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/bitops.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>
110 #include <linux/highmem.h>
111 #include <linux/init.h>
112 #include <linux/module.h>
113 #include <linux/netpoll.h>
114 #include <linux/rcupdate.h>
115 #include <linux/delay.h>
116 #include <net/iw_handler.h>
117 #include <asm/current.h>
118 #include <linux/audit.h>
119 #include <linux/dmaengine.h>
120 #include <linux/err.h>
121 #include <linux/ctype.h>
122 #include <linux/if_arp.h>
123 #include <linux/if_vlan.h>
124 #include <linux/ip.h>
126 #include <net/mpls.h>
127 #include <linux/ipv6.h>
128 #include <linux/in.h>
129 #include <linux/jhash.h>
130 #include <linux/random.h>
131 #include <trace/events/napi.h>
132 #include <trace/events/net.h>
133 #include <trace/events/skb.h>
134 #include <trace/events/qdisc.h>
135 #include <linux/inetdevice.h>
136 #include <linux/cpu_rmap.h>
137 #include <linux/static_key.h>
138 #include <linux/hashtable.h>
139 #include <linux/vmalloc.h>
140 #include <linux/if_macvlan.h>
141 #include <linux/errqueue.h>
142 #include <linux/hrtimer.h>
143 #include <linux/netfilter_netdev.h>
144 #include <linux/crash_dump.h>
145 #include <linux/sctp.h>
146 #include <net/udp_tunnel.h>
147 #include <linux/net_namespace.h>
148 #include <linux/indirect_call_wrapper.h>
149 #include <net/devlink.h>
150 #include <linux/pm_runtime.h>
151 #include <linux/prandom.h>
152 #include <linux/once_lite.h>
155 #include "net-sysfs.h"
158 static DEFINE_SPINLOCK(ptype_lock);
159 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly;
160 struct list_head ptype_all __read_mostly; /* Taps */
162 static int netif_rx_internal(struct sk_buff *skb);
163 static int call_netdevice_notifiers_info(unsigned long val,
164 struct netdev_notifier_info *info);
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 netdev_name_node_del(name_node);
349 kfree(name_node->name);
350 netdev_name_node_free(name_node);
353 int netdev_name_node_alt_destroy(struct net_device *dev, const char *name)
355 struct netdev_name_node *name_node;
356 struct net *net = dev_net(dev);
358 name_node = netdev_name_node_lookup(net, name);
361 /* lookup might have found our primary name or a name belonging
364 if (name_node == dev->name_node || name_node->dev != dev)
367 __netdev_name_node_alt_destroy(name_node);
372 static void netdev_name_node_alt_flush(struct net_device *dev)
374 struct netdev_name_node *name_node, *tmp;
376 list_for_each_entry_safe(name_node, tmp, &dev->name_node->list, list)
377 __netdev_name_node_alt_destroy(name_node);
380 /* Device list insertion */
381 static void list_netdevice(struct net_device *dev)
383 struct net *net = dev_net(dev);
387 write_lock(&dev_base_lock);
388 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head);
389 netdev_name_node_add(net, dev->name_node);
390 hlist_add_head_rcu(&dev->index_hlist,
391 dev_index_hash(net, dev->ifindex));
392 write_unlock(&dev_base_lock);
394 dev_base_seq_inc(net);
397 /* Device list removal
398 * caller must respect a RCU grace period before freeing/reusing dev
400 static void unlist_netdevice(struct net_device *dev, bool lock)
404 /* Unlink dev from the device chain */
406 write_lock(&dev_base_lock);
407 list_del_rcu(&dev->dev_list);
408 netdev_name_node_del(dev->name_node);
409 hlist_del_rcu(&dev->index_hlist);
411 write_unlock(&dev_base_lock);
413 dev_base_seq_inc(dev_net(dev));
420 static RAW_NOTIFIER_HEAD(netdev_chain);
423 * Device drivers call our routines to queue packets here. We empty the
424 * queue in the local softnet handler.
427 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data);
428 EXPORT_PER_CPU_SYMBOL(softnet_data);
430 #ifdef CONFIG_LOCKDEP
432 * register_netdevice() inits txq->_xmit_lock and sets lockdep class
433 * according to dev->type
435 static const unsigned short netdev_lock_type[] = {
436 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25,
437 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET,
438 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM,
439 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP,
440 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD,
441 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25,
442 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP,
443 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD,
444 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI,
445 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE,
446 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET,
447 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL,
448 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM,
449 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE,
450 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE};
452 static const char *const netdev_lock_name[] = {
453 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25",
454 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET",
455 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM",
456 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP",
457 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD",
458 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25",
459 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP",
460 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD",
461 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI",
462 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE",
463 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET",
464 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL",
465 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM",
466 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE",
467 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"};
469 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)];
470 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)];
472 static inline unsigned short netdev_lock_pos(unsigned short dev_type)
476 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++)
477 if (netdev_lock_type[i] == dev_type)
479 /* the last key is used by default */
480 return ARRAY_SIZE(netdev_lock_type) - 1;
483 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
484 unsigned short dev_type)
488 i = netdev_lock_pos(dev_type);
489 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i],
490 netdev_lock_name[i]);
493 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
497 i = netdev_lock_pos(dev->type);
498 lockdep_set_class_and_name(&dev->addr_list_lock,
499 &netdev_addr_lock_key[i],
500 netdev_lock_name[i]);
503 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
504 unsigned short dev_type)
508 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
513 /*******************************************************************************
515 * Protocol management and registration routines
517 *******************************************************************************/
521 * Add a protocol ID to the list. Now that the input handler is
522 * smarter we can dispense with all the messy stuff that used to be
525 * BEWARE!!! Protocol handlers, mangling input packets,
526 * MUST BE last in hash buckets and checking protocol handlers
527 * MUST start from promiscuous ptype_all chain in net_bh.
528 * It is true now, do not change it.
529 * Explanation follows: if protocol handler, mangling packet, will
530 * be the first on list, it is not able to sense, that packet
531 * is cloned and should be copied-on-write, so that it will
532 * change it and subsequent readers will get broken packet.
536 static inline struct list_head *ptype_head(const struct packet_type *pt)
538 if (pt->type == htons(ETH_P_ALL))
539 return pt->dev ? &pt->dev->ptype_all : &ptype_all;
541 return pt->dev ? &pt->dev->ptype_specific :
542 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK];
546 * dev_add_pack - add packet handler
547 * @pt: packet type declaration
549 * Add a protocol handler to the networking stack. The passed &packet_type
550 * is linked into kernel lists and may not be freed until it has been
551 * removed from the kernel lists.
553 * This call does not sleep therefore it can not
554 * guarantee all CPU's that are in middle of receiving packets
555 * will see the new packet type (until the next received packet).
558 void dev_add_pack(struct packet_type *pt)
560 struct list_head *head = ptype_head(pt);
562 spin_lock(&ptype_lock);
563 list_add_rcu(&pt->list, head);
564 spin_unlock(&ptype_lock);
566 EXPORT_SYMBOL(dev_add_pack);
569 * __dev_remove_pack - remove packet handler
570 * @pt: packet type declaration
572 * Remove a protocol handler that was previously added to the kernel
573 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
574 * from the kernel lists and can be freed or reused once this function
577 * The packet type might still be in use by receivers
578 * and must not be freed until after all the CPU's have gone
579 * through a quiescent state.
581 void __dev_remove_pack(struct packet_type *pt)
583 struct list_head *head = ptype_head(pt);
584 struct packet_type *pt1;
586 spin_lock(&ptype_lock);
588 list_for_each_entry(pt1, head, list) {
590 list_del_rcu(&pt->list);
595 pr_warn("dev_remove_pack: %p not found\n", pt);
597 spin_unlock(&ptype_lock);
599 EXPORT_SYMBOL(__dev_remove_pack);
602 * dev_remove_pack - remove packet handler
603 * @pt: packet type declaration
605 * Remove a protocol handler that was previously added to the kernel
606 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
607 * from the kernel lists and can be freed or reused once this function
610 * This call sleeps to guarantee that no CPU is looking at the packet
613 void dev_remove_pack(struct packet_type *pt)
615 __dev_remove_pack(pt);
619 EXPORT_SYMBOL(dev_remove_pack);
622 /*******************************************************************************
624 * Device Interface Subroutines
626 *******************************************************************************/
629 * dev_get_iflink - get 'iflink' value of a interface
630 * @dev: targeted interface
632 * Indicates the ifindex the interface is linked to.
633 * Physical interfaces have the same 'ifindex' and 'iflink' values.
636 int dev_get_iflink(const struct net_device *dev)
638 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink)
639 return dev->netdev_ops->ndo_get_iflink(dev);
643 EXPORT_SYMBOL(dev_get_iflink);
646 * dev_fill_metadata_dst - Retrieve tunnel egress information.
647 * @dev: targeted interface
650 * For better visibility of tunnel traffic OVS needs to retrieve
651 * egress tunnel information for a packet. Following API allows
652 * user to get this info.
654 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb)
656 struct ip_tunnel_info *info;
658 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst)
661 info = skb_tunnel_info_unclone(skb);
664 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX)))
667 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb);
669 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst);
671 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack)
673 int k = stack->num_paths++;
675 if (WARN_ON_ONCE(k >= NET_DEVICE_PATH_STACK_MAX))
678 return &stack->path[k];
681 int dev_fill_forward_path(const struct net_device *dev, const u8 *daddr,
682 struct net_device_path_stack *stack)
684 const struct net_device *last_dev;
685 struct net_device_path_ctx ctx = {
688 struct net_device_path *path;
691 memcpy(ctx.daddr, daddr, sizeof(ctx.daddr));
692 stack->num_paths = 0;
693 while (ctx.dev && ctx.dev->netdev_ops->ndo_fill_forward_path) {
695 path = dev_fwd_path(stack);
699 memset(path, 0, sizeof(struct net_device_path));
700 ret = ctx.dev->netdev_ops->ndo_fill_forward_path(&ctx, path);
704 if (WARN_ON_ONCE(last_dev == ctx.dev))
711 path = dev_fwd_path(stack);
714 path->type = DEV_PATH_ETHERNET;
719 EXPORT_SYMBOL_GPL(dev_fill_forward_path);
722 * __dev_get_by_name - find a device by its name
723 * @net: the applicable net namespace
724 * @name: name to find
726 * Find an interface by name. Must be called under RTNL semaphore
727 * or @dev_base_lock. If the name is found a pointer to the device
728 * is returned. If the name is not found then %NULL is returned. The
729 * reference counters are not incremented so the caller must be
730 * careful with locks.
733 struct net_device *__dev_get_by_name(struct net *net, const char *name)
735 struct netdev_name_node *node_name;
737 node_name = netdev_name_node_lookup(net, name);
738 return node_name ? node_name->dev : NULL;
740 EXPORT_SYMBOL(__dev_get_by_name);
743 * dev_get_by_name_rcu - find a device by its name
744 * @net: the applicable net namespace
745 * @name: name to find
747 * Find an interface by name.
748 * If the name is found a pointer to the device is returned.
749 * If the name is not found then %NULL is returned.
750 * The reference counters are not incremented so the caller must be
751 * careful with locks. The caller must hold RCU lock.
754 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name)
756 struct netdev_name_node *node_name;
758 node_name = netdev_name_node_lookup_rcu(net, name);
759 return node_name ? node_name->dev : NULL;
761 EXPORT_SYMBOL(dev_get_by_name_rcu);
764 * dev_get_by_name - find a device by its name
765 * @net: the applicable net namespace
766 * @name: name to find
768 * Find an interface by name. This can be called from any
769 * context and does its own locking. The returned handle has
770 * the usage count incremented and the caller must use dev_put() to
771 * release it when it is no longer needed. %NULL is returned if no
772 * matching device is found.
775 struct net_device *dev_get_by_name(struct net *net, const char *name)
777 struct net_device *dev;
780 dev = dev_get_by_name_rcu(net, name);
785 EXPORT_SYMBOL(dev_get_by_name);
788 * __dev_get_by_index - find a device by its ifindex
789 * @net: the applicable net namespace
790 * @ifindex: index of device
792 * Search for an interface by index. Returns %NULL if the device
793 * is not found or a pointer to the device. The device has not
794 * had its reference counter increased so the caller must be careful
795 * about locking. The caller must hold either the RTNL semaphore
799 struct net_device *__dev_get_by_index(struct net *net, int ifindex)
801 struct net_device *dev;
802 struct hlist_head *head = dev_index_hash(net, ifindex);
804 hlist_for_each_entry(dev, head, index_hlist)
805 if (dev->ifindex == ifindex)
810 EXPORT_SYMBOL(__dev_get_by_index);
813 * dev_get_by_index_rcu - find a device by its ifindex
814 * @net: the applicable net namespace
815 * @ifindex: index of device
817 * Search for an interface by index. Returns %NULL if the device
818 * is not found or a pointer to the device. The device has not
819 * had its reference counter increased so the caller must be careful
820 * about locking. The caller must hold RCU lock.
823 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex)
825 struct net_device *dev;
826 struct hlist_head *head = dev_index_hash(net, ifindex);
828 hlist_for_each_entry_rcu(dev, head, index_hlist)
829 if (dev->ifindex == ifindex)
834 EXPORT_SYMBOL(dev_get_by_index_rcu);
838 * dev_get_by_index - find a device by its ifindex
839 * @net: the applicable net namespace
840 * @ifindex: index of device
842 * Search for an interface by index. Returns NULL if the device
843 * is not found or a pointer to the device. The device returned has
844 * had a reference added and the pointer is safe until the user calls
845 * dev_put to indicate they have finished with it.
848 struct net_device *dev_get_by_index(struct net *net, int ifindex)
850 struct net_device *dev;
853 dev = dev_get_by_index_rcu(net, ifindex);
858 EXPORT_SYMBOL(dev_get_by_index);
861 * dev_get_by_napi_id - find a device by napi_id
862 * @napi_id: ID of the NAPI struct
864 * Search for an interface by NAPI ID. Returns %NULL if the device
865 * is not found or a pointer to the device. The device has not had
866 * its reference counter increased so the caller must be careful
867 * about locking. The caller must hold RCU lock.
870 struct net_device *dev_get_by_napi_id(unsigned int napi_id)
872 struct napi_struct *napi;
874 WARN_ON_ONCE(!rcu_read_lock_held());
876 if (napi_id < MIN_NAPI_ID)
879 napi = napi_by_id(napi_id);
881 return napi ? napi->dev : NULL;
883 EXPORT_SYMBOL(dev_get_by_napi_id);
886 * netdev_get_name - get a netdevice name, knowing its ifindex.
887 * @net: network namespace
888 * @name: a pointer to the buffer where the name will be stored.
889 * @ifindex: the ifindex of the interface to get the name from.
891 int netdev_get_name(struct net *net, char *name, int ifindex)
893 struct net_device *dev;
896 down_read(&devnet_rename_sem);
899 dev = dev_get_by_index_rcu(net, ifindex);
905 strcpy(name, dev->name);
910 up_read(&devnet_rename_sem);
915 * dev_getbyhwaddr_rcu - find a device by its hardware address
916 * @net: the applicable net namespace
917 * @type: media type of device
918 * @ha: hardware address
920 * Search for an interface by MAC address. Returns NULL if the device
921 * is not found or a pointer to the device.
922 * The caller must hold RCU or RTNL.
923 * The returned device has not had its ref count increased
924 * and the caller must therefore be careful about locking
928 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type,
931 struct net_device *dev;
933 for_each_netdev_rcu(net, dev)
934 if (dev->type == type &&
935 !memcmp(dev->dev_addr, ha, dev->addr_len))
940 EXPORT_SYMBOL(dev_getbyhwaddr_rcu);
942 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type)
944 struct net_device *dev, *ret = NULL;
947 for_each_netdev_rcu(net, dev)
948 if (dev->type == type) {
956 EXPORT_SYMBOL(dev_getfirstbyhwtype);
959 * __dev_get_by_flags - find any device with given flags
960 * @net: the applicable net namespace
961 * @if_flags: IFF_* values
962 * @mask: bitmask of bits in if_flags to check
964 * Search for any interface with the given flags. Returns NULL if a device
965 * is not found or a pointer to the device. Must be called inside
966 * rtnl_lock(), and result refcount is unchanged.
969 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags,
972 struct net_device *dev, *ret;
977 for_each_netdev(net, dev) {
978 if (((dev->flags ^ if_flags) & mask) == 0) {
985 EXPORT_SYMBOL(__dev_get_by_flags);
988 * dev_valid_name - check if name is okay for network device
991 * Network device names need to be valid file names to
992 * allow sysfs to work. We also disallow any kind of
995 bool dev_valid_name(const char *name)
999 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ)
1001 if (!strcmp(name, ".") || !strcmp(name, ".."))
1005 if (*name == '/' || *name == ':' || isspace(*name))
1011 EXPORT_SYMBOL(dev_valid_name);
1014 * __dev_alloc_name - allocate a name for a device
1015 * @net: network namespace to allocate the device name in
1016 * @name: name format string
1017 * @buf: scratch buffer and result name string
1019 * Passed a format string - eg "lt%d" it will try and find a suitable
1020 * id. It scans list of devices to build up a free map, then chooses
1021 * the first empty slot. The caller must hold the dev_base or rtnl lock
1022 * while allocating the name and adding the device in order to avoid
1024 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1025 * Returns the number of the unit assigned or a negative errno code.
1028 static int __dev_alloc_name(struct net *net, const char *name, char *buf)
1032 const int max_netdevices = 8*PAGE_SIZE;
1033 unsigned long *inuse;
1034 struct net_device *d;
1036 if (!dev_valid_name(name))
1039 p = strchr(name, '%');
1042 * Verify the string as this thing may have come from
1043 * the user. There must be either one "%d" and no other "%"
1046 if (p[1] != 'd' || strchr(p + 2, '%'))
1049 /* Use one page as a bit array of possible slots */
1050 inuse = (unsigned long *) get_zeroed_page(GFP_ATOMIC);
1054 for_each_netdev(net, d) {
1055 struct netdev_name_node *name_node;
1056 list_for_each_entry(name_node, &d->name_node->list, list) {
1057 if (!sscanf(name_node->name, name, &i))
1059 if (i < 0 || i >= max_netdevices)
1062 /* avoid cases where sscanf is not exact inverse of printf */
1063 snprintf(buf, IFNAMSIZ, name, i);
1064 if (!strncmp(buf, name_node->name, IFNAMSIZ))
1065 __set_bit(i, inuse);
1067 if (!sscanf(d->name, name, &i))
1069 if (i < 0 || i >= max_netdevices)
1072 /* avoid cases where sscanf is not exact inverse of printf */
1073 snprintf(buf, IFNAMSIZ, name, i);
1074 if (!strncmp(buf, d->name, IFNAMSIZ))
1075 __set_bit(i, inuse);
1078 i = find_first_zero_bit(inuse, max_netdevices);
1079 free_page((unsigned long) inuse);
1082 snprintf(buf, IFNAMSIZ, name, i);
1083 if (!netdev_name_in_use(net, buf))
1086 /* It is possible to run out of possible slots
1087 * when the name is long and there isn't enough space left
1088 * for the digits, or if all bits are used.
1093 static int dev_alloc_name_ns(struct net *net,
1094 struct net_device *dev,
1101 ret = __dev_alloc_name(net, name, buf);
1103 strscpy(dev->name, buf, IFNAMSIZ);
1108 * dev_alloc_name - allocate a name for a device
1110 * @name: name format string
1112 * Passed a format string - eg "lt%d" it will try and find a suitable
1113 * id. It scans list of devices to build up a free map, then chooses
1114 * the first empty slot. The caller must hold the dev_base or rtnl lock
1115 * while allocating the name and adding the device in order to avoid
1117 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1118 * Returns the number of the unit assigned or a negative errno code.
1121 int dev_alloc_name(struct net_device *dev, const char *name)
1123 return dev_alloc_name_ns(dev_net(dev), dev, name);
1125 EXPORT_SYMBOL(dev_alloc_name);
1127 static int dev_get_valid_name(struct net *net, struct net_device *dev,
1132 if (!dev_valid_name(name))
1135 if (strchr(name, '%'))
1136 return dev_alloc_name_ns(net, dev, name);
1137 else if (netdev_name_in_use(net, name))
1139 else if (dev->name != name)
1140 strscpy(dev->name, name, IFNAMSIZ);
1146 * dev_change_name - change name of a device
1148 * @newname: name (or format string) must be at least IFNAMSIZ
1150 * Change name of a device, can pass format strings "eth%d".
1153 int dev_change_name(struct net_device *dev, const char *newname)
1155 unsigned char old_assign_type;
1156 char oldname[IFNAMSIZ];
1162 BUG_ON(!dev_net(dev));
1166 down_write(&devnet_rename_sem);
1168 if (strncmp(newname, dev->name, IFNAMSIZ) == 0) {
1169 up_write(&devnet_rename_sem);
1173 memcpy(oldname, dev->name, IFNAMSIZ);
1175 err = dev_get_valid_name(net, dev, newname);
1177 up_write(&devnet_rename_sem);
1181 if (oldname[0] && !strchr(oldname, '%'))
1182 netdev_info(dev, "renamed from %s%s\n", oldname,
1183 dev->flags & IFF_UP ? " (while UP)" : "");
1185 old_assign_type = dev->name_assign_type;
1186 dev->name_assign_type = NET_NAME_RENAMED;
1189 ret = device_rename(&dev->dev, dev->name);
1191 memcpy(dev->name, oldname, IFNAMSIZ);
1192 dev->name_assign_type = old_assign_type;
1193 up_write(&devnet_rename_sem);
1197 up_write(&devnet_rename_sem);
1199 netdev_adjacent_rename_links(dev, oldname);
1201 write_lock(&dev_base_lock);
1202 netdev_name_node_del(dev->name_node);
1203 write_unlock(&dev_base_lock);
1207 write_lock(&dev_base_lock);
1208 netdev_name_node_add(net, dev->name_node);
1209 write_unlock(&dev_base_lock);
1211 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev);
1212 ret = notifier_to_errno(ret);
1215 /* err >= 0 after dev_alloc_name() or stores the first errno */
1218 down_write(&devnet_rename_sem);
1219 memcpy(dev->name, oldname, IFNAMSIZ);
1220 memcpy(oldname, newname, IFNAMSIZ);
1221 dev->name_assign_type = old_assign_type;
1222 old_assign_type = NET_NAME_RENAMED;
1225 netdev_err(dev, "name change rollback failed: %d\n",
1234 * dev_set_alias - change ifalias of a device
1236 * @alias: name up to IFALIASZ
1237 * @len: limit of bytes to copy from info
1239 * Set ifalias for a device,
1241 int dev_set_alias(struct net_device *dev, const char *alias, size_t len)
1243 struct dev_ifalias *new_alias = NULL;
1245 if (len >= IFALIASZ)
1249 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL);
1253 memcpy(new_alias->ifalias, alias, len);
1254 new_alias->ifalias[len] = 0;
1257 mutex_lock(&ifalias_mutex);
1258 new_alias = rcu_replace_pointer(dev->ifalias, new_alias,
1259 mutex_is_locked(&ifalias_mutex));
1260 mutex_unlock(&ifalias_mutex);
1263 kfree_rcu(new_alias, rcuhead);
1267 EXPORT_SYMBOL(dev_set_alias);
1270 * dev_get_alias - get ifalias of a device
1272 * @name: buffer to store name of ifalias
1273 * @len: size of buffer
1275 * get ifalias for a device. Caller must make sure dev cannot go
1276 * away, e.g. rcu read lock or own a reference count to device.
1278 int dev_get_alias(const struct net_device *dev, char *name, size_t len)
1280 const struct dev_ifalias *alias;
1284 alias = rcu_dereference(dev->ifalias);
1286 ret = snprintf(name, len, "%s", alias->ifalias);
1293 * netdev_features_change - device changes features
1294 * @dev: device to cause notification
1296 * Called to indicate a device has changed features.
1298 void netdev_features_change(struct net_device *dev)
1300 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev);
1302 EXPORT_SYMBOL(netdev_features_change);
1305 * netdev_state_change - device changes state
1306 * @dev: device to cause notification
1308 * Called to indicate a device has changed state. This function calls
1309 * the notifier chains for netdev_chain and sends a NEWLINK message
1310 * to the routing socket.
1312 void netdev_state_change(struct net_device *dev)
1314 if (dev->flags & IFF_UP) {
1315 struct netdev_notifier_change_info change_info = {
1319 call_netdevice_notifiers_info(NETDEV_CHANGE,
1321 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL, 0, NULL);
1324 EXPORT_SYMBOL(netdev_state_change);
1327 * __netdev_notify_peers - notify network peers about existence of @dev,
1328 * to be called when rtnl lock is already held.
1329 * @dev: network device
1331 * Generate traffic such that interested network peers are aware of
1332 * @dev, such as by generating a gratuitous ARP. This may be used when
1333 * a device wants to inform the rest of the network about some sort of
1334 * reconfiguration such as a failover event or virtual machine
1337 void __netdev_notify_peers(struct net_device *dev)
1340 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
1341 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev);
1343 EXPORT_SYMBOL(__netdev_notify_peers);
1346 * netdev_notify_peers - notify network peers about existence of @dev
1347 * @dev: network device
1349 * Generate traffic such that interested network peers are aware of
1350 * @dev, such as by generating a gratuitous ARP. This may be used when
1351 * a device wants to inform the rest of the network about some sort of
1352 * reconfiguration such as a failover event or virtual machine
1355 void netdev_notify_peers(struct net_device *dev)
1358 __netdev_notify_peers(dev);
1361 EXPORT_SYMBOL(netdev_notify_peers);
1363 static int napi_threaded_poll(void *data);
1365 static int napi_kthread_create(struct napi_struct *n)
1369 /* Create and wake up the kthread once to put it in
1370 * TASK_INTERRUPTIBLE mode to avoid the blocked task
1371 * warning and work with loadavg.
1373 n->thread = kthread_run(napi_threaded_poll, n, "napi/%s-%d",
1374 n->dev->name, n->napi_id);
1375 if (IS_ERR(n->thread)) {
1376 err = PTR_ERR(n->thread);
1377 pr_err("kthread_run failed with err %d\n", err);
1384 static int __dev_open(struct net_device *dev, struct netlink_ext_ack *extack)
1386 const struct net_device_ops *ops = dev->netdev_ops;
1390 dev_addr_check(dev);
1392 if (!netif_device_present(dev)) {
1393 /* may be detached because parent is runtime-suspended */
1394 if (dev->dev.parent)
1395 pm_runtime_resume(dev->dev.parent);
1396 if (!netif_device_present(dev))
1400 /* Block netpoll from trying to do any rx path servicing.
1401 * If we don't do this there is a chance ndo_poll_controller
1402 * or ndo_poll may be running while we open the device
1404 netpoll_poll_disable(dev);
1406 ret = call_netdevice_notifiers_extack(NETDEV_PRE_UP, dev, extack);
1407 ret = notifier_to_errno(ret);
1411 set_bit(__LINK_STATE_START, &dev->state);
1413 if (ops->ndo_validate_addr)
1414 ret = ops->ndo_validate_addr(dev);
1416 if (!ret && ops->ndo_open)
1417 ret = ops->ndo_open(dev);
1419 netpoll_poll_enable(dev);
1422 clear_bit(__LINK_STATE_START, &dev->state);
1424 dev->flags |= IFF_UP;
1425 dev_set_rx_mode(dev);
1427 add_device_randomness(dev->dev_addr, dev->addr_len);
1434 * dev_open - prepare an interface for use.
1435 * @dev: device to open
1436 * @extack: netlink extended ack
1438 * Takes a device from down to up state. The device's private open
1439 * function is invoked and then the multicast lists are loaded. Finally
1440 * the device is moved into the up state and a %NETDEV_UP message is
1441 * sent to the netdev notifier chain.
1443 * Calling this function on an active interface is a nop. On a failure
1444 * a negative errno code is returned.
1446 int dev_open(struct net_device *dev, struct netlink_ext_ack *extack)
1450 if (dev->flags & IFF_UP)
1453 ret = __dev_open(dev, extack);
1457 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL);
1458 call_netdevice_notifiers(NETDEV_UP, dev);
1462 EXPORT_SYMBOL(dev_open);
1464 static void __dev_close_many(struct list_head *head)
1466 struct net_device *dev;
1471 list_for_each_entry(dev, head, close_list) {
1472 /* Temporarily disable netpoll until the interface is down */
1473 netpoll_poll_disable(dev);
1475 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev);
1477 clear_bit(__LINK_STATE_START, &dev->state);
1479 /* Synchronize to scheduled poll. We cannot touch poll list, it
1480 * can be even on different cpu. So just clear netif_running().
1482 * dev->stop() will invoke napi_disable() on all of it's
1483 * napi_struct instances on this device.
1485 smp_mb__after_atomic(); /* Commit netif_running(). */
1488 dev_deactivate_many(head);
1490 list_for_each_entry(dev, head, close_list) {
1491 const struct net_device_ops *ops = dev->netdev_ops;
1494 * Call the device specific close. This cannot fail.
1495 * Only if device is UP
1497 * We allow it to be called even after a DETACH hot-plug
1503 dev->flags &= ~IFF_UP;
1504 netpoll_poll_enable(dev);
1508 static void __dev_close(struct net_device *dev)
1512 list_add(&dev->close_list, &single);
1513 __dev_close_many(&single);
1517 void dev_close_many(struct list_head *head, bool unlink)
1519 struct net_device *dev, *tmp;
1521 /* Remove the devices that don't need to be closed */
1522 list_for_each_entry_safe(dev, tmp, head, close_list)
1523 if (!(dev->flags & IFF_UP))
1524 list_del_init(&dev->close_list);
1526 __dev_close_many(head);
1528 list_for_each_entry_safe(dev, tmp, head, close_list) {
1529 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL);
1530 call_netdevice_notifiers(NETDEV_DOWN, dev);
1532 list_del_init(&dev->close_list);
1535 EXPORT_SYMBOL(dev_close_many);
1538 * dev_close - shutdown an interface.
1539 * @dev: device to shutdown
1541 * This function moves an active device into down state. A
1542 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device
1543 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier
1546 void dev_close(struct net_device *dev)
1548 if (dev->flags & IFF_UP) {
1551 list_add(&dev->close_list, &single);
1552 dev_close_many(&single, true);
1556 EXPORT_SYMBOL(dev_close);
1560 * dev_disable_lro - disable Large Receive Offload on a device
1563 * Disable Large Receive Offload (LRO) on a net device. Must be
1564 * called under RTNL. This is needed if received packets may be
1565 * forwarded to another interface.
1567 void dev_disable_lro(struct net_device *dev)
1569 struct net_device *lower_dev;
1570 struct list_head *iter;
1572 dev->wanted_features &= ~NETIF_F_LRO;
1573 netdev_update_features(dev);
1575 if (unlikely(dev->features & NETIF_F_LRO))
1576 netdev_WARN(dev, "failed to disable LRO!\n");
1578 netdev_for_each_lower_dev(dev, lower_dev, iter)
1579 dev_disable_lro(lower_dev);
1581 EXPORT_SYMBOL(dev_disable_lro);
1584 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device
1587 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be
1588 * called under RTNL. This is needed if Generic XDP is installed on
1591 static void dev_disable_gro_hw(struct net_device *dev)
1593 dev->wanted_features &= ~NETIF_F_GRO_HW;
1594 netdev_update_features(dev);
1596 if (unlikely(dev->features & NETIF_F_GRO_HW))
1597 netdev_WARN(dev, "failed to disable GRO_HW!\n");
1600 const char *netdev_cmd_to_name(enum netdev_cmd cmd)
1603 case NETDEV_##val: \
1604 return "NETDEV_" __stringify(val);
1606 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER)
1607 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE)
1608 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE)
1609 N(POST_INIT) N(PRE_UNINIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN)
1610 N(CHANGEUPPER) N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA)
1611 N(BONDING_INFO) N(PRECHANGEUPPER) N(CHANGELOWERSTATE)
1612 N(UDP_TUNNEL_PUSH_INFO) N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN)
1613 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO)
1614 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO)
1615 N(PRE_CHANGEADDR) N(OFFLOAD_XSTATS_ENABLE) N(OFFLOAD_XSTATS_DISABLE)
1616 N(OFFLOAD_XSTATS_REPORT_USED) N(OFFLOAD_XSTATS_REPORT_DELTA)
1620 return "UNKNOWN_NETDEV_EVENT";
1622 EXPORT_SYMBOL_GPL(netdev_cmd_to_name);
1624 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val,
1625 struct net_device *dev)
1627 struct netdev_notifier_info info = {
1631 return nb->notifier_call(nb, val, &info);
1634 static int call_netdevice_register_notifiers(struct notifier_block *nb,
1635 struct net_device *dev)
1639 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev);
1640 err = notifier_to_errno(err);
1644 if (!(dev->flags & IFF_UP))
1647 call_netdevice_notifier(nb, NETDEV_UP, dev);
1651 static void call_netdevice_unregister_notifiers(struct notifier_block *nb,
1652 struct net_device *dev)
1654 if (dev->flags & IFF_UP) {
1655 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1657 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1659 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1662 static int call_netdevice_register_net_notifiers(struct notifier_block *nb,
1665 struct net_device *dev;
1668 for_each_netdev(net, dev) {
1669 err = call_netdevice_register_notifiers(nb, dev);
1676 for_each_netdev_continue_reverse(net, dev)
1677 call_netdevice_unregister_notifiers(nb, dev);
1681 static void call_netdevice_unregister_net_notifiers(struct notifier_block *nb,
1684 struct net_device *dev;
1686 for_each_netdev(net, dev)
1687 call_netdevice_unregister_notifiers(nb, dev);
1690 static int dev_boot_phase = 1;
1693 * register_netdevice_notifier - register a network notifier block
1696 * Register a notifier to be called when network device events occur.
1697 * The notifier passed is linked into the kernel structures and must
1698 * not be reused until it has been unregistered. A negative errno code
1699 * is returned on a failure.
1701 * When registered all registration and up events are replayed
1702 * to the new notifier to allow device to have a race free
1703 * view of the network device list.
