOFT_FEATURES; 10232 10233 if (dev->udp_tunnel_nic_info) { 10234 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10235 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10236 } 10237 10238 dev->wanted_features = dev->features & dev->hw_features; 10239 10240 if (!(dev->flags & IFF_LOOPBACK)) 10241 dev->hw_features |= NETIF_F_NOCACHE_COPY; 10242 10243 /* If IPv4 TCP segmentation offload is supported we should also 10244 * allow the device to enable segmenting the frame with the option 10245 * of ignoring a static IP ID value. This doesn't enable the 10246 * feature itself but allows the user to enable it later. 10247 */ 10248 if (dev->hw_features & NETIF_F_TSO) 10249 dev->hw_features |= NETIF_F_TSO_MANGLEID; 10250 if (dev->vlan_features & NETIF_F_TSO) 10251 dev->vlan_features |= NETIF_F_TSO_MANGLEID; 10252 if (dev->mpls_features & NETIF_F_TSO) 10253 dev->mpls_features |= NETIF_F_TSO_MANGLEID; 10254 if (dev->hw_enc_features & NETIF_F_TSO) 10255 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID; 10256 10257 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices. 10258 */ 10259 dev->vlan_features |= NETIF_F_HIGHDMA; 10260 10261 /* Make NETIF_F_SG inheritable to tunnel devices. 10262 */ 10263 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL; 10264 10265 /* Make NETIF_F_SG inheritable to MPLS. 10266 */ 10267 dev->mpls_features |= NETIF_F_SG; 10268 10269 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev); 10270 ret = notifier_to_errno(ret); 10271 if (ret) 10272 goto err_ifindex_release; 10273 10274 ret = netdev_register_kobject(dev); 10275 10276 WRITE_ONCE(dev->reg_state, ret ? NETREG_UNREGISTERED : NETREG_REGISTERED); 10277 10278 if (ret) 10279 goto err_uninit_notify; 10280 10281 __netdev_update_features(dev); 10282 10283 /* 10284 * Default initial state at registry is that the 10285 * device is present. 10286 */ 10287 10288 set_bit(__LINK_STATE_PRESENT, &dev->state); 10289 10290 linkwatch_init_dev(dev); 10291 10292 dev_init_scheduler(dev); 10293 10294 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL); 10295 list_netdevice(dev); 10296 10297 add_device_randomness(dev->dev_addr, dev->addr_len); 10298 10299 /* If the device has permanent device address, driver should 10300 * set dev_addr and also addr_assign_type should be set to 10301 * NET_ADDR_PERM (default value). 10302 */ 10303 if (dev->addr_assign_type == NET_ADDR_PERM) 10304 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len); 10305 10306 /* Notify protocols, that a new device appeared. */ 10307 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev); 10308 ret = notifier_to_errno(ret); 10309 if (ret) { 10310 /* Expect explicit free_netdev() on failure */ 10311 dev->needs_free_netdev = false; 10312 unregister_netdevice_queue(dev, NULL); 10313 goto out; 10314 } 10315 /* 10316 * Prevent userspace races by waiting until the network 10317 * device is fully setup before sending notifications. 10318 */ 10319 if (!dev->rtnl_link_ops || 10320 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 10321 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 10322 10323 out: 10324 return ret; 10325 10326 err_uninit_notify: 10327 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 10328 err_ifindex_release: 10329 dev_index_release(net, dev->ifindex); 10330 err_free_pcpu: 10331 netdev_do_free_pcpu_stats(dev); 10332 err_uninit: 10333 if (dev->netdev_ops->ndo_uninit) 10334 dev->netdev_ops->ndo_uninit(dev); 10335 if (dev->priv_destructor) 10336 dev->priv_destructor(dev); 10337 err_free_name: 10338 netdev_name_node_free(dev->name_node); 10339 goto out; 10340 } 10341 EXPORT_SYMBOL(register_netdevice); 10342 10343 /** 10344 * init_dummy_netdev - init a dummy network device for NAPI 10345 * @dev: device to init 10346 * 10347 * This takes a network device structure and initialize the minimum 10348 * amount of fields so it can be used to schedule NAPI polls without 10349 * registering a full blown interface. This is to be used by drivers 10350 * that need to tie several hardware interfaces to a single NAPI 10351 * poll scheduler due to HW limitations. 10352 */ 10353 void init_dummy_netdev(struct net_device *dev) 10354 { 10355 /* Clear everything. Note we don't initialize spinlocks 10356 * are they aren't supposed to be taken by any of the 10357 * NAPI code and this dummy netdev is supposed to be 10358 * only ever used for NAPI polls 10359 */ 10360 memset(dev, 0, sizeof(struct net_device)); 10361 10362 /* make sure we BUG if trying to hit standard 10363 * register/unregister code path 10364 */ 10365 dev->reg_state = NETREG_DUMMY; 10366 10367 /* NAPI wants this */ 10368 INIT_LIST_HEAD(&dev->napi_list); 10369 10370 /* a dummy interface is started by default */ 10371 set_bit(__LINK_STATE_PRESENT, &dev->state); 10372 set_bit(__LINK_STATE_START, &dev->state); 10373 10374 /* napi_busy_loop stats accounting wants this */ 10375 dev_net_set(dev, &init_net); 10376 10377 /* Note : We dont allocate pcpu_refcnt for dummy devices, 10378 * because users of this 'device' dont need to change 10379 * its refcount. 10380 */ 10381 } 10382 EXPORT_SYMBOL_GPL(init_dummy_netdev); 10383 10384 10385 /** 10386 * register_netdev - register a network device 10387 * @dev: device to register 10388 * 10389 * Take a completed network device structure and add it to the kernel 10390 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier 10391 * chain. 0 is returned on success. A negative errno code is returned 10392 * on a failure to set up the device, or if the name is a duplicate. 10393 * 10394 * This is a wrapper around register_netdevice that takes the rtnl semaphore 10395 * and expands the device name if you passed a format string to 10396 * alloc_netdev. 10397 */ 10398 int register_netdev(struct net_device *dev) 10399 { 10400 int err; 10401 10402 if (rtnl_lock_killable()) 10403 return -EINTR; 10404 err = register_netdevice(dev); 10405 rtnl_unlock(); 10406 return err; 10407 } 10408 EXPORT_SYMBOL(register_netdev); 10409 10410 int netdev_refcnt_read(const struct net_device *dev) 10411 { 10412 #ifdef CONFIG_PCPU_DEV_REFCNT 10413 int i, refcnt = 0; 10414 10415 for_each_possible_cpu(i) 10416 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i); 10417 return refcnt; 10418 #else 10419 return refcount_read(&dev->dev_refcnt); 10420 #endif 10421 } 10422 EXPORT_SYMBOL(netdev_refcnt_read); 10423 10424 int netdev_unregister_timeout_secs __read_mostly = 10; 10425 10426 #define WAIT_REFS_MIN_MSECS 1 10427 #define WAIT_REFS_MAX_MSECS 250 10428 /** 10429 * netdev_wait_allrefs_any - wait until all references are gone. 10430 * @list: list of net_devices to wait on 10431 * 10432 * This is called when unregistering network devices. 