1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * NET3 Protocol independent device support routines.
5 * Derived from the non IP parts of dev.c 1.0.19
7 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
8 * Mark Evans, <evansmp@uhura.aston.ac.uk>
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>
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
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
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
47 * Rudi Cilibrasi : Pass the right thing to
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
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
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
71 #include <linux/uaccess.h>
72 #include <linux/bitmap.h>
73 #include <linux/capability.h>
74 #include <linux/cpu.h>
75 #include <linux/types.h>
76 #include <linux/kernel.h>
77 #include <linux/hash.h>
78 #include <linux/slab.h>
79 #include <linux/sched.h>
80 #include <linux/sched/mm.h>
81 #include <linux/mutex.h>
82 #include <linux/rwsem.h>
83 #include <linux/string.h>
85 #include <linux/socket.h>
86 #include <linux/sockios.h>
87 #include <linux/errno.h>
88 #include <linux/interrupt.h>
89 #include <linux/if_ether.h>
90 #include <linux/netdevice.h>
91 #include <linux/etherdevice.h>
92 #include <linux/ethtool.h>
93 #include <linux/skbuff.h>
94 #include <linux/kthread.h>
95 #include <linux/bpf.h>
96 #include <linux/bpf_trace.h>
97 #include <net/net_namespace.h>
99 #include <net/busy_poll.h>
100 #include <linux/rtnetlink.h>
101 #include <linux/stat.h>
104 #include <net/dst_metadata.h>
106 #include <net/pkt_sched.h>
107 #include <net/pkt_cls.h>
108 #include <net/checksum.h>
109 #include <net/xfrm.h>
111 #include <linux/highmem.h>
112 #include <linux/init.h>
113 #include <linux/module.h>
114 #include <linux/netpoll.h>
115 #include <linux/rcupdate.h>
116 #include <linux/delay.h>
117 #include <net/iw_handler.h>
118 #include <asm/current.h>
119 #include <linux/audit.h>
120 #include <linux/dmaengine.h>
121 #include <linux/err.h>
122 #include <linux/ctype.h>
123 #include <linux/if_arp.h>
124 #include <linux/if_vlan.h>
125 #include <linux/ip.h>
127 #include <net/mpls.h>
128 #include <linux/ipv6.h>
129 #include <linux/in.h>
130 #include <linux/jhash.h>
131 #include <linux/random.h>
132 #include <trace/events/napi.h>
133 #include <trace/events/net.h>
134 #include <trace/events/skb.h>
135 #include <trace/events/qdisc.h>
136 #include <trace/events/xdp.h>
137 #include <linux/inetdevice.h>
138 #include <linux/cpu_rmap.h>
139 #include <linux/static_key.h>
140 #include <linux/hashtable.h>
141 #include <linux/vmalloc.h>
142 #include <linux/if_macvlan.h>
143 #include <linux/errqueue.h>
144 #include <linux/hrtimer.h>
145 #include <linux/netfilter_netdev.h>
146 #include <linux/crash_dump.h>
147 #include <linux/sctp.h>
148 #include <net/udp_tunnel.h>
149 #include <linux/net_namespace.h>
150 #include <linux/indirect_call_wrapper.h>
151 #include <net/devlink.h>
152 #include <linux/pm_runtime.h>
153 #include <linux/prandom.h>
154 #include <linux/once_lite.h>
155 #include <net/netdev_rx_queue.h>
158 #include "net-sysfs.h"
160 static DEFINE_SPINLOCK(ptype_lock);
161 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly;
162 struct list_head ptype_all __read_mostly; /* Taps */
164 static int netif_rx_internal(struct sk_buff *skb);
165 static int call_netdevice_notifiers_extack(unsigned long val,
166 struct net_device *dev,
167 struct netlink_ext_ack *extack);
168 static struct napi_struct *napi_by_id(unsigned int napi_id);
171 * The @dev_base_head list is protected by @dev_base_lock and the rtnl
174 * Pure readers hold dev_base_lock for reading, or rcu_read_lock()
176 * Writers must hold the rtnl semaphore while they loop through the
177 * dev_base_head list, and hold dev_base_lock for writing when they do the
178 * actual updates. This allows pure readers to access the list even
179 * while a writer is preparing to update it.
181 * To put it another way, dev_base_lock is held for writing only to
182 * protect against pure readers; the rtnl semaphore provides the
183 * protection against other writers.
185 * See, for example usages, register_netdevice() and
186 * unregister_netdevice(), which must be called with the rtnl
189 DEFINE_RWLOCK(dev_base_lock);
190 EXPORT_SYMBOL(dev_base_lock);
192 static DEFINE_MUTEX(ifalias_mutex);
194 /* protects napi_hash addition/deletion and napi_gen_id */
195 static DEFINE_SPINLOCK(napi_hash_lock);
197 static unsigned int napi_gen_id = NR_CPUS;
198 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8);
200 static DECLARE_RWSEM(devnet_rename_sem);
202 static inline void dev_base_seq_inc(struct net *net)
204 while (++net->dev_base_seq == 0)
208 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name)
210 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ));
212 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)];
215 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex)
217 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)];
220 static inline void rps_lock_irqsave(struct softnet_data *sd,
221 unsigned long *flags)
223 if (IS_ENABLED(CONFIG_RPS))
224 spin_lock_irqsave(&sd->input_pkt_queue.lock, *flags);
225 else if (!IS_ENABLED(CONFIG_PREEMPT_RT))
226 local_irq_save(*flags);
229 static inline void rps_lock_irq_disable(struct softnet_data *sd)
231 if (IS_ENABLED(CONFIG_RPS))
232 spin_lock_irq(&sd->input_pkt_queue.lock);
233 else if (!IS_ENABLED(CONFIG_PREEMPT_RT))
237 static inline void rps_unlock_irq_restore(struct softnet_data *sd,
238 unsigned long *flags)
240 if (IS_ENABLED(CONFIG_RPS))
241 spin_unlock_irqrestore(&sd->input_pkt_queue.lock, *flags);
242 else if (!IS_ENABLED(CONFIG_PREEMPT_RT))
243 local_irq_restore(*flags);
246 static inline void rps_unlock_irq_enable(struct softnet_data *sd)
248 if (IS_ENABLED(CONFIG_RPS))
249 spin_unlock_irq(&sd->input_pkt_queue.lock);
250 else if (!IS_ENABLED(CONFIG_PREEMPT_RT))
254 static struct netdev_name_node *netdev_name_node_alloc(struct net_device *dev,
257 struct netdev_name_node *name_node;
259 name_node = kmalloc(sizeof(*name_node), GFP_KERNEL);
262 INIT_HLIST_NODE(&name_node->hlist);
263 name_node->dev = dev;
264 name_node->name = name;
268 static struct netdev_name_node *
269 netdev_name_node_head_alloc(struct net_device *dev)
271 struct netdev_name_node *name_node;
273 name_node = netdev_name_node_alloc(dev, dev->name);
276 INIT_LIST_HEAD(&name_node->list);
280 static void netdev_name_node_free(struct netdev_name_node *name_node)
285 static void netdev_name_node_add(struct net *net,
286 struct netdev_name_node *name_node)
288 hlist_add_head_rcu(&name_node->hlist,
289 dev_name_hash(net, name_node->name));
292 static void netdev_name_node_del(struct netdev_name_node *name_node)
294 hlist_del_rcu(&name_node->hlist);
297 static struct netdev_name_node *netdev_name_node_lookup(struct net *net,
300 struct hlist_head *head = dev_name_hash(net, name);
301 struct netdev_name_node *name_node;
303 hlist_for_each_entry(name_node, head, hlist)
304 if (!strcmp(name_node->name, name))
309 static struct netdev_name_node *netdev_name_node_lookup_rcu(struct net *net,
312 struct hlist_head *head = dev_name_hash(net, name);
313 struct netdev_name_node *name_node;
315 hlist_for_each_entry_rcu(name_node, head, hlist)
316 if (!strcmp(name_node->name, name))
321 bool netdev_name_in_use(struct net *net, const char *name)
323 return netdev_name_node_lookup(net, name);
325 EXPORT_SYMBOL(netdev_name_in_use);
327 int netdev_name_node_alt_create(struct net_device *dev, const char *name)
329 struct netdev_name_node *name_node;
330 struct net *net = dev_net(dev);
332 name_node = netdev_name_node_lookup(net, name);
335 name_node = netdev_name_node_alloc(dev, name);
338 netdev_name_node_add(net, name_node);
339 /* The node that holds dev->name acts as a head of per-device list. */
340 list_add_tail(&name_node->list, &dev->name_node->list);
345 static void __netdev_name_node_alt_destroy(struct netdev_name_node *name_node)
347 list_del(&name_node->list);
348 kfree(name_node->name);
349 netdev_name_node_free(name_node);
352 int netdev_name_node_alt_destroy(struct net_device *dev, const char *name)
354 struct netdev_name_node *name_node;
355 struct net *net = dev_net(dev);
357 name_node = netdev_name_node_lookup(net, name);
360 /* lookup might have found our primary name or a name belonging
363 if (name_node == dev->name_node || name_node->dev != dev)
366 netdev_name_node_del(name_node);
368 __netdev_name_node_alt_destroy(name_node);
373 static void netdev_name_node_alt_flush(struct net_device *dev)
375 struct netdev_name_node *name_node, *tmp;
377 list_for_each_entry_safe(name_node, tmp, &dev->name_node->list, list)
378 __netdev_name_node_alt_destroy(name_node);
381 /* Device list insertion */
382 static void list_netdevice(struct net_device *dev)
384 struct netdev_name_node *name_node;
385 struct net *net = dev_net(dev);
389 write_lock(&dev_base_lock);
390 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head);
391 netdev_name_node_add(net, dev->name_node);
392 hlist_add_head_rcu(&dev->index_hlist,
393 dev_index_hash(net, dev->ifindex));
394 write_unlock(&dev_base_lock);
396 netdev_for_each_altname(dev, name_node)
397 netdev_name_node_add(net, name_node);
399 /* We reserved the ifindex, this can't fail */
400 WARN_ON(xa_store(&net->dev_by_index, dev->ifindex, dev, GFP_KERNEL));
402 dev_base_seq_inc(net);
405 /* Device list removal
406 * caller must respect a RCU grace period before freeing/reusing dev
408 static void unlist_netdevice(struct net_device *dev, bool lock)
410 struct netdev_name_node *name_node;
411 struct net *net = dev_net(dev);
415 xa_erase(&net->dev_by_index, dev->ifindex);
417 netdev_for_each_altname(dev, name_node)
418 netdev_name_node_del(name_node);
420 /* Unlink dev from the device chain */
422 write_lock(&dev_base_lock);
423 list_del_rcu(&dev->dev_list);
424 netdev_name_node_del(dev->name_node);
425 hlist_del_rcu(&dev->index_hlist);
427 write_unlock(&dev_base_lock);
429 dev_base_seq_inc(dev_net(dev));
436 static RAW_NOTIFIER_HEAD(netdev_chain);
439 * Device drivers call our routines to queue packets here. We empty the
440 * queue in the local softnet handler.
443 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data);
444 EXPORT_PER_CPU_SYMBOL(softnet_data);
446 #ifdef CONFIG_LOCKDEP
448 * register_netdevice() inits txq->_xmit_lock and sets lockdep class
449 * according to dev->type
451 static const unsigned short netdev_lock_type[] = {
452 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25,
453 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET,
454 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM,
455 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP,
456 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD,
457 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25,
458 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP,
459 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD,
460 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI,
461 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE,
462 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET,
463 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL,
464 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM,
465 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE,
466 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE};
468 static const char *const netdev_lock_name[] = {
469 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25",
470 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET",
471 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM",
472 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP",
473 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD",
474 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25",
475 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP",
476 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD",
477 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI",
478 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE",
479 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET",
480 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL",
481 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM",
482 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE",
483 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"};
485 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)];
486 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)];
488 static inline unsigned short netdev_lock_pos(unsigned short dev_type)
492 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++)
493 if (netdev_lock_type[i] == dev_type)
495 /* the last key is used by default */
496 return ARRAY_SIZE(netdev_lock_type) - 1;
499 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
500 unsigned short dev_type)
504 i = netdev_lock_pos(dev_type);
505 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i],
506 netdev_lock_name[i]);
509 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
513 i = netdev_lock_pos(dev->type);
514 lockdep_set_class_and_name(&dev->addr_list_lock,
515 &netdev_addr_lock_key[i],
516 netdev_lock_name[i]);
519 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
520 unsigned short dev_type)
524 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
529 /*******************************************************************************
531 * Protocol management and registration routines
533 *******************************************************************************/
537 * Add a protocol ID to the list. Now that the input handler is
538 * smarter we can dispense with all the messy stuff that used to be
541 * BEWARE!!! Protocol handlers, mangling input packets,
542 * MUST BE last in hash buckets and checking protocol handlers
543 * MUST start from promiscuous ptype_all chain in net_bh.
544 * It is true now, do not change it.
545 * Explanation follows: if protocol handler, mangling packet, will
546 * be the first on list, it is not able to sense, that packet
547 * is cloned and should be copied-on-write, so that it will
548 * change it and subsequent readers will get broken packet.
552 static inline struct list_head *ptype_head(const struct packet_type *pt)
554 if (pt->type == htons(ETH_P_ALL))
555 return pt->dev ? &pt->dev->ptype_all : &ptype_all;
557 return pt->dev ? &pt->dev->ptype_specific :
558 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK];
562 * dev_add_pack - add packet handler
563 * @pt: packet type declaration
565 * Add a protocol handler to the networking stack. The passed &packet_type
566 * is linked into kernel lists and may not be freed until it has been
567 * removed from the kernel lists.
569 * This call does not sleep therefore it can not
570 * guarantee all CPU's that are in middle of receiving packets
571 * will see the new packet type (until the next received packet).
574 void dev_add_pack(struct packet_type *pt)
576 struct list_head *head = ptype_head(pt);
578 spin_lock(&ptype_lock);
579 list_add_rcu(&pt->list, head);
580 spin_unlock(&ptype_lock);
582 EXPORT_SYMBOL(dev_add_pack);
585 * __dev_remove_pack - remove packet handler
586 * @pt: packet type declaration
588 * Remove a protocol handler that was previously added to the kernel
589 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
590 * from the kernel lists and can be freed or reused once this function
593 * The packet type might still be in use by receivers
594 * and must not be freed until after all the CPU's have gone
595 * through a quiescent state.
597 void __dev_remove_pack(struct packet_type *pt)
599 struct list_head *head = ptype_head(pt);
600 struct packet_type *pt1;
602 spin_lock(&ptype_lock);
604 list_for_each_entry(pt1, head, list) {
606 list_del_rcu(&pt->list);
611 pr_warn("dev_remove_pack: %p not found\n", pt);
613 spin_unlock(&ptype_lock);
615 EXPORT_SYMBOL(__dev_remove_pack);
618 * dev_remove_pack - remove packet handler
619 * @pt: packet type declaration
621 * Remove a protocol handler that was previously added to the kernel
622 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
623 * from the kernel lists and can be freed or reused once this function
626 * This call sleeps to guarantee that no CPU is looking at the packet
629 void dev_remove_pack(struct packet_type *pt)
631 __dev_remove_pack(pt);
635 EXPORT_SYMBOL(dev_remove_pack);
638 /*******************************************************************************
640 * Device Interface Subroutines
642 *******************************************************************************/
645 * dev_get_iflink - get 'iflink' value of a interface
646 * @dev: targeted interface
648 * Indicates the ifindex the interface is linked to.
649 * Physical interfaces have the same 'ifindex' and 'iflink' values.
652 int dev_get_iflink(const struct net_device *dev)
654 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink)
655 return dev->netdev_ops->ndo_get_iflink(dev);
659 EXPORT_SYMBOL(dev_get_iflink);
662 * dev_fill_metadata_dst - Retrieve tunnel egress information.
663 * @dev: targeted interface
666 * For better visibility of tunnel traffic OVS needs to retrieve
667 * egress tunnel information for a packet. Following API allows
668 * user to get this info.
670 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb)
672 struct ip_tunnel_info *info;
674 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst)
677 info = skb_tunnel_info_unclone(skb);
680 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX)))
683 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb);
685 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst);
687 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack)
689 int k = stack->num_paths++;
691 if (WARN_ON_ONCE(k >= NET_DEVICE_PATH_STACK_MAX))
694 return &stack->path[k];
697 int dev_fill_forward_path(const struct net_device *dev, const u8 *daddr,
698 struct net_device_path_stack *stack)
700 const struct net_device *last_dev;
701 struct net_device_path_ctx ctx = {
704 struct net_device_path *path;
707 memcpy(ctx.daddr, daddr, sizeof(ctx.daddr));
708 stack->num_paths = 0;
709 while (ctx.dev && ctx.dev->netdev_ops->ndo_fill_forward_path) {
711 path = dev_fwd_path(stack);
715 memset(path, 0, sizeof(struct net_device_path));
716 ret = ctx.dev->netdev_ops->ndo_fill_forward_path(&ctx, path);
720 if (WARN_ON_ONCE(last_dev == ctx.dev))
727 path = dev_fwd_path(stack);
730 path->type = DEV_PATH_ETHERNET;
735 EXPORT_SYMBOL_GPL(dev_fill_forward_path);
738 * __dev_get_by_name - find a device by its name
739 * @net: the applicable net namespace
740 * @name: name to find
742 * Find an interface by name. Must be called under RTNL semaphore
743 * or @dev_base_lock. If the name is found a pointer to the device
744 * is returned. If the name is not found then %NULL is returned. The
745 * reference counters are not incremented so the caller must be
746 * careful with locks.
749 struct net_device *__dev_get_by_name(struct net *net, const char *name)
751 struct netdev_name_node *node_name;
753 node_name = netdev_name_node_lookup(net, name);
754 return node_name ? node_name->dev : NULL;
756 EXPORT_SYMBOL(__dev_get_by_name);
759 * dev_get_by_name_rcu - find a device by its name
760 * @net: the applicable net namespace
761 * @name: name to find
763 * Find an interface by name.
764 * If the name is found a pointer to the device is returned.
765 * If the name is not found then %NULL is returned.
766 * The reference counters are not incremented so the caller must be
767 * careful with locks. The caller must hold RCU lock.
770 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name)
772 struct netdev_name_node *node_name;
774 node_name = netdev_name_node_lookup_rcu(net, name);
775 return node_name ? node_name->dev : NULL;
777 EXPORT_SYMBOL(dev_get_by_name_rcu);
779 /* Deprecated for new users, call netdev_get_by_name() instead */
780 struct net_device *dev_get_by_name(struct net *net, const char *name)
782 struct net_device *dev;
785 dev = dev_get_by_name_rcu(net, name);
790 EXPORT_SYMBOL(dev_get_by_name);
793 * netdev_get_by_name() - find a device by its name
794 * @net: the applicable net namespace
795 * @name: name to find
796 * @tracker: tracking object for the acquired reference
797 * @gfp: allocation flags for the tracker
799 * Find an interface by name. This can be called from any
800 * context and does its own locking. The returned handle has
801 * the usage count incremented and the caller must use netdev_put() to
802 * release it when it is no longer needed. %NULL is returned if no
803 * matching device is found.
805 struct net_device *netdev_get_by_name(struct net *net, const char *name,
806 netdevice_tracker *tracker, gfp_t gfp)
808 struct net_device *dev;
810 dev = dev_get_by_name(net, name);
812 netdev_tracker_alloc(dev, tracker, gfp);
815 EXPORT_SYMBOL(netdev_get_by_name);
818 * __dev_get_by_index - find a device by its ifindex
819 * @net: the applicable net namespace
820 * @ifindex: index of device
822 * Search for an interface by index. Returns %NULL if the device
823 * is not found or a pointer to the device. The device has not
824 * had its reference counter increased so the caller must be careful
825 * about locking. The caller must hold either the RTNL semaphore
829 struct net_device *__dev_get_by_index(struct net *net, int ifindex)
831 struct net_device *dev;
832 struct hlist_head *head = dev_index_hash(net, ifindex);
834 hlist_for_each_entry(dev, head, index_hlist)
835 if (dev->ifindex == ifindex)
840 EXPORT_SYMBOL(__dev_get_by_index);
843 * dev_get_by_index_rcu - find a device by its ifindex
844 * @net: the applicable net namespace
845 * @ifindex: index of device
847 * Search for an interface by index. Returns %NULL if the device
848 * is not found or a pointer to the device. The device has not
849 * had its reference counter increased so the caller must be careful
850 * about locking. The caller must hold RCU lock.
853 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex)
855 struct net_device *dev;
856 struct hlist_head *head = dev_index_hash(net, ifindex);
858 hlist_for_each_entry_rcu(dev, head, index_hlist)
859 if (dev->ifindex == ifindex)
864 EXPORT_SYMBOL(dev_get_by_index_rcu);
866 /* Deprecated for new users, call netdev_get_by_index() instead */
867 struct net_device *dev_get_by_index(struct net *net, int ifindex)
869 struct net_device *dev;
872 dev = dev_get_by_index_rcu(net, ifindex);
877 EXPORT_SYMBOL(dev_get_by_index);
880 * netdev_get_by_index() - find a device by its ifindex
881 * @net: the applicable net namespace
882 * @ifindex: index of device
883 * @tracker: tracking object for the acquired reference
884 * @gfp: allocation flags for the tracker
886 * Search for an interface by index. Returns NULL if the device
887 * is not found or a pointer to the device. The device returned has
888 * had a reference added and the pointer is safe until the user calls
889 * netdev_put() to indicate they have finished with it.
891 struct net_device *netdev_get_by_index(struct net *net, int ifindex,
892 netdevice_tracker *tracker, gfp_t gfp)
894 struct net_device *dev;
896 dev = dev_get_by_index(net, ifindex);
898 netdev_tracker_alloc(dev, tracker, gfp);
901 EXPORT_SYMBOL(netdev_get_by_index);
904 * dev_get_by_napi_id - find a device by napi_id
905 * @napi_id: ID of the NAPI struct
907 * Search for an interface by NAPI ID. Returns %NULL if the device
908 * is not found or a pointer to the device. The device has not had
909 * its reference counter increased so the caller must be careful
910 * about locking. The caller must hold RCU lock.
913 struct net_device *dev_get_by_napi_id(unsigned int napi_id)
915 struct napi_struct *napi;
917 WARN_ON_ONCE(!rcu_read_lock_held());
919 if (napi_id < MIN_NAPI_ID)
922 napi = napi_by_id(napi_id);
924 return napi ? napi->dev : NULL;
926 EXPORT_SYMBOL(dev_get_by_napi_id);
929 * netdev_get_name - get a netdevice name, knowing its ifindex.
930 * @net: network namespace
931 * @name: a pointer to the buffer where the name will be stored.
932 * @ifindex: the ifindex of the interface to get the name from.
934 int netdev_get_name(struct net *net, char *name, int ifindex)
936 struct net_device *dev;
939 down_read(&devnet_rename_sem);
942 dev = dev_get_by_index_rcu(net, ifindex);
948 strcpy(name, dev->name);
953 up_read(&devnet_rename_sem);
958 * dev_getbyhwaddr_rcu - find a device by its hardware address
959 * @net: the applicable net namespace
960 * @type: media type of device
961 * @ha: hardware address
963 * Search for an interface by MAC address. Returns NULL if the device
964 * is not found or a pointer to the device.
965 * The caller must hold RCU or RTNL.
966 * The returned device has not had its ref count increased
967 * and the caller must therefore be careful about locking
971 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type,
974 struct net_device *dev;
976 for_each_netdev_rcu(net, dev)
977 if (dev->type == type &&
978 !memcmp(dev->dev_addr, ha, dev->addr_len))
983 EXPORT_SYMBOL(dev_getbyhwaddr_rcu);
985 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type)
987 struct net_device *dev, *ret = NULL;
990 for_each_netdev_rcu(net, dev)
991 if (dev->type == type) {
999 EXPORT_SYMBOL(dev_getfirstbyhwtype);
1002 * __dev_get_by_flags - find any device with given flags
1003 * @net: the applicable net namespace
1004 * @if_flags: IFF_* values
1005 * @mask: bitmask of bits in if_flags to check
1007 * Search for any interface with the given flags. Returns NULL if a device
1008 * is not found or a pointer to the device. Must be called inside
1009 * rtnl_lock(), and result refcount is unchanged.
1012 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags,
1013 unsigned short mask)
1015 struct net_device *dev, *ret;
1020 for_each_netdev(net, dev) {
1021 if (((dev->flags ^ if_flags) & mask) == 0) {
1028 EXPORT_SYMBOL(__dev_get_by_flags);
1031 * dev_valid_name - check if name is okay for network device
1032 * @name: name string
1034 * Network device names need to be valid file names to
1035 * allow sysfs to work. We also disallow any kind of
1038 bool dev_valid_name(const char *name)
1042 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ)
1044 if (!strcmp(name, ".") || !strcmp(name, ".."))
1048 if (*name == '/' || *name == ':' || isspace(*name))
1054 EXPORT_SYMBOL(dev_valid_name);
1057 * __dev_alloc_name - allocate a name for a device
1058 * @net: network namespace to allocate the device name in
1059 * @name: name format string
1060 * @buf: scratch buffer and result name string
1062 * Passed a format string - eg "lt%d" it will try and find a suitable
1063 * id. It scans list of devices to build up a free map, then chooses
1064 * the first empty slot. The caller must hold the dev_base or rtnl lock
1065 * while allocating the name and adding the device in order to avoid
1067 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1068 * Returns the number of the unit assigned or a negative errno code.
1071 static int __dev_alloc_name(struct net *net, const char *name, char *buf)
1075 const int max_netdevices = 8*PAGE_SIZE;
1076 unsigned long *inuse;
1077 struct net_device *d;
1079 if (!dev_valid_name(name))
1082 p = strchr(name, '%');
1085 * Verify the string as this thing may have come from
1086 * the user. There must be either one "%d" and no other "%"
1089 if (p[1] != 'd' || strchr(p + 2, '%'))
1092 /* Use one page as a bit array of possible slots */
1093 inuse = bitmap_zalloc(max_netdevices, GFP_ATOMIC);
1097 for_each_netdev(net, d) {
1098 struct netdev_name_node *name_node;
1100 netdev_for_each_altname(d, name_node) {
1101 if (!sscanf(name_node->name, name, &i))
1103 if (i < 0 || i >= max_netdevices)
1106 /* avoid cases where sscanf is not exact inverse of printf */
1107 snprintf(buf, IFNAMSIZ, name, i);
1108 if (!strncmp(buf, name_node->name, IFNAMSIZ))
1109 __set_bit(i, inuse);
1111 if (!sscanf(d->name, name, &i))
1113 if (i < 0 || i >= max_netdevices)
1116 /* avoid cases where sscanf is not exact inverse of printf */
1117 snprintf(buf, IFNAMSIZ, name, i);
1118 if (!strncmp(buf, d->name, IFNAMSIZ))
1119 __set_bit(i, inuse);
1122 i = find_first_zero_bit(inuse, max_netdevices);
1126 snprintf(buf, IFNAMSIZ, name, i);
1127 if (!netdev_name_in_use(net, buf))
1130 /* It is possible to run out of possible slots
1131 * when the name is long and there isn't enough space left
1132 * for the digits, or if all bits are used.
1137 static int dev_prep_valid_name(struct net *net, struct net_device *dev,
1138 const char *want_name, char *out_name)
1142 if (!dev_valid_name(want_name))
1145 if (strchr(want_name, '%')) {
1146 ret = __dev_alloc_name(net, want_name, out_name);
1147 return ret < 0 ? ret : 0;
1148 } else if (netdev_name_in_use(net, want_name)) {
1150 } else if (out_name != want_name) {
1151 strscpy(out_name, want_name, IFNAMSIZ);
1157 static int dev_alloc_name_ns(struct net *net,
1158 struct net_device *dev,
1165 ret = __dev_alloc_name(net, name, buf);
1167 strscpy(dev->name, buf, IFNAMSIZ);
1172 * dev_alloc_name - allocate a name for a device
1174 * @name: name format string
1176 * Passed a format string - eg "lt%d" it will try and find a suitable
1177 * id. It scans list of devices to build up a free map, then chooses
1178 * the first empty slot. The caller must hold the dev_base or rtnl lock
1179 * while allocating the name and adding the device in order to avoid
1181 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1182 * Returns the number of the unit assigned or a negative errno code.
1185 int dev_alloc_name(struct net_device *dev, const char *name)
1187 return dev_alloc_name_ns(dev_net(dev), dev, name);
1189 EXPORT_SYMBOL(dev_alloc_name);
1191 static int dev_get_valid_name(struct net *net, struct net_device *dev,
1197 ret = dev_prep_valid_name(net, dev, name, buf);
1199 strscpy(dev->name, buf, IFNAMSIZ);
1204 * dev_change_name - change name of a device
1206 * @newname: name (or format string) must be at least IFNAMSIZ
1208 * Change name of a device, can pass format strings "eth%d".
1211 int dev_change_name(struct net_device *dev, const char *newname)
1213 unsigned char old_assign_type;
1214 char oldname[IFNAMSIZ];
1220 BUG_ON(!dev_net(dev));
1224 down_write(&devnet_rename_sem);
1226 if (strncmp(newname, dev->name, IFNAMSIZ) == 0) {
1227 up_write(&devnet_rename_sem);
1231 memcpy(oldname, dev->name, IFNAMSIZ);
1233 err = dev_get_valid_name(net, dev, newname);
1235 up_write(&devnet_rename_sem);
1239 if (oldname[0] && !strchr(oldname, '%'))
1240 netdev_info(dev, "renamed from %s%s\n", oldname,
1241 dev->flags & IFF_UP ? " (while UP)" : "");
1243 old_assign_type = dev->name_assign_type;
1244 dev->name_assign_type = NET_NAME_RENAMED;
1247 ret = device_rename(&dev->dev, dev->name);
1249 memcpy(dev->name, oldname, IFNAMSIZ);
1250 dev->name_assign_type = old_assign_type;
1251 up_write(&devnet_rename_sem);
1255 up_write(&devnet_rename_sem);
1257 netdev_adjacent_rename_links(dev, oldname);
1259 write_lock(&dev_base_lock);
1260 netdev_name_node_del(dev->name_node);
1261 write_unlock(&dev_base_lock);
1265 write_lock(&dev_base_lock);
1266 netdev_name_node_add(net, dev->name_node);
1267 write_unlock(&dev_base_lock);
1269 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev);
1270 ret = notifier_to_errno(ret);
1273 /* err >= 0 after dev_alloc_name() or stores the first errno */
1276 down_write(&devnet_rename_sem);
1277 memcpy(dev->name, oldname, IFNAMSIZ);
1278 memcpy(oldname, newname, IFNAMSIZ);
1279 dev->name_assign_type = old_assign_type;
1280 old_assign_type = NET_NAME_RENAMED;
1283 netdev_err(dev, "name change rollback failed: %d\n",
1292 * dev_set_alias - change ifalias of a device
1294 * @alias: name up to IFALIASZ
1295 * @len: limit of bytes to copy from info
1297 * Set ifalias for a device,
1299 int dev_set_alias(struct net_device *dev, const char *alias, size_t len)
1301 struct dev_ifalias *new_alias = NULL;
1303 if (len >= IFALIASZ)
1307 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL);
1311 memcpy(new_alias->ifalias, alias, len);
1312 new_alias->ifalias[len] = 0;
1315 mutex_lock(&ifalias_mutex);
1316 new_alias = rcu_replace_pointer(dev->ifalias, new_alias,
1317 mutex_is_locked(&ifalias_mutex));
1318 mutex_unlock(&ifalias_mutex);
1321 kfree_rcu(new_alias, rcuhead);
1325 EXPORT_SYMBOL(dev_set_alias);
1328 * dev_get_alias - get ifalias of a device
1330 * @name: buffer to store name of ifalias
1331 * @len: size of buffer
1333 * get ifalias for a device. Caller must make sure dev cannot go
1334 * away, e.g. rcu read lock or own a reference count to device.
1336 int dev_get_alias(const struct net_device *dev, char *name, size_t len)
1338 const struct dev_ifalias *alias;
1342 alias = rcu_dereference(dev->ifalias);
1344 ret = snprintf(name, len, "%s", alias->ifalias);
1351 * netdev_features_change - device changes features
1352 * @dev: device to cause notification
1354 * Called to indicate a device has changed features.
1356 void netdev_features_change(struct net_device *dev)
1358 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev);
1360 EXPORT_SYMBOL(netdev_features_change);
1363 * netdev_state_change - device changes state
1364 * @dev: device to cause notification
1366 * Called to indicate a device has changed state. This function calls
1367 * the notifier chains for netdev_chain and sends a NEWLINK message
1368 * to the routing socket.
1370 void netdev_state_change(struct net_device *dev)
1372 if (dev->flags & IFF_UP) {
1373 struct netdev_notifier_change_info change_info = {
1377 call_netdevice_notifiers_info(NETDEV_CHANGE,
1379 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL, 0, NULL);
1382 EXPORT_SYMBOL(netdev_state_change);
1385 * __netdev_notify_peers - notify network peers about existence of @dev,
1386 * to be called when rtnl lock is already held.
1387 * @dev: network device
1389 * Generate traffic such that interested network peers are aware of
1390 * @dev, such as by generating a gratuitous ARP. This may be used when
1391 * a device wants to inform the rest of the network about some sort of
1392 * reconfiguration such as a failover event or virtual machine
1395 void __netdev_notify_peers(struct net_device *dev)
1398 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
1399 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev);
1401 EXPORT_SYMBOL(__netdev_notify_peers);
1404 * netdev_notify_peers - notify network peers about existence of @dev
1405 * @dev: network device
1407 * Generate traffic such that interested network peers are aware of
1408 * @dev, such as by generating a gratuitous ARP. This may be used when
1409 * a device wants to inform the rest of the network about some sort of
1410 * reconfiguration such as a failover event or virtual machine
1413 void netdev_notify_peers(struct net_device *dev)
1416 __netdev_notify_peers(dev);
1419 EXPORT_SYMBOL(netdev_notify_peers);
1421 static int napi_threaded_poll(void *data);
1423 static int napi_kthread_create(struct napi_struct *n)
1427 /* Create and wake up the kthread once to put it in
1428 * TASK_INTERRUPTIBLE mode to avoid the blocked task
1429 * warning and work with loadavg.
1431 n->thread = kthread_run(napi_threaded_poll, n, "napi/%s-%d",
1432 n->dev->name, n->napi_id);
1433 if (IS_ERR(n->thread)) {
1434 err = PTR_ERR(n->thread);
1435 pr_err("kthread_run failed with err %d\n", err);
1442 static int __dev_open(struct net_device *dev, struct netlink_ext_ack *extack)
1444 const struct net_device_ops *ops = dev->netdev_ops;
1448 dev_addr_check(dev);
1450 if (!netif_device_present(dev)) {
1451 /* may be detached because parent is runtime-suspended */
1452 if (dev->dev.parent)
1453 pm_runtime_resume(dev->dev.parent);
1454 if (!netif_device_present(dev))
1458 /* Block netpoll from trying to do any rx path servicing.
1459 * If we don't do this there is a chance ndo_poll_controller
1460 * or ndo_poll may be running while we open the device
1462 netpoll_poll_disable(dev);
1464 ret = call_netdevice_notifiers_extack(NETDEV_PRE_UP, dev, extack);
1465 ret = notifier_to_errno(ret);
1469 set_bit(__LINK_STATE_START, &dev->state);
1471 if (ops->ndo_validate_addr)
1472 ret = ops->ndo_validate_addr(dev);
1474 if (!ret && ops->ndo_open)
1475 ret = ops->ndo_open(dev);
1477 netpoll_poll_enable(dev);
1480 clear_bit(__LINK_STATE_START, &dev->state);
1482 dev->flags |= IFF_UP;
1483 dev_set_rx_mode(dev);
1485 add_device_randomness(dev->dev_addr, dev->addr_len);
1492 * dev_open - prepare an interface for use.