1706 int register_netdevice_notifier(struct notifier_block *nb)
1711 /* Close race with setup_net() and cleanup_net() */
1712 down_write(&pernet_ops_rwsem);
1714 err = raw_notifier_chain_register(&netdev_chain, nb);
1720 err = call_netdevice_register_net_notifiers(nb, net);
1727 up_write(&pernet_ops_rwsem);
1731 for_each_net_continue_reverse(net)
1732 call_netdevice_unregister_net_notifiers(nb, net);
1734 raw_notifier_chain_unregister(&netdev_chain, nb);
1737 EXPORT_SYMBOL(register_netdevice_notifier);
1740 * unregister_netdevice_notifier - unregister a network notifier block
1743 * Unregister a notifier previously registered by
1744 * register_netdevice_notifier(). The notifier is unlinked into the
1745 * kernel structures and may then be reused. A negative errno code
1746 * is returned on a failure.
1748 * After unregistering unregister and down device events are synthesized
1749 * for all devices on the device list to the removed notifier to remove
1750 * the need for special case cleanup code.
1753 int unregister_netdevice_notifier(struct notifier_block *nb)
1758 /* Close race with setup_net() and cleanup_net() */
1759 down_write(&pernet_ops_rwsem);
1761 err = raw_notifier_chain_unregister(&netdev_chain, nb);
1766 call_netdevice_unregister_net_notifiers(nb, net);
1770 up_write(&pernet_ops_rwsem);
1773 EXPORT_SYMBOL(unregister_netdevice_notifier);
1775 static int __register_netdevice_notifier_net(struct net *net,
1776 struct notifier_block *nb,
1777 bool ignore_call_fail)
1781 err = raw_notifier_chain_register(&net->netdev_chain, nb);
1787 err = call_netdevice_register_net_notifiers(nb, net);
1788 if (err && !ignore_call_fail)
1789 goto chain_unregister;
1794 raw_notifier_chain_unregister(&net->netdev_chain, nb);
1798 static int __unregister_netdevice_notifier_net(struct net *net,
1799 struct notifier_block *nb)
1803 err = raw_notifier_chain_unregister(&net->netdev_chain, nb);
1807 call_netdevice_unregister_net_notifiers(nb, net);
1812 * register_netdevice_notifier_net - register a per-netns network notifier block
1813 * @net: network namespace
1816 * Register a notifier to be called when network device events occur.
1817 * The notifier passed is linked into the kernel structures and must
1818 * not be reused until it has been unregistered. A negative errno code
1819 * is returned on a failure.
1821 * When registered all registration and up events are replayed
1822 * to the new notifier to allow device to have a race free
1823 * view of the network device list.
1826 int register_netdevice_notifier_net(struct net *net, struct notifier_block *nb)
1831 err = __register_netdevice_notifier_net(net, nb, false);
1835 EXPORT_SYMBOL(register_netdevice_notifier_net);
1838 * unregister_netdevice_notifier_net - unregister a per-netns
1839 * network notifier block
1840 * @net: network namespace
1843 * Unregister a notifier previously registered by
1844 * register_netdevice_notifier_net(). The notifier is unlinked from the
1845 * kernel structures and may then be reused. A negative errno code
1846 * is returned on a failure.
1848 * After unregistering unregister and down device events are synthesized
1849 * for all devices on the device list to the removed notifier to remove
1850 * the need for special case cleanup code.
1853 int unregister_netdevice_notifier_net(struct net *net,
1854 struct notifier_block *nb)
1859 err = __unregister_netdevice_notifier_net(net, nb);
1863 EXPORT_SYMBOL(unregister_netdevice_notifier_net);
1865 static void __move_netdevice_notifier_net(struct net *src_net,
1866 struct net *dst_net,
1867 struct notifier_block *nb)
1869 __unregister_netdevice_notifier_net(src_net, nb);
1870 __register_netdevice_notifier_net(dst_net, nb, true);
1873 int register_netdevice_notifier_dev_net(struct net_device *dev,
1874 struct notifier_block *nb,
1875 struct netdev_net_notifier *nn)
1880 err = __register_netdevice_notifier_net(dev_net(dev), nb, false);
1883 list_add(&nn->list, &dev->net_notifier_list);
1888 EXPORT_SYMBOL(register_netdevice_notifier_dev_net);
1890 int unregister_netdevice_notifier_dev_net(struct net_device *dev,
1891 struct notifier_block *nb,
1892 struct netdev_net_notifier *nn)
1897 list_del(&nn->list);
1898 err = __unregister_netdevice_notifier_net(dev_net(dev), nb);
1902 EXPORT_SYMBOL(unregister_netdevice_notifier_dev_net);
1904 static void move_netdevice_notifiers_dev_net(struct net_device *dev,
1907 struct netdev_net_notifier *nn;
1909 list_for_each_entry(nn, &dev->net_notifier_list, list)
1910 __move_netdevice_notifier_net(dev_net(dev), net, nn->nb);
1914 * call_netdevice_notifiers_info - call all network notifier blocks
1915 * @val: value passed unmodified to notifier function
1916 * @info: notifier information data
1918 * Call all network notifier blocks. Parameters and return value
1919 * are as for raw_notifier_call_chain().
1922 static int call_netdevice_notifiers_info(unsigned long val,
1923 struct netdev_notifier_info *info)
1925 struct net *net = dev_net(info->dev);
1930 /* Run per-netns notifier block chain first, then run the global one.
1931 * Hopefully, one day, the global one is going to be removed after
1932 * all notifier block registrators get converted to be per-netns.
1934 ret = raw_notifier_call_chain(&net->netdev_chain, val, info);
1935 if (ret & NOTIFY_STOP_MASK)
1937 return raw_notifier_call_chain(&netdev_chain, val, info);
1941 * call_netdevice_notifiers_info_robust - call per-netns notifier blocks
1942 * for and rollback on error
1943 * @val_up: value passed unmodified to notifier function
1944 * @val_down: value passed unmodified to the notifier function when
1945 * recovering from an error on @val_up
1946 * @info: notifier information data
1948 * Call all per-netns network notifier blocks, but not notifier blocks on
1949 * the global notifier chain. Parameters and return value are as for
1950 * raw_notifier_call_chain_robust().
1954 call_netdevice_notifiers_info_robust(unsigned long val_up,
1955 unsigned long val_down,
1956 struct netdev_notifier_info *info)
1958 struct net *net = dev_net(info->dev);
1962 return raw_notifier_call_chain_robust(&net->netdev_chain,
1963 val_up, val_down, info);
1966 static int call_netdevice_notifiers_extack(unsigned long val,
1967 struct net_device *dev,
1968 struct netlink_ext_ack *extack)
1970 struct netdev_notifier_info info = {
1975 return call_netdevice_notifiers_info(val, &info);
1979 * call_netdevice_notifiers - call all network notifier blocks
1980 * @val: value passed unmodified to notifier function
1981 * @dev: net_device pointer passed unmodified to notifier function
1983 * Call all network notifier blocks. Parameters and return value
1984 * are as for raw_notifier_call_chain().
1987 int call_netdevice_notifiers(unsigned long val, struct net_device *dev)
1989 return call_netdevice_notifiers_extack(val, dev, NULL);
1991 EXPORT_SYMBOL(call_netdevice_notifiers);
1994 * call_netdevice_notifiers_mtu - call all network notifier blocks
1995 * @val: value passed unmodified to notifier function
1996 * @dev: net_device pointer passed unmodified to notifier function
1997 * @arg: additional u32 argument passed to the notifier function
1999 * Call all network notifier blocks. Parameters and return value
2000 * are as for raw_notifier_call_chain().
2002 static int call_netdevice_notifiers_mtu(unsigned long val,
2003 struct net_device *dev, u32 arg)
2005 struct netdev_notifier_info_ext info = {
2010 BUILD_BUG_ON(offsetof(struct netdev_notifier_info_ext, info) != 0);
2012 return call_netdevice_notifiers_info(val, &info.info);
2015 #ifdef CONFIG_NET_INGRESS
2016 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key);
2018 void net_inc_ingress_queue(void)
2020 static_branch_inc(&ingress_needed_key);
2022 EXPORT_SYMBOL_GPL(net_inc_ingress_queue);
2024 void net_dec_ingress_queue(void)
2026 static_branch_dec(&ingress_needed_key);
2028 EXPORT_SYMBOL_GPL(net_dec_ingress_queue);
2031 #ifdef CONFIG_NET_EGRESS
2032 static DEFINE_STATIC_KEY_FALSE(egress_needed_key);
2034 void net_inc_egress_queue(void)
2036 static_branch_inc(&egress_needed_key);
2038 EXPORT_SYMBOL_GPL(net_inc_egress_queue);
2040 void net_dec_egress_queue(void)
2042 static_branch_dec(&egress_needed_key);
2044 EXPORT_SYMBOL_GPL(net_dec_egress_queue);
2047 DEFINE_STATIC_KEY_FALSE(netstamp_needed_key);
2048 EXPORT_SYMBOL(netstamp_needed_key);
2049 #ifdef CONFIG_JUMP_LABEL
2050 static atomic_t netstamp_needed_deferred;
2051 static atomic_t netstamp_wanted;
2052 static void netstamp_clear(struct work_struct *work)
2054 int deferred = atomic_xchg(&netstamp_needed_deferred, 0);
2057 wanted = atomic_add_return(deferred, &netstamp_wanted);
2059 static_branch_enable(&netstamp_needed_key);
2061 static_branch_disable(&netstamp_needed_key);
2063 static DECLARE_WORK(netstamp_work, netstamp_clear);
2066 void net_enable_timestamp(void)
2068 #ifdef CONFIG_JUMP_LABEL
2069 int wanted = atomic_read(&netstamp_wanted);
2071 while (wanted > 0) {
2072 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted + 1))
2075 atomic_inc(&netstamp_needed_deferred);
2076 schedule_work(&netstamp_work);
2078 static_branch_inc(&netstamp_needed_key);
2081 EXPORT_SYMBOL(net_enable_timestamp);
2083 void net_disable_timestamp(void)
2085 #ifdef CONFIG_JUMP_LABEL
2086 int wanted = atomic_read(&netstamp_wanted);
2088 while (wanted > 1) {
2089 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted - 1))
2092 atomic_dec(&netstamp_needed_deferred);
2093 schedule_work(&netstamp_work);
2095 static_branch_dec(&netstamp_needed_key);
2098 EXPORT_SYMBOL(net_disable_timestamp);
2100 static inline void net_timestamp_set(struct sk_buff *skb)
2103 skb->mono_delivery_time = 0;
2104 if (static_branch_unlikely(&netstamp_needed_key))
2105 skb->tstamp = ktime_get_real();
2108 #define net_timestamp_check(COND, SKB) \
2109 if (static_branch_unlikely(&netstamp_needed_key)) { \
2110 if ((COND) && !(SKB)->tstamp) \
2111 (SKB)->tstamp = ktime_get_real(); \
2114 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb)
2116 return __is_skb_forwardable(dev, skb, true);
2118 EXPORT_SYMBOL_GPL(is_skb_forwardable);
2120 static int __dev_forward_skb2(struct net_device *dev, struct sk_buff *skb,
2123 int ret = ____dev_forward_skb(dev, skb, check_mtu);
2126 skb->protocol = eth_type_trans(skb, dev);
2127 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
2133 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
2135 return __dev_forward_skb2(dev, skb, true);
2137 EXPORT_SYMBOL_GPL(__dev_forward_skb);
2140 * dev_forward_skb - loopback an skb to another netif
2142 * @dev: destination network device
2143 * @skb: buffer to forward
2146 * NET_RX_SUCCESS (no congestion)
2147 * NET_RX_DROP (packet was dropped, but freed)
2149 * dev_forward_skb can be used for injecting an skb from the
2150 * start_xmit function of one device into the receive queue
2151 * of another device.
2153 * The receiving device may be in another namespace, so
2154 * we have to clear all information in the skb that could
2155 * impact namespace isolation.
2157 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
2159 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb);
2161 EXPORT_SYMBOL_GPL(dev_forward_skb);
2163 int dev_forward_skb_nomtu(struct net_device *dev, struct sk_buff *skb)
2165 return __dev_forward_skb2(dev, skb, false) ?: netif_rx_internal(skb);
2168 static inline int deliver_skb(struct sk_buff *skb,
2169 struct packet_type *pt_prev,
2170 struct net_device *orig_dev)
2172 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
2174 refcount_inc(&skb->users);
2175 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
2178 static inline void deliver_ptype_list_skb(struct sk_buff *skb,
2179 struct packet_type **pt,
2180 struct net_device *orig_dev,
2182 struct list_head *ptype_list)
2184 struct packet_type *ptype, *pt_prev = *pt;
2186 list_for_each_entry_rcu(ptype, ptype_list, list) {
2187 if (ptype->type != type)
2190 deliver_skb(skb, pt_prev, orig_dev);
2196 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb)
2198 if (!ptype->af_packet_priv || !skb->sk)
2201 if (ptype->id_match)
2202 return ptype->id_match(ptype, skb->sk);
2203 else if ((struct sock *)ptype->af_packet_priv == skb->sk)
2210 * dev_nit_active - return true if any network interface taps are in use
2212 * @dev: network device to check for the presence of taps
2214 bool dev_nit_active(struct net_device *dev)
2216 return !list_empty(&ptype_all) || !list_empty(&dev->ptype_all);
2218 EXPORT_SYMBOL_GPL(dev_nit_active);
2221 * Support routine. Sends outgoing frames to any network
2222 * taps currently in use.
2225 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
2227 struct packet_type *ptype;
2228 struct sk_buff *skb2 = NULL;
2229 struct packet_type *pt_prev = NULL;
2230 struct list_head *ptype_list = &ptype_all;
2234 list_for_each_entry_rcu(ptype, ptype_list, list) {
2235 if (ptype->ignore_outgoing)
2238 /* Never send packets back to the socket
2239 * they originated from - MvS (miquels@drinkel.ow.org)
2241 if (skb_loop_sk(ptype, skb))
2245 deliver_skb(skb2, pt_prev, skb->dev);
2250 /* need to clone skb, done only once */
2251 skb2 = skb_clone(skb, GFP_ATOMIC);
2255 net_timestamp_set(skb2);
2257 /* skb->nh should be correctly
2258 * set by sender, so that the second statement is
2259 * just protection against buggy protocols.
2261 skb_reset_mac_header(skb2);
2263 if (skb_network_header(skb2) < skb2->data ||
2264 skb_network_header(skb2) > skb_tail_pointer(skb2)) {
2265 net_crit_ratelimited("protocol %04x is buggy, dev %s\n",
2266 ntohs(skb2->protocol),
2268 skb_reset_network_header(skb2);
2271 skb2->transport_header = skb2->network_header;
2272 skb2->pkt_type = PACKET_OUTGOING;
2276 if (ptype_list == &ptype_all) {
2277 ptype_list = &dev->ptype_all;
2282 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC))
2283 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
2289 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit);
2292 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change
2293 * @dev: Network device
2294 * @txq: number of queues available
2296 * If real_num_tx_queues is changed the tc mappings may no longer be
2297 * valid. To resolve this verify the tc mapping remains valid and if
2298 * not NULL the mapping. With no priorities mapping to this
2299 * offset/count pair it will no longer be used. In the worst case TC0
2300 * is invalid nothing can be done so disable priority mappings. If is
2301 * expected that drivers will fix this mapping if they can before
2302 * calling netif_set_real_num_tx_queues.
2304 static void netif_setup_tc(struct net_device *dev, unsigned int txq)
2307 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2309 /* If TC0 is invalidated disable TC mapping */
2310 if (tc->offset + tc->count > txq) {
2311 netdev_warn(dev, "Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n");
2316 /* Invalidated prio to tc mappings set to TC0 */
2317 for (i = 1; i < TC_BITMASK + 1; i++) {
2318 int q = netdev_get_prio_tc_map(dev, i);
2320 tc = &dev->tc_to_txq[q];
2321 if (tc->offset + tc->count > txq) {
2322 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",
2324 netdev_set_prio_tc_map(dev, i, 0);
2329 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq)
2332 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2335 /* walk through the TCs and see if it falls into any of them */
2336 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) {
2337 if ((txq - tc->offset) < tc->count)
2341 /* didn't find it, just return -1 to indicate no match */
2347 EXPORT_SYMBOL(netdev_txq_to_tc);
2350 static struct static_key xps_needed __read_mostly;
2351 static struct static_key xps_rxqs_needed __read_mostly;
2352 static DEFINE_MUTEX(xps_map_mutex);
2353 #define xmap_dereference(P) \
2354 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex))
2356 static bool remove_xps_queue(struct xps_dev_maps *dev_maps,
2357 struct xps_dev_maps *old_maps, int tci, u16 index)
2359 struct xps_map *map = NULL;
2363 map = xmap_dereference(dev_maps->attr_map[tci]);
2367 for (pos = map->len; pos--;) {
2368 if (map->queues[pos] != index)
2372 map->queues[pos] = map->queues[--map->len];
2377 RCU_INIT_POINTER(old_maps->attr_map[tci], NULL);
2378 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL);
2379 kfree_rcu(map, rcu);
2386 static bool remove_xps_queue_cpu(struct net_device *dev,
2387 struct xps_dev_maps *dev_maps,
2388 int cpu, u16 offset, u16 count)
2390 int num_tc = dev_maps->num_tc;
2391 bool active = false;
2394 for (tci = cpu * num_tc; num_tc--; tci++) {
2397 for (i = count, j = offset; i--; j++) {
2398 if (!remove_xps_queue(dev_maps, NULL, tci, j))
2408 static void reset_xps_maps(struct net_device *dev,
2409 struct xps_dev_maps *dev_maps,
2410 enum xps_map_type type)
2412 static_key_slow_dec_cpuslocked(&xps_needed);
2413 if (type == XPS_RXQS)
2414 static_key_slow_dec_cpuslocked(&xps_rxqs_needed);
2416 RCU_INIT_POINTER(dev->xps_maps[type], NULL);
2418 kfree_rcu(dev_maps, rcu);
2421 static void clean_xps_maps(struct net_device *dev, enum xps_map_type type,
2422 u16 offset, u16 count)
2424 struct xps_dev_maps *dev_maps;
2425 bool active = false;
2428 dev_maps = xmap_dereference(dev->xps_maps[type]);
2432 for (j = 0; j < dev_maps->nr_ids; j++)
2433 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, count);
2435 reset_xps_maps(dev, dev_maps, type);
2437 if (type == XPS_CPUS) {
2438 for (i = offset + (count - 1); count--; i--)
2439 netdev_queue_numa_node_write(
2440 netdev_get_tx_queue(dev, i), NUMA_NO_NODE);
2444 static void netif_reset_xps_queues(struct net_device *dev, u16 offset,
2447 if (!static_key_false(&xps_needed))
2451 mutex_lock(&xps_map_mutex);
2453 if (static_key_false(&xps_rxqs_needed))
2454 clean_xps_maps(dev, XPS_RXQS, offset, count);
2456 clean_xps_maps(dev, XPS_CPUS, offset, count);
2458 mutex_unlock(&xps_map_mutex);
2462 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index)
2464 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index);
2467 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index,
2468 u16 index, bool is_rxqs_map)
2470 struct xps_map *new_map;
2471 int alloc_len = XPS_MIN_MAP_ALLOC;
2474 for (pos = 0; map && pos < map->len; pos++) {
2475 if (map->queues[pos] != index)
2480 /* Need to add tx-queue to this CPU's/rx-queue's existing map */
2482 if (pos < map->alloc_len)
2485 alloc_len = map->alloc_len * 2;
2488 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's
2492 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL);
2494 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL,
2495 cpu_to_node(attr_index));
2499 for (i = 0; i < pos; i++)
2500 new_map->queues[i] = map->queues[i];
2501 new_map->alloc_len = alloc_len;
2507 /* Copy xps maps at a given index */
2508 static void xps_copy_dev_maps(struct xps_dev_maps *dev_maps,
2509 struct xps_dev_maps *new_dev_maps, int index,
2510 int tc, bool skip_tc)
2512 int i, tci = index * dev_maps->num_tc;
2513 struct xps_map *map;
2515 /* copy maps belonging to foreign traffic classes */
2516 for (i = 0; i < dev_maps->num_tc; i++, tci++) {
2517 if (i == tc && skip_tc)
2520 /* fill in the new device map from the old device map */
2521 map = xmap_dereference(dev_maps->attr_map[tci]);
2522 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map);
2526 /* Must be called under cpus_read_lock */
2527 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask,
2528 u16 index, enum xps_map_type type)
2530 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL, *old_dev_maps = NULL;
2531 const unsigned long *online_mask = NULL;
2532 bool active = false, copy = false;
2533 int i, j, tci, numa_node_id = -2;
2534 int maps_sz, num_tc = 1, tc = 0;
2535 struct xps_map *map, *new_map;
2536 unsigned int nr_ids;
2538 WARN_ON_ONCE(index >= dev->num_tx_queues);
2541 /* Do not allow XPS on subordinate device directly */
2542 num_tc = dev->num_tc;
2546 /* If queue belongs to subordinate dev use its map */
2547 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev;
2549 tc = netdev_txq_to_tc(dev, index);
2554 mutex_lock(&xps_map_mutex);
2556 dev_maps = xmap_dereference(dev->xps_maps[type]);
2557 if (type == XPS_RXQS) {
2558 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues);
2559 nr_ids = dev->num_rx_queues;
2561 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc);
2562 if (num_possible_cpus() > 1)
2563 online_mask = cpumask_bits(cpu_online_mask);
2564 nr_ids = nr_cpu_ids;
2567 if (maps_sz < L1_CACHE_BYTES)
2568 maps_sz = L1_CACHE_BYTES;
2570 /* The old dev_maps could be larger or smaller than the one we're
2571 * setting up now, as dev->num_tc or nr_ids could have been updated in
2572 * between. We could try to be smart, but let's be safe instead and only
2573 * copy foreign traffic classes if the two map sizes match.
2576 dev_maps->num_tc == num_tc && dev_maps->nr_ids == nr_ids)
2579 /* allocate memory for queue storage */
2580 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids),
2582 if (!new_dev_maps) {
2583 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL);
2584 if (!new_dev_maps) {
2585 mutex_unlock(&xps_map_mutex);
2589 new_dev_maps->nr_ids = nr_ids;
2590 new_dev_maps->num_tc = num_tc;
2593 tci = j * num_tc + tc;
2594 map = copy ? xmap_dereference(dev_maps->attr_map[tci]) : NULL;
2596 map = expand_xps_map(map, j, index, type == XPS_RXQS);
2600 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map);
2604 goto out_no_new_maps;
2607 /* Increment static keys at most once per type */
2608 static_key_slow_inc_cpuslocked(&xps_needed);
2609 if (type == XPS_RXQS)
2610 static_key_slow_inc_cpuslocked(&xps_rxqs_needed);
2613 for (j = 0; j < nr_ids; j++) {
2614 bool skip_tc = false;
2616 tci = j * num_tc + tc;
2617 if (netif_attr_test_mask(j, mask, nr_ids) &&
2618 netif_attr_test_online(j, online_mask, nr_ids)) {
2619 /* add tx-queue to CPU/rx-queue maps */
2624 map = xmap_dereference(new_dev_maps->attr_map[tci]);
2625 while ((pos < map->len) && (map->queues[pos] != index))
2628 if (pos == map->len)
2629 map->queues[map->len++] = index;
2631 if (type == XPS_CPUS) {
2632 if (numa_node_id == -2)
2633 numa_node_id = cpu_to_node(j);
2634 else if (numa_node_id != cpu_to_node(j))
2641 xps_copy_dev_maps(dev_maps, new_dev_maps, j, tc,
2645 rcu_assign_pointer(dev->xps_maps[type], new_dev_maps);
2647 /* Cleanup old maps */
2649 goto out_no_old_maps;
2651 for (j = 0; j < dev_maps->nr_ids; j++) {
2652 for (i = num_tc, tci = j * dev_maps->num_tc; i--; tci++) {
2653 map = xmap_dereference(dev_maps->attr_map[tci]);
2658 new_map = xmap_dereference(new_dev_maps->attr_map[tci]);
2663 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL);
2664 kfree_rcu(map, rcu);
2668 old_dev_maps = dev_maps;
2671 dev_maps = new_dev_maps;
2675 if (type == XPS_CPUS)
2676 /* update Tx queue numa node */
2677 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index),
2678 (numa_node_id >= 0) ?
2679 numa_node_id : NUMA_NO_NODE);
2684 /* removes tx-queue from unused CPUs/rx-queues */
2685 for (j = 0; j < dev_maps->nr_ids; j++) {
2686 tci = j * dev_maps->num_tc;
2688 for (i = 0; i < dev_maps->num_tc; i++, tci++) {
2690 netif_attr_test_mask(j, mask, dev_maps->nr_ids) &&
2691 netif_attr_test_online(j, online_mask, dev_maps->nr_ids))
2694 active |= remove_xps_queue(dev_maps,
2695 copy ? old_dev_maps : NULL,
2701 kfree_rcu(old_dev_maps, rcu);
2703 /* free map if not active */
2705 reset_xps_maps(dev, dev_maps, type);
2708 mutex_unlock(&xps_map_mutex);
2712 /* remove any maps that we added */
2713 for (j = 0; j < nr_ids; j++) {
2714 for (i = num_tc, tci = j * num_tc; i--; tci++) {
2715 new_map = xmap_dereference(new_dev_maps->attr_map[tci]);
2717 xmap_dereference(dev_maps->attr_map[tci]) :
2719 if (new_map && new_map != map)
2724 mutex_unlock(&xps_map_mutex);
2726 kfree(new_dev_maps);
2729 EXPORT_SYMBOL_GPL(__netif_set_xps_queue);
2731 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask,
2737 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, XPS_CPUS);
2742 EXPORT_SYMBOL(netif_set_xps_queue);
2745 static void netdev_unbind_all_sb_channels(struct net_device *dev)
2747 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues];
2749 /* Unbind any subordinate channels */
2750 while (txq-- != &dev->_tx[0]) {
2752 netdev_unbind_sb_channel(dev, txq->sb_dev);
2756 void netdev_reset_tc(struct net_device *dev)
2759 netif_reset_xps_queues_gt(dev, 0);
2761 netdev_unbind_all_sb_channels(dev);
2763 /* Reset TC configuration of device */
2765 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq));
2766 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map));
2768 EXPORT_SYMBOL(netdev_reset_tc);
2770 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset)
2772 if (tc >= dev->num_tc)
2776 netif_reset_xps_queues(dev, offset, count);
2778 dev->tc_to_txq[tc].count = count;
2779 dev->tc_to_txq[tc].offset = offset;
2782 EXPORT_SYMBOL(netdev_set_tc_queue);
2784 int netdev_set_num_tc(struct net_device *dev, u8 num_tc)
2786 if (num_tc > TC_MAX_QUEUE)
2790 netif_reset_xps_queues_gt(dev, 0);
2792 netdev_unbind_all_sb_channels(dev);
2794 dev->num_tc = num_tc;
2797 EXPORT_SYMBOL(netdev_set_num_tc);
2799 void netdev_unbind_sb_channel(struct net_device *dev,
2800 struct net_device *sb_dev)
2802 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues];
2805 netif_reset_xps_queues_gt(sb_dev, 0);
2807 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq));
2808 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map));
2810 while (txq-- != &dev->_tx[0]) {
2811 if (txq->sb_dev == sb_dev)
2815 EXPORT_SYMBOL(netdev_unbind_sb_channel);
2817 int netdev_bind_sb_channel_queue(struct net_device *dev,
2818 struct net_device *sb_dev,
2819 u8 tc, u16 count, u16 offset)
2821 /* Make certain the sb_dev and dev are already configured */
2822 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc)
2825 /* We cannot hand out queues we don't have */
2826 if ((offset + count) > dev->real_num_tx_queues)
2829 /* Record the mapping */
2830 sb_dev->tc_to_txq[tc].count = count;
2831 sb_dev->tc_to_txq[tc].offset = offset;
2833 /* Provide a way for Tx queue to find the tc_to_txq map or
2834 * XPS map for itself.
2837 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev;
2841 EXPORT_SYMBOL(netdev_bind_sb_channel_queue);
2843 int netdev_set_sb_channel(struct net_device *dev, u16 channel)
2845 /* Do not use a multiqueue device to represent a subordinate channel */
2846 if (netif_is_multiqueue(dev))
2849 /* We allow channels 1 - 32767 to be used for subordinate channels.
2850 * Channel 0 is meant to be "native" mode and used only to represent
2851 * the main root device. We allow writing 0 to reset the device back
2852 * to normal mode after being used as a subordinate channel.
2854 if (channel > S16_MAX)
2857 dev->num_tc = -channel;
2861 EXPORT_SYMBOL(netdev_set_sb_channel);
2864 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues
2865 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed.
2867 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq)
2872 disabling = txq < dev->real_num_tx_queues;
2874 if (txq < 1 || txq > dev->num_tx_queues)
2877 if (dev->reg_state == NETREG_REGISTERED ||
2878 dev->reg_state == NETREG_UNREGISTERING) {
2881 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues,
2887 netif_setup_tc(dev, txq);
2889 dev_qdisc_change_real_num_tx(dev, txq);
2891 dev->real_num_tx_queues = txq;
2895 qdisc_reset_all_tx_gt(dev, txq);
2897 netif_reset_xps_queues_gt(dev, txq);
2901 dev->real_num_tx_queues = txq;
2906 EXPORT_SYMBOL(netif_set_real_num_tx_queues);
2910 * netif_set_real_num_rx_queues - set actual number of RX queues used
2911 * @dev: Network device
2912 * @rxq: Actual number of RX queues
2914 * This must be called either with the rtnl_lock held or before
2915 * registration of the net device. Returns 0 on success, or a
2916 * negative error code. If called before registration, it always
2919 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq)
2923 if (rxq < 1 || rxq > dev->num_rx_queues)
2926 if (dev->reg_state == NETREG_REGISTERED) {
2929 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues,
2935 dev->real_num_rx_queues = rxq;
2938 EXPORT_SYMBOL(netif_set_real_num_rx_queues);
2942 * netif_set_real_num_queues - set actual number of RX and TX queues used
2943 * @dev: Network device
2944 * @txq: Actual number of TX queues
2945 * @rxq: Actual number of RX queues
2947 * Set the real number of both TX and RX queues.
2948 * Does nothing if the number of queues is already correct.
2950 int netif_set_real_num_queues(struct net_device *dev,
2951 unsigned int txq, unsigned int rxq)
2953 unsigned int old_rxq = dev->real_num_rx_queues;
2956 if (txq < 1 || txq > dev->num_tx_queues ||
2957 rxq < 1 || rxq > dev->num_rx_queues)
2960 /* Start from increases, so the error path only does decreases -
2961 * decreases can't fail.
2963 if (rxq > dev->real_num_rx_queues) {
2964 err = netif_set_real_num_rx_queues(dev, rxq);
2968 if (txq > dev->real_num_tx_queues) {
2969 err = netif_set_real_num_tx_queues(dev, txq);
2973 if (rxq < dev->real_num_rx_queues)
2974 WARN_ON(netif_set_real_num_rx_queues(dev, rxq));
2975 if (txq < dev->real_num_tx_queues)
2976 WARN_ON(netif_set_real_num_tx_queues(dev, txq));
2980 WARN_ON(netif_set_real_num_rx_queues(dev, old_rxq));
2983 EXPORT_SYMBOL(netif_set_real_num_queues);
2986 * netif_set_tso_max_size() - set the max size of TSO frames supported
2987 * @dev: netdev to update
2988 * @size: max skb->len of a TSO frame
2990 * Set the limit on the size of TSO super-frames the device can handle.
2991 * Unless explicitly set the stack will assume the value of
2992 * %GSO_LEGACY_MAX_SIZE.
2994 void netif_set_tso_max_size(struct net_device *dev, unsigned int size)
2996 dev->tso_max_size = min(GSO_MAX_SIZE, size);
2997 if (size < READ_ONCE(dev->gso_max_size))
2998 netif_set_gso_max_size(dev, size);
2999 if (size < READ_ONCE(dev->gso_ipv4_max_size))
3000 netif_set_gso_ipv4_max_size(dev, size);
3002 EXPORT_SYMBOL(netif_set_tso_max_size);
3005 * netif_set_tso_max_segs() - set the max number of segs supported for TSO
3006 * @dev: netdev to update
3007 * @segs: max number of TCP segments
3009 * Set the limit on the number of TCP segments the device can generate from
3010 * a single TSO super-frame.
3011 * Unless explicitly set the stack will assume the value of %GSO_MAX_SEGS.
3013 void netif_set_tso_max_segs(struct net_device *dev, unsigned int segs)
3015 dev->tso_max_segs = segs;
3016 if (segs < READ_ONCE(dev->gso_max_segs))
3017 netif_set_gso_max_segs(dev, segs);
3019 EXPORT_SYMBOL(netif_set_tso_max_segs);
3022 * netif_inherit_tso_max() - copy all TSO limits from a lower device to an upper
3023 * @to: netdev to update
3024 * @from: netdev from which to copy the limits
3026 void netif_inherit_tso_max(struct net_device *to, const struct net_device *from)
3028 netif_set_tso_max_size(to, from->tso_max_size);
3029 netif_set_tso_max_segs(to, from->tso_max_segs);
3031 EXPORT_SYMBOL(netif_inherit_tso_max);
3034 * netif_get_num_default_rss_queues - default number of RSS queues
3036 * Default value is the number of physical cores if there are only 1 or 2, or
3037 * divided by 2 if there are more.