10433 * 10434 * Any protocol or device that holds a reference should register 10435 * for netdevice notification, and cleanup and put back the 10436 * reference if they receive an UNREGISTER event. 10437 * We can get stuck here if buggy protocols don't correctly 10438 * call dev_put. 10439 */ 10440 static struct net_device *netdev_wait_allrefs_any(struct list_head *list) 10441 { 10442 unsigned long rebroadcast_time, warning_time; 10443 struct net_device *dev; 10444 int wait = 0; 10445 10446 rebroadcast_time = warning_time = jiffies; 10447 10448 list_for_each_entry(dev, list, todo_list) 10449 if (netdev_refcnt_read(dev) == 1) 10450 return dev; 10451 10452 while (true) { 10453 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) { 10454 rtnl_lock(); 10455 10456 /* Rebroadcast unregister notification */ 10457 list_for_each_entry(dev, list, todo_list) 10458 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 10459 10460 __rtnl_unlock(); 10461 rcu_barrier(); 10462 rtnl_lock(); 10463 10464 list_for_each_entry(dev, list, todo_list) 10465 if (test_bit(__LINK_STATE_LINKWATCH_PENDING, 10466 &dev->state)) { 10467 /* We must not have linkwatch events 10468 * pending on unregister. If this 10469 * happens, we simply run the queue 10470 * unscheduled, resulting in a noop 10471 * for this device. 10472 */ 10473 linkwatch_run_queue(); 10474 break; 10475 } 10476 10477 __rtnl_unlock(); 10478 10479 rebroadcast_time = jiffies; 10480 } 10481 10482 if (!wait) { 10483 rcu_barrier(); 10484 wait = WAIT_REFS_MIN_MSECS; 10485 } else { 10486 msleep(wait); 10487 wait = min(wait << 1, WAIT_REFS_MAX_MSECS); 10488 } 10489 10490 list_for_each_entry(dev, list, todo_list) 10491 if (netdev_refcnt_read(dev) == 1) 10492 return dev; 10493 10494 if (time_after(jiffies, warning_time + 10495 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) { 10496 list_for_each_entry(dev, list, todo_list) { 10497 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n", 10498 dev->name, netdev_refcnt_read(dev)); 10499 ref_tracker_dir_print(&dev->refcnt_tracker, 10); 10500 } 10501 10502 warning_time = jiffies; 10503 } 10504 } 10505 } 10506 10507 /* The sequence is: 10508 * 10509 * rtnl_lock(); 10510 * ... 10511 * register_netdevice(x1); 10512 * register_netdevice(x2); 10513 * ... 10514 * unregister_netdevice(y1); 10515 * unregister_netdevice(y2); 10516 * ... 10517 * rtnl_unlock(); 10518 * free_netdev(y1); 10519 * free_netdev(y2); 10520 * 10521 * We are invoked by rtnl_unlock(). 10522 * This allows us to deal with problems: 10523 * 1) We can delete sysfs objects which invoke hotplug 10524 * without deadlocking with linkwatch via keventd. 10525 * 2) Since we run with the RTNL semaphore not held, we can sleep 10526 * safely in order to wait for the netdev refcnt to drop to zero. 10527 * 10528 * We must not return until all unregister events added during 10529 * the interval the lock was held have been completed. 10530 */ 10531 void netdev_run_todo(void) 10532 { 10533 struct net_device *dev, *tmp; 10534 struct list_head list; 10535 int cnt; 10536 #ifdef CONFIG_LOCKDEP 10537 struct list_head unlink_list; 10538 10539 list_replace_init(&net_unlink_list, &unlink_list); 10540 10541 while (!list_empty(&unlink_list)) { 10542 struct net_device *dev = list_first_entry(&unlink_list, 10543 struct net_device, 10544 unlink_list); 10545 list_del_init(&dev->unlink_list); 10546 dev->nested_level = dev->lower_level - 1; 10547 } 10548 #endif 10549 10550 /* Snapshot list, allow later requests */ 10551 list_replace_init(&net_todo_list, &list); 10552 10553 __rtnl_unlock(); 10554 10555 /* Wait for rcu callbacks to finish before next phase */ 10556 if (!list_empty(&list)) 10557 rcu_barrier(); 10558 10559 list_for_each_entry_safe(dev, tmp, &list, todo_list) { 10560 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) { 10561 netdev_WARN(dev, "run_todo but not unregistering\n"); 10562 list_del(&dev->todo_list); 10563 continue; 10564 } 10565 10566 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERED); 10567 linkwatch_sync_dev(dev); 10568 } 10569 10570 cnt = 0; 10571 while (!list_empty(&list)) { 10572 dev = netdev_wait_allrefs_any(&list); 10573 list_del(&dev->todo_list); 10574 10575 /* paranoia */ 10576 BUG_ON(netdev_refcnt_read(dev) != 1); 10577 BUG_ON(!list_empty(&dev->ptype_all)); 10578 BUG_ON(!list_empty(&dev->ptype_specific)); 10579 WARN_ON(rcu_access_pointer(dev->ip_ptr)); 10580 WARN_ON(rcu_access_pointer(dev->ip6_ptr)); 10581 10582 netdev_do_free_pcpu_stats(dev); 10583 if (dev->priv_destructor) 10584 dev->priv_destructor(dev); 10585 if (dev->needs_free_netdev) 10586 free_netdev(dev); 10587 10588 cnt++; 10589 10590 /* Free network device */ 10591 kobject_put(&dev->dev.kobj); 10592 } 10593 if (cnt && atomic_sub_and_test(cnt, &dev_unreg_count)) 10594 wake_up(&netdev_unregistering_wq); 10595 } 10596 10597 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has 10598 * all the same fields in the same order as net_device_stats, with only 10599 * the type differing, but rtnl_link_stats64 may have additional fields 10600 * at the end for newer counters. 10601 */ 10602 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64, 10603 const struct net_device_stats *netdev_stats) 10604 { 10605 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t); 10606 const atomic_long_t *src = (atomic_long_t *)netdev_stats; 10607 u64 *dst = (u64 *)stats64; 10608 10609 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64)); 10610 for (i = 0; i < n; i++) 10611 dst[i] = (unsigned long)atomic_long_read(&src[i]); 10612 /* zero out counters that only exist in rtnl_link_stats64 */ 10613 memset((char *)stats64 + n * sizeof(u64), 0, 10614 sizeof(*stats64) - n * sizeof(u64)); 10615 } 10616 EXPORT_SYMBOL(netdev_stats_to_stats64); 10617 10618 static __cold struct net_device_core_stats __percpu *netdev_core_stats_alloc( 10619 struct net_device *dev) 10620 { 10621 struct net_device_core_stats __percpu *p; 10622 10623 p = alloc_percpu_gfp(struct net_device_core_stats, 10624 GFP_ATOMIC | __GFP_NOWARN); 10625 10626 if (p && cmpxchg(&dev->core_stats, NULL, p)) 10627 free_percpu(p); 10628 10629 /* This READ_ONCE() pairs with the cmpxchg() above */ 10630 return READ_ONCE(dev->core_stats); 10631 } 10632 10633 noinline void netdev_core_stats_inc(struct net_device *dev, u32 offset) 10634 { 10635 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 10636 struct net_device_core_stats __percpu *p = READ_ONCE(dev->core_stats); 10637 unsigned long __percpu *field; 10638 10639 if (unlikely(!p)) { 10640 p = netdev_core_stats_alloc(dev); 10641 if (!p) 10642 return; 10643 } 10644 10645 field = (__force unsigned long __percpu *)((__force void *)p + offset); 10646 this_cpu_inc(*field); 10647 } 10648 EXPORT_SYMBOL_GPL(netdev_core_stats_inc); 10649 10650 /** 10651 * dev_get_stats - get network device statistics 10652 * @dev: device to get statistics from 10653 * @storage: place to store stats 10654 * 10655 * Get network statistics from device. Return @storage. 