1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * NET3 Protocol independent device support routines.
5 * Derived from the non IP parts of dev.c 1.0.19
7 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
8 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Florian la Roche <rzsfl@rz.uni-sb.de>
12 * Alan Cox <gw4pts@gw4pts.ampr.org>
13 * David Hinds <dahinds@users.sourceforge.net>
14 * Alexey Kuznetsov <kuznet@ms2.inr.ac.ru>
15 * Adam Sulmicki <adam@cfar.umd.edu>
16 * Pekka Riikonen <priikone@poesidon.pspt.fi>
19 * D.J. Barrow : Fixed bug where dev->refcnt gets set
20 * to 2 if register_netdev gets called
21 * before net_dev_init & also removed a
22 * few lines of code in the process.
23 * Alan Cox : device private ioctl copies fields back.
24 * Alan Cox : Transmit queue code does relevant
25 * stunts to keep the queue safe.
26 * Alan Cox : Fixed double lock.
27 * Alan Cox : Fixed promisc NULL pointer trap
28 * ???????? : Support the full private ioctl range
29 * Alan Cox : Moved ioctl permission check into
31 * Tim Kordas : SIOCADDMULTI/SIOCDELMULTI
32 * Alan Cox : 100 backlog just doesn't cut it when
33 * you start doing multicast video 8)
34 * Alan Cox : Rewrote net_bh and list manager.
35 * Alan Cox : Fix ETH_P_ALL echoback lengths.
36 * Alan Cox : Took out transmit every packet pass
37 * Saved a few bytes in the ioctl handler
38 * Alan Cox : Network driver sets packet type before
39 * calling netif_rx. Saves a function
41 * Alan Cox : Hashed net_bh()
42 * Richard Kooijman: Timestamp fixes.
43 * Alan Cox : Wrong field in SIOCGIFDSTADDR
44 * Alan Cox : Device lock protection.
45 * Alan Cox : Fixed nasty side effect of device close
47 * Rudi Cilibrasi : Pass the right thing to
49 * Dave Miller : 32bit quantity for the device lock to
50 * make it work out on a Sparc.
51 * Bjorn Ekwall : Added KERNELD hack.
52 * Alan Cox : Cleaned up the backlog initialise.
53 * Craig Metz : SIOCGIFCONF fix if space for under
55 * Thomas Bogendoerfer : Return ENODEV for dev_open, if there
56 * is no device open function.
57 * Andi Kleen : Fix error reporting for SIOCGIFCONF
58 * Michael Chastain : Fix signed/unsigned for SIOCGIFCONF
59 * Cyrus Durgin : Cleaned for KMOD
60 * Adam Sulmicki : Bug Fix : Network Device Unload
61 * A network device unload needs to purge
63 * Paul Rusty Russell : SIOCSIFNAME
64 * Pekka Riikonen : Netdev boot-time settings code
65 * Andrew Morton : Make unregister_netdevice wait
66 * indefinitely on dev->refcnt
67 * J Hadi Salim : - Backlog queue sampling
68 * - netif_rx() feedback
71 #include <linux/uaccess.h>
72 #include <linux/bitmap.h>
73 #include <linux/capability.h>
74 #include <linux/cpu.h>
75 #include <linux/types.h>
76 #include <linux/kernel.h>
77 #include <linux/hash.h>
78 #include <linux/slab.h>
79 #include <linux/sched.h>
80 #include <linux/sched/mm.h>
81 #include <linux/mutex.h>
82 #include <linux/rwsem.h>
83 #include <linux/string.h>
85 #include <linux/socket.h>
86 #include <linux/sockios.h>
87 #include <linux/errno.h>
88 #include <linux/interrupt.h>
89 #include <linux/if_ether.h>
90 #include <linux/netdevice.h>
91 #include <linux/etherdevice.h>
92 #include <linux/ethtool.h>
93 #include <linux/skbuff.h>
94 #include <linux/kthread.h>
95 #include <linux/bpf.h>
96 #include <linux/bpf_trace.h>
97 #include <net/net_namespace.h>
99 #include <net/busy_poll.h>
100 #include <linux/rtnetlink.h>
101 #include <linux/stat.h>
104 #include <net/dst_metadata.h>
106 #include <net/pkt_sched.h>
107 #include <net/pkt_cls.h>
108 #include <net/checksum.h>
109 #include <net/xfrm.h>
111 #include <linux/highmem.h>
112 #include <linux/init.h>
113 #include <linux/module.h>
114 #include <linux/netpoll.h>
115 #include <linux/rcupdate.h>
116 #include <linux/delay.h>
117 #include <net/iw_handler.h>
118 #include <asm/current.h>
119 #include <linux/audit.h>
120 #include <linux/dmaengine.h>
121 #include <linux/err.h>
122 #include <linux/ctype.h>
123 #include <linux/if_arp.h>
124 #include <linux/if_vlan.h>
125 #include <linux/ip.h>
127 #include <net/mpls.h>
128 #include <linux/ipv6.h>
129 #include <linux/in.h>
130 #include <linux/jhash.h>
131 #include <linux/random.h>
132 #include <trace/events/napi.h>
133 #include <trace/events/net.h>
134 #include <trace/events/skb.h>
135 #include <trace/events/qdisc.h>
136 #include <trace/events/xdp.h>
137 #include <linux/inetdevice.h>
138 #include <linux/cpu_rmap.h>
139 #include <linux/static_key.h>
140 #include <linux/hashtable.h>
141 #include <linux/vmalloc.h>
142 #include <linux/if_macvlan.h>
143 #include <linux/errqueue.h>
144 #include <linux/hrtimer.h>
145 #include <linux/netfilter_netdev.h>
146 #include <linux/crash_dump.h>
147 #include <linux/sctp.h>
148 #include <net/udp_tunnel.h>
149 #include <linux/net_namespace.h>
150 #include <linux/indirect_call_wrapper.h>
151 #include <net/devlink.h>
152 #include <linux/pm_runtime.h>
153 #include <linux/prandom.h>
154 #include <linux/once_lite.h>
155 #include <net/netdev_rx_queue.h>
158 #include "net-sysfs.h"
160 static DEFINE_SPINLOCK(ptype_lock);
161 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly;
162 struct list_head ptype_all __read_mostly; /* Taps */
164 static int netif_rx_internal(struct sk_buff *skb);
165 static int call_netdevice_notifiers_extack(unsigned long val,
166 struct net_device *dev,
167 struct netlink_ext_ack *extack);
170 * The @dev_base_head list is protected by @dev_base_lock and the rtnl
173 * Pure readers hold dev_base_lock for reading, or rcu_read_lock()
175 * Writers must hold the rtnl semaphore while they loop through the
176 * dev_base_head list, and hold dev_base_lock for writing when they do the
177 * actual updates. This allows pure readers to access the list even
178 * while a writer is preparing to update it.
180 * To put it another way, dev_base_lock is held for writing only to
181 * protect against pure readers; the rtnl semaphore provides the
182 * protection against other writers.
184 * See, for example usages, register_netdevice() and
185 * unregister_netdevice(), which must be called with the rtnl
188 DEFINE_RWLOCK(dev_base_lock);
189 EXPORT_SYMBOL(dev_base_lock);
191 static DEFINE_MUTEX(ifalias_mutex);
193 /* protects napi_hash addition/deletion and napi_gen_id */
194 static DEFINE_SPINLOCK(napi_hash_lock);
196 static unsigned int napi_gen_id = NR_CPUS;
197 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8);
199 static DECLARE_RWSEM(devnet_rename_sem);
201 static inline void dev_base_seq_inc(struct net *net)
203 while (++net->dev_base_seq == 0)
207 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name)
209 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ));
211 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)];
214 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex)
216 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)];
219 static inline void rps_lock_irqsave(struct softnet_data *sd,
220 unsigned long *flags)
222 if (IS_ENABLED(CONFIG_RPS))
223 spin_lock_irqsave(&sd->input_pkt_queue.lock, *flags);
224 else if (!IS_ENABLED(CONFIG_PREEMPT_RT))
225 local_irq_save(*flags);
228 static inline void rps_lock_irq_disable(struct softnet_data *sd)
230 if (IS_ENABLED(CONFIG_RPS))
231 spin_lock_irq(&sd->input_pkt_queue.lock);
232 else if (!IS_ENABLED(CONFIG_PREEMPT_RT))
236 static inline void rps_unlock_irq_restore(struct softnet_data *sd,
237 unsigned long *flags)
239 if (IS_ENABLED(CONFIG_RPS))
240 spin_unlock_irqrestore(&sd->input_pkt_queue.lock, *flags);
241 else if (!IS_ENABLED(CONFIG_PREEMPT_RT))
242 local_irq_restore(*flags);
245 static inline void rps_unlock_irq_enable(struct softnet_data *sd)
247 if (IS_ENABLED(CONFIG_RPS))
248 spin_unlock_irq(&sd->input_pkt_queue.lock);
249 else if (!IS_ENABLED(CONFIG_PREEMPT_RT))
253 static struct netdev_name_node *netdev_name_node_alloc(struct net_device *dev,
256 struct netdev_name_node *name_node;
258 name_node = kmalloc(sizeof(*name_node), GFP_KERNEL);
261 INIT_HLIST_NODE(&name_node->hlist);
262 name_node->dev = dev;
263 name_node->name = name;
267 static struct netdev_name_node *
268 netdev_name_node_head_alloc(struct net_device *dev)
270 struct netdev_name_node *name_node;
272 name_node = netdev_name_node_alloc(dev, dev->name);
275 INIT_LIST_HEAD(&name_node->list);
279 static void netdev_name_node_free(struct netdev_name_node *name_node)
284 static void netdev_name_node_add(struct net *net,
285 struct netdev_name_node *name_node)
287 hlist_add_head_rcu(&name_node->hlist,
288 dev_name_hash(net, name_node->name));
291 static void netdev_name_node_del(struct netdev_name_node *name_node)
293 hlist_del_rcu(&name_node->hlist);
296 static struct netdev_name_node *netdev_name_node_lookup(struct net *net,
299 struct hlist_head *head = dev_name_hash(net, name);
300 struct netdev_name_node *name_node;
302 hlist_for_each_entry(name_node, head, hlist)
303 if (!strcmp(name_node->name, name))
308 static struct netdev_name_node *netdev_name_node_lookup_rcu(struct net *net,
311 struct hlist_head *head = dev_name_hash(net, name);
312 struct netdev_name_node *name_node;
314 hlist_for_each_entry_rcu(name_node, head, hlist)
315 if (!strcmp(name_node->name, name))
320 bool netdev_name_in_use(struct net *net, const char *name)
322 return netdev_name_node_lookup(net, name);
324 EXPORT_SYMBOL(netdev_name_in_use);
326 int netdev_name_node_alt_create(struct net_device *dev, const char *name)
328 struct netdev_name_node *name_node;
329 struct net *net = dev_net(dev);
331 name_node = netdev_name_node_lookup(net, name);
334 name_node = netdev_name_node_alloc(dev, name);
337 netdev_name_node_add(net, name_node);
338 /* The node that holds dev->name acts as a head of per-device list. */
339 list_add_tail_rcu(&name_node->list, &dev->name_node->list);
344 static void __netdev_name_node_alt_destroy(struct netdev_name_node *name_node)
346 list_del(&name_node->list);
347 kfree(name_node->name);
348 netdev_name_node_free(name_node);
351 int netdev_name_node_alt_destroy(struct net_device *dev, const char *name)
353 struct netdev_name_node *name_node;
354 struct net *net = dev_net(dev);
356 name_node = netdev_name_node_lookup(net, name);
359 /* lookup might have found our primary name or a name belonging
362 if (name_node == dev->name_node || name_node->dev != dev)
365 netdev_name_node_del(name_node);
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 netdev_name_node *name_node;
384 struct net *net = dev_net(dev);
388 write_lock(&dev_base_lock);
389 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head);
390 netdev_name_node_add(net, dev->name_node);
391 hlist_add_head_rcu(&dev->index_hlist,
392 dev_index_hash(net, dev->ifindex));
393 write_unlock(&dev_base_lock);
395 netdev_for_each_altname(dev, name_node)
396 netdev_name_node_add(net, name_node);
398 /* We reserved the ifindex, this can't fail */
399 WARN_ON(xa_store(&net->dev_by_index, dev->ifindex, dev, GFP_KERNEL));
401 dev_base_seq_inc(net);
404 /* Device list removal
405 * caller must respect a RCU grace period before freeing/reusing dev
407 static void unlist_netdevice(struct net_device *dev, bool lock)
409 struct netdev_name_node *name_node;
410 struct net *net = dev_net(dev);
414 xa_erase(&net->dev_by_index, dev->ifindex);
416 netdev_for_each_altname(dev, name_node)
417 netdev_name_node_del(name_node);
419 /* Unlink dev from the device chain */
421 write_lock(&dev_base_lock);
422 list_del_rcu(&dev->dev_list);
423 netdev_name_node_del(dev->name_node);
424 hlist_del_rcu(&dev->index_hlist);
426 write_unlock(&dev_base_lock);
428 dev_base_seq_inc(dev_net(dev));
435 static RAW_NOTIFIER_HEAD(netdev_chain);
438 * Device drivers call our routines to queue packets here. We empty the
439 * queue in the local softnet handler.
442 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data);
443 EXPORT_PER_CPU_SYMBOL(softnet_data);
445 #ifdef CONFIG_LOCKDEP
447 * register_netdevice() inits txq->_xmit_lock and sets lockdep class
448 * according to dev->type
450 static const unsigned short netdev_lock_type[] = {
451 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25,
452 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET,
453 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM,
454 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP,
455 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD,
456 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25,
457 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP,
458 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD,
459 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI,
460 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE,
461 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET,
462 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL,
463 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM,
464 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE,
465 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE};
467 static const char *const netdev_lock_name[] = {
468 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25",
469 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET",
470 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM",
471 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP",
472 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD",
473 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25",
474 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP",
475 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD",
476 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI",
477 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE",
478 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET",
479 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL",
480 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM",
481 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE",
482 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"};
484 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)];
485 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)];
487 static inline unsigned short netdev_lock_pos(unsigned short dev_type)
491 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++)
492 if (netdev_lock_type[i] == dev_type)
494 /* the last key is used by default */
495 return ARRAY_SIZE(netdev_lock_type) - 1;
498 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
499 unsigned short dev_type)
503 i = netdev_lock_pos(dev_type);
504 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i],
505 netdev_lock_name[i]);
508 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
512 i = netdev_lock_pos(dev->type);
513 lockdep_set_class_and_name(&dev->addr_list_lock,
514 &netdev_addr_lock_key[i],
515 netdev_lock_name[i]);
518 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
519 unsigned short dev_type)
523 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
528 /*******************************************************************************
530 * Protocol management and registration routines
532 *******************************************************************************/
536 * Add a protocol ID to the list. Now that the input handler is
537 * smarter we can dispense with all the messy stuff that used to be
540 * BEWARE!!! Protocol handlers, mangling input packets,
541 * MUST BE last in hash buckets and checking protocol handlers
542 * MUST start from promiscuous ptype_all chain in net_bh.
543 * It is true now, do not change it.
544 * Explanation follows: if protocol handler, mangling packet, will
545 * be the first on list, it is not able to sense, that packet
546 * is cloned and should be copied-on-write, so that it will
547 * change it and subsequent readers will get broken packet.
551 static inline struct list_head *ptype_head(const struct packet_type *pt)
553 if (pt->type == htons(ETH_P_ALL))
554 return pt->dev ? &pt->dev->ptype_all : &ptype_all;
556 return pt->dev ? &pt->dev->ptype_specific :
557 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK];
561 * dev_add_pack - add packet handler
562 * @pt: packet type declaration
564 * Add a protocol handler to the networking stack. The passed &packet_type
565 * is linked into kernel lists and may not be freed until it has been
566 * removed from the kernel lists.
568 * This call does not sleep therefore it can not
569 * guarantee all CPU's that are in middle of receiving packets
570 * will see the new packet type (until the next received packet).
573 void dev_add_pack(struct packet_type *pt)
575 struct list_head *head = ptype_head(pt);
577 spin_lock(&ptype_lock);
578 list_add_rcu(&pt->list, head);
579 spin_unlock(&ptype_lock);
581 EXPORT_SYMBOL(dev_add_pack);
584 * __dev_remove_pack - remove packet handler
585 * @pt: packet type declaration
587 * Remove a protocol handler that was previously added to the kernel
588 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
589 * from the kernel lists and can be freed or reused once this function
592 * The packet type might still be in use by receivers
593 * and must not be freed until after all the CPU's have gone
594 * through a quiescent state.
596 void __dev_remove_pack(struct packet_type *pt)
598 struct list_head *head = ptype_head(pt);
599 struct packet_type *pt1;
601 spin_lock(&ptype_lock);
603 list_for_each_entry(pt1, head, list) {
605 list_del_rcu(&pt->list);
610 pr_warn("dev_remove_pack: %p not found\n", pt);
612 spin_unlock(&ptype_lock);
614 EXPORT_SYMBOL(__dev_remove_pack);
617 * dev_remove_pack - remove packet handler
618 * @pt: packet type declaration
620 * Remove a protocol handler that was previously added to the kernel
621 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
622 * from the kernel lists and can be freed or reused once this function
625 * This call sleeps to guarantee that no CPU is looking at the packet
628 void dev_remove_pack(struct packet_type *pt)
630 __dev_remove_pack(pt);
634 EXPORT_SYMBOL(dev_remove_pack);
637 /*******************************************************************************
639 * Device Interface Subroutines
641 *******************************************************************************/
644 * dev_get_iflink - get 'iflink' value of a interface
645 * @dev: targeted interface
647 * Indicates the ifindex the interface is linked to.
648 * Physical interfaces have the same 'ifindex' and 'iflink' values.
651 int dev_get_iflink(const struct net_device *dev)
653 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink)
654 return dev->netdev_ops->ndo_get_iflink(dev);
658 EXPORT_SYMBOL(dev_get_iflink);
661 * dev_fill_metadata_dst - Retrieve tunnel egress information.
662 * @dev: targeted interface
665 * For better visibility of tunnel traffic OVS needs to retrieve
666 * egress tunnel information for a packet. Following API allows
667 * user to get this info.
669 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb)
671 struct ip_tunnel_info *info;
673 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst)
676 info = skb_tunnel_info_unclone(skb);
679 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX)))
682 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb);
684 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst);
686 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack)
688 int k = stack->num_paths++;
690 if (WARN_ON_ONCE(k >= NET_DEVICE_PATH_STACK_MAX))
693 return &stack->path[k];
696 int dev_fill_forward_path(const struct net_device *dev, const u8 *daddr,
697 struct net_device_path_stack *stack)
699 const struct net_device *last_dev;
700 struct net_device_path_ctx ctx = {
703 struct net_device_path *path;
706 memcpy(ctx.daddr, daddr, sizeof(ctx.daddr));
707 stack->num_paths = 0;
708 while (ctx.dev && ctx.dev->netdev_ops->ndo_fill_forward_path) {
710 path = dev_fwd_path(stack);
714 memset(path, 0, sizeof(struct net_device_path));
715 ret = ctx.dev->netdev_ops->ndo_fill_forward_path(&ctx, path);
719 if (WARN_ON_ONCE(last_dev == ctx.dev))
726 path = dev_fwd_path(stack);
729 path->type = DEV_PATH_ETHERNET;
734 EXPORT_SYMBOL_GPL(dev_fill_forward_path);
737 * __dev_get_by_name - find a device by its name
738 * @net: the applicable net namespace
739 * @name: name to find
741 * Find an interface by name. Must be called under RTNL semaphore
742 * or @dev_base_lock. If the name is found a pointer to the device
743 * is returned. If the name is not found then %NULL is returned. The
744 * reference counters are not incremented so the caller must be
745 * careful with locks.
748 struct net_device *__dev_get_by_name(struct net *net, const char *name)
750 struct netdev_name_node *node_name;
752 node_name = netdev_name_node_lookup(net, name);
753 return node_name ? node_name->dev : NULL;
755 EXPORT_SYMBOL(__dev_get_by_name);
758 * dev_get_by_name_rcu - find a device by its name
759 * @net: the applicable net namespace
760 * @name: name to find
762 * Find an interface by name.
763 * If the name is found a pointer to the device is returned.
764 * If the name is not found then %NULL is returned.
765 * The reference counters are not incremented so the caller must be
766 * careful with locks. The caller must hold RCU lock.
769 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name)
771 struct netdev_name_node *node_name;
773 node_name = netdev_name_node_lookup_rcu(net, name);
774 return node_name ? node_name->dev : NULL;
776 EXPORT_SYMBOL(dev_get_by_name_rcu);
778 /* Deprecated for new users, call netdev_get_by_name() instead */
779 struct net_device *dev_get_by_name(struct net *net, const char *name)
781 struct net_device *dev;
784 dev = dev_get_by_name_rcu(net, name);
789 EXPORT_SYMBOL(dev_get_by_name);
792 * netdev_get_by_name() - find a device by its name
793 * @net: the applicable net namespace
794 * @name: name to find
795 * @tracker: tracking object for the acquired reference
796 * @gfp: allocation flags for the tracker
798 * Find an interface by name. This can be called from any
799 * context and does its own locking. The returned handle has
800 * the usage count incremented and the caller must use netdev_put() to
801 * release it when it is no longer needed. %NULL is returned if no
802 * matching device is found.
804 struct net_device *netdev_get_by_name(struct net *net, const char *name,
805 netdevice_tracker *tracker, gfp_t gfp)
807 struct net_device *dev;
809 dev = dev_get_by_name(net, name);
811 netdev_tracker_alloc(dev, tracker, gfp);
814 EXPORT_SYMBOL(netdev_get_by_name);
817 * __dev_get_by_index - find a device by its ifindex
818 * @net: the applicable net namespace
819 * @ifindex: index of device
821 * Search for an interface by index. Returns %NULL if the device
822 * is not found or a pointer to the device. The device has not
823 * had its reference counter increased so the caller must be careful
824 * about locking. The caller must hold either the RTNL semaphore
828 struct net_device *__dev_get_by_index(struct net *net, int ifindex)
830 struct net_device *dev;
831 struct hlist_head *head = dev_index_hash(net, ifindex);
833 hlist_for_each_entry(dev, head, index_hlist)
834 if (dev->ifindex == ifindex)
839 EXPORT_SYMBOL(__dev_get_by_index);
842 * dev_get_by_index_rcu - find a device by its ifindex
843 * @net: the applicable net namespace
844 * @ifindex: index of device
846 * Search for an interface by index. Returns %NULL if the device
847 * is not found or a pointer to the device. The device has not
848 * had its reference counter increased so the caller must be careful
849 * about locking. The caller must hold RCU lock.
852 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex)
854 struct net_device *dev;
855 struct hlist_head *head = dev_index_hash(net, ifindex);
857 hlist_for_each_entry_rcu(dev, head, index_hlist)
858 if (dev->ifindex == ifindex)
863 EXPORT_SYMBOL(dev_get_by_index_rcu);
865 /* Deprecated for new users, call netdev_get_by_index() instead */
866 struct net_device *dev_get_by_index(struct net *net, int ifindex)
868 struct net_device *dev;
871 dev = dev_get_by_index_rcu(net, ifindex);
876 EXPORT_SYMBOL(dev_get_by_index);
879 * netdev_get_by_index() - find a device by its ifindex
880 * @net: the applicable net namespace
881 * @ifindex: index of device
882 * @tracker: tracking object for the acquired reference
883 * @gfp: allocation flags for the tracker
885 * Search for an interface by index. Returns NULL if the device
886 * is not found or a pointer to the device. The device returned has
887 * had a reference added and the pointer is safe until the user calls
888 * netdev_put() to indicate they have finished with it.
890 struct net_device *netdev_get_by_index(struct net *net, int ifindex,
891 netdevice_tracker *tracker, gfp_t gfp)
893 struct net_device *dev;
895 dev = dev_get_by_index(net, ifindex);
897 netdev_tracker_alloc(dev, tracker, gfp);
900 EXPORT_SYMBOL(netdev_get_by_index);
903 * dev_get_by_napi_id - find a device by napi_id
904 * @napi_id: ID of the NAPI struct
906 * Search for an interface by NAPI ID. Returns %NULL if the device
907 * is not found or a pointer to the device. The device has not had
908 * its reference counter increased so the caller must be careful
909 * about locking. The caller must hold RCU lock.
912 struct net_device *dev_get_by_napi_id(unsigned int napi_id)
914 struct napi_struct *napi;
916 WARN_ON_ONCE(!rcu_read_lock_held());
918 if (napi_id < MIN_NAPI_ID)
921 napi = napi_by_id(napi_id);
923 return napi ? napi->dev : NULL;
925 EXPORT_SYMBOL(dev_get_by_napi_id);
928 * netdev_get_name - get a netdevice name, knowing its ifindex.
929 * @net: network namespace
930 * @name: a pointer to the buffer where the name will be stored.
931 * @ifindex: the ifindex of the interface to get the name from.
933 int netdev_get_name(struct net *net, char *name, int ifindex)
935 struct net_device *dev;
938 down_read(&devnet_rename_sem);
941 dev = dev_get_by_index_rcu(net, ifindex);
947 strcpy(name, dev->name);
952 up_read(&devnet_rename_sem);
957 * dev_getbyhwaddr_rcu - find a device by its hardware address
958 * @net: the applicable net namespace
959 * @type: media type of device
960 * @ha: hardware address
962 * Search for an interface by MAC address. Returns NULL if the device
963 * is not found or a pointer to the device.
964 * The caller must hold RCU or RTNL.