3039 int netif_get_num_default_rss_queues(void)
3044 if (unlikely(is_kdump_kernel() || !zalloc_cpumask_var(&cpus, GFP_KERNEL)))
3047 cpumask_copy(cpus, cpu_online_mask);
3048 for_each_cpu(cpu, cpus) {
3050 cpumask_andnot(cpus, cpus, topology_sibling_cpumask(cpu));
3052 free_cpumask_var(cpus);
3054 return count > 2 ? DIV_ROUND_UP(count, 2) : count;
3056 EXPORT_SYMBOL(netif_get_num_default_rss_queues);
3058 static void __netif_reschedule(struct Qdisc *q)
3060 struct softnet_data *sd;
3061 unsigned long flags;
3063 local_irq_save(flags);
3064 sd = this_cpu_ptr(&softnet_data);
3065 q->next_sched = NULL;
3066 *sd->output_queue_tailp = q;
3067 sd->output_queue_tailp = &q->next_sched;
3068 raise_softirq_irqoff(NET_TX_SOFTIRQ);
3069 local_irq_restore(flags);
3072 void __netif_schedule(struct Qdisc *q)
3074 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
3075 __netif_reschedule(q);
3077 EXPORT_SYMBOL(__netif_schedule);
3079 struct dev_kfree_skb_cb {
3080 enum skb_drop_reason reason;
3083 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb)
3085 return (struct dev_kfree_skb_cb *)skb->cb;
3088 void netif_schedule_queue(struct netdev_queue *txq)
3091 if (!netif_xmit_stopped(txq)) {
3092 struct Qdisc *q = rcu_dereference(txq->qdisc);
3094 __netif_schedule(q);
3098 EXPORT_SYMBOL(netif_schedule_queue);
3100 void netif_tx_wake_queue(struct netdev_queue *dev_queue)
3102 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) {
3106 q = rcu_dereference(dev_queue->qdisc);
3107 __netif_schedule(q);
3111 EXPORT_SYMBOL(netif_tx_wake_queue);
3113 void dev_kfree_skb_irq_reason(struct sk_buff *skb, enum skb_drop_reason reason)
3115 unsigned long flags;
3120 if (likely(refcount_read(&skb->users) == 1)) {
3122 refcount_set(&skb->users, 0);
3123 } else if (likely(!refcount_dec_and_test(&skb->users))) {
3126 get_kfree_skb_cb(skb)->reason = reason;
3127 local_irq_save(flags);
3128 skb->next = __this_cpu_read(softnet_data.completion_queue);
3129 __this_cpu_write(softnet_data.completion_queue, skb);
3130 raise_softirq_irqoff(NET_TX_SOFTIRQ);
3131 local_irq_restore(flags);
3133 EXPORT_SYMBOL(dev_kfree_skb_irq_reason);
3135 void dev_kfree_skb_any_reason(struct sk_buff *skb, enum skb_drop_reason reason)
3137 if (in_hardirq() || irqs_disabled())
3138 dev_kfree_skb_irq_reason(skb, reason);
3140 kfree_skb_reason(skb, reason);
3142 EXPORT_SYMBOL(dev_kfree_skb_any_reason);
3146 * netif_device_detach - mark device as removed
3147 * @dev: network device
3149 * Mark device as removed from system and therefore no longer available.
3151 void netif_device_detach(struct net_device *dev)
3153 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) &&
3154 netif_running(dev)) {
3155 netif_tx_stop_all_queues(dev);
3158 EXPORT_SYMBOL(netif_device_detach);
3161 * netif_device_attach - mark device as attached
3162 * @dev: network device
3164 * Mark device as attached from system and restart if needed.
3166 void netif_device_attach(struct net_device *dev)
3168 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) &&
3169 netif_running(dev)) {
3170 netif_tx_wake_all_queues(dev);
3171 __netdev_watchdog_up(dev);
3174 EXPORT_SYMBOL(netif_device_attach);
3177 * Returns a Tx hash based on the given packet descriptor a Tx queues' number
3178 * to be used as a distribution range.
3180 static u16 skb_tx_hash(const struct net_device *dev,
3181 const struct net_device *sb_dev,
3182 struct sk_buff *skb)
3186 u16 qcount = dev->real_num_tx_queues;
3189 u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
3191 qoffset = sb_dev->tc_to_txq[tc].offset;
3192 qcount = sb_dev->tc_to_txq[tc].count;
3193 if (unlikely(!qcount)) {
3194 net_warn_ratelimited("%s: invalid qcount, qoffset %u for tc %u\n",
3195 sb_dev->name, qoffset, tc);
3197 qcount = dev->real_num_tx_queues;
3201 if (skb_rx_queue_recorded(skb)) {
3202 hash = skb_get_rx_queue(skb);
3203 if (hash >= qoffset)
3205 while (unlikely(hash >= qcount))
3207 return hash + qoffset;
3210 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset;
3213 static void skb_warn_bad_offload(const struct sk_buff *skb)
3215 static const netdev_features_t null_features;
3216 struct net_device *dev = skb->dev;
3217 const char *name = "";
3219 if (!net_ratelimit())
3223 if (dev->dev.parent)
3224 name = dev_driver_string(dev->dev.parent);
3226 name = netdev_name(dev);
3228 skb_dump(KERN_WARNING, skb, false);
3229 WARN(1, "%s: caps=(%pNF, %pNF)\n",
3230 name, dev ? &dev->features : &null_features,
3231 skb->sk ? &skb->sk->sk_route_caps : &null_features);
3235 * Invalidate hardware checksum when packet is to be mangled, and
3236 * complete checksum manually on outgoing path.
3238 int skb_checksum_help(struct sk_buff *skb)
3241 int ret = 0, offset;
3243 if (skb->ip_summed == CHECKSUM_COMPLETE)
3244 goto out_set_summed;
3246 if (unlikely(skb_is_gso(skb))) {
3247 skb_warn_bad_offload(skb);
3251 /* Before computing a checksum, we should make sure no frag could
3252 * be modified by an external entity : checksum could be wrong.
3254 if (skb_has_shared_frag(skb)) {
3255 ret = __skb_linearize(skb);
3260 offset = skb_checksum_start_offset(skb);
3262 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) {
3263 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false);
3266 csum = skb_checksum(skb, offset, skb->len - offset, 0);
3268 offset += skb->csum_offset;
3269 if (WARN_ON_ONCE(offset + sizeof(__sum16) > skb_headlen(skb))) {
3270 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false);
3273 ret = skb_ensure_writable(skb, offset + sizeof(__sum16));
3277 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0;
3279 skb->ip_summed = CHECKSUM_NONE;
3283 EXPORT_SYMBOL(skb_checksum_help);
3285 int skb_crc32c_csum_help(struct sk_buff *skb)
3288 int ret = 0, offset, start;
3290 if (skb->ip_summed != CHECKSUM_PARTIAL)
3293 if (unlikely(skb_is_gso(skb)))
3296 /* Before computing a checksum, we should make sure no frag could
3297 * be modified by an external entity : checksum could be wrong.
3299 if (unlikely(skb_has_shared_frag(skb))) {
3300 ret = __skb_linearize(skb);
3304 start = skb_checksum_start_offset(skb);
3305 offset = start + offsetof(struct sctphdr, checksum);
3306 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) {
3311 ret = skb_ensure_writable(skb, offset + sizeof(__le32));
3315 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start,
3316 skb->len - start, ~(__u32)0,
3318 *(__le32 *)(skb->data + offset) = crc32c_csum;
3319 skb->ip_summed = CHECKSUM_NONE;
3320 skb->csum_not_inet = 0;
3325 __be16 skb_network_protocol(struct sk_buff *skb, int *depth)
3327 __be16 type = skb->protocol;
3329 /* Tunnel gso handlers can set protocol to ethernet. */
3330 if (type == htons(ETH_P_TEB)) {
3333 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr))))
3336 eth = (struct ethhdr *)skb->data;
3337 type = eth->h_proto;
3340 return __vlan_get_protocol(skb, type, depth);
3343 /* openvswitch calls this on rx path, so we need a different check.
3345 static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
3348 return skb->ip_summed != CHECKSUM_PARTIAL &&
3349 skb->ip_summed != CHECKSUM_UNNECESSARY;
3351 return skb->ip_summed == CHECKSUM_NONE;
3355 * __skb_gso_segment - Perform segmentation on skb.
3356 * @skb: buffer to segment
3357 * @features: features for the output path (see dev->features)
3358 * @tx_path: whether it is called in TX path
3360 * This function segments the given skb and returns a list of segments.
3362 * It may return NULL if the skb requires no segmentation. This is
3363 * only possible when GSO is used for verifying header integrity.
3365 * Segmentation preserves SKB_GSO_CB_OFFSET bytes of previous skb cb.
3367 struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
3368 netdev_features_t features, bool tx_path)
3370 struct sk_buff *segs;
3372 if (unlikely(skb_needs_check(skb, tx_path))) {
3375 /* We're going to init ->check field in TCP or UDP header */
3376 err = skb_cow_head(skb, 0);
3378 return ERR_PTR(err);
3381 /* Only report GSO partial support if it will enable us to
3382 * support segmentation on this frame without needing additional
3385 if (features & NETIF_F_GSO_PARTIAL) {
3386 netdev_features_t partial_features = NETIF_F_GSO_ROBUST;
3387 struct net_device *dev = skb->dev;
3389 partial_features |= dev->features & dev->gso_partial_features;
3390 if (!skb_gso_ok(skb, features | partial_features))
3391 features &= ~NETIF_F_GSO_PARTIAL;
3394 BUILD_BUG_ON(SKB_GSO_CB_OFFSET +
3395 sizeof(*SKB_GSO_CB(skb)) > sizeof(skb->cb));
3397 SKB_GSO_CB(skb)->mac_offset = skb_headroom(skb);
3398 SKB_GSO_CB(skb)->encap_level = 0;
3400 skb_reset_mac_header(skb);
3401 skb_reset_mac_len(skb);
3403 segs = skb_mac_gso_segment(skb, features);
3405 if (segs != skb && unlikely(skb_needs_check(skb, tx_path) && !IS_ERR(segs)))
3406 skb_warn_bad_offload(skb);
3410 EXPORT_SYMBOL(__skb_gso_segment);
3412 /* Take action when hardware reception checksum errors are detected. */
3414 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb)
3416 netdev_err(dev, "hw csum failure\n");
3417 skb_dump(KERN_ERR, skb, true);
3421 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb)
3423 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb);
3425 EXPORT_SYMBOL(netdev_rx_csum_fault);
3428 /* XXX: check that highmem exists at all on the given machine. */
3429 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
3431 #ifdef CONFIG_HIGHMEM
3434 if (!(dev->features & NETIF_F_HIGHDMA)) {
3435 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
3436 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
3438 if (PageHighMem(skb_frag_page(frag)))
3446 /* If MPLS offload request, verify we are testing hardware MPLS features
3447 * instead of standard features for the netdev.
3449 #if IS_ENABLED(CONFIG_NET_MPLS_GSO)
3450 static netdev_features_t net_mpls_features(struct sk_buff *skb,
3451 netdev_features_t features,
3454 if (eth_p_mpls(type))
3455 features &= skb->dev->mpls_features;
3460 static netdev_features_t net_mpls_features(struct sk_buff *skb,
3461 netdev_features_t features,
3468 static netdev_features_t harmonize_features(struct sk_buff *skb,
3469 netdev_features_t features)
3473 type = skb_network_protocol(skb, NULL);
3474 features = net_mpls_features(skb, features, type);
3476 if (skb->ip_summed != CHECKSUM_NONE &&
3477 !can_checksum_protocol(features, type)) {
3478 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
3480 if (illegal_highdma(skb->dev, skb))
3481 features &= ~NETIF_F_SG;
3486 netdev_features_t passthru_features_check(struct sk_buff *skb,
3487 struct net_device *dev,
3488 netdev_features_t features)
3492 EXPORT_SYMBOL(passthru_features_check);
3494 static netdev_features_t dflt_features_check(struct sk_buff *skb,
3495 struct net_device *dev,
3496 netdev_features_t features)
3498 return vlan_features_check(skb, features);
3501 static netdev_features_t gso_features_check(const struct sk_buff *skb,
3502 struct net_device *dev,
3503 netdev_features_t features)
3505 u16 gso_segs = skb_shinfo(skb)->gso_segs;
3507 if (gso_segs > READ_ONCE(dev->gso_max_segs))
3508 return features & ~NETIF_F_GSO_MASK;
3510 if (!skb_shinfo(skb)->gso_type) {
3511 skb_warn_bad_offload(skb);
3512 return features & ~NETIF_F_GSO_MASK;
3515 /* Support for GSO partial features requires software
3516 * intervention before we can actually process the packets
3517 * so we need to strip support for any partial features now
3518 * and we can pull them back in after we have partially
3519 * segmented the frame.
3521 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL))
3522 features &= ~dev->gso_partial_features;
3524 /* Make sure to clear the IPv4 ID mangling feature if the
3525 * IPv4 header has the potential to be fragmented.
3527 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
3528 struct iphdr *iph = skb->encapsulation ?
3529 inner_ip_hdr(skb) : ip_hdr(skb);
3531 if (!(iph->frag_off & htons(IP_DF)))
3532 features &= ~NETIF_F_TSO_MANGLEID;
3538 netdev_features_t netif_skb_features(struct sk_buff *skb)
3540 struct net_device *dev = skb->dev;
3541 netdev_features_t features = dev->features;
3543 if (skb_is_gso(skb))
3544 features = gso_features_check(skb, dev, features);
3546 /* If encapsulation offload request, verify we are testing
3547 * hardware encapsulation features instead of standard
3548 * features for the netdev
3550 if (skb->encapsulation)
3551 features &= dev->hw_enc_features;
3553 if (skb_vlan_tagged(skb))
3554 features = netdev_intersect_features(features,
3555 dev->vlan_features |
3556 NETIF_F_HW_VLAN_CTAG_TX |
3557 NETIF_F_HW_VLAN_STAG_TX);
3559 if (dev->netdev_ops->ndo_features_check)
3560 features &= dev->netdev_ops->ndo_features_check(skb, dev,
3563 features &= dflt_features_check(skb, dev, features);
3565 return harmonize_features(skb, features);
3567 EXPORT_SYMBOL(netif_skb_features);
3569 static int xmit_one(struct sk_buff *skb, struct net_device *dev,
3570 struct netdev_queue *txq, bool more)
3575 if (dev_nit_active(dev))
3576 dev_queue_xmit_nit(skb, dev);
3579 trace_net_dev_start_xmit(skb, dev);
3580 rc = netdev_start_xmit(skb, dev, txq, more);
3581 trace_net_dev_xmit(skb, rc, dev, len);
3586 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev,
3587 struct netdev_queue *txq, int *ret)
3589 struct sk_buff *skb = first;
3590 int rc = NETDEV_TX_OK;
3593 struct sk_buff *next = skb->next;
3595 skb_mark_not_on_list(skb);
3596 rc = xmit_one(skb, dev, txq, next != NULL);
3597 if (unlikely(!dev_xmit_complete(rc))) {
3603 if (netif_tx_queue_stopped(txq) && skb) {
3604 rc = NETDEV_TX_BUSY;
3614 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb,
3615 netdev_features_t features)
3617 if (skb_vlan_tag_present(skb) &&
3618 !vlan_hw_offload_capable(features, skb->vlan_proto))
3619 skb = __vlan_hwaccel_push_inside(skb);
3623 int skb_csum_hwoffload_help(struct sk_buff *skb,
3624 const netdev_features_t features)
3626 if (unlikely(skb_csum_is_sctp(skb)))
3627 return !!(features & NETIF_F_SCTP_CRC) ? 0 :
3628 skb_crc32c_csum_help(skb);
3630 if (features & NETIF_F_HW_CSUM)
3633 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) {
3634 switch (skb->csum_offset) {
3635 case offsetof(struct tcphdr, check):
3636 case offsetof(struct udphdr, check):
3641 return skb_checksum_help(skb);
3643 EXPORT_SYMBOL(skb_csum_hwoffload_help);
3645 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again)
3647 netdev_features_t features;
3649 features = netif_skb_features(skb);
3650 skb = validate_xmit_vlan(skb, features);
3654 skb = sk_validate_xmit_skb(skb, dev);
3658 if (netif_needs_gso(skb, features)) {
3659 struct sk_buff *segs;
3661 segs = skb_gso_segment(skb, features);
3669 if (skb_needs_linearize(skb, features) &&
3670 __skb_linearize(skb))
3673 /* If packet is not checksummed and device does not
3674 * support checksumming for this protocol, complete
3675 * checksumming here.
3677 if (skb->ip_summed == CHECKSUM_PARTIAL) {
3678 if (skb->encapsulation)
3679 skb_set_inner_transport_header(skb,
3680 skb_checksum_start_offset(skb));
3682 skb_set_transport_header(skb,
3683 skb_checksum_start_offset(skb));
3684 if (skb_csum_hwoffload_help(skb, features))
3689 skb = validate_xmit_xfrm(skb, features, again);
3696 dev_core_stats_tx_dropped_inc(dev);
3700 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again)
3702 struct sk_buff *next, *head = NULL, *tail;
3704 for (; skb != NULL; skb = next) {
3706 skb_mark_not_on_list(skb);
3708 /* in case skb wont be segmented, point to itself */
3711 skb = validate_xmit_skb(skb, dev, again);
3719 /* If skb was segmented, skb->prev points to
3720 * the last segment. If not, it still contains skb.
3726 EXPORT_SYMBOL_GPL(validate_xmit_skb_list);
3728 static void qdisc_pkt_len_init(struct sk_buff *skb)
3730 const struct skb_shared_info *shinfo = skb_shinfo(skb);
3732 qdisc_skb_cb(skb)->pkt_len = skb->len;
3734 /* To get more precise estimation of bytes sent on wire,
3735 * we add to pkt_len the headers size of all segments
3737 if (shinfo->gso_size && skb_transport_header_was_set(skb)) {
3738 u16 gso_segs = shinfo->gso_segs;
3739 unsigned int hdr_len;
3741 /* mac layer + network layer */
3742 hdr_len = skb_transport_offset(skb);
3744 /* + transport layer */
3745 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) {
3746 const struct tcphdr *th;
3747 struct tcphdr _tcphdr;
3749 th = skb_header_pointer(skb, hdr_len,
3750 sizeof(_tcphdr), &_tcphdr);
3752 hdr_len += __tcp_hdrlen(th);
3754 struct udphdr _udphdr;
3756 if (skb_header_pointer(skb, hdr_len,
3757 sizeof(_udphdr), &_udphdr))
3758 hdr_len += sizeof(struct udphdr);
3761 if (shinfo->gso_type & SKB_GSO_DODGY)
3762 gso_segs = DIV_ROUND_UP(skb->len - hdr_len,
3765 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len;
3769 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q,
3770 struct sk_buff **to_free,
3771 struct netdev_queue *txq)
3775 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK;
3776 if (rc == NET_XMIT_SUCCESS)
3777 trace_qdisc_enqueue(q, txq, skb);
3781 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
3782 struct net_device *dev,
3783 struct netdev_queue *txq)
3785 spinlock_t *root_lock = qdisc_lock(q);
3786 struct sk_buff *to_free = NULL;
3790 qdisc_calculate_pkt_len(skb, q);
3792 if (q->flags & TCQ_F_NOLOCK) {
3793 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) &&
3794 qdisc_run_begin(q)) {
3795 /* Retest nolock_qdisc_is_empty() within the protection
3796 * of q->seqlock to protect from racing with requeuing.
3798 if (unlikely(!nolock_qdisc_is_empty(q))) {
3799 rc = dev_qdisc_enqueue(skb, q, &to_free, txq);
3806 qdisc_bstats_cpu_update(q, skb);
3807 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) &&
3808 !nolock_qdisc_is_empty(q))
3812 return NET_XMIT_SUCCESS;
3815 rc = dev_qdisc_enqueue(skb, q, &to_free, txq);
3819 if (unlikely(to_free))
3820 kfree_skb_list_reason(to_free,
3821 SKB_DROP_REASON_QDISC_DROP);
3826 * Heuristic to force contended enqueues to serialize on a
3827 * separate lock before trying to get qdisc main lock.
3828 * This permits qdisc->running owner to get the lock more
3829 * often and dequeue packets faster.
3830 * On PREEMPT_RT it is possible to preempt the qdisc owner during xmit
3831 * and then other tasks will only enqueue packets. The packets will be
3832 * sent after the qdisc owner is scheduled again. To prevent this
3833 * scenario the task always serialize on the lock.
3835 contended = qdisc_is_running(q) || IS_ENABLED(CONFIG_PREEMPT_RT);
3836 if (unlikely(contended))
3837 spin_lock(&q->busylock);
3839 spin_lock(root_lock);
3840 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3841 __qdisc_drop(skb, &to_free);
3843 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
3844 qdisc_run_begin(q)) {
3846 * This is a work-conserving queue; there are no old skbs
3847 * waiting to be sent out; and the qdisc is not running -
3848 * xmit the skb directly.
3851 qdisc_bstats_update(q, skb);
3853 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) {
3854 if (unlikely(contended)) {
3855 spin_unlock(&q->busylock);
3862 rc = NET_XMIT_SUCCESS;
3864 rc = dev_qdisc_enqueue(skb, q, &to_free, txq);
3865 if (qdisc_run_begin(q)) {
3866 if (unlikely(contended)) {
3867 spin_unlock(&q->busylock);
3874 spin_unlock(root_lock);
3875 if (unlikely(to_free))
3876 kfree_skb_list_reason(to_free, SKB_DROP_REASON_QDISC_DROP);
3877 if (unlikely(contended))
3878 spin_unlock(&q->busylock);
3882 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO)
3883 static void skb_update_prio(struct sk_buff *skb)
3885 const struct netprio_map *map;
3886 const struct sock *sk;
3887 unsigned int prioidx;
3891 map = rcu_dereference_bh(skb->dev->priomap);
3894 sk = skb_to_full_sk(skb);
3898 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data);
3900 if (prioidx < map->priomap_len)
3901 skb->priority = map->priomap[prioidx];
3904 #define skb_update_prio(skb)
3908 * dev_loopback_xmit - loop back @skb
3909 * @net: network namespace this loopback is happening in
3910 * @sk: sk needed to be a netfilter okfn
3911 * @skb: buffer to transmit
3913 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb)
3915 skb_reset_mac_header(skb);
3916 __skb_pull(skb, skb_network_offset(skb));
3917 skb->pkt_type = PACKET_LOOPBACK;
3918 if (skb->ip_summed == CHECKSUM_NONE)
3919 skb->ip_summed = CHECKSUM_UNNECESSARY;
3920 DEBUG_NET_WARN_ON_ONCE(!skb_dst(skb));
3925 EXPORT_SYMBOL(dev_loopback_xmit);
3927 #ifdef CONFIG_NET_EGRESS
3928 static struct sk_buff *
3929 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
3931 #ifdef CONFIG_NET_CLS_ACT
3932 struct mini_Qdisc *miniq = rcu_dereference_bh(dev->miniq_egress);
3933 struct tcf_result cl_res;
3938 /* qdisc_skb_cb(skb)->pkt_len was already set by the caller. */
3939 tc_skb_cb(skb)->mru = 0;
3940 tc_skb_cb(skb)->post_ct = false;
3941 mini_qdisc_bstats_cpu_update(miniq, skb);
3943 switch (tcf_classify(skb, miniq->block, miniq->filter_list, &cl_res, false)) {
3945 case TC_ACT_RECLASSIFY:
3946 skb->tc_index = TC_H_MIN(cl_res.classid);
3949 mini_qdisc_qstats_cpu_drop(miniq);
3950 *ret = NET_XMIT_DROP;
3951 kfree_skb_reason(skb, SKB_DROP_REASON_TC_EGRESS);
3956 *ret = NET_XMIT_SUCCESS;
3959 case TC_ACT_REDIRECT:
3960 /* No need to push/pop skb's mac_header here on egress! */
3961 skb_do_redirect(skb);
3962 *ret = NET_XMIT_SUCCESS;
3967 #endif /* CONFIG_NET_CLS_ACT */
3972 static struct netdev_queue *
3973 netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb)
3975 int qm = skb_get_queue_mapping(skb);
3977 return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm));
3980 static bool netdev_xmit_txqueue_skipped(void)
3982 return __this_cpu_read(softnet_data.xmit.skip_txqueue);
3985 void netdev_xmit_skip_txqueue(bool skip)
3987 __this_cpu_write(softnet_data.xmit.skip_txqueue, skip);
3989 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue);
3990 #endif /* CONFIG_NET_EGRESS */
3993 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb,
3994 struct xps_dev_maps *dev_maps, unsigned int tci)
3996 int tc = netdev_get_prio_tc_map(dev, skb->priority);
3997 struct xps_map *map;
3998 int queue_index = -1;
4000 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids)
4003 tci *= dev_maps->num_tc;
4006 map = rcu_dereference(dev_maps->attr_map[tci]);
4009 queue_index = map->queues[0];
4011 queue_index = map->queues[reciprocal_scale(
4012 skb_get_hash(skb), map->len)];
4013 if (unlikely(queue_index >= dev->real_num_tx_queues))
4020 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev,
4021 struct sk_buff *skb)
4024 struct xps_dev_maps *dev_maps;
4025 struct sock *sk = skb->sk;
4026 int queue_index = -1;
4028 if (!static_key_false(&xps_needed))
4032 if (!static_key_false(&xps_rxqs_needed))
4035 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]);
4037 int tci = sk_rx_queue_get(sk);
4040 queue_index = __get_xps_queue_idx(dev, skb, dev_maps,
4045 if (queue_index < 0) {
4046 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]);
4048 unsigned int tci = skb->sender_cpu - 1;
4050 queue_index = __get_xps_queue_idx(dev, skb, dev_maps,
4062 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb,
4063 struct net_device *sb_dev)
4067 EXPORT_SYMBOL(dev_pick_tx_zero);
4069 u16 dev_pick_tx_cpu_id(struct net_device *dev, struct sk_buff *skb,
4070 struct net_device *sb_dev)
4072 return (u16)raw_smp_processor_id() % dev->real_num_tx_queues;
4074 EXPORT_SYMBOL(dev_pick_tx_cpu_id);
4076 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb,
4077 struct net_device *sb_dev)
4079 struct sock *sk = skb->sk;
4080 int queue_index = sk_tx_queue_get(sk);
4082 sb_dev = sb_dev ? : dev;
4084 if (queue_index < 0 || skb->ooo_okay ||
4085 queue_index >= dev->real_num_tx_queues) {
4086 int new_index = get_xps_queue(dev, sb_dev, skb);
4089 new_index = skb_tx_hash(dev, sb_dev, skb);
4091 if (queue_index != new_index && sk &&
4093 rcu_access_pointer(sk->sk_dst_cache))
4094 sk_tx_queue_set(sk, new_index);
4096 queue_index = new_index;
4101 EXPORT_SYMBOL(netdev_pick_tx);
4103 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev,
4104 struct sk_buff *skb,
4105 struct net_device *sb_dev)
4107 int queue_index = 0;
4110 u32 sender_cpu = skb->sender_cpu - 1;
4112 if (sender_cpu >= (u32)NR_CPUS)
4113 skb->sender_cpu = raw_smp_processor_id() + 1;
4116 if (dev->real_num_tx_queues != 1) {
4117 const struct net_device_ops *ops = dev->netdev_ops;
4119 if (ops->ndo_select_queue)
4120 queue_index = ops->ndo_select_queue(dev, skb, sb_dev);
4122 queue_index = netdev_pick_tx(dev, skb, sb_dev);
4124 queue_index = netdev_cap_txqueue(dev, queue_index);
4127 skb_set_queue_mapping(skb, queue_index);
4128 return netdev_get_tx_queue(dev, queue_index);
4132 * __dev_queue_xmit() - transmit a buffer
4133 * @skb: buffer to transmit
4134 * @sb_dev: suboordinate device used for L2 forwarding offload
4136 * Queue a buffer for transmission to a network device. The caller must
4137 * have set the device and priority and built the buffer before calling
4138 * this function. The function can be called from an interrupt.
4140 * When calling this method, interrupts MUST be enabled. This is because
4141 * the BH enable code must have IRQs enabled so that it will not deadlock.
4143 * Regardless of the return value, the skb is consumed, so it is currently
4144 * difficult to retry a send to this method. (You can bump the ref count
4145 * before sending to hold a reference for retry if you are careful.)
4148 * * 0 - buffer successfully transmitted
4149 * * positive qdisc return code - NET_XMIT_DROP etc.
4150 * * negative errno - other errors
4152 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev)
4154 struct net_device *dev = skb->dev;
4155 struct netdev_queue *txq = NULL;
4160 skb_reset_mac_header(skb);
4161 skb_assert_len(skb);
4163 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP))
4164 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED);
4166 /* Disable soft irqs for various locks below. Also
4167 * stops preemption for RCU.
4171 skb_update_prio(skb);
4173 qdisc_pkt_len_init(skb);
4174 #ifdef CONFIG_NET_CLS_ACT
4175 skb->tc_at_ingress = 0;
4177 #ifdef CONFIG_NET_EGRESS
4178 if (static_branch_unlikely(&egress_needed_key)) {
4179 if (nf_hook_egress_active()) {
4180 skb = nf_hook_egress(skb, &rc, dev);
4185 netdev_xmit_skip_txqueue(false);
4187 nf_skip_egress(skb, true);
4188 skb = sch_handle_egress(skb, &rc, dev);
4191 nf_skip_egress(skb, false);
4193 if (netdev_xmit_txqueue_skipped())
4194 txq = netdev_tx_queue_mapping(dev, skb);
4197 /* If device/qdisc don't need skb->dst, release it right now while
4198 * its hot in this cpu cache.
4200 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
4206 txq = netdev_core_pick_tx(dev, skb, sb_dev);
4208 q = rcu_dereference_bh(txq->qdisc);
4210 trace_net_dev_queue(skb);
4212 rc = __dev_xmit_skb(skb, q, dev, txq);
4216 /* The device has no queue. Common case for software devices:
4217 * loopback, all the sorts of tunnels...
4219 * Really, it is unlikely that netif_tx_lock protection is necessary
4220 * here. (f.e. loopback and IP tunnels are clean ignoring statistics
4222 * However, it is possible, that they rely on protection
4225 * Check this and shot the lock. It is not prone from deadlocks.
4226 *Either shot noqueue qdisc, it is even simpler 8)
4228 if (dev->flags & IFF_UP) {
4229 int cpu = smp_processor_id(); /* ok because BHs are off */
4231 /* Other cpus might concurrently change txq->xmit_lock_owner
4232 * to -1 or to their cpu id, but not to our id.
4234 if (READ_ONCE(txq->xmit_lock_owner) != cpu) {
4235 if (dev_xmit_recursion())
4236 goto recursion_alert;
4238 skb = validate_xmit_skb(skb, dev, &again);
4242 HARD_TX_LOCK(dev, txq, cpu);
4244 if (!netif_xmit_stopped(txq)) {
4245 dev_xmit_recursion_inc();
4246 skb = dev_hard_start_xmit(skb, dev, txq, &rc);
4247 dev_xmit_recursion_dec();
4248 if (dev_xmit_complete(rc)) {
4249 HARD_TX_UNLOCK(dev, txq);
4253 HARD_TX_UNLOCK(dev, txq);
4254 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
4257 /* Recursion is detected! It is possible,
4261 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
4267 rcu_read_unlock_bh();
4269 dev_core_stats_tx_dropped_inc(dev);
4270 kfree_skb_list(skb);
4273 rcu_read_unlock_bh();
4276 EXPORT_SYMBOL(__dev_queue_xmit);
4278 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id)
4280 struct net_device *dev = skb->dev;
4281 struct sk_buff *orig_skb = skb;
4282 struct netdev_queue *txq;
4283 int ret = NETDEV_TX_BUSY;
4286 if (unlikely(!netif_running(dev) ||
4287 !netif_carrier_ok(dev)))
4290 skb = validate_xmit_skb_list(skb, dev, &again);
4291 if (skb != orig_skb)
4294 skb_set_queue_mapping(skb, queue_id);
4295 txq = skb_get_tx_queue(dev, skb);
4299 dev_xmit_recursion_inc();
4300 HARD_TX_LOCK(dev, txq, smp_processor_id());
4301 if (!netif_xmit_frozen_or_drv_stopped(txq))
4302 ret = netdev_start_xmit(skb, dev, txq, false);
4303 HARD_TX_UNLOCK(dev, txq);
4304 dev_xmit_recursion_dec();
4309 dev_core_stats_tx_dropped_inc(dev);
4310 kfree_skb_list(skb);
4311 return NET_XMIT_DROP;
4313 EXPORT_SYMBOL(__dev_direct_xmit);
4315 /*************************************************************************
4317 *************************************************************************/
4319 int netdev_max_backlog __read_mostly = 1000;
4320 EXPORT_SYMBOL(netdev_max_backlog);
4322 int netdev_tstamp_prequeue __read_mostly = 1;
4323 unsigned int sysctl_skb_defer_max __read_mostly = 64;
4324 int netdev_budget __read_mostly = 300;
4325 /* Must be at least 2 jiffes to guarantee 1 jiffy timeout */
4326 unsigned int __read_mostly netdev_budget_usecs = 2 * USEC_PER_SEC / HZ;
4327 int weight_p __read_mostly = 64; /* old backlog weight */
4328 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */
4329 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */
4330 int dev_rx_weight __read_mostly = 64;
4331 int dev_tx_weight __read_mostly = 64;
4333 /* Called with irq disabled */
4334 static inline void ____napi_schedule(struct softnet_data *sd,
4335 struct napi_struct *napi)
4337 struct task_struct *thread;
4339 lockdep_assert_irqs_disabled();
4341 if (test_bit(NAPI_STATE_THREADED, &napi->state)) {
4342 /* Paired with smp_mb__before_atomic() in
4343 * napi_enable()/dev_set_threaded().
4344 * Use READ_ONCE() to guarantee a complete
4345 * read on napi->thread. Only call
4346 * wake_up_process() when it's not NULL.
4348 thread = READ_ONCE(napi->thread);
4350 /* Avoid doing set_bit() if the thread is in
4351 * INTERRUPTIBLE state, cause napi_thread_wait()
4352 * makes sure to proceed with napi polling
4353 * if the thread is explicitly woken from here.