1493 * @dev: device to open
1494 * @extack: netlink extended ack
1496 * Takes a device from down to up state. The device's private open
1497 * function is invoked and then the multicast lists are loaded. Finally
1498 * the device is moved into the up state and a %NETDEV_UP message is
1499 * sent to the netdev notifier chain.
1501 * Calling this function on an active interface is a nop. On a failure
1502 * a negative errno code is returned.
1504 int dev_open(struct net_device *dev, struct netlink_ext_ack *extack)
1508 if (dev->flags & IFF_UP)
1511 ret = __dev_open(dev, extack);
1515 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL);
1516 call_netdevice_notifiers(NETDEV_UP, dev);
1520 EXPORT_SYMBOL(dev_open);
1522 static void __dev_close_many(struct list_head *head)
1524 struct net_device *dev;
1529 list_for_each_entry(dev, head, close_list) {
1530 /* Temporarily disable netpoll until the interface is down */
1531 netpoll_poll_disable(dev);
1533 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev);
1535 clear_bit(__LINK_STATE_START, &dev->state);
1537 /* Synchronize to scheduled poll. We cannot touch poll list, it
1538 * can be even on different cpu. So just clear netif_running().
1540 * dev->stop() will invoke napi_disable() on all of it's
1541 * napi_struct instances on this device.
1543 smp_mb__after_atomic(); /* Commit netif_running(). */
1546 dev_deactivate_many(head);
1548 list_for_each_entry(dev, head, close_list) {
1549 const struct net_device_ops *ops = dev->netdev_ops;
1552 * Call the device specific close. This cannot fail.
1553 * Only if device is UP
1555 * We allow it to be called even after a DETACH hot-plug
1561 dev->flags &= ~IFF_UP;
1562 netpoll_poll_enable(dev);
1566 static void __dev_close(struct net_device *dev)
1570 list_add(&dev->close_list, &single);
1571 __dev_close_many(&single);
1575 void dev_close_many(struct list_head *head, bool unlink)
1577 struct net_device *dev, *tmp;
1579 /* Remove the devices that don't need to be closed */
1580 list_for_each_entry_safe(dev, tmp, head, close_list)
1581 if (!(dev->flags & IFF_UP))
1582 list_del_init(&dev->close_list);
1584 __dev_close_many(head);
1586 list_for_each_entry_safe(dev, tmp, head, close_list) {
1587 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL);
1588 call_netdevice_notifiers(NETDEV_DOWN, dev);
1590 list_del_init(&dev->close_list);
1593 EXPORT_SYMBOL(dev_close_many);
1596 * dev_close - shutdown an interface.
1597 * @dev: device to shutdown
1599 * This function moves an active device into down state. A
1600 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device
1601 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier
1604 void dev_close(struct net_device *dev)
1606 if (dev->flags & IFF_UP) {
1609 list_add(&dev->close_list, &single);
1610 dev_close_many(&single, true);
1614 EXPORT_SYMBOL(dev_close);
1618 * dev_disable_lro - disable Large Receive Offload on a device
1621 * Disable Large Receive Offload (LRO) on a net device. Must be
1622 * called under RTNL. This is needed if received packets may be
1623 * forwarded to another interface.
1625 void dev_disable_lro(struct net_device *dev)
1627 struct net_device *lower_dev;
1628 struct list_head *iter;
1630 dev->wanted_features &= ~NETIF_F_LRO;
1631 netdev_update_features(dev);
1633 if (unlikely(dev->features & NETIF_F_LRO))
1634 netdev_WARN(dev, "failed to disable LRO!\n");
1636 netdev_for_each_lower_dev(dev, lower_dev, iter)
1637 dev_disable_lro(lower_dev);
1639 EXPORT_SYMBOL(dev_disable_lro);
1642 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device
1645 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be
1646 * called under RTNL. This is needed if Generic XDP is installed on
1649 static void dev_disable_gro_hw(struct net_device *dev)
1651 dev->wanted_features &= ~NETIF_F_GRO_HW;
1652 netdev_update_features(dev);
1654 if (unlikely(dev->features & NETIF_F_GRO_HW))
1655 netdev_WARN(dev, "failed to disable GRO_HW!\n");
1658 const char *netdev_cmd_to_name(enum netdev_cmd cmd)
1661 case NETDEV_##val: \
1662 return "NETDEV_" __stringify(val);
1664 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER)
1665 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE)
1666 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE)
1667 N(POST_INIT) N(PRE_UNINIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN)
1668 N(CHANGEUPPER) N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA)
1669 N(BONDING_INFO) N(PRECHANGEUPPER) N(CHANGELOWERSTATE)
1670 N(UDP_TUNNEL_PUSH_INFO) N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN)
1671 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO)
1672 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO)
1673 N(PRE_CHANGEADDR) N(OFFLOAD_XSTATS_ENABLE) N(OFFLOAD_XSTATS_DISABLE)
1674 N(OFFLOAD_XSTATS_REPORT_USED) N(OFFLOAD_XSTATS_REPORT_DELTA)
1678 return "UNKNOWN_NETDEV_EVENT";
1680 EXPORT_SYMBOL_GPL(netdev_cmd_to_name);
1682 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val,
1683 struct net_device *dev)
1685 struct netdev_notifier_info info = {
1689 return nb->notifier_call(nb, val, &info);
1692 static int call_netdevice_register_notifiers(struct notifier_block *nb,
1693 struct net_device *dev)
1697 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev);
1698 err = notifier_to_errno(err);
1702 if (!(dev->flags & IFF_UP))
1705 call_netdevice_notifier(nb, NETDEV_UP, dev);
1709 static void call_netdevice_unregister_notifiers(struct notifier_block *nb,
1710 struct net_device *dev)
1712 if (dev->flags & IFF_UP) {
1713 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1715 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1717 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1720 static int call_netdevice_register_net_notifiers(struct notifier_block *nb,
1723 struct net_device *dev;
1726 for_each_netdev(net, dev) {
1727 err = call_netdevice_register_notifiers(nb, dev);
1734 for_each_netdev_continue_reverse(net, dev)
1735 call_netdevice_unregister_notifiers(nb, dev);
1739 static void call_netdevice_unregister_net_notifiers(struct notifier_block *nb,
1742 struct net_device *dev;
1744 for_each_netdev(net, dev)
1745 call_netdevice_unregister_notifiers(nb, dev);
1748 static int dev_boot_phase = 1;
1751 * register_netdevice_notifier - register a network notifier block
1754 * Register a notifier to be called when network device events occur.
1755 * The notifier passed is linked into the kernel structures and must
1756 * not be reused until it has been unregistered. A negative errno code
1757 * is returned on a failure.
1759 * When registered all registration and up events are replayed
1760 * to the new notifier to allow device to have a race free
1761 * view of the network device list.
1764 int register_netdevice_notifier(struct notifier_block *nb)
1769 /* Close race with setup_net() and cleanup_net() */
1770 down_write(&pernet_ops_rwsem);
1772 err = raw_notifier_chain_register(&netdev_chain, nb);
1778 err = call_netdevice_register_net_notifiers(nb, net);
1785 up_write(&pernet_ops_rwsem);
1789 for_each_net_continue_reverse(net)
1790 call_netdevice_unregister_net_notifiers(nb, net);
1792 raw_notifier_chain_unregister(&netdev_chain, nb);
1795 EXPORT_SYMBOL(register_netdevice_notifier);
1798 * unregister_netdevice_notifier - unregister a network notifier block
1801 * Unregister a notifier previously registered by
1802 * register_netdevice_notifier(). The notifier is unlinked into the
1803 * kernel structures and may then be reused. A negative errno code
1804 * is returned on a failure.
1806 * After unregistering unregister and down device events are synthesized
1807 * for all devices on the device list to the removed notifier to remove
1808 * the need for special case cleanup code.
1811 int unregister_netdevice_notifier(struct notifier_block *nb)
1816 /* Close race with setup_net() and cleanup_net() */
1817 down_write(&pernet_ops_rwsem);
1819 err = raw_notifier_chain_unregister(&netdev_chain, nb);
1824 call_netdevice_unregister_net_notifiers(nb, net);
1828 up_write(&pernet_ops_rwsem);
1831 EXPORT_SYMBOL(unregister_netdevice_notifier);
1833 static int __register_netdevice_notifier_net(struct net *net,
1834 struct notifier_block *nb,
1835 bool ignore_call_fail)
1839 err = raw_notifier_chain_register(&net->netdev_chain, nb);
1845 err = call_netdevice_register_net_notifiers(nb, net);
1846 if (err && !ignore_call_fail)
1847 goto chain_unregister;
1852 raw_notifier_chain_unregister(&net->netdev_chain, nb);
1856 static int __unregister_netdevice_notifier_net(struct net *net,
1857 struct notifier_block *nb)
1861 err = raw_notifier_chain_unregister(&net->netdev_chain, nb);
1865 call_netdevice_unregister_net_notifiers(nb, net);
1870 * register_netdevice_notifier_net - register a per-netns network notifier block
1871 * @net: network namespace
1874 * Register a notifier to be called when network device events occur.
1875 * The notifier passed is linked into the kernel structures and must
1876 * not be reused until it has been unregistered. A negative errno code
1877 * is returned on a failure.
1879 * When registered all registration and up events are replayed
1880 * to the new notifier to allow device to have a race free
1881 * view of the network device list.
1884 int register_netdevice_notifier_net(struct net *net, struct notifier_block *nb)
1889 err = __register_netdevice_notifier_net(net, nb, false);
1893 EXPORT_SYMBOL(register_netdevice_notifier_net);
1896 * unregister_netdevice_notifier_net - unregister a per-netns
1897 * network notifier block
1898 * @net: network namespace
1901 * Unregister a notifier previously registered by
1902 * register_netdevice_notifier_net(). The notifier is unlinked from the
1903 * kernel structures and may then be reused. A negative errno code
1904 * is returned on a failure.
1906 * After unregistering unregister and down device events are synthesized
1907 * for all devices on the device list to the removed notifier to remove
1908 * the need for special case cleanup code.
1911 int unregister_netdevice_notifier_net(struct net *net,
1912 struct notifier_block *nb)
1917 err = __unregister_netdevice_notifier_net(net, nb);
1921 EXPORT_SYMBOL(unregister_netdevice_notifier_net);
1923 static void __move_netdevice_notifier_net(struct net *src_net,
1924 struct net *dst_net,
1925 struct notifier_block *nb)
1927 __unregister_netdevice_notifier_net(src_net, nb);
1928 __register_netdevice_notifier_net(dst_net, nb, true);
1931 int register_netdevice_notifier_dev_net(struct net_device *dev,
1932 struct notifier_block *nb,
1933 struct netdev_net_notifier *nn)
1938 err = __register_netdevice_notifier_net(dev_net(dev), nb, false);
1941 list_add(&nn->list, &dev->net_notifier_list);
1946 EXPORT_SYMBOL(register_netdevice_notifier_dev_net);
1948 int unregister_netdevice_notifier_dev_net(struct net_device *dev,
1949 struct notifier_block *nb,
1950 struct netdev_net_notifier *nn)
1955 list_del(&nn->list);
1956 err = __unregister_netdevice_notifier_net(dev_net(dev), nb);
1960 EXPORT_SYMBOL(unregister_netdevice_notifier_dev_net);
1962 static void move_netdevice_notifiers_dev_net(struct net_device *dev,
1965 struct netdev_net_notifier *nn;
1967 list_for_each_entry(nn, &dev->net_notifier_list, list)
1968 __move_netdevice_notifier_net(dev_net(dev), net, nn->nb);
1972 * call_netdevice_notifiers_info - call all network notifier blocks
1973 * @val: value passed unmodified to notifier function
1974 * @info: notifier information data
1976 * Call all network notifier blocks. Parameters and return value
1977 * are as for raw_notifier_call_chain().
1980 int call_netdevice_notifiers_info(unsigned long val,
1981 struct netdev_notifier_info *info)
1983 struct net *net = dev_net(info->dev);
1988 /* Run per-netns notifier block chain first, then run the global one.
1989 * Hopefully, one day, the global one is going to be removed after
1990 * all notifier block registrators get converted to be per-netns.
1992 ret = raw_notifier_call_chain(&net->netdev_chain, val, info);
1993 if (ret & NOTIFY_STOP_MASK)
1995 return raw_notifier_call_chain(&netdev_chain, val, info);
1999 * call_netdevice_notifiers_info_robust - call per-netns notifier blocks
2000 * for and rollback on error
2001 * @val_up: value passed unmodified to notifier function
2002 * @val_down: value passed unmodified to the notifier function when
2003 * recovering from an error on @val_up
2004 * @info: notifier information data
2006 * Call all per-netns network notifier blocks, but not notifier blocks on
2007 * the global notifier chain. Parameters and return value are as for
2008 * raw_notifier_call_chain_robust().
2012 call_netdevice_notifiers_info_robust(unsigned long val_up,
2013 unsigned long val_down,
2014 struct netdev_notifier_info *info)
2016 struct net *net = dev_net(info->dev);
2020 return raw_notifier_call_chain_robust(&net->netdev_chain,
2021 val_up, val_down, info);
2024 static int call_netdevice_notifiers_extack(unsigned long val,
2025 struct net_device *dev,
2026 struct netlink_ext_ack *extack)
2028 struct netdev_notifier_info info = {
2033 return call_netdevice_notifiers_info(val, &info);
2037 * call_netdevice_notifiers - call all network notifier blocks
2038 * @val: value passed unmodified to notifier function
2039 * @dev: net_device pointer passed unmodified to notifier function
2041 * Call all network notifier blocks. Parameters and return value
2042 * are as for raw_notifier_call_chain().
2045 int call_netdevice_notifiers(unsigned long val, struct net_device *dev)
2047 return call_netdevice_notifiers_extack(val, dev, NULL);
2049 EXPORT_SYMBOL(call_netdevice_notifiers);
2052 * call_netdevice_notifiers_mtu - call all network notifier blocks
2053 * @val: value passed unmodified to notifier function
2054 * @dev: net_device pointer passed unmodified to notifier function
2055 * @arg: additional u32 argument passed to the notifier function
2057 * Call all network notifier blocks. Parameters and return value
2058 * are as for raw_notifier_call_chain().
2060 static int call_netdevice_notifiers_mtu(unsigned long val,
2061 struct net_device *dev, u32 arg)
2063 struct netdev_notifier_info_ext info = {
2068 BUILD_BUG_ON(offsetof(struct netdev_notifier_info_ext, info) != 0);
2070 return call_netdevice_notifiers_info(val, &info.info);
2073 #ifdef CONFIG_NET_INGRESS
2074 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key);
2076 void net_inc_ingress_queue(void)
2078 static_branch_inc(&ingress_needed_key);
2080 EXPORT_SYMBOL_GPL(net_inc_ingress_queue);
2082 void net_dec_ingress_queue(void)
2084 static_branch_dec(&ingress_needed_key);
2086 EXPORT_SYMBOL_GPL(net_dec_ingress_queue);
2089 #ifdef CONFIG_NET_EGRESS
2090 static DEFINE_STATIC_KEY_FALSE(egress_needed_key);
2092 void net_inc_egress_queue(void)
2094 static_branch_inc(&egress_needed_key);
2096 EXPORT_SYMBOL_GPL(net_inc_egress_queue);
2098 void net_dec_egress_queue(void)
2100 static_branch_dec(&egress_needed_key);
2102 EXPORT_SYMBOL_GPL(net_dec_egress_queue);
2105 DEFINE_STATIC_KEY_FALSE(netstamp_needed_key);
2106 EXPORT_SYMBOL(netstamp_needed_key);
2107 #ifdef CONFIG_JUMP_LABEL
2108 static atomic_t netstamp_needed_deferred;
2109 static atomic_t netstamp_wanted;
2110 static void netstamp_clear(struct work_struct *work)
2112 int deferred = atomic_xchg(&netstamp_needed_deferred, 0);
2115 wanted = atomic_add_return(deferred, &netstamp_wanted);
2117 static_branch_enable(&netstamp_needed_key);
2119 static_branch_disable(&netstamp_needed_key);
2121 static DECLARE_WORK(netstamp_work, netstamp_clear);
2124 void net_enable_timestamp(void)
2126 #ifdef CONFIG_JUMP_LABEL
2127 int wanted = atomic_read(&netstamp_wanted);
2129 while (wanted > 0) {
2130 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted + 1))
2133 atomic_inc(&netstamp_needed_deferred);
2134 schedule_work(&netstamp_work);
2136 static_branch_inc(&netstamp_needed_key);
2139 EXPORT_SYMBOL(net_enable_timestamp);
2141 void net_disable_timestamp(void)
2143 #ifdef CONFIG_JUMP_LABEL
2144 int wanted = atomic_read(&netstamp_wanted);
2146 while (wanted > 1) {
2147 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted - 1))
2150 atomic_dec(&netstamp_needed_deferred);
2151 schedule_work(&netstamp_work);
2153 static_branch_dec(&netstamp_needed_key);
2156 EXPORT_SYMBOL(net_disable_timestamp);
2158 static inline void net_timestamp_set(struct sk_buff *skb)
2161 skb->mono_delivery_time = 0;
2162 if (static_branch_unlikely(&netstamp_needed_key))
2163 skb->tstamp = ktime_get_real();
2166 #define net_timestamp_check(COND, SKB) \
2167 if (static_branch_unlikely(&netstamp_needed_key)) { \
2168 if ((COND) && !(SKB)->tstamp) \
2169 (SKB)->tstamp = ktime_get_real(); \
2172 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb)
2174 return __is_skb_forwardable(dev, skb, true);
2176 EXPORT_SYMBOL_GPL(is_skb_forwardable);
2178 static int __dev_forward_skb2(struct net_device *dev, struct sk_buff *skb,
2181 int ret = ____dev_forward_skb(dev, skb, check_mtu);
2184 skb->protocol = eth_type_trans(skb, dev);
2185 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
2191 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
2193 return __dev_forward_skb2(dev, skb, true);
2195 EXPORT_SYMBOL_GPL(__dev_forward_skb);
2198 * dev_forward_skb - loopback an skb to another netif
2200 * @dev: destination network device
2201 * @skb: buffer to forward
2204 * NET_RX_SUCCESS (no congestion)
2205 * NET_RX_DROP (packet was dropped, but freed)
2207 * dev_forward_skb can be used for injecting an skb from the
2208 * start_xmit function of one device into the receive queue
2209 * of another device.
2211 * The receiving device may be in another namespace, so
2212 * we have to clear all information in the skb that could
2213 * impact namespace isolation.
2215 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
2217 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb);
2219 EXPORT_SYMBOL_GPL(dev_forward_skb);
2221 int dev_forward_skb_nomtu(struct net_device *dev, struct sk_buff *skb)
2223 return __dev_forward_skb2(dev, skb, false) ?: netif_rx_internal(skb);
2226 static inline int deliver_skb(struct sk_buff *skb,
2227 struct packet_type *pt_prev,
2228 struct net_device *orig_dev)
2230 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
2232 refcount_inc(&skb->users);
2233 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
2236 static inline void deliver_ptype_list_skb(struct sk_buff *skb,
2237 struct packet_type **pt,
2238 struct net_device *orig_dev,
2240 struct list_head *ptype_list)
2242 struct packet_type *ptype, *pt_prev = *pt;
2244 list_for_each_entry_rcu(ptype, ptype_list, list) {
2245 if (ptype->type != type)
2248 deliver_skb(skb, pt_prev, orig_dev);
2254 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb)
2256 if (!ptype->af_packet_priv || !skb->sk)
2259 if (ptype->id_match)
2260 return ptype->id_match(ptype, skb->sk);
2261 else if ((struct sock *)ptype->af_packet_priv == skb->sk)
2268 * dev_nit_active - return true if any network interface taps are in use
2270 * @dev: network device to check for the presence of taps
2272 bool dev_nit_active(struct net_device *dev)
2274 return !list_empty(&ptype_all) || !list_empty(&dev->ptype_all);
2276 EXPORT_SYMBOL_GPL(dev_nit_active);
2279 * Support routine. Sends outgoing frames to any network
2280 * taps currently in use.
2283 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
2285 struct packet_type *ptype;
2286 struct sk_buff *skb2 = NULL;
2287 struct packet_type *pt_prev = NULL;
2288 struct list_head *ptype_list = &ptype_all;
2292 list_for_each_entry_rcu(ptype, ptype_list, list) {
2293 if (ptype->ignore_outgoing)
2296 /* Never send packets back to the socket
2297 * they originated from - MvS (miquels@drinkel.ow.org)
2299 if (skb_loop_sk(ptype, skb))
2303 deliver_skb(skb2, pt_prev, skb->dev);
2308 /* need to clone skb, done only once */
2309 skb2 = skb_clone(skb, GFP_ATOMIC);
2313 net_timestamp_set(skb2);
2315 /* skb->nh should be correctly
2316 * set by sender, so that the second statement is
2317 * just protection against buggy protocols.
2319 skb_reset_mac_header(skb2);
2321 if (skb_network_header(skb2) < skb2->data ||
2322 skb_network_header(skb2) > skb_tail_pointer(skb2)) {
2323 net_crit_ratelimited("protocol %04x is buggy, dev %s\n",
2324 ntohs(skb2->protocol),
2326 skb_reset_network_header(skb2);
2329 skb2->transport_header = skb2->network_header;
2330 skb2->pkt_type = PACKET_OUTGOING;
2334 if (ptype_list == &ptype_all) {
2335 ptype_list = &dev->ptype_all;
2340 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC))
2341 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
2347 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit);
2350 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change
2351 * @dev: Network device
2352 * @txq: number of queues available
2354 * If real_num_tx_queues is changed the tc mappings may no longer be
2355 * valid. To resolve this verify the tc mapping remains valid and if
2356 * not NULL the mapping. With no priorities mapping to this
2357 * offset/count pair it will no longer be used. In the worst case TC0
2358 * is invalid nothing can be done so disable priority mappings. If is
2359 * expected that drivers will fix this mapping if they can before
2360 * calling netif_set_real_num_tx_queues.
2362 static void netif_setup_tc(struct net_device *dev, unsigned int txq)
2365 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2367 /* If TC0 is invalidated disable TC mapping */
2368 if (tc->offset + tc->count > txq) {
2369 netdev_warn(dev, "Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n");
2374 /* Invalidated prio to tc mappings set to TC0 */
2375 for (i = 1; i < TC_BITMASK + 1; i++) {
2376 int q = netdev_get_prio_tc_map(dev, i);
2378 tc = &dev->tc_to_txq[q];
2379 if (tc->offset + tc->count > txq) {
2380 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",
2382 netdev_set_prio_tc_map(dev, i, 0);
2387 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq)
2390 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2393 /* walk through the TCs and see if it falls into any of them */
2394 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) {
2395 if ((txq - tc->offset) < tc->count)
2399 /* didn't find it, just return -1 to indicate no match */
2405 EXPORT_SYMBOL(netdev_txq_to_tc);
2408 static struct static_key xps_needed __read_mostly;
2409 static struct static_key xps_rxqs_needed __read_mostly;
2410 static DEFINE_MUTEX(xps_map_mutex);
2411 #define xmap_dereference(P) \
2412 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex))
2414 static bool remove_xps_queue(struct xps_dev_maps *dev_maps,
2415 struct xps_dev_maps *old_maps, int tci, u16 index)
2417 struct xps_map *map = NULL;
2420 map = xmap_dereference(dev_maps->attr_map[tci]);
2424 for (pos = map->len; pos--;) {
2425 if (map->queues[pos] != index)
2429 map->queues[pos] = map->queues[--map->len];
2434 RCU_INIT_POINTER(old_maps->attr_map[tci], NULL);
2435 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL);
2436 kfree_rcu(map, rcu);
2443 static bool remove_xps_queue_cpu(struct net_device *dev,
2444 struct xps_dev_maps *dev_maps,
2445 int cpu, u16 offset, u16 count)
2447 int num_tc = dev_maps->num_tc;
2448 bool active = false;
2451 for (tci = cpu * num_tc; num_tc--; tci++) {
2454 for (i = count, j = offset; i--; j++) {
2455 if (!remove_xps_queue(dev_maps, NULL, tci, j))
2465 static void reset_xps_maps(struct net_device *dev,
2466 struct xps_dev_maps *dev_maps,
2467 enum xps_map_type type)
2469 static_key_slow_dec_cpuslocked(&xps_needed);
2470 if (type == XPS_RXQS)
2471 static_key_slow_dec_cpuslocked(&xps_rxqs_needed);
2473 RCU_INIT_POINTER(dev->xps_maps[type], NULL);
2475 kfree_rcu(dev_maps, rcu);
2478 static void clean_xps_maps(struct net_device *dev, enum xps_map_type type,
2479 u16 offset, u16 count)
2481 struct xps_dev_maps *dev_maps;
2482 bool active = false;
2485 dev_maps = xmap_dereference(dev->xps_maps[type]);
2489 for (j = 0; j < dev_maps->nr_ids; j++)
2490 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, count);
2492 reset_xps_maps(dev, dev_maps, type);
2494 if (type == XPS_CPUS) {
2495 for (i = offset + (count - 1); count--; i--)
2496 netdev_queue_numa_node_write(
2497 netdev_get_tx_queue(dev, i), NUMA_NO_NODE);
2501 static void netif_reset_xps_queues(struct net_device *dev, u16 offset,
2504 if (!static_key_false(&xps_needed))
2508 mutex_lock(&xps_map_mutex);
2510 if (static_key_false(&xps_rxqs_needed))
2511 clean_xps_maps(dev, XPS_RXQS, offset, count);
2513 clean_xps_maps(dev, XPS_CPUS, offset, count);
2515 mutex_unlock(&xps_map_mutex);
2519 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index)
2521 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index);
2524 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index,
2525 u16 index, bool is_rxqs_map)
2527 struct xps_map *new_map;
2528 int alloc_len = XPS_MIN_MAP_ALLOC;
2531 for (pos = 0; map && pos < map->len; pos++) {
2532 if (map->queues[pos] != index)
2537 /* Need to add tx-queue to this CPU's/rx-queue's existing map */
2539 if (pos < map->alloc_len)
2542 alloc_len = map->alloc_len * 2;
2545 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's
2549 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL);
2551 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL,
2552 cpu_to_node(attr_index));
2556 for (i = 0; i < pos; i++)
2557 new_map->queues[i] = map->queues[i];
2558 new_map->alloc_len = alloc_len;
2564 /* Copy xps maps at a given index */
2565 static void xps_copy_dev_maps(struct xps_dev_maps *dev_maps,
2566 struct xps_dev_maps *new_dev_maps, int index,
2567 int tc, bool skip_tc)
2569 int i, tci = index * dev_maps->num_tc;
2570 struct xps_map *map;
2572 /* copy maps belonging to foreign traffic classes */
2573 for (i = 0; i < dev_maps->num_tc; i++, tci++) {
2574 if (i == tc && skip_tc)
2577 /* fill in the new device map from the old device map */
2578 map = xmap_dereference(dev_maps->attr_map[tci]);
2579 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map);
2583 /* Must be called under cpus_read_lock */
2584 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask,
2585 u16 index, enum xps_map_type type)
2587 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL, *old_dev_maps = NULL;
2588 const unsigned long *online_mask = NULL;
2589 bool active = false, copy = false;
2590 int i, j, tci, numa_node_id = -2;
2591 int maps_sz, num_tc = 1, tc = 0;
2592 struct xps_map *map, *new_map;
2593 unsigned int nr_ids;
2595 WARN_ON_ONCE(index >= dev->num_tx_queues);
2598 /* Do not allow XPS on subordinate device directly */
2599 num_tc = dev->num_tc;
2603 /* If queue belongs to subordinate dev use its map */
2604 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev;
2606 tc = netdev_txq_to_tc(dev, index);
2611 mutex_lock(&xps_map_mutex);
2613 dev_maps = xmap_dereference(dev->xps_maps[type]);
2614 if (type == XPS_RXQS) {
2615 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues);
2616 nr_ids = dev->num_rx_queues;
2618 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc);
2619 if (num_possible_cpus() > 1)
2620 online_mask = cpumask_bits(cpu_online_mask);
2621 nr_ids = nr_cpu_ids;
2624 if (maps_sz < L1_CACHE_BYTES)
2625 maps_sz = L1_CACHE_BYTES;
2627 /* The old dev_maps could be larger or smaller than the one we're
2628 * setting up now, as dev->num_tc or nr_ids could have been updated in
2629 * between. We could try to be smart, but let's be safe instead and only
2630 * copy foreign traffic classes if the two map sizes match.
2633 dev_maps->num_tc == num_tc && dev_maps->nr_ids == nr_ids)
2636 /* allocate memory for queue storage */
2637 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids),
2639 if (!new_dev_maps) {
2640 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL);
2641 if (!new_dev_maps) {
2642 mutex_unlock(&xps_map_mutex);
2646 new_dev_maps->nr_ids = nr_ids;
2647 new_dev_maps->num_tc = num_tc;
2650 tci = j * num_tc + tc;
2651 map = copy ? xmap_dereference(dev_maps->attr_map[tci]) : NULL;
2653 map = expand_xps_map(map, j, index, type == XPS_RXQS);
2657 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map);
2661 goto out_no_new_maps;
2664 /* Increment static keys at most once per type */
2665 static_key_slow_inc_cpuslocked(&xps_needed);
2666 if (type == XPS_RXQS)
2667 static_key_slow_inc_cpuslocked(&xps_rxqs_needed);
2670 for (j = 0; j < nr_ids; j++) {
2671 bool skip_tc = false;
2673 tci = j * num_tc + tc;
2674 if (netif_attr_test_mask(j, mask, nr_ids) &&
2675 netif_attr_test_online(j, online_mask, nr_ids)) {
2676 /* add tx-queue to CPU/rx-queue maps */
2681 map = xmap_dereference(new_dev_maps->attr_map[tci]);
2682 while ((pos < map->len) && (map->queues[pos] != index))
2685 if (pos == map->len)
2686 map->queues[map->len++] = index;
2688 if (type == XPS_CPUS) {
2689 if (numa_node_id == -2)
2690 numa_node_id = cpu_to_node(j);
2691 else if (numa_node_id != cpu_to_node(j))
2698 xps_copy_dev_maps(dev_maps, new_dev_maps, j, tc,
2702 rcu_assign_pointer(dev->xps_maps[type], new_dev_maps);
2704 /* Cleanup old maps */
2706 goto out_no_old_maps;
2708 for (j = 0; j < dev_maps->nr_ids; j++) {
2709 for (i = num_tc, tci = j * dev_maps->num_tc; i--; tci++) {
2710 map = xmap_dereference(dev_maps->attr_map[tci]);
2715 new_map = xmap_dereference(new_dev_maps->attr_map[tci]);
2720 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL);
2721 kfree_rcu(map, rcu);
2725 old_dev_maps = dev_maps;
2728 dev_maps = new_dev_maps;
2732 if (type == XPS_CPUS)
2733 /* update Tx queue numa node */
2734 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index),
2735 (numa_node_id >= 0) ?
2736 numa_node_id : NUMA_NO_NODE);
2741 /* removes tx-queue from unused CPUs/rx-queues */
2742 for (j = 0; j < dev_maps->nr_ids; j++) {
2743 tci = j * dev_maps->num_tc;
2745 for (i = 0; i < dev_maps->num_tc; i++, tci++) {
2747 netif_attr_test_mask(j, mask, dev_maps->nr_ids) &&
2748 netif_attr_test_online(j, online_mask, dev_maps->nr_ids))
2751 active |= remove_xps_queue(dev_maps,
2752 copy ? old_dev_maps : NULL,
2758 kfree_rcu(old_dev_maps, rcu);
2760 /* free map if not active */
2762 reset_xps_maps(dev, dev_maps, type);
2765 mutex_unlock(&xps_map_mutex);
2769 /* remove any maps that we added */
2770 for (j = 0; j < nr_ids; j++) {
2771 for (i = num_tc, tci = j * num_tc; i--; tci++) {
2772 new_map = xmap_dereference(new_dev_maps->attr_map[tci]);
2774 xmap_dereference(dev_maps->attr_map[tci]) :
2776 if (new_map && new_map != map)
2781 mutex_unlock(&xps_map_mutex);
2783 kfree(new_dev_maps);
2786 EXPORT_SYMBOL_GPL(__netif_set_xps_queue);
2788 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask,
2794 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, XPS_CPUS);
2799 EXPORT_SYMBOL(netif_set_xps_queue);
2802 static void netdev_unbind_all_sb_channels(struct net_device *dev)
2804 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues];
2806 /* Unbind any subordinate channels */
2807 while (txq-- != &dev->_tx[0]) {
2809 netdev_unbind_sb_channel(dev, txq->sb_dev);
2813 void netdev_reset_tc(struct net_device *dev)
2816 netif_reset_xps_queues_gt(dev, 0);
2818 netdev_unbind_all_sb_channels(dev);
2820 /* Reset TC configuration of device */
2822 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq));
2823 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map));
2825 EXPORT_SYMBOL(netdev_reset_tc);
2827 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset)
2829 if (tc >= dev->num_tc)
2833 netif_reset_xps_queues(dev, offset, count);
2835 dev->tc_to_txq[tc].count = count;
2836 dev->tc_to_txq[tc].offset = offset;
2839 EXPORT_SYMBOL(netdev_set_tc_queue);
2841 int netdev_set_num_tc(struct net_device *dev, u8 num_tc)
2843 if (num_tc > TC_MAX_QUEUE)
2847 netif_reset_xps_queues_gt(dev, 0);
2849 netdev_unbind_all_sb_channels(dev);
2851 dev->num_tc = num_tc;
2854 EXPORT_SYMBOL(netdev_set_num_tc);
2856 void netdev_unbind_sb_channel(struct net_device *dev,
2857 struct net_device *sb_dev)
2859 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues];
2862 netif_reset_xps_queues_gt(sb_dev, 0);
2864 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq));
2865 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map));
2867 while (txq-- != &dev->_tx[0]) {
2868 if (txq->sb_dev == sb_dev)
2872 EXPORT_SYMBOL(netdev_unbind_sb_channel);
2874 int netdev_bind_sb_channel_queue(struct net_device *dev,
2875 struct net_device *sb_dev,
2876 u8 tc, u16 count, u16 offset)
2878 /* Make certain the sb_dev and dev are already configured */
2879 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc)
2882 /* We cannot hand out queues we don't have */
2883 if ((offset + count) > dev->real_num_tx_queues)
2886 /* Record the mapping */
2887 sb_dev->tc_to_txq[tc].count = count;
2888 sb_dev->tc_to_txq[tc].offset = offset;
2890 /* Provide a way for Tx queue to find the tc_to_txq map or
2891 * XPS map for itself.
2894 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev;
2898 EXPORT_SYMBOL(netdev_bind_sb_channel_queue);
2900 int netdev_set_sb_channel(struct net_device *dev, u16 channel)
2902 /* Do not use a multiqueue device to represent a subordinate channel */
2903 if (netif_is_multiqueue(dev))
2906 /* We allow channels 1 - 32767 to be used for subordinate channels.
2907 * Channel 0 is meant to be "native" mode and used only to represent
2908 * the main root device. We allow writing 0 to reset the device back
2909 * to normal mode after being used as a subordinate channel.
2911 if (channel > S16_MAX)
2914 dev->num_tc = -channel;
2918 EXPORT_SYMBOL(netdev_set_sb_channel);
2921 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues
2922 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed.