965 * The returned device has not had its ref count increased
966 * and the caller must therefore be careful about locking
970 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type,
973 struct net_device *dev;
975 for_each_netdev_rcu(net, dev)
976 if (dev->type == type &&
977 !memcmp(dev->dev_addr, ha, dev->addr_len))
982 EXPORT_SYMBOL(dev_getbyhwaddr_rcu);
984 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type)
986 struct net_device *dev, *ret = NULL;
989 for_each_netdev_rcu(net, dev)
990 if (dev->type == type) {
998 EXPORT_SYMBOL(dev_getfirstbyhwtype);
1001 * __dev_get_by_flags - find any device with given flags
1002 * @net: the applicable net namespace
1003 * @if_flags: IFF_* values
1004 * @mask: bitmask of bits in if_flags to check
1006 * Search for any interface with the given flags. Returns NULL if a device
1007 * is not found or a pointer to the device. Must be called inside
1008 * rtnl_lock(), and result refcount is unchanged.
1011 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags,
1012 unsigned short mask)
1014 struct net_device *dev, *ret;
1019 for_each_netdev(net, dev) {
1020 if (((dev->flags ^ if_flags) & mask) == 0) {
1027 EXPORT_SYMBOL(__dev_get_by_flags);
1030 * dev_valid_name - check if name is okay for network device
1031 * @name: name string
1033 * Network device names need to be valid file names to
1034 * allow sysfs to work. We also disallow any kind of
1037 bool dev_valid_name(const char *name)
1041 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ)
1043 if (!strcmp(name, ".") || !strcmp(name, ".."))
1047 if (*name == '/' || *name == ':' || isspace(*name))
1053 EXPORT_SYMBOL(dev_valid_name);
1056 * __dev_alloc_name - allocate a name for a device
1057 * @net: network namespace to allocate the device name in
1058 * @name: name format string
1059 * @res: result name string
1061 * Passed a format string - eg "lt%d" it will try and find a suitable
1062 * id. It scans list of devices to build up a free map, then chooses
1063 * the first empty slot. The caller must hold the dev_base or rtnl lock
1064 * while allocating the name and adding the device in order to avoid
1066 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1067 * Returns the number of the unit assigned or a negative errno code.
1070 static int __dev_alloc_name(struct net *net, const char *name, char *res)
1074 const int max_netdevices = 8*PAGE_SIZE;
1075 unsigned long *inuse;
1076 struct net_device *d;
1079 /* Verify the string as this thing may have come from the user.
1080 * There must be one "%d" and no other "%" characters.
1082 p = strchr(name, '%');
1083 if (!p || p[1] != 'd' || strchr(p + 2, '%'))
1086 /* Use one page as a bit array of possible slots */
1087 inuse = bitmap_zalloc(max_netdevices, GFP_ATOMIC);
1091 for_each_netdev(net, d) {
1092 struct netdev_name_node *name_node;
1094 netdev_for_each_altname(d, name_node) {
1095 if (!sscanf(name_node->name, name, &i))
1097 if (i < 0 || i >= max_netdevices)
1100 /* avoid cases where sscanf is not exact inverse of printf */
1101 snprintf(buf, IFNAMSIZ, name, i);
1102 if (!strncmp(buf, name_node->name, IFNAMSIZ))
1103 __set_bit(i, inuse);
1105 if (!sscanf(d->name, name, &i))
1107 if (i < 0 || i >= max_netdevices)
1110 /* avoid cases where sscanf is not exact inverse of printf */
1111 snprintf(buf, IFNAMSIZ, name, i);
1112 if (!strncmp(buf, d->name, IFNAMSIZ))
1113 __set_bit(i, inuse);
1116 i = find_first_zero_bit(inuse, max_netdevices);
1118 if (i == max_netdevices)
1121 /* 'res' and 'name' could overlap, use 'buf' as an intermediate buffer */
1122 strscpy(buf, name, IFNAMSIZ);
1123 snprintf(res, IFNAMSIZ, buf, i);
1127 /* Returns negative errno or allocated unit id (see __dev_alloc_name()) */
1128 static int dev_prep_valid_name(struct net *net, struct net_device *dev,
1129 const char *want_name, char *out_name,
1132 if (!dev_valid_name(want_name))
1135 if (strchr(want_name, '%'))
1136 return __dev_alloc_name(net, want_name, out_name);
1138 if (netdev_name_in_use(net, want_name))
1140 if (out_name != want_name)
1141 strscpy(out_name, want_name, IFNAMSIZ);
1146 * dev_alloc_name - allocate a name for a device
1148 * @name: name format string
1150 * Passed a format string - eg "lt%d" it will try and find a suitable
1151 * id. It scans list of devices to build up a free map, then chooses
1152 * the first empty slot. The caller must hold the dev_base or rtnl lock
1153 * while allocating the name and adding the device in order to avoid
1155 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1156 * Returns the number of the unit assigned or a negative errno code.
1159 int dev_alloc_name(struct net_device *dev, const char *name)
1161 return dev_prep_valid_name(dev_net(dev), dev, name, dev->name, ENFILE);
1163 EXPORT_SYMBOL(dev_alloc_name);
1165 static int dev_get_valid_name(struct net *net, struct net_device *dev,
1170 ret = dev_prep_valid_name(net, dev, name, dev->name, EEXIST);
1171 return ret < 0 ? ret : 0;
1175 * dev_change_name - change name of a device
1177 * @newname: name (or format string) must be at least IFNAMSIZ
1179 * Change name of a device, can pass format strings "eth%d".
1182 int dev_change_name(struct net_device *dev, const char *newname)
1184 unsigned char old_assign_type;
1185 char oldname[IFNAMSIZ];
1191 BUG_ON(!dev_net(dev));
1195 down_write(&devnet_rename_sem);
1197 if (strncmp(newname, dev->name, IFNAMSIZ) == 0) {
1198 up_write(&devnet_rename_sem);
1202 memcpy(oldname, dev->name, IFNAMSIZ);
1204 err = dev_get_valid_name(net, dev, newname);
1206 up_write(&devnet_rename_sem);
1210 if (oldname[0] && !strchr(oldname, '%'))
1211 netdev_info(dev, "renamed from %s%s\n", oldname,
1212 dev->flags & IFF_UP ? " (while UP)" : "");
1214 old_assign_type = dev->name_assign_type;
1215 dev->name_assign_type = NET_NAME_RENAMED;
1218 ret = device_rename(&dev->dev, dev->name);
1220 memcpy(dev->name, oldname, IFNAMSIZ);
1221 dev->name_assign_type = old_assign_type;
1222 up_write(&devnet_rename_sem);
1226 up_write(&devnet_rename_sem);
1228 netdev_adjacent_rename_links(dev, oldname);
1230 write_lock(&dev_base_lock);
1231 netdev_name_node_del(dev->name_node);
1232 write_unlock(&dev_base_lock);
1236 write_lock(&dev_base_lock);
1237 netdev_name_node_add(net, dev->name_node);
1238 write_unlock(&dev_base_lock);
1240 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev);
1241 ret = notifier_to_errno(ret);
1244 /* err >= 0 after dev_alloc_name() or stores the first errno */
1247 down_write(&devnet_rename_sem);
1248 memcpy(dev->name, oldname, IFNAMSIZ);
1249 memcpy(oldname, newname, IFNAMSIZ);
1250 dev->name_assign_type = old_assign_type;
1251 old_assign_type = NET_NAME_RENAMED;
1254 netdev_err(dev, "name change rollback failed: %d\n",
1263 * dev_set_alias - change ifalias of a device
1265 * @alias: name up to IFALIASZ
1266 * @len: limit of bytes to copy from info
1268 * Set ifalias for a device,
1270 int dev_set_alias(struct net_device *dev, const char *alias, size_t len)
1272 struct dev_ifalias *new_alias = NULL;
1274 if (len >= IFALIASZ)
1278 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL);
1282 memcpy(new_alias->ifalias, alias, len);
1283 new_alias->ifalias[len] = 0;
1286 mutex_lock(&ifalias_mutex);
1287 new_alias = rcu_replace_pointer(dev->ifalias, new_alias,
1288 mutex_is_locked(&ifalias_mutex));
1289 mutex_unlock(&ifalias_mutex);
1292 kfree_rcu(new_alias, rcuhead);
1296 EXPORT_SYMBOL(dev_set_alias);
1299 * dev_get_alias - get ifalias of a device
1301 * @name: buffer to store name of ifalias
1302 * @len: size of buffer
1304 * get ifalias for a device. Caller must make sure dev cannot go
1305 * away, e.g. rcu read lock or own a reference count to device.
1307 int dev_get_alias(const struct net_device *dev, char *name, size_t len)
1309 const struct dev_ifalias *alias;
1313 alias = rcu_dereference(dev->ifalias);
1315 ret = snprintf(name, len, "%s", alias->ifalias);
1322 * netdev_features_change - device changes features
1323 * @dev: device to cause notification
1325 * Called to indicate a device has changed features.
1327 void netdev_features_change(struct net_device *dev)
1329 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev);
1331 EXPORT_SYMBOL(netdev_features_change);
1334 * netdev_state_change - device changes state
1335 * @dev: device to cause notification
1337 * Called to indicate a device has changed state. This function calls
1338 * the notifier chains for netdev_chain and sends a NEWLINK message
1339 * to the routing socket.
1341 void netdev_state_change(struct net_device *dev)
1343 if (dev->flags & IFF_UP) {
1344 struct netdev_notifier_change_info change_info = {
1348 call_netdevice_notifiers_info(NETDEV_CHANGE,
1350 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL, 0, NULL);
1353 EXPORT_SYMBOL(netdev_state_change);
1356 * __netdev_notify_peers - notify network peers about existence of @dev,
1357 * to be called when rtnl lock is already held.
1358 * @dev: network device
1360 * Generate traffic such that interested network peers are aware of
1361 * @dev, such as by generating a gratuitous ARP. This may be used when
1362 * a device wants to inform the rest of the network about some sort of
1363 * reconfiguration such as a failover event or virtual machine
1366 void __netdev_notify_peers(struct net_device *dev)
1369 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
1370 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev);
1372 EXPORT_SYMBOL(__netdev_notify_peers);
1375 * netdev_notify_peers - notify network peers about existence of @dev
1376 * @dev: network device
1378 * Generate traffic such that interested network peers are aware of
1379 * @dev, such as by generating a gratuitous ARP. This may be used when
1380 * a device wants to inform the rest of the network about some sort of
1381 * reconfiguration such as a failover event or virtual machine
1384 void netdev_notify_peers(struct net_device *dev)
1387 __netdev_notify_peers(dev);
1390 EXPORT_SYMBOL(netdev_notify_peers);
1392 static int napi_threaded_poll(void *data);
1394 static int napi_kthread_create(struct napi_struct *n)
1398 /* Create and wake up the kthread once to put it in
1399 * TASK_INTERRUPTIBLE mode to avoid the blocked task
1400 * warning and work with loadavg.
1402 n->thread = kthread_run(napi_threaded_poll, n, "napi/%s-%d",
1403 n->dev->name, n->napi_id);
1404 if (IS_ERR(n->thread)) {
1405 err = PTR_ERR(n->thread);
1406 pr_err("kthread_run failed with err %d\n", err);
1413 static int __dev_open(struct net_device *dev, struct netlink_ext_ack *extack)
1415 const struct net_device_ops *ops = dev->netdev_ops;
1419 dev_addr_check(dev);
1421 if (!netif_device_present(dev)) {
1422 /* may be detached because parent is runtime-suspended */
1423 if (dev->dev.parent)
1424 pm_runtime_resume(dev->dev.parent);
1425 if (!netif_device_present(dev))
1429 /* Block netpoll from trying to do any rx path servicing.
1430 * If we don't do this there is a chance ndo_poll_controller
1431 * or ndo_poll may be running while we open the device
1433 netpoll_poll_disable(dev);
1435 ret = call_netdevice_notifiers_extack(NETDEV_PRE_UP, dev, extack);
1436 ret = notifier_to_errno(ret);
1440 set_bit(__LINK_STATE_START, &dev->state);
1442 if (ops->ndo_validate_addr)
1443 ret = ops->ndo_validate_addr(dev);
1445 if (!ret && ops->ndo_open)
1446 ret = ops->ndo_open(dev);
1448 netpoll_poll_enable(dev);
1451 clear_bit(__LINK_STATE_START, &dev->state);
1453 dev->flags |= IFF_UP;
1454 dev_set_rx_mode(dev);
1456 add_device_randomness(dev->dev_addr, dev->addr_len);
1463 * dev_open - prepare an interface for use.
1464 * @dev: device to open
1465 * @extack: netlink extended ack
1467 * Takes a device from down to up state. The device's private open
1468 * function is invoked and then the multicast lists are loaded. Finally
1469 * the device is moved into the up state and a %NETDEV_UP message is
1470 * sent to the netdev notifier chain.
1472 * Calling this function on an active interface is a nop. On a failure
1473 * a negative errno code is returned.
1475 int dev_open(struct net_device *dev, struct netlink_ext_ack *extack)
1479 if (dev->flags & IFF_UP)
1482 ret = __dev_open(dev, extack);
1486 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL);
1487 call_netdevice_notifiers(NETDEV_UP, dev);
1491 EXPORT_SYMBOL(dev_open);
1493 static void __dev_close_many(struct list_head *head)
1495 struct net_device *dev;
1500 list_for_each_entry(dev, head, close_list) {
1501 /* Temporarily disable netpoll until the interface is down */
1502 netpoll_poll_disable(dev);
1504 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev);
1506 clear_bit(__LINK_STATE_START, &dev->state);
1508 /* Synchronize to scheduled poll. We cannot touch poll list, it
1509 * can be even on different cpu. So just clear netif_running().
1511 * dev->stop() will invoke napi_disable() on all of it's
1512 * napi_struct instances on this device.
1514 smp_mb__after_atomic(); /* Commit netif_running(). */
1517 dev_deactivate_many(head);
1519 list_for_each_entry(dev, head, close_list) {
1520 const struct net_device_ops *ops = dev->netdev_ops;
1523 * Call the device specific close. This cannot fail.
1524 * Only if device is UP
1526 * We allow it to be called even after a DETACH hot-plug
1532 dev->flags &= ~IFF_UP;
1533 netpoll_poll_enable(dev);
1537 static void __dev_close(struct net_device *dev)
1541 list_add(&dev->close_list, &single);
1542 __dev_close_many(&single);
1546 void dev_close_many(struct list_head *head, bool unlink)
1548 struct net_device *dev, *tmp;
1550 /* Remove the devices that don't need to be closed */
1551 list_for_each_entry_safe(dev, tmp, head, close_list)
1552 if (!(dev->flags & IFF_UP))
1553 list_del_init(&dev->close_list);
1555 __dev_close_many(head);
1557 list_for_each_entry_safe(dev, tmp, head, close_list) {
1558 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL);
1559 call_netdevice_notifiers(NETDEV_DOWN, dev);
1561 list_del_init(&dev->close_list);
1564 EXPORT_SYMBOL(dev_close_many);
1567 * dev_close - shutdown an interface.
1568 * @dev: device to shutdown
1570 * This function moves an active device into down state. A
1571 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device
1572 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier
1575 void dev_close(struct net_device *dev)
1577 if (dev->flags & IFF_UP) {
1580 list_add(&dev->close_list, &single);
1581 dev_close_many(&single, true);
1585 EXPORT_SYMBOL(dev_close);
1589 * dev_disable_lro - disable Large Receive Offload on a device
1592 * Disable Large Receive Offload (LRO) on a net device. Must be
1593 * called under RTNL. This is needed if received packets may be
1594 * forwarded to another interface.
1596 void dev_disable_lro(struct net_device *dev)
1598 struct net_device *lower_dev;
1599 struct list_head *iter;
1601 dev->wanted_features &= ~NETIF_F_LRO;
1602 netdev_update_features(dev);
1604 if (unlikely(dev->features & NETIF_F_LRO))
1605 netdev_WARN(dev, "failed to disable LRO!\n");
1607 netdev_for_each_lower_dev(dev, lower_dev, iter)
1608 dev_disable_lro(lower_dev);
1610 EXPORT_SYMBOL(dev_disable_lro);
1613 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device
1616 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be
1617 * called under RTNL. This is needed if Generic XDP is installed on
1620 static void dev_disable_gro_hw(struct net_device *dev)
1622 dev->wanted_features &= ~NETIF_F_GRO_HW;
1623 netdev_update_features(dev);
1625 if (unlikely(dev->features & NETIF_F_GRO_HW))
1626 netdev_WARN(dev, "failed to disable GRO_HW!\n");
1629 const char *netdev_cmd_to_name(enum netdev_cmd cmd)
1632 case NETDEV_##val: \
1633 return "NETDEV_" __stringify(val);
1635 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER)
1636 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE)
1637 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE)
1638 N(POST_INIT) N(PRE_UNINIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN)
1639 N(CHANGEUPPER) N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA)
1640 N(BONDING_INFO) N(PRECHANGEUPPER) N(CHANGELOWERSTATE)
1641 N(UDP_TUNNEL_PUSH_INFO) N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN)
1642 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO)
1643 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO)
1644 N(PRE_CHANGEADDR) N(OFFLOAD_XSTATS_ENABLE) N(OFFLOAD_XSTATS_DISABLE)
1645 N(OFFLOAD_XSTATS_REPORT_USED) N(OFFLOAD_XSTATS_REPORT_DELTA)
1649 return "UNKNOWN_NETDEV_EVENT";
1651 EXPORT_SYMBOL_GPL(netdev_cmd_to_name);
1653 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val,
1654 struct net_device *dev)
1656 struct netdev_notifier_info info = {
1660 return nb->notifier_call(nb, val, &info);
1663 static int call_netdevice_register_notifiers(struct notifier_block *nb,
1664 struct net_device *dev)
1668 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev);
1669 err = notifier_to_errno(err);
1673 if (!(dev->flags & IFF_UP))
1676 call_netdevice_notifier(nb, NETDEV_UP, dev);
1680 static void call_netdevice_unregister_notifiers(struct notifier_block *nb,
1681 struct net_device *dev)
1683 if (dev->flags & IFF_UP) {
1684 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1686 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1688 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1691 static int call_netdevice_register_net_notifiers(struct notifier_block *nb,
1694 struct net_device *dev;
1697 for_each_netdev(net, dev) {
1698 err = call_netdevice_register_notifiers(nb, dev);
1705 for_each_netdev_continue_reverse(net, dev)
1706 call_netdevice_unregister_notifiers(nb, dev);
1710 static void call_netdevice_unregister_net_notifiers(struct notifier_block *nb,
1713 struct net_device *dev;
1715 for_each_netdev(net, dev)
1716 call_netdevice_unregister_notifiers(nb, dev);
1719 static int dev_boot_phase = 1;
1722 * register_netdevice_notifier - register a network notifier block
1725 * Register a notifier to be called when network device events occur.
1726 * The notifier passed is linked into the kernel structures and must
1727 * not be reused until it has been unregistered. A negative errno code
1728 * is returned on a failure.
1730 * When registered all registration and up events are replayed
1731 * to the new notifier to allow device to have a race free
1732 * view of the network device list.
1735 int register_netdevice_notifier(struct notifier_block *nb)
1740 /* Close race with setup_net() and cleanup_net() */
1741 down_write(&pernet_ops_rwsem);
1743 err = raw_notifier_chain_register(&netdev_chain, nb);
1749 err = call_netdevice_register_net_notifiers(nb, net);
1756 up_write(&pernet_ops_rwsem);
1760 for_each_net_continue_reverse(net)
1761 call_netdevice_unregister_net_notifiers(nb, net);
1763 raw_notifier_chain_unregister(&netdev_chain, nb);
1766 EXPORT_SYMBOL(register_netdevice_notifier);
1769 * unregister_netdevice_notifier - unregister a network notifier block
1772 * Unregister a notifier previously registered by
1773 * register_netdevice_notifier(). The notifier is unlinked into the
1774 * kernel structures and may then be reused. A negative errno code
1775 * is returned on a failure.
1777 * After unregistering unregister and down device events are synthesized
1778 * for all devices on the device list to the removed notifier to remove
1779 * the need for special case cleanup code.
1782 int unregister_netdevice_notifier(struct notifier_block *nb)
1787 /* Close race with setup_net() and cleanup_net() */
1788 down_write(&pernet_ops_rwsem);
1790 err = raw_notifier_chain_unregister(&netdev_chain, nb);
1795 call_netdevice_unregister_net_notifiers(nb, net);
1799 up_write(&pernet_ops_rwsem);
1802 EXPORT_SYMBOL(unregister_netdevice_notifier);
1804 static int __register_netdevice_notifier_net(struct net *net,
1805 struct notifier_block *nb,
1806 bool ignore_call_fail)
1810 err = raw_notifier_chain_register(&net->netdev_chain, nb);
1816 err = call_netdevice_register_net_notifiers(nb, net);
1817 if (err && !ignore_call_fail)
1818 goto chain_unregister;
1823 raw_notifier_chain_unregister(&net->netdev_chain, nb);
1827 static int __unregister_netdevice_notifier_net(struct net *net,
1828 struct notifier_block *nb)
1832 err = raw_notifier_chain_unregister(&net->netdev_chain, nb);
1836 call_netdevice_unregister_net_notifiers(nb, net);
1841 * register_netdevice_notifier_net - register a per-netns network notifier block
1842 * @net: network namespace
1845 * Register a notifier to be called when network device events occur.
1846 * The notifier passed is linked into the kernel structures and must
1847 * not be reused until it has been unregistered. A negative errno code
1848 * is returned on a failure.
1850 * When registered all registration and up events are replayed
1851 * to the new notifier to allow device to have a race free
1852 * view of the network device list.
1855 int register_netdevice_notifier_net(struct net *net, struct notifier_block *nb)
1860 err = __register_netdevice_notifier_net(net, nb, false);
1864 EXPORT_SYMBOL(register_netdevice_notifier_net);
1867 * unregister_netdevice_notifier_net - unregister a per-netns
1868 * network notifier block
1869 * @net: network namespace
1872 * Unregister a notifier previously registered by
1873 * register_netdevice_notifier_net(). The notifier is unlinked from the
1874 * kernel structures and may then be reused. A negative errno code
1875 * is returned on a failure.
1877 * After unregistering unregister and down device events are synthesized
1878 * for all devices on the device list to the removed notifier to remove
1879 * the need for special case cleanup code.
1882 int unregister_netdevice_notifier_net(struct net *net,
1883 struct notifier_block *nb)
1888 err = __unregister_netdevice_notifier_net(net, nb);
1892 EXPORT_SYMBOL(unregister_netdevice_notifier_net);
1894 static void __move_netdevice_notifier_net(struct net *src_net,
1895 struct net *dst_net,
1896 struct notifier_block *nb)
1898 __unregister_netdevice_notifier_net(src_net, nb);
1899 __register_netdevice_notifier_net(dst_net, nb, true);
1902 int register_netdevice_notifier_dev_net(struct net_device *dev,
1903 struct notifier_block *nb,
1904 struct netdev_net_notifier *nn)
1909 err = __register_netdevice_notifier_net(dev_net(dev), nb, false);
1912 list_add(&nn->list, &dev->net_notifier_list);
1917 EXPORT_SYMBOL(register_netdevice_notifier_dev_net);
1919 int unregister_netdevice_notifier_dev_net(struct net_device *dev,
1920 struct notifier_block *nb,
1921 struct netdev_net_notifier *nn)
1926 list_del(&nn->list);
1927 err = __unregister_netdevice_notifier_net(dev_net(dev), nb);
1931 EXPORT_SYMBOL(unregister_netdevice_notifier_dev_net);
1933 static void move_netdevice_notifiers_dev_net(struct net_device *dev,
1936 struct netdev_net_notifier *nn;
1938 list_for_each_entry(nn, &dev->net_notifier_list, list)
1939 __move_netdevice_notifier_net(dev_net(dev), net, nn->nb);
1943 * call_netdevice_notifiers_info - call all network notifier blocks
1944 * @val: value passed unmodified to notifier function
1945 * @info: notifier information data
1947 * Call all network notifier blocks. Parameters and return value
1948 * are as for raw_notifier_call_chain().
1951 int call_netdevice_notifiers_info(unsigned long val,
1952 struct netdev_notifier_info *info)
1954 struct net *net = dev_net(info->dev);
1959 /* Run per-netns notifier block chain first, then run the global one.
1960 * Hopefully, one day, the global one is going to be removed after
1961 * all notifier block registrators get converted to be per-netns.
1963 ret = raw_notifier_call_chain(&net->netdev_chain, val, info);
1964 if (ret & NOTIFY_STOP_MASK)
1966 return raw_notifier_call_chain(&netdev_chain, val, info);
1970 * call_netdevice_notifiers_info_robust - call per-netns notifier blocks
1971 * for and rollback on error
1972 * @val_up: value passed unmodified to notifier function
1973 * @val_down: value passed unmodified to the notifier function when
1974 * recovering from an error on @val_up
1975 * @info: notifier information data
1977 * Call all per-netns network notifier blocks, but not notifier blocks on
1978 * the global notifier chain. Parameters and return value are as for
1979 * raw_notifier_call_chain_robust().
1983 call_netdevice_notifiers_info_robust(unsigned long val_up,
1984 unsigned long val_down,
1985 struct netdev_notifier_info *info)
1987 struct net *net = dev_net(info->dev);
1991 return raw_notifier_call_chain_robust(&net->netdev_chain,
1992 val_up, val_down, info);
1995 static int call_netdevice_notifiers_extack(unsigned long val,
1996 struct net_device *dev,
1997 struct netlink_ext_ack *extack)
1999 struct netdev_notifier_info info = {
2004 return call_netdevice_notifiers_info(val, &info);
2008 * call_netdevice_notifiers - call all network notifier blocks
2009 * @val: value passed unmodified to notifier function
2010 * @dev: net_device pointer passed unmodified to notifier function
2012 * Call all network notifier blocks. Parameters and return value
2013 * are as for raw_notifier_call_chain().
