1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * NET3 Protocol independent device support routines. 4 * 5 * Derived from the non IP parts of dev.c 1.0.19 6 * Authors: Ross Biro 7 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 8 * Mark Evans, <evansmp@uhura.aston.ac.uk> 9 * 10 * Additional Authors: 11 * Florian la Roche <rzsfl@rz.uni-sb.de> 12 * Alan Cox <gw4pts@gw4pts.ampr.org> 13 * David Hinds <dahinds@users.sourceforge.net> 14 * Alexey Kuznetsov <kuznet@ms2.inr.ac.ru> 15 * Adam Sulmicki <adam@cfar.umd.edu> 16 * Pekka Riikonen <priikone@poesidon.pspt.fi> 17 * 18 * Changes: 19 * D.J. Barrow : Fixed bug where dev->refcnt gets set 20 * to 2 if register_netdev gets called 21 * before net_dev_init & also removed a 22 * few lines of code in the process. 23 * Alan Cox : device private ioctl copies fields back. 24 * Alan Cox : Transmit queue code does relevant 25 * stunts to keep the queue safe. 26 * Alan Cox : Fixed double lock. 27 * Alan Cox : Fixed promisc NULL pointer trap 28 * ???????? : Support the full private ioctl range 29 * Alan Cox : Moved ioctl permission check into 30 * drivers 31 * Tim Kordas : SIOCADDMULTI/SIOCDELMULTI 32 * Alan Cox : 100 backlog just doesn't cut it when 33 * you start doing multicast video 8) 34 * Alan Cox : Rewrote net_bh and list manager. 35 * Alan Cox : Fix ETH_P_ALL echoback lengths. 36 * Alan Cox : Took out transmit every packet pass 37 * Saved a few bytes in the ioctl handler 38 * Alan Cox : Network driver sets packet type before 39 * calling netif_rx. Saves a function 40 * call a packet. 41 * Alan Cox : Hashed net_bh() 42 * Richard Kooijman: Timestamp fixes. 43 * Alan Cox : Wrong field in SIOCGIFDSTADDR 44 * Alan Cox : Device lock protection. 45 * Alan Cox : Fixed nasty side effect of device close 46 * changes. 47 * Rudi Cilibrasi : Pass the right thing to 48 * set_mac_address() 49 * Dave Miller : 32bit quantity for the device lock to 50 * make it work out on a Sparc. 51 * Bjorn Ekwall : Added KERNELD hack. 52 * Alan Cox : Cleaned up the backlog initialise. 53 * Craig Metz : SIOCGIFCONF fix if space for under 54 * 1 device. 55 * Thomas Bogendoerfer : Return ENODEV for dev_open, if there 56 * is no device open function. 57 * Andi Kleen : Fix error reporting for SIOCGIFCONF 58 * Michael Chastain : Fix signed/unsigned for SIOCGIFCONF 59 * Cyrus Durgin : Cleaned for KMOD 60 * Adam Sulmicki : Bug Fix : Network Device Unload 61 * A network device unload needs to purge 62 * the backlog queue. 63 * Paul Rusty Russell : SIOCSIFNAME 64 * Pekka Riikonen : Netdev boot-time settings code 65 * Andrew Morton : Make unregister_netdevice wait 66 * indefinitely on dev->refcnt 67 * J Hadi Salim : - Backlog queue sampling 68 * - netif_rx() feedback 69 */ 70 71 #include <linux/uaccess.h> 72 #include <linux/bitmap.h> 73 #include <linux/capability.h> 74 #include <linux/cpu.h> 75 #include <linux/types.h> 76 #include <linux/kernel.h> 77 #include <linux/hash.h> 78 #include <linux/slab.h> 79 #include <linux/sched.h> 80 #include <linux/sched/isolation.h> 81 #include <linux/sched/mm.h> 82 #include <linux/smpboot.h> 83 #include <linux/mutex.h> 84 #include <linux/rwsem.h> 85 #include <linux/string.h> 86 #include <linux/mm.h> 87 #include <linux/socket.h> 88 #include <linux/sockios.h> 89 #include <linux/errno.h> 90 #include <linux/interrupt.h> 91 #include <linux/if_ether.h> 92 #include <linux/netdevice.h> 93 #include <linux/etherdevice.h> 94 #include <linux/ethtool.h> 95 #include <linux/ethtool_netlink.h> 96 #include <linux/skbuff.h> 97 #include <linux/kthread.h> 98 #include <linux/bpf.h> 99 #include <linux/bpf_trace.h> 100 #include <net/net_namespace.h> 101 #include <net/sock.h> 102 #include <net/busy_poll.h> 103 #include <linux/rtnetlink.h> 104 #include <linux/stat.h> 105 #include <net/dsa.h> 106 #include <net/dst.h> 107 #include <net/dst_metadata.h> 108 #include <net/gro.h> 109 #include <net/netdev_queues.h> 110 #include <net/pkt_sched.h> 111 #include <net/pkt_cls.h> 112 #include <net/checksum.h> 113 #include <net/xfrm.h> 114 #include <net/tcx.h> 115 #include <linux/highmem.h> 116 #include <linux/init.h> 117 #include <linux/module.h> 118 #include <linux/netpoll.h> 119 #include <linux/rcupdate.h> 120 #include <linux/delay.h> 121 #include <net/iw_handler.h> 122 #include <asm/current.h> 123 #include <linux/audit.h> 124 #include <linux/dmaengine.h> 125 #include <linux/err.h> 126 #include <linux/ctype.h> 127 #include <linux/if_arp.h> 128 #include <linux/if_vlan.h> 129 #include <linux/ip.h> 130 #include <net/ip.h> 131 #include <net/mpls.h> 132 #include <linux/ipv6.h> 133 #include <linux/in.h> 134 #include <linux/jhash.h> 135 #include <linux/random.h> 136 #include <trace/events/napi.h> 137 #include <trace/events/net.h> 138 #include <trace/events/skb.h> 139 #include <trace/events/qdisc.h> 140 #include <trace/events/xdp.h> 141 #include <linux/inetdevice.h> 142 #include <linux/cpu_rmap.h> 143 #include <linux/static_key.h> 144 #include <linux/hashtable.h> 145 #include <linux/vmalloc.h> 146 #include <linux/if_macvlan.h> 147 #include <linux/errqueue.h> 148 #include <linux/hrtimer.h> 149 #include <linux/netfilter_netdev.h> 150 #include <linux/crash_dump.h> 151 #include <linux/sctp.h> 152 #include <net/udp_tunnel.h> 153 #include <linux/net_namespace.h> 154 #include <linux/indirect_call_wrapper.h> 155 #include <net/devlink.h> 156 #include <linux/pm_runtime.h> 157 #include <linux/prandom.h> 158 #include <linux/once_lite.h> 159 #include <net/netdev_lock.h> 160 #include <net/netdev_rx_queue.h> 161 #include <net/page_pool/types.h> 162 #include <net/page_pool/helpers.h> 163 #include <net/page_pool/memory_provider.h> 164 #include <net/rps.h> 165 #include <linux/phy_link_topology.h> 166 167 #include "dev.h" 168 #include "devmem.h" 169 #include "net-sysfs.h" 170 171 static DEFINE_SPINLOCK(ptype_lock); 172 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly; 173 174 static int netif_rx_internal(struct sk_buff *skb); 175 static int call_netdevice_notifiers_extack(unsigned long val, 176 struct net_device *dev, 177 struct netlink_ext_ack *extack); 178 179 static DEFINE_MUTEX(ifalias_mutex); 180 181 /* protects napi_hash addition/deletion and napi_gen_id */ 182 static DEFINE_SPINLOCK(napi_hash_lock); 183 184 static unsigned int napi_gen_id = NR_CPUS; 185 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8); 186 187 static inline void dev_base_seq_inc(struct net *net) 188 { 189 unsigned int val = net->dev_base_seq + 1; 190 191 WRITE_ONCE(net->dev_base_seq, val ?: 1); 192 } 193 194 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name) 195 { 196 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ)); 197 198 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)]; 199 } 200 201 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex) 202 { 203 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)]; 204 } 205 206 #ifndef CONFIG_PREEMPT_RT 207 208 static DEFINE_STATIC_KEY_FALSE(use_backlog_threads_key); 209 210 static int __init setup_backlog_napi_threads(char *arg) 211 { 212 static_branch_enable(&use_backlog_threads_key); 213 return 0; 214 } 215 early_param("thread_backlog_napi", setup_backlog_napi_threads); 216 217 static bool use_backlog_threads(void) 218 { 219 return static_branch_unlikely(&use_backlog_threads_key); 220 } 221 222 #else 223 224 static bool use_backlog_threads(void) 225 { 226 return true; 227 } 228 229 #endif 230 231 static inline void backlog_lock_irq_save(struct softnet_data *sd, 232 unsigned long *flags) 233 { 234 if (IS_ENABLED(CONFIG_PREEMPT_RT)) { 235 spin_lock_irqsave(&sd->input_pkt_queue.lock, *flags); 236 } else { 237 local_irq_save(*flags); 238 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 239 spin_lock(&sd->input_pkt_queue.lock); 240 } 241 } 242 243 static inline void backlog_lock_irq_disable(struct softnet_data *sd) 244 { 245 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 246 spin_lock_irq(&sd->input_pkt_queue.lock); 247 else 248 local_irq_disable(); 249 } 250 251 static inline void backlog_unlock_irq_restore(struct softnet_data *sd, 252 unsigned long flags) 253 { 254 if (IS_ENABLED(CONFIG_PREEMPT_RT)) { 255 spin_unlock_irqrestore(&sd->input_pkt_queue.lock, flags); 256 } else { 257 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 258 spin_unlock(&sd->input_pkt_queue.lock); 259 local_irq_restore(flags); 260 } 261 } 262 263 static inline void backlog_unlock_irq_enable(struct softnet_data *sd) 264 { 265 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 266 spin_unlock_irq(&sd->input_pkt_queue.lock); 267 else 268 local_irq_enable(); 269 } 270 271 static struct netdev_name_node *netdev_name_node_alloc(struct net_device *dev, 272 const char *name) 273 { 274 struct netdev_name_node *name_node; 275 276 name_node = kmalloc_obj(*name_node); 277 if (!name_node) 278 return NULL; 279 INIT_HLIST_NODE(&name_node->hlist); 280 name_node->dev = dev; 281 name_node->name = name; 282 return name_node; 283 } 284 285 static struct netdev_name_node * 286 netdev_name_node_head_alloc(struct net_device *dev) 287 { 288 struct netdev_name_node *name_node; 289 290 name_node = netdev_name_node_alloc(dev, dev->name); 291 if (!name_node) 292 return NULL; 293 INIT_LIST_HEAD(&name_node->list); 294 return name_node; 295 } 296 297 static void netdev_name_node_free(struct netdev_name_node *name_node) 298 { 299 kfree(name_node); 300 } 301 302 static void netdev_name_node_add(struct net *net, 303 struct netdev_name_node *name_node) 304 { 305 hlist_add_head_rcu(&name_node->hlist, 306 dev_name_hash(net, name_node->name)); 307 } 308 309 static void netdev_name_node_del(struct netdev_name_node *name_node) 310 { 311 hlist_del_rcu(&name_node->hlist); 312 } 313 314 static struct netdev_name_node *netdev_name_node_lookup(struct net *net, 315 const char *name) 316 { 317 struct hlist_head *head = dev_name_hash(net, name); 318 struct netdev_name_node *name_node; 319 320 hlist_for_each_entry(name_node, head, hlist) 321 if (!strcmp(name_node->name, name)) 322 return name_node; 323 return NULL; 324 } 325 326 static struct netdev_name_node *netdev_name_node_lookup_rcu(struct net *net, 327 const char *name) 328 { 329 struct hlist_head *head = dev_name_hash(net, name); 330 struct netdev_name_node *name_node; 331 332 hlist_for_each_entry_rcu(name_node, head, hlist) 333 if (!strcmp(name_node->name, name)) 334 return name_node; 335 return NULL; 336 } 337 338 bool netdev_name_in_use(struct net *net, const char *name) 339 { 340 return netdev_name_node_lookup(net, name); 341 } 342 EXPORT_SYMBOL(netdev_name_in_use); 343 344 int netdev_name_node_alt_create(struct net_device *dev, const char *name) 345 { 346 struct netdev_name_node *name_node; 347 struct net *net = dev_net(dev); 348 349 name_node = netdev_name_node_lookup(net, name); 350 if (name_node) 351 return -EEXIST; 352 name_node = netdev_name_node_alloc(dev, name); 353 if (!name_node) 354 return -ENOMEM; 355 netdev_name_node_add(net, name_node); 356 /* The node that holds dev->name acts as a head of per-device list. */ 357 list_add_tail_rcu(&name_node->list, &dev->name_node->list); 358 359 return 0; 360 } 361 362 static void netdev_name_node_alt_free(struct rcu_head *head) 363 { 364 struct netdev_name_node *name_node = 365 container_of(head, struct netdev_name_node, rcu); 366 367 kfree(name_node->name); 368 netdev_name_node_free(name_node); 369 } 370 371 static void __netdev_name_node_alt_destroy(struct netdev_name_node *name_node) 372 { 373 netdev_name_node_del(name_node); 374 list_del_rcu(&name_node->list); 375 call_rcu(&name_node->rcu, netdev_name_node_alt_free); 376 } 377 378 int netdev_name_node_alt_destroy(struct net_device *dev, const char *name) 379 { 380 struct netdev_name_node *name_node; 381 struct net *net = dev_net(dev); 382 383 name_node = netdev_name_node_lookup(net, name); 384 if (!name_node) 385 return -ENOENT; 386 /* lookup might have found our primary name or a name belonging 387 * to another device. 388 */ 389 if (name_node == dev->name_node || name_node->dev != dev) 390 return -EINVAL; 391 392 __netdev_name_node_alt_destroy(name_node); 393 return 0; 394 } 395 396 static void netdev_name_node_alt_flush(struct net_device *dev) 397 { 398 struct netdev_name_node *name_node, *tmp; 399 400 list_for_each_entry_safe(name_node, tmp, &dev->name_node->list, list) { 401 list_del(&name_node->list); 402 netdev_name_node_alt_free(&name_node->rcu); 403 } 404 } 405 406 /* Device list insertion */ 407 static void list_netdevice(struct net_device *dev) 408 { 409 struct netdev_name_node *name_node; 410 struct net *net = dev_net(dev); 411 412 ASSERT_RTNL(); 413 414 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head); 415 netdev_name_node_add(net, dev->name_node); 416 hlist_add_head_rcu(&dev->index_hlist, 417 dev_index_hash(net, dev->ifindex)); 418 419 netdev_for_each_altname(dev, name_node) 420 netdev_name_node_add(net, name_node); 421 422 /* We reserved the ifindex, this can't fail */ 423 WARN_ON(xa_store(&net->dev_by_index, dev->ifindex, dev, GFP_KERNEL)); 424 425 dev_base_seq_inc(net); 426 } 427 428 /* Device list removal 429 * caller must respect a RCU grace period before freeing/reusing dev 430 */ 431 static void unlist_netdevice(struct net_device *dev) 432 { 433 struct netdev_name_node *name_node; 434 struct net *net = dev_net(dev); 435 436 ASSERT_RTNL(); 437 438 xa_erase(&net->dev_by_index, dev->ifindex); 439 440 netdev_for_each_altname(dev, name_node) 441 netdev_name_node_del(name_node); 442 443 /* Unlink dev from the device chain */ 444 list_del_rcu(&dev->dev_list); 445 netdev_name_node_del(dev->name_node); 446 hlist_del_rcu(&dev->index_hlist); 447 448 dev_base_seq_inc(dev_net(dev)); 449 } 450 451 /* 452 * Our notifier list 453 */ 454 455 static RAW_NOTIFIER_HEAD(netdev_chain); 456 457 /* 458 * Device drivers call our routines to queue packets here. We empty the 459 * queue in the local softnet handler. 460 */ 461 462 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data) = { 463 .process_queue_bh_lock = INIT_LOCAL_LOCK(process_queue_bh_lock), 464 }; 465 EXPORT_PER_CPU_SYMBOL(softnet_data); 466 467 /* Page_pool has a lockless array/stack to alloc/recycle pages. 468 * PP consumers must pay attention to run APIs in the appropriate context 469 * (e.g. NAPI context). 470 */ 471 DEFINE_PER_CPU(struct page_pool_bh, system_page_pool) = { 472 .bh_lock = INIT_LOCAL_LOCK(bh_lock), 473 }; 474 475 #ifdef CONFIG_LOCKDEP 476 /* 477 * register_netdevice() inits txq->_xmit_lock and sets lockdep class 478 * according to dev->type 479 */ 480 static const unsigned short netdev_lock_type[] = { 481 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25, 482 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET, 483 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM, 484 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP, 485 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD, 486 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25, 487 ARPHRD_CAN, ARPHRD_MCTP, 488 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP, 489 ARPHRD_RAWHDLC, ARPHRD_RAWIP, 490 ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD, 491 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI, 492 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE, 493 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET, 494 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL, 495 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM, 496 ARPHRD_IEEE80211_RADIOTAP, 497 ARPHRD_IEEE802154, ARPHRD_IEEE802154_MONITOR, 498 ARPHRD_PHONET, ARPHRD_PHONET_PIPE, 499 ARPHRD_CAIF, ARPHRD_IP6GRE, ARPHRD_NETLINK, ARPHRD_6LOWPAN, 500 ARPHRD_VSOCKMON, 501 ARPHRD_VOID, ARPHRD_NONE}; 502 503 static const char *const netdev_lock_name[] = { 504 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25", 505 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET", 506 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM", 507 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP", 508 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD", 509 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25", 510 "_xmit_CAN", "_xmit_MCTP", 511 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP", 512 "_xmit_RAWHDLC", "_xmit_RAWIP", 513 "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD", 514 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI", 515 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE", 516 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET", 517 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL", 518 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM", 519 "_xmit_IEEE80211_RADIOTAP", 520 "_xmit_IEEE802154", "_xmit_IEEE802154_MONITOR", 521 "_xmit_PHONET", "_xmit_PHONET_PIPE", 522 "_xmit_CAIF", "_xmit_IP6GRE", "_xmit_NETLINK", "_xmit_6LOWPAN", 523 "_xmit_VSOCKMON", 524 "_xmit_VOID", "_xmit_NONE"}; 525 526 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)]; 527 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)]; 528 529 static inline unsigned short netdev_lock_pos(unsigned short dev_type) 530 { 531 int i; 532 533 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++) 534 if (netdev_lock_type[i] == dev_type) 535 return i; 536 /* the last key is used by default */ 537 WARN_ONCE(1, "netdev_lock_pos() could not find dev_type=%u\n", dev_type); 538 return ARRAY_SIZE(netdev_lock_type) - 1; 539 } 540 541 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 542 unsigned short dev_type) 543 { 544 int i; 545 546 i = netdev_lock_pos(dev_type); 547 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i], 548 netdev_lock_name[i]); 549 } 550 551 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 552 { 553 int i; 554 555 i = netdev_lock_pos(dev->type); 556 lockdep_set_class_and_name(&dev->addr_list_lock, 557 &netdev_addr_lock_key[i], 558 netdev_lock_name[i]); 559 } 560 #else 561 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 562 unsigned short dev_type) 563 { 564 } 565 566 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 567 { 568 } 569 #endif 570 571 /******************************************************************************* 572 * 573 * Protocol management and registration routines 574 * 575 *******************************************************************************/ 576 577 578 /* 579 * Add a protocol ID to the list. Now that the input handler is 580 * smarter we can dispense with all the messy stuff that used to be 581 * here. 582 * 583 * BEWARE!!! Protocol handlers, mangling input packets, 584 * MUST BE last in hash buckets and checking protocol handlers 585 * MUST start from promiscuous ptype_all chain in net_bh. 586 * It is true now, do not change it. 587 * Explanation follows: if protocol handler, mangling packet, will 588 * be the first on list, it is not able to sense, that packet 589 * is cloned and should be copied-on-write, so that it will 590 * change it and subsequent readers will get broken packet. 591 * --ANK (980803) 592 */ 593 594 static inline struct list_head *ptype_head(const struct packet_type *pt) 595 { 596 if (pt->type == htons(ETH_P_ALL)) { 597 if (!pt->af_packet_net && !pt->dev) 598 return NULL; 599 600 return pt->dev ? &pt->dev->ptype_all : 601 &pt->af_packet_net->ptype_all; 602 } 603 604 if (pt->dev) 605 return &pt->dev->ptype_specific; 606 607 return pt->af_packet_net ? &pt->af_packet_net->ptype_specific : 608 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK]; 609 } 610 611 /** 612 * dev_add_pack - add packet handler 613 * @pt: packet type declaration 614 * 615 * Add a protocol handler to the networking stack. The passed &packet_type 616 * is linked into kernel lists and may not be freed until it has been 617 * removed from the kernel lists. 618 * 619 * This call does not sleep therefore it can not 620 * guarantee all CPU's that are in middle of receiving packets 621 * will see the new packet type (until the next received packet). 622 */ 623 624 void dev_add_pack(struct packet_type *pt) 625 { 626 struct list_head *head = ptype_head(pt); 627 628 if (WARN_ON_ONCE(!head)) 629 return; 630 631 spin_lock(&ptype_lock); 632 list_add_rcu(&pt->list, head); 633 spin_unlock(&ptype_lock); 634 } 635 EXPORT_SYMBOL(dev_add_pack); 636 637 /** 638 * __dev_remove_pack - remove packet handler 639 * @pt: packet type declaration 640 * 641 * Remove a protocol handler that was previously added to the kernel 642 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 643 * from the kernel lists and can be freed or reused once this function 644 * returns. 645 * 646 * The packet type might still be in use by receivers 647 * and must not be freed until after all the CPU's have gone 648 * through a quiescent state. 649 */ 650 void __dev_remove_pack(struct packet_type *pt) 651 { 652 struct list_head *head = ptype_head(pt); 653 struct packet_type *pt1; 654 655 if (!head) 656 return; 657 658 spin_lock(&ptype_lock); 659 660 list_for_each_entry(pt1, head, list) { 661 if (pt == pt1) { 662 list_del_rcu(&pt->list); 663 goto out; 664 } 665 } 666 667 pr_warn("dev_remove_pack: %p not found\n", pt); 668 out: 669 spin_unlock(&ptype_lock); 670 } 671 EXPORT_SYMBOL(__dev_remove_pack); 672 673 /** 674 * dev_remove_pack - remove packet handler 675 * @pt: packet type declaration 676 * 677 * Remove a protocol handler that was previously added to the kernel 678 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 679 * from the kernel lists and can be freed or reused once this function 680 * returns. 681 * 682 * This call sleeps to guarantee that no CPU is looking at the packet 683 * type after return. 684 */ 685 void dev_remove_pack(struct packet_type *pt) 686 { 687 __dev_remove_pack(pt); 688 689 synchronize_net(); 690 } 691 EXPORT_SYMBOL(dev_remove_pack); 692 693 694 /******************************************************************************* 695 * 696 * Device Interface Subroutines 697 * 698 *******************************************************************************/ 699 700 /** 701 * dev_get_iflink - get 'iflink' value of a interface 702 * @dev: targeted interface 703 * 704 * Indicates the ifindex the interface is linked to. 705 * Physical interfaces have the same 'ifindex' and 'iflink' values. 706 */ 707 708 int dev_get_iflink(const struct net_device *dev) 709 { 710 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink) 711 return dev->netdev_ops->ndo_get_iflink(dev); 712 713 return READ_ONCE(dev->ifindex); 714 } 715 EXPORT_SYMBOL(dev_get_iflink); 716 717 /** 718 * dev_fill_metadata_dst - Retrieve tunnel egress information. 719 * @dev: targeted interface 720 * @skb: The packet. 721 * 722 * For better visibility of tunnel traffic OVS needs to retrieve 723 * egress tunnel information for a packet. Following API allows 724 * user to get this info. 725 */ 726 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb) 727 { 728 struct ip_tunnel_info *info; 729 730 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst) 731 return -EINVAL; 732 733 info = skb_tunnel_info_unclone(skb); 734 if (!info) 735 return -ENOMEM; 736 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX))) 737 return -EINVAL; 738 739 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb); 740 } 741 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst); 742 743 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack) 744 { 745 if (stack->num_paths + 1 > NET_DEVICE_PATH_STACK_MAX) 746 return NULL; 747 748 return &stack->path[stack->num_paths]; 749 } 750 751 void dev_fill_forward_path_release(struct net_device_path_stack *stack) 752 { 753 struct net_device_path *path; 754 int k; 755 756 if (stack->num_paths == 0) 757 return; 758 759 for (k = stack->num_paths - 1; k >= 0; k--) { 760 path = &stack->path[k]; 761 switch (path->type) { 762 case DEV_PATH_TUN: 763 dst_release(path->tun.dst); 764 break; 765 default: 766 break; 767 } 768 } 769 } 770 EXPORT_SYMBOL_GPL(dev_fill_forward_path_release); 771 772 int dev_fill_forward_path(struct net_device_path_ctx *ctx, 773 struct net_device_path_stack *stack) 774 { 775 const struct net_device *last_dev; 776 struct net_device_path *path; 777 int ret = 0; 778 779 stack->num_paths = 0; 780 while (ctx->dev && ctx->dev->netdev_ops->ndo_fill_forward_path) { 781 last_dev = ctx->dev; 782 path = dev_fwd_path(stack); 783 if (!path) 784 goto err_out; 785 786 memset(path, 0, sizeof(struct net_device_path)); 787 ret = ctx->dev->netdev_ops->ndo_fill_forward_path(ctx, path); 788 if (ret < 0) 789 goto err_out; 790 791 stack->num_paths++; 792 if (WARN_ON_ONCE(last_dev == ctx->dev)) 793 goto err_out; 794 } 795 796 if (!ctx->dev) 797 return ret; 798 799 path = dev_fwd_path(stack); 800 if (!path) 801 goto err_out; 802 803 path->type = DEV_PATH_ETHERNET; 804 path->dev = ctx->dev; 805 stack->num_paths++; 806 807 return 0; 808 err_out: 809 dev_fill_forward_path_release(stack); 810 811 return -1; 812 } 813 EXPORT_SYMBOL_GPL(dev_fill_forward_path); 814 815 /* must be called under rcu_read_lock(), as we dont take a reference */ 816 static struct napi_struct *napi_by_id(unsigned int napi_id) 817 { 818 unsigned int hash = napi_id % HASH_SIZE(napi_hash); 819 struct napi_struct *napi; 820 821 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node) 822 if (napi->napi_id == napi_id) 823 return napi; 824 825 return NULL; 826 } 827 828 /* must be called under rcu_read_lock(), as we dont take a reference */ 829 static struct napi_struct * 830 netdev_napi_by_id(struct net *net, unsigned int napi_id) 831 { 832 struct napi_struct *napi; 833 834 napi = napi_by_id(napi_id); 835 if (!napi) 836 return NULL; 837 838 if (WARN_ON_ONCE(!napi->dev)) 839 return NULL; 840 if (!net_eq(net, dev_net(napi->dev))) 841 return NULL; 842 843 return napi; 844 } 845 846 /** 847 * netdev_napi_by_id_lock() - find a device by NAPI ID and lock it 848 * @net: the applicable net namespace 849 * @napi_id: ID of a NAPI of a target device 850 * 851 * Find a NAPI instance with @napi_id. Lock its device. 852 * The device must be in %NETREG_REGISTERED state for lookup to succeed. 853 * netdev_unlock() must be called to release it. 854 * 855 * Return: pointer to NAPI, its device with lock held, NULL if not found. 856 */ 857 struct napi_struct * 858 netdev_napi_by_id_lock(struct net *net, unsigned int napi_id) 859 { 860 struct napi_struct *napi; 861 struct net_device *dev; 862 863 rcu_read_lock(); 864 napi = netdev_napi_by_id(net, napi_id); 865 if (!napi || READ_ONCE(napi->dev->reg_state) != NETREG_REGISTERED) { 866 rcu_read_unlock(); 867 return NULL; 868 } 869 870 dev = napi->dev; 871 dev_hold(dev); 872 rcu_read_unlock(); 873 874 dev = __netdev_put_lock(dev, net); 875 if (!dev) 876 return NULL; 877 878 rcu_read_lock(); 879 napi = netdev_napi_by_id(net, napi_id); 880 if (napi && napi->dev != dev) 881 napi = NULL; 882 rcu_read_unlock(); 883 884 if (!napi) 885 netdev_unlock(dev); 886 return napi; 887 } 888 889 /** 890 * __dev_get_by_name - find a device by its name 891 * @net: the applicable net namespace 892 * @name: name to find 893 * 894 * Find an interface by name. Must be called under RTNL semaphore. 895 * If the name is found a pointer to the device is returned. 896 * If the name is not found then %NULL is returned. The 897 * reference counters are not incremented so the caller must be 898 * careful with locks. 899 */ 900 901 struct net_device *__dev_get_by_name(struct net *net, const char *name) 902 { 903 struct netdev_name_node *node_name; 904 905 node_name = netdev_name_node_lookup(net, name); 906 return node_name ? node_name->dev : NULL; 907 } 908 EXPORT_SYMBOL(__dev_get_by_name); 909 910 /** 911 * dev_get_by_name_rcu - find a device by its name 912 * @net: the applicable net namespace 913 * @name: name to find 914 * 915 * Find an interface by name. 916 * If the name is found a pointer to the device is returned. 917 * If the name is not found then %NULL is returned. 918 * The reference counters are not incremented so the caller must be 919 * careful with locks. The caller must hold RCU lock. 920 */ 921 922 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name) 923 { 924 struct netdev_name_node *node_name; 925 926 node_name = netdev_name_node_lookup_rcu(net, name); 927 return node_name ? node_name->dev : NULL; 928 } 929 EXPORT_SYMBOL(dev_get_by_name_rcu); 930 931 /* Deprecated for new users, call netdev_get_by_name() instead */ 932 struct net_device *dev_get_by_name(struct net *net, const char *name) 933 { 934 struct net_device *dev; 935 936 rcu_read_lock(); 937 dev = dev_get_by_name_rcu(net, name); 938 dev_hold(dev); 939 rcu_read_unlock(); 940 return dev; 941 } 942 EXPORT_SYMBOL(dev_get_by_name); 943 944 /** 945 * netdev_get_by_name() - find a device by its name 946 * @net: the applicable net namespace 947 * @name: name to find 948 * @tracker: tracking object for the acquired reference 949 * @gfp: allocation flags for the tracker 950 * 951 * Find an interface by name. This can be called from any 952 * context and does its own locking. The returned handle has 953 * the usage count incremented and the caller must use netdev_put() to 954 * release it when it is no longer needed. %NULL is returned if no 955 * matching device is found. 956 */ 957 struct net_device *netdev_get_by_name(struct net *net, const char *name, 958 netdevice_tracker *tracker, gfp_t gfp) 959 { 960 struct net_device *dev; 961 962 dev = dev_get_by_name(net, name); 963 if (dev) 964 netdev_tracker_alloc(dev, tracker, gfp); 965 return dev; 966 } 967 EXPORT_SYMBOL(netdev_get_by_name); 968 969 /** 970 * __dev_get_by_index - find a device by its ifindex 971 * @net: the applicable net namespace 972 * @ifindex: index of device 973 * 974 * Search for an interface by index. Returns %NULL if the device 975 * is not found or a pointer to the device. The device has not 976 * had its reference counter increased so the caller must be careful 977 * about locking. The caller must hold the RTNL semaphore. 978 */ 979 980 struct net_device *__dev_get_by_index(struct net *net, int ifindex) 981 { 982 struct net_device *dev; 983 struct hlist_head *head = dev_index_hash(net, ifindex); 984 985 hlist_for_each_entry(dev, head, index_hlist) 986 if (dev->ifindex == ifindex) 987 return dev; 988 989 return NULL; 990 } 991 EXPORT_SYMBOL(__dev_get_by_index); 992 993 /** 994 * dev_get_by_index_rcu - find a device by its ifindex 995 * @net: the applicable net namespace 996 * @ifindex: index of device 997 * 998 * Search for an interface by index. Returns %NULL if the device 999 * is not found or a pointer to the device. The device has not 1000 * had its reference counter increased so the caller must be careful 1001 * about locking. The caller must hold RCU lock. 1002 */ 1003 1004 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex) 1005 { 1006 struct net_device *dev; 1007 struct hlist_head *head = dev_index_hash(net, ifindex); 1008 1009 hlist_for_each_entry_rcu(dev, head, index_hlist) 1010 if (dev->ifindex == ifindex) 1011 return dev; 1012 1013 return NULL; 1014 } 1015 EXPORT_SYMBOL(dev_get_by_index_rcu); 1016 1017 /* Deprecated for new users, call netdev_get_by_index() instead */ 1018 struct net_device *dev_get_by_index(struct net *net, int ifindex) 1019 { 1020 struct net_device *dev; 1021 1022 rcu_read_lock(); 1023 dev = dev_get_by_index_rcu(net, ifindex); 1024 dev_hold(dev); 1025 rcu_read_unlock(); 1026 return dev; 1027 } 1028 EXPORT_SYMBOL(dev_get_by_index); 1029 1030 /** 1031 * netdev_get_by_index() - find a device by its ifindex 1032 * @net: the applicable net namespace 1033 * @ifindex: index of device 1034 * @tracker: tracking object for the acquired reference 1035 * @gfp: allocation flags for the tracker 1036 * 1037 * Search for an interface by index. Returns NULL if the device 1038 * is not found or a pointer to the device. The device returned has 1039 * had a reference added and the pointer is safe until the user calls 1040 * netdev_put() to indicate they have finished with it. 1041 */ 1042 struct net_device *netdev_get_by_index(struct net *net, int ifindex, 1043 netdevice_tracker *tracker, gfp_t gfp) 1044 { 1045 struct net_device *dev; 1046 1047 dev = dev_get_by_index(net, ifindex); 1048 if (dev) 1049 netdev_tracker_alloc(dev, tracker, gfp); 1050 return dev; 1051 } 1052 EXPORT_SYMBOL(netdev_get_by_index); 1053 1054 /** 1055 * dev_get_by_napi_id - find a device by napi_id 1056 * @napi_id: ID of the NAPI struct 1057 * 1058 * Search for an interface by NAPI ID. Returns %NULL if the device 1059 * is not found or a pointer to the device. The device has not had 1060 * its reference counter increased so the caller must be careful 1061 * about locking. The caller must hold RCU lock. 1062 */ 1063 struct net_device *dev_get_by_napi_id(unsigned int napi_id) 1064 { 1065 struct napi_struct *napi; 1066 1067 WARN_ON_ONCE(!rcu_read_lock_held()); 1068 1069 if (!napi_id_valid(napi_id)) 1070 return NULL; 1071 1072 napi = napi_by_id(napi_id); 1073 1074 return napi ? napi->dev : NULL; 1075 } 1076 1077 /* Release the held reference on the net_device, and if the net_device 1078 * is still registered try to lock the instance lock. If device is being 1079 * unregistered NULL will be returned (but the reference has been released, 1080 * either way!) 1081 * 1082 * This helper is intended for locking net_device after it has been looked up 1083 * using a lockless lookup helper. Lock prevents the instance from going away. 1084 */ 1085 struct net_device * 1086 netdev_put_lock(struct net_device *dev, struct net *net, 1087 netdevice_tracker *tracker) 1088 { 1089 netdev_lock(dev); 1090 if (dev->reg_state > NETREG_REGISTERED || 1091 dev->moving_ns || !net_eq(dev_net(dev), net)) { 1092 netdev_unlock(dev); 1093 netdev_put(dev, tracker); 1094 return NULL; 1095 } 1096 netdev_put(dev, tracker); 1097 return dev; 1098 } 1099 1100 static struct net_device * 1101 __netdev_put_lock_ops_compat(struct net_device *dev, struct net *net) 1102 { 1103 netdev_lock_ops_compat(dev); 1104 if (dev->reg_state > NETREG_REGISTERED || 1105 dev->moving_ns || !net_eq(dev_net(dev), net)) { 1106 netdev_unlock_ops_compat(dev); 1107 dev_put(dev); 1108 return NULL; 1109 } 1110 dev_put(dev); 1111 return dev; 1112 } 1113 1114 /** 1115 * netdev_get_by_index_lock() - find a device by its ifindex 1116 * @net: the applicable net namespace 1117 * @ifindex: index of device 1118 * 1119 * Search for an interface by index. If a valid device 1120 * with @ifindex is found it will be returned with netdev->lock held. 1121 * netdev_unlock() must be called to release it. 1122 * 1123 * Return: pointer to a device with lock held, NULL if not found. 1124 */ 1125 struct net_device *netdev_get_by_index_lock(struct net *net, int ifindex) 1126 { 1127 struct net_device *dev; 1128 1129 dev = dev_get_by_index(net, ifindex); 1130 if (!dev) 1131 return NULL; 1132 1133 return __netdev_put_lock(dev, net); 1134 } 1135 1136 struct net_device * 1137 netdev_get_by_index_lock_ops_compat(struct net *net, int ifindex) 1138 { 1139 struct net_device *dev; 1140 1141 dev = dev_get_by_index(net, ifindex); 1142 if (!dev) 1143 return NULL; 1144 1145 return __netdev_put_lock_ops_compat(dev, net); 1146 } 1147 1148 struct net_device * 1149 netdev_xa_find_lock(struct net *net, struct net_device *dev, 1150 unsigned long *index) 1151 { 1152 if (dev) 1153 netdev_unlock(dev); 1154 1155 do { 1156 rcu_read_lock(); 1157 dev = xa_find(&net->dev_by_index, index, ULONG_MAX, XA_PRESENT); 1158 if (!dev) { 1159 rcu_read_unlock(); 1160 return NULL; 1161 } 1162 dev_hold(dev); 1163 rcu_read_unlock(); 1164 1165 dev = __netdev_put_lock(dev, net); 1166 if (dev) 1167 return dev; 1168 1169 (*index)++; 1170 } while (true); 1171 } 1172 1173 struct net_device * 1174 netdev_xa_find_lock_ops_compat(struct net *net, struct net_device *dev, 1175 unsigned long *index) 1176 { 1177 if (dev) 1178 netdev_unlock_ops_compat(dev); 1179 1180 do { 1181 rcu_read_lock(); 1182 dev = xa_find(&net->dev_by_index, index, ULONG_MAX, XA_PRESENT); 1183 if (!dev) { 1184 rcu_read_unlock(); 1185 return NULL; 1186 } 1187 dev_hold(dev); 1188 rcu_read_unlock(); 1189 1190 dev = __netdev_put_lock_ops_compat(dev, net); 1191 if (dev) 1192 return dev; 1193 1194 (*index)++; 1195 } while (true); 1196 } 1197 1198 static DEFINE_SEQLOCK(netdev_rename_lock); 1199 1200 void netdev_copy_name(struct net_device *dev, char *name) 1201 { 1202 unsigned int seq; 1203 1204 do { 1205 seq = read_seqbegin(&netdev_rename_lock); 1206 strscpy(name, dev->name, IFNAMSIZ); 1207 } while (read_seqretry(&netdev_rename_lock, seq)); 1208 } 1209 1210 /** 1211 * netdev_get_name - get a netdevice name, knowing its ifindex. 1212 * @net: network namespace 1213 * @name: a pointer to the buffer where the name will be stored. 1214 * @ifindex: the ifindex of the interface to get the name from. 1215 */ 1216 int netdev_get_name(struct net *net, char *name, int ifindex) 1217 { 1218 struct net_device *dev; 1219 int ret; 1220 1221 rcu_read_lock(); 1222 1223 dev = dev_get_by_index_rcu(net, ifindex); 1224 if (!dev) { 1225 ret = -ENODEV; 1226 goto out; 1227 } 1228 1229 netdev_copy_name(dev, name); 1230 1231 ret = 0; 1232 out: 1233 rcu_read_unlock(); 1234 return ret; 1235 } 1236 1237 static bool dev_addr_cmp(struct net_device *dev, unsigned short type, 1238 const char *ha) 1239 { 1240 return dev->type == type && !memcmp(dev->dev_addr, ha, dev->addr_len); 1241 } 1242 1243 /** 1244 * dev_getbyhwaddr_rcu - find a device by its hardware address 1245 * @net: the applicable net namespace 1246 * @type: media type of device 1247 * @ha: hardware address 1248 * 1249 * Search for an interface by MAC address. Returns NULL if the device 1250 * is not found or a pointer to the device. 1251 * The caller must hold RCU. 1252 * The returned device has not had its ref count increased 1253 * and the caller must therefore be careful about locking 1254 * 1255 */ 1256 1257 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type, 1258 const char *ha) 1259 { 1260 struct net_device *dev; 1261 1262 for_each_netdev_rcu(net, dev) 1263 if (dev_addr_cmp(dev, type, ha)) 1264 return dev; 1265 1266 return NULL; 1267 } 1268 EXPORT_SYMBOL(dev_getbyhwaddr_rcu); 1269 1270 /** 1271 * dev_getbyhwaddr() - find a device by its hardware address 1272 * @net: the applicable net namespace 1273 * @type: media type of device 1274 * @ha: hardware address 1275 * 1276 * Similar to dev_getbyhwaddr_rcu(), but the owner needs to hold 1277 * rtnl_lock. 1278 * 1279 * Context: rtnl_lock() must be held. 1280 * Return: pointer to the net_device, or NULL if not found 1281 */ 1282 struct net_device *dev_getbyhwaddr(struct net *net, unsigned short type, 1283 const char *ha) 1284 { 1285 struct net_device *dev; 1286 1287 ASSERT_RTNL(); 1288 for_each_netdev(net, dev) 1289 if (dev_addr_cmp(dev, type, ha)) 1290 return dev; 1291 1292 return NULL; 1293 } 1294 EXPORT_SYMBOL(dev_getbyhwaddr); 1295 1296 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type) 1297 { 1298 struct net_device *dev, *ret = NULL; 1299 1300 rcu_read_lock(); 1301 for_each_netdev_rcu(net, dev) 1302 if (dev->type == type) { 1303 dev_hold(dev); 1304 ret = dev; 1305 break; 1306 } 1307 rcu_read_unlock(); 1308 return ret; 1309 } 1310 EXPORT_SYMBOL(dev_getfirstbyhwtype); 1311 1312 /** 1313 * netdev_get_by_flags_rcu - find any device with given flags 1314 * @net: the applicable net namespace 1315 * @tracker: tracking object for the acquired reference 1316 * @if_flags: IFF_* values 1317 * @mask: bitmask of bits in if_flags to check 1318 * 1319 * Search for any interface with the given flags. 1320 * 1321 * Context: rcu_read_lock() must be held. 1322 * Returns: NULL if a device is not found or a pointer to the device. 1323 */ 1324 struct net_device *netdev_get_by_flags_rcu(struct net *net, netdevice_tracker *tracker, 1325 unsigned short if_flags, unsigned short mask) 1326 { 1327 struct net_device *dev; 1328 1329 for_each_netdev_rcu(net, dev) { 1330 if (((READ_ONCE(dev->flags) ^ if_flags) & mask) == 0) { 1331 netdev_hold(dev, tracker, GFP_ATOMIC); 1332 return dev; 1333 } 1334 } 1335 1336 return NULL; 1337 } 1338 1339 /** 1340 * dev_valid_name - check if name is okay for network device 1341 * @name: name string 1342 * 1343 * Network device names need to be valid file names to 1344 * allow sysfs to work. We also disallow any kind of 1345 * whitespace. 1346 */ 1347 bool dev_valid_name(const char *name) 1348 { 1349 if (*name == '\0') 1350 return false; 1351 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ) 1352 return false; 1353 if (!strcmp(name, ".") || !strcmp(name, "..")) 1354 return false; 1355 1356 while (*name) { 1357 if (*name == '/' || *name == ':' || isspace(*name)) 1358 return false; 1359 name++; 1360 } 1361 return true; 1362 } 1363 EXPORT_SYMBOL(dev_valid_name); 1364 1365 /** 1366 * __dev_alloc_name - allocate a name for a device 1367 * @net: network namespace to allocate the device name in 1368 * @name: name format string 1369 * @res: result name string 1370 * 1371 * Passed a format string - eg "lt%d" it will try and find a suitable 1372 * id. It scans list of devices to build up a free map, then chooses 1373 * the first empty slot. The caller must hold the dev_base or rtnl lock 1374 * while allocating the name and adding the device in order to avoid 1375 * duplicates. 1376 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1377 * Returns the number of the unit assigned or a negative errno code. 1378 */ 1379 1380 static int __dev_alloc_name(struct net *net, const char *name, char *res) 1381 { 1382 int i = 0; 1383 const char *p; 1384 const int max_netdevices = 8*PAGE_SIZE; 1385 unsigned long *inuse; 1386 struct net_device *d; 1387 char buf[IFNAMSIZ]; 1388 1389 /* Verify the string as this thing may have come from the user. 1390 * There must be one "%d" and no other "%" characters. 1391 */ 1392 p = strchr(name, '%'); 1393 if (!p || p[1] != 'd' || strchr(p + 2, '%')) 1394 return -EINVAL; 1395 1396 /* Use one page as a bit array of possible slots */ 1397 inuse = bitmap_zalloc(max_netdevices, GFP_ATOMIC); 1398 if (!inuse) 1399 return -ENOMEM; 1400 1401 for_each_netdev(net, d) { 1402 struct netdev_name_node *name_node; 1403 1404 netdev_for_each_altname(d, name_node) { 1405 if (!sscanf(name_node->name, name, &i)) 1406 continue; 1407 if (i < 0 || i >= max_netdevices) 1408 continue; 1409 1410 /* avoid cases where sscanf is not exact inverse of printf */ 1411 snprintf(buf, IFNAMSIZ, name, i); 1412 if (!strncmp(buf, name_node->name, IFNAMSIZ)) 1413 __set_bit(i, inuse); 1414 } 1415 if (!sscanf(d->name, name, &i)) 1416 continue; 1417 if (i < 0 || i >= max_netdevices) 1418 continue; 1419 1420 /* avoid cases where sscanf is not exact inverse of printf */ 1421 snprintf(buf, IFNAMSIZ, name, i); 1422 if (!strncmp(buf, d->name, IFNAMSIZ)) 1423 __set_bit(i, inuse); 1424 } 1425 1426 i = find_first_zero_bit(inuse, max_netdevices); 1427 bitmap_free(inuse); 1428 if (i == max_netdevices) 1429 return -ENFILE; 1430 1431 /* 'res' and 'name' could overlap, use 'buf' as an intermediate buffer */ 1432 strscpy(buf, name, IFNAMSIZ); 1433 snprintf(res, IFNAMSIZ, buf, i); 1434 return i; 1435 } 1436 1437 /* Returns negative errno or allocated unit id (see __dev_alloc_name()) */ 1438 static int dev_prep_valid_name(struct net *net, struct net_device *dev, 1439 const char *want_name, char *out_name, 1440 int dup_errno) 1441 { 1442 if (!dev_valid_name(want_name)) 1443 return -EINVAL; 1444 1445 if (strchr(want_name, '%')) 1446 return __dev_alloc_name(net, want_name, out_name); 1447 1448 if (netdev_name_in_use(net, want_name)) 1449 return -dup_errno; 1450 if (out_name != want_name) 1451 strscpy(out_name, want_name, IFNAMSIZ); 1452 return 0; 1453 } 1454 1455 /** 1456 * dev_alloc_name - allocate a name for a device 1457 * @dev: device 1458 * @name: name format string 1459 * 1460 * Passed a format string - eg "lt%d" it will try and find a suitable 1461 * id. It scans list of devices to build up a free map, then chooses 1462 * the first empty slot. The caller must hold the dev_base or rtnl lock 1463 * while allocating the name and adding the device in order to avoid 1464 * duplicates. 1465 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1466 * Returns the number of the unit assigned or a negative errno code. 1467 */ 1468 1469 int dev_alloc_name(struct net_device *dev, const char *name) 1470 { 1471 return dev_prep_valid_name(dev_net(dev), dev, name, dev->name, ENFILE); 1472 } 1473 EXPORT_SYMBOL(dev_alloc_name); 1474 1475 static int dev_get_valid_name(struct net *net, struct net_device *dev, 1476 const char *name) 1477 { 1478 int ret; 1479 1480 ret = dev_prep_valid_name(net, dev, name, dev->name, EEXIST); 1481 return ret < 0 ? ret : 0; 1482 } 1483 1484 int netif_change_name(struct net_device *dev, const char *newname) 1485 { 1486 struct net *net = dev_net(dev); 1487 unsigned char old_assign_type; 1488 char oldname[IFNAMSIZ]; 1489 int err = 0; 1490 int ret; 1491 1492 ASSERT_RTNL_NET(net); 1493 1494 if (!strncmp(newname, dev->name, IFNAMSIZ)) 1495 return 0; 1496 1497 memcpy(oldname, dev->name, IFNAMSIZ); 1498 1499 write_seqlock_bh(&netdev_rename_lock); 1500 err = dev_get_valid_name(net, dev, newname); 1501 write_sequnlock_bh(&netdev_rename_lock); 1502 1503 if (err < 0) 1504 return err; 1505 1506 if (oldname[0] && !strchr(oldname, '%')) 1507 netdev_info(dev, "renamed from %s%s\n", oldname, 1508 dev->flags & IFF_UP ? " (while UP)" : ""); 1509 1510 old_assign_type = dev->name_assign_type; 1511 WRITE_ONCE(dev->name_assign_type, NET_NAME_RENAMED); 1512 1513 rollback: 1514 ret = device_rename(&dev->dev, dev->name); 1515 if (ret) { 1516 write_seqlock_bh(&netdev_rename_lock); 1517 memcpy(dev->name, oldname, IFNAMSIZ); 1518 write_sequnlock_bh(&netdev_rename_lock); 1519 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1520 return ret; 1521 } 1522 1523 netdev_adjacent_rename_links(dev, oldname); 1524 1525 netdev_name_node_del(dev->name_node); 1526 1527 synchronize_net(); 1528 1529 netdev_name_node_add(net, dev->name_node); 1530 1531 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev); 1532 ret = notifier_to_errno(ret); 1533 1534 if (ret) { 1535 /* err >= 0 after dev_alloc_name() or stores the first errno */ 1536 if (err >= 0) { 1537 err = ret; 1538 write_seqlock_bh(&netdev_rename_lock); 1539 memcpy(dev->name, oldname, IFNAMSIZ); 1540 write_sequnlock_bh(&netdev_rename_lock); 1541 memcpy(oldname, newname, IFNAMSIZ); 1542 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1543 old_assign_type = NET_NAME_RENAMED; 1544 goto rollback; 1545 } else { 1546 netdev_err(dev, "name change rollback failed: %d\n", 1547 ret); 1548 } 1549 } 1550 1551 return err; 1552 } 1553 1554 int netif_set_alias(struct net_device *dev, const char *alias, size_t len) 1555 { 1556 struct dev_ifalias *new_alias = NULL; 1557 1558 if (len >= IFALIASZ) 1559 return -EINVAL; 1560 1561 if (len) { 1562 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL); 1563 if (!new_alias) 1564 return -ENOMEM; 1565 1566 memcpy(new_alias->ifalias, alias, len); 1567 new_alias->ifalias[len] = 0; 1568 } 1569 1570 mutex_lock(&ifalias_mutex); 1571 new_alias = rcu_replace_pointer(dev->ifalias, new_alias, 1572 mutex_is_locked(&ifalias_mutex)); 1573 mutex_unlock(&ifalias_mutex); 1574 1575 if (new_alias) 1576 kfree_rcu(new_alias, rcuhead); 1577 1578 return len; 1579 } 1580 1581 /** 1582 * dev_get_alias - get ifalias of a device 1583 * @dev: device 1584 * @name: buffer to store name of ifalias 1585 * @len: size of buffer 1586 * 1587 * get ifalias for a device. Caller must make sure dev cannot go 1588 * away, e.g. rcu read lock or own a reference count to device. 1589 */ 1590 int dev_get_alias(const struct net_device *dev, char *name, size_t len) 1591 { 1592 const struct dev_ifalias *alias; 1593 int ret = 0; 1594 1595 rcu_read_lock(); 1596 alias = rcu_dereference(dev->ifalias); 1597 if (alias) 1598 ret = snprintf(name, len, "%s", alias->ifalias); 1599 rcu_read_unlock(); 1600 1601 return ret; 1602 } 1603 1604 /** 1605 * netdev_features_change - device changes features 1606 * @dev: device to cause notification 1607 * 1608 * Called to indicate a device has changed features. 1609 */ 1610 void netdev_features_change(struct net_device *dev) 1611 { 1612 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev); 1613 } 1614 EXPORT_SYMBOL(netdev_features_change); 1615 1616 void netif_state_change(struct net_device *dev) 1617 { 1618 netdev_assert_locked_ops_compat_or_invisible(dev); 1619 1620 if (dev->flags & IFF_UP) { 1621 struct netdev_notifier_change_info change_info = { 1622 .info.dev = dev, 1623 }; 1624 1625 call_netdevice_notifiers_info(NETDEV_CHANGE, 1626 &change_info.info); 1627 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL, 0, NULL); 1628 } 1629 } 1630 1631 /** 1632 * __netdev_notify_peers - notify network peers about existence of @dev, 1633 * to be called when rtnl lock is already held. 1634 * @dev: network device 1635 * 1636 * Generate traffic such that interested network peers are aware of 1637 * @dev, such as by generating a gratuitous ARP. This may be used when 1638 * a device wants to inform the rest of the network about some sort of 1639 * reconfiguration such as a failover event or virtual machine 1640 * migration. 1641 */ 1642 void __netdev_notify_peers(struct net_device *dev) 1643 { 1644 ASSERT_RTNL(); 1645 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev); 1646 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev); 1647 } 1648 EXPORT_SYMBOL(__netdev_notify_peers); 1649 1650 /** 1651 * netdev_notify_peers - notify network peers about existence of @dev 1652 * @dev: network device 1653 * 1654 * Generate traffic such that interested network peers are aware of 1655 * @dev, such as by generating a gratuitous ARP. This may be used when 1656 * a device wants to inform the rest of the network about some sort of 1657 * reconfiguration such as a failover event or virtual machine 1658 * migration. 