10656 * The device driver may provide its own method by setting 10657 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats; 10658 * otherwise the internal statistics structure is used. 10659 */ 10660 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev, 10661 struct rtnl_link_stats64 *storage) 10662 { 10663 const struct net_device_ops *ops = dev->netdev_ops; 10664 const struct net_device_core_stats __percpu *p; 10665 10666 if (ops->ndo_get_stats64) { 10667 memset(storage, 0, sizeof(*storage)); 10668 ops->ndo_get_stats64(dev, storage); 10669 } else if (ops->ndo_get_stats) { 10670 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev)); 10671 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_TSTATS) { 10672 dev_get_tstats64(dev, storage); 10673 } else { 10674 netdev_stats_to_stats64(storage, &dev->stats); 10675 } 10676 10677 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 10678 p = READ_ONCE(dev->core_stats); 10679 if (p) { 10680 const struct net_device_core_stats *core_stats; 10681 int i; 10682 10683 for_each_possible_cpu(i) { 10684 core_stats = per_cpu_ptr(p, i); 10685 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped); 10686 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped); 10687 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler); 10688 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped); 10689 } 10690 } 10691 return storage; 10692 } 10693 EXPORT_SYMBOL(dev_get_stats); 10694 10695 /** 10696 * dev_fetch_sw_netstats - get per-cpu network device statistics 10697 * @s: place to store stats 10698 * @netstats: per-cpu network stats to read from 10699 * 10700 * Read per-cpu network statistics and populate the related fields in @s. 10701 */ 10702 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s, 10703 const struct pcpu_sw_netstats __percpu *netstats) 10704 { 10705 int cpu; 10706 10707 for_each_possible_cpu(cpu) { 10708 u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 10709 const struct pcpu_sw_netstats *stats; 10710 unsigned int start; 10711 10712 stats = per_cpu_ptr(netstats, cpu); 10713 do { 10714 start = u64_stats_fetch_begin(&stats->syncp); 10715 rx_packets = u64_stats_read(&stats->rx_packets); 10716 rx_bytes = u64_stats_read(&stats->rx_bytes); 10717 tx_packets = u64_stats_read(&stats->tx_packets); 10718 tx_bytes = u64_stats_read(&stats->tx_bytes); 10719 } while (u64_stats_fetch_retry(&stats->syncp, start)); 10720 10721 s->rx_packets += rx_packets; 10722 s->rx_bytes += rx_bytes; 10723 s->tx_packets += tx_packets; 10724 s->tx_bytes += tx_bytes; 10725 } 10726 } 10727 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats); 10728 10729 /** 10730 * dev_get_tstats64 - ndo_get_stats64 implementation 10731 * @dev: device to get statistics from 10732 * @s: place to store stats 10733 * 10734 * Populate @s from dev->stats and dev->tstats. Can be used as 10735 * ndo_get_stats64() callback. 10736 */ 10737 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s) 10738 { 10739 netdev_stats_to_stats64(s, &dev->stats); 10740 dev_fetch_sw_netstats(s, dev->tstats); 10741 } 10742 EXPORT_SYMBOL_GPL(dev_get_tstats64); 10743 10744 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev) 10745 { 10746 struct netdev_queue *queue = dev_ingress_queue(dev); 10747 10748 #ifdef CONFIG_NET_CLS_ACT 10749 if (queue) 10750 return queue; 10751 queue = kzalloc(sizeof(*queue), GFP_KERNEL); 10752 if (!queue) 10753 return NULL; 10754 netdev_init_one_queue(dev, queue, NULL); 10755 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc); 10756 RCU_INIT_POINTER(queue->qdisc_sleeping, &noop_qdisc); 10757 rcu_assign_pointer(dev->ingress_queue, queue); 10758 #endif 10759 return queue; 10760 } 10761 10762 static const struct ethtool_ops default_ethtool_ops; 10763 10764 void netdev_set_default_ethtool_ops(struct net_device *dev, 10765 const struct ethtool_ops *ops) 10766 { 10767 if (dev->ethtool_ops == &default_ethtool_ops) 10768 dev->ethtool_ops = ops; 10769 } 10770 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops); 10771 10772 /** 10773 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default 10774 * @dev: netdev to enable the IRQ coalescing on 10775 * 10776 * Sets a conservative default for SW IRQ coalescing. Users can use 10777 * sysfs attributes to override the default values. 10778 */ 10779 void netdev_sw_irq_coalesce_default_on(struct net_device *dev) 10780 { 10781 WARN_ON(dev->reg_state == NETREG_REGISTERED); 10782 10783 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) { 10784 dev->gro_flush_timeout = 20000; 10785 dev->napi_defer_hard_irqs = 1; 10786 } 10787 } 10788 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on); 10789 10790 void netdev_freemem(struct net_device *dev) 10791 { 10792 char *addr = (char *)dev - dev->padded; 10793 10794 kvfree(addr); 10795 } 10796 10797 /** 10798 * alloc_netdev_mqs - allocate network device 10799 * @sizeof_priv: size of private data to allocate space for 10800 * @name: device name format string 10801 * @name_assign_type: origin of device name 10802 * @setup: callback to initialize device 10803 * @txqs: the number of TX subqueues to allocate 10804 * @rxqs: the number of RX subqueues to allocate 10805 * 10806 * Allocates a struct net_device with private data area for driver use 10807 * and performs basic initialization. Also allocates subqueue structs 10808 * for each queue on the device. 