2016 int call_netdevice_notifiers(unsigned long val, struct net_device *dev)
2018 return call_netdevice_notifiers_extack(val, dev, NULL);
2020 EXPORT_SYMBOL(call_netdevice_notifiers);
2023 * call_netdevice_notifiers_mtu - call all network notifier blocks
2024 * @val: value passed unmodified to notifier function
2025 * @dev: net_device pointer passed unmodified to notifier function
2026 * @arg: additional u32 argument passed to the notifier function
2028 * Call all network notifier blocks. Parameters and return value
2029 * are as for raw_notifier_call_chain().
2031 static int call_netdevice_notifiers_mtu(unsigned long val,
2032 struct net_device *dev, u32 arg)
2034 struct netdev_notifier_info_ext info = {
2039 BUILD_BUG_ON(offsetof(struct netdev_notifier_info_ext, info) != 0);
2041 return call_netdevice_notifiers_info(val, &info.info);
2044 #ifdef CONFIG_NET_INGRESS
2045 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key);
2047 void net_inc_ingress_queue(void)
2049 static_branch_inc(&ingress_needed_key);
2051 EXPORT_SYMBOL_GPL(net_inc_ingress_queue);
2053 void net_dec_ingress_queue(void)
2055 static_branch_dec(&ingress_needed_key);
2057 EXPORT_SYMBOL_GPL(net_dec_ingress_queue);
2060 #ifdef CONFIG_NET_EGRESS
2061 static DEFINE_STATIC_KEY_FALSE(egress_needed_key);
2063 void net_inc_egress_queue(void)
2065 static_branch_inc(&egress_needed_key);
2067 EXPORT_SYMBOL_GPL(net_inc_egress_queue);
2069 void net_dec_egress_queue(void)
2071 static_branch_dec(&egress_needed_key);
2073 EXPORT_SYMBOL_GPL(net_dec_egress_queue);
2076 DEFINE_STATIC_KEY_FALSE(netstamp_needed_key);
2077 EXPORT_SYMBOL(netstamp_needed_key);
2078 #ifdef CONFIG_JUMP_LABEL
2079 static atomic_t netstamp_needed_deferred;
2080 static atomic_t netstamp_wanted;
2081 static void netstamp_clear(struct work_struct *work)
2083 int deferred = atomic_xchg(&netstamp_needed_deferred, 0);
2086 wanted = atomic_add_return(deferred, &netstamp_wanted);
2088 static_branch_enable(&netstamp_needed_key);
2090 static_branch_disable(&netstamp_needed_key);
2092 static DECLARE_WORK(netstamp_work, netstamp_clear);
2095 void net_enable_timestamp(void)
2097 #ifdef CONFIG_JUMP_LABEL
2098 int wanted = atomic_read(&netstamp_wanted);
2100 while (wanted > 0) {
2101 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted + 1))
2104 atomic_inc(&netstamp_needed_deferred);
2105 schedule_work(&netstamp_work);
2107 static_branch_inc(&netstamp_needed_key);
2110 EXPORT_SYMBOL(net_enable_timestamp);
2112 void net_disable_timestamp(void)
2114 #ifdef CONFIG_JUMP_LABEL
2115 int wanted = atomic_read(&netstamp_wanted);
2117 while (wanted > 1) {
2118 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted - 1))
2121 atomic_dec(&netstamp_needed_deferred);
2122 schedule_work(&netstamp_work);
2124 static_branch_dec(&netstamp_needed_key);
2127 EXPORT_SYMBOL(net_disable_timestamp);
2129 static inline void net_timestamp_set(struct sk_buff *skb)
2132 skb->mono_delivery_time = 0;
2133 if (static_branch_unlikely(&netstamp_needed_key))
2134 skb->tstamp = ktime_get_real();
2137 #define net_timestamp_check(COND, SKB) \
2138 if (static_branch_unlikely(&netstamp_needed_key)) { \
2139 if ((COND) && !(SKB)->tstamp) \
2140 (SKB)->tstamp = ktime_get_real(); \
2143 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb)
2145 return __is_skb_forwardable(dev, skb, true);
2147 EXPORT_SYMBOL_GPL(is_skb_forwardable);
2149 static int __dev_forward_skb2(struct net_device *dev, struct sk_buff *skb,
2152 int ret = ____dev_forward_skb(dev, skb, check_mtu);
2155 skb->protocol = eth_type_trans(skb, dev);
2156 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
2162 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
2164 return __dev_forward_skb2(dev, skb, true);
2166 EXPORT_SYMBOL_GPL(__dev_forward_skb);
2169 * dev_forward_skb - loopback an skb to another netif
2171 * @dev: destination network device
2172 * @skb: buffer to forward
2175 * NET_RX_SUCCESS (no congestion)
2176 * NET_RX_DROP (packet was dropped, but freed)
2178 * dev_forward_skb can be used for injecting an skb from the
2179 * start_xmit function of one device into the receive queue
2180 * of another device.
2182 * The receiving device may be in another namespace, so
2183 * we have to clear all information in the skb that could
2184 * impact namespace isolation.
2186 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
2188 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb);
2190 EXPORT_SYMBOL_GPL(dev_forward_skb);
2192 int dev_forward_skb_nomtu(struct net_device *dev, struct sk_buff *skb)
2194 return __dev_forward_skb2(dev, skb, false) ?: netif_rx_internal(skb);
2197 static inline int deliver_skb(struct sk_buff *skb,
2198 struct packet_type *pt_prev,
2199 struct net_device *orig_dev)
2201 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
2203 refcount_inc(&skb->users);
2204 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
2207 static inline void deliver_ptype_list_skb(struct sk_buff *skb,
2208 struct packet_type **pt,
2209 struct net_device *orig_dev,
2211 struct list_head *ptype_list)
2213 struct packet_type *ptype, *pt_prev = *pt;
2215 list_for_each_entry_rcu(ptype, ptype_list, list) {
2216 if (ptype->type != type)
2219 deliver_skb(skb, pt_prev, orig_dev);
2225 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb)
2227 if (!ptype->af_packet_priv || !skb->sk)
2230 if (ptype->id_match)
2231 return ptype->id_match(ptype, skb->sk);
2232 else if ((struct sock *)ptype->af_packet_priv == skb->sk)
2239 * dev_nit_active - return true if any network interface taps are in use
2241 * @dev: network device to check for the presence of taps
2243 bool dev_nit_active(struct net_device *dev)
2245 return !list_empty(&ptype_all) || !list_empty(&dev->ptype_all);
2247 EXPORT_SYMBOL_GPL(dev_nit_active);
2250 * Support routine. Sends outgoing frames to any network
2251 * taps currently in use.
2254 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
2256 struct packet_type *ptype;
2257 struct sk_buff *skb2 = NULL;
2258 struct packet_type *pt_prev = NULL;
2259 struct list_head *ptype_list = &ptype_all;
2263 list_for_each_entry_rcu(ptype, ptype_list, list) {
2264 if (READ_ONCE(ptype->ignore_outgoing))
2267 /* Never send packets back to the socket
2268 * they originated from - MvS (miquels@drinkel.ow.org)
2270 if (skb_loop_sk(ptype, skb))
2274 deliver_skb(skb2, pt_prev, skb->dev);
2279 /* need to clone skb, done only once */
2280 skb2 = skb_clone(skb, GFP_ATOMIC);
2284 net_timestamp_set(skb2);
2286 /* skb->nh should be correctly
2287 * set by sender, so that the second statement is
2288 * just protection against buggy protocols.
2290 skb_reset_mac_header(skb2);
2292 if (skb_network_header(skb2) < skb2->data ||
2293 skb_network_header(skb2) > skb_tail_pointer(skb2)) {
2294 net_crit_ratelimited("protocol %04x is buggy, dev %s\n",
2295 ntohs(skb2->protocol),
2297 skb_reset_network_header(skb2);
2300 skb2->transport_header = skb2->network_header;
2301 skb2->pkt_type = PACKET_OUTGOING;
2305 if (ptype_list == &ptype_all) {
2306 ptype_list = &dev->ptype_all;
2311 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC))
2312 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
2318 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit);
2321 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change
2322 * @dev: Network device
2323 * @txq: number of queues available
2325 * If real_num_tx_queues is changed the tc mappings may no longer be
2326 * valid. To resolve this verify the tc mapping remains valid and if
2327 * not NULL the mapping. With no priorities mapping to this
2328 * offset/count pair it will no longer be used. In the worst case TC0
2329 * is invalid nothing can be done so disable priority mappings. If is
2330 * expected that drivers will fix this mapping if they can before
2331 * calling netif_set_real_num_tx_queues.
2333 static void netif_setup_tc(struct net_device *dev, unsigned int txq)
2336 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2338 /* If TC0 is invalidated disable TC mapping */
2339 if (tc->offset + tc->count > txq) {
2340 netdev_warn(dev, "Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n");
2345 /* Invalidated prio to tc mappings set to TC0 */
2346 for (i = 1; i < TC_BITMASK + 1; i++) {
2347 int q = netdev_get_prio_tc_map(dev, i);
2349 tc = &dev->tc_to_txq[q];
2350 if (tc->offset + tc->count > txq) {
2351 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",
2353 netdev_set_prio_tc_map(dev, i, 0);
2358 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq)
2361 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2364 /* walk through the TCs and see if it falls into any of them */
2365 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) {
2366 if ((txq - tc->offset) < tc->count)
2370 /* didn't find it, just return -1 to indicate no match */
2376 EXPORT_SYMBOL(netdev_txq_to_tc);
2379 static struct static_key xps_needed __read_mostly;
2380 static struct static_key xps_rxqs_needed __read_mostly;
2381 static DEFINE_MUTEX(xps_map_mutex);
2382 #define xmap_dereference(P) \
2383 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex))
2385 static bool remove_xps_queue(struct xps_dev_maps *dev_maps,
2386 struct xps_dev_maps *old_maps, int tci, u16 index)
2388 struct xps_map *map = NULL;
2391 map = xmap_dereference(dev_maps->attr_map[tci]);
2395 for (pos = map->len; pos--;) {
2396 if (map->queues[pos] != index)
2400 map->queues[pos] = map->queues[--map->len];
2405 RCU_INIT_POINTER(old_maps->attr_map[tci], NULL);
2406 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL);
2407 kfree_rcu(map, rcu);
2414 static bool remove_xps_queue_cpu(struct net_device *dev,
2415 struct xps_dev_maps *dev_maps,
2416 int cpu, u16 offset, u16 count)
2418 int num_tc = dev_maps->num_tc;
2419 bool active = false;
2422 for (tci = cpu * num_tc; num_tc--; tci++) {
2425 for (i = count, j = offset; i--; j++) {
2426 if (!remove_xps_queue(dev_maps, NULL, tci, j))
2436 static void reset_xps_maps(struct net_device *dev,
2437 struct xps_dev_maps *dev_maps,
2438 enum xps_map_type type)
2440 static_key_slow_dec_cpuslocked(&xps_needed);
2441 if (type == XPS_RXQS)
2442 static_key_slow_dec_cpuslocked(&xps_rxqs_needed);
2444 RCU_INIT_POINTER(dev->xps_maps[type], NULL);
2446 kfree_rcu(dev_maps, rcu);
2449 static void clean_xps_maps(struct net_device *dev, enum xps_map_type type,
2450 u16 offset, u16 count)
2452 struct xps_dev_maps *dev_maps;
2453 bool active = false;
2456 dev_maps = xmap_dereference(dev->xps_maps[type]);
2460 for (j = 0; j < dev_maps->nr_ids; j++)
2461 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, count);
2463 reset_xps_maps(dev, dev_maps, type);
2465 if (type == XPS_CPUS) {
2466 for (i = offset + (count - 1); count--; i--)
2467 netdev_queue_numa_node_write(
2468 netdev_get_tx_queue(dev, i), NUMA_NO_NODE);
2472 static void netif_reset_xps_queues(struct net_device *dev, u16 offset,
2475 if (!static_key_false(&xps_needed))
2479 mutex_lock(&xps_map_mutex);
2481 if (static_key_false(&xps_rxqs_needed))
2482 clean_xps_maps(dev, XPS_RXQS, offset, count);
2484 clean_xps_maps(dev, XPS_CPUS, offset, count);
2486 mutex_unlock(&xps_map_mutex);
2490 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index)
2492 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index);
2495 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index,
2496 u16 index, bool is_rxqs_map)
2498 struct xps_map *new_map;
2499 int alloc_len = XPS_MIN_MAP_ALLOC;
2502 for (pos = 0; map && pos < map->len; pos++) {
2503 if (map->queues[pos] != index)
2508 /* Need to add tx-queue to this CPU's/rx-queue's existing map */
2510 if (pos < map->alloc_len)
2513 alloc_len = map->alloc_len * 2;
2516 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's
2520 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL);
2522 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL,
2523 cpu_to_node(attr_index));
2527 for (i = 0; i < pos; i++)
2528 new_map->queues[i] = map->queues[i];
2529 new_map->alloc_len = alloc_len;
2535 /* Copy xps maps at a given index */
2536 static void xps_copy_dev_maps(struct xps_dev_maps *dev_maps,
2537 struct xps_dev_maps *new_dev_maps, int index,
2538 int tc, bool skip_tc)
2540 int i, tci = index * dev_maps->num_tc;
2541 struct xps_map *map;
2543 /* copy maps belonging to foreign traffic classes */
2544 for (i = 0; i < dev_maps->num_tc; i++, tci++) {
2545 if (i == tc && skip_tc)
2548 /* fill in the new device map from the old device map */
2549 map = xmap_dereference(dev_maps->attr_map[tci]);
2550 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map);
2554 /* Must be called under cpus_read_lock */
2555 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask,
2556 u16 index, enum xps_map_type type)
2558 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL, *old_dev_maps = NULL;
2559 const unsigned long *online_mask = NULL;
2560 bool active = false, copy = false;
2561 int i, j, tci, numa_node_id = -2;
2562 int maps_sz, num_tc = 1, tc = 0;
2563 struct xps_map *map, *new_map;
2564 unsigned int nr_ids;
2566 WARN_ON_ONCE(index >= dev->num_tx_queues);
2569 /* Do not allow XPS on subordinate device directly */
2570 num_tc = dev->num_tc;
2574 /* If queue belongs to subordinate dev use its map */
2575 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev;
2577 tc = netdev_txq_to_tc(dev, index);
2582 mutex_lock(&xps_map_mutex);
2584 dev_maps = xmap_dereference(dev->xps_maps[type]);
2585 if (type == XPS_RXQS) {
2586 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues);
2587 nr_ids = dev->num_rx_queues;
2589 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc);
2590 if (num_possible_cpus() > 1)
2591 online_mask = cpumask_bits(cpu_online_mask);
2592 nr_ids = nr_cpu_ids;
2595 if (maps_sz < L1_CACHE_BYTES)
2596 maps_sz = L1_CACHE_BYTES;
2598 /* The old dev_maps could be larger or smaller than the one we're
2599 * setting up now, as dev->num_tc or nr_ids could have been updated in
2600 * between. We could try to be smart, but let's be safe instead and only
2601 * copy foreign traffic classes if the two map sizes match.
2604 dev_maps->num_tc == num_tc && dev_maps->nr_ids == nr_ids)
2607 /* allocate memory for queue storage */
2608 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids),
2610 if (!new_dev_maps) {
2611 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL);
2612 if (!new_dev_maps) {
2613 mutex_unlock(&xps_map_mutex);
2617 new_dev_maps->nr_ids = nr_ids;
2618 new_dev_maps->num_tc = num_tc;
2621 tci = j * num_tc + tc;
2622 map = copy ? xmap_dereference(dev_maps->attr_map[tci]) : NULL;
2624 map = expand_xps_map(map, j, index, type == XPS_RXQS);
2628 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map);
2632 goto out_no_new_maps;
2635 /* Increment static keys at most once per type */
2636 static_key_slow_inc_cpuslocked(&xps_needed);
2637 if (type == XPS_RXQS)
2638 static_key_slow_inc_cpuslocked(&xps_rxqs_needed);
2641 for (j = 0; j < nr_ids; j++) {
2642 bool skip_tc = false;
2644 tci = j * num_tc + tc;
2645 if (netif_attr_test_mask(j, mask, nr_ids) &&
2646 netif_attr_test_online(j, online_mask, nr_ids)) {
2647 /* add tx-queue to CPU/rx-queue maps */
2652 map = xmap_dereference(new_dev_maps->attr_map[tci]);
2653 while ((pos < map->len) && (map->queues[pos] != index))
2656 if (pos == map->len)
2657 map->queues[map->len++] = index;
2659 if (type == XPS_CPUS) {
2660 if (numa_node_id == -2)
2661 numa_node_id = cpu_to_node(j);
2662 else if (numa_node_id != cpu_to_node(j))
2669 xps_copy_dev_maps(dev_maps, new_dev_maps, j, tc,
2673 rcu_assign_pointer(dev->xps_maps[type], new_dev_maps);
2675 /* Cleanup old maps */
2677 goto out_no_old_maps;
2679 for (j = 0; j < dev_maps->nr_ids; j++) {
2680 for (i = num_tc, tci = j * dev_maps->num_tc; i--; tci++) {
2681 map = xmap_dereference(dev_maps->attr_map[tci]);
2686 new_map = xmap_dereference(new_dev_maps->attr_map[tci]);
2691 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL);
2692 kfree_rcu(map, rcu);
2696 old_dev_maps = dev_maps;
2699 dev_maps = new_dev_maps;
2703 if (type == XPS_CPUS)
2704 /* update Tx queue numa node */
2705 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index),
2706 (numa_node_id >= 0) ?
2707 numa_node_id : NUMA_NO_NODE);
2712 /* removes tx-queue from unused CPUs/rx-queues */
2713 for (j = 0; j < dev_maps->nr_ids; j++) {
2714 tci = j * dev_maps->num_tc;
2716 for (i = 0; i < dev_maps->num_tc; i++, tci++) {
2718 netif_attr_test_mask(j, mask, dev_maps->nr_ids) &&
2719 netif_attr_test_online(j, online_mask, dev_maps->nr_ids))
2722 active |= remove_xps_queue(dev_maps,
2723 copy ? old_dev_maps : NULL,
2729 kfree_rcu(old_dev_maps, rcu);
2731 /* free map if not active */
2733 reset_xps_maps(dev, dev_maps, type);
2736 mutex_unlock(&xps_map_mutex);
2740 /* remove any maps that we added */
2741 for (j = 0; j < nr_ids; j++) {
2742 for (i = num_tc, tci = j * num_tc; i--; tci++) {
2743 new_map = xmap_dereference(new_dev_maps->attr_map[tci]);
2745 xmap_dereference(dev_maps->attr_map[tci]) :
2747 if (new_map && new_map != map)
2752 mutex_unlock(&xps_map_mutex);
2754 kfree(new_dev_maps);
2757 EXPORT_SYMBOL_GPL(__netif_set_xps_queue);
2759 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask,
2765 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, XPS_CPUS);
2770 EXPORT_SYMBOL(netif_set_xps_queue);
2773 static void netdev_unbind_all_sb_channels(struct net_device *dev)
2775 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues];
2777 /* Unbind any subordinate channels */
2778 while (txq-- != &dev->_tx[0]) {
2780 netdev_unbind_sb_channel(dev, txq->sb_dev);
2784 void netdev_reset_tc(struct net_device *dev)
2787 netif_reset_xps_queues_gt(dev, 0);
2789 netdev_unbind_all_sb_channels(dev);
2791 /* Reset TC configuration of device */
2793 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq));
2794 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map));
2796 EXPORT_SYMBOL(netdev_reset_tc);
2798 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset)
2800 if (tc >= dev->num_tc)
2804 netif_reset_xps_queues(dev, offset, count);
2806 dev->tc_to_txq[tc].count = count;
2807 dev->tc_to_txq[tc].offset = offset;
2810 EXPORT_SYMBOL(netdev_set_tc_queue);
2812 int netdev_set_num_tc(struct net_device *dev, u8 num_tc)
2814 if (num_tc > TC_MAX_QUEUE)
2818 netif_reset_xps_queues_gt(dev, 0);
2820 netdev_unbind_all_sb_channels(dev);
2822 dev->num_tc = num_tc;
2825 EXPORT_SYMBOL(netdev_set_num_tc);
2827 void netdev_unbind_sb_channel(struct net_device *dev,
2828 struct net_device *sb_dev)
2830 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues];
2833 netif_reset_xps_queues_gt(sb_dev, 0);
2835 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq));
2836 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map));
2838 while (txq-- != &dev->_tx[0]) {
2839 if (txq->sb_dev == sb_dev)
2843 EXPORT_SYMBOL(netdev_unbind_sb_channel);
2845 int netdev_bind_sb_channel_queue(struct net_device *dev,
2846 struct net_device *sb_dev,
2847 u8 tc, u16 count, u16 offset)
2849 /* Make certain the sb_dev and dev are already configured */
2850 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc)
2853 /* We cannot hand out queues we don't have */
2854 if ((offset + count) > dev->real_num_tx_queues)
2857 /* Record the mapping */
2858 sb_dev->tc_to_txq[tc].count = count;
2859 sb_dev->tc_to_txq[tc].offset = offset;
2861 /* Provide a way for Tx queue to find the tc_to_txq map or
2862 * XPS map for itself.
2865 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev;
2869 EXPORT_SYMBOL(netdev_bind_sb_channel_queue);
2871 int netdev_set_sb_channel(struct net_device *dev, u16 channel)
2873 /* Do not use a multiqueue device to represent a subordinate channel */
2874 if (netif_is_multiqueue(dev))
2877 /* We allow channels 1 - 32767 to be used for subordinate channels.
2878 * Channel 0 is meant to be "native" mode and used only to represent
2879 * the main root device. We allow writing 0 to reset the device back
2880 * to normal mode after being used as a subordinate channel.
2882 if (channel > S16_MAX)
2885 dev->num_tc = -channel;
2889 EXPORT_SYMBOL(netdev_set_sb_channel);
2892 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues
2893 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed.
2895 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq)
2900 disabling = txq < dev->real_num_tx_queues;
2902 if (txq < 1 || txq > dev->num_tx_queues)
2905 if (dev->reg_state == NETREG_REGISTERED ||
2906 dev->reg_state == NETREG_UNREGISTERING) {
2909 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues,
2915 netif_setup_tc(dev, txq);
2917 dev_qdisc_change_real_num_tx(dev, txq);
2919 dev->real_num_tx_queues = txq;
2923 qdisc_reset_all_tx_gt(dev, txq);
2925 netif_reset_xps_queues_gt(dev, txq);
2929 dev->real_num_tx_queues = txq;
2934 EXPORT_SYMBOL(netif_set_real_num_tx_queues);
2938 * netif_set_real_num_rx_queues - set actual number of RX queues used
2939 * @dev: Network device
2940 * @rxq: Actual number of RX queues
2942 * This must be called either with the rtnl_lock held or before
2943 * registration of the net device. Returns 0 on success, or a
2944 * negative error code. If called before registration, it always
2947 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq)
2951 if (rxq < 1 || rxq > dev->num_rx_queues)
2954 if (dev->reg_state == NETREG_REGISTERED) {
2957 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues,
2963 dev->real_num_rx_queues = rxq;
2966 EXPORT_SYMBOL(netif_set_real_num_rx_queues);
2970 * netif_set_real_num_queues - set actual number of RX and TX queues used
2971 * @dev: Network device
2972 * @txq: Actual number of TX queues
2973 * @rxq: Actual number of RX queues
2975 * Set the real number of both TX and RX queues.
2976 * Does nothing if the number of queues is already correct.
2978 int netif_set_real_num_queues(struct net_device *dev,
2979 unsigned int txq, unsigned int rxq)
2981 unsigned int old_rxq = dev->real_num_rx_queues;
2984 if (txq < 1 || txq > dev->num_tx_queues ||
2985 rxq < 1 || rxq > dev->num_rx_queues)
2988 /* Start from increases, so the error path only does decreases -
2989 * decreases can't fail.
2991 if (rxq > dev->real_num_rx_queues) {
2992 err = netif_set_real_num_rx_queues(dev, rxq);
2996 if (txq > dev->real_num_tx_queues) {
2997 err = netif_set_real_num_tx_queues(dev, txq);
3001 if (rxq < dev->real_num_rx_queues)
3002 WARN_ON(netif_set_real_num_rx_queues(dev, rxq));
3003 if (txq < dev->real_num_tx_queues)
3004 WARN_ON(netif_set_real_num_tx_queues(dev, txq));
3008 WARN_ON(netif_set_real_num_rx_queues(dev, old_rxq));
3011 EXPORT_SYMBOL(netif_set_real_num_queues);
3014 * netif_set_tso_max_size() - set the max size of TSO frames supported
3015 * @dev: netdev to update
3016 * @size: max skb->len of a TSO frame
3018 * Set the limit on the size of TSO super-frames the device can handle.
3019 * Unless explicitly set the stack will assume the value of
3020 * %GSO_LEGACY_MAX_SIZE.