1659 */ 1660 void netdev_notify_peers(struct net_device *dev) 1661 { 1662 rtnl_lock(); 1663 __netdev_notify_peers(dev); 1664 rtnl_unlock(); 1665 } 1666 EXPORT_SYMBOL(netdev_notify_peers); 1667 1668 static int napi_threaded_poll(void *data); 1669 1670 static int napi_kthread_create(struct napi_struct *n) 1671 { 1672 int err = 0; 1673 1674 /* Create and wake up the kthread once to put it in 1675 * TASK_INTERRUPTIBLE mode to avoid the blocked task 1676 * warning and work with loadavg. 1677 */ 1678 n->thread = kthread_run(napi_threaded_poll, n, "napi/%s-%d", 1679 n->dev->name, n->napi_id); 1680 if (IS_ERR(n->thread)) { 1681 err = PTR_ERR(n->thread); 1682 pr_err("kthread_run failed with err %d\n", err); 1683 n->thread = NULL; 1684 } 1685 1686 return err; 1687 } 1688 1689 static int __dev_open(struct net_device *dev, struct netlink_ext_ack *extack) 1690 { 1691 const struct net_device_ops *ops = dev->netdev_ops; 1692 int ret; 1693 1694 ASSERT_RTNL(); 1695 dev_addr_check(dev); 1696 1697 if (!netif_device_present(dev)) { 1698 /* may be detached because parent is runtime-suspended */ 1699 if (dev->dev.parent) 1700 pm_runtime_resume(dev->dev.parent); 1701 if (!netif_device_present(dev)) 1702 return -ENODEV; 1703 } 1704 1705 /* Block netpoll from trying to do any rx path servicing. 1706 * If we don't do this there is a chance ndo_poll_controller 1707 * or ndo_poll may be running while we open the device 1708 */ 1709 netpoll_poll_disable(dev); 1710 1711 ret = call_netdevice_notifiers_extack(NETDEV_PRE_UP, dev, extack); 1712 ret = notifier_to_errno(ret); 1713 if (ret) 1714 return ret; 1715 1716 set_bit(__LINK_STATE_START, &dev->state); 1717 1718 netdev_assert_locked_ops_compat(dev); 1719 1720 if (ops->ndo_validate_addr) 1721 ret = ops->ndo_validate_addr(dev); 1722 1723 if (!ret && ops->ndo_open) 1724 ret = ops->ndo_open(dev); 1725 1726 netpoll_poll_enable(dev); 1727 1728 if (ret) 1729 clear_bit(__LINK_STATE_START, &dev->state); 1730 else { 1731 netif_set_up(dev, true); 1732 dev_set_rx_mode(dev); 1733 dev_activate(dev); 1734 add_device_randomness(dev->dev_addr, dev->addr_len); 1735 } 1736 1737 return ret; 1738 } 1739 1740 int netif_open(struct net_device *dev, struct netlink_ext_ack *extack) 1741 { 1742 int ret; 1743 1744 if (dev->flags & IFF_UP) 1745 return 0; 1746 1747 ret = __dev_open(dev, extack); 1748 if (ret < 0) 1749 return ret; 1750 1751 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1752 call_netdevice_notifiers(NETDEV_UP, dev); 1753 1754 return ret; 1755 } 1756 EXPORT_SYMBOL(netif_open); 1757 1758 static void __dev_close_many(struct list_head *head) 1759 { 1760 struct net_device *dev; 1761 1762 ASSERT_RTNL(); 1763 might_sleep(); 1764 1765 list_for_each_entry(dev, head, close_list) { 1766 /* Temporarily disable netpoll until the interface is down */ 1767 netpoll_poll_disable(dev); 1768 1769 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev); 1770 1771 clear_bit(__LINK_STATE_START, &dev->state); 1772 1773 /* Synchronize to scheduled poll. We cannot touch poll list, it 1774 * can be even on different cpu. So just clear netif_running(). 1775 * 1776 * dev->stop() will invoke napi_disable() on all of it's 1777 * napi_struct instances on this device. 1778 */ 1779 smp_mb__after_atomic(); /* Commit netif_running(). */ 1780 } 1781 1782 dev_deactivate_many(head, true); 1783 1784 list_for_each_entry(dev, head, close_list) { 1785 const struct net_device_ops *ops = dev->netdev_ops; 1786 1787 /* 1788 * Call the device specific close. This cannot fail. 1789 * Only if device is UP 1790 * 1791 * We allow it to be called even after a DETACH hot-plug 1792 * event. 1793 */ 1794 1795 netdev_assert_locked_ops_compat(dev); 1796 1797 if (ops->ndo_stop) 1798 ops->ndo_stop(dev); 1799 1800 netif_rx_mode_cancel_retry(dev); 1801 netif_set_up(dev, false); 1802 netpoll_poll_enable(dev); 1803 } 1804 } 1805 1806 static void __dev_close(struct net_device *dev) 1807 { 1808 LIST_HEAD(single); 1809 1810 list_add(&dev->close_list, &single); 1811 __dev_close_many(&single); 1812 list_del(&single); 1813 } 1814 1815 void netif_close_many(struct list_head *head, bool unlink) 1816 { 1817 struct net_device *dev, *tmp; 1818 1819 /* Remove the devices that don't need to be closed */ 1820 list_for_each_entry_safe(dev, tmp, head, close_list) 1821 if (!(dev->flags & IFF_UP)) 1822 list_del_init(&dev->close_list); 1823 1824 __dev_close_many(head); 1825 1826 list_for_each_entry_safe(dev, tmp, head, close_list) { 1827 netdev_assert_locked_ops_compat(dev); 1828 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1829 call_netdevice_notifiers(NETDEV_DOWN, dev); 1830 if (unlink) 1831 list_del_init(&dev->close_list); 1832 } 1833 } 1834 EXPORT_SYMBOL_NS_GPL(netif_close_many, "NETDEV_INTERNAL"); 1835 1836 void netif_close(struct net_device *dev) 1837 { 1838 if (dev->flags & IFF_UP) { 1839 LIST_HEAD(single); 1840 1841 list_add(&dev->close_list, &single); 1842 netif_close_many(&single, true); 1843 list_del(&single); 1844 } 1845 } 1846 EXPORT_SYMBOL(netif_close); 1847 1848 void netif_disable_lro(struct net_device *dev) 1849 { 1850 struct net_device *lower_dev; 1851 struct list_head *iter; 1852 1853 dev->wanted_features &= ~NETIF_F_LRO; 1854 netdev_update_features(dev); 1855 1856 if (unlikely(dev->features & NETIF_F_LRO)) 1857 netdev_WARN(dev, "failed to disable LRO!\n"); 1858 1859 netdev_for_each_lower_dev(dev, lower_dev, iter) { 1860 netdev_lock_ops(lower_dev); 1861 netif_disable_lro(lower_dev); 1862 netdev_unlock_ops(lower_dev); 1863 } 1864 } 1865 1866 /** 1867 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device 1868 * @dev: device 1869 * 1870 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be 1871 * called under RTNL. This is needed if Generic XDP is installed on 1872 * the device. 1873 */ 1874 static void dev_disable_gro_hw(struct net_device *dev) 1875 { 1876 dev->wanted_features &= ~NETIF_F_GRO_HW; 1877 netdev_update_features(dev); 1878 1879 if (unlikely(dev->features & NETIF_F_GRO_HW)) 1880 netdev_WARN(dev, "failed to disable GRO_HW!\n"); 1881 } 1882 1883 const char *netdev_cmd_to_name(enum netdev_cmd cmd) 1884 { 1885 #define N(val) \ 1886 case NETDEV_##val: \ 1887 return "NETDEV_" __stringify(val); 1888 switch (cmd) { 1889 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER) 1890 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE) 1891 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE) 1892 N(POST_INIT) N(PRE_UNINIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) 1893 N(CHANGEUPPER) N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) 1894 N(BONDING_INFO) N(PRECHANGEUPPER) N(CHANGELOWERSTATE) 1895 N(UDP_TUNNEL_PUSH_INFO) N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN) 1896 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO) 1897 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO) 1898 N(PRE_CHANGEADDR) N(OFFLOAD_XSTATS_ENABLE) N(OFFLOAD_XSTATS_DISABLE) 1899 N(OFFLOAD_XSTATS_REPORT_USED) N(OFFLOAD_XSTATS_REPORT_DELTA) 1900 N(XDP_FEAT_CHANGE) 1901 } 1902 #undef N 1903 return "UNKNOWN_NETDEV_EVENT"; 1904 } 1905 EXPORT_SYMBOL_GPL(netdev_cmd_to_name); 1906 1907 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val, 1908 struct net_device *dev) 1909 { 1910 struct netdev_notifier_info info = { 1911 .dev = dev, 1912 }; 1913 1914 return nb->notifier_call(nb, val, &info); 1915 } 1916 1917 static int call_netdevice_register_notifiers(struct notifier_block *nb, 1918 struct net_device *dev) 1919 { 1920 int err; 1921 1922 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev); 1923 err = notifier_to_errno(err); 1924 if (err) 1925 return err; 1926 1927 if (!(dev->flags & IFF_UP)) 1928 return 0; 1929 1930 call_netdevice_notifier(nb, NETDEV_UP, dev); 1931 return 0; 1932 } 1933 1934 static void call_netdevice_unregister_notifiers(struct notifier_block *nb, 1935 struct net_device *dev) 1936 { 1937 if (dev->flags & IFF_UP) { 1938 netdev_lock_ops(dev); 1939 call_netdevice_notifier(nb, NETDEV_GOING_DOWN, 1940 dev); 1941 call_netdevice_notifier(nb, NETDEV_DOWN, dev); 1942 netdev_unlock_ops(dev); 1943 } 1944 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev); 1945 } 1946 1947 static int call_netdevice_register_net_notifiers(struct notifier_block *nb, 1948 struct net *net) 1949 { 1950 struct net_device *dev; 1951 int err; 1952 1953 for_each_netdev(net, dev) { 1954 netdev_lock_ops(dev); 1955 err = call_netdevice_register_notifiers(nb, dev); 1956 netdev_unlock_ops(dev); 1957 if (err) 1958 goto rollback; 1959 } 1960 return 0; 1961 1962 rollback: 1963 for_each_netdev_continue_reverse(net, dev) 1964 call_netdevice_unregister_notifiers(nb, dev); 1965 return err; 1966 } 1967 1968 static void call_netdevice_unregister_net_notifiers(struct notifier_block *nb, 1969 struct net *net) 1970 { 1971 struct net_device *dev; 1972 1973 for_each_netdev(net, dev) 1974 call_netdevice_unregister_notifiers(nb, dev); 1975 } 1976 1977 static int dev_boot_phase = 1; 1978 1979 /** 1980 * register_netdevice_notifier - register a network notifier block 1981 * @nb: notifier 1982 * 1983 * Register a notifier to be called when network device events occur. 1984 * The notifier passed is linked into the kernel structures and must 1985 * not be reused until it has been unregistered. A negative errno code 1986 * is returned on a failure. 1987 * 1988 * When registered all registration and up events are replayed 1989 * to the new notifier to allow device to have a race free 1990 * view of the network device list. 1991 */ 1992 1993 int register_netdevice_notifier(struct notifier_block *nb) 1994 { 1995 struct net *net; 1996 int err; 1997 1998 /* Close race with setup_net() and cleanup_net() */ 1999 down_write(&pernet_ops_rwsem); 2000 2001 /* When RTNL is removed, we need protection for netdev_chain. */ 2002 rtnl_lock(); 2003 2004 err = raw_notifier_chain_register(&netdev_chain, nb); 2005 if (err) 2006 goto unlock; 2007 if (dev_boot_phase) 2008 goto unlock; 2009 for_each_net(net) { 2010 __rtnl_net_lock(net); 2011 err = call_netdevice_register_net_notifiers(nb, net); 2012 __rtnl_net_unlock(net); 2013 if (err) 2014 goto rollback; 2015 } 2016 2017 unlock: 2018 rtnl_unlock(); 2019 up_write(&pernet_ops_rwsem); 2020 return err; 2021 2022 rollback: 2023 for_each_net_continue_reverse(net) { 2024 __rtnl_net_lock(net); 2025 call_netdevice_unregister_net_notifiers(nb, net); 2026 __rtnl_net_unlock(net); 2027 } 2028 2029 raw_notifier_chain_unregister(&netdev_chain, nb); 2030 goto unlock; 2031 } 2032 EXPORT_SYMBOL(register_netdevice_notifier); 2033 2034 /** 2035 * unregister_netdevice_notifier - unregister a network notifier block 2036 * @nb: notifier 2037 * 2038 * Unregister a notifier previously registered by 2039 * register_netdevice_notifier(). The notifier is unlinked into the 2040 * kernel structures and may then be reused. A negative errno code 2041 * is returned on a failure. 2042 * 2043 * After unregistering unregister and down device events are synthesized 2044 * for all devices on the device list to the removed notifier to remove 2045 * the need for special case cleanup code. 2046 */ 2047 2048 int unregister_netdevice_notifier(struct notifier_block *nb) 2049 { 2050 struct net *net; 2051 int err; 2052 2053 /* Close race with setup_net() and cleanup_net() */ 2054 down_write(&pernet_ops_rwsem); 2055 rtnl_lock(); 2056 err = raw_notifier_chain_unregister(&netdev_chain, nb); 2057 if (err) 2058 goto unlock; 2059 2060 for_each_net(net) { 2061 __rtnl_net_lock(net); 2062 call_netdevice_unregister_net_notifiers(nb, net); 2063 __rtnl_net_unlock(net); 2064 } 2065 2066 unlock: 2067 rtnl_unlock(); 2068 up_write(&pernet_ops_rwsem); 2069 return err; 2070 } 2071 EXPORT_SYMBOL(unregister_netdevice_notifier); 2072 2073 static int __register_netdevice_notifier_net(struct net *net, 2074 struct notifier_block *nb, 2075 bool ignore_call_fail) 2076 { 2077 int err; 2078 2079 err = raw_notifier_chain_register(&net->netdev_chain, nb); 2080 if (err) 2081 return err; 2082 if (dev_boot_phase) 2083 return 0; 2084 2085 err = call_netdevice_register_net_notifiers(nb, net); 2086 if (err && !ignore_call_fail) 2087 goto chain_unregister; 2088 2089 return 0; 2090 2091 chain_unregister: 2092 raw_notifier_chain_unregister(&net->netdev_chain, nb); 2093 return err; 2094 } 2095 2096 static int __unregister_netdevice_notifier_net(struct net *net, 2097 struct notifier_block *nb) 2098 { 2099 int err; 2100 2101 err = raw_notifier_chain_unregister(&net->netdev_chain, nb); 2102 if (err) 2103 return err; 2104 2105 call_netdevice_unregister_net_notifiers(nb, net); 2106 return 0; 2107 } 2108 2109 /** 2110 * register_netdevice_notifier_net - register a per-netns network notifier block 2111 * @net: network namespace 2112 * @nb: notifier 2113 * 2114 * Register a notifier to be called when network device events occur. 2115 * The notifier passed is linked into the kernel structures and must 2116 * not be reused until it has been unregistered. A negative errno code 2117 * is returned on a failure. 2118 * 2119 * When registered all registration and up events are replayed 2120 * to the new notifier to allow device to have a race free 2121 * view of the network device list. 2122 */ 2123 2124 int register_netdevice_notifier_net(struct net *net, struct notifier_block *nb) 2125 { 2126 int err; 2127 2128 rtnl_net_lock(net); 2129 err = __register_netdevice_notifier_net(net, nb, false); 2130 rtnl_net_unlock(net); 2131 2132 return err; 2133 } 2134 EXPORT_SYMBOL(register_netdevice_notifier_net); 2135 2136 /** 2137 * unregister_netdevice_notifier_net - unregister a per-netns 2138 * network notifier block 2139 * @net: network namespace 2140 * @nb: notifier 2141 * 2142 * Unregister a notifier previously registered by 2143 * register_netdevice_notifier_net(). The notifier is unlinked from the 2144 * kernel structures and may then be reused. A negative errno code 2145 * is returned on a failure. 2146 * 2147 * After unregistering unregister and down device events are synthesized 2148 * for all devices on the device list to the removed notifier to remove 2149 * the need for special case cleanup code. 2150 */ 2151 2152 int unregister_netdevice_notifier_net(struct net *net, 2153 struct notifier_block *nb) 2154 { 2155 int err; 2156 2157 rtnl_net_lock(net); 2158 err = __unregister_netdevice_notifier_net(net, nb); 2159 rtnl_net_unlock(net); 2160 2161 return err; 2162 } 2163 EXPORT_SYMBOL(unregister_netdevice_notifier_net); 2164 2165 static void __move_netdevice_notifier_net(struct net *src_net, 2166 struct net *dst_net, 2167 struct notifier_block *nb) 2168 { 2169 __unregister_netdevice_notifier_net(src_net, nb); 2170 __register_netdevice_notifier_net(dst_net, nb, true); 2171 } 2172 2173 static void rtnl_net_dev_lock(struct net_device *dev) 2174 { 2175 bool again; 2176 2177 do { 2178 struct net *net; 2179 2180 again = false; 2181 2182 /* netns might be being dismantled. */ 2183 rcu_read_lock(); 2184 net = dev_net_rcu(dev); 2185 net_passive_inc(net); 2186 rcu_read_unlock(); 2187 2188 rtnl_net_lock(net); 2189 2190 #ifdef CONFIG_NET_NS 2191 /* dev might have been moved to another netns. */ 2192 if (!net_eq(net, rcu_access_pointer(dev->nd_net.net))) { 2193 rtnl_net_unlock(net); 2194 net_passive_dec(net); 2195 again = true; 2196 } 2197 #endif 2198 } while (again); 2199 } 2200 2201 static void rtnl_net_dev_unlock(struct net_device *dev) 2202 { 2203 struct net *net = dev_net(dev); 2204 2205 rtnl_net_unlock(net); 2206 net_passive_dec(net); 2207 } 2208 2209 int register_netdevice_notifier_dev_net(struct net_device *dev, 2210 struct notifier_block *nb, 2211 struct netdev_net_notifier *nn) 2212 { 2213 int err; 2214 2215 rtnl_net_dev_lock(dev); 2216 err = __register_netdevice_notifier_net(dev_net(dev), nb, false); 2217 if (!err) { 2218 nn->nb = nb; 2219 list_add(&nn->list, &dev->net_notifier_list); 2220 } 2221 rtnl_net_dev_unlock(dev); 2222 2223 return err; 2224 } 2225 EXPORT_SYMBOL(register_netdevice_notifier_dev_net); 2226 2227 int unregister_netdevice_notifier_dev_net(struct net_device *dev, 2228 struct notifier_block *nb, 2229 struct netdev_net_notifier *nn) 2230 { 2231 int err; 2232 2233 rtnl_net_dev_lock(dev); 2234 list_del(&nn->list); 2235 err = __unregister_netdevice_notifier_net(dev_net(dev), nb); 2236 rtnl_net_dev_unlock(dev); 2237 2238 return err; 2239 } 2240 EXPORT_SYMBOL(unregister_netdevice_notifier_dev_net); 2241 2242 static void move_netdevice_notifiers_dev_net(struct net_device *dev, 2243 struct net *net) 2244 { 2245 struct netdev_net_notifier *nn; 2246 2247 list_for_each_entry(nn, &dev->net_notifier_list, list) 2248 __move_netdevice_notifier_net(dev_net(dev), net, nn->nb); 2249 } 2250 2251 /** 2252 * call_netdevice_notifiers_info - call all network notifier blocks 2253 * @val: value passed unmodified to notifier function 2254 * @info: notifier information data 2255 * 2256 * Call all network notifier blocks. Parameters and return value 2257 * are as for raw_notifier_call_chain(). 2258 */ 2259 2260 int call_netdevice_notifiers_info(unsigned long val, 2261 struct netdev_notifier_info *info) 2262 { 2263 struct net *net = dev_net(info->dev); 2264 int ret; 2265 2266 ASSERT_RTNL(); 2267 2268 /* Run per-netns notifier block chain first, then run the global one. 2269 * Hopefully, one day, the global one is going to be removed after 2270 * all notifier block registrators get converted to be per-netns. 2271 */ 2272 ret = raw_notifier_call_chain(&net->netdev_chain, val, info); 2273 if (ret & NOTIFY_STOP_MASK) 2274 return ret; 2275 return raw_notifier_call_chain(&netdev_chain, val, info); 2276 } 2277 2278 /** 2279 * call_netdevice_notifiers_info_robust - call per-netns notifier blocks 2280 * for and rollback on error 2281 * @val_up: value passed unmodified to notifier function 2282 * @val_down: value passed unmodified to the notifier function when 2283 * recovering from an error on @val_up 2284 * @info: notifier information data 2285 * 2286 * Call all per-netns network notifier blocks, but not notifier blocks on 2287 * the global notifier chain. Parameters and return value are as for 2288 * raw_notifier_call_chain_robust(). 2289 */ 2290 2291 static int 2292 call_netdevice_notifiers_info_robust(unsigned long val_up, 2293 unsigned long val_down, 2294 struct netdev_notifier_info *info) 2295 { 2296 struct net *net = dev_net(info->dev); 2297 2298 ASSERT_RTNL(); 2299 2300 return raw_notifier_call_chain_robust(&net->netdev_chain, 2301 val_up, val_down, info); 2302 } 2303 2304 static int call_netdevice_notifiers_extack(unsigned long val, 2305 struct net_device *dev, 2306 struct netlink_ext_ack *extack) 2307 { 2308 struct netdev_notifier_info info = { 2309 .dev = dev, 2310 .extack = extack, 2311 }; 2312 2313 return call_netdevice_notifiers_info(val, &info); 2314 } 2315 2316 /** 2317 * call_netdevice_notifiers - call all network notifier blocks 2318 * @val: value passed unmodified to notifier function 2319 * @dev: net_device pointer passed unmodified to notifier function 2320 * 2321 * Call all network notifier blocks. Parameters and return value 2322 * are as for raw_notifier_call_chain(). 2323 */ 2324 2325 int call_netdevice_notifiers(unsigned long val, struct net_device *dev) 2326 { 2327 return call_netdevice_notifiers_extack(val, dev, NULL); 2328 } 2329 EXPORT_SYMBOL(call_netdevice_notifiers); 2330 2331 /** 2332 * call_netdevice_notifiers_mtu - call all network notifier blocks 2333 * @val: value passed unmodified to notifier function 2334 * @dev: net_device pointer passed unmodified to notifier function 2335 * @arg: additional u32 argument passed to the notifier function 2336 * 2337 * Call all network notifier blocks. Parameters and return value 2338 * are as for raw_notifier_call_chain(). 2339 */ 2340 static int call_netdevice_notifiers_mtu(unsigned long val, 2341 struct net_device *dev, u32 arg) 2342 { 2343 struct netdev_notifier_info_ext info = { 2344 .info.dev = dev, 2345 .ext.mtu = arg, 2346 }; 2347 2348 BUILD_BUG_ON(offsetof(struct netdev_notifier_info_ext, info) != 0); 2349 2350 return call_netdevice_notifiers_info(val, &info.info); 2351 } 2352 2353 #ifdef CONFIG_NET_INGRESS 2354 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key); 2355 2356 void net_inc_ingress_queue(void) 2357 { 2358 static_branch_inc(&ingress_needed_key); 2359 } 2360 EXPORT_SYMBOL_GPL(net_inc_ingress_queue); 2361 2362 void net_dec_ingress_queue(void) 2363 { 2364 static_branch_dec(&ingress_needed_key); 2365 } 2366 EXPORT_SYMBOL_GPL(net_dec_ingress_queue); 2367 #endif 2368 2369 #ifdef CONFIG_NET_EGRESS 2370 static DEFINE_STATIC_KEY_FALSE(egress_needed_key); 2371 2372 void net_inc_egress_queue(void) 2373 { 2374 static_branch_inc(&egress_needed_key); 2375 } 2376 EXPORT_SYMBOL_GPL(net_inc_egress_queue); 2377 2378 void net_dec_egress_queue(void) 2379 { 2380 static_branch_dec(&egress_needed_key); 2381 } 2382 EXPORT_SYMBOL_GPL(net_dec_egress_queue); 2383 #endif 2384 2385 #ifdef CONFIG_NET_CLS_ACT 2386 DEFINE_STATIC_KEY_FALSE(tcf_sw_enabled_key); 2387 EXPORT_SYMBOL(tcf_sw_enabled_key); 2388 #endif 2389 2390 DEFINE_STATIC_KEY_FALSE(netstamp_needed_key); 2391 EXPORT_SYMBOL(netstamp_needed_key); 2392 #ifdef CONFIG_JUMP_LABEL 2393 static atomic_t netstamp_needed_deferred; 2394 static atomic_t netstamp_wanted; 2395 static void netstamp_clear(struct work_struct *work) 2396 { 2397 int deferred = atomic_xchg(&netstamp_needed_deferred, 0); 2398 int wanted; 2399 2400 wanted = atomic_add_return(deferred, &netstamp_wanted); 2401 if (wanted > 0) 2402 static_branch_enable(&netstamp_needed_key); 2403 else 2404 static_branch_disable(&netstamp_needed_key); 2405 } 2406 static DECLARE_WORK(netstamp_work, netstamp_clear); 2407 #endif 2408 2409 void net_enable_timestamp(void) 2410 { 2411 #ifdef CONFIG_JUMP_LABEL 2412 int wanted = atomic_read(&netstamp_wanted); 2413 2414 while (wanted > 0) { 2415 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted + 1)) 2416 return; 2417 } 2418 atomic_inc(&netstamp_needed_deferred); 2419 schedule_work(&netstamp_work); 2420 #else 2421 static_branch_inc(&netstamp_needed_key); 2422 #endif 2423 } 2424 EXPORT_SYMBOL(net_enable_timestamp); 2425 2426 void net_disable_timestamp(void) 2427 { 2428 #ifdef CONFIG_JUMP_LABEL 2429 int wanted = atomic_read(&netstamp_wanted); 2430 2431 while (wanted > 1) { 2432 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted - 1)) 2433 return; 2434 } 2435 atomic_dec(&netstamp_needed_deferred); 2436 schedule_work(&netstamp_work); 2437 #else 2438 static_branch_dec(&netstamp_needed_key); 2439 #endif 2440 } 2441 EXPORT_SYMBOL(net_disable_timestamp); 2442 2443 static inline void net_timestamp_set(struct sk_buff *skb) 2444 { 2445 skb->tstamp = 0; 2446 skb->tstamp_type = SKB_CLOCK_REALTIME; 2447 if (static_branch_unlikely(&netstamp_needed_key)) 2448 skb->tstamp = ktime_get_real(); 2449 } 2450 2451 #define net_timestamp_check(COND, SKB) \ 2452 if (static_branch_unlikely(&netstamp_needed_key)) { \ 2453 if ((COND) && !(SKB)->tstamp) \ 2454 (SKB)->tstamp = ktime_get_real(); \ 2455 } \ 2456 2457 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb) 2458 { 2459 return __is_skb_forwardable(dev, skb, true); 2460 } 2461 EXPORT_SYMBOL_GPL(is_skb_forwardable); 2462 2463 static int __dev_forward_skb2(struct net_device *dev, struct sk_buff *skb, 2464 bool check_mtu) 2465 { 2466 int ret = ____dev_forward_skb(dev, skb, check_mtu); 2467 2468 if (likely(!ret)) { 2469 skb->protocol = eth_type_trans(skb, dev); 2470 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN); 2471 } 2472 2473 return ret; 2474 } 2475 2476 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2477 { 2478 return __dev_forward_skb2(dev, skb, true); 2479 } 2480 EXPORT_SYMBOL_GPL(__dev_forward_skb); 2481 2482 /** 2483 * dev_forward_skb - loopback an skb to another netif 2484 * 2485 * @dev: destination network device 2486 * @skb: buffer to forward 2487 * 2488 * return values: 2489 * NET_RX_SUCCESS (no congestion) 2490 * NET_RX_DROP (packet was dropped, but freed) 2491 * 2492 * dev_forward_skb can be used for injecting an skb from the 2493 * start_xmit function of one device into the receive queue 2494 * of another device. 2495 * 2496 * The receiving device may be in another namespace, so 2497 * we have to clear all information in the skb that could 2498 * impact namespace isolation. 2499 */ 2500 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2501 { 2502 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb); 2503 } 2504 EXPORT_SYMBOL_GPL(dev_forward_skb); 2505 2506 int dev_forward_skb_nomtu(struct net_device *dev, struct sk_buff *skb) 2507 { 2508 return __dev_forward_skb2(dev, skb, false) ?: netif_rx_internal(skb); 2509 } 2510 2511 static int deliver_skb(struct sk_buff *skb, 2512 struct packet_type *pt_prev, 2513 struct net_device *orig_dev) 2514 { 2515 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 2516 return -ENOMEM; 2517 refcount_inc(&skb->users); 2518 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 2519 } 2520 2521 static inline void deliver_ptype_list_skb(struct sk_buff *skb, 2522 struct packet_type **pt, 2523 struct net_device *orig_dev, 2524 __be16 type, 2525 struct list_head *ptype_list) 2526 { 2527 struct packet_type *ptype, *pt_prev = *pt; 2528 2529 list_for_each_entry_rcu(ptype, ptype_list, list) { 2530 if (ptype->type != type) 2531 continue; 2532 if (unlikely(pt_prev)) 2533 deliver_skb(skb, pt_prev, orig_dev); 2534 pt_prev = ptype; 2535 } 2536 *pt = pt_prev; 2537 } 2538 2539 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb) 2540 { 2541 if (!ptype->af_packet_priv || !skb->sk) 2542 return false; 2543 2544 if (ptype->id_match) 2545 return ptype->id_match(ptype, skb->sk); 2546 else if ((struct sock *)ptype->af_packet_priv == skb->sk) 2547 return true; 2548 2549 return false; 2550 } 2551 2552 /** 2553 * dev_nit_active_rcu - return true if any network interface taps are in use 2554 * 2555 * The caller must hold the RCU lock 2556 * 2557 * @dev: network device to check for the presence of taps 2558 */ 2559 bool dev_nit_active_rcu(const struct net_device *dev) 2560 { 2561 /* Callers may hold either RCU or RCU BH lock */ 2562 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 2563 2564 return !list_empty(&dev_net(dev)->ptype_all) || 2565 !list_empty(&dev->ptype_all); 2566 } 2567 EXPORT_SYMBOL_GPL(dev_nit_active_rcu); 2568 2569 /* 2570 * Support routine. Sends outgoing frames to any network 2571 * taps currently in use. 2572 */ 2573 2574 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev) 2575 { 2576 struct packet_type *ptype, *pt_prev = NULL; 2577 struct list_head *ptype_list; 2578 struct sk_buff *skb2 = NULL; 2579 2580 rcu_read_lock(); 2581 ptype_list = &dev_net_rcu(dev)->ptype_all; 2582 again: 2583 list_for_each_entry_rcu(ptype, ptype_list, list) { 2584 if (READ_ONCE(ptype->ignore_outgoing)) 2585 continue; 2586 2587 /* Never send packets back to the socket 2588 * they originated from - MvS (miquels@drinkel.ow.org) 2589 */ 2590 if (skb_loop_sk(ptype, skb)) 2591 continue; 2592 2593 if (unlikely(pt_prev)) { 2594 deliver_skb(skb2, pt_prev, skb->dev); 2595 pt_prev = ptype; 2596 continue; 2597 } 2598 2599 /* need to clone skb, done only once */ 2600 skb2 = skb_clone(skb, GFP_ATOMIC); 2601 if (!skb2) 2602 goto out_unlock; 2603 2604 net_timestamp_set(skb2); 2605 2606 /* skb->nh should be correctly 2607 * set by sender, so that the second statement is 2608 * just protection against buggy protocols. 2609 */ 2610 skb_reset_mac_header(skb2); 2611 2612 if (skb_network_header(skb2) < skb2->data || 2613 skb_network_header(skb2) > skb_tail_pointer(skb2)) { 2614 net_crit_ratelimited("protocol %04x is buggy, dev %s\n", 2615 ntohs(skb2->protocol), 2616 dev->name); 2617 skb_reset_network_header(skb2); 2618 } 2619 2620 skb2->transport_header = skb2->network_header; 2621 skb2->pkt_type = PACKET_OUTGOING; 2622 pt_prev = ptype; 2623 } 2624 2625 if (ptype_list != &dev->ptype_all) { 2626 ptype_list = &dev->ptype_all; 2627 goto again; 2628 } 2629 out_unlock: 2630 if (pt_prev) { 2631 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC)) 2632 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev); 2633 else 2634 kfree_skb(skb2); 2635 } 2636 rcu_read_unlock(); 2637 } 2638 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit); 2639 2640 /** 2641 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change 2642 * @dev: Network device 2643 * @txq: number of queues available 2644 * 2645 * If real_num_tx_queues is changed the tc mappings may no longer be 2646 * valid. To resolve this verify the tc mapping remains valid and if 2647 * not NULL the mapping. With no priorities mapping to this 2648 * offset/count pair it will no longer be used. In the worst case TC0 2649 * is invalid nothing can be done so disable priority mappings. If is 2650 * expected that drivers will fix this mapping if they can before 2651 * calling netif_set_real_num_tx_queues. 2652 */ 2653 static void netif_setup_tc(struct net_device *dev, unsigned int txq) 2654 { 2655 struct netdev_tc_txq res; 2656 int i; 2657 2658 res.combined = READ_ONCE(dev->tc_to_txq[0].combined); 2659 2660 /* If TC0 is invalidated disable TC mapping */ 2661 if (res.offset + res.count > txq) { 2662 netdev_warn(dev, "Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n"); 2663 dev->num_tc = 0; 2664 return; 2665 } 2666 2667 /* Invalidated prio to tc mappings set to TC0 */ 2668 for (i = 1; i < TC_BITMASK + 1; i++) { 2669 int q = netdev_get_prio_tc_map(dev, i); 2670 2671 res.combined = READ_ONCE(dev->tc_to_txq[q].combined); 2672 if (res.offset + res.count > txq) { 2673 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", 2674 i, q); 2675 netdev_set_prio_tc_map(dev, i, 0); 2676 } 2677 } 2678 } 2679 2680 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq) 2681 { 2682 if (dev->num_tc) { 2683 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2684 int i; 2685 2686 /* walk through the TCs and see if it falls into any of them */ 2687 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) { 2688 struct netdev_tc_txq res; 2689 2690 res.combined = READ_ONCE(tc->combined); 2691 if ((txq - res.offset) < res.count) 2692 return i; 2693 } 2694 2695 /* didn't find it, just return -1 to indicate no match */ 2696 return -1; 2697 } 2698 2699 return 0; 2700 } 2701 EXPORT_SYMBOL(netdev_txq_to_tc); 2702 2703 #ifdef CONFIG_XPS 2704 static struct static_key xps_needed __read_mostly; 2705 static struct static_key xps_rxqs_needed __read_mostly; 2706 static DEFINE_MUTEX(xps_map_mutex); 2707 #define xmap_dereference(P) \ 2708 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex)) 2709 2710 static bool remove_xps_queue(struct xps_dev_maps *dev_maps, 2711 struct xps_dev_maps *old_maps, int tci, u16 index) 2712 { 2713 struct xps_map *map = NULL; 2714 int pos; 2715 2716 map = xmap_dereference(dev_maps->attr_map[tci]); 2717 if (!map) 2718 return false; 2719 2720 for (pos = map->len; pos--;) { 2721 if (map->queues[pos] != index) 2722 continue; 2723 2724 if (map->len > 1) { 2725 map->queues[pos] = map->queues[--map->len]; 2726 break; 2727 } 2728 2729 if (old_maps) 2730 RCU_INIT_POINTER(old_maps->attr_map[tci], NULL); 2731 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2732 kfree_rcu(map, rcu); 2733 return false; 2734 } 2735 2736 return true; 2737 } 2738 2739 static bool remove_xps_queue_cpu(struct net_device *dev, 2740 struct xps_dev_maps *dev_maps, 2741 int cpu, u16 offset, u16 count) 2742 { 2743 int num_tc = dev_maps->num_tc; 2744 bool active = false; 2745 int tci; 2746 2747 for (tci = cpu * num_tc; num_tc--; tci++) { 2748 int i, j; 2749 2750 for (i = count, j = offset; i--; j++) { 2751 if (!remove_xps_queue(dev_maps, NULL, tci, j)) 2752 break; 2753 } 2754 2755 active |= i < 0; 2756 } 2757 2758 return active; 2759 } 2760 2761 static void reset_xps_maps(struct net_device *dev, 2762 struct xps_dev_maps *dev_maps, 2763 enum xps_map_type type) 2764 { 2765 static_key_slow_dec_cpuslocked(&xps_needed); 2766 if (type == XPS_RXQS) 2767 static_key_slow_dec_cpuslocked(&xps_rxqs_needed); 2768 2769 RCU_INIT_POINTER(dev->xps_maps[type], NULL); 2770 2771 kfree_rcu(dev_maps, rcu); 2772 } 2773 2774 static void clean_xps_maps(struct net_device *dev, enum xps_map_type type, 2775 u16 offset, u16 count) 2776 { 2777 struct xps_dev_maps *dev_maps; 2778 bool active = false; 2779 int i, j; 2780 2781 dev_maps = xmap_dereference(dev->xps_maps[type]); 2782 if (!dev_maps) 2783 return; 2784 2785 for (j = 0; j < dev_maps->nr_ids; j++) 2786 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, count); 2787 if (!active) 2788 reset_xps_maps(dev, dev_maps, type); 2789 2790 if (type == XPS_CPUS) { 2791 for (i = offset + (count - 1); count--; i--) 2792 netdev_queue_numa_node_write( 2793 netdev_get_tx_queue(dev, i), NUMA_NO_NODE); 2794 } 2795 } 2796 2797 static void netif_reset_xps_queues(struct net_device *dev, u16 offset, 2798 u16 count) 2799 { 2800 if (!static_key_false(&xps_needed)) 2801 return; 2802 2803 cpus_read_lock(); 2804 mutex_lock(&xps_map_mutex); 2805 2806 if (static_key_false(&xps_rxqs_needed)) 2807 clean_xps_maps(dev, XPS_RXQS, offset, count); 2808 2809 clean_xps_maps(dev, XPS_CPUS, offset, count); 2810 2811 mutex_unlock(&xps_map_mutex); 2812 cpus_read_unlock(); 2813 } 2814 2815 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index) 2816 { 2817 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index); 2818 } 2819 2820 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index, 2821 u16 index, bool is_rxqs_map) 2822 { 2823 struct xps_map *new_map; 2824 int alloc_len = XPS_MIN_MAP_ALLOC; 2825 int i, pos; 2826 2827 for (pos = 0; map && pos < map->len; pos++) { 2828 if (map->queues[pos] != index) 2829 continue; 2830 return map; 2831 } 2832 2833 /* Need to add tx-queue to this CPU's/rx-queue's existing map */ 2834 if (map) { 2835 if (pos < map->alloc_len) 2836 return map; 2837 2838 alloc_len = map->alloc_len * 2; 2839 } 2840 2841 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's 2842 * map 2843 */ 2844 if (is_rxqs_map) 2845 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL); 2846 else 2847 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL, 2848 cpu_to_node(attr_index)); 2849 if (!new_map) 2850 return NULL; 2851 2852 for (i = 0; i < pos; i++) 2853 new_map->queues[i] = map->queues[i]; 2854 new_map->alloc_len = alloc_len; 2855 new_map->len = pos; 2856 2857 return new_map; 2858 } 2859 2860 /* Copy xps maps at a given index */ 2861 static void xps_copy_dev_maps(struct xps_dev_maps *dev_maps, 2862 struct xps_dev_maps *new_dev_maps, int index, 2863 int tc, bool skip_tc) 2864 { 2865 int i, tci = index * dev_maps->num_tc; 2866 struct xps_map *map; 2867 2868 /* copy maps belonging to foreign traffic classes */ 2869 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2870 if (i == tc && skip_tc) 2871 continue; 2872 2873 /* fill in the new device map from the old device map */ 2874 map = xmap_dereference(dev_maps->attr_map[tci]); 2875 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2876 } 2877 } 2878 2879 /* Must be called under cpus_read_lock */ 2880 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask, 2881 u16 index, enum xps_map_type type) 2882 { 2883 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL, *old_dev_maps = NULL; 2884 const unsigned long *online_mask = NULL; 2885 bool active = false, copy = false; 2886 int i, j, tci, numa_node_id = -2; 2887 int maps_sz, num_tc = 1, tc = 0; 2888 struct xps_map *map, *new_map; 2889 unsigned int nr_ids; 2890 2891 WARN_ON_ONCE(index >= dev->num_tx_queues); 2892 2893 if (dev->num_tc) { 2894 /* Do not allow XPS on subordinate device directly */ 2895 num_tc = dev->num_tc; 2896 if (num_tc < 0) 2897 return -EINVAL; 2898 2899 /* If queue belongs to subordinate dev use its map */ 2900 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev; 2901 2902 tc = netdev_txq_to_tc(dev, index); 2903 if (tc < 0) 2904 return -EINVAL; 2905 } 2906 2907 mutex_lock(&xps_map_mutex); 2908 2909 dev_maps = xmap_dereference(dev->xps_maps[type]); 2910 if (type == XPS_RXQS) { 2911 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues); 2912 nr_ids = dev->num_rx_queues; 2913 } else { 2914 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc); 2915 if (num_possible_cpus() > 1) 2916 online_mask = cpumask_bits(cpu_online_mask); 2917 nr_ids = nr_cpu_ids; 2918 } 2919 2920 if (maps_sz < L1_CACHE_BYTES) 2921 maps_sz = L1_CACHE_BYTES; 2922 2923 /* The old dev_maps could be larger or smaller than the one we're 2924 * setting up now, as dev->num_tc or nr_ids could have been updated in 2925 * between. We could try to be smart, but let's be safe instead and only 2926 * copy foreign traffic classes if the two map sizes match. 2927 */ 2928 if (dev_maps && 2929 dev_maps->num_tc == num_tc && dev_maps->nr_ids == nr_ids) 2930 copy = true; 2931 2932 /* allocate memory for queue storage */ 2933 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids), 2934 j < nr_ids;) { 2935 if (!new_dev_maps) { 2936 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL); 2937 if (!new_dev_maps) { 2938 mutex_unlock(&xps_map_mutex); 2939 return -ENOMEM; 2940 } 2941 2942 new_dev_maps->nr_ids = nr_ids; 2943 new_dev_maps->num_tc = num_tc; 2944 } 2945 2946 tci = j * num_tc + tc; 2947 map = copy ? xmap_dereference(dev_maps->attr_map[tci]) : NULL; 2948 2949 map = expand_xps_map(map, j, index, type == XPS_RXQS); 2950 if (!map) 2951 goto error; 2952 2953 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2954 } 2955 2956 if (!new_dev_maps) 2957 goto out_no_new_maps; 2958 2959 if (!dev_maps) { 2960 /* Increment static keys at most once per type */ 2961 static_key_slow_inc_cpuslocked(&xps_needed); 2962 if (type == XPS_RXQS) 2963 static_key_slow_inc_cpuslocked(&xps_rxqs_needed); 2964 } 2965 2966 for (j = 0; j < nr_ids; j++) { 2967 bool skip_tc = false; 2968 2969 tci = j * num_tc + tc; 2970 if (netif_attr_test_mask(j, mask, nr_ids) && 2971 netif_attr_test_online(j, online_mask, nr_ids)) { 2972 /* add tx-queue to CPU/rx-queue maps */ 2973 int pos = 0; 2974 2975 skip_tc = true; 2976 2977 map = xmap_dereference(new_dev_maps->attr_map[tci]); 2978 while ((pos < map->len) && (map->queues[pos] != index)) 2979 pos++; 2980 2981 if (pos == map->len) 2982 map->queues[map->len++] = index; 2983 #ifdef CONFIG_NUMA 2984 if (type == XPS_CPUS) { 2985 if (numa_node_id == -2) 2986 numa_node_id = cpu_to_node(j); 2987 else if (numa_node_id != cpu_to_node(j)) 2988 numa_node_id = -1; 2989 } 2990 #endif 2991 } 2992 2993 if (copy) 2994 xps_copy_dev_maps(dev_maps, new_dev_maps, j, tc, 2995 skip_tc); 2996 } 2997 2998 rcu_assign_pointer(dev->xps_maps[type], new_dev_maps); 2999 3000 /* Cleanup old maps */ 3001 if (!dev_maps) 3002 goto out_no_old_maps; 3003 3004 for (j = 0; j < dev_maps->nr_ids; j++) { 3005 for (i = num_tc, tci = j * dev_maps->num_tc; i--; tci++) { 3006 map = xmap_dereference(dev_maps->attr_map[tci]); 3007 if (!map) 3008 continue; 3009 3010 if (copy) { 3011 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 3012 if (map == new_map) 3013 continue; 3014 } 3015 3016 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 3017 kfree_rcu(map, rcu); 3018 } 3019 } 3020 3021 old_dev_maps = dev_maps; 3022 3023 out_no_old_maps: 3024 dev_maps = new_dev_maps; 3025 active = true; 3026 3027 out_no_new_maps: 3028 if (type == XPS_CPUS) 3029 /* update Tx queue numa node */ 3030 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index), 3031 (numa_node_id >= 0) ? 3032 numa_node_id : NUMA_NO_NODE); 3033 3034 if (!dev_maps) 3035 goto out_no_maps; 3036 3037 /* removes tx-queue from unused CPUs/rx-queues */ 3038 for (j = 0; j < dev_maps->nr_ids; j++) { 3039 tci = j * dev_maps->num_tc; 3040 3041 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 3042 if (i == tc && 3043 netif_attr_test_mask(j, mask, dev_maps->nr_ids) && 3044 netif_attr_test_online(j, online_mask, dev_maps->nr_ids)) 3045 continue; 3046 3047 active |= remove_xps_queue(dev_maps, 3048 copy ? old_dev_maps : NULL, 3049 tci, index); 3050 } 3051 } 3052 3053 if (old_dev_maps) 3054 kfree_rcu(old_dev_maps, rcu); 3055 3056 /* free map if not active */ 3057 if (!active) 3058 reset_xps_maps(dev, dev_maps, type); 3059 3060 out_no_maps: 3061 mutex_unlock(&xps_map_mutex); 3062 3063 return 0; 3064 error: 3065 /* remove any maps that we added */ 3066 for (j = 0; j < nr_ids; j++) { 3067 for (i = num_tc, tci = j * num_tc; i--; tci++) { 3068 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 3069 map = copy ? 3070 xmap_dereference(dev_maps->attr_map[tci]) : 3071 NULL; 3072 if (new_map && new_map != map) 3073 kfree(new_map); 3074 } 3075 } 3076 3077 mutex_unlock(&xps_map_mutex); 3078 3079 kfree(new_dev_maps); 3080 return -ENOMEM; 3081 } 3082 EXPORT_SYMBOL_GPL(__netif_set_xps_queue); 3083 3084 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, 3085 u16 index) 3086 { 3087 int ret; 3088 3089 cpus_read_lock(); 3090 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, XPS_CPUS); 3091 cpus_read_unlock(); 3092 3093 return ret; 3094 } 3095 EXPORT_SYMBOL(netif_set_xps_queue); 3096 3097 #endif 3098 static void netdev_unbind_all_sb_channels(struct net_device *dev) 3099 { 3100 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 3101 3102 /* Unbind any subordinate channels */ 3103 while (txq-- != &dev->_tx[0]) { 3104 if (txq->sb_dev) 3105 netdev_unbind_sb_channel(dev, txq->sb_dev); 3106 } 3107 } 3108 3109 void netdev_reset_tc(struct net_device *dev) 3110 { 3111 int i; 3112 3113 #ifdef CONFIG_XPS 3114 netif_reset_xps_queues_gt(dev, 0); 3115 #endif 3116 netdev_unbind_all_sb_channels(dev); 3117 3118 /* Reset TC configuration of device */ 3119 dev->num_tc = 0; 3120 for (i = 0; i < TC_MAX_QUEUE; i++) 3121 WRITE_ONCE(dev->tc_to_txq[i].combined, 0); 3122 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map)); 3123 } 3124 EXPORT_SYMBOL(netdev_reset_tc); 3125 3126 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset) 3127 { 3128 struct netdev_tc_txq res = { 3129 .count = count, 3130 .offset = offset, 3131 }; 3132 3133 if (tc >= dev->num_tc) 3134 return -EINVAL; 3135 3136 #ifdef CONFIG_XPS 3137 netif_reset_xps_queues(dev, offset, count); 3138 #endif 3139 WRITE_ONCE(dev->tc_to_txq[tc].combined, res.combined); 3140 return 0; 3141 } 3142 EXPORT_SYMBOL(netdev_set_tc_queue); 3143 3144 int netdev_set_num_tc(struct net_device *dev, u8 num_tc) 3145 { 3146 if (num_tc > TC_MAX_QUEUE) 3147 return -EINVAL; 3148 3149 #ifdef CONFIG_XPS 3150 netif_reset_xps_queues_gt(dev, 0); 3151 #endif 3152 netdev_unbind_all_sb_channels(dev); 3153 3154 dev->num_tc = num_tc; 3155 return 0; 3156 } 3157 EXPORT_SYMBOL(netdev_set_num_tc); 3158 3159 void netdev_unbind_sb_channel(struct net_device *dev, 3160 struct net_device *sb_dev) 3161 { 3162 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 3163 int i; 3164 3165 #ifdef CONFIG_XPS 3166 netif_reset_xps_queues_gt(sb_dev, 0); 3167 #endif 3168 for (i = 0; i < TC_MAX_QUEUE; i++) 3169 WRITE_ONCE(sb_dev->tc_to_txq[i].combined, 0); 3170 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map)); 3171 3172 while (txq-- != &dev->_tx[0]) { 3173 if (txq->sb_dev == sb_dev) 3174 txq->sb_dev = NULL; 3175 } 3176 } 3177 EXPORT_SYMBOL(netdev_unbind_sb_channel); 3178 3179 int netdev_bind_sb_channel_queue(struct net_device *dev, 3180 struct net_device *sb_dev, 3181 u8 tc, u16 count, u16 offset) 3182 { 3183 /* Make certain the sb_dev and dev are already configured */ 3184 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc) 3185 return -EINVAL; 3186 3187 /* We cannot hand out queues we don't have */ 3188 if ((offset + count) > dev->real_num_tx_queues) 3189 return -EINVAL; 3190 3191 /* Record the mapping */ 3192 struct netdev_tc_txq res = { 3193 .count = count, 3194 .offset = offset, 3195 }; 3196 3197 WRITE_ONCE(sb_dev->tc_to_txq[tc].combined, res.combined); 3198 3199 /* Provide a way for Tx queue to find the tc_to_txq map or 3200 * XPS map for itself. 3201 */ 3202 while (count--) 3203 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev; 3204 3205 return 0; 3206 } 3207 EXPORT_SYMBOL(netdev_bind_sb_channel_queue); 3208 3209 int netdev_set_sb_channel(struct net_device *dev, u16 channel) 3210 { 3211 /* Do not use a multiqueue device to represent a subordinate channel */ 3212 if (netif_is_multiqueue(dev)) 3213 return -ENODEV; 3214 3215 /* We allow channels 1 - 32767 to be used for subordinate channels. 3216 * Channel 0 is meant to be "native" mode and used only to represent 3217 * the main root device. We allow writing 0 to reset the device back 3218 * to normal mode after being used as a subordinate channel. 3219 */ 3220 if (channel > S16_MAX) 3221 return -EINVAL; 3222 3223 dev->num_tc = -channel; 3224 3225 return 0; 3226 } 3227 EXPORT_SYMBOL(netdev_set_sb_channel); 3228 3229 /* 3230 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues 3231 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed. 3232 */ 3233 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq) 3234 { 3235 bool disabling; 3236 int rc; 3237 3238 disabling = txq < dev->real_num_tx_queues; 3239 3240 if (txq < 1 || txq > dev->num_tx_queues) 3241 return -EINVAL; 3242 3243 if (dev->reg_state == NETREG_REGISTERED || 3244 dev->reg_state == NETREG_UNREGISTERING) { 3245 netdev_assert_locked_ops_compat(dev); 3246 3247 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues, 3248 txq); 3249 if (rc) 3250 return rc; 3251 3252 if (dev->num_tc) 3253 netif_setup_tc(dev, txq); 3254 3255 net_shaper_set_real_num_tx_queues(dev, txq); 3256 3257 dev_qdisc_change_real_num_tx(dev, txq); 3258 3259 dev->real_num_tx_queues = txq; 3260 3261 if (disabling) { 3262 synchronize_net(); 3263 qdisc_reset_all_tx_gt(dev, txq); 3264 #ifdef CONFIG_XPS 3265 netif_reset_xps_queues_gt(dev, txq); 3266 #endif 3267 } 3268 } else { 3269 dev->real_num_tx_queues = txq; 3270 } 3271 3272 return 0; 3273 } 3274 EXPORT_SYMBOL(netif_set_real_num_tx_queues); 3275 3276 /** 3277 * netif_set_real_num_rx_queues - set actual number of RX queues used 3278 * @dev: Network device 3279 * @rxq: Actual number of RX queues 3280 * 3281 * This must be called either with the rtnl_lock held or before 3282 * registration of the net device. Returns 0 on success, or a 3283 * negative error code. If called before registration, it always 3284 * succeeds. 3285 */ 3286 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq) 3287 { 3288 int rc; 3289 3290 if (rxq < 1 || rxq > dev->num_rx_queues) 3291 return -EINVAL; 3292 3293 if (dev->reg_state == NETREG_REGISTERED) { 3294 netdev_assert_locked_ops_compat(dev); 3295 3296 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues, 3297 rxq); 3298 if (rc) 3299 return rc; 3300 } 3301 3302 dev->real_num_rx_queues = rxq; 3303 return 0; 3304 } 3305 EXPORT_SYMBOL(netif_set_real_num_rx_queues); 3306 3307 /** 3308 * netif_set_real_num_queues - set actual number of RX and TX queues used 3309 * @dev: Network device 3310 * @txq: Actual number of TX queues 3311 * @rxq: Actual number of RX queues 3312 * 3313 * Set the real number of both TX and RX queues. 3314 * Does nothing if the number of queues is already correct. 3315 */ 3316 int netif_set_real_num_queues(struct net_device *dev, 3317 unsigned int txq, unsigned int rxq) 3318 { 3319 unsigned int old_rxq = dev->real_num_rx_queues; 3320 int err; 3321 3322 if (txq < 1 || txq > dev->num_tx_queues || 3323 rxq < 1 || rxq > dev->num_rx_queues) 3324 return -EINVAL; 3325 3326 /* Start from increases, so the error path only does decreases - 3327 * decreases can't fail. 3328 */ 3329 if (rxq > dev->real_num_rx_queues) { 3330 err = netif_set_real_num_rx_queues(dev, rxq); 3331 if (err) 3332 return err; 3333 } 3334 if (txq > dev->real_num_tx_queues) { 3335 err = netif_set_real_num_tx_queues(dev, txq); 3336 if (err) 3337 goto undo_rx; 3338 } 3339 if (rxq < dev->real_num_rx_queues) 3340 WARN_ON(netif_set_real_num_rx_queues(dev, rxq)); 3341 if (txq < dev->real_num_tx_queues) 3342 WARN_ON(netif_set_real_num_tx_queues(dev, txq)); 3343 3344 return 0; 3345 undo_rx: 3346 WARN_ON(netif_set_real_num_rx_queues(dev, old_rxq)); 3347 return err; 3348 } 3349 EXPORT_SYMBOL(netif_set_real_num_queues); 3350 3351 /** 3352 * netif_set_tso_max_size() - set the max size of TSO frames supported 3353 * @dev: netdev to update 3354 * @size: max skb->len of a TSO frame 3355 * 3356 * Set the limit on the size of TSO super-frames the device can handle. 3357 * Unless explicitly set the stack will assume the value of 3358 * %GSO_LEGACY_MAX_SIZE. 3359 */ 3360 void netif_set_tso_max_size(struct net_device *dev, unsigned int size) 3361 { 3362 dev->tso_max_size = min(GSO_MAX_SIZE, size); 3363 if (size < READ_ONCE(dev->gso_max_size)) 3364 netif_set_gso_max_size(dev, size); 3365 if (size < READ_ONCE(dev->gso_ipv4_max_size)) 3366 netif_set_gso_ipv4_max_size(dev, size); 3367 } 3368 EXPORT_SYMBOL(netif_set_tso_max_size); 3369 3370 /** 3371 * netif_set_tso_max_segs() - set the max number of segs supported for TSO 3372 * @dev: netdev to update 3373 * @segs: max number of TCP segments 3374 * 3375 * Set the limit on the number of TCP segments the device can generate from 3376 * a single TSO super-frame. 