4355 if (READ_ONCE(thread->__state) != TASK_INTERRUPTIBLE)
4356 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state);
4357 wake_up_process(thread);
4362 list_add_tail(&napi->poll_list, &sd->poll_list);
4363 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
4368 /* One global table that all flow-based protocols share. */
4369 struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
4370 EXPORT_SYMBOL(rps_sock_flow_table);
4371 u32 rps_cpu_mask __read_mostly;
4372 EXPORT_SYMBOL(rps_cpu_mask);
4374 struct static_key_false rps_needed __read_mostly;
4375 EXPORT_SYMBOL(rps_needed);
4376 struct static_key_false rfs_needed __read_mostly;
4377 EXPORT_SYMBOL(rfs_needed);
4379 static struct rps_dev_flow *
4380 set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
4381 struct rps_dev_flow *rflow, u16 next_cpu)
4383 if (next_cpu < nr_cpu_ids) {
4384 #ifdef CONFIG_RFS_ACCEL
4385 struct netdev_rx_queue *rxqueue;
4386 struct rps_dev_flow_table *flow_table;
4387 struct rps_dev_flow *old_rflow;
4392 /* Should we steer this flow to a different hardware queue? */
4393 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
4394 !(dev->features & NETIF_F_NTUPLE))
4396 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
4397 if (rxq_index == skb_get_rx_queue(skb))
4400 rxqueue = dev->_rx + rxq_index;
4401 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4404 flow_id = skb_get_hash(skb) & flow_table->mask;
4405 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
4406 rxq_index, flow_id);
4410 rflow = &flow_table->flows[flow_id];
4412 if (old_rflow->filter == rflow->filter)
4413 old_rflow->filter = RPS_NO_FILTER;
4417 per_cpu(softnet_data, next_cpu).input_queue_head;
4420 rflow->cpu = next_cpu;
4425 * get_rps_cpu is called from netif_receive_skb and returns the target
4426 * CPU from the RPS map of the receiving queue for a given skb.
4427 * rcu_read_lock must be held on entry.
4429 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
4430 struct rps_dev_flow **rflowp)
4432 const struct rps_sock_flow_table *sock_flow_table;
4433 struct netdev_rx_queue *rxqueue = dev->_rx;
4434 struct rps_dev_flow_table *flow_table;
4435 struct rps_map *map;
4440 if (skb_rx_queue_recorded(skb)) {
4441 u16 index = skb_get_rx_queue(skb);
4443 if (unlikely(index >= dev->real_num_rx_queues)) {
4444 WARN_ONCE(dev->real_num_rx_queues > 1,
4445 "%s received packet on queue %u, but number "
4446 "of RX queues is %u\n",
4447 dev->name, index, dev->real_num_rx_queues);
4453 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */
4455 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4456 map = rcu_dereference(rxqueue->rps_map);
4457 if (!flow_table && !map)
4460 skb_reset_network_header(skb);
4461 hash = skb_get_hash(skb);
4465 sock_flow_table = rcu_dereference(rps_sock_flow_table);
4466 if (flow_table && sock_flow_table) {
4467 struct rps_dev_flow *rflow;
4471 /* First check into global flow table if there is a match */
4472 ident = sock_flow_table->ents[hash & sock_flow_table->mask];
4473 if ((ident ^ hash) & ~rps_cpu_mask)
4476 next_cpu = ident & rps_cpu_mask;
4478 /* OK, now we know there is a match,
4479 * we can look at the local (per receive queue) flow table
4481 rflow = &flow_table->flows[hash & flow_table->mask];
4485 * If the desired CPU (where last recvmsg was done) is
4486 * different from current CPU (one in the rx-queue flow
4487 * table entry), switch if one of the following holds:
4488 * - Current CPU is unset (>= nr_cpu_ids).
4489 * - Current CPU is offline.
4490 * - The current CPU's queue tail has advanced beyond the
4491 * last packet that was enqueued using this table entry.
4492 * This guarantees that all previous packets for the flow
4493 * have been dequeued, thus preserving in order delivery.
4495 if (unlikely(tcpu != next_cpu) &&
4496 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) ||
4497 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
4498 rflow->last_qtail)) >= 0)) {
4500 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
4503 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) {
4513 tcpu = map->cpus[reciprocal_scale(hash, map->len)];
4514 if (cpu_online(tcpu)) {
4524 #ifdef CONFIG_RFS_ACCEL
4527 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
4528 * @dev: Device on which the filter was set
4529 * @rxq_index: RX queue index
4530 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
4531 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
4533 * Drivers that implement ndo_rx_flow_steer() should periodically call
4534 * this function for each installed filter and remove the filters for
4535 * which it returns %true.
4537 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
4538 u32 flow_id, u16 filter_id)
4540 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
4541 struct rps_dev_flow_table *flow_table;
4542 struct rps_dev_flow *rflow;
4547 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4548 if (flow_table && flow_id <= flow_table->mask) {
4549 rflow = &flow_table->flows[flow_id];
4550 cpu = READ_ONCE(rflow->cpu);
4551 if (rflow->filter == filter_id && cpu < nr_cpu_ids &&
4552 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
4553 rflow->last_qtail) <
4554 (int)(10 * flow_table->mask)))
4560 EXPORT_SYMBOL(rps_may_expire_flow);
4562 #endif /* CONFIG_RFS_ACCEL */
4564 /* Called from hardirq (IPI) context */
4565 static void rps_trigger_softirq(void *data)
4567 struct softnet_data *sd = data;
4569 ____napi_schedule(sd, &sd->backlog);
4573 #endif /* CONFIG_RPS */
4575 /* Called from hardirq (IPI) context */
4576 static void trigger_rx_softirq(void *data)
4578 struct softnet_data *sd = data;
4580 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
4581 smp_store_release(&sd->defer_ipi_scheduled, 0);
4585 * After we queued a packet into sd->input_pkt_queue,
4586 * we need to make sure this queue is serviced soon.
4588 * - If this is another cpu queue, link it to our rps_ipi_list,
4589 * and make sure we will process rps_ipi_list from net_rx_action().
4590 * As we do not know yet if we are called from net_rx_action(),
4591 * we have to raise NET_RX_SOFTIRQ. This might change in the future.
4593 * - If this is our own queue, NAPI schedule our backlog.
4594 * Note that this also raises NET_RX_SOFTIRQ.
4596 static void napi_schedule_rps(struct softnet_data *sd)
4598 struct softnet_data *mysd = this_cpu_ptr(&softnet_data);
4602 sd->rps_ipi_next = mysd->rps_ipi_list;
4603 mysd->rps_ipi_list = sd;
4605 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
4608 #endif /* CONFIG_RPS */
4609 __napi_schedule_irqoff(&mysd->backlog);
4612 #ifdef CONFIG_NET_FLOW_LIMIT
4613 int netdev_flow_limit_table_len __read_mostly = (1 << 12);
4616 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
4618 #ifdef CONFIG_NET_FLOW_LIMIT
4619 struct sd_flow_limit *fl;
4620 struct softnet_data *sd;
4621 unsigned int old_flow, new_flow;
4623 if (qlen < (READ_ONCE(netdev_max_backlog) >> 1))
4626 sd = this_cpu_ptr(&softnet_data);
4629 fl = rcu_dereference(sd->flow_limit);
4631 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1);
4632 old_flow = fl->history[fl->history_head];
4633 fl->history[fl->history_head] = new_flow;
4636 fl->history_head &= FLOW_LIMIT_HISTORY - 1;
4638 if (likely(fl->buckets[old_flow]))
4639 fl->buckets[old_flow]--;
4641 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) {
4653 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
4654 * queue (may be a remote CPU queue).
4656 static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
4657 unsigned int *qtail)
4659 enum skb_drop_reason reason;
4660 struct softnet_data *sd;
4661 unsigned long flags;
4664 reason = SKB_DROP_REASON_NOT_SPECIFIED;
4665 sd = &per_cpu(softnet_data, cpu);
4667 rps_lock_irqsave(sd, &flags);
4668 if (!netif_running(skb->dev))
4670 qlen = skb_queue_len(&sd->input_pkt_queue);
4671 if (qlen <= READ_ONCE(netdev_max_backlog) && !skb_flow_limit(skb, qlen)) {
4674 __skb_queue_tail(&sd->input_pkt_queue, skb);
4675 input_queue_tail_incr_save(sd, qtail);
4676 rps_unlock_irq_restore(sd, &flags);
4677 return NET_RX_SUCCESS;
4680 /* Schedule NAPI for backlog device
4681 * We can use non atomic operation since we own the queue lock
4683 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state))
4684 napi_schedule_rps(sd);
4687 reason = SKB_DROP_REASON_CPU_BACKLOG;
4691 rps_unlock_irq_restore(sd, &flags);
4693 dev_core_stats_rx_dropped_inc(skb->dev);
4694 kfree_skb_reason(skb, reason);
4698 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb)
4700 struct net_device *dev = skb->dev;
4701 struct netdev_rx_queue *rxqueue;
4705 if (skb_rx_queue_recorded(skb)) {
4706 u16 index = skb_get_rx_queue(skb);
4708 if (unlikely(index >= dev->real_num_rx_queues)) {
4709 WARN_ONCE(dev->real_num_rx_queues > 1,
4710 "%s received packet on queue %u, but number "
4711 "of RX queues is %u\n",
4712 dev->name, index, dev->real_num_rx_queues);
4714 return rxqueue; /* Return first rxqueue */
4721 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp,
4722 struct bpf_prog *xdp_prog)
4724 void *orig_data, *orig_data_end, *hard_start;
4725 struct netdev_rx_queue *rxqueue;
4726 bool orig_bcast, orig_host;
4727 u32 mac_len, frame_sz;
4728 __be16 orig_eth_type;
4733 /* The XDP program wants to see the packet starting at the MAC
4736 mac_len = skb->data - skb_mac_header(skb);
4737 hard_start = skb->data - skb_headroom(skb);
4739 /* SKB "head" area always have tailroom for skb_shared_info */
4740 frame_sz = (void *)skb_end_pointer(skb) - hard_start;
4741 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
4743 rxqueue = netif_get_rxqueue(skb);
4744 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq);
4745 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len,
4746 skb_headlen(skb) + mac_len, true);
4748 orig_data_end = xdp->data_end;
4749 orig_data = xdp->data;
4750 eth = (struct ethhdr *)xdp->data;
4751 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr);
4752 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest);
4753 orig_eth_type = eth->h_proto;
4755 act = bpf_prog_run_xdp(xdp_prog, xdp);
4757 /* check if bpf_xdp_adjust_head was used */
4758 off = xdp->data - orig_data;
4761 __skb_pull(skb, off);
4763 __skb_push(skb, -off);
4765 skb->mac_header += off;
4766 skb_reset_network_header(skb);
4769 /* check if bpf_xdp_adjust_tail was used */
4770 off = xdp->data_end - orig_data_end;
4772 skb_set_tail_pointer(skb, xdp->data_end - xdp->data);
4773 skb->len += off; /* positive on grow, negative on shrink */
4776 /* check if XDP changed eth hdr such SKB needs update */
4777 eth = (struct ethhdr *)xdp->data;
4778 if ((orig_eth_type != eth->h_proto) ||
4779 (orig_host != ether_addr_equal_64bits(eth->h_dest,
4780 skb->dev->dev_addr)) ||
4781 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) {
4782 __skb_push(skb, ETH_HLEN);
4783 skb->pkt_type = PACKET_HOST;
4784 skb->protocol = eth_type_trans(skb, skb->dev);
4787 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull
4788 * before calling us again on redirect path. We do not call do_redirect
4789 * as we leave that up to the caller.
4791 * Caller is responsible for managing lifetime of skb (i.e. calling
4792 * kfree_skb in response to actions it cannot handle/XDP_DROP).
4797 __skb_push(skb, mac_len);
4800 metalen = xdp->data - xdp->data_meta;
4802 skb_metadata_set(skb, metalen);
4809 static u32 netif_receive_generic_xdp(struct sk_buff *skb,
4810 struct xdp_buff *xdp,
4811 struct bpf_prog *xdp_prog)
4815 /* Reinjected packets coming from act_mirred or similar should
4816 * not get XDP generic processing.
4818 if (skb_is_redirected(skb))
4821 /* XDP packets must be linear and must have sufficient headroom
4822 * of XDP_PACKET_HEADROOM bytes. This is the guarantee that also
4823 * native XDP provides, thus we need to do it here as well.
4825 if (skb_cloned(skb) || skb_is_nonlinear(skb) ||
4826 skb_headroom(skb) < XDP_PACKET_HEADROOM) {
4827 int hroom = XDP_PACKET_HEADROOM - skb_headroom(skb);
4828 int troom = skb->tail + skb->data_len - skb->end;
4830 /* In case we have to go down the path and also linearize,
4831 * then lets do the pskb_expand_head() work just once here.
4833 if (pskb_expand_head(skb,
4834 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0,
4835 troom > 0 ? troom + 128 : 0, GFP_ATOMIC))
4837 if (skb_linearize(skb))
4841 act = bpf_prog_run_generic_xdp(skb, xdp, xdp_prog);
4848 bpf_warn_invalid_xdp_action(skb->dev, xdp_prog, act);
4851 trace_xdp_exception(skb->dev, xdp_prog, act);
4862 /* When doing generic XDP we have to bypass the qdisc layer and the
4863 * network taps in order to match in-driver-XDP behavior. This also means
4864 * that XDP packets are able to starve other packets going through a qdisc,
4865 * and DDOS attacks will be more effective. In-driver-XDP use dedicated TX
4866 * queues, so they do not have this starvation issue.
4868 void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog)
4870 struct net_device *dev = skb->dev;
4871 struct netdev_queue *txq;
4872 bool free_skb = true;
4875 txq = netdev_core_pick_tx(dev, skb, NULL);
4876 cpu = smp_processor_id();
4877 HARD_TX_LOCK(dev, txq, cpu);
4878 if (!netif_xmit_frozen_or_drv_stopped(txq)) {
4879 rc = netdev_start_xmit(skb, dev, txq, 0);
4880 if (dev_xmit_complete(rc))
4883 HARD_TX_UNLOCK(dev, txq);
4885 trace_xdp_exception(dev, xdp_prog, XDP_TX);
4886 dev_core_stats_tx_dropped_inc(dev);
4891 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key);
4893 int do_xdp_generic(struct bpf_prog *xdp_prog, struct sk_buff *skb)
4896 struct xdp_buff xdp;
4900 act = netif_receive_generic_xdp(skb, &xdp, xdp_prog);
4901 if (act != XDP_PASS) {
4904 err = xdp_do_generic_redirect(skb->dev, skb,
4910 generic_xdp_tx(skb, xdp_prog);
4918 kfree_skb_reason(skb, SKB_DROP_REASON_XDP);
4921 EXPORT_SYMBOL_GPL(do_xdp_generic);
4923 static int netif_rx_internal(struct sk_buff *skb)
4927 net_timestamp_check(READ_ONCE(netdev_tstamp_prequeue), skb);
4929 trace_netif_rx(skb);
4932 if (static_branch_unlikely(&rps_needed)) {
4933 struct rps_dev_flow voidflow, *rflow = &voidflow;
4938 cpu = get_rps_cpu(skb->dev, skb, &rflow);
4940 cpu = smp_processor_id();
4942 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4950 ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail);
4956 * __netif_rx - Slightly optimized version of netif_rx
4957 * @skb: buffer to post
4959 * This behaves as netif_rx except that it does not disable bottom halves.
4960 * As a result this function may only be invoked from the interrupt context
4961 * (either hard or soft interrupt).
4963 int __netif_rx(struct sk_buff *skb)
4967 lockdep_assert_once(hardirq_count() | softirq_count());
4969 trace_netif_rx_entry(skb);
4970 ret = netif_rx_internal(skb);
4971 trace_netif_rx_exit(ret);
4974 EXPORT_SYMBOL(__netif_rx);
4977 * netif_rx - post buffer to the network code
4978 * @skb: buffer to post
4980 * This function receives a packet from a device driver and queues it for
4981 * the upper (protocol) levels to process via the backlog NAPI device. It
4982 * always succeeds. The buffer may be dropped during processing for
4983 * congestion control or by the protocol layers.
4984 * The network buffer is passed via the backlog NAPI device. Modern NIC
4985 * driver should use NAPI and GRO.
4986 * This function can used from interrupt and from process context. The
4987 * caller from process context must not disable interrupts before invoking
4991 * NET_RX_SUCCESS (no congestion)
4992 * NET_RX_DROP (packet was dropped)
4995 int netif_rx(struct sk_buff *skb)
4997 bool need_bh_off = !(hardirq_count() | softirq_count());
5002 trace_netif_rx_entry(skb);
5003 ret = netif_rx_internal(skb);
5004 trace_netif_rx_exit(ret);
5009 EXPORT_SYMBOL(netif_rx);
5011 static __latent_entropy void net_tx_action(struct softirq_action *h)
5013 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
5015 if (sd->completion_queue) {
5016 struct sk_buff *clist;
5018 local_irq_disable();
5019 clist = sd->completion_queue;
5020 sd->completion_queue = NULL;
5024 struct sk_buff *skb = clist;
5026 clist = clist->next;
5028 WARN_ON(refcount_read(&skb->users));
5029 if (likely(get_kfree_skb_cb(skb)->reason == SKB_CONSUMED))
5030 trace_consume_skb(skb, net_tx_action);
5032 trace_kfree_skb(skb, net_tx_action,
5033 get_kfree_skb_cb(skb)->reason);
5035 if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
5038 __kfree_skb_defer(skb);
5042 if (sd->output_queue) {
5045 local_irq_disable();
5046 head = sd->output_queue;
5047 sd->output_queue = NULL;
5048 sd->output_queue_tailp = &sd->output_queue;
5054 struct Qdisc *q = head;
5055 spinlock_t *root_lock = NULL;
5057 head = head->next_sched;
5059 /* We need to make sure head->next_sched is read
5060 * before clearing __QDISC_STATE_SCHED
5062 smp_mb__before_atomic();
5064 if (!(q->flags & TCQ_F_NOLOCK)) {
5065 root_lock = qdisc_lock(q);
5066 spin_lock(root_lock);
5067 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED,
5069 /* There is a synchronize_net() between
5070 * STATE_DEACTIVATED flag being set and
5071 * qdisc_reset()/some_qdisc_is_busy() in
5072 * dev_deactivate(), so we can safely bail out
5073 * early here to avoid data race between
5074 * qdisc_deactivate() and some_qdisc_is_busy()
5075 * for lockless qdisc.
5077 clear_bit(__QDISC_STATE_SCHED, &q->state);
5081 clear_bit(__QDISC_STATE_SCHED, &q->state);
5084 spin_unlock(root_lock);
5090 xfrm_dev_backlog(sd);
5093 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE)
5094 /* This hook is defined here for ATM LANE */
5095 int (*br_fdb_test_addr_hook)(struct net_device *dev,
5096 unsigned char *addr) __read_mostly;
5097 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook);
5100 static inline struct sk_buff *
5101 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
5102 struct net_device *orig_dev, bool *another)
5104 #ifdef CONFIG_NET_CLS_ACT
5105 struct mini_Qdisc *miniq = rcu_dereference_bh(skb->dev->miniq_ingress);
5106 struct tcf_result cl_res;
5108 /* If there's at least one ingress present somewhere (so
5109 * we get here via enabled static key), remaining devices
5110 * that are not configured with an ingress qdisc will bail
5117 *ret = deliver_skb(skb, *pt_prev, orig_dev);
5121 qdisc_skb_cb(skb)->pkt_len = skb->len;
5122 tc_skb_cb(skb)->mru = 0;
5123 tc_skb_cb(skb)->post_ct = false;
5124 skb->tc_at_ingress = 1;
5125 mini_qdisc_bstats_cpu_update(miniq, skb);
5127 switch (tcf_classify(skb, miniq->block, miniq->filter_list, &cl_res, false)) {
5129 case TC_ACT_RECLASSIFY:
5130 skb->tc_index = TC_H_MIN(cl_res.classid);
5133 mini_qdisc_qstats_cpu_drop(miniq);
5134 kfree_skb_reason(skb, SKB_DROP_REASON_TC_INGRESS);
5141 *ret = NET_RX_SUCCESS;
5143 case TC_ACT_REDIRECT:
5144 /* skb_mac_header check was done by cls/act_bpf, so
5145 * we can safely push the L2 header back before
5146 * redirecting to another netdev
5148 __skb_push(skb, skb->mac_len);
5149 if (skb_do_redirect(skb) == -EAGAIN) {
5150 __skb_pull(skb, skb->mac_len);
5154 *ret = NET_RX_SUCCESS;
5156 case TC_ACT_CONSUMED:
5157 *ret = NET_RX_SUCCESS;
5162 #endif /* CONFIG_NET_CLS_ACT */
5167 * netdev_is_rx_handler_busy - check if receive handler is registered
5168 * @dev: device to check
5170 * Check if a receive handler is already registered for a given device.
5171 * Return true if there one.
5173 * The caller must hold the rtnl_mutex.
5175 bool netdev_is_rx_handler_busy(struct net_device *dev)
5178 return dev && rtnl_dereference(dev->rx_handler);
5180 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy);
5183 * netdev_rx_handler_register - register receive handler
5184 * @dev: device to register a handler for
5185 * @rx_handler: receive handler to register
5186 * @rx_handler_data: data pointer that is used by rx handler
5188 * Register a receive handler for a device. This handler will then be
5189 * called from __netif_receive_skb. A negative errno code is returned
5192 * The caller must hold the rtnl_mutex.
5194 * For a general description of rx_handler, see enum rx_handler_result.
5196 int netdev_rx_handler_register(struct net_device *dev,
5197 rx_handler_func_t *rx_handler,
5198 void *rx_handler_data)
5200 if (netdev_is_rx_handler_busy(dev))
5203 if (dev->priv_flags & IFF_NO_RX_HANDLER)
5206 /* Note: rx_handler_data must be set before rx_handler */
5207 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data);
5208 rcu_assign_pointer(dev->rx_handler, rx_handler);
5212 EXPORT_SYMBOL_GPL(netdev_rx_handler_register);
5215 * netdev_rx_handler_unregister - unregister receive handler
5216 * @dev: device to unregister a handler from
5218 * Unregister a receive handler from a device.
5220 * The caller must hold the rtnl_mutex.
5222 void netdev_rx_handler_unregister(struct net_device *dev)
5226 RCU_INIT_POINTER(dev->rx_handler, NULL);
5227 /* a reader seeing a non NULL rx_handler in a rcu_read_lock()
5228 * section has a guarantee to see a non NULL rx_handler_data
5232 RCU_INIT_POINTER(dev->rx_handler_data, NULL);
5234 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister);
5237 * Limit the use of PFMEMALLOC reserves to those protocols that implement
5238 * the special handling of PFMEMALLOC skbs.
5240 static bool skb_pfmemalloc_protocol(struct sk_buff *skb)
5242 switch (skb->protocol) {
5243 case htons(ETH_P_ARP):
5244 case htons(ETH_P_IP):
5245 case htons(ETH_P_IPV6):
5246 case htons(ETH_P_8021Q):
5247 case htons(ETH_P_8021AD):
5254 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev,
5255 int *ret, struct net_device *orig_dev)
5257 if (nf_hook_ingress_active(skb)) {
5261 *ret = deliver_skb(skb, *pt_prev, orig_dev);
5266 ingress_retval = nf_hook_ingress(skb);
5268 return ingress_retval;
5273 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc,
5274 struct packet_type **ppt_prev)
5276 struct packet_type *ptype, *pt_prev;
5277 rx_handler_func_t *rx_handler;
5278 struct sk_buff *skb = *pskb;
5279 struct net_device *orig_dev;
5280 bool deliver_exact = false;
5281 int ret = NET_RX_DROP;
5284 net_timestamp_check(!READ_ONCE(netdev_tstamp_prequeue), skb);
5286 trace_netif_receive_skb(skb);
5288 orig_dev = skb->dev;
5290 skb_reset_network_header(skb);
5291 if (!skb_transport_header_was_set(skb))
5292 skb_reset_transport_header(skb);
5293 skb_reset_mac_len(skb);
5298 skb->skb_iif = skb->dev->ifindex;
5300 __this_cpu_inc(softnet_data.processed);
5302 if (static_branch_unlikely(&generic_xdp_needed_key)) {
5306 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
5309 if (ret2 != XDP_PASS) {
5315 if (eth_type_vlan(skb->protocol)) {
5316 skb = skb_vlan_untag(skb);
5321 if (skb_skip_tc_classify(skb))
5327 list_for_each_entry_rcu(ptype, &ptype_all, list) {
5329 ret = deliver_skb(skb, pt_prev, orig_dev);
5333 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) {
5335 ret = deliver_skb(skb, pt_prev, orig_dev);
5340 #ifdef CONFIG_NET_INGRESS
5341 if (static_branch_unlikely(&ingress_needed_key)) {
5342 bool another = false;
5344 nf_skip_egress(skb, true);
5345 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev,
5352 nf_skip_egress(skb, false);
5353 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0)
5357 skb_reset_redirect(skb);
5359 if (pfmemalloc && !skb_pfmemalloc_protocol(skb))
5362 if (skb_vlan_tag_present(skb)) {
5364 ret = deliver_skb(skb, pt_prev, orig_dev);
5367 if (vlan_do_receive(&skb))
5369 else if (unlikely(!skb))
5373 rx_handler = rcu_dereference(skb->dev->rx_handler);
5376 ret = deliver_skb(skb, pt_prev, orig_dev);
5379 switch (rx_handler(&skb)) {
5380 case RX_HANDLER_CONSUMED:
5381 ret = NET_RX_SUCCESS;
5383 case RX_HANDLER_ANOTHER:
5385 case RX_HANDLER_EXACT:
5386 deliver_exact = true;
5388 case RX_HANDLER_PASS:
5395 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) {
5397 if (skb_vlan_tag_get_id(skb)) {
5398 /* Vlan id is non 0 and vlan_do_receive() above couldn't
5401 skb->pkt_type = PACKET_OTHERHOST;
5402 } else if (eth_type_vlan(skb->protocol)) {
5403 /* Outer header is 802.1P with vlan 0, inner header is
5404 * 802.1Q or 802.1AD and vlan_do_receive() above could
5405 * not find vlan dev for vlan id 0.
5407 __vlan_hwaccel_clear_tag(skb);
5408 skb = skb_vlan_untag(skb);
5411 if (vlan_do_receive(&skb))
5412 /* After stripping off 802.1P header with vlan 0
5413 * vlan dev is found for inner header.
5416 else if (unlikely(!skb))
5419 /* We have stripped outer 802.1P vlan 0 header.
5420 * But could not find vlan dev.
5421 * check again for vlan id to set OTHERHOST.
5425 /* Note: we might in the future use prio bits
5426 * and set skb->priority like in vlan_do_receive()
5427 * For the time being, just ignore Priority Code Point
5429 __vlan_hwaccel_clear_tag(skb);
5432 type = skb->protocol;
5434 /* deliver only exact match when indicated */
5435 if (likely(!deliver_exact)) {
5436 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
5437 &ptype_base[ntohs(type) &
5441 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
5442 &orig_dev->ptype_specific);
5444 if (unlikely(skb->dev != orig_dev)) {
5445 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
5446 &skb->dev->ptype_specific);
5450 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
5452 *ppt_prev = pt_prev;
5456 dev_core_stats_rx_dropped_inc(skb->dev);
5458 dev_core_stats_rx_nohandler_inc(skb->dev);
5459 kfree_skb_reason(skb, SKB_DROP_REASON_UNHANDLED_PROTO);
5460 /* Jamal, now you will not able to escape explaining
5461 * me how you were going to use this. :-)
5467 /* The invariant here is that if *ppt_prev is not NULL
5468 * then skb should also be non-NULL.
5470 * Apparently *ppt_prev assignment above holds this invariant due to
5471 * skb dereferencing near it.
5477 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc)
5479 struct net_device *orig_dev = skb->dev;
5480 struct packet_type *pt_prev = NULL;
5483 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev);
5485 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb,
5486 skb->dev, pt_prev, orig_dev);
5491 * netif_receive_skb_core - special purpose version of netif_receive_skb
5492 * @skb: buffer to process
5494 * More direct receive version of netif_receive_skb(). It should
5495 * only be used by callers that have a need to skip RPS and Generic XDP.
5496 * Caller must also take care of handling if ``(page_is_)pfmemalloc``.
5498 * This function may only be called from softirq context and interrupts
5499 * should be enabled.
5501 * Return values (usually ignored):
5502 * NET_RX_SUCCESS: no congestion
5503 * NET_RX_DROP: packet was dropped
5505 int netif_receive_skb_core(struct sk_buff *skb)
5510 ret = __netif_receive_skb_one_core(skb, false);
5515 EXPORT_SYMBOL(netif_receive_skb_core);
5517 static inline void __netif_receive_skb_list_ptype(struct list_head *head,
5518 struct packet_type *pt_prev,
5519 struct net_device *orig_dev)
5521 struct sk_buff *skb, *next;
5525 if (list_empty(head))
5527 if (pt_prev->list_func != NULL)
5528 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv,
5529 ip_list_rcv, head, pt_prev, orig_dev);
5531 list_for_each_entry_safe(skb, next, head, list) {
5532 skb_list_del_init(skb);
5533 pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
5537 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc)
5539 /* Fast-path assumptions:
5540 * - There is no RX handler.
5541 * - Only one packet_type matches.
5542 * If either of these fails, we will end up doing some per-packet
5543 * processing in-line, then handling the 'last ptype' for the whole
5544 * sublist. This can't cause out-of-order delivery to any single ptype,
5545 * because the 'last ptype' must be constant across the sublist, and all
5546 * other ptypes are handled per-packet.
5548 /* Current (common) ptype of sublist */
5549 struct packet_type *pt_curr = NULL;
5550 /* Current (common) orig_dev of sublist */
5551 struct net_device *od_curr = NULL;
5552 struct list_head sublist;
5553 struct sk_buff *skb, *next;
5555 INIT_LIST_HEAD(&sublist);
5556 list_for_each_entry_safe(skb, next, head, list) {
5557 struct net_device *orig_dev = skb->dev;
5558 struct packet_type *pt_prev = NULL;
5560 skb_list_del_init(skb);
5561 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev);
5564 if (pt_curr != pt_prev || od_curr != orig_dev) {
5565 /* dispatch old sublist */
5566 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr);
5567 /* start new sublist */
5568 INIT_LIST_HEAD(&sublist);
5572 list_add_tail(&skb->list, &sublist);
5575 /* dispatch final sublist */
5576 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr);
5579 static int __netif_receive_skb(struct sk_buff *skb)
5583 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) {
5584 unsigned int noreclaim_flag;
5587 * PFMEMALLOC skbs are special, they should
5588 * - be delivered to SOCK_MEMALLOC sockets only
5589 * - stay away from userspace
5590 * - have bounded memory usage
5592 * Use PF_MEMALLOC as this saves us from propagating the allocation
5593 * context down to all allocation sites.
5595 noreclaim_flag = memalloc_noreclaim_save();
5596 ret = __netif_receive_skb_one_core(skb, true);
5597 memalloc_noreclaim_restore(noreclaim_flag);
5599 ret = __netif_receive_skb_one_core(skb, false);
5604 static void __netif_receive_skb_list(struct list_head *head)
5606 unsigned long noreclaim_flag = 0;
5607 struct sk_buff *skb, *next;
5608 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */
5610 list_for_each_entry_safe(skb, next, head, list) {
5611 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) {
5612 struct list_head sublist;
5614 /* Handle the previous sublist */
5615 list_cut_before(&sublist, head, &skb->list);
5616 if (!list_empty(&sublist))
5617 __netif_receive_skb_list_core(&sublist, pfmemalloc);
5618 pfmemalloc = !pfmemalloc;
5619 /* See comments in __netif_receive_skb */
5621 noreclaim_flag = memalloc_noreclaim_save();
5623 memalloc_noreclaim_restore(noreclaim_flag);
5626 /* Handle the remaining sublist */
5627 if (!list_empty(head))
5628 __netif_receive_skb_list_core(head, pfmemalloc);
5629 /* Restore pflags */
5631 memalloc_noreclaim_restore(noreclaim_flag);
5634 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp)
5636 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog);
5637 struct bpf_prog *new = xdp->prog;
5640 switch (xdp->command) {
5641 case XDP_SETUP_PROG:
5642 rcu_assign_pointer(dev->xdp_prog, new);
5647 static_branch_dec(&generic_xdp_needed_key);
5648 } else if (new && !old) {
5649 static_branch_inc(&generic_xdp_needed_key);
5650 dev_disable_lro(dev);
5651 dev_disable_gro_hw(dev);
5663 static int netif_receive_skb_internal(struct sk_buff *skb)
5667 net_timestamp_check(READ_ONCE(netdev_tstamp_prequeue), skb);
5669 if (skb_defer_rx_timestamp(skb))
5670 return NET_RX_SUCCESS;
5674 if (static_branch_unlikely(&rps_needed)) {
5675 struct rps_dev_flow voidflow, *rflow = &voidflow;
5676 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
5679 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
5685 ret = __netif_receive_skb(skb);
5690 void netif_receive_skb_list_internal(struct list_head *head)
5692 struct sk_buff *skb, *next;
5693 struct list_head sublist;
5695 INIT_LIST_HEAD(&sublist);
5696 list_for_each_entry_safe(skb, next, head, list) {
5697 net_timestamp_check(READ_ONCE(netdev_tstamp_prequeue), skb);
5698 skb_list_del_init(skb);
5699 if (!skb_defer_rx_timestamp(skb))
5700 list_add_tail(&skb->list, &sublist);
5702 list_splice_init(&sublist, head);
5706 if (static_branch_unlikely(&rps_needed)) {
5707 list_for_each_entry_safe(skb, next, head, list) {
5708 struct rps_dev_flow voidflow, *rflow = &voidflow;
5709 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
5712 /* Will be handled, remove from list */
5713 skb_list_del_init(skb);
5714 enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
5719 __netif_receive_skb_list(head);
5724 * netif_receive_skb - process receive buffer from network
5725 * @skb: buffer to process
5727 * netif_receive_skb() is the main receive data processing function.
5728 * It always succeeds. The buffer may be dropped during processing
5729 * for congestion control or by the protocol layers.
5731 * This function may only be called from softirq context and interrupts
5732 * should be enabled.
5734 * Return values (usually ignored):
5735 * NET_RX_SUCCESS: no congestion
5736 * NET_RX_DROP: packet was dropped
5738 int netif_receive_skb(struct sk_buff *skb)
5742 trace_netif_receive_skb_entry(skb);
5744 ret = netif_receive_skb_internal(skb);
5745 trace_netif_receive_skb_exit(ret);
5749 EXPORT_SYMBOL(netif_receive_skb);
5752 * netif_receive_skb_list - process many receive buffers from network
5753 * @head: list of skbs to process.
5755 * Since return value of netif_receive_skb() is normally ignored, and
5756 * wouldn't be meaningful for a list, this function returns void.