2924 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq)
2929 disabling = txq < dev->real_num_tx_queues;
2931 if (txq < 1 || txq > dev->num_tx_queues)
2934 if (dev->reg_state == NETREG_REGISTERED ||
2935 dev->reg_state == NETREG_UNREGISTERING) {
2938 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues,
2944 netif_setup_tc(dev, txq);
2946 dev_qdisc_change_real_num_tx(dev, txq);
2948 dev->real_num_tx_queues = txq;
2952 qdisc_reset_all_tx_gt(dev, txq);
2954 netif_reset_xps_queues_gt(dev, txq);
2958 dev->real_num_tx_queues = txq;
2963 EXPORT_SYMBOL(netif_set_real_num_tx_queues);
2967 * netif_set_real_num_rx_queues - set actual number of RX queues used
2968 * @dev: Network device
2969 * @rxq: Actual number of RX queues
2971 * This must be called either with the rtnl_lock held or before
2972 * registration of the net device. Returns 0 on success, or a
2973 * negative error code. If called before registration, it always
2976 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq)
2980 if (rxq < 1 || rxq > dev->num_rx_queues)
2983 if (dev->reg_state == NETREG_REGISTERED) {
2986 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues,
2992 dev->real_num_rx_queues = rxq;
2995 EXPORT_SYMBOL(netif_set_real_num_rx_queues);
2999 * netif_set_real_num_queues - set actual number of RX and TX queues used
3000 * @dev: Network device
3001 * @txq: Actual number of TX queues
3002 * @rxq: Actual number of RX queues
3004 * Set the real number of both TX and RX queues.
3005 * Does nothing if the number of queues is already correct.
3007 int netif_set_real_num_queues(struct net_device *dev,
3008 unsigned int txq, unsigned int rxq)
3010 unsigned int old_rxq = dev->real_num_rx_queues;
3013 if (txq < 1 || txq > dev->num_tx_queues ||
3014 rxq < 1 || rxq > dev->num_rx_queues)
3017 /* Start from increases, so the error path only does decreases -
3018 * decreases can't fail.
3020 if (rxq > dev->real_num_rx_queues) {
3021 err = netif_set_real_num_rx_queues(dev, rxq);
3025 if (txq > dev->real_num_tx_queues) {
3026 err = netif_set_real_num_tx_queues(dev, txq);
3030 if (rxq < dev->real_num_rx_queues)
3031 WARN_ON(netif_set_real_num_rx_queues(dev, rxq));
3032 if (txq < dev->real_num_tx_queues)
3033 WARN_ON(netif_set_real_num_tx_queues(dev, txq));
3037 WARN_ON(netif_set_real_num_rx_queues(dev, old_rxq));
3040 EXPORT_SYMBOL(netif_set_real_num_queues);
3043 * netif_set_tso_max_size() - set the max size of TSO frames supported
3044 * @dev: netdev to update
3045 * @size: max skb->len of a TSO frame
3047 * Set the limit on the size of TSO super-frames the device can handle.
3048 * Unless explicitly set the stack will assume the value of
3049 * %GSO_LEGACY_MAX_SIZE.
3051 void netif_set_tso_max_size(struct net_device *dev, unsigned int size)
3053 dev->tso_max_size = min(GSO_MAX_SIZE, size);
3054 if (size < READ_ONCE(dev->gso_max_size))
3055 netif_set_gso_max_size(dev, size);
3056 if (size < READ_ONCE(dev->gso_ipv4_max_size))
3057 netif_set_gso_ipv4_max_size(dev, size);
3059 EXPORT_SYMBOL(netif_set_tso_max_size);
3062 * netif_set_tso_max_segs() - set the max number of segs supported for TSO
3063 * @dev: netdev to update
3064 * @segs: max number of TCP segments
3066 * Set the limit on the number of TCP segments the device can generate from
3067 * a single TSO super-frame.
3068 * Unless explicitly set the stack will assume the value of %GSO_MAX_SEGS.
3070 void netif_set_tso_max_segs(struct net_device *dev, unsigned int segs)
3072 dev->tso_max_segs = segs;
3073 if (segs < READ_ONCE(dev->gso_max_segs))
3074 netif_set_gso_max_segs(dev, segs);
3076 EXPORT_SYMBOL(netif_set_tso_max_segs);
3079 * netif_inherit_tso_max() - copy all TSO limits from a lower device to an upper
3080 * @to: netdev to update
3081 * @from: netdev from which to copy the limits
3083 void netif_inherit_tso_max(struct net_device *to, const struct net_device *from)
3085 netif_set_tso_max_size(to, from->tso_max_size);
3086 netif_set_tso_max_segs(to, from->tso_max_segs);
3088 EXPORT_SYMBOL(netif_inherit_tso_max);
3091 * netif_get_num_default_rss_queues - default number of RSS queues
3093 * Default value is the number of physical cores if there are only 1 or 2, or
3094 * divided by 2 if there are more.
3096 int netif_get_num_default_rss_queues(void)
3101 if (unlikely(is_kdump_kernel() || !zalloc_cpumask_var(&cpus, GFP_KERNEL)))
3104 cpumask_copy(cpus, cpu_online_mask);
3105 for_each_cpu(cpu, cpus) {
3107 cpumask_andnot(cpus, cpus, topology_sibling_cpumask(cpu));
3109 free_cpumask_var(cpus);
3111 return count > 2 ? DIV_ROUND_UP(count, 2) : count;
3113 EXPORT_SYMBOL(netif_get_num_default_rss_queues);
3115 static void __netif_reschedule(struct Qdisc *q)
3117 struct softnet_data *sd;
3118 unsigned long flags;
3120 local_irq_save(flags);
3121 sd = this_cpu_ptr(&softnet_data);
3122 q->next_sched = NULL;
3123 *sd->output_queue_tailp = q;
3124 sd->output_queue_tailp = &q->next_sched;
3125 raise_softirq_irqoff(NET_TX_SOFTIRQ);
3126 local_irq_restore(flags);
3129 void __netif_schedule(struct Qdisc *q)
3131 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
3132 __netif_reschedule(q);
3134 EXPORT_SYMBOL(__netif_schedule);
3136 struct dev_kfree_skb_cb {
3137 enum skb_drop_reason reason;
3140 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb)
3142 return (struct dev_kfree_skb_cb *)skb->cb;
3145 void netif_schedule_queue(struct netdev_queue *txq)
3148 if (!netif_xmit_stopped(txq)) {
3149 struct Qdisc *q = rcu_dereference(txq->qdisc);
3151 __netif_schedule(q);
3155 EXPORT_SYMBOL(netif_schedule_queue);
3157 void netif_tx_wake_queue(struct netdev_queue *dev_queue)
3159 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) {
3163 q = rcu_dereference(dev_queue->qdisc);
3164 __netif_schedule(q);
3168 EXPORT_SYMBOL(netif_tx_wake_queue);
3170 void dev_kfree_skb_irq_reason(struct sk_buff *skb, enum skb_drop_reason reason)
3172 unsigned long flags;
3177 if (likely(refcount_read(&skb->users) == 1)) {
3179 refcount_set(&skb->users, 0);
3180 } else if (likely(!refcount_dec_and_test(&skb->users))) {
3183 get_kfree_skb_cb(skb)->reason = reason;
3184 local_irq_save(flags);
3185 skb->next = __this_cpu_read(softnet_data.completion_queue);
3186 __this_cpu_write(softnet_data.completion_queue, skb);
3187 raise_softirq_irqoff(NET_TX_SOFTIRQ);
3188 local_irq_restore(flags);
3190 EXPORT_SYMBOL(dev_kfree_skb_irq_reason);
3192 void dev_kfree_skb_any_reason(struct sk_buff *skb, enum skb_drop_reason reason)
3194 if (in_hardirq() || irqs_disabled())
3195 dev_kfree_skb_irq_reason(skb, reason);
3197 kfree_skb_reason(skb, reason);
3199 EXPORT_SYMBOL(dev_kfree_skb_any_reason);
3203 * netif_device_detach - mark device as removed
3204 * @dev: network device
3206 * Mark device as removed from system and therefore no longer available.
3208 void netif_device_detach(struct net_device *dev)
3210 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) &&
3211 netif_running(dev)) {
3212 netif_tx_stop_all_queues(dev);
3215 EXPORT_SYMBOL(netif_device_detach);
3218 * netif_device_attach - mark device as attached
3219 * @dev: network device
3221 * Mark device as attached from system and restart if needed.
3223 void netif_device_attach(struct net_device *dev)
3225 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) &&
3226 netif_running(dev)) {
3227 netif_tx_wake_all_queues(dev);
3228 __netdev_watchdog_up(dev);
3231 EXPORT_SYMBOL(netif_device_attach);
3234 * Returns a Tx hash based on the given packet descriptor a Tx queues' number
3235 * to be used as a distribution range.
3237 static u16 skb_tx_hash(const struct net_device *dev,
3238 const struct net_device *sb_dev,
3239 struct sk_buff *skb)
3243 u16 qcount = dev->real_num_tx_queues;
3246 u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
3248 qoffset = sb_dev->tc_to_txq[tc].offset;
3249 qcount = sb_dev->tc_to_txq[tc].count;
3250 if (unlikely(!qcount)) {
3251 net_warn_ratelimited("%s: invalid qcount, qoffset %u for tc %u\n",
3252 sb_dev->name, qoffset, tc);
3254 qcount = dev->real_num_tx_queues;
3258 if (skb_rx_queue_recorded(skb)) {
3259 DEBUG_NET_WARN_ON_ONCE(qcount == 0);
3260 hash = skb_get_rx_queue(skb);
3261 if (hash >= qoffset)
3263 while (unlikely(hash >= qcount))
3265 return hash + qoffset;
3268 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset;
3271 void skb_warn_bad_offload(const struct sk_buff *skb)
3273 static const netdev_features_t null_features;
3274 struct net_device *dev = skb->dev;
3275 const char *name = "";
3277 if (!net_ratelimit())
3281 if (dev->dev.parent)
3282 name = dev_driver_string(dev->dev.parent);
3284 name = netdev_name(dev);
3286 skb_dump(KERN_WARNING, skb, false);
3287 WARN(1, "%s: caps=(%pNF, %pNF)\n",
3288 name, dev ? &dev->features : &null_features,
3289 skb->sk ? &skb->sk->sk_route_caps : &null_features);
3293 * Invalidate hardware checksum when packet is to be mangled, and
3294 * complete checksum manually on outgoing path.
3296 int skb_checksum_help(struct sk_buff *skb)
3299 int ret = 0, offset;
3301 if (skb->ip_summed == CHECKSUM_COMPLETE)
3302 goto out_set_summed;
3304 if (unlikely(skb_is_gso(skb))) {
3305 skb_warn_bad_offload(skb);
3309 /* Before computing a checksum, we should make sure no frag could
3310 * be modified by an external entity : checksum could be wrong.
3312 if (skb_has_shared_frag(skb)) {
3313 ret = __skb_linearize(skb);
3318 offset = skb_checksum_start_offset(skb);
3320 if (unlikely(offset >= skb_headlen(skb))) {
3321 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false);
3322 WARN_ONCE(true, "offset (%d) >= skb_headlen() (%u)\n",
3323 offset, skb_headlen(skb));
3326 csum = skb_checksum(skb, offset, skb->len - offset, 0);
3328 offset += skb->csum_offset;
3329 if (unlikely(offset + sizeof(__sum16) > skb_headlen(skb))) {
3330 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false);
3331 WARN_ONCE(true, "offset+2 (%zu) > skb_headlen() (%u)\n",
3332 offset + sizeof(__sum16), skb_headlen(skb));
3335 ret = skb_ensure_writable(skb, offset + sizeof(__sum16));
3339 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0;
3341 skb->ip_summed = CHECKSUM_NONE;
3345 EXPORT_SYMBOL(skb_checksum_help);
3347 int skb_crc32c_csum_help(struct sk_buff *skb)
3350 int ret = 0, offset, start;
3352 if (skb->ip_summed != CHECKSUM_PARTIAL)
3355 if (unlikely(skb_is_gso(skb)))
3358 /* Before computing a checksum, we should make sure no frag could
3359 * be modified by an external entity : checksum could be wrong.
3361 if (unlikely(skb_has_shared_frag(skb))) {
3362 ret = __skb_linearize(skb);
3366 start = skb_checksum_start_offset(skb);
3367 offset = start + offsetof(struct sctphdr, checksum);
3368 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) {
3373 ret = skb_ensure_writable(skb, offset + sizeof(__le32));
3377 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start,
3378 skb->len - start, ~(__u32)0,
3380 *(__le32 *)(skb->data + offset) = crc32c_csum;
3381 skb_reset_csum_not_inet(skb);
3386 __be16 skb_network_protocol(struct sk_buff *skb, int *depth)
3388 __be16 type = skb->protocol;
3390 /* Tunnel gso handlers can set protocol to ethernet. */
3391 if (type == htons(ETH_P_TEB)) {
3394 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr))))
3397 eth = (struct ethhdr *)skb->data;
3398 type = eth->h_proto;
3401 return vlan_get_protocol_and_depth(skb, type, depth);
3405 /* Take action when hardware reception checksum errors are detected. */
3407 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb)
3409 netdev_err(dev, "hw csum failure\n");
3410 skb_dump(KERN_ERR, skb, true);
3414 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb)
3416 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb);
3418 EXPORT_SYMBOL(netdev_rx_csum_fault);
3421 /* XXX: check that highmem exists at all on the given machine. */
3422 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
3424 #ifdef CONFIG_HIGHMEM
3427 if (!(dev->features & NETIF_F_HIGHDMA)) {
3428 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
3429 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
3431 if (PageHighMem(skb_frag_page(frag)))
3439 /* If MPLS offload request, verify we are testing hardware MPLS features
3440 * instead of standard features for the netdev.
3442 #if IS_ENABLED(CONFIG_NET_MPLS_GSO)
3443 static netdev_features_t net_mpls_features(struct sk_buff *skb,
3444 netdev_features_t features,
3447 if (eth_p_mpls(type))
3448 features &= skb->dev->mpls_features;
3453 static netdev_features_t net_mpls_features(struct sk_buff *skb,
3454 netdev_features_t features,
3461 static netdev_features_t harmonize_features(struct sk_buff *skb,
3462 netdev_features_t features)
3466 type = skb_network_protocol(skb, NULL);
3467 features = net_mpls_features(skb, features, type);
3469 if (skb->ip_summed != CHECKSUM_NONE &&
3470 !can_checksum_protocol(features, type)) {
3471 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
3473 if (illegal_highdma(skb->dev, skb))
3474 features &= ~NETIF_F_SG;
3479 netdev_features_t passthru_features_check(struct sk_buff *skb,
3480 struct net_device *dev,
3481 netdev_features_t features)
3485 EXPORT_SYMBOL(passthru_features_check);
3487 static netdev_features_t dflt_features_check(struct sk_buff *skb,
3488 struct net_device *dev,
3489 netdev_features_t features)
3491 return vlan_features_check(skb, features);
3494 static netdev_features_t gso_features_check(const struct sk_buff *skb,
3495 struct net_device *dev,
3496 netdev_features_t features)
3498 u16 gso_segs = skb_shinfo(skb)->gso_segs;
3500 if (gso_segs > READ_ONCE(dev->gso_max_segs))
3501 return features & ~NETIF_F_GSO_MASK;
3503 if (unlikely(skb->len >= READ_ONCE(dev->gso_max_size)))
3504 return features & ~NETIF_F_GSO_MASK;
3506 if (!skb_shinfo(skb)->gso_type) {
3507 skb_warn_bad_offload(skb);
3508 return features & ~NETIF_F_GSO_MASK;
3511 /* Support for GSO partial features requires software
3512 * intervention before we can actually process the packets
3513 * so we need to strip support for any partial features now
3514 * and we can pull them back in after we have partially
3515 * segmented the frame.
3517 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL))
3518 features &= ~dev->gso_partial_features;
3520 /* Make sure to clear the IPv4 ID mangling feature if the
3521 * IPv4 header has the potential to be fragmented.
3523 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
3524 struct iphdr *iph = skb->encapsulation ?
3525 inner_ip_hdr(skb) : ip_hdr(skb);
3527 if (!(iph->frag_off & htons(IP_DF)))
3528 features &= ~NETIF_F_TSO_MANGLEID;
3534 netdev_features_t netif_skb_features(struct sk_buff *skb)
3536 struct net_device *dev = skb->dev;
3537 netdev_features_t features = dev->features;
3539 if (skb_is_gso(skb))
3540 features = gso_features_check(skb, dev, features);
3542 /* If encapsulation offload request, verify we are testing
3543 * hardware encapsulation features instead of standard
3544 * features for the netdev
3546 if (skb->encapsulation)
3547 features &= dev->hw_enc_features;
3549 if (skb_vlan_tagged(skb))
3550 features = netdev_intersect_features(features,
3551 dev->vlan_features |
3552 NETIF_F_HW_VLAN_CTAG_TX |
3553 NETIF_F_HW_VLAN_STAG_TX);
3555 if (dev->netdev_ops->ndo_features_check)
3556 features &= dev->netdev_ops->ndo_features_check(skb, dev,
3559 features &= dflt_features_check(skb, dev, features);
3561 return harmonize_features(skb, features);
3563 EXPORT_SYMBOL(netif_skb_features);
3565 static int xmit_one(struct sk_buff *skb, struct net_device *dev,
3566 struct netdev_queue *txq, bool more)
3571 if (dev_nit_active(dev))
3572 dev_queue_xmit_nit(skb, dev);
3575 trace_net_dev_start_xmit(skb, dev);
3576 rc = netdev_start_xmit(skb, dev, txq, more);
3577 trace_net_dev_xmit(skb, rc, dev, len);
3582 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev,
3583 struct netdev_queue *txq, int *ret)
3585 struct sk_buff *skb = first;
3586 int rc = NETDEV_TX_OK;
3589 struct sk_buff *next = skb->next;
3591 skb_mark_not_on_list(skb);
3592 rc = xmit_one(skb, dev, txq, next != NULL);
3593 if (unlikely(!dev_xmit_complete(rc))) {
3599 if (netif_tx_queue_stopped(txq) && skb) {
3600 rc = NETDEV_TX_BUSY;
3610 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb,
3611 netdev_features_t features)
3613 if (skb_vlan_tag_present(skb) &&
3614 !vlan_hw_offload_capable(features, skb->vlan_proto))
3615 skb = __vlan_hwaccel_push_inside(skb);
3619 int skb_csum_hwoffload_help(struct sk_buff *skb,
3620 const netdev_features_t features)
3622 if (unlikely(skb_csum_is_sctp(skb)))
3623 return !!(features & NETIF_F_SCTP_CRC) ? 0 :
3624 skb_crc32c_csum_help(skb);
3626 if (features & NETIF_F_HW_CSUM)
3629 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) {
3630 switch (skb->csum_offset) {
3631 case offsetof(struct tcphdr, check):
3632 case offsetof(struct udphdr, check):
3637 return skb_checksum_help(skb);
3639 EXPORT_SYMBOL(skb_csum_hwoffload_help);
3641 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again)
3643 netdev_features_t features;
3645 features = netif_skb_features(skb);
3646 skb = validate_xmit_vlan(skb, features);
3650 skb = sk_validate_xmit_skb(skb, dev);
3654 if (netif_needs_gso(skb, features)) {
3655 struct sk_buff *segs;
3657 segs = skb_gso_segment(skb, features);
3665 if (skb_needs_linearize(skb, features) &&
3666 __skb_linearize(skb))
3669 /* If packet is not checksummed and device does not
3670 * support checksumming for this protocol, complete
3671 * checksumming here.
3673 if (skb->ip_summed == CHECKSUM_PARTIAL) {
3674 if (skb->encapsulation)
3675 skb_set_inner_transport_header(skb,
3676 skb_checksum_start_offset(skb));
3678 skb_set_transport_header(skb,
3679 skb_checksum_start_offset(skb));
3680 if (skb_csum_hwoffload_help(skb, features))
3685 skb = validate_xmit_xfrm(skb, features, again);
3692 dev_core_stats_tx_dropped_inc(dev);
3696 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again)
3698 struct sk_buff *next, *head = NULL, *tail;
3700 for (; skb != NULL; skb = next) {
3702 skb_mark_not_on_list(skb);
3704 /* in case skb wont be segmented, point to itself */
3707 skb = validate_xmit_skb(skb, dev, again);
3715 /* If skb was segmented, skb->prev points to
3716 * the last segment. If not, it still contains skb.
3722 EXPORT_SYMBOL_GPL(validate_xmit_skb_list);
3724 static void qdisc_pkt_len_init(struct sk_buff *skb)
3726 const struct skb_shared_info *shinfo = skb_shinfo(skb);
3728 qdisc_skb_cb(skb)->pkt_len = skb->len;
3730 /* To get more precise estimation of bytes sent on wire,
3731 * we add to pkt_len the headers size of all segments
3733 if (shinfo->gso_size && skb_transport_header_was_set(skb)) {
3734 u16 gso_segs = shinfo->gso_segs;
3735 unsigned int hdr_len;
3737 /* mac layer + network layer */
3738 hdr_len = skb_transport_offset(skb);
3740 /* + transport layer */
3741 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) {
3742 const struct tcphdr *th;
3743 struct tcphdr _tcphdr;
3745 th = skb_header_pointer(skb, hdr_len,
3746 sizeof(_tcphdr), &_tcphdr);
3748 hdr_len += __tcp_hdrlen(th);
3750 struct udphdr _udphdr;
3752 if (skb_header_pointer(skb, hdr_len,
3753 sizeof(_udphdr), &_udphdr))
3754 hdr_len += sizeof(struct udphdr);
3757 if (shinfo->gso_type & SKB_GSO_DODGY)
3758 gso_segs = DIV_ROUND_UP(skb->len - hdr_len,
3761 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len;
3765 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q,
3766 struct sk_buff **to_free,
3767 struct netdev_queue *txq)
3771 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK;
3772 if (rc == NET_XMIT_SUCCESS)
3773 trace_qdisc_enqueue(q, txq, skb);
3777 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
3778 struct net_device *dev,
3779 struct netdev_queue *txq)
3781 spinlock_t *root_lock = qdisc_lock(q);
3782 struct sk_buff *to_free = NULL;
3786 qdisc_calculate_pkt_len(skb, q);
3788 if (q->flags & TCQ_F_NOLOCK) {
3789 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) &&
3790 qdisc_run_begin(q)) {
3791 /* Retest nolock_qdisc_is_empty() within the protection
3792 * of q->seqlock to protect from racing with requeuing.
3794 if (unlikely(!nolock_qdisc_is_empty(q))) {
3795 rc = dev_qdisc_enqueue(skb, q, &to_free, txq);
3802 qdisc_bstats_cpu_update(q, skb);
3803 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) &&
3804 !nolock_qdisc_is_empty(q))
3808 return NET_XMIT_SUCCESS;
3811 rc = dev_qdisc_enqueue(skb, q, &to_free, txq);
3815 if (unlikely(to_free))
3816 kfree_skb_list_reason(to_free,
3817 SKB_DROP_REASON_QDISC_DROP);
3822 * Heuristic to force contended enqueues to serialize on a
3823 * separate lock before trying to get qdisc main lock.
3824 * This permits qdisc->running owner to get the lock more
3825 * often and dequeue packets faster.
3826 * On PREEMPT_RT it is possible to preempt the qdisc owner during xmit
3827 * and then other tasks will only enqueue packets. The packets will be
3828 * sent after the qdisc owner is scheduled again. To prevent this
3829 * scenario the task always serialize on the lock.
3831 contended = qdisc_is_running(q) || IS_ENABLED(CONFIG_PREEMPT_RT);
3832 if (unlikely(contended))
3833 spin_lock(&q->busylock);
3835 spin_lock(root_lock);
3836 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3837 __qdisc_drop(skb, &to_free);
3839 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
3840 qdisc_run_begin(q)) {
3842 * This is a work-conserving queue; there are no old skbs
3843 * waiting to be sent out; and the qdisc is not running -
3844 * xmit the skb directly.
3847 qdisc_bstats_update(q, skb);
3849 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) {
3850 if (unlikely(contended)) {
3851 spin_unlock(&q->busylock);
3858 rc = NET_XMIT_SUCCESS;
3860 rc = dev_qdisc_enqueue(skb, q, &to_free, txq);
3861 if (qdisc_run_begin(q)) {
3862 if (unlikely(contended)) {
3863 spin_unlock(&q->busylock);
3870 spin_unlock(root_lock);
3871 if (unlikely(to_free))
3872 kfree_skb_list_reason(to_free, SKB_DROP_REASON_QDISC_DROP);
3873 if (unlikely(contended))
3874 spin_unlock(&q->busylock);
3878 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO)
3879 static void skb_update_prio(struct sk_buff *skb)
3881 const struct netprio_map *map;
3882 const struct sock *sk;
3883 unsigned int prioidx;
3887 map = rcu_dereference_bh(skb->dev->priomap);
3890 sk = skb_to_full_sk(skb);
3894 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data);
3896 if (prioidx < map->priomap_len)
3897 skb->priority = map->priomap[prioidx];
3900 #define skb_update_prio(skb)
3904 * dev_loopback_xmit - loop back @skb
3905 * @net: network namespace this loopback is happening in
3906 * @sk: sk needed to be a netfilter okfn
3907 * @skb: buffer to transmit
3909 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb)
3911 skb_reset_mac_header(skb);
3912 __skb_pull(skb, skb_network_offset(skb));
3913 skb->pkt_type = PACKET_LOOPBACK;
3914 if (skb->ip_summed == CHECKSUM_NONE)
3915 skb->ip_summed = CHECKSUM_UNNECESSARY;
3916 DEBUG_NET_WARN_ON_ONCE(!skb_dst(skb));
3921 EXPORT_SYMBOL(dev_loopback_xmit);
3923 #ifdef CONFIG_NET_EGRESS
3924 static struct netdev_queue *
3925 netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb)
3927 int qm = skb_get_queue_mapping(skb);
3929 return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm));
3932 static bool netdev_xmit_txqueue_skipped(void)
3934 return __this_cpu_read(softnet_data.xmit.skip_txqueue);
3937 void netdev_xmit_skip_txqueue(bool skip)
3939 __this_cpu_write(softnet_data.xmit.skip_txqueue, skip);
3941 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue);
3942 #endif /* CONFIG_NET_EGRESS */
3944 #ifdef CONFIG_NET_XGRESS
3945 static int tc_run(struct tcx_entry *entry, struct sk_buff *skb)
3947 int ret = TC_ACT_UNSPEC;
3948 #ifdef CONFIG_NET_CLS_ACT
3949 struct mini_Qdisc *miniq = rcu_dereference_bh(entry->miniq);
3950 struct tcf_result res;
3955 tc_skb_cb(skb)->mru = 0;
3956 tc_skb_cb(skb)->post_ct = false;
3958 mini_qdisc_bstats_cpu_update(miniq, skb);
3959 ret = tcf_classify(skb, miniq->block, miniq->filter_list, &res, false);
3960 /* Only tcf related quirks below. */
3963 mini_qdisc_qstats_cpu_drop(miniq);
3966 case TC_ACT_RECLASSIFY:
3967 skb->tc_index = TC_H_MIN(res.classid);
3970 #endif /* CONFIG_NET_CLS_ACT */
3974 static DEFINE_STATIC_KEY_FALSE(tcx_needed_key);
3978 static_branch_inc(&tcx_needed_key);
3983 static_branch_dec(&tcx_needed_key);
3986 static __always_inline enum tcx_action_base
3987 tcx_run(const struct bpf_mprog_entry *entry, struct sk_buff *skb,
3988 const bool needs_mac)
3990 const struct bpf_mprog_fp *fp;
3991 const struct bpf_prog *prog;
3995 __skb_push(skb, skb->mac_len);
3996 bpf_mprog_foreach_prog(entry, fp, prog) {
3997 bpf_compute_data_pointers(skb);
3998 ret = bpf_prog_run(prog, skb);
3999 if (ret != TCX_NEXT)
4003 __skb_pull(skb, skb->mac_len);
4004 return tcx_action_code(skb, ret);
4007 static __always_inline struct sk_buff *
4008 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
4009 struct net_device *orig_dev, bool *another)
4011 struct bpf_mprog_entry *entry = rcu_dereference_bh(skb->dev->tcx_ingress);
4017 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4021 qdisc_skb_cb(skb)->pkt_len = skb->len;
4022 tcx_set_ingress(skb, true);
4024 if (static_branch_unlikely(&tcx_needed_key)) {
4025 sch_ret = tcx_run(entry, skb, true);
4026 if (sch_ret != TC_ACT_UNSPEC)
4027 goto ingress_verdict;
4029 sch_ret = tc_run(tcx_entry(entry), skb);
4032 case TC_ACT_REDIRECT:
4033 /* skb_mac_header check was done by BPF, so we can safely
4034 * push the L2 header back before redirecting to another
4037 __skb_push(skb, skb->mac_len);
4038 if (skb_do_redirect(skb) == -EAGAIN) {
4039 __skb_pull(skb, skb->mac_len);
4043 *ret = NET_RX_SUCCESS;
4046 kfree_skb_reason(skb, SKB_DROP_REASON_TC_INGRESS);
4049 /* used by tc_run */
4055 case TC_ACT_CONSUMED:
4056 *ret = NET_RX_SUCCESS;
4063 static __always_inline struct sk_buff *
4064 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
4066 struct bpf_mprog_entry *entry = rcu_dereference_bh(dev->tcx_egress);
4072 /* qdisc_skb_cb(skb)->pkt_len & tcx_set_ingress() was
4073 * already set by the caller.
4075 if (static_branch_unlikely(&tcx_needed_key)) {
4076 sch_ret = tcx_run(entry, skb, false);
4077 if (sch_ret != TC_ACT_UNSPEC)
4078 goto egress_verdict;
4080 sch_ret = tc_run(tcx_entry(entry), skb);
4083 case TC_ACT_REDIRECT:
4084 /* No need to push/pop skb's mac_header here on egress! */
4085 skb_do_redirect(skb);
4086 *ret = NET_XMIT_SUCCESS;
4089 kfree_skb_reason(skb, SKB_DROP_REASON_TC_EGRESS);
4090 *ret = NET_XMIT_DROP;
4092 /* used by tc_run */
4098 case TC_ACT_CONSUMED:
4099 *ret = NET_XMIT_SUCCESS;
4106 static __always_inline struct sk_buff *
4107 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
4108 struct net_device *orig_dev, bool *another)
4113 static __always_inline struct sk_buff *
4114 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
4118 #endif /* CONFIG_NET_XGRESS */
4121 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb,
4122 struct xps_dev_maps *dev_maps, unsigned int tci)
4124 int tc = netdev_get_prio_tc_map(dev, skb->priority);
4125 struct xps_map *map;
4126 int queue_index = -1;
4128 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids)
4131 tci *= dev_maps->num_tc;
4134 map = rcu_dereference(dev_maps->attr_map[tci]);
4137 queue_index = map->queues[0];
4139 queue_index = map->queues[reciprocal_scale(
4140 skb_get_hash(skb), map->len)];
4141 if (unlikely(queue_index >= dev->real_num_tx_queues))
4148 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev,
4149 struct sk_buff *skb)
4152 struct xps_dev_maps *dev_maps;
4153 struct sock *sk = skb->sk;
4154 int queue_index = -1;
4156 if (!static_key_false(&xps_needed))
4160 if (!static_key_false(&xps_rxqs_needed))
4163 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]);
4165 int tci = sk_rx_queue_get(sk);
4168 queue_index = __get_xps_queue_idx(dev, skb, dev_maps,
4173 if (queue_index < 0) {
4174 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]);
4176 unsigned int tci = skb->sender_cpu - 1;
4178 queue_index = __get_xps_queue_idx(dev, skb, dev_maps,
4190 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb,
4191 struct net_device *sb_dev)
4195 EXPORT_SYMBOL(dev_pick_tx_zero);
4197 u16 dev_pick_tx_cpu_id(struct net_device *dev, struct sk_buff *skb,
4198 struct net_device *sb_dev)
4200 return (u16)raw_smp_processor_id() % dev->real_num_tx_queues;
4202 EXPORT_SYMBOL(dev_pick_tx_cpu_id);
4204 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb,
4205 struct net_device *sb_dev)
4207 struct sock *sk = skb->sk;
4208 int queue_index = sk_tx_queue_get(sk);
4210 sb_dev = sb_dev ? : dev;
4212 if (queue_index < 0 || skb->ooo_okay ||
4213 queue_index >= dev->real_num_tx_queues) {
4214 int new_index = get_xps_queue(dev, sb_dev, skb);
4217 new_index = skb_tx_hash(dev, sb_dev, skb);
4219 if (queue_index != new_index && sk &&
4221 rcu_access_pointer(sk->sk_dst_cache))
4222 sk_tx_queue_set(sk, new_index);
4224 queue_index = new_index;
4229 EXPORT_SYMBOL(netdev_pick_tx);
4231 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev,
4232 struct sk_buff *skb,
4233 struct net_device *sb_dev)
4235 int queue_index = 0;
4238 u32 sender_cpu = skb->sender_cpu - 1;
4240 if (sender_cpu >= (u32)NR_CPUS)
4241 skb->sender_cpu = raw_smp_processor_id() + 1;
4244 if (dev->real_num_tx_queues != 1) {
4245 const struct net_device_ops *ops = dev->netdev_ops;
4247 if (ops->ndo_select_queue)
4248 queue_index = ops->ndo_select_queue(dev, skb, sb_dev);
4250 queue_index = netdev_pick_tx(dev, skb, sb_dev);
4252 queue_index = netdev_cap_txqueue(dev, queue_index);
4255 skb_set_queue_mapping(skb, queue_index);
4256 return netdev_get_tx_queue(dev, queue_index);
4260 * __dev_queue_xmit() - transmit a buffer
4261 * @skb: buffer to transmit
4262 * @sb_dev: suboordinate device used for L2 forwarding offload
4264 * Queue a buffer for transmission to a network device. The caller must
4265 * have set the device and priority and built the buffer before calling
4266 * this function. The function can be called from an interrupt.
4268 * When calling this method, interrupts MUST be enabled. This is because
4269 * the BH enable code must have IRQs enabled so that it will not deadlock.
4271 * Regardless of the return value, the skb is consumed, so it is currently
4272 * difficult to retry a send to this method. (You can bump the ref count
4273 * before sending to hold a reference for retry if you are careful.)
4276 * * 0 - buffer successfully transmitted
4277 * * positive qdisc return code - NET_XMIT_DROP etc.
4278 * * negative errno - other errors
4280 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev)
4282 struct net_device *dev = skb->dev;
4283 struct netdev_queue *txq = NULL;
4288 skb_reset_mac_header(skb);
4289 skb_assert_len(skb);
4291 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP))
4292 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED);
4294 /* Disable soft irqs for various locks below. Also
4295 * stops preemption for RCU.
4299 skb_update_prio(skb);
4301 qdisc_pkt_len_init(skb);
4302 tcx_set_ingress(skb, false);
4303 #ifdef CONFIG_NET_EGRESS
4304 if (static_branch_unlikely(&egress_needed_key)) {
4305 if (nf_hook_egress_active()) {
4306 skb = nf_hook_egress(skb, &rc, dev);
4311 netdev_xmit_skip_txqueue(false);
4313 nf_skip_egress(skb, true);
4314 skb = sch_handle_egress(skb, &rc, dev);
4317 nf_skip_egress(skb, false);
4319 if (netdev_xmit_txqueue_skipped())
4320 txq = netdev_tx_queue_mapping(dev, skb);
4323 /* If device/qdisc don't need skb->dst, release it right now while
4324 * its hot in this cpu cache.
4326 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
4332 txq = netdev_core_pick_tx(dev, skb, sb_dev);
4334 q = rcu_dereference_bh(txq->qdisc);
4336 trace_net_dev_queue(skb);
4338 rc = __dev_xmit_skb(skb, q, dev, txq);
4342 /* The device has no queue. Common case for software devices:
4343 * loopback, all the sorts of tunnels...
4345 * Really, it is unlikely that netif_tx_lock protection is necessary
4346 * here. (f.e. loopback and IP tunnels are clean ignoring statistics
4348 * However, it is possible, that they rely on protection
4351 * Check this and shot the lock. It is not prone from deadlocks.