3377 * Unless explicitly set the stack will assume the value of %GSO_MAX_SEGS. 3378 */ 3379 void netif_set_tso_max_segs(struct net_device *dev, unsigned int segs) 3380 { 3381 dev->tso_max_segs = segs; 3382 if (segs < READ_ONCE(dev->gso_max_segs)) 3383 netif_set_gso_max_segs(dev, segs); 3384 } 3385 EXPORT_SYMBOL(netif_set_tso_max_segs); 3386 3387 /** 3388 * netif_inherit_tso_max() - copy all TSO limits from a lower device to an upper 3389 * @to: netdev to update 3390 * @from: netdev from which to copy the limits 3391 */ 3392 void netif_inherit_tso_max(struct net_device *to, const struct net_device *from) 3393 { 3394 netif_set_tso_max_size(to, from->tso_max_size); 3395 netif_set_tso_max_segs(to, from->tso_max_segs); 3396 } 3397 EXPORT_SYMBOL(netif_inherit_tso_max); 3398 3399 /** 3400 * netif_get_num_default_rss_queues - default number of RSS queues 3401 * 3402 * Default value is the number of physical cores if there are only 1 or 2, or 3403 * divided by 2 if there are more. 3404 */ 3405 int netif_get_num_default_rss_queues(void) 3406 { 3407 cpumask_var_t cpus; 3408 int cpu, count = 0; 3409 3410 if (unlikely(is_kdump_kernel() || !zalloc_cpumask_var(&cpus, GFP_KERNEL))) 3411 return 1; 3412 3413 cpumask_copy(cpus, cpu_online_mask); 3414 for_each_cpu(cpu, cpus) { 3415 ++count; 3416 cpumask_andnot(cpus, cpus, topology_sibling_cpumask(cpu)); 3417 } 3418 free_cpumask_var(cpus); 3419 3420 return count > 2 ? DIV_ROUND_UP(count, 2) : count; 3421 } 3422 EXPORT_SYMBOL(netif_get_num_default_rss_queues); 3423 3424 static void __netif_reschedule(struct Qdisc *q) 3425 { 3426 struct softnet_data *sd; 3427 unsigned long flags; 3428 3429 local_irq_save(flags); 3430 sd = this_cpu_ptr(&softnet_data); 3431 q->next_sched = NULL; 3432 *sd->output_queue_tailp = q; 3433 sd->output_queue_tailp = &q->next_sched; 3434 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3435 local_irq_restore(flags); 3436 } 3437 3438 void __netif_schedule(struct Qdisc *q) 3439 { 3440 /* If q->defer_list is not empty, at least one thread is 3441 * in __dev_xmit_skb() before llist_del_all(&q->defer_list). 3442 * This thread will attempt to run the queue. 3443 */ 3444 if (!llist_empty(&q->defer_list)) 3445 return; 3446 3447 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state)) 3448 __netif_reschedule(q); 3449 } 3450 EXPORT_SYMBOL(__netif_schedule); 3451 3452 struct dev_kfree_skb_cb { 3453 enum skb_drop_reason reason; 3454 }; 3455 3456 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb) 3457 { 3458 return (struct dev_kfree_skb_cb *)skb->cb; 3459 } 3460 3461 void netif_schedule_queue(struct netdev_queue *txq) 3462 { 3463 rcu_read_lock(); 3464 if (!netif_xmit_stopped(txq)) { 3465 struct Qdisc *q = rcu_dereference(txq->qdisc); 3466 3467 __netif_schedule(q); 3468 } 3469 rcu_read_unlock(); 3470 } 3471 EXPORT_SYMBOL(netif_schedule_queue); 3472 3473 void netif_tx_wake_queue(struct netdev_queue *dev_queue) 3474 { 3475 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) { 3476 struct Qdisc *q; 3477 3478 rcu_read_lock(); 3479 q = rcu_dereference(dev_queue->qdisc); 3480 __netif_schedule(q); 3481 rcu_read_unlock(); 3482 } 3483 } 3484 EXPORT_SYMBOL(netif_tx_wake_queue); 3485 3486 void dev_kfree_skb_irq_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3487 { 3488 unsigned long flags; 3489 3490 if (unlikely(!skb)) 3491 return; 3492 3493 if (likely(refcount_read(&skb->users) == 1)) { 3494 smp_rmb(); 3495 refcount_set(&skb->users, 0); 3496 } else if (likely(!refcount_dec_and_test(&skb->users))) { 3497 return; 3498 } 3499 get_kfree_skb_cb(skb)->reason = reason; 3500 local_irq_save(flags); 3501 skb->next = __this_cpu_read(softnet_data.completion_queue); 3502 __this_cpu_write(softnet_data.completion_queue, skb); 3503 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3504 local_irq_restore(flags); 3505 } 3506 EXPORT_SYMBOL(dev_kfree_skb_irq_reason); 3507 3508 void dev_kfree_skb_any_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3509 { 3510 if (in_hardirq() || irqs_disabled()) 3511 dev_kfree_skb_irq_reason(skb, reason); 3512 else 3513 kfree_skb_reason(skb, reason); 3514 } 3515 EXPORT_SYMBOL(dev_kfree_skb_any_reason); 3516 3517 3518 /** 3519 * netif_device_detach - mark device as removed 3520 * @dev: network device 3521 * 3522 * Mark device as removed from system and therefore no longer available. 3523 */ 3524 void netif_device_detach(struct net_device *dev) 3525 { 3526 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) && 3527 netif_running(dev)) { 3528 netif_tx_stop_all_queues(dev); 3529 } 3530 } 3531 EXPORT_SYMBOL(netif_device_detach); 3532 3533 /** 3534 * netif_device_attach - mark device as attached 3535 * @dev: network device 3536 * 3537 * Mark device as attached from system and restart if needed. 3538 */ 3539 void netif_device_attach(struct net_device *dev) 3540 { 3541 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) && 3542 netif_running(dev)) { 3543 netif_tx_wake_all_queues(dev); 3544 netdev_watchdog_up(dev); 3545 } 3546 } 3547 EXPORT_SYMBOL(netif_device_attach); 3548 3549 /* 3550 * Returns a Tx hash based on the given packet descriptor a Tx queues' number 3551 * to be used as a distribution range. 3552 */ 3553 static u16 skb_tx_hash(const struct net_device *dev, 3554 const struct net_device *sb_dev, 3555 struct sk_buff *skb) 3556 { 3557 u32 hash; 3558 u16 qoffset = 0; 3559 u16 qcount = dev->real_num_tx_queues; 3560 3561 if (dev->num_tc) { 3562 u8 tc = netdev_get_prio_tc_map(dev, skb->priority); 3563 struct netdev_tc_txq res; 3564 3565 res.combined = READ_ONCE(sb_dev->tc_to_txq[tc].combined); 3566 qoffset = res.offset; 3567 qcount = res.count; 3568 if (unlikely(!qcount)) { 3569 net_warn_ratelimited("%s: invalid qcount, qoffset %u for tc %u\n", 3570 sb_dev->name, qoffset, tc); 3571 qoffset = 0; 3572 qcount = dev->real_num_tx_queues; 3573 } 3574 } 3575 3576 if (skb_rx_queue_recorded(skb)) { 3577 DEBUG_NET_WARN_ON_ONCE(qcount == 0); 3578 hash = skb_get_rx_queue(skb); 3579 if (hash >= qoffset) 3580 hash -= qoffset; 3581 while (unlikely(hash >= qcount)) 3582 hash -= qcount; 3583 return hash + qoffset; 3584 } 3585 3586 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset; 3587 } 3588 3589 void skb_warn_bad_offload(const struct sk_buff *skb) 3590 { 3591 static const netdev_features_t null_features; 3592 struct net_device *dev = skb->dev; 3593 const char *name = ""; 3594 3595 if (!net_ratelimit()) 3596 return; 3597 3598 if (dev) { 3599 if (dev->dev.parent) 3600 name = dev_driver_string(dev->dev.parent); 3601 else 3602 name = netdev_name(dev); 3603 } 3604 skb_dump(KERN_WARNING, skb, false); 3605 WARN(1, "%s: caps=(%pNF, %pNF)\n", 3606 name, dev ? &dev->features : &null_features, 3607 skb->sk ? &skb->sk->sk_route_caps : &null_features); 3608 } 3609 3610 /* 3611 * Invalidate hardware checksum when packet is to be mangled, and 3612 * complete checksum manually on outgoing path. 3613 */ 3614 int skb_checksum_help(struct sk_buff *skb) 3615 { 3616 __wsum csum; 3617 int ret = 0, offset; 3618 3619 if (skb->ip_summed == CHECKSUM_COMPLETE) 3620 goto out_set_summed; 3621 3622 if (unlikely(skb_is_gso(skb))) { 3623 skb_warn_bad_offload(skb); 3624 return -EINVAL; 3625 } 3626 3627 if (!skb_frags_readable(skb)) { 3628 return -EFAULT; 3629 } 3630 3631 /* Before computing a checksum, we should make sure no frag could 3632 * be modified by an external entity : checksum could be wrong. 3633 */ 3634 if (skb_has_shared_frag(skb)) { 3635 ret = __skb_linearize(skb); 3636 if (ret) 3637 goto out; 3638 } 3639 3640 offset = skb_checksum_start_offset(skb); 3641 ret = -EINVAL; 3642 if (unlikely(offset >= skb_headlen(skb))) { 3643 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3644 WARN_ONCE(true, "offset (%d) >= skb_headlen() (%u)\n", 3645 offset, skb_headlen(skb)); 3646 goto out; 3647 } 3648 csum = skb_checksum(skb, offset, skb->len - offset, 0); 3649 3650 offset += skb->csum_offset; 3651 if (unlikely(offset + sizeof(__sum16) > skb_headlen(skb))) { 3652 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3653 WARN_ONCE(true, "offset+2 (%zu) > skb_headlen() (%u)\n", 3654 offset + sizeof(__sum16), skb_headlen(skb)); 3655 goto out; 3656 } 3657 ret = skb_ensure_writable(skb, offset + sizeof(__sum16)); 3658 if (ret) 3659 goto out; 3660 3661 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0; 3662 out_set_summed: 3663 skb->ip_summed = CHECKSUM_NONE; 3664 out: 3665 return ret; 3666 } 3667 EXPORT_SYMBOL(skb_checksum_help); 3668 3669 #ifdef CONFIG_NET_CRC32C 3670 int skb_crc32c_csum_help(struct sk_buff *skb) 3671 { 3672 u32 crc; 3673 int ret = 0, offset, start; 3674 3675 if (skb->ip_summed != CHECKSUM_PARTIAL) 3676 goto out; 3677 3678 if (unlikely(skb_is_gso(skb))) 3679 goto out; 3680 3681 /* Before computing a checksum, we should make sure no frag could 3682 * be modified by an external entity : checksum could be wrong. 3683 */ 3684 if (unlikely(skb_has_shared_frag(skb))) { 3685 ret = __skb_linearize(skb); 3686 if (ret) 3687 goto out; 3688 } 3689 start = skb_checksum_start_offset(skb); 3690 offset = start + offsetof(struct sctphdr, checksum); 3691 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) { 3692 ret = -EINVAL; 3693 goto out; 3694 } 3695 3696 ret = skb_ensure_writable(skb, offset + sizeof(__le32)); 3697 if (ret) 3698 goto out; 3699 3700 crc = ~skb_crc32c(skb, start, skb->len - start, ~0); 3701 *(__le32 *)(skb->data + offset) = cpu_to_le32(crc); 3702 skb_reset_csum_not_inet(skb); 3703 out: 3704 return ret; 3705 } 3706 EXPORT_SYMBOL(skb_crc32c_csum_help); 3707 #endif /* CONFIG_NET_CRC32C */ 3708 3709 __be16 skb_network_protocol(struct sk_buff *skb, int *depth) 3710 { 3711 __be16 type = skb->protocol; 3712 3713 /* Tunnel gso handlers can set protocol to ethernet. */ 3714 if (type == htons(ETH_P_TEB)) { 3715 struct ethhdr *eth; 3716 3717 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr)))) 3718 return 0; 3719 3720 eth = (struct ethhdr *)skb->data; 3721 type = eth->h_proto; 3722 } 3723 3724 return vlan_get_protocol_and_depth(skb, type, depth); 3725 } 3726 3727 3728 /* Take action when hardware reception checksum errors are detected. */ 3729 #ifdef CONFIG_BUG 3730 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3731 { 3732 netdev_err(dev, "hw csum failure\n"); 3733 skb_dump(KERN_ERR, skb, true); 3734 dump_stack(); 3735 } 3736 3737 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3738 { 3739 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb); 3740 } 3741 EXPORT_SYMBOL(netdev_rx_csum_fault); 3742 #endif 3743 3744 /* XXX: check that highmem exists at all on the given machine. */ 3745 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb) 3746 { 3747 #ifdef CONFIG_HIGHMEM 3748 int i; 3749 3750 if (!(dev->features & NETIF_F_HIGHDMA)) { 3751 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 3752 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 3753 struct page *page = skb_frag_page(frag); 3754 3755 if (page && PageHighMem(page)) 3756 return 1; 3757 } 3758 } 3759 #endif 3760 return 0; 3761 } 3762 3763 /* If MPLS offload request, verify we are testing hardware MPLS features 3764 * instead of standard features for the netdev. 3765 */ 3766 #if IS_ENABLED(CONFIG_NET_MPLS_GSO) 3767 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3768 netdev_features_t features, 3769 __be16 type) 3770 { 3771 if (eth_p_mpls(type)) 3772 features &= skb->dev->mpls_features; 3773 3774 return features; 3775 } 3776 #else 3777 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3778 netdev_features_t features, 3779 __be16 type) 3780 { 3781 return features; 3782 } 3783 #endif 3784 3785 static netdev_features_t harmonize_features(struct sk_buff *skb, 3786 netdev_features_t features) 3787 { 3788 __be16 type; 3789 3790 type = skb_network_protocol(skb, NULL); 3791 features = net_mpls_features(skb, features, type); 3792 3793 if (skb->ip_summed != CHECKSUM_NONE && 3794 !can_checksum_protocol(features, type)) { 3795 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3796 } 3797 if (illegal_highdma(skb->dev, skb)) 3798 features &= ~NETIF_F_SG; 3799 3800 return features; 3801 } 3802 3803 netdev_features_t passthru_features_check(struct sk_buff *skb, 3804 struct net_device *dev, 3805 netdev_features_t features) 3806 { 3807 return features; 3808 } 3809 EXPORT_SYMBOL(passthru_features_check); 3810 3811 static netdev_features_t dflt_features_check(struct sk_buff *skb, 3812 struct net_device *dev, 3813 netdev_features_t features) 3814 { 3815 return vlan_features_check(skb, features); 3816 } 3817 3818 static bool skb_gso_has_extension_hdr(const struct sk_buff *skb) 3819 { 3820 if (!skb->encapsulation) 3821 return ((skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6 || 3822 (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4 && 3823 vlan_get_protocol(skb) == htons(ETH_P_IPV6))) && 3824 skb_transport_header_was_set(skb) && 3825 skb_network_header_len(skb) != sizeof(struct ipv6hdr)); 3826 else 3827 return (!skb_inner_network_header_was_set(skb) || 3828 ((skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6 || 3829 (skb_shinfo(skb)->gso_type & SKB_GSO_UDP_L4 && 3830 inner_ip_hdr(skb)->version == 6)) && 3831 skb_inner_network_header_len(skb) != sizeof(struct ipv6hdr))); 3832 } 3833 3834 static netdev_features_t gso_features_check(const struct sk_buff *skb, 3835 struct net_device *dev, 3836 netdev_features_t features) 3837 { 3838 u16 gso_segs = skb_shinfo(skb)->gso_segs; 3839 3840 if (gso_segs > READ_ONCE(dev->gso_max_segs)) 3841 return features & ~NETIF_F_GSO_MASK; 3842 3843 if (unlikely(skb->len >= netif_get_gso_max_size(dev, skb))) 3844 return features & ~NETIF_F_GSO_MASK; 3845 3846 if (!skb_shinfo(skb)->gso_type) { 3847 skb_warn_bad_offload(skb); 3848 return features & ~NETIF_F_GSO_MASK; 3849 } 3850 3851 /* Support for GSO partial features requires software 3852 * intervention before we can actually process the packets 3853 * so we need to strip support for any partial features now 3854 * and we can pull them back in after we have partially 3855 * segmented the frame. 3856 */ 3857 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)) 3858 features &= ~dev->gso_partial_features; 3859 3860 /* Make sure to clear the IPv4 ID mangling feature if the IPv4 header 3861 * has the potential to be fragmented so that TSO does not generate 3862 * segments with the same ID. For encapsulated packets, the ID mangling 3863 * feature is guaranteed not to use the same ID for the outer IPv4 3864 * headers of the generated segments if the headers have the potential 3865 * to be fragmented, so there is no need to clear the IPv4 ID mangling 3866 * feature (see the section about NETIF_F_TSO_MANGLEID in 3867 * segmentation-offloads.rst). 3868 */ 3869 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) { 3870 const struct iphdr *iph; 3871 struct iphdr _iph; 3872 int nhoff = skb->encapsulation ? 3873 skb_inner_network_offset(skb) : 3874 skb_network_offset(skb); 3875 3876 iph = skb_header_pointer(skb, nhoff, sizeof(_iph), &_iph); 3877 3878 if (!iph || !(iph->frag_off & htons(IP_DF))) 3879 features &= ~dev->mangleid_features; 3880 } 3881 3882 /* NETIF_F_IPV6_CSUM does not support IPv6 extension headers, 3883 * so neither does TSO that depends on it. 3884 */ 3885 if (features & NETIF_F_IPV6_CSUM && 3886 skb_gso_has_extension_hdr(skb)) 3887 features &= ~(NETIF_F_IPV6_CSUM | NETIF_F_TSO6 | NETIF_F_GSO_UDP_L4); 3888 3889 return features; 3890 } 3891 3892 netdev_features_t netif_skb_features(struct sk_buff *skb) 3893 { 3894 struct net_device *dev = skb->dev; 3895 netdev_features_t features = dev->features; 3896 3897 if (skb_is_gso(skb)) 3898 features = gso_features_check(skb, dev, features); 3899 3900 /* If encapsulation offload request, verify we are testing 3901 * hardware encapsulation features instead of standard 3902 * features for the netdev 3903 */ 3904 if (skb->encapsulation) 3905 features &= dev->hw_enc_features; 3906 3907 if (skb_vlan_tagged(skb)) 3908 features = netdev_intersect_features(features, 3909 dev->vlan_features | 3910 NETIF_F_HW_VLAN_CTAG_TX | 3911 NETIF_F_HW_VLAN_STAG_TX); 3912 3913 if (dev->netdev_ops->ndo_features_check) 3914 features &= dev->netdev_ops->ndo_features_check(skb, dev, 3915 features); 3916 else 3917 features &= dflt_features_check(skb, dev, features); 3918 3919 return harmonize_features(skb, features); 3920 } 3921 EXPORT_SYMBOL(netif_skb_features); 3922 3923 static int xmit_one(struct sk_buff *skb, struct net_device *dev, 3924 struct netdev_queue *txq, bool more) 3925 { 3926 unsigned int len; 3927 int rc; 3928 3929 if (dev_nit_active_rcu(dev)) 3930 dev_queue_xmit_nit(skb, dev); 3931 3932 len = skb->len; 3933 trace_net_dev_start_xmit(skb, dev); 3934 rc = netdev_start_xmit(skb, dev, txq, more); 3935 trace_net_dev_xmit(skb, rc, dev, len); 3936 3937 return rc; 3938 } 3939 3940 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev, 3941 struct netdev_queue *txq, int *ret) 3942 { 3943 struct sk_buff *skb = first; 3944 int rc = NETDEV_TX_OK; 3945 3946 while (skb) { 3947 struct sk_buff *next = skb->next; 3948 3949 skb_mark_not_on_list(skb); 3950 rc = xmit_one(skb, dev, txq, next != NULL); 3951 if (unlikely(!dev_xmit_complete(rc))) { 3952 skb->next = next; 3953 goto out; 3954 } 3955 3956 skb = next; 3957 if (netif_tx_queue_stopped(txq) && skb) { 3958 rc = NETDEV_TX_BUSY; 3959 break; 3960 } 3961 } 3962 3963 out: 3964 *ret = rc; 3965 return skb; 3966 } 3967 3968 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb, 3969 netdev_features_t features) 3970 { 3971 if (skb_vlan_tag_present(skb) && 3972 !vlan_hw_offload_capable(features, skb->vlan_proto)) 3973 skb = __vlan_hwaccel_push_inside(skb); 3974 return skb; 3975 } 3976 3977 int skb_csum_hwoffload_help(struct sk_buff *skb, 3978 const netdev_features_t features) 3979 { 3980 if (unlikely(skb_csum_is_sctp(skb))) 3981 return !!(features & NETIF_F_SCTP_CRC) ? 0 : 3982 skb_crc32c_csum_help(skb); 3983 3984 if (features & NETIF_F_HW_CSUM) 3985 return 0; 3986 3987 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) { 3988 if (vlan_get_protocol(skb) == htons(ETH_P_IPV6) && 3989 skb_network_header_len(skb) != sizeof(struct ipv6hdr)) 3990 goto sw_checksum; 3991 3992 switch (skb->csum_offset) { 3993 case offsetof(struct tcphdr, check): 3994 case offsetof(struct udphdr, check): 3995 return 0; 3996 } 3997 } 3998 3999 sw_checksum: 4000 return skb_checksum_help(skb); 4001 } 4002 EXPORT_SYMBOL(skb_csum_hwoffload_help); 4003 4004 /* Checks if this SKB belongs to an HW offloaded socket 4005 * and whether any SW fallbacks are required based on dev. 4006 * Check decrypted mark in case skb_orphan() cleared socket. 4007 */ 4008 static struct sk_buff *sk_validate_xmit_skb(struct sk_buff *skb, 4009 struct net_device *dev) 4010 { 4011 #ifdef CONFIG_SOCK_VALIDATE_XMIT 4012 struct sk_buff *(*sk_validate)(struct sock *sk, struct net_device *dev, 4013 struct sk_buff *skb); 4014 struct sock *sk = skb->sk; 4015 4016 sk_validate = NULL; 4017 if (sk) { 4018 if (sk_fullsock(sk)) 4019 sk_validate = sk->sk_validate_xmit_skb; 4020 else if (sk_is_inet(sk) && sk->sk_state == TCP_TIME_WAIT) 4021 sk_validate = inet_twsk(sk)->tw_validate_xmit_skb; 4022 } 4023 4024 if (sk_validate) { 4025 skb = sk_validate(sk, dev, skb); 4026 } else if (unlikely(skb_is_decrypted(skb))) { 4027 pr_warn_ratelimited("unencrypted skb with no associated socket - dropping\n"); 4028 kfree_skb(skb); 4029 skb = NULL; 4030 } 4031 #endif 4032 4033 return skb; 4034 } 4035 4036 static struct sk_buff *validate_xmit_unreadable_skb(struct sk_buff *skb, 4037 struct net_device *dev) 4038 { 4039 struct skb_shared_info *shinfo; 4040 struct net_iov *niov; 4041 4042 if (likely(skb_frags_readable(skb) || 4043 dev->netmem_tx == NETMEM_TX_NO_DMA)) 4044 goto out; 4045 4046 if (dev->netmem_tx == NETMEM_TX_NONE) 4047 goto out_free; 4048 4049 shinfo = skb_shinfo(skb); 4050 4051 if (shinfo->nr_frags > 0) { 4052 niov = netmem_to_net_iov(skb_frag_netmem(&shinfo->frags[0])); 4053 if (net_is_devmem_iov(niov) && 4054 READ_ONCE(net_devmem_iov_binding(niov)->dev) != dev) 4055 goto out_free; 4056 } 4057 4058 out: 4059 return skb; 4060 4061 out_free: 4062 kfree_skb(skb); 4063 return NULL; 4064 } 4065 4066 /* Returns the skb on success, NULL if dropped, or ERR_PTR(-EINPROGRESS) 4067 * if stolen by async xfrm crypto (delivered via xfrm_dev_resume()). 4068 */ 4069 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again) 4070 { 4071 netdev_features_t features; 4072 4073 skb = validate_xmit_unreadable_skb(skb, dev); 4074 if (unlikely(!skb)) 4075 goto out_null; 4076 4077 features = netif_skb_features(skb); 4078 skb = validate_xmit_vlan(skb, features); 4079 if (unlikely(!skb)) 4080 goto out_null; 4081 4082 skb = sk_validate_xmit_skb(skb, dev); 4083 if (unlikely(!skb)) 4084 goto out_null; 4085 4086 if (netif_needs_gso(skb, features)) { 4087 struct sk_buff *segs; 4088 4089 segs = skb_gso_segment(skb, features); 4090 if (IS_ERR(segs)) { 4091 goto out_kfree_skb; 4092 } else if (segs) { 4093 consume_skb(skb); 4094 skb = segs; 4095 } 4096 } else { 4097 if (skb_needs_linearize(skb, features) && 4098 __skb_linearize(skb)) 4099 goto out_kfree_skb; 4100 4101 /* If packet is not checksummed and device does not 4102 * support checksumming for this protocol, complete 4103 * checksumming here. 4104 */ 4105 if (skb->ip_summed == CHECKSUM_PARTIAL) { 4106 if (skb->encapsulation) 4107 skb_set_inner_transport_header(skb, 4108 skb_checksum_start_offset(skb)); 4109 else 4110 skb_set_transport_header(skb, 4111 skb_checksum_start_offset(skb)); 4112 if (skb_csum_hwoffload_help(skb, features)) 4113 goto out_kfree_skb; 4114 } 4115 } 4116 4117 skb = validate_xmit_xfrm(skb, features, again); 4118 4119 return skb; 4120 4121 out_kfree_skb: 4122 kfree_skb(skb); 4123 out_null: 4124 dev_core_stats_tx_dropped_inc(dev); 4125 return NULL; 4126 } 4127 4128 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again) 4129 { 4130 struct sk_buff *next, *head = NULL, *tail; 4131 4132 for (; skb != NULL; skb = next) { 4133 next = skb->next; 4134 skb_mark_not_on_list(skb); 4135 4136 /* in case skb won't be segmented, point to itself */ 4137 skb->prev = skb; 4138 4139 skb = validate_xmit_skb(skb, dev, again); 4140 if (IS_ERR_OR_NULL(skb)) 4141 continue; 4142 4143 if (!head) 4144 head = skb; 4145 else 4146 tail->next = skb; 4147 /* If skb was segmented, skb->prev points to 4148 * the last segment. If not, it still contains skb. 4149 */ 4150 tail = skb->prev; 4151 } 4152 return head; 4153 } 4154 EXPORT_SYMBOL_GPL(validate_xmit_skb_list); 4155 4156 static enum skb_drop_reason qdisc_pkt_len_segs_init(struct sk_buff *skb) 4157 { 4158 struct skb_shared_info *shinfo = skb_shinfo(skb); 4159 unsigned int hdr_len, tlen; 4160 u16 gso_segs; 4161 4162 qdisc_skb_cb(skb)->pkt_len = skb->len; 4163 if (!shinfo->gso_size) { 4164 qdisc_skb_cb(skb)->pkt_segs = 1; 4165 return SKB_NOT_DROPPED_YET; 4166 } 4167 4168 qdisc_skb_cb(skb)->pkt_segs = gso_segs = shinfo->gso_segs; 4169 4170 /* To get more precise estimation of bytes sent on wire, 4171 * we add to pkt_len the headers size of all segments 4172 */ 4173 4174 /* mac layer + network layer */ 4175 if (!skb->encapsulation) { 4176 if (unlikely(!skb_transport_header_was_set(skb))) 4177 return SKB_NOT_DROPPED_YET; 4178 hdr_len = skb_transport_offset(skb); 4179 } else { 4180 hdr_len = skb_inner_transport_offset(skb); 4181 } 4182 /* + transport layer */ 4183 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) { 4184 const struct tcphdr *th; 4185 4186 if (!pskb_may_pull(skb, hdr_len + sizeof(struct tcphdr))) 4187 return SKB_DROP_REASON_SKB_BAD_GSO; 4188 4189 th = (const struct tcphdr *)(skb->data + hdr_len); 4190 tlen = __tcp_hdrlen(th); 4191 if (tlen < sizeof(*th)) 4192 return SKB_DROP_REASON_SKB_BAD_GSO; 4193 hdr_len += tlen; 4194 if (!pskb_may_pull(skb, hdr_len)) 4195 return SKB_DROP_REASON_SKB_BAD_GSO; 4196 } else if (shinfo->gso_type & SKB_GSO_UDP_L4) { 4197 if (!pskb_may_pull(skb, hdr_len + sizeof(struct udphdr))) 4198 return SKB_DROP_REASON_SKB_BAD_GSO; 4199 hdr_len += sizeof(struct udphdr); 4200 } 4201 4202 /* prior pskb_may_pull() might have changed skb->head. */ 4203 shinfo = skb_shinfo(skb); 4204 if (unlikely(shinfo->gso_type & SKB_GSO_DODGY)) { 4205 int payload = skb->len - hdr_len; 4206 4207 /* Malicious packet. */ 4208 if (payload <= 0) 4209 return SKB_DROP_REASON_SKB_BAD_GSO; 4210 gso_segs = DIV_ROUND_UP(payload, shinfo->gso_size); 4211 shinfo->gso_segs = gso_segs; 4212 qdisc_skb_cb(skb)->pkt_segs = gso_segs; 4213 } 4214 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len; 4215 return SKB_NOT_DROPPED_YET; 4216 } 4217 4218 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q, 4219 struct sk_buff **to_free, 4220 struct netdev_queue *txq) 4221 { 4222 int rc; 4223 4224 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK; 4225 if (rc == NET_XMIT_SUCCESS) 4226 trace_qdisc_enqueue(q, txq, skb); 4227 return rc; 4228 } 4229 4230 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q, 4231 struct net_device *dev, 4232 struct netdev_queue *txq) 4233 { 4234 struct sk_buff *next, *to_free = NULL, *to_free2 = NULL; 4235 spinlock_t *root_lock = qdisc_lock(q); 4236 struct llist_node *ll_list, *first_n; 4237 unsigned long defer_count = 0; 4238 int rc; 4239 4240 qdisc_calculate_pkt_len(skb, q); 4241 4242 tcf_set_qdisc_drop_reason(skb, QDISC_DROP_GENERIC); 4243 4244 if (q->flags & TCQ_F_NOLOCK) { 4245 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) && 4246 qdisc_run_begin(q)) { 4247 /* Retest nolock_qdisc_is_empty() within the protection 4248 * of q->seqlock to protect from racing with requeuing. 4249 */ 4250 if (unlikely(!nolock_qdisc_is_empty(q))) { 4251 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4252 __qdisc_run(q); 4253 to_free2 = qdisc_run_end(q); 4254 4255 goto free_skbs; 4256 } 4257 4258 qdisc_bstats_cpu_update(q, skb); 4259 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) && 4260 !nolock_qdisc_is_empty(q)) 4261 __qdisc_run(q); 4262 4263 to_free2 = qdisc_run_end(q); 4264 rc = NET_XMIT_SUCCESS; 4265 goto free_skbs; 4266 } 4267 4268 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4269 to_free2 = qdisc_run(q); 4270 goto free_skbs; 4271 } 4272 4273 /* Open code llist_add(&skb->ll_node, &q->defer_list) + queue limit. 4274 * In the try_cmpxchg() loop, we want to increment q->defer_count 4275 * at most once to limit the number of skbs in defer_list. 4276 * We perform the defer_count increment only if the list is not empty, 4277 * because some arches have slow atomic_long_inc_return(). 4278 */ 4279 first_n = READ_ONCE(q->defer_list.first); 4280 do { 4281 if (first_n && !defer_count) { 4282 defer_count = atomic_long_inc_return(&q->defer_count); 4283 if (unlikely(defer_count > READ_ONCE(net_hotdata.qdisc_max_burst))) { 4284 kfree_skb_reason(skb, SKB_DROP_REASON_QDISC_BURST_DROP); 4285 return NET_XMIT_DROP; 4286 } 4287 } 4288 skb->ll_node.next = first_n; 4289 } while (!try_cmpxchg(&q->defer_list.first, &first_n, &skb->ll_node)); 4290 4291 /* If defer_list was not empty, we know the cpu which queued 4292 * the first skb will process the whole list for us. 4293 */ 4294 if (first_n) 4295 return NET_XMIT_SUCCESS; 4296 4297 spin_lock(root_lock); 4298 4299 ll_list = llist_del_all(&q->defer_list); 4300 /* There is a small race because we clear defer_count not atomically 4301 * with the prior llist_del_all(). This means defer_list could grow 4302 * over qdisc_max_burst. 4303 */ 4304 atomic_long_set(&q->defer_count, 0); 4305 4306 ll_list = llist_reverse_order(ll_list); 4307 4308 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) { 4309 llist_for_each_entry_safe(skb, next, ll_list, ll_node) 4310 __qdisc_drop(skb, &to_free); 4311 rc = NET_XMIT_DROP; 4312 goto unlock; 4313 } 4314 if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && 4315 !llist_next(ll_list) && qdisc_run_begin(q)) { 4316 /* 4317 * This is a work-conserving queue; there are no old skbs 4318 * waiting to be sent out; and the qdisc is not running - 4319 * xmit the skb directly. 4320 */ 4321 4322 DEBUG_NET_WARN_ON_ONCE(skb != llist_entry(ll_list, 4323 struct sk_buff, 4324 ll_node)); 4325 qdisc_bstats_update(q, skb); 4326 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) 4327 __qdisc_run(q); 4328 to_free2 = qdisc_run_end(q); 4329 rc = NET_XMIT_SUCCESS; 4330 } else { 4331 int count = 0; 4332 4333 llist_for_each_entry_safe(skb, next, ll_list, ll_node) { 4334 if (next) { 4335 prefetch(next); 4336 prefetch(&next->priority); 4337 skb_mark_not_on_list(skb); 4338 } 4339 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4340 count++; 4341 } 4342 to_free2 = qdisc_run(q); 4343 if (count != 1) 4344 rc = NET_XMIT_SUCCESS; 4345 } 4346 unlock: 4347 spin_unlock(root_lock); 4348 4349 free_skbs: 4350 tcf_kfree_skb_list(to_free, q, txq, dev); 4351 tcf_kfree_skb_list(to_free2, q, txq, dev); 4352 return rc; 4353 } 4354 4355 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO) 4356 static void skb_update_prio(struct sk_buff *skb) 4357 { 4358 const struct netprio_map *map; 4359 const struct sock *sk; 4360 unsigned int prioidx; 4361 4362 if (skb->priority) 4363 return; 4364 map = rcu_dereference_bh(skb->dev->priomap); 4365 if (!map) 4366 return; 4367 sk = skb_to_full_sk(skb); 4368 if (!sk) 4369 return; 4370 4371 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data); 4372 4373 if (prioidx < map->priomap_len) 4374 skb->priority = map->priomap[prioidx]; 4375 } 4376 #else 4377 #define skb_update_prio(skb) 4378 #endif 4379 4380 /** 4381 * dev_loopback_xmit - loop back @skb 4382 * @net: network namespace this loopback is happening in 4383 * @sk: sk needed to be a netfilter okfn 4384 * @skb: buffer to transmit 4385 */ 4386 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb) 4387 { 4388 skb_reset_mac_header(skb); 4389 __skb_pull(skb, skb_network_offset(skb)); 4390 skb->pkt_type = PACKET_LOOPBACK; 4391 if (skb->ip_summed == CHECKSUM_NONE) 4392 skb->ip_summed = CHECKSUM_UNNECESSARY; 4393 DEBUG_NET_WARN_ON_ONCE(!skb_dst(skb)); 4394 skb_dst_force(skb); 4395 netif_rx(skb); 4396 return 0; 4397 } 4398 EXPORT_SYMBOL(dev_loopback_xmit); 4399 4400 #ifdef CONFIG_NET_EGRESS 4401 static struct netdev_queue * 4402 netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb) 4403 { 4404 int qm = skb_get_queue_mapping(skb); 4405 4406 return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm)); 4407 } 4408 4409 #ifndef CONFIG_PREEMPT_RT 4410 static bool netdev_xmit_txqueue_skipped(void) 4411 { 4412 return __this_cpu_read(softnet_data.xmit.skip_txqueue); 4413 } 4414 4415 void netdev_xmit_skip_txqueue(bool skip) 4416 { 4417 __this_cpu_write(softnet_data.xmit.skip_txqueue, skip); 4418 } 4419 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4420 4421 #else 4422 static bool netdev_xmit_txqueue_skipped(void) 4423 { 4424 return current->net_xmit.skip_txqueue; 4425 } 4426 4427 void netdev_xmit_skip_txqueue(bool skip) 4428 { 4429 current->net_xmit.skip_txqueue = skip; 4430 } 4431 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4432 #endif 4433 #endif /* CONFIG_NET_EGRESS */ 4434 4435 #ifdef CONFIG_NET_XGRESS 4436 static int tc_run(struct tcx_entry *entry, struct sk_buff *skb, 4437 enum skb_drop_reason *drop_reason) 4438 { 4439 int ret = TC_ACT_UNSPEC; 4440 #ifdef CONFIG_NET_CLS_ACT 4441 struct mini_Qdisc *miniq = rcu_dereference_bh(entry->miniq); 4442 struct tcf_result res; 4443 4444 if (!miniq) 4445 return ret; 4446 4447 /* Global bypass */ 4448 if (!static_branch_likely(&tcf_sw_enabled_key)) 4449 return ret; 4450 4451 /* Block-wise bypass */ 4452 if (tcf_block_bypass_sw(miniq->block)) 4453 return ret; 4454 4455 tc_skb_cb(skb)->mru = 0; 4456 qdisc_skb_cb(skb)->post_ct = false; 4457 tcf_set_drop_reason(skb, *drop_reason); 4458 4459 mini_qdisc_bstats_cpu_update(miniq, skb); 4460 ret = tcf_classify(skb, miniq->block, miniq->filter_list, &res, false); 4461 /* Only tcf related quirks below. */ 4462 switch (ret) { 4463 case TC_ACT_SHOT: 4464 *drop_reason = tcf_get_drop_reason(skb); 4465 mini_qdisc_qstats_cpu_drop(miniq); 4466 break; 4467 case TC_ACT_OK: 4468 case TC_ACT_RECLASSIFY: 4469 skb->tc_index = TC_H_MIN(res.classid); 4470 break; 4471 } 4472 #endif /* CONFIG_NET_CLS_ACT */ 4473 return ret; 4474 } 4475 4476 static DEFINE_STATIC_KEY_FALSE(tcx_needed_key); 4477 4478 void tcx_inc(void) 4479 { 4480 static_branch_inc(&tcx_needed_key); 4481 } 4482 4483 void tcx_dec(void) 4484 { 4485 static_branch_dec(&tcx_needed_key); 4486 } 4487 4488 static __always_inline enum tcx_action_base 4489 tcx_run(const struct bpf_mprog_entry *entry, struct sk_buff *skb, 4490 const bool needs_mac) 4491 { 4492 const struct bpf_mprog_fp *fp; 4493 const struct bpf_prog *prog; 4494 int ret = TCX_NEXT; 4495 4496 if (needs_mac) 4497 __skb_push(skb, skb->mac_len); 4498 bpf_mprog_foreach_prog(entry, fp, prog) { 4499 bpf_compute_data_pointers(skb); 4500 ret = bpf_prog_run(prog, skb); 4501 if (ret != TCX_NEXT) 4502 break; 4503 } 4504 if (needs_mac) 4505 __skb_pull(skb, skb->mac_len); 4506 return tcx_action_code(skb, ret); 4507 } 4508 4509 static __always_inline struct sk_buff * 4510 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4511 struct net_device *orig_dev, bool *another) 4512 { 4513 struct bpf_mprog_entry *entry = rcu_dereference_bh(skb->dev->tcx_ingress); 4514 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_INGRESS; 4515 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4516 int sch_ret; 4517 4518 if (!entry) 4519 return skb; 4520 4521 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4522 if (unlikely(*pt_prev)) { 4523 *ret = deliver_skb(skb, *pt_prev, orig_dev); 4524 *pt_prev = NULL; 4525 } 4526 4527 qdisc_pkt_len_segs_init(skb); 4528 tcx_set_ingress(skb, true); 4529 4530 if (static_branch_unlikely(&tcx_needed_key)) { 4531 sch_ret = tcx_run(entry, skb, true); 4532 if (sch_ret != TC_ACT_UNSPEC) 4533 goto ingress_verdict; 4534 } 4535 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4536 ingress_verdict: 4537 switch (sch_ret) { 4538 case TC_ACT_REDIRECT: 4539 /* skb_mac_header check was done by BPF, so we can safely 4540 * push the L2 header back before redirecting to another 4541 * netdev. 4542 */ 4543 __skb_push(skb, skb->mac_len); 4544 if (skb_do_redirect(skb) == -EAGAIN) { 4545 __skb_pull(skb, skb->mac_len); 4546 *another = true; 4547 break; 4548 } 4549 *ret = NET_RX_SUCCESS; 4550 bpf_net_ctx_clear(bpf_net_ctx); 4551 return NULL; 4552 case TC_ACT_SHOT: 4553 kfree_skb_reason(skb, drop_reason); 4554 *ret = NET_RX_DROP; 4555 bpf_net_ctx_clear(bpf_net_ctx); 4556 return NULL; 4557 /* used by tc_run */ 4558 case TC_ACT_STOLEN: 4559 case TC_ACT_QUEUED: 4560 case TC_ACT_TRAP: 4561 consume_skb(skb); 4562 fallthrough; 4563 case TC_ACT_CONSUMED: 4564 *ret = NET_RX_SUCCESS; 4565 bpf_net_ctx_clear(bpf_net_ctx); 4566 return NULL; 4567 } 4568 bpf_net_ctx_clear(bpf_net_ctx); 4569 4570 return skb; 4571 } 4572 4573 static __always_inline struct sk_buff * 4574 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4575 { 4576 struct bpf_mprog_entry *entry = rcu_dereference_bh(dev->tcx_egress); 4577 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_EGRESS; 4578 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4579 int sch_ret; 4580 4581 if (!entry) 4582 return skb; 4583 4584 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4585 4586 /* qdisc_skb_cb(skb)->pkt_len & tcx_set_ingress() was 4587 * already set by the caller. 4588 */ 4589 if (static_branch_unlikely(&tcx_needed_key)) { 4590 sch_ret = tcx_run(entry, skb, false); 4591 if (sch_ret != TC_ACT_UNSPEC) 4592 goto egress_verdict; 4593 } 4594 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4595 egress_verdict: 4596 switch (sch_ret) { 4597 case TC_ACT_REDIRECT: 4598 /* No need to push/pop skb's mac_header here on egress! */ 4599 skb_do_redirect(skb); 4600 *ret = NET_XMIT_SUCCESS; 4601 bpf_net_ctx_clear(bpf_net_ctx); 4602 return NULL; 4603 case TC_ACT_SHOT: 4604 kfree_skb_reason(skb, drop_reason); 4605 *ret = NET_XMIT_DROP; 4606 bpf_net_ctx_clear(bpf_net_ctx); 4607 return NULL; 4608 /* used by tc_run */ 4609 case TC_ACT_STOLEN: 4610 case TC_ACT_QUEUED: 4611 case TC_ACT_TRAP: 4612 consume_skb(skb); 4613 fallthrough; 4614 case TC_ACT_CONSUMED: 4615 *ret = NET_XMIT_SUCCESS; 4616 bpf_net_ctx_clear(bpf_net_ctx); 4617 return NULL; 4618 } 4619 bpf_net_ctx_clear(bpf_net_ctx); 4620 4621 return skb; 4622 } 4623 #else 4624 static __always_inline struct sk_buff * 4625 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4626 struct net_device *orig_dev, bool *another) 4627 { 4628 return skb; 4629 } 4630 4631 static __always_inline struct sk_buff * 4632 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4633 { 4634 return skb; 4635 } 4636 #endif /* CONFIG_NET_XGRESS */ 4637 4638 #ifdef CONFIG_XPS 4639 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb, 4640 struct xps_dev_maps *dev_maps, unsigned int tci) 4641 { 4642 int tc = netdev_get_prio_tc_map(dev, skb->priority); 4643 struct xps_map *map; 4644 int queue_index = -1; 4645 4646 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids) 4647 return queue_index; 4648 4649 tci *= dev_maps->num_tc; 4650 tci += tc; 4651 4652 map = rcu_dereference(dev_maps->attr_map[tci]); 4653 if (map) { 4654 if (map->len == 1) 4655 queue_index = map->queues[0]; 4656 else 4657 queue_index = map->queues[reciprocal_scale( 4658 skb_get_hash(skb), map->len)]; 4659 if (unlikely(queue_index >= dev->real_num_tx_queues)) 4660 queue_index = -1; 4661 } 4662 return queue_index; 4663 } 4664 #endif 4665 4666 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev, 4667 struct sk_buff *skb) 4668 { 4669 #ifdef CONFIG_XPS 4670 struct xps_dev_maps *dev_maps; 4671 struct sock *sk = skb->sk; 4672 int queue_index = -1; 4673 4674 if (!static_key_false(&xps_needed)) 4675 return -1; 4676 4677 rcu_read_lock(); 4678 if (!static_key_false(&xps_rxqs_needed)) 4679 goto get_cpus_map; 4680 4681 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]); 4682 if (dev_maps) { 4683 int tci = sk_rx_queue_get(sk); 4684 4685 if (tci >= 0) 4686 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4687 tci); 4688 } 4689 4690 get_cpus_map: 4691 if (queue_index < 0) { 4692 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]); 4693 if (dev_maps) { 4694 unsigned int tci = skb->sender_cpu - 1; 4695 4696 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4697 tci); 4698 } 4699 } 4700 rcu_read_unlock(); 4701 4702 return queue_index; 4703 #else 4704 return -1; 4705 #endif 4706 } 4707 4708 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb, 4709 struct net_device *sb_dev) 4710 { 4711 return 0; 4712 } 4713 EXPORT_SYMBOL(dev_pick_tx_zero); 4714 4715 int sk_tx_queue_get(const struct sock *sk) 4716 { 4717 int resel, val; 4718 4719 if (!sk) 4720 return -1; 4721 /* Paired with WRITE_ONCE() in sk_tx_queue_clear() 4722 * and sk_tx_queue_set(). 4723 */ 4724 val = READ_ONCE(sk->sk_tx_queue_mapping); 4725 4726 if (val == NO_QUEUE_MAPPING) 4727 return -1; 4728 4729 if (!sk_fullsock(sk)) 4730 return val; 4731 4732 resel = READ_ONCE(sock_net(sk)->core.sysctl_txq_reselection); 4733 if (resel && time_is_before_jiffies( 4734 READ_ONCE(sk->sk_tx_queue_mapping_jiffies) + resel)) 4735 return -1; 4736 4737 return val; 4738 } 4739 EXPORT_SYMBOL(sk_tx_queue_get); 4740 4741 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb, 4742 struct net_device *sb_dev) 4743 { 4744 struct sock *sk = skb->sk; 4745 int queue_index = sk_tx_queue_get(sk); 4746 4747 sb_dev = sb_dev ? : dev; 4748 4749 if (queue_index < 0 || skb->ooo_okay || 4750 queue_index >= dev->real_num_tx_queues) { 4751 int new_index = get_xps_queue(dev, sb_dev, skb); 4752 4753 if (new_index < 0) 4754 new_index = skb_tx_hash(dev, sb_dev, skb); 4755 4756 if (sk && sk_fullsock(sk) && 4757 rcu_access_pointer(sk->sk_dst_cache)) 4758 sk_tx_queue_set(sk, new_index); 4759 4760 queue_index = new_index; 4761 } 4762 4763 return queue_index; 4764 } 4765 EXPORT_SYMBOL(netdev_pick_tx); 4766 4767 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev, 4768 struct sk_buff *skb, 4769 struct net_device *sb_dev) 4770 { 4771 int queue_index = 0; 4772 4773 #ifdef CONFIG_XPS 4774 u32 sender_cpu = skb->sender_cpu - 1; 4775 4776 if (sender_cpu >= (u32)NR_CPUS) 4777 skb->sender_cpu = raw_smp_processor_id() + 1; 4778 #endif 4779 4780 if (dev->real_num_tx_queues != 1) { 4781 const struct net_device_ops *ops = dev->netdev_ops; 4782 4783 if (ops->ndo_select_queue) 4784 queue_index = ops->ndo_select_queue(dev, skb, sb_dev); 4785 else 4786 queue_index = netdev_pick_tx(dev, skb, sb_dev); 4787 4788 queue_index = netdev_cap_txqueue(dev, queue_index); 4789 } 4790 4791 skb_set_queue_mapping(skb, queue_index); 4792 return netdev_get_tx_queue(dev, queue_index); 4793 } 4794 4795 /** 4796 * __dev_queue_xmit() - transmit a buffer 4797 * @skb: buffer to transmit 4798 * @sb_dev: suboordinate device used for L2 forwarding offload 4799 * 4800 * Queue a buffer for transmission to a network device. The caller must 4801 * have set the device and priority and built the buffer before calling 4802 * this function. The function can be called from an interrupt. 4803 * 4804 * When calling this method, interrupts MUST be enabled. This is because 4805 * the BH enable code must have IRQs enabled so that it will not deadlock. 4806 * 4807 * Regardless of the return value, the skb is consumed, so it is currently 4808 * difficult to retry a send to this method. (You can bump the ref count 4809 * before sending to hold a reference for retry if you are careful.) 4810 * 4811 * Return: 4812 * * 0 - buffer successfully transmitted 4813 * * positive qdisc return code - NET_XMIT_DROP etc. 4814 * * negative errno - other errors 4815 */ 4816 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev) 4817 { 4818 struct net_device *dev = skb->dev; 4819 struct netdev_queue *txq = NULL; 4820 enum skb_drop_reason reason; 4821 int cpu, rc = -ENOMEM; 4822 bool again = false; 4823 struct Qdisc *q; 4824 4825 skb_reset_mac_header(skb); 4826 skb_assert_len(skb); 4827 4828 if (unlikely(skb_shinfo(skb)->tx_flags & 4829 (SKBTX_SCHED_TSTAMP | SKBTX_BPF))) 4830 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED); 4831 4832 reason = qdisc_pkt_len_segs_init(skb); 4833 if (unlikely(reason)) { 4834 dev_core_stats_tx_dropped_inc(dev); 4835 kfree_skb_reason(skb, reason); 4836 return -EINVAL; 4837 } 4838 /* Disable soft irqs for various locks below. Also 4839 * stops preemption for RCU. 4840 */ 4841 rcu_read_lock_bh(); 4842 4843 skb_update_prio(skb); 4844 4845 tcx_set_ingress(skb, false); 4846 #ifdef CONFIG_NET_EGRESS 4847 if (static_branch_unlikely(&egress_needed_key)) { 4848 if (nf_hook_egress_active()) { 4849 skb = nf_hook_egress(skb, &rc, dev); 4850 if (!skb) 4851 goto out; 4852 } 4853 4854 netdev_xmit_skip_txqueue(false); 4855 4856 nf_skip_egress(skb, true); 4857 skb = sch_handle_egress(skb, &rc, dev); 4858 if (!skb) 4859 goto out; 4860 nf_skip_egress(skb, false); 4861 4862 if (netdev_xmit_txqueue_skipped()) 4863 txq = netdev_tx_queue_mapping(dev, skb); 4864 } 4865 #endif 4866 /* If device/qdisc don't need skb->dst, release it right now while 4867 * its hot in this cpu cache. 4868 */ 4869 if (dev->priv_flags & IFF_XMIT_DST_RELEASE) 4870 skb_dst_drop(skb); 4871 else 4872 skb_dst_force(skb); 4873 4874 if (!txq) 4875 txq = netdev_core_pick_tx(dev, skb, sb_dev); 4876 4877 q = rcu_dereference_bh(txq->qdisc); 4878 4879 trace_net_dev_queue(skb); 4880 if (q->enqueue) { 4881 rc = __dev_xmit_skb(skb, q, dev, txq); 4882 goto out; 4883 } 4884 4885 /* The device has no queue. Common case for software devices: 4886 * loopback, all the sorts of tunnels... 4887 4888 * Really, it is unlikely that netif_tx_lock protection is necessary 4889 * here. (f.e. loopback and IP tunnels are clean ignoring statistics 4890 * counters.) 4891 * However, it is possible, that they rely on protection 4892 * made by us here. 4893 4894 * Check this and shot the lock. It is not prone from deadlocks. 4895 *Either shot noqueue qdisc, it is even simpler 8) 4896 */ 4897 if (unlikely(!(dev->flags & IFF_UP))) { 4898 reason = SKB_DROP_REASON_DEV_READY; 4899 goto drop; 4900 } 4901 4902 cpu = smp_processor_id(); /* ok because BHs are off */ 4903 4904 if (likely(!netif_tx_owned(txq, cpu))) { 4905 bool is_list = false; 4906 4907 if (dev_xmit_recursion()) 4908 goto recursion_alert; 4909 4910 skb = validate_xmit_skb(skb, dev, &again); 4911 if (IS_ERR_OR_NULL(skb)) { 4912 if (PTR_ERR(skb) == -EINPROGRESS) 4913 rc = NET_XMIT_SUCCESS; 4914 goto out; 4915 } 4916 4917 HARD_TX_LOCK(dev, txq, cpu); 4918 4919 if (!netif_xmit_stopped(txq)) { 4920 is_list = !!skb->next; 4921 4922 dev_xmit_recursion_inc(); 4923 skb = dev_hard_start_xmit(skb, dev, txq, &rc); 4924 dev_xmit_recursion_dec(); 4925 4926 /* GSO segments a single SKB into a list of frames. 4927 * TCP expects error to mean none of the data was sent. 4928 */ 4929 if (is_list) 4930 rc = NETDEV_TX_OK; 4931 } 4932 HARD_TX_UNLOCK(dev, txq); 4933 if (!skb) /* xmit completed */ 4934 goto out; 4935 4936 net_crit_ratelimited("Virtual device %s asks to queue packet!\n", 4937 dev->name); 4938 /* NETDEV_TX_BUSY or queue was stopped */ 4939 if (!is_list) 4940 rc = -ENETDOWN; 4941 } else { 4942 /* Recursion is detected! It is possible unfortunately. */ 4943 recursion_alert: 4944 net_crit_ratelimited("Dead loop on virtual device %s (net %llu), fix it urgently!\n", 4945 dev->name, dev_net(dev)->net_cookie); 4946 4947 rc = -ENETDOWN; 4948 } 4949 4950 reason = SKB_DROP_REASON_RECURSION_LIMIT; 4951 drop: 4952 rcu_read_unlock_bh(); 4953 4954 dev_core_stats_tx_dropped_inc(dev); 4955 kfree_skb_list_reason(skb, reason); 4956 return rc; 4957 out: 4958 rcu_read_unlock_bh(); 4959 return rc; 4960 } 4961 EXPORT_SYMBOL(__dev_queue_xmit); 4962 4963 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id) 4964 { 4965 struct net_device *dev = skb->dev; 4966 struct sk_buff *orig_skb = skb; 4967 struct netdev_queue *txq; 4968 int ret = NETDEV_TX_BUSY; 4969 bool again = false; 4970 4971 if (unlikely(!netif_running(dev) || 4972 !netif_carrier_ok(dev))) 4973 goto drop; 4974 4975 skb = validate_xmit_skb_list(skb, dev, &again); 4976 if (skb != orig_skb) 4977 goto drop; 4978 4979 skb_set_queue_mapping(skb, queue_id); 4980 txq = skb_get_tx_queue(dev, skb); 4981 4982 local_bh_disable(); 4983 4984 dev_xmit_recursion_inc(); 4985 HARD_TX_LOCK(dev, txq, smp_processor_id()); 4986 if (!netif_xmit_frozen_or_drv_stopped(txq)) 4987 ret = netdev_start_xmit(skb, dev, txq, false); 4988 HARD_TX_UNLOCK(dev, txq); 4989 dev_xmit_recursion_dec(); 4990 4991 local_bh_enable(); 4992 return ret; 4993 drop: 4994 dev_core_stats_tx_dropped_inc(dev); 4995 kfree_skb_list(skb); 4996 return NET_XMIT_DROP; 4997 } 4998 EXPORT_SYMBOL(__dev_direct_xmit); 4999 5000 /************************************************************************* 5001 * Receiver routines 5002 *************************************************************************/ 5003 static DEFINE_PER_CPU(struct task_struct *, backlog_napi); 5004 5005 int weight_p __read_mostly = 64; /* old backlog weight */ 5006 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */ 5007 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */ 5008 5009 /* Called with irq disabled */ 5010 static inline void ____napi_schedule(struct softnet_data *sd, 5011 struct napi_struct *napi) 5012 { 5013 struct task_struct *thread; 5014 5015 lockdep_assert_irqs_disabled(); 5016 5017 if (test_bit(NAPI_STATE_THREADED, &napi->state)) { 5018 /* Paired with smp_mb__before_atomic() in 5019 * napi_enable()/netif_set_threaded(). 5020 * Use READ_ONCE() to guarantee a complete 5021 * read on napi->thread. Only call 5022 * wake_up_process() when it's not NULL. 5023 */ 5024 thread = READ_ONCE(napi->thread); 5025 if (thread) { 5026 if (use_backlog_threads() && thread == raw_cpu_read(backlog_napi)) 5027 goto use_local_napi; 5028 5029 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 5030 wake_up_process(thread); 5031 return; 5032 } 5033 } 5034 5035 use_local_napi: 5036 DEBUG_NET_WARN_ON_ONCE(!list_empty(&napi->poll_list)); 5037 list_add_tail(&napi->poll_list, &sd->poll_list); 5038 WRITE_ONCE(napi->list_owner, smp_processor_id()); 5039 /* If not called from net_rx_action() 5040 * we have to raise NET_RX_SOFTIRQ. 5041 */ 5042 if (!sd->in_net_rx_action) 5043 raise_softirq_irqoff(NET_RX_SOFTIRQ); 5044 } 5045 5046 #ifdef CONFIG_RPS 5047 5048 struct static_key_false rps_needed __read_mostly; 5049 EXPORT_SYMBOL(rps_needed); 5050 struct static_key_false rfs_needed __read_mostly; 5051 EXPORT_SYMBOL(rfs_needed); 5052 5053 static u32 rfs_slot(u32 hash, rps_tag_ptr tag_ptr) 5054 { 5055 return hash_32(hash, rps_tag_to_log(tag_ptr)); 5056 } 5057 5058 #ifdef CONFIG_RFS_ACCEL 5059 /** 5060 * rps_flow_is_active - check whether the flow is recently active. 5061 * @rflow: Specific flow to check activity. 5062 * @log: ilog2(hashsize). 5063 * @cpu: CPU saved in @rflow. 5064 * 5065 * If the CPU has processed many packets since the flow's last activity 5066 * (beyond 10 times the table size), the flow is considered stale. 5067 * 5068 * Return: true if flow was recently active. 5069 */ 5070 static bool rps_flow_is_active(struct rps_dev_flow *rflow, 5071 u8 log, 5072 unsigned int cpu) 5073 { 5074 unsigned int flow_last_active; 5075 unsigned int sd_input_head; 5076 5077 if (cpu >= nr_cpu_ids) 5078 return false; 5079 5080 sd_input_head = READ_ONCE(per_cpu(softnet_data, cpu).input_queue_head); 5081 flow_last_active = READ_ONCE(rflow->last_qtail); 5082 5083 return (int)(sd_input_head - flow_last_active) < 5084 (int)(10 << log); 5085 } 5086 #endif 5087 5088 static struct rps_dev_flow * 5089 set_rps_cpu(struct net_device *dev, struct sk_buff *skb, 5090 struct rps_dev_flow *rflow, u16 next_cpu, u32 hash) 5091 { 5092 if (next_cpu < nr_cpu_ids) { 5093 u32 head; 5094 #ifdef CONFIG_RFS_ACCEL 5095 struct netdev_rx_queue *rxqueue; 5096 struct rps_dev_flow *flow_table; 5097 struct rps_dev_flow *old_rflow; 5098 struct rps_dev_flow *tmp_rflow; 5099 rps_tag_ptr q_tag_ptr; 5100 unsigned int tmp_cpu; 5101 u16 rxq_index; 5102 u32 flow_id; 5103 int rc; 5104 5105 /* Should we steer this flow to a different hardware queue? */ 5106 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap || 5107 !