5758 * This function may only be called from softirq context and interrupts
5759 * should be enabled.
5761 void netif_receive_skb_list(struct list_head *head)
5763 struct sk_buff *skb;
5765 if (list_empty(head))
5767 if (trace_netif_receive_skb_list_entry_enabled()) {
5768 list_for_each_entry(skb, head, list)
5769 trace_netif_receive_skb_list_entry(skb);
5771 netif_receive_skb_list_internal(head);
5772 trace_netif_receive_skb_list_exit(0);
5774 EXPORT_SYMBOL(netif_receive_skb_list);
5776 static DEFINE_PER_CPU(struct work_struct, flush_works);
5778 /* Network device is going away, flush any packets still pending */
5779 static void flush_backlog(struct work_struct *work)
5781 struct sk_buff *skb, *tmp;
5782 struct softnet_data *sd;
5785 sd = this_cpu_ptr(&softnet_data);
5787 rps_lock_irq_disable(sd);
5788 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) {
5789 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
5790 __skb_unlink(skb, &sd->input_pkt_queue);
5791 dev_kfree_skb_irq(skb);
5792 input_queue_head_incr(sd);
5795 rps_unlock_irq_enable(sd);
5797 skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
5798 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
5799 __skb_unlink(skb, &sd->process_queue);
5801 input_queue_head_incr(sd);
5807 static bool flush_required(int cpu)
5809 #if IS_ENABLED(CONFIG_RPS)
5810 struct softnet_data *sd = &per_cpu(softnet_data, cpu);
5813 rps_lock_irq_disable(sd);
5815 /* as insertion into process_queue happens with the rps lock held,
5816 * process_queue access may race only with dequeue
5818 do_flush = !skb_queue_empty(&sd->input_pkt_queue) ||
5819 !skb_queue_empty_lockless(&sd->process_queue);
5820 rps_unlock_irq_enable(sd);
5824 /* without RPS we can't safely check input_pkt_queue: during a
5825 * concurrent remote skb_queue_splice() we can detect as empty both
5826 * input_pkt_queue and process_queue even if the latter could end-up
5827 * containing a lot of packets.
5832 static void flush_all_backlogs(void)
5834 static cpumask_t flush_cpus;
5837 /* since we are under rtnl lock protection we can use static data
5838 * for the cpumask and avoid allocating on stack the possibly
5845 cpumask_clear(&flush_cpus);
5846 for_each_online_cpu(cpu) {
5847 if (flush_required(cpu)) {
5848 queue_work_on(cpu, system_highpri_wq,
5849 per_cpu_ptr(&flush_works, cpu));
5850 cpumask_set_cpu(cpu, &flush_cpus);
5854 /* we can have in flight packet[s] on the cpus we are not flushing,
5855 * synchronize_net() in unregister_netdevice_many() will take care of
5858 for_each_cpu(cpu, &flush_cpus)
5859 flush_work(per_cpu_ptr(&flush_works, cpu));
5864 static void net_rps_send_ipi(struct softnet_data *remsd)
5868 struct softnet_data *next = remsd->rps_ipi_next;
5870 if (cpu_online(remsd->cpu))
5871 smp_call_function_single_async(remsd->cpu, &remsd->csd);
5878 * net_rps_action_and_irq_enable sends any pending IPI's for rps.
5879 * Note: called with local irq disabled, but exits with local irq enabled.
5881 static void net_rps_action_and_irq_enable(struct softnet_data *sd)
5884 struct softnet_data *remsd = sd->rps_ipi_list;
5887 sd->rps_ipi_list = NULL;
5891 /* Send pending IPI's to kick RPS processing on remote cpus. */
5892 net_rps_send_ipi(remsd);
5898 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd)
5901 return sd->rps_ipi_list != NULL;
5907 static int process_backlog(struct napi_struct *napi, int quota)
5909 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
5913 /* Check if we have pending ipi, its better to send them now,
5914 * not waiting net_rx_action() end.
5916 if (sd_has_rps_ipi_waiting(sd)) {
5917 local_irq_disable();
5918 net_rps_action_and_irq_enable(sd);
5921 napi->weight = READ_ONCE(dev_rx_weight);
5923 struct sk_buff *skb;
5925 while ((skb = __skb_dequeue(&sd->process_queue))) {
5927 __netif_receive_skb(skb);
5929 input_queue_head_incr(sd);
5930 if (++work >= quota)
5935 rps_lock_irq_disable(sd);
5936 if (skb_queue_empty(&sd->input_pkt_queue)) {
5938 * Inline a custom version of __napi_complete().
5939 * only current cpu owns and manipulates this napi,
5940 * and NAPI_STATE_SCHED is the only possible flag set
5942 * We can use a plain write instead of clear_bit(),
5943 * and we dont need an smp_mb() memory barrier.
5948 skb_queue_splice_tail_init(&sd->input_pkt_queue,
5949 &sd->process_queue);
5951 rps_unlock_irq_enable(sd);
5958 * __napi_schedule - schedule for receive
5959 * @n: entry to schedule
5961 * The entry's receive function will be scheduled to run.
5962 * Consider using __napi_schedule_irqoff() if hard irqs are masked.
5964 void __napi_schedule(struct napi_struct *n)
5966 unsigned long flags;
5968 local_irq_save(flags);
5969 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
5970 local_irq_restore(flags);
5972 EXPORT_SYMBOL(__napi_schedule);
5975 * napi_schedule_prep - check if napi can be scheduled
5978 * Test if NAPI routine is already running, and if not mark
5979 * it as running. This is used as a condition variable to
5980 * insure only one NAPI poll instance runs. We also make
5981 * sure there is no pending NAPI disable.
5983 bool napi_schedule_prep(struct napi_struct *n)
5985 unsigned long new, val = READ_ONCE(n->state);
5988 if (unlikely(val & NAPIF_STATE_DISABLE))
5990 new = val | NAPIF_STATE_SCHED;
5992 /* Sets STATE_MISSED bit if STATE_SCHED was already set
5993 * This was suggested by Alexander Duyck, as compiler
5994 * emits better code than :
5995 * if (val & NAPIF_STATE_SCHED)
5996 * new |= NAPIF_STATE_MISSED;
5998 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED *
6000 } while (!try_cmpxchg(&n->state, &val, new));
6002 return !(val & NAPIF_STATE_SCHED);
6004 EXPORT_SYMBOL(napi_schedule_prep);
6007 * __napi_schedule_irqoff - schedule for receive
6008 * @n: entry to schedule
6010 * Variant of __napi_schedule() assuming hard irqs are masked.
6012 * On PREEMPT_RT enabled kernels this maps to __napi_schedule()
6013 * because the interrupt disabled assumption might not be true
6014 * due to force-threaded interrupts and spinlock substitution.
6016 void __napi_schedule_irqoff(struct napi_struct *n)
6018 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
6019 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
6023 EXPORT_SYMBOL(__napi_schedule_irqoff);
6025 bool napi_complete_done(struct napi_struct *n, int work_done)
6027 unsigned long flags, val, new, timeout = 0;
6031 * 1) Don't let napi dequeue from the cpu poll list
6032 * just in case its running on a different cpu.
6033 * 2) If we are busy polling, do nothing here, we have
6034 * the guarantee we will be called later.
6036 if (unlikely(n->state & (NAPIF_STATE_NPSVC |
6037 NAPIF_STATE_IN_BUSY_POLL)))
6042 timeout = READ_ONCE(n->dev->gro_flush_timeout);
6043 n->defer_hard_irqs_count = READ_ONCE(n->dev->napi_defer_hard_irqs);
6045 if (n->defer_hard_irqs_count > 0) {
6046 n->defer_hard_irqs_count--;
6047 timeout = READ_ONCE(n->dev->gro_flush_timeout);
6051 if (n->gro_bitmask) {
6052 /* When the NAPI instance uses a timeout and keeps postponing
6053 * it, we need to bound somehow the time packets are kept in
6056 napi_gro_flush(n, !!timeout);
6061 if (unlikely(!list_empty(&n->poll_list))) {
6062 /* If n->poll_list is not empty, we need to mask irqs */
6063 local_irq_save(flags);
6064 list_del_init(&n->poll_list);
6065 local_irq_restore(flags);
6068 val = READ_ONCE(n->state);
6070 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED));
6072 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED |
6073 NAPIF_STATE_SCHED_THREADED |
6074 NAPIF_STATE_PREFER_BUSY_POLL);
6076 /* If STATE_MISSED was set, leave STATE_SCHED set,
6077 * because we will call napi->poll() one more time.
6078 * This C code was suggested by Alexander Duyck to help gcc.
6080 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED *
6082 } while (!try_cmpxchg(&n->state, &val, new));
6084 if (unlikely(val & NAPIF_STATE_MISSED)) {
6090 hrtimer_start(&n->timer, ns_to_ktime(timeout),
6091 HRTIMER_MODE_REL_PINNED);
6094 EXPORT_SYMBOL(napi_complete_done);
6096 /* must be called under rcu_read_lock(), as we dont take a reference */
6097 static struct napi_struct *napi_by_id(unsigned int napi_id)
6099 unsigned int hash = napi_id % HASH_SIZE(napi_hash);
6100 struct napi_struct *napi;
6102 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node)
6103 if (napi->napi_id == napi_id)
6109 #if defined(CONFIG_NET_RX_BUSY_POLL)
6111 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule)
6113 if (!skip_schedule) {
6114 gro_normal_list(napi);
6115 __napi_schedule(napi);
6119 if (napi->gro_bitmask) {
6120 /* flush too old packets
6121 * If HZ < 1000, flush all packets.
6123 napi_gro_flush(napi, HZ >= 1000);
6126 gro_normal_list(napi);
6127 clear_bit(NAPI_STATE_SCHED, &napi->state);
6130 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, bool prefer_busy_poll,
6133 bool skip_schedule = false;
6134 unsigned long timeout;
6137 /* Busy polling means there is a high chance device driver hard irq
6138 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was
6139 * set in napi_schedule_prep().
6140 * Since we are about to call napi->poll() once more, we can safely
6141 * clear NAPI_STATE_MISSED.
6143 * Note: x86 could use a single "lock and ..." instruction
6144 * to perform these two clear_bit()
6146 clear_bit(NAPI_STATE_MISSED, &napi->state);
6147 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state);
6151 if (prefer_busy_poll) {
6152 napi->defer_hard_irqs_count = READ_ONCE(napi->dev->napi_defer_hard_irqs);
6153 timeout = READ_ONCE(napi->dev->gro_flush_timeout);
6154 if (napi->defer_hard_irqs_count && timeout) {
6155 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED);
6156 skip_schedule = true;
6160 /* All we really want here is to re-enable device interrupts.
6161 * Ideally, a new ndo_busy_poll_stop() could avoid another round.
6163 rc = napi->poll(napi, budget);
6164 /* We can't gro_normal_list() here, because napi->poll() might have
6165 * rearmed the napi (napi_complete_done()) in which case it could
6166 * already be running on another CPU.
6168 trace_napi_poll(napi, rc, budget);
6169 netpoll_poll_unlock(have_poll_lock);
6171 __busy_poll_stop(napi, skip_schedule);
6175 void napi_busy_loop(unsigned int napi_id,
6176 bool (*loop_end)(void *, unsigned long),
6177 void *loop_end_arg, bool prefer_busy_poll, u16 budget)
6179 unsigned long start_time = loop_end ? busy_loop_current_time() : 0;
6180 int (*napi_poll)(struct napi_struct *napi, int budget);
6181 void *have_poll_lock = NULL;
6182 struct napi_struct *napi;
6189 napi = napi_by_id(napi_id);
6199 unsigned long val = READ_ONCE(napi->state);
6201 /* If multiple threads are competing for this napi,
6202 * we avoid dirtying napi->state as much as we can.
6204 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED |
6205 NAPIF_STATE_IN_BUSY_POLL)) {
6206 if (prefer_busy_poll)
6207 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state);
6210 if (cmpxchg(&napi->state, val,
6211 val | NAPIF_STATE_IN_BUSY_POLL |
6212 NAPIF_STATE_SCHED) != val) {
6213 if (prefer_busy_poll)
6214 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state);
6217 have_poll_lock = netpoll_poll_lock(napi);
6218 napi_poll = napi->poll;
6220 work = napi_poll(napi, budget);
6221 trace_napi_poll(napi, work, budget);
6222 gro_normal_list(napi);
6225 __NET_ADD_STATS(dev_net(napi->dev),
6226 LINUX_MIB_BUSYPOLLRXPACKETS, work);
6229 if (!loop_end || loop_end(loop_end_arg, start_time))
6232 if (unlikely(need_resched())) {
6234 busy_poll_stop(napi, have_poll_lock, prefer_busy_poll, budget);
6238 if (loop_end(loop_end_arg, start_time))
6245 busy_poll_stop(napi, have_poll_lock, prefer_busy_poll, budget);
6250 EXPORT_SYMBOL(napi_busy_loop);
6252 #endif /* CONFIG_NET_RX_BUSY_POLL */
6254 static void napi_hash_add(struct napi_struct *napi)
6256 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state))
6259 spin_lock(&napi_hash_lock);
6261 /* 0..NR_CPUS range is reserved for sender_cpu use */
6263 if (unlikely(++napi_gen_id < MIN_NAPI_ID))
6264 napi_gen_id = MIN_NAPI_ID;
6265 } while (napi_by_id(napi_gen_id));
6266 napi->napi_id = napi_gen_id;
6268 hlist_add_head_rcu(&napi->napi_hash_node,
6269 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]);
6271 spin_unlock(&napi_hash_lock);
6274 /* Warning : caller is responsible to make sure rcu grace period
6275 * is respected before freeing memory containing @napi
6277 static void napi_hash_del(struct napi_struct *napi)
6279 spin_lock(&napi_hash_lock);
6281 hlist_del_init_rcu(&napi->napi_hash_node);
6283 spin_unlock(&napi_hash_lock);
6286 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer)
6288 struct napi_struct *napi;
6290 napi = container_of(timer, struct napi_struct, timer);
6292 /* Note : we use a relaxed variant of napi_schedule_prep() not setting
6293 * NAPI_STATE_MISSED, since we do not react to a device IRQ.
6295 if (!napi_disable_pending(napi) &&
6296 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) {
6297 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state);
6298 __napi_schedule_irqoff(napi);
6301 return HRTIMER_NORESTART;
6304 static void init_gro_hash(struct napi_struct *napi)
6308 for (i = 0; i < GRO_HASH_BUCKETS; i++) {
6309 INIT_LIST_HEAD(&napi->gro_hash[i].list);
6310 napi->gro_hash[i].count = 0;
6312 napi->gro_bitmask = 0;
6315 int dev_set_threaded(struct net_device *dev, bool threaded)
6317 struct napi_struct *napi;
6320 if (dev->threaded == threaded)
6324 list_for_each_entry(napi, &dev->napi_list, dev_list) {
6325 if (!napi->thread) {
6326 err = napi_kthread_create(napi);
6335 dev->threaded = threaded;
6337 /* Make sure kthread is created before THREADED bit
6340 smp_mb__before_atomic();
6342 /* Setting/unsetting threaded mode on a napi might not immediately
6343 * take effect, if the current napi instance is actively being
6344 * polled. In this case, the switch between threaded mode and
6345 * softirq mode will happen in the next round of napi_schedule().
6346 * This should not cause hiccups/stalls to the live traffic.
6348 list_for_each_entry(napi, &dev->napi_list, dev_list) {
6350 set_bit(NAPI_STATE_THREADED, &napi->state);
6352 clear_bit(NAPI_STATE_THREADED, &napi->state);
6357 EXPORT_SYMBOL(dev_set_threaded);
6359 void netif_napi_add_weight(struct net_device *dev, struct napi_struct *napi,
6360 int (*poll)(struct napi_struct *, int), int weight)
6362 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state)))
6365 INIT_LIST_HEAD(&napi->poll_list);
6366 INIT_HLIST_NODE(&napi->napi_hash_node);
6367 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
6368 napi->timer.function = napi_watchdog;
6369 init_gro_hash(napi);
6371 INIT_LIST_HEAD(&napi->rx_list);
6374 if (weight > NAPI_POLL_WEIGHT)
6375 netdev_err_once(dev, "%s() called with weight %d\n", __func__,
6377 napi->weight = weight;
6379 #ifdef CONFIG_NETPOLL
6380 napi->poll_owner = -1;
6382 set_bit(NAPI_STATE_SCHED, &napi->state);
6383 set_bit(NAPI_STATE_NPSVC, &napi->state);
6384 list_add_rcu(&napi->dev_list, &dev->napi_list);
6385 napi_hash_add(napi);
6386 napi_get_frags_check(napi);
6387 /* Create kthread for this napi if dev->threaded is set.
6388 * Clear dev->threaded if kthread creation failed so that
6389 * threaded mode will not be enabled in napi_enable().
6391 if (dev->threaded && napi_kthread_create(napi))
6394 EXPORT_SYMBOL(netif_napi_add_weight);
6396 void napi_disable(struct napi_struct *n)
6398 unsigned long val, new;
6401 set_bit(NAPI_STATE_DISABLE, &n->state);
6403 val = READ_ONCE(n->state);
6405 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) {
6406 usleep_range(20, 200);
6407 val = READ_ONCE(n->state);
6410 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC;
6411 new &= ~(NAPIF_STATE_THREADED | NAPIF_STATE_PREFER_BUSY_POLL);
6412 } while (!try_cmpxchg(&n->state, &val, new));
6414 hrtimer_cancel(&n->timer);
6416 clear_bit(NAPI_STATE_DISABLE, &n->state);
6418 EXPORT_SYMBOL(napi_disable);
6421 * napi_enable - enable NAPI scheduling
6424 * Resume NAPI from being scheduled on this context.
6425 * Must be paired with napi_disable.
6427 void napi_enable(struct napi_struct *n)
6429 unsigned long new, val = READ_ONCE(n->state);
6432 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val));
6434 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC);
6435 if (n->dev->threaded && n->thread)
6436 new |= NAPIF_STATE_THREADED;
6437 } while (!try_cmpxchg(&n->state, &val, new));
6439 EXPORT_SYMBOL(napi_enable);
6441 static void flush_gro_hash(struct napi_struct *napi)
6445 for (i = 0; i < GRO_HASH_BUCKETS; i++) {
6446 struct sk_buff *skb, *n;
6448 list_for_each_entry_safe(skb, n, &napi->gro_hash[i].list, list)
6450 napi->gro_hash[i].count = 0;
6454 /* Must be called in process context */
6455 void __netif_napi_del(struct napi_struct *napi)
6457 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state))
6460 napi_hash_del(napi);
6461 list_del_rcu(&napi->dev_list);
6462 napi_free_frags(napi);
6464 flush_gro_hash(napi);
6465 napi->gro_bitmask = 0;
6468 kthread_stop(napi->thread);
6469 napi->thread = NULL;
6472 EXPORT_SYMBOL(__netif_napi_del);
6474 static int __napi_poll(struct napi_struct *n, bool *repoll)
6480 /* This NAPI_STATE_SCHED test is for avoiding a race
6481 * with netpoll's poll_napi(). Only the entity which
6482 * obtains the lock and sees NAPI_STATE_SCHED set will
6483 * actually make the ->poll() call. Therefore we avoid
6484 * accidentally calling ->poll() when NAPI is not scheduled.
6487 if (test_bit(NAPI_STATE_SCHED, &n->state)) {
6488 work = n->poll(n, weight);
6489 trace_napi_poll(n, work, weight);
6492 if (unlikely(work > weight))
6493 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n",
6494 n->poll, work, weight);
6496 if (likely(work < weight))
6499 /* Drivers must not modify the NAPI state if they
6500 * consume the entire weight. In such cases this code
6501 * still "owns" the NAPI instance and therefore can
6502 * move the instance around on the list at-will.
6504 if (unlikely(napi_disable_pending(n))) {
6509 /* The NAPI context has more processing work, but busy-polling
6510 * is preferred. Exit early.
6512 if (napi_prefer_busy_poll(n)) {
6513 if (napi_complete_done(n, work)) {
6514 /* If timeout is not set, we need to make sure
6515 * that the NAPI is re-scheduled.
6522 if (n->gro_bitmask) {
6523 /* flush too old packets
6524 * If HZ < 1000, flush all packets.
6526 napi_gro_flush(n, HZ >= 1000);
6531 /* Some drivers may have called napi_schedule
6532 * prior to exhausting their budget.
6534 if (unlikely(!list_empty(&n->poll_list))) {
6535 pr_warn_once("%s: Budget exhausted after napi rescheduled\n",
6536 n->dev ? n->dev->name : "backlog");
6545 static int napi_poll(struct napi_struct *n, struct list_head *repoll)
6547 bool do_repoll = false;
6551 list_del_init(&n->poll_list);
6553 have = netpoll_poll_lock(n);
6555 work = __napi_poll(n, &do_repoll);
6558 list_add_tail(&n->poll_list, repoll);
6560 netpoll_poll_unlock(have);
6565 static int napi_thread_wait(struct napi_struct *napi)
6569 set_current_state(TASK_INTERRUPTIBLE);
6571 while (!kthread_should_stop()) {
6572 /* Testing SCHED_THREADED bit here to make sure the current
6573 * kthread owns this napi and could poll on this napi.
6574 * Testing SCHED bit is not enough because SCHED bit might be
6575 * set by some other busy poll thread or by napi_disable().
6577 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state) || woken) {
6578 WARN_ON(!list_empty(&napi->poll_list));
6579 __set_current_state(TASK_RUNNING);
6584 /* woken being true indicates this thread owns this napi. */
6586 set_current_state(TASK_INTERRUPTIBLE);
6588 __set_current_state(TASK_RUNNING);
6593 static int napi_threaded_poll(void *data)
6595 struct napi_struct *napi = data;
6598 while (!napi_thread_wait(napi)) {
6600 bool repoll = false;
6604 have = netpoll_poll_lock(napi);
6605 __napi_poll(napi, &repoll);
6606 netpoll_poll_unlock(have);
6619 static void skb_defer_free_flush(struct softnet_data *sd)
6621 struct sk_buff *skb, *next;
6623 /* Paired with WRITE_ONCE() in skb_attempt_defer_free() */
6624 if (!READ_ONCE(sd->defer_list))
6627 spin_lock_irq(&sd->defer_lock);
6628 skb = sd->defer_list;
6629 sd->defer_list = NULL;
6630 sd->defer_count = 0;
6631 spin_unlock_irq(&sd->defer_lock);
6633 while (skb != NULL) {
6635 napi_consume_skb(skb, 1);
6640 static __latent_entropy void net_rx_action(struct softirq_action *h)
6642 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
6643 unsigned long time_limit = jiffies +
6644 usecs_to_jiffies(READ_ONCE(netdev_budget_usecs));
6645 int budget = READ_ONCE(netdev_budget);
6649 sd->in_net_rx_action = true;
6650 local_irq_disable();
6651 list_splice_init(&sd->poll_list, &list);
6655 struct napi_struct *n;
6657 skb_defer_free_flush(sd);
6659 if (list_empty(&list)) {
6660 sd->in_net_rx_action = false;
6661 if (!sd_has_rps_ipi_waiting(sd) && list_empty(&repoll))
6666 n = list_first_entry(&list, struct napi_struct, poll_list);
6667 budget -= napi_poll(n, &repoll);
6669 /* If softirq window is exhausted then punt.
6670 * Allow this to run for 2 jiffies since which will allow
6671 * an average latency of 1.5/HZ.
6673 if (unlikely(budget <= 0 ||
6674 time_after_eq(jiffies, time_limit))) {
6680 local_irq_disable();
6682 list_splice_tail_init(&sd->poll_list, &list);
6683 list_splice_tail(&repoll, &list);
6684 list_splice(&list, &sd->poll_list);
6685 if (!list_empty(&sd->poll_list))
6686 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
6688 sd->in_net_rx_action = false;
6690 net_rps_action_and_irq_enable(sd);
6694 struct netdev_adjacent {
6695 struct net_device *dev;
6696 netdevice_tracker dev_tracker;
6698 /* upper master flag, there can only be one master device per list */
6701 /* lookup ignore flag */
6704 /* counter for the number of times this device was added to us */
6707 /* private field for the users */
6710 struct list_head list;
6711 struct rcu_head rcu;
6714 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev,
6715 struct list_head *adj_list)
6717 struct netdev_adjacent *adj;
6719 list_for_each_entry(adj, adj_list, list) {
6720 if (adj->dev == adj_dev)
6726 static int ____netdev_has_upper_dev(struct net_device *upper_dev,
6727 struct netdev_nested_priv *priv)
6729 struct net_device *dev = (struct net_device *)priv->data;
6731 return upper_dev == dev;
6735 * netdev_has_upper_dev - Check if device is linked to an upper device
6737 * @upper_dev: upper device to check
6739 * Find out if a device is linked to specified upper device and return true
6740 * in case it is. Note that this checks only immediate upper device,
6741 * not through a complete stack of devices. The caller must hold the RTNL lock.
6743 bool netdev_has_upper_dev(struct net_device *dev,
6744 struct net_device *upper_dev)
6746 struct netdev_nested_priv priv = {
6747 .data = (void *)upper_dev,
6752 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev,
6755 EXPORT_SYMBOL(netdev_has_upper_dev);
6758 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device
6760 * @upper_dev: upper device to check
6762 * Find out if a device is linked to specified upper device and return true
6763 * in case it is. Note that this checks the entire upper device chain.
6764 * The caller must hold rcu lock.
6767 bool netdev_has_upper_dev_all_rcu(struct net_device *dev,
6768 struct net_device *upper_dev)
6770 struct netdev_nested_priv priv = {
6771 .data = (void *)upper_dev,
6774 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev,
6777 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu);
6780 * netdev_has_any_upper_dev - Check if device is linked to some device
6783 * Find out if a device is linked to an upper device and return true in case
6784 * it is. The caller must hold the RTNL lock.
6786 bool netdev_has_any_upper_dev(struct net_device *dev)
6790 return !list_empty(&dev->adj_list.upper);
6792 EXPORT_SYMBOL(netdev_has_any_upper_dev);
6795 * netdev_master_upper_dev_get - Get master upper device
6798 * Find a master upper device and return pointer to it or NULL in case
6799 * it's not there. The caller must hold the RTNL lock.
6801 struct net_device *netdev_master_upper_dev_get(struct net_device *dev)
6803 struct netdev_adjacent *upper;
6807 if (list_empty(&dev->adj_list.upper))
6810 upper = list_first_entry(&dev->adj_list.upper,
6811 struct netdev_adjacent, list);
6812 if (likely(upper->master))
6816 EXPORT_SYMBOL(netdev_master_upper_dev_get);
6818 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev)
6820 struct netdev_adjacent *upper;
6824 if (list_empty(&dev->adj_list.upper))
6827 upper = list_first_entry(&dev->adj_list.upper,
6828 struct netdev_adjacent, list);
6829 if (likely(upper->master) && !upper->ignore)
6835 * netdev_has_any_lower_dev - Check if device is linked to some device
6838 * Find out if a device is linked to a lower device and return true in case
6839 * it is. The caller must hold the RTNL lock.
6841 static bool netdev_has_any_lower_dev(struct net_device *dev)
6845 return !list_empty(&dev->adj_list.lower);
6848 void *netdev_adjacent_get_private(struct list_head *adj_list)
6850 struct netdev_adjacent *adj;
6852 adj = list_entry(adj_list, struct netdev_adjacent, list);
6854 return adj->private;
6856 EXPORT_SYMBOL(netdev_adjacent_get_private);
6859 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list
6861 * @iter: list_head ** of the current position
6863 * Gets the next device from the dev's upper list, starting from iter
6864 * position. The caller must hold RCU read lock.
6866 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev,
6867 struct list_head **iter)
6869 struct netdev_adjacent *upper;
6871 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
6873 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6875 if (&upper->list == &dev->adj_list.upper)
6878 *iter = &upper->list;
6882 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu);
6884 static struct net_device *__netdev_next_upper_dev(struct net_device *dev,
6885 struct list_head **iter,
6888 struct netdev_adjacent *upper;
6890 upper = list_entry((*iter)->next, struct netdev_adjacent, list);
6892 if (&upper->list == &dev->adj_list.upper)
6895 *iter = &upper->list;
6896 *ignore = upper->ignore;
6901 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev,
6902 struct list_head **iter)
6904 struct netdev_adjacent *upper;
6906 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
6908 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6910 if (&upper->list == &dev->adj_list.upper)
6913 *iter = &upper->list;
6918 static int __netdev_walk_all_upper_dev(struct net_device *dev,
6919 int (*fn)(struct net_device *dev,
6920 struct netdev_nested_priv *priv),
6921 struct netdev_nested_priv *priv)
6923 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
6924 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
6929 iter = &dev->adj_list.upper;
6933 ret = fn(now, priv);
6940 udev = __netdev_next_upper_dev(now, &iter, &ignore);
6947 niter = &udev->adj_list.upper;
6948 dev_stack[cur] = now;
6949 iter_stack[cur++] = iter;
6956 next = dev_stack[--cur];
6957 niter = iter_stack[cur];
6967 int netdev_walk_all_upper_dev_rcu(struct net_device *dev,
6968 int (*fn)(struct net_device *dev,
6969 struct netdev_nested_priv *priv),
6970 struct netdev_nested_priv *priv)
6972 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
6973 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
6977 iter = &dev->adj_list.upper;
6981 ret = fn(now, priv);
6988 udev = netdev_next_upper_dev_rcu(now, &iter);
6993 niter = &udev->adj_list.upper;
6994 dev_stack[cur] = now;
6995 iter_stack[cur++] = iter;
7002 next = dev_stack[--cur];
7003 niter = iter_stack[cur];
7012 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu);
7014 static bool __netdev_has_upper_dev(struct net_device *dev,
7015 struct net_device *upper_dev)
7017 struct netdev_nested_priv priv = {
7019 .data = (void *)upper_dev,
7024 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev,
7029 * netdev_lower_get_next_private - Get the next ->private from the
7030 * lower neighbour list
7032 * @iter: list_head ** of the current position
7034 * Gets the next netdev_adjacent->private from the dev's lower neighbour
7035 * list, starting from iter position. The caller must hold either hold the
7036 * RTNL lock or its own locking that guarantees that the neighbour lower
7037 * list will remain unchanged.
7039 void *netdev_lower_get_next_private(struct net_device *dev,
7040 struct list_head **iter)
7042 struct netdev_adjacent *lower;
7044 lower = list_entry(*iter, struct netdev_adjacent, list);
7046 if (&lower->list == &dev->adj_list.lower)
7049 *iter = lower->list.next;
7051 return lower->private;
7053 EXPORT_SYMBOL(netdev_lower_get_next_private);
7056 * netdev_lower_get_next_private_rcu - Get the next ->private from the
7057 * lower neighbour list, RCU
7060 * @iter: list_head ** of the current position
7062 * Gets the next netdev_adjacent->private from the dev's lower neighbour
7063 * list, starting from iter position. The caller must hold RCU read lock.
7065 void *netdev_lower_get_next_private_rcu(struct net_device *dev,
7066 struct list_head **iter)
7068 struct netdev_adjacent *lower;
7070 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
7072 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7074 if (&lower->list == &dev->adj_list.lower)
7077 *iter = &lower->list;
7079 return lower->private;
7081 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu);
7084 * netdev_lower_get_next - Get the next device from the lower neighbour
7087 * @iter: list_head ** of the current position
7089 * Gets the next netdev_adjacent from the dev's lower neighbour
7090 * list, starting from iter position. The caller must hold RTNL lock or
7091 * its own locking that guarantees that the neighbour lower
7092 * list will remain unchanged.
7094 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter)
7096 struct netdev_adjacent *lower;
7098 lower = list_entry(*iter, struct netdev_adjacent, list);
7100 if (&lower->list == &dev->adj_list.lower)
7103 *iter = lower->list.next;
7107 EXPORT_SYMBOL(netdev_lower_get_next);
7109 static struct net_device *netdev_next_lower_dev(struct net_device *dev,
7110 struct list_head **iter)
7112 struct netdev_adjacent *lower;
7114 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
7116 if (&lower->list == &dev->adj_list.lower)
7119 *iter = &lower->list;
7124 static struct net_device *__netdev_next_lower_dev(struct net_device *dev,
7125 struct list_head **iter,
7128 struct netdev_adjacent *lower;
7130 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
7132 if (&lower->list == &dev->adj_list.lower)
7135 *iter = &lower->list;
7136 *ignore = lower->ignore;
7141 int netdev_walk_all_lower_dev(struct net_device *dev,
7142 int (*fn)(struct net_device *dev,
7143 struct netdev_nested_priv *priv),
7144 struct netdev_nested_priv *priv)
7146 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7147 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7151 iter = &dev->adj_list.lower;
7155 ret = fn(now, priv);
7162 ldev = netdev_next_lower_dev(now, &iter);
7167 niter = &ldev->adj_list.lower;
7168 dev_stack[cur] = now;
7169 iter_stack[cur++] = iter;
7176 next = dev_stack[--cur];
7177 niter = iter_stack[cur];
7186 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev);
7188 static int __netdev_walk_all_lower_dev(struct net_device *dev,
7189 int (*fn)(struct net_device *dev,
7190 struct netdev_nested_priv *priv),
7191 struct netdev_nested_priv *priv)
7193 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7194 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7199 iter = &dev->adj_list.lower;
7203 ret = fn(now, priv);
7210 ldev = __netdev_next_lower_dev(now, &iter, &ignore);
7217 niter = &ldev->adj_list.lower;
7218 dev_stack[cur] = now;
7219 iter_stack[cur++] = iter;
7226 next = dev_stack[--cur];
7227 niter = iter_stack[cur];
7237 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev,
7238 struct list_head **iter)
7240 struct netdev_adjacent *lower;
7242 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7243 if (&lower->list == &dev->adj_list.lower)
7246 *iter = &lower->list;
7250 EXPORT_SYMBOL(netdev_next_lower_dev_rcu);
7252 static u8 __netdev_upper_depth(struct net_device *dev)
7254 struct net_device *udev;
7255 struct list_head *iter;
7259 for (iter = &dev->adj_list.upper,
7260 udev = __netdev_next_upper_dev(dev, &iter, &ignore);
7262 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) {
7265 if (max_depth < udev->upper_level)
7266 max_depth = udev->upper_level;
7272 static u8 __netdev_lower_depth(struct net_device *dev)
7274 struct net_device *ldev;
7275 struct list_head *iter;
7279 for (iter = &dev->adj_list.lower,
7280 ldev = __netdev_next_lower_dev(dev, &iter, &ignore);
7282 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) {
7285 if (max_depth < ldev->lower_level)
7286 max_depth = ldev->lower_level;
7292 static int __netdev_update_upper_level(struct net_device *dev,
7293 struct netdev_nested_priv *__unused)
7295 dev->upper_level = __netdev_upper_depth(dev) + 1;
7299 #ifdef CONFIG_LOCKDEP
7300 static LIST_HEAD(net_unlink_list);
7302 static void net_unlink_todo(struct net_device *dev)
7304 if (list_empty(&dev->unlink_list))
7305 list_add_tail(&dev->unlink_list, &net_unlink_list);
7309 static int __netdev_update_lower_level(struct net_device *dev,
7310 struct netdev_nested_priv *priv)
7312 dev->lower_level = __netdev_lower_depth(dev) + 1;
7314 #ifdef CONFIG_LOCKDEP
7318 if (priv->flags & NESTED_SYNC_IMM)
7319 dev->nested_level = dev->lower_level - 1;
7320 if (priv->flags & NESTED_SYNC_TODO)
7321 net_unlink_todo(dev);
7326 int netdev_walk_all_lower_dev_rcu(struct net_device *dev,
7327 int (*fn)(struct net_device *dev,
7328 struct netdev_nested_priv *priv),
7329 struct netdev_nested_priv *priv)
7331 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7332 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7336 iter = &dev->adj_list.lower;
7340 ret = fn(now, priv);
7347 ldev = netdev_next_lower_dev_rcu(now, &iter);
7352 niter = &ldev->adj_list.lower;
7353 dev_stack[cur] = now;
7354 iter_stack[cur++] = iter;
7361 next = dev_stack[--cur];
7362 niter = iter_stack[cur];
7371 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu);
7374 * netdev_lower_get_first_private_rcu - Get the first ->private from the
7375 * lower neighbour list, RCU
7379 * Gets the first netdev_adjacent->private from the dev's lower neighbour
7380 * list. The caller must hold RCU read lock.