4352 *Either shot noqueue qdisc, it is even simpler 8)
4354 if (dev->flags & IFF_UP) {
4355 int cpu = smp_processor_id(); /* ok because BHs are off */
4357 /* Other cpus might concurrently change txq->xmit_lock_owner
4358 * to -1 or to their cpu id, but not to our id.
4360 if (READ_ONCE(txq->xmit_lock_owner) != cpu) {
4361 if (dev_xmit_recursion())
4362 goto recursion_alert;
4364 skb = validate_xmit_skb(skb, dev, &again);
4368 HARD_TX_LOCK(dev, txq, cpu);
4370 if (!netif_xmit_stopped(txq)) {
4371 dev_xmit_recursion_inc();
4372 skb = dev_hard_start_xmit(skb, dev, txq, &rc);
4373 dev_xmit_recursion_dec();
4374 if (dev_xmit_complete(rc)) {
4375 HARD_TX_UNLOCK(dev, txq);
4379 HARD_TX_UNLOCK(dev, txq);
4380 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
4383 /* Recursion is detected! It is possible,
4387 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
4393 rcu_read_unlock_bh();
4395 dev_core_stats_tx_dropped_inc(dev);
4396 kfree_skb_list(skb);
4399 rcu_read_unlock_bh();
4402 EXPORT_SYMBOL(__dev_queue_xmit);
4404 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id)
4406 struct net_device *dev = skb->dev;
4407 struct sk_buff *orig_skb = skb;
4408 struct netdev_queue *txq;
4409 int ret = NETDEV_TX_BUSY;
4412 if (unlikely(!netif_running(dev) ||
4413 !netif_carrier_ok(dev)))
4416 skb = validate_xmit_skb_list(skb, dev, &again);
4417 if (skb != orig_skb)
4420 skb_set_queue_mapping(skb, queue_id);
4421 txq = skb_get_tx_queue(dev, skb);
4425 dev_xmit_recursion_inc();
4426 HARD_TX_LOCK(dev, txq, smp_processor_id());
4427 if (!netif_xmit_frozen_or_drv_stopped(txq))
4428 ret = netdev_start_xmit(skb, dev, txq, false);
4429 HARD_TX_UNLOCK(dev, txq);
4430 dev_xmit_recursion_dec();
4435 dev_core_stats_tx_dropped_inc(dev);
4436 kfree_skb_list(skb);
4437 return NET_XMIT_DROP;
4439 EXPORT_SYMBOL(__dev_direct_xmit);
4441 /*************************************************************************
4443 *************************************************************************/
4445 int netdev_max_backlog __read_mostly = 1000;
4446 EXPORT_SYMBOL(netdev_max_backlog);
4448 int netdev_tstamp_prequeue __read_mostly = 1;
4449 unsigned int sysctl_skb_defer_max __read_mostly = 64;
4450 int netdev_budget __read_mostly = 300;
4451 /* Must be at least 2 jiffes to guarantee 1 jiffy timeout */
4452 unsigned int __read_mostly netdev_budget_usecs = 2 * USEC_PER_SEC / HZ;
4453 int weight_p __read_mostly = 64; /* old backlog weight */
4454 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */
4455 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */
4456 int dev_rx_weight __read_mostly = 64;
4457 int dev_tx_weight __read_mostly = 64;
4459 /* Called with irq disabled */
4460 static inline void ____napi_schedule(struct softnet_data *sd,
4461 struct napi_struct *napi)
4463 struct task_struct *thread;
4465 lockdep_assert_irqs_disabled();
4467 if (test_bit(NAPI_STATE_THREADED, &napi->state)) {
4468 /* Paired with smp_mb__before_atomic() in
4469 * napi_enable()/dev_set_threaded().
4470 * Use READ_ONCE() to guarantee a complete
4471 * read on napi->thread. Only call
4472 * wake_up_process() when it's not NULL.
4474 thread = READ_ONCE(napi->thread);
4476 /* Avoid doing set_bit() if the thread is in
4477 * INTERRUPTIBLE state, cause napi_thread_wait()
4478 * makes sure to proceed with napi polling
4479 * if the thread is explicitly woken from here.
4481 if (READ_ONCE(thread->__state) != TASK_INTERRUPTIBLE)
4482 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state);
4483 wake_up_process(thread);
4488 list_add_tail(&napi->poll_list, &sd->poll_list);
4489 WRITE_ONCE(napi->list_owner, smp_processor_id());
4490 /* If not called from net_rx_action()
4491 * we have to raise NET_RX_SOFTIRQ.
4493 if (!sd->in_net_rx_action)
4494 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
4499 /* One global table that all flow-based protocols share. */
4500 struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
4501 EXPORT_SYMBOL(rps_sock_flow_table);
4502 u32 rps_cpu_mask __read_mostly;
4503 EXPORT_SYMBOL(rps_cpu_mask);
4505 struct static_key_false rps_needed __read_mostly;
4506 EXPORT_SYMBOL(rps_needed);
4507 struct static_key_false rfs_needed __read_mostly;
4508 EXPORT_SYMBOL(rfs_needed);
4510 static struct rps_dev_flow *
4511 set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
4512 struct rps_dev_flow *rflow, u16 next_cpu)
4514 if (next_cpu < nr_cpu_ids) {
4515 #ifdef CONFIG_RFS_ACCEL
4516 struct netdev_rx_queue *rxqueue;
4517 struct rps_dev_flow_table *flow_table;
4518 struct rps_dev_flow *old_rflow;
4523 /* Should we steer this flow to a different hardware queue? */
4524 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
4525 !(dev->features & NETIF_F_NTUPLE))
4527 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
4528 if (rxq_index == skb_get_rx_queue(skb))
4531 rxqueue = dev->_rx + rxq_index;
4532 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4535 flow_id = skb_get_hash(skb) & flow_table->mask;
4536 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
4537 rxq_index, flow_id);
4541 rflow = &flow_table->flows[flow_id];
4543 if (old_rflow->filter == rflow->filter)
4544 old_rflow->filter = RPS_NO_FILTER;
4548 per_cpu(softnet_data, next_cpu).input_queue_head;
4551 rflow->cpu = next_cpu;
4556 * get_rps_cpu is called from netif_receive_skb and returns the target
4557 * CPU from the RPS map of the receiving queue for a given skb.
4558 * rcu_read_lock must be held on entry.
4560 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
4561 struct rps_dev_flow **rflowp)
4563 const struct rps_sock_flow_table *sock_flow_table;
4564 struct netdev_rx_queue *rxqueue = dev->_rx;
4565 struct rps_dev_flow_table *flow_table;
4566 struct rps_map *map;
4571 if (skb_rx_queue_recorded(skb)) {
4572 u16 index = skb_get_rx_queue(skb);
4574 if (unlikely(index >= dev->real_num_rx_queues)) {
4575 WARN_ONCE(dev->real_num_rx_queues > 1,
4576 "%s received packet on queue %u, but number "
4577 "of RX queues is %u\n",
4578 dev->name, index, dev->real_num_rx_queues);
4584 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */
4586 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4587 map = rcu_dereference(rxqueue->rps_map);
4588 if (!flow_table && !map)
4591 skb_reset_network_header(skb);
4592 hash = skb_get_hash(skb);
4596 sock_flow_table = rcu_dereference(rps_sock_flow_table);
4597 if (flow_table && sock_flow_table) {
4598 struct rps_dev_flow *rflow;
4602 /* First check into global flow table if there is a match.
4603 * This READ_ONCE() pairs with WRITE_ONCE() from rps_record_sock_flow().
4605 ident = READ_ONCE(sock_flow_table->ents[hash & sock_flow_table->mask]);
4606 if ((ident ^ hash) & ~rps_cpu_mask)
4609 next_cpu = ident & rps_cpu_mask;
4611 /* OK, now we know there is a match,
4612 * we can look at the local (per receive queue) flow table
4614 rflow = &flow_table->flows[hash & flow_table->mask];
4618 * If the desired CPU (where last recvmsg was done) is
4619 * different from current CPU (one in the rx-queue flow
4620 * table entry), switch if one of the following holds:
4621 * - Current CPU is unset (>= nr_cpu_ids).
4622 * - Current CPU is offline.
4623 * - The current CPU's queue tail has advanced beyond the
4624 * last packet that was enqueued using this table entry.
4625 * This guarantees that all previous packets for the flow
4626 * have been dequeued, thus preserving in order delivery.
4628 if (unlikely(tcpu != next_cpu) &&
4629 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) ||
4630 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
4631 rflow->last_qtail)) >= 0)) {
4633 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
4636 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) {
4646 tcpu = map->cpus[reciprocal_scale(hash, map->len)];
4647 if (cpu_online(tcpu)) {
4657 #ifdef CONFIG_RFS_ACCEL
4660 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
4661 * @dev: Device on which the filter was set
4662 * @rxq_index: RX queue index
4663 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
4664 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
4666 * Drivers that implement ndo_rx_flow_steer() should periodically call
4667 * this function for each installed filter and remove the filters for
4668 * which it returns %true.
4670 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
4671 u32 flow_id, u16 filter_id)
4673 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
4674 struct rps_dev_flow_table *flow_table;
4675 struct rps_dev_flow *rflow;
4680 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4681 if (flow_table && flow_id <= flow_table->mask) {
4682 rflow = &flow_table->flows[flow_id];
4683 cpu = READ_ONCE(rflow->cpu);
4684 if (rflow->filter == filter_id && cpu < nr_cpu_ids &&
4685 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
4686 rflow->last_qtail) <
4687 (int)(10 * flow_table->mask)))
4693 EXPORT_SYMBOL(rps_may_expire_flow);
4695 #endif /* CONFIG_RFS_ACCEL */
4697 /* Called from hardirq (IPI) context */
4698 static void rps_trigger_softirq(void *data)
4700 struct softnet_data *sd = data;
4702 ____napi_schedule(sd, &sd->backlog);
4706 #endif /* CONFIG_RPS */
4709 * After we queued a packet into sd->input_pkt_queue,
4710 * we need to make sure this queue is serviced soon.
4712 * - If this is another cpu queue, link it to our rps_ipi_list,
4713 * and make sure we will process rps_ipi_list from net_rx_action().
4715 * - If this is our own queue, NAPI schedule our backlog.
4716 * Note that this also raises NET_RX_SOFTIRQ.
4718 static void napi_schedule_rps(struct softnet_data *sd)
4720 struct softnet_data *mysd = this_cpu_ptr(&softnet_data);
4724 sd->rps_ipi_next = mysd->rps_ipi_list;
4725 mysd->rps_ipi_list = sd;
4727 /* If not called from net_rx_action() or napi_threaded_poll()
4728 * we have to raise NET_RX_SOFTIRQ.
4730 if (!mysd->in_net_rx_action && !mysd->in_napi_threaded_poll)
4731 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
4734 #endif /* CONFIG_RPS */
4735 __napi_schedule_irqoff(&mysd->backlog);
4738 #ifdef CONFIG_NET_FLOW_LIMIT
4739 int netdev_flow_limit_table_len __read_mostly = (1 << 12);
4742 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
4744 #ifdef CONFIG_NET_FLOW_LIMIT
4745 struct sd_flow_limit *fl;
4746 struct softnet_data *sd;
4747 unsigned int old_flow, new_flow;
4749 if (qlen < (READ_ONCE(netdev_max_backlog) >> 1))
4752 sd = this_cpu_ptr(&softnet_data);
4755 fl = rcu_dereference(sd->flow_limit);
4757 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1);
4758 old_flow = fl->history[fl->history_head];
4759 fl->history[fl->history_head] = new_flow;
4762 fl->history_head &= FLOW_LIMIT_HISTORY - 1;
4764 if (likely(fl->buckets[old_flow]))
4765 fl->buckets[old_flow]--;
4767 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) {
4779 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
4780 * queue (may be a remote CPU queue).
4782 static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
4783 unsigned int *qtail)
4785 enum skb_drop_reason reason;
4786 struct softnet_data *sd;
4787 unsigned long flags;
4790 reason = SKB_DROP_REASON_NOT_SPECIFIED;
4791 sd = &per_cpu(softnet_data, cpu);
4793 rps_lock_irqsave(sd, &flags);
4794 if (!netif_running(skb->dev))
4796 qlen = skb_queue_len(&sd->input_pkt_queue);
4797 if (qlen <= READ_ONCE(netdev_max_backlog) && !skb_flow_limit(skb, qlen)) {
4800 __skb_queue_tail(&sd->input_pkt_queue, skb);
4801 input_queue_tail_incr_save(sd, qtail);
4802 rps_unlock_irq_restore(sd, &flags);
4803 return NET_RX_SUCCESS;
4806 /* Schedule NAPI for backlog device
4807 * We can use non atomic operation since we own the queue lock
4809 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state))
4810 napi_schedule_rps(sd);
4813 reason = SKB_DROP_REASON_CPU_BACKLOG;
4817 rps_unlock_irq_restore(sd, &flags);
4819 dev_core_stats_rx_dropped_inc(skb->dev);
4820 kfree_skb_reason(skb, reason);
4824 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb)
4826 struct net_device *dev = skb->dev;
4827 struct netdev_rx_queue *rxqueue;
4831 if (skb_rx_queue_recorded(skb)) {
4832 u16 index = skb_get_rx_queue(skb);
4834 if (unlikely(index >= dev->real_num_rx_queues)) {
4835 WARN_ONCE(dev->real_num_rx_queues > 1,
4836 "%s received packet on queue %u, but number "
4837 "of RX queues is %u\n",
4838 dev->name, index, dev->real_num_rx_queues);
4840 return rxqueue; /* Return first rxqueue */
4847 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp,
4848 struct bpf_prog *xdp_prog)
4850 void *orig_data, *orig_data_end, *hard_start;
4851 struct netdev_rx_queue *rxqueue;
4852 bool orig_bcast, orig_host;
4853 u32 mac_len, frame_sz;
4854 __be16 orig_eth_type;
4859 /* The XDP program wants to see the packet starting at the MAC
4862 mac_len = skb->data - skb_mac_header(skb);
4863 hard_start = skb->data - skb_headroom(skb);
4865 /* SKB "head" area always have tailroom for skb_shared_info */
4866 frame_sz = (void *)skb_end_pointer(skb) - hard_start;
4867 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
4869 rxqueue = netif_get_rxqueue(skb);
4870 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq);
4871 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len,
4872 skb_headlen(skb) + mac_len, true);
4874 orig_data_end = xdp->data_end;
4875 orig_data = xdp->data;
4876 eth = (struct ethhdr *)xdp->data;
4877 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr);
4878 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest);
4879 orig_eth_type = eth->h_proto;
4881 act = bpf_prog_run_xdp(xdp_prog, xdp);
4883 /* check if bpf_xdp_adjust_head was used */
4884 off = xdp->data - orig_data;
4887 __skb_pull(skb, off);
4889 __skb_push(skb, -off);
4891 skb->mac_header += off;
4892 skb_reset_network_header(skb);
4895 /* check if bpf_xdp_adjust_tail was used */
4896 off = xdp->data_end - orig_data_end;
4898 skb_set_tail_pointer(skb, xdp->data_end - xdp->data);
4899 skb->len += off; /* positive on grow, negative on shrink */
4902 /* check if XDP changed eth hdr such SKB needs update */
4903 eth = (struct ethhdr *)xdp->data;
4904 if ((orig_eth_type != eth->h_proto) ||
4905 (orig_host != ether_addr_equal_64bits(eth->h_dest,
4906 skb->dev->dev_addr)) ||
4907 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) {
4908 __skb_push(skb, ETH_HLEN);
4909 skb->pkt_type = PACKET_HOST;
4910 skb->protocol = eth_type_trans(skb, skb->dev);
4913 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull
4914 * before calling us again on redirect path. We do not call do_redirect
4915 * as we leave that up to the caller.
4917 * Caller is responsible for managing lifetime of skb (i.e. calling
4918 * kfree_skb in response to actions it cannot handle/XDP_DROP).
4923 __skb_push(skb, mac_len);
4926 metalen = xdp->data - xdp->data_meta;
4928 skb_metadata_set(skb, metalen);
4935 static u32 netif_receive_generic_xdp(struct sk_buff *skb,
4936 struct xdp_buff *xdp,
4937 struct bpf_prog *xdp_prog)
4941 /* Reinjected packets coming from act_mirred or similar should
4942 * not get XDP generic processing.
4944 if (skb_is_redirected(skb))
4947 /* XDP packets must be linear and must have sufficient headroom
4948 * of XDP_PACKET_HEADROOM bytes. This is the guarantee that also
4949 * native XDP provides, thus we need to do it here as well.
4951 if (skb_cloned(skb) || skb_is_nonlinear(skb) ||
4952 skb_headroom(skb) < XDP_PACKET_HEADROOM) {
4953 int hroom = XDP_PACKET_HEADROOM - skb_headroom(skb);
4954 int troom = skb->tail + skb->data_len - skb->end;
4956 /* In case we have to go down the path and also linearize,
4957 * then lets do the pskb_expand_head() work just once here.
4959 if (pskb_expand_head(skb,
4960 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0,
4961 troom > 0 ? troom + 128 : 0, GFP_ATOMIC))
4963 if (skb_linearize(skb))
4967 act = bpf_prog_run_generic_xdp(skb, xdp, xdp_prog);
4974 bpf_warn_invalid_xdp_action(skb->dev, xdp_prog, act);
4977 trace_xdp_exception(skb->dev, xdp_prog, act);
4988 /* When doing generic XDP we have to bypass the qdisc layer and the
4989 * network taps in order to match in-driver-XDP behavior. This also means
4990 * that XDP packets are able to starve other packets going through a qdisc,
4991 * and DDOS attacks will be more effective. In-driver-XDP use dedicated TX
4992 * queues, so they do not have this starvation issue.
4994 void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog)
4996 struct net_device *dev = skb->dev;
4997 struct netdev_queue *txq;
4998 bool free_skb = true;
5001 txq = netdev_core_pick_tx(dev, skb, NULL);
5002 cpu = smp_processor_id();
5003 HARD_TX_LOCK(dev, txq, cpu);
5004 if (!netif_xmit_frozen_or_drv_stopped(txq)) {
5005 rc = netdev_start_xmit(skb, dev, txq, 0);
5006 if (dev_xmit_complete(rc))
5009 HARD_TX_UNLOCK(dev, txq);
5011 trace_xdp_exception(dev, xdp_prog, XDP_TX);
5012 dev_core_stats_tx_dropped_inc(dev);
5017 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key);
5019 int do_xdp_generic(struct bpf_prog *xdp_prog, struct sk_buff *skb)
5022 struct xdp_buff xdp;
5026 act = netif_receive_generic_xdp(skb, &xdp, xdp_prog);
5027 if (act != XDP_PASS) {
5030 err = xdp_do_generic_redirect(skb->dev, skb,
5036 generic_xdp_tx(skb, xdp_prog);
5044 kfree_skb_reason(skb, SKB_DROP_REASON_XDP);
5047 EXPORT_SYMBOL_GPL(do_xdp_generic);
5049 static int netif_rx_internal(struct sk_buff *skb)
5053 net_timestamp_check(READ_ONCE(netdev_tstamp_prequeue), skb);
5055 trace_netif_rx(skb);
5058 if (static_branch_unlikely(&rps_needed)) {
5059 struct rps_dev_flow voidflow, *rflow = &voidflow;
5064 cpu = get_rps_cpu(skb->dev, skb, &rflow);
5066 cpu = smp_processor_id();
5068 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
5076 ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail);
5082 * __netif_rx - Slightly optimized version of netif_rx
5083 * @skb: buffer to post
5085 * This behaves as netif_rx except that it does not disable bottom halves.
5086 * As a result this function may only be invoked from the interrupt context
5087 * (either hard or soft interrupt).
5089 int __netif_rx(struct sk_buff *skb)
5093 lockdep_assert_once(hardirq_count() | softirq_count());
5095 trace_netif_rx_entry(skb);
5096 ret = netif_rx_internal(skb);
5097 trace_netif_rx_exit(ret);
5100 EXPORT_SYMBOL(__netif_rx);
5103 * netif_rx - post buffer to the network code
5104 * @skb: buffer to post
5106 * This function receives a packet from a device driver and queues it for
5107 * the upper (protocol) levels to process via the backlog NAPI device. It
5108 * always succeeds. The buffer may be dropped during processing for
5109 * congestion control or by the protocol layers.
5110 * The network buffer is passed via the backlog NAPI device. Modern NIC
5111 * driver should use NAPI and GRO.
5112 * This function can used from interrupt and from process context. The
5113 * caller from process context must not disable interrupts before invoking
5117 * NET_RX_SUCCESS (no congestion)
5118 * NET_RX_DROP (packet was dropped)
5121 int netif_rx(struct sk_buff *skb)
5123 bool need_bh_off = !(hardirq_count() | softirq_count());
5128 trace_netif_rx_entry(skb);
5129 ret = netif_rx_internal(skb);
5130 trace_netif_rx_exit(ret);
5135 EXPORT_SYMBOL(netif_rx);
5137 static __latent_entropy void net_tx_action(struct softirq_action *h)
5139 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
5141 if (sd->completion_queue) {
5142 struct sk_buff *clist;
5144 local_irq_disable();
5145 clist = sd->completion_queue;
5146 sd->completion_queue = NULL;
5150 struct sk_buff *skb = clist;
5152 clist = clist->next;
5154 WARN_ON(refcount_read(&skb->users));
5155 if (likely(get_kfree_skb_cb(skb)->reason == SKB_CONSUMED))
5156 trace_consume_skb(skb, net_tx_action);
5158 trace_kfree_skb(skb, net_tx_action,
5159 get_kfree_skb_cb(skb)->reason);
5161 if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
5164 __napi_kfree_skb(skb,
5165 get_kfree_skb_cb(skb)->reason);
5169 if (sd->output_queue) {
5172 local_irq_disable();
5173 head = sd->output_queue;
5174 sd->output_queue = NULL;
5175 sd->output_queue_tailp = &sd->output_queue;
5181 struct Qdisc *q = head;
5182 spinlock_t *root_lock = NULL;
5184 head = head->next_sched;
5186 /* We need to make sure head->next_sched is read
5187 * before clearing __QDISC_STATE_SCHED
5189 smp_mb__before_atomic();
5191 if (!(q->flags & TCQ_F_NOLOCK)) {
5192 root_lock = qdisc_lock(q);
5193 spin_lock(root_lock);
5194 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED,
5196 /* There is a synchronize_net() between
5197 * STATE_DEACTIVATED flag being set and
5198 * qdisc_reset()/some_qdisc_is_busy() in
5199 * dev_deactivate(), so we can safely bail out
5200 * early here to avoid data race between
5201 * qdisc_deactivate() and some_qdisc_is_busy()
5202 * for lockless qdisc.
5204 clear_bit(__QDISC_STATE_SCHED, &q->state);
5208 clear_bit(__QDISC_STATE_SCHED, &q->state);
5211 spin_unlock(root_lock);
5217 xfrm_dev_backlog(sd);
5220 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE)
5221 /* This hook is defined here for ATM LANE */
5222 int (*br_fdb_test_addr_hook)(struct net_device *dev,
5223 unsigned char *addr) __read_mostly;
5224 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook);
5228 * netdev_is_rx_handler_busy - check if receive handler is registered
5229 * @dev: device to check
5231 * Check if a receive handler is already registered for a given device.
5232 * Return true if there one.
5234 * The caller must hold the rtnl_mutex.
5236 bool netdev_is_rx_handler_busy(struct net_device *dev)
5239 return dev && rtnl_dereference(dev->rx_handler);
5241 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy);
5244 * netdev_rx_handler_register - register receive handler
5245 * @dev: device to register a handler for
5246 * @rx_handler: receive handler to register
5247 * @rx_handler_data: data pointer that is used by rx handler
5249 * Register a receive handler for a device. This handler will then be
5250 * called from __netif_receive_skb. A negative errno code is returned
5253 * The caller must hold the rtnl_mutex.
5255 * For a general description of rx_handler, see enum rx_handler_result.
5257 int netdev_rx_handler_register(struct net_device *dev,
5258 rx_handler_func_t *rx_handler,
5259 void *rx_handler_data)
5261 if (netdev_is_rx_handler_busy(dev))
5264 if (dev->priv_flags & IFF_NO_RX_HANDLER)
5267 /* Note: rx_handler_data must be set before rx_handler */
5268 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data);
5269 rcu_assign_pointer(dev->rx_handler, rx_handler);
5273 EXPORT_SYMBOL_GPL(netdev_rx_handler_register);
5276 * netdev_rx_handler_unregister - unregister receive handler
5277 * @dev: device to unregister a handler from
5279 * Unregister a receive handler from a device.
5281 * The caller must hold the rtnl_mutex.
5283 void netdev_rx_handler_unregister(struct net_device *dev)
5287 RCU_INIT_POINTER(dev->rx_handler, NULL);
5288 /* a reader seeing a non NULL rx_handler in a rcu_read_lock()
5289 * section has a guarantee to see a non NULL rx_handler_data
5293 RCU_INIT_POINTER(dev->rx_handler_data, NULL);
5295 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister);
5298 * Limit the use of PFMEMALLOC reserves to those protocols that implement
5299 * the special handling of PFMEMALLOC skbs.
5301 static bool skb_pfmemalloc_protocol(struct sk_buff *skb)
5303 switch (skb->protocol) {
5304 case htons(ETH_P_ARP):
5305 case htons(ETH_P_IP):
5306 case htons(ETH_P_IPV6):
5307 case htons(ETH_P_8021Q):
5308 case htons(ETH_P_8021AD):
5315 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev,
5316 int *ret, struct net_device *orig_dev)
5318 if (nf_hook_ingress_active(skb)) {
5322 *ret = deliver_skb(skb, *pt_prev, orig_dev);
5327 ingress_retval = nf_hook_ingress(skb);
5329 return ingress_retval;
5334 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc,
5335 struct packet_type **ppt_prev)
5337 struct packet_type *ptype, *pt_prev;
5338 rx_handler_func_t *rx_handler;
5339 struct sk_buff *skb = *pskb;
5340 struct net_device *orig_dev;
5341 bool deliver_exact = false;
5342 int ret = NET_RX_DROP;
5345 net_timestamp_check(!READ_ONCE(netdev_tstamp_prequeue), skb);
5347 trace_netif_receive_skb(skb);
5349 orig_dev = skb->dev;
5351 skb_reset_network_header(skb);
5352 if (!skb_transport_header_was_set(skb))
5353 skb_reset_transport_header(skb);
5354 skb_reset_mac_len(skb);
5359 skb->skb_iif = skb->dev->ifindex;
5361 __this_cpu_inc(softnet_data.processed);
5363 if (static_branch_unlikely(&generic_xdp_needed_key)) {
5367 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
5370 if (ret2 != XDP_PASS) {
5376 if (eth_type_vlan(skb->protocol)) {
5377 skb = skb_vlan_untag(skb);
5382 if (skb_skip_tc_classify(skb))
5388 list_for_each_entry_rcu(ptype, &ptype_all, list) {
5390 ret = deliver_skb(skb, pt_prev, orig_dev);
5394 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) {
5396 ret = deliver_skb(skb, pt_prev, orig_dev);
5401 #ifdef CONFIG_NET_INGRESS
5402 if (static_branch_unlikely(&ingress_needed_key)) {
5403 bool another = false;
5405 nf_skip_egress(skb, true);
5406 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev,
5413 nf_skip_egress(skb, false);
5414 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0)
5418 skb_reset_redirect(skb);
5420 if (pfmemalloc && !skb_pfmemalloc_protocol(skb))
5423 if (skb_vlan_tag_present(skb)) {
5425 ret = deliver_skb(skb, pt_prev, orig_dev);
5428 if (vlan_do_receive(&skb))
5430 else if (unlikely(!skb))
5434 rx_handler = rcu_dereference(skb->dev->rx_handler);
5437 ret = deliver_skb(skb, pt_prev, orig_dev);
5440 switch (rx_handler(&skb)) {
5441 case RX_HANDLER_CONSUMED:
5442 ret = NET_RX_SUCCESS;
5444 case RX_HANDLER_ANOTHER:
5446 case RX_HANDLER_EXACT:
5447 deliver_exact = true;
5449 case RX_HANDLER_PASS:
5456 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) {
5458 if (skb_vlan_tag_get_id(skb)) {
5459 /* Vlan id is non 0 and vlan_do_receive() above couldn't
5462 skb->pkt_type = PACKET_OTHERHOST;
5463 } else if (eth_type_vlan(skb->protocol)) {
5464 /* Outer header is 802.1P with vlan 0, inner header is
5465 * 802.1Q or 802.1AD and vlan_do_receive() above could
5466 * not find vlan dev for vlan id 0.
5468 __vlan_hwaccel_clear_tag(skb);
5469 skb = skb_vlan_untag(skb);
5472 if (vlan_do_receive(&skb))
5473 /* After stripping off 802.1P header with vlan 0
5474 * vlan dev is found for inner header.
5477 else if (unlikely(!skb))
5480 /* We have stripped outer 802.1P vlan 0 header.
5481 * But could not find vlan dev.
5482 * check again for vlan id to set OTHERHOST.
5486 /* Note: we might in the future use prio bits
5487 * and set skb->priority like in vlan_do_receive()
5488 * For the time being, just ignore Priority Code Point
5490 __vlan_hwaccel_clear_tag(skb);
5493 type = skb->protocol;
5495 /* deliver only exact match when indicated */
5496 if (likely(!deliver_exact)) {
5497 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
5498 &ptype_base[ntohs(type) &
5502 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
5503 &orig_dev->ptype_specific);
5505 if (unlikely(skb->dev != orig_dev)) {
5506 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
5507 &skb->dev->ptype_specific);
5511 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
5513 *ppt_prev = pt_prev;
5517 dev_core_stats_rx_dropped_inc(skb->dev);
5519 dev_core_stats_rx_nohandler_inc(skb->dev);
5520 kfree_skb_reason(skb, SKB_DROP_REASON_UNHANDLED_PROTO);
5521 /* Jamal, now you will not able to escape explaining
5522 * me how you were going to use this. :-)
5528 /* The invariant here is that if *ppt_prev is not NULL
5529 * then skb should also be non-NULL.
5531 * Apparently *ppt_prev assignment above holds this invariant due to
5532 * skb dereferencing near it.
5538 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc)
5540 struct net_device *orig_dev = skb->dev;
5541 struct packet_type *pt_prev = NULL;
5544 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev);
5546 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb,
5547 skb->dev, pt_prev, orig_dev);
5552 * netif_receive_skb_core - special purpose version of netif_receive_skb
5553 * @skb: buffer to process
5555 * More direct receive version of netif_receive_skb(). It should
5556 * only be used by callers that have a need to skip RPS and Generic XDP.
5557 * Caller must also take care of handling if ``(page_is_)pfmemalloc``.
5559 * This function may only be called from softirq context and interrupts
5560 * should be enabled.
5562 * Return values (usually ignored):
5563 * NET_RX_SUCCESS: no congestion
5564 * NET_RX_DROP: packet was dropped
5566 int netif_receive_skb_core(struct sk_buff *skb)
5571 ret = __netif_receive_skb_one_core(skb, false);
5576 EXPORT_SYMBOL(netif_receive_skb_core);
5578 static inline void __netif_receive_skb_list_ptype(struct list_head *head,
5579 struct packet_type *pt_prev,
5580 struct net_device *orig_dev)
5582 struct sk_buff *skb, *next;
5586 if (list_empty(head))
5588 if (pt_prev->list_func != NULL)
5589 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv,
5590 ip_list_rcv, head, pt_prev, orig_dev);
5592 list_for_each_entry_safe(skb, next, head, list) {
5593 skb_list_del_init(skb);
5594 pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
5598 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc)
5600 /* Fast-path assumptions:
5601 * - There is no RX handler.
5602 * - Only one packet_type matches.
5603 * If either of these fails, we will end up doing some per-packet
5604 * processing in-line, then handling the 'last ptype' for the whole
5605 * sublist. This can't cause out-of-order delivery to any single ptype,
5606 * because the 'last ptype' must be constant across the sublist, and all
5607 * other ptypes are handled per-packet.
5609 /* Current (common) ptype of sublist */
5610 struct packet_type *pt_curr = NULL;
5611 /* Current (common) orig_dev of sublist */
5612 struct net_device *od_curr = NULL;
5613 struct list_head sublist;
5614 struct sk_buff *skb, *next;
5616 INIT_LIST_HEAD(&sublist);
5617 list_for_each_entry_safe(skb, next, head, list) {
5618 struct net_device *orig_dev = skb->dev;
5619 struct packet_type *pt_prev = NULL;
5621 skb_list_del_init(skb);
5622 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev);
5625 if (pt_curr != pt_prev || od_curr != orig_dev) {
5626 /* dispatch old sublist */
5627 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr);
5628 /* start new sublist */
5629 INIT_LIST_HEAD(&sublist);
5633 list_add_tail(&skb->list, &sublist);
5636 /* dispatch final sublist */
5637 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr);
5640 static int __netif_receive_skb(struct sk_buff *skb)
5644 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) {
5645 unsigned int noreclaim_flag;
5648 * PFMEMALLOC skbs are special, they should
5649 * - be delivered to SOCK_MEMALLOC sockets only
5650 * - stay away from userspace
5651 * - have bounded memory usage
5653 * Use PF_MEMALLOC as this saves us from propagating the allocation
5654 * context down to all allocation sites.
5656 noreclaim_flag = memalloc_noreclaim_save();
5657 ret = __netif_receive_skb_one_core(skb, true);
5658 memalloc_noreclaim_restore(noreclaim_flag);
5660 ret = __netif_receive_skb_one_core(skb, false);
5665 static void __netif_receive_skb_list(struct list_head *head)
5667 unsigned long noreclaim_flag = 0;
5668 struct sk_buff *skb, *next;
5669 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */
5671 list_for_each_entry_safe(skb, next, head, list) {
5672 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) {
5673 struct list_head sublist;
5675 /* Handle the previous sublist */
5676 list_cut_before(&sublist, head, &skb->list);
5677 if (!list_empty(&sublist))
5678 __netif_receive_skb_list_core(&sublist, pfmemalloc);
5679 pfmemalloc = !pfmemalloc;
5680 /* See comments in __netif_receive_skb */
5682 noreclaim_flag = memalloc_noreclaim_save();
5684 memalloc_noreclaim_restore(noreclaim_flag);
5687 /* Handle the remaining sublist */
5688 if (!list_empty(head))
5689 __netif_receive_skb_list_core(head, pfmemalloc);
5690 /* Restore pflags */
5692 memalloc_noreclaim_restore(noreclaim_flag);
5695 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp)
5697 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog);
5698 struct bpf_prog *new = xdp->prog;
5701 switch (xdp->command) {
5702 case XDP_SETUP_PROG:
5703 rcu_assign_pointer(dev->xdp_prog, new);
5708 static_branch_dec(&generic_xdp_needed_key);
5709 } else if (new && !old) {
5710 static_branch_inc(&generic_xdp_needed_key);
5711 dev_disable_lro(dev);
5712 dev_disable_gro_hw(dev);
5724 static int netif_receive_skb_internal(struct sk_buff *skb)
5728 net_timestamp_check(READ_ONCE(netdev_tstamp_prequeue), skb);
5730 if (skb_defer_rx_timestamp(skb))
5731 return NET_RX_SUCCESS;
5735 if (static_branch_unlikely(&rps_needed)) {
5736 struct rps_dev_flow voidflow, *rflow = &voidflow;
5737 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
5740 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
5746 ret = __netif_receive_skb(skb);
5751 void netif_receive_skb_list_internal(struct list_head *head)
5753 struct sk_buff *skb, *next;
5754 struct list_head sublist;
5756 INIT_LIST_HEAD(&sublist);
5757 list_for_each_entry_safe(skb, next, head, list) {
5758 net_timestamp_check(READ_ONCE(netdev_tstamp_prequeue), skb);
5759 skb_list_del_init(skb);
5760 if (!skb_defer_rx_timestamp(skb))
5761 list_add_tail(&skb->list, &sublist);
5763 list_splice_init(&sublist, head);
5767 if (static_branch_unlikely(&rps_needed)) {
5768 list_for_each_entry_safe(skb, next, head, list) {
5769 struct rps_dev_flow voidflow, *rflow = &voidflow;
5770 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
5773 /* Will be handled, remove from list */
5774 skb_list_del_init(skb);
5775 enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
5780 __netif_receive_skb_list(head);
5785 * netif_receive_skb - process receive buffer from network
5786 * @skb: buffer to process
5788 * netif_receive_skb() is the main receive data processing function.