(dev->features & NETIF_F_NTUPLE)) 5108 goto out; 5109 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu); 5110 if (rxq_index == skb_get_rx_queue(skb)) 5111 goto out; 5112 5113 rxqueue = dev->_rx + rxq_index; 5114 q_tag_ptr = READ_ONCE(rxqueue->rps_flow_table); 5115 if (!q_tag_ptr) 5116 goto out; 5117 5118 flow_id = rfs_slot(hash, q_tag_ptr); 5119 flow_table = rps_tag_to_table(q_tag_ptr); 5120 tmp_rflow = flow_table + flow_id; 5121 tmp_cpu = READ_ONCE(tmp_rflow->cpu); 5122 5123 if (READ_ONCE(tmp_rflow->filter) != RPS_NO_FILTER) { 5124 if (rps_flow_is_active(tmp_rflow, 5125 rps_tag_to_log(q_tag_ptr), 5126 tmp_cpu)) { 5127 if (hash != READ_ONCE(tmp_rflow->hash) || 5128 next_cpu == tmp_cpu) 5129 goto out; 5130 } 5131 } 5132 5133 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb, 5134 rxq_index, flow_id); 5135 if (rc < 0) 5136 goto out; 5137 5138 old_rflow = rflow; 5139 rflow = tmp_rflow; 5140 WRITE_ONCE(rflow->filter, rc); 5141 WRITE_ONCE(rflow->hash, hash); 5142 5143 if (old_rflow->filter == rc) 5144 WRITE_ONCE(old_rflow->filter, RPS_NO_FILTER); 5145 out: 5146 #endif 5147 head = READ_ONCE(per_cpu(softnet_data, next_cpu).input_queue_head); 5148 rps_input_queue_tail_save(&rflow->last_qtail, head); 5149 } 5150 5151 WRITE_ONCE(rflow->cpu, next_cpu); 5152 return rflow; 5153 } 5154 5155 /* 5156 * get_rps_cpu is called from netif_receive_skb and returns the target 5157 * CPU from the RPS map of the receiving queue for a given skb. 5158 * rcu_read_lock must be held on entry. 5159 */ 5160 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb, 5161 struct rps_dev_flow **rflowp) 5162 { 5163 struct netdev_rx_queue *rxqueue = dev->_rx; 5164 rps_tag_ptr global_tag_ptr, q_tag_ptr; 5165 struct rps_map *map; 5166 int cpu = -1; 5167 u32 tcpu; 5168 u32 hash; 5169 5170 if (skb_rx_queue_recorded(skb)) { 5171 u16 index = skb_get_rx_queue(skb); 5172 5173 if (unlikely(index >= dev->real_num_rx_queues)) { 5174 WARN_ONCE(dev->real_num_rx_queues > 1, 5175 "%s received packet on queue %u, but number " 5176 "of RX queues is %u\n", 5177 dev->name, index, dev->real_num_rx_queues); 5178 goto done; 5179 } 5180 rxqueue += index; 5181 } 5182 5183 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */ 5184 5185 q_tag_ptr = READ_ONCE(rxqueue->rps_flow_table); 5186 map = rcu_dereference(rxqueue->rps_map); 5187 if (!q_tag_ptr && !map) 5188 goto done; 5189 5190 skb_reset_network_header(skb); 5191 hash = skb_get_hash(skb); 5192 if (!hash) 5193 goto done; 5194 5195 global_tag_ptr = READ_ONCE(net_hotdata.rps_sock_flow_table); 5196 if (q_tag_ptr && global_tag_ptr) { 5197 struct rps_sock_flow_table *sock_flow_table; 5198 struct rps_dev_flow *flow_table; 5199 struct rps_dev_flow *rflow; 5200 u32 next_cpu; 5201 u32 flow_id; 5202 u32 ident; 5203 5204 /* First check into global flow table if there is a match. 5205 * This READ_ONCE() pairs with WRITE_ONCE() from rps_record_sock_flow(). 5206 */ 5207 flow_id = hash & rps_tag_to_mask(global_tag_ptr); 5208 sock_flow_table = rps_tag_to_table(global_tag_ptr); 5209 ident = READ_ONCE(sock_flow_table[flow_id].ent); 5210 if ((ident ^ hash) & ~net_hotdata.rps_cpu_mask) 5211 goto try_rps; 5212 5213 next_cpu = ident & net_hotdata.rps_cpu_mask; 5214 5215 /* OK, now we know there is a match, 5216 * we can look at the local (per receive queue) flow table 5217 */ 5218 flow_id = rfs_slot(hash, q_tag_ptr); 5219 flow_table = rps_tag_to_table(q_tag_ptr); 5220 rflow = flow_table + flow_id; 5221 tcpu = rflow->cpu; 5222 5223 /* 5224 * If the desired CPU (where last recvmsg was done) is 5225 * different from current CPU (one in the rx-queue flow 5226 * table entry), switch if one of the following holds: 5227 * - Current CPU is unset (>= nr_cpu_ids). 5228 * - Current CPU is offline. 5229 * - The current CPU's queue tail has advanced beyond the 5230 * last packet that was enqueued using this table entry. 5231 * This guarantees that all previous packets for the flow 5232 * have been dequeued, thus preserving in order delivery. 5233 */ 5234 if (unlikely(tcpu != next_cpu) && 5235 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) || 5236 ((int)(READ_ONCE(per_cpu(softnet_data, tcpu).input_queue_head) - 5237 rflow->last_qtail)) >= 0)) { 5238 tcpu = next_cpu; 5239 rflow = set_rps_cpu(dev, skb, rflow, next_cpu, hash); 5240 } 5241 5242 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) { 5243 *rflowp = rflow; 5244 cpu = tcpu; 5245 goto done; 5246 } 5247 } 5248 5249 try_rps: 5250 5251 if (map) { 5252 tcpu = map->cpus[reciprocal_scale(hash, map->len)]; 5253 if (cpu_online(tcpu)) { 5254 cpu = tcpu; 5255 goto done; 5256 } 5257 } 5258 5259 done: 5260 return cpu; 5261 } 5262 5263 #ifdef CONFIG_RFS_ACCEL 5264 5265 /** 5266 * rps_may_expire_flow - check whether an RFS hardware filter may be removed 5267 * @dev: Device on which the filter was set 5268 * @rxq_index: RX queue index 5269 * @flow_id: Flow ID passed to ndo_rx_flow_steer() 5270 * @filter_id: Filter ID returned by ndo_rx_flow_steer() 5271 * 5272 * Drivers that implement ndo_rx_flow_steer() should periodically call 5273 * this function for each installed filter and remove the filters for 5274 * which it returns %true. 5275 */ 5276 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index, 5277 u32 flow_id, u16 filter_id) 5278 { 5279 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index; 5280 struct rps_dev_flow *flow_table; 5281 struct rps_dev_flow *rflow; 5282 rps_tag_ptr q_tag_ptr; 5283 bool expire = true; 5284 u8 log; 5285 5286 rcu_read_lock(); 5287 q_tag_ptr = READ_ONCE(rxqueue->rps_flow_table); 5288 log = rps_tag_to_log(q_tag_ptr); 5289 if (q_tag_ptr && flow_id < (1UL << log)) { 5290 unsigned int cpu; 5291 5292 flow_table = rps_tag_to_table(q_tag_ptr); 5293 rflow = flow_table + flow_id; 5294 cpu = READ_ONCE(rflow->cpu); 5295 if (READ_ONCE(rflow->filter) == filter_id && 5296 rps_flow_is_active(rflow, log, cpu)) 5297 expire = false; 5298 } 5299 rcu_read_unlock(); 5300 return expire; 5301 } 5302 EXPORT_SYMBOL(rps_may_expire_flow); 5303 5304 #endif /* CONFIG_RFS_ACCEL */ 5305 5306 /* Called from hardirq (IPI) context */ 5307 static void rps_trigger_softirq(void *data) 5308 { 5309 struct softnet_data *sd = data; 5310 5311 ____napi_schedule(sd, &sd->backlog); 5312 /* Pairs with READ_ONCE() in softnet_seq_show() */ 5313 WRITE_ONCE(sd->received_rps, sd->received_rps + 1); 5314 } 5315 5316 #endif /* CONFIG_RPS */ 5317 5318 /* Called from hardirq (IPI) context */ 5319 static void trigger_rx_softirq(void *data) 5320 { 5321 struct softnet_data *sd = data; 5322 5323 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 5324 smp_store_release(&sd->defer_ipi_scheduled, 0); 5325 } 5326 5327 /* 5328 * After we queued a packet into sd->input_pkt_queue, 5329 * we need to make sure this queue is serviced soon. 5330 * 5331 * - If this is another cpu queue, link it to our rps_ipi_list, 5332 * and make sure we will process rps_ipi_list from net_rx_action(). 5333 * 5334 * - If this is our own queue, NAPI schedule our backlog. 5335 * Note that this also raises NET_RX_SOFTIRQ. 5336 */ 5337 static void napi_schedule_rps(struct softnet_data *sd) 5338 { 5339 struct softnet_data *mysd = this_cpu_ptr(&softnet_data); 5340 5341 #ifdef CONFIG_RPS 5342 if (sd != mysd) { 5343 if (use_backlog_threads()) { 5344 __napi_schedule_irqoff(&sd->backlog); 5345 return; 5346 } 5347 5348 sd->rps_ipi_next = mysd->rps_ipi_list; 5349 mysd->rps_ipi_list = sd; 5350 5351 /* If not called from net_rx_action() or napi_threaded_poll() 5352 * we have to raise NET_RX_SOFTIRQ. 5353 */ 5354 if (!mysd->in_net_rx_action && !mysd->in_napi_threaded_poll) 5355 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 5356 return; 5357 } 5358 #endif /* CONFIG_RPS */ 5359 __napi_schedule_irqoff(&mysd->backlog); 5360 } 5361 5362 void kick_defer_list_purge(unsigned int cpu) 5363 { 5364 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 5365 unsigned long flags; 5366 5367 if (use_backlog_threads()) { 5368 backlog_lock_irq_save(sd, &flags); 5369 5370 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) 5371 __napi_schedule_irqoff(&sd->backlog); 5372 5373 backlog_unlock_irq_restore(sd, flags); 5374 5375 } else if (!cmpxchg(&sd->defer_ipi_scheduled, 0, 1)) { 5376 smp_call_function_single_async(cpu, &sd->defer_csd); 5377 } 5378 } 5379 5380 #ifdef CONFIG_NET_FLOW_LIMIT 5381 int netdev_flow_limit_table_len __read_mostly = (1 << 12); 5382 #endif 5383 5384 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen, 5385 int max_backlog) 5386 { 5387 #ifdef CONFIG_NET_FLOW_LIMIT 5388 unsigned int old_flow, new_flow; 5389 const struct softnet_data *sd; 5390 struct sd_flow_limit *fl; 5391 5392 if (likely(qlen < (max_backlog >> 1))) 5393 return false; 5394 5395 sd = this_cpu_ptr(&softnet_data); 5396 5397 rcu_read_lock(); 5398 fl = rcu_dereference(sd->flow_limit); 5399 if (fl) { 5400 new_flow = hash_32(skb_get_hash(skb), fl->log_buckets); 5401 old_flow = fl->history[fl->history_head]; 5402 fl->history[fl->history_head] = new_flow; 5403 5404 fl->history_head++; 5405 fl->history_head &= FLOW_LIMIT_HISTORY - 1; 5406 5407 if (likely(fl->buckets[old_flow])) 5408 fl->buckets[old_flow]--; 5409 5410 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) { 5411 /* Pairs with READ_ONCE() in softnet_seq_show() */ 5412 WRITE_ONCE(fl->count, fl->count + 1); 5413 rcu_read_unlock(); 5414 return true; 5415 } 5416 } 5417 rcu_read_unlock(); 5418 #endif 5419 return false; 5420 } 5421 5422 /* 5423 * enqueue_to_backlog is called to queue an skb to a per CPU backlog 5424 * queue (may be a remote CPU queue). 5425 */ 5426 static int enqueue_to_backlog(struct sk_buff *skb, int cpu, 5427 unsigned int *qtail) 5428 { 5429 enum skb_drop_reason reason; 5430 struct softnet_data *sd; 5431 unsigned long flags; 5432 unsigned int qlen; 5433 int max_backlog; 5434 u32 tail; 5435 5436 reason = SKB_DROP_REASON_DEV_READY; 5437 5438 sd = &per_cpu(softnet_data, cpu); 5439 5440 qlen = skb_queue_len_lockless(&sd->input_pkt_queue); 5441 max_backlog = READ_ONCE(net_hotdata.max_backlog); 5442 if (unlikely(qlen > max_backlog) || 5443 skb_flow_limit(skb, qlen, max_backlog)) 5444 goto cpu_backlog_drop; 5445 backlog_lock_irq_save(sd, &flags); 5446 qlen = skb_queue_len(&sd->input_pkt_queue); 5447 if (likely(qlen <= max_backlog)) { 5448 if (unlikely(!netif_running(skb->dev))) { 5449 backlog_unlock_irq_restore(sd, flags); 5450 goto bad_dev; 5451 } 5452 if (!qlen) { 5453 /* Schedule NAPI for backlog device. We can use 5454 * non atomic operation as we own the queue lock. 5455 */ 5456 if (!__test_and_set_bit(NAPI_STATE_SCHED, 5457 &sd->backlog.state)) 5458 napi_schedule_rps(sd); 5459 } 5460 __skb_queue_tail(&sd->input_pkt_queue, skb); 5461 tail = rps_input_queue_tail_incr(sd); 5462 backlog_unlock_irq_restore(sd, flags); 5463 5464 /* save the tail outside of the critical section */ 5465 rps_input_queue_tail_save(qtail, tail); 5466 return NET_RX_SUCCESS; 5467 } 5468 5469 backlog_unlock_irq_restore(sd, flags); 5470 5471 cpu_backlog_drop: 5472 reason = SKB_DROP_REASON_CPU_BACKLOG; 5473 numa_drop_add(&sd->drop_counters, 1); 5474 bad_dev: 5475 dev_core_stats_rx_dropped_inc(skb->dev); 5476 kfree_skb_reason(skb, reason); 5477 return NET_RX_DROP; 5478 } 5479 5480 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb) 5481 { 5482 struct net_device *dev = skb->dev; 5483 struct netdev_rx_queue *rxqueue; 5484 5485 rxqueue = dev->_rx; 5486 5487 if (skb_rx_queue_recorded(skb)) { 5488 u16 index = skb_get_rx_queue(skb); 5489 5490 if (unlikely(index >= dev->real_num_rx_queues)) { 5491 WARN_ONCE(dev->real_num_rx_queues > 1, 5492 "%s received packet on queue %u, but number " 5493 "of RX queues is %u\n", 5494 dev->name, index, dev->real_num_rx_queues); 5495 5496 return rxqueue; /* Return first rxqueue */ 5497 } 5498 rxqueue += index; 5499 } 5500 return rxqueue; 5501 } 5502 5503 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, 5504 const struct bpf_prog *xdp_prog) 5505 { 5506 void *orig_data, *orig_data_end, *hard_start; 5507 struct netdev_rx_queue *rxqueue; 5508 bool orig_bcast, orig_host; 5509 u32 mac_len, frame_sz; 5510 __be16 orig_eth_type; 5511 struct ethhdr *eth; 5512 u32 metalen, act; 5513 int off; 5514 5515 /* The XDP program wants to see the packet starting at the MAC 5516 * header. 5517 */ 5518 mac_len = skb->data - skb_mac_header(skb); 5519 hard_start = skb->data - skb_headroom(skb); 5520 5521 /* SKB "head" area always have tailroom for skb_shared_info */ 5522 frame_sz = (void *)skb_end_pointer(skb) - hard_start; 5523 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 5524 5525 rxqueue = netif_get_rxqueue(skb); 5526 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq); 5527 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len, 5528 skb_headlen(skb) + mac_len, true); 5529 if (skb_is_nonlinear(skb)) { 5530 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 5531 xdp_buff_set_frags_flag(xdp); 5532 } else { 5533 xdp_buff_clear_frags_flag(xdp); 5534 } 5535 5536 orig_data_end = xdp->data_end; 5537 orig_data = xdp->data; 5538 eth = (struct ethhdr *)xdp->data; 5539 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr); 5540 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest); 5541 orig_eth_type = eth->h_proto; 5542 5543 act = bpf_prog_run_xdp(xdp_prog, xdp); 5544 5545 /* check if bpf_xdp_adjust_head was used */ 5546 off = xdp->data - orig_data; 5547 if (off) { 5548 if (off > 0) 5549 __skb_pull(skb, off); 5550 else if (off < 0) 5551 __skb_push(skb, -off); 5552 5553 skb->mac_header += off; 5554 skb_reset_network_header(skb); 5555 } 5556 5557 /* check if bpf_xdp_adjust_tail was used */ 5558 off = xdp->data_end - orig_data_end; 5559 if (off != 0) { 5560 skb_set_tail_pointer(skb, xdp->data_end - xdp->data); 5561 skb->len += off; /* positive on grow, negative on shrink */ 5562 } 5563 5564 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 5565 * (e.g. bpf_xdp_adjust_tail). Remove the old fragment contribution 5566 * from skb->len before updating data_len, then add the new one back. 5567 */ 5568 skb->len -= skb->data_len; 5569 if (xdp_buff_has_frags(xdp)) { 5570 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 5571 skb->len += skb->data_len; 5572 } else { 5573 skb->data_len = 0; 5574 } 5575 5576 /* check if XDP changed eth hdr such SKB needs update */ 5577 eth = (struct ethhdr *)xdp->data; 5578 if ((orig_eth_type != eth->h_proto) || 5579 (orig_host != ether_addr_equal_64bits(eth->h_dest, 5580 skb->dev->dev_addr)) || 5581 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) { 5582 __skb_push(skb, ETH_HLEN); 5583 skb->pkt_type = PACKET_HOST; 5584 skb->protocol = eth_type_trans(skb, skb->dev); 5585 } 5586 5587 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull 5588 * before calling us again on redirect path. We do not call do_redirect 5589 * as we leave that up to the caller. 5590 * 5591 * Caller is responsible for managing lifetime of skb (i.e. calling 5592 * kfree_skb in response to actions it cannot handle/XDP_DROP). 5593 */ 5594 switch (act) { 5595 case XDP_REDIRECT: 5596 case XDP_TX: 5597 __skb_push(skb, mac_len); 5598 break; 5599 case XDP_PASS: 5600 metalen = xdp->data - xdp->data_meta; 5601 if (metalen) 5602 skb_metadata_set(skb, metalen); 5603 break; 5604 } 5605 5606 return act; 5607 } 5608 5609 static int 5610 netif_skb_check_for_xdp(struct sk_buff **pskb, const struct bpf_prog *prog) 5611 { 5612 struct sk_buff *skb = *pskb; 5613 int err, hroom, troom; 5614 5615 local_lock_nested_bh(&system_page_pool.bh_lock); 5616 err = skb_cow_data_for_xdp(this_cpu_read(system_page_pool.pool), pskb, prog); 5617 local_unlock_nested_bh(&system_page_pool.bh_lock); 5618 if (!err) 5619 return 0; 5620 5621 /* In case we have to go down the path and also linearize, 5622 * then lets do the pskb_expand_head() work just once here. 5623 */ 5624 hroom = XDP_PACKET_HEADROOM - skb_headroom(skb); 5625 troom = skb->tail + skb->data_len - skb->end; 5626 err = pskb_expand_head(skb, 5627 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0, 5628 troom > 0 ? troom + 128 : 0, GFP_ATOMIC); 5629 if (err) 5630 return err; 5631 5632 return skb_linearize(skb); 5633 } 5634 5635 static u32 netif_receive_generic_xdp(struct sk_buff **pskb, 5636 struct xdp_buff *xdp, 5637 const struct bpf_prog *xdp_prog) 5638 { 5639 struct sk_buff *skb = *pskb; 5640 u32 mac_len, act = XDP_DROP; 5641 5642 /* Reinjected packets coming from act_mirred or similar should 5643 * not get XDP generic processing. 5644 */ 5645 if (skb_is_redirected(skb)) 5646 return XDP_PASS; 5647 5648 /* XDP packets must have sufficient headroom of XDP_PACKET_HEADROOM 5649 * bytes. This is the guarantee that also native XDP provides, 5650 * thus we need to do it here as well. 5651 */ 5652 mac_len = skb->data - skb_mac_header(skb); 5653 __skb_push(skb, mac_len); 5654 5655 if (skb_cloned(skb) || skb_is_nonlinear(skb) || 5656 skb_headroom(skb) < XDP_PACKET_HEADROOM) { 5657 if (netif_skb_check_for_xdp(pskb, xdp_prog)) 5658 goto do_drop; 5659 } 5660 5661 __skb_pull(*pskb, mac_len); 5662 5663 act = bpf_prog_run_generic_xdp(*pskb, xdp, xdp_prog); 5664 switch (act) { 5665 case XDP_REDIRECT: 5666 case XDP_TX: 5667 case XDP_PASS: 5668 break; 5669 default: 5670 bpf_warn_invalid_xdp_action((*pskb)->dev, xdp_prog, act); 5671 fallthrough; 5672 case XDP_ABORTED: 5673 trace_xdp_exception((*pskb)->dev, xdp_prog, act); 5674 fallthrough; 5675 case XDP_DROP: 5676 do_drop: 5677 kfree_skb(*pskb); 5678 break; 5679 } 5680 5681 return act; 5682 } 5683 5684 /* When doing generic XDP we have to bypass the qdisc layer and the 5685 * network taps in order to match in-driver-XDP behavior. This also means 5686 * that XDP packets are able to starve other packets going through a qdisc, 5687 * and DDOS attacks will be more effective. In-driver-XDP use dedicated TX 5688 * queues, so they do not have this starvation issue. 5689 */ 5690 void generic_xdp_tx(struct sk_buff *skb, const struct bpf_prog *xdp_prog) 5691 { 5692 struct net_device *dev = skb->dev; 5693 struct netdev_queue *txq; 5694 bool free_skb = true; 5695 int cpu, rc; 5696 5697 txq = netdev_core_pick_tx(dev, skb, NULL); 5698 cpu = smp_processor_id(); 5699 HARD_TX_LOCK(dev, txq, cpu); 5700 if (!netif_xmit_frozen_or_drv_stopped(txq)) { 5701 rc = netdev_start_xmit(skb, dev, txq, 0); 5702 if (dev_xmit_complete(rc)) 5703 free_skb = false; 5704 } 5705 HARD_TX_UNLOCK(dev, txq); 5706 if (free_skb) { 5707 trace_xdp_exception(dev, xdp_prog, XDP_TX); 5708 dev_core_stats_tx_dropped_inc(dev); 5709 kfree_skb(skb); 5710 } 5711 } 5712 5713 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key); 5714 5715 int do_xdp_generic(const struct bpf_prog *xdp_prog, struct sk_buff **pskb) 5716 { 5717 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 5718 5719 if (xdp_prog) { 5720 struct xdp_buff xdp; 5721 u32 act; 5722 int err; 5723 5724 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 5725 act = netif_receive_generic_xdp(pskb, &xdp, xdp_prog); 5726 if (act != XDP_PASS) { 5727 switch (act) { 5728 case XDP_REDIRECT: 5729 err = xdp_do_generic_redirect((*pskb)->dev, *pskb, 5730 &xdp, xdp_prog); 5731 if (err) 5732 goto out_redir; 5733 break; 5734 case XDP_TX: 5735 generic_xdp_tx(*pskb, xdp_prog); 5736 break; 5737 } 5738 bpf_net_ctx_clear(bpf_net_ctx); 5739 return XDP_DROP; 5740 } 5741 bpf_net_ctx_clear(bpf_net_ctx); 5742 } 5743 return XDP_PASS; 5744 out_redir: 5745 bpf_net_ctx_clear(bpf_net_ctx); 5746 kfree_skb_reason(*pskb, SKB_DROP_REASON_XDP); 5747 return XDP_DROP; 5748 } 5749 EXPORT_SYMBOL_GPL(do_xdp_generic); 5750 5751 static int netif_rx_internal(struct sk_buff *skb) 5752 { 5753 int ret; 5754 5755 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5756 5757 trace_netif_rx(skb); 5758 5759 #ifdef CONFIG_RPS 5760 if (static_branch_unlikely(&rps_needed)) { 5761 struct rps_dev_flow voidflow, *rflow = &voidflow; 5762 int cpu; 5763 5764 rcu_read_lock(); 5765 5766 cpu = get_rps_cpu(skb->dev, skb, &rflow); 5767 if (cpu < 0) 5768 cpu = smp_processor_id(); 5769 5770 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5771 5772 rcu_read_unlock(); 5773 } else 5774 #endif 5775 { 5776 unsigned int qtail; 5777 5778 ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail); 5779 } 5780 return ret; 5781 } 5782 5783 /** 5784 * __netif_rx - Slightly optimized version of netif_rx 5785 * @skb: buffer to post 5786 * 5787 * This behaves as netif_rx except that it does not disable bottom halves. 5788 * As a result this function may only be invoked from the interrupt context 5789 * (either hard or soft interrupt). 5790 */ 5791 int __netif_rx(struct sk_buff *skb) 5792 { 5793 int ret; 5794 5795 lockdep_assert_once(hardirq_count() | softirq_count()); 5796 5797 trace_netif_rx_entry(skb); 5798 ret = netif_rx_internal(skb); 5799 trace_netif_rx_exit(ret); 5800 return ret; 5801 } 5802 EXPORT_SYMBOL(__netif_rx); 5803 5804 /** 5805 * netif_rx - post buffer to the network code 5806 * @skb: buffer to post 5807 * 5808 * This function receives a packet from a device driver and queues it for 5809 * the upper (protocol) levels to process via the backlog NAPI device. It 5810 * always succeeds. The buffer may be dropped during processing for 5811 * congestion control or by the protocol layers. 5812 * The network buffer is passed via the backlog NAPI device. Modern NIC 5813 * driver should use NAPI and GRO. 5814 * This function can used from interrupt and from process context. The 5815 * caller from process context must not disable interrupts before invoking 5816 * this function. 5817 * 5818 * return values: 5819 * NET_RX_SUCCESS (no congestion) 5820 * NET_RX_DROP (packet was dropped) 5821 * 5822 */ 5823 int netif_rx(struct sk_buff *skb) 5824 { 5825 bool need_bh_off = !(hardirq_count() | softirq_count()); 5826 int ret; 5827 5828 if (need_bh_off) 5829 local_bh_disable(); 5830 trace_netif_rx_entry(skb); 5831 ret = netif_rx_internal(skb); 5832 trace_netif_rx_exit(ret); 5833 if (need_bh_off) 5834 local_bh_enable(); 5835 return ret; 5836 } 5837 EXPORT_SYMBOL(netif_rx); 5838 5839 static __latent_entropy void net_tx_action(void) 5840 { 5841 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 5842 5843 if (sd->completion_queue) { 5844 struct sk_buff *clist; 5845 5846 local_irq_disable(); 5847 clist = sd->completion_queue; 5848 sd->completion_queue = NULL; 5849 local_irq_enable(); 5850 5851 while (clist) { 5852 struct sk_buff *skb = clist; 5853 5854 clist = clist->next; 5855 5856 WARN_ON(refcount_read(&skb->users)); 5857 if (likely(get_kfree_skb_cb(skb)->reason == SKB_CONSUMED)) 5858 trace_consume_skb(skb, net_tx_action); 5859 else 5860 trace_kfree_skb(skb, net_tx_action, 5861 get_kfree_skb_cb(skb)->reason, NULL); 5862 5863 if (skb->fclone != SKB_FCLONE_UNAVAILABLE) 5864 __kfree_skb(skb); 5865 else 5866 __napi_kfree_skb(skb, 5867 get_kfree_skb_cb(skb)->reason); 5868 } 5869 } 5870 5871 if (sd->output_queue) { 5872 struct Qdisc *head; 5873 5874 local_irq_disable(); 5875 head = sd->output_queue; 5876 sd->output_queue = NULL; 5877 sd->output_queue_tailp = &sd->output_queue; 5878 local_irq_enable(); 5879 5880 rcu_read_lock(); 5881 5882 while (head) { 5883 spinlock_t *root_lock = NULL; 5884 struct sk_buff *to_free; 5885 struct Qdisc *q = head; 5886 5887 head = head->next_sched; 5888 5889 /* We need to make sure head->next_sched is read 5890 * before clearing __QDISC_STATE_SCHED 5891 */ 5892 smp_mb__before_atomic(); 5893 5894 if (!(q->flags & TCQ_F_NOLOCK)) { 5895 root_lock = qdisc_lock(q); 5896 spin_lock(root_lock); 5897 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, 5898 &q->state))) { 5899 /* There is a synchronize_net() between 5900 * STATE_DEACTIVATED flag being set and 5901 * qdisc_reset()/some_qdisc_is_busy() in 5902 * dev_deactivate(), so we can safely bail out 5903 * early here to avoid data race between 5904 * qdisc_deactivate() and some_qdisc_is_busy() 5905 * for lockless qdisc. 5906 */ 5907 clear_bit(__QDISC_STATE_SCHED, &q->state); 5908 continue; 5909 } 5910 5911 clear_bit(__QDISC_STATE_SCHED, &q->state); 5912 to_free = qdisc_run(q); 5913 if (root_lock) 5914 spin_unlock(root_lock); 5915 tcf_kfree_skb_list(to_free, q, NULL, qdisc_dev(q)); 5916 } 5917 5918 rcu_read_unlock(); 5919 } 5920 5921 xfrm_dev_backlog(sd); 5922 } 5923 5924 /** 5925 * netdev_is_rx_handler_busy - check if receive handler is registered 5926 * @dev: device to check 5927 * 5928 * Check if a receive handler is already registered for a given device. 5929 * Return true if there one. 5930 * 5931 * The caller must hold the rtnl_mutex. 5932 */ 5933 bool netdev_is_rx_handler_busy(struct net_device *dev) 5934 { 5935 ASSERT_RTNL(); 5936 return dev && rtnl_dereference(dev->rx_handler); 5937 } 5938 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy); 5939 5940 /** 5941 * netdev_rx_handler_register - register receive handler 5942 * @dev: device to register a handler for 5943 * @rx_handler: receive handler to register 5944 * @rx_handler_data: data pointer that is used by rx handler 5945 * 5946 * Register a receive handler for a device. This handler will then be 5947 * called from __netif_receive_skb. A negative errno code is returned 5948 * on a failure. 5949 * 5950 * The caller must hold the rtnl_mutex. 5951 * 5952 * For a general description of rx_handler, see enum rx_handler_result. 5953 */ 5954 int netdev_rx_handler_register(struct net_device *dev, 5955 rx_handler_func_t *rx_handler, 5956 void *rx_handler_data) 5957 { 5958 if (netdev_is_rx_handler_busy(dev)) 5959 return -EBUSY; 5960 5961 if (dev->priv_flags & IFF_NO_RX_HANDLER) 5962 return -EINVAL; 5963 5964 /* Note: rx_handler_data must be set before rx_handler */ 5965 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data); 5966 rcu_assign_pointer(dev->rx_handler, rx_handler); 5967 5968 return 0; 5969 } 5970 EXPORT_SYMBOL_GPL(netdev_rx_handler_register); 5971 5972 /** 5973 * netdev_rx_handler_unregister - unregister receive handler 5974 * @dev: device to unregister a handler from 5975 * 5976 * Unregister a receive handler from a device. 5977 * 5978 * The caller must hold the rtnl_mutex. 5979 */ 5980 void netdev_rx_handler_unregister(struct net_device *dev) 5981 { 5982 5983 ASSERT_RTNL(); 5984 RCU_INIT_POINTER(dev->rx_handler, NULL); 5985 /* a reader seeing a non NULL rx_handler in a rcu_read_lock() 5986 * section has a guarantee to see a non NULL rx_handler_data 5987 * as well. 5988 */ 5989 synchronize_net(); 5990 RCU_INIT_POINTER(dev->rx_handler_data, NULL); 5991 } 5992 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister); 5993 5994 /* 5995 * Limit the use of PFMEMALLOC reserves to those protocols that implement 5996 * the special handling of PFMEMALLOC skbs. 5997 */ 5998 static bool skb_pfmemalloc_protocol(struct sk_buff *skb) 5999 { 6000 switch (skb->protocol) { 6001 case htons(ETH_P_ARP): 6002 case htons(ETH_P_IP): 6003 case htons(ETH_P_IPV6): 6004 case htons(ETH_P_8021Q): 6005 case htons(ETH_P_8021AD): 6006 return true; 6007 default: 6008 return false; 6009 } 6010 } 6011 6012 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev, 6013 int *ret, struct net_device *orig_dev) 6014 { 6015 if (nf_hook_ingress_active(skb)) { 6016 int ingress_retval; 6017 6018 if (unlikely(*pt_prev)) { 6019 *ret = deliver_skb(skb, *pt_prev, orig_dev); 6020 *pt_prev = NULL; 6021 } 6022 6023 rcu_read_lock(); 6024 ingress_retval = nf_hook_ingress(skb); 6025 rcu_read_unlock(); 6026 return ingress_retval; 6027 } 6028 return 0; 6029 } 6030 6031 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc, 6032 struct packet_type **ppt_prev) 6033 { 6034 enum skb_drop_reason drop_reason = SKB_DROP_REASON_UNHANDLED_PROTO; 6035 struct packet_type *ptype, *pt_prev; 6036 rx_handler_func_t *rx_handler; 6037 struct sk_buff *skb = *pskb; 6038 struct net_device *orig_dev; 6039 bool deliver_exact = false; 6040 int ret = NET_RX_DROP; 6041 __be16 type; 6042 6043 net_timestamp_check(!READ_ONCE(net_hotdata.tstamp_prequeue), skb); 6044 6045 trace_netif_receive_skb(skb); 6046 6047 orig_dev = skb->dev; 6048 6049 skb_reset_network_header(skb); 6050 #if !defined(CONFIG_DEBUG_NET) 6051 /* We plan to no longer reset the transport header here. 6052 * Give some time to fuzzers and dev build to catch bugs 6053 * in network stacks. 6054 */ 6055 if (!skb_transport_header_was_set(skb)) 6056 skb_reset_transport_header(skb); 6057 #endif 6058 skb_reset_mac_len(skb); 6059 6060 pt_prev = NULL; 6061 6062 another_round: 6063 skb->skb_iif = skb->dev->ifindex; 6064 6065 __this_cpu_inc(softnet_data.processed); 6066 6067 if (static_branch_unlikely(&generic_xdp_needed_key)) { 6068 int ret2; 6069 6070 migrate_disable(); 6071 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), 6072 &skb); 6073 migrate_enable(); 6074 6075 if (ret2 != XDP_PASS) { 6076 ret = NET_RX_DROP; 6077 goto out; 6078 } 6079 } 6080 6081 if (eth_type_vlan(skb->protocol)) { 6082 skb = skb_vlan_untag(skb); 6083 if (unlikely(!skb)) 6084 goto out; 6085 } 6086 6087 if (skb_skip_tc_classify(skb)) 6088 goto skip_classify; 6089 6090 if (pfmemalloc) 6091 goto skip_taps; 6092 6093 list_for_each_entry_rcu(ptype, &dev_net_rcu(skb->dev)->ptype_all, 6094 list) { 6095 if (unlikely(pt_prev)) 6096 ret = deliver_skb(skb, pt_prev, orig_dev); 6097 pt_prev = ptype; 6098 } 6099 6100 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) { 6101 if (unlikely(pt_prev)) 6102 ret = deliver_skb(skb, pt_prev, orig_dev); 6103 pt_prev = ptype; 6104 } 6105 6106 skip_taps: 6107 #ifdef CONFIG_NET_INGRESS 6108 if (static_branch_unlikely(&ingress_needed_key)) { 6109 bool another = false; 6110 6111 nf_skip_egress(skb, true); 6112 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev, 6113 &another); 6114 if (another) 6115 goto another_round; 6116 if (!skb) 6117 goto out; 6118 6119 nf_skip_egress(skb, false); 6120 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0) 6121 goto out; 6122 } 6123 #endif 6124 skb_reset_redirect(skb); 6125 skip_classify: 6126 if (pfmemalloc && !skb_pfmemalloc_protocol(skb)) { 6127 drop_reason = SKB_DROP_REASON_PFMEMALLOC; 6128 goto drop; 6129 } 6130 6131 if (skb_vlan_tag_present(skb)) { 6132 if (unlikely(pt_prev)) { 6133 ret = deliver_skb(skb, pt_prev, orig_dev); 6134 pt_prev = NULL; 6135 } 6136 if (vlan_do_receive(&skb)) 6137 goto another_round; 6138 else if (unlikely(!skb)) 6139 goto out; 6140 } 6141 6142 rx_handler = rcu_dereference(skb->dev->rx_handler); 6143 if (rx_handler) { 6144 if (unlikely(pt_prev)) { 6145 ret = deliver_skb(skb, pt_prev, orig_dev); 6146 pt_prev = NULL; 6147 } 6148 switch (rx_handler(&skb)) { 6149 case RX_HANDLER_CONSUMED: 6150 ret = NET_RX_SUCCESS; 6151 goto out; 6152 case RX_HANDLER_ANOTHER: 6153 goto another_round; 6154 case RX_HANDLER_EXACT: 6155 deliver_exact = true; 6156 break; 6157 case RX_HANDLER_PASS: 6158 break; 6159 default: 6160 BUG(); 6161 } 6162 } 6163 6164 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) { 6165 check_vlan_id: 6166 if (skb_vlan_tag_get_id(skb)) { 6167 /* Vlan id is non 0 and vlan_do_receive() above couldn't 6168 * find vlan device. 6169 */ 6170 skb->pkt_type = PACKET_OTHERHOST; 6171 } else if (eth_type_vlan(skb->protocol)) { 6172 /* Outer header is 802.1P with vlan 0, inner header is 6173 * 802.1Q or 802.1AD and vlan_do_receive() above could 6174 * not find vlan dev for vlan id 0. 6175 */ 6176 __vlan_hwaccel_clear_tag(skb); 6177 skb = skb_vlan_untag(skb); 6178 if (unlikely(!skb)) 6179 goto out; 6180 if (vlan_do_receive(&skb)) 6181 /* After stripping off 802.1P header with vlan 0 6182 * vlan dev is found for inner header. 6183 */ 6184 goto another_round; 6185 else if (unlikely(!skb)) 6186 goto out; 6187 else 6188 /* We have stripped outer 802.1P vlan 0 header. 6189 * But could not find vlan dev. 6190 * check again for vlan id to set OTHERHOST. 6191 */ 6192 goto check_vlan_id; 6193 } 6194 /* Note: we might in the future use prio bits 6195 * and set skb->priority like in vlan_do_receive() 6196 * For the time being, just ignore Priority Code Point 6197 */ 6198 __vlan_hwaccel_clear_tag(skb); 6199 } 6200 6201 type = skb->protocol; 6202 6203 /* deliver only exact match when indicated */ 6204 if (likely(!deliver_exact)) { 6205 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6206 &ptype_base[ntohs(type) & 6207 PTYPE_HASH_MASK]); 6208 6209 /* orig_dev and skb->dev could belong to different netns; 6210 * Even in such case we need to traverse only the list 6211 * coming from skb->dev, as the ptype owner (packet socket) 6212 * will use dev_net(skb->dev) to do namespace filtering. 6213 */ 6214 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6215 &dev_net_rcu(skb->dev)->ptype_specific); 6216 } 6217 6218 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6219 &orig_dev->ptype_specific); 6220 6221 if (unlikely(skb->dev != orig_dev)) { 6222 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 6223 &skb->dev->ptype_specific); 6224 } 6225 6226 if (pt_prev) { 6227 *ppt_prev = pt_prev; 6228 } else { 6229 drop: 6230 if (!deliver_exact) 6231 dev_core_stats_rx_dropped_inc(skb->dev); 6232 else 6233 dev_core_stats_rx_nohandler_inc(skb->dev); 6234 6235 kfree_skb_reason(skb, drop_reason); 6236 /* Jamal, now you will not able to escape explaining 6237 * me how you were going to use this. :-) 6238 */ 6239 ret = NET_RX_DROP; 6240 } 6241 6242 out: 6243 /* The invariant here is that if *ppt_prev is not NULL 6244 * then skb should also be non-NULL. 6245 * 6246 * Apparently *ppt_prev assignment above holds this invariant due to 6247 * skb dereferencing near it. 6248 */ 6249 *pskb = skb; 6250 return ret; 6251 } 6252 6253 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc) 6254 { 6255 struct net_device *orig_dev = skb->dev; 6256 struct packet_type *pt_prev = NULL; 6257 int ret; 6258 6259 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 6260 if (pt_prev) 6261 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb, 6262 skb->dev, pt_prev, orig_dev); 6263 return ret; 6264 } 6265 6266 /** 6267 * netif_receive_skb_core - special purpose version of netif_receive_skb 6268 * @skb: buffer to process 6269 * 6270 * More direct receive version of netif_receive_skb(). It should 6271 * only be used by callers that have a need to skip RPS and Generic XDP. 6272 * Caller must also take care of handling if ``(page_is_)pfmemalloc``. 6273 * 6274 * This function may only be called from softirq context and interrupts 6275 * should be enabled. 6276 * 6277 * Return values (usually ignored): 6278 * NET_RX_SUCCESS: no congestion 6279 * NET_RX_DROP: packet was dropped 6280 */ 6281 int netif_receive_skb_core(struct sk_buff *skb) 6282 { 6283 int ret; 6284 6285 rcu_read_lock(); 6286 ret = __netif_receive_skb_one_core(skb, false); 6287 rcu_read_unlock(); 6288 6289 return ret; 6290 } 6291 EXPORT_SYMBOL(netif_receive_skb_core); 6292 6293 static inline void __netif_receive_skb_list_ptype(struct list_head *head, 6294 struct packet_type *pt_prev, 6295 struct net_device *orig_dev) 6296 { 6297 struct sk_buff *skb, *next; 6298 6299 if (!pt_prev) 6300 return; 6301 if (list_empty(head)) 6302 return; 6303 if (pt_prev->list_func != NULL) 6304 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv, 6305 ip_list_rcv, head, pt_prev, orig_dev); 6306 else 6307 list_for_each_entry_safe(skb, next, head, list) { 6308 skb_list_del_init(skb); 6309 pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 6310 } 6311 } 6312 6313 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc) 6314 { 6315 /* Fast-path assumptions: 6316 * - There is no RX handler. 6317 * - Only one packet_type matches. 6318 * If either of these fails, we will end up doing some per-packet 6319 * processing in-line, then handling the 'last ptype' for the whole 6320 * sublist. This can't cause out-of-order delivery to any single ptype, 6321 * because the 'last ptype' must be constant across the sublist, and all 6322 * other ptypes are handled per-packet. 6323 */ 6324 /* Current (common) ptype of sublist */ 6325 struct packet_type *pt_curr = NULL; 6326 /* Current (common) orig_dev of sublist */ 6327 struct net_device *od_curr = NULL; 6328 struct sk_buff *skb, *next; 6329 LIST_HEAD(sublist); 6330 6331 list_for_each_entry_safe(skb, next, head, list) { 6332 struct net_device *orig_dev = skb->dev; 6333 struct packet_type *pt_prev = NULL; 6334 6335 skb_list_del_init(skb); 6336 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 6337 if (!pt_prev) 6338 continue; 6339 if (pt_curr != pt_prev || od_curr != orig_dev) { 6340 /* dispatch old sublist */ 6341 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 6342 /* start new sublist */ 6343 INIT_LIST_HEAD(&sublist); 6344 pt_curr = pt_prev; 6345 od_curr = orig_dev; 6346 } 6347 list_add_tail(&skb->list, &sublist); 6348 } 6349 6350 /* dispatch final sublist */ 6351 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 6352 } 6353 6354 static int __netif_receive_skb(struct sk_buff *skb) 6355 { 6356 int ret; 6357 6358 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) { 6359 unsigned int noreclaim_flag; 6360 6361 /* 6362 * PFMEMALLOC skbs are special, they should 6363 * - be delivered to SOCK_MEMALLOC sockets only 6364 * - stay away from userspace 6365 * - have bounded memory usage 6366 * 6367 * Use PF_MEMALLOC as this saves us from propagating the allocation 6368 * context down to all allocation sites. 6369 */ 6370 noreclaim_flag = memalloc_noreclaim_save(); 6371 ret = __netif_receive_skb_one_core(skb, true); 6372 memalloc_noreclaim_restore(noreclaim_flag); 6373 } else 6374 ret = __netif_receive_skb_one_core(skb, false); 6375 6376 return ret; 6377 } 6378 6379 static void __netif_receive_skb_list(struct list_head *head) 6380 { 6381 unsigned long noreclaim_flag = 0; 6382 struct sk_buff *skb, *next; 6383 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */ 6384 6385 list_for_each_entry_safe(skb, next, head, list) { 6386 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) { 6387 struct list_head sublist; 6388 6389 /* Handle the previous sublist */ 6390 list_cut_before(&sublist, head, &skb->list); 6391 if (!list_empty(&sublist)) 6392 __netif_receive_skb_list_core(&sublist, pfmemalloc); 6393 pfmemalloc = !pfmemalloc; 6394 /* See comments in __netif_receive_skb */ 6395 if (pfmemalloc) 6396 noreclaim_flag = memalloc_noreclaim_save(); 6397 else 6398 memalloc_noreclaim_restore(noreclaim_flag); 6399 } 6400 } 6401 /* Handle the remaining sublist */ 6402 if (!list_empty(head)) 6403 __netif_receive_skb_list_core(head, pfmemalloc); 6404 /* Restore pflags */ 6405 if (pfmemalloc) 6406 memalloc_noreclaim_restore(noreclaim_flag); 6407 } 6408 6409 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp) 6410 { 6411 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog); 6412 struct bpf_prog *new = xdp->prog; 6413 int ret = 0; 6414 6415 switch (xdp->command) { 6416 case XDP_SETUP_PROG: 6417 rcu_assign_pointer(dev->xdp_prog, new); 6418 if (old) 6419 bpf_prog_put(old); 6420 6421 if (old && !new) { 6422 static_branch_dec(&generic_xdp_needed_key); 6423 } else if (new && !old) { 6424 static_branch_inc(&generic_xdp_needed_key); 6425 netif_disable_lro(dev); 6426 dev_disable_gro_hw(dev); 6427 } 6428 break; 6429 6430 default: 6431 ret = -EINVAL; 6432 break; 6433 } 6434 6435 return ret; 6436 } 6437 6438 static int netif_receive_skb_internal(struct sk_buff *skb) 6439 { 6440 int ret; 6441 6442 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 6443 6444 if (skb_defer_rx_timestamp(skb)) 6445 return NET_RX_SUCCESS; 6446 6447 rcu_read_lock(); 6448 #ifdef CONFIG_RPS 6449 if (static_branch_unlikely(&rps_needed)) { 6450 struct rps_dev_flow voidflow, *rflow = &voidflow; 6451 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6452 6453 if (cpu >= 0) { 6454 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6455 rcu_read_unlock(); 6456 return ret; 6457 } 6458 } 6459 #endif 6460 ret = __netif_receive_skb(skb); 6461 rcu_read_unlock(); 6462 return ret; 6463 } 6464 6465 void netif_receive_skb_list_internal(struct list_head *head) 6466 { 6467 struct sk_buff *skb, *next; 6468 LIST_HEAD(sublist); 6469 6470 list_for_each_entry_safe(skb, next, head, list) { 6471 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), 6472 skb); 6473 skb_list_del_init(skb); 6474 if (!skb_defer_rx_timestamp(skb)) 6475 list_add_tail(&skb->list, &sublist); 6476 } 6477 list_splice_init(&sublist, head); 6478 6479 rcu_read_lock(); 6480 #ifdef CONFIG_RPS 6481 if (static_branch_unlikely(&rps_needed)) { 6482 list_for_each_entry_safe(skb, next, head, list) { 6483 struct rps_dev_flow voidflow, *rflow = &voidflow; 6484 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6485 6486 if (cpu >= 0) { 6487 /* Will be handled, remove from list */ 6488 skb_list_del_init(skb); 6489 enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6490 } 6491 } 6492 } 6493 #endif 6494 __netif_receive_skb_list(head); 6495 rcu_read_unlock(); 6496 } 6497 6498 /** 6499 * netif_receive_skb - process receive buffer from network 6500 * @skb: buffer to process 6501 * 6502 * netif_receive_skb() is the main receive data processing function. 6503 * It always succeeds. The buffer may be dropped during processing 6504 * for congestion control or by the protocol layers. 6505 * 6506 * This function may only be called from softirq context and interrupts 6507 * should be enabled. 6508 * 6509 * Return values (usually ignored): 6510 * NET_RX_SUCCESS: no congestion 6511 * NET_RX_DROP: packet was dropped 6512 */ 6513 int netif_receive_skb(struct sk_buff *skb) 6514 { 6515 int ret; 6516 6517 trace_netif_receive_skb_entry(skb); 6518 6519 ret = netif_receive_skb_internal(skb); 6520 trace_netif_receive_skb_exit(ret); 6521 6522 return ret; 6523 } 6524 EXPORT_SYMBOL(netif_receive_skb); 6525 6526 /** 6527 * netif_receive_skb_list - process many receive buffers from network 6528 * @head: list of skbs to process. 6529 * 6530 * Since return value of netif_receive_skb() is normally ignored, and 6531 * wouldn't be meaningful for a list, this function returns void. 6532 * 6533 * This function may only be called from softirq context and interrupts 6534 * should be enabled. 6535 */ 6536 void netif_receive_skb_list(struct list_head *head) 6537 { 6538 struct sk_buff *skb; 6539 6540 if (list_empty(head)) 6541 return; 6542 if (trace_netif_receive_skb_list_entry_enabled()) { 6543 list_for_each_entry(skb, head, list) 6544 trace_netif_receive_skb_list_entry(skb); 6545 } 6546 netif_receive_skb_list_internal(head); 6547 trace_netif_receive_skb_list_exit(0); 6548 } 6549 EXPORT_SYMBOL(netif_receive_skb_list); 6550 6551 /* Network device is going away, flush any packets still pending */ 6552 static void flush_backlog(struct work_struct *work) 6553 { 6554 struct sk_buff *skb, *tmp; 6555 struct sk_buff_head list; 6556 struct softnet_data *sd; 6557 6558 __skb_queue_head_init(&list); 6559 local_bh_disable(); 6560 sd = this_cpu_ptr(&softnet_data); 6561 6562 backlog_lock_irq_disable(sd); 6563 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) { 6564 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6565 __skb_unlink(skb, &sd->input_pkt_queue); 6566 __skb_queue_tail(&list, skb); 6567 rps_input_queue_head_incr(sd); 6568 } 6569 } 6570 backlog_unlock_irq_enable(sd); 6571 6572 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6573 skb_queue_walk_safe(&sd->process_queue, skb, tmp) { 6574 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6575 __skb_unlink(skb, &sd->process_queue); 6576 __skb_queue_tail(&list, skb); 6577 rps_input_queue_head_incr(sd); 6578 } 6579 } 6580 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6581 local_bh_enable(); 6582 6583 __skb_queue_purge_reason(&list, SKB_DROP_REASON_DEV_READY); 6584 } 6585 6586 static bool flush_required(int cpu) 6587 { 6588 #if IS_ENABLED(CONFIG_RPS) 6589 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 6590 bool do_flush; 6591 6592 backlog_lock_irq_disable(sd); 6593 6594 /* as insertion into process_queue happens with the rps lock held, 6595 * process_queue access may race only with dequeue 6596 */ 6597 do_flush = !skb_queue_empty(&sd->input_pkt_queue) || 6598 !skb_queue_empty_lockless(&sd->process_queue); 6599 backlog_unlock_irq_enable(sd); 6600 6601 return do_flush; 6602 #endif 6603 /* without RPS we can't safely check input_pkt_queue: during a 6604 * concurrent remote skb_queue_splice() we can detect as empty both 6605 * input_pkt_queue and process_queue even if the latter could end-up 6606 * containing a lot of packets. 6607 */ 6608 return true; 6609 } 6610 6611 struct flush_backlogs { 6612 cpumask_t flush_cpus; 6613 struct work_struct w[]; 6614 }; 6615 6616 static struct flush_backlogs *flush_backlogs_alloc(void) 6617 { 6618 return kmalloc_flex(struct flush_backlogs, w, nr_cpu_ids); 6619 } 6620 6621 static struct flush_backlogs *flush_backlogs_fallback; 6622 static DEFINE_MUTEX(flush_backlogs_mutex); 6623 6624 static void flush_all_backlogs(void) 6625 { 6626 struct flush_backlogs *ptr = flush_backlogs_alloc(); 6627 unsigned int cpu; 6628 6629 if (!ptr) { 6630 mutex_lock(&flush_backlogs_mutex); 6631 ptr = flush_backlogs_fallback; 6632 } 6633 cpumask_clear(&ptr->flush_cpus); 6634 6635 cpus_read_lock(); 6636 6637 for_each_online_cpu(cpu) { 6638 if (flush_required(cpu)) { 6639 INIT_WORK(&ptr->w[cpu], flush_backlog); 6640 queue_work_on(cpu, system_highpri_wq, &ptr->w[cpu]); 6641 __cpumask_set_cpu(cpu, &ptr->flush_cpus); 6642 } 6643 } 6644 6645 /* we can have in flight packet[s] on the cpus we are not flushing, 6646 * synchronize_net() in unregister_netdevice_many() will take care of 6647 * them. 6648 */ 6649 for_each_cpu(cpu, &ptr->flush_cpus) 6650 flush_work(&ptr->w[cpu]); 6651 6652 cpus_read_unlock(); 6653 6654 if (ptr != flush_backlogs_fallback) 6655 kfree(ptr); 6656 else 6657 mutex_unlock(&flush_backlogs_mutex); 6658 } 6659 6660 static void net_rps_send_ipi(struct softnet_data *remsd) 6661 { 6662 #ifdef CONFIG_RPS 6663 while (remsd) { 6664 struct softnet_data *next = remsd->rps_ipi_next; 6665 6666 if (cpu_online(remsd->cpu)) 6667 smp_call_function_single_async(remsd->cpu, &remsd->csd); 6668 remsd = next; 6669 } 6670 #endif 6671 } 6672 6673 /* 6674 * net_rps_action_and_irq_enable sends any pending IPI's for rps. 6675 * Note: called with local irq disabled, but exits with local irq enabled. 6676 */ 6677 static void net_rps_action_and_irq_enable(struct softnet_data *sd) 6678 { 6679 #ifdef CONFIG_RPS 6680 struct softnet_data *remsd = sd->rps_ipi_list; 6681 6682 if (!use_backlog_threads() && remsd) { 6683 sd->rps_ipi_list = NULL; 6684 6685 local_irq_enable(); 6686 6687 /* Send pending IPI's to kick RPS processing on remote cpus. */ 6688 net_rps_send_ipi(remsd); 6689 } else 6690 #endif 6691 local_irq_enable(); 6692 } 6693 6694 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd) 6695 { 6696 #ifdef CONFIG_RPS 6697 return !use_backlog_threads() && sd->rps_ipi_list; 6698 #else 6699 return false; 6700 #endif 6701 } 6702 6703 static int process_backlog(struct napi_struct *napi, int quota) 6704 { 6705 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog); 6706 bool again = true; 6707 int work = 0; 6708 6709 /* Check if we have pending ipi, its better to send them now, 6710 * not waiting net_rx_action() end. 6711 */ 6712 if (sd_has_rps_ipi_waiting(sd)) { 6713 local_irq_disable(); 6714 net_rps_action_and_irq_enable(sd); 6715 } 6716 6717 napi->weight = READ_ONCE(net_hotdata.dev_rx_weight); 6718 while (again) { 6719 struct sk_buff *skb; 6720 6721 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6722 while ((skb = __skb_dequeue(&sd->process_queue))) { 6723 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6724 rcu_read_lock(); 6725 __netif_receive_skb(skb); 6726 rcu_read_unlock(); 6727 if (++work >= quota) { 6728 rps_input_queue_head_add(sd, work); 6729 return work; 6730 } 6731 6732 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6733 } 6734 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6735 6736 backlog_lock_irq_disable(sd); 6737 if (skb_queue_empty(&sd->input_pkt_queue)) { 6738 /* 6739 * Inline a custom version of __napi_complete(). 6740 * only current cpu owns and manipulates this napi, 6741 * and NAPI_STATE_SCHED is the only possible flag set 6742 * on backlog. 6743 * We can use a plain write instead of clear_bit(), 6744 * and we dont need an smp_mb() memory barrier. 6745 */ 6746 napi->state &= NAPIF_STATE_THREADED; 6747 again = false; 6748 } else { 6749 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6750 skb_queue_splice_tail_init(&sd->input_pkt_queue, 6751 &sd->process_queue); 6752 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6753 } 6754 backlog_unlock_irq_enable(sd); 6755 } 6756 6757 if (work) 6758 rps_input_queue_head_add(sd, work); 6759 return work; 6760 } 6761 6762 /** 6763 * __napi_schedule - schedule for receive 6764 * @n: entry to schedule 6765 * 6766 * The entry's receive function will be scheduled to run. 6767 * Consider using __napi_schedule_irqoff() if hard irqs are masked. 6768 */ 6769 void __napi_schedule(struct napi_struct *n) 6770 { 6771 unsigned long flags; 6772 6773 local_irq_save(flags); 6774 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6775 local_irq_restore(flags); 6776 } 6777 EXPORT_SYMBOL(__napi_schedule); 6778 6779 /** 6780 * napi_schedule_prep - check if napi can be scheduled 6781 * @n: napi context 6782 * 6783 * Test if NAPI routine is already running, and if not mark 6784 * it as running. This is used as a condition variable to 6785 * insure only one NAPI poll instance runs. We also make 6786 * sure there is no pending NAPI disable. 6787 */ 6788 bool napi_schedule_prep(struct napi_struct *n) 6789 { 6790 unsigned long new, val = READ_ONCE(n->state); 6791 6792 do { 6793 if (unlikely(val & NAPIF_STATE_DISABLE)) 6794 return false; 6795 new = val | NAPIF_STATE_SCHED; 6796 6797 /* Sets STATE_MISSED bit if STATE_SCHED was already set 6798 * This was suggested by Alexander Duyck, as compiler 6799 * emits better code than : 6800 * if (val & NAPIF_STATE_SCHED) 6801 * new |= NAPIF_STATE_MISSED; 6802 */ 6803 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED * 6804 NAPIF_STATE_MISSED; 6805 } while (!try_cmpxchg(&n->state, &val, new)); 6806 6807 return !(val & NAPIF_STATE_SCHED); 6808 } 6809 EXPORT_SYMBOL(napi_schedule_prep); 6810 6811 /** 6812 * __napi_schedule_irqoff - schedule for receive 6813 * @n: entry to schedule 6814 * 6815 * Variant of __napi_schedule() assuming hard irqs are masked. 6816 * 6817 * On PREEMPT_RT enabled kernels this maps to __napi_schedule() 6818 * because the interrupt disabled assumption might not be true 6819 * due to force-threaded interrupts and spinlock substitution. 