10809 */ 10810 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name, 10811 unsigned char name_assign_type, 10812 void (*setup)(struct net_device *), 10813 unsigned int txqs, unsigned int rxqs) 10814 { 10815 struct net_device *dev; 10816 unsigned int alloc_size; 10817 struct net_device *p; 10818 10819 BUG_ON(strlen(name) >= sizeof(dev->name)); 10820 10821 if (txqs < 1) { 10822 pr_err("alloc_netdev: Unable to allocate device with zero queues\n"); 10823 return NULL; 10824 } 10825 10826 if (rxqs < 1) { 10827 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n"); 10828 return NULL; 10829 } 10830 10831 alloc_size = sizeof(struct net_device); 10832 if (sizeof_priv) { 10833 /* ensure 32-byte alignment of private area */ 10834 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN); 10835 alloc_size += sizeof_priv; 10836 } 10837 /* ensure 32-byte alignment of whole construct */ 10838 alloc_size += NETDEV_ALIGN - 1; 10839 10840 p = kvzalloc(alloc_size, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10841 if (!p) 10842 return NULL; 10843 10844 dev = PTR_ALIGN(p, NETDEV_ALIGN); 10845 dev->padded = (char *)dev - (char *)p; 10846 10847 ref_tracker_dir_init(&dev->refcnt_tracker, 128, name); 10848 #ifdef CONFIG_PCPU_DEV_REFCNT 10849 dev->pcpu_refcnt = alloc_percpu(int); 10850 if (!dev->pcpu_refcnt) 10851 goto free_dev; 10852 __dev_hold(dev); 10853 #else 10854 refcount_set(&dev->dev_refcnt, 1); 10855 #endif 10856 10857 if (dev_addr_init(dev)) 10858 goto free_pcpu; 10859 10860 dev_mc_init(dev); 10861 dev_uc_init(dev); 10862 10863 dev_net_set(dev, &init_net); 10864 10865 dev->gso_max_size = GSO_LEGACY_MAX_SIZE; 10866 dev->xdp_zc_max_segs = 1; 10867 dev->gso_max_segs = GSO_MAX_SEGS; 10868 dev->gro_max_size = GRO_LEGACY_MAX_SIZE; 10869 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE; 10870 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE; 10871 dev->tso_max_size = TSO_LEGACY_MAX_SIZE; 10872 dev->tso_max_segs = TSO_MAX_SEGS; 10873 dev->upper_level = 1; 10874 dev->lower_level = 1; 10875 #ifdef CONFIG_LOCKDEP 10876 dev->nested_level = 0; 10877 INIT_LIST_HEAD(&dev->unlink_list); 10878 #endif 10879 10880 INIT_LIST_HEAD(&dev->napi_list); 10881 INIT_LIST_HEAD(&dev->unreg_list); 10882 INIT_LIST_HEAD(&dev->close_list); 10883 INIT_LIST_HEAD(&dev->link_watch_list); 10884 INIT_LIST_HEAD(&dev->adj_list.upper); 10885 INIT_LIST_HEAD(&dev->adj_list.lower); 10886 INIT_LIST_HEAD(&dev->ptype_all); 10887 INIT_LIST_HEAD(&dev->ptype_specific); 10888 INIT_LIST_HEAD(&dev->net_notifier_list); 10889 #ifdef CONFIG_NET_SCHED 10890 hash_init(dev->qdisc_hash); 10891 #endif 10892 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 10893 setup(dev); 10894 10895 if (!dev->tx_queue_len) { 10896 dev->priv_flags |= IFF_NO_QUEUE; 10897 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN; 10898 } 10899 10900 dev->num_tx_queues = txqs; 10901 dev->real_num_tx_queues = txqs; 10902 if (netif_alloc_netdev_queues(dev)) 10903 goto free_all; 10904 10905 dev->num_rx_queues = rxqs; 10906 dev->real_num_rx_queues = rxqs; 10907 if (netif_alloc_rx_queues(dev)) 10908 goto free_all; 10909 10910 strcpy(dev->name, name); 10911 dev->name_assign_type = name_assign_type; 10912 dev->group = INIT_NETDEV_GROUP; 10913 if (!dev->ethtool_ops) 10914 dev->ethtool_ops = &default_ethtool_ops; 10915 10916 nf_hook_netdev_init(dev); 10917 10918 return dev; 10919 10920 free_all: 10921 free_netdev(dev); 10922 return NULL; 10923 10924 free_pcpu: 10925 #ifdef CONFIG_PCPU_DEV_REFCNT 10926 free_percpu(dev->pcpu_refcnt); 10927 free_dev: 10928 #endif 10929 netdev_freemem(dev); 10930 return NULL; 10931 } 10932 EXPORT_SYMBOL(alloc_netdev_mqs); 10933 10934 /** 10935 * free_netdev - free network device 10936 * @dev: device 10937 * 10938 * This function does the last stage of destroying an allocated device 10939 * interface. The reference to the device object is released. If this 10940 * is the last reference then it will be freed.Must be called in process 10941 * context. 10942 */ 10943 void free_netdev(struct net_device *dev) 10944 { 10945 struct napi_struct *p, *n; 10946 10947 might_sleep(); 10948 10949 /* When called immediately after register_netdevice() failed the unwind 10950 * handling may still be dismantling the device. Handle that case by 10951 * deferring the free. 10952 */ 10953 if (dev->reg_state == NETREG_UNREGISTERING) { 10954 ASSERT_RTNL(); 10955 dev->needs_free_netdev = true; 10956 return; 10957 } 10958 10959 netif_free_tx_queues(dev); 10960 netif_free_rx_queues(dev); 10961 10962 kfree(rcu_dereference_protected(dev->ingress_queue, 1)); 10963 10964 /* Flush device addresses */ 10965 dev_addr_flush(dev); 10966 10967 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list) 10968 netif_napi_del(p); 10969 10970 ref_tracker_dir_exit(&dev->refcnt_tracker); 10971 #ifdef CONFIG_PCPU_DEV_REFCNT 10972 free_percpu(dev->pcpu_refcnt); 10973 dev->pcpu_refcnt = NULL; 10974 #endif 10975 free_percpu(dev->core_stats); 10976 dev->core_stats = NULL; 10977 free_percpu(dev->xdp_bulkq); 10978 dev->xdp_bulkq = NULL; 10979 10980 /* Compatibility with error handling in drivers */ 10981 if (dev->reg_state == NETREG_UNINITIALIZED) { 10982 netdev_freemem(dev); 10983 return; 10984 } 10985 10986 BUG_ON(dev->reg_state != NETREG_UNREGISTERED); 10987 WRITE_ONCE(dev->reg_state, NETREG_RELEASED); 10988 10989 /* will free via device release */ 10990 put_device(&dev->dev); 10991 } 10992 EXPORT_SYMBOL(free_netdev); 10993 10994 /** 10995 * synchronize_net - Synchronize with packet receive processing 10996 * 10997 * Wait for packets currently being received to be done. 10998 * Does not block later packets from starting. 10999 */ 11000 void synchronize_net(void) 11001 { 11002 might_sleep(); 11003 if (rtnl_is_locked()) 11004 synchronize_rcu_expedited(); 11005 else 11006 synchronize_rcu(); 11007 } 11008 EXPORT_SYMBOL(synchronize_net); 11009 11010 /** 11011 * unregister_netdevice_queue - remove device from the kernel 11012 * @dev: device 11013 * @head: list 11014 * 11015 * This function shuts down a device interface and removes it 11016 * from the kernel tables. 11017 * If head not NULL, device is queued to be unregistered later. 