7382 void *netdev_lower_get_first_private_rcu(struct net_device *dev)
7384 struct netdev_adjacent *lower;
7386 lower = list_first_or_null_rcu(&dev->adj_list.lower,
7387 struct netdev_adjacent, list);
7389 return lower->private;
7392 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu);
7395 * netdev_master_upper_dev_get_rcu - Get master upper device
7398 * Find a master upper device and return pointer to it or NULL in case
7399 * it's not there. The caller must hold the RCU read lock.
7401 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev)
7403 struct netdev_adjacent *upper;
7405 upper = list_first_or_null_rcu(&dev->adj_list.upper,
7406 struct netdev_adjacent, list);
7407 if (upper && likely(upper->master))
7411 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu);
7413 static int netdev_adjacent_sysfs_add(struct net_device *dev,
7414 struct net_device *adj_dev,
7415 struct list_head *dev_list)
7417 char linkname[IFNAMSIZ+7];
7419 sprintf(linkname, dev_list == &dev->adj_list.upper ?
7420 "upper_%s" : "lower_%s", adj_dev->name);
7421 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj),
7424 static void netdev_adjacent_sysfs_del(struct net_device *dev,
7426 struct list_head *dev_list)
7428 char linkname[IFNAMSIZ+7];
7430 sprintf(linkname, dev_list == &dev->adj_list.upper ?
7431 "upper_%s" : "lower_%s", name);
7432 sysfs_remove_link(&(dev->dev.kobj), linkname);
7435 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev,
7436 struct net_device *adj_dev,
7437 struct list_head *dev_list)
7439 return (dev_list == &dev->adj_list.upper ||
7440 dev_list == &dev->adj_list.lower) &&
7441 net_eq(dev_net(dev), dev_net(adj_dev));
7444 static int __netdev_adjacent_dev_insert(struct net_device *dev,
7445 struct net_device *adj_dev,
7446 struct list_head *dev_list,
7447 void *private, bool master)
7449 struct netdev_adjacent *adj;
7452 adj = __netdev_find_adj(adj_dev, dev_list);
7456 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n",
7457 dev->name, adj_dev->name, adj->ref_nr);
7462 adj = kmalloc(sizeof(*adj), GFP_KERNEL);
7467 adj->master = master;
7469 adj->private = private;
7470 adj->ignore = false;
7471 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL);
7473 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n",
7474 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name);
7476 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) {
7477 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list);
7482 /* Ensure that master link is always the first item in list. */
7484 ret = sysfs_create_link(&(dev->dev.kobj),
7485 &(adj_dev->dev.kobj), "master");
7487 goto remove_symlinks;
7489 list_add_rcu(&adj->list, dev_list);
7491 list_add_tail_rcu(&adj->list, dev_list);
7497 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
7498 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
7500 netdev_put(adj_dev, &adj->dev_tracker);
7506 static void __netdev_adjacent_dev_remove(struct net_device *dev,
7507 struct net_device *adj_dev,
7509 struct list_head *dev_list)
7511 struct netdev_adjacent *adj;
7513 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n",
7514 dev->name, adj_dev->name, ref_nr);
7516 adj = __netdev_find_adj(adj_dev, dev_list);
7519 pr_err("Adjacency does not exist for device %s from %s\n",
7520 dev->name, adj_dev->name);
7525 if (adj->ref_nr > ref_nr) {
7526 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n",
7527 dev->name, adj_dev->name, ref_nr,
7528 adj->ref_nr - ref_nr);
7529 adj->ref_nr -= ref_nr;
7534 sysfs_remove_link(&(dev->dev.kobj), "master");
7536 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
7537 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
7539 list_del_rcu(&adj->list);
7540 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n",
7541 adj_dev->name, dev->name, adj_dev->name);
7542 netdev_put(adj_dev, &adj->dev_tracker);
7543 kfree_rcu(adj, rcu);
7546 static int __netdev_adjacent_dev_link_lists(struct net_device *dev,
7547 struct net_device *upper_dev,
7548 struct list_head *up_list,
7549 struct list_head *down_list,
7550 void *private, bool master)
7554 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list,
7559 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list,
7562 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list);
7569 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev,
7570 struct net_device *upper_dev,
7572 struct list_head *up_list,
7573 struct list_head *down_list)
7575 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list);
7576 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list);
7579 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev,
7580 struct net_device *upper_dev,
7581 void *private, bool master)
7583 return __netdev_adjacent_dev_link_lists(dev, upper_dev,
7584 &dev->adj_list.upper,
7585 &upper_dev->adj_list.lower,
7589 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev,
7590 struct net_device *upper_dev)
7592 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1,
7593 &dev->adj_list.upper,
7594 &upper_dev->adj_list.lower);
7597 static int __netdev_upper_dev_link(struct net_device *dev,
7598 struct net_device *upper_dev, bool master,
7599 void *upper_priv, void *upper_info,
7600 struct netdev_nested_priv *priv,
7601 struct netlink_ext_ack *extack)
7603 struct netdev_notifier_changeupper_info changeupper_info = {
7608 .upper_dev = upper_dev,
7611 .upper_info = upper_info,
7613 struct net_device *master_dev;
7618 if (dev == upper_dev)
7621 /* To prevent loops, check if dev is not upper device to upper_dev. */
7622 if (__netdev_has_upper_dev(upper_dev, dev))
7625 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV)
7629 if (__netdev_has_upper_dev(dev, upper_dev))
7632 master_dev = __netdev_master_upper_dev_get(dev);
7634 return master_dev == upper_dev ? -EEXIST : -EBUSY;
7637 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
7638 &changeupper_info.info);
7639 ret = notifier_to_errno(ret);
7643 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv,
7648 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
7649 &changeupper_info.info);
7650 ret = notifier_to_errno(ret);
7654 __netdev_update_upper_level(dev, NULL);
7655 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL);
7657 __netdev_update_lower_level(upper_dev, priv);
7658 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level,
7664 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
7670 * netdev_upper_dev_link - Add a link to the upper device
7672 * @upper_dev: new upper device
7673 * @extack: netlink extended ack
7675 * Adds a link to device which is upper to this one. The caller must hold
7676 * the RTNL lock. On a failure a negative errno code is returned.
7677 * On success the reference counts are adjusted and the function
7680 int netdev_upper_dev_link(struct net_device *dev,
7681 struct net_device *upper_dev,
7682 struct netlink_ext_ack *extack)
7684 struct netdev_nested_priv priv = {
7685 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO,
7689 return __netdev_upper_dev_link(dev, upper_dev, false,
7690 NULL, NULL, &priv, extack);
7692 EXPORT_SYMBOL(netdev_upper_dev_link);
7695 * netdev_master_upper_dev_link - Add a master link to the upper device
7697 * @upper_dev: new upper device
7698 * @upper_priv: upper device private
7699 * @upper_info: upper info to be passed down via notifier
7700 * @extack: netlink extended ack
7702 * Adds a link to device which is upper to this one. In this case, only
7703 * one master upper device can be linked, although other non-master devices
7704 * might be linked as well. The caller must hold the RTNL lock.
7705 * On a failure a negative errno code is returned. On success the reference
7706 * counts are adjusted and the function returns zero.
7708 int netdev_master_upper_dev_link(struct net_device *dev,
7709 struct net_device *upper_dev,
7710 void *upper_priv, void *upper_info,
7711 struct netlink_ext_ack *extack)
7713 struct netdev_nested_priv priv = {
7714 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO,
7718 return __netdev_upper_dev_link(dev, upper_dev, true,
7719 upper_priv, upper_info, &priv, extack);
7721 EXPORT_SYMBOL(netdev_master_upper_dev_link);
7723 static void __netdev_upper_dev_unlink(struct net_device *dev,
7724 struct net_device *upper_dev,
7725 struct netdev_nested_priv *priv)
7727 struct netdev_notifier_changeupper_info changeupper_info = {
7731 .upper_dev = upper_dev,
7737 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev;
7739 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
7740 &changeupper_info.info);
7742 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
7744 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
7745 &changeupper_info.info);
7747 __netdev_update_upper_level(dev, NULL);
7748 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL);
7750 __netdev_update_lower_level(upper_dev, priv);
7751 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level,
7756 * netdev_upper_dev_unlink - Removes a link to upper device
7758 * @upper_dev: new upper device
7760 * Removes a link to device which is upper to this one. The caller must hold
7763 void netdev_upper_dev_unlink(struct net_device *dev,
7764 struct net_device *upper_dev)
7766 struct netdev_nested_priv priv = {
7767 .flags = NESTED_SYNC_TODO,
7771 __netdev_upper_dev_unlink(dev, upper_dev, &priv);
7773 EXPORT_SYMBOL(netdev_upper_dev_unlink);
7775 static void __netdev_adjacent_dev_set(struct net_device *upper_dev,
7776 struct net_device *lower_dev,
7779 struct netdev_adjacent *adj;
7781 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower);
7785 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper);
7790 static void netdev_adjacent_dev_disable(struct net_device *upper_dev,
7791 struct net_device *lower_dev)
7793 __netdev_adjacent_dev_set(upper_dev, lower_dev, true);
7796 static void netdev_adjacent_dev_enable(struct net_device *upper_dev,
7797 struct net_device *lower_dev)
7799 __netdev_adjacent_dev_set(upper_dev, lower_dev, false);
7802 int netdev_adjacent_change_prepare(struct net_device *old_dev,
7803 struct net_device *new_dev,
7804 struct net_device *dev,
7805 struct netlink_ext_ack *extack)
7807 struct netdev_nested_priv priv = {
7816 if (old_dev && new_dev != old_dev)
7817 netdev_adjacent_dev_disable(dev, old_dev);
7818 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv,
7821 if (old_dev && new_dev != old_dev)
7822 netdev_adjacent_dev_enable(dev, old_dev);
7828 EXPORT_SYMBOL(netdev_adjacent_change_prepare);
7830 void netdev_adjacent_change_commit(struct net_device *old_dev,
7831 struct net_device *new_dev,
7832 struct net_device *dev)
7834 struct netdev_nested_priv priv = {
7835 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO,
7839 if (!new_dev || !old_dev)
7842 if (new_dev == old_dev)
7845 netdev_adjacent_dev_enable(dev, old_dev);
7846 __netdev_upper_dev_unlink(old_dev, dev, &priv);
7848 EXPORT_SYMBOL(netdev_adjacent_change_commit);
7850 void netdev_adjacent_change_abort(struct net_device *old_dev,
7851 struct net_device *new_dev,
7852 struct net_device *dev)
7854 struct netdev_nested_priv priv = {
7862 if (old_dev && new_dev != old_dev)
7863 netdev_adjacent_dev_enable(dev, old_dev);
7865 __netdev_upper_dev_unlink(new_dev, dev, &priv);
7867 EXPORT_SYMBOL(netdev_adjacent_change_abort);
7870 * netdev_bonding_info_change - Dispatch event about slave change
7872 * @bonding_info: info to dispatch
7874 * Send NETDEV_BONDING_INFO to netdev notifiers with info.
7875 * The caller must hold the RTNL lock.
7877 void netdev_bonding_info_change(struct net_device *dev,
7878 struct netdev_bonding_info *bonding_info)
7880 struct netdev_notifier_bonding_info info = {
7884 memcpy(&info.bonding_info, bonding_info,
7885 sizeof(struct netdev_bonding_info));
7886 call_netdevice_notifiers_info(NETDEV_BONDING_INFO,
7889 EXPORT_SYMBOL(netdev_bonding_info_change);
7891 static int netdev_offload_xstats_enable_l3(struct net_device *dev,
7892 struct netlink_ext_ack *extack)
7894 struct netdev_notifier_offload_xstats_info info = {
7896 .info.extack = extack,
7897 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3,
7902 dev->offload_xstats_l3 = kzalloc(sizeof(*dev->offload_xstats_l3),
7904 if (!dev->offload_xstats_l3)
7907 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE,
7908 NETDEV_OFFLOAD_XSTATS_DISABLE,
7910 err = notifier_to_errno(rc);
7917 kfree(dev->offload_xstats_l3);
7918 dev->offload_xstats_l3 = NULL;
7922 int netdev_offload_xstats_enable(struct net_device *dev,
7923 enum netdev_offload_xstats_type type,
7924 struct netlink_ext_ack *extack)
7928 if (netdev_offload_xstats_enabled(dev, type))
7932 case NETDEV_OFFLOAD_XSTATS_TYPE_L3:
7933 return netdev_offload_xstats_enable_l3(dev, extack);
7939 EXPORT_SYMBOL(netdev_offload_xstats_enable);
7941 static void netdev_offload_xstats_disable_l3(struct net_device *dev)
7943 struct netdev_notifier_offload_xstats_info info = {
7945 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3,
7948 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE,
7950 kfree(dev->offload_xstats_l3);
7951 dev->offload_xstats_l3 = NULL;
7954 int netdev_offload_xstats_disable(struct net_device *dev,
7955 enum netdev_offload_xstats_type type)
7959 if (!netdev_offload_xstats_enabled(dev, type))
7963 case NETDEV_OFFLOAD_XSTATS_TYPE_L3:
7964 netdev_offload_xstats_disable_l3(dev);
7971 EXPORT_SYMBOL(netdev_offload_xstats_disable);
7973 static void netdev_offload_xstats_disable_all(struct net_device *dev)
7975 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3);
7978 static struct rtnl_hw_stats64 *
7979 netdev_offload_xstats_get_ptr(const struct net_device *dev,
7980 enum netdev_offload_xstats_type type)
7983 case NETDEV_OFFLOAD_XSTATS_TYPE_L3:
7984 return dev->offload_xstats_l3;
7991 bool netdev_offload_xstats_enabled(const struct net_device *dev,
7992 enum netdev_offload_xstats_type type)
7996 return netdev_offload_xstats_get_ptr(dev, type);
7998 EXPORT_SYMBOL(netdev_offload_xstats_enabled);
8000 struct netdev_notifier_offload_xstats_ru {
8004 struct netdev_notifier_offload_xstats_rd {
8005 struct rtnl_hw_stats64 stats;
8009 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest,
8010 const struct rtnl_hw_stats64 *src)
8012 dest->rx_packets += src->rx_packets;
8013 dest->tx_packets += src->tx_packets;
8014 dest->rx_bytes += src->rx_bytes;
8015 dest->tx_bytes += src->tx_bytes;
8016 dest->rx_errors += src->rx_errors;
8017 dest->tx_errors += src->tx_errors;
8018 dest->rx_dropped += src->rx_dropped;
8019 dest->tx_dropped += src->tx_dropped;
8020 dest->multicast += src->multicast;
8023 static int netdev_offload_xstats_get_used(struct net_device *dev,
8024 enum netdev_offload_xstats_type type,
8026 struct netlink_ext_ack *extack)
8028 struct netdev_notifier_offload_xstats_ru report_used = {};
8029 struct netdev_notifier_offload_xstats_info info = {
8031 .info.extack = extack,
8033 .report_used = &report_used,
8037 WARN_ON(!netdev_offload_xstats_enabled(dev, type));
8038 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED,
8040 *p_used = report_used.used;
8041 return notifier_to_errno(rc);
8044 static int netdev_offload_xstats_get_stats(struct net_device *dev,
8045 enum netdev_offload_xstats_type type,
8046 struct rtnl_hw_stats64 *p_stats,
8048 struct netlink_ext_ack *extack)
8050 struct netdev_notifier_offload_xstats_rd report_delta = {};
8051 struct netdev_notifier_offload_xstats_info info = {
8053 .info.extack = extack,
8055 .report_delta = &report_delta,
8057 struct rtnl_hw_stats64 *stats;
8060 stats = netdev_offload_xstats_get_ptr(dev, type);
8061 if (WARN_ON(!stats))
8064 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA,
8067 /* Cache whatever we got, even if there was an error, otherwise the
8068 * successful stats retrievals would get lost.
8070 netdev_hw_stats64_add(stats, &report_delta.stats);
8074 *p_used = report_delta.used;
8076 return notifier_to_errno(rc);
8079 int netdev_offload_xstats_get(struct net_device *dev,
8080 enum netdev_offload_xstats_type type,
8081 struct rtnl_hw_stats64 *p_stats, bool *p_used,
8082 struct netlink_ext_ack *extack)
8087 return netdev_offload_xstats_get_stats(dev, type, p_stats,
8090 return netdev_offload_xstats_get_used(dev, type, p_used,
8093 EXPORT_SYMBOL(netdev_offload_xstats_get);
8096 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta,
8097 const struct rtnl_hw_stats64 *stats)
8099 report_delta->used = true;
8100 netdev_hw_stats64_add(&report_delta->stats, stats);
8102 EXPORT_SYMBOL(netdev_offload_xstats_report_delta);
8105 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used)
8107 report_used->used = true;
8109 EXPORT_SYMBOL(netdev_offload_xstats_report_used);
8111 void netdev_offload_xstats_push_delta(struct net_device *dev,
8112 enum netdev_offload_xstats_type type,
8113 const struct rtnl_hw_stats64 *p_stats)
8115 struct rtnl_hw_stats64 *stats;
8119 stats = netdev_offload_xstats_get_ptr(dev, type);
8120 if (WARN_ON(!stats))
8123 netdev_hw_stats64_add(stats, p_stats);
8125 EXPORT_SYMBOL(netdev_offload_xstats_push_delta);
8128 * netdev_get_xmit_slave - Get the xmit slave of master device
8131 * @all_slaves: assume all the slaves are active
8133 * The reference counters are not incremented so the caller must be
8134 * careful with locks. The caller must hold RCU lock.
8135 * %NULL is returned if no slave is found.
8138 struct net_device *netdev_get_xmit_slave(struct net_device *dev,
8139 struct sk_buff *skb,
8142 const struct net_device_ops *ops = dev->netdev_ops;
8144 if (!ops->ndo_get_xmit_slave)
8146 return ops->ndo_get_xmit_slave(dev, skb, all_slaves);
8148 EXPORT_SYMBOL(netdev_get_xmit_slave);
8150 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev,
8153 const struct net_device_ops *ops = dev->netdev_ops;
8155 if (!ops->ndo_sk_get_lower_dev)
8157 return ops->ndo_sk_get_lower_dev(dev, sk);
8161 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket
8165 * %NULL is returned if no lower device is found.
8168 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev,
8171 struct net_device *lower;
8173 lower = netdev_sk_get_lower_dev(dev, sk);
8176 lower = netdev_sk_get_lower_dev(dev, sk);
8181 EXPORT_SYMBOL(netdev_sk_get_lowest_dev);
8183 static void netdev_adjacent_add_links(struct net_device *dev)
8185 struct netdev_adjacent *iter;
8187 struct net *net = dev_net(dev);
8189 list_for_each_entry(iter, &dev->adj_list.upper, list) {
8190 if (!net_eq(net, dev_net(iter->dev)))
8192 netdev_adjacent_sysfs_add(iter->dev, dev,
8193 &iter->dev->adj_list.lower);
8194 netdev_adjacent_sysfs_add(dev, iter->dev,
8195 &dev->adj_list.upper);
8198 list_for_each_entry(iter, &dev->adj_list.lower, list) {
8199 if (!net_eq(net, dev_net(iter->dev)))
8201 netdev_adjacent_sysfs_add(iter->dev, dev,
8202 &iter->dev->adj_list.upper);
8203 netdev_adjacent_sysfs_add(dev, iter->dev,
8204 &dev->adj_list.lower);
8208 static void netdev_adjacent_del_links(struct net_device *dev)
8210 struct netdev_adjacent *iter;
8212 struct net *net = dev_net(dev);
8214 list_for_each_entry(iter, &dev->adj_list.upper, list) {
8215 if (!net_eq(net, dev_net(iter->dev)))
8217 netdev_adjacent_sysfs_del(iter->dev, dev->name,
8218 &iter->dev->adj_list.lower);
8219 netdev_adjacent_sysfs_del(dev, iter->dev->name,
8220 &dev->adj_list.upper);
8223 list_for_each_entry(iter, &dev->adj_list.lower, list) {
8224 if (!net_eq(net, dev_net(iter->dev)))
8226 netdev_adjacent_sysfs_del(iter->dev, dev->name,
8227 &iter->dev->adj_list.upper);
8228 netdev_adjacent_sysfs_del(dev, iter->dev->name,
8229 &dev->adj_list.lower);
8233 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname)
8235 struct netdev_adjacent *iter;
8237 struct net *net = dev_net(dev);
8239 list_for_each_entry(iter, &dev->adj_list.upper, list) {
8240 if (!net_eq(net, dev_net(iter->dev)))
8242 netdev_adjacent_sysfs_del(iter->dev, oldname,
8243 &iter->dev->adj_list.lower);
8244 netdev_adjacent_sysfs_add(iter->dev, dev,
8245 &iter->dev->adj_list.lower);
8248 list_for_each_entry(iter, &dev->adj_list.lower, list) {
8249 if (!net_eq(net, dev_net(iter->dev)))
8251 netdev_adjacent_sysfs_del(iter->dev, oldname,
8252 &iter->dev->adj_list.upper);
8253 netdev_adjacent_sysfs_add(iter->dev, dev,
8254 &iter->dev->adj_list.upper);
8258 void *netdev_lower_dev_get_private(struct net_device *dev,
8259 struct net_device *lower_dev)
8261 struct netdev_adjacent *lower;
8265 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower);
8269 return lower->private;
8271 EXPORT_SYMBOL(netdev_lower_dev_get_private);
8275 * netdev_lower_state_changed - Dispatch event about lower device state change
8276 * @lower_dev: device
8277 * @lower_state_info: state to dispatch
8279 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info.
8280 * The caller must hold the RTNL lock.
8282 void netdev_lower_state_changed(struct net_device *lower_dev,
8283 void *lower_state_info)
8285 struct netdev_notifier_changelowerstate_info changelowerstate_info = {
8286 .info.dev = lower_dev,
8290 changelowerstate_info.lower_state_info = lower_state_info;
8291 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE,
8292 &changelowerstate_info.info);
8294 EXPORT_SYMBOL(netdev_lower_state_changed);
8296 static void dev_change_rx_flags(struct net_device *dev, int flags)
8298 const struct net_device_ops *ops = dev->netdev_ops;
8300 if (ops->ndo_change_rx_flags)
8301 ops->ndo_change_rx_flags(dev, flags);
8304 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify)
8306 unsigned int old_flags = dev->flags;
8312 dev->flags |= IFF_PROMISC;
8313 dev->promiscuity += inc;
8314 if (dev->promiscuity == 0) {
8317 * If inc causes overflow, untouch promisc and return error.
8320 dev->flags &= ~IFF_PROMISC;
8322 dev->promiscuity -= inc;
8323 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n");
8327 if (dev->flags != old_flags) {
8328 netdev_info(dev, "%s promiscuous mode\n",
8329 dev->flags & IFF_PROMISC ? "entered" : "left");
8330 if (audit_enabled) {
8331 current_uid_gid(&uid, &gid);
8332 audit_log(audit_context(), GFP_ATOMIC,
8333 AUDIT_ANOM_PROMISCUOUS,
8334 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u",
8335 dev->name, (dev->flags & IFF_PROMISC),
8336 (old_flags & IFF_PROMISC),
8337 from_kuid(&init_user_ns, audit_get_loginuid(current)),
8338 from_kuid(&init_user_ns, uid),
8339 from_kgid(&init_user_ns, gid),
8340 audit_get_sessionid(current));
8343 dev_change_rx_flags(dev, IFF_PROMISC);
8346 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL);
8351 * dev_set_promiscuity - update promiscuity count on a device
8355 * Add or remove promiscuity from a device. While the count in the device
8356 * remains above zero the interface remains promiscuous. Once it hits zero
8357 * the device reverts back to normal filtering operation. A negative inc
8358 * value is used to drop promiscuity on the device.
8359 * Return 0 if successful or a negative errno code on error.
8361 int dev_set_promiscuity(struct net_device *dev, int inc)
8363 unsigned int old_flags = dev->flags;
8366 err = __dev_set_promiscuity(dev, inc, true);
8369 if (dev->flags != old_flags)
8370 dev_set_rx_mode(dev);
8373 EXPORT_SYMBOL(dev_set_promiscuity);
8375 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify)
8377 unsigned int old_flags = dev->flags, old_gflags = dev->gflags;
8381 dev->flags |= IFF_ALLMULTI;
8382 dev->allmulti += inc;
8383 if (dev->allmulti == 0) {
8386 * If inc causes overflow, untouch allmulti and return error.
8389 dev->flags &= ~IFF_ALLMULTI;
8391 dev->allmulti -= inc;
8392 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n");
8396 if (dev->flags ^ old_flags) {
8397 netdev_info(dev, "%s allmulticast mode\n",
8398 dev->flags & IFF_ALLMULTI ? "entered" : "left");
8399 dev_change_rx_flags(dev, IFF_ALLMULTI);
8400 dev_set_rx_mode(dev);
8402 __dev_notify_flags(dev, old_flags,
8403 dev->gflags ^ old_gflags, 0, NULL);
8409 * dev_set_allmulti - update allmulti count on a device
8413 * Add or remove reception of all multicast frames to a device. While the
8414 * count in the device remains above zero the interface remains listening
8415 * to all interfaces. Once it hits zero the device reverts back to normal
8416 * filtering operation. A negative @inc value is used to drop the counter
8417 * when releasing a resource needing all multicasts.
8418 * Return 0 if successful or a negative errno code on error.
8421 int dev_set_allmulti(struct net_device *dev, int inc)
8423 return __dev_set_allmulti(dev, inc, true);
8425 EXPORT_SYMBOL(dev_set_allmulti);
8428 * Upload unicast and multicast address lists to device and
8429 * configure RX filtering. When the device doesn't support unicast
8430 * filtering it is put in promiscuous mode while unicast addresses
8433 void __dev_set_rx_mode(struct net_device *dev)
8435 const struct net_device_ops *ops = dev->netdev_ops;
8437 /* dev_open will call this function so the list will stay sane. */
8438 if (!(dev->flags&IFF_UP))
8441 if (!netif_device_present(dev))
8444 if (!(dev->priv_flags & IFF_UNICAST_FLT)) {
8445 /* Unicast addresses changes may only happen under the rtnl,
8446 * therefore calling __dev_set_promiscuity here is safe.
8448 if (!netdev_uc_empty(dev) && !dev->uc_promisc) {
8449 __dev_set_promiscuity(dev, 1, false);
8450 dev->uc_promisc = true;
8451 } else if (netdev_uc_empty(dev) && dev->uc_promisc) {
8452 __dev_set_promiscuity(dev, -1, false);
8453 dev->uc_promisc = false;
8457 if (ops->ndo_set_rx_mode)
8458 ops->ndo_set_rx_mode(dev);
8461 void dev_set_rx_mode(struct net_device *dev)
8463 netif_addr_lock_bh(dev);
8464 __dev_set_rx_mode(dev);
8465 netif_addr_unlock_bh(dev);
8469 * dev_get_flags - get flags reported to userspace
8472 * Get the combination of flag bits exported through APIs to userspace.
8474 unsigned int dev_get_flags(const struct net_device *dev)
8478 flags = (dev->flags & ~(IFF_PROMISC |
8483 (dev->gflags & (IFF_PROMISC |
8486 if (netif_running(dev)) {
8487 if (netif_oper_up(dev))
8488 flags |= IFF_RUNNING;
8489 if (netif_carrier_ok(dev))
8490 flags |= IFF_LOWER_UP;
8491 if (netif_dormant(dev))
8492 flags |= IFF_DORMANT;
8497 EXPORT_SYMBOL(dev_get_flags);
8499 int __dev_change_flags(struct net_device *dev, unsigned int flags,
8500 struct netlink_ext_ack *extack)
8502 unsigned int old_flags = dev->flags;
8508 * Set the flags on our device.
8511 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP |
8512 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL |
8514 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC |
8518 * Load in the correct multicast list now the flags have changed.
8521 if ((old_flags ^ flags) & IFF_MULTICAST)
8522 dev_change_rx_flags(dev, IFF_MULTICAST);
8524 dev_set_rx_mode(dev);
8527 * Have we downed the interface. We handle IFF_UP ourselves
8528 * according to user attempts to set it, rather than blindly
8533 if ((old_flags ^ flags) & IFF_UP) {
8534 if (old_flags & IFF_UP)
8537 ret = __dev_open(dev, extack);
8540 if ((flags ^ dev->gflags) & IFF_PROMISC) {
8541 int inc = (flags & IFF_PROMISC) ? 1 : -1;
8542 unsigned int old_flags = dev->flags;
8544 dev->gflags ^= IFF_PROMISC;
8546 if (__dev_set_promiscuity(dev, inc, false) >= 0)
8547 if (dev->flags != old_flags)
8548 dev_set_rx_mode(dev);
8551 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI
8552 * is important. Some (broken) drivers set IFF_PROMISC, when
8553 * IFF_ALLMULTI is requested not asking us and not reporting.
8555 if ((flags ^ dev->gflags) & IFF_ALLMULTI) {
8556 int inc = (flags & IFF_ALLMULTI) ? 1 : -1;
8558 dev->gflags ^= IFF_ALLMULTI;
8559 __dev_set_allmulti(dev, inc, false);
8565 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags,
8566 unsigned int gchanges, u32 portid,
8567 const struct nlmsghdr *nlh)
8569 unsigned int changes = dev->flags ^ old_flags;
8572 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh);
8574 if (changes & IFF_UP) {
8575 if (dev->flags & IFF_UP)
8576 call_netdevice_notifiers(NETDEV_UP, dev);
8578 call_netdevice_notifiers(NETDEV_DOWN, dev);
8581 if (dev->flags & IFF_UP &&
8582 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) {
8583 struct netdev_notifier_change_info change_info = {
8587 .flags_changed = changes,
8590 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info);
8595 * dev_change_flags - change device settings
8597 * @flags: device state flags
8598 * @extack: netlink extended ack
8600 * Change settings on device based state flags. The flags are
8601 * in the userspace exported format.
8603 int dev_change_flags(struct net_device *dev, unsigned int flags,
8604 struct netlink_ext_ack *extack)
8607 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags;
8609 ret = __dev_change_flags(dev, flags, extack);
8613 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags);
8614 __dev_notify_flags(dev, old_flags, changes, 0, NULL);
8617 EXPORT_SYMBOL(dev_change_flags);
8619 int __dev_set_mtu(struct net_device *dev, int new_mtu)
8621 const struct net_device_ops *ops = dev->netdev_ops;
8623 if (ops->ndo_change_mtu)
8624 return ops->ndo_change_mtu(dev, new_mtu);
8626 /* Pairs with all the lockless reads of dev->mtu in the stack */
8627 WRITE_ONCE(dev->mtu, new_mtu);
8630 EXPORT_SYMBOL(__dev_set_mtu);
8632 int dev_validate_mtu(struct net_device *dev, int new_mtu,
8633 struct netlink_ext_ack *extack)
8635 /* MTU must be positive, and in range */
8636 if (new_mtu < 0 || new_mtu < dev->min_mtu) {
8637 NL_SET_ERR_MSG(extack, "mtu less than device minimum");
8641 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) {
8642 NL_SET_ERR_MSG(extack, "mtu greater than device maximum");
8649 * dev_set_mtu_ext - Change maximum transfer unit
8651 * @new_mtu: new transfer unit
8652 * @extack: netlink extended ack
8654 * Change the maximum transfer size of the network device.
8656 int dev_set_mtu_ext(struct net_device *dev, int new_mtu,
8657 struct netlink_ext_ack *extack)
8661 if (new_mtu == dev->mtu)
8664 err = dev_validate_mtu(dev, new_mtu, extack);
8668 if (!netif_device_present(dev))
8671 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev);
8672 err = notifier_to_errno(err);
8676 orig_mtu = dev->mtu;
8677 err = __dev_set_mtu(dev, new_mtu);
8680 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev,
8682 err = notifier_to_errno(err);
8684 /* setting mtu back and notifying everyone again,
8685 * so that they have a chance to revert changes.