6820 */ 6821 void __napi_schedule_irqoff(struct napi_struct *n) 6822 { 6823 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6824 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6825 else 6826 __napi_schedule(n); 6827 } 6828 EXPORT_SYMBOL(__napi_schedule_irqoff); 6829 6830 bool napi_complete_done(struct napi_struct *n, int work_done) 6831 { 6832 unsigned long flags, val, new, timeout = 0; 6833 bool ret = true; 6834 6835 /* 6836 * 1) Don't let napi dequeue from the cpu poll list 6837 * just in case its running on a different cpu. 6838 * 2) If we are busy polling, do nothing here, we have 6839 * the guarantee we will be called later. 6840 */ 6841 if (unlikely(n->state & (NAPIF_STATE_NPSVC | 6842 NAPIF_STATE_IN_BUSY_POLL))) 6843 return false; 6844 6845 if (work_done) { 6846 if (n->gro.bitmask) 6847 timeout = napi_get_gro_flush_timeout(n); 6848 n->defer_hard_irqs_count = napi_get_defer_hard_irqs(n); 6849 } 6850 if (n->defer_hard_irqs_count > 0) { 6851 n->defer_hard_irqs_count--; 6852 timeout = napi_get_gro_flush_timeout(n); 6853 if (timeout) 6854 ret = false; 6855 } 6856 6857 /* 6858 * When the NAPI instance uses a timeout and keeps postponing 6859 * it, we need to bound somehow the time packets are kept in 6860 * the GRO layer. 6861 */ 6862 gro_flush_normal(&n->gro, !!timeout); 6863 6864 if (unlikely(!list_empty(&n->poll_list))) { 6865 /* If n->poll_list is not empty, we need to mask irqs */ 6866 local_irq_save(flags); 6867 list_del_init(&n->poll_list); 6868 local_irq_restore(flags); 6869 } 6870 WRITE_ONCE(n->list_owner, -1); 6871 6872 val = READ_ONCE(n->state); 6873 do { 6874 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED)); 6875 6876 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED | 6877 NAPIF_STATE_SCHED_THREADED | 6878 NAPIF_STATE_PREFER_BUSY_POLL); 6879 6880 /* If STATE_MISSED was set, leave STATE_SCHED set, 6881 * because we will call napi->poll() one more time. 6882 * This C code was suggested by Alexander Duyck to help gcc. 6883 */ 6884 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED * 6885 NAPIF_STATE_SCHED; 6886 } while (!try_cmpxchg(&n->state, &val, new)); 6887 6888 if (unlikely(val & NAPIF_STATE_MISSED)) { 6889 __napi_schedule(n); 6890 return false; 6891 } 6892 6893 if (timeout) 6894 hrtimer_start(&n->timer, ns_to_ktime(timeout), 6895 HRTIMER_MODE_REL_PINNED); 6896 return ret; 6897 } 6898 EXPORT_SYMBOL(napi_complete_done); 6899 6900 static void skb_defer_free_flush(void) 6901 { 6902 struct llist_node *free_list; 6903 struct sk_buff *skb, *next; 6904 struct skb_defer_node *sdn; 6905 int node; 6906 6907 for_each_node(node) { 6908 sdn = this_cpu_ptr(net_hotdata.skb_defer_nodes) + node; 6909 6910 if (llist_empty(&sdn->defer_list)) 6911 continue; 6912 atomic_long_set(&sdn->defer_count, 0); 6913 free_list = llist_del_all(&sdn->defer_list); 6914 6915 llist_for_each_entry_safe(skb, next, free_list, ll_node) { 6916 prefetch(next); 6917 napi_consume_skb(skb, 1); 6918 } 6919 } 6920 } 6921 6922 #if defined(CONFIG_NET_RX_BUSY_POLL) 6923 6924 enum { 6925 NAPI_F_PREFER_BUSY_POLL = 1, 6926 NAPI_F_END_ON_RESCHED = 2, 6927 }; 6928 6929 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, 6930 unsigned flags, u16 budget) 6931 { 6932 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6933 unsigned long timeout = 0; 6934 int rc; 6935 6936 /* Busy polling means there is a high chance device driver hard irq 6937 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was 6938 * set in napi_schedule_prep(). 6939 * Since we either call napi->poll() once more or start the timer, 6940 * we can safely clear NAPI_STATE_MISSED. 6941 * 6942 * Note: x86 could use a single "lock and ..." instruction 6943 * to perform these two clear_bit() 6944 */ 6945 clear_bit(NAPI_STATE_MISSED, &napi->state); 6946 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state); 6947 6948 local_bh_disable(); 6949 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6950 6951 if (flags & NAPI_F_PREFER_BUSY_POLL) { 6952 napi->defer_hard_irqs_count = napi_get_defer_hard_irqs(napi); 6953 if (napi->defer_hard_irqs_count) 6954 timeout = napi_get_gro_flush_timeout(napi); 6955 } 6956 if (timeout) { 6957 netpoll_poll_unlock(have_poll_lock); 6958 6959 /* Drain aged GRO packets before clearing SCHED since the NAPI 6960 * won't run again until after the timer fires. When HZ < 1000, 6961 * GRO age comparison is too coarse, so flush everything. 6962 */ 6963 gro_flush_normal(&napi->gro, HZ >= 1000); 6964 6965 clear_bit(NAPI_STATE_SCHED, &napi->state); 6966 hrtimer_start(&napi->timer, ns_to_ktime(timeout), 6967 HRTIMER_MODE_REL_PINNED); 6968 } else { 6969 /* Use driver poll to re-enable device interrupts. */ 6970 rc = napi->poll(napi, budget); 6971 /* Unless rc == budget we no longer own the NAPI instance, 6972 * IRQ may fire on another CPU, poll this NAPI, and enter GRO. 6973 */ 6974 trace_napi_poll(napi, rc, budget); 6975 netpoll_poll_unlock(have_poll_lock); 6976 if (rc == budget) { 6977 gro_normal_list(&napi->gro); 6978 __napi_schedule(napi); 6979 } 6980 } 6981 6982 bpf_net_ctx_clear(bpf_net_ctx); 6983 local_bh_enable(); 6984 } 6985 6986 static void __napi_busy_loop(unsigned int napi_id, 6987 bool (*loop_end)(void *, unsigned long), 6988 void *loop_end_arg, unsigned flags, u16 budget) 6989 { 6990 unsigned long start_time = loop_end ? busy_loop_current_time() : 0; 6991 int (*napi_poll)(struct napi_struct *napi, int budget); 6992 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6993 void *have_poll_lock = NULL; 6994 struct napi_struct *napi; 6995 6996 WARN_ON_ONCE(!rcu_read_lock_held()); 6997 6998 restart: 6999 napi_poll = NULL; 7000 7001 napi = napi_by_id(napi_id); 7002 if (!napi) 7003 return; 7004 7005 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 7006 preempt_disable(); 7007 for (;;) { 7008 int work = 0; 7009 7010 local_bh_disable(); 7011 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7012 if (!napi_poll) { 7013 unsigned long val = READ_ONCE(napi->state); 7014 7015 /* If multiple threads are competing for this napi, 7016 * we avoid dirtying napi->state as much as we can. 7017 */ 7018 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED | 7019 NAPIF_STATE_IN_BUSY_POLL)) { 7020 if (flags & NAPI_F_PREFER_BUSY_POLL) 7021 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 7022 goto count; 7023 } 7024 if (cmpxchg(&napi->state, val, 7025 val | NAPIF_STATE_IN_BUSY_POLL | 7026 NAPIF_STATE_SCHED) != val) { 7027 if (flags & NAPI_F_PREFER_BUSY_POLL) 7028 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 7029 goto count; 7030 } 7031 have_poll_lock = netpoll_poll_lock(napi); 7032 napi_poll = napi->poll; 7033 } 7034 work = napi_poll(napi, budget); 7035 trace_napi_poll(napi, work, budget); 7036 gro_normal_list(&napi->gro); 7037 count: 7038 if (work > 0) 7039 __NET_ADD_STATS(dev_net(napi->dev), 7040 LINUX_MIB_BUSYPOLLRXPACKETS, work); 7041 skb_defer_free_flush(); 7042 bpf_net_ctx_clear(bpf_net_ctx); 7043 local_bh_enable(); 7044 7045 if (!loop_end || loop_end(loop_end_arg, start_time)) 7046 break; 7047 7048 if (unlikely(need_resched())) { 7049 if (flags & NAPI_F_END_ON_RESCHED) 7050 break; 7051 if (napi_poll) 7052 busy_poll_stop(napi, have_poll_lock, flags, budget); 7053 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 7054 preempt_enable(); 7055 rcu_read_unlock(); 7056 cond_resched(); 7057 rcu_read_lock(); 7058 if (loop_end(loop_end_arg, start_time)) 7059 return; 7060 goto restart; 7061 } 7062 cpu_relax(); 7063 } 7064 if (napi_poll) 7065 busy_poll_stop(napi, have_poll_lock, flags, budget); 7066 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 7067 preempt_enable(); 7068 } 7069 7070 void napi_busy_loop_rcu(unsigned int napi_id, 7071 bool (*loop_end)(void *, unsigned long), 7072 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 7073 { 7074 unsigned flags = NAPI_F_END_ON_RESCHED; 7075 7076 if (prefer_busy_poll) 7077 flags |= NAPI_F_PREFER_BUSY_POLL; 7078 7079 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 7080 } 7081 7082 void napi_busy_loop(unsigned int napi_id, 7083 bool (*loop_end)(void *, unsigned long), 7084 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 7085 { 7086 unsigned flags = prefer_busy_poll ? NAPI_F_PREFER_BUSY_POLL : 0; 7087 7088 rcu_read_lock(); 7089 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 7090 rcu_read_unlock(); 7091 } 7092 EXPORT_SYMBOL(napi_busy_loop); 7093 7094 void napi_suspend_irqs(unsigned int napi_id) 7095 { 7096 struct napi_struct *napi; 7097 7098 rcu_read_lock(); 7099 napi = napi_by_id(napi_id); 7100 if (napi) { 7101 unsigned long timeout = napi_get_irq_suspend_timeout(napi); 7102 7103 if (timeout) 7104 hrtimer_start(&napi->timer, ns_to_ktime(timeout), 7105 HRTIMER_MODE_REL_PINNED); 7106 } 7107 rcu_read_unlock(); 7108 } 7109 7110 void napi_resume_irqs(unsigned int napi_id) 7111 { 7112 struct napi_struct *napi; 7113 7114 rcu_read_lock(); 7115 napi = napi_by_id(napi_id); 7116 if (napi) { 7117 /* If irq_suspend_timeout is set to 0 between the call to 7118 * napi_suspend_irqs and now, the original value still 7119 * determines the safety timeout as intended and napi_watchdog 7120 * will resume irq processing. 7121 */ 7122 if (napi_get_irq_suspend_timeout(napi)) { 7123 local_bh_disable(); 7124 napi_schedule(napi); 7125 local_bh_enable(); 7126 } 7127 } 7128 rcu_read_unlock(); 7129 } 7130 7131 #endif /* CONFIG_NET_RX_BUSY_POLL */ 7132 7133 static void __napi_hash_add_with_id(struct napi_struct *napi, 7134 unsigned int napi_id) 7135 { 7136 napi->gro.cached_napi_id = napi_id; 7137 7138 WRITE_ONCE(napi->napi_id, napi_id); 7139 hlist_add_head_rcu(&napi->napi_hash_node, 7140 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]); 7141 } 7142 7143 static void napi_hash_add_with_id(struct napi_struct *napi, 7144 unsigned int napi_id) 7145 { 7146 unsigned long flags; 7147 7148 spin_lock_irqsave(&napi_hash_lock, flags); 7149 WARN_ON_ONCE(napi_by_id(napi_id)); 7150 __napi_hash_add_with_id(napi, napi_id); 7151 spin_unlock_irqrestore(&napi_hash_lock, flags); 7152 } 7153 7154 static void napi_hash_add(struct napi_struct *napi) 7155 { 7156 unsigned long flags; 7157 7158 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state)) 7159 return; 7160 7161 spin_lock_irqsave(&napi_hash_lock, flags); 7162 7163 /* 0..NR_CPUS range is reserved for sender_cpu use */ 7164 do { 7165 if (unlikely(!napi_id_valid(++napi_gen_id))) 7166 napi_gen_id = MIN_NAPI_ID; 7167 } while (napi_by_id(napi_gen_id)); 7168 7169 __napi_hash_add_with_id(napi, napi_gen_id); 7170 7171 spin_unlock_irqrestore(&napi_hash_lock, flags); 7172 } 7173 7174 /* Warning : caller is responsible to make sure rcu grace period 7175 * is respected before freeing memory containing @napi 7176 */ 7177 static void napi_hash_del(struct napi_struct *napi) 7178 { 7179 unsigned long flags; 7180 7181 spin_lock_irqsave(&napi_hash_lock, flags); 7182 7183 hlist_del_init_rcu(&napi->napi_hash_node); 7184 7185 spin_unlock_irqrestore(&napi_hash_lock, flags); 7186 } 7187 7188 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer) 7189 { 7190 struct napi_struct *napi; 7191 7192 napi = container_of(timer, struct napi_struct, timer); 7193 7194 /* Note : we use a relaxed variant of napi_schedule_prep() not setting 7195 * NAPI_STATE_MISSED, since we do not react to a device IRQ. 7196 */ 7197 if (!napi_disable_pending(napi) && 7198 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) { 7199 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 7200 __napi_schedule_irqoff(napi); 7201 } 7202 7203 return HRTIMER_NORESTART; 7204 } 7205 7206 static void napi_stop_kthread(struct napi_struct *napi) 7207 { 7208 unsigned long val, new; 7209 7210 /* Wait until the napi STATE_THREADED is unset. */ 7211 while (true) { 7212 val = READ_ONCE(napi->state); 7213 7214 /* If napi kthread own this napi or the napi is idle, 7215 * STATE_THREADED can be unset here. 7216 */ 7217 if ((val & NAPIF_STATE_SCHED_THREADED) || 7218 !(val & NAPIF_STATE_SCHED)) { 7219 new = val & (~(NAPIF_STATE_THREADED | 7220 NAPIF_STATE_THREADED_BUSY_POLL)); 7221 } else { 7222 msleep(20); 7223 continue; 7224 } 7225 7226 if (try_cmpxchg(&napi->state, &val, new)) 7227 break; 7228 } 7229 7230 /* Once STATE_THREADED is unset, wait for SCHED_THREADED to be unset by 7231 * the kthread. 7232 */ 7233 while (true) { 7234 if (!test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) 7235 break; 7236 7237 msleep(20); 7238 } 7239 7240 kthread_stop(napi->thread); 7241 napi->thread = NULL; 7242 } 7243 7244 static void napi_set_threaded_state(struct napi_struct *napi, 7245 enum netdev_napi_threaded threaded_mode) 7246 { 7247 bool threaded = threaded_mode != NETDEV_NAPI_THREADED_DISABLED; 7248 bool busy_poll = threaded_mode == NETDEV_NAPI_THREADED_BUSY_POLL; 7249 7250 assign_bit(NAPI_STATE_THREADED, &napi->state, threaded); 7251 assign_bit(NAPI_STATE_THREADED_BUSY_POLL, &napi->state, busy_poll); 7252 } 7253 7254 int napi_set_threaded(struct napi_struct *napi, 7255 enum netdev_napi_threaded threaded) 7256 { 7257 if (threaded) { 7258 if (!napi->thread) { 7259 int err = napi_kthread_create(napi); 7260 7261 if (err) 7262 return err; 7263 } 7264 } 7265 7266 if (napi->config) 7267 napi->config->threaded = threaded; 7268 7269 /* Setting/unsetting threaded mode on a napi might not immediately 7270 * take effect, if the current napi instance is actively being 7271 * polled. In this case, the switch between threaded mode and 7272 * softirq mode will happen in the next round of napi_schedule(). 7273 * This should not cause hiccups/stalls to the live traffic. 7274 */ 7275 if (!threaded && napi->thread) { 7276 napi_stop_kthread(napi); 7277 } else { 7278 /* Make sure kthread is created before THREADED bit is set. */ 7279 smp_mb__before_atomic(); 7280 napi_set_threaded_state(napi, threaded); 7281 } 7282 7283 return 0; 7284 } 7285 7286 int netif_set_threaded(struct net_device *dev, 7287 enum netdev_napi_threaded threaded) 7288 { 7289 struct napi_struct *napi; 7290 int i, err = 0; 7291 7292 netdev_assert_locked_or_invisible(dev); 7293 7294 if (threaded) { 7295 list_for_each_entry(napi, &dev->napi_list, dev_list) { 7296 if (!napi->thread) { 7297 err = napi_kthread_create(napi); 7298 if (err) { 7299 threaded = NETDEV_NAPI_THREADED_DISABLED; 7300 break; 7301 } 7302 } 7303 } 7304 } 7305 7306 WRITE_ONCE(dev->threaded, threaded); 7307 7308 /* The error should not occur as the kthreads are already created. */ 7309 list_for_each_entry(napi, &dev->napi_list, dev_list) 7310 WARN_ON_ONCE(napi_set_threaded(napi, threaded)); 7311 7312 /* Override the config for all NAPIs even if currently not listed */ 7313 for (i = 0; i < dev->num_napi_configs; i++) 7314 dev->napi_config[i].threaded = threaded; 7315 7316 return err; 7317 } 7318 7319 /** 7320 * netif_threaded_enable() - enable threaded NAPIs 7321 * @dev: net_device instance 7322 * 7323 * Enable threaded mode for the NAPI instances of the device. This may be useful 7324 * for devices where multiple NAPI instances get scheduled by a single 7325 * interrupt. Threaded NAPI allows moving the NAPI processing to cores other 7326 * than the core where IRQ is mapped. 7327 * 7328 * This function should be called before @dev is registered. 7329 */ 7330 void netif_threaded_enable(struct net_device *dev) 7331 { 7332 WARN_ON_ONCE(netif_set_threaded(dev, NETDEV_NAPI_THREADED_ENABLED)); 7333 } 7334 EXPORT_SYMBOL(netif_threaded_enable); 7335 7336 /** 7337 * netif_queue_set_napi - Associate queue with the napi 7338 * @dev: device to which NAPI and queue belong 7339 * @queue_index: Index of queue 7340 * @type: queue type as RX or TX 7341 * @napi: NAPI context, pass NULL to clear previously set NAPI 7342 * 7343 * Set queue with its corresponding napi context. This should be done after 7344 * registering the NAPI handler for the queue-vector and the queues have been 7345 * mapped to the corresponding interrupt vector. 7346 */ 7347 void netif_queue_set_napi(struct net_device *dev, unsigned int queue_index, 7348 enum netdev_queue_type type, struct napi_struct *napi) 7349 { 7350 struct netdev_rx_queue *rxq; 7351 struct netdev_queue *txq; 7352 7353 if (WARN_ON_ONCE(napi && !napi->dev)) 7354 return; 7355 netdev_assert_locked_ops_compat_or_invisible(dev); 7356 7357 switch (type) { 7358 case NETDEV_QUEUE_TYPE_RX: 7359 rxq = __netif_get_rx_queue(dev, queue_index); 7360 rxq->napi = napi; 7361 return; 7362 case NETDEV_QUEUE_TYPE_TX: 7363 txq = netdev_get_tx_queue(dev, queue_index); 7364 txq->napi = napi; 7365 return; 7366 default: 7367 return; 7368 } 7369 } 7370 EXPORT_SYMBOL(netif_queue_set_napi); 7371 7372 static void 7373 netif_napi_irq_notify(struct irq_affinity_notify *notify, 7374 const cpumask_t *mask) 7375 { 7376 struct napi_struct *napi = 7377 container_of(notify, struct napi_struct, notify); 7378 #ifdef CONFIG_RFS_ACCEL 7379 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 7380 int err; 7381 #endif 7382 7383 if (napi->config && napi->dev->irq_affinity_auto) 7384 cpumask_copy(&napi->config->affinity_mask, mask); 7385 7386 #ifdef CONFIG_RFS_ACCEL 7387 if (napi->dev->rx_cpu_rmap_auto) { 7388 err = cpu_rmap_update(rmap, napi->napi_rmap_idx, mask); 7389 if (err) 7390 netdev_warn(napi->dev, "RMAP update failed (%d)\n", 7391 err); 7392 } 7393 #endif 7394 } 7395 7396 #ifdef CONFIG_RFS_ACCEL 7397 static void netif_napi_affinity_release(struct kref *ref) 7398 { 7399 struct napi_struct *napi = 7400 container_of(ref, struct napi_struct, notify.kref); 7401 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 7402 7403 netdev_assert_locked(napi->dev); 7404 WARN_ON(test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, 7405 &napi->state)); 7406 7407 if (!napi->dev->rx_cpu_rmap_auto) 7408 return; 7409 rmap->obj[napi->napi_rmap_idx] = NULL; 7410 napi->napi_rmap_idx = -1; 7411 cpu_rmap_put(rmap); 7412 } 7413 7414 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 7415 { 7416 if (dev->rx_cpu_rmap_auto) 7417 return 0; 7418 7419 dev->rx_cpu_rmap = alloc_irq_cpu_rmap(num_irqs); 7420 if (!dev->rx_cpu_rmap) 7421 return -ENOMEM; 7422 7423 dev->rx_cpu_rmap_auto = true; 7424 return 0; 7425 } 7426 EXPORT_SYMBOL(netif_enable_cpu_rmap); 7427 7428 static void netif_del_cpu_rmap(struct net_device *dev) 7429 { 7430 struct cpu_rmap *rmap = dev->rx_cpu_rmap; 7431 7432 if (!dev->rx_cpu_rmap_auto) 7433 return; 7434 7435 /* Free the rmap */ 7436 cpu_rmap_put(rmap); 7437 dev->rx_cpu_rmap = NULL; 7438 dev->rx_cpu_rmap_auto = false; 7439 } 7440 7441 #else 7442 static void netif_napi_affinity_release(struct kref *ref) 7443 { 7444 } 7445 7446 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 7447 { 7448 return 0; 7449 } 7450 EXPORT_SYMBOL(netif_enable_cpu_rmap); 7451 7452 static void netif_del_cpu_rmap(struct net_device *dev) 7453 { 7454 } 7455 #endif 7456 7457 void netif_set_affinity_auto(struct net_device *dev) 7458 { 7459 unsigned int i, maxqs, numa; 7460 7461 maxqs = max(dev->num_tx_queues, dev->num_rx_queues); 7462 numa = dev_to_node(&dev->dev); 7463 7464 for (i = 0; i < maxqs; i++) 7465 cpumask_set_cpu(cpumask_local_spread(i, numa), 7466 &dev->napi_config[i].affinity_mask); 7467 7468 dev->irq_affinity_auto = true; 7469 } 7470 EXPORT_SYMBOL(netif_set_affinity_auto); 7471 7472 void netif_napi_set_irq_locked(struct napi_struct *napi, int irq) 7473 { 7474 int rc; 7475 7476 netdev_assert_locked_or_invisible(napi->dev); 7477 7478 if (napi->irq == irq) 7479 return; 7480 7481 /* Remove existing resources */ 7482 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7483 irq_set_affinity_notifier(napi->irq, NULL); 7484 7485 napi->irq = irq; 7486 if (irq < 0 || 7487 (!napi->dev->rx_cpu_rmap_auto && !napi->dev->irq_affinity_auto)) 7488 return; 7489 7490 /* Abort for buggy drivers */ 7491 if (napi->dev->irq_affinity_auto && WARN_ON_ONCE(!napi->config)) 7492 return; 7493 7494 #ifdef CONFIG_RFS_ACCEL 7495 if (napi->dev->rx_cpu_rmap_auto) { 7496 rc = cpu_rmap_add(napi->dev->rx_cpu_rmap, napi); 7497 if (rc < 0) 7498 return; 7499 7500 cpu_rmap_get(napi->dev->rx_cpu_rmap); 7501 napi->napi_rmap_idx = rc; 7502 } 7503 #endif 7504 7505 /* Use core IRQ notifier */ 7506 napi->notify.notify = netif_napi_irq_notify; 7507 napi->notify.release = netif_napi_affinity_release; 7508 rc = irq_set_affinity_notifier(irq, &napi->notify); 7509 if (rc) { 7510 netdev_warn(napi->dev, "Unable to set IRQ notifier (%d)\n", 7511 rc); 7512 goto put_rmap; 7513 } 7514 7515 set_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state); 7516 return; 7517 7518 put_rmap: 7519 #ifdef CONFIG_RFS_ACCEL 7520 if (napi->dev->rx_cpu_rmap_auto) { 7521 napi->dev->rx_cpu_rmap->obj[napi->napi_rmap_idx] = NULL; 7522 cpu_rmap_put(napi->dev->rx_cpu_rmap); 7523 napi->napi_rmap_idx = -1; 7524 } 7525 #endif 7526 napi->notify.notify = NULL; 7527 napi->notify.release = NULL; 7528 } 7529 EXPORT_SYMBOL(netif_napi_set_irq_locked); 7530 7531 static void napi_restore_config(struct napi_struct *n) 7532 { 7533 n->defer_hard_irqs = n->config->defer_hard_irqs; 7534 n->gro_flush_timeout = n->config->gro_flush_timeout; 7535 n->irq_suspend_timeout = n->config->irq_suspend_timeout; 7536 7537 if (n->dev->irq_affinity_auto && 7538 test_bit(NAPI_STATE_HAS_NOTIFIER, &n->state)) 7539 irq_set_affinity(n->irq, &n->config->affinity_mask); 7540 7541 /* a NAPI ID might be stored in the config, if so use it. if not, use 7542 * napi_hash_add to generate one for us. 7543 */ 7544 if (n->config->napi_id) { 7545 napi_hash_add_with_id(n, n->config->napi_id); 7546 } else { 7547 napi_hash_add(n); 7548 n->config->napi_id = n->napi_id; 7549 } 7550 7551 WARN_ON_ONCE(napi_set_threaded(n, n->config->threaded)); 7552 } 7553 7554 static void napi_save_config(struct napi_struct *n) 7555 { 7556 n->config->defer_hard_irqs = n->defer_hard_irqs; 7557 n->config->gro_flush_timeout = n->gro_flush_timeout; 7558 n->config->irq_suspend_timeout = n->irq_suspend_timeout; 7559 napi_hash_del(n); 7560 } 7561 7562 /* Netlink wants the NAPI list to be sorted by ID, if adding a NAPI which will 7563 * inherit an existing ID try to insert it at the right position. 7564 */ 7565 static void 7566 netif_napi_dev_list_add(struct net_device *dev, struct napi_struct *napi) 7567 { 7568 unsigned int new_id, pos_id; 7569 struct list_head *higher; 7570 struct napi_struct *pos; 7571 7572 new_id = UINT_MAX; 7573 if (napi->config && napi->config->napi_id) 7574 new_id = napi->config->napi_id; 7575 7576 higher = &dev->napi_list; 7577 list_for_each_entry(pos, &dev->napi_list, dev_list) { 7578 if (napi_id_valid(pos->napi_id)) 7579 pos_id = pos->napi_id; 7580 else if (pos->config) 7581 pos_id = pos->config->napi_id; 7582 else 7583 pos_id = UINT_MAX; 7584 7585 if (pos_id <= new_id) 7586 break; 7587 higher = &pos->dev_list; 7588 } 7589 list_add_rcu(&napi->dev_list, higher); /* adds after higher */ 7590 } 7591 7592 /* Double check that napi_get_frags() allocates skbs with 7593 * skb->head being backed by slab, not a page fragment. 7594 * This is to make sure bug fixed in 3226b158e67c 7595 * ("net: avoid 32 x truesize under-estimation for tiny skbs") 7596 * does not accidentally come back. 7597 */ 7598 static void napi_get_frags_check(struct napi_struct *napi) 7599 { 7600 struct sk_buff *skb; 7601 7602 local_bh_disable(); 7603 skb = napi_get_frags(napi); 7604 WARN_ON_ONCE(skb && skb->head_frag); 7605 napi_free_frags(napi); 7606 local_bh_enable(); 7607 } 7608 7609 void netif_napi_add_weight_locked(struct net_device *dev, 7610 struct napi_struct *napi, 7611 int (*poll)(struct napi_struct *, int), 7612 int weight) 7613 { 7614 netdev_assert_locked(dev); 7615 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state))) 7616 return; 7617 7618 INIT_LIST_HEAD(&napi->poll_list); 7619 INIT_HLIST_NODE(&napi->napi_hash_node); 7620 hrtimer_setup(&napi->timer, napi_watchdog, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); 7621 gro_init(&napi->gro); 7622 napi->skb = NULL; 7623 napi->poll = poll; 7624 if (weight > NAPI_POLL_WEIGHT) 7625 netdev_err_once(dev, "%s() called with weight %d\n", __func__, 7626 weight); 7627 napi->weight = weight; 7628 napi->dev = dev; 7629 #ifdef CONFIG_NETPOLL 7630 napi->poll_owner = -1; 7631 #endif 7632 napi->list_owner = -1; 7633 set_bit(NAPI_STATE_SCHED, &napi->state); 7634 set_bit(NAPI_STATE_NPSVC, &napi->state); 7635 netif_napi_dev_list_add(dev, napi); 7636 7637 /* default settings from sysfs are applied to all NAPIs. any per-NAPI 7638 * configuration will be loaded in napi_enable 7639 */ 7640 napi_set_defer_hard_irqs(napi, READ_ONCE(dev->napi_defer_hard_irqs)); 7641 napi_set_gro_flush_timeout(napi, READ_ONCE(dev->gro_flush_timeout)); 7642 7643 napi_get_frags_check(napi); 7644 /* Create kthread for this napi if dev->threaded is set. 7645 * Clear dev->threaded if kthread creation failed so that 7646 * threaded mode will not be enabled in napi_enable(). 7647 */ 7648 if (napi_get_threaded_config(dev, napi)) 7649 if (napi_kthread_create(napi)) 7650 dev->threaded = NETDEV_NAPI_THREADED_DISABLED; 7651 netif_napi_set_irq_locked(napi, -1); 7652 } 7653 EXPORT_SYMBOL(netif_napi_add_weight_locked); 7654 7655 void napi_disable_locked(struct napi_struct *n) 7656 { 7657 unsigned long val, new; 7658 7659 might_sleep(); 7660 netdev_assert_locked(n->dev); 7661 7662 set_bit(NAPI_STATE_DISABLE, &n->state); 7663 7664 val = READ_ONCE(n->state); 7665 do { 7666 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) { 7667 usleep_range(20, 200); 7668 val = READ_ONCE(n->state); 7669 } 7670 7671 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC; 7672 new &= ~(NAPIF_STATE_THREADED | 7673 NAPIF_STATE_THREADED_BUSY_POLL | 7674 NAPIF_STATE_PREFER_BUSY_POLL); 7675 } while (!try_cmpxchg(&n->state, &val, new)); 7676 7677 hrtimer_cancel(&n->timer); 7678 7679 if (n->config) 7680 napi_save_config(n); 7681 else 7682 napi_hash_del(n); 7683 7684 clear_bit(NAPI_STATE_DISABLE, &n->state); 7685 } 7686 EXPORT_SYMBOL(napi_disable_locked); 7687 7688 /** 7689 * napi_disable() - prevent NAPI from scheduling 7690 * @n: NAPI context 7691 * 7692 * Stop NAPI from being scheduled on this context. 7693 * Waits till any outstanding processing completes. 7694 * Takes netdev_lock() for associated net_device. 7695 */ 7696 void napi_disable(struct napi_struct *n) 7697 { 7698 netdev_lock(n->dev); 7699 napi_disable_locked(n); 7700 netdev_unlock(n->dev); 7701 } 7702 EXPORT_SYMBOL(napi_disable); 7703 7704 void napi_enable_locked(struct napi_struct *n) 7705 { 7706 unsigned long new, val = READ_ONCE(n->state); 7707 7708 if (n->config) 7709 napi_restore_config(n); 7710 else 7711 napi_hash_add(n); 7712 7713 do { 7714 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val)); 7715 7716 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC); 7717 if (n->dev->threaded && n->thread) 7718 new |= NAPIF_STATE_THREADED; 7719 } while (!try_cmpxchg(&n->state, &val, new)); 7720 } 7721 EXPORT_SYMBOL(napi_enable_locked); 7722 7723 /** 7724 * napi_enable() - enable NAPI scheduling 7725 * @n: NAPI context 7726 * 7727 * Enable scheduling of a NAPI instance. 7728 * Must be paired with napi_disable(). 7729 * Takes netdev_lock() for associated net_device. 7730 */ 7731 void napi_enable(struct napi_struct *n) 7732 { 7733 netdev_lock(n->dev); 7734 napi_enable_locked(n); 7735 netdev_unlock(n->dev); 7736 } 7737 EXPORT_SYMBOL(napi_enable); 7738 7739 /* Must be called in process context */ 7740 void __netif_napi_del_locked(struct napi_struct *napi) 7741 { 7742 netdev_assert_locked(napi->dev); 7743 7744 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state)) 7745 return; 7746 7747 /* Make sure NAPI is disabled (or was never enabled). */ 7748 WARN_ON(!test_bit(NAPI_STATE_SCHED, &napi->state)); 7749 7750 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7751 irq_set_affinity_notifier(napi->irq, NULL); 7752 7753 if (napi->config) { 7754 napi->index = -1; 7755 napi->config = NULL; 7756 } 7757 7758 list_del_rcu(&napi->dev_list); 7759 napi_free_frags(napi); 7760 7761 gro_cleanup(&napi->gro); 7762 7763 if (napi->thread) { 7764 kthread_stop(napi->thread); 7765 napi->thread = NULL; 7766 } 7767 } 7768 EXPORT_SYMBOL(__netif_napi_del_locked); 7769 7770 static int __napi_poll(struct napi_struct *n, bool *repoll) 7771 { 7772 int work, weight; 7773 7774 weight = n->weight; 7775 7776 /* This NAPI_STATE_SCHED test is for avoiding a race 7777 * with netpoll's poll_napi(). Only the entity which 7778 * obtains the lock and sees NAPI_STATE_SCHED set will 7779 * actually make the ->poll() call. Therefore we avoid 7780 * accidentally calling ->poll() when NAPI is not scheduled. 7781 */ 7782 work = 0; 7783 if (napi_is_scheduled(n)) { 7784 work = n->poll(n, weight); 7785 trace_napi_poll(n, work, weight); 7786 7787 xdp_do_check_flushed(n); 7788 } 7789 7790 if (unlikely(work > weight)) 7791 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n", 7792 n->poll, work, weight); 7793 7794 if (likely(work < weight)) 7795 return work; 7796 7797 /* Drivers must not modify the NAPI state if they 7798 * consume the entire weight. In such cases this code 7799 * still "owns" the NAPI instance and therefore can 7800 * move the instance around on the list at-will. 7801 */ 7802 if (unlikely(napi_disable_pending(n))) { 7803 napi_complete(n); 7804 return work; 7805 } 7806 7807 /* The NAPI context has more processing work, but busy-polling 7808 * is preferred. Exit early. 7809 */ 7810 if (napi_prefer_busy_poll(n)) { 7811 if (napi_complete_done(n, work)) { 7812 /* If timeout is not set, we need to make sure 7813 * that the NAPI is re-scheduled. 7814 */ 7815 napi_schedule(n); 7816 } 7817 return work; 7818 } 7819 7820 /* Flush too old packets. If HZ < 1000, flush all packets */ 7821 gro_flush_normal(&n->gro, HZ >= 1000); 7822 7823 /* Some drivers may have called napi_schedule 7824 * prior to exhausting their budget. 7825 */ 7826 if (unlikely(!list_empty(&n->poll_list))) { 7827 pr_warn_once("%s: Budget exhausted after napi rescheduled\n", 7828 n->dev ? n->dev->name : "backlog"); 7829 return work; 7830 } 7831 7832 *repoll = true; 7833 7834 return work; 7835 } 7836 7837 static int napi_poll(struct napi_struct *n, struct list_head *repoll) 7838 { 7839 bool do_repoll = false; 7840 void *have; 7841 int work; 7842 7843 list_del_init(&n->poll_list); 7844 7845 have = netpoll_poll_lock(n); 7846 7847 work = __napi_poll(n, &do_repoll); 7848 7849 if (do_repoll) { 7850 #if defined(CONFIG_DEBUG_NET) 7851 if (unlikely(!napi_is_scheduled(n))) 7852 pr_crit("repoll requested for device %s %ps but napi is not scheduled.\n", 7853 n->dev->name, n->poll); 7854 #endif 7855 list_add_tail(&n->poll_list, repoll); 7856 } 7857 netpoll_poll_unlock(have); 7858 7859 return work; 7860 } 7861 7862 static int napi_thread_wait(struct napi_struct *napi) 7863 { 7864 set_current_state(TASK_INTERRUPTIBLE); 7865 7866 while (!kthread_should_stop()) { 7867 /* Testing SCHED_THREADED bit here to make sure the current 7868 * kthread owns this napi and could poll on this napi. 7869 * Testing SCHED bit is not enough because SCHED bit might be 7870 * set by some other busy poll thread or by napi_disable(). 7871 */ 7872 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) { 7873 WARN_ON(!list_empty(&napi->poll_list)); 7874 __set_current_state(TASK_RUNNING); 7875 return 0; 7876 } 7877 7878 schedule(); 7879 set_current_state(TASK_INTERRUPTIBLE); 7880 } 7881 __set_current_state(TASK_RUNNING); 7882 7883 return -1; 7884 } 7885 7886 static void napi_threaded_poll_loop(struct napi_struct *napi, 7887 unsigned long *busy_poll_last_qs) 7888 { 7889 unsigned long last_qs = busy_poll_last_qs ? *busy_poll_last_qs : jiffies; 7890 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7891 struct softnet_data *sd; 7892 7893 for (;;) { 7894 bool repoll = false; 7895 void *have; 7896 7897 local_bh_disable(); 7898 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7899 7900 sd = this_cpu_ptr(&softnet_data); 7901 sd->in_napi_threaded_poll = true; 7902 7903 have = netpoll_poll_lock(napi); 7904 __napi_poll(napi, &repoll); 7905 netpoll_poll_unlock(have); 7906 7907 sd->in_napi_threaded_poll = false; 7908 barrier(); 7909 7910 if (sd_has_rps_ipi_waiting(sd)) { 7911 local_irq_disable(); 7912 net_rps_action_and_irq_enable(sd); 7913 } 7914 skb_defer_free_flush(); 7915 bpf_net_ctx_clear(bpf_net_ctx); 7916 7917 /* When busy poll is enabled, the old packets are not flushed in 7918 * napi_complete_done. So flush them here. 7919 */ 7920 if (busy_poll_last_qs) 7921 gro_flush_normal(&napi->gro, HZ >= 1000); 7922 local_bh_enable(); 7923 7924 /* Call cond_resched here to avoid watchdog warnings. */ 7925 if (repoll || busy_poll_last_qs) { 7926 rcu_softirq_qs_periodic(last_qs); 7927 cond_resched(); 7928 } 7929 7930 if (!repoll) 7931 break; 7932 } 7933 7934 if (busy_poll_last_qs) 7935 *busy_poll_last_qs = last_qs; 7936 } 7937 7938 static int napi_threaded_poll(void *data) 7939 { 7940 struct napi_struct *napi = data; 7941 unsigned long last_qs = jiffies; 7942 bool want_busy_poll; 7943 bool in_busy_poll; 7944 unsigned long val; 7945 7946 while (!napi_thread_wait(napi)) { 7947 val = READ_ONCE(napi->state); 7948 7949 want_busy_poll = val & NAPIF_STATE_THREADED_BUSY_POLL; 7950 in_busy_poll = val & NAPIF_STATE_IN_BUSY_POLL; 7951 7952 if (unlikely(val & NAPIF_STATE_DISABLE)) 7953 want_busy_poll = false; 7954 7955 if (want_busy_poll != in_busy_poll) 7956 assign_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state, 7957 want_busy_poll); 7958 7959 napi_threaded_poll_loop(napi, want_busy_poll ? &last_qs : NULL); 7960 } 7961 7962 return 0; 7963 } 7964 7965 static __latent_entropy void net_rx_action(void) 7966 { 7967 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 7968 unsigned long time_limit = jiffies + 7969 usecs_to_jiffies(READ_ONCE(net_hotdata.netdev_budget_usecs)); 7970 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7971 int budget = READ_ONCE(net_hotdata.netdev_budget); 7972 LIST_HEAD(list); 7973 LIST_HEAD(repoll); 7974 7975 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7976 start: 7977 sd->in_net_rx_action = true; 7978 local_irq_disable(); 7979 list_splice_init(&sd->poll_list, &list); 7980 local_irq_enable(); 7981 7982 for (;;) { 7983 struct napi_struct *n; 7984 7985 skb_defer_free_flush(); 7986 7987 if (list_empty(&list)) { 7988 if (list_empty(&repoll)) { 7989 sd->in_net_rx_action = false; 7990 barrier(); 7991 /* We need to check if ____napi_schedule() 7992 * had refilled poll_list while 7993 * sd->in_net_rx_action was true. 7994 */ 7995 if (!list_empty(&sd->poll_list)) 7996 goto start; 7997 if (!sd_has_rps_ipi_waiting(sd)) 7998 goto end; 7999 } 8000 break; 8001 } 8002 8003 n = list_first_entry(&list, struct napi_struct, poll_list); 8004 budget -= napi_poll(n, &repoll); 8005 8006 /* If softirq window is exhausted then punt. 8007 * Allow this to run for 2 jiffies since which will allow 8008 * an average latency of 1.5/HZ. 8009 */ 8010 if (unlikely(budget <= 0 || 8011 time_after_eq(jiffies, time_limit))) { 8012 /* Pairs with READ_ONCE() in softnet_seq_show() */ 8013 WRITE_ONCE(sd->time_squeeze, sd->time_squeeze + 1); 8014 break; 8015 } 8016 } 8017 8018 local_irq_disable(); 8019 8020 list_splice_tail_init(&sd->poll_list, &list); 8021 list_splice_tail(&repoll, &list); 8022 list_splice(&list, &sd->poll_list); 8023 if (!list_empty(&sd->poll_list)) 8024 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 8025 else 8026 sd->in_net_rx_action = false; 8027 8028 net_rps_action_and_irq_enable(sd); 8029 end: 8030 bpf_net_ctx_clear(bpf_net_ctx); 8031 } 8032 8033 struct netdev_adjacent { 8034 struct net_device *dev; 8035 netdevice_tracker dev_tracker; 8036 8037 /* upper master flag, there can only be one master device per list */ 8038 bool master; 8039 8040 /* lookup ignore flag */ 8041 bool ignore; 8042 8043 /* counter for the number of times this device was added to us */ 8044 u16 ref_nr; 8045 8046 /* private field for the users */ 8047 void *private; 8048 8049 struct list_head list; 8050 struct rcu_head rcu; 8051 }; 8052 8053 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev, 8054 struct list_head *adj_list) 8055 { 8056 struct netdev_adjacent *adj; 8057 8058 list_for_each_entry(adj, adj_list, list) { 8059 if (adj->dev == adj_dev) 8060 return adj; 8061 } 8062 return NULL; 8063 } 8064 8065 static int ____netdev_has_upper_dev(struct net_device *upper_dev, 8066 struct netdev_nested_priv *priv) 8067 { 8068 struct net_device *dev = (struct net_device *)priv->data; 8069 8070 return upper_dev == dev; 8071 } 8072 8073 /** 8074 * netdev_has_upper_dev - Check if device is linked to an upper device 8075 * @dev: device 8076 * @upper_dev: upper device to check 8077 * 8078 * Find out if a device is linked to specified upper device and return true 8079 * in case it is. Note that this checks only immediate upper device, 8080 * not through a complete stack of devices. The caller must hold the RTNL lock. 8081 */ 8082 bool netdev_has_upper_dev(struct net_device *dev, 8083 struct net_device *upper_dev) 8084 { 8085 struct netdev_nested_priv priv = { 8086 .data = (void *)upper_dev, 8087 }; 8088 8089 ASSERT_RTNL(); 8090 8091 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 8092 &priv); 8093 } 8094 EXPORT_SYMBOL(netdev_has_upper_dev); 8095 8096 /** 8097 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device 8098 * @dev: device 8099 * @upper_dev: upper device to check 8100 * 8101 * Find out if a device is linked to specified upper device and return true 8102 * in case it is. Note that this checks the entire upper device chain. 8103 * The caller must hold rcu lock. 8104 */ 8105 8106 bool netdev_has_upper_dev_all_rcu(struct net_device *dev, 8107 struct net_device *upper_dev) 8108 { 8109 struct netdev_nested_priv priv = { 8110 .data = (void *)upper_dev, 8111 }; 8112 8113 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 8114 &priv); 8115 } 8116 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu); 8117 8118 /** 8119 * netdev_has_any_upper_dev - Check if device is linked to some device 8120 * @dev: device 8121 * 8122 * Find out if a device is linked to an upper device and return true in case 8123 * it is. The caller must hold the RTNL lock. 8124 */ 8125 bool netdev_has_any_upper_dev(struct net_device *dev) 8126 { 8127 ASSERT_RTNL(); 8128 8129 return !list_empty(&dev->adj_list.upper); 8130 } 8131 EXPORT_SYMBOL(netdev_has_any_upper_dev); 8132 8133 /** 8134 * netdev_master_upper_dev_get - Get master upper device 8135 * @dev: device 8136 * 8137 * Find a master upper device and return pointer to it or NULL in case 8138 * it's not there. The caller must hold the RTNL lock. 8139 */ 8140 struct net_device *netdev_master_upper_dev_get(struct net_device *dev) 8141 { 8142 struct netdev_adjacent *upper; 8143 8144 ASSERT_RTNL(); 8145 8146 if (list_empty(&dev->adj_list.upper)) 8147 return NULL; 8148 8149 upper = list_first_entry(&dev->adj_list.upper, 8150 struct netdev_adjacent, list); 8151 if (likely(upper->master)) 8152 return upper->dev; 8153 return NULL; 8154 } 8155 EXPORT_SYMBOL(netdev_master_upper_dev_get); 8156 8157 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev) 8158 { 8159 struct netdev_adjacent *upper; 8160 8161 ASSERT_RTNL(); 8162 8163 if (list_empty(&dev->adj_list.upper)) 8164 return NULL; 8165 8166 upper = list_first_entry(&dev->adj_list.upper, 8167 struct netdev_adjacent, list); 8168 if (likely(upper->master) && !upper->ignore) 8169 return upper->dev; 8170 return NULL; 8171 } 8172 8173 /** 8174 * netdev_has_any_lower_dev - Check if device is linked to some device 8175 * @dev: device 8176 * 8177 * Find out if a device is linked to a lower device and return true in case 8178 * it is. The caller must hold the RTNL lock. 8179 */ 8180 static bool netdev_has_any_lower_dev(struct net_device *dev) 8181 { 8182 ASSERT_RTNL(); 8183 8184 return !list_empty(&dev->adj_list.lower); 8185 } 8186 8187 void *netdev_adjacent_get_private(struct list_head *adj_list) 8188 { 8189 struct netdev_adjacent *adj; 8190 8191 adj = list_entry(adj_list, struct netdev_adjacent, list); 8192 8193 return adj->private; 8194 } 8195 EXPORT_SYMBOL(netdev_adjacent_get_private); 8196 8197 /** 8198 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list 8199 * @dev: device 8200 * @iter: list_head ** of the current position 8201 * 8202 * Gets the next device from the dev's upper list, starting from iter 8203 * position. The caller must hold RCU read lock. 8204 */ 8205 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev, 8206 struct list_head **iter) 8207 { 8208 struct netdev_adjacent *upper; 8209 8210 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held() && 8211 !lockdep_rtnl_is_held()); 8212 8213 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8214 8215 if (&upper->list == &dev->adj_list.upper) 8216 return NULL; 8217 8218 *iter = &upper->list; 8219 8220 return upper->dev; 8221 } 8222 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu); 8223 8224 static struct net_device *__netdev_next_upper_dev(struct net_device *dev, 8225 struct list_head **iter, 8226 bool *ignore) 8227 { 8228 struct netdev_adjacent *upper; 8229 8230 upper = list_entry((*iter)->next, struct netdev_adjacent, list); 8231 8232 if (&upper->list == &dev->adj_list.upper) 8233 return NULL; 8234 8235 *iter = &upper->list; 8236 *ignore = upper->ignore; 8237 8238 return upper->dev; 8239 } 8240 8241 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev, 8242 struct list_head **iter) 8243 { 8244 struct netdev_adjacent *upper; 8245 8246 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 8247 8248 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8249 8250 if (&upper->list == &dev->adj_list.upper) 8251 return NULL; 8252 8253 *iter = &upper->list; 8254 8255 return upper->dev; 8256 } 8257 8258 static int __netdev_walk_all_upper_dev(struct net_device *dev, 8259 int (*fn)(struct net_device *dev, 8260 struct netdev_nested_priv *priv), 8261 struct netdev_nested_priv *priv) 8262 { 8263 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8264 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8265 int ret, cur = 0; 8266 bool ignore; 8267 8268 now = dev; 8269 iter = &dev->adj_list.upper; 8270 8271 while (1) { 8272 if (now != dev) { 8273 ret = fn(now, priv); 8274 if (ret) 8275 return ret; 8276 } 8277 8278 next = NULL; 8279 while (1) { 8280 udev = __netdev_next_upper_dev(now, &iter, &ignore); 8281 if (!udev) 8282 break; 8283 if (ignore) 8284 continue; 8285 8286 next = udev; 8287 niter = &udev->adj_list.upper; 8288 dev_stack[cur] = now; 8289 iter_stack[cur++] = iter; 8290 break; 8291 } 8292 8293 if (!next) { 8294 if (!cur) 8295 return 0; 8296 next = dev_stack[--cur]; 8297 niter = iter_stack[cur]; 8298 } 8299 8300 now = next; 8301 iter = niter; 8302 } 8303 8304 return 0; 8305 } 8306 8307 int netdev_walk_all_upper_dev_rcu(struct net_device *dev, 8308 int (*fn)(struct net_device *dev, 8309 struct netdev_nested_priv *priv), 8310 struct netdev_nested_priv *priv) 8311 { 8312 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8313 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8314 int ret, cur = 0; 8315 8316 now = dev; 8317 iter = &dev->adj_list.upper; 8318 8319 while (1) { 8320 if (now != dev) { 8321 ret = fn(now, priv); 8322 if (ret) 8323 return ret; 8324 } 8325 8326 next = NULL; 8327 while (1) { 8328 udev = netdev_next_upper_dev_rcu(now, &iter); 8329 if (!udev) 8330 break; 8331 8332 next = udev; 8333 niter = &udev->adj_list.upper; 8334 dev_stack[cur] = now; 8335 iter_stack[cur++] = iter; 8336 break; 8337 } 8338 8339 if (!next) { 8340 if (!cur) 8341 return 0; 8342 next = dev_stack[--cur]; 8343 niter = iter_stack[cur]; 8344 } 8345 8346 now = next; 8347 iter = niter; 8348 } 8349 8350 return 0; 8351 } 8352 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu); 8353 8354 static bool __netdev_has_upper_dev(struct net_device *dev, 8355 struct net_device *upper_dev) 8356 { 8357 struct netdev_nested_priv priv = { 8358 .flags = 0, 8359 .data = (void *)upper_dev, 8360 }; 8361 8362 ASSERT_RTNL(); 8363 8364 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev, 8365 &priv); 8366 } 8367 8368 /** 8369 * netdev_lower_get_next_private - Get the next ->private from the 8370 * lower neighbour list 8371 * @dev: device 8372 * @iter: list_head ** of the current position 8373 * 8374 * Gets the next netdev_adjacent->private from the dev's lower neighbour 8375 * list, starting from iter position. The caller must hold either hold the 8376 * RTNL lock or its own locking that guarantees that the neighbour lower 8377 * list will remain unchanged. 8378 */ 8379 void *netdev_lower_get_next_private(struct net_device *dev, 8380 struct list_head **iter) 8381 { 8382 struct netdev_adjacent *lower; 8383 8384 lower = list_entry(*iter, struct netdev_adjacent, list); 8385 8386 if (&lower->list == &dev->adj_list.lower) 8387 return NULL; 8388 8389 *iter = lower->list.next; 8390 8391 return lower->private; 8392 } 8393 EXPORT_SYMBOL(netdev_lower_get_next_private); 8394 8395 /** 8396 * netdev_lower_get_next_private_rcu - Get the next ->private from the 8397 * lower neighbour list, RCU 8398 * variant 8399 * @dev: device 8400 * @iter: list_head ** of the current position 8401 * 8402 * Gets the next netdev_adjacent->private from the dev's lower neighbour 8403 * list, starting from iter position. The caller must hold RCU read lock. 8404 */ 8405 void *netdev_lower_get_next_private_rcu(struct net_device *dev, 8406 struct list_head **iter) 8407 { 8408 struct netdev_adjacent *lower; 8409 8410 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 8411 8412 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8413 8414 if (&lower->list == &dev->adj_list.lower) 8415 return NULL; 8416 8417 *iter = &lower->list; 8418 8419 return lower->private; 8420 } 8421 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu); 8422 8423 /** 8424 * netdev_lower_get_next - Get the next device from the lower neighbour 8425 * list 8426 * @dev: device 8427 * @iter: list_head ** of the current position 8428 * 8429 * Gets the next netdev_adjacent from the dev's lower neighbour 8430 * list, starting from iter position. The caller must hold RTNL lock or 8431 * its own locking that guarantees that the neighbour lower 8432 * list will remain unchanged. 