11018 * 11019 * Callers must hold the rtnl semaphore. You may want 11020 * unregister_netdev() instead of this. 11021 */ 11022 11023 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head) 11024 { 11025 ASSERT_RTNL(); 11026 11027 if (head) { 11028 list_move_tail(&dev->unreg_list, head); 11029 } else { 11030 LIST_HEAD(single); 11031 11032 list_add(&dev->unreg_list, &single); 11033 unregister_netdevice_many(&single); 11034 } 11035 } 11036 EXPORT_SYMBOL(unregister_netdevice_queue); 11037 11038 void unregister_netdevice_many_notify(struct list_head *head, 11039 u32 portid, const struct nlmsghdr *nlh) 11040 { 11041 struct net_device *dev, *tmp; 11042 LIST_HEAD(close_head); 11043 int cnt = 0; 11044 11045 BUG_ON(dev_boot_phase); 11046 ASSERT_RTNL(); 11047 11048 if (list_empty(head)) 11049 return; 11050 11051 list_for_each_entry_safe(dev, tmp, head, unreg_list) { 11052 /* Some devices call without registering 11053 * for initialization unwind. Remove those 11054 * devices and proceed with the remaining. 11055 */ 11056 if (dev->reg_state == NETREG_UNINITIALIZED) { 11057 pr_debug("unregister_netdevice: device %s/%p never was registered\n", 11058 dev->name, dev); 11059 11060 WARN_ON(1); 11061 list_del(&dev->unreg_list); 11062 continue; 11063 } 11064 dev->dismantle = true; 11065 BUG_ON(dev->reg_state != NETREG_REGISTERED); 11066 } 11067 11068 /* If device is running, close it first. */ 11069 list_for_each_entry(dev, head, unreg_list) 11070 list_add_tail(&dev->close_list, &close_head); 11071 dev_close_many(&close_head, true); 11072 11073 list_for_each_entry(dev, head, unreg_list) { 11074 /* And unlink it from device chain. */ 11075 unlist_netdevice(dev); 11076 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERING); 11077 } 11078 flush_all_backlogs(); 11079 11080 synchronize_net(); 11081 11082 list_for_each_entry(dev, head, unreg_list) { 11083 struct sk_buff *skb = NULL; 11084 11085 /* Shutdown queueing discipline. */ 11086 dev_shutdown(dev); 11087 dev_tcx_uninstall(dev); 11088 dev_xdp_uninstall(dev); 11089 bpf_dev_bound_netdev_unregister(dev); 11090 11091 netdev_offload_xstats_disable_all(dev); 11092 11093 /* Notify protocols, that we are about to destroy 11094 * this device. They should clean all the things. 11095 */ 11096 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11097 11098 if (!dev->rtnl_link_ops || 11099 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 11100 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0, 11101 GFP_KERNEL, NULL, 0, 11102 portid, nlh); 11103 11104 /* 11105 * Flush the unicast and multicast chains 11106 */ 11107 dev_uc_flush(dev); 11108 dev_mc_flush(dev); 11109 11110 netdev_name_node_alt_flush(dev); 11111 netdev_name_node_free(dev->name_node); 11112 11113 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11114 11115 if (dev->netdev_ops->ndo_uninit) 11116 dev->netdev_ops->ndo_uninit(dev); 11117 11118 if (skb) 11119 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh); 11120 11121 /* Notifier chain MUST detach us all upper devices. */ 11122 WARN_ON(netdev_has_any_upper_dev(dev)); 11123 WARN_ON(netdev_has_any_lower_dev(dev)); 11124 11125 /* Remove entries from kobject tree */ 11126 netdev_unregister_kobject(dev); 11127 #ifdef CONFIG_XPS 11128 /* Remove XPS queueing entries */ 11129 netif_reset_xps_queues_gt(dev, 0); 11130 #endif 11131 } 11132 11133 synchronize_net(); 11134 11135 list_for_each_entry(dev, head, unreg_list) { 11136 netdev_put(dev, &dev->dev_registered_tracker); 11137 net_set_todo(dev); 11138 cnt++; 11139 } 11140 atomic_add(cnt, &dev_unreg_count); 11141 11142 list_del(head); 11143 } 11144 11145 /** 11146 * unregister_netdevice_many - unregister many devices 11147 * @head: list of devices 11148 * 11149 * Note: As most callers use a stack allocated list_head, 11150 * we force a list_del() to make sure stack wont be corrupted later. 11151 */ 11152 void unregister_netdevice_many(struct list_head *head) 11153 { 11154 unregister_netdevice_many_notify(head, 0, NULL); 11155 } 11156 EXPORT_SYMBOL(unregister_netdevice_many); 11157 11158 /** 11159 * unregister_netdev - remove device from the kernel 11160 * @dev: device 11161 * 11162 * This function shuts down a device interface and removes it 11163 * from the kernel tables. 11164 * 11165 * This is just a wrapper for unregister_netdevice that takes 11166 * the rtnl semaphore. In general you want to use this and not 11167 * unregister_netdevice. 11168 */ 11169 void unregister_netdev(struct net_device *dev) 11170 { 11171 rtnl_lock(); 11172 unregister_netdevice(dev); 11173 rtnl_unlock(); 11174 } 11175 EXPORT_SYMBOL(unregister_netdev); 11176 11177 /** 11178 * __dev_change_net_namespace - move device to different nethost namespace 11179 * @dev: device 11180 * @net: network namespace 11181 * @pat: If not NULL name pattern to try if the current device name 11182 * is already taken in the destination network namespace. 11183 * @new_ifindex: If not zero, specifies device index in the target 11184 * namespace. 11185 * 11186 * This function shuts down a device interface and moves it 11187 * to a new network namespace. On success 0 is returned, on 11188 * a failure a netagive errno code is returned. 11189 * 11190 * Callers must hold the rtnl semaphore. 11191 */ 11192 11193 int __dev_change_net_namespace(struct net_device *dev, struct net *net, 11194 const char *pat, int new_ifindex) 11195 { 11196 struct netdev_name_node *name_node; 11197 struct net *net_old = dev_net(dev); 11198 char new_name[IFNAMSIZ] = {}; 11199 int err, new_nsid; 11200 11201 ASSERT_RTNL(); 11202 11203 /* Don't allow namespace local devices to be moved. */ 11204 err = -EINVAL; 11205 if (dev->features & NETIF_F_NETNS_LOCAL) 11206 goto out; 11207 11208 /* Ensure the device has been registrered */ 11209 if (dev->reg_state != NETREG_REGISTERED) 11210 goto out; 11211 11212 /* Get out if there is nothing todo */ 11213 err = 0; 11214 if (net_eq(net_old, net)) 11215 goto out; 11216 11217 /* Pick the destination device name, and ensure 11218 * we can use it in the destination network namespace. 11219 */ 11220 err = -EEXIST; 11221 if (netdev_name_in_use(net, dev->name)) { 11222 /* We get here if we can't use the current device name */ 11223 if (!pat) 11224 goto out; 11225 err = dev_prep_valid_name(net, dev, pat, new_name, EEXIST); 11226 if (err < 0) 11227 goto out; 11228 } 11229 /* Check that none of the altnames conflicts. */ 11230 err = -EEXIST; 11231 netdev_for_each_altname(dev, name_node) 11232 if (netdev_name_in_use(net, name_node->name)) 11233 goto out; 11234 11235 /* Check that new_ifindex isn't used yet. */ 11236 if (new_ifindex) { 11237 err = dev_index_reserve(net, new_ifindex); 11238 if (err < 0) 11239 goto out; 11240 } else { 11241 /* If there is an ifindex conflict assign a new one */ 11242 err = dev_index_reserve(net, dev->ifindex); 11243 if (err == -EBUSY) 11244 err = dev_index_reserve(net, 0); 11245 if (err < 0) 11246 goto out; 11247 new_ifindex = err; 11248 } 11249 11250 /* 11251 * And now a mini version of register_netdevice unregister_netdevice. 