8687 __dev_set_mtu(dev, orig_mtu);
8688 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev,
8695 int dev_set_mtu(struct net_device *dev, int new_mtu)
8697 struct netlink_ext_ack extack;
8700 memset(&extack, 0, sizeof(extack));
8701 err = dev_set_mtu_ext(dev, new_mtu, &extack);
8702 if (err && extack._msg)
8703 net_err_ratelimited("%s: %s\n", dev->name, extack._msg);
8706 EXPORT_SYMBOL(dev_set_mtu);
8709 * dev_change_tx_queue_len - Change TX queue length of a netdevice
8711 * @new_len: new tx queue length
8713 int dev_change_tx_queue_len(struct net_device *dev, unsigned long new_len)
8715 unsigned int orig_len = dev->tx_queue_len;
8718 if (new_len != (unsigned int)new_len)
8721 if (new_len != orig_len) {
8722 dev->tx_queue_len = new_len;
8723 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev);
8724 res = notifier_to_errno(res);
8727 res = dev_qdisc_change_tx_queue_len(dev);
8735 netdev_err(dev, "refused to change device tx_queue_len\n");
8736 dev->tx_queue_len = orig_len;
8741 * dev_set_group - Change group this device belongs to
8743 * @new_group: group this device should belong to
8745 void dev_set_group(struct net_device *dev, int new_group)
8747 dev->group = new_group;
8751 * dev_pre_changeaddr_notify - Call NETDEV_PRE_CHANGEADDR.
8753 * @addr: new address
8754 * @extack: netlink extended ack
8756 int dev_pre_changeaddr_notify(struct net_device *dev, const char *addr,
8757 struct netlink_ext_ack *extack)
8759 struct netdev_notifier_pre_changeaddr_info info = {
8761 .info.extack = extack,
8766 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info);
8767 return notifier_to_errno(rc);
8769 EXPORT_SYMBOL(dev_pre_changeaddr_notify);
8772 * dev_set_mac_address - Change Media Access Control Address
8775 * @extack: netlink extended ack
8777 * Change the hardware (MAC) address of the device
8779 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa,
8780 struct netlink_ext_ack *extack)
8782 const struct net_device_ops *ops = dev->netdev_ops;
8785 if (!ops->ndo_set_mac_address)
8787 if (sa->sa_family != dev->type)
8789 if (!netif_device_present(dev))
8791 err = dev_pre_changeaddr_notify(dev, sa->sa_data, extack);
8794 err = ops->ndo_set_mac_address(dev, sa);
8797 dev->addr_assign_type = NET_ADDR_SET;
8798 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
8799 add_device_randomness(dev->dev_addr, dev->addr_len);
8802 EXPORT_SYMBOL(dev_set_mac_address);
8804 static DECLARE_RWSEM(dev_addr_sem);
8806 int dev_set_mac_address_user(struct net_device *dev, struct sockaddr *sa,
8807 struct netlink_ext_ack *extack)
8811 down_write(&dev_addr_sem);
8812 ret = dev_set_mac_address(dev, sa, extack);
8813 up_write(&dev_addr_sem);
8816 EXPORT_SYMBOL(dev_set_mac_address_user);
8818 int dev_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name)
8820 size_t size = sizeof(sa->sa_data_min);
8821 struct net_device *dev;
8824 down_read(&dev_addr_sem);
8827 dev = dev_get_by_name_rcu(net, dev_name);
8833 memset(sa->sa_data, 0, size);
8835 memcpy(sa->sa_data, dev->dev_addr,
8836 min_t(size_t, size, dev->addr_len));
8837 sa->sa_family = dev->type;
8841 up_read(&dev_addr_sem);
8844 EXPORT_SYMBOL(dev_get_mac_address);
8847 * dev_change_carrier - Change device carrier
8849 * @new_carrier: new value
8851 * Change device carrier
8853 int dev_change_carrier(struct net_device *dev, bool new_carrier)
8855 const struct net_device_ops *ops = dev->netdev_ops;
8857 if (!ops->ndo_change_carrier)
8859 if (!netif_device_present(dev))
8861 return ops->ndo_change_carrier(dev, new_carrier);
8865 * dev_get_phys_port_id - Get device physical port ID
8869 * Get device physical port ID
8871 int dev_get_phys_port_id(struct net_device *dev,
8872 struct netdev_phys_item_id *ppid)
8874 const struct net_device_ops *ops = dev->netdev_ops;
8876 if (!ops->ndo_get_phys_port_id)
8878 return ops->ndo_get_phys_port_id(dev, ppid);
8882 * dev_get_phys_port_name - Get device physical port name
8885 * @len: limit of bytes to copy to name
8887 * Get device physical port name
8889 int dev_get_phys_port_name(struct net_device *dev,
8890 char *name, size_t len)
8892 const struct net_device_ops *ops = dev->netdev_ops;
8895 if (ops->ndo_get_phys_port_name) {
8896 err = ops->ndo_get_phys_port_name(dev, name, len);
8897 if (err != -EOPNOTSUPP)
8900 return devlink_compat_phys_port_name_get(dev, name, len);
8904 * dev_get_port_parent_id - Get the device's port parent identifier
8905 * @dev: network device
8906 * @ppid: pointer to a storage for the port's parent identifier
8907 * @recurse: allow/disallow recursion to lower devices
8909 * Get the devices's port parent identifier
8911 int dev_get_port_parent_id(struct net_device *dev,
8912 struct netdev_phys_item_id *ppid,
8915 const struct net_device_ops *ops = dev->netdev_ops;
8916 struct netdev_phys_item_id first = { };
8917 struct net_device *lower_dev;
8918 struct list_head *iter;
8921 if (ops->ndo_get_port_parent_id) {
8922 err = ops->ndo_get_port_parent_id(dev, ppid);
8923 if (err != -EOPNOTSUPP)
8927 err = devlink_compat_switch_id_get(dev, ppid);
8928 if (!recurse || err != -EOPNOTSUPP)
8931 netdev_for_each_lower_dev(dev, lower_dev, iter) {
8932 err = dev_get_port_parent_id(lower_dev, ppid, true);
8937 else if (memcmp(&first, ppid, sizeof(*ppid)))
8943 EXPORT_SYMBOL(dev_get_port_parent_id);
8946 * netdev_port_same_parent_id - Indicate if two network devices have
8947 * the same port parent identifier
8948 * @a: first network device
8949 * @b: second network device
8951 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b)
8953 struct netdev_phys_item_id a_id = { };
8954 struct netdev_phys_item_id b_id = { };
8956 if (dev_get_port_parent_id(a, &a_id, true) ||
8957 dev_get_port_parent_id(b, &b_id, true))
8960 return netdev_phys_item_id_same(&a_id, &b_id);
8962 EXPORT_SYMBOL(netdev_port_same_parent_id);
8965 * dev_change_proto_down - set carrier according to proto_down.
8968 * @proto_down: new value
8970 int dev_change_proto_down(struct net_device *dev, bool proto_down)
8972 if (!(dev->priv_flags & IFF_CHANGE_PROTO_DOWN))
8974 if (!netif_device_present(dev))
8977 netif_carrier_off(dev);
8979 netif_carrier_on(dev);
8980 dev->proto_down = proto_down;
8985 * dev_change_proto_down_reason - proto down reason
8988 * @mask: proto down mask
8989 * @value: proto down value
8991 void dev_change_proto_down_reason(struct net_device *dev, unsigned long mask,
8997 dev->proto_down_reason = value;
8999 for_each_set_bit(b, &mask, 32) {
9000 if (value & (1 << b))
9001 dev->proto_down_reason |= BIT(b);
9003 dev->proto_down_reason &= ~BIT(b);
9008 struct bpf_xdp_link {
9009 struct bpf_link link;
9010 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */
9014 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags)
9016 if (flags & XDP_FLAGS_HW_MODE)
9018 if (flags & XDP_FLAGS_DRV_MODE)
9019 return XDP_MODE_DRV;
9020 if (flags & XDP_FLAGS_SKB_MODE)
9021 return XDP_MODE_SKB;
9022 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB;
9025 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode)
9029 return generic_xdp_install;
9032 return dev->netdev_ops->ndo_bpf;
9038 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev,
9039 enum bpf_xdp_mode mode)
9041 return dev->xdp_state[mode].link;
9044 static struct bpf_prog *dev_xdp_prog(struct net_device *dev,
9045 enum bpf_xdp_mode mode)
9047 struct bpf_xdp_link *link = dev_xdp_link(dev, mode);
9050 return link->link.prog;
9051 return dev->xdp_state[mode].prog;
9054 u8 dev_xdp_prog_count(struct net_device *dev)
9059 for (i = 0; i < __MAX_XDP_MODE; i++)
9060 if (dev->xdp_state[i].prog || dev->xdp_state[i].link)
9064 EXPORT_SYMBOL_GPL(dev_xdp_prog_count);
9066 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode)
9068 struct bpf_prog *prog = dev_xdp_prog(dev, mode);
9070 return prog ? prog->aux->id : 0;
9073 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode,
9074 struct bpf_xdp_link *link)
9076 dev->xdp_state[mode].link = link;
9077 dev->xdp_state[mode].prog = NULL;
9080 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode,
9081 struct bpf_prog *prog)
9083 dev->xdp_state[mode].link = NULL;
9084 dev->xdp_state[mode].prog = prog;
9087 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode,
9088 bpf_op_t bpf_op, struct netlink_ext_ack *extack,
9089 u32 flags, struct bpf_prog *prog)
9091 struct netdev_bpf xdp;
9094 memset(&xdp, 0, sizeof(xdp));
9095 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG;
9096 xdp.extack = extack;
9100 /* Drivers assume refcnt is already incremented (i.e, prog pointer is
9101 * "moved" into driver), so they don't increment it on their own, but
9102 * they do decrement refcnt when program is detached or replaced.
9103 * Given net_device also owns link/prog, we need to bump refcnt here
9104 * to prevent drivers from underflowing it.
9108 err = bpf_op(dev, &xdp);
9115 if (mode != XDP_MODE_HW)
9116 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog);
9121 static void dev_xdp_uninstall(struct net_device *dev)
9123 struct bpf_xdp_link *link;
9124 struct bpf_prog *prog;
9125 enum bpf_xdp_mode mode;
9130 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) {
9131 prog = dev_xdp_prog(dev, mode);
9135 bpf_op = dev_xdp_bpf_op(dev, mode);
9139 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL));
9141 /* auto-detach link from net device */
9142 link = dev_xdp_link(dev, mode);
9148 dev_xdp_set_link(dev, mode, NULL);
9152 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack,
9153 struct bpf_xdp_link *link, struct bpf_prog *new_prog,
9154 struct bpf_prog *old_prog, u32 flags)
9156 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES);
9157 struct bpf_prog *cur_prog;
9158 struct net_device *upper;
9159 struct list_head *iter;
9160 enum bpf_xdp_mode mode;
9166 /* either link or prog attachment, never both */
9167 if (link && (new_prog || old_prog))
9169 /* link supports only XDP mode flags */
9170 if (link && (flags & ~XDP_FLAGS_MODES)) {
9171 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment");
9174 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */
9175 if (num_modes > 1) {
9176 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set");
9179 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */
9180 if (!num_modes && dev_xdp_prog_count(dev) > 1) {
9181 NL_SET_ERR_MSG(extack,
9182 "More than one program loaded, unset mode is ambiguous");
9185 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */
9186 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) {
9187 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified");
9191 mode = dev_xdp_mode(dev, flags);
9192 /* can't replace attached link */
9193 if (dev_xdp_link(dev, mode)) {
9194 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link");
9198 /* don't allow if an upper device already has a program */
9199 netdev_for_each_upper_dev_rcu(dev, upper, iter) {
9200 if (dev_xdp_prog_count(upper) > 0) {
9201 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program");
9206 cur_prog = dev_xdp_prog(dev, mode);
9207 /* can't replace attached prog with link */
9208 if (link && cur_prog) {
9209 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link");
9212 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) {
9213 NL_SET_ERR_MSG(extack, "Active program does not match expected");
9217 /* put effective new program into new_prog */
9219 new_prog = link->link.prog;
9222 bool offload = mode == XDP_MODE_HW;
9223 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB
9224 ? XDP_MODE_DRV : XDP_MODE_SKB;
9226 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) {
9227 NL_SET_ERR_MSG(extack, "XDP program already attached");
9230 if (!offload && dev_xdp_prog(dev, other_mode)) {
9231 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time");
9234 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) {
9235 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported");
9238 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) {
9239 NL_SET_ERR_MSG(extack, "Program bound to different device");
9242 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) {
9243 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device");
9246 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) {
9247 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device");
9252 /* don't call drivers if the effective program didn't change */
9253 if (new_prog != cur_prog) {
9254 bpf_op = dev_xdp_bpf_op(dev, mode);
9256 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode");
9260 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog);
9266 dev_xdp_set_link(dev, mode, link);
9268 dev_xdp_set_prog(dev, mode, new_prog);
9270 bpf_prog_put(cur_prog);
9275 static int dev_xdp_attach_link(struct net_device *dev,
9276 struct netlink_ext_ack *extack,
9277 struct bpf_xdp_link *link)
9279 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags);
9282 static int dev_xdp_detach_link(struct net_device *dev,
9283 struct netlink_ext_ack *extack,
9284 struct bpf_xdp_link *link)
9286 enum bpf_xdp_mode mode;
9291 mode = dev_xdp_mode(dev, link->flags);
9292 if (dev_xdp_link(dev, mode) != link)
9295 bpf_op = dev_xdp_bpf_op(dev, mode);
9296 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL));
9297 dev_xdp_set_link(dev, mode, NULL);
9301 static void bpf_xdp_link_release(struct bpf_link *link)
9303 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9307 /* if racing with net_device's tear down, xdp_link->dev might be
9308 * already NULL, in which case link was already auto-detached
9310 if (xdp_link->dev) {
9311 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link));
9312 xdp_link->dev = NULL;
9318 static int bpf_xdp_link_detach(struct bpf_link *link)
9320 bpf_xdp_link_release(link);
9324 static void bpf_xdp_link_dealloc(struct bpf_link *link)
9326 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9331 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link,
9332 struct seq_file *seq)
9334 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9339 ifindex = xdp_link->dev->ifindex;
9342 seq_printf(seq, "ifindex:\t%u\n", ifindex);
9345 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link,
9346 struct bpf_link_info *info)
9348 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9353 ifindex = xdp_link->dev->ifindex;
9356 info->xdp.ifindex = ifindex;
9360 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog,
9361 struct bpf_prog *old_prog)
9363 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9364 enum bpf_xdp_mode mode;
9370 /* link might have been auto-released already, so fail */
9371 if (!xdp_link->dev) {
9376 if (old_prog && link->prog != old_prog) {
9380 old_prog = link->prog;
9381 if (old_prog->type != new_prog->type ||
9382 old_prog->expected_attach_type != new_prog->expected_attach_type) {
9387 if (old_prog == new_prog) {
9388 /* no-op, don't disturb drivers */
9389 bpf_prog_put(new_prog);
9393 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags);
9394 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode);
9395 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL,
9396 xdp_link->flags, new_prog);
9400 old_prog = xchg(&link->prog, new_prog);
9401 bpf_prog_put(old_prog);
9408 static const struct bpf_link_ops bpf_xdp_link_lops = {
9409 .release = bpf_xdp_link_release,
9410 .dealloc = bpf_xdp_link_dealloc,
9411 .detach = bpf_xdp_link_detach,
9412 .show_fdinfo = bpf_xdp_link_show_fdinfo,
9413 .fill_link_info = bpf_xdp_link_fill_link_info,
9414 .update_prog = bpf_xdp_link_update,
9417 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog)
9419 struct net *net = current->nsproxy->net_ns;
9420 struct bpf_link_primer link_primer;
9421 struct bpf_xdp_link *link;
9422 struct net_device *dev;
9426 dev = dev_get_by_index(net, attr->link_create.target_ifindex);
9432 link = kzalloc(sizeof(*link), GFP_USER);
9438 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog);
9440 link->flags = attr->link_create.flags;
9442 err = bpf_link_prime(&link->link, &link_primer);
9448 err = dev_xdp_attach_link(dev, NULL, link);
9453 bpf_link_cleanup(&link_primer);
9457 fd = bpf_link_settle(&link_primer);
9458 /* link itself doesn't hold dev's refcnt to not complicate shutdown */
9471 * dev_change_xdp_fd - set or clear a bpf program for a device rx path
9473 * @extack: netlink extended ack
9474 * @fd: new program fd or negative value to clear
9475 * @expected_fd: old program fd that userspace expects to replace or clear
9476 * @flags: xdp-related flags
9478 * Set or clear a bpf program for a device
9480 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack,
9481 int fd, int expected_fd, u32 flags)
9483 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags);
9484 struct bpf_prog *new_prog = NULL, *old_prog = NULL;
9490 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP,
9491 mode != XDP_MODE_SKB);
9492 if (IS_ERR(new_prog))
9493 return PTR_ERR(new_prog);
9496 if (expected_fd >= 0) {
9497 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP,
9498 mode != XDP_MODE_SKB);
9499 if (IS_ERR(old_prog)) {
9500 err = PTR_ERR(old_prog);
9506 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags);
9509 if (err && new_prog)
9510 bpf_prog_put(new_prog);
9512 bpf_prog_put(old_prog);
9517 * dev_new_index - allocate an ifindex
9518 * @net: the applicable net namespace
9520 * Returns a suitable unique value for a new device interface
9521 * number. The caller must hold the rtnl semaphore or the
9522 * dev_base_lock to be sure it remains unique.
9524 static int dev_new_index(struct net *net)
9526 int ifindex = net->ifindex;
9531 if (!__dev_get_by_index(net, ifindex))
9532 return net->ifindex = ifindex;
9536 /* Delayed registration/unregisteration */
9537 LIST_HEAD(net_todo_list);
9538 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq);
9540 static void net_set_todo(struct net_device *dev)
9542 list_add_tail(&dev->todo_list, &net_todo_list);
9543 atomic_inc(&dev_net(dev)->dev_unreg_count);
9546 static netdev_features_t netdev_sync_upper_features(struct net_device *lower,
9547 struct net_device *upper, netdev_features_t features)
9549 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
9550 netdev_features_t feature;
9553 for_each_netdev_feature(upper_disables, feature_bit) {
9554 feature = __NETIF_F_BIT(feature_bit);
9555 if (!(upper->wanted_features & feature)
9556 && (features & feature)) {
9557 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n",
9558 &feature, upper->name);
9559 features &= ~feature;
9566 static void netdev_sync_lower_features(struct net_device *upper,
9567 struct net_device *lower, netdev_features_t features)
9569 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
9570 netdev_features_t feature;
9573 for_each_netdev_feature(upper_disables, feature_bit) {
9574 feature = __NETIF_F_BIT(feature_bit);
9575 if (!(features & feature) && (lower->features & feature)) {
9576 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n",
9577 &feature, lower->name);
9578 lower->wanted_features &= ~feature;
9579 __netdev_update_features(lower);
9581 if (unlikely(lower->features & feature))
9582 netdev_WARN(upper, "failed to disable %pNF on %s!\n",
9583 &feature, lower->name);
9585 netdev_features_change(lower);
9590 static netdev_features_t netdev_fix_features(struct net_device *dev,
9591 netdev_features_t features)
9593 /* Fix illegal checksum combinations */
9594 if ((features & NETIF_F_HW_CSUM) &&
9595 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
9596 netdev_warn(dev, "mixed HW and IP checksum settings.\n");
9597 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
9600 /* TSO requires that SG is present as well. */
9601 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) {
9602 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n");
9603 features &= ~NETIF_F_ALL_TSO;
9606 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) &&
9607 !(features & NETIF_F_IP_CSUM)) {
9608 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n");
9609 features &= ~NETIF_F_TSO;
9610 features &= ~NETIF_F_TSO_ECN;
9613 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) &&
9614 !(features & NETIF_F_IPV6_CSUM)) {
9615 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n");
9616 features &= ~NETIF_F_TSO6;
9619 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */
9620 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO))
9621 features &= ~NETIF_F_TSO_MANGLEID;
9623 /* TSO ECN requires that TSO is present as well. */
9624 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN)
9625 features &= ~NETIF_F_TSO_ECN;
9627 /* Software GSO depends on SG. */
9628 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) {
9629 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n");
9630 features &= ~NETIF_F_GSO;
9633 /* GSO partial features require GSO partial be set */
9634 if ((features & dev->gso_partial_features) &&
9635 !(features & NETIF_F_GSO_PARTIAL)) {
9637 "Dropping partially supported GSO features since no GSO partial.\n");
9638 features &= ~dev->gso_partial_features;
9641 if (!(features & NETIF_F_RXCSUM)) {
9642 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet
9643 * successfully merged by hardware must also have the
9644 * checksum verified by hardware. If the user does not
9645 * want to enable RXCSUM, logically, we should disable GRO_HW.
9647 if (features & NETIF_F_GRO_HW) {
9648 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n");
9649 features &= ~NETIF_F_GRO_HW;
9653 /* LRO/HW-GRO features cannot be combined with RX-FCS */
9654 if (features & NETIF_F_RXFCS) {
9655 if (features & NETIF_F_LRO) {
9656 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n");
9657 features &= ~NETIF_F_LRO;
9660 if (features & NETIF_F_GRO_HW) {
9661 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n");
9662 features &= ~NETIF_F_GRO_HW;
9666 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) {
9667 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n");
9668 features &= ~NETIF_F_LRO;
9671 if (features & NETIF_F_HW_TLS_TX) {
9672 bool ip_csum = (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) ==
9673 (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM);
9674 bool hw_csum = features & NETIF_F_HW_CSUM;
9676 if (!ip_csum && !hw_csum) {
9677 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n");
9678 features &= ~NETIF_F_HW_TLS_TX;
9682 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) {
9683 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n");
9684 features &= ~NETIF_F_HW_TLS_RX;
9690 int __netdev_update_features(struct net_device *dev)
9692 struct net_device *upper, *lower;
9693 netdev_features_t features;
9694 struct list_head *iter;
9699 features = netdev_get_wanted_features(dev);
9701 if (dev->netdev_ops->ndo_fix_features)
9702 features = dev->netdev_ops->ndo_fix_features(dev, features);
9704 /* driver might be less strict about feature dependencies */
9705 features = netdev_fix_features(dev, features);
9707 /* some features can't be enabled if they're off on an upper device */
9708 netdev_for_each_upper_dev_rcu(dev, upper, iter)
9709 features = netdev_sync_upper_features(dev, upper, features);
9711 if (dev->features == features)
9714 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n",
9715 &dev->features, &features);
9717 if (dev->netdev_ops->ndo_set_features)
9718 err = dev->netdev_ops->ndo_set_features(dev, features);
9722 if (unlikely(err < 0)) {
9724 "set_features() failed (%d); wanted %pNF, left %pNF\n",
9725 err, &features, &dev->features);
9726 /* return non-0 since some features might have changed and
9727 * it's better to fire a spurious notification than miss it
9733 /* some features must be disabled on lower devices when disabled
9734 * on an upper device (think: bonding master or bridge)
9736 netdev_for_each_lower_dev(dev, lower, iter)
9737 netdev_sync_lower_features(dev, lower, features);
9740 netdev_features_t diff = features ^ dev->features;
9742 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) {
9743 /* udp_tunnel_{get,drop}_rx_info both need
9744 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the
9745 * device, or they won't do anything.
9746 * Thus we need to update dev->features
9747 * *before* calling udp_tunnel_get_rx_info,
9748 * but *after* calling udp_tunnel_drop_rx_info.
9750 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) {
9751 dev->features = features;
9752 udp_tunnel_get_rx_info(dev);
9754 udp_tunnel_drop_rx_info(dev);
9758 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) {
9759 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) {
9760 dev->features = features;
9761 err |= vlan_get_rx_ctag_filter_info(dev);
9763 vlan_drop_rx_ctag_filter_info(dev);
9767 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) {
9768 if (features & NETIF_F_HW_VLAN_STAG_FILTER) {
9769 dev->features = features;
9770 err |= vlan_get_rx_stag_filter_info(dev);
9772 vlan_drop_rx_stag_filter_info(dev);
9776 dev->features = features;
9779 return err < 0 ? 0 : 1;
9783 * netdev_update_features - recalculate device features
9784 * @dev: the device to check
9786 * Recalculate dev->features set and send notifications if it
9787 * has changed. Should be called after driver or hardware dependent
9788 * conditions might have changed that influence the features.
9790 void netdev_update_features(struct net_device *dev)
9792 if (__netdev_update_features(dev))
9793 netdev_features_change(dev);
9795 EXPORT_SYMBOL(netdev_update_features);
9798 * netdev_change_features - recalculate device features
9799 * @dev: the device to check
9801 * Recalculate dev->features set and send notifications even
9802 * if they have not changed. Should be called instead of
9803 * netdev_update_features() if also dev->vlan_features might
9804 * have changed to allow the changes to be propagated to stacked
9807 void netdev_change_features(struct net_device *dev)
9809 __netdev_update_features(dev);
9810 netdev_features_change(dev);
9812 EXPORT_SYMBOL(netdev_change_features);
9815 * netif_stacked_transfer_operstate - transfer operstate
9816 * @rootdev: the root or lower level device to transfer state from
9817 * @dev: the device to transfer operstate to
9819 * Transfer operational state from root to device. This is normally
9820 * called when a stacking relationship exists between the root
9821 * device and the device(a leaf device).
9823 void netif_stacked_transfer_operstate(const struct net_device *rootdev,
9824 struct net_device *dev)
9826 if (rootdev->operstate == IF_OPER_DORMANT)
9827 netif_dormant_on(dev);
9829 netif_dormant_off(dev);
9831 if (rootdev->operstate == IF_OPER_TESTING)
9832 netif_testing_on(dev);
9834 netif_testing_off(dev);
9836 if (netif_carrier_ok(rootdev))
9837 netif_carrier_on(dev);
9839 netif_carrier_off(dev);
9841 EXPORT_SYMBOL(netif_stacked_transfer_operstate);
9843 static int netif_alloc_rx_queues(struct net_device *dev)
9845 unsigned int i, count = dev->num_rx_queues;
9846 struct netdev_rx_queue *rx;
9847 size_t sz = count * sizeof(*rx);
9852 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL);
9858 for (i = 0; i < count; i++) {
9861 /* XDP RX-queue setup */
9862 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0);
9869 /* Rollback successful reg's and free other resources */
9871 xdp_rxq_info_unreg(&rx[i].xdp_rxq);
9877 static void netif_free_rx_queues(struct net_device *dev)
9879 unsigned int i, count = dev->num_rx_queues;
9881 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */
9885 for (i = 0; i < count; i++)
9886 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq);
9891 static void netdev_init_one_queue(struct net_device *dev,
9892 struct netdev_queue *queue, void *_unused)
9894 /* Initialize queue lock */
9895 spin_lock_init(&queue->_xmit_lock);
9896 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type);
9897 queue->xmit_lock_owner = -1;
9898 netdev_queue_numa_node_write(queue, NUMA_NO_NODE);
9901 dql_init(&queue->dql, HZ);
9905 static void netif_free_tx_queues(struct net_device *dev)
9910 static int netif_alloc_netdev_queues(struct net_device *dev)
9912 unsigned int count = dev->num_tx_queues;
9913 struct netdev_queue *tx;
9914 size_t sz = count * sizeof(*tx);
9916 if (count < 1 || count > 0xffff)
9919 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL);
9925 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL);
9926 spin_lock_init(&dev->tx_global_lock);
9931 void netif_tx_stop_all_queues(struct net_device *dev)
9935 for (i = 0; i < dev->num_tx_queues; i++) {
9936 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
9938 netif_tx_stop_queue(txq);
9941 EXPORT_SYMBOL(netif_tx_stop_all_queues);
9944 * register_netdevice() - register a network device
9945 * @dev: device to register
9947 * Take a prepared network device structure and make it externally accessible.
9948 * A %NETDEV_REGISTER message is sent to the netdev notifier chain.
9949 * Callers must hold the rtnl lock - you may want register_netdev()
9952 int register_netdevice(struct net_device *dev)
9955 struct net *net = dev_net(dev);
9957 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE <
9958 NETDEV_FEATURE_COUNT);
9959 BUG_ON(dev_boot_phase);
9964 /* When net_device's are persistent, this will be fatal. */
9965 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
9968 ret = ethtool_check_ops(dev->ethtool_ops);
9972 spin_lock_init(&dev->addr_list_lock);
9973 netdev_set_addr_lockdep_class(dev);
9975 ret = dev_get_valid_name(net, dev, dev->name);
9980 dev->name_node = netdev_name_node_head_alloc(dev);
9981 if (!dev->name_node)
9984 /* Init, if this function is available */
9985 if (dev->netdev_ops->ndo_init) {
9986 ret = dev->netdev_ops->ndo_init(dev);
9994 if (((dev->hw_features | dev->features) &
9995 NETIF_F_HW_VLAN_CTAG_FILTER) &&
9996 (!dev->netdev_ops->ndo_vlan_rx_add_vid ||
9997 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) {
9998 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n");
10005 dev->ifindex = dev_new_index(net);
10006 else if (__dev_get_by_index(net, dev->ifindex))
10009 /* Transfer changeable features to wanted_features and enable
10010 * software offloads (GSO and GRO).
10012 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF);
10013 dev->features |= NETIF_F_SOFT_FEATURES;
10015 if (dev->udp_tunnel_nic_info) {
10016 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT;
10017 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT;
10020 dev->wanted_features = dev->features & dev->hw_features;
10022 if (!(dev->flags & IFF_LOOPBACK))
10023 dev->hw_features |= NETIF_F_NOCACHE_COPY;
10025 /* If IPv4 TCP segmentation offload is supported we should also
10026 * allow the device to enable segmenting the frame with the option
10027 * of ignoring a static IP ID value. This doesn't enable the
10028 * feature itself but allows the user to enable it later.
10030 if (dev->hw_features & NETIF_F_TSO)
10031 dev->hw_features |= NETIF_F_TSO_MANGLEID;
10032 if (dev->vlan_features & NETIF_F_TSO)
10033 dev->vlan_features |= NETIF_F_TSO_MANGLEID;
10034 if (dev->mpls_features & NETIF_F_TSO)
10035 dev->mpls_features |= NETIF_F_TSO_MANGLEID;
10036 if (dev->hw_enc_features & NETIF_F_TSO)
10037 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID;
10039 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices.
10041 dev->vlan_features |= NETIF_F_HIGHDMA;
10043 /* Make NETIF_F_SG inheritable to tunnel devices.
10045 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL;
10047 /* Make NETIF_F_SG inheritable to MPLS.
10049 dev->mpls_features |= NETIF_F_SG;
10051 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev);
10052 ret = notifier_to_errno(ret);
10056 ret = netdev_register_kobject(dev);
10057 write_lock(&dev_base_lock);
10058 dev->reg_state = ret ? NETREG_UNREGISTERED : NETREG_REGISTERED;
10059 write_unlock(&dev_base_lock);
10061 goto err_uninit_notify;
10063 __netdev_update_features(dev);
10066 * Default initial state at registry is that the
10067 * device is present.
10070 set_bit(__LINK_STATE_PRESENT, &dev->state);
10072 linkwatch_init_dev(dev);
10074 dev_init_scheduler(dev);
10076 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL);
10077 list_netdevice(dev);
10079 add_device_randomness(dev->dev_addr, dev->addr_len);
10081 /* If the device has permanent device address, driver should
10082 * set dev_addr and also addr_assign_type should be set to
10083 * NET_ADDR_PERM (default value).
10085 if (dev->addr_assign_type == NET_ADDR_PERM)
10086 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len);
10088 /* Notify protocols, that a new device appeared. */
10089 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
10090 ret = notifier_to_errno(ret);
10092 /* Expect explicit free_netdev() on failure */
10093 dev->needs_free_netdev = false;
10094 unregister_netdevice_queue(dev, NULL);
10098 * Prevent userspace races by waiting until the network
10099 * device is fully setup before sending notifications.
10101 if (!dev->rtnl_link_ops ||
10102 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
10103 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL);
10109 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev);
10111 if (dev->netdev_ops->ndo_uninit)
10112 dev->netdev_ops->ndo_uninit(dev);
10113 if (dev->priv_destructor)
10114 dev->priv_destructor(dev);
10116 netdev_name_node_free(dev->name_node);
10119 EXPORT_SYMBOL(register_netdevice);
10122 * init_dummy_netdev - init a dummy network device for NAPI
10123 * @dev: device to init
10125 * This takes a network device structure and initialize the minimum
10126 * amount of fields so it can be used to schedule NAPI polls without
10127 * registering a full blown interface. This is to be used by drivers
10128 * that need to tie several hardware interfaces to a single NAPI
10129 * poll scheduler due to HW limitations.
10131 int init_dummy_netdev(struct net_device *dev)
10133 /* Clear everything. Note we don't initialize spinlocks
10134 * are they aren't supposed to be taken by any of the
10135 * NAPI code and this dummy netdev is supposed to be
10136 * only ever used for NAPI polls
10138 memset(dev, 0, sizeof(struct net_device));
10140 /* make sure we BUG if trying to hit standard
10141 * register/unregister code path
10143 dev->reg_state = NETREG_DUMMY;
10145 /* NAPI wants this */
10146 INIT_LIST_HEAD(&dev->napi_list);
10148 /* a dummy interface is started by default */
10149 set_bit(__LINK_STATE_PRESENT, &dev->state);
10150 set_bit(__LINK_STATE_START, &dev->state);
10152 /* napi_busy_loop stats accounting wants this */
10153 dev_net_set(dev, &init_net);
10155 /* Note : We dont allocate pcpu_refcnt for dummy devices,
10156 * because users of this 'device' dont need to change
10162 EXPORT_SYMBOL_GPL(init_dummy_netdev);
10166 * register_netdev - register a network device
10167 * @dev: device to register
10169 * Take a completed network device structure and add it to the kernel
10170 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
10171 * chain. 0 is returned on success. A negative errno code is returned
10172 * on a failure to set up the device, or if the name is a duplicate.