3022 void netif_set_tso_max_size(struct net_device *dev, unsigned int size)
3024 dev->tso_max_size = min(GSO_MAX_SIZE, size);
3025 if (size < READ_ONCE(dev->gso_max_size))
3026 netif_set_gso_max_size(dev, size);
3027 if (size < READ_ONCE(dev->gso_ipv4_max_size))
3028 netif_set_gso_ipv4_max_size(dev, size);
3030 EXPORT_SYMBOL(netif_set_tso_max_size);
3033 * netif_set_tso_max_segs() - set the max number of segs supported for TSO
3034 * @dev: netdev to update
3035 * @segs: max number of TCP segments
3037 * Set the limit on the number of TCP segments the device can generate from
3038 * a single TSO super-frame.
3039 * Unless explicitly set the stack will assume the value of %GSO_MAX_SEGS.
3041 void netif_set_tso_max_segs(struct net_device *dev, unsigned int segs)
3043 dev->tso_max_segs = segs;
3044 if (segs < READ_ONCE(dev->gso_max_segs))
3045 netif_set_gso_max_segs(dev, segs);
3047 EXPORT_SYMBOL(netif_set_tso_max_segs);
3050 * netif_inherit_tso_max() - copy all TSO limits from a lower device to an upper
3051 * @to: netdev to update
3052 * @from: netdev from which to copy the limits
3054 void netif_inherit_tso_max(struct net_device *to, const struct net_device *from)
3056 netif_set_tso_max_size(to, from->tso_max_size);
3057 netif_set_tso_max_segs(to, from->tso_max_segs);
3059 EXPORT_SYMBOL(netif_inherit_tso_max);
3062 * netif_get_num_default_rss_queues - default number of RSS queues
3064 * Default value is the number of physical cores if there are only 1 or 2, or
3065 * divided by 2 if there are more.
3067 int netif_get_num_default_rss_queues(void)
3072 if (unlikely(is_kdump_kernel() || !zalloc_cpumask_var(&cpus, GFP_KERNEL)))
3075 cpumask_copy(cpus, cpu_online_mask);
3076 for_each_cpu(cpu, cpus) {
3078 cpumask_andnot(cpus, cpus, topology_sibling_cpumask(cpu));
3080 free_cpumask_var(cpus);
3082 return count > 2 ? DIV_ROUND_UP(count, 2) : count;
3084 EXPORT_SYMBOL(netif_get_num_default_rss_queues);
3086 static void __netif_reschedule(struct Qdisc *q)
3088 struct softnet_data *sd;
3089 unsigned long flags;
3091 local_irq_save(flags);
3092 sd = this_cpu_ptr(&softnet_data);
3093 q->next_sched = NULL;
3094 *sd->output_queue_tailp = q;
3095 sd->output_queue_tailp = &q->next_sched;
3096 raise_softirq_irqoff(NET_TX_SOFTIRQ);
3097 local_irq_restore(flags);
3100 void __netif_schedule(struct Qdisc *q)
3102 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
3103 __netif_reschedule(q);
3105 EXPORT_SYMBOL(__netif_schedule);
3107 struct dev_kfree_skb_cb {
3108 enum skb_drop_reason reason;
3111 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb)
3113 return (struct dev_kfree_skb_cb *)skb->cb;
3116 void netif_schedule_queue(struct netdev_queue *txq)
3119 if (!netif_xmit_stopped(txq)) {
3120 struct Qdisc *q = rcu_dereference(txq->qdisc);
3122 __netif_schedule(q);
3126 EXPORT_SYMBOL(netif_schedule_queue);
3128 void netif_tx_wake_queue(struct netdev_queue *dev_queue)
3130 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) {
3134 q = rcu_dereference(dev_queue->qdisc);
3135 __netif_schedule(q);
3139 EXPORT_SYMBOL(netif_tx_wake_queue);
3141 void dev_kfree_skb_irq_reason(struct sk_buff *skb, enum skb_drop_reason reason)
3143 unsigned long flags;
3148 if (likely(refcount_read(&skb->users) == 1)) {
3150 refcount_set(&skb->users, 0);
3151 } else if (likely(!refcount_dec_and_test(&skb->users))) {
3154 get_kfree_skb_cb(skb)->reason = reason;
3155 local_irq_save(flags);
3156 skb->next = __this_cpu_read(softnet_data.completion_queue);
3157 __this_cpu_write(softnet_data.completion_queue, skb);
3158 raise_softirq_irqoff(NET_TX_SOFTIRQ);
3159 local_irq_restore(flags);
3161 EXPORT_SYMBOL(dev_kfree_skb_irq_reason);
3163 void dev_kfree_skb_any_reason(struct sk_buff *skb, enum skb_drop_reason reason)
3165 if (in_hardirq() || irqs_disabled())
3166 dev_kfree_skb_irq_reason(skb, reason);
3168 kfree_skb_reason(skb, reason);
3170 EXPORT_SYMBOL(dev_kfree_skb_any_reason);
3174 * netif_device_detach - mark device as removed
3175 * @dev: network device
3177 * Mark device as removed from system and therefore no longer available.
3179 void netif_device_detach(struct net_device *dev)
3181 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) &&
3182 netif_running(dev)) {
3183 netif_tx_stop_all_queues(dev);
3186 EXPORT_SYMBOL(netif_device_detach);
3189 * netif_device_attach - mark device as attached
3190 * @dev: network device
3192 * Mark device as attached from system and restart if needed.
3194 void netif_device_attach(struct net_device *dev)
3196 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) &&
3197 netif_running(dev)) {
3198 netif_tx_wake_all_queues(dev);
3199 __netdev_watchdog_up(dev);
3202 EXPORT_SYMBOL(netif_device_attach);
3205 * Returns a Tx hash based on the given packet descriptor a Tx queues' number
3206 * to be used as a distribution range.
3208 static u16 skb_tx_hash(const struct net_device *dev,
3209 const struct net_device *sb_dev,
3210 struct sk_buff *skb)
3214 u16 qcount = dev->real_num_tx_queues;
3217 u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
3219 qoffset = sb_dev->tc_to_txq[tc].offset;
3220 qcount = sb_dev->tc_to_txq[tc].count;
3221 if (unlikely(!qcount)) {
3222 net_warn_ratelimited("%s: invalid qcount, qoffset %u for tc %u\n",
3223 sb_dev->name, qoffset, tc);
3225 qcount = dev->real_num_tx_queues;
3229 if (skb_rx_queue_recorded(skb)) {
3230 DEBUG_NET_WARN_ON_ONCE(qcount == 0);
3231 hash = skb_get_rx_queue(skb);
3232 if (hash >= qoffset)
3234 while (unlikely(hash >= qcount))
3236 return hash + qoffset;
3239 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset;
3242 void skb_warn_bad_offload(const struct sk_buff *skb)
3244 static const netdev_features_t null_features;
3245 struct net_device *dev = skb->dev;
3246 const char *name = "";
3248 if (!net_ratelimit())
3252 if (dev->dev.parent)
3253 name = dev_driver_string(dev->dev.parent);
3255 name = netdev_name(dev);
3257 skb_dump(KERN_WARNING, skb, false);
3258 WARN(1, "%s: caps=(%pNF, %pNF)\n",
3259 name, dev ? &dev->features : &null_features,
3260 skb->sk ? &skb->sk->sk_route_caps : &null_features);
3264 * Invalidate hardware checksum when packet is to be mangled, and
3265 * complete checksum manually on outgoing path.
3267 int skb_checksum_help(struct sk_buff *skb)
3270 int ret = 0, offset;
3272 if (skb->ip_summed == CHECKSUM_COMPLETE)
3273 goto out_set_summed;
3275 if (unlikely(skb_is_gso(skb))) {
3276 skb_warn_bad_offload(skb);
3280 /* Before computing a checksum, we should make sure no frag could
3281 * be modified by an external entity : checksum could be wrong.
3283 if (skb_has_shared_frag(skb)) {
3284 ret = __skb_linearize(skb);
3289 offset = skb_checksum_start_offset(skb);
3291 if (unlikely(offset >= skb_headlen(skb))) {
3292 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false);
3293 WARN_ONCE(true, "offset (%d) >= skb_headlen() (%u)\n",
3294 offset, skb_headlen(skb));
3297 csum = skb_checksum(skb, offset, skb->len - offset, 0);
3299 offset += skb->csum_offset;
3300 if (unlikely(offset + sizeof(__sum16) > skb_headlen(skb))) {
3301 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false);
3302 WARN_ONCE(true, "offset+2 (%zu) > skb_headlen() (%u)\n",
3303 offset + sizeof(__sum16), skb_headlen(skb));
3306 ret = skb_ensure_writable(skb, offset + sizeof(__sum16));
3310 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0;
3312 skb->ip_summed = CHECKSUM_NONE;
3316 EXPORT_SYMBOL(skb_checksum_help);
3318 int skb_crc32c_csum_help(struct sk_buff *skb)
3321 int ret = 0, offset, start;
3323 if (skb->ip_summed != CHECKSUM_PARTIAL)
3326 if (unlikely(skb_is_gso(skb)))
3329 /* Before computing a checksum, we should make sure no frag could
3330 * be modified by an external entity : checksum could be wrong.
3332 if (unlikely(skb_has_shared_frag(skb))) {
3333 ret = __skb_linearize(skb);
3337 start = skb_checksum_start_offset(skb);
3338 offset = start + offsetof(struct sctphdr, checksum);
3339 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) {
3344 ret = skb_ensure_writable(skb, offset + sizeof(__le32));
3348 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start,
3349 skb->len - start, ~(__u32)0,
3351 *(__le32 *)(skb->data + offset) = crc32c_csum;
3352 skb_reset_csum_not_inet(skb);
3357 __be16 skb_network_protocol(struct sk_buff *skb, int *depth)
3359 __be16 type = skb->protocol;
3361 /* Tunnel gso handlers can set protocol to ethernet. */
3362 if (type == htons(ETH_P_TEB)) {
3365 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr))))
3368 eth = (struct ethhdr *)skb->data;
3369 type = eth->h_proto;
3372 return vlan_get_protocol_and_depth(skb, type, depth);
3376 /* Take action when hardware reception checksum errors are detected. */
3378 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb)
3380 netdev_err(dev, "hw csum failure\n");
3381 skb_dump(KERN_ERR, skb, true);
3385 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb)
3387 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb);
3389 EXPORT_SYMBOL(netdev_rx_csum_fault);
3392 /* XXX: check that highmem exists at all on the given machine. */
3393 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
3395 #ifdef CONFIG_HIGHMEM
3398 if (!(dev->features & NETIF_F_HIGHDMA)) {
3399 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
3400 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
3402 if (PageHighMem(skb_frag_page(frag)))
3410 /* If MPLS offload request, verify we are testing hardware MPLS features
3411 * instead of standard features for the netdev.
3413 #if IS_ENABLED(CONFIG_NET_MPLS_GSO)
3414 static netdev_features_t net_mpls_features(struct sk_buff *skb,
3415 netdev_features_t features,
3418 if (eth_p_mpls(type))
3419 features &= skb->dev->mpls_features;
3424 static netdev_features_t net_mpls_features(struct sk_buff *skb,
3425 netdev_features_t features,
3432 static netdev_features_t harmonize_features(struct sk_buff *skb,
3433 netdev_features_t features)
3437 type = skb_network_protocol(skb, NULL);
3438 features = net_mpls_features(skb, features, type);
3440 if (skb->ip_summed != CHECKSUM_NONE &&
3441 !can_checksum_protocol(features, type)) {
3442 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
3444 if (illegal_highdma(skb->dev, skb))
3445 features &= ~NETIF_F_SG;
3450 netdev_features_t passthru_features_check(struct sk_buff *skb,
3451 struct net_device *dev,
3452 netdev_features_t features)
3456 EXPORT_SYMBOL(passthru_features_check);
3458 static netdev_features_t dflt_features_check(struct sk_buff *skb,
3459 struct net_device *dev,
3460 netdev_features_t features)
3462 return vlan_features_check(skb, features);
3465 static netdev_features_t gso_features_check(const struct sk_buff *skb,
3466 struct net_device *dev,
3467 netdev_features_t features)
3469 u16 gso_segs = skb_shinfo(skb)->gso_segs;
3471 if (gso_segs > READ_ONCE(dev->gso_max_segs))
3472 return features & ~NETIF_F_GSO_MASK;
3474 if (unlikely(skb->len >= READ_ONCE(dev->gso_max_size)))
3475 return features & ~NETIF_F_GSO_MASK;
3477 if (!skb_shinfo(skb)->gso_type) {
3478 skb_warn_bad_offload(skb);
3479 return features & ~NETIF_F_GSO_MASK;
3482 /* Support for GSO partial features requires software
3483 * intervention before we can actually process the packets
3484 * so we need to strip support for any partial features now
3485 * and we can pull them back in after we have partially
3486 * segmented the frame.
3488 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL))
3489 features &= ~dev->gso_partial_features;
3491 /* Make sure to clear the IPv4 ID mangling feature if the
3492 * IPv4 header has the potential to be fragmented.
3494 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
3495 struct iphdr *iph = skb->encapsulation ?
3496 inner_ip_hdr(skb) : ip_hdr(skb);
3498 if (!(iph->frag_off & htons(IP_DF)))
3499 features &= ~NETIF_F_TSO_MANGLEID;
3505 netdev_features_t netif_skb_features(struct sk_buff *skb)
3507 struct net_device *dev = skb->dev;
3508 netdev_features_t features = dev->features;
3510 if (skb_is_gso(skb))
3511 features = gso_features_check(skb, dev, features);
3513 /* If encapsulation offload request, verify we are testing
3514 * hardware encapsulation features instead of standard
3515 * features for the netdev
3517 if (skb->encapsulation)
3518 features &= dev->hw_enc_features;
3520 if (skb_vlan_tagged(skb))
3521 features = netdev_intersect_features(features,
3522 dev->vlan_features |
3523 NETIF_F_HW_VLAN_CTAG_TX |
3524 NETIF_F_HW_VLAN_STAG_TX);
3526 if (dev->netdev_ops->ndo_features_check)
3527 features &= dev->netdev_ops->ndo_features_check(skb, dev,
3530 features &= dflt_features_check(skb, dev, features);
3532 return harmonize_features(skb, features);
3534 EXPORT_SYMBOL(netif_skb_features);
3536 static int xmit_one(struct sk_buff *skb, struct net_device *dev,
3537 struct netdev_queue *txq, bool more)
3542 if (dev_nit_active(dev))
3543 dev_queue_xmit_nit(skb, dev);
3546 trace_net_dev_start_xmit(skb, dev);
3547 rc = netdev_start_xmit(skb, dev, txq, more);
3548 trace_net_dev_xmit(skb, rc, dev, len);
3553 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev,
3554 struct netdev_queue *txq, int *ret)
3556 struct sk_buff *skb = first;
3557 int rc = NETDEV_TX_OK;
3560 struct sk_buff *next = skb->next;
3562 skb_mark_not_on_list(skb);
3563 rc = xmit_one(skb, dev, txq, next != NULL);
3564 if (unlikely(!dev_xmit_complete(rc))) {
3570 if (netif_tx_queue_stopped(txq) && skb) {
3571 rc = NETDEV_TX_BUSY;
3581 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb,
3582 netdev_features_t features)
3584 if (skb_vlan_tag_present(skb) &&
3585 !vlan_hw_offload_capable(features, skb->vlan_proto))
3586 skb = __vlan_hwaccel_push_inside(skb);
3590 int skb_csum_hwoffload_help(struct sk_buff *skb,
3591 const netdev_features_t features)
3593 if (unlikely(skb_csum_is_sctp(skb)))
3594 return !!(features & NETIF_F_SCTP_CRC) ? 0 :
3595 skb_crc32c_csum_help(skb);
3597 if (features & NETIF_F_HW_CSUM)
3600 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) {
3601 switch (skb->csum_offset) {
3602 case offsetof(struct tcphdr, check):
3603 case offsetof(struct udphdr, check):
3608 return skb_checksum_help(skb);
3610 EXPORT_SYMBOL(skb_csum_hwoffload_help);
3612 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again)
3614 netdev_features_t features;
3616 features = netif_skb_features(skb);
3617 skb = validate_xmit_vlan(skb, features);
3621 skb = sk_validate_xmit_skb(skb, dev);
3625 if (netif_needs_gso(skb, features)) {
3626 struct sk_buff *segs;
3628 segs = skb_gso_segment(skb, features);
3636 if (skb_needs_linearize(skb, features) &&
3637 __skb_linearize(skb))
3640 /* If packet is not checksummed and device does not
3641 * support checksumming for this protocol, complete
3642 * checksumming here.
3644 if (skb->ip_summed == CHECKSUM_PARTIAL) {
3645 if (skb->encapsulation)
3646 skb_set_inner_transport_header(skb,
3647 skb_checksum_start_offset(skb));
3649 skb_set_transport_header(skb,
3650 skb_checksum_start_offset(skb));
3651 if (skb_csum_hwoffload_help(skb, features))
3656 skb = validate_xmit_xfrm(skb, features, again);
3663 dev_core_stats_tx_dropped_inc(dev);
3667 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again)
3669 struct sk_buff *next, *head = NULL, *tail;
3671 for (; skb != NULL; skb = next) {
3673 skb_mark_not_on_list(skb);
3675 /* in case skb wont be segmented, point to itself */
3678 skb = validate_xmit_skb(skb, dev, again);
3686 /* If skb was segmented, skb->prev points to
3687 * the last segment. If not, it still contains skb.
3693 EXPORT_SYMBOL_GPL(validate_xmit_skb_list);
3695 static void qdisc_pkt_len_init(struct sk_buff *skb)
3697 const struct skb_shared_info *shinfo = skb_shinfo(skb);
3699 qdisc_skb_cb(skb)->pkt_len = skb->len;
3701 /* To get more precise estimation of bytes sent on wire,
3702 * we add to pkt_len the headers size of all segments
3704 if (shinfo->gso_size && skb_transport_header_was_set(skb)) {
3705 u16 gso_segs = shinfo->gso_segs;
3706 unsigned int hdr_len;
3708 /* mac layer + network layer */
3709 hdr_len = skb_transport_offset(skb);
3711 /* + transport layer */
3712 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) {
3713 const struct tcphdr *th;
3714 struct tcphdr _tcphdr;
3716 th = skb_header_pointer(skb, hdr_len,
3717 sizeof(_tcphdr), &_tcphdr);
3719 hdr_len += __tcp_hdrlen(th);
3721 struct udphdr _udphdr;
3723 if (skb_header_pointer(skb, hdr_len,
3724 sizeof(_udphdr), &_udphdr))
3725 hdr_len += sizeof(struct udphdr);
3728 if (shinfo->gso_type & SKB_GSO_DODGY)
3729 gso_segs = DIV_ROUND_UP(skb->len - hdr_len,
3732 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len;
3736 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q,
3737 struct sk_buff **to_free,
3738 struct netdev_queue *txq)
3742 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK;
3743 if (rc == NET_XMIT_SUCCESS)
3744 trace_qdisc_enqueue(q, txq, skb);
3748 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
3749 struct net_device *dev,
3750 struct netdev_queue *txq)
3752 spinlock_t *root_lock = qdisc_lock(q);
3753 struct sk_buff *to_free = NULL;
3757 qdisc_calculate_pkt_len(skb, q);
3759 tcf_set_drop_reason(skb, SKB_DROP_REASON_QDISC_DROP);
3761 if (q->flags & TCQ_F_NOLOCK) {
3762 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) &&
3763 qdisc_run_begin(q)) {
3764 /* Retest nolock_qdisc_is_empty() within the protection
3765 * of q->seqlock to protect from racing with requeuing.
3767 if (unlikely(!nolock_qdisc_is_empty(q))) {
3768 rc = dev_qdisc_enqueue(skb, q, &to_free, txq);
3775 qdisc_bstats_cpu_update(q, skb);
3776 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) &&
3777 !nolock_qdisc_is_empty(q))
3781 return NET_XMIT_SUCCESS;
3784 rc = dev_qdisc_enqueue(skb, q, &to_free, txq);
3788 if (unlikely(to_free))
3789 kfree_skb_list_reason(to_free,
3790 tcf_get_drop_reason(to_free));
3794 if (unlikely(READ_ONCE(q->owner) == smp_processor_id())) {
3795 kfree_skb_reason(skb, SKB_DROP_REASON_TC_RECLASSIFY_LOOP);
3796 return NET_XMIT_DROP;
3799 * Heuristic to force contended enqueues to serialize on a
3800 * separate lock before trying to get qdisc main lock.
3801 * This permits qdisc->running owner to get the lock more
3802 * often and dequeue packets faster.
3803 * On PREEMPT_RT it is possible to preempt the qdisc owner during xmit
3804 * and then other tasks will only enqueue packets. The packets will be
3805 * sent after the qdisc owner is scheduled again. To prevent this
3806 * scenario the task always serialize on the lock.
3808 contended = qdisc_is_running(q) || IS_ENABLED(CONFIG_PREEMPT_RT);
3809 if (unlikely(contended))
3810 spin_lock(&q->busylock);
3812 spin_lock(root_lock);
3813 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3814 __qdisc_drop(skb, &to_free);
3816 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
3817 qdisc_run_begin(q)) {
3819 * This is a work-conserving queue; there are no old skbs
3820 * waiting to be sent out; and the qdisc is not running -
3821 * xmit the skb directly.
3824 qdisc_bstats_update(q, skb);
3826 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) {
3827 if (unlikely(contended)) {
3828 spin_unlock(&q->busylock);
3835 rc = NET_XMIT_SUCCESS;
3837 WRITE_ONCE(q->owner, smp_processor_id());
3838 rc = dev_qdisc_enqueue(skb, q, &to_free, txq);
3839 WRITE_ONCE(q->owner, -1);
3840 if (qdisc_run_begin(q)) {
3841 if (unlikely(contended)) {
3842 spin_unlock(&q->busylock);
3849 spin_unlock(root_lock);
3850 if (unlikely(to_free))
3851 kfree_skb_list_reason(to_free,
3852 tcf_get_drop_reason(to_free));
3853 if (unlikely(contended))
3854 spin_unlock(&q->busylock);
3858 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO)
3859 static void skb_update_prio(struct sk_buff *skb)
3861 const struct netprio_map *map;
3862 const struct sock *sk;
3863 unsigned int prioidx;
3867 map = rcu_dereference_bh(skb->dev->priomap);
3870 sk = skb_to_full_sk(skb);
3874 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data);
3876 if (prioidx < map->priomap_len)
3877 skb->priority = map->priomap[prioidx];
3880 #define skb_update_prio(skb)
3884 * dev_loopback_xmit - loop back @skb
3885 * @net: network namespace this loopback is happening in
3886 * @sk: sk needed to be a netfilter okfn
3887 * @skb: buffer to transmit
3889 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb)
3891 skb_reset_mac_header(skb);
3892 __skb_pull(skb, skb_network_offset(skb));
3893 skb->pkt_type = PACKET_LOOPBACK;
3894 if (skb->ip_summed == CHECKSUM_NONE)
3895 skb->ip_summed = CHECKSUM_UNNECESSARY;
3896 DEBUG_NET_WARN_ON_ONCE(!skb_dst(skb));
3901 EXPORT_SYMBOL(dev_loopback_xmit);
3903 #ifdef CONFIG_NET_EGRESS
3904 static struct netdev_queue *
3905 netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb)
3907 int qm = skb_get_queue_mapping(skb);
3909 return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm));
3912 static bool netdev_xmit_txqueue_skipped(void)
3914 return __this_cpu_read(softnet_data.xmit.skip_txqueue);
3917 void netdev_xmit_skip_txqueue(bool skip)
3919 __this_cpu_write(softnet_data.xmit.skip_txqueue, skip);
3921 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue);
3922 #endif /* CONFIG_NET_EGRESS */
3924 #ifdef CONFIG_NET_XGRESS
3925 static int tc_run(struct tcx_entry *entry, struct sk_buff *skb,
3926 enum skb_drop_reason *drop_reason)
3928 int ret = TC_ACT_UNSPEC;
3929 #ifdef CONFIG_NET_CLS_ACT
3930 struct mini_Qdisc *miniq = rcu_dereference_bh(entry->miniq);
3931 struct tcf_result res;
3936 tc_skb_cb(skb)->mru = 0;
3937 tc_skb_cb(skb)->post_ct = false;
3938 tcf_set_drop_reason(skb, *drop_reason);
3940 mini_qdisc_bstats_cpu_update(miniq, skb);
3941 ret = tcf_classify(skb, miniq->block, miniq->filter_list, &res, false);
3942 /* Only tcf related quirks below. */
3945 *drop_reason = tcf_get_drop_reason(skb);
3946 mini_qdisc_qstats_cpu_drop(miniq);
3949 case TC_ACT_RECLASSIFY:
3950 skb->tc_index = TC_H_MIN(res.classid);
3953 #endif /* CONFIG_NET_CLS_ACT */
3957 static DEFINE_STATIC_KEY_FALSE(tcx_needed_key);
3961 static_branch_inc(&tcx_needed_key);
3966 static_branch_dec(&tcx_needed_key);
3969 static __always_inline enum tcx_action_base
3970 tcx_run(const struct bpf_mprog_entry *entry, struct sk_buff *skb,
3971 const bool needs_mac)
3973 const struct bpf_mprog_fp *fp;
3974 const struct bpf_prog *prog;
3978 __skb_push(skb, skb->mac_len);
3979 bpf_mprog_foreach_prog(entry, fp, prog) {
3980 bpf_compute_data_pointers(skb);
3981 ret = bpf_prog_run(prog, skb);
3982 if (ret != TCX_NEXT)
3986 __skb_pull(skb, skb->mac_len);
3987 return tcx_action_code(skb, ret);
3990 static __always_inline struct sk_buff *
3991 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
3992 struct net_device *orig_dev, bool *another)
3994 struct bpf_mprog_entry *entry = rcu_dereference_bh(skb->dev->tcx_ingress);
3995 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_INGRESS;
4001 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4005 qdisc_skb_cb(skb)->pkt_len = skb->len;
4006 tcx_set_ingress(skb, true);
4008 if (static_branch_unlikely(&tcx_needed_key)) {
4009 sch_ret = tcx_run(entry, skb, true);
4010 if (sch_ret != TC_ACT_UNSPEC)
4011 goto ingress_verdict;
4013 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason);
4016 case TC_ACT_REDIRECT:
4017 /* skb_mac_header check was done by BPF, so we can safely
4018 * push the L2 header back before redirecting to another
4021 __skb_push(skb, skb->mac_len);
4022 if (skb_do_redirect(skb) == -EAGAIN) {
4023 __skb_pull(skb, skb->mac_len);
4027 *ret = NET_RX_SUCCESS;
4030 kfree_skb_reason(skb, drop_reason);
4033 /* used by tc_run */
4039 case TC_ACT_CONSUMED:
4040 *ret = NET_RX_SUCCESS;
4047 static __always_inline struct sk_buff *
4048 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
4050 struct bpf_mprog_entry *entry = rcu_dereference_bh(dev->tcx_egress);
4051 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_EGRESS;
4057 /* qdisc_skb_cb(skb)->pkt_len & tcx_set_ingress() was
4058 * already set by the caller.