11252 */ 11253 11254 /* If device is running close it first. */ 11255 dev_close(dev); 11256 11257 /* And unlink it from device chain */ 11258 unlist_netdevice(dev); 11259 11260 synchronize_net(); 11261 11262 /* Shutdown queueing discipline. */ 11263 dev_shutdown(dev); 11264 11265 /* Notify protocols, that we are about to destroy 11266 * this device. They should clean all the things. 11267 * 11268 * Note that dev->reg_state stays at NETREG_REGISTERED. 11269 * This is wanted because this way 8021q and macvlan know 11270 * the device is just moving and can keep their slaves up. 11271 */ 11272 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11273 rcu_barrier(); 11274 11275 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL); 11276 11277 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid, 11278 new_ifindex); 11279 11280 /* 11281 * Flush the unicast and multicast chains 11282 */ 11283 dev_uc_flush(dev); 11284 dev_mc_flush(dev); 11285 11286 /* Send a netdev-removed uevent to the old namespace */ 11287 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE); 11288 netdev_adjacent_del_links(dev); 11289 11290 /* Move per-net netdevice notifiers that are following the netdevice */ 11291 move_netdevice_notifiers_dev_net(dev, net); 11292 11293 /* Actually switch the network namespace */ 11294 dev_net_set(dev, net); 11295 dev->ifindex = new_ifindex; 11296 11297 if (new_name[0]) /* Rename the netdev to prepared name */ 11298 strscpy(dev->name, new_name, IFNAMSIZ); 11299 11300 /* Fixup kobjects */ 11301 dev_set_uevent_suppress(&dev->dev, 1); 11302 err = device_rename(&dev->dev, dev->name); 11303 dev_set_uevent_suppress(&dev->dev, 0); 11304 WARN_ON(err); 11305 11306 /* Send a netdev-add uevent to the new namespace */ 11307 kobject_uevent(&dev->dev.kobj, KOBJ_ADD); 11308 netdev_adjacent_add_links(dev); 11309 11310 /* Adapt owner in case owning user namespace of target network 11311 * namespace is different from the original one. 11312 */ 11313 err = netdev_change_owner(dev, net_old, net); 11314 WARN_ON(err); 11315 11316 /* Add the device back in the hashes */ 11317 list_netdevice(dev); 11318 11319 /* Notify protocols, that a new device appeared. */ 11320 call_netdevice_notifiers(NETDEV_REGISTER, dev); 11321 11322 /* 11323 * Prevent userspace races by waiting until the network 11324 * device is fully setup before sending notifications. 11325 */ 11326 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 11327 11328 synchronize_net(); 11329 err = 0; 11330 out: 11331 return err; 11332 } 11333 EXPORT_SYMBOL_GPL(__dev_change_net_namespace); 11334 11335 static int dev_cpu_dead(unsigned int oldcpu) 11336 { 11337 struct sk_buff **list_skb; 11338 struct sk_buff *skb; 11339 unsigned int cpu; 11340 struct softnet_data *sd, *oldsd, *remsd = NULL; 11341 11342 local_irq_disable(); 11343 cpu = smp_processor_id(); 11344 sd = &per_cpu(softnet_data, cpu); 11345 oldsd = &per_cpu(softnet_data, oldcpu); 11346 11347 /* Find end of our completion_queue. */ 11348 list_skb = &sd->completion_queue; 11349 while (*list_skb) 11350 list_skb = &(*list_skb)->next; 11351 /* Append completion queue from offline CPU. */ 11352 *list_skb = oldsd->completion_queue; 11353 oldsd->completion_queue = NULL; 11354 11355 /* Append output queue from offline CPU. */ 11356 if (oldsd->output_queue) { 11357 *sd->output_queue_tailp = oldsd->output_queue; 11358 sd->output_queue_tailp = oldsd->output_queue_tailp; 11359 oldsd->output_queue = NULL; 11360 oldsd->output_queue_tailp = &oldsd->output_queue; 11361 } 11362 /* Append NAPI poll list from offline CPU, with one exception : 11363 * process_backlog() must be called by cpu owning percpu backlog. 11364 * We properly handle process_queue & input_pkt_queue later. 11365 */ 11366 while (!list_empty(&oldsd->poll_list)) { 11367 struct napi_struct *napi = list_first_entry(&oldsd->poll_list, 11368 struct napi_struct, 11369 poll_list); 11370 11371 list_del_init(&napi->poll_list); 11372 if (napi->poll == process_backlog) 11373 napi->state = 0; 11374 else 11375 ____napi_schedule(sd, napi); 11376 } 11377 11378 raise_softirq_irqoff(NET_TX_SOFTIRQ); 11379 local_irq_enable(); 11380 11381 #ifdef CONFIG_RPS 11382 remsd = oldsd->rps_ipi_list; 11383 oldsd->rps_ipi_list = NULL; 11384 #endif 11385 /* send out pending IPI's on offline CPU */ 11386 net_rps_send_ipi(remsd); 11387 11388 /* Process offline CPU's input_pkt_queue */ 11389 while ((skb = __skb_dequeue(&oldsd->process_queue))) { 11390 netif_rx(skb); 11391 input_queue_head_incr(oldsd); 11392 } 11393 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) { 11394 netif_rx(skb); 11395 input_queue_head_incr(oldsd); 11396 } 11397 11398 return 0; 11399 } 11400 11401 /** 11402 * netdev_increment_features - increment feature set by one 11403 * @all: current feature set 11404 * @one: new feature set 11405 * @mask: mask feature set 11406 * 11407 * Computes a new feature set after adding a device with feature set 11408 * @one to the master device with current feature set @all. Will not 11409 * enable anything that is off in @mask. Returns the new feature set. 11410 */ 11411 netdev_features_t netdev_increment_features(netdev_features_t all, 11412 netdev_features_t one, netdev_features_t mask) 11413 { 11414 if (mask & NETIF_F_HW_CSUM) 11415 mask |= NETIF_F_CSUM_MASK; 11416 mask |= NETIF_F_VLAN_CHALLENGED; 11417 11418 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask; 11419 all &= one | ~NETIF_F_ALL_FOR_ALL; 11420 11421 /* If one device supports hw checksumming, set for all. */ 11422 if (all & NETIF_F_HW_CSUM) 11423 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM); 11424 11425 return all; 11426 } 11427 EXPORT_SYMBOL(netdev_increment_features); 11428 11429 static struct hlist_head * __net_init netdev_create_hash(void) 11430 { 11431 int i; 11432 struct hlist_head *hash; 11433 11434 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL); 11435 if (hash != NULL) 11436 for (i = 0; i < NETDEV_HASHENTRIES; i++) 