8433 */ 8434 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter) 8435 { 8436 struct netdev_adjacent *lower; 8437 8438 lower = list_entry(*iter, struct netdev_adjacent, list); 8439 8440 if (&lower->list == &dev->adj_list.lower) 8441 return NULL; 8442 8443 *iter = lower->list.next; 8444 8445 return lower->dev; 8446 } 8447 EXPORT_SYMBOL(netdev_lower_get_next); 8448 8449 static struct net_device *netdev_next_lower_dev(struct net_device *dev, 8450 struct list_head **iter) 8451 { 8452 struct netdev_adjacent *lower; 8453 8454 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 8455 8456 if (&lower->list == &dev->adj_list.lower) 8457 return NULL; 8458 8459 *iter = &lower->list; 8460 8461 return lower->dev; 8462 } 8463 8464 static struct net_device *__netdev_next_lower_dev(struct net_device *dev, 8465 struct list_head **iter, 8466 bool *ignore) 8467 { 8468 struct netdev_adjacent *lower; 8469 8470 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 8471 8472 if (&lower->list == &dev->adj_list.lower) 8473 return NULL; 8474 8475 *iter = &lower->list; 8476 *ignore = lower->ignore; 8477 8478 return lower->dev; 8479 } 8480 8481 int netdev_walk_all_lower_dev(struct net_device *dev, 8482 int (*fn)(struct net_device *dev, 8483 struct netdev_nested_priv *priv), 8484 struct netdev_nested_priv *priv) 8485 { 8486 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8487 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8488 int ret, cur = 0; 8489 8490 now = dev; 8491 iter = &dev->adj_list.lower; 8492 8493 while (1) { 8494 if (now != dev) { 8495 ret = fn(now, priv); 8496 if (ret) 8497 return ret; 8498 } 8499 8500 next = NULL; 8501 while (1) { 8502 ldev = netdev_next_lower_dev(now, &iter); 8503 if (!ldev) 8504 break; 8505 8506 next = ldev; 8507 niter = &ldev->adj_list.lower; 8508 dev_stack[cur] = now; 8509 iter_stack[cur++] = iter; 8510 break; 8511 } 8512 8513 if (!next) { 8514 if (!cur) 8515 return 0; 8516 next = dev_stack[--cur]; 8517 niter = iter_stack[cur]; 8518 } 8519 8520 now = next; 8521 iter = niter; 8522 } 8523 8524 return 0; 8525 } 8526 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev); 8527 8528 static int __netdev_walk_all_lower_dev(struct net_device *dev, 8529 int (*fn)(struct net_device *dev, 8530 struct netdev_nested_priv *priv), 8531 struct netdev_nested_priv *priv) 8532 { 8533 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8534 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8535 int ret, cur = 0; 8536 bool ignore; 8537 8538 now = dev; 8539 iter = &dev->adj_list.lower; 8540 8541 while (1) { 8542 if (now != dev) { 8543 ret = fn(now, priv); 8544 if (ret) 8545 return ret; 8546 } 8547 8548 next = NULL; 8549 while (1) { 8550 ldev = __netdev_next_lower_dev(now, &iter, &ignore); 8551 if (!ldev) 8552 break; 8553 if (ignore) 8554 continue; 8555 8556 next = ldev; 8557 niter = &ldev->adj_list.lower; 8558 dev_stack[cur] = now; 8559 iter_stack[cur++] = iter; 8560 break; 8561 } 8562 8563 if (!next) { 8564 if (!cur) 8565 return 0; 8566 next = dev_stack[--cur]; 8567 niter = iter_stack[cur]; 8568 } 8569 8570 now = next; 8571 iter = niter; 8572 } 8573 8574 return 0; 8575 } 8576 8577 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev, 8578 struct list_head **iter) 8579 { 8580 struct netdev_adjacent *lower; 8581 8582 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8583 if (&lower->list == &dev->adj_list.lower) 8584 return NULL; 8585 8586 *iter = &lower->list; 8587 8588 return lower->dev; 8589 } 8590 EXPORT_SYMBOL(netdev_next_lower_dev_rcu); 8591 8592 static u8 __netdev_upper_depth(struct net_device *dev) 8593 { 8594 struct net_device *udev; 8595 struct list_head *iter; 8596 u8 max_depth = 0; 8597 bool ignore; 8598 8599 for (iter = &dev->adj_list.upper, 8600 udev = __netdev_next_upper_dev(dev, &iter, &ignore); 8601 udev; 8602 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) { 8603 if (ignore) 8604 continue; 8605 if (max_depth < udev->upper_level) 8606 max_depth = udev->upper_level; 8607 } 8608 8609 return max_depth; 8610 } 8611 8612 static u8 __netdev_lower_depth(struct net_device *dev) 8613 { 8614 struct net_device *ldev; 8615 struct list_head *iter; 8616 u8 max_depth = 0; 8617 bool ignore; 8618 8619 for (iter = &dev->adj_list.lower, 8620 ldev = __netdev_next_lower_dev(dev, &iter, &ignore); 8621 ldev; 8622 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) { 8623 if (ignore) 8624 continue; 8625 if (max_depth < ldev->lower_level) 8626 max_depth = ldev->lower_level; 8627 } 8628 8629 return max_depth; 8630 } 8631 8632 static int __netdev_update_upper_level(struct net_device *dev, 8633 struct netdev_nested_priv *__unused) 8634 { 8635 dev->upper_level = __netdev_upper_depth(dev) + 1; 8636 return 0; 8637 } 8638 8639 #ifdef CONFIG_LOCKDEP 8640 static LIST_HEAD(net_unlink_list); 8641 8642 static void net_unlink_todo(struct net_device *dev) 8643 { 8644 if (list_empty(&dev->unlink_list)) 8645 list_add_tail(&dev->unlink_list, &net_unlink_list); 8646 } 8647 #endif 8648 8649 static int __netdev_update_lower_level(struct net_device *dev, 8650 struct netdev_nested_priv *priv) 8651 { 8652 dev->lower_level = __netdev_lower_depth(dev) + 1; 8653 8654 #ifdef CONFIG_LOCKDEP 8655 if (!priv) 8656 return 0; 8657 8658 if (priv->flags & NESTED_SYNC_IMM) 8659 dev->nested_level = dev->lower_level - 1; 8660 if (priv->flags & NESTED_SYNC_TODO) 8661 net_unlink_todo(dev); 8662 #endif 8663 return 0; 8664 } 8665 8666 int netdev_walk_all_lower_dev_rcu(struct net_device *dev, 8667 int (*fn)(struct net_device *dev, 8668 struct netdev_nested_priv *priv), 8669 struct netdev_nested_priv *priv) 8670 { 8671 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8672 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8673 int ret, cur = 0; 8674 8675 now = dev; 8676 iter = &dev->adj_list.lower; 8677 8678 while (1) { 8679 if (now != dev) { 8680 ret = fn(now, priv); 8681 if (ret) 8682 return ret; 8683 } 8684 8685 next = NULL; 8686 while (1) { 8687 ldev = netdev_next_lower_dev_rcu(now, &iter); 8688 if (!ldev) 8689 break; 8690 8691 next = ldev; 8692 niter = &ldev->adj_list.lower; 8693 dev_stack[cur] = now; 8694 iter_stack[cur++] = iter; 8695 break; 8696 } 8697 8698 if (!next) { 8699 if (!cur) 8700 return 0; 8701 next = dev_stack[--cur]; 8702 niter = iter_stack[cur]; 8703 } 8704 8705 now = next; 8706 iter = niter; 8707 } 8708 8709 return 0; 8710 } 8711 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu); 8712 8713 /** 8714 * netdev_lower_get_first_private_rcu - Get the first ->private from the 8715 * lower neighbour list, RCU 8716 * variant 8717 * @dev: device 8718 * 8719 * Gets the first netdev_adjacent->private from the dev's lower neighbour 8720 * list. The caller must hold RCU read lock. 8721 */ 8722 void *netdev_lower_get_first_private_rcu(struct net_device *dev) 8723 { 8724 struct netdev_adjacent *lower; 8725 8726 lower = list_first_or_null_rcu(&dev->adj_list.lower, 8727 struct netdev_adjacent, list); 8728 if (lower) 8729 return lower->private; 8730 return NULL; 8731 } 8732 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu); 8733 8734 /** 8735 * netdev_master_upper_dev_get_rcu - Get master upper device 8736 * @dev: device 8737 * 8738 * Find a master upper device and return pointer to it or NULL in case 8739 * it's not there. The caller must hold the RCU read lock. 8740 */ 8741 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev) 8742 { 8743 struct netdev_adjacent *upper; 8744 8745 upper = list_first_or_null_rcu(&dev->adj_list.upper, 8746 struct netdev_adjacent, list); 8747 if (upper && likely(upper->master)) 8748 return upper->dev; 8749 return NULL; 8750 } 8751 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu); 8752 8753 static int netdev_adjacent_sysfs_add(struct net_device *dev, 8754 struct net_device *adj_dev, 8755 struct list_head *dev_list) 8756 { 8757 char linkname[IFNAMSIZ+7]; 8758 8759 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8760 "upper_%s" : "lower_%s", adj_dev->name); 8761 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj), 8762 linkname); 8763 } 8764 static void netdev_adjacent_sysfs_del(struct net_device *dev, 8765 char *name, 8766 struct list_head *dev_list) 8767 { 8768 char linkname[IFNAMSIZ+7]; 8769 8770 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8771 "upper_%s" : "lower_%s", name); 8772 sysfs_remove_link(&(dev->dev.kobj), linkname); 8773 } 8774 8775 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev, 8776 struct net_device *adj_dev, 8777 struct list_head *dev_list) 8778 { 8779 return (dev_list == &dev->adj_list.upper || 8780 dev_list == &dev->adj_list.lower) && 8781 net_eq(dev_net(dev), dev_net(adj_dev)); 8782 } 8783 8784 static int __netdev_adjacent_dev_insert(struct net_device *dev, 8785 struct net_device *adj_dev, 8786 struct list_head *dev_list, 8787 void *private, bool master) 8788 { 8789 struct netdev_adjacent *adj; 8790 int ret; 8791 8792 adj = __netdev_find_adj(adj_dev, dev_list); 8793 8794 if (adj) { 8795 adj->ref_nr += 1; 8796 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n", 8797 dev->name, adj_dev->name, adj->ref_nr); 8798 8799 return 0; 8800 } 8801 8802 adj = kmalloc_obj(*adj); 8803 if (!adj) 8804 return -ENOMEM; 8805 8806 adj->dev = adj_dev; 8807 adj->master = master; 8808 adj->ref_nr = 1; 8809 adj->private = private; 8810 adj->ignore = false; 8811 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL); 8812 8813 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n", 8814 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name); 8815 8816 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) { 8817 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list); 8818 if (ret) 8819 goto free_adj; 8820 } 8821 8822 /* Ensure that master link is always the first item in list. */ 8823 if (master) { 8824 ret = sysfs_create_link(&(dev->dev.kobj), 8825 &(adj_dev->dev.kobj), "master"); 8826 if (ret) 8827 goto remove_symlinks; 8828 8829 list_add_rcu(&adj->list, dev_list); 8830 } else { 8831 list_add_tail_rcu(&adj->list, dev_list); 8832 } 8833 8834 return 0; 8835 8836 remove_symlinks: 8837 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8838 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8839 free_adj: 8840 netdev_put(adj_dev, &adj->dev_tracker); 8841 kfree(adj); 8842 8843 return ret; 8844 } 8845 8846 static void __netdev_adjacent_dev_remove(struct net_device *dev, 8847 struct net_device *adj_dev, 8848 u16 ref_nr, 8849 struct list_head *dev_list) 8850 { 8851 struct netdev_adjacent *adj; 8852 8853 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n", 8854 dev->name, adj_dev->name, ref_nr); 8855 8856 adj = __netdev_find_adj(adj_dev, dev_list); 8857 8858 if (!adj) { 8859 pr_err("Adjacency does not exist for device %s from %s\n", 8860 dev->name, adj_dev->name); 8861 WARN_ON(1); 8862 return; 8863 } 8864 8865 if (adj->ref_nr > ref_nr) { 8866 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n", 8867 dev->name, adj_dev->name, ref_nr, 8868 adj->ref_nr - ref_nr); 8869 adj->ref_nr -= ref_nr; 8870 return; 8871 } 8872 8873 if (adj->master) 8874 sysfs_remove_link(&(dev->dev.kobj), "master"); 8875 8876 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8877 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8878 8879 list_del_rcu(&adj->list); 8880 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n", 8881 adj_dev->name, dev->name, adj_dev->name); 8882 netdev_put(adj_dev, &adj->dev_tracker); 8883 kfree_rcu(adj, rcu); 8884 } 8885 8886 static int __netdev_adjacent_dev_link_lists(struct net_device *dev, 8887 struct net_device *upper_dev, 8888 struct list_head *up_list, 8889 struct list_head *down_list, 8890 void *private, bool master) 8891 { 8892 int ret; 8893 8894 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list, 8895 private, master); 8896 if (ret) 8897 return ret; 8898 8899 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list, 8900 private, false); 8901 if (ret) { 8902 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list); 8903 return ret; 8904 } 8905 8906 return 0; 8907 } 8908 8909 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev, 8910 struct net_device *upper_dev, 8911 u16 ref_nr, 8912 struct list_head *up_list, 8913 struct list_head *down_list) 8914 { 8915 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list); 8916 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list); 8917 } 8918 8919 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev, 8920 struct net_device *upper_dev, 8921 void *private, bool master) 8922 { 8923 return __netdev_adjacent_dev_link_lists(dev, upper_dev, 8924 &dev->adj_list.upper, 8925 &upper_dev->adj_list.lower, 8926 private, master); 8927 } 8928 8929 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev, 8930 struct net_device *upper_dev) 8931 { 8932 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1, 8933 &dev->adj_list.upper, 8934 &upper_dev->adj_list.lower); 8935 } 8936 8937 static int __netdev_upper_dev_link(struct net_device *dev, 8938 struct net_device *upper_dev, bool master, 8939 void *upper_priv, void *upper_info, 8940 struct netdev_nested_priv *priv, 8941 struct netlink_ext_ack *extack) 8942 { 8943 struct netdev_notifier_changeupper_info changeupper_info = { 8944 .info = { 8945 .dev = dev, 8946 .extack = extack, 8947 }, 8948 .upper_dev = upper_dev, 8949 .master = master, 8950 .linking = true, 8951 .upper_info = upper_info, 8952 }; 8953 struct net_device *master_dev; 8954 int ret = 0; 8955 8956 ASSERT_RTNL(); 8957 8958 if (dev == upper_dev) 8959 return -EBUSY; 8960 8961 /* To prevent loops, check if dev is not upper device to upper_dev. */ 8962 if (__netdev_has_upper_dev(upper_dev, dev)) 8963 return -EBUSY; 8964 8965 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV) 8966 return -EMLINK; 8967 8968 if (!master) { 8969 if (__netdev_has_upper_dev(dev, upper_dev)) 8970 return -EEXIST; 8971 } else { 8972 master_dev = __netdev_master_upper_dev_get(dev); 8973 if (master_dev) 8974 return master_dev == upper_dev ? -EEXIST : -EBUSY; 8975 } 8976 8977 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8978 &changeupper_info.info); 8979 ret = notifier_to_errno(ret); 8980 if (ret) 8981 return ret; 8982 8983 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv, 8984 master); 8985 if (ret) 8986 return ret; 8987 8988 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8989 &changeupper_info.info); 8990 ret = notifier_to_errno(ret); 8991 if (ret) 8992 goto rollback; 8993 8994 __netdev_update_upper_level(dev, NULL); 8995 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8996 8997 __netdev_update_lower_level(upper_dev, priv); 8998 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8999 priv); 9000 9001 return 0; 9002 9003 rollback: 9004 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 9005 9006 return ret; 9007 } 9008 9009 /** 9010 * netdev_upper_dev_link - Add a link to the upper device 9011 * @dev: device 9012 * @upper_dev: new upper device 9013 * @extack: netlink extended ack 9014 * 9015 * Adds a link to device which is upper to this one. The caller must hold 9016 * the RTNL lock. On a failure a negative errno code is returned. 9017 * On success the reference counts are adjusted and the function 9018 * returns zero. 9019 */ 9020 int netdev_upper_dev_link(struct net_device *dev, 9021 struct net_device *upper_dev, 9022 struct netlink_ext_ack *extack) 9023 { 9024 struct netdev_nested_priv priv = { 9025 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 9026 .data = NULL, 9027 }; 9028 9029 return __netdev_upper_dev_link(dev, upper_dev, false, 9030 NULL, NULL, &priv, extack); 9031 } 9032 EXPORT_SYMBOL(netdev_upper_dev_link); 9033 9034 /** 9035 * netdev_master_upper_dev_link - Add a master link to the upper device 9036 * @dev: device 9037 * @upper_dev: new upper device 9038 * @upper_priv: upper device private 9039 * @upper_info: upper info to be passed down via notifier 9040 * @extack: netlink extended ack 9041 * 9042 * Adds a link to device which is upper to this one. In this case, only 9043 * one master upper device can be linked, although other non-master devices 9044 * might be linked as well. The caller must hold the RTNL lock. 9045 * On a failure a negative errno code is returned. On success the reference 9046 * counts are adjusted and the function returns zero. 9047 */ 9048 int netdev_master_upper_dev_link(struct net_device *dev, 9049 struct net_device *upper_dev, 9050 void *upper_priv, void *upper_info, 9051 struct netlink_ext_ack *extack) 9052 { 9053 struct netdev_nested_priv priv = { 9054 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 9055 .data = NULL, 9056 }; 9057 9058 return __netdev_upper_dev_link(dev, upper_dev, true, 9059 upper_priv, upper_info, &priv, extack); 9060 } 9061 EXPORT_SYMBOL(netdev_master_upper_dev_link); 9062 9063 static void __netdev_upper_dev_unlink(struct net_device *dev, 9064 struct net_device *upper_dev, 9065 struct netdev_nested_priv *priv) 9066 { 9067 struct netdev_notifier_changeupper_info changeupper_info = { 9068 .info = { 9069 .dev = dev, 9070 }, 9071 .upper_dev = upper_dev, 9072 .linking = false, 9073 }; 9074 9075 ASSERT_RTNL(); 9076 9077 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev; 9078 9079 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 9080 &changeupper_info.info); 9081 9082 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 9083 9084 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 9085 &changeupper_info.info); 9086 9087 __netdev_update_upper_level(dev, NULL); 9088 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 9089 9090 __netdev_update_lower_level(upper_dev, priv); 9091 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 9092 priv); 9093 } 9094 9095 /** 9096 * netdev_upper_dev_unlink - Removes a link to upper device 9097 * @dev: device 9098 * @upper_dev: new upper device 9099 * 9100 * Removes a link to device which is upper to this one. The caller must hold 9101 * the RTNL lock. 9102 */ 9103 void netdev_upper_dev_unlink(struct net_device *dev, 9104 struct net_device *upper_dev) 9105 { 9106 struct netdev_nested_priv priv = { 9107 .flags = NESTED_SYNC_TODO, 9108 .data = NULL, 9109 }; 9110 9111 __netdev_upper_dev_unlink(dev, upper_dev, &priv); 9112 } 9113 EXPORT_SYMBOL(netdev_upper_dev_unlink); 9114 9115 static void __netdev_adjacent_dev_set(struct net_device *upper_dev, 9116 struct net_device *lower_dev, 9117 bool val) 9118 { 9119 struct netdev_adjacent *adj; 9120 9121 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower); 9122 if (adj) 9123 adj->ignore = val; 9124 9125 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper); 9126 if (adj) 9127 adj->ignore = val; 9128 } 9129 9130 static void netdev_adjacent_dev_disable(struct net_device *upper_dev, 9131 struct net_device *lower_dev) 9132 { 9133 __netdev_adjacent_dev_set(upper_dev, lower_dev, true); 9134 } 9135 9136 static void netdev_adjacent_dev_enable(struct net_device *upper_dev, 9137 struct net_device *lower_dev) 9138 { 9139 __netdev_adjacent_dev_set(upper_dev, lower_dev, false); 9140 } 9141 9142 int netdev_adjacent_change_prepare(struct net_device *old_dev, 9143 struct net_device *new_dev, 9144 struct net_device *dev, 9145 struct netlink_ext_ack *extack) 9146 { 9147 struct netdev_nested_priv priv = { 9148 .flags = 0, 9149 .data = NULL, 9150 }; 9151 int err; 9152 9153 if (!new_dev) 9154 return 0; 9155 9156 if (old_dev && new_dev != old_dev) 9157 netdev_adjacent_dev_disable(dev, old_dev); 9158 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv, 9159 extack); 9160 if (err) { 9161 if (old_dev && new_dev != old_dev) 9162 netdev_adjacent_dev_enable(dev, old_dev); 9163 return err; 9164 } 9165 9166 return 0; 9167 } 9168 EXPORT_SYMBOL(netdev_adjacent_change_prepare); 9169 9170 void netdev_adjacent_change_commit(struct net_device *old_dev, 9171 struct net_device *new_dev, 9172 struct net_device *dev) 9173 { 9174 struct netdev_nested_priv priv = { 9175 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 9176 .data = NULL, 9177 }; 9178 9179 if (!new_dev || !old_dev) 9180 return; 9181 9182 if (new_dev == old_dev) 9183 return; 9184 9185 netdev_adjacent_dev_enable(dev, old_dev); 9186 __netdev_upper_dev_unlink(old_dev, dev, &priv); 9187 } 9188 EXPORT_SYMBOL(netdev_adjacent_change_commit); 9189 9190 void netdev_adjacent_change_abort(struct net_device *old_dev, 9191 struct net_device *new_dev, 9192 struct net_device *dev) 9193 { 9194 struct netdev_nested_priv priv = { 9195 .flags = 0, 9196 .data = NULL, 9197 }; 9198 9199 if (!new_dev) 9200 return; 9201 9202 if (old_dev && new_dev != old_dev) 9203 netdev_adjacent_dev_enable(dev, old_dev); 9204 9205 __netdev_upper_dev_unlink(new_dev, dev, &priv); 9206 } 9207 EXPORT_SYMBOL(netdev_adjacent_change_abort); 9208 9209 /** 9210 * netdev_bonding_info_change - Dispatch event about slave change 9211 * @dev: device 9212 * @bonding_info: info to dispatch 9213 * 9214 * Send NETDEV_BONDING_INFO to netdev notifiers with info. 9215 * The caller must hold the RTNL lock. 9216 */ 9217 void netdev_bonding_info_change(struct net_device *dev, 9218 struct netdev_bonding_info *bonding_info) 9219 { 9220 struct netdev_notifier_bonding_info info = { 9221 .info.dev = dev, 9222 }; 9223 9224 memcpy(&info.bonding_info, bonding_info, 9225 sizeof(struct netdev_bonding_info)); 9226 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, 9227 &info.info); 9228 } 9229 EXPORT_SYMBOL(netdev_bonding_info_change); 9230 9231 static int netdev_offload_xstats_enable_l3(struct net_device *dev, 9232 struct netlink_ext_ack *extack) 9233 { 9234 struct netdev_notifier_offload_xstats_info info = { 9235 .info.dev = dev, 9236 .info.extack = extack, 9237 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 9238 }; 9239 int err; 9240 int rc; 9241 9242 dev->offload_xstats_l3 = kzalloc_obj(*dev->offload_xstats_l3); 9243 if (!dev->offload_xstats_l3) 9244 return -ENOMEM; 9245 9246 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE, 9247 NETDEV_OFFLOAD_XSTATS_DISABLE, 9248 &info.info); 9249 err = notifier_to_errno(rc); 9250 if (err) 9251 goto free_stats; 9252 9253 return 0; 9254 9255 free_stats: 9256 kfree(dev->offload_xstats_l3); 9257 dev->offload_xstats_l3 = NULL; 9258 return err; 9259 } 9260 9261 int netdev_offload_xstats_enable(struct net_device *dev, 9262 enum netdev_offload_xstats_type type, 9263 struct netlink_ext_ack *extack) 9264 { 9265 ASSERT_RTNL(); 9266 9267 if (netdev_offload_xstats_enabled(dev, type)) 9268 return -EALREADY; 9269 9270 switch (type) { 9271 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 9272 return netdev_offload_xstats_enable_l3(dev, extack); 9273 } 9274 9275 WARN_ON(1); 9276 return -EINVAL; 9277 } 9278 EXPORT_SYMBOL(netdev_offload_xstats_enable); 9279 9280 static void netdev_offload_xstats_disable_l3(struct net_device *dev) 9281 { 9282 struct netdev_notifier_offload_xstats_info info = { 9283 .info.dev = dev, 9284 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 9285 }; 9286 9287 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE, 9288 &info.info); 9289 kfree(dev->offload_xstats_l3); 9290 dev->offload_xstats_l3 = NULL; 9291 } 9292 9293 int netdev_offload_xstats_disable(struct net_device *dev, 9294 enum netdev_offload_xstats_type type) 9295 { 9296 ASSERT_RTNL(); 9297 9298 if (!netdev_offload_xstats_enabled(dev, type)) 9299 return -EALREADY; 9300 9301 switch (type) { 9302 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 9303 netdev_offload_xstats_disable_l3(dev); 9304 return 0; 9305 } 9306 9307 WARN_ON(1); 9308 return -EINVAL; 9309 } 9310 EXPORT_SYMBOL(netdev_offload_xstats_disable); 9311 9312 static void netdev_offload_xstats_disable_all(struct net_device *dev) 9313 { 9314 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3); 9315 } 9316 9317 static struct rtnl_hw_stats64 * 9318 netdev_offload_xstats_get_ptr(const struct net_device *dev, 9319 enum netdev_offload_xstats_type type) 9320 { 9321 switch (type) { 9322 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 9323 return dev->offload_xstats_l3; 9324 } 9325 9326 WARN_ON(1); 9327 return NULL; 9328 } 9329 9330 bool netdev_offload_xstats_enabled(const struct net_device *dev, 9331 enum netdev_offload_xstats_type type) 9332 { 9333 ASSERT_RTNL(); 9334 9335 return netdev_offload_xstats_get_ptr(dev, type); 9336 } 9337 EXPORT_SYMBOL(netdev_offload_xstats_enabled); 9338 9339 struct netdev_notifier_offload_xstats_ru { 9340 bool used; 9341 }; 9342 9343 struct netdev_notifier_offload_xstats_rd { 9344 struct rtnl_hw_stats64 stats; 9345 bool used; 9346 }; 9347 9348 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest, 9349 const struct rtnl_hw_stats64 *src) 9350 { 9351 dest->rx_packets += src->rx_packets; 9352 dest->tx_packets += src->tx_packets; 9353 dest->rx_bytes += src->rx_bytes; 9354 dest->tx_bytes += src->tx_bytes; 9355 dest->rx_errors += src->rx_errors; 9356 dest->tx_errors += src->tx_errors; 9357 dest->rx_dropped += src->rx_dropped; 9358 dest->tx_dropped += src->tx_dropped; 9359 dest->multicast += src->multicast; 9360 } 9361 9362 static int netdev_offload_xstats_get_used(struct net_device *dev, 9363 enum netdev_offload_xstats_type type, 9364 bool *p_used, 9365 struct netlink_ext_ack *extack) 9366 { 9367 struct netdev_notifier_offload_xstats_ru report_used = {}; 9368 struct netdev_notifier_offload_xstats_info info = { 9369 .info.dev = dev, 9370 .info.extack = extack, 9371 .type = type, 9372 .report_used = &report_used, 9373 }; 9374 int rc; 9375 9376 WARN_ON(!netdev_offload_xstats_enabled(dev, type)); 9377 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED, 9378 &info.info); 9379 *p_used = report_used.used; 9380 return notifier_to_errno(rc); 9381 } 9382 9383 static int netdev_offload_xstats_get_stats(struct net_device *dev, 9384 enum netdev_offload_xstats_type type, 9385 struct rtnl_hw_stats64 *p_stats, 9386 bool *p_used, 9387 struct netlink_ext_ack *extack) 9388 { 9389 struct netdev_notifier_offload_xstats_rd report_delta = {}; 9390 struct netdev_notifier_offload_xstats_info info = { 9391 .info.dev = dev, 9392 .info.extack = extack, 9393 .type = type, 9394 .report_delta = &report_delta, 9395 }; 9396 struct rtnl_hw_stats64 *stats; 9397 int rc; 9398 9399 stats = netdev_offload_xstats_get_ptr(dev, type); 9400 if (WARN_ON(!stats)) 9401 return -EINVAL; 9402 9403 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA, 9404 &info.info); 9405 9406 /* Cache whatever we got, even if there was an error, otherwise the 9407 * successful stats retrievals would get lost. 9408 */ 9409 netdev_hw_stats64_add(stats, &report_delta.stats); 9410 9411 if (p_stats) 9412 *p_stats = *stats; 9413 *p_used = report_delta.used; 9414 9415 return notifier_to_errno(rc); 9416 } 9417 9418 int netdev_offload_xstats_get(struct net_device *dev, 9419 enum netdev_offload_xstats_type type, 9420 struct rtnl_hw_stats64 *p_stats, bool *p_used, 9421 struct netlink_ext_ack *extack) 9422 { 9423 ASSERT_RTNL(); 9424 9425 if (p_stats) 9426 return netdev_offload_xstats_get_stats(dev, type, p_stats, 9427 p_used, extack); 9428 else 9429 return netdev_offload_xstats_get_used(dev, type, p_used, 9430 extack); 9431 } 9432 EXPORT_SYMBOL(netdev_offload_xstats_get); 9433 9434 void 9435 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta, 9436 const struct rtnl_hw_stats64 *stats) 9437 { 9438 report_delta->used = true; 9439 netdev_hw_stats64_add(&report_delta->stats, stats); 9440 } 9441 EXPORT_SYMBOL(netdev_offload_xstats_report_delta); 9442 9443 void 9444 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used) 9445 { 9446 report_used->used = true; 9447 } 9448 EXPORT_SYMBOL(netdev_offload_xstats_report_used); 9449 9450 void netdev_offload_xstats_push_delta(struct net_device *dev, 9451 enum netdev_offload_xstats_type type, 9452 const struct rtnl_hw_stats64 *p_stats) 9453 { 9454 struct rtnl_hw_stats64 *stats; 9455 9456 ASSERT_RTNL(); 9457 9458 stats = netdev_offload_xstats_get_ptr(dev, type); 9459 if (WARN_ON(!stats)) 9460 return; 9461 9462 netdev_hw_stats64_add(stats, p_stats); 9463 } 9464 EXPORT_SYMBOL(netdev_offload_xstats_push_delta); 9465 9466 /** 9467 * netdev_get_xmit_slave - Get the xmit slave of master device 9468 * @dev: device 9469 * @skb: The packet 9470 * @all_slaves: assume all the slaves are active 9471 * 9472 * The reference counters are not incremented so the caller must be 9473 * careful with locks. The caller must hold RCU lock. 9474 * %NULL is returned if no slave is found. 9475 */ 9476 9477 struct net_device *netdev_get_xmit_slave(struct net_device *dev, 9478 struct sk_buff *skb, 9479 bool all_slaves) 9480 { 9481 const struct net_device_ops *ops = dev->netdev_ops; 9482 9483 if (!ops->ndo_get_xmit_slave) 9484 return NULL; 9485 return ops->ndo_get_xmit_slave(dev, skb, all_slaves); 9486 } 9487 EXPORT_SYMBOL(netdev_get_xmit_slave); 9488 9489 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev, 9490 struct sock *sk) 9491 { 9492 const struct net_device_ops *ops = dev->netdev_ops; 9493 9494 if (!ops->ndo_sk_get_lower_dev) 9495 return NULL; 9496 return ops->ndo_sk_get_lower_dev(dev, sk); 9497 } 9498 9499 /** 9500 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket 9501 * @dev: device 9502 * @sk: the socket 9503 * 9504 * %NULL is returned if no lower device is found. 9505 */ 9506 9507 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev, 9508 struct sock *sk) 9509 { 9510 struct net_device *lower; 9511 9512 lower = netdev_sk_get_lower_dev(dev, sk); 9513 while (lower) { 9514 dev = lower; 9515 lower = netdev_sk_get_lower_dev(dev, sk); 9516 } 9517 9518 return dev; 9519 } 9520 EXPORT_SYMBOL(netdev_sk_get_lowest_dev); 9521 9522 static void netdev_adjacent_add_links(struct net_device *dev) 9523 { 9524 struct netdev_adjacent *iter; 9525 9526 struct net *net = dev_net(dev); 9527 9528 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9529 if (!net_eq(net, dev_net(iter->dev))) 9530 continue; 9531 netdev_adjacent_sysfs_add(iter->dev, dev, 9532 &iter->dev->adj_list.lower); 9533 netdev_adjacent_sysfs_add(dev, iter->dev, 9534 &dev->adj_list.upper); 9535 } 9536 9537 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9538 if (!net_eq(net, dev_net(iter->dev))) 9539 continue; 9540 netdev_adjacent_sysfs_add(iter->dev, dev, 9541 &iter->dev->adj_list.upper); 9542 netdev_adjacent_sysfs_add(dev, iter->dev, 9543 &dev->adj_list.lower); 9544 } 9545 } 9546 9547 static void netdev_adjacent_del_links(struct net_device *dev) 9548 { 9549 struct netdev_adjacent *iter; 9550 9551 struct net *net = dev_net(dev); 9552 9553 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9554 if (!net_eq(net, dev_net(iter->dev))) 9555 continue; 9556 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9557 &iter->dev->adj_list.lower); 9558 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9559 &dev->adj_list.upper); 9560 } 9561 9562 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9563 if (!net_eq(net, dev_net(iter->dev))) 9564 continue; 9565 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9566 &iter->dev->adj_list.upper); 9567 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9568 &dev->adj_list.lower); 9569 } 9570 } 9571 9572 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname) 9573 { 9574 struct netdev_adjacent *iter; 9575 9576 struct net *net = dev_net(dev); 9577 9578 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9579 if (!net_eq(net, dev_net(iter->dev))) 9580 continue; 9581 netdev_adjacent_sysfs_del(iter->dev, oldname, 9582 &iter->dev->adj_list.lower); 9583 netdev_adjacent_sysfs_add(iter->dev, dev, 9584 &iter->dev->adj_list.lower); 9585 } 9586 9587 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9588 if (!net_eq(net, dev_net(iter->dev))) 9589 continue; 9590 netdev_adjacent_sysfs_del(iter->dev, oldname, 9591 &iter->dev->adj_list.upper); 9592 netdev_adjacent_sysfs_add(iter->dev, dev, 9593 &iter->dev->adj_list.upper); 9594 } 9595 } 9596 9597 void *netdev_lower_dev_get_private(struct net_device *dev, 9598 struct net_device *lower_dev) 9599 { 9600 struct netdev_adjacent *lower; 9601 9602 if (!lower_dev) 9603 return NULL; 9604 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower); 9605 if (!lower) 9606 return NULL; 9607 9608 return lower->private; 9609 } 9610 EXPORT_SYMBOL(netdev_lower_dev_get_private); 9611 9612 9613 /** 9614 * netdev_lower_state_changed - Dispatch event about lower device state change 9615 * @lower_dev: device 9616 * @lower_state_info: state to dispatch 9617 * 9618 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info. 9619 * The caller must hold the RTNL lock. 9620 */ 9621 void netdev_lower_state_changed(struct net_device *lower_dev, 9622 void *lower_state_info) 9623 { 9624 struct netdev_notifier_changelowerstate_info changelowerstate_info = { 9625 .info.dev = lower_dev, 9626 }; 9627 9628 ASSERT_RTNL(); 9629 changelowerstate_info.lower_state_info = lower_state_info; 9630 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, 9631 &changelowerstate_info.info); 9632 } 9633 EXPORT_SYMBOL(netdev_lower_state_changed); 9634 9635 static void dev_change_rx_flags(struct net_device *dev, int flags) 9636 { 9637 const struct net_device_ops *ops = dev->netdev_ops; 9638 9639 if (ops->ndo_change_rx_flags) 9640 ops->ndo_change_rx_flags(dev, flags); 9641 } 9642 9643 int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify) 9644 { 9645 unsigned int old_flags = dev->flags; 9646 unsigned int promiscuity, flags; 9647 kuid_t uid; 9648 kgid_t gid; 9649 9650 netdev_assert_locked_ops_compat(dev); 9651 9652 promiscuity = dev->promiscuity + inc; 9653 if (promiscuity == 0) { 9654 /* 9655 * Avoid overflow. 9656 * If inc causes overflow, untouch promisc and return error. 9657 */ 9658 if (unlikely(inc > 0)) { 9659 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n"); 9660 return -EOVERFLOW; 9661 } 9662 flags = old_flags & ~IFF_PROMISC; 9663 } else { 9664 flags = old_flags | IFF_PROMISC; 9665 } 9666 WRITE_ONCE(dev->promiscuity, promiscuity); 9667 if (flags != old_flags) { 9668 WRITE_ONCE(dev->flags, flags); 9669 netdev_info(dev, "%s promiscuous mode\n", 9670 dev->flags & IFF_PROMISC ? "entered" : "left"); 9671 if (audit_enabled) { 9672 current_uid_gid(&uid, &gid); 9673 audit_log(audit_context(), GFP_ATOMIC, 9674 AUDIT_ANOM_PROMISCUOUS, 9675 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u", 9676 dev->name, (dev->flags & IFF_PROMISC), 9677 (old_flags & IFF_PROMISC), 9678 from_kuid(&init_user_ns, audit_get_loginuid(current)), 9679 from_kuid(&init_user_ns, uid), 9680 from_kgid(&init_user_ns, gid), 9681 audit_get_sessionid(current)); 9682 } 9683 9684 dev_change_rx_flags(dev, IFF_PROMISC); 9685 } 9686 if (notify) 9687 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL); 9688 return 0; 9689 } 9690 9691 int netif_set_promiscuity(struct net_device *dev, int inc) 9692 { 9693 unsigned int old_flags = dev->flags; 9694 int err; 9695 9696 err = __dev_set_promiscuity(dev, inc, true); 9697 if (err < 0) 9698 return err; 9699 if (dev->flags != old_flags) 9700 dev_set_rx_mode(dev); 9701 return err; 9702 } 9703 9704 int netif_set_allmulti(struct net_device *dev, int inc, bool notify) 9705 { 9706 unsigned int old_flags = dev->flags, old_gflags = dev->gflags; 9707 unsigned int allmulti, flags; 9708 9709 netdev_assert_locked_ops_compat(dev); 9710 9711 allmulti = dev->allmulti + inc; 9712 if (allmulti == 0) { 9713 /* 9714 * Avoid overflow. 9715 * If inc causes overflow, untouch allmulti and return error. 9716 */ 9717 if (unlikely(inc > 0)) { 9718 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n"); 9719 return -EOVERFLOW; 9720 } 9721 flags = old_flags & ~IFF_ALLMULTI; 9722 } else { 9723 flags = old_flags | IFF_ALLMULTI; 9724 } 9725 WRITE_ONCE(dev->allmulti, allmulti); 9726 if (flags != old_flags) { 9727 WRITE_ONCE(dev->flags, flags); 9728 netdev_info(dev, "%s allmulticast mode\n", 9729 dev->flags & IFF_ALLMULTI ? "entered" : "left"); 9730 dev_change_rx_flags(dev, IFF_ALLMULTI); 9731 dev_set_rx_mode(dev); 9732 if (notify) 9733 __dev_notify_flags(dev, old_flags, 9734 dev->gflags ^ old_gflags, 0, NULL); 9735 } 9736 return 0; 9737 } 9738 9739 9740 /** 9741 * netif_get_flags() - get flags reported to userspace 9742 * @dev: device 9743 * 9744 * Get the combination of flag bits exported through APIs to userspace. 9745 */ 9746 unsigned int netif_get_flags(const struct net_device *dev) 9747 { 9748 unsigned int flags; 9749 9750 flags = (READ_ONCE(dev->flags) & ~(IFF_PROMISC | 9751 IFF_ALLMULTI | 9752 IFF_RUNNING | 9753 IFF_LOWER_UP | 9754 IFF_DORMANT)) | 9755 (READ_ONCE(dev->gflags) & (IFF_PROMISC | 9756 IFF_ALLMULTI)); 9757 9758 if (netif_running(dev)) { 9759 if (netif_oper_up(dev)) 9760 flags |= IFF_RUNNING; 9761 if (netif_carrier_ok(dev)) 9762 flags |= IFF_LOWER_UP; 9763 if (netif_dormant(dev)) 9764 flags |= IFF_DORMANT; 9765 } 9766 9767 return flags; 9768 } 9769 EXPORT_SYMBOL(netif_get_flags); 9770 9771 int __dev_change_flags(struct net_device *dev, unsigned int flags, 9772 struct netlink_ext_ack *extack) 9773 { 9774 unsigned int old_flags = dev->flags; 9775 int ret; 9776 9777 netdev_assert_locked_ops_compat(dev); 9778 9779 /* 9780 * Set the flags on our device. 9781 */ 9782 9783 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP | 9784 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL | 9785 IFF_AUTOMEDIA)) | 9786 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC | 9787 IFF_ALLMULTI)); 9788 9789 /* 9790 * Load in the correct multicast list now the flags have changed. 9791 */ 9792 9793 if ((old_flags ^ flags) & IFF_MULTICAST) 9794 dev_change_rx_flags(dev, IFF_MULTICAST); 9795 9796 dev_set_rx_mode(dev); 9797 9798 /* 9799 * Have we downed the interface. We handle IFF_UP ourselves 9800 * according to user attempts to set it, rather than blindly 9801 * setting it. 9802 */ 9803 9804 ret = 0; 9805 if ((old_flags ^ flags) & IFF_UP) { 9806 if (old_flags & IFF_UP) 9807 __dev_close(dev); 9808 else 9809 ret = __dev_open(dev, extack); 9810 } 9811 9812 if ((flags ^ dev->gflags) & IFF_PROMISC) { 9813 int inc = (flags & IFF_PROMISC) ? 1 : -1; 9814 old_flags = dev->flags; 9815 9816 dev->gflags ^= IFF_PROMISC; 9817 9818 if (__dev_set_promiscuity(dev, inc, false) >= 0) 9819 if (dev->flags != old_flags) 9820 dev_set_rx_mode(dev); 9821 } 9822 9823 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI 9824 * is important. Some (broken) drivers set IFF_PROMISC, when 9825 * IFF_ALLMULTI is requested not asking us and not reporting. 9826 */ 9827 if ((flags ^ dev->gflags) & IFF_ALLMULTI) { 9828 int inc = (flags & IFF_ALLMULTI) ? 1 : -1; 9829 9830 dev->gflags ^= IFF_ALLMULTI; 9831 netif_set_allmulti(dev, inc, false); 9832 } 9833 9834 return ret; 9835 } 9836 9837 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags, 9838 unsigned int gchanges, u32 portid, 9839 const struct nlmsghdr *nlh) 9840 { 9841 unsigned int changes = dev->flags ^ old_flags; 9842 9843 netdev_assert_locked_ops_compat(dev); 9844 9845 if (gchanges) 9846 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh); 9847 9848 if (changes & IFF_UP) { 9849 if (dev->flags & IFF_UP) 9850 call_netdevice_notifiers(NETDEV_UP, dev); 9851 else 9852 call_netdevice_notifiers(NETDEV_DOWN, dev); 9853 } 9854 9855 if (dev->flags & IFF_UP && 9856 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) { 9857 struct netdev_notifier_change_info change_info = { 9858 .info = { 9859 .dev = dev, 9860 }, 9861 .flags_changed = changes, 9862 }; 9863 9864 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info); 9865 } 9866 } 9867 9868 int netif_change_flags(struct net_device *dev, unsigned int flags, 9869 struct netlink_ext_ack *extack) 9870 { 9871 int ret; 9872 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags; 9873 9874 ret = __dev_change_flags(dev, flags, extack); 9875 if (ret < 0) 9876 return ret; 9877 9878 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags); 9879 __dev_notify_flags(dev, old_flags, changes, 0, NULL); 9880 return ret; 9881 } 9882 EXPORT_SYMBOL(netif_change_flags); 9883 9884 int __netif_set_mtu(struct net_device *dev, int new_mtu) 9885 { 9886 const struct net_device_ops *ops = dev->netdev_ops; 9887 9888 if (ops->ndo_change_mtu) 9889 return ops->ndo_change_mtu(dev, new_mtu); 9890 9891 /* Pairs with all the lockless reads of dev->mtu in the stack */ 9892 WRITE_ONCE(dev->mtu, new_mtu); 9893 return 0; 9894 } 9895 EXPORT_SYMBOL_NS_GPL(__netif_set_mtu, "NETDEV_INTERNAL"); 9896 9897 int dev_validate_mtu(struct net_device *dev, int new_mtu, 9898 struct netlink_ext_ack *extack) 9899 { 9900 /* MTU must be positive, and in range */ 9901 if (new_mtu < 0 || new_mtu < dev->min_mtu) { 9902 NL_SET_ERR_MSG(extack, "mtu less than device minimum"); 9903 return -EINVAL; 9904 } 9905 9906 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) { 9907 NL_SET_ERR_MSG(extack, "mtu greater than device maximum"); 9908 return -EINVAL; 9909 } 9910 return 0; 9911 } 9912 9913 /** 9914 * netif_set_mtu_ext() - Change maximum transfer unit 9915 * @dev: device 9916 * @new_mtu: new transfer unit 9917 * @extack: netlink extended ack 9918 * 9919 * Change the maximum transfer size of the network device. 9920 * 9921 * Return: 0 on success, -errno on failure. 9922 */ 9923 int netif_set_mtu_ext(struct net_device *dev, int new_mtu, 9924 struct netlink_ext_ack *extack) 9925 { 9926 int err, orig_mtu; 9927 9928 netdev_assert_locked_ops_compat(dev); 9929 9930 if (new_mtu == dev->mtu) 9931 return 0; 9932 9933 err = dev_validate_mtu(dev, new_mtu, extack); 9934 if (err) 9935 return err; 9936 9937 if (!netif_device_present(dev)) 9938 return -ENODEV; 9939 9940 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev); 9941 err = notifier_to_errno(err); 9942 if (err) 9943 return err; 9944 9945 orig_mtu = dev->mtu; 9946 err = __netif_set_mtu(dev, new_mtu); 9947 9948 if (!err) { 9949 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9950 orig_mtu); 9951 err = notifier_to_errno(err); 9952 if (err) { 9953 /* setting mtu back and notifying everyone again, 9954 * so that they have a chance to revert changes. 9955 */ 9956 __netif_set_mtu(dev, orig_mtu); 9957 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9958 new_mtu); 9959 } 9960 } 9961 return err; 9962 } 9963 9964 int netif_set_mtu(struct net_device *dev, int new_mtu) 9965 { 9966 struct netlink_ext_ack extack; 9967 int err; 9968 9969 memset(&extack, 0, sizeof(extack)); 9970 err = netif_set_mtu_ext(dev, new_mtu, &extack); 9971 if (err && extack._msg) 9972 net_err_ratelimited("%s: %s\n", dev->name, extack._msg); 9973 return err; 9974 } 9975 EXPORT_SYMBOL(netif_set_mtu); 9976 9977 int netif_change_tx_queue_len(struct net_device *dev, unsigned long new_len) 9978 { 9979 unsigned int orig_len = dev->tx_queue_len; 9980 int res; 9981 9982 if (new_len != (unsigned int)new_len) 9983 return -ERANGE; 9984 9985 if (new_len != orig_len) { 9986 WRITE_ONCE(dev->tx_queue_len, new_len); 9987 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev); 9988 res = notifier_to_errno(res); 9989 if (res) 9990 goto err_rollback; 9991 res = dev_qdisc_change_tx_queue_len(dev); 9992 if (res) 9993 goto err_rollback; 9994 } 9995 9996 return 0; 9997 9998 err_rollback: 9999 netdev_err(dev, "refused to change device tx_queue_len\n"); 10000 WRITE_ONCE(dev->tx_queue_len, orig_len); 10001 return res; 10002 } 10003 10004 void netif_set_group(struct net_device *dev, int new_group) 10005 { 10006 dev->group = new_group; 10007 } 10008 10009 /** 10010 * netif_pre_changeaddr_notify() - Call NETDEV_PRE_CHANGEADDR. 10011 * @dev: device 10012 * @addr: new address 10013 * @extack: netlink extended ack 10014 * 10015 * Return: 0 on success, -errno on failure. 10016 */ 10017 int netif_pre_changeaddr_notify(struct net_device *dev, const char *addr, 10018 struct netlink_ext_ack *extack) 10019 { 10020 struct netdev_notifier_pre_changeaddr_info info = { 10021 .info.dev = dev, 10022 .info.extack = extack, 10023 .dev_addr = addr, 10024 }; 10025 int rc; 10026 10027 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info); 10028 return notifier_to_errno(rc); 10029 } 10030 EXPORT_SYMBOL_NS_GPL(netif_pre_changeaddr_notify, "NETDEV_INTERNAL"); 10031 10032 int netif_set_mac_address(struct net_device *dev, struct sockaddr_storage *ss, 10033 struct netlink_ext_ack *extack) 10034 { 10035 const struct net_device_ops *ops = dev->netdev_ops; 10036 int err; 10037 10038 if (!ops->ndo_set_mac_address) 10039 return -EOPNOTSUPP; 10040 if (ss->ss_family != dev->type) 10041 return -EINVAL; 10042 if (!netif_device_present(dev)) 10043 return -ENODEV; 10044 err = netif_pre_changeaddr_notify(dev, ss->__data, extack); 10045 if (err) 10046 return err; 10047 if (memcmp(dev->dev_addr, ss->__data, dev->addr_len)) { 10048 err = ops->ndo_set_mac_address(dev, ss); 10049 if (err) 10050 return err; 10051 } 10052 dev->addr_assign_type = NET_ADDR_SET; 10053 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev); 10054 add_device_randomness(dev->dev_addr, dev->addr_len); 10055 return 0; 10056 } 10057 10058 DECLARE_RWSEM(dev_addr_sem); 10059 10060 /* "sa" is a true struct sockaddr with limited "sa_data" member. */ 10061 int netif_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name) 10062 { 10063 size_t size = sizeof(sa->sa_data); 10064 struct net_device *dev; 10065 int ret = 0; 10066 10067 down_read(&dev_addr_sem); 10068 rcu_read_lock(); 10069 10070 dev = dev_get_by_name_rcu(net, dev_name); 10071 if (!dev) { 10072 ret = -ENODEV; 10073 goto unlock; 10074 } 10075 if (!dev->addr_len) 10076 memset(sa->sa_data, 0, size); 10077 else 10078 memcpy(sa->sa_data, dev->dev_addr, 10079 min_t(size_t, size, dev->addr_len)); 10080 sa->sa_family = dev->type; 10081 10082 unlock: 10083 rcu_read_unlock(); 10084 up_read(&dev_addr_sem); 10085 return ret; 10086 } 10087 EXPORT_SYMBOL_NS_GPL(netif_get_mac_address, "NETDEV_INTERNAL"); 10088 10089 int netif_change_carrier(struct net_device *dev, bool new_carrier) 10090 { 10091 const struct net_device_ops *ops = dev->netdev_ops; 10092 10093 if (!ops->ndo_change_carrier) 10094 return -EOPNOTSUPP; 10095 if (!netif_device_present(dev)) 10096 return -ENODEV; 10097 return ops->ndo_change_carrier(dev, new_carrier); 10098 } 10099 10100 /** 10101 * dev_get_phys_port_id - Get device physical port ID 10102 * @dev: device 10103 * @ppid: port ID 10104 * 10105 * Get device physical port ID 10106 */ 10107 int dev_get_phys_port_id(struct net_device *dev, 10108 struct netdev_phys_item_id *ppid) 10109 { 10110 const struct net_device_ops *ops = dev->netdev_ops; 10111 10112 if (!ops->ndo_get_phys_port_id) 10113 return -EOPNOTSUPP; 10114 return ops->ndo_get_phys_port_id(dev, ppid); 10115 } 10116 10117 /** 10118 * dev_get_phys_port_name - Get device physical port name 10119 * @dev: device 10120 * @name: port name 10121 * @len: limit of bytes to copy to name 10122 * 10123 * Get device physical port name 10124 */ 10125 int dev_get_phys_port_name(struct net_device *dev, 10126 char *name, size_t len) 10127 { 10128 const struct net_device_ops *ops = dev->netdev_ops; 10129 int err; 10130 10131 if (ops->ndo_get_phys_port_name) { 10132 err = ops->ndo_get_phys_port_name(dev, name, len); 10133 if (err != -EOPNOTSUPP) 10134 return err; 10135 } 10136 return devlink_compat_phys_port_name_get(dev, name, len); 10137 } 10138 10139 /** 10140 * netif_get_port_parent_id() - Get the device's port parent identifier 10141 * @dev: network device 10142 * @ppid: pointer to a storage for the port's parent identifier 10143 * @recurse: allow/disallow recursion to lower devices 10144 * 10145 * Get the devices's port parent identifier. 10146 * 10147 * Return: 0 on success, -errno on failure. 10148 */ 10149 int netif_get_port_parent_id(struct net_device *dev, 10150 struct netdev_phys_item_id *ppid, bool recurse) 10151 { 10152 const struct net_device_ops *ops = dev->netdev_ops; 10153 struct netdev_phys_item_id first = { }; 10154 struct net_device *lower_dev; 10155 struct list_head *iter; 10156 int err; 10157 10158 if (ops->ndo_get_port_parent_id) { 10159 err = ops->ndo_get_port_parent_id(dev, ppid); 10160 if (err != -EOPNOTSUPP) 10161 return err; 10162 } 10163 10164 err = devlink_compat_switch_id_get(dev, ppid); 10165 if (!recurse || err != -EOPNOTSUPP) 10166 return err; 10167 10168 netdev_for_each_lower_dev(dev, lower_dev, iter) { 10169 err = netif_get_port_parent_id(lower_dev, ppid, true); 10170 if (err) 10171 break; 10172 if (!first.id_len) 10173 first = *ppid; 10174 else if (memcmp(&first, ppid, sizeof(*ppid))) 10175 return -EOPNOTSUPP; 10176 } 10177 10178 return err; 10179 } 10180 EXPORT_SYMBOL(netif_get_port_parent_id); 10181 10182 /** 10183 * netdev_port_same_parent_id - Indicate if two network devices have 10184 * the same port parent identifier 10185 * @a: first network device 10186 * @b: second network device 10187 */ 10188 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b) 10189 { 10190 struct netdev_phys_item_id a_id = { }; 10191 struct netdev_phys_item_id b_id = { }; 10192 10193 if (netif_get_port_parent_id(a, &a_id, true) || 10194 netif_get_port_parent_id(b, &b_id, true)) 10195 return false; 10196 10197 return netdev_phys_item_id_same(&a_id, &b_id); 10198 } 10199 EXPORT_SYMBOL(netdev_port_same_parent_id); 10200 10201 static struct net_device *dev_get_iflink_dev(struct net_device *dev) 10202 { 10203 struct net *net; 10204 10205 ASSERT_RTNL(); 10206 10207 if (!dev->netdev_ops->ndo_get_iflink || !dev->rtnl_link_ops || 10208 !dev->rtnl_link_ops->get_link_net) 10209 return dev; 10210 10211 net = dev->rtnl_link_ops->get_link_net(dev); 10212 return __dev_get_by_index(net, dev_get_iflink(dev)); 10213 } 10214 10215 int netif_change_proto_down(struct net_device *dev, bool proto_down) 10216 { 10217 struct net_device *iflink_dev; 10218 10219 if (!dev->change_proto_down) 10220 return -EOPNOTSUPP; 10221 if (!netif_device_present(dev)) 10222 return -ENODEV; 10223 iflink_dev = dev_get_iflink_dev(dev); 10224 if (!iflink_dev) 10225 return -ENODEV; 10226 WRITE_ONCE(dev->proto_down, proto_down); 10227 if (proto_down) 10228 netif_carrier_off(dev); 10229 else if (dev == iflink_dev || netif_carrier_ok(iflink_dev)) 10230 netif_carrier_on(dev); 10231 return 0; 10232 } 10233 10234 /** 10235 * netdev_change_proto_down_reason_locked - proto down reason 10236 * 10237 * @dev: device 10238 * @mask: proto down mask 10239 * @value: proto down value 10240 */ 10241 void netdev_change_proto_down_reason_locked(struct net_device *dev, 10242 unsigned long mask, u32 value) 10243 { 10244 u32 proto_down_reason; 10245 int b; 10246 10247 if (!mask) { 10248 proto_down_reason = value; 10249 } else { 10250 proto_down_reason = dev->proto_down_reason; 10251 for_each_set_bit(b, &mask, 32) { 10252 if (value & (1 << b)) 10253 proto_down_reason |= BIT(b); 10254 else 10255 proto_down_reason &= ~BIT(b); 10256 } 10257 } 10258 WRITE_ONCE(dev->proto_down_reason, proto_down_reason); 10259 } 10260 10261 struct bpf_xdp_link { 10262 struct bpf_link link; 10263 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */ 10264 int flags; 10265 }; 10266 10267 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags) 10268 { 10269 if (flags & XDP_FLAGS_HW_MODE) 10270 return XDP_MODE_HW; 10271 if (flags & XDP_FLAGS_DRV_MODE) 10272 return XDP_MODE_DRV; 10273 if (flags & XDP_FLAGS_SKB_MODE) 10274 return XDP_MODE_SKB; 10275 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB; 10276 } 10277 10278 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode) 10279 { 10280 switch (mode) { 10281 case XDP_MODE_SKB: 10282 return generic_xdp_install; 10283 case XDP_MODE_DRV: 10284 case XDP_MODE_HW: 10285 return dev->netdev_ops->ndo_bpf; 10286 default: 10287 return NULL; 10288 } 10289 } 10290 10291 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev, 10292 enum bpf_xdp_mode mode) 10293 { 10294 return dev->xdp_state[mode].link; 10295 } 10296 10297 static struct bpf_prog *dev_xdp_prog(struct net_device *dev, 10298 enum bpf_xdp_mode mode) 10299 { 10300 struct bpf_xdp_link *link = dev_xdp_link(dev, mode); 10301 10302 if (link) 10303 return link->link.prog; 10304 return dev->xdp_state[mode].prog; 10305 } 10306 10307 u8 dev_xdp_prog_count(struct net_device *dev) 10308 { 10309 u8 count = 0; 10310 int i; 10311 10312 for (i = 0; i < __MAX_XDP_MODE; i++) 10313 if (dev->xdp_state[i].prog || dev->xdp_state[i].link) 10314 count++; 10315 return count; 10316 } 10317 EXPORT_SYMBOL_GPL(dev_xdp_prog_count); 10318 10319 u8 dev_xdp_sb_prog_count(struct net_device *dev) 10320 { 10321 u8 count = 0; 10322 int i; 10323 10324 for (i = 0; i < __MAX_XDP_MODE; i++) 10325 if (dev->xdp_state[i].prog && 10326 !dev->xdp_state[i].prog->aux->xdp_has_frags) 10327 count++; 10328 return count; 10329 } 10330 10331 int netif_xdp_propagate(struct net_device *dev, struct netdev_bpf *bpf) 10332 { 10333 if (!dev->netdev_ops->ndo_bpf) 10334 return -EOPNOTSUPP; 10335 10336 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 10337 bpf->command == XDP_SETUP_PROG && 10338 bpf->prog && !bpf->prog->aux->xdp_has_frags) { 10339 NL_SET_ERR_MSG(bpf->extack, 10340 "unable to propagate XDP to device using tcp-data-split"); 10341 return -EBUSY; 10342 } 10343 10344 if (dev_get_min_mp_channel_count(dev)) { 10345 NL_SET_ERR_MSG(bpf->extack, "unable to propagate XDP to device using memory provider"); 10346 return -EBUSY; 10347 } 10348 10349 return dev->netdev_ops->ndo_bpf(dev, bpf); 10350 } 10351 EXPORT_SYMBOL_GPL(netif_xdp_propagate); 10352 10353 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode) 10354 { 10355 struct bpf_prog *prog = dev_xdp_prog(dev, mode); 10356 10357 return prog ? prog->aux->id : 0; 10358 } 10359 10360 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode, 10361 struct bpf_xdp_link *link) 10362 { 10363 dev->xdp_state[mode].link = link; 10364 dev->xdp_state[mode].prog = NULL; 10365 } 10366 10367 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode, 10368 struct bpf_prog *prog) 10369 { 10370 dev->xdp_state[mode].link = NULL; 10371 dev->xdp_state[mode].prog = prog; 10372 } 10373 10374 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode, 10375 bpf_op_t bpf_op, struct netlink_ext_ack *extack, 10376 u32 flags, struct bpf_prog *prog) 10377 { 10378 struct netdev_bpf xdp; 10379 int err; 10380 10381 netdev_assert_locked_ops_compat(dev); 10382 10383 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 10384 prog && !prog->aux->xdp_has_frags) { 10385 NL_SET_ERR_MSG(extack, "unable to install XDP to device using tcp-data-split"); 10386 return -EBUSY; 10387 } 10388 10389 if (dev_get_min_mp_channel_count(dev)) { 10390 NL_SET_ERR_MSG(extack, "unable to install XDP to device using memory provider"); 10391 return -EBUSY; 10392 } 10393 10394 memset(&xdp, 0, sizeof(xdp)); 10395 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG; 10396 xdp.extack = extack; 10397 xdp.flags = flags; 10398 xdp.prog = prog; 10399 10400 /* Drivers assume refcnt is already incremented (i.e, prog pointer is 10401 * "moved" into driver), so they don't increment it on their own, but 10402 * they do decrement refcnt when program is detached or replaced. 