4060 if (static_branch_unlikely(&tcx_needed_key)) {
4061 sch_ret = tcx_run(entry, skb, false);
4062 if (sch_ret != TC_ACT_UNSPEC)
4063 goto egress_verdict;
4065 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason);
4068 case TC_ACT_REDIRECT:
4069 /* No need to push/pop skb's mac_header here on egress! */
4070 skb_do_redirect(skb);
4071 *ret = NET_XMIT_SUCCESS;
4074 kfree_skb_reason(skb, drop_reason);
4075 *ret = NET_XMIT_DROP;
4077 /* used by tc_run */
4083 case TC_ACT_CONSUMED:
4084 *ret = NET_XMIT_SUCCESS;
4091 static __always_inline struct sk_buff *
4092 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
4093 struct net_device *orig_dev, bool *another)
4098 static __always_inline struct sk_buff *
4099 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
4103 #endif /* CONFIG_NET_XGRESS */
4106 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb,
4107 struct xps_dev_maps *dev_maps, unsigned int tci)
4109 int tc = netdev_get_prio_tc_map(dev, skb->priority);
4110 struct xps_map *map;
4111 int queue_index = -1;
4113 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids)
4116 tci *= dev_maps->num_tc;
4119 map = rcu_dereference(dev_maps->attr_map[tci]);
4122 queue_index = map->queues[0];
4124 queue_index = map->queues[reciprocal_scale(
4125 skb_get_hash(skb), map->len)];
4126 if (unlikely(queue_index >= dev->real_num_tx_queues))
4133 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev,
4134 struct sk_buff *skb)
4137 struct xps_dev_maps *dev_maps;
4138 struct sock *sk = skb->sk;
4139 int queue_index = -1;
4141 if (!static_key_false(&xps_needed))
4145 if (!static_key_false(&xps_rxqs_needed))
4148 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]);
4150 int tci = sk_rx_queue_get(sk);
4153 queue_index = __get_xps_queue_idx(dev, skb, dev_maps,
4158 if (queue_index < 0) {
4159 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]);
4161 unsigned int tci = skb->sender_cpu - 1;
4163 queue_index = __get_xps_queue_idx(dev, skb, dev_maps,
4175 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb,
4176 struct net_device *sb_dev)
4180 EXPORT_SYMBOL(dev_pick_tx_zero);
4182 u16 dev_pick_tx_cpu_id(struct net_device *dev, struct sk_buff *skb,
4183 struct net_device *sb_dev)
4185 return (u16)raw_smp_processor_id() % dev->real_num_tx_queues;
4187 EXPORT_SYMBOL(dev_pick_tx_cpu_id);
4189 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb,
4190 struct net_device *sb_dev)
4192 struct sock *sk = skb->sk;
4193 int queue_index = sk_tx_queue_get(sk);
4195 sb_dev = sb_dev ? : dev;
4197 if (queue_index < 0 || skb->ooo_okay ||
4198 queue_index >= dev->real_num_tx_queues) {
4199 int new_index = get_xps_queue(dev, sb_dev, skb);
4202 new_index = skb_tx_hash(dev, sb_dev, skb);
4204 if (queue_index != new_index && sk &&
4206 rcu_access_pointer(sk->sk_dst_cache))
4207 sk_tx_queue_set(sk, new_index);
4209 queue_index = new_index;
4214 EXPORT_SYMBOL(netdev_pick_tx);
4216 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev,
4217 struct sk_buff *skb,
4218 struct net_device *sb_dev)
4220 int queue_index = 0;
4223 u32 sender_cpu = skb->sender_cpu - 1;
4225 if (sender_cpu >= (u32)NR_CPUS)
4226 skb->sender_cpu = raw_smp_processor_id() + 1;
4229 if (dev->real_num_tx_queues != 1) {
4230 const struct net_device_ops *ops = dev->netdev_ops;
4232 if (ops->ndo_select_queue)
4233 queue_index = ops->ndo_select_queue(dev, skb, sb_dev);
4235 queue_index = netdev_pick_tx(dev, skb, sb_dev);
4237 queue_index = netdev_cap_txqueue(dev, queue_index);
4240 skb_set_queue_mapping(skb, queue_index);
4241 return netdev_get_tx_queue(dev, queue_index);
4245 * __dev_queue_xmit() - transmit a buffer
4246 * @skb: buffer to transmit
4247 * @sb_dev: suboordinate device used for L2 forwarding offload
4249 * Queue a buffer for transmission to a network device. The caller must
4250 * have set the device and priority and built the buffer before calling
4251 * this function. The function can be called from an interrupt.
4253 * When calling this method, interrupts MUST be enabled. This is because
4254 * the BH enable code must have IRQs enabled so that it will not deadlock.
4256 * Regardless of the return value, the skb is consumed, so it is currently
4257 * difficult to retry a send to this method. (You can bump the ref count
4258 * before sending to hold a reference for retry if you are careful.)
4261 * * 0 - buffer successfully transmitted
4262 * * positive qdisc return code - NET_XMIT_DROP etc.
4263 * * negative errno - other errors
4265 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev)
4267 struct net_device *dev = skb->dev;
4268 struct netdev_queue *txq = NULL;
4273 skb_reset_mac_header(skb);
4274 skb_assert_len(skb);
4276 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP))
4277 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED);
4279 /* Disable soft irqs for various locks below. Also
4280 * stops preemption for RCU.
4284 skb_update_prio(skb);
4286 qdisc_pkt_len_init(skb);
4287 tcx_set_ingress(skb, false);
4288 #ifdef CONFIG_NET_EGRESS
4289 if (static_branch_unlikely(&egress_needed_key)) {
4290 if (nf_hook_egress_active()) {
4291 skb = nf_hook_egress(skb, &rc, dev);
4296 netdev_xmit_skip_txqueue(false);
4298 nf_skip_egress(skb, true);
4299 skb = sch_handle_egress(skb, &rc, dev);
4302 nf_skip_egress(skb, false);
4304 if (netdev_xmit_txqueue_skipped())
4305 txq = netdev_tx_queue_mapping(dev, skb);
4308 /* If device/qdisc don't need skb->dst, release it right now while
4309 * its hot in this cpu cache.
4311 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
4317 txq = netdev_core_pick_tx(dev, skb, sb_dev);
4319 q = rcu_dereference_bh(txq->qdisc);
4321 trace_net_dev_queue(skb);
4323 rc = __dev_xmit_skb(skb, q, dev, txq);
4327 /* The device has no queue. Common case for software devices:
4328 * loopback, all the sorts of tunnels...
4330 * Really, it is unlikely that netif_tx_lock protection is necessary
4331 * here. (f.e. loopback and IP tunnels are clean ignoring statistics
4333 * However, it is possible, that they rely on protection
4336 * Check this and shot the lock. It is not prone from deadlocks.
4337 *Either shot noqueue qdisc, it is even simpler 8)
4339 if (dev->flags & IFF_UP) {
4340 int cpu = smp_processor_id(); /* ok because BHs are off */
4342 /* Other cpus might concurrently change txq->xmit_lock_owner
4343 * to -1 or to their cpu id, but not to our id.
4345 if (READ_ONCE(txq->xmit_lock_owner) != cpu) {
4346 if (dev_xmit_recursion())
4347 goto recursion_alert;
4349 skb = validate_xmit_skb(skb, dev, &again);
4353 HARD_TX_LOCK(dev, txq, cpu);
4355 if (!netif_xmit_stopped(txq)) {
4356 dev_xmit_recursion_inc();
4357 skb = dev_hard_start_xmit(skb, dev, txq, &rc);
4358 dev_xmit_recursion_dec();
4359 if (dev_xmit_complete(rc)) {
4360 HARD_TX_UNLOCK(dev, txq);
4364 HARD_TX_UNLOCK(dev, txq);
4365 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
4368 /* Recursion is detected! It is possible,
4372 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
4378 rcu_read_unlock_bh();
4380 dev_core_stats_tx_dropped_inc(dev);
4381 kfree_skb_list(skb);
4384 rcu_read_unlock_bh();
4387 EXPORT_SYMBOL(__dev_queue_xmit);
4389 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id)
4391 struct net_device *dev = skb->dev;
4392 struct sk_buff *orig_skb = skb;
4393 struct netdev_queue *txq;
4394 int ret = NETDEV_TX_BUSY;
4397 if (unlikely(!netif_running(dev) ||
4398 !netif_carrier_ok(dev)))
4401 skb = validate_xmit_skb_list(skb, dev, &again);
4402 if (skb != orig_skb)
4405 skb_set_queue_mapping(skb, queue_id);
4406 txq = skb_get_tx_queue(dev, skb);
4410 dev_xmit_recursion_inc();
4411 HARD_TX_LOCK(dev, txq, smp_processor_id());
4412 if (!netif_xmit_frozen_or_drv_stopped(txq))
4413 ret = netdev_start_xmit(skb, dev, txq, false);
4414 HARD_TX_UNLOCK(dev, txq);
4415 dev_xmit_recursion_dec();
4420 dev_core_stats_tx_dropped_inc(dev);
4421 kfree_skb_list(skb);
4422 return NET_XMIT_DROP;
4424 EXPORT_SYMBOL(__dev_direct_xmit);
4426 /*************************************************************************
4428 *************************************************************************/
4430 int netdev_max_backlog __read_mostly = 1000;
4431 EXPORT_SYMBOL(netdev_max_backlog);
4433 int netdev_tstamp_prequeue __read_mostly = 1;
4434 unsigned int sysctl_skb_defer_max __read_mostly = 64;
4435 int netdev_budget __read_mostly = 300;
4436 /* Must be at least 2 jiffes to guarantee 1 jiffy timeout */
4437 unsigned int __read_mostly netdev_budget_usecs = 2 * USEC_PER_SEC / HZ;
4438 int weight_p __read_mostly = 64; /* old backlog weight */
4439 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */
4440 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */
4441 int dev_rx_weight __read_mostly = 64;
4442 int dev_tx_weight __read_mostly = 64;
4444 /* Called with irq disabled */
4445 static inline void ____napi_schedule(struct softnet_data *sd,
4446 struct napi_struct *napi)
4448 struct task_struct *thread;
4450 lockdep_assert_irqs_disabled();
4452 if (test_bit(NAPI_STATE_THREADED, &napi->state)) {
4453 /* Paired with smp_mb__before_atomic() in
4454 * napi_enable()/dev_set_threaded().
4455 * Use READ_ONCE() to guarantee a complete
4456 * read on napi->thread. Only call
4457 * wake_up_process() when it's not NULL.
4459 thread = READ_ONCE(napi->thread);
4461 /* Avoid doing set_bit() if the thread is in
4462 * INTERRUPTIBLE state, cause napi_thread_wait()
4463 * makes sure to proceed with napi polling
4464 * if the thread is explicitly woken from here.
4466 if (READ_ONCE(thread->__state) != TASK_INTERRUPTIBLE)
4467 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state);
4468 wake_up_process(thread);
4473 list_add_tail(&napi->poll_list, &sd->poll_list);
4474 WRITE_ONCE(napi->list_owner, smp_processor_id());
4475 /* If not called from net_rx_action()
4476 * we have to raise NET_RX_SOFTIRQ.
4478 if (!sd->in_net_rx_action)
4479 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
4484 /* One global table that all flow-based protocols share. */
4485 struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
4486 EXPORT_SYMBOL(rps_sock_flow_table);
4487 u32 rps_cpu_mask __read_mostly;
4488 EXPORT_SYMBOL(rps_cpu_mask);
4490 struct static_key_false rps_needed __read_mostly;
4491 EXPORT_SYMBOL(rps_needed);
4492 struct static_key_false rfs_needed __read_mostly;
4493 EXPORT_SYMBOL(rfs_needed);
4495 static struct rps_dev_flow *
4496 set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
4497 struct rps_dev_flow *rflow, u16 next_cpu)
4499 if (next_cpu < nr_cpu_ids) {
4500 #ifdef CONFIG_RFS_ACCEL
4501 struct netdev_rx_queue *rxqueue;
4502 struct rps_dev_flow_table *flow_table;
4503 struct rps_dev_flow *old_rflow;
4508 /* Should we steer this flow to a different hardware queue? */
4509 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
4510 !(dev->features & NETIF_F_NTUPLE))
4512 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
4513 if (rxq_index == skb_get_rx_queue(skb))
4516 rxqueue = dev->_rx + rxq_index;
4517 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4520 flow_id = skb_get_hash(skb) & flow_table->mask;
4521 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
4522 rxq_index, flow_id);
4526 rflow = &flow_table->flows[flow_id];
4528 if (old_rflow->filter == rflow->filter)
4529 old_rflow->filter = RPS_NO_FILTER;
4533 per_cpu(softnet_data, next_cpu).input_queue_head;
4536 rflow->cpu = next_cpu;
4541 * get_rps_cpu is called from netif_receive_skb and returns the target
4542 * CPU from the RPS map of the receiving queue for a given skb.
4543 * rcu_read_lock must be held on entry.
4545 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
4546 struct rps_dev_flow **rflowp)
4548 const struct rps_sock_flow_table *sock_flow_table;
4549 struct netdev_rx_queue *rxqueue = dev->_rx;
4550 struct rps_dev_flow_table *flow_table;
4551 struct rps_map *map;
4556 if (skb_rx_queue_recorded(skb)) {
4557 u16 index = skb_get_rx_queue(skb);
4559 if (unlikely(index >= dev->real_num_rx_queues)) {
4560 WARN_ONCE(dev->real_num_rx_queues > 1,
4561 "%s received packet on queue %u, but number "
4562 "of RX queues is %u\n",
4563 dev->name, index, dev->real_num_rx_queues);
4569 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */
4571 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4572 map = rcu_dereference(rxqueue->rps_map);
4573 if (!flow_table && !map)
4576 skb_reset_network_header(skb);
4577 hash = skb_get_hash(skb);
4581 sock_flow_table = rcu_dereference(rps_sock_flow_table);
4582 if (flow_table && sock_flow_table) {
4583 struct rps_dev_flow *rflow;
4587 /* First check into global flow table if there is a match.
4588 * This READ_ONCE() pairs with WRITE_ONCE() from rps_record_sock_flow().
4590 ident = READ_ONCE(sock_flow_table->ents[hash & sock_flow_table->mask]);
4591 if ((ident ^ hash) & ~rps_cpu_mask)
4594 next_cpu = ident & rps_cpu_mask;
4596 /* OK, now we know there is a match,
4597 * we can look at the local (per receive queue) flow table
4599 rflow = &flow_table->flows[hash & flow_table->mask];
4603 * If the desired CPU (where last recvmsg was done) is
4604 * different from current CPU (one in the rx-queue flow
4605 * table entry), switch if one of the following holds:
4606 * - Current CPU is unset (>= nr_cpu_ids).
4607 * - Current CPU is offline.
4608 * - The current CPU's queue tail has advanced beyond the
4609 * last packet that was enqueued using this table entry.
4610 * This guarantees that all previous packets for the flow
4611 * have been dequeued, thus preserving in order delivery.
4613 if (unlikely(tcpu != next_cpu) &&
4614 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) ||
4615 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
4616 rflow->last_qtail)) >= 0)) {
4618 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
4621 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) {
4631 tcpu = map->cpus[reciprocal_scale(hash, map->len)];
4632 if (cpu_online(tcpu)) {
4642 #ifdef CONFIG_RFS_ACCEL
4645 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
4646 * @dev: Device on which the filter was set
4647 * @rxq_index: RX queue index
4648 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
4649 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
4651 * Drivers that implement ndo_rx_flow_steer() should periodically call
4652 * this function for each installed filter and remove the filters for
4653 * which it returns %true.
4655 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
4656 u32 flow_id, u16 filter_id)
4658 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
4659 struct rps_dev_flow_table *flow_table;
4660 struct rps_dev_flow *rflow;
4665 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4666 if (flow_table && flow_id <= flow_table->mask) {
4667 rflow = &flow_table->flows[flow_id];
4668 cpu = READ_ONCE(rflow->cpu);
4669 if (rflow->filter == filter_id && cpu < nr_cpu_ids &&
4670 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
4671 rflow->last_qtail) <
4672 (int)(10 * flow_table->mask)))
4678 EXPORT_SYMBOL(rps_may_expire_flow);
4680 #endif /* CONFIG_RFS_ACCEL */
4682 /* Called from hardirq (IPI) context */
4683 static void rps_trigger_softirq(void *data)
4685 struct softnet_data *sd = data;
4687 ____napi_schedule(sd, &sd->backlog);
4691 #endif /* CONFIG_RPS */
4693 /* Called from hardirq (IPI) context */
4694 static void trigger_rx_softirq(void *data)
4696 struct softnet_data *sd = data;
4698 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
4699 smp_store_release(&sd->defer_ipi_scheduled, 0);
4703 * After we queued a packet into sd->input_pkt_queue,
4704 * we need to make sure this queue is serviced soon.
4706 * - If this is another cpu queue, link it to our rps_ipi_list,
4707 * and make sure we will process rps_ipi_list from net_rx_action().
4709 * - If this is our own queue, NAPI schedule our backlog.
4710 * Note that this also raises NET_RX_SOFTIRQ.
4712 static void napi_schedule_rps(struct softnet_data *sd)
4714 struct softnet_data *mysd = this_cpu_ptr(&softnet_data);
4718 sd->rps_ipi_next = mysd->rps_ipi_list;
4719 mysd->rps_ipi_list = sd;
4721 /* If not called from net_rx_action() or napi_threaded_poll()
4722 * we have to raise NET_RX_SOFTIRQ.
4724 if (!mysd->in_net_rx_action && !mysd->in_napi_threaded_poll)
4725 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
4728 #endif /* CONFIG_RPS */
4729 __napi_schedule_irqoff(&mysd->backlog);
4732 #ifdef CONFIG_NET_FLOW_LIMIT
4733 int netdev_flow_limit_table_len __read_mostly = (1 << 12);
4736 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
4738 #ifdef CONFIG_NET_FLOW_LIMIT
4739 struct sd_flow_limit *fl;
4740 struct softnet_data *sd;
4741 unsigned int old_flow, new_flow;
4743 if (qlen < (READ_ONCE(netdev_max_backlog) >> 1))
4746 sd = this_cpu_ptr(&softnet_data);
4749 fl = rcu_dereference(sd->flow_limit);
4751 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1);
4752 old_flow = fl->history[fl->history_head];
4753 fl->history[fl->history_head] = new_flow;
4756 fl->history_head &= FLOW_LIMIT_HISTORY - 1;
4758 if (likely(fl->buckets[old_flow]))
4759 fl->buckets[old_flow]--;
4761 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) {
4773 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
4774 * queue (may be a remote CPU queue).
4776 static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
4777 unsigned int *qtail)
4779 enum skb_drop_reason reason;
4780 struct softnet_data *sd;
4781 unsigned long flags;
4784 reason = SKB_DROP_REASON_NOT_SPECIFIED;
4785 sd = &per_cpu(softnet_data, cpu);
4787 rps_lock_irqsave(sd, &flags);
4788 if (!netif_running(skb->dev))
4790 qlen = skb_queue_len(&sd->input_pkt_queue);
4791 if (qlen <= READ_ONCE(netdev_max_backlog) && !skb_flow_limit(skb, qlen)) {
4794 __skb_queue_tail(&sd->input_pkt_queue, skb);
4795 input_queue_tail_incr_save(sd, qtail);
4796 rps_unlock_irq_restore(sd, &flags);
4797 return NET_RX_SUCCESS;
4800 /* Schedule NAPI for backlog device
4801 * We can use non atomic operation since we own the queue lock
4803 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state))
4804 napi_schedule_rps(sd);
4807 reason = SKB_DROP_REASON_CPU_BACKLOG;
4811 rps_unlock_irq_restore(sd, &flags);
4813 dev_core_stats_rx_dropped_inc(skb->dev);
4814 kfree_skb_reason(skb, reason);
4818 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb)
4820 struct net_device *dev = skb->dev;
4821 struct netdev_rx_queue *rxqueue;
4825 if (skb_rx_queue_recorded(skb)) {
4826 u16 index = skb_get_rx_queue(skb);
4828 if (unlikely(index >= dev->real_num_rx_queues)) {
4829 WARN_ONCE(dev->real_num_rx_queues > 1,
4830 "%s received packet on queue %u, but number "
4831 "of RX queues is %u\n",
4832 dev->name, index, dev->real_num_rx_queues);
4834 return rxqueue; /* Return first rxqueue */
4841 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp,
4842 struct bpf_prog *xdp_prog)
4844 void *orig_data, *orig_data_end, *hard_start;
4845 struct netdev_rx_queue *rxqueue;
4846 bool orig_bcast, orig_host;
4847 u32 mac_len, frame_sz;
4848 __be16 orig_eth_type;
4853 /* The XDP program wants to see the packet starting at the MAC
4856 mac_len = skb->data - skb_mac_header(skb);
4857 hard_start = skb->data - skb_headroom(skb);
4859 /* SKB "head" area always have tailroom for skb_shared_info */
4860 frame_sz = (void *)skb_end_pointer(skb) - hard_start;
4861 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
4863 rxqueue = netif_get_rxqueue(skb);
4864 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq);
4865 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len,
4866 skb_headlen(skb) + mac_len, true);
4868 orig_data_end = xdp->data_end;
4869 orig_data = xdp->data;
4870 eth = (struct ethhdr *)xdp->data;
4871 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr);
4872 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest);
4873 orig_eth_type = eth->h_proto;
4875 act = bpf_prog_run_xdp(xdp_prog, xdp);
4877 /* check if bpf_xdp_adjust_head was used */
4878 off = xdp->data - orig_data;
4881 __skb_pull(skb, off);
4883 __skb_push(skb, -off);
4885 skb->mac_header += off;
4886 skb_reset_network_header(skb);
4889 /* check if bpf_xdp_adjust_tail was used */
4890 off = xdp->data_end - orig_data_end;
4892 skb_set_tail_pointer(skb, xdp->data_end - xdp->data);
4893 skb->len += off; /* positive on grow, negative on shrink */
4896 /* check if XDP changed eth hdr such SKB needs update */
4897 eth = (struct ethhdr *)xdp->data;
4898 if ((orig_eth_type != eth->h_proto) ||
4899 (orig_host != ether_addr_equal_64bits(eth->h_dest,
4900 skb->dev->dev_addr)) ||
4901 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) {
4902 __skb_push(skb, ETH_HLEN);
4903 skb->pkt_type = PACKET_HOST;
4904 skb->protocol = eth_type_trans(skb, skb->dev);
4907 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull
4908 * before calling us again on redirect path. We do not call do_redirect
4909 * as we leave that up to the caller.
4911 * Caller is responsible for managing lifetime of skb (i.e. calling
4912 * kfree_skb in response to actions it cannot handle/XDP_DROP).
4917 __skb_push(skb, mac_len);
4920 metalen = xdp->data - xdp->data_meta;
4922 skb_metadata_set(skb, metalen);
4929 static u32 netif_receive_generic_xdp(struct sk_buff *skb,
4930 struct xdp_buff *xdp,
4931 struct bpf_prog *xdp_prog)
4935 /* Reinjected packets coming from act_mirred or similar should
4936 * not get XDP generic processing.
4938 if (skb_is_redirected(skb))
4941 /* XDP packets must be linear and must have sufficient headroom
4942 * of XDP_PACKET_HEADROOM bytes. This is the guarantee that also
4943 * native XDP provides, thus we need to do it here as well.
4945 if (skb_cloned(skb) || skb_is_nonlinear(skb) ||
4946 skb_headroom(skb) < XDP_PACKET_HEADROOM) {
4947 int hroom = XDP_PACKET_HEADROOM - skb_headroom(skb);
4948 int troom = skb->tail + skb->data_len - skb->end;
4950 /* In case we have to go down the path and also linearize,
4951 * then lets do the pskb_expand_head() work just once here.