11437 INIT_HLIST_HEAD(&hash[i]); 11438 11439 return hash; 11440 } 11441 11442 /* Initialize per network namespace state */ 11443 static int __net_init netdev_init(struct net *net) 11444 { 11445 BUILD_BUG_ON(GRO_HASH_BUCKETS > 11446 8 * sizeof_field(struct napi_struct, gro_bitmask)); 11447 11448 INIT_LIST_HEAD(&net->dev_base_head); 11449 11450 net->dev_name_head = netdev_create_hash(); 11451 if (net->dev_name_head == NULL) 11452 goto err_name; 11453 11454 net->dev_index_head = netdev_create_hash(); 11455 if (net->dev_index_head == NULL) 11456 goto err_idx; 11457 11458 xa_init_flags(&net->dev_by_index, XA_FLAGS_ALLOC1); 11459 11460 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain); 11461 11462 return 0; 11463 11464 err_idx: 11465 kfree(net->dev_name_head); 11466 err_name: 11467 return -ENOMEM; 11468 } 11469 11470 /** 11471 * netdev_drivername - network driver for the device 11472 * @dev: network device 11473 * 11474 * Determine network driver for device. 11475 */ 11476 const char *netdev_drivername(const struct net_device *dev) 11477 { 11478 const struct device_driver *driver; 11479 const struct device *parent; 11480 const char *empty = ""; 11481 11482 parent = dev->dev.parent; 11483 if (!parent) 11484 return empty; 11485 11486 driver = parent->driver; 11487 if (driver && driver->name) 11488 return driver->name; 11489 return empty; 11490 } 11491 11492 static void __netdev_printk(const char *level, const struct net_device *dev, 11493 struct va_format *vaf) 11494 { 11495 if (dev && dev->dev.parent) { 11496 dev_printk_emit(level[1] - '0', 11497 dev->dev.parent, 11498 "%s %s %s%s: %pV", 11499 dev_driver_string(dev->dev.parent), 11500 dev_name(dev->dev.parent), 11501 netdev_name(dev), netdev_reg_state(dev), 11502 vaf); 11503 } else if (dev) { 11504 printk("%s%s%s: %pV", 11505 level, netdev_name(dev), netdev_reg_state(dev), vaf); 11506 } else { 11507 printk("%s(NULL net_device): %pV", level, vaf); 11508 } 11509 } 11510 11511 void netdev_printk(const char *level, const struct net_device *dev, 11512 const char *format, ...) 11513 { 11514 struct va_format vaf; 11515 va_list args; 11516 11517 va_start(args, format); 11518 11519 vaf.fmt = format; 11520 vaf.va = &args; 11521 11522 __netdev_printk(level, dev, &vaf); 11523 11524 va_end(args); 11525 } 11526 EXPORT_SYMBOL(netdev_printk); 11527 11528 #define define_netdev_printk_level(func, level) \ 11529 void func(const struct net_device *dev, const char *fmt, ...) \ 11530 { \ 11531 struct va_format vaf; \ 11532 va_list args; \ 11533 \ 11534 va_start(args, fmt); \ 11535 \ 11536 vaf.fmt = fmt; \ 11537 vaf.va = &args; \ 11538 \ 11539 __netdev_printk(level, dev, &vaf); \ 11540 \ 11541 va_end(args); \ 11542 } \ 11543 EXPORT_SYMBOL(func); 11544 11545 define_netdev_printk_level(netdev_emerg, KERN_EMERG); 11546 define_netdev_printk_level(netdev_alert, KERN_ALERT); 11547 define_netdev_printk_level(netdev_crit, KERN_CRIT); 11548 define_netdev_printk_level(netdev_err, KERN_ERR); 11549 define_netdev_printk_level(netdev_warn, KERN_WARNING); 11550 define_netdev_printk_level(netdev_notice, KERN_NOTICE); 11551 define_netdev_printk_level(netdev_info, KERN_INFO); 11552 11553 static void __net_exit netdev_exit(struct net *net) 11554 { 11555 kfree(net->dev_name_head); 11556 kfree(net->dev_index_head); 11557 xa_destroy(&net->dev_by_index); 11558 if (net != &init_net) 11559 WARN_ON_ONCE(!list_empty(&net->dev_base_head)); 11560 } 11561 11562 static struct pernet_operations __net_initdata netdev_net_ops = { 11563 .init = netdev_init, 11564 .exit = netdev_exit, 11565 }; 11566 11567 static void __net_exit default_device_exit_net(struct net *net) 11568 { 11569 struct netdev_name_node *name_node, *tmp; 11570 struct net_device *dev, *aux; 11571 /* 11572 * Push all migratable network devices back to the 11573 * initial network namespace 11574 */ 11575 ASSERT_RTNL(); 11576 for_each_netdev_safe(net, dev, aux) { 11577 int err; 11578 char fb_name[IFNAMSIZ]; 11579 11580 /* Ignore unmoveable devices (i.e. loopback) */ 11581 if (dev->features & NETIF_F_NETNS_LOCAL) 11582 continue; 11583 11584 /* Leave virtual devices for the generic cleanup */ 11585 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund) 11586 continue; 11587 11588 /* Push remaining network devices to init_net */ 11589 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex); 11590 if (netdev_name_in_use(&init_net, fb_name)) 11591 snprintf(fb_name, IFNAMSIZ, "dev%%d"); 11592 11593 netdev_for_each_altname_safe(dev, name_node, tmp) 11594 if (netdev_name_in_use(&init_net, name_node->name)) 11595 __netdev_name_node_alt_destroy(name_node); 11596 11597 err = dev_change_net_namespace(dev, &init_net, fb_name); 11598 if (err) { 11599 pr_emerg("%s: failed to move %s to init_net: %d\n", 11600 __func__, dev->name, err); 11601 BUG(); 11602 } 11603 } 11604 } 11605 11606 static void __net_exit default_device_exit_batch(struct list_head *net_list) 11607 { 11608 /* At exit all network devices most be removed from a network 11609 * namespace. Do this in the reverse order of registration. 11610 * Do this across as many network namespaces as possible to 11611 * improve batching efficiency. 11612 */ 11613 struct net_device *dev; 11614 struct net *net; 11615 LIST_HEAD(dev_kill_list); 11616 11617 rtnl_lock(); 11618 list_for_each_entry(net, net_list, exit_list) { 11619 default_device_exit_net(net); 11620 cond_resched(); 11621 } 11622 11623 list_for_each_entry(net, net_list, exit_list) { 11624 for_each_netdev_reverse(net, dev) { 11625 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink) 11626 dev->rtnl_link_ops->dellink(dev, &dev_kill_list); 11627 else 11628 unregister_netdevice_queue(dev, &dev_kill_list); 11629 } 11630 } 11631 unregister_netdevice_many(&dev_kill_list); 11632 rtnl_unlock(); 11633 } 11634 11635 static struct pernet_operations __net_initdata default_device_ops = { 11636 .exit_batch = default_device_exit_batch, 11637 }; 11638 11639 static void __init net_dev_struct_check(void) 11640 { 11641 /* TX read-mostly hotpath */ 11642 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, priv_flags); 11643 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, netdev_ops); 11644 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, header_ops); 11645 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, _tx); 11646 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, real_num_tx_queues); 