5789 * It always succeeds. The buffer may be dropped during processing
5790 * for congestion control or by the protocol layers.
5792 * This function may only be called from softirq context and interrupts
5793 * should be enabled.
5795 * Return values (usually ignored):
5796 * NET_RX_SUCCESS: no congestion
5797 * NET_RX_DROP: packet was dropped
5799 int netif_receive_skb(struct sk_buff *skb)
5803 trace_netif_receive_skb_entry(skb);
5805 ret = netif_receive_skb_internal(skb);
5806 trace_netif_receive_skb_exit(ret);
5810 EXPORT_SYMBOL(netif_receive_skb);
5813 * netif_receive_skb_list - process many receive buffers from network
5814 * @head: list of skbs to process.
5816 * Since return value of netif_receive_skb() is normally ignored, and
5817 * wouldn't be meaningful for a list, this function returns void.
5819 * This function may only be called from softirq context and interrupts
5820 * should be enabled.
5822 void netif_receive_skb_list(struct list_head *head)
5824 struct sk_buff *skb;
5826 if (list_empty(head))
5828 if (trace_netif_receive_skb_list_entry_enabled()) {
5829 list_for_each_entry(skb, head, list)
5830 trace_netif_receive_skb_list_entry(skb);
5832 netif_receive_skb_list_internal(head);
5833 trace_netif_receive_skb_list_exit(0);
5835 EXPORT_SYMBOL(netif_receive_skb_list);
5837 static DEFINE_PER_CPU(struct work_struct, flush_works);
5839 /* Network device is going away, flush any packets still pending */
5840 static void flush_backlog(struct work_struct *work)
5842 struct sk_buff *skb, *tmp;
5843 struct softnet_data *sd;
5846 sd = this_cpu_ptr(&softnet_data);
5848 rps_lock_irq_disable(sd);
5849 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) {
5850 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
5851 __skb_unlink(skb, &sd->input_pkt_queue);
5852 dev_kfree_skb_irq(skb);
5853 input_queue_head_incr(sd);
5856 rps_unlock_irq_enable(sd);
5858 skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
5859 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
5860 __skb_unlink(skb, &sd->process_queue);
5862 input_queue_head_incr(sd);
5868 static bool flush_required(int cpu)
5870 #if IS_ENABLED(CONFIG_RPS)
5871 struct softnet_data *sd = &per_cpu(softnet_data, cpu);
5874 rps_lock_irq_disable(sd);
5876 /* as insertion into process_queue happens with the rps lock held,
5877 * process_queue access may race only with dequeue
5879 do_flush = !skb_queue_empty(&sd->input_pkt_queue) ||
5880 !skb_queue_empty_lockless(&sd->process_queue);
5881 rps_unlock_irq_enable(sd);
5885 /* without RPS we can't safely check input_pkt_queue: during a
5886 * concurrent remote skb_queue_splice() we can detect as empty both
5887 * input_pkt_queue and process_queue even if the latter could end-up
5888 * containing a lot of packets.
5893 static void flush_all_backlogs(void)
5895 static cpumask_t flush_cpus;
5898 /* since we are under rtnl lock protection we can use static data
5899 * for the cpumask and avoid allocating on stack the possibly
5906 cpumask_clear(&flush_cpus);
5907 for_each_online_cpu(cpu) {
5908 if (flush_required(cpu)) {
5909 queue_work_on(cpu, system_highpri_wq,
5910 per_cpu_ptr(&flush_works, cpu));
5911 cpumask_set_cpu(cpu, &flush_cpus);
5915 /* we can have in flight packet[s] on the cpus we are not flushing,
5916 * synchronize_net() in unregister_netdevice_many() will take care of
5919 for_each_cpu(cpu, &flush_cpus)
5920 flush_work(per_cpu_ptr(&flush_works, cpu));
5925 static void net_rps_send_ipi(struct softnet_data *remsd)
5929 struct softnet_data *next = remsd->rps_ipi_next;
5931 if (cpu_online(remsd->cpu))
5932 smp_call_function_single_async(remsd->cpu, &remsd->csd);
5939 * net_rps_action_and_irq_enable sends any pending IPI's for rps.
5940 * Note: called with local irq disabled, but exits with local irq enabled.
5942 static void net_rps_action_and_irq_enable(struct softnet_data *sd)
5945 struct softnet_data *remsd = sd->rps_ipi_list;
5948 sd->rps_ipi_list = NULL;
5952 /* Send pending IPI's to kick RPS processing on remote cpus. */
5953 net_rps_send_ipi(remsd);
5959 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd)
5962 return sd->rps_ipi_list != NULL;
5968 static int process_backlog(struct napi_struct *napi, int quota)
5970 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
5974 /* Check if we have pending ipi, its better to send them now,
5975 * not waiting net_rx_action() end.
5977 if (sd_has_rps_ipi_waiting(sd)) {
5978 local_irq_disable();
5979 net_rps_action_and_irq_enable(sd);
5982 napi->weight = READ_ONCE(dev_rx_weight);
5984 struct sk_buff *skb;
5986 while ((skb = __skb_dequeue(&sd->process_queue))) {
5988 __netif_receive_skb(skb);
5990 input_queue_head_incr(sd);
5991 if (++work >= quota)
5996 rps_lock_irq_disable(sd);
5997 if (skb_queue_empty(&sd->input_pkt_queue)) {
5999 * Inline a custom version of __napi_complete().
6000 * only current cpu owns and manipulates this napi,
6001 * and NAPI_STATE_SCHED is the only possible flag set
6003 * We can use a plain write instead of clear_bit(),
6004 * and we dont need an smp_mb() memory barrier.
6009 skb_queue_splice_tail_init(&sd->input_pkt_queue,
6010 &sd->process_queue);
6012 rps_unlock_irq_enable(sd);
6019 * __napi_schedule - schedule for receive
6020 * @n: entry to schedule
6022 * The entry's receive function will be scheduled to run.
6023 * Consider using __napi_schedule_irqoff() if hard irqs are masked.
6025 void __napi_schedule(struct napi_struct *n)
6027 unsigned long flags;
6029 local_irq_save(flags);
6030 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
6031 local_irq_restore(flags);
6033 EXPORT_SYMBOL(__napi_schedule);
6036 * napi_schedule_prep - check if napi can be scheduled
6039 * Test if NAPI routine is already running, and if not mark
6040 * it as running. This is used as a condition variable to
6041 * insure only one NAPI poll instance runs. We also make
6042 * sure there is no pending NAPI disable.
6044 bool napi_schedule_prep(struct napi_struct *n)
6046 unsigned long new, val = READ_ONCE(n->state);
6049 if (unlikely(val & NAPIF_STATE_DISABLE))
6051 new = val | NAPIF_STATE_SCHED;
6053 /* Sets STATE_MISSED bit if STATE_SCHED was already set
6054 * This was suggested by Alexander Duyck, as compiler
6055 * emits better code than :
6056 * if (val & NAPIF_STATE_SCHED)
6057 * new |= NAPIF_STATE_MISSED;
6059 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED *
6061 } while (!try_cmpxchg(&n->state, &val, new));
6063 return !(val & NAPIF_STATE_SCHED);
6065 EXPORT_SYMBOL(napi_schedule_prep);
6068 * __napi_schedule_irqoff - schedule for receive
6069 * @n: entry to schedule
6071 * Variant of __napi_schedule() assuming hard irqs are masked.
6073 * On PREEMPT_RT enabled kernels this maps to __napi_schedule()
6074 * because the interrupt disabled assumption might not be true
6075 * due to force-threaded interrupts and spinlock substitution.
6077 void __napi_schedule_irqoff(struct napi_struct *n)
6079 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
6080 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
6084 EXPORT_SYMBOL(__napi_schedule_irqoff);
6086 bool napi_complete_done(struct napi_struct *n, int work_done)
6088 unsigned long flags, val, new, timeout = 0;
6092 * 1) Don't let napi dequeue from the cpu poll list
6093 * just in case its running on a different cpu.
6094 * 2) If we are busy polling, do nothing here, we have
6095 * the guarantee we will be called later.
6097 if (unlikely(n->state & (NAPIF_STATE_NPSVC |
6098 NAPIF_STATE_IN_BUSY_POLL)))
6103 timeout = READ_ONCE(n->dev->gro_flush_timeout);
6104 n->defer_hard_irqs_count = READ_ONCE(n->dev->napi_defer_hard_irqs);
6106 if (n->defer_hard_irqs_count > 0) {
6107 n->defer_hard_irqs_count--;
6108 timeout = READ_ONCE(n->dev->gro_flush_timeout);
6112 if (n->gro_bitmask) {
6113 /* When the NAPI instance uses a timeout and keeps postponing
6114 * it, we need to bound somehow the time packets are kept in
6117 napi_gro_flush(n, !!timeout);
6122 if (unlikely(!list_empty(&n->poll_list))) {
6123 /* If n->poll_list is not empty, we need to mask irqs */
6124 local_irq_save(flags);
6125 list_del_init(&n->poll_list);
6126 local_irq_restore(flags);
6128 WRITE_ONCE(n->list_owner, -1);
6130 val = READ_ONCE(n->state);
6132 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED));
6134 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED |
6135 NAPIF_STATE_SCHED_THREADED |
6136 NAPIF_STATE_PREFER_BUSY_POLL);
6138 /* If STATE_MISSED was set, leave STATE_SCHED set,
6139 * because we will call napi->poll() one more time.
6140 * This C code was suggested by Alexander Duyck to help gcc.
6142 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED *
6144 } while (!try_cmpxchg(&n->state, &val, new));
6146 if (unlikely(val & NAPIF_STATE_MISSED)) {
6152 hrtimer_start(&n->timer, ns_to_ktime(timeout),
6153 HRTIMER_MODE_REL_PINNED);
6156 EXPORT_SYMBOL(napi_complete_done);
6158 /* must be called under rcu_read_lock(), as we dont take a reference */
6159 static struct napi_struct *napi_by_id(unsigned int napi_id)
6161 unsigned int hash = napi_id % HASH_SIZE(napi_hash);
6162 struct napi_struct *napi;
6164 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node)
6165 if (napi->napi_id == napi_id)
6171 #if defined(CONFIG_NET_RX_BUSY_POLL)
6173 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule)
6175 if (!skip_schedule) {
6176 gro_normal_list(napi);
6177 __napi_schedule(napi);
6181 if (napi->gro_bitmask) {
6182 /* flush too old packets
6183 * If HZ < 1000, flush all packets.
6185 napi_gro_flush(napi, HZ >= 1000);
6188 gro_normal_list(napi);
6189 clear_bit(NAPI_STATE_SCHED, &napi->state);
6192 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, bool prefer_busy_poll,
6195 bool skip_schedule = false;
6196 unsigned long timeout;
6199 /* Busy polling means there is a high chance device driver hard irq
6200 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was
6201 * set in napi_schedule_prep().
6202 * Since we are about to call napi->poll() once more, we can safely
6203 * clear NAPI_STATE_MISSED.
6205 * Note: x86 could use a single "lock and ..." instruction
6206 * to perform these two clear_bit()
6208 clear_bit(NAPI_STATE_MISSED, &napi->state);
6209 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state);
6213 if (prefer_busy_poll) {
6214 napi->defer_hard_irqs_count = READ_ONCE(napi->dev->napi_defer_hard_irqs);
6215 timeout = READ_ONCE(napi->dev->gro_flush_timeout);
6216 if (napi->defer_hard_irqs_count && timeout) {
6217 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED);
6218 skip_schedule = true;
6222 /* All we really want here is to re-enable device interrupts.
6223 * Ideally, a new ndo_busy_poll_stop() could avoid another round.
6225 rc = napi->poll(napi, budget);
6226 /* We can't gro_normal_list() here, because napi->poll() might have
6227 * rearmed the napi (napi_complete_done()) in which case it could
6228 * already be running on another CPU.
6230 trace_napi_poll(napi, rc, budget);
6231 netpoll_poll_unlock(have_poll_lock);
6233 __busy_poll_stop(napi, skip_schedule);
6237 void napi_busy_loop(unsigned int napi_id,
6238 bool (*loop_end)(void *, unsigned long),
6239 void *loop_end_arg, bool prefer_busy_poll, u16 budget)
6241 unsigned long start_time = loop_end ? busy_loop_current_time() : 0;
6242 int (*napi_poll)(struct napi_struct *napi, int budget);
6243 void *have_poll_lock = NULL;
6244 struct napi_struct *napi;
6251 napi = napi_by_id(napi_id);
6255 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
6262 unsigned long val = READ_ONCE(napi->state);
6264 /* If multiple threads are competing for this napi,
6265 * we avoid dirtying napi->state as much as we can.
6267 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED |
6268 NAPIF_STATE_IN_BUSY_POLL)) {
6269 if (prefer_busy_poll)
6270 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state);
6273 if (cmpxchg(&napi->state, val,
6274 val | NAPIF_STATE_IN_BUSY_POLL |
6275 NAPIF_STATE_SCHED) != val) {
6276 if (prefer_busy_poll)
6277 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state);
6280 have_poll_lock = netpoll_poll_lock(napi);
6281 napi_poll = napi->poll;
6283 work = napi_poll(napi, budget);
6284 trace_napi_poll(napi, work, budget);
6285 gro_normal_list(napi);
6288 __NET_ADD_STATS(dev_net(napi->dev),
6289 LINUX_MIB_BUSYPOLLRXPACKETS, work);
6292 if (!loop_end || loop_end(loop_end_arg, start_time))
6295 if (unlikely(need_resched())) {
6297 busy_poll_stop(napi, have_poll_lock, prefer_busy_poll, budget);
6298 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
6302 if (loop_end(loop_end_arg, start_time))
6309 busy_poll_stop(napi, have_poll_lock, prefer_busy_poll, budget);
6310 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
6315 EXPORT_SYMBOL(napi_busy_loop);
6317 #endif /* CONFIG_NET_RX_BUSY_POLL */
6319 static void napi_hash_add(struct napi_struct *napi)
6321 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state))
6324 spin_lock(&napi_hash_lock);
6326 /* 0..NR_CPUS range is reserved for sender_cpu use */
6328 if (unlikely(++napi_gen_id < MIN_NAPI_ID))
6329 napi_gen_id = MIN_NAPI_ID;
6330 } while (napi_by_id(napi_gen_id));
6331 napi->napi_id = napi_gen_id;
6333 hlist_add_head_rcu(&napi->napi_hash_node,
6334 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]);
6336 spin_unlock(&napi_hash_lock);
6339 /* Warning : caller is responsible to make sure rcu grace period
6340 * is respected before freeing memory containing @napi
6342 static void napi_hash_del(struct napi_struct *napi)
6344 spin_lock(&napi_hash_lock);
6346 hlist_del_init_rcu(&napi->napi_hash_node);
6348 spin_unlock(&napi_hash_lock);
6351 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer)
6353 struct napi_struct *napi;
6355 napi = container_of(timer, struct napi_struct, timer);
6357 /* Note : we use a relaxed variant of napi_schedule_prep() not setting
6358 * NAPI_STATE_MISSED, since we do not react to a device IRQ.
6360 if (!napi_disable_pending(napi) &&
6361 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) {
6362 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state);
6363 __napi_schedule_irqoff(napi);
6366 return HRTIMER_NORESTART;
6369 static void init_gro_hash(struct napi_struct *napi)
6373 for (i = 0; i < GRO_HASH_BUCKETS; i++) {
6374 INIT_LIST_HEAD(&napi->gro_hash[i].list);
6375 napi->gro_hash[i].count = 0;
6377 napi->gro_bitmask = 0;
6380 int dev_set_threaded(struct net_device *dev, bool threaded)
6382 struct napi_struct *napi;
6385 if (dev->threaded == threaded)
6389 list_for_each_entry(napi, &dev->napi_list, dev_list) {
6390 if (!napi->thread) {
6391 err = napi_kthread_create(napi);
6400 dev->threaded = threaded;
6402 /* Make sure kthread is created before THREADED bit
6405 smp_mb__before_atomic();
6407 /* Setting/unsetting threaded mode on a napi might not immediately
6408 * take effect, if the current napi instance is actively being
6409 * polled. In this case, the switch between threaded mode and
6410 * softirq mode will happen in the next round of napi_schedule().
6411 * This should not cause hiccups/stalls to the live traffic.
6413 list_for_each_entry(napi, &dev->napi_list, dev_list)
6414 assign_bit(NAPI_STATE_THREADED, &napi->state, threaded);
6418 EXPORT_SYMBOL(dev_set_threaded);
6420 void netif_napi_add_weight(struct net_device *dev, struct napi_struct *napi,
6421 int (*poll)(struct napi_struct *, int), int weight)
6423 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state)))
6426 INIT_LIST_HEAD(&napi->poll_list);
6427 INIT_HLIST_NODE(&napi->napi_hash_node);
6428 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
6429 napi->timer.function = napi_watchdog;
6430 init_gro_hash(napi);
6432 INIT_LIST_HEAD(&napi->rx_list);
6435 if (weight > NAPI_POLL_WEIGHT)
6436 netdev_err_once(dev, "%s() called with weight %d\n", __func__,
6438 napi->weight = weight;
6440 #ifdef CONFIG_NETPOLL
6441 napi->poll_owner = -1;
6443 napi->list_owner = -1;
6444 set_bit(NAPI_STATE_SCHED, &napi->state);
6445 set_bit(NAPI_STATE_NPSVC, &napi->state);
6446 list_add_rcu(&napi->dev_list, &dev->napi_list);
6447 napi_hash_add(napi);
6448 napi_get_frags_check(napi);
6449 /* Create kthread for this napi if dev->threaded is set.
6450 * Clear dev->threaded if kthread creation failed so that
6451 * threaded mode will not be enabled in napi_enable().
6453 if (dev->threaded && napi_kthread_create(napi))
6456 EXPORT_SYMBOL(netif_napi_add_weight);
6458 void napi_disable(struct napi_struct *n)
6460 unsigned long val, new;
6463 set_bit(NAPI_STATE_DISABLE, &n->state);
6465 val = READ_ONCE(n->state);
6467 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) {
6468 usleep_range(20, 200);
6469 val = READ_ONCE(n->state);
6472 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC;
6473 new &= ~(NAPIF_STATE_THREADED | NAPIF_STATE_PREFER_BUSY_POLL);
6474 } while (!try_cmpxchg(&n->state, &val, new));
6476 hrtimer_cancel(&n->timer);
6478 clear_bit(NAPI_STATE_DISABLE, &n->state);
6480 EXPORT_SYMBOL(napi_disable);
6483 * napi_enable - enable NAPI scheduling
6486 * Resume NAPI from being scheduled on this context.
6487 * Must be paired with napi_disable.
6489 void napi_enable(struct napi_struct *n)
6491 unsigned long new, val = READ_ONCE(n->state);
6494 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val));
6496 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC);
6497 if (n->dev->threaded && n->thread)
6498 new |= NAPIF_STATE_THREADED;
6499 } while (!try_cmpxchg(&n->state, &val, new));
6501 EXPORT_SYMBOL(napi_enable);
6503 static void flush_gro_hash(struct napi_struct *napi)
6507 for (i = 0; i < GRO_HASH_BUCKETS; i++) {
6508 struct sk_buff *skb, *n;
6510 list_for_each_entry_safe(skb, n, &napi->gro_hash[i].list, list)
6512 napi->gro_hash[i].count = 0;
6516 /* Must be called in process context */
6517 void __netif_napi_del(struct napi_struct *napi)
6519 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state))
6522 napi_hash_del(napi);
6523 list_del_rcu(&napi->dev_list);
6524 napi_free_frags(napi);
6526 flush_gro_hash(napi);
6527 napi->gro_bitmask = 0;
6530 kthread_stop(napi->thread);
6531 napi->thread = NULL;
6534 EXPORT_SYMBOL(__netif_napi_del);
6536 static int __napi_poll(struct napi_struct *n, bool *repoll)
6542 /* This NAPI_STATE_SCHED test is for avoiding a race
6543 * with netpoll's poll_napi(). Only the entity which
6544 * obtains the lock and sees NAPI_STATE_SCHED set will
6545 * actually make the ->poll() call. Therefore we avoid
6546 * accidentally calling ->poll() when NAPI is not scheduled.
6549 if (test_bit(NAPI_STATE_SCHED, &n->state)) {
6550 work = n->poll(n, weight);
6551 trace_napi_poll(n, work, weight);
6554 if (unlikely(work > weight))
6555 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n",
6556 n->poll, work, weight);
6558 if (likely(work < weight))
6561 /* Drivers must not modify the NAPI state if they
6562 * consume the entire weight. In such cases this code
6563 * still "owns" the NAPI instance and therefore can
6564 * move the instance around on the list at-will.
6566 if (unlikely(napi_disable_pending(n))) {
6571 /* The NAPI context has more processing work, but busy-polling
6572 * is preferred. Exit early.
6574 if (napi_prefer_busy_poll(n)) {
6575 if (napi_complete_done(n, work)) {
6576 /* If timeout is not set, we need to make sure
6577 * that the NAPI is re-scheduled.
6584 if (n->gro_bitmask) {
6585 /* flush too old packets
6586 * If HZ < 1000, flush all packets.
6588 napi_gro_flush(n, HZ >= 1000);
6593 /* Some drivers may have called napi_schedule
6594 * prior to exhausting their budget.
6596 if (unlikely(!list_empty(&n->poll_list))) {
6597 pr_warn_once("%s: Budget exhausted after napi rescheduled\n",
6598 n->dev ? n->dev->name : "backlog");
6607 static int napi_poll(struct napi_struct *n, struct list_head *repoll)
6609 bool do_repoll = false;
6613 list_del_init(&n->poll_list);
6615 have = netpoll_poll_lock(n);
6617 work = __napi_poll(n, &do_repoll);
6620 list_add_tail(&n->poll_list, repoll);
6622 netpoll_poll_unlock(have);
6627 static int napi_thread_wait(struct napi_struct *napi)
6631 set_current_state(TASK_INTERRUPTIBLE);
6633 while (!kthread_should_stop()) {
6634 /* Testing SCHED_THREADED bit here to make sure the current
6635 * kthread owns this napi and could poll on this napi.
6636 * Testing SCHED bit is not enough because SCHED bit might be
6637 * set by some other busy poll thread or by napi_disable().
6639 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state) || woken) {
6640 WARN_ON(!list_empty(&napi->poll_list));
6641 __set_current_state(TASK_RUNNING);
6646 /* woken being true indicates this thread owns this napi. */
6648 set_current_state(TASK_INTERRUPTIBLE);
6650 __set_current_state(TASK_RUNNING);
6655 static void skb_defer_free_flush(struct softnet_data *sd)
6657 struct sk_buff *skb, *next;
6659 /* Paired with WRITE_ONCE() in skb_attempt_defer_free() */
6660 if (!READ_ONCE(sd->defer_list))
6663 spin_lock(&sd->defer_lock);
6664 skb = sd->defer_list;
6665 sd->defer_list = NULL;
6666 sd->defer_count = 0;
6667 spin_unlock(&sd->defer_lock);
6669 while (skb != NULL) {
6671 napi_consume_skb(skb, 1);
6676 #ifndef CONFIG_PREEMPT_RT
6678 /* Called from hardirq (IPI) context */
6679 static void trigger_rx_softirq(void *data)
6681 struct softnet_data *sd = data;
6683 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
6684 smp_store_release(&sd->defer_ipi_scheduled, 0);
6689 static void trigger_rx_softirq(struct work_struct *defer_work)
6691 struct softnet_data *sd;
6693 sd = container_of(defer_work, struct softnet_data, defer_work);
6694 smp_store_release(&sd->defer_ipi_scheduled, 0);
6696 skb_defer_free_flush(sd);
6702 static int napi_threaded_poll(void *data)
6704 struct napi_struct *napi = data;
6705 struct softnet_data *sd;
6708 while (!napi_thread_wait(napi)) {
6710 bool repoll = false;
6713 sd = this_cpu_ptr(&softnet_data);
6714 sd->in_napi_threaded_poll = true;
6716 have = netpoll_poll_lock(napi);
6717 __napi_poll(napi, &repoll);
6718 netpoll_poll_unlock(have);
6720 sd->in_napi_threaded_poll = false;
6723 if (sd_has_rps_ipi_waiting(sd)) {
6724 local_irq_disable();
6725 net_rps_action_and_irq_enable(sd);
6727 skb_defer_free_flush(sd);
6739 static __latent_entropy void net_rx_action(struct softirq_action *h)
6741 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
6742 unsigned long time_limit = jiffies +
6743 usecs_to_jiffies(READ_ONCE(netdev_budget_usecs));
6744 int budget = READ_ONCE(netdev_budget);
6749 sd->in_net_rx_action = true;
6750 local_irq_disable();
6751 list_splice_init(&sd->poll_list, &list);
6755 struct napi_struct *n;
6757 skb_defer_free_flush(sd);
6759 if (list_empty(&list)) {
6760 if (list_empty(&repoll)) {
6761 sd->in_net_rx_action = false;
6763 /* We need to check if ____napi_schedule()
6764 * had refilled poll_list while
6765 * sd->in_net_rx_action was true.
6767 if (!list_empty(&sd->poll_list))
6769 if (!sd_has_rps_ipi_waiting(sd))
6775 n = list_first_entry(&list, struct napi_struct, poll_list);
6776 budget -= napi_poll(n, &repoll);
6778 /* If softirq window is exhausted then punt.
6779 * Allow this to run for 2 jiffies since which will allow
6780 * an average latency of 1.5/HZ.
6782 if (unlikely(budget <= 0 ||
6783 time_after_eq(jiffies, time_limit))) {
6789 local_irq_disable();
6791 list_splice_tail_init(&sd->poll_list, &list);
6792 list_splice_tail(&repoll, &list);
6793 list_splice(&list, &sd->poll_list);
6794 if (!list_empty(&sd->poll_list))
6795 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
6797 sd->in_net_rx_action = false;
6799 net_rps_action_and_irq_enable(sd);
6803 struct netdev_adjacent {
6804 struct net_device *dev;
6805 netdevice_tracker dev_tracker;
6807 /* upper master flag, there can only be one master device per list */
6810 /* lookup ignore flag */
6813 /* counter for the number of times this device was added to us */
6816 /* private field for the users */
6819 struct list_head list;
6820 struct rcu_head rcu;
6823 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev,
6824 struct list_head *adj_list)
6826 struct netdev_adjacent *adj;
6828 list_for_each_entry(adj, adj_list, list) {
6829 if (adj->dev == adj_dev)
6835 static int ____netdev_has_upper_dev(struct net_device *upper_dev,
6836 struct netdev_nested_priv *priv)
6838 struct net_device *dev = (struct net_device *)priv->data;
6840 return upper_dev == dev;
6844 * netdev_has_upper_dev - Check if device is linked to an upper device
6846 * @upper_dev: upper device to check
6848 * Find out if a device is linked to specified upper device and return true
6849 * in case it is. Note that this checks only immediate upper device,
6850 * not through a complete stack of devices. The caller must hold the RTNL lock.
6852 bool netdev_has_upper_dev(struct net_device *dev,
6853 struct net_device *upper_dev)
6855 struct netdev_nested_priv priv = {
6856 .data = (void *)upper_dev,
6861 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev,
6864 EXPORT_SYMBOL(netdev_has_upper_dev);
6867 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device
6869 * @upper_dev: upper device to check
6871 * Find out if a device is linked to specified upper device and return true
6872 * in case it is. Note that this checks the entire upper device chain.
6873 * The caller must hold rcu lock.
6876 bool netdev_has_upper_dev_all_rcu(struct net_device *dev,
6877 struct net_device *upper_dev)
6879 struct netdev_nested_priv priv = {
6880 .data = (void *)upper_dev,
6883 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev,
6886 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu);
6889 * netdev_has_any_upper_dev - Check if device is linked to some device
6892 * Find out if a device is linked to an upper device and return true in case
6893 * it is. The caller must hold the RTNL lock.
6895 bool netdev_has_any_upper_dev(struct net_device *dev)
6899 return !list_empty(&dev->adj_list.upper);
6901 EXPORT_SYMBOL(netdev_has_any_upper_dev);
6904 * netdev_master_upper_dev_get - Get master upper device
6907 * Find a master upper device and return pointer to it or NULL in case
6908 * it's not there. The caller must hold the RTNL lock.
6910 struct net_device *netdev_master_upper_dev_get(struct net_device *dev)
6912 struct netdev_adjacent *upper;
6916 if (list_empty(&dev->adj_list.upper))
6919 upper = list_first_entry(&dev->adj_list.upper,
6920 struct netdev_adjacent, list);
6921 if (likely(upper->master))
6925 EXPORT_SYMBOL(netdev_master_upper_dev_get);
6927 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev)
6929 struct netdev_adjacent *upper;
6933 if (list_empty(&dev->adj_list.upper))
6936 upper = list_first_entry(&dev->adj_list.upper,
6937 struct netdev_adjacent, list);
6938 if (likely(upper->master) && !upper->ignore)
6944 * netdev_has_any_lower_dev - Check if device is linked to some device
6947 * Find out if a device is linked to a lower device and return true in case
6948 * it is. The caller must hold the RTNL lock.
6950 static bool netdev_has_any_lower_dev(struct net_device *dev)
6954 return !list_empty(&dev->adj_list.lower);
6957 void *netdev_adjacent_get_private(struct list_head *adj_list)
6959 struct netdev_adjacent *adj;
6961 adj = list_entry(adj_list, struct netdev_adjacent, list);
6963 return adj->private;
6965 EXPORT_SYMBOL(netdev_adjacent_get_private);
6968 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list
6970 * @iter: list_head ** of the current position
6972 * Gets the next device from the dev's upper list, starting from iter
6973 * position. The caller must hold RCU read lock.
6975 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev,
6976 struct list_head **iter)
6978 struct netdev_adjacent *upper;
6980 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
6982 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6984 if (&upper->list == &dev->adj_list.upper)
6987 *iter = &upper->list;
6991 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu);
6993 static struct net_device *__netdev_next_upper_dev(struct net_device *dev,
6994 struct list_head **iter,
6997 struct netdev_adjacent *upper;
6999 upper = list_entry((*iter)->next, struct netdev_adjacent, list);
7001 if (&upper->list == &dev->adj_list.upper)
7004 *iter = &upper->list;
7005 *ignore = upper->ignore;
7010 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev,
7011 struct list_head **iter)
7013 struct netdev_adjacent *upper;
7015 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
7017 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7019 if (&upper->list == &dev->adj_list.upper)
7022 *iter = &upper->list;
7027 static int __netdev_walk_all_upper_dev(struct net_device *dev,
7028 int (*fn)(struct net_device *dev,
7029 struct netdev_nested_priv *priv),
7030 struct netdev_nested_priv *priv)
7032 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7033 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7038 iter = &dev->adj_list.upper;
7042 ret = fn(now, priv);
7049 udev = __netdev_next_upper_dev(now, &iter, &ignore);
7056 niter = &udev->adj_list.upper;
7057 dev_stack[cur] = now;
7058 iter_stack[cur++] = iter;
7065 next = dev_stack[--cur];
7066 niter = iter_stack[cur];
7076 int netdev_walk_all_upper_dev_rcu(struct net_device *dev,
7077 int (*fn)(struct net_device *dev,
7078 struct netdev_nested_priv *priv),
7079 struct netdev_nested_priv *priv)
7081 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7082 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7086 iter = &dev->adj_list.upper;
7090 ret = fn(now, priv);
7097 udev = netdev_next_upper_dev_rcu(now, &iter);
7102 niter = &udev->adj_list.upper;
7103 dev_stack[cur] = now;
7104 iter_stack[cur++] = iter;
7111 next = dev_stack[--cur];
7112 niter = iter_stack[cur];
7121 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu);
7123 static bool __netdev_has_upper_dev(struct net_device *dev,
7124 struct net_device *upper_dev)
7126 struct netdev_nested_priv priv = {
7128 .data = (void *)upper_dev,
7133 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev,
7138 * netdev_lower_get_next_private - Get the next ->private from the
7139 * lower neighbour list
7141 * @iter: list_head ** of the current position
7143 * Gets the next netdev_adjacent->private from the dev's lower neighbour
7144 * list, starting from iter position. The caller must hold either hold the
7145 * RTNL lock or its own locking that guarantees that the neighbour lower
7146 * list will remain unchanged.
7148 void *netdev_lower_get_next_private(struct net_device *dev,
7149 struct list_head **iter)
7151 struct netdev_adjacent *lower;
7153 lower = list_entry(*iter, struct netdev_adjacent, list);
7155 if (&lower->list == &dev->adj_list.lower)
7158 *iter = lower->list.next;
7160 return lower->private;
7162 EXPORT_SYMBOL(netdev_lower_get_next_private);
7165 * netdev_lower_get_next_private_rcu - Get the next ->private from the
7166 * lower neighbour list, RCU
7169 * @iter: list_head ** of the current position
7171 * Gets the next netdev_adjacent->private from the dev's lower neighbour
7172 * list, starting from iter position. The caller must hold RCU read lock.
7174 void *netdev_lower_get_next_private_rcu(struct net_device *dev,
7175 struct list_head **iter)
7177 struct netdev_adjacent *lower;
7179 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
7181 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7183 if (&lower->list == &dev->adj_list.lower)
7186 *iter = &lower->list;
7188 return lower->private;
7190 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu);
7193 * netdev_lower_get_next - Get the next device from the lower neighbour
7196 * @iter: list_head ** of the current position
7198 * Gets the next netdev_adjacent from the dev's lower neighbour
7199 * list, starting from iter position. The caller must hold RTNL lock or
7200 * its own locking that guarantees that the neighbour lower
7201 * list will remain unchanged.