4953 if (pskb_expand_head(skb,
4954 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0,
4955 troom > 0 ? troom + 128 : 0, GFP_ATOMIC))
4957 if (skb_linearize(skb))
4961 act = bpf_prog_run_generic_xdp(skb, xdp, xdp_prog);
4968 bpf_warn_invalid_xdp_action(skb->dev, xdp_prog, act);
4971 trace_xdp_exception(skb->dev, xdp_prog, act);
4982 /* When doing generic XDP we have to bypass the qdisc layer and the
4983 * network taps in order to match in-driver-XDP behavior. This also means
4984 * that XDP packets are able to starve other packets going through a qdisc,
4985 * and DDOS attacks will be more effective. In-driver-XDP use dedicated TX
4986 * queues, so they do not have this starvation issue.
4988 void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog)
4990 struct net_device *dev = skb->dev;
4991 struct netdev_queue *txq;
4992 bool free_skb = true;
4995 txq = netdev_core_pick_tx(dev, skb, NULL);
4996 cpu = smp_processor_id();
4997 HARD_TX_LOCK(dev, txq, cpu);
4998 if (!netif_xmit_frozen_or_drv_stopped(txq)) {
4999 rc = netdev_start_xmit(skb, dev, txq, 0);
5000 if (dev_xmit_complete(rc))
5003 HARD_TX_UNLOCK(dev, txq);
5005 trace_xdp_exception(dev, xdp_prog, XDP_TX);
5006 dev_core_stats_tx_dropped_inc(dev);
5011 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key);
5013 int do_xdp_generic(struct bpf_prog *xdp_prog, struct sk_buff *skb)
5016 struct xdp_buff xdp;
5020 act = netif_receive_generic_xdp(skb, &xdp, xdp_prog);
5021 if (act != XDP_PASS) {
5024 err = xdp_do_generic_redirect(skb->dev, skb,
5030 generic_xdp_tx(skb, xdp_prog);
5038 kfree_skb_reason(skb, SKB_DROP_REASON_XDP);
5041 EXPORT_SYMBOL_GPL(do_xdp_generic);
5043 static int netif_rx_internal(struct sk_buff *skb)
5047 net_timestamp_check(READ_ONCE(netdev_tstamp_prequeue), skb);
5049 trace_netif_rx(skb);
5052 if (static_branch_unlikely(&rps_needed)) {
5053 struct rps_dev_flow voidflow, *rflow = &voidflow;
5058 cpu = get_rps_cpu(skb->dev, skb, &rflow);
5060 cpu = smp_processor_id();
5062 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
5070 ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail);
5076 * __netif_rx - Slightly optimized version of netif_rx
5077 * @skb: buffer to post
5079 * This behaves as netif_rx except that it does not disable bottom halves.
5080 * As a result this function may only be invoked from the interrupt context
5081 * (either hard or soft interrupt).
5083 int __netif_rx(struct sk_buff *skb)
5087 lockdep_assert_once(hardirq_count() | softirq_count());
5089 trace_netif_rx_entry(skb);
5090 ret = netif_rx_internal(skb);
5091 trace_netif_rx_exit(ret);
5094 EXPORT_SYMBOL(__netif_rx);
5097 * netif_rx - post buffer to the network code
5098 * @skb: buffer to post
5100 * This function receives a packet from a device driver and queues it for
5101 * the upper (protocol) levels to process via the backlog NAPI device. It
5102 * always succeeds. The buffer may be dropped during processing for
5103 * congestion control or by the protocol layers.
5104 * The network buffer is passed via the backlog NAPI device. Modern NIC
5105 * driver should use NAPI and GRO.
5106 * This function can used from interrupt and from process context. The
5107 * caller from process context must not disable interrupts before invoking
5111 * NET_RX_SUCCESS (no congestion)
5112 * NET_RX_DROP (packet was dropped)
5115 int netif_rx(struct sk_buff *skb)
5117 bool need_bh_off = !(hardirq_count() | softirq_count());
5122 trace_netif_rx_entry(skb);
5123 ret = netif_rx_internal(skb);
5124 trace_netif_rx_exit(ret);
5129 EXPORT_SYMBOL(netif_rx);
5131 static __latent_entropy void net_tx_action(struct softirq_action *h)
5133 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
5135 if (sd->completion_queue) {
5136 struct sk_buff *clist;
5138 local_irq_disable();
5139 clist = sd->completion_queue;
5140 sd->completion_queue = NULL;
5144 struct sk_buff *skb = clist;
5146 clist = clist->next;
5148 WARN_ON(refcount_read(&skb->users));
5149 if (likely(get_kfree_skb_cb(skb)->reason == SKB_CONSUMED))
5150 trace_consume_skb(skb, net_tx_action);
5152 trace_kfree_skb(skb, net_tx_action,
5153 get_kfree_skb_cb(skb)->reason);
5155 if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
5158 __napi_kfree_skb(skb,
5159 get_kfree_skb_cb(skb)->reason);
5163 if (sd->output_queue) {
5166 local_irq_disable();
5167 head = sd->output_queue;
5168 sd->output_queue = NULL;
5169 sd->output_queue_tailp = &sd->output_queue;
5175 struct Qdisc *q = head;
5176 spinlock_t *root_lock = NULL;
5178 head = head->next_sched;
5180 /* We need to make sure head->next_sched is read
5181 * before clearing __QDISC_STATE_SCHED
5183 smp_mb__before_atomic();
5185 if (!(q->flags & TCQ_F_NOLOCK)) {
5186 root_lock = qdisc_lock(q);
5187 spin_lock(root_lock);
5188 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED,
5190 /* There is a synchronize_net() between
5191 * STATE_DEACTIVATED flag being set and
5192 * qdisc_reset()/some_qdisc_is_busy() in
5193 * dev_deactivate(), so we can safely bail out
5194 * early here to avoid data race between
5195 * qdisc_deactivate() and some_qdisc_is_busy()
5196 * for lockless qdisc.
5198 clear_bit(__QDISC_STATE_SCHED, &q->state);
5202 clear_bit(__QDISC_STATE_SCHED, &q->state);
5205 spin_unlock(root_lock);
5211 xfrm_dev_backlog(sd);
5214 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE)
5215 /* This hook is defined here for ATM LANE */
5216 int (*br_fdb_test_addr_hook)(struct net_device *dev,
5217 unsigned char *addr) __read_mostly;
5218 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook);
5222 * netdev_is_rx_handler_busy - check if receive handler is registered
5223 * @dev: device to check
5225 * Check if a receive handler is already registered for a given device.
5226 * Return true if there one.
5228 * The caller must hold the rtnl_mutex.
5230 bool netdev_is_rx_handler_busy(struct net_device *dev)
5233 return dev && rtnl_dereference(dev->rx_handler);
5235 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy);
5238 * netdev_rx_handler_register - register receive handler
5239 * @dev: device to register a handler for
5240 * @rx_handler: receive handler to register
5241 * @rx_handler_data: data pointer that is used by rx handler
5243 * Register a receive handler for a device. This handler will then be
5244 * called from __netif_receive_skb. A negative errno code is returned
5247 * The caller must hold the rtnl_mutex.
5249 * For a general description of rx_handler, see enum rx_handler_result.
5251 int netdev_rx_handler_register(struct net_device *dev,
5252 rx_handler_func_t *rx_handler,
5253 void *rx_handler_data)
5255 if (netdev_is_rx_handler_busy(dev))
5258 if (dev->priv_flags & IFF_NO_RX_HANDLER)
5261 /* Note: rx_handler_data must be set before rx_handler */
5262 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data);
5263 rcu_assign_pointer(dev->rx_handler, rx_handler);
5267 EXPORT_SYMBOL_GPL(netdev_rx_handler_register);
5270 * netdev_rx_handler_unregister - unregister receive handler
5271 * @dev: device to unregister a handler from
5273 * Unregister a receive handler from a device.
5275 * The caller must hold the rtnl_mutex.
5277 void netdev_rx_handler_unregister(struct net_device *dev)
5281 RCU_INIT_POINTER(dev->rx_handler, NULL);
5282 /* a reader seeing a non NULL rx_handler in a rcu_read_lock()
5283 * section has a guarantee to see a non NULL rx_handler_data
5287 RCU_INIT_POINTER(dev->rx_handler_data, NULL);
5289 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister);
5292 * Limit the use of PFMEMALLOC reserves to those protocols that implement
5293 * the special handling of PFMEMALLOC skbs.
5295 static bool skb_pfmemalloc_protocol(struct sk_buff *skb)
5297 switch (skb->protocol) {
5298 case htons(ETH_P_ARP):
5299 case htons(ETH_P_IP):
5300 case htons(ETH_P_IPV6):
5301 case htons(ETH_P_8021Q):
5302 case htons(ETH_P_8021AD):
5309 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev,
5310 int *ret, struct net_device *orig_dev)
5312 if (nf_hook_ingress_active(skb)) {
5316 *ret = deliver_skb(skb, *pt_prev, orig_dev);
5321 ingress_retval = nf_hook_ingress(skb);
5323 return ingress_retval;
5328 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc,
5329 struct packet_type **ppt_prev)
5331 struct packet_type *ptype, *pt_prev;
5332 rx_handler_func_t *rx_handler;
5333 struct sk_buff *skb = *pskb;
5334 struct net_device *orig_dev;
5335 bool deliver_exact = false;
5336 int ret = NET_RX_DROP;
5339 net_timestamp_check(!READ_ONCE(netdev_tstamp_prequeue), skb);
5341 trace_netif_receive_skb(skb);
5343 orig_dev = skb->dev;
5345 skb_reset_network_header(skb);
5346 if (!skb_transport_header_was_set(skb))
5347 skb_reset_transport_header(skb);
5348 skb_reset_mac_len(skb);
5353 skb->skb_iif = skb->dev->ifindex;
5355 __this_cpu_inc(softnet_data.processed);
5357 if (static_branch_unlikely(&generic_xdp_needed_key)) {
5361 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
5364 if (ret2 != XDP_PASS) {
5370 if (eth_type_vlan(skb->protocol)) {
5371 skb = skb_vlan_untag(skb);
5376 if (skb_skip_tc_classify(skb))
5382 list_for_each_entry_rcu(ptype, &ptype_all, list) {
5384 ret = deliver_skb(skb, pt_prev, orig_dev);
5388 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) {
5390 ret = deliver_skb(skb, pt_prev, orig_dev);
5395 #ifdef CONFIG_NET_INGRESS
5396 if (static_branch_unlikely(&ingress_needed_key)) {
5397 bool another = false;
5399 nf_skip_egress(skb, true);
5400 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev,
5407 nf_skip_egress(skb, false);
5408 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0)
5412 skb_reset_redirect(skb);
5414 if (pfmemalloc && !skb_pfmemalloc_protocol(skb))
5417 if (skb_vlan_tag_present(skb)) {
5419 ret = deliver_skb(skb, pt_prev, orig_dev);
5422 if (vlan_do_receive(&skb))
5424 else if (unlikely(!skb))
5428 rx_handler = rcu_dereference(skb->dev->rx_handler);
5431 ret = deliver_skb(skb, pt_prev, orig_dev);
5434 switch (rx_handler(&skb)) {
5435 case RX_HANDLER_CONSUMED:
5436 ret = NET_RX_SUCCESS;
5438 case RX_HANDLER_ANOTHER:
5440 case RX_HANDLER_EXACT:
5441 deliver_exact = true;
5443 case RX_HANDLER_PASS:
5450 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) {
5452 if (skb_vlan_tag_get_id(skb)) {
5453 /* Vlan id is non 0 and vlan_do_receive() above couldn't
5456 skb->pkt_type = PACKET_OTHERHOST;
5457 } else if (eth_type_vlan(skb->protocol)) {
5458 /* Outer header is 802.1P with vlan 0, inner header is
5459 * 802.1Q or 802.1AD and vlan_do_receive() above could
5460 * not find vlan dev for vlan id 0.
5462 __vlan_hwaccel_clear_tag(skb);
5463 skb = skb_vlan_untag(skb);
5466 if (vlan_do_receive(&skb))
5467 /* After stripping off 802.1P header with vlan 0
5468 * vlan dev is found for inner header.
5471 else if (unlikely(!skb))
5474 /* We have stripped outer 802.1P vlan 0 header.
5475 * But could not find vlan dev.
5476 * check again for vlan id to set OTHERHOST.
5480 /* Note: we might in the future use prio bits
5481 * and set skb->priority like in vlan_do_receive()
5482 * For the time being, just ignore Priority Code Point
5484 __vlan_hwaccel_clear_tag(skb);
5487 type = skb->protocol;
5489 /* deliver only exact match when indicated */
5490 if (likely(!deliver_exact)) {
5491 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
5492 &ptype_base[ntohs(type) &
5496 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
5497 &orig_dev->ptype_specific);
5499 if (unlikely(skb->dev != orig_dev)) {
5500 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
5501 &skb->dev->ptype_specific);
5505 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
5507 *ppt_prev = pt_prev;
5511 dev_core_stats_rx_dropped_inc(skb->dev);
5513 dev_core_stats_rx_nohandler_inc(skb->dev);
5514 kfree_skb_reason(skb, SKB_DROP_REASON_UNHANDLED_PROTO);
5515 /* Jamal, now you will not able to escape explaining
5516 * me how you were going to use this. :-)
5522 /* The invariant here is that if *ppt_prev is not NULL
5523 * then skb should also be non-NULL.
5525 * Apparently *ppt_prev assignment above holds this invariant due to
5526 * skb dereferencing near it.
5532 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc)
5534 struct net_device *orig_dev = skb->dev;
5535 struct packet_type *pt_prev = NULL;
5538 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev);
5540 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb,
5541 skb->dev, pt_prev, orig_dev);
5546 * netif_receive_skb_core - special purpose version of netif_receive_skb
5547 * @skb: buffer to process
5549 * More direct receive version of netif_receive_skb(). It should
5550 * only be used by callers that have a need to skip RPS and Generic XDP.
5551 * Caller must also take care of handling if ``(page_is_)pfmemalloc``.
5553 * This function may only be called from softirq context and interrupts
5554 * should be enabled.
5556 * Return values (usually ignored):
5557 * NET_RX_SUCCESS: no congestion
5558 * NET_RX_DROP: packet was dropped
5560 int netif_receive_skb_core(struct sk_buff *skb)
5565 ret = __netif_receive_skb_one_core(skb, false);
5570 EXPORT_SYMBOL(netif_receive_skb_core);
5572 static inline void __netif_receive_skb_list_ptype(struct list_head *head,
5573 struct packet_type *pt_prev,
5574 struct net_device *orig_dev)
5576 struct sk_buff *skb, *next;
5580 if (list_empty(head))
5582 if (pt_prev->list_func != NULL)
5583 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv,
5584 ip_list_rcv, head, pt_prev, orig_dev);
5586 list_for_each_entry_safe(skb, next, head, list) {
5587 skb_list_del_init(skb);
5588 pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
5592 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc)
5594 /* Fast-path assumptions:
5595 * - There is no RX handler.
5596 * - Only one packet_type matches.
5597 * If either of these fails, we will end up doing some per-packet
5598 * processing in-line, then handling the 'last ptype' for the whole
5599 * sublist. This can't cause out-of-order delivery to any single ptype,
5600 * because the 'last ptype' must be constant across the sublist, and all
5601 * other ptypes are handled per-packet.
5603 /* Current (common) ptype of sublist */
5604 struct packet_type *pt_curr = NULL;
5605 /* Current (common) orig_dev of sublist */
5606 struct net_device *od_curr = NULL;
5607 struct list_head sublist;
5608 struct sk_buff *skb, *next;
5610 INIT_LIST_HEAD(&sublist);
5611 list_for_each_entry_safe(skb, next, head, list) {
5612 struct net_device *orig_dev = skb->dev;
5613 struct packet_type *pt_prev = NULL;
5615 skb_list_del_init(skb);
5616 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev);
5619 if (pt_curr != pt_prev || od_curr != orig_dev) {
5620 /* dispatch old sublist */
5621 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr);
5622 /* start new sublist */
5623 INIT_LIST_HEAD(&sublist);
5627 list_add_tail(&skb->list, &sublist);
5630 /* dispatch final sublist */
5631 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr);
5634 static int __netif_receive_skb(struct sk_buff *skb)
5638 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) {
5639 unsigned int noreclaim_flag;
5642 * PFMEMALLOC skbs are special, they should
5643 * - be delivered to SOCK_MEMALLOC sockets only
5644 * - stay away from userspace
5645 * - have bounded memory usage
5647 * Use PF_MEMALLOC as this saves us from propagating the allocation
5648 * context down to all allocation sites.
5650 noreclaim_flag = memalloc_noreclaim_save();
5651 ret = __netif_receive_skb_one_core(skb, true);
5652 memalloc_noreclaim_restore(noreclaim_flag);
5654 ret = __netif_receive_skb_one_core(skb, false);
5659 static void __netif_receive_skb_list(struct list_head *head)
5661 unsigned long noreclaim_flag = 0;
5662 struct sk_buff *skb, *next;
5663 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */
5665 list_for_each_entry_safe(skb, next, head, list) {
5666 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) {
5667 struct list_head sublist;
5669 /* Handle the previous sublist */
5670 list_cut_before(&sublist, head, &skb->list);
5671 if (!list_empty(&sublist))
5672 __netif_receive_skb_list_core(&sublist, pfmemalloc);
5673 pfmemalloc = !pfmemalloc;
5674 /* See comments in __netif_receive_skb */
5676 noreclaim_flag = memalloc_noreclaim_save();
5678 memalloc_noreclaim_restore(noreclaim_flag);
5681 /* Handle the remaining sublist */
5682 if (!list_empty(head))
5683 __netif_receive_skb_list_core(head, pfmemalloc);
5684 /* Restore pflags */
5686 memalloc_noreclaim_restore(noreclaim_flag);
5689 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp)
5691 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog);
5692 struct bpf_prog *new = xdp->prog;
5695 switch (xdp->command) {
5696 case XDP_SETUP_PROG:
5697 rcu_assign_pointer(dev->xdp_prog, new);
5702 static_branch_dec(&generic_xdp_needed_key);
5703 } else if (new && !old) {
5704 static_branch_inc(&generic_xdp_needed_key);
5705 dev_disable_lro(dev);
5706 dev_disable_gro_hw(dev);
5718 static int netif_receive_skb_internal(struct sk_buff *skb)
5722 net_timestamp_check(READ_ONCE(netdev_tstamp_prequeue), skb);
5724 if (skb_defer_rx_timestamp(skb))
5725 return NET_RX_SUCCESS;
5729 if (static_branch_unlikely(&rps_needed)) {
5730 struct rps_dev_flow voidflow, *rflow = &voidflow;
5731 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
5734 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
5740 ret = __netif_receive_skb(skb);
5745 void netif_receive_skb_list_internal(struct list_head *head)
5747 struct sk_buff *skb, *next;
5748 struct list_head sublist;
5750 INIT_LIST_HEAD(&sublist);
5751 list_for_each_entry_safe(skb, next, head, list) {
5752 net_timestamp_check(READ_ONCE(netdev_tstamp_prequeue), skb);
5753 skb_list_del_init(skb);
5754 if (!skb_defer_rx_timestamp(skb))
5755 list_add_tail(&skb->list, &sublist);
5757 list_splice_init(&sublist, head);
5761 if (static_branch_unlikely(&rps_needed)) {
5762 list_for_each_entry_safe(skb, next, head, list) {
5763 struct rps_dev_flow voidflow, *rflow = &voidflow;
5764 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
5767 /* Will be handled, remove from list */
5768 skb_list_del_init(skb);
5769 enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
5774 __netif_receive_skb_list(head);
5779 * netif_receive_skb - process receive buffer from network
5780 * @skb: buffer to process
5782 * netif_receive_skb() is the main receive data processing function.
5783 * It always succeeds. The buffer may be dropped during processing
5784 * for congestion control or by the protocol layers.
5786 * This function may only be called from softirq context and interrupts
5787 * should be enabled.
5789 * Return values (usually ignored):
5790 * NET_RX_SUCCESS: no congestion
5791 * NET_RX_DROP: packet was dropped
5793 int netif_receive_skb(struct sk_buff *skb)
5797 trace_netif_receive_skb_entry(skb);
5799 ret = netif_receive_skb_internal(skb);
5800 trace_netif_receive_skb_exit(ret);
5804 EXPORT_SYMBOL(netif_receive_skb);
5807 * netif_receive_skb_list - process many receive buffers from network
5808 * @head: list of skbs to process.
5810 * Since return value of netif_receive_skb() is normally ignored, and
5811 * wouldn't be meaningful for a list, this function returns void.
5813 * This function may only be called from softirq context and interrupts
5814 * should be enabled.
5816 void netif_receive_skb_list(struct list_head *head)
5818 struct sk_buff *skb;
5820 if (list_empty(head))
5822 if (trace_netif_receive_skb_list_entry_enabled()) {
5823 list_for_each_entry(skb, head, list)
5824 trace_netif_receive_skb_list_entry(skb);
5826 netif_receive_skb_list_internal(head);
5827 trace_netif_receive_skb_list_exit(0);
5829 EXPORT_SYMBOL(netif_receive_skb_list);
5831 static DEFINE_PER_CPU(struct work_struct, flush_works);
5833 /* Network device is going away, flush any packets still pending */
5834 static void flush_backlog(struct work_struct *work)
5836 struct sk_buff *skb, *tmp;
5837 struct softnet_data *sd;
5840 sd = this_cpu_ptr(&softnet_data);
5842 rps_lock_irq_disable(sd);
5843 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) {
5844 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
5845 __skb_unlink(skb, &sd->input_pkt_queue);
5846 dev_kfree_skb_irq(skb);
5847 input_queue_head_incr(sd);
5850 rps_unlock_irq_enable(sd);
5852 skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
5853 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
5854 __skb_unlink(skb, &sd->process_queue);
5856 input_queue_head_incr(sd);
5862 static bool flush_required(int cpu)
5864 #if IS_ENABLED(CONFIG_RPS)
5865 struct softnet_data *sd = &per_cpu(softnet_data, cpu);
5868 rps_lock_irq_disable(sd);
5870 /* as insertion into process_queue happens with the rps lock held,
5871 * process_queue access may race only with dequeue
5873 do_flush = !skb_queue_empty(&sd->input_pkt_queue) ||
5874 !skb_queue_empty_lockless(&sd->process_queue);
5875 rps_unlock_irq_enable(sd);
5879 /* without RPS we can't safely check input_pkt_queue: during a
5880 * concurrent remote skb_queue_splice() we can detect as empty both
5881 * input_pkt_queue and process_queue even if the latter could end-up
5882 * containing a lot of packets.
5887 static void flush_all_backlogs(void)
5889 static cpumask_t flush_cpus;
5892 /* since we are under rtnl lock protection we can use static data
5893 * for the cpumask and avoid allocating on stack the possibly
5900 cpumask_clear(&flush_cpus);
5901 for_each_online_cpu(cpu) {
5902 if (flush_required(cpu)) {
5903 queue_work_on(cpu, system_highpri_wq,
5904 per_cpu_ptr(&flush_works, cpu));
5905 cpumask_set_cpu(cpu, &flush_cpus);
5909 /* we can have in flight packet[s] on the cpus we are not flushing,
5910 * synchronize_net() in unregister_netdevice_many() will take care of
5913 for_each_cpu(cpu, &flush_cpus)
5914 flush_work(per_cpu_ptr(&flush_works, cpu));
5919 static void net_rps_send_ipi(struct softnet_data *remsd)
5923 struct softnet_data *next = remsd->rps_ipi_next;
5925 if (cpu_online(remsd->cpu))
5926 smp_call_function_single_async(remsd->cpu, &remsd->csd);
5933 * net_rps_action_and_irq_enable sends any pending IPI's for rps.
5934 * Note: called with local irq disabled, but exits with local irq enabled.
5936 static void net_rps_action_and_irq_enable(struct softnet_data *sd)
5939 struct softnet_data *remsd = sd->rps_ipi_list;
5942 sd->rps_ipi_list = NULL;
5946 /* Send pending IPI's to kick RPS processing on remote cpus. */
5947 net_rps_send_ipi(remsd);
5953 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd)
5956 return sd->rps_ipi_list != NULL;
5962 static int process_backlog(struct napi_struct *napi, int quota)
5964 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
5968 /* Check if we have pending ipi, its better to send them now,
5969 * not waiting net_rx_action() end.
5971 if (sd_has_rps_ipi_waiting(sd)) {
5972 local_irq_disable();
5973 net_rps_action_and_irq_enable(sd);
5976 napi->weight = READ_ONCE(dev_rx_weight);
5978 struct sk_buff *skb;
5980 while ((skb = __skb_dequeue(&sd->process_queue))) {
5982 __netif_receive_skb(skb);
5984 input_queue_head_incr(sd);
5985 if (++work >= quota)
5990 rps_lock_irq_disable(sd);
5991 if (skb_queue_empty(&sd->input_pkt_queue)) {
5993 * Inline a custom version of __napi_complete().
5994 * only current cpu owns and manipulates this napi,
5995 * and NAPI_STATE_SCHED is the only possible flag set
5997 * We can use a plain write instead of clear_bit(),
5998 * and we dont need an smp_mb() memory barrier.
6003 skb_queue_splice_tail_init(&sd->input_pkt_queue,
6004 &sd->process_queue);
6006 rps_unlock_irq_enable(sd);
6013 * __napi_schedule - schedule for receive
6014 * @n: entry to schedule
6016 * The entry's receive function will be scheduled to run.
6017 * Consider using __napi_schedule_irqoff() if hard irqs are masked.
6019 void __napi_schedule(struct napi_struct *n)
6021 unsigned long flags;
6023 local_irq_save(flags);
6024 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
6025 local_irq_restore(flags);
6027 EXPORT_SYMBOL(__napi_schedule);
6030 * napi_schedule_prep - check if napi can be scheduled
6033 * Test if NAPI routine is already running, and if not mark
6034 * it as running. This is used as a condition variable to
6035 * insure only one NAPI poll instance runs. We also make
6036 * sure there is no pending NAPI disable.