11647 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_size); 11648 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_ipv4_max_size); 11649 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_segs); 11650 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_partial_features); 11651 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, num_tc); 11652 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, mtu); 11653 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, needed_headroom); 11654 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tc_to_txq); 11655 #ifdef CONFIG_XPS 11656 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, xps_maps); 11657 #endif 11658 #ifdef CONFIG_NETFILTER_EGRESS 11659 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, nf_hooks_egress); 11660 #endif 11661 #ifdef CONFIG_NET_XGRESS 11662 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tcx_egress); 11663 #endif 11664 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_tx, 160); 11665 11666 /* TXRX read-mostly hotpath */ 11667 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, lstats); 11668 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, flags); 11669 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, hard_header_len); 11670 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, features); 11671 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, ip6_ptr); 11672 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_txrx, 38); 11673 11674 /* RX read-mostly hotpath */ 11675 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ptype_specific); 11676 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ifindex); 11677 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, real_num_rx_queues); 11678 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, _rx); 11679 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_flush_timeout); 11680 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, napi_defer_hard_irqs); 11681 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_max_size); 11682 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_ipv4_max_size); 11683 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler); 11684 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler_data); 11685 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, nd_net); 11686 #ifdef CONFIG_NETPOLL 11687 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, npinfo); 11688 #endif 11689 #ifdef CONFIG_NET_XGRESS 11690 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, tcx_ingress); 11691 #endif 11692 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_rx, 104); 11693 } 11694 11695 /* 11696 * Initialize the DEV module. At boot time this walks the device list and 11697 * unhooks any devices that fail to initialise (normally hardware not 11698 * present) and leaves us with a valid list of present and active devices. 11699 * 11700 */ 11701 11702 /* We allocate 256 pages for each CPU if PAGE_SHIFT is 12 */ 11703 #define SYSTEM_PERCPU_PAGE_POOL_SIZE ((1 << 20) / PAGE_SIZE) 11704 11705 static int net_page_pool_create(int cpuid) 11706 { 11707 #if IS_ENABLED(CONFIG_PAGE_POOL) 11708 struct page_pool_params page_pool_params = { 11709 .pool_size = SYSTEM_PERCPU_PAGE_POOL_SIZE, 11710 .flags = PP_FLAG_SYSTEM_POOL, 11711 .nid = NUMA_NO_NODE, 11712 }; 11713 struct page_pool *pp_ptr; 11714 11715 pp_ptr = page_pool_create_percpu(&page_pool_params, cpuid); 11716 if (IS_ERR(pp_ptr)) 11717 return -ENOMEM; 11718 11719 per_cpu(system_page_pool, cpuid) = pp_ptr; 11720 #endif 11721 return 0; 11722 } 11723 11724 /* 11725 * This is called single threaded during boot, so no need 11726 * to take the rtnl semaphore. 11727 */ 11728 static int __init net_dev_init(void) 11729 { 11730 int i, rc = -ENOMEM; 11731 11732 BUG_ON(!dev_boot_phase); 11733 11734 net_dev_struct_check(); 11735 11736 if (dev_proc_init()) 11737 goto out; 11738 11739 if (netdev_kobject_init()) 11740 goto out; 11741 11742 for (i = 0; i < PTYPE_HASH_SIZE; i++) 11743 INIT_LIST_HEAD(&ptype_base[i]); 11744 11745 if (register_pernet_subsys(&netdev_net_ops)) 11746 goto out; 11747 11748 /* 11749 * Initialise the packet receive queues. 11750 */ 11751 11752 for_each_possible_cpu(i) { 11753 struct work_struct *flush = per_cpu_ptr(&flush_works, i); 11754 struct softnet_data *sd = &per_cpu(softnet_data, i); 11755 11756 INIT_WORK(flush, flush_backlog); 11757 11758 skb_queue_head_init(&sd->input_pkt_queue); 11759 skb_queue_head_init(&sd->process_queue); 11760 #ifdef CONFIG_XFRM_OFFLOAD 11761 skb_queue_head_init(&sd->xfrm_backlog); 11762 #endif 11763 INIT_LIST_HEAD(&sd->poll_list); 11764 sd->output_queue_tailp = &sd->output_queue; 11765 #ifdef CONFIG_RPS 11766 INIT_CSD(&sd->csd, rps_trigger_softirq, sd); 11767 sd->cpu = i; 11768 #endif 11769 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd); 11770 spin_lock_init(&sd->defer_lock); 11771 11772 init_gro_hash(&sd->backlog); 11773 sd->backlog.poll = process_backlog; 11774 sd->backlog.weight = weight_p; 11775 11776 if (net_page_pool_create(i)) 11777 goto out; 11778 } 11779 11780 dev_boot_phase = 0; 11781 11782 /* The loopback device is special if any other network devices 11783 * is present in a network namespace the loopback device must 11784 * be present. Since we now dynamically allocate and free the 11785 * loopback device ensure this invariant is maintained by 11786 * keeping the loopback device as the first device on the 11787 * list of network devices. Ensuring the loopback devices 11788 * is the first device that appears and the last network device 11789 * that disappears. 11790 */ 11791 if (register_pernet_device(&loopback_net_ops)) 11792 goto out; 11793 11794 if (register_pernet_device(&default_device_ops)) 11795 goto out; 11796 11797 open_softirq(NET_TX_SOFTIRQ, net_tx_action); 11798 open_softirq(NET_RX_SOFTIRQ, net_rx_action); 11799 11800 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead", 11801 NULL, dev_cpu_dead); 11802 WARN_ON(rc < 0); 11803 rc = 0; 11804 out: 11805 if (rc < 0) { 11806 for_each_possible_cpu(i) { 11807 struct page_pool *pp_ptr; 11808 11809 pp_ptr = per_cpu(system_page_pool, i); 11810 if (!pp_ptr) 11811 continue; 11812 11813 page_pool_destroy(pp_ptr); 11814 per_cpu(system_page_pool, i) = NULL; 11815 } 11816 } 11817 11818 return rc; 11819 } 11820 11821 subsys_initcall(net_dev_init); 11822