10174 * This is a wrapper around register_netdevice that takes the rtnl semaphore
10175 * and expands the device name if you passed a format string to
10178 int register_netdev(struct net_device *dev)
10182 if (rtnl_lock_killable())
10184 err = register_netdevice(dev);
10188 EXPORT_SYMBOL(register_netdev);
10190 int netdev_refcnt_read(const struct net_device *dev)
10192 #ifdef CONFIG_PCPU_DEV_REFCNT
10195 for_each_possible_cpu(i)
10196 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
10199 return refcount_read(&dev->dev_refcnt);
10202 EXPORT_SYMBOL(netdev_refcnt_read);
10204 int netdev_unregister_timeout_secs __read_mostly = 10;
10206 #define WAIT_REFS_MIN_MSECS 1
10207 #define WAIT_REFS_MAX_MSECS 250
10209 * netdev_wait_allrefs_any - wait until all references are gone.
10210 * @list: list of net_devices to wait on
10212 * This is called when unregistering network devices.
10214 * Any protocol or device that holds a reference should register
10215 * for netdevice notification, and cleanup and put back the
10216 * reference if they receive an UNREGISTER event.
10217 * We can get stuck here if buggy protocols don't correctly
10220 static struct net_device *netdev_wait_allrefs_any(struct list_head *list)
10222 unsigned long rebroadcast_time, warning_time;
10223 struct net_device *dev;
10226 rebroadcast_time = warning_time = jiffies;
10228 list_for_each_entry(dev, list, todo_list)
10229 if (netdev_refcnt_read(dev) == 1)
10233 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
10236 /* Rebroadcast unregister notification */
10237 list_for_each_entry(dev, list, todo_list)
10238 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
10244 list_for_each_entry(dev, list, todo_list)
10245 if (test_bit(__LINK_STATE_LINKWATCH_PENDING,
10247 /* We must not have linkwatch events
10248 * pending on unregister. If this
10249 * happens, we simply run the queue
10250 * unscheduled, resulting in a noop
10253 linkwatch_run_queue();
10259 rebroadcast_time = jiffies;
10264 wait = WAIT_REFS_MIN_MSECS;
10267 wait = min(wait << 1, WAIT_REFS_MAX_MSECS);
10270 list_for_each_entry(dev, list, todo_list)
10271 if (netdev_refcnt_read(dev) == 1)
10274 if (time_after(jiffies, warning_time +
10275 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) {
10276 list_for_each_entry(dev, list, todo_list) {
10277 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n",
10278 dev->name, netdev_refcnt_read(dev));
10279 ref_tracker_dir_print(&dev->refcnt_tracker, 10);
10282 warning_time = jiffies;
10287 /* The sequence is:
10291 * register_netdevice(x1);
10292 * register_netdevice(x2);
10294 * unregister_netdevice(y1);
10295 * unregister_netdevice(y2);
10301 * We are invoked by rtnl_unlock().
10302 * This allows us to deal with problems:
10303 * 1) We can delete sysfs objects which invoke hotplug
10304 * without deadlocking with linkwatch via keventd.
10305 * 2) Since we run with the RTNL semaphore not held, we can sleep
10306 * safely in order to wait for the netdev refcnt to drop to zero.
10308 * We must not return until all unregister events added during
10309 * the interval the lock was held have been completed.
10311 void netdev_run_todo(void)
10313 struct net_device *dev, *tmp;
10314 struct list_head list;
10315 #ifdef CONFIG_LOCKDEP
10316 struct list_head unlink_list;
10318 list_replace_init(&net_unlink_list, &unlink_list);
10320 while (!list_empty(&unlink_list)) {
10321 struct net_device *dev = list_first_entry(&unlink_list,
10324 list_del_init(&dev->unlink_list);
10325 dev->nested_level = dev->lower_level - 1;
10329 /* Snapshot list, allow later requests */
10330 list_replace_init(&net_todo_list, &list);
10334 /* Wait for rcu callbacks to finish before next phase */
10335 if (!list_empty(&list))
10338 list_for_each_entry_safe(dev, tmp, &list, todo_list) {
10339 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) {
10340 netdev_WARN(dev, "run_todo but not unregistering\n");
10341 list_del(&dev->todo_list);
10345 write_lock(&dev_base_lock);
10346 dev->reg_state = NETREG_UNREGISTERED;
10347 write_unlock(&dev_base_lock);
10348 linkwatch_forget_dev(dev);
10351 while (!list_empty(&list)) {
10352 dev = netdev_wait_allrefs_any(&list);
10353 list_del(&dev->todo_list);
10356 BUG_ON(netdev_refcnt_read(dev) != 1);
10357 BUG_ON(!list_empty(&dev->ptype_all));
10358 BUG_ON(!list_empty(&dev->ptype_specific));
10359 WARN_ON(rcu_access_pointer(dev->ip_ptr));
10360 WARN_ON(rcu_access_pointer(dev->ip6_ptr));
10362 if (dev->priv_destructor)
10363 dev->priv_destructor(dev);
10364 if (dev->needs_free_netdev)
10367 if (atomic_dec_and_test(&dev_net(dev)->dev_unreg_count))
10368 wake_up(&netdev_unregistering_wq);
10370 /* Free network device */
10371 kobject_put(&dev->dev.kobj);
10375 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has
10376 * all the same fields in the same order as net_device_stats, with only
10377 * the type differing, but rtnl_link_stats64 may have additional fields
10378 * at the end for newer counters.
10380 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
10381 const struct net_device_stats *netdev_stats)
10383 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t);
10384 const atomic_long_t *src = (atomic_long_t *)netdev_stats;
10385 u64 *dst = (u64 *)stats64;
10387 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64));
10388 for (i = 0; i < n; i++)
10389 dst[i] = (unsigned long)atomic_long_read(&src[i]);
10390 /* zero out counters that only exist in rtnl_link_stats64 */
10391 memset((char *)stats64 + n * sizeof(u64), 0,
10392 sizeof(*stats64) - n * sizeof(u64));
10394 EXPORT_SYMBOL(netdev_stats_to_stats64);
10396 struct net_device_core_stats __percpu *netdev_core_stats_alloc(struct net_device *dev)
10398 struct net_device_core_stats __percpu *p;
10400 p = alloc_percpu_gfp(struct net_device_core_stats,
10401 GFP_ATOMIC | __GFP_NOWARN);
10403 if (p && cmpxchg(&dev->core_stats, NULL, p))
10406 /* This READ_ONCE() pairs with the cmpxchg() above */
10407 return READ_ONCE(dev->core_stats);
10409 EXPORT_SYMBOL(netdev_core_stats_alloc);
10412 * dev_get_stats - get network device statistics
10413 * @dev: device to get statistics from
10414 * @storage: place to store stats
10416 * Get network statistics from device. Return @storage.
10417 * The device driver may provide its own method by setting
10418 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats;
10419 * otherwise the internal statistics structure is used.
10421 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev,
10422 struct rtnl_link_stats64 *storage)
10424 const struct net_device_ops *ops = dev->netdev_ops;
10425 const struct net_device_core_stats __percpu *p;
10427 if (ops->ndo_get_stats64) {
10428 memset(storage, 0, sizeof(*storage));
10429 ops->ndo_get_stats64(dev, storage);
10430 } else if (ops->ndo_get_stats) {
10431 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev));
10433 netdev_stats_to_stats64(storage, &dev->stats);
10436 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */
10437 p = READ_ONCE(dev->core_stats);
10439 const struct net_device_core_stats *core_stats;
10442 for_each_possible_cpu(i) {
10443 core_stats = per_cpu_ptr(p, i);
10444 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped);
10445 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped);
10446 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler);
10447 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped);
10452 EXPORT_SYMBOL(dev_get_stats);
10455 * dev_fetch_sw_netstats - get per-cpu network device statistics
10456 * @s: place to store stats
10457 * @netstats: per-cpu network stats to read from
10459 * Read per-cpu network statistics and populate the related fields in @s.
10461 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s,
10462 const struct pcpu_sw_netstats __percpu *netstats)
10466 for_each_possible_cpu(cpu) {
10467 u64 rx_packets, rx_bytes, tx_packets, tx_bytes;
10468 const struct pcpu_sw_netstats *stats;
10469 unsigned int start;
10471 stats = per_cpu_ptr(netstats, cpu);
10473 start = u64_stats_fetch_begin(&stats->syncp);
10474 rx_packets = u64_stats_read(&stats->rx_packets);
10475 rx_bytes = u64_stats_read(&stats->rx_bytes);
10476 tx_packets = u64_stats_read(&stats->tx_packets);
10477 tx_bytes = u64_stats_read(&stats->tx_bytes);
10478 } while (u64_stats_fetch_retry(&stats->syncp, start));
10480 s->rx_packets += rx_packets;
10481 s->rx_bytes += rx_bytes;
10482 s->tx_packets += tx_packets;
10483 s->tx_bytes += tx_bytes;
10486 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats);
10489 * dev_get_tstats64 - ndo_get_stats64 implementation
10490 * @dev: device to get statistics from
10491 * @s: place to store stats
10493 * Populate @s from dev->stats and dev->tstats. Can be used as
10494 * ndo_get_stats64() callback.
10496 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s)
10498 netdev_stats_to_stats64(s, &dev->stats);
10499 dev_fetch_sw_netstats(s, dev->tstats);
10501 EXPORT_SYMBOL_GPL(dev_get_tstats64);
10503 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev)
10505 struct netdev_queue *queue = dev_ingress_queue(dev);
10507 #ifdef CONFIG_NET_CLS_ACT
10510 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
10513 netdev_init_one_queue(dev, queue, NULL);
10514 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc);
10515 queue->qdisc_sleeping = &noop_qdisc;
10516 rcu_assign_pointer(dev->ingress_queue, queue);
10521 static const struct ethtool_ops default_ethtool_ops;
10523 void netdev_set_default_ethtool_ops(struct net_device *dev,
10524 const struct ethtool_ops *ops)
10526 if (dev->ethtool_ops == &default_ethtool_ops)
10527 dev->ethtool_ops = ops;
10529 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops);
10532 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default
10533 * @dev: netdev to enable the IRQ coalescing on
10535 * Sets a conservative default for SW IRQ coalescing. Users can use
10536 * sysfs attributes to override the default values.
10538 void netdev_sw_irq_coalesce_default_on(struct net_device *dev)
10540 WARN_ON(dev->reg_state == NETREG_REGISTERED);
10542 dev->gro_flush_timeout = 20000;
10543 dev->napi_defer_hard_irqs = 1;
10545 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on);
10547 void netdev_freemem(struct net_device *dev)
10549 char *addr = (char *)dev - dev->padded;
10555 * alloc_netdev_mqs - allocate network device
10556 * @sizeof_priv: size of private data to allocate space for
10557 * @name: device name format string
10558 * @name_assign_type: origin of device name
10559 * @setup: callback to initialize device
10560 * @txqs: the number of TX subqueues to allocate
10561 * @rxqs: the number of RX subqueues to allocate
10563 * Allocates a struct net_device with private data area for driver use
10564 * and performs basic initialization. Also allocates subqueue structs
10565 * for each queue on the device.
10567 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
10568 unsigned char name_assign_type,
10569 void (*setup)(struct net_device *),
10570 unsigned int txqs, unsigned int rxqs)
10572 struct net_device *dev;
10573 unsigned int alloc_size;
10574 struct net_device *p;
10576 BUG_ON(strlen(name) >= sizeof(dev->name));
10579 pr_err("alloc_netdev: Unable to allocate device with zero queues\n");
10584 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n");
10588 alloc_size = sizeof(struct net_device);
10590 /* ensure 32-byte alignment of private area */
10591 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
10592 alloc_size += sizeof_priv;
10594 /* ensure 32-byte alignment of whole construct */
10595 alloc_size += NETDEV_ALIGN - 1;
10597 p = kvzalloc(alloc_size, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL);
10601 dev = PTR_ALIGN(p, NETDEV_ALIGN);
10602 dev->padded = (char *)dev - (char *)p;
10604 ref_tracker_dir_init(&dev->refcnt_tracker, 128);
10605 #ifdef CONFIG_PCPU_DEV_REFCNT
10606 dev->pcpu_refcnt = alloc_percpu(int);
10607 if (!dev->pcpu_refcnt)
10611 refcount_set(&dev->dev_refcnt, 1);
10614 if (dev_addr_init(dev))
10620 dev_net_set(dev, &init_net);
10622 dev->gso_max_size = GSO_LEGACY_MAX_SIZE;
10623 dev->gso_max_segs = GSO_MAX_SEGS;
10624 dev->gro_max_size = GRO_LEGACY_MAX_SIZE;
10625 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE;
10626 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE;
10627 dev->tso_max_size = TSO_LEGACY_MAX_SIZE;
10628 dev->tso_max_segs = TSO_MAX_SEGS;
10629 dev->upper_level = 1;
10630 dev->lower_level = 1;
10631 #ifdef CONFIG_LOCKDEP
10632 dev->nested_level = 0;
10633 INIT_LIST_HEAD(&dev->unlink_list);
10636 INIT_LIST_HEAD(&dev->napi_list);
10637 INIT_LIST_HEAD(&dev->unreg_list);
10638 INIT_LIST_HEAD(&dev->close_list);
10639 INIT_LIST_HEAD(&dev->link_watch_list);
10640 INIT_LIST_HEAD(&dev->adj_list.upper);
10641 INIT_LIST_HEAD(&dev->adj_list.lower);
10642 INIT_LIST_HEAD(&dev->ptype_all);
10643 INIT_LIST_HEAD(&dev->ptype_specific);
10644 INIT_LIST_HEAD(&dev->net_notifier_list);
10645 #ifdef CONFIG_NET_SCHED
10646 hash_init(dev->qdisc_hash);
10648 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM;
10651 if (!dev->tx_queue_len) {
10652 dev->priv_flags |= IFF_NO_QUEUE;
10653 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN;
10656 dev->num_tx_queues = txqs;
10657 dev->real_num_tx_queues = txqs;
10658 if (netif_alloc_netdev_queues(dev))
10661 dev->num_rx_queues = rxqs;
10662 dev->real_num_rx_queues = rxqs;
10663 if (netif_alloc_rx_queues(dev))
10666 strcpy(dev->name, name);
10667 dev->name_assign_type = name_assign_type;
10668 dev->group = INIT_NETDEV_GROUP;
10669 if (!dev->ethtool_ops)
10670 dev->ethtool_ops = &default_ethtool_ops;
10672 nf_hook_netdev_init(dev);
10681 #ifdef CONFIG_PCPU_DEV_REFCNT
10682 free_percpu(dev->pcpu_refcnt);
10685 netdev_freemem(dev);
10688 EXPORT_SYMBOL(alloc_netdev_mqs);
10691 * free_netdev - free network device
10694 * This function does the last stage of destroying an allocated device
10695 * interface. The reference to the device object is released. If this
10696 * is the last reference then it will be freed.Must be called in process
10699 void free_netdev(struct net_device *dev)
10701 struct napi_struct *p, *n;
10705 /* When called immediately after register_netdevice() failed the unwind
10706 * handling may still be dismantling the device. Handle that case by
10707 * deferring the free.
10709 if (dev->reg_state == NETREG_UNREGISTERING) {
10711 dev->needs_free_netdev = true;
10715 netif_free_tx_queues(dev);
10716 netif_free_rx_queues(dev);
10718 kfree(rcu_dereference_protected(dev->ingress_queue, 1));
10720 /* Flush device addresses */
10721 dev_addr_flush(dev);
10723 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
10726 ref_tracker_dir_exit(&dev->refcnt_tracker);
10727 #ifdef CONFIG_PCPU_DEV_REFCNT
10728 free_percpu(dev->pcpu_refcnt);
10729 dev->pcpu_refcnt = NULL;
10731 free_percpu(dev->core_stats);
10732 dev->core_stats = NULL;
10733 free_percpu(dev->xdp_bulkq);
10734 dev->xdp_bulkq = NULL;
10736 /* Compatibility with error handling in drivers */
10737 if (dev->reg_state == NETREG_UNINITIALIZED) {
10738 netdev_freemem(dev);
10742 BUG_ON(dev->reg_state != NETREG_UNREGISTERED);
10743 dev->reg_state = NETREG_RELEASED;
10745 /* will free via device release */
10746 put_device(&dev->dev);
10748 EXPORT_SYMBOL(free_netdev);
10751 * synchronize_net - Synchronize with packet receive processing
10753 * Wait for packets currently being received to be done.
10754 * Does not block later packets from starting.
10756 void synchronize_net(void)
10759 if (rtnl_is_locked())
10760 synchronize_rcu_expedited();
10764 EXPORT_SYMBOL(synchronize_net);
10767 * unregister_netdevice_queue - remove device from the kernel
10771 * This function shuts down a device interface and removes it
10772 * from the kernel tables.
10773 * If head not NULL, device is queued to be unregistered later.
10775 * Callers must hold the rtnl semaphore. You may want
10776 * unregister_netdev() instead of this.
10779 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head)
10784 list_move_tail(&dev->unreg_list, head);
10788 list_add(&dev->unreg_list, &single);
10789 unregister_netdevice_many(&single);
10792 EXPORT_SYMBOL(unregister_netdevice_queue);
10794 void unregister_netdevice_many_notify(struct list_head *head,
10795 u32 portid, const struct nlmsghdr *nlh)
10797 struct net_device *dev, *tmp;
10798 LIST_HEAD(close_head);
10800 BUG_ON(dev_boot_phase);
10803 if (list_empty(head))
10806 list_for_each_entry_safe(dev, tmp, head, unreg_list) {
10807 /* Some devices call without registering
10808 * for initialization unwind. Remove those
10809 * devices and proceed with the remaining.
10811 if (dev->reg_state == NETREG_UNINITIALIZED) {
10812 pr_debug("unregister_netdevice: device %s/%p never was registered\n",
10816 list_del(&dev->unreg_list);
10819 dev->dismantle = true;
10820 BUG_ON(dev->reg_state != NETREG_REGISTERED);
10823 /* If device is running, close it first. */
10824 list_for_each_entry(dev, head, unreg_list)
10825 list_add_tail(&dev->close_list, &close_head);
10826 dev_close_many(&close_head, true);
10828 list_for_each_entry(dev, head, unreg_list) {
10829 /* And unlink it from device chain. */
10830 write_lock(&dev_base_lock);
10831 unlist_netdevice(dev, false);
10832 dev->reg_state = NETREG_UNREGISTERING;
10833 write_unlock(&dev_base_lock);
10835 flush_all_backlogs();
10839 list_for_each_entry(dev, head, unreg_list) {
10840 struct sk_buff *skb = NULL;
10842 /* Shutdown queueing discipline. */
10845 dev_xdp_uninstall(dev);
10846 bpf_dev_bound_netdev_unregister(dev);
10848 netdev_offload_xstats_disable_all(dev);
10850 /* Notify protocols, that we are about to destroy
10851 * this device. They should clean all the things.
10853 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
10855 if (!dev->rtnl_link_ops ||
10856 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
10857 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0,
10858 GFP_KERNEL, NULL, 0,
10859 portid, nlmsg_seq(nlh));
10862 * Flush the unicast and multicast chains
10867 netdev_name_node_alt_flush(dev);
10868 netdev_name_node_free(dev->name_node);
10870 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev);
10872 if (dev->netdev_ops->ndo_uninit)
10873 dev->netdev_ops->ndo_uninit(dev);
10876 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh);
10878 /* Notifier chain MUST detach us all upper devices. */
10879 WARN_ON(netdev_has_any_upper_dev(dev));
10880 WARN_ON(netdev_has_any_lower_dev(dev));
10882 /* Remove entries from kobject tree */
10883 netdev_unregister_kobject(dev);
10885 /* Remove XPS queueing entries */
10886 netif_reset_xps_queues_gt(dev, 0);
10892 list_for_each_entry(dev, head, unreg_list) {
10893 netdev_put(dev, &dev->dev_registered_tracker);
10901 * unregister_netdevice_many - unregister many devices
10902 * @head: list of devices
10904 * Note: As most callers use a stack allocated list_head,
10905 * we force a list_del() to make sure stack wont be corrupted later.
10907 void unregister_netdevice_many(struct list_head *head)
10909 unregister_netdevice_many_notify(head, 0, NULL);
10911 EXPORT_SYMBOL(unregister_netdevice_many);
10914 * unregister_netdev - remove device from the kernel
10917 * This function shuts down a device interface and removes it
10918 * from the kernel tables.
10920 * This is just a wrapper for unregister_netdevice that takes
10921 * the rtnl semaphore. In general you want to use this and not
10922 * unregister_netdevice.
10924 void unregister_netdev(struct net_device *dev)
10927 unregister_netdevice(dev);
10930 EXPORT_SYMBOL(unregister_netdev);
10933 * __dev_change_net_namespace - move device to different nethost namespace
10935 * @net: network namespace
10936 * @pat: If not NULL name pattern to try if the current device name
10937 * is already taken in the destination network namespace.
10938 * @new_ifindex: If not zero, specifies device index in the target
10941 * This function shuts down a device interface and moves it
10942 * to a new network namespace. On success 0 is returned, on
10943 * a failure a netagive errno code is returned.
10945 * Callers must hold the rtnl semaphore.
10948 int __dev_change_net_namespace(struct net_device *dev, struct net *net,
10949 const char *pat, int new_ifindex)
10951 struct net *net_old = dev_net(dev);
10956 /* Don't allow namespace local devices to be moved. */
10958 if (dev->features & NETIF_F_NETNS_LOCAL)
10961 /* Ensure the device has been registrered */
10962 if (dev->reg_state != NETREG_REGISTERED)
10965 /* Get out if there is nothing todo */
10967 if (net_eq(net_old, net))
10970 /* Pick the destination device name, and ensure
10971 * we can use it in the destination network namespace.
10974 if (netdev_name_in_use(net, dev->name)) {
10975 /* We get here if we can't use the current device name */
10978 err = dev_get_valid_name(net, dev, pat);
10983 /* Check that new_ifindex isn't used yet. */
10985 if (new_ifindex && __dev_get_by_index(net, new_ifindex))
10989 * And now a mini version of register_netdevice unregister_netdevice.
10992 /* If device is running close it first. */
10995 /* And unlink it from device chain */
10996 unlist_netdevice(dev, true);
11000 /* Shutdown queueing discipline. */
11003 /* Notify protocols, that we are about to destroy
11004 * this device. They should clean all the things.
11006 * Note that dev->reg_state stays at NETREG_REGISTERED.
11007 * This is wanted because this way 8021q and macvlan know
11008 * the device is just moving and can keep their slaves up.
11010 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
11013 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL);
11014 /* If there is an ifindex conflict assign a new one */
11015 if (!new_ifindex) {
11016 if (__dev_get_by_index(net, dev->ifindex))
11017 new_ifindex = dev_new_index(net);
11019 new_ifindex = dev->ifindex;
11022 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid,
11026 * Flush the unicast and multicast chains
11031 /* Send a netdev-removed uevent to the old namespace */
11032 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE);
11033 netdev_adjacent_del_links(dev);
11035 /* Move per-net netdevice notifiers that are following the netdevice */
11036 move_netdevice_notifiers_dev_net(dev, net);
11038 /* Actually switch the network namespace */
11039 dev_net_set(dev, net);
11040 dev->ifindex = new_ifindex;
11042 /* Send a netdev-add uevent to the new namespace */
11043 kobject_uevent(&dev->dev.kobj, KOBJ_ADD);
11044 netdev_adjacent_add_links(dev);
11046 /* Fixup kobjects */
11047 err = device_rename(&dev->dev, dev->name);
11050 /* Adapt owner in case owning user namespace of target network
11051 * namespace is different from the original one.
11053 err = netdev_change_owner(dev, net_old, net);
11056 /* Add the device back in the hashes */
11057 list_netdevice(dev);
11059 /* Notify protocols, that a new device appeared. */
11060 call_netdevice_notifiers(NETDEV_REGISTER, dev);
11063 * Prevent userspace races by waiting until the network
11064 * device is fully setup before sending notifications.
11066 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL);
11073 EXPORT_SYMBOL_GPL(__dev_change_net_namespace);
11075 static int dev_cpu_dead(unsigned int oldcpu)
11077 struct sk_buff **list_skb;
11078 struct sk_buff *skb;
11080 struct softnet_data *sd, *oldsd, *remsd = NULL;
11082 local_irq_disable();
11083 cpu = smp_processor_id();
11084 sd = &per_cpu(softnet_data, cpu);
11085 oldsd = &per_cpu(softnet_data, oldcpu);
11087 /* Find end of our completion_queue. */
11088 list_skb = &sd->completion_queue;
11090 list_skb = &(*list_skb)->next;
11091 /* Append completion queue from offline CPU. */
11092 *list_skb = oldsd->completion_queue;
11093 oldsd->completion_queue = NULL;
11095 /* Append output queue from offline CPU. */
11096 if (oldsd->output_queue) {
11097 *sd->output_queue_tailp = oldsd->output_queue;
11098 sd->output_queue_tailp = oldsd->output_queue_tailp;
11099 oldsd->output_queue = NULL;
11100 oldsd->output_queue_tailp = &oldsd->output_queue;
11102 /* Append NAPI poll list from offline CPU, with one exception :
11103 * process_backlog() must be called by cpu owning percpu backlog.
11104 * We properly handle process_queue & input_pkt_queue later.
11106 while (!list_empty(&oldsd->poll_list)) {
11107 struct napi_struct *napi = list_first_entry(&oldsd->poll_list,
11108 struct napi_struct,
11111 list_del_init(&napi->poll_list);
11112 if (napi->poll == process_backlog)
11115 ____napi_schedule(sd, napi);
11118 raise_softirq_irqoff(NET_TX_SOFTIRQ);
11119 local_irq_enable();
11122 remsd = oldsd->rps_ipi_list;
11123 oldsd->rps_ipi_list = NULL;
11125 /* send out pending IPI's on offline CPU */
11126 net_rps_send_ipi(remsd);
11128 /* Process offline CPU's input_pkt_queue */
11129 while ((skb = __skb_dequeue(&oldsd->process_queue))) {
11131 input_queue_head_incr(oldsd);
11133 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) {
11135 input_queue_head_incr(oldsd);
11142 * netdev_increment_features - increment feature set by one
11143 * @all: current feature set
11144 * @one: new feature set
11145 * @mask: mask feature set
11147 * Computes a new feature set after adding a device with feature set
11148 * @one to the master device with current feature set @all. Will not
11149 * enable anything that is off in @mask. Returns the new feature set.
11151 netdev_features_t netdev_increment_features(netdev_features_t all,
11152 netdev_features_t one, netdev_features_t mask)
11154 if (mask & NETIF_F_HW_CSUM)
11155 mask |= NETIF_F_CSUM_MASK;
11156 mask |= NETIF_F_VLAN_CHALLENGED;
11158 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask;
11159 all &= one | ~NETIF_F_ALL_FOR_ALL;
11161 /* If one device supports hw checksumming, set for all. */
11162 if (all & NETIF_F_HW_CSUM)
11163 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM);
11167 EXPORT_SYMBOL(netdev_increment_features);
11169 static struct hlist_head * __net_init netdev_create_hash(void)
11172 struct hlist_head *hash;
11174 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL);
11176 for (i = 0; i < NETDEV_HASHENTRIES; i++)
11177 INIT_HLIST_HEAD(&hash[i]);
11182 /* Initialize per network namespace state */
11183 static int __net_init netdev_init(struct net *net)
11185 BUILD_BUG_ON(GRO_HASH_BUCKETS >
11186 8 * sizeof_field(struct napi_struct, gro_bitmask));
11188 INIT_LIST_HEAD(&net->dev_base_head);
11190 net->dev_name_head = netdev_create_hash();
11191 if (net->dev_name_head == NULL)
11194 net->dev_index_head = netdev_create_hash();
11195 if (net->dev_index_head == NULL)
11198 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain);
11203 kfree(net->dev_name_head);
11209 * netdev_drivername - network driver for the device
11210 * @dev: network device
11212 * Determine network driver for device.
11214 const char *netdev_drivername(const struct net_device *dev)
11216 const struct device_driver *driver;
11217 const struct device *parent;
11218 const char *empty = "";
11220 parent = dev->dev.parent;
11224 driver = parent->driver;
11225 if (driver && driver->name)
11226 return driver->name;
11230 static void __netdev_printk(const char *level, const struct net_device *dev,
11231 struct va_format *vaf)
11233 if (dev && dev->dev.parent) {
11234 dev_printk_emit(level[1] - '0',
11237 dev_driver_string(dev->dev.parent),
11238 dev_name(dev->dev.parent),
11239 netdev_name(dev), netdev_reg_state(dev),
11242 printk("%s%s%s: %pV",
11243 level, netdev_name(dev), netdev_reg_state(dev), vaf);
11245 printk("%s(NULL net_device): %pV", level, vaf);
11249 void netdev_printk(const char *level, const struct net_device *dev,
11250 const char *format, ...)
11252 struct va_format vaf;
11255 va_start(args, format);
11260 __netdev_printk(level, dev, &vaf);
11264 EXPORT_SYMBOL(netdev_printk);
11266 #define define_netdev_printk_level(func, level) \
11267 void func(const struct net_device *dev, const char *fmt, ...) \
11269 struct va_format vaf; \
11272 va_start(args, fmt); \
11277 __netdev_printk(level, dev, &vaf); \
11281 EXPORT_SYMBOL(func);
11283 define_netdev_printk_level(netdev_emerg, KERN_EMERG);
11284 define_netdev_printk_level(netdev_alert, KERN_ALERT);
11285 define_netdev_printk_level(netdev_crit, KERN_CRIT);
11286 define_netdev_printk_level(netdev_err, KERN_ERR);
11287 define_netdev_printk_level(netdev_warn, KERN_WARNING);
11288 define_netdev_printk_level(netdev_notice, KERN_NOTICE);
11289 define_netdev_printk_level(netdev_info, KERN_INFO);
11291 static void __net_exit netdev_exit(struct net *net)
11293 kfree(net->dev_name_head);
11294 kfree(net->dev_index_head);
11295 if (net != &init_net)
11296 WARN_ON_ONCE(!list_empty(&net->dev_base_head));
11299 static struct pernet_operations __net_initdata netdev_net_ops = {
11300 .init = netdev_init,
11301 .exit = netdev_exit,
11304 static void __net_exit default_device_exit_net(struct net *net)
11306 struct net_device *dev, *aux;
11308 * Push all migratable network devices back to the
11309 * initial network namespace
11312 for_each_netdev_safe(net, dev, aux) {
11314 char fb_name[IFNAMSIZ];
11316 /* Ignore unmoveable devices (i.e. loopback) */
11317 if (dev->features & NETIF_F_NETNS_LOCAL)
11320 /* Leave virtual devices for the generic cleanup */
11321 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund)
11324 /* Push remaining network devices to init_net */
11325 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex);
11326 if (netdev_name_in_use(&init_net, fb_name))
11327 snprintf(fb_name, IFNAMSIZ, "dev%%d");
11328 err = dev_change_net_namespace(dev, &init_net, fb_name);
11330 pr_emerg("%s: failed to move %s to init_net: %d\n",
11331 __func__, dev->name, err);
11337 static void __net_exit default_device_exit_batch(struct list_head *net_list)
11339 /* At exit all network devices most be removed from a network
11340 * namespace. Do this in the reverse order of registration.
11341 * Do this across as many network namespaces as possible to
11342 * improve batching efficiency.
11344 struct net_device *dev;
11346 LIST_HEAD(dev_kill_list);
11349 list_for_each_entry(net, net_list, exit_list) {
11350 default_device_exit_net(net);
11354 list_for_each_entry(net, net_list, exit_list) {
11355 for_each_netdev_reverse(net, dev) {
11356 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink)
11357 dev->rtnl_link_ops->dellink(dev, &dev_kill_list);
11359 unregister_netdevice_queue(dev, &dev_kill_list);
11362 unregister_netdevice_many(&dev_kill_list);
11366 static struct pernet_operations __net_initdata default_device_ops = {
11367 .exit_batch = default_device_exit_batch,
11371 * Initialize the DEV module. At boot time this walks the device list and
11372 * unhooks any devices that fail to initialise (normally hardware not
11373 * present) and leaves us with a valid list of present and active devices.
11378 * This is called single threaded during boot, so no need
11379 * to take the rtnl semaphore.
11381 static int __init net_dev_init(void)
11383 int i, rc = -ENOMEM;
11385 BUG_ON(!dev_boot_phase);
11387 if (dev_proc_init())
11390 if (netdev_kobject_init())
11393 INIT_LIST_HEAD(&ptype_all);
11394 for (i = 0; i < PTYPE_HASH_SIZE; i++)
11395 INIT_LIST_HEAD(&ptype_base[i]);
11397 if (register_pernet_subsys(&netdev_net_ops))
11401 * Initialise the packet receive queues.
11404 for_each_possible_cpu(i) {
11405 struct work_struct *flush = per_cpu_ptr(&flush_works, i);
11406 struct softnet_data *sd = &per_cpu(softnet_data, i);
11408 INIT_WORK(flush, flush_backlog);
11410 skb_queue_head_init(&sd->input_pkt_queue);
11411 skb_queue_head_init(&sd->process_queue);
11412 #ifdef CONFIG_XFRM_OFFLOAD
11413 skb_queue_head_init(&sd->xfrm_backlog);
11415 INIT_LIST_HEAD(&sd->poll_list);
11416 sd->output_queue_tailp = &sd->output_queue;
11418 INIT_CSD(&sd->csd, rps_trigger_softirq, sd);
11421 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd);
11422 spin_lock_init(&sd->defer_lock);
11424 init_gro_hash(&sd->backlog);
11425 sd->backlog.poll = process_backlog;
11426 sd->backlog.weight = weight_p;
11429 dev_boot_phase = 0;
11431 /* The loopback device is special if any other network devices
11432 * is present in a network namespace the loopback device must
11433 * be present. Since we now dynamically allocate and free the
11434 * loopback device ensure this invariant is maintained by
11435 * keeping the loopback device as the first device on the
11436 * list of network devices. Ensuring the loopback devices
11437 * is the first device that appears and the last network device
11440 if (register_pernet_device(&loopback_net_ops))
11443 if (register_pernet_device(&default_device_ops))
11446 open_softirq(NET_TX_SOFTIRQ, net_tx_action);
11447 open_softirq(NET_RX_SOFTIRQ, net_rx_action);
11449 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead",
11450 NULL, dev_cpu_dead);
11457 subsys_initcall(net_dev_init);