6038 bool napi_schedule_prep(struct napi_struct *n)
6040 unsigned long new, val = READ_ONCE(n->state);
6043 if (unlikely(val & NAPIF_STATE_DISABLE))
6045 new = val | NAPIF_STATE_SCHED;
6047 /* Sets STATE_MISSED bit if STATE_SCHED was already set
6048 * This was suggested by Alexander Duyck, as compiler
6049 * emits better code than :
6050 * if (val & NAPIF_STATE_SCHED)
6051 * new |= NAPIF_STATE_MISSED;
6053 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED *
6055 } while (!try_cmpxchg(&n->state, &val, new));
6057 return !(val & NAPIF_STATE_SCHED);
6059 EXPORT_SYMBOL(napi_schedule_prep);
6062 * __napi_schedule_irqoff - schedule for receive
6063 * @n: entry to schedule
6065 * Variant of __napi_schedule() assuming hard irqs are masked.
6067 * On PREEMPT_RT enabled kernels this maps to __napi_schedule()
6068 * because the interrupt disabled assumption might not be true
6069 * due to force-threaded interrupts and spinlock substitution.
6071 void __napi_schedule_irqoff(struct napi_struct *n)
6073 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
6074 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
6078 EXPORT_SYMBOL(__napi_schedule_irqoff);
6080 bool napi_complete_done(struct napi_struct *n, int work_done)
6082 unsigned long flags, val, new, timeout = 0;
6086 * 1) Don't let napi dequeue from the cpu poll list
6087 * just in case its running on a different cpu.
6088 * 2) If we are busy polling, do nothing here, we have
6089 * the guarantee we will be called later.
6091 if (unlikely(n->state & (NAPIF_STATE_NPSVC |
6092 NAPIF_STATE_IN_BUSY_POLL)))
6097 timeout = READ_ONCE(n->dev->gro_flush_timeout);
6098 n->defer_hard_irqs_count = READ_ONCE(n->dev->napi_defer_hard_irqs);
6100 if (n->defer_hard_irqs_count > 0) {
6101 n->defer_hard_irqs_count--;
6102 timeout = READ_ONCE(n->dev->gro_flush_timeout);
6106 if (n->gro_bitmask) {
6107 /* When the NAPI instance uses a timeout and keeps postponing
6108 * it, we need to bound somehow the time packets are kept in
6111 napi_gro_flush(n, !!timeout);
6116 if (unlikely(!list_empty(&n->poll_list))) {
6117 /* If n->poll_list is not empty, we need to mask irqs */
6118 local_irq_save(flags);
6119 list_del_init(&n->poll_list);
6120 local_irq_restore(flags);
6122 WRITE_ONCE(n->list_owner, -1);
6124 val = READ_ONCE(n->state);
6126 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED));
6128 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED |
6129 NAPIF_STATE_SCHED_THREADED |
6130 NAPIF_STATE_PREFER_BUSY_POLL);
6132 /* If STATE_MISSED was set, leave STATE_SCHED set,
6133 * because we will call napi->poll() one more time.
6134 * This C code was suggested by Alexander Duyck to help gcc.
6136 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED *
6138 } while (!try_cmpxchg(&n->state, &val, new));
6140 if (unlikely(val & NAPIF_STATE_MISSED)) {
6146 hrtimer_start(&n->timer, ns_to_ktime(timeout),
6147 HRTIMER_MODE_REL_PINNED);
6150 EXPORT_SYMBOL(napi_complete_done);
6152 /* must be called under rcu_read_lock(), as we dont take a reference */
6153 struct napi_struct *napi_by_id(unsigned int napi_id)
6155 unsigned int hash = napi_id % HASH_SIZE(napi_hash);
6156 struct napi_struct *napi;
6158 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node)
6159 if (napi->napi_id == napi_id)
6165 #if defined(CONFIG_NET_RX_BUSY_POLL)
6167 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule)
6169 if (!skip_schedule) {
6170 gro_normal_list(napi);
6171 __napi_schedule(napi);
6175 if (napi->gro_bitmask) {
6176 /* flush too old packets
6177 * If HZ < 1000, flush all packets.
6179 napi_gro_flush(napi, HZ >= 1000);
6182 gro_normal_list(napi);
6183 clear_bit(NAPI_STATE_SCHED, &napi->state);
6186 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, bool prefer_busy_poll,
6189 bool skip_schedule = false;
6190 unsigned long timeout;
6193 /* Busy polling means there is a high chance device driver hard irq
6194 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was
6195 * set in napi_schedule_prep().
6196 * Since we are about to call napi->poll() once more, we can safely
6197 * clear NAPI_STATE_MISSED.
6199 * Note: x86 could use a single "lock and ..." instruction
6200 * to perform these two clear_bit()
6202 clear_bit(NAPI_STATE_MISSED, &napi->state);
6203 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state);
6207 if (prefer_busy_poll) {
6208 napi->defer_hard_irqs_count = READ_ONCE(napi->dev->napi_defer_hard_irqs);
6209 timeout = READ_ONCE(napi->dev->gro_flush_timeout);
6210 if (napi->defer_hard_irqs_count && timeout) {
6211 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED);
6212 skip_schedule = true;
6216 /* All we really want here is to re-enable device interrupts.
6217 * Ideally, a new ndo_busy_poll_stop() could avoid another round.
6219 rc = napi->poll(napi, budget);
6220 /* We can't gro_normal_list() here, because napi->poll() might have
6221 * rearmed the napi (napi_complete_done()) in which case it could
6222 * already be running on another CPU.
6224 trace_napi_poll(napi, rc, budget);
6225 netpoll_poll_unlock(have_poll_lock);
6227 __busy_poll_stop(napi, skip_schedule);
6231 void napi_busy_loop(unsigned int napi_id,
6232 bool (*loop_end)(void *, unsigned long),
6233 void *loop_end_arg, bool prefer_busy_poll, u16 budget)
6235 unsigned long start_time = loop_end ? busy_loop_current_time() : 0;
6236 int (*napi_poll)(struct napi_struct *napi, int budget);
6237 void *have_poll_lock = NULL;
6238 struct napi_struct *napi;
6245 napi = napi_by_id(napi_id);
6249 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
6256 unsigned long val = READ_ONCE(napi->state);
6258 /* If multiple threads are competing for this napi,
6259 * we avoid dirtying napi->state as much as we can.
6261 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED |
6262 NAPIF_STATE_IN_BUSY_POLL)) {
6263 if (prefer_busy_poll)
6264 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state);
6267 if (cmpxchg(&napi->state, val,
6268 val | NAPIF_STATE_IN_BUSY_POLL |
6269 NAPIF_STATE_SCHED) != val) {
6270 if (prefer_busy_poll)
6271 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state);
6274 have_poll_lock = netpoll_poll_lock(napi);
6275 napi_poll = napi->poll;
6277 work = napi_poll(napi, budget);
6278 trace_napi_poll(napi, work, budget);
6279 gro_normal_list(napi);
6282 __NET_ADD_STATS(dev_net(napi->dev),
6283 LINUX_MIB_BUSYPOLLRXPACKETS, work);
6286 if (!loop_end || loop_end(loop_end_arg, start_time))
6289 if (unlikely(need_resched())) {
6291 busy_poll_stop(napi, have_poll_lock, prefer_busy_poll, budget);
6292 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
6296 if (loop_end(loop_end_arg, start_time))
6303 busy_poll_stop(napi, have_poll_lock, prefer_busy_poll, budget);
6304 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
6309 EXPORT_SYMBOL(napi_busy_loop);
6311 #endif /* CONFIG_NET_RX_BUSY_POLL */
6313 static void napi_hash_add(struct napi_struct *napi)
6315 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state))
6318 spin_lock(&napi_hash_lock);
6320 /* 0..NR_CPUS range is reserved for sender_cpu use */
6322 if (unlikely(++napi_gen_id < MIN_NAPI_ID))
6323 napi_gen_id = MIN_NAPI_ID;
6324 } while (napi_by_id(napi_gen_id));
6325 napi->napi_id = napi_gen_id;
6327 hlist_add_head_rcu(&napi->napi_hash_node,
6328 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]);
6330 spin_unlock(&napi_hash_lock);
6333 /* Warning : caller is responsible to make sure rcu grace period
6334 * is respected before freeing memory containing @napi
6336 static void napi_hash_del(struct napi_struct *napi)
6338 spin_lock(&napi_hash_lock);
6340 hlist_del_init_rcu(&napi->napi_hash_node);
6342 spin_unlock(&napi_hash_lock);
6345 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer)
6347 struct napi_struct *napi;
6349 napi = container_of(timer, struct napi_struct, timer);
6351 /* Note : we use a relaxed variant of napi_schedule_prep() not setting
6352 * NAPI_STATE_MISSED, since we do not react to a device IRQ.
6354 if (!napi_disable_pending(napi) &&
6355 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) {
6356 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state);
6357 __napi_schedule_irqoff(napi);
6360 return HRTIMER_NORESTART;
6363 static void init_gro_hash(struct napi_struct *napi)
6367 for (i = 0; i < GRO_HASH_BUCKETS; i++) {
6368 INIT_LIST_HEAD(&napi->gro_hash[i].list);
6369 napi->gro_hash[i].count = 0;
6371 napi->gro_bitmask = 0;
6374 int dev_set_threaded(struct net_device *dev, bool threaded)
6376 struct napi_struct *napi;
6379 if (dev->threaded == threaded)
6383 list_for_each_entry(napi, &dev->napi_list, dev_list) {
6384 if (!napi->thread) {
6385 err = napi_kthread_create(napi);
6394 dev->threaded = threaded;
6396 /* Make sure kthread is created before THREADED bit
6399 smp_mb__before_atomic();
6401 /* Setting/unsetting threaded mode on a napi might not immediately
6402 * take effect, if the current napi instance is actively being
6403 * polled. In this case, the switch between threaded mode and
6404 * softirq mode will happen in the next round of napi_schedule().
6405 * This should not cause hiccups/stalls to the live traffic.
6407 list_for_each_entry(napi, &dev->napi_list, dev_list)
6408 assign_bit(NAPI_STATE_THREADED, &napi->state, threaded);
6412 EXPORT_SYMBOL(dev_set_threaded);
6415 * netif_queue_set_napi - Associate queue with the napi
6416 * @dev: device to which NAPI and queue belong
6417 * @queue_index: Index of queue
6418 * @type: queue type as RX or TX
6419 * @napi: NAPI context, pass NULL to clear previously set NAPI
6421 * Set queue with its corresponding napi context. This should be done after
6422 * registering the NAPI handler for the queue-vector and the queues have been
6423 * mapped to the corresponding interrupt vector.
6425 void netif_queue_set_napi(struct net_device *dev, unsigned int queue_index,
6426 enum netdev_queue_type type, struct napi_struct *napi)
6428 struct netdev_rx_queue *rxq;
6429 struct netdev_queue *txq;
6431 if (WARN_ON_ONCE(napi && !napi->dev))
6433 if (dev->reg_state >= NETREG_REGISTERED)
6437 case NETDEV_QUEUE_TYPE_RX:
6438 rxq = __netif_get_rx_queue(dev, queue_index);
6441 case NETDEV_QUEUE_TYPE_TX:
6442 txq = netdev_get_tx_queue(dev, queue_index);
6449 EXPORT_SYMBOL(netif_queue_set_napi);
6451 void netif_napi_add_weight(struct net_device *dev, struct napi_struct *napi,
6452 int (*poll)(struct napi_struct *, int), int weight)
6454 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state)))
6457 INIT_LIST_HEAD(&napi->poll_list);
6458 INIT_HLIST_NODE(&napi->napi_hash_node);
6459 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
6460 napi->timer.function = napi_watchdog;
6461 init_gro_hash(napi);
6463 INIT_LIST_HEAD(&napi->rx_list);
6466 if (weight > NAPI_POLL_WEIGHT)
6467 netdev_err_once(dev, "%s() called with weight %d\n", __func__,
6469 napi->weight = weight;
6471 #ifdef CONFIG_NETPOLL
6472 napi->poll_owner = -1;
6474 napi->list_owner = -1;
6475 set_bit(NAPI_STATE_SCHED, &napi->state);
6476 set_bit(NAPI_STATE_NPSVC, &napi->state);
6477 list_add_rcu(&napi->dev_list, &dev->napi_list);
6478 napi_hash_add(napi);
6479 napi_get_frags_check(napi);
6480 /* Create kthread for this napi if dev->threaded is set.
6481 * Clear dev->threaded if kthread creation failed so that
6482 * threaded mode will not be enabled in napi_enable().
6484 if (dev->threaded && napi_kthread_create(napi))
6486 netif_napi_set_irq(napi, -1);
6488 EXPORT_SYMBOL(netif_napi_add_weight);
6490 void napi_disable(struct napi_struct *n)
6492 unsigned long val, new;
6495 set_bit(NAPI_STATE_DISABLE, &n->state);
6497 val = READ_ONCE(n->state);
6499 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) {
6500 usleep_range(20, 200);
6501 val = READ_ONCE(n->state);
6504 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC;
6505 new &= ~(NAPIF_STATE_THREADED | NAPIF_STATE_PREFER_BUSY_POLL);
6506 } while (!try_cmpxchg(&n->state, &val, new));
6508 hrtimer_cancel(&n->timer);
6510 clear_bit(NAPI_STATE_DISABLE, &n->state);
6512 EXPORT_SYMBOL(napi_disable);
6515 * napi_enable - enable NAPI scheduling
6518 * Resume NAPI from being scheduled on this context.
6519 * Must be paired with napi_disable.
6521 void napi_enable(struct napi_struct *n)
6523 unsigned long new, val = READ_ONCE(n->state);
6526 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val));
6528 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC);
6529 if (n->dev->threaded && n->thread)
6530 new |= NAPIF_STATE_THREADED;
6531 } while (!try_cmpxchg(&n->state, &val, new));
6533 EXPORT_SYMBOL(napi_enable);
6535 static void flush_gro_hash(struct napi_struct *napi)
6539 for (i = 0; i < GRO_HASH_BUCKETS; i++) {
6540 struct sk_buff *skb, *n;
6542 list_for_each_entry_safe(skb, n, &napi->gro_hash[i].list, list)
6544 napi->gro_hash[i].count = 0;
6548 /* Must be called in process context */
6549 void __netif_napi_del(struct napi_struct *napi)
6551 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state))
6554 napi_hash_del(napi);
6555 list_del_rcu(&napi->dev_list);
6556 napi_free_frags(napi);
6558 flush_gro_hash(napi);
6559 napi->gro_bitmask = 0;
6562 kthread_stop(napi->thread);
6563 napi->thread = NULL;
6566 EXPORT_SYMBOL(__netif_napi_del);
6568 static int __napi_poll(struct napi_struct *n, bool *repoll)
6574 /* This NAPI_STATE_SCHED test is for avoiding a race
6575 * with netpoll's poll_napi(). Only the entity which
6576 * obtains the lock and sees NAPI_STATE_SCHED set will
6577 * actually make the ->poll() call. Therefore we avoid
6578 * accidentally calling ->poll() when NAPI is not scheduled.
6581 if (napi_is_scheduled(n)) {
6582 work = n->poll(n, weight);
6583 trace_napi_poll(n, work, weight);
6585 xdp_do_check_flushed(n);
6588 if (unlikely(work > weight))
6589 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n",
6590 n->poll, work, weight);
6592 if (likely(work < weight))
6595 /* Drivers must not modify the NAPI state if they
6596 * consume the entire weight. In such cases this code
6597 * still "owns" the NAPI instance and therefore can
6598 * move the instance around on the list at-will.
6600 if (unlikely(napi_disable_pending(n))) {
6605 /* The NAPI context has more processing work, but busy-polling
6606 * is preferred. Exit early.
6608 if (napi_prefer_busy_poll(n)) {
6609 if (napi_complete_done(n, work)) {
6610 /* If timeout is not set, we need to make sure
6611 * that the NAPI is re-scheduled.
6618 if (n->gro_bitmask) {
6619 /* flush too old packets
6620 * If HZ < 1000, flush all packets.
6622 napi_gro_flush(n, HZ >= 1000);
6627 /* Some drivers may have called napi_schedule
6628 * prior to exhausting their budget.
6630 if (unlikely(!list_empty(&n->poll_list))) {
6631 pr_warn_once("%s: Budget exhausted after napi rescheduled\n",
6632 n->dev ? n->dev->name : "backlog");
6641 static int napi_poll(struct napi_struct *n, struct list_head *repoll)
6643 bool do_repoll = false;
6647 list_del_init(&n->poll_list);
6649 have = netpoll_poll_lock(n);
6651 work = __napi_poll(n, &do_repoll);
6654 list_add_tail(&n->poll_list, repoll);
6656 netpoll_poll_unlock(have);
6661 static int napi_thread_wait(struct napi_struct *napi)
6665 set_current_state(TASK_INTERRUPTIBLE);
6667 while (!kthread_should_stop()) {
6668 /* Testing SCHED_THREADED bit here to make sure the current
6669 * kthread owns this napi and could poll on this napi.
6670 * Testing SCHED bit is not enough because SCHED bit might be
6671 * set by some other busy poll thread or by napi_disable().
6673 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state) || woken) {
6674 WARN_ON(!list_empty(&napi->poll_list));
6675 __set_current_state(TASK_RUNNING);
6680 /* woken being true indicates this thread owns this napi. */
6682 set_current_state(TASK_INTERRUPTIBLE);
6684 __set_current_state(TASK_RUNNING);
6689 static void skb_defer_free_flush(struct softnet_data *sd)
6691 struct sk_buff *skb, *next;
6693 /* Paired with WRITE_ONCE() in skb_attempt_defer_free() */
6694 if (!READ_ONCE(sd->defer_list))
6697 spin_lock(&sd->defer_lock);
6698 skb = sd->defer_list;
6699 sd->defer_list = NULL;
6700 sd->defer_count = 0;
6701 spin_unlock(&sd->defer_lock);
6703 while (skb != NULL) {
6705 napi_consume_skb(skb, 1);
6710 static int napi_threaded_poll(void *data)
6712 struct napi_struct *napi = data;
6713 struct softnet_data *sd;
6716 while (!napi_thread_wait(napi)) {
6717 unsigned long last_qs = jiffies;
6720 bool repoll = false;
6723 sd = this_cpu_ptr(&softnet_data);
6724 sd->in_napi_threaded_poll = true;
6726 have = netpoll_poll_lock(napi);
6727 __napi_poll(napi, &repoll);
6728 netpoll_poll_unlock(have);
6730 sd->in_napi_threaded_poll = false;
6733 if (sd_has_rps_ipi_waiting(sd)) {
6734 local_irq_disable();
6735 net_rps_action_and_irq_enable(sd);
6737 skb_defer_free_flush(sd);
6743 rcu_softirq_qs_periodic(last_qs);
6750 static __latent_entropy void net_rx_action(struct softirq_action *h)
6752 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
6753 unsigned long time_limit = jiffies +
6754 usecs_to_jiffies(READ_ONCE(netdev_budget_usecs));
6755 int budget = READ_ONCE(netdev_budget);
6760 sd->in_net_rx_action = true;
6761 local_irq_disable();
6762 list_splice_init(&sd->poll_list, &list);
6766 struct napi_struct *n;
6768 skb_defer_free_flush(sd);
6770 if (list_empty(&list)) {
6771 if (list_empty(&repoll)) {
6772 sd->in_net_rx_action = false;
6774 /* We need to check if ____napi_schedule()
6775 * had refilled poll_list while
6776 * sd->in_net_rx_action was true.
6778 if (!list_empty(&sd->poll_list))
6780 if (!sd_has_rps_ipi_waiting(sd))
6786 n = list_first_entry(&list, struct napi_struct, poll_list);
6787 budget -= napi_poll(n, &repoll);
6789 /* If softirq window is exhausted then punt.
6790 * Allow this to run for 2 jiffies since which will allow
6791 * an average latency of 1.5/HZ.
6793 if (unlikely(budget <= 0 ||
6794 time_after_eq(jiffies, time_limit))) {
6800 local_irq_disable();
6802 list_splice_tail_init(&sd->poll_list, &list);
6803 list_splice_tail(&repoll, &list);
6804 list_splice(&list, &sd->poll_list);
6805 if (!list_empty(&sd->poll_list))
6806 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
6808 sd->in_net_rx_action = false;
6810 net_rps_action_and_irq_enable(sd);
6814 struct netdev_adjacent {
6815 struct net_device *dev;
6816 netdevice_tracker dev_tracker;
6818 /* upper master flag, there can only be one master device per list */
6821 /* lookup ignore flag */
6824 /* counter for the number of times this device was added to us */
6827 /* private field for the users */
6830 struct list_head list;
6831 struct rcu_head rcu;
6834 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev,
6835 struct list_head *adj_list)
6837 struct netdev_adjacent *adj;
6839 list_for_each_entry(adj, adj_list, list) {
6840 if (adj->dev == adj_dev)
6846 static int ____netdev_has_upper_dev(struct net_device *upper_dev,
6847 struct netdev_nested_priv *priv)
6849 struct net_device *dev = (struct net_device *)priv->data;
6851 return upper_dev == dev;
6855 * netdev_has_upper_dev - Check if device is linked to an upper device
6857 * @upper_dev: upper device to check
6859 * Find out if a device is linked to specified upper device and return true
6860 * in case it is. Note that this checks only immediate upper device,
6861 * not through a complete stack of devices. The caller must hold the RTNL lock.
6863 bool netdev_has_upper_dev(struct net_device *dev,
6864 struct net_device *upper_dev)
6866 struct netdev_nested_priv priv = {
6867 .data = (void *)upper_dev,
6872 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev,
6875 EXPORT_SYMBOL(netdev_has_upper_dev);
6878 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device
6880 * @upper_dev: upper device to check
6882 * Find out if a device is linked to specified upper device and return true
6883 * in case it is. Note that this checks the entire upper device chain.
6884 * The caller must hold rcu lock.
6887 bool netdev_has_upper_dev_all_rcu(struct net_device *dev,
6888 struct net_device *upper_dev)
6890 struct netdev_nested_priv priv = {
6891 .data = (void *)upper_dev,
6894 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev,
6897 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu);
6900 * netdev_has_any_upper_dev - Check if device is linked to some device
6903 * Find out if a device is linked to an upper device and return true in case
6904 * it is. The caller must hold the RTNL lock.
6906 bool netdev_has_any_upper_dev(struct net_device *dev)
6910 return !list_empty(&dev->adj_list.upper);
6912 EXPORT_SYMBOL(netdev_has_any_upper_dev);
6915 * netdev_master_upper_dev_get - Get master upper device
6918 * Find a master upper device and return pointer to it or NULL in case
6919 * it's not there. The caller must hold the RTNL lock.
6921 struct net_device *netdev_master_upper_dev_get(struct net_device *dev)
6923 struct netdev_adjacent *upper;
6927 if (list_empty(&dev->adj_list.upper))
6930 upper = list_first_entry(&dev->adj_list.upper,
6931 struct netdev_adjacent, list);
6932 if (likely(upper->master))
6936 EXPORT_SYMBOL(netdev_master_upper_dev_get);
6938 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev)
6940 struct netdev_adjacent *upper;
6944 if (list_empty(&dev->adj_list.upper))
6947 upper = list_first_entry(&dev->adj_list.upper,
6948 struct netdev_adjacent, list);
6949 if (likely(upper->master) && !upper->ignore)
6955 * netdev_has_any_lower_dev - Check if device is linked to some device
6958 * Find out if a device is linked to a lower device and return true in case
6959 * it is. The caller must hold the RTNL lock.
6961 static bool netdev_has_any_lower_dev(struct net_device *dev)
6965 return !list_empty(&dev->adj_list.lower);
6968 void *netdev_adjacent_get_private(struct list_head *adj_list)
6970 struct netdev_adjacent *adj;
6972 adj = list_entry(adj_list, struct netdev_adjacent, list);
6974 return adj->private;
6976 EXPORT_SYMBOL(netdev_adjacent_get_private);
6979 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list
6981 * @iter: list_head ** of the current position
6983 * Gets the next device from the dev's upper list, starting from iter
6984 * position. The caller must hold RCU read lock.