7203 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter)
7205 struct netdev_adjacent *lower;
7207 lower = list_entry(*iter, struct netdev_adjacent, list);
7209 if (&lower->list == &dev->adj_list.lower)
7212 *iter = lower->list.next;
7216 EXPORT_SYMBOL(netdev_lower_get_next);
7218 static struct net_device *netdev_next_lower_dev(struct net_device *dev,
7219 struct list_head **iter)
7221 struct netdev_adjacent *lower;
7223 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
7225 if (&lower->list == &dev->adj_list.lower)
7228 *iter = &lower->list;
7233 static struct net_device *__netdev_next_lower_dev(struct net_device *dev,
7234 struct list_head **iter,
7237 struct netdev_adjacent *lower;
7239 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
7241 if (&lower->list == &dev->adj_list.lower)
7244 *iter = &lower->list;
7245 *ignore = lower->ignore;
7250 int netdev_walk_all_lower_dev(struct net_device *dev,
7251 int (*fn)(struct net_device *dev,
7252 struct netdev_nested_priv *priv),
7253 struct netdev_nested_priv *priv)
7255 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7256 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7260 iter = &dev->adj_list.lower;
7264 ret = fn(now, priv);
7271 ldev = netdev_next_lower_dev(now, &iter);
7276 niter = &ldev->adj_list.lower;
7277 dev_stack[cur] = now;
7278 iter_stack[cur++] = iter;
7285 next = dev_stack[--cur];
7286 niter = iter_stack[cur];
7295 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev);
7297 static int __netdev_walk_all_lower_dev(struct net_device *dev,
7298 int (*fn)(struct net_device *dev,
7299 struct netdev_nested_priv *priv),
7300 struct netdev_nested_priv *priv)
7302 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7303 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7308 iter = &dev->adj_list.lower;
7312 ret = fn(now, priv);
7319 ldev = __netdev_next_lower_dev(now, &iter, &ignore);
7326 niter = &ldev->adj_list.lower;
7327 dev_stack[cur] = now;
7328 iter_stack[cur++] = iter;
7335 next = dev_stack[--cur];
7336 niter = iter_stack[cur];
7346 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev,
7347 struct list_head **iter)
7349 struct netdev_adjacent *lower;
7351 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7352 if (&lower->list == &dev->adj_list.lower)
7355 *iter = &lower->list;
7359 EXPORT_SYMBOL(netdev_next_lower_dev_rcu);
7361 static u8 __netdev_upper_depth(struct net_device *dev)
7363 struct net_device *udev;
7364 struct list_head *iter;
7368 for (iter = &dev->adj_list.upper,
7369 udev = __netdev_next_upper_dev(dev, &iter, &ignore);
7371 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) {
7374 if (max_depth < udev->upper_level)
7375 max_depth = udev->upper_level;
7381 static u8 __netdev_lower_depth(struct net_device *dev)
7383 struct net_device *ldev;
7384 struct list_head *iter;
7388 for (iter = &dev->adj_list.lower,
7389 ldev = __netdev_next_lower_dev(dev, &iter, &ignore);
7391 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) {
7394 if (max_depth < ldev->lower_level)
7395 max_depth = ldev->lower_level;
7401 static int __netdev_update_upper_level(struct net_device *dev,
7402 struct netdev_nested_priv *__unused)
7404 dev->upper_level = __netdev_upper_depth(dev) + 1;
7408 #ifdef CONFIG_LOCKDEP
7409 static LIST_HEAD(net_unlink_list);
7411 static void net_unlink_todo(struct net_device *dev)
7413 if (list_empty(&dev->unlink_list))
7414 list_add_tail(&dev->unlink_list, &net_unlink_list);
7418 static int __netdev_update_lower_level(struct net_device *dev,
7419 struct netdev_nested_priv *priv)
7421 dev->lower_level = __netdev_lower_depth(dev) + 1;
7423 #ifdef CONFIG_LOCKDEP
7427 if (priv->flags & NESTED_SYNC_IMM)
7428 dev->nested_level = dev->lower_level - 1;
7429 if (priv->flags & NESTED_SYNC_TODO)
7430 net_unlink_todo(dev);
7435 int netdev_walk_all_lower_dev_rcu(struct net_device *dev,
7436 int (*fn)(struct net_device *dev,
7437 struct netdev_nested_priv *priv),
7438 struct netdev_nested_priv *priv)
7440 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7441 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7445 iter = &dev->adj_list.lower;
7449 ret = fn(now, priv);
7456 ldev = netdev_next_lower_dev_rcu(now, &iter);
7461 niter = &ldev->adj_list.lower;
7462 dev_stack[cur] = now;
7463 iter_stack[cur++] = iter;
7470 next = dev_stack[--cur];
7471 niter = iter_stack[cur];
7480 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu);
7483 * netdev_lower_get_first_private_rcu - Get the first ->private from the
7484 * lower neighbour list, RCU
7488 * Gets the first netdev_adjacent->private from the dev's lower neighbour
7489 * list. The caller must hold RCU read lock.
7491 void *netdev_lower_get_first_private_rcu(struct net_device *dev)
7493 struct netdev_adjacent *lower;
7495 lower = list_first_or_null_rcu(&dev->adj_list.lower,
7496 struct netdev_adjacent, list);
7498 return lower->private;
7501 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu);
7504 * netdev_master_upper_dev_get_rcu - Get master upper device
7507 * Find a master upper device and return pointer to it or NULL in case
7508 * it's not there. The caller must hold the RCU read lock.
7510 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev)
7512 struct netdev_adjacent *upper;
7514 upper = list_first_or_null_rcu(&dev->adj_list.upper,
7515 struct netdev_adjacent, list);
7516 if (upper && likely(upper->master))
7520 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu);
7522 static int netdev_adjacent_sysfs_add(struct net_device *dev,
7523 struct net_device *adj_dev,
7524 struct list_head *dev_list)
7526 char linkname[IFNAMSIZ+7];
7528 sprintf(linkname, dev_list == &dev->adj_list.upper ?
7529 "upper_%s" : "lower_%s", adj_dev->name);
7530 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj),
7533 static void netdev_adjacent_sysfs_del(struct net_device *dev,
7535 struct list_head *dev_list)
7537 char linkname[IFNAMSIZ+7];
7539 sprintf(linkname, dev_list == &dev->adj_list.upper ?
7540 "upper_%s" : "lower_%s", name);
7541 sysfs_remove_link(&(dev->dev.kobj), linkname);
7544 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev,
7545 struct net_device *adj_dev,
7546 struct list_head *dev_list)
7548 return (dev_list == &dev->adj_list.upper ||
7549 dev_list == &dev->adj_list.lower) &&
7550 net_eq(dev_net(dev), dev_net(adj_dev));
7553 static int __netdev_adjacent_dev_insert(struct net_device *dev,
7554 struct net_device *adj_dev,
7555 struct list_head *dev_list,
7556 void *private, bool master)
7558 struct netdev_adjacent *adj;
7561 adj = __netdev_find_adj(adj_dev, dev_list);
7565 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n",
7566 dev->name, adj_dev->name, adj->ref_nr);
7571 adj = kmalloc(sizeof(*adj), GFP_KERNEL);
7576 adj->master = master;
7578 adj->private = private;
7579 adj->ignore = false;
7580 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL);
7582 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n",
7583 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name);
7585 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) {
7586 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list);
7591 /* Ensure that master link is always the first item in list. */
7593 ret = sysfs_create_link(&(dev->dev.kobj),
7594 &(adj_dev->dev.kobj), "master");
7596 goto remove_symlinks;
7598 list_add_rcu(&adj->list, dev_list);
7600 list_add_tail_rcu(&adj->list, dev_list);
7606 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
7607 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
7609 netdev_put(adj_dev, &adj->dev_tracker);
7615 static void __netdev_adjacent_dev_remove(struct net_device *dev,
7616 struct net_device *adj_dev,
7618 struct list_head *dev_list)
7620 struct netdev_adjacent *adj;
7622 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n",
7623 dev->name, adj_dev->name, ref_nr);
7625 adj = __netdev_find_adj(adj_dev, dev_list);
7628 pr_err("Adjacency does not exist for device %s from %s\n",
7629 dev->name, adj_dev->name);
7634 if (adj->ref_nr > ref_nr) {
7635 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n",
7636 dev->name, adj_dev->name, ref_nr,
7637 adj->ref_nr - ref_nr);
7638 adj->ref_nr -= ref_nr;
7643 sysfs_remove_link(&(dev->dev.kobj), "master");
7645 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
7646 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
7648 list_del_rcu(&adj->list);
7649 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n",
7650 adj_dev->name, dev->name, adj_dev->name);
7651 netdev_put(adj_dev, &adj->dev_tracker);
7652 kfree_rcu(adj, rcu);
7655 static int __netdev_adjacent_dev_link_lists(struct net_device *dev,
7656 struct net_device *upper_dev,
7657 struct list_head *up_list,
7658 struct list_head *down_list,
7659 void *private, bool master)
7663 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list,
7668 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list,
7671 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list);
7678 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev,
7679 struct net_device *upper_dev,
7681 struct list_head *up_list,
7682 struct list_head *down_list)
7684 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list);
7685 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list);
7688 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev,
7689 struct net_device *upper_dev,
7690 void *private, bool master)
7692 return __netdev_adjacent_dev_link_lists(dev, upper_dev,
7693 &dev->adj_list.upper,
7694 &upper_dev->adj_list.lower,
7698 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev,
7699 struct net_device *upper_dev)
7701 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1,
7702 &dev->adj_list.upper,
7703 &upper_dev->adj_list.lower);
7706 static int __netdev_upper_dev_link(struct net_device *dev,
7707 struct net_device *upper_dev, bool master,
7708 void *upper_priv, void *upper_info,
7709 struct netdev_nested_priv *priv,
7710 struct netlink_ext_ack *extack)
7712 struct netdev_notifier_changeupper_info changeupper_info = {
7717 .upper_dev = upper_dev,
7720 .upper_info = upper_info,
7722 struct net_device *master_dev;
7727 if (dev == upper_dev)
7730 /* To prevent loops, check if dev is not upper device to upper_dev. */
7731 if (__netdev_has_upper_dev(upper_dev, dev))
7734 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV)
7738 if (__netdev_has_upper_dev(dev, upper_dev))
7741 master_dev = __netdev_master_upper_dev_get(dev);
7743 return master_dev == upper_dev ? -EEXIST : -EBUSY;
7746 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
7747 &changeupper_info.info);
7748 ret = notifier_to_errno(ret);
7752 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv,
7757 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
7758 &changeupper_info.info);
7759 ret = notifier_to_errno(ret);
7763 __netdev_update_upper_level(dev, NULL);
7764 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL);
7766 __netdev_update_lower_level(upper_dev, priv);
7767 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level,
7773 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
7779 * netdev_upper_dev_link - Add a link to the upper device
7781 * @upper_dev: new upper device
7782 * @extack: netlink extended ack
7784 * Adds a link to device which is upper to this one. The caller must hold
7785 * the RTNL lock. On a failure a negative errno code is returned.
7786 * On success the reference counts are adjusted and the function
7789 int netdev_upper_dev_link(struct net_device *dev,
7790 struct net_device *upper_dev,
7791 struct netlink_ext_ack *extack)
7793 struct netdev_nested_priv priv = {
7794 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO,
7798 return __netdev_upper_dev_link(dev, upper_dev, false,
7799 NULL, NULL, &priv, extack);
7801 EXPORT_SYMBOL(netdev_upper_dev_link);
7804 * netdev_master_upper_dev_link - Add a master link to the upper device
7806 * @upper_dev: new upper device
7807 * @upper_priv: upper device private
7808 * @upper_info: upper info to be passed down via notifier
7809 * @extack: netlink extended ack
7811 * Adds a link to device which is upper to this one. In this case, only
7812 * one master upper device can be linked, although other non-master devices
7813 * might be linked as well. The caller must hold the RTNL lock.
7814 * On a failure a negative errno code is returned. On success the reference
7815 * counts are adjusted and the function returns zero.
7817 int netdev_master_upper_dev_link(struct net_device *dev,
7818 struct net_device *upper_dev,
7819 void *upper_priv, void *upper_info,
7820 struct netlink_ext_ack *extack)
7822 struct netdev_nested_priv priv = {
7823 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO,
7827 return __netdev_upper_dev_link(dev, upper_dev, true,
7828 upper_priv, upper_info, &priv, extack);
7830 EXPORT_SYMBOL(netdev_master_upper_dev_link);
7832 static void __netdev_upper_dev_unlink(struct net_device *dev,
7833 struct net_device *upper_dev,
7834 struct netdev_nested_priv *priv)
7836 struct netdev_notifier_changeupper_info changeupper_info = {
7840 .upper_dev = upper_dev,
7846 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev;
7848 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
7849 &changeupper_info.info);
7851 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
7853 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
7854 &changeupper_info.info);
7856 __netdev_update_upper_level(dev, NULL);
7857 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL);
7859 __netdev_update_lower_level(upper_dev, priv);
7860 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level,
7865 * netdev_upper_dev_unlink - Removes a link to upper device
7867 * @upper_dev: new upper device
7869 * Removes a link to device which is upper to this one. The caller must hold
7872 void netdev_upper_dev_unlink(struct net_device *dev,
7873 struct net_device *upper_dev)
7875 struct netdev_nested_priv priv = {
7876 .flags = NESTED_SYNC_TODO,
7880 __netdev_upper_dev_unlink(dev, upper_dev, &priv);
7882 EXPORT_SYMBOL(netdev_upper_dev_unlink);
7884 static void __netdev_adjacent_dev_set(struct net_device *upper_dev,
7885 struct net_device *lower_dev,
7888 struct netdev_adjacent *adj;
7890 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower);
7894 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper);
7899 static void netdev_adjacent_dev_disable(struct net_device *upper_dev,
7900 struct net_device *lower_dev)
7902 __netdev_adjacent_dev_set(upper_dev, lower_dev, true);
7905 static void netdev_adjacent_dev_enable(struct net_device *upper_dev,
7906 struct net_device *lower_dev)
7908 __netdev_adjacent_dev_set(upper_dev, lower_dev, false);
7911 int netdev_adjacent_change_prepare(struct net_device *old_dev,
7912 struct net_device *new_dev,
7913 struct net_device *dev,
7914 struct netlink_ext_ack *extack)
7916 struct netdev_nested_priv priv = {
7925 if (old_dev && new_dev != old_dev)
7926 netdev_adjacent_dev_disable(dev, old_dev);
7927 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv,
7930 if (old_dev && new_dev != old_dev)
7931 netdev_adjacent_dev_enable(dev, old_dev);
7937 EXPORT_SYMBOL(netdev_adjacent_change_prepare);
7939 void netdev_adjacent_change_commit(struct net_device *old_dev,
7940 struct net_device *new_dev,
7941 struct net_device *dev)
7943 struct netdev_nested_priv priv = {
7944 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO,
7948 if (!new_dev || !old_dev)
7951 if (new_dev == old_dev)
7954 netdev_adjacent_dev_enable(dev, old_dev);
7955 __netdev_upper_dev_unlink(old_dev, dev, &priv);
7957 EXPORT_SYMBOL(netdev_adjacent_change_commit);
7959 void netdev_adjacent_change_abort(struct net_device *old_dev,
7960 struct net_device *new_dev,
7961 struct net_device *dev)
7963 struct netdev_nested_priv priv = {
7971 if (old_dev && new_dev != old_dev)
7972 netdev_adjacent_dev_enable(dev, old_dev);
7974 __netdev_upper_dev_unlink(new_dev, dev, &priv);
7976 EXPORT_SYMBOL(netdev_adjacent_change_abort);
7979 * netdev_bonding_info_change - Dispatch event about slave change
7981 * @bonding_info: info to dispatch
7983 * Send NETDEV_BONDING_INFO to netdev notifiers with info.
7984 * The caller must hold the RTNL lock.
7986 void netdev_bonding_info_change(struct net_device *dev,
7987 struct netdev_bonding_info *bonding_info)
7989 struct netdev_notifier_bonding_info info = {
7993 memcpy(&info.bonding_info, bonding_info,
7994 sizeof(struct netdev_bonding_info));
7995 call_netdevice_notifiers_info(NETDEV_BONDING_INFO,
7998 EXPORT_SYMBOL(netdev_bonding_info_change);
8000 static int netdev_offload_xstats_enable_l3(struct net_device *dev,
8001 struct netlink_ext_ack *extack)
8003 struct netdev_notifier_offload_xstats_info info = {
8005 .info.extack = extack,
8006 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3,
8011 dev->offload_xstats_l3 = kzalloc(sizeof(*dev->offload_xstats_l3),
8013 if (!dev->offload_xstats_l3)
8016 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE,
8017 NETDEV_OFFLOAD_XSTATS_DISABLE,
8019 err = notifier_to_errno(rc);
8026 kfree(dev->offload_xstats_l3);
8027 dev->offload_xstats_l3 = NULL;
8031 int netdev_offload_xstats_enable(struct net_device *dev,
8032 enum netdev_offload_xstats_type type,
8033 struct netlink_ext_ack *extack)
8037 if (netdev_offload_xstats_enabled(dev, type))
8041 case NETDEV_OFFLOAD_XSTATS_TYPE_L3:
8042 return netdev_offload_xstats_enable_l3(dev, extack);
8048 EXPORT_SYMBOL(netdev_offload_xstats_enable);
8050 static void netdev_offload_xstats_disable_l3(struct net_device *dev)
8052 struct netdev_notifier_offload_xstats_info info = {
8054 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3,
8057 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE,
8059 kfree(dev->offload_xstats_l3);
8060 dev->offload_xstats_l3 = NULL;
8063 int netdev_offload_xstats_disable(struct net_device *dev,
8064 enum netdev_offload_xstats_type type)
8068 if (!netdev_offload_xstats_enabled(dev, type))
8072 case NETDEV_OFFLOAD_XSTATS_TYPE_L3:
8073 netdev_offload_xstats_disable_l3(dev);
8080 EXPORT_SYMBOL(netdev_offload_xstats_disable);
8082 static void netdev_offload_xstats_disable_all(struct net_device *dev)
8084 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3);
8087 static struct rtnl_hw_stats64 *
8088 netdev_offload_xstats_get_ptr(const struct net_device *dev,
8089 enum netdev_offload_xstats_type type)
8092 case NETDEV_OFFLOAD_XSTATS_TYPE_L3:
8093 return dev->offload_xstats_l3;
8100 bool netdev_offload_xstats_enabled(const struct net_device *dev,
8101 enum netdev_offload_xstats_type type)
8105 return netdev_offload_xstats_get_ptr(dev, type);
8107 EXPORT_SYMBOL(netdev_offload_xstats_enabled);
8109 struct netdev_notifier_offload_xstats_ru {
8113 struct netdev_notifier_offload_xstats_rd {
8114 struct rtnl_hw_stats64 stats;
8118 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest,
8119 const struct rtnl_hw_stats64 *src)
8121 dest->rx_packets += src->rx_packets;
8122 dest->tx_packets += src->tx_packets;
8123 dest->rx_bytes += src->rx_bytes;
8124 dest->tx_bytes += src->tx_bytes;
8125 dest->rx_errors += src->rx_errors;
8126 dest->tx_errors += src->tx_errors;
8127 dest->rx_dropped += src->rx_dropped;
8128 dest->tx_dropped += src->tx_dropped;
8129 dest->multicast += src->multicast;
8132 static int netdev_offload_xstats_get_used(struct net_device *dev,
8133 enum netdev_offload_xstats_type type,
8135 struct netlink_ext_ack *extack)
8137 struct netdev_notifier_offload_xstats_ru report_used = {};
8138 struct netdev_notifier_offload_xstats_info info = {
8140 .info.extack = extack,
8142 .report_used = &report_used,
8146 WARN_ON(!netdev_offload_xstats_enabled(dev, type));
8147 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED,
8149 *p_used = report_used.used;
8150 return notifier_to_errno(rc);
8153 static int netdev_offload_xstats_get_stats(struct net_device *dev,
8154 enum netdev_offload_xstats_type type,
8155 struct rtnl_hw_stats64 *p_stats,
8157 struct netlink_ext_ack *extack)
8159 struct netdev_notifier_offload_xstats_rd report_delta = {};
8160 struct netdev_notifier_offload_xstats_info info = {
8162 .info.extack = extack,
8164 .report_delta = &report_delta,
8166 struct rtnl_hw_stats64 *stats;
8169 stats = netdev_offload_xstats_get_ptr(dev, type);
8170 if (WARN_ON(!stats))
8173 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA,
8176 /* Cache whatever we got, even if there was an error, otherwise the
8177 * successful stats retrievals would get lost.
8179 netdev_hw_stats64_add(stats, &report_delta.stats);
8183 *p_used = report_delta.used;
8185 return notifier_to_errno(rc);
8188 int netdev_offload_xstats_get(struct net_device *dev,
8189 enum netdev_offload_xstats_type type,
8190 struct rtnl_hw_stats64 *p_stats, bool *p_used,
8191 struct netlink_ext_ack *extack)
8196 return netdev_offload_xstats_get_stats(dev, type, p_stats,
8199 return netdev_offload_xstats_get_used(dev, type, p_used,
8202 EXPORT_SYMBOL(netdev_offload_xstats_get);
8205 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta,
8206 const struct rtnl_hw_stats64 *stats)
8208 report_delta->used = true;
8209 netdev_hw_stats64_add(&report_delta->stats, stats);
8211 EXPORT_SYMBOL(netdev_offload_xstats_report_delta);
8214 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used)
8216 report_used->used = true;
8218 EXPORT_SYMBOL(netdev_offload_xstats_report_used);
8220 void netdev_offload_xstats_push_delta(struct net_device *dev,
8221 enum netdev_offload_xstats_type type,
8222 const struct rtnl_hw_stats64 *p_stats)
8224 struct rtnl_hw_stats64 *stats;
8228 stats = netdev_offload_xstats_get_ptr(dev, type);
8229 if (WARN_ON(!stats))
8232 netdev_hw_stats64_add(stats, p_stats);
8234 EXPORT_SYMBOL(netdev_offload_xstats_push_delta);
8237 * netdev_get_xmit_slave - Get the xmit slave of master device
8240 * @all_slaves: assume all the slaves are active
8242 * The reference counters are not incremented so the caller must be
8243 * careful with locks. The caller must hold RCU lock.
8244 * %NULL is returned if no slave is found.
8247 struct net_device *netdev_get_xmit_slave(struct net_device *dev,
8248 struct sk_buff *skb,
8251 const struct net_device_ops *ops = dev->netdev_ops;
8253 if (!ops->ndo_get_xmit_slave)
8255 return ops->ndo_get_xmit_slave(dev, skb, all_slaves);
8257 EXPORT_SYMBOL(netdev_get_xmit_slave);
8259 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev,
8262 const struct net_device_ops *ops = dev->netdev_ops;
8264 if (!ops->ndo_sk_get_lower_dev)
8266 return ops->ndo_sk_get_lower_dev(dev, sk);
8270 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket
8274 * %NULL is returned if no lower device is found.
8277 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev,
8280 struct net_device *lower;
8282 lower = netdev_sk_get_lower_dev(dev, sk);
8285 lower = netdev_sk_get_lower_dev(dev, sk);
8290 EXPORT_SYMBOL(netdev_sk_get_lowest_dev);
8292 static void netdev_adjacent_add_links(struct net_device *dev)
8294 struct netdev_adjacent *iter;
8296 struct net *net = dev_net(dev);
8298 list_for_each_entry(iter, &dev->adj_list.upper, list) {
8299 if (!net_eq(net, dev_net(iter->dev)))
8301 netdev_adjacent_sysfs_add(iter->dev, dev,
8302 &iter->dev->adj_list.lower);
8303 netdev_adjacent_sysfs_add(dev, iter->dev,
8304 &dev->adj_list.upper);
8307 list_for_each_entry(iter, &dev->adj_list.lower, list) {
8308 if (!net_eq(net, dev_net(iter->dev)))
8310 netdev_adjacent_sysfs_add(iter->dev, dev,
8311 &iter->dev->adj_list.upper);
8312 netdev_adjacent_sysfs_add(dev, iter->dev,
8313 &dev->adj_list.lower);
8317 static void netdev_adjacent_del_links(struct net_device *dev)
8319 struct netdev_adjacent *iter;
8321 struct net *net = dev_net(dev);
8323 list_for_each_entry(iter, &dev->adj_list.upper, list) {
8324 if (!net_eq(net, dev_net(iter->dev)))
8326 netdev_adjacent_sysfs_del(iter->dev, dev->name,
8327 &iter->dev->adj_list.lower);
8328 netdev_adjacent_sysfs_del(dev, iter->dev->name,
8329 &dev->adj_list.upper);
8332 list_for_each_entry(iter, &dev->adj_list.lower, list) {
8333 if (!net_eq(net, dev_net(iter->dev)))
8335 netdev_adjacent_sysfs_del(iter->dev, dev->name,
8336 &iter->dev->adj_list.upper);
8337 netdev_adjacent_sysfs_del(dev, iter->dev->name,
8338 &dev->adj_list.lower);
8342 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname)
8344 struct netdev_adjacent *iter;
8346 struct net *net = dev_net(dev);
8348 list_for_each_entry(iter, &dev->adj_list.upper, list) {
8349 if (!net_eq(net, dev_net(iter->dev)))
8351 netdev_adjacent_sysfs_del(iter->dev, oldname,
8352 &iter->dev->adj_list.lower);
8353 netdev_adjacent_sysfs_add(iter->dev, dev,
8354 &iter->dev->adj_list.lower);
8357 list_for_each_entry(iter, &dev->adj_list.lower, list) {
8358 if (!net_eq(net, dev_net(iter->dev)))
8360 netdev_adjacent_sysfs_del(iter->dev, oldname,
8361 &iter->dev->adj_list.upper);
8362 netdev_adjacent_sysfs_add(iter->dev, dev,
8363 &iter->dev->adj_list.upper);
8367 void *netdev_lower_dev_get_private(struct net_device *dev,
8368 struct net_device *lower_dev)
8370 struct netdev_adjacent *lower;
8374 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower);
8378 return lower->private;
8380 EXPORT_SYMBOL(netdev_lower_dev_get_private);
8384 * netdev_lower_state_changed - Dispatch event about lower device state change
8385 * @lower_dev: device
8386 * @lower_state_info: state to dispatch
8388 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info.
8389 * The caller must hold the RTNL lock.
8391 void netdev_lower_state_changed(struct net_device *lower_dev,
8392 void *lower_state_info)
8394 struct netdev_notifier_changelowerstate_info changelowerstate_info = {
8395 .info.dev = lower_dev,
8399 changelowerstate_info.lower_state_info = lower_state_info;
8400 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE,
8401 &changelowerstate_info.info);
8403 EXPORT_SYMBOL(netdev_lower_state_changed);
8405 static void dev_change_rx_flags(struct net_device *dev, int flags)
8407 const struct net_device_ops *ops = dev->netdev_ops;
8409 if (ops->ndo_change_rx_flags)
8410 ops->ndo_change_rx_flags(dev, flags);
8413 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify)
8415 unsigned int old_flags = dev->flags;
8421 dev->flags |= IFF_PROMISC;
8422 dev->promiscuity += inc;
8423 if (dev->promiscuity == 0) {
8426 * If inc causes overflow, untouch promisc and return error.
8429 dev->flags &= ~IFF_PROMISC;
8431 dev->promiscuity -= inc;
8432 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n");
8436 if (dev->flags != old_flags) {
8437 netdev_info(dev, "%s promiscuous mode\n",
8438 dev->flags & IFF_PROMISC ? "entered" : "left");
8439 if (audit_enabled) {
8440 current_uid_gid(&uid, &gid);
8441 audit_log(audit_context(), GFP_ATOMIC,
8442 AUDIT_ANOM_PROMISCUOUS,
8443 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u",
8444 dev->name, (dev->flags & IFF_PROMISC),
8445 (old_flags & IFF_PROMISC),
8446 from_kuid(&init_user_ns, audit_get_loginuid(current)),
8447 from_kuid(&init_user_ns, uid),
8448 from_kgid(&init_user_ns, gid),
8449 audit_get_sessionid(current));
8452 dev_change_rx_flags(dev, IFF_PROMISC);
8455 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL);
8460 * dev_set_promiscuity - update promiscuity count on a device
8464 * Add or remove promiscuity from a device. While the count in the device
8465 * remains above zero the interface remains promiscuous. Once it hits zero
8466 * the device reverts back to normal filtering operation. A negative inc
8467 * value is used to drop promiscuity on the device.
8468 * Return 0 if successful or a negative errno code on error.
8470 int dev_set_promiscuity(struct net_device *dev, int inc)
8472 unsigned int old_flags = dev->flags;
8475 err = __dev_set_promiscuity(dev, inc, true);
8478 if (dev->flags != old_flags)
8479 dev_set_rx_mode(dev);
8482 EXPORT_SYMBOL(dev_set_promiscuity);
8484 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify)
8486 unsigned int old_flags = dev->flags, old_gflags = dev->gflags;
8490 dev->flags |= IFF_ALLMULTI;
8491 dev->allmulti += inc;
8492 if (dev->allmulti == 0) {
8495 * If inc causes overflow, untouch allmulti and return error.
8498 dev->flags &= ~IFF_ALLMULTI;
8500 dev->allmulti -= inc;
8501 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n");
8505 if (dev->flags ^ old_flags) {
8506 netdev_info(dev, "%s allmulticast mode\n",
8507 dev->flags & IFF_ALLMULTI ? "entered" : "left");
8508 dev_change_rx_flags(dev, IFF_ALLMULTI);
8509 dev_set_rx_mode(dev);
8511 __dev_notify_flags(dev, old_flags,
8512 dev->gflags ^ old_gflags, 0, NULL);
8518 * dev_set_allmulti - update allmulti count on a device
8522 * Add or remove reception of all multicast frames to a device. While the
8523 * count in the device remains above zero the interface remains listening
8524 * to all interfaces. Once it hits zero the device reverts back to normal
8525 * filtering operation. A negative @inc value is used to drop the counter
8526 * when releasing a resource needing all multicasts.
8527 * Return 0 if successful or a negative errno code on error.
8530 int dev_set_allmulti(struct net_device *dev, int inc)
8532 return __dev_set_allmulti(dev, inc, true);
8534 EXPORT_SYMBOL(dev_set_allmulti);
8537 * Upload unicast and multicast address lists to device and
8538 * configure RX filtering. When the device doesn't support unicast
8539 * filtering it is put in promiscuous mode while unicast addresses
8542 void __dev_set_rx_mode(struct net_device *dev)
8544 const struct net_device_ops *ops = dev->netdev_ops;
8546 /* dev_open will call this function so the list will stay sane. */
8547 if (!(dev->flags&IFF_UP))
8550 if (!netif_device_present(dev))
8553 if (!(dev->priv_flags & IFF_UNICAST_FLT)) {
8554 /* Unicast addresses changes may only happen under the rtnl,
8555 * therefore calling __dev_set_promiscuity here is safe.
8557 if (!netdev_uc_empty(dev) && !dev->uc_promisc) {
8558 __dev_set_promiscuity(dev, 1, false);
8559 dev->uc_promisc = true;
8560 } else if (netdev_uc_empty(dev) && dev->uc_promisc) {
8561 __dev_set_promiscuity(dev, -1, false);
8562 dev->uc_promisc = false;
8566 if (ops->ndo_set_rx_mode)
8567 ops->ndo_set_rx_mode(dev);
8570 void dev_set_rx_mode(struct net_device *dev)
8572 netif_addr_lock_bh(dev);
8573 __dev_set_rx_mode(dev);
8574 netif_addr_unlock_bh(dev);
8578 * dev_get_flags - get flags reported to userspace
8581 * Get the combination of flag bits exported through APIs to userspace.
8583 unsigned int dev_get_flags(const struct net_device *dev)
8587 flags = (dev->flags & ~(IFF_PROMISC |
8592 (dev->gflags & (IFF_PROMISC |
8595 if (netif_running(dev)) {
8596 if (netif_oper_up(dev))
8597 flags |= IFF_RUNNING;
8598 if (netif_carrier_ok(dev))
8599 flags |= IFF_LOWER_UP;
8600 if (netif_dormant(dev))
8601 flags |= IFF_DORMANT;
8606 EXPORT_SYMBOL(dev_get_flags);
8608 int __dev_change_flags(struct net_device *dev, unsigned int flags,
8609 struct netlink_ext_ack *extack)
8611 unsigned int old_flags = dev->flags;
8617 * Set the flags on our device.
8620 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP |
8621 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL |
8623 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC |
8627 * Load in the correct multicast list now the flags have changed.
8630 if ((old_flags ^ flags) & IFF_MULTICAST)
8631 dev_change_rx_flags(dev, IFF_MULTICAST);
8633 dev_set_rx_mode(dev);
8636 * Have we downed the interface. We handle IFF_UP ourselves
8637 * according to user attempts to set it, rather than blindly
8642 if ((old_flags ^ flags) & IFF_UP) {
8643 if (old_flags & IFF_UP)
8646 ret = __dev_open(dev, extack);
8649 if ((flags ^ dev->gflags) & IFF_PROMISC) {
8650 int inc = (flags & IFF_PROMISC) ? 1 : -1;
8651 unsigned int old_flags = dev->flags;
8653 dev->gflags ^= IFF_PROMISC;
8655 if (__dev_set_promiscuity(dev, inc, false) >= 0)
8656 if (dev->flags != old_flags)
8657 dev_set_rx_mode(dev);
8660 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI
8661 * is important. Some (broken) drivers set IFF_PROMISC, when
8662 * IFF_ALLMULTI is requested not asking us and not reporting.
8664 if ((flags ^ dev->gflags) & IFF_ALLMULTI) {
8665 int inc = (flags & IFF_ALLMULTI) ? 1 : -1;
8667 dev->gflags ^= IFF_ALLMULTI;
8668 __dev_set_allmulti(dev, inc, false);
8674 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags,
8675 unsigned int gchanges, u32 portid,
8676 const struct nlmsghdr *nlh)
8678 unsigned int changes = dev->flags ^ old_flags;
8681 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh);
8683 if (changes & IFF_UP) {
8684 if (dev->flags & IFF_UP)
8685 call_netdevice_notifiers(NETDEV_UP, dev);
8687 call_netdevice_notifiers(NETDEV_DOWN, dev);
8690 if (dev->flags & IFF_UP &&
8691 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) {
8692 struct netdev_notifier_change_info change_info = {
8696 .flags_changed = changes,
8699 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info);
8704 * dev_change_flags - change device settings
8706 * @flags: device state flags
8707 * @extack: netlink extended ack
8709 * Change settings on device based state flags. The flags are
8710 * in the userspace exported format.
8712 int dev_change_flags(struct net_device *dev, unsigned int flags,
8713 struct netlink_ext_ack *extack)
8716 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags;
8718 ret = __dev_change_flags(dev, flags, extack);
8722 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags);
8723 __dev_notify_flags(dev, old_flags, changes, 0, NULL);
8726 EXPORT_SYMBOL(dev_change_flags);
8728 int __dev_set_mtu(struct net_device *dev, int new_mtu)
8730 const struct net_device_ops *ops = dev->netdev_ops;
8732 if (ops->ndo_change_mtu)
8733 return ops->ndo_change_mtu(dev, new_mtu);
8735 /* Pairs with all the lockless reads of dev->mtu in the stack */
8736 WRITE_ONCE(dev->mtu, new_mtu);
8739 EXPORT_SYMBOL(__dev_set_mtu);
8741 int dev_validate_mtu(struct net_device *dev, int new_mtu,
8742 struct netlink_ext_ack *extack)
8744 /* MTU must be positive, and in range */
8745 if (new_mtu < 0 || new_mtu < dev->min_mtu) {
8746 NL_SET_ERR_MSG(extack, "mtu less than device minimum");
8750 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) {
8751 NL_SET_ERR_MSG(extack, "mtu greater than device maximum");
8758 * dev_set_mtu_ext - Change maximum transfer unit
8760 * @new_mtu: new transfer unit
8761 * @extack: netlink extended ack
8763 * Change the maximum transfer size of the network device.
8765 int dev_set_mtu_ext(struct net_device *dev, int new_mtu,
8766 struct netlink_ext_ack *extack)
8770 if (new_mtu == dev->mtu)
8773 err = dev_validate_mtu(dev, new_mtu, extack);
8777 if (!netif_device_present(dev))
8780 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev);
8781 err = notifier_to_errno(err);
8785 orig_mtu = dev->mtu;
8786 err = __dev_set_mtu(dev, new_mtu);
8789 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev,
8791 err = notifier_to_errno(err);
8793 /* setting mtu back and notifying everyone again,
8794 * so that they have a chance to revert changes.