10403 * Given net_device also owns link/prog, we need to bump refcnt here 10404 * to prevent drivers from underflowing it. 10405 */ 10406 if (prog) 10407 bpf_prog_inc(prog); 10408 err = bpf_op(dev, &xdp); 10409 if (err) { 10410 if (prog) 10411 bpf_prog_put(prog); 10412 return err; 10413 } 10414 10415 if (mode != XDP_MODE_HW) 10416 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog); 10417 10418 return 0; 10419 } 10420 10421 static void dev_xdp_uninstall(struct net_device *dev) 10422 { 10423 struct bpf_xdp_link *link; 10424 struct bpf_prog *prog; 10425 enum bpf_xdp_mode mode; 10426 bpf_op_t bpf_op; 10427 10428 ASSERT_RTNL(); 10429 10430 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) { 10431 prog = dev_xdp_prog(dev, mode); 10432 if (!prog) 10433 continue; 10434 10435 bpf_op = dev_xdp_bpf_op(dev, mode); 10436 if (!bpf_op) 10437 continue; 10438 10439 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 10440 10441 /* auto-detach link from net device */ 10442 link = dev_xdp_link(dev, mode); 10443 if (link) 10444 link->dev = NULL; 10445 else 10446 bpf_prog_put(prog); 10447 10448 dev_xdp_set_link(dev, mode, NULL); 10449 } 10450 } 10451 10452 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack, 10453 struct bpf_xdp_link *link, struct bpf_prog *new_prog, 10454 struct bpf_prog *old_prog, u32 flags) 10455 { 10456 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES); 10457 struct bpf_prog *cur_prog; 10458 struct net_device *upper; 10459 struct list_head *iter; 10460 enum bpf_xdp_mode mode; 10461 bpf_op_t bpf_op; 10462 int err; 10463 10464 ASSERT_RTNL(); 10465 10466 /* either link or prog attachment, never both */ 10467 if (link && (new_prog || old_prog)) 10468 return -EINVAL; 10469 /* link supports only XDP mode flags */ 10470 if (link && (flags & ~XDP_FLAGS_MODES)) { 10471 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment"); 10472 return -EINVAL; 10473 } 10474 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */ 10475 if (num_modes > 1) { 10476 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set"); 10477 return -EINVAL; 10478 } 10479 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */ 10480 if (!num_modes && dev_xdp_prog_count(dev) > 1) { 10481 NL_SET_ERR_MSG(extack, 10482 "More than one program loaded, unset mode is ambiguous"); 10483 return -EINVAL; 10484 } 10485 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */ 10486 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) { 10487 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified"); 10488 return -EINVAL; 10489 } 10490 10491 mode = dev_xdp_mode(dev, flags); 10492 /* can't replace attached link */ 10493 if (dev_xdp_link(dev, mode)) { 10494 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link"); 10495 return -EBUSY; 10496 } 10497 10498 /* don't allow if an upper device already has a program */ 10499 netdev_for_each_upper_dev_rcu(dev, upper, iter) { 10500 if (dev_xdp_prog_count(upper) > 0) { 10501 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program"); 10502 return -EEXIST; 10503 } 10504 } 10505 10506 cur_prog = dev_xdp_prog(dev, mode); 10507 /* can't replace attached prog with link */ 10508 if (link && cur_prog) { 10509 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link"); 10510 return -EBUSY; 10511 } 10512 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) { 10513 NL_SET_ERR_MSG(extack, "Active program does not match expected"); 10514 return -EEXIST; 10515 } 10516 10517 /* put effective new program into new_prog */ 10518 if (link) 10519 new_prog = link->link.prog; 10520 10521 if (new_prog) { 10522 bool offload = mode == XDP_MODE_HW; 10523 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB 10524 ? XDP_MODE_DRV : XDP_MODE_SKB; 10525 10526 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) { 10527 NL_SET_ERR_MSG(extack, "XDP program already attached"); 10528 return -EBUSY; 10529 } 10530 if (!offload && dev_xdp_prog(dev, other_mode)) { 10531 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time"); 10532 return -EEXIST; 10533 } 10534 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) { 10535 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported"); 10536 return -EINVAL; 10537 } 10538 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) { 10539 NL_SET_ERR_MSG(extack, "Program bound to different device"); 10540 return -EINVAL; 10541 } 10542 if (bpf_prog_is_dev_bound(new_prog->aux) && mode == XDP_MODE_SKB) { 10543 NL_SET_ERR_MSG(extack, "Can't attach device-bound programs in generic mode"); 10544 return -EINVAL; 10545 } 10546 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) { 10547 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device"); 10548 return -EINVAL; 10549 } 10550 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) { 10551 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device"); 10552 return -EINVAL; 10553 } 10554 } 10555 10556 /* don't call drivers if the effective program didn't change */ 10557 if (new_prog != cur_prog) { 10558 bpf_op = dev_xdp_bpf_op(dev, mode); 10559 if (!bpf_op) { 10560 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode"); 10561 return -EOPNOTSUPP; 10562 } 10563 10564 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog); 10565 if (err) 10566 return err; 10567 } 10568 10569 if (link) 10570 dev_xdp_set_link(dev, mode, link); 10571 else 10572 dev_xdp_set_prog(dev, mode, new_prog); 10573 if (cur_prog) 10574 bpf_prog_put(cur_prog); 10575 10576 return 0; 10577 } 10578 10579 static int dev_xdp_attach_link(struct net_device *dev, 10580 struct netlink_ext_ack *extack, 10581 struct bpf_xdp_link *link) 10582 { 10583 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags); 10584 } 10585 10586 static int dev_xdp_detach_link(struct net_device *dev, 10587 struct netlink_ext_ack *extack, 10588 struct bpf_xdp_link *link) 10589 { 10590 enum bpf_xdp_mode mode; 10591 bpf_op_t bpf_op; 10592 10593 ASSERT_RTNL(); 10594 10595 mode = dev_xdp_mode(dev, link->flags); 10596 if (dev_xdp_link(dev, mode) != link) 10597 return -EINVAL; 10598 10599 bpf_op = dev_xdp_bpf_op(dev, mode); 10600 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 10601 dev_xdp_set_link(dev, mode, NULL); 10602 return 0; 10603 } 10604 10605 static void bpf_xdp_link_release(struct bpf_link *link) 10606 { 10607 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10608 10609 rtnl_lock(); 10610 10611 /* if racing with net_device's tear down, xdp_link->dev might be 10612 * already NULL, in which case link was already auto-detached 10613 */ 10614 if (xdp_link->dev) { 10615 netdev_lock_ops(xdp_link->dev); 10616 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link)); 10617 netdev_unlock_ops(xdp_link->dev); 10618 xdp_link->dev = NULL; 10619 } 10620 10621 rtnl_unlock(); 10622 } 10623 10624 static int bpf_xdp_link_detach(struct bpf_link *link) 10625 { 10626 bpf_xdp_link_release(link); 10627 return 0; 10628 } 10629 10630 static void bpf_xdp_link_dealloc(struct bpf_link *link) 10631 { 10632 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10633 10634 kfree(xdp_link); 10635 } 10636 10637 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link, 10638 struct seq_file *seq) 10639 { 10640 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10641 u32 ifindex = 0; 10642 10643 rtnl_lock(); 10644 if (xdp_link->dev) 10645 ifindex = xdp_link->dev->ifindex; 10646 rtnl_unlock(); 10647 10648 seq_printf(seq, "ifindex:\t%u\n", ifindex); 10649 } 10650 10651 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link, 10652 struct bpf_link_info *info) 10653 { 10654 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10655 u32 ifindex = 0; 10656 10657 rtnl_lock(); 10658 if (xdp_link->dev) 10659 ifindex = xdp_link->dev->ifindex; 10660 rtnl_unlock(); 10661 10662 info->xdp.ifindex = ifindex; 10663 return 0; 10664 } 10665 10666 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog, 10667 struct bpf_prog *old_prog) 10668 { 10669 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10670 enum bpf_xdp_mode mode; 10671 bpf_op_t bpf_op; 10672 int err = 0; 10673 10674 rtnl_lock(); 10675 10676 /* link might have been auto-released already, so fail */ 10677 if (!xdp_link->dev) { 10678 err = -ENOLINK; 10679 goto out_unlock; 10680 } 10681 10682 if (old_prog && link->prog != old_prog) { 10683 err = -EPERM; 10684 goto out_unlock; 10685 } 10686 old_prog = link->prog; 10687 if (old_prog->type != new_prog->type || 10688 old_prog->expected_attach_type != new_prog->expected_attach_type) { 10689 err = -EINVAL; 10690 goto out_unlock; 10691 } 10692 10693 if (old_prog == new_prog) { 10694 /* no-op, don't disturb drivers */ 10695 bpf_prog_put(new_prog); 10696 goto out_unlock; 10697 } 10698 10699 netdev_lock_ops(xdp_link->dev); 10700 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags); 10701 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode); 10702 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL, 10703 xdp_link->flags, new_prog); 10704 netdev_unlock_ops(xdp_link->dev); 10705 if (err) 10706 goto out_unlock; 10707 10708 old_prog = xchg(&link->prog, new_prog); 10709 bpf_prog_put(old_prog); 10710 10711 out_unlock: 10712 rtnl_unlock(); 10713 return err; 10714 } 10715 10716 static const struct bpf_link_ops bpf_xdp_link_lops = { 10717 .release = bpf_xdp_link_release, 10718 .dealloc = bpf_xdp_link_dealloc, 10719 .detach = bpf_xdp_link_detach, 10720 .show_fdinfo = bpf_xdp_link_show_fdinfo, 10721 .fill_link_info = bpf_xdp_link_fill_link_info, 10722 .update_prog = bpf_xdp_link_update, 10723 }; 10724 10725 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog) 10726 { 10727 struct net *net = current->nsproxy->net_ns; 10728 struct bpf_link_primer link_primer; 10729 struct netlink_ext_ack extack = {}; 10730 struct bpf_xdp_link *link; 10731 struct net_device *dev; 10732 int err, fd; 10733 10734 rtnl_lock(); 10735 dev = dev_get_by_index(net, attr->link_create.target_ifindex); 10736 if (!dev) { 10737 rtnl_unlock(); 10738 return -EINVAL; 10739 } 10740 10741 link = kzalloc_obj(*link, GFP_USER); 10742 if (!link) { 10743 err = -ENOMEM; 10744 goto unlock; 10745 } 10746 10747 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog, 10748 attr->link_create.attach_type); 10749 link->dev = dev; 10750 link->flags = attr->link_create.flags; 10751 10752 err = bpf_link_prime(&link->link, &link_primer); 10753 if (err) { 10754 kfree(link); 10755 goto unlock; 10756 } 10757 10758 netdev_lock_ops(dev); 10759 err = dev_xdp_attach_link(dev, &extack, link); 10760 netdev_unlock_ops(dev); 10761 rtnl_unlock(); 10762 10763 if (err) { 10764 link->dev = NULL; 10765 bpf_link_cleanup(&link_primer); 10766 trace_bpf_xdp_link_attach_failed(extack._msg); 10767 goto out_put_dev; 10768 } 10769 10770 fd = bpf_link_settle(&link_primer); 10771 /* link itself doesn't hold dev's refcnt to not complicate shutdown */ 10772 dev_put(dev); 10773 return fd; 10774 10775 unlock: 10776 rtnl_unlock(); 10777 10778 out_put_dev: 10779 dev_put(dev); 10780 return err; 10781 } 10782 10783 /** 10784 * dev_change_xdp_fd - set or clear a bpf program for a device rx path 10785 * @dev: device 10786 * @extack: netlink extended ack 10787 * @fd: new program fd or negative value to clear 10788 * @expected_fd: old program fd that userspace expects to replace or clear 10789 * @flags: xdp-related flags 10790 * 10791 * Set or clear a bpf program for a device 10792 */ 10793 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, 10794 int fd, int expected_fd, u32 flags) 10795 { 10796 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags); 10797 struct bpf_prog *new_prog = NULL, *old_prog = NULL; 10798 int err; 10799 10800 ASSERT_RTNL(); 10801 10802 if (fd >= 0) { 10803 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP, 10804 mode != XDP_MODE_SKB); 10805 if (IS_ERR(new_prog)) 10806 return PTR_ERR(new_prog); 10807 } 10808 10809 if (expected_fd >= 0) { 10810 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP, 10811 mode != XDP_MODE_SKB); 10812 if (IS_ERR(old_prog)) { 10813 err = PTR_ERR(old_prog); 10814 old_prog = NULL; 10815 goto err_out; 10816 } 10817 } 10818 10819 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags); 10820 10821 err_out: 10822 if (err && new_prog) 10823 bpf_prog_put(new_prog); 10824 if (old_prog) 10825 bpf_prog_put(old_prog); 10826 return err; 10827 } 10828 10829 u32 dev_get_min_mp_channel_count(const struct net_device *dev) 10830 { 10831 int i; 10832 10833 netdev_assert_locked_ops_compat(dev); 10834 10835 for (i = dev->real_num_rx_queues - 1; i >= 0; i--) 10836 if (dev->_rx[i].mp_params.mp_priv) 10837 /* The channel count is the idx plus 1. */ 10838 return i + 1; 10839 10840 return 0; 10841 } 10842 10843 /** 10844 * dev_index_reserve() - allocate an ifindex in a namespace 10845 * @net: the applicable net namespace 10846 * @ifindex: requested ifindex, pass %0 to get one allocated 10847 * 10848 * Allocate a ifindex for a new device. Caller must either use the ifindex 10849 * to store the device (via list_netdevice()) or call dev_index_release() 10850 * to give the index up. 10851 * 10852 * Return: a suitable unique value for a new device interface number or -errno. 10853 */ 10854 static int dev_index_reserve(struct net *net, u32 ifindex) 10855 { 10856 int err; 10857 10858 if (ifindex > INT_MAX) { 10859 DEBUG_NET_WARN_ON_ONCE(1); 10860 return -EINVAL; 10861 } 10862 10863 if (!ifindex) 10864 err = xa_alloc_cyclic(&net->dev_by_index, &ifindex, NULL, 10865 xa_limit_31b, &net->ifindex, GFP_KERNEL); 10866 else 10867 err = xa_insert(&net->dev_by_index, ifindex, NULL, GFP_KERNEL); 10868 if (err < 0) 10869 return err; 10870 10871 return ifindex; 10872 } 10873 10874 static void dev_index_release(struct net *net, int ifindex) 10875 { 10876 /* Expect only unused indexes, unlist_netdevice() removes the used */ 10877 WARN_ON(xa_erase(&net->dev_by_index, ifindex)); 10878 } 10879 10880 static bool from_cleanup_net(void) 10881 { 10882 #ifdef CONFIG_NET_NS 10883 return current == READ_ONCE(cleanup_net_task); 10884 #else 10885 return false; 10886 #endif 10887 } 10888 10889 /* Delayed registration/unregisteration */ 10890 LIST_HEAD(net_todo_list); 10891 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq); 10892 atomic_t dev_unreg_count = ATOMIC_INIT(0); 10893 10894 static void net_set_todo(struct net_device *dev) 10895 { 10896 list_add_tail(&dev->todo_list, &net_todo_list); 10897 } 10898 10899 static netdev_features_t netdev_sync_upper_features(struct net_device *lower, 10900 struct net_device *upper, netdev_features_t features) 10901 { 10902 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10903 netdev_features_t feature; 10904 int feature_bit; 10905 10906 for_each_netdev_feature(upper_disables, feature_bit) { 10907 feature = __NETIF_F_BIT(feature_bit); 10908 if (!(upper->wanted_features & feature) 10909 && (features & feature)) { 10910 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n", 10911 &feature, upper->name); 10912 features &= ~feature; 10913 } 10914 } 10915 10916 return features; 10917 } 10918 10919 static void netdev_sync_lower_features(struct net_device *upper, 10920 struct net_device *lower, netdev_features_t features) 10921 { 10922 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10923 netdev_features_t feature; 10924 int feature_bit; 10925 10926 for_each_netdev_feature(upper_disables, feature_bit) { 10927 feature = __NETIF_F_BIT(feature_bit); 10928 if (!(features & feature) && (lower->features & feature)) { 10929 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n", 10930 &feature, lower->name); 10931 netdev_lock_ops(lower); 10932 lower->wanted_features &= ~feature; 10933 __netdev_update_features(lower); 10934 10935 if (unlikely(lower->features & feature)) 10936 netdev_WARN(upper, "failed to disable %pNF on %s!\n", 10937 &feature, lower->name); 10938 else 10939 netdev_features_change(lower); 10940 netdev_unlock_ops(lower); 10941 } 10942 } 10943 } 10944 10945 static bool netdev_has_ip_or_hw_csum(netdev_features_t features) 10946 { 10947 netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 10948 bool ip_csum = (features & ip_csum_mask) == ip_csum_mask; 10949 bool hw_csum = features & NETIF_F_HW_CSUM; 10950 10951 return ip_csum || hw_csum; 10952 } 10953 10954 static netdev_features_t netdev_fix_features(struct net_device *dev, 10955 netdev_features_t features) 10956 { 10957 /* Fix illegal checksum combinations */ 10958 if ((features & NETIF_F_HW_CSUM) && 10959 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) { 10960 netdev_warn(dev, "mixed HW and IP checksum settings.\n"); 10961 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM); 10962 } 10963 10964 /* TSO requires that SG is present as well. */ 10965 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) { 10966 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n"); 10967 features &= ~NETIF_F_ALL_TSO; 10968 } 10969 10970 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) && 10971 !(features & NETIF_F_IP_CSUM)) { 10972 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n"); 10973 features &= ~NETIF_F_TSO; 10974 features &= ~NETIF_F_TSO_ECN; 10975 } 10976 10977 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) && 10978 !(features & NETIF_F_IPV6_CSUM)) { 10979 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n"); 10980 features &= ~NETIF_F_TSO6; 10981 } 10982 10983 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */ 10984 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO)) 10985 features &= ~NETIF_F_TSO_MANGLEID; 10986 10987 /* TSO ECN requires that TSO is present as well. */ 10988 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN) 10989 features &= ~NETIF_F_TSO_ECN; 10990 10991 /* Software GSO depends on SG. */ 10992 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) { 10993 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n"); 10994 features &= ~NETIF_F_GSO; 10995 } 10996 10997 /* GSO partial features require GSO partial be set */ 10998 if ((features & dev->gso_partial_features) && 10999 !(features & NETIF_F_GSO_PARTIAL)) { 11000 netdev_dbg(dev, 11001 "Dropping partially supported GSO features since no GSO partial.\n"); 11002 features &= ~dev->gso_partial_features; 11003 } 11004 11005 if (!(features & NETIF_F_RXCSUM)) { 11006 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet 11007 * successfully merged by hardware must also have the 11008 * checksum verified by hardware. If the user does not 11009 * want to enable RXCSUM, logically, we should disable GRO_HW. 11010 */ 11011 if (features & NETIF_F_GRO_HW) { 11012 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n"); 11013 features &= ~NETIF_F_GRO_HW; 11014 } 11015 } 11016 11017 /* LRO/HW-GRO features cannot be combined with RX-FCS */ 11018 if (features & NETIF_F_RXFCS) { 11019 if (features & NETIF_F_LRO) { 11020 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n"); 11021 features &= ~NETIF_F_LRO; 11022 } 11023 11024 if (features & NETIF_F_GRO_HW) { 11025 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n"); 11026 features &= ~NETIF_F_GRO_HW; 11027 } 11028 } 11029 11030 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) { 11031 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n"); 11032 features &= ~NETIF_F_LRO; 11033 } 11034 11035 if ((features & NETIF_F_HW_TLS_TX) && !netdev_has_ip_or_hw_csum(features)) { 11036 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n"); 11037 features &= ~NETIF_F_HW_TLS_TX; 11038 } 11039 11040 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) { 11041 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n"); 11042 features &= ~NETIF_F_HW_TLS_RX; 11043 } 11044 11045 if ((features & NETIF_F_GSO_UDP_L4) && !netdev_has_ip_or_hw_csum(features)) { 11046 netdev_dbg(dev, "Dropping USO feature since no CSUM feature.\n"); 11047 features &= ~NETIF_F_GSO_UDP_L4; 11048 } 11049 11050 return features; 11051 } 11052 11053 int __netdev_update_features(struct net_device *dev) 11054 { 11055 struct net_device *upper, *lower; 11056 netdev_features_t features; 11057 struct list_head *iter; 11058 int err = -1; 11059 11060 ASSERT_RTNL(); 11061 netdev_assert_locked_ops_compat(dev); 11062 11063 features = netdev_get_wanted_features(dev); 11064 11065 if (dev->netdev_ops->ndo_fix_features) 11066 features = dev->netdev_ops->ndo_fix_features(dev, features); 11067 11068 /* driver might be less strict about feature dependencies */ 11069 features = netdev_fix_features(dev, features); 11070 11071 /* some features can't be enabled if they're off on an upper device */ 11072 netdev_for_each_upper_dev_rcu(dev, upper, iter) 11073 features = netdev_sync_upper_features(dev, upper, features); 11074 11075 if (dev->features == features) 11076 goto sync_lower; 11077 11078 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n", 11079 &dev->features, &features); 11080 11081 if (dev->netdev_ops->ndo_set_features) 11082 err = dev->netdev_ops->ndo_set_features(dev, features); 11083 else 11084 err = 0; 11085 11086 if (unlikely(err < 0)) { 11087 netdev_err(dev, 11088 "set_features() failed (%d); wanted %pNF, left %pNF\n", 11089 err, &features, &dev->features); 11090 /* return non-0 since some features might have changed and 11091 * it's better to fire a spurious notification than miss it 11092 */ 11093 return -1; 11094 } 11095 11096 sync_lower: 11097 /* some features must be disabled on lower devices when disabled 11098 * on an upper device (think: bonding master or bridge) 11099 */ 11100 netdev_for_each_lower_dev(dev, lower, iter) 11101 netdev_sync_lower_features(dev, lower, features); 11102 11103 if (!err) { 11104 netdev_features_t diff = features ^ dev->features; 11105 11106 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) { 11107 /* udp_tunnel_{get,drop}_rx_info both need 11108 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the 11109 * device, or they won't do anything. 11110 * Thus we need to update dev->features 11111 * *before* calling udp_tunnel_get_rx_info, 11112 * but *after* calling udp_tunnel_drop_rx_info. 11113 */ 11114 udp_tunnel_nic_lock(dev); 11115 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) { 11116 dev->features = features; 11117 udp_tunnel_get_rx_info(dev); 11118 } else { 11119 udp_tunnel_drop_rx_info(dev); 11120 } 11121 udp_tunnel_nic_unlock(dev); 11122 } 11123 11124 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) { 11125 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) { 11126 dev->features = features; 11127 err |= vlan_get_rx_ctag_filter_info(dev); 11128 } else { 11129 vlan_drop_rx_ctag_filter_info(dev); 11130 } 11131 } 11132 11133 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) { 11134 if (features & NETIF_F_HW_VLAN_STAG_FILTER) { 11135 dev->features = features; 11136 err |= vlan_get_rx_stag_filter_info(dev); 11137 } else { 11138 vlan_drop_rx_stag_filter_info(dev); 11139 } 11140 } 11141 11142 dev->features = features; 11143 } 11144 11145 return err < 0 ? 0 : 1; 11146 } 11147 11148 /** 11149 * netdev_update_features - recalculate device features 11150 * @dev: the device to check 11151 * 11152 * Recalculate dev->features set and send notifications if it 11153 * has changed. Should be called after driver or hardware dependent 11154 * conditions might have changed that influence the features. 11155 */ 11156 void netdev_update_features(struct net_device *dev) 11157 { 11158 if (__netdev_update_features(dev)) 11159 netdev_features_change(dev); 11160 } 11161 EXPORT_SYMBOL(netdev_update_features); 11162 11163 /** 11164 * netdev_change_features - recalculate device features 11165 * @dev: the device to check 11166 * 11167 * Recalculate dev->features set and send notifications even 11168 * if they have not changed. Should be called instead of 11169 * netdev_update_features() if also dev->vlan_features might 11170 * have changed to allow the changes to be propagated to stacked 11171 * VLAN devices. 11172 */ 11173 void netdev_change_features(struct net_device *dev) 11174 { 11175 __netdev_update_features(dev); 11176 netdev_features_change(dev); 11177 } 11178 EXPORT_SYMBOL(netdev_change_features); 11179 11180 /** 11181 * netif_stacked_transfer_operstate - transfer operstate 11182 * @rootdev: the root or lower level device to transfer state from 11183 * @dev: the device to transfer operstate to 11184 * 11185 * Transfer operational state from root to device. This is normally 11186 * called when a stacking relationship exists between the root 11187 * device and the device(a leaf device). 11188 */ 11189 void netif_stacked_transfer_operstate(const struct net_device *rootdev, 11190 struct net_device *dev) 11191 { 11192 if (rootdev->operstate == IF_OPER_DORMANT) 11193 netif_dormant_on(dev); 11194 else 11195 netif_dormant_off(dev); 11196 11197 if (rootdev->operstate == IF_OPER_TESTING) 11198 netif_testing_on(dev); 11199 else 11200 netif_testing_off(dev); 11201 11202 if (netif_carrier_ok(rootdev)) 11203 netif_carrier_on(dev); 11204 else 11205 netif_carrier_off(dev); 11206 } 11207 EXPORT_SYMBOL(netif_stacked_transfer_operstate); 11208 11209 static int netif_alloc_rx_queues(struct net_device *dev) 11210 { 11211 unsigned int i, count = dev->num_rx_queues; 11212 struct netdev_rx_queue *rx; 11213 size_t sz = count * sizeof(*rx); 11214 int err = 0; 11215 11216 BUG_ON(count < 1); 11217 11218 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11219 if (!rx) 11220 return -ENOMEM; 11221 11222 dev->_rx = rx; 11223 11224 for (i = 0; i < count; i++) { 11225 rx[i].dev = dev; 11226 11227 /* XDP RX-queue setup */ 11228 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0); 11229 if (err < 0) 11230 goto err_rxq_info; 11231 } 11232 return 0; 11233 11234 err_rxq_info: 11235 /* Rollback successful reg's and free other resources */ 11236 while (i--) 11237 xdp_rxq_info_unreg(&rx[i].xdp_rxq); 11238 kvfree(dev->_rx); 11239 dev->_rx = NULL; 11240 return err; 11241 } 11242 11243 static void netif_free_rx_queues(struct net_device *dev) 11244 { 11245 unsigned int i, count = dev->num_rx_queues; 11246 11247 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */ 11248 if (!dev->_rx) 11249 return; 11250 11251 for (i = 0; i < count; i++) 11252 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq); 11253 11254 kvfree(dev->_rx); 11255 } 11256 11257 static void netdev_init_one_queue(struct net_device *dev, 11258 struct netdev_queue *queue, void *_unused) 11259 { 11260 /* Initialize queue lock */ 11261 spin_lock_init(&queue->_xmit_lock); 11262 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type); 11263 queue->xmit_lock_owner = -1; 11264 netdev_queue_numa_node_write(queue, NUMA_NO_NODE); 11265 queue->dev = dev; 11266 #ifdef CONFIG_BQL 11267 dql_init(&queue->dql, HZ); 11268 #endif 11269 } 11270 11271 static void netif_free_tx_queues(struct net_device *dev) 11272 { 11273 kvfree(dev->_tx); 11274 } 11275 11276 static int netif_alloc_netdev_queues(struct net_device *dev) 11277 { 11278 unsigned int count = dev->num_tx_queues; 11279 struct netdev_queue *tx; 11280 size_t sz = count * sizeof(*tx); 11281 11282 if (count < 1 || count > 0xffff) 11283 return -EINVAL; 11284 11285 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11286 if (!tx) 11287 return -ENOMEM; 11288 11289 dev->_tx = tx; 11290 11291 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL); 11292 spin_lock_init(&dev->tx_global_lock); 11293 spin_lock_init(&dev->watchdog_lock); 11294 dev->watchdog_ref_held = false; 11295 return 0; 11296 } 11297 11298 void netif_tx_stop_all_queues(struct net_device *dev) 11299 { 11300 unsigned int i; 11301 11302 for (i = 0; i < dev->num_tx_queues; i++) { 11303 struct netdev_queue *txq = netdev_get_tx_queue(dev, i); 11304 11305 netif_tx_stop_queue(txq); 11306 } 11307 } 11308 EXPORT_SYMBOL(netif_tx_stop_all_queues); 11309 11310 static int netdev_do_alloc_pcpu_stats(struct net_device *dev) 11311 { 11312 void __percpu *v; 11313 11314 /* Drivers implementing ndo_get_peer_dev must support tstat 11315 * accounting, so that skb_do_redirect() can bump the dev's 11316 * RX stats upon network namespace switch. 11317 */ 11318 if (dev->netdev_ops->ndo_get_peer_dev && 11319 dev->pcpu_stat_type != NETDEV_PCPU_STAT_TSTATS) 11320 return -EOPNOTSUPP; 11321 11322 switch (dev->pcpu_stat_type) { 11323 case NETDEV_PCPU_STAT_NONE: 11324 return 0; 11325 case NETDEV_PCPU_STAT_LSTATS: 11326 v = dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats); 11327 break; 11328 case NETDEV_PCPU_STAT_TSTATS: 11329 v = dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats); 11330 break; 11331 case NETDEV_PCPU_STAT_DSTATS: 11332 v = dev->dstats = netdev_alloc_pcpu_stats(struct pcpu_dstats); 11333 break; 11334 default: 11335 return -EINVAL; 11336 } 11337 11338 return v ? 0 : -ENOMEM; 11339 } 11340 11341 static void netdev_do_free_pcpu_stats(struct net_device *dev) 11342 { 11343 switch (dev->pcpu_stat_type) { 11344 case NETDEV_PCPU_STAT_NONE: 11345 return; 11346 case NETDEV_PCPU_STAT_LSTATS: 11347 free_percpu(dev->lstats); 11348 break; 11349 case NETDEV_PCPU_STAT_TSTATS: 11350 free_percpu(dev->tstats); 11351 break; 11352 case NETDEV_PCPU_STAT_DSTATS: 11353 free_percpu(dev->dstats); 11354 break; 11355 } 11356 } 11357 11358 static void netdev_free_phy_link_topology(struct net_device *dev) 11359 { 11360 struct phy_link_topology *topo = dev->link_topo; 11361 11362 if (IS_ENABLED(CONFIG_PHYLIB) && topo) { 11363 xa_destroy(&topo->phys); 11364 kfree(topo); 11365 dev->link_topo = NULL; 11366 } 11367 } 11368 11369 /** 11370 * register_netdevice() - register a network device 11371 * @dev: device to register 11372 * 11373 * Take a prepared network device structure and make it externally accessible. 11374 * A %NETDEV_REGISTER message is sent to the netdev notifier chain. 11375 * Callers must hold the rtnl lock - you may want register_netdev() 11376 * instead of this. 11377 */ 11378 int register_netdevice(struct net_device *dev) 11379 { 11380 int ret; 11381 struct net *net = dev_net(dev); 11382 11383 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE < 11384 NETDEV_FEATURE_COUNT); 11385 BUG_ON(dev_boot_phase); 11386 ASSERT_RTNL(); 11387 11388 might_sleep(); 11389 11390 /* When net_device's are persistent, this will be fatal. */ 11391 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED); 11392 BUG_ON(!net); 11393 11394 ret = ethtool_check_ops(dev->ethtool_ops); 11395 if (ret) 11396 return ret; 11397 11398 /* rss ctx ID 0 is reserved for the default context, start from 1 */ 11399 xa_init_flags(&dev->ethtool->rss_ctx, XA_FLAGS_ALLOC1); 11400 mutex_init(&dev->ethtool->rss_lock); 11401 11402 spin_lock_init(&dev->addr_list_lock); 11403 netdev_set_addr_lockdep_class(dev); 11404 11405 ret = dev_get_valid_name(net, dev, dev->name); 11406 if (ret < 0) 11407 goto out; 11408 11409 ret = -ENOMEM; 11410 dev->name_node = netdev_name_node_head_alloc(dev); 11411 if (!dev->name_node) 11412 goto out; 11413 11414 /* Init, if this function is available */ 11415 if (dev->netdev_ops->ndo_init) { 11416 ret = dev->netdev_ops->ndo_init(dev); 11417 if (ret) { 11418 if (ret > 0) 11419 ret = -EIO; 11420 goto err_free_name; 11421 } 11422 } 11423 11424 if (((dev->hw_features | dev->features) & 11425 NETIF_F_HW_VLAN_CTAG_FILTER) && 11426 (!dev->netdev_ops->ndo_vlan_rx_add_vid || 11427 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) { 11428 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n"); 11429 ret = -EINVAL; 11430 goto err_uninit; 11431 } 11432 11433 if (netdev_need_ops_lock(dev) && 11434 dev->netdev_ops->ndo_set_rx_mode && 11435 !dev->netdev_ops->ndo_set_rx_mode_async) 11436 netdev_WARN(dev, "ops-locked drivers should use ndo_set_rx_mode_async\n"); 11437 11438 ret = netdev_do_alloc_pcpu_stats(dev); 11439 if (ret) 11440 goto err_uninit; 11441 11442 ret = dev_index_reserve(net, dev->ifindex); 11443 if (ret < 0) 11444 goto err_free_pcpu; 11445 dev->ifindex = ret; 11446 11447 /* Transfer changeable features to wanted_features and enable 11448 * software offloads (GSO and GRO). 11449 */ 11450 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF); 11451 dev->features |= NETIF_F_SOFT_FEATURES; 11452 11453 if (dev->udp_tunnel_nic_info) { 11454 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT; 11455 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT; 11456 } 11457 11458 dev->wanted_features = dev->features & dev->hw_features; 11459 11460 if (!(dev->flags & IFF_LOOPBACK)) 11461 dev->hw_features |= NETIF_F_NOCACHE_COPY; 11462 11463 /* If IPv4 TCP segmentation offload is supported we should also 11464 * allow the device to enable segmenting the frame with the option 11465 * of ignoring a static IP ID value. This doesn't enable the 11466 * feature itself but allows the user to enable it later. 11467 */ 11468 if (dev->hw_features & NETIF_F_TSO) 11469 dev->hw_features |= NETIF_F_TSO_MANGLEID; 11470 if (dev->vlan_features & NETIF_F_TSO) 11471 dev->vlan_features |= NETIF_F_TSO_MANGLEID; 11472 if (dev->mpls_features & NETIF_F_TSO) 11473 dev->mpls_features |= NETIF_F_TSO_MANGLEID; 11474 if (dev->hw_enc_features & NETIF_F_TSO) 11475 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID; 11476 11477 /* TSO_MANGLEID belongs in mangleid_features by definition */ 11478 dev->mangleid_features |= NETIF_F_TSO_MANGLEID; 11479 11480 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices. 11481 */ 11482 dev->vlan_features |= NETIF_F_HIGHDMA; 11483 11484 /* Make NETIF_F_SG inheritable to tunnel devices. 11485 */ 11486 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL; 11487 11488 /* Make NETIF_F_SG inheritable to MPLS. 11489 */ 11490 dev->mpls_features |= NETIF_F_SG; 11491 11492 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev); 11493 ret = notifier_to_errno(ret); 11494 if (ret) 11495 goto err_ifindex_release; 11496 11497 ret = netdev_register_kobject(dev); 11498 11499 netdev_lock(dev); 11500 WRITE_ONCE(dev->reg_state, ret ? NETREG_UNREGISTERED : NETREG_REGISTERED); 11501 netdev_unlock(dev); 11502 11503 if (ret) 11504 goto err_uninit_notify; 11505 11506 netdev_lock_ops(dev); 11507 __netdev_update_features(dev); 11508 netdev_unlock_ops(dev); 11509 11510 /* 11511 * Default initial state at registry is that the 11512 * device is present. 11513 */ 11514 11515 set_bit(__LINK_STATE_PRESENT, &dev->state); 11516 11517 linkwatch_init_dev(dev); 11518 11519 dev_init_scheduler(dev); 11520 11521 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL); 11522 list_netdevice(dev); 11523 11524 add_device_randomness(dev->dev_addr, dev->addr_len); 11525 11526 /* If the device has permanent device address, driver should 11527 * set dev_addr and also addr_assign_type should be set to 11528 * NET_ADDR_PERM (default value). 11529 */ 11530 if (dev->addr_assign_type == NET_ADDR_PERM) 11531 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len); 11532 11533 /* Notify protocols, that a new device appeared. */ 11534 netdev_lock_ops(dev); 11535 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev); 11536 netdev_unlock_ops(dev); 11537 ret = notifier_to_errno(ret); 11538 if (ret) { 11539 /* Expect explicit free_netdev() on failure */ 11540 dev->needs_free_netdev = false; 11541 unregister_netdevice_queue(dev, NULL); 11542 goto out; 11543 } 11544 /* 11545 * Prevent userspace races by waiting until the network 11546 * device is fully setup before sending notifications. 11547 */ 11548 netdev_uevent_add(dev); 11549 if (!(dev->rtnl_link_ops && dev->rtnl_link_initializing)) 11550 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 11551 11552 out: 11553 return ret; 11554 11555 err_uninit_notify: 11556 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11557 err_ifindex_release: 11558 dev_index_release(net, dev->ifindex); 11559 err_free_pcpu: 11560 netdev_do_free_pcpu_stats(dev); 11561 err_uninit: 11562 if (dev->netdev_ops->ndo_uninit) 11563 dev->netdev_ops->ndo_uninit(dev); 11564 if (dev->priv_destructor) 11565 dev->priv_destructor(dev); 11566 err_free_name: 11567 netdev_name_node_free(dev->name_node); 11568 goto out; 11569 } 11570 EXPORT_SYMBOL(register_netdevice); 11571 11572 /* Initialize the core of a dummy net device. 11573 * The setup steps dummy netdevs need which normal netdevs get by going 11574 * through register_netdevice(). 11575 */ 11576 static void init_dummy_netdev(struct net_device *dev) 11577 { 11578 /* make sure we BUG if trying to hit standard 11579 * register/unregister code path 11580 */ 11581 dev->reg_state = NETREG_DUMMY; 11582 11583 /* a dummy interface is started by default */ 11584 set_bit(__LINK_STATE_PRESENT, &dev->state); 11585 set_bit(__LINK_STATE_START, &dev->state); 11586 11587 /* Note : We dont allocate pcpu_refcnt for dummy devices, 11588 * because users of this 'device' dont need to change 11589 * its refcount. 11590 */ 11591 } 11592 11593 /** 11594 * register_netdev - register a network device 11595 * @dev: device to register 11596 * 11597 * Take a completed network device structure and add it to the kernel 11598 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier 11599 * chain. 0 is returned on success. A negative errno code is returned 11600 * on a failure to set up the device, or if the name is a duplicate. 11601 * 11602 * This is a wrapper around register_netdevice that takes the rtnl semaphore 11603 * and expands the device name if you passed a format string to 11604 * alloc_netdev. 11605 */ 11606 int register_netdev(struct net_device *dev) 11607 { 11608 struct net *net = dev_net(dev); 11609 int err; 11610 11611 if (rtnl_net_lock_killable(net)) 11612 return -EINTR; 11613 11614 err = register_netdevice(dev); 11615 11616 rtnl_net_unlock(net); 11617 11618 return err; 11619 } 11620 EXPORT_SYMBOL(register_netdev); 11621 11622 int netdev_refcnt_read(const struct net_device *dev) 11623 { 11624 #ifdef CONFIG_PCPU_DEV_REFCNT 11625 int i, refcnt = 0; 11626 11627 for_each_possible_cpu(i) 11628 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i); 11629 return refcnt; 11630 #else 11631 return refcount_read(&dev->dev_refcnt); 11632 #endif 11633 } 11634 EXPORT_SYMBOL(netdev_refcnt_read); 11635 11636 int netdev_unregister_timeout_secs __read_mostly = 10; 11637 11638 #define WAIT_REFS_MIN_MSECS 1 11639 #define WAIT_REFS_MAX_MSECS 250 11640 /** 11641 * netdev_wait_allrefs_any - wait until all references are gone. 11642 * @list: list of net_devices to wait on 11643 * 11644 * This is called when unregistering network devices. 11645 * 11646 * Any protocol or device that holds a reference should register 11647 * for netdevice notification, and cleanup and put back the 11648 * reference if they receive an UNREGISTER event. 11649 * We can get stuck here if buggy protocols don't correctly 11650 * call dev_put. 11651 */ 11652 static struct net_device *netdev_wait_allrefs_any(struct list_head *list) 11653 { 11654 unsigned long rebroadcast_time, warning_time; 11655 struct net_device *dev; 11656 int wait = 0; 11657 11658 rebroadcast_time = warning_time = jiffies; 11659 11660 list_for_each_entry(dev, list, todo_list) 11661 if (netdev_refcnt_read(dev) == 1) 11662 return dev; 11663 11664 while (true) { 11665 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) { 11666 rtnl_lock(); 11667 11668 /* Rebroadcast unregister notification */ 11669 list_for_each_entry(dev, list, todo_list) { 11670 struct net *net = dev_net(dev); 11671 11672 __rtnl_net_lock(net); 11673 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11674 __rtnl_net_unlock(net); 11675 } 11676 11677 __rtnl_unlock(); 11678 rcu_barrier(); 11679 rtnl_lock(); 11680 11681 list_for_each_entry(dev, list, todo_list) 11682 if (test_bit(__LINK_STATE_LINKWATCH_PENDING, 11683 &dev->state)) { 11684 /* We must not have linkwatch events 11685 * pending on unregister. If this 11686 * happens, we simply run the queue 11687 * unscheduled, resulting in a noop 11688 * for this device. 11689 */ 11690 linkwatch_run_queue(); 11691 break; 11692 } 11693 11694 __rtnl_unlock(); 11695 11696 rebroadcast_time = jiffies; 11697 } 11698 11699 rcu_barrier(); 11700 11701 if (!wait) { 11702 wait = WAIT_REFS_MIN_MSECS; 11703 } else { 11704 msleep(wait); 11705 wait = min(wait << 1, WAIT_REFS_MAX_MSECS); 11706 } 11707 11708 list_for_each_entry(dev, list, todo_list) 11709 if (netdev_refcnt_read(dev) == 1) 11710 return dev; 11711 11712 if (time_after(jiffies, warning_time + 11713 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) { 11714 list_for_each_entry(dev, list, todo_list) { 11715 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n", 11716 dev->name, netdev_refcnt_read(dev)); 11717 ref_tracker_dir_print(&dev->refcnt_tracker, 10); 11718 } 11719 11720 warning_time = jiffies; 11721 } 11722 } 11723 } 11724 11725 /* The sequence is: 11726 * 11727 * rtnl_lock(); 11728 * ... 11729 * register_netdevice(x1); 11730 * register_netdevice(x2); 11731 * ... 11732 * unregister_netdevice(y1); 11733 * unregister_netdevice(y2); 11734 * ... 11735 * rtnl_unlock(); 11736 * free_netdev(y1); 11737 * free_netdev(y2); 11738 * 11739 * We are invoked by rtnl_unlock(). 11740 * This allows us to deal with problems: 11741 * 1) We can delete sysfs objects which invoke hotplug 11742 * without deadlocking with linkwatch via keventd. 11743 * 2) Since we run with the RTNL semaphore not held, we can sleep 11744 * safely in order to wait for the netdev refcnt to drop to zero. 11745 * 11746 * We must not return until all unregister events added during 11747 * the interval the lock was held have been completed. 11748 */ 11749 void netdev_run_todo(void) 11750 { 11751 struct net_device *dev, *tmp; 11752 struct list_head list; 11753 int cnt; 11754 #ifdef CONFIG_LOCKDEP 11755 struct list_head unlink_list; 11756 11757 list_replace_init(&net_unlink_list, &unlink_list); 11758 11759 while (!list_empty(&unlink_list)) { 11760 dev = list_first_entry(&unlink_list, struct net_device, 11761 unlink_list); 11762 list_del_init(&dev->unlink_list); 11763 dev->nested_level = dev->lower_level - 1; 11764 } 11765 #endif 11766 11767 /* Snapshot list, allow later requests */ 11768 list_replace_init(&net_todo_list, &list); 11769 11770 __rtnl_unlock(); 11771 11772 /* Wait for rcu callbacks to finish before next phase */ 11773 if (!list_empty(&list)) 11774 rcu_barrier(); 11775 11776 list_for_each_entry_safe(dev, tmp, &list, todo_list) { 11777 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) { 11778 netdev_WARN(dev, "run_todo but not unregistering\n"); 11779 list_del(&dev->todo_list); 11780 continue; 11781 } 11782 11783 netdev_lock(dev); 11784 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERED); 11785 netdev_unlock(dev); 11786 linkwatch_sync_dev(dev); 11787 } 11788 11789 cnt = 0; 11790 while (!list_empty(&list)) { 11791 dev = netdev_wait_allrefs_any(&list); 11792 list_del(&dev->todo_list); 11793 11794 /* paranoia */ 11795 BUG_ON(netdev_refcnt_read(dev) != 1); 11796 BUG_ON(!list_empty(&dev->ptype_all)); 11797 BUG_ON(!list_empty(&dev->ptype_specific)); 11798 WARN_ON(rcu_access_pointer(dev->ip_ptr)); 11799 WARN_ON(rcu_access_pointer(dev->ip6_ptr)); 11800 11801 netdev_name_node_alt_flush(dev); 11802 netdev_name_node_free(dev->name_node); 11803 netdev_do_free_pcpu_stats(dev); 11804 if (dev->priv_destructor) 11805 dev->priv_destructor(dev); 11806 if (dev->needs_free_netdev) 11807 free_netdev(dev); 11808 11809 cnt++; 11810 11811 /* Free network device */ 11812 kobject_put(&dev->dev.kobj); 11813 } 11814 if (cnt && atomic_sub_and_test(cnt, &dev_unreg_count)) 11815 wake_up(&netdev_unregistering_wq); 11816 } 11817 11818 /* Collate per-cpu network dstats statistics 11819 * 11820 * Read per-cpu network statistics from dev->dstats and populate the related 11821 * fields in @s. 11822 */ 11823 static void dev_fetch_dstats(struct rtnl_link_stats64 *s, 11824 const struct pcpu_dstats __percpu *dstats) 11825 { 11826 int cpu; 11827 11828 for_each_possible_cpu(cpu) { 11829 u64 rx_packets, rx_bytes, rx_drops; 11830 u64 tx_packets, tx_bytes, tx_drops; 11831 const struct pcpu_dstats *stats; 11832 unsigned int start; 11833 11834 stats = per_cpu_ptr(dstats, cpu); 11835 do { 11836 start = u64_stats_fetch_begin(&stats->syncp); 11837 rx_packets = u64_stats_read(&stats->rx_packets); 11838 rx_bytes = u64_stats_read(&stats->rx_bytes); 11839 rx_drops = u64_stats_read(&stats->rx_drops); 11840 tx_packets = u64_stats_read(&stats->tx_packets); 11841 tx_bytes = u64_stats_read(&stats->tx_bytes); 11842 tx_drops = u64_stats_read(&stats->tx_drops); 11843 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11844 11845 s->rx_packets += rx_packets; 11846 s->rx_bytes += rx_bytes; 11847 s->rx_dropped += rx_drops; 11848 s->tx_packets += tx_packets; 11849 s->tx_bytes += tx_bytes; 11850 s->tx_dropped += tx_drops; 11851 } 11852 } 11853 11854 /* ndo_get_stats64 implementation for dtstats-based accounting. 11855 * 11856 * Populate @s from dev->stats and dev->dstats. This is used internally by the 11857 * core for NETDEV_PCPU_STAT_DSTAT-type stats collection. 11858 */ 11859 static void dev_get_dstats64(const struct net_device *dev, 11860 struct rtnl_link_stats64 *s) 11861 { 11862 netdev_stats_to_stats64(s, &dev->stats); 11863 dev_fetch_dstats(s, dev->dstats); 11864 } 11865 11866 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has 11867 * all the same fields in the same order as net_device_stats, with only 11868 * the type differing, but rtnl_link_stats64 may have additional fields 11869 * at the end for newer counters. 11870 */ 11871 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64, 11872 const struct net_device_stats *netdev_stats) 11873 { 11874 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t); 11875 const atomic_long_t *src = (atomic_long_t *)netdev_stats; 11876 u64 *dst = (u64 *)stats64; 11877 11878 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64)); 11879 for (i = 0; i < n; i++) 11880 dst[i] = (unsigned long)atomic_long_read(&src[i]); 11881 /* zero out counters that only exist in rtnl_link_stats64 */ 11882 memset((char *)stats64 + n * sizeof(u64), 0, 11883 sizeof(*stats64) - n * sizeof(u64)); 11884 } 11885 EXPORT_SYMBOL(netdev_stats_to_stats64); 11886 11887 static __cold struct net_device_core_stats __percpu *netdev_core_stats_alloc( 11888 struct net_device *dev) 11889 { 11890 struct net_device_core_stats __percpu *p; 11891 11892 p = alloc_percpu_gfp(struct net_device_core_stats, 11893 GFP_ATOMIC | __GFP_NOWARN); 11894 11895 if (p && cmpxchg(&dev->core_stats, NULL, p)) 11896 free_percpu(p); 11897 11898 /* This READ_ONCE() pairs with the cmpxchg() above */ 11899 return READ_ONCE(dev->core_stats); 11900 } 11901 11902 noinline void netdev_core_stats_inc(struct net_device *dev, u32 offset) 11903 { 11904 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11905 struct net_device_core_stats __percpu *p = READ_ONCE(dev->core_stats); 11906 unsigned long __percpu *field; 11907 11908 if (unlikely(!p)) { 11909 p = netdev_core_stats_alloc(dev); 11910 if (!p) 11911 return; 11912 } 11913 11914 field = (unsigned long __percpu *)((void __percpu *)p + offset); 11915 this_cpu_inc(*field); 11916 } 11917 EXPORT_SYMBOL_GPL(netdev_core_stats_inc); 11918 11919 /** 11920 * dev_get_stats - get network device statistics 11921 * @dev: device to get statistics from 11922 * @storage: place to store stats 11923 * 11924 * Get network statistics from device. Return @storage. 11925 * The device driver may provide its own method by setting 11926 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats; 11927 * otherwise the internal statistics structure is used. 11928 */ 11929 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev, 11930 struct rtnl_link_stats64 *storage) 11931 { 11932 const struct net_device_ops *ops = dev->netdev_ops; 11933 const struct net_device_core_stats __percpu *p; 11934 11935 /* 11936 * IPv{4,6} and udp tunnels share common stat helpers and use 11937 * different stat type (NETDEV_PCPU_STAT_TSTATS vs 11938 * NETDEV_PCPU_STAT_DSTATS). Ensure the accounting is consistent. 11939 */ 11940 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_bytes) != 11941 offsetof(struct pcpu_dstats, rx_bytes)); 11942 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_packets) != 11943 offsetof(struct pcpu_dstats, rx_packets)); 11944 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_bytes) != 11945 offsetof(struct pcpu_dstats, tx_bytes)); 11946 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_packets) != 11947 offsetof(struct pcpu_dstats, tx_packets)); 11948 11949 if (ops->ndo_get_stats64) { 11950 memset(storage, 0, sizeof(*storage)); 11951 ops->ndo_get_stats64(dev, storage); 11952 } else if (ops->ndo_get_stats) { 11953 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev)); 11954 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_TSTATS) { 11955 dev_get_tstats64(dev, storage); 11956 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_DSTATS) { 11957 dev_get_dstats64(dev, storage); 11958 } else { 11959 netdev_stats_to_stats64(storage, &dev->stats); 11960 } 11961 11962 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11963 p = READ_ONCE(dev->core_stats); 11964 if (p) { 11965 const struct net_device_core_stats *core_stats; 11966 int i; 11967 11968 for_each_possible_cpu(i) { 11969 core_stats = per_cpu_ptr(p, i); 11970 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped); 11971 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped); 11972 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler); 11973 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped); 11974 } 11975 } 11976 return storage; 11977 } 11978 EXPORT_SYMBOL(dev_get_stats); 11979 11980 /** 11981 * dev_fetch_sw_netstats - get per-cpu network device statistics 11982 * @s: place to store stats 11983 * @netstats: per-cpu network stats to read from 11984 * 11985 * Read per-cpu network statistics and populate the related fields in @s. 11986 */ 11987 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s, 11988 const struct pcpu_sw_netstats __percpu *netstats) 11989 { 11990 int cpu; 11991 11992 for_each_possible_cpu(cpu) { 11993 u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 11994 const struct pcpu_sw_netstats *stats; 11995 unsigned int start; 11996 11997 stats = per_cpu_ptr(netstats, cpu); 11998 do { 11999 start = u64_stats_fetch_begin(&stats->syncp); 12000 rx_packets = u64_stats_read(&stats->rx_packets); 12001 rx_bytes = u64_stats_read(&stats->rx_bytes); 12002 tx_packets = u64_stats_read(&stats->tx_packets); 12003 tx_bytes = u64_stats_read(&stats->tx_bytes); 12004 } while (u64_stats_fetch_retry(&stats->syncp, start)); 12005 12006 s->rx_packets += rx_packets; 12007 s->rx_bytes += rx_bytes; 12008 s->tx_packets += tx_packets; 12009 s->tx_bytes += tx_bytes; 12010 } 12011 } 12012 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats); 12013 12014 /** 12015 * dev_get_tstats64 - ndo_get_stats64 implementation 12016 * @dev: device to get statistics from 12017 * @s: place to store stats 12018 * 12019 * Populate @s from dev->stats and dev->tstats. Can be used as 12020 * ndo_get_stats64() callback. 12021 */ 12022 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s) 12023 { 12024 netdev_stats_to_stats64(s, &dev->stats); 12025 dev_fetch_sw_netstats(s, dev->tstats); 12026 } 12027 EXPORT_SYMBOL_GPL(dev_get_tstats64); 12028 12029 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev) 12030 { 12031 struct netdev_queue *queue = dev_ingress_queue(dev); 12032 12033 #ifdef CONFIG_NET_CLS_ACT 12034 if (queue) 12035 return queue; 12036 queue = kzalloc_obj(*queue); 12037 if (!queue) 12038 return NULL; 12039 netdev_init_one_queue(dev, queue, NULL); 12040 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc); 12041 RCU_INIT_POINTER(queue->qdisc_sleeping, &noop_qdisc); 12042 rcu_assign_pointer(dev->ingress_queue, queue); 12043 #endif 12044 return queue; 12045 } 12046 12047 static const struct ethtool_ops default_ethtool_ops; 12048 12049 void netdev_set_default_ethtool_ops(struct net_device *dev, 12050 const struct ethtool_ops *ops) 12051 { 12052 if (dev->ethtool_ops == &default_ethtool_ops) 12053 dev->ethtool_ops = ops; 12054 } 12055 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops); 12056 12057 /** 12058 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default 12059 * @dev: netdev to enable the IRQ coalescing on 12060 * 12061 * Sets a conservative default for SW IRQ coalescing. Users can use 12062 * sysfs attributes to override the default values. 12063 */ 12064 void netdev_sw_irq_coalesce_default_on(struct net_device *dev) 12065 { 12066 WARN_ON(dev->reg_state == NETREG_REGISTERED); 12067 12068 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) { 12069 netdev_set_gro_flush_timeout(dev, 20000); 12070 netdev_set_defer_hard_irqs(dev, 1); 12071 } 12072 } 12073 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on); 12074 12075 /** 12076 * alloc_netdev_mqs - allocate network device 12077 * @sizeof_priv: size of private data to allocate space for 12078 * @name: device name format string 12079 * @name_assign_type: origin of device name 12080 * @setup: callback to initialize device 12081 * @txqs: the number of TX subqueues to allocate 12082 * @rxqs: the number of RX subqueues to allocate 12083 * 12084 * Allocates a struct net_device with private data area for driver use 12085 * and performs basic initialization. Also allocates subqueue structs 12086 * for each queue on the device. 12087 */ 12088 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name, 12089 unsigned char name_assign_type, 12090 void (*setup)(struct net_device *), 12091 unsigned int txqs, unsigned int rxqs) 12092 { 12093 struct net_device *dev; 12094 size_t napi_config_sz; 12095 unsigned int maxqs; 12096 12097 BUG_ON(strlen(name) >= sizeof(dev->name)); 12098 12099 if (txqs < 1) { 12100 pr_err("alloc_netdev: Unable to allocate device with zero queues\n"); 12101 return NULL; 12102 } 12103 12104 if (rxqs < 1) { 12105 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n"); 12106 return NULL; 12107 } 12108 12109 maxqs = max(txqs, rxqs); 12110 12111 dev = kvzalloc_flex(*dev, priv, sizeof_priv, 12112 GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 12113 if (!dev) 12114 return NULL; 12115 12116 dev->priv_len = sizeof_priv; 12117 12118 ref_tracker_dir_init(&dev->refcnt_tracker, 128, "netdev"); 12119 #ifdef CONFIG_PCPU_DEV_REFCNT 12120 dev->pcpu_refcnt = alloc_percpu(int); 12121 if (!dev->pcpu_refcnt) 12122 goto free_dev; 12123 __dev_hold(dev); 12124 #else 12125 refcount_set(&dev->dev_refcnt, 1); 12126 #endif 12127 12128 if (dev_addr_init(dev)) 12129 goto free_pcpu; 12130 12131 dev_mc_init(dev); 12132 dev_uc_init(dev); 12133 12134 dev_net_set(dev, &init_net); 12135 12136 dev->gso_max_size = GSO_LEGACY_MAX_SIZE; 12137 dev->xdp_zc_max_segs = 1; 12138 dev->gso_max_segs = GSO_MAX_SEGS; 12139 dev->gro_max_size = GRO_LEGACY_MAX_SIZE; 12140 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE; 12141 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE; 12142 dev->tso_max_size = TSO_LEGACY_MAX_SIZE; 12143 dev->tso_max_segs = TSO_MAX_SEGS; 12144 dev->upper_level = 1; 12145 dev->lower_level = 1; 12146 #ifdef CONFIG_LOCKDEP 12147 dev->nested_level = 0; 12148 INIT_LIST_HEAD(&dev->unlink_list); 12149 #endif 12150 12151 INIT_LIST_HEAD(&dev->napi_list); 12152 INIT_LIST_HEAD(&dev->unreg_list); 12153 #ifdef CONFIG_DEBUG_NET_SMALL_RTNL 12154 INIT_LIST_HEAD(&dev->unreg_list_net); 12155 #endif 12156 INIT_LIST_HEAD(&dev->close_list); 12157 INIT_LIST_HEAD(&dev->link_watch_list); 12158 INIT_LIST_HEAD(&dev->adj_list.upper); 12159 INIT_LIST_HEAD(&dev->adj_list.lower); 12160 INIT_LIST_HEAD(&dev->ptype_all); 12161 INIT_LIST_HEAD(&dev->ptype_specific); 12162 INIT_LIST_HEAD(&dev->net_notifier_list); 12163 INIT_LIST_HEAD(&dev->work_node); 12164 #ifdef CONFIG_NET_SCHED 12165 hash_init(dev->qdisc_hash); 12166 #endif 12167 12168 mutex_init(&dev->lock); 12169 netif_rx_mode_init(dev); 12170 12171 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 12172 setup(dev); 12173 12174 if (!dev->tx_queue_len) { 12175 dev->priv_flags |= IFF_NO_QUEUE; 12176 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN; 12177 } 12178 12179 dev->num_tx_queues = txqs; 12180 dev->real_num_tx_queues = txqs; 12181 if (netif_alloc_netdev_queues(dev)) 12182 goto free_all; 12183 12184 dev->num_rx_queues = rxqs; 12185 dev->real_num_rx_queues = rxqs; 12186 if (netif_alloc_rx_queues(dev)) 12187 goto free_all; 12188 dev->ethtool = kzalloc_obj(*dev->ethtool, GFP_KERNEL_ACCOUNT); 12189 if (!dev->ethtool) 12190 goto free_all; 12191 12192 dev->cfg = kzalloc_obj(*dev->cfg, GFP_KERNEL_ACCOUNT); 12193 if (!dev->cfg) 12194 goto free_all; 12195 dev->cfg_pending = dev->cfg; 12196 12197 dev->num_napi_configs = maxqs; 12198 napi_config_sz = array_size(maxqs, sizeof(*dev->napi_config)); 12199 dev->napi_config = kvzalloc(napi_config_sz, GFP_KERNEL_ACCOUNT); 12200 if (!dev->napi_config) 12201 goto free_all; 12202 12203 strscpy(dev->name, name); 12204 dev->name_assign_type = name_assign_type; 12205 dev->group = INIT_NETDEV_GROUP; 12206 if (!dev->ethtool_ops) 12207 dev->ethtool_ops = &default_ethtool_ops; 12208 12209 nf_hook_netdev_init(dev); 12210 12211 return dev; 12212 12213 free_all: 12214 free_netdev(dev); 12215 return NULL; 12216 12217 free_pcpu: 12218 #ifdef CONFIG_PCPU_DEV_REFCNT 12219 free_percpu(dev->pcpu_refcnt); 12220 free_dev: 12221 #endif 12222 kvfree(dev); 12223 return NULL; 12224 } 12225 EXPORT_SYMBOL(alloc_netdev_mqs); 12226 12227 static void netdev_napi_exit(struct net_device *dev) 12228 { 12229 if (!list_empty(&dev->napi_list)) { 12230 struct napi_struct *p, *n; 12231 12232 netdev_lock(dev); 12233 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list) 12234 __netif_napi_del_locked(p); 12235 netdev_unlock(dev); 12236 12237 synchronize_net(); 12238 } 12239 12240 kvfree(dev->napi_config); 12241 } 12242 12243 /** 12244 * free_netdev - free network device 12245 * @dev: device 12246 * 12247 * This function does the last stage of destroying an allocated device 12248 * interface. The reference to the device object is released. If this 12249 * is the last reference then it will be freed.Must be called in process 12250 * context. 12251 */ 12252 void free_netdev(struct net_device *dev) 12253 { 12254 might_sleep(); 12255 12256 /* When called immediately after register_netdevice() failed the unwind 12257 * handling may still be dismantling the device. Handle that case by 12258 * deferring the free. 12259 */ 12260 if (dev->reg_state == NETREG_UNREGISTERING) { 12261 ASSERT_RTNL(); 12262 dev->needs_free_netdev = true; 12263 return; 12264 } 12265 12266 WARN_ON(dev->cfg != dev->cfg_pending); 12267 kfree(dev->cfg); 12268 kfree(dev->ethtool); 12269 netif_free_tx_queues(dev); 12270 netif_free_rx_queues(dev); 12271 12272 kfree(rcu_dereference_protected(dev->ingress_queue, 1)); 12273 12274 __hw_addr_flush(&dev->rx_mode_addr_cache); 12275 12276 /* Flush device addresses */ 12277 dev_addr_flush(dev); 12278 12279 netdev_napi_exit(dev); 12280 12281 netif_del_cpu_rmap(dev); 12282 12283 ref_tracker_dir_exit(&dev->refcnt_tracker); 12284 #ifdef CONFIG_PCPU_DEV_REFCNT 12285 free_percpu(dev->pcpu_refcnt); 12286 dev->pcpu_refcnt = NULL; 12287 #endif 12288 free_percpu(dev->core_stats); 12289 dev->core_stats = NULL; 12290 free_percpu(dev->xdp_bulkq); 12291 dev->xdp_bulkq = NULL; 12292 12293 netdev_free_phy_link_topology(dev); 12294 12295 mutex_destroy(&dev->lock); 12296 12297 /* Compatibility with error handling in drivers */ 12298 if (dev->reg_state == NETREG_UNINITIALIZED || 12299 dev->reg_state == NETREG_DUMMY) { 12300 kvfree(dev); 12301 return; 12302 } 12303 12304 BUG_ON(dev->reg_state != NETREG_UNREGISTERED); 12305 WRITE_ONCE(dev->reg_state, NETREG_RELEASED); 12306 12307 /* will free via device release */ 12308 put_device(&dev->dev); 12309 } 12310 EXPORT_SYMBOL(free_netdev); 12311 12312 /** 12313 * alloc_netdev_dummy - Allocate and initialize a dummy net device. 12314 * @sizeof_priv: size of private data to allocate space for 12315 * 12316 * Return: the allocated net_device on success, NULL otherwise 12317 */ 12318 struct net_device *alloc_netdev_dummy(int sizeof_priv) 12319 { 12320 return alloc_netdev(sizeof_priv, "dummy#", NET_NAME_UNKNOWN, 12321 init_dummy_netdev); 12322 } 12323 EXPORT_SYMBOL_GPL(alloc_netdev_dummy); 12324 12325 /** 12326 * synchronize_net - Synchronize with packet receive processing 12327 * 12328 * Wait for packets currently being received to be done. 12329 * Does not block later packets from starting. 12330 */ 12331 void synchronize_net(void) 12332 { 12333 might_sleep(); 12334 if (from_cleanup_net() || rtnl_is_locked()) 12335 synchronize_rcu_expedited(); 12336 else 12337 synchronize_rcu(); 12338 } 12339 EXPORT_SYMBOL(synchronize_net); 12340 12341 static void netdev_rss_contexts_free(struct net_device *dev) 12342 { 12343 struct ethtool_rxfh_context *ctx; 12344 unsigned long context; 12345 12346 mutex_lock(&dev->ethtool->rss_lock); 12347 xa_for_each(&dev->ethtool->rss_ctx, context, ctx) { 12348 xa_erase(&dev->ethtool->rss_ctx, context); 12349 dev->ethtool_ops->remove_rxfh_context(dev, ctx, context, NULL); 12350 kfree(ctx); 12351 } 12352 xa_destroy(&dev->ethtool->rss_ctx); 12353 mutex_unlock(&dev->ethtool->rss_lock); 12354 } 12355 12356 /** 12357 * unregister_netdevice_queue - remove device from the kernel 12358 * @dev: device 12359 * @head: list 12360 * 12361 * This function shuts down a device interface and removes it 12362 * from the kernel tables. 12363 * If head not NULL, device is queued to be unregistered later. 12364 * 12365 * Callers must hold the rtnl semaphore. You may want 12366 * unregister_netdev() instead of this. 12367 */ 12368 12369 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head) 12370 { 12371 ASSERT_RTNL(); 12372 12373 #ifdef CONFIG_DEBUG_NET_SMALL_RTNL 12374 DEBUG_NET_WARN_ON_ONCE(!list_empty(&dev->unreg_list_net)); 12375 #endif 12376 12377 if (head) { 12378 list_move_tail(&dev->unreg_list, head); 12379 } else { 12380 LIST_HEAD(single); 12381 12382 list_add(&dev->unreg_list, &single); 12383 unregister_netdevice_many(&single); 12384 } 12385 } 12386 EXPORT_SYMBOL(unregister_netdevice_queue); 12387 12388 static void dev_memory_provider_uninstall(struct net_device *dev) 12389 { 12390 unsigned int i; 12391 12392 for (i = 0; i < dev->real_num_rx_queues; i++) { 12393 struct netdev_rx_queue *rxq = &dev->_rx[i]; 12394 12395 __netif_mp_uninstall_rxq(rxq, &rxq->mp_params); 12396 } 12397 } 12398 12399 /* devices must be UP and netdev_lock()'d */ 12400 static void netif_close_many_and_unlock(struct list_head *close_head) 12401 { 12402 struct net_device *dev, *tmp; 12403 12404 netif_close_many(close_head, false); 12405 12406 /* ... now unlock them */ 12407 list_for_each_entry_safe(dev, tmp, close_head, close_list) { 12408 netdev_unlock(dev); 12409 list_del_init(&dev->close_list); 12410 } 12411 } 12412 12413 static void netif_close_many_and_unlock_cond(struct list_head *close_head) 12414 { 12415 #ifdef CONFIG_LOCKDEP 12416 /* We can only track up to MAX_LOCK_DEPTH locks per task. 12417 * 12418 * Reserve half the available slots for additional locks possibly 12419 * taken by notifiers and (soft)irqs. 12420 */ 12421 unsigned int limit = MAX_LOCK_DEPTH / 2; 12422 12423 if (lockdep_depth(current) > limit) 12424 netif_close_many_and_unlock(close_head); 12425 #endif 12426 } 12427 12428 bool unregister_netdevice_queued(const struct net_device *dev) 12429 { 12430 ASSERT_RTNL(); 12431 return !list_empty(&dev->unreg_list); 12432 } 12433 12434 void unregister_netdevice_many_notify(struct list_head *head, 12435 u32 portid, const struct nlmsghdr *nlh) 12436 { 12437 struct net_device *dev, *tmp; 12438 LIST_HEAD(close_head); 12439 int cnt = 0; 12440 12441 BUG_ON(dev_boot_phase); 12442 ASSERT_RTNL(); 12443 12444 if (list_empty(head)) 12445 return; 12446 12447 list_for_each_entry_safe(dev, tmp, head, unreg_list) { 12448 /* Some devices call without registering 12449 * for initialization unwind. Remove those 12450 * devices and proceed with the remaining. 12451 */ 12452 if (dev->reg_state == NETREG_UNINITIALIZED) { 12453 pr_debug("unregister_netdevice: device %s/%p never was registered\n", 12454 dev->name, dev); 12455 12456 WARN_ON(1); 12457 list_del(&dev->unreg_list); 12458 continue; 12459 } 12460 dev->dismantle = true; 12461 BUG_ON(dev->reg_state != NETREG_REGISTERED); 12462 } 12463 12464 /* If device is running, close it first. Start with ops locked... */ 12465 list_for_each_entry(dev, head, unreg_list) { 12466 if (!(dev->flags & IFF_UP)) 12467 continue; 12468 if (netdev_need_ops_lock(dev)) { 12469 list_add_tail(&dev->close_list, &close_head); 12470 netdev_lock(dev); 12471 } 12472 netif_close_many_and_unlock_cond(&close_head); 12473 } 12474 netif_close_many_and_unlock(&close_head); 12475 /* ... now go over the rest. */ 12476 list_for_each_entry(dev, head, unreg_list) { 12477 if (!netdev_need_ops_lock(dev)) 12478 list_add_tail(&dev->close_list, &close_head); 12479 } 12480 netif_close_many(&close_head, true); 12481 12482 list_for_each_entry(dev, head, unreg_list) { 12483 /* And unlink it from device chain. */ 12484 unlist_netdevice(dev); 12485 netdev_lock(dev); 12486 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERING); 12487 netdev_unlock(dev); 12488 } 12489 flush_all_backlogs(); 12490 12491 synchronize_net(); 12492 12493 list_for_each_entry(dev, head, unreg_list) { 12494 struct sk_buff *skb = NULL; 12495 12496 /* Shutdown queueing discipline. */ 12497 netdev_lock_ops(dev); 12498 dev_shutdown(dev); 12499 dev_tcx_uninstall(dev); 12500 dev_xdp_uninstall(dev); 12501 dev_memory_provider_uninstall(dev); 12502 netdev_work_cancel_all(dev); 12503 netdev_unlock_ops(dev); 12504 bpf_dev_bound_netdev_unregister(dev); 12505 12506 netdev_offload_xstats_disable_all(dev); 12507 12508 /* Notify protocols, that we are about to destroy 12509 * this device. They should clean all the things. 12510 */ 12511 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 12512 12513 if (!(dev->rtnl_link_ops && dev->rtnl_link_initializing)) 12514 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0, 12515 GFP_KERNEL, NULL, 0, 12516 portid, nlh); 12517 12518 /* 12519 * Flush the unicast and multicast chains 12520 */ 12521 dev_uc_flush(dev); 12522 dev_mc_flush(dev); 12523 12524 12525 netdev_rss_contexts_free(dev); 12526 12527 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 12528 12529 if (dev->netdev_ops->ndo_uninit) 12530 dev->netdev_ops->ndo_uninit(dev); 12531 12532 mutex_destroy(&dev->ethtool->rss_lock); 12533 12534 net_shaper_flush_netdev(dev); 12535 12536 if (skb) 12537 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh); 12538 12539 /* Notifier chain MUST detach us all upper devices. */ 12540 WARN_ON(netdev_has_any_upper_dev(dev)); 12541 WARN_ON(netdev_has_any_lower_dev(dev)); 12542 12543 /* Remove entries from kobject tree */ 12544 netdev_unregister_kobject(dev); 12545 #ifdef CONFIG_XPS 12546 /* Remove XPS queueing entries */ 12547 netif_reset_xps_queues_gt(dev, 0); 12548 #endif 12549 } 12550 12551 synchronize_net(); 12552 12553 list_for_each_entry(dev, head, unreg_list) { 12554 #ifdef CONFIG_DEBUG_NET_SMALL_RTNL 12555 struct net *net = dev_net(dev); 12556 12557 /* spin_lock() can be moved outside of the loop 12558 * once the per-netns RTNL conversion completes. 12559 */ 12560 spin_lock(&net->dev_unreg_lock); 12561 list_del(&dev->unreg_list_net); 12562 spin_unlock(&net->dev_unreg_lock); 12563 #endif 12564 netdev_put(dev, &dev->dev_registered_tracker); 12565 net_set_todo(dev); 12566 cnt++; 12567 } 12568 atomic_add(cnt, &dev_unreg_count); 12569 12570 list_del(head); 12571 } 12572 12573 /** 12574 * unregister_netdevice_many - unregister many devices 12575 * @head: list of devices 12576 * 12577 * Note: As most callers use a stack allocated list_head, 12578 * we force a list_del() to make sure stack won't be corrupted later. 12579 */ 12580 void unregister_netdevice_many(struct list_head *head) 12581 { 12582 unregister_netdevice_many_notify(head, 0, NULL); 12583 } 12584 EXPORT_SYMBOL(unregister_netdevice_many); 12585 12586 #ifdef CONFIG_DEBUG_NET_SMALL_RTNL 12587 void unregister_netdevice_queue_net(struct net *net, struct net_device *dev, 12588 struct list_head *head) 12589 { 12590 netdev_lock(dev); 12591 12592 if (net_eq(dev_net(dev), net)) { 12593 netdev_unlock(dev); 12594 unregister_netdevice_queue(dev, head); 12595 return; 12596 } 12597 12598 net = dev_net(dev); 12599 12600 spin_lock(&net->dev_unreg_lock); 12601 12602 DEBUG_NET_WARN_ON_ONCE(!list_empty(&dev->unreg_list)); 12603 DEBUG_NET_WARN_ON_ONCE(!list_empty(&dev->unreg_list_net)); 12604 12605 list_add_tail(&dev->unreg_list_net, &net->dev_unreg_head); 12606 rtnl_net_queue_work(net); 12607 12608 spin_unlock(&net->dev_unreg_lock); 12609 12610 netdev_unlock(dev); 12611 } 12612 EXPORT_SYMBOL(unregister_netdevice_queue_net); 12613 12614 void unregister_netdevice_queue_many_net(struct net *net, struct list_head *head) 12615 { 12616 struct net_device *dev, *tmp; 12617 12618 spin_lock(&net->dev_unreg_lock); 12619 list_for_each_entry_safe(dev, tmp, head, unreg_list) { 12620 /* Once all cross-netns unregister_netdevice_queue() is 12621 * converted to _net() (or for debugging), remove this check. 12622 */ 12623 if (!net_eq(dev_net(dev), net)) 12624 continue; 12625 12626 DEBUG_NET_WARN_ONCE(!net_eq(dev_net(dev), net), 12627 "%s was unregistered from a different netns.\n", 12628 dev->name); 12629 12630 list_del_init(&dev->unreg_list); 12631 list_move_tail(&dev->unreg_list_net, &net->dev_unreg_head); 12632 } 12633 spin_unlock(&net->dev_unreg_lock); 12634 } 12635 12636 static void unregister_netdevice_move_net(struct net *net_old, 12637 struct net *net, 12638 struct net_device *dev) 12639 { 12640 if (net_old > net) { 12641 spin_lock(&net->dev_unreg_lock); 12642 spin_lock_nested(&net_old->dev_unreg_lock, SINGLE_DEPTH_NESTING); 12643 } else { 12644 spin_lock(&net_old->dev_unreg_lock); 12645 spin_lock_nested(&net->dev_unreg_lock, SINGLE_DEPTH_NESTING); 12646 } 12647 12648 if (!list_empty(&dev->unreg_list_net)) { 12649 list_del(&dev->unreg_list_net); 12650 list_add_tail(&dev->unreg_list_net, &net->dev_unreg_head); 12651 } 12652 12653 spin_unlock(&net_old->dev_unreg_lock); 12654 spin_unlock(&net->dev_unreg_lock); 12655 } 12656 12657 void unregister_netdevice_many_net(struct net *net) 12658 { 12659 struct net_device *dev, *tmp; 12660 LIST_HEAD(unreg_head_net); 12661 LIST_HEAD(unreg_head); 12662 12663 spin_lock(&net->dev_unreg_lock); 12664 list_splice_init(&net->dev_unreg_head, &unreg_head_net); 12665 spin_unlock(&net->dev_unreg_lock); 12666 12667 list_for_each_entry_safe(dev, tmp, &unreg_head_net, unreg_list_net) { 12668 list_del_init(&dev->unreg_list_net); 12669 list_add_tail(&dev->unreg_list, &unreg_head); 12670 } 12671 12672 unregister_netdevice_many(&unreg_head); 12673 } 12674 #endif 12675 12676 /** 12677 * unregister_netdev - remove device from the kernel 12678 * @dev: device 12679 * 12680 * This function shuts down a device interface and removes it 12681 * from the kernel tables. 12682 * 12683 * This is just a wrapper for unregister_netdevice that takes 12684 * the rtnl semaphore. In general you want to use this and not 12685 * unregister_netdevice. 12686 */ 12687 void unregister_netdev(struct net_device *dev) 12688 { 12689 rtnl_net_dev_lock(dev); 12690 unregister_netdevice(dev); 12691 rtnl_net_dev_unlock(dev); 12692 } 12693 EXPORT_SYMBOL(unregister_netdev); 12694 12695 int __dev_change_net_namespace(struct net_device *dev, struct net *net, 12696 const char *pat, int new_ifindex, 12697 struct netlink_ext_ack *extack) 12698 { 12699 struct netdev_name_node *name_node; 12700 struct net *net_old = dev_net(dev); 12701 char new_name[IFNAMSIZ] = {}; 12702 int err, new_nsid; 12703 12704 ASSERT_RTNL(); 12705 12706 /* Don't allow namespace local devices to be moved. */ 12707 err = -EINVAL; 12708 if (dev->netns_immutable) { 12709 NL_SET_ERR_MSG(extack, "The interface netns is immutable"); 12710 goto out; 12711 } 12712 12713 /* Ensure the device has been registered */ 12714 if (dev->reg_state != NETREG_REGISTERED) { 12715 NL_SET_ERR_MSG(extack, "The interface isn't registered"); 12716 goto out; 12717 } 12718 12719 /* Get out if there is nothing todo */ 12720 err = 0; 12721 if (net_eq(net_old, net)) 12722 goto out; 12723 12724 /* Pick the destination device name, and ensure 12725 * we can use it in the destination network namespace. 12726 */ 12727 err = -EEXIST; 12728 if (netdev_name_in_use(net, dev->name)) { 12729 /* We get here if we can't use the current device name */ 12730 if (!pat) { 12731 NL_SET_ERR_MSG(extack, 12732 "An interface with the same name exists in the target netns"); 12733 goto out; 12734 } 12735 err = dev_prep_valid_name(net, dev, pat, new_name, EEXIST); 12736 if (err < 0) { 12737 NL_SET_ERR_MSG_FMT(extack, 12738 "Unable to use '%s' for the new interface name in the target netns", 12739 pat); 12740 goto out; 12741 } 12742 } 12743 /* Check that none of the altnames conflicts. */ 12744 err = -EEXIST; 12745 netdev_for_each_altname(dev, name_node) { 12746 if (netdev_name_in_use(net, name_node->name)) { 12747 NL_SET_ERR_MSG_FMT(extack, 12748 "An interface with the altname %s exists in the target netns", 12749 name_node->name); 12750 goto out; 12751 } 12752 } 12753 12754 /* Check that new_ifindex isn't used yet. */ 12755 if (new_ifindex) { 12756 err = dev_index_reserve(net, new_ifindex); 12757 if (err < 0) { 12758 NL_SET_ERR_MSG_FMT(extack, 12759 "The ifindex %d is not available in the target netns", 12760 new_ifindex); 12761 goto out; 12762 } 12763 } else { 12764 /* If there is an ifindex conflict assign a new one */ 12765 err = dev_index_reserve(net, dev->ifindex); 12766 if (err == -EBUSY) 12767 err = dev_index_reserve(net, 0); 12768 if (err < 0) { 12769 NL_SET_ERR_MSG(extack, 12770 "Unable to allocate a new ifindex in the target netns"); 12771 goto out; 12772 } 12773 new_ifindex = err; 12774 } 12775 12776 /* 12777 * And now a mini version of register_netdevice unregister_netdevice. 12778 */ 12779 12780 netdev_lock_ops(dev); 12781 /* If device is running close it first. */ 12782 netif_close(dev); 12783 /* And unlink it from device chain */ 12784 unlist_netdevice(dev); 12785 12786 if (!netdev_need_ops_lock(dev)) 12787 netdev_lock(dev); 12788 dev->moving_ns = true; 12789 netdev_unlock(dev); 12790 12791 synchronize_net(); 12792 12793 /* Shutdown queueing discipline. */ 12794 netdev_lock_ops(dev); 12795 dev_shutdown(dev); 12796 netdev_unlock_ops(dev); 12797 12798 /* Notify protocols, that we are about to destroy 12799 * this device. They should clean all the things. 12800 * 12801 * Note that dev->reg_state stays at NETREG_REGISTERED. 12802 * This is wanted because this way 8021q and macvlan know 12803 * the device is just moving and can keep their slaves up. 12804 */ 12805 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 12806 rcu_barrier(); 12807 12808 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL); 12809 12810 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid, 12811 new_ifindex); 12812 12813 /* 12814 * Flush the unicast and multicast chains 12815 */ 12816 dev_uc_flush(dev); 12817 dev_mc_flush(dev); 12818 12819 /* Send a netdev-removed uevent to the old namespace */ 12820 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE); 12821 netdev_adjacent_del_links(dev); 12822 12823 /* Move per-net netdevice notifiers that are following the netdevice */ 12824 move_netdevice_notifiers_dev_net(dev, net); 12825 12826 /* Actually switch the network namespace */ 12827 netdev_lock(dev); 12828 dev_net_set(dev, net); 12829 netdev_unlock(dev); 12830 dev->ifindex = new_ifindex; 12831 12832 #ifdef CONFIG_DEBUG_NET_SMALL_RTNL 12833 unregister_netdevice_move_net(net_old, net, dev); 12834 #endif 12835 12836 if (new_name[0]) { 12837 /* Rename the netdev to prepared name */ 12838 write_seqlock_bh(&netdev_rename_lock); 12839 strscpy(dev->name, new_name, IFNAMSIZ); 12840 write_sequnlock_bh(&netdev_rename_lock); 12841 } 12842 12843 /* Fixup kobjects */ 12844 dev_set_uevent_suppress(&dev->dev, 1); 12845 err = device_rename(&dev->dev, dev->name); 12846 dev_set_uevent_suppress(&dev->dev, 0); 12847 WARN_ON(err); 12848 12849 /* Send a netdev-add uevent to the new namespace */ 12850 kobject_uevent(&dev->dev.kobj, KOBJ_ADD); 12851 netdev_adjacent_add_links(dev); 12852 12853 /* Adapt owner in case owning user namespace of target network 12854 * namespace is different from the original one. 12855 */ 12856 err = netdev_change_owner(dev, net_old, net); 12857 WARN_ON(err); 12858 12859 netdev_lock(dev); 12860 dev->moving_ns = false; 12861 if (!netdev_need_ops_lock(dev)) 12862 netdev_unlock(dev); 12863 12864 /* Add the device back in the hashes */ 12865 list_netdevice(dev); 12866 /* Notify protocols, that a new device appeared. */ 12867 call_netdevice_notifiers(NETDEV_REGISTER, dev); 12868 netdev_unlock_ops(dev); 12869 12870 /* 12871 * Prevent userspace races by waiting until the network 12872 * device is fully setup before sending notifications. 12873 */ 12874 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 12875 12876 synchronize_net(); 12877 err = 0; 12878 out: 12879 return err; 12880 } 12881 12882 static int dev_cpu_dead(unsigned int oldcpu) 12883 { 12884 struct sk_buff **list_skb; 12885 struct sk_buff *skb; 12886 unsigned int cpu; 12887 struct softnet_data *sd, *oldsd, *remsd = NULL; 12888 12889 local_irq_disable(); 12890 cpu = smp_processor_id(); 12891 sd = &per_cpu(softnet_data, cpu); 12892 oldsd = &per_cpu(softnet_data, oldcpu); 12893 12894 /* Find end of our completion_queue. */ 12895 list_skb = &sd->completion_queue; 12896 while (*list_skb) 12897 list_skb = &(*list_skb)->next; 12898 /* Append completion queue from offline CPU. */ 12899 *list_skb = oldsd->completion_queue; 12900 oldsd->completion_queue = NULL; 12901 12902 /* Append output queue from offline CPU. */ 12903 if (oldsd->output_queue) { 12904 *sd->output_queue_tailp = oldsd->output_queue; 12905 sd->output_queue_tailp = oldsd->output_queue_tailp; 12906 oldsd->output_queue = NULL; 12907 oldsd->output_queue_tailp = &oldsd->output_queue; 12908 } 12909 /* Append NAPI poll list from offline CPU, with one exception : 12910 * process_backlog() must be called by cpu owning percpu backlog. 12911 * We properly handle process_queue & input_pkt_queue later. 12912 */ 12913 while (!list_empty(&oldsd->poll_list)) { 12914 struct napi_struct *napi = list_first_entry(&oldsd->poll_list, 12915 struct napi_struct, 12916 poll_list); 12917 12918 list_del_init(&napi->poll_list); 12919 if (napi->poll == process_backlog) 12920 napi->state &= NAPIF_STATE_THREADED; 12921 else 12922 ____napi_schedule(sd, napi); 12923 } 12924 12925 raise_softirq_irqoff(NET_TX_SOFTIRQ); 12926 local_irq_enable(); 12927 12928 if (!use_backlog_threads()) { 12929 #ifdef CONFIG_RPS 12930 remsd = oldsd->rps_ipi_list; 12931 oldsd->rps_ipi_list = NULL; 12932 #endif 12933 /* send out pending IPI's on offline CPU */ 12934 net_rps_send_ipi(remsd); 12935 } 12936 12937 /* Process offline CPU's input_pkt_queue */ 12938 while ((skb = __skb_dequeue(&oldsd->process_queue))) { 12939 netif_rx(skb); 12940 rps_input_queue_head_incr(oldsd); 12941 } 12942 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) { 12943 netif_rx(skb); 12944 rps_input_queue_head_incr(oldsd); 12945 } 12946 12947 return 0; 12948 } 12949 12950 /** 12951 * netdev_increment_features - increment feature set by one 12952 * @all: current feature set 12953 * @one: new feature set 12954 * @mask: mask feature set 12955 * 12956 * Computes a new feature set after adding a device with feature set 12957 * @one to the master device with current feature set @all. Will not 12958 * enable anything that is off in @mask. Returns the new feature set. 12959 */ 12960 netdev_features_t netdev_increment_features(netdev_features_t all, 12961 netdev_features_t one, netdev_features_t mask) 12962 { 12963 if (mask & NETIF_F_HW_CSUM) 12964 mask |= NETIF_F_CSUM_MASK; 12965 mask |= NETIF_F_VLAN_CHALLENGED; 12966 12967 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask; 12968 all &= one | ~NETIF_F_ALL_FOR_ALL; 12969 12970 /* If one device supports hw checksumming, set for all. */ 12971 if (all & NETIF_F_HW_CSUM) 12972 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM); 12973 12974 return all; 12975 } 12976 EXPORT_SYMBOL(netdev_increment_features); 12977 12978 /** 12979 * netdev_compute_master_upper_features - compute feature from lowers 12980 * @dev: the upper device 12981 * @update_header: whether to update upper device's header_len/headroom/tailroom 12982 * 12983 * Recompute the upper device's feature based on all lower devices. 12984 */ 12985 void netdev_compute_master_upper_features(struct net_device *dev, bool update_header) 12986 { 12987 unsigned int dst_release_flag = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 12988 netdev_features_t gso_partial_features = MASTER_UPPER_DEV_GSO_PARTIAL_FEATURES; 12989 netdev_features_t xfrm_features = MASTER_UPPER_DEV_XFRM_FEATURES; 12990 netdev_features_t mpls_features = MASTER_UPPER_DEV_MPLS_FEATURES; 12991 netdev_features_t vlan_features = MASTER_UPPER_DEV_VLAN_FEATURES; 12992 netdev_features_t enc_features = MASTER_UPPER_DEV_ENC_FEATURES; 12993 unsigned short max_header_len = ETH_HLEN; 12994 unsigned int tso_max_size = TSO_MAX_SIZE; 12995 unsigned short max_headroom = 0; 12996 unsigned short max_tailroom = 0; 12997 u16 tso_max_segs = TSO_MAX_SEGS; 12998 struct net_device *lower_dev; 12999 struct list_head *iter; 13000 13001 mpls_features = netdev_base_features(mpls_features); 13002 vlan_features = netdev_base_features(vlan_features); 13003 enc_features = netdev_base_features(enc_features); 13004 13005 netdev_for_each_lower_dev(dev, lower_dev, iter) { 13006 gso_partial_features = netdev_increment_features(gso_partial_features, 13007 lower_dev->gso_partial_features, 13008 MASTER_UPPER_DEV_GSO_PARTIAL_FEATURES); 13009 13010 vlan_features = netdev_increment_features(vlan_features, 13011 lower_dev->vlan_features, 13012 MASTER_UPPER_DEV_VLAN_FEATURES); 13013 13014 enc_features = netdev_increment_features(enc_features, 13015 lower_dev->hw_enc_features, 13016 MASTER_UPPER_DEV_ENC_FEATURES); 13017 13018 if (IS_ENABLED(CONFIG_XFRM_OFFLOAD)) 13019 xfrm_features = netdev_increment_features(xfrm_features, 13020 lower_dev->hw_enc_features, 13021 MASTER_UPPER_DEV_XFRM_FEATURES); 13022 13023 mpls_features = netdev_increment_features(mpls_features, 13024 lower_dev->mpls_features, 13025 MASTER_UPPER_DEV_MPLS_FEATURES); 13026 13027 dst_release_flag &= lower_dev->priv_flags; 13028 13029 if (update_header) { 13030 max_header_len = max(max_header_len, lower_dev->hard_header_len); 13031 max_headroom = max(max_headroom, lower_dev->needed_headroom); 13032 max_tailroom = max(max_tailroom, lower_dev->needed_tailroom); 13033 } 13034 13035 tso_max_size = min(tso_max_size, lower_dev->tso_max_size); 13036 tso_max_segs = min(tso_max_segs, lower_dev->tso_max_segs); 13037 } 13038 13039 dev->gso_partial_features = gso_partial_features; 13040 dev->vlan_features = vlan_features; 13041 dev->hw_enc_features = enc_features | NETIF_F_GSO_ENCAP_ALL | 13042 NETIF_F_HW_VLAN_CTAG_TX | 13043 NETIF_F_HW_VLAN_STAG_TX; 13044 if (IS_ENABLED(CONFIG_XFRM_OFFLOAD)) 13045 dev->hw_enc_features |= xfrm_features; 13046 dev->mpls_features = mpls_features; 13047 13048 dev->priv_flags &= ~IFF_XMIT_DST_RELEASE; 13049 if ((dev->priv_flags & IFF_XMIT_DST_RELEASE_PERM) && 13050 dst_release_flag == (IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM)) 13051 dev->priv_flags |= IFF_XMIT_DST_RELEASE; 13052 13053 if (update_header) { 13054 dev->hard_header_len = max_header_len; 13055 dev->needed_headroom = max_headroom; 13056 dev->needed_tailroom = max_tailroom; 13057 } 13058 13059 netif_set_tso_max_segs(dev, tso_max_segs); 13060 netif_set_tso_max_size(dev, tso_max_size); 13061 13062 netdev_change_features(dev); 13063 } 13064 EXPORT_SYMBOL(netdev_compute_master_upper_features); 13065 13066 static struct hlist_head * __net_init netdev_create_hash(void) 13067 { 13068 int i; 13069 struct hlist_head *hash; 13070 13071 hash = kmalloc_objs(*hash, NETDEV_HASHENTRIES); 13072 if (hash != NULL) 13073 for (i = 0; i < NETDEV_HASHENTRIES; i++) 13074 INIT_HLIST_HEAD(&hash[i]); 13075 13076 return hash; 13077 } 13078 13079 /* Initialize per network namespace state */ 13080 static int __net_init netdev_init(struct net *net) 13081 { 13082 BUILD_BUG_ON(GRO_HASH_BUCKETS > 13083 BITS_PER_BYTE * sizeof_field(struct gro_node, bitmask)); 13084 13085 INIT_LIST_HEAD(&net->dev_base_head); 13086 13087 net->dev_name_head = netdev_create_hash(); 13088 if (net->dev_name_head == NULL) 13089 goto err_name; 13090 13091 net->dev_index_head = netdev_create_hash(); 13092 if (net->dev_index_head == NULL) 13093 goto err_idx; 13094 13095 xa_init_flags(&net->dev_by_index, XA_FLAGS_ALLOC1); 13096 13097 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain); 13098 13099 return 0; 13100 13101 err_idx: 13102 kfree(net->dev_name_head); 13103 err_name: 13104 return -ENOMEM; 13105 } 13106 13107 /** 13108 * netdev_drivername - network driver for the device 13109 * @dev: network device 13110 * 13111 * Determine network driver for device. 13112 */ 13113 const char *netdev_drivername(const struct net_device *dev) 13114 { 13115 const struct device_driver *driver; 13116 const struct device *parent; 13117 const char *empty = ""; 13118 13119 parent = dev->dev.parent; 13120 if (!parent) 13121 return empty; 13122 13123 driver = parent->driver; 13124 if (driver && driver->name) 13125 return driver->name; 13126 return empty; 13127 } 13128 13129 static void __netdev_printk(const char *level, const struct net_device *dev, 13130 struct va_format *vaf) 13131 { 13132 if (dev && dev->dev.parent) { 13133 dev_printk_emit(level[1] - '0', 13134 dev->dev.parent, 13135 "%s %s %s%s: %pV", 13136 dev_driver_string(dev->dev.parent), 13137 dev_name(dev->dev.parent), 13138 netdev_name(dev), netdev_reg_state(dev), 13139 vaf); 13140 } else if (dev) { 13141 printk("%s%s%s: %pV", 13142 level, netdev_name(dev), netdev_reg_state(dev), vaf); 13143 } else { 13144 printk("%s(NULL net_device): %pV", level, vaf); 13145 } 13146 } 13147 13148 void netdev_printk(const char *level, const struct net_device *dev, 13149 const char *format, ...) 13150 { 13151 struct va_format vaf; 13152 va_list args; 13153 13154 va_start(args, format); 13155 13156 vaf.fmt = format; 13157 vaf.va = &args; 13158 13159 __netdev_printk(level, dev, &vaf); 13160 13161 va_end(args); 13162 } 13163 EXPORT_SYMBOL(netdev_printk); 13164 13165 #define define_netdev_printk_level(func, level) \ 13166 void func(const struct net_device *dev, const char *fmt, ...) \ 13167 { \ 13168 struct va_format vaf; \ 13169 va_list args; \ 13170 \ 13171 va_start(args, fmt); \ 13172 \ 13173 vaf.fmt = fmt; \ 13174 vaf.va = &args; \ 13175 \ 13176 __netdev_printk(level, dev, &vaf); \ 13177 \ 13178 va_end(args); \ 13179 } \ 13180 EXPORT_SYMBOL(func); 13181 13182 define_netdev_printk_level(netdev_emerg, KERN_EMERG); 13183 define_netdev_printk_level(netdev_alert, KERN_ALERT); 13184 define_netdev_printk_level(netdev_crit, KERN_CRIT); 13185 define_netdev_printk_level(netdev_err, KERN_ERR); 13186 define_netdev_printk_level(netdev_warn, KERN_WARNING); 13187 define_netdev_printk_level(netdev_notice, KERN_NOTICE); 13188 define_netdev_printk_level(netdev_info, KERN_INFO); 13189 13190 static void __net_exit netdev_exit(struct net *net) 13191 { 13192 kfree(net->dev_name_head); 13193 kfree(net->dev_index_head); 13194 xa_destroy(&net->dev_by_index); 13195 if (net != &init_net) 13196 WARN_ON_ONCE(!list_empty(&net->dev_base_head)); 13197 } 13198 13199 static struct pernet_operations __net_initdata netdev_net_ops = { 13200 .init = netdev_init, 13201 .exit = netdev_exit, 13202 }; 13203 13204 static void __net_exit default_device_exit_net(struct net *net) 13205 { 13206 struct netdev_name_node *name_node, *tmp; 13207 struct net_device *dev, *aux; 13208 /* 13209 * Push all migratable network devices back to the 13210 * initial network namespace 13211 */ 13212 13213 for_each_netdev_safe(net, dev, aux) { 13214 int err; 13215 char fb_name[IFNAMSIZ]; 13216 13217 /* Ignore unmoveable devices (i.e. loopback) */ 13218 if (dev->netns_immutable) 13219 continue; 13220 13221 /* Leave virtual devices for the generic cleanup */ 13222 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund) 13223 continue; 13224 13225 /* Push remaining network devices to init_net */ 13226 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex); 13227 if (netdev_name_in_use(&init_net, fb_name)) 13228 snprintf(fb_name, IFNAMSIZ, "dev%%d"); 13229 13230 netdev_for_each_altname_safe(dev, name_node, tmp) 13231 if (netdev_name_in_use(&init_net, name_node->name)) 13232 __netdev_name_node_alt_destroy(name_node); 13233 13234 err = dev_change_net_namespace(dev, &init_net, fb_name); 13235 if (err) { 13236 pr_emerg("%s: failed to move %s to init_net: %d\n", 13237 __func__, dev->name, err); 13238 BUG(); 13239 } 13240 } 13241 } 13242 13243 static void __net_exit default_device_exit_batch(struct list_head *net_list) 13244 { 13245 /* At exit all network devices most be removed from a network 13246 * namespace. Do this in the reverse order of registration. 13247 * Do this across as many network namespaces as possible to 13248 * improve batching efficiency. 13249 */ 13250 struct net_device *dev; 13251 struct net *net; 13252 LIST_HEAD(dev_kill_list); 13253 13254 rtnl_lock(); 13255 13256 __rtnl_net_lock(&init_net); 13257 13258 list_for_each_entry(net, net_list, exit_list) { 13259 __rtnl_net_lock(net); 13260 default_device_exit_net(net); 13261 __rtnl_net_unlock(net); 13262 13263 cond_resched(); 13264 } 13265 13266 __rtnl_net_unlock(&init_net); 13267 13268 list_for_each_entry(net, net_list, exit_list) { 13269 __rtnl_net_lock(net); 13270 13271 for_each_netdev_reverse(net, dev) { 13272 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink) 13273 dev->rtnl_link_ops->dellink(dev, &dev_kill_list); 13274 else 13275 unregister_netdevice_queue(dev, &dev_kill_list); 13276 } 13277 13278 unregister_netdevice_queue_many_net(net, &dev_kill_list); 13279 __rtnl_net_unlock(net); 13280 } 13281 unregister_netdevice_many(&dev_kill_list); 13282 rtnl_unlock(); 13283 13284 rtnl_net_flush_workqueue(); 13285 } 13286 13287 static struct pernet_operations __net_initdata default_device_ops = { 13288 .exit_batch = default_device_exit_batch, 13289 }; 13290 13291 static void __init net_dev_struct_check(void) 13292 { 13293 /* TX read-mostly hotpath */ 13294 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, priv_flags_fast); 13295 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, netdev_ops); 13296 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, header_ops); 13297 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, _tx); 13298 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, real_num_tx_queues); 13299 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_size); 13300 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_ipv4_max_size); 13301 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_segs); 13302 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_partial_features); 13303 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, num_tc); 13304 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, mtu); 13305 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, needed_headroom); 13306 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tc_to_txq); 13307 #ifdef CONFIG_XPS 13308 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, xps_maps); 13309 #endif 13310 #ifdef CONFIG_NETFILTER_EGRESS 13311 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, nf_hooks_egress); 13312 #endif 13313 #ifdef CONFIG_NET_XGRESS 13314 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tcx_egress); 13315 #endif 13316 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_tx, 160); 13317 13318 /* TXRX read-mostly hotpath */ 13319 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, lstats); 13320 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, state); 13321 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, flags); 13322 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, hard_header_len); 13323 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, features); 13324 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, ip6_ptr); 13325 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_txrx, 46); 13326 13327 /* RX read-mostly hotpath */ 13328 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ptype_specific); 13329 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ifindex); 13330 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, real_num_rx_queues); 13331 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, _rx); 13332 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_max_size); 13333 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_ipv4_max_size); 13334 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler); 13335 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler_data); 13336 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, nd_net); 13337 #ifdef CONFIG_NETPOLL 13338 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, npinfo); 13339 #endif 13340 #ifdef CONFIG_NET_XGRESS 13341 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, tcx_ingress); 13342 #endif 13343 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_rx, 92); 13344 } 13345 13346 /* 13347 * Initialize the DEV module. At boot time this walks the device list and 13348 * unhooks any devices that fail to initialise (normally hardware not 13349 * present) and leaves us with a valid list of present and active devices. 13350 * 13351 */ 13352 13353 /* We allocate 256 pages for each CPU if PAGE_SHIFT is 12 */ 13354 #define SYSTEM_PERCPU_PAGE_POOL_SIZE ((1 << 20) / PAGE_SIZE) 13355 13356 static int net_page_pool_create(int cpuid) 13357 { 13358 #if IS_ENABLED(CONFIG_PAGE_POOL) 13359 struct page_pool_params page_pool_params = { 13360 .pool_size = SYSTEM_PERCPU_PAGE_POOL_SIZE, 13361 .flags = PP_FLAG_SYSTEM_POOL, 13362 .nid = cpu_to_mem(cpuid), 13363 }; 13364 struct page_pool *pp_ptr; 13365 int err; 13366 13367 pp_ptr = page_pool_create_percpu(&page_pool_params, cpuid); 13368 if (IS_ERR(pp_ptr)) 13369 return -ENOMEM; 13370 13371 err = xdp_reg_page_pool(pp_ptr); 13372 if (err) { 13373 page_pool_destroy(pp_ptr); 13374 return err; 13375 } 13376 13377 per_cpu(system_page_pool.pool, cpuid) = pp_ptr; 13378 #endif 13379 return 0; 13380 } 13381 13382 static int backlog_napi_should_run(unsigned int cpu) 13383 { 13384 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 13385 struct napi_struct *napi = &sd->backlog; 13386 13387 return test_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 13388 } 13389 13390 static void run_backlog_napi(unsigned int cpu) 13391 { 13392 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 13393 13394 napi_threaded_poll_loop(&sd->backlog, NULL); 13395 } 13396 13397 static void backlog_napi_setup(unsigned int cpu) 13398 { 13399 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 13400 struct napi_struct *napi = &sd->backlog; 13401 13402 napi->thread = this_cpu_read(backlog_napi); 13403 set_bit(NAPI_STATE_THREADED, &napi->state); 13404 } 13405 13406 static struct smp_hotplug_thread backlog_threads = { 13407 .store = &backlog_napi, 13408 .thread_should_run = backlog_napi_should_run, 13409 .thread_fn = run_backlog_napi, 13410 .thread_comm = "backlog_napi/%u", 13411 .setup = backlog_napi_setup, 13412 }; 13413 13414 /* 13415 * This is called single threaded during boot, so no need 13416 * to take the rtnl semaphore. 13417 */ 13418 static int __init net_dev_init(void) 13419 { 13420 int i, rc = -ENOMEM; 13421 13422 BUG_ON(!dev_boot_phase); 13423 13424 net_dev_struct_check(); 13425 13426 if (dev_proc_init()) 13427 goto out; 13428 13429 if (netdev_kobject_init()) 13430 goto out; 13431 13432 for (i = 0; i < PTYPE_HASH_SIZE; i++) 13433 INIT_LIST_HEAD(&ptype_base[i]); 13434 13435 if (register_pernet_subsys(&netdev_net_ops)) 13436 goto out; 13437 13438 /* 13439 * Initialise the packet receive queues. 13440 */ 13441 13442 flush_backlogs_fallback = flush_backlogs_alloc(); 13443 if (!flush_backlogs_fallback) 13444 goto out; 13445 13446 for_each_possible_cpu(i) { 13447 struct softnet_data *sd = &per_cpu(softnet_data, i); 13448 13449 skb_queue_head_init(&sd->input_pkt_queue); 13450 skb_queue_head_init(&sd->process_queue); 13451 #ifdef CONFIG_XFRM_OFFLOAD 13452 skb_queue_head_init(&sd->xfrm_backlog); 13453 #endif 13454 INIT_LIST_HEAD(&sd->poll_list); 13455 sd->output_queue_tailp = &sd->output_queue; 13456 #ifdef CONFIG_RPS 13457 INIT_CSD(&sd->csd, rps_trigger_softirq, sd); 13458 sd->cpu = i; 13459 #endif 13460 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd); 13461 13462 gro_init(&sd->backlog.gro); 13463 sd->backlog.poll = process_backlog; 13464 sd->backlog.weight = weight_p; 13465 INIT_LIST_HEAD(&sd->backlog.poll_list); 13466 13467 if (net_page_pool_create(i)) 13468 goto out; 13469 } 13470 net_hotdata.skb_defer_nodes = 13471 __alloc_percpu(sizeof(struct skb_defer_node) * nr_node_ids, 13472 __alignof__(struct skb_defer_node)); 13473 if (!net_hotdata.skb_defer_nodes) 13474 goto out; 13475 if (use_backlog_threads()) 13476 smpboot_register_percpu_thread(&backlog_threads); 13477 13478 dev_boot_phase = 0; 13479 13480 /* The loopback device is special if any other network devices 13481 * is present in a network namespace the loopback device must 13482 * be present. Since we now dynamically allocate and free the 13483 * loopback device ensure this invariant is maintained by 13484 * keeping the loopback device as the first device on the 13485 * list of network devices. Ensuring the loopback devices 13486 * is the first device that appears and the last network device 13487 * that disappears. 13488 */ 13489 if (register_pernet_device(&loopback_net_ops)) 13490 goto out; 13491 13492 if (register_pernet_device(&default_device_ops)) 13493 goto out; 13494 13495 open_softirq(NET_TX_SOFTIRQ, net_tx_action); 13496 open_softirq(NET_RX_SOFTIRQ, net_rx_action); 13497 13498 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead", 13499 NULL, dev_cpu_dead); 13500 WARN_ON(rc < 0); 13501 rc = 0; 13502 13503 /* avoid static key IPIs to isolated CPUs */ 13504 if (housekeeping_enabled(HK_TYPE_MISC)) 13505 net_enable_timestamp(); 13506 out: 13507 if (rc < 0) { 13508 for_each_possible_cpu(i) { 13509 struct page_pool *pp_ptr; 13510 13511 pp_ptr = per_cpu(system_page_pool.pool, i); 13512 if (!pp_ptr) 13513 continue; 13514 13515 xdp_unreg_page_pool(pp_ptr); 13516 page_pool_destroy(pp_ptr); 13517 per_cpu(system_page_pool.pool, i) = NULL; 13518 } 13519 } 13520 13521 return rc; 13522 } 13523 13524 subsys_initcall(net_dev_init); 13525