6986 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev,
6987 struct list_head **iter)
6989 struct netdev_adjacent *upper;
6991 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
6993 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6995 if (&upper->list == &dev->adj_list.upper)
6998 *iter = &upper->list;
7002 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu);
7004 static struct net_device *__netdev_next_upper_dev(struct net_device *dev,
7005 struct list_head **iter,
7008 struct netdev_adjacent *upper;
7010 upper = list_entry((*iter)->next, struct netdev_adjacent, list);
7012 if (&upper->list == &dev->adj_list.upper)
7015 *iter = &upper->list;
7016 *ignore = upper->ignore;
7021 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev,
7022 struct list_head **iter)
7024 struct netdev_adjacent *upper;
7026 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
7028 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7030 if (&upper->list == &dev->adj_list.upper)
7033 *iter = &upper->list;
7038 static int __netdev_walk_all_upper_dev(struct net_device *dev,
7039 int (*fn)(struct net_device *dev,
7040 struct netdev_nested_priv *priv),
7041 struct netdev_nested_priv *priv)
7043 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7044 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7049 iter = &dev->adj_list.upper;
7053 ret = fn(now, priv);
7060 udev = __netdev_next_upper_dev(now, &iter, &ignore);
7067 niter = &udev->adj_list.upper;
7068 dev_stack[cur] = now;
7069 iter_stack[cur++] = iter;
7076 next = dev_stack[--cur];
7077 niter = iter_stack[cur];
7087 int netdev_walk_all_upper_dev_rcu(struct net_device *dev,
7088 int (*fn)(struct net_device *dev,
7089 struct netdev_nested_priv *priv),
7090 struct netdev_nested_priv *priv)
7092 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7093 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7097 iter = &dev->adj_list.upper;
7101 ret = fn(now, priv);
7108 udev = netdev_next_upper_dev_rcu(now, &iter);
7113 niter = &udev->adj_list.upper;
7114 dev_stack[cur] = now;
7115 iter_stack[cur++] = iter;
7122 next = dev_stack[--cur];
7123 niter = iter_stack[cur];
7132 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu);
7134 static bool __netdev_has_upper_dev(struct net_device *dev,
7135 struct net_device *upper_dev)
7137 struct netdev_nested_priv priv = {
7139 .data = (void *)upper_dev,
7144 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev,
7149 * netdev_lower_get_next_private - Get the next ->private from the
7150 * lower neighbour list
7152 * @iter: list_head ** of the current position
7154 * Gets the next netdev_adjacent->private from the dev's lower neighbour
7155 * list, starting from iter position. The caller must hold either hold the
7156 * RTNL lock or its own locking that guarantees that the neighbour lower
7157 * list will remain unchanged.
7159 void *netdev_lower_get_next_private(struct net_device *dev,
7160 struct list_head **iter)
7162 struct netdev_adjacent *lower;
7164 lower = list_entry(*iter, struct netdev_adjacent, list);
7166 if (&lower->list == &dev->adj_list.lower)
7169 *iter = lower->list.next;
7171 return lower->private;
7173 EXPORT_SYMBOL(netdev_lower_get_next_private);
7176 * netdev_lower_get_next_private_rcu - Get the next ->private from the
7177 * lower neighbour list, RCU
7180 * @iter: list_head ** of the current position
7182 * Gets the next netdev_adjacent->private from the dev's lower neighbour
7183 * list, starting from iter position. The caller must hold RCU read lock.
7185 void *netdev_lower_get_next_private_rcu(struct net_device *dev,
7186 struct list_head **iter)
7188 struct netdev_adjacent *lower;
7190 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
7192 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7194 if (&lower->list == &dev->adj_list.lower)
7197 *iter = &lower->list;
7199 return lower->private;
7201 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu);
7204 * netdev_lower_get_next - Get the next device from the lower neighbour
7207 * @iter: list_head ** of the current position
7209 * Gets the next netdev_adjacent from the dev's lower neighbour
7210 * list, starting from iter position. The caller must hold RTNL lock or
7211 * its own locking that guarantees that the neighbour lower
7212 * list will remain unchanged.
7214 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter)
7216 struct netdev_adjacent *lower;
7218 lower = list_entry(*iter, struct netdev_adjacent, list);
7220 if (&lower->list == &dev->adj_list.lower)
7223 *iter = lower->list.next;
7227 EXPORT_SYMBOL(netdev_lower_get_next);
7229 static struct net_device *netdev_next_lower_dev(struct net_device *dev,
7230 struct list_head **iter)
7232 struct netdev_adjacent *lower;
7234 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
7236 if (&lower->list == &dev->adj_list.lower)
7239 *iter = &lower->list;
7244 static struct net_device *__netdev_next_lower_dev(struct net_device *dev,
7245 struct list_head **iter,
7248 struct netdev_adjacent *lower;
7250 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
7252 if (&lower->list == &dev->adj_list.lower)
7255 *iter = &lower->list;
7256 *ignore = lower->ignore;
7261 int netdev_walk_all_lower_dev(struct net_device *dev,
7262 int (*fn)(struct net_device *dev,
7263 struct netdev_nested_priv *priv),
7264 struct netdev_nested_priv *priv)
7266 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7267 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7271 iter = &dev->adj_list.lower;
7275 ret = fn(now, priv);
7282 ldev = netdev_next_lower_dev(now, &iter);
7287 niter = &ldev->adj_list.lower;
7288 dev_stack[cur] = now;
7289 iter_stack[cur++] = iter;
7296 next = dev_stack[--cur];
7297 niter = iter_stack[cur];
7306 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev);
7308 static int __netdev_walk_all_lower_dev(struct net_device *dev,
7309 int (*fn)(struct net_device *dev,
7310 struct netdev_nested_priv *priv),
7311 struct netdev_nested_priv *priv)
7313 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7314 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7319 iter = &dev->adj_list.lower;
7323 ret = fn(now, priv);
7330 ldev = __netdev_next_lower_dev(now, &iter, &ignore);
7337 niter = &ldev->adj_list.lower;
7338 dev_stack[cur] = now;
7339 iter_stack[cur++] = iter;
7346 next = dev_stack[--cur];
7347 niter = iter_stack[cur];
7357 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev,
7358 struct list_head **iter)
7360 struct netdev_adjacent *lower;
7362 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7363 if (&lower->list == &dev->adj_list.lower)
7366 *iter = &lower->list;
7370 EXPORT_SYMBOL(netdev_next_lower_dev_rcu);
7372 static u8 __netdev_upper_depth(struct net_device *dev)
7374 struct net_device *udev;
7375 struct list_head *iter;
7379 for (iter = &dev->adj_list.upper,
7380 udev = __netdev_next_upper_dev(dev, &iter, &ignore);
7382 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) {
7385 if (max_depth < udev->upper_level)
7386 max_depth = udev->upper_level;
7392 static u8 __netdev_lower_depth(struct net_device *dev)
7394 struct net_device *ldev;
7395 struct list_head *iter;
7399 for (iter = &dev->adj_list.lower,
7400 ldev = __netdev_next_lower_dev(dev, &iter, &ignore);
7402 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) {
7405 if (max_depth < ldev->lower_level)
7406 max_depth = ldev->lower_level;
7412 static int __netdev_update_upper_level(struct net_device *dev,
7413 struct netdev_nested_priv *__unused)
7415 dev->upper_level = __netdev_upper_depth(dev) + 1;
7419 #ifdef CONFIG_LOCKDEP
7420 static LIST_HEAD(net_unlink_list);
7422 static void net_unlink_todo(struct net_device *dev)
7424 if (list_empty(&dev->unlink_list))
7425 list_add_tail(&dev->unlink_list, &net_unlink_list);
7429 static int __netdev_update_lower_level(struct net_device *dev,
7430 struct netdev_nested_priv *priv)
7432 dev->lower_level = __netdev_lower_depth(dev) + 1;
7434 #ifdef CONFIG_LOCKDEP
7438 if (priv->flags & NESTED_SYNC_IMM)
7439 dev->nested_level = dev->lower_level - 1;
7440 if (priv->flags & NESTED_SYNC_TODO)
7441 net_unlink_todo(dev);
7446 int netdev_walk_all_lower_dev_rcu(struct net_device *dev,
7447 int (*fn)(struct net_device *dev,
7448 struct netdev_nested_priv *priv),
7449 struct netdev_nested_priv *priv)
7451 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7452 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7456 iter = &dev->adj_list.lower;
7460 ret = fn(now, priv);
7467 ldev = netdev_next_lower_dev_rcu(now, &iter);
7472 niter = &ldev->adj_list.lower;
7473 dev_stack[cur] = now;
7474 iter_stack[cur++] = iter;
7481 next = dev_stack[--cur];
7482 niter = iter_stack[cur];
7491 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu);
7494 * netdev_lower_get_first_private_rcu - Get the first ->private from the
7495 * lower neighbour list, RCU
7499 * Gets the first netdev_adjacent->private from the dev's lower neighbour
7500 * list. The caller must hold RCU read lock.
7502 void *netdev_lower_get_first_private_rcu(struct net_device *dev)
7504 struct netdev_adjacent *lower;
7506 lower = list_first_or_null_rcu(&dev->adj_list.lower,
7507 struct netdev_adjacent, list);
7509 return lower->private;
7512 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu);
7515 * netdev_master_upper_dev_get_rcu - Get master upper device
7518 * Find a master upper device and return pointer to it or NULL in case
7519 * it's not there. The caller must hold the RCU read lock.
7521 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev)
7523 struct netdev_adjacent *upper;
7525 upper = list_first_or_null_rcu(&dev->adj_list.upper,
7526 struct netdev_adjacent, list);
7527 if (upper && likely(upper->master))
7531 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu);
7533 static int netdev_adjacent_sysfs_add(struct net_device *dev,
7534 struct net_device *adj_dev,
7535 struct list_head *dev_list)
7537 char linkname[IFNAMSIZ+7];
7539 sprintf(linkname, dev_list == &dev->adj_list.upper ?
7540 "upper_%s" : "lower_%s", adj_dev->name);
7541 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj),
7544 static void netdev_adjacent_sysfs_del(struct net_device *dev,
7546 struct list_head *dev_list)
7548 char linkname[IFNAMSIZ+7];
7550 sprintf(linkname, dev_list == &dev->adj_list.upper ?
7551 "upper_%s" : "lower_%s", name);
7552 sysfs_remove_link(&(dev->dev.kobj), linkname);
7555 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev,
7556 struct net_device *adj_dev,
7557 struct list_head *dev_list)
7559 return (dev_list == &dev->adj_list.upper ||
7560 dev_list == &dev->adj_list.lower) &&
7561 net_eq(dev_net(dev), dev_net(adj_dev));
7564 static int __netdev_adjacent_dev_insert(struct net_device *dev,
7565 struct net_device *adj_dev,
7566 struct list_head *dev_list,
7567 void *private, bool master)
7569 struct netdev_adjacent *adj;
7572 adj = __netdev_find_adj(adj_dev, dev_list);
7576 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n",
7577 dev->name, adj_dev->name, adj->ref_nr);
7582 adj = kmalloc(sizeof(*adj), GFP_KERNEL);
7587 adj->master = master;
7589 adj->private = private;
7590 adj->ignore = false;
7591 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL);
7593 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n",
7594 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name);
7596 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) {
7597 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list);
7602 /* Ensure that master link is always the first item in list. */
7604 ret = sysfs_create_link(&(dev->dev.kobj),
7605 &(adj_dev->dev.kobj), "master");
7607 goto remove_symlinks;
7609 list_add_rcu(&adj->list, dev_list);
7611 list_add_tail_rcu(&adj->list, dev_list);
7617 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
7618 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
7620 netdev_put(adj_dev, &adj->dev_tracker);
7626 static void __netdev_adjacent_dev_remove(struct net_device *dev,
7627 struct net_device *adj_dev,
7629 struct list_head *dev_list)
7631 struct netdev_adjacent *adj;
7633 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n",
7634 dev->name, adj_dev->name, ref_nr);
7636 adj = __netdev_find_adj(adj_dev, dev_list);
7639 pr_err("Adjacency does not exist for device %s from %s\n",
7640 dev->name, adj_dev->name);
7645 if (adj->ref_nr > ref_nr) {
7646 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n",
7647 dev->name, adj_dev->name, ref_nr,
7648 adj->ref_nr - ref_nr);
7649 adj->ref_nr -= ref_nr;
7654 sysfs_remove_link(&(dev->dev.kobj), "master");
7656 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
7657 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
7659 list_del_rcu(&adj->list);
7660 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n",
7661 adj_dev->name, dev->name, adj_dev->name);
7662 netdev_put(adj_dev, &adj->dev_tracker);
7663 kfree_rcu(adj, rcu);
7666 static int __netdev_adjacent_dev_link_lists(struct net_device *dev,
7667 struct net_device *upper_dev,
7668 struct list_head *up_list,
7669 struct list_head *down_list,
7670 void *private, bool master)
7674 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list,
7679 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list,
7682 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list);
7689 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev,
7690 struct net_device *upper_dev,
7692 struct list_head *up_list,
7693 struct list_head *down_list)
7695 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list);
7696 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list);
7699 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev,
7700 struct net_device *upper_dev,
7701 void *private, bool master)
7703 return __netdev_adjacent_dev_link_lists(dev, upper_dev,
7704 &dev->adj_list.upper,
7705 &upper_dev->adj_list.lower,
7709 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev,
7710 struct net_device *upper_dev)
7712 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1,
7713 &dev->adj_list.upper,
7714 &upper_dev->adj_list.lower);
7717 static int __netdev_upper_dev_link(struct net_device *dev,
7718 struct net_device *upper_dev, bool master,
7719 void *upper_priv, void *upper_info,
7720 struct netdev_nested_priv *priv,
7721 struct netlink_ext_ack *extack)
7723 struct netdev_notifier_changeupper_info changeupper_info = {
7728 .upper_dev = upper_dev,
7731 .upper_info = upper_info,
7733 struct net_device *master_dev;
7738 if (dev == upper_dev)
7741 /* To prevent loops, check if dev is not upper device to upper_dev. */
7742 if (__netdev_has_upper_dev(upper_dev, dev))
7745 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV)
7749 if (__netdev_has_upper_dev(dev, upper_dev))
7752 master_dev = __netdev_master_upper_dev_get(dev);
7754 return master_dev == upper_dev ? -EEXIST : -EBUSY;
7757 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
7758 &changeupper_info.info);
7759 ret = notifier_to_errno(ret);
7763 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv,
7768 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
7769 &changeupper_info.info);
7770 ret = notifier_to_errno(ret);
7774 __netdev_update_upper_level(dev, NULL);
7775 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL);
7777 __netdev_update_lower_level(upper_dev, priv);
7778 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level,
7784 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
7790 * netdev_upper_dev_link - Add a link to the upper device
7792 * @upper_dev: new upper device
7793 * @extack: netlink extended ack
7795 * Adds a link to device which is upper to this one. The caller must hold
7796 * the RTNL lock. On a failure a negative errno code is returned.
7797 * On success the reference counts are adjusted and the function
7800 int netdev_upper_dev_link(struct net_device *dev,
7801 struct net_device *upper_dev,
7802 struct netlink_ext_ack *extack)
7804 struct netdev_nested_priv priv = {
7805 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO,
7809 return __netdev_upper_dev_link(dev, upper_dev, false,
7810 NULL, NULL, &priv, extack);
7812 EXPORT_SYMBOL(netdev_upper_dev_link);
7815 * netdev_master_upper_dev_link - Add a master link to the upper device
7817 * @upper_dev: new upper device
7818 * @upper_priv: upper device private
7819 * @upper_info: upper info to be passed down via notifier
7820 * @extack: netlink extended ack
7822 * Adds a link to device which is upper to this one. In this case, only
7823 * one master upper device can be linked, although other non-master devices
7824 * might be linked as well. The caller must hold the RTNL lock.
7825 * On a failure a negative errno code is returned. On success the reference
7826 * counts are adjusted and the function returns zero.
7828 int netdev_master_upper_dev_link(struct net_device *dev,
7829 struct net_device *upper_dev,
7830 void *upper_priv, void *upper_info,
7831 struct netlink_ext_ack *extack)
7833 struct netdev_nested_priv priv = {
7834 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO,
7838 return __netdev_upper_dev_link(dev, upper_dev, true,
7839 upper_priv, upper_info, &priv, extack);
7841 EXPORT_SYMBOL(netdev_master_upper_dev_link);
7843 static void __netdev_upper_dev_unlink(struct net_device *dev,
7844 struct net_device *upper_dev,
7845 struct netdev_nested_priv *priv)
7847 struct netdev_notifier_changeupper_info changeupper_info = {
7851 .upper_dev = upper_dev,
7857 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev;
7859 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
7860 &changeupper_info.info);
7862 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
7864 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
7865 &changeupper_info.info);
7867 __netdev_update_upper_level(dev, NULL);
7868 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL);
7870 __netdev_update_lower_level(upper_dev, priv);
7871 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level,
7876 * netdev_upper_dev_unlink - Removes a link to upper device
7878 * @upper_dev: new upper device
7880 * Removes a link to device which is upper to this one. The caller must hold
7883 void netdev_upper_dev_unlink(struct net_device *dev,
7884 struct net_device *upper_dev)
7886 struct netdev_nested_priv priv = {
7887 .flags = NESTED_SYNC_TODO,
7891 __netdev_upper_dev_unlink(dev, upper_dev, &priv);
7893 EXPORT_SYMBOL(netdev_upper_dev_unlink);
7895 static void __netdev_adjacent_dev_set(struct net_device *upper_dev,
7896 struct net_device *lower_dev,
7899 struct netdev_adjacent *adj;
7901 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower);
7905 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper);
7910 static void netdev_adjacent_dev_disable(struct net_device *upper_dev,
7911 struct net_device *lower_dev)
7913 __netdev_adjacent_dev_set(upper_dev, lower_dev, true);
7916 static void netdev_adjacent_dev_enable(struct net_device *upper_dev,
7917 struct net_device *lower_dev)
7919 __netdev_adjacent_dev_set(upper_dev, lower_dev, false);
7922 int netdev_adjacent_change_prepare(struct net_device *old_dev,
7923 struct net_device *new_dev,
7924 struct net_device *dev,
7925 struct netlink_ext_ack *extack)
7927 struct netdev_nested_priv priv = {
7936 if (old_dev && new_dev != old_dev)
7937 netdev_adjacent_dev_disable(dev, old_dev);
7938 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv,
7941 if (old_dev && new_dev != old_dev)
7942 netdev_adjacent_dev_enable(dev, old_dev);
7948 EXPORT_SYMBOL(netdev_adjacent_change_prepare);
7950 void netdev_adjacent_change_commit(struct net_device *old_dev,
7951 struct net_device *new_dev,
7952 struct net_device *dev)
7954 struct netdev_nested_priv priv = {
7955 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO,
7959 if (!new_dev || !old_dev)
7962 if (new_dev == old_dev)
7965 netdev_adjacent_dev_enable(dev, old_dev);
7966 __netdev_upper_dev_unlink(old_dev, dev, &priv);
7968 EXPORT_SYMBOL(netdev_adjacent_change_commit);
7970 void netdev_adjacent_change_abort(struct net_device *old_dev,
7971 struct net_device *new_dev,
7972 struct net_device *dev)
7974 struct netdev_nested_priv priv = {
7982 if (old_dev && new_dev != old_dev)
7983 netdev_adjacent_dev_enable(dev, old_dev);
7985 __netdev_upper_dev_unlink(new_dev, dev, &priv);
7987 EXPORT_SYMBOL(netdev_adjacent_change_abort);
7990 * netdev_bonding_info_change - Dispatch event about slave change
7992 * @bonding_info: info to dispatch
7994 * Send NETDEV_BONDING_INFO to netdev notifiers with info.
7995 * The caller must hold the RTNL lock.
7997 void netdev_bonding_info_change(struct net_device *dev,
7998 struct netdev_bonding_info *bonding_info)
8000 struct netdev_notifier_bonding_info info = {
8004 memcpy(&info.bonding_info, bonding_info,
8005 sizeof(struct netdev_bonding_info));
8006 call_netdevice_notifiers_info(NETDEV_BONDING_INFO,
8009 EXPORT_SYMBOL(netdev_bonding_info_change);
8011 static int netdev_offload_xstats_enable_l3(struct net_device *dev,
8012 struct netlink_ext_ack *extack)
8014 struct netdev_notifier_offload_xstats_info info = {
8016 .info.extack = extack,
8017 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3,
8022 dev->offload_xstats_l3 = kzalloc(sizeof(*dev->offload_xstats_l3),
8024 if (!dev->offload_xstats_l3)
8027 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE,
8028 NETDEV_OFFLOAD_XSTATS_DISABLE,
8030 err = notifier_to_errno(rc);
8037 kfree(dev->offload_xstats_l3);
8038 dev->offload_xstats_l3 = NULL;
8042 int netdev_offload_xstats_enable(struct net_device *dev,
8043 enum netdev_offload_xstats_type type,
8044 struct netlink_ext_ack *extack)
8048 if (netdev_offload_xstats_enabled(dev, type))
8052 case NETDEV_OFFLOAD_XSTATS_TYPE_L3:
8053 return netdev_offload_xstats_enable_l3(dev, extack);
8059 EXPORT_SYMBOL(netdev_offload_xstats_enable);
8061 static void netdev_offload_xstats_disable_l3(struct net_device *dev)
8063 struct netdev_notifier_offload_xstats_info info = {
8065 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3,
8068 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE,
8070 kfree(dev->offload_xstats_l3);
8071 dev->offload_xstats_l3 = NULL;
8074 int netdev_offload_xstats_disable(struct net_device *dev,
8075 enum netdev_offload_xstats_type type)
8079 if (!netdev_offload_xstats_enabled(dev, type))
8083 case NETDEV_OFFLOAD_XSTATS_TYPE_L3:
8084 netdev_offload_xstats_disable_l3(dev);
8091 EXPORT_SYMBOL(netdev_offload_xstats_disable);
8093 static void netdev_offload_xstats_disable_all(struct net_device *dev)
8095 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3);
8098 static struct rtnl_hw_stats64 *
8099 netdev_offload_xstats_get_ptr(const struct net_device *dev,
8100 enum netdev_offload_xstats_type type)
8103 case NETDEV_OFFLOAD_XSTATS_TYPE_L3:
8104 return dev->offload_xstats_l3;
8111 bool netdev_offload_xstats_enabled(const struct net_device *dev,
8112 enum netdev_offload_xstats_type type)
8116 return netdev_offload_xstats_get_ptr(dev, type);
8118 EXPORT_SYMBOL(netdev_offload_xstats_enabled);
8120 struct netdev_notifier_offload_xstats_ru {
8124 struct netdev_notifier_offload_xstats_rd {
8125 struct rtnl_hw_stats64 stats;
8129 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest,
8130 const struct rtnl_hw_stats64 *src)
8132 dest->rx_packets += src->rx_packets;
8133 dest->tx_packets += src->tx_packets;
8134 dest->rx_bytes += src->rx_bytes;
8135 dest->tx_bytes += src->tx_bytes;
8136 dest->rx_errors += src->rx_errors;
8137 dest->tx_errors += src->tx_errors;
8138 dest->rx_dropped += src->rx_dropped;
8139 dest->tx_dropped += src->tx_dropped;
8140 dest->multicast += src->multicast;
8143 static int netdev_offload_xstats_get_used(struct net_device *dev,
8144 enum netdev_offload_xstats_type type,
8146 struct netlink_ext_ack *extack)
8148 struct netdev_notifier_offload_xstats_ru report_used = {};
8149 struct netdev_notifier_offload_xstats_info info = {
8151 .info.extack = extack,
8153 .report_used = &report_used,
8157 WARN_ON(!netdev_offload_xstats_enabled(dev, type));
8158 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED,
8160 *p_used = report_used.used;
8161 return notifier_to_errno(rc);
8164 static int netdev_offload_xstats_get_stats(struct net_device *dev,
8165 enum netdev_offload_xstats_type type,
8166 struct rtnl_hw_stats64 *p_stats,
8168 struct netlink_ext_ack *extack)
8170 struct netdev_notifier_offload_xstats_rd report_delta = {};
8171 struct netdev_notifier_offload_xstats_info info = {
8173 .info.extack = extack,
8175 .report_delta = &report_delta,
8177 struct rtnl_hw_stats64 *stats;
8180 stats = netdev_offload_xstats_get_ptr(dev, type);
8181 if (WARN_ON(!stats))
8184 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA,
8187 /* Cache whatever we got, even if there was an error, otherwise the
8188 * successful stats retrievals would get lost.
8190 netdev_hw_stats64_add(stats, &report_delta.stats);
8194 *p_used = report_delta.used;
8196 return notifier_to_errno(rc);
8199 int netdev_offload_xstats_get(struct net_device *dev,
8200 enum netdev_offload_xstats_type type,
8201 struct rtnl_hw_stats64 *p_stats, bool *p_used,
8202 struct netlink_ext_ack *extack)
8207 return netdev_offload_xstats_get_stats(dev, type, p_stats,
8210 return netdev_offload_xstats_get_used(dev, type, p_used,
8213 EXPORT_SYMBOL(netdev_offload_xstats_get);
8216 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta,
8217 const struct rtnl_hw_stats64 *stats)
8219 report_delta->used = true;
8220 netdev_hw_stats64_add(&report_delta->stats, stats);
8222 EXPORT_SYMBOL(netdev_offload_xstats_report_delta);
8225 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used)
8227 report_used->used = true;
8229 EXPORT_SYMBOL(netdev_offload_xstats_report_used);
8231 void netdev_offload_xstats_push_delta(struct net_device *dev,
8232 enum netdev_offload_xstats_type type,
8233 const struct rtnl_hw_stats64 *p_stats)
8235 struct rtnl_hw_stats64 *stats;
8239 stats = netdev_offload_xstats_get_ptr(dev, type);
8240 if (WARN_ON(!stats))
8243 netdev_hw_stats64_add(stats, p_stats);
8245 EXPORT_SYMBOL(netdev_offload_xstats_push_delta);
8248 * netdev_get_xmit_slave - Get the xmit slave of master device
8251 * @all_slaves: assume all the slaves are active
8253 * The reference counters are not incremented so the caller must be
8254 * careful with locks. The caller must hold RCU lock.
8255 * %NULL is returned if no slave is found.
8258 struct net_device *netdev_get_xmit_slave(struct net_device *dev,
8259 struct sk_buff *skb,
8262 const struct net_device_ops *ops = dev->netdev_ops;
8264 if (!ops->ndo_get_xmit_slave)
8266 return ops->ndo_get_xmit_slave(dev, skb, all_slaves);
8268 EXPORT_SYMBOL(netdev_get_xmit_slave);
8270 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev,
8273 const struct net_device_ops *ops = dev->netdev_ops;
8275 if (!ops->