8796 __dev_set_mtu(dev, orig_mtu);
8797 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev,
8804 int dev_set_mtu(struct net_device *dev, int new_mtu)
8806 struct netlink_ext_ack extack;
8809 memset(&extack, 0, sizeof(extack));
8810 err = dev_set_mtu_ext(dev, new_mtu, &extack);
8811 if (err && extack._msg)
8812 net_err_ratelimited("%s: %s\n", dev->name, extack._msg);
8815 EXPORT_SYMBOL(dev_set_mtu);
8818 * dev_change_tx_queue_len - Change TX queue length of a netdevice
8820 * @new_len: new tx queue length
8822 int dev_change_tx_queue_len(struct net_device *dev, unsigned long new_len)
8824 unsigned int orig_len = dev->tx_queue_len;
8827 if (new_len != (unsigned int)new_len)
8830 if (new_len != orig_len) {
8831 dev->tx_queue_len = new_len;
8832 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev);
8833 res = notifier_to_errno(res);
8836 res = dev_qdisc_change_tx_queue_len(dev);
8844 netdev_err(dev, "refused to change device tx_queue_len\n");
8845 dev->tx_queue_len = orig_len;
8850 * dev_set_group - Change group this device belongs to
8852 * @new_group: group this device should belong to
8854 void dev_set_group(struct net_device *dev, int new_group)
8856 dev->group = new_group;
8860 * dev_pre_changeaddr_notify - Call NETDEV_PRE_CHANGEADDR.
8862 * @addr: new address
8863 * @extack: netlink extended ack
8865 int dev_pre_changeaddr_notify(struct net_device *dev, const char *addr,
8866 struct netlink_ext_ack *extack)
8868 struct netdev_notifier_pre_changeaddr_info info = {
8870 .info.extack = extack,
8875 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info);
8876 return notifier_to_errno(rc);
8878 EXPORT_SYMBOL(dev_pre_changeaddr_notify);
8881 * dev_set_mac_address - Change Media Access Control Address
8884 * @extack: netlink extended ack
8886 * Change the hardware (MAC) address of the device
8888 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa,
8889 struct netlink_ext_ack *extack)
8891 const struct net_device_ops *ops = dev->netdev_ops;
8894 if (!ops->ndo_set_mac_address)
8896 if (sa->sa_family != dev->type)
8898 if (!netif_device_present(dev))
8900 err = dev_pre_changeaddr_notify(dev, sa->sa_data, extack);
8903 if (memcmp(dev->dev_addr, sa->sa_data, dev->addr_len)) {
8904 err = ops->ndo_set_mac_address(dev, sa);
8908 dev->addr_assign_type = NET_ADDR_SET;
8909 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
8910 add_device_randomness(dev->dev_addr, dev->addr_len);
8913 EXPORT_SYMBOL(dev_set_mac_address);
8915 static DECLARE_RWSEM(dev_addr_sem);
8917 int dev_set_mac_address_user(struct net_device *dev, struct sockaddr *sa,
8918 struct netlink_ext_ack *extack)
8922 down_write(&dev_addr_sem);
8923 ret = dev_set_mac_address(dev, sa, extack);
8924 up_write(&dev_addr_sem);
8927 EXPORT_SYMBOL(dev_set_mac_address_user);
8929 int dev_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name)
8931 size_t size = sizeof(sa->sa_data_min);
8932 struct net_device *dev;
8935 down_read(&dev_addr_sem);
8938 dev = dev_get_by_name_rcu(net, dev_name);
8944 memset(sa->sa_data, 0, size);
8946 memcpy(sa->sa_data, dev->dev_addr,
8947 min_t(size_t, size, dev->addr_len));
8948 sa->sa_family = dev->type;
8952 up_read(&dev_addr_sem);
8955 EXPORT_SYMBOL(dev_get_mac_address);
8958 * dev_change_carrier - Change device carrier
8960 * @new_carrier: new value
8962 * Change device carrier
8964 int dev_change_carrier(struct net_device *dev, bool new_carrier)
8966 const struct net_device_ops *ops = dev->netdev_ops;
8968 if (!ops->ndo_change_carrier)
8970 if (!netif_device_present(dev))
8972 return ops->ndo_change_carrier(dev, new_carrier);
8976 * dev_get_phys_port_id - Get device physical port ID
8980 * Get device physical port ID
8982 int dev_get_phys_port_id(struct net_device *dev,
8983 struct netdev_phys_item_id *ppid)
8985 const struct net_device_ops *ops = dev->netdev_ops;
8987 if (!ops->ndo_get_phys_port_id)
8989 return ops->ndo_get_phys_port_id(dev, ppid);
8993 * dev_get_phys_port_name - Get device physical port name
8996 * @len: limit of bytes to copy to name
8998 * Get device physical port name
9000 int dev_get_phys_port_name(struct net_device *dev,
9001 char *name, size_t len)
9003 const struct net_device_ops *ops = dev->netdev_ops;
9006 if (ops->ndo_get_phys_port_name) {
9007 err = ops->ndo_get_phys_port_name(dev, name, len);
9008 if (err != -EOPNOTSUPP)
9011 return devlink_compat_phys_port_name_get(dev, name, len);
9015 * dev_get_port_parent_id - Get the device's port parent identifier
9016 * @dev: network device
9017 * @ppid: pointer to a storage for the port's parent identifier
9018 * @recurse: allow/disallow recursion to lower devices
9020 * Get the devices's port parent identifier
9022 int dev_get_port_parent_id(struct net_device *dev,
9023 struct netdev_phys_item_id *ppid,
9026 const struct net_device_ops *ops = dev->netdev_ops;
9027 struct netdev_phys_item_id first = { };
9028 struct net_device *lower_dev;
9029 struct list_head *iter;
9032 if (ops->ndo_get_port_parent_id) {
9033 err = ops->ndo_get_port_parent_id(dev, ppid);
9034 if (err != -EOPNOTSUPP)
9038 err = devlink_compat_switch_id_get(dev, ppid);
9039 if (!recurse || err != -EOPNOTSUPP)
9042 netdev_for_each_lower_dev(dev, lower_dev, iter) {
9043 err = dev_get_port_parent_id(lower_dev, ppid, true);
9048 else if (memcmp(&first, ppid, sizeof(*ppid)))
9054 EXPORT_SYMBOL(dev_get_port_parent_id);
9057 * netdev_port_same_parent_id - Indicate if two network devices have
9058 * the same port parent identifier
9059 * @a: first network device
9060 * @b: second network device
9062 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b)
9064 struct netdev_phys_item_id a_id = { };
9065 struct netdev_phys_item_id b_id = { };
9067 if (dev_get_port_parent_id(a, &a_id, true) ||
9068 dev_get_port_parent_id(b, &b_id, true))
9071 return netdev_phys_item_id_same(&a_id, &b_id);
9073 EXPORT_SYMBOL(netdev_port_same_parent_id);
9076 * dev_change_proto_down - set carrier according to proto_down.
9079 * @proto_down: new value
9081 int dev_change_proto_down(struct net_device *dev, bool proto_down)
9083 if (!(dev->priv_flags & IFF_CHANGE_PROTO_DOWN))
9085 if (!netif_device_present(dev))
9088 netif_carrier_off(dev);
9090 netif_carrier_on(dev);
9091 dev->proto_down = proto_down;
9096 * dev_change_proto_down_reason - proto down reason
9099 * @mask: proto down mask
9100 * @value: proto down value
9102 void dev_change_proto_down_reason(struct net_device *dev, unsigned long mask,
9108 dev->proto_down_reason = value;
9110 for_each_set_bit(b, &mask, 32) {
9111 if (value & (1 << b))
9112 dev->proto_down_reason |= BIT(b);
9114 dev->proto_down_reason &= ~BIT(b);
9119 struct bpf_xdp_link {
9120 struct bpf_link link;
9121 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */
9125 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags)
9127 if (flags & XDP_FLAGS_HW_MODE)
9129 if (flags & XDP_FLAGS_DRV_MODE)
9130 return XDP_MODE_DRV;
9131 if (flags & XDP_FLAGS_SKB_MODE)
9132 return XDP_MODE_SKB;
9133 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB;
9136 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode)
9140 return generic_xdp_install;
9143 return dev->netdev_ops->ndo_bpf;
9149 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev,
9150 enum bpf_xdp_mode mode)
9152 return dev->xdp_state[mode].link;
9155 static struct bpf_prog *dev_xdp_prog(struct net_device *dev,
9156 enum bpf_xdp_mode mode)
9158 struct bpf_xdp_link *link = dev_xdp_link(dev, mode);
9161 return link->link.prog;
9162 return dev->xdp_state[mode].prog;
9165 u8 dev_xdp_prog_count(struct net_device *dev)
9170 for (i = 0; i < __MAX_XDP_MODE; i++)
9171 if (dev->xdp_state[i].prog || dev->xdp_state[i].link)
9175 EXPORT_SYMBOL_GPL(dev_xdp_prog_count);
9177 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode)
9179 struct bpf_prog *prog = dev_xdp_prog(dev, mode);
9181 return prog ? prog->aux->id : 0;
9184 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode,
9185 struct bpf_xdp_link *link)
9187 dev->xdp_state[mode].link = link;
9188 dev->xdp_state[mode].prog = NULL;
9191 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode,
9192 struct bpf_prog *prog)
9194 dev->xdp_state[mode].link = NULL;
9195 dev->xdp_state[mode].prog = prog;
9198 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode,
9199 bpf_op_t bpf_op, struct netlink_ext_ack *extack,
9200 u32 flags, struct bpf_prog *prog)
9202 struct netdev_bpf xdp;
9205 memset(&xdp, 0, sizeof(xdp));
9206 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG;
9207 xdp.extack = extack;
9211 /* Drivers assume refcnt is already incremented (i.e, prog pointer is
9212 * "moved" into driver), so they don't increment it on their own, but
9213 * they do decrement refcnt when program is detached or replaced.
9214 * Given net_device also owns link/prog, we need to bump refcnt here
9215 * to prevent drivers from underflowing it.
9219 err = bpf_op(dev, &xdp);
9226 if (mode != XDP_MODE_HW)
9227 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog);
9232 static void dev_xdp_uninstall(struct net_device *dev)
9234 struct bpf_xdp_link *link;
9235 struct bpf_prog *prog;
9236 enum bpf_xdp_mode mode;
9241 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) {
9242 prog = dev_xdp_prog(dev, mode);
9246 bpf_op = dev_xdp_bpf_op(dev, mode);
9250 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL));
9252 /* auto-detach link from net device */
9253 link = dev_xdp_link(dev, mode);
9259 dev_xdp_set_link(dev, mode, NULL);
9263 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack,
9264 struct bpf_xdp_link *link, struct bpf_prog *new_prog,
9265 struct bpf_prog *old_prog, u32 flags)
9267 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES);
9268 struct bpf_prog *cur_prog;
9269 struct net_device *upper;
9270 struct list_head *iter;
9271 enum bpf_xdp_mode mode;
9277 /* either link or prog attachment, never both */
9278 if (link && (new_prog || old_prog))
9280 /* link supports only XDP mode flags */
9281 if (link && (flags & ~XDP_FLAGS_MODES)) {
9282 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment");
9285 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */
9286 if (num_modes > 1) {
9287 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set");
9290 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */
9291 if (!num_modes && dev_xdp_prog_count(dev) > 1) {
9292 NL_SET_ERR_MSG(extack,
9293 "More than one program loaded, unset mode is ambiguous");
9296 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */
9297 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) {
9298 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified");
9302 mode = dev_xdp_mode(dev, flags);
9303 /* can't replace attached link */
9304 if (dev_xdp_link(dev, mode)) {
9305 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link");
9309 /* don't allow if an upper device already has a program */
9310 netdev_for_each_upper_dev_rcu(dev, upper, iter) {
9311 if (dev_xdp_prog_count(upper) > 0) {
9312 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program");
9317 cur_prog = dev_xdp_prog(dev, mode);
9318 /* can't replace attached prog with link */
9319 if (link && cur_prog) {
9320 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link");
9323 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) {
9324 NL_SET_ERR_MSG(extack, "Active program does not match expected");
9328 /* put effective new program into new_prog */
9330 new_prog = link->link.prog;
9333 bool offload = mode == XDP_MODE_HW;
9334 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB
9335 ? XDP_MODE_DRV : XDP_MODE_SKB;
9337 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) {
9338 NL_SET_ERR_MSG(extack, "XDP program already attached");
9341 if (!offload && dev_xdp_prog(dev, other_mode)) {
9342 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time");
9345 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) {
9346 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported");
9349 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) {
9350 NL_SET_ERR_MSG(extack, "Program bound to different device");
9353 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) {
9354 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device");
9357 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) {
9358 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device");
9363 /* don't call drivers if the effective program didn't change */
9364 if (new_prog != cur_prog) {
9365 bpf_op = dev_xdp_bpf_op(dev, mode);
9367 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode");
9371 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog);
9377 dev_xdp_set_link(dev, mode, link);
9379 dev_xdp_set_prog(dev, mode, new_prog);
9381 bpf_prog_put(cur_prog);
9386 static int dev_xdp_attach_link(struct net_device *dev,
9387 struct netlink_ext_ack *extack,
9388 struct bpf_xdp_link *link)
9390 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags);
9393 static int dev_xdp_detach_link(struct net_device *dev,
9394 struct netlink_ext_ack *extack,
9395 struct bpf_xdp_link *link)
9397 enum bpf_xdp_mode mode;
9402 mode = dev_xdp_mode(dev, link->flags);
9403 if (dev_xdp_link(dev, mode) != link)
9406 bpf_op = dev_xdp_bpf_op(dev, mode);
9407 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL));
9408 dev_xdp_set_link(dev, mode, NULL);
9412 static void bpf_xdp_link_release(struct bpf_link *link)
9414 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9418 /* if racing with net_device's tear down, xdp_link->dev might be
9419 * already NULL, in which case link was already auto-detached
9421 if (xdp_link->dev) {
9422 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link));
9423 xdp_link->dev = NULL;
9429 static int bpf_xdp_link_detach(struct bpf_link *link)
9431 bpf_xdp_link_release(link);
9435 static void bpf_xdp_link_dealloc(struct bpf_link *link)
9437 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9442 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link,
9443 struct seq_file *seq)
9445 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9450 ifindex = xdp_link->dev->ifindex;
9453 seq_printf(seq, "ifindex:\t%u\n", ifindex);
9456 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link,
9457 struct bpf_link_info *info)
9459 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9464 ifindex = xdp_link->dev->ifindex;
9467 info->xdp.ifindex = ifindex;
9471 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog,
9472 struct bpf_prog *old_prog)
9474 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9475 enum bpf_xdp_mode mode;
9481 /* link might have been auto-released already, so fail */
9482 if (!xdp_link->dev) {
9487 if (old_prog && link->prog != old_prog) {
9491 old_prog = link->prog;
9492 if (old_prog->type != new_prog->type ||
9493 old_prog->expected_attach_type != new_prog->expected_attach_type) {
9498 if (old_prog == new_prog) {
9499 /* no-op, don't disturb drivers */
9500 bpf_prog_put(new_prog);
9504 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags);
9505 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode);
9506 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL,
9507 xdp_link->flags, new_prog);
9511 old_prog = xchg(&link->prog, new_prog);
9512 bpf_prog_put(old_prog);
9519 static const struct bpf_link_ops bpf_xdp_link_lops = {
9520 .release = bpf_xdp_link_release,
9521 .dealloc = bpf_xdp_link_dealloc,
9522 .detach = bpf_xdp_link_detach,
9523 .show_fdinfo = bpf_xdp_link_show_fdinfo,
9524 .fill_link_info = bpf_xdp_link_fill_link_info,
9525 .update_prog = bpf_xdp_link_update,
9528 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog)
9530 struct net *net = current->nsproxy->net_ns;
9531 struct bpf_link_primer link_primer;
9532 struct netlink_ext_ack extack = {};
9533 struct bpf_xdp_link *link;
9534 struct net_device *dev;
9538 dev = dev_get_by_index(net, attr->link_create.target_ifindex);
9544 link = kzalloc(sizeof(*link), GFP_USER);
9550 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog);
9552 link->flags = attr->link_create.flags;
9554 err = bpf_link_prime(&link->link, &link_primer);
9560 err = dev_xdp_attach_link(dev, &extack, link);
9565 bpf_link_cleanup(&link_primer);
9566 trace_bpf_xdp_link_attach_failed(extack._msg);
9570 fd = bpf_link_settle(&link_primer);
9571 /* link itself doesn't hold dev's refcnt to not complicate shutdown */
9584 * dev_change_xdp_fd - set or clear a bpf program for a device rx path
9586 * @extack: netlink extended ack
9587 * @fd: new program fd or negative value to clear
9588 * @expected_fd: old program fd that userspace expects to replace or clear
9589 * @flags: xdp-related flags
9591 * Set or clear a bpf program for a device
9593 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack,
9594 int fd, int expected_fd, u32 flags)
9596 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags);
9597 struct bpf_prog *new_prog = NULL, *old_prog = NULL;
9603 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP,
9604 mode != XDP_MODE_SKB);
9605 if (IS_ERR(new_prog))
9606 return PTR_ERR(new_prog);
9609 if (expected_fd >= 0) {
9610 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP,
9611 mode != XDP_MODE_SKB);
9612 if (IS_ERR(old_prog)) {
9613 err = PTR_ERR(old_prog);
9619 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags);
9622 if (err && new_prog)
9623 bpf_prog_put(new_prog);
9625 bpf_prog_put(old_prog);
9630 * dev_index_reserve() - allocate an ifindex in a namespace
9631 * @net: the applicable net namespace
9632 * @ifindex: requested ifindex, pass %0 to get one allocated
9634 * Allocate a ifindex for a new device. Caller must either use the ifindex
9635 * to store the device (via list_netdevice()) or call dev_index_release()
9636 * to give the index up.
9638 * Return: a suitable unique value for a new device interface number or -errno.
9640 static int dev_index_reserve(struct net *net, u32 ifindex)
9644 if (ifindex > INT_MAX) {
9645 DEBUG_NET_WARN_ON_ONCE(1);
9650 err = xa_alloc_cyclic(&net->dev_by_index, &ifindex, NULL,
9651 xa_limit_31b, &net->ifindex, GFP_KERNEL);
9653 err = xa_insert(&net->dev_by_index, ifindex, NULL, GFP_KERNEL);
9660 static void dev_index_release(struct net *net, int ifindex)
9662 /* Expect only unused indexes, unlist_netdevice() removes the used */
9663 WARN_ON(xa_erase(&net->dev_by_index, ifindex));
9666 /* Delayed registration/unregisteration */
9667 LIST_HEAD(net_todo_list);
9668 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq);
9670 static void net_set_todo(struct net_device *dev)
9672 list_add_tail(&dev->todo_list, &net_todo_list);
9673 atomic_inc(&dev_net(dev)->dev_unreg_count);
9676 static netdev_features_t netdev_sync_upper_features(struct net_device *lower,
9677 struct net_device *upper, netdev_features_t features)
9679 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
9680 netdev_features_t feature;
9683 for_each_netdev_feature(upper_disables, feature_bit) {
9684 feature = __NETIF_F_BIT(feature_bit);
9685 if (!(upper->wanted_features & feature)
9686 && (features & feature)) {
9687 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n",
9688 &feature, upper->name);
9689 features &= ~feature;
9696 static void netdev_sync_lower_features(struct net_device *upper,
9697 struct net_device *lower, netdev_features_t features)
9699 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
9700 netdev_features_t feature;
9703 for_each_netdev_feature(upper_disables, feature_bit) {
9704 feature = __NETIF_F_BIT(feature_bit);
9705 if (!(features & feature) && (lower->features & feature)) {
9706 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n",
9707 &feature, lower->name);
9708 lower->wanted_features &= ~feature;
9709 __netdev_update_features(lower);
9711 if (unlikely(lower->features & feature))
9712 netdev_WARN(upper, "failed to disable %pNF on %s!\n",
9713 &feature, lower->name);
9715 netdev_features_change(lower);
9720 static netdev_features_t netdev_fix_features(struct net_device *dev,
9721 netdev_features_t features)
9723 /* Fix illegal checksum combinations */
9724 if ((features & NETIF_F_HW_CSUM) &&
9725 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
9726 netdev_warn(dev, "mixed HW and IP checksum settings.\n");
9727 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
9730 /* TSO requires that SG is present as well. */
9731 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) {
9732 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n");
9733 features &= ~NETIF_F_ALL_TSO;
9736 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) &&
9737 !(features & NETIF_F_IP_CSUM)) {
9738 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n");
9739 features &= ~NETIF_F_TSO;
9740 features &= ~NETIF_F_TSO_ECN;
9743 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) &&
9744 !(features & NETIF_F_IPV6_CSUM)) {
9745 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n");
9746 features &= ~NETIF_F_TSO6;
9749 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */
9750 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO))
9751 features &= ~NETIF_F_TSO_MANGLEID;
9753 /* TSO ECN requires that TSO is present as well. */
9754 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN)
9755 features &= ~NETIF_F_TSO_ECN;
9757 /* Software GSO depends on SG. */
9758 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) {
9759 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n");
9760 features &= ~NETIF_F_GSO;
9763 /* GSO partial features require GSO partial be set */
9764 if ((features & dev->gso_partial_features) &&
9765 !(features & NETIF_F_GSO_PARTIAL)) {
9767 "Dropping partially supported GSO features since no GSO partial.\n");
9768 features &= ~dev->gso_partial_features;
9771 if (!(features & NETIF_F_RXCSUM)) {
9772 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet
9773 * successfully merged by hardware must also have the
9774 * checksum verified by hardware. If the user does not
9775 * want to enable RXCSUM, logically, we should disable GRO_HW.
9777 if (features & NETIF_F_GRO_HW) {
9778 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n");
9779 features &= ~NETIF_F_GRO_HW;
9783 /* LRO/HW-GRO features cannot be combined with RX-FCS */
9784 if (features & NETIF_F_RXFCS) {
9785 if (features & NETIF_F_LRO) {
9786 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n");
9787 features &= ~NETIF_F_LRO;
9790 if (features & NETIF_F_GRO_HW) {
9791 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n");
9792 features &= ~NETIF_F_GRO_HW;
9796 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) {
9797 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n");
9798 features &= ~NETIF_F_LRO;
9801 if (features & NETIF_F_HW_TLS_TX) {
9802 bool ip_csum = (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) ==
9803 (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM);
9804 bool hw_csum = features & NETIF_F_HW_CSUM;
9806 if (!ip_csum && !hw_csum) {
9807 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n");
9808 features &= ~NETIF_F_HW_TLS_TX;
9812 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) {
9813 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n");
9814 features &= ~NETIF_F_HW_TLS_RX;
9820 int __netdev_update_features(struct net_device *dev)
9822 struct net_device *upper, *lower;
9823 netdev_features_t features;
9824 struct list_head *iter;
9829 features = netdev_get_wanted_features(dev);
9831 if (dev->netdev_ops->ndo_fix_features)
9832 features = dev->netdev_ops->ndo_fix_features(dev, features);
9834 /* driver might be less strict about feature dependencies */
9835 features = netdev_fix_features(dev, features);
9837 /* some features can't be enabled if they're off on an upper device */
9838 netdev_for_each_upper_dev_rcu(dev, upper, iter)
9839 features = netdev_sync_upper_features(dev, upper, features);
9841 if (dev->features == features)
9844 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n",
9845 &dev->features, &features);
9847 if (dev->netdev_ops->ndo_set_features)
9848 err = dev->netdev_ops->ndo_set_features(dev, features);
9852 if (unlikely(err < 0)) {
9854 "set_features() failed (%d); wanted %pNF, left %pNF\n",
9855 err, &features, &dev->features);
9856 /* return non-0 since some features might have changed and
9857 * it's better to fire a spurious notification than miss it
9863 /* some features must be disabled on lower devices when disabled
9864 * on an upper device (think: bonding master or bridge)
9866 netdev_for_each_lower_dev(dev, lower, iter)
9867 netdev_sync_lower_features(dev, lower, features);
9870 netdev_features_t diff = features ^ dev->features;
9872 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) {
9873 /* udp_tunnel_{get,drop}_rx_info both need
9874 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the
9875 * device, or they won't do anything.
9876 * Thus we need to update dev->features
9877 * *before* calling udp_tunnel_get_rx_info,
9878 * but *after* calling udp_tunnel_drop_rx_info.
9880 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) {
9881 dev->features = features;
9882 udp_tunnel_get_rx_info(dev);
9884 udp_tunnel_drop_rx_info(dev);
9888 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) {
9889 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) {
9890 dev->features = features;
9891 err |= vlan_get_rx_ctag_filter_info(dev);
9893 vlan_drop_rx_ctag_filter_info(dev);
9897 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) {
9898 if (features & NETIF_F_HW_VLAN_STAG_FILTER) {
9899 dev->features = features;
9900 err |= vlan_get_rx_stag_filter_info(dev);
9902 vlan_drop_rx_stag_filter_info(dev);
9906 dev->features = features;
9909 return err < 0 ? 0 : 1;
9913 * netdev_update_features - recalculate device features
9914 * @dev: the device to check
9916 * Recalculate dev->features set and send notifications if it
9917 * has changed. Should be called after driver or hardware dependent
9918 * conditions might have changed that influence the features.
9920 void netdev_update_features(struct net_device *dev)
9922 if (__netdev_update_features(dev))
9923 netdev_features_change(dev);
9925 EXPORT_SYMBOL(netdev_update_features);
9928 * netdev_change_features - recalculate device features
9929 * @dev: the device to check
9931 * Recalculate dev->features set and send notifications even
9932 * if they have not changed. Should be called instead of
9933 * netdev_update_features() if also dev->vlan_features might
9934 * have changed to allow the changes to be propagated to stacked
9937 void netdev_change_features(struct net_device *dev)
9939 __netdev_update_features(dev);
9940 netdev_features_change(dev);
9942 EXPORT_SYMBOL(netdev_change_features);
9945 * netif_stacked_transfer_operstate - transfer operstate
9946 * @rootdev: the root or lower level device to transfer state from
9947 * @dev: the device to transfer operstate to
9949 * Transfer operational state from root to device. This is normally
9950 * called when a stacking relationship exists between the root
9951 * device and the device(a leaf device).
9953 void netif_stacked_transfer_operstate(const struct net_device *rootdev,
9954 struct net_device *dev)
9956 if (rootdev->operstate == IF_OPER_DORMANT)
9957 netif_dormant_on(dev);
9959 netif_dormant_off(dev);
9961 if (rootdev->operstate == IF_OPER_TESTING)
9962 netif_testing_on(dev);
9964 netif_testing_off(dev);
9966 if (netif_carrier_ok(rootdev))
9967 netif_carrier_on(dev);
9969 netif_carrier_off(dev);
9971 EXPORT_SYMBOL(netif_stacked_transfer_operstate);
9973 static int netif_alloc_rx_queues(struct net_device *dev)
9975 unsigned int i, count = dev->num_rx_queues;
9976 struct netdev_rx_queue *rx;
9977 size_t sz = count * sizeof(*rx);
9982 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL);
9988 for (i = 0; i < count; i++) {
9991 /* XDP RX-queue setup */
9992 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0);
9999 /* Rollback successful reg's and free other resources */
10001 xdp_rxq_info_unreg(&rx[i].xdp_rxq);
10007 static void netif_free_rx_queues(struct net_device *dev)
10009 unsigned int i, count = dev->num_rx_queues;
10011 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */
10015 for (i = 0; i < count; i++)
10016 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq);
10021 static void netdev_init_one_queue(struct net_device *dev,
10022 struct netdev_queue *queue, void *_unused)
10024 /* Initialize queue lock */
10025 spin_lock_init(&queue->_xmit_lock);
10026 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type);
10027 queue->xmit_lock_owner = -1;
10028 netdev_queue_numa_node_write(queue, NUMA_NO_NODE);
10031 dql_init(&queue->dql, HZ);
10035 static void netif_free_tx_queues(struct net_device *dev)
10040 static int netif_alloc_netdev_queues(struct net_device *dev)
10042 unsigned int count = dev->num_tx_queues;
10043 struct netdev_queue *tx;
10044 size_t sz = count * sizeof(*tx);
10046 if (count < 1 || count > 0xffff)
10049 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL);
10055 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL);
10056 spin_lock_init(&dev->tx_global_lock);
10061 void netif_tx_stop_all_queues(struct net_device *dev)
10065 for (i = 0; i < dev->num_tx_queues; i++) {
10066 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
10068 netif_tx_stop_queue(txq);
10071 EXPORT_SYMBOL(netif_tx_stop_all_queues);
10073 static int netdev_do_alloc_pcpu_stats(struct net_device *dev)
10077 /* Drivers implementing ndo_get_peer_dev must support tstat
10078 * accounting, so that skb_do_redirect() can bump the dev's
10079 * RX stats upon network namespace switch.
10081 if (dev->netdev_ops->ndo_get_peer_dev &&
10082 dev->pcpu_stat_type != NETDEV_PCPU_STAT_TSTATS)
10083 return -EOPNOTSUPP;
10085 switch (dev->pcpu_stat_type) {
10086 case NETDEV_PCPU_STAT_NONE:
10088 case NETDEV_PCPU_STAT_LSTATS:
10089 v = dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats);
10091 case NETDEV_PCPU_STAT_TSTATS:
10092 v = dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats);
10094 case NETDEV_PCPU_STAT_DSTATS:
10095 v = dev->dstats = netdev_alloc_pcpu_stats(struct pcpu_dstats);
10101 return v ? 0 : -ENOMEM;
10104 static void netdev_do_free_pcpu_stats(struct net_device *dev)
10106 switch (dev->pcpu_stat_type) {
10107 case NETDEV_PCPU_STAT_NONE:
10109 case NETDEV_PCPU_STAT_LSTATS:
10110 free_percpu(dev->lstats);
10112 case NETDEV_PCPU_STAT_TSTATS:
10113 free_percpu(dev->tstats);
10115 case NETDEV_PCPU_STAT_DSTATS:
10116 free_percpu(dev->dstats);
10122 * register_netdevice() - register a network device
10123 * @dev: device to register
10125 * Take a prepared network device structure and make it externally accessible.
10126 * A %NETDEV_REGISTER message is sent to the netdev notifier chain.
10127 * Callers must hold the rtnl lock - you may want register_netdev()
10130 int register_netdevice(struct net_device *dev)
10133 struct net *net = dev_net(dev);
10135 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE <
10136 NETDEV_FEATURE_COUNT);
10137 BUG_ON(dev_boot_phase);
10142 /* When net_device's are persistent, this will be fatal. */
10143 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
10146 ret = ethtool_check_ops(dev->ethtool_ops);
10150 spin_lock_init(&dev->addr_list_lock);
10151 netdev_set_addr_lockdep_class(dev);
10153 ret = dev_get_valid_name(net, dev, dev->name);
10158 dev->name_node = netdev_name_node_head_alloc(dev);
10159 if (!dev->name_node)
10162 /* Init, if this function is available */
10163 if (dev->netdev_ops->ndo_init) {
10164 ret = dev->netdev_ops->ndo_init(dev);
10168 goto err_free_name;
10172 if (((dev->hw_features | dev->features) &
10173 NETIF_F_HW_VLAN_CTAG_FILTER) &&
10174 (!dev->netdev_ops->ndo_vlan_rx_add_vid ||
10175 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) {
10176 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n");
10181 ret = netdev_do_alloc_pcpu_stats(dev);
10185 ret = dev_index_reserve(net, dev->ifindex);
10187 goto err_free_pcpu;
10188 dev->ifindex = ret;
10190 /* Transfer changeable features to wanted_features and enable
10191 * software offloads (GSO and GRO).
10193 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF);
10194 dev->features |= NETIF_F_SOFT_FEATURES;
10196 if (dev->udp_tunnel_nic_info) {
10197 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT;
10198 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT;
10201 dev->wanted_features = dev->features & dev->hw_features;
10203 if (!(dev->flags & IFF_LOOPBACK))
10204 dev->hw_features |= NETIF_F_NOCACHE_COPY;
10206 /* If IPv4 TCP segmentation offload is supported we should also
10207 * allow the device to enable segmenting the frame with the option
10208 * of ignoring a static IP ID value. This doesn't enable the
10209 * feature itself but allows the user to enable it later.
10211 if (dev->hw_features & NETIF_F_TSO)
10212 dev->hw_features |= NETIF_F_TSO_MANGLEID;
10213 if (dev->vlan_features & NETIF_F_TSO)
10214 dev->vlan_features |= NETIF_F_TSO_MANGLEID;
10215 if (dev->mpls_features & NETIF_F_TSO)
10216 dev->mpls_features |= NETIF_F_TSO_MANGLEID;
10217 if (dev->hw_enc_features & NETIF_F_TSO)
10218 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID;
10220 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices.
10222 dev->vlan_features |= NETIF_F_HIGHDMA;
10224 /* Make NETIF_F_SG inheritable to tunnel devices.
10226 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL;
10228 /* Make NETIF_F_SG inheritable to MPLS.
10230 dev->mpls_features |= NETIF_F_SG;
10232 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev);
10233 ret = notifier_to_errno(ret);
10235 goto err_ifindex_release;
10237 ret = netdev_register_kobject(dev);
10238 write_lock(&dev_base_lock);
10239 dev->reg_state = ret ? NETREG_UNREGISTERED : NETREG_REGISTERED;
10240 write_unlock(&dev_base_lock);
10242 goto err_uninit_notify;
10244 __netdev_update_features(dev);
10247 * Default initial state at registry is that the
10248 * device is present.
10251 set_bit(__LINK_STATE_PRESENT, &dev->state);
10253 linkwatch_init_dev(dev);
10255 dev_init_scheduler(dev);
10257 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL);
10258 list_netdevice(dev);
10260 add_device_randomness(dev->dev_addr, dev->addr_len);
10262 /* If the device has permanent device address, driver should
10263 * set dev_addr and also addr_assign_type should be set to
10264 * NET_ADDR_PERM (default value).
10266 if (dev->addr_assign_type == NET_ADDR_PERM)
10267 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len);
10269 /* Notify protocols, that a new device appeared. */
10270 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
10271 ret = notifier_to_errno(ret);
10273 /* Expect explicit free_netdev() on failure */
10274 dev->needs_free_netdev = false;
10275 unregister_netdevice_queue(dev, NULL);
10279 * Prevent userspace races by waiting until the network
10280 * device is fully setup before sending notifications.
10282 if (!dev->rtnl_link_ops ||
10283 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
10284 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL);
10290 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev);
10291 err_ifindex_release:
10292 dev_index_release(net, dev->ifindex);
10294 netdev_do_free_pcpu_stats(dev);
10296 if (dev->netdev_ops->ndo_uninit)
10297 dev->netdev_ops->ndo_uninit(dev);
10298 if (dev->priv_destructor)
10299 dev->priv_destructor(dev);
10301 netdev_name_node_free(dev->name_node);
10304 EXPORT_SYMBOL(register_netdevice);
10307 * init_dummy_netdev - init a dummy network device for NAPI
10308 * @dev: device to init
10310 * This takes a network device structure and initialize the minimum
10311 * amount of fields so it can be used to schedule NAPI polls without
10312 * registering a full blown interface. This is to be used by drivers
10313 * that need to tie several hardware interfaces to a single NAPI
10314 * poll scheduler due to HW limitations.
10316 int init_dummy_netdev(struct net_device *dev)
10318 /* Clear everything. Note we don't initialize spinlocks
10319 * are they aren't supposed to be taken by any of the
10320 * NAPI code and this dummy netdev is supposed to be
10321 * only ever used for NAPI polls
10323 memset(dev, 0, sizeof(struct net_device));
10325 /* make sure we BUG if trying to hit standard
10326 * register/unregister code path
10328 dev->reg_state = NETREG_DUMMY;
10330 /* NAPI wants this */
10331 INIT_LIST_HEAD(&dev->napi_list);
10333 /* a dummy interface is started by default */
10334 set_bit(__LINK_STATE_PRESENT, &dev->state);
10335 set_bit(__LINK_STATE_START, &dev->state);
10337 /* napi_busy_loop stats accounting wants this */
10338 dev_net_set(dev, &init_net);
10340 /* Note : We dont allocate pcpu_refcnt for dummy devices,
10341 * because users of this 'device' dont need to change
10347 EXPORT_SYMBOL_GPL(init_dummy_netdev);
10351 * register_netdev - register a network device
10352 * @dev: device to register
10354 * Take a completed network device structure and add it to the kernel
10355 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
10356 * chain. 0 is returned on success. A negative errno code is returned
10357 * on a failure to set up the device, or if the name is a duplicate.
10359 * This is a wrapper around register_netdevice that takes the rtnl semaphore
10360 * and expands the device name if you passed a format string to
10363 int register_netdev(struct net_device *dev)
10367 if (rtnl_lock_killable())
10369 err = register_netdevice(dev);
10373 EXPORT_SYMBOL(register_netdev);
10375 int netdev_refcnt_read(const struct net_device *dev)
10377 #ifdef CONFIG_PCPU_DEV_REFCNT
10380 for_each_possible_cpu(i)
10381 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
10384 return refcount_read(&dev->dev_refcnt);
10387 EXPORT_SYMBOL(netdev_refcnt_read);
10389 int netdev_unregister_timeout_secs __read_mostly = 10;
10391 #define WAIT_REFS_MIN_MSECS 1
10392 #define WAIT_REFS_MAX_MSECS 250
10394 * netdev_wait_allrefs_any - wait until all references are gone.
10395 * @list: list of net_devices to wait on
10397 * This is called when unregistering network devices.
10399 * Any protocol or device that holds a reference should register
10400 * for netdevice notification, and cleanup and put back the
10401 * reference if they receive an UNREGISTER event.
10402 * We can get stuck here if buggy protocols don't correctly
10405 static struct net_device *netdev_wait_allrefs_any(struct list_head *list)
10407 unsigned long rebroadcast_time, warning_time;
10408 struct net_device *dev;
10411 rebroadcast_time = warning_time = jiffies;
10413 list_for_each_entry(dev, list, todo_list)
10414 if (netdev_refcnt_read(dev) == 1)
10418 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
10421 /* Rebroadcast unregister notification */
10422 list_for_each_entry(dev, list, todo_list)
10423 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
10429 list_for_each_entry(dev, list, todo_list)
10430 if (test_bit(__LINK_STATE_LINKWATCH_PENDING,
10432 /* We must not have linkwatch events
10433 * pending on unregister. If this
10434 * happens, we simply run the queue
10435 * unscheduled, resulting in a noop
10438 linkwatch_run_queue();
10444 rebroadcast_time = jiffies;
10449 wait = WAIT_REFS_MIN_MSECS;
10452 wait = min(wait << 1, WAIT_REFS_MAX_MSECS);
10455 list_for_each_entry(dev, list, todo_list)
10456 if (netdev_refcnt_read(dev) == 1)
10459 if (time_after(jiffies, warning_time +
10460 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) {
10461 list_for_each_entry(dev, list, todo_list) {
10462 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n",
10463 dev->name, netdev_refcnt_read(dev));
10464 ref_tracker_dir_print(&dev->refcnt_tracker, 10);
10467 warning_time = jiffies;
10472 /* The sequence is:
10476 * register_netdevice(x1);
10477 * register_netdevice(x2);
10479 * unregister_netdevice(y1);
10480 * unregister_netdevice(y2);
10486 * We are invoked by rtnl_unlock().
10487 * This allows us to deal with problems:
10488 * 1) We can delete sysfs objects which invoke hotplug
10489 * without deadlocking with linkwatch via keventd.
10490 * 2) Since we run with the RTNL semaphore not held, we can sleep
10491 * safely in order to wait for the netdev refcnt to drop to zero.
10493 * We must not return until all unregister events added during
10494 * the interval the lock was held have been completed.
10496 void netdev_run_todo(void)
10498 struct net_device *dev, *tmp;
10499 struct list_head list;
10500 #ifdef CONFIG_LOCKDEP
10501 struct list_head unlink_list;
10503 list_replace_init(&net_unlink_list, &unlink_list);
10505 while (!list_empty(&unlink_list)) {
10506 struct net_device *dev = list_first_entry(&unlink_list,
10509 list_del_init(&dev->unlink_list);
10510 dev->nested_level = dev->lower_level - 1;
10514 /* Snapshot list, allow later requests */
10515 list_replace_init(&net_todo_list, &list);
10519 /* Wait for rcu callbacks to finish before next phase */
10520 if (!list_empty(&list))
10523 list_for_each_entry_safe(dev, tmp, &list, todo_list) {
10524 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) {
10525 netdev_WARN(dev, "run_todo but not unregistering\n");
10526 list_del(&dev->todo_list);
10530 write_lock(&dev_base_lock);
10531 dev->reg_state = NETREG_UNREGISTERED;
10532 write_unlock(&dev_base_lock);
10533 linkwatch_forget_dev(dev);
10536 while (!list_empty(&list)) {
10537 dev = netdev_wait_allrefs_any(&list);
10538 list_del(&dev->todo_list);
10541 BUG_ON(netdev_refcnt_read(dev) != 1);
10542 BUG_ON(!list_empty(&dev->ptype_all));
10543 BUG_ON(!list_empty(&dev->ptype_specific));
10544 WARN_ON(rcu_access_pointer(dev->ip_ptr));
10545 WARN_ON(rcu_access_pointer(dev->ip6_ptr));
10547 netdev_do_free_pcpu_stats(dev);
10548 if (dev->priv_destructor)
10549 dev->priv_destructor(dev);
10550 if (dev->needs_free_netdev)
10553 if (atomic_dec_and_test(&dev_net(dev)->dev_unreg_count))
10554 wake_up(&netdev_unregistering_wq);
10556 /* Free network device */
10557 kobject_put(&dev->dev.kobj);
10561 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has
10562 * all the same fields in the same order as net_device_stats, with only
10563 * the type differing, but rtnl_link_stats64 may have additional fields
10564 * at the end for newer counters.
10566 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
10567 const struct net_device_stats *netdev_stats)
10569 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t);
10570 const atomic_long_t *src = (atomic_long_t *)netdev_stats;
10571 u64 *dst = (u64 *)stats64;
10573 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64));
10574 for (i = 0; i < n; i++)
10575 dst[i] = (unsigned long)atomic_long_read(&src[i]);
10576 /* zero out counters that only exist in rtnl_link_stats64 */
10577 memset((char *)stats64 + n * sizeof(u64), 0,
10578 sizeof(*stats64) - n * sizeof(u64));
10580 EXPORT_SYMBOL(netdev_stats_to_stats64);
10582 struct net_device_core_stats __percpu *netdev_core_stats_alloc(struct net_device *dev)
10584 struct net_device_core_stats __percpu *p;
10586 p = alloc_percpu_gfp(struct net_device_core_stats,
10587 GFP_ATOMIC | __GFP_NOWARN);
10589 if (p && cmpxchg(&dev->core_stats, NULL, p))
10592 /* This READ_ONCE() pairs with the cmpxchg() above */
10593 return READ_ONCE(dev->core_stats);
10595 EXPORT_SYMBOL(netdev_core_stats_alloc);
10598 * dev_get_stats - get network device statistics
10599 * @dev: device to get statistics from
10600 * @storage: place to store stats
10602 * Get network statistics from device. Return @storage.
10603 * The device driver may provide its own method by setting
10604 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats;
10605 * otherwise the internal statistics structure is used.
10607 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev,
10608 struct rtnl_link_stats64 *storage)
10610 const struct net_device_ops *ops = dev->netdev_ops;
10611 const struct net_device_core_stats __percpu *p;
10613 if (ops->ndo_get_stats64) {
10614 memset(storage, 0, sizeof(*storage));
10615 ops->ndo_get_stats64(dev, storage);
10616 } else if (ops->ndo_get_stats) {
10617 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev));
10619 netdev_stats_to_stats64(storage, &dev->stats);
10622 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */
10623 p = READ_ONCE(dev->core_stats);
10625 const struct net_device_core_stats *core_stats;
10628 for_each_possible_cpu(i) {
10629 core_stats = per_cpu_ptr(p, i);
10630 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped);
10631 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped);
10632 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler);
10633 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped);
10638 EXPORT_SYMBOL(dev_get_stats);
10641 * dev_fetch_sw_netstats - get per-cpu network device statistics
10642 * @s: place to store stats
10643 * @netstats: per-cpu network stats to read from
10645 * Read per-cpu network statistics and populate the related fields in @s.
10647 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s,
10648 const struct pcpu_sw_netstats __percpu *netstats)
10652 for_each_possible_cpu(cpu) {
10653 u64 rx_packets, rx_bytes, tx_packets, tx_bytes;
10654 const struct pcpu_sw_netstats *stats;
10655 unsigned int start;
10657 stats = per_cpu_ptr(netstats, cpu);
10659 start = u64_stats_fetch_begin(&stats->syncp);
10660 rx_packets = u64_stats_read(&stats->rx_packets);
10661 rx_bytes = u64_stats_read(&stats->rx_bytes);
10662 tx_packets = u64_stats_read(&stats->tx_packets);
10663 tx_bytes = u64_stats_read(&stats->tx_bytes);
10664 } while (u64_stats_fetch_retry(&stats->syncp, start));
10666 s->rx_packets += rx_packets;
10667 s->rx_bytes += rx_bytes;
10668 s->tx_packets += tx_packets;
10669 s->tx_bytes += tx_bytes;
10672 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats);
10675 * dev_get_tstats64 - ndo_get_stats64 implementation
10676 * @dev: device to get statistics from
10677 * @s: place to store stats
10679 * Populate @s from dev->stats and dev->tstats. Can be used as
10680 * ndo_get_stats64() callback.
10682 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s)
10684 netdev_stats_to_stats64(s, &dev->stats);
10685 dev_fetch_sw_netstats(s, dev->tstats);
10687 EXPORT_SYMBOL_GPL(dev_get_tstats64);
10689 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev)
10691 struct netdev_queue *queue = dev_ingress_queue(dev);
10693 #ifdef CONFIG_NET_CLS_ACT
10696 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
10699 netdev_init_one_queue(dev, queue, NULL);
10700 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc);
10701 RCU_INIT_POINTER(queue->qdisc_sleeping, &noop_qdisc);
10702 rcu_assign_pointer(dev->ingress_queue, queue);
10707 static const struct ethtool_ops default_ethtool_ops;
10709 void netdev_set_default_ethtool_ops(struct net_device *dev,
10710 const struct ethtool_ops *ops)
10712 if (dev->ethtool_ops == &default_ethtool_ops)
10713 dev->ethtool_ops = ops;
10715 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops);
10718 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default
10719 * @dev: netdev to enable the IRQ coalescing on
10721 * Sets a conservative default for SW IRQ coalescing. Users can use
10722 * sysfs attributes to override the default values.
10724 void netdev_sw_irq_coalesce_default_on(struct net_device *dev)
10726 WARN_ON(dev->reg_state == NETREG_REGISTERED);
10728 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
10729 dev->gro_flush_timeout = 20000;
10730 dev->napi_defer_hard_irqs = 1;
10733 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on);
10735 void netdev_freemem(struct net_device *dev)
10737 char *addr = (char *)dev - dev->padded;
10743 * alloc_netdev_mqs - allocate network device
10744 * @sizeof_priv: size of private data to allocate space for
10745 * @name: device name format string
10746 * @name_assign_type: origin of device name
10747 * @setup: callback to initialize device
10748 * @txqs: the number of TX subqueues to allocate
10749 * @rxqs: the number of RX subqueues to allocate
10751 * Allocates a struct net_device with private data area for driver use
10752 * and performs basic initialization. Also allocates subqueue structs
10753 * for each queue on the device.
10755 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
10756 unsigned char name_assign_type,
10757 void (*setup)(struct net_device *),
10758 unsigned int txqs, unsigned int rxqs)
10760 struct net_device *dev;
10761 unsigned int alloc_size;
10762 struct net_device *p;
10764 BUG_ON(strlen(name) >= sizeof(dev->name));
10767 pr_err("alloc_netdev: Unable to allocate device with zero queues\n");
10772 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n");
10776 alloc_size = sizeof(struct net_device);
10778 /* ensure 32-byte alignment of private area */
10779 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
10780 alloc_size += sizeof_priv;
10782 /* ensure 32-byte alignment of whole construct */
10783 alloc_size += NETDEV_ALIGN - 1;
10785 p = kvzalloc(alloc_size, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL);
10789 dev = PTR_ALIGN(p, NETDEV_ALIGN);
10790 dev->padded = (char *)dev - (char *)p;
10792 ref_tracker_dir_init(&dev->refcnt_tracker, 128, name);
10793 #ifdef CONFIG_PCPU_DEV_REFCNT
10794 dev->pcpu_refcnt = alloc_percpu(int);
10795 if (!dev->pcpu_refcnt)
10799 refcount_set(&dev->dev_refcnt, 1);
10802 if (dev_addr_init(dev))
10808 dev_net_set(dev, &init_net);
10810 dev->gso_max_size = GSO_LEGACY_MAX_SIZE;
10811 dev->xdp_zc_max_segs = 1;
10812 dev->gso_max_segs = GSO_MAX_SEGS;
10813 dev->gro_max_size = GRO_LEGACY_MAX_SIZE;
10814 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE;
10815 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE;
10816 dev->tso_max_size = TSO_LEGACY_MAX_SIZE;
10817 dev->tso_max_segs = TSO_MAX_SEGS;
10818 dev->upper_level = 1;
10819 dev->lower_level = 1;
10820 #ifdef CONFIG_LOCKDEP
10821 dev->nested_level = 0;
10822 INIT_LIST_HEAD(&dev->unlink_list);
10825 INIT_LIST_HEAD(&dev->napi_list);
10826 INIT_LIST_HEAD(&dev->unreg_list);
10827 INIT_LIST_HEAD(&dev->close_list);
10828 INIT_LIST_HEAD(&dev->link_watch_list);
10829 INIT_LIST_HEAD(&dev->adj_list.upper);
10830 INIT_LIST_HEAD(&dev->adj_list.lower);
10831 INIT_LIST_HEAD(&dev->ptype_all);
10832 INIT_LIST_HEAD(&dev->ptype_specific);
10833 INIT_LIST_HEAD(&dev->net_notifier_list);
10834 #ifdef CONFIG_NET_SCHED
10835 hash_init(dev->qdisc_hash);
10837 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM;
10840 if (!dev->tx_queue_len) {
10841 dev->priv_flags |= IFF_NO_QUEUE;
10842 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN;
10845 dev->num_tx_queues = txqs;
10846 dev->real_num_tx_queues = txqs;
10847 if (netif_alloc_netdev_queues(dev))
10850 dev->num_rx_queues = rxqs;
10851 dev->real_num_rx_queues = rxqs;
10852 if (netif_alloc_rx_queues(dev))
10855 strcpy(dev->name, name);
10856 dev->name_assign_type = name_assign_type;
10857 dev->group = INIT_NETDEV_GROUP;
10858 if (!dev->ethtool_ops)
10859 dev->ethtool_ops = &default_ethtool_ops;
10861 nf_hook_netdev_init(dev);
10870 #ifdef CONFIG_PCPU_DEV_REFCNT
10871 free_percpu(dev->pcpu_refcnt);
10874 netdev_freemem(dev);
10877 EXPORT_SYMBOL(alloc_netdev_mqs);
10880 * free_netdev - free network device
10883 * This function does the last stage of destroying an allocated device
10884 * interface. The reference to the device object is released. If this
10885 * is the last reference then it will be freed.Must be called in process
10888 void free_netdev(struct net_device *dev)
10890 struct napi_struct *p, *n;
10894 /* When called immediately after register_netdevice() failed the unwind
10895 * handling may still be dismantling the device. Handle that case by
10896 * deferring the free.
10898 if (dev->reg_state == NETREG_UNREGISTERING) {
10900 dev->needs_free_netdev = true;
10904 netif_free_tx_queues(dev);
10905 netif_free_rx_queues(dev);
10907 kfree(rcu_dereference_protected(dev->ingress_queue, 1));
10909 /* Flush device addresses */
10910 dev_addr_flush(dev);
10912 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
10915 ref_tracker_dir_exit(&dev->refcnt_tracker);
10916 #ifdef CONFIG_PCPU_DEV_REFCNT
10917 free_percpu(dev->pcpu_refcnt);
10918 dev->pcpu_refcnt = NULL;
10920 free_percpu(dev->core_stats);
10921 dev->core_stats = NULL;
10922 free_percpu(dev->xdp_bulkq);
10923 dev->xdp_bulkq = NULL;
10925 /* Compatibility with error handling in drivers */
10926 if (dev->reg_state == NETREG_UNINITIALIZED) {
10927 netdev_freemem(dev);
10931 BUG_ON(dev->reg_state != NETREG_UNREGISTERED);
10932 dev->reg_state = NETREG_RELEASED;
10934 /* will free via device release */
10935 put_device(&dev->dev);
10937 EXPORT_SYMBOL(free_netdev);
10940 * synchronize_net - Synchronize with packet receive processing
10942 * Wait for packets currently being received to be done.
10943 * Does not block later packets from starting.
10945 void synchronize_net(void)
10948 if (rtnl_is_locked())
10949 synchronize_rcu_expedited();
10953 EXPORT_SYMBOL(synchronize_net);
10956 * unregister_netdevice_queue - remove device from the kernel
10960 * This function shuts down a device interface and removes it
10961 * from the kernel tables.
10962 * If head not NULL, device is queued to be unregistered later.
10964 * Callers must hold the rtnl semaphore. You may want
10965 * unregister_netdev() instead of this.
10968 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head)
10973 list_move_tail(&dev->unreg_list, head);
10977 list_add(&dev->unreg_list, &single);
10978 unregister_netdevice_many(&single);
10981 EXPORT_SYMBOL(unregister_netdevice_queue);
10983 void unregister_netdevice_many_notify(struct list_head *head,
10984 u32 portid, const struct nlmsghdr *nlh)
10986 struct net_device *dev, *tmp;
10987 LIST_HEAD(close_head);
10989 BUG_ON(dev_boot_phase);
10992 if (list_empty(head))
10995 list_for_each_entry_safe(dev, tmp, head, unreg_list) {
10996 /* Some devices call without registering
10997 * for initialization unwind. Remove those
10998 * devices and proceed with the remaining.
11000 if (dev->reg_state == NETREG_UNINITIALIZED) {
11001 pr_debug("unregister_netdevice: device %s/%p never was registered\n",
11005 list_del(&dev->unreg_list);
11008 dev->dismantle = true;
11009 BUG_ON(dev->reg_state != NETREG_REGISTERED);
11012 /* If device is running, close it first. */
11013 list_for_each_entry(dev, head, unreg_list)
11014 list_add_tail(&dev->close_list, &close_head);
11015 dev_close_many(&close_head, true);
11017 list_for_each_entry(dev, head, unreg_list) {
11018 /* And unlink it from device chain. */
11019 write_lock(&dev_base_lock);
11020 unlist_netdevice(dev, false);
11021 dev->reg_state = NETREG_UNREGISTERING;
11022 write_unlock(&dev_base_lock);
11024 flush_all_backlogs();
11028 list_for_each_entry(dev, head, unreg_list) {
11029 struct sk_buff *skb = NULL;
11031 /* Shutdown queueing discipline. */
11033 dev_tcx_uninstall(dev);
11034 dev_xdp_uninstall(dev);
11035 bpf_dev_bound_netdev_unregister(dev);
11037 netdev_offload_xstats_disable_all(dev);
11039 /* Notify protocols, that we are about to destroy
11040 * this device. They should clean all the things.
11042 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
11044 if (!dev->rtnl_link_ops ||
11045 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
11046 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0,
11047 GFP_KERNEL, NULL, 0,
11051 * Flush the unicast and multicast chains
11056 netdev_name_node_alt_flush(dev);
11057 netdev_name_node_free(dev->name_node);
11059 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev);
11061 if (dev->netdev_ops->ndo_uninit)
11062 dev->netdev_ops->ndo_uninit(dev);
11065 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh);
11067 /* Notifier chain MUST detach us all upper devices. */
11068 WARN_ON(netdev_has_any_upper_dev(dev));
11069 WARN_ON(netdev_has_any_lower_dev(dev));
11071 /* Remove entries from kobject tree */
11072 netdev_unregister_kobject(dev);
11074 /* Remove XPS queueing entries */
11075 netif_reset_xps_queues_gt(dev, 0);
11081 list_for_each_entry(dev, head, unreg_list) {
11082 netdev_put(dev, &dev->dev_registered_tracker);
11090 * unregister_netdevice_many - unregister many devices
11091 * @head: list of devices
11093 * Note: As most callers use a stack allocated list_head,
11094 * we force a list_del() to make sure stack wont be corrupted later.
11096 void unregister_netdevice_many(struct list_head *head)
11098 unregister_netdevice_many_notify(head, 0, NULL);
11100 EXPORT_SYMBOL(unregister_netdevice_many);
11103 * unregister_netdev - remove device from the kernel
11106 * This function shuts down a device interface and removes it
11107 * from the kernel tables.
11109 * This is just a wrapper for unregister_netdevice that takes
11110 * the rtnl semaphore. In general you want to use this and not
11111 * unregister_netdevice.
11113 void unregister_netdev(struct net_device *dev)
11116 unregister_netdevice(dev);
11119 EXPORT_SYMBOL(unregister_netdev);
11122 * __dev_change_net_namespace - move device to different nethost namespace
11124 * @net: network namespace
11125 * @pat: If not NULL name pattern to try if the current device name
11126 * is already taken in the destination network namespace.
11127 * @new_ifindex: If not zero, specifies device index in the target
11130 * This function shuts down a device interface and moves it
11131 * to a new network namespace. On success 0 is returned, on
11132 * a failure a netagive errno code is returned.
11134 * Callers must hold the rtnl semaphore.
11137 int __dev_change_net_namespace(struct net_device *dev, struct net *net,
11138 const char *pat, int new_ifindex)
11140 struct netdev_name_node *name_node;
11141 struct net *net_old = dev_net(dev);
11142 char new_name[IFNAMSIZ] = {};
11147 /* Don't allow namespace local devices to be moved. */
11149 if (dev->features & NETIF_F_NETNS_LOCAL)
11152 /* Ensure the device has been registrered */
11153 if (dev->reg_state != NETREG_REGISTERED)
11156 /* Get out if there is nothing todo */
11158 if (net_eq(net_old, net))
11161 /* Pick the destination device name, and ensure
11162 * we can use it in the destination network namespace.
11165 if (netdev_name_in_use(net, dev->name)) {
11166 /* We get here if we can't use the current device name */
11169 err = dev_prep_valid_name(net, dev, pat, new_name);
11173 /* Check that none of the altnames conflicts. */
11175 netdev_for_each_altname(dev, name_node)
11176 if (netdev_name_in_use(net, name_node->name))
11179 /* Check that new_ifindex isn't used yet. */
11181 err = dev_index_reserve(net, new_ifindex);
11185 /* If there is an ifindex conflict assign a new one */
11186 err = dev_index_reserve(net, dev->ifindex);
11188 err = dev_index_reserve(net, 0);
11195 * And now a mini version of register_netdevice unregister_netdevice.
11198 /* If device is running close it first. */
11201 /* And unlink it from device chain */
11202 unlist_netdevice(dev, true);
11206 /* Shutdown queueing discipline. */
11209 /* Notify protocols, that we are about to destroy
11210 * this device. They should clean all the things.
11212 * Note that dev->reg_state stays at NETREG_REGISTERED.
11213 * This is wanted because this way 8021q and macvlan know
11214 * the device is just moving and can keep their slaves up.
11216 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
11219 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL);
11221 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid,
11225 * Flush the unicast and multicast chains
11230 /* Send a netdev-removed uevent to the old namespace */
11231 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE);
11232 netdev_adjacent_del_links(dev);
11234 /* Move per-net netdevice notifiers that are following the netdevice */
11235 move_netdevice_notifiers_dev_net(dev, net);
11237 /* Actually switch the network namespace */
11238 dev_net_set(dev, net);
11239 dev->ifindex = new_ifindex;
11241 /* Send a netdev-add uevent to the new namespace */
11242 kobject_uevent(&dev->dev.kobj, KOBJ_ADD);
11243 netdev_adjacent_add_links(dev);
11245 if (new_name[0]) /* Rename the netdev to prepared name */
11246 strscpy(dev->name, new_name, IFNAMSIZ);
11248 /* Fixup kobjects */
11249 err = device_rename(&dev->dev, dev->name);
11252 /* Adapt owner in case owning user namespace of target network
11253 * namespace is different from the original one.
11255 err = netdev_change_owner(dev, net_old, net);
11258 /* Add the device back in the hashes */
11259 list_netdevice(dev);
11261 /* Notify protocols, that a new device appeared. */
11262 call_netdevice_notifiers(NETDEV_REGISTER, dev);
11265 * Prevent userspace races by waiting until the network
11266 * device is fully setup before sending notifications.
11268 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL);
11275 EXPORT_SYMBOL_GPL(__dev_change_net_namespace);
11277 static int dev_cpu_dead(unsigned int oldcpu)
11279 struct sk_buff **list_skb;
11280 struct sk_buff *skb;
11282 struct softnet_data *sd, *oldsd, *remsd = NULL;
11284 local_irq_disable();
11285 cpu = smp_processor_id();
11286 sd = &per_cpu(softnet_data, cpu);
11287 oldsd = &per_cpu(softnet_data, oldcpu);
11289 /* Find end of our completion_queue. */
11290 list_skb = &sd->completion_queue;
11292 list_skb = &(*list_skb)->next;
11293 /* Append completion queue from offline CPU. */
11294 *list_skb = oldsd->completion_queue;
11295 oldsd->completion_queue = NULL;
11297 /* Append output queue from offline CPU. */
11298 if (oldsd->output_queue) {
11299 *sd->output_queue_tailp = oldsd->output_queue;
11300 sd->output_queue_tailp = oldsd->output_queue_tailp;
11301 oldsd->output_queue = NULL;
11302 oldsd->output_queue_tailp = &oldsd->output_queue;
11304 /* Append NAPI poll list from offline CPU, with one exception :
11305 * process_backlog() must be called by cpu owning percpu backlog.
11306 * We properly handle process_queue & input_pkt_queue later.
11308 while (!list_empty(&oldsd->poll_list)) {
11309 struct napi_struct *napi = list_first_entry(&oldsd->poll_list,
11310 struct napi_struct,
11313 list_del_init(&napi->poll_list);
11314 if (napi->poll == process_backlog)
11317 ____napi_schedule(sd, napi);
11320 raise_softirq_irqoff(NET_TX_SOFTIRQ);
11321 local_irq_enable();
11324 remsd = oldsd->rps_ipi_list;
11325 oldsd->rps_ipi_list = NULL;
11327 /* send out pending IPI's on offline CPU */
11328 net_rps_send_ipi(remsd);
11330 /* Process offline CPU's input_pkt_queue */
11331 while ((skb = __skb_dequeue(&oldsd->process_queue))) {
11333 input_queue_head_incr(oldsd);
11335 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) {
11337 input_queue_head_incr(oldsd);
11344 * netdev_increment_features - increment feature set by one
11345 * @all: current feature set
11346 * @one: new feature set
11347 * @mask: mask feature set
11349 * Computes a new feature set after adding a device with feature set
11350 * @one to the master device with current feature set @all. Will not
11351 * enable anything that is off in @mask. Returns the new feature set.
11353 netdev_features_t netdev_increment_features(netdev_features_t all,
11354 netdev_features_t one, netdev_features_t mask)
11356 if (mask & NETIF_F_HW_CSUM)
11357 mask |= NETIF_F_CSUM_MASK;
11358 mask |= NETIF_F_VLAN_CHALLENGED;
11360 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask;
11361 all &= one | ~NETIF_F_ALL_FOR_ALL;
11363 /* If one device supports hw checksumming, set for all. */
11364 if (all & NETIF_F_HW_CSUM)
11365 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM);
11369 EXPORT_SYMBOL(netdev_increment_features);
11371 static struct hlist_head * __net_init netdev_create_hash(void)
11374 struct hlist_head *hash;
11376 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL);
11378 for (i = 0; i < NETDEV_HASHENTRIES; i++)
11379 INIT_HLIST_HEAD(&hash[i]);
11384 /* Initialize per network namespace state */
11385 static int __net_init netdev_init(struct net *net)
11387 BUILD_BUG_ON(GRO_HASH_BUCKETS >
11388 8 * sizeof_field(struct napi_struct, gro_bitmask));
11390 INIT_LIST_HEAD(&net->dev_base_head);
11392 net->dev_name_head = netdev_create_hash();
11393 if (net->dev_name_head == NULL)
11396 net->dev_index_head = netdev_create_hash();
11397 if (net->dev_index_head == NULL)
11400 xa_init_flags(&net->dev_by_index, XA_FLAGS_ALLOC1);
11402 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain);
11407 kfree(net->dev_name_head);
11413 * netdev_drivername - network driver for the device
11414 * @dev: network device
11416 * Determine network driver for device.
11418 const char *netdev_drivername(const struct net_device *dev)
11420 const struct device_driver *driver;
11421 const struct device *parent;
11422 const char *empty = "";
11424 parent = dev->dev.parent;
11428 driver = parent->driver;
11429 if (driver && driver->name)
11430 return driver->name;
11434 static void __netdev_printk(const char *level, const struct net_device *dev,
11435 struct va_format *vaf)
11437 if (dev && dev->dev.parent) {
11438 dev_printk_emit(level[1] - '0',
11441 dev_driver_string(dev->dev.parent),
11442 dev_name(dev->dev.parent),
11443 netdev_name(dev), netdev_reg_state(dev),
11446 printk("%s%s%s: %pV",
11447 level, netdev_name(dev), netdev_reg_state(dev), vaf);
11449 printk("%s(NULL net_device): %pV", level, vaf);
11453 void netdev_printk(const char *level, const struct net_device *dev,
11454 const char *format, ...)
11456 struct va_format vaf;
11459 va_start(args, format);
11464 __netdev_printk(level, dev, &vaf);
11468 EXPORT_SYMBOL(netdev_printk);
11470 #define define_netdev_printk_level(func, level) \
11471 void func(const struct net_device *dev, const char *fmt, ...) \
11473 struct va_format vaf; \
11476 va_start(args, fmt); \
11481 __netdev_printk(level, dev, &vaf); \
11485 EXPORT_SYMBOL(func);
11487 define_netdev_printk_level(netdev_emerg, KERN_EMERG);
11488 define_netdev_printk_level(netdev_alert, KERN_ALERT);
11489 define_netdev_printk_level(netdev_crit, KERN_CRIT);
11490 define_netdev_printk_level(netdev_err, KERN_ERR);
11491 define_netdev_printk_level(netdev_warn, KERN_WARNING);
11492 define_netdev_printk_level(netdev_notice, KERN_NOTICE);
11493 define_netdev_printk_level(netdev_info, KERN_INFO);
11495 static void __net_exit netdev_exit(struct net *net)
11497 kfree(net->dev_name_head);
11498 kfree(net->dev_index_head);
11499 xa_destroy(&net->dev_by_index);
11500 if (net != &init_net)
11501 WARN_ON_ONCE(!list_empty(&net->dev_base_head));
11504 static struct pernet_operations __net_initdata netdev_net_ops = {
11505 .init = netdev_init,
11506 .exit = netdev_exit,
11509 static void __net_exit default_device_exit_net(struct net *net)
11511 struct netdev_name_node *name_node, *tmp;
11512 struct net_device *dev, *aux;
11514 * Push all migratable network devices back to the
11515 * initial network namespace
11518 for_each_netdev_safe(net, dev, aux) {
11520 char fb_name[IFNAMSIZ];
11522 /* Ignore unmoveable devices (i.e. loopback) */
11523 if (dev->features & NETIF_F_NETNS_LOCAL)
11526 /* Leave virtual devices for the generic cleanup */
11527 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund)
11530 /* Push remaining network devices to init_net */
11531 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex);
11532 if (netdev_name_in_use(&init_net, fb_name))
11533 snprintf(fb_name, IFNAMSIZ, "dev%%d");
11535 netdev_for_each_altname_safe(dev, name_node, tmp)
11536 if (netdev_name_in_use(&init_net, name_node->name)) {
11537 netdev_name_node_del(name_node);
11539 __netdev_name_node_alt_destroy(name_node);
11542 err = dev_change_net_namespace(dev, &init_net, fb_name);
11544 pr_emerg("%s: failed to move %s to init_net: %d\n",
11545 __func__, dev->name, err);
11551 static void __net_exit default_device_exit_batch(struct list_head *net_list)
11553 /* At exit all network devices most be removed from a network
11554 * namespace. Do this in the reverse order of registration.
11555 * Do this across as many network namespaces as possible to
11556 * improve batching efficiency.
11558 struct net_device *dev;
11560 LIST_HEAD(dev_kill_list);
11563 list_for_each_entry(net, net_list, exit_list) {
11564 default_device_exit_net(net);
11568 list_for_each_entry(net, net_list, exit_list) {
11569 for_each_netdev_reverse(net, dev) {
11570 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink)
11571 dev->rtnl_link_ops->dellink(dev, &dev_kill_list);
11573 unregister_netdevice_queue(dev, &dev_kill_list);
11576 unregister_netdevice_many(&dev_kill_list);
11580 static struct pernet_operations __net_initdata default_device_ops = {
11581 .exit_batch = default_device_exit_batch,
11585 * Initialize the DEV module. At boot time this walks the device list and
11586 * unhooks any devices that fail to initialise (normally hardware not
11587 * present) and leaves us with a valid list of present and active devices.
11592 * This is called single threaded during boot, so no need
11593 * to take the rtnl semaphore.
11595 static int __init net_dev_init(void)
11597 int i, rc = -ENOMEM;
11599 BUG_ON(!dev_boot_phase);
11601 if (dev_proc_init())
11604 if (netdev_kobject_init())
11607 INIT_LIST_HEAD(&ptype_all);
11608 for (i = 0; i < PTYPE_HASH_SIZE; i++)
11609 INIT_LIST_HEAD(&ptype_base[i]);
11611 if (register_pernet_subsys(&netdev_net_ops))
11615 * Initialise the packet receive queues.
11618 for_each_possible_cpu(i) {
11619 struct work_struct *flush = per_cpu_ptr(&flush_works, i);
11620 struct softnet_data *sd = &per_cpu(softnet_data, i);
11622 INIT_WORK(flush, flush_backlog);
11624 skb_queue_head_init(&sd->input_pkt_queue);
11625 skb_queue_head_init(&sd->process_queue);
11626 #ifdef CONFIG_XFRM_OFFLOAD
11627 skb_queue_head_init(&sd->xfrm_backlog);
11629 INIT_LIST_HEAD(&sd->poll_list);
11630 sd->output_queue_tailp = &sd->output_queue;
11632 INIT_CSD(&sd->csd, rps_trigger_softirq, sd);
11635 #ifndef CONFIG_PREEMPT_RT
11636 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd);
11638 INIT_WORK(&sd->defer_work, trigger_rx_softirq);
11640 spin_lock_init(&sd->defer_lock);
11642 init_gro_hash(&sd->backlog);
11643 sd->backlog.poll = process_backlog;
11644 sd->backlog.weight = weight_p;
11647 dev_boot_phase = 0;
11649 /* The loopback device is special if any other network devices
11650 * is present in a network namespace the loopback device must
11651 * be present. Since we now dynamically allocate and free the
11652 * loopback device ensure this invariant is maintained by
11653 * keeping the loopback device as the first device on the
11654 * list of network devices. Ensuring the loopback devices
11655 * is the first device that appears and the last network device
11658 if (register_pernet_device(&loopback_net_ops))
11661 if (register_pernet_device(&default_device_ops))
11664 open_softirq(NET_TX_SOFTIRQ, net_tx_action);
11665 open_softirq(NET_RX_SOFTIRQ, net_rx_action);
11667 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead",
11668 NULL, dev_cpu_dead);
11675 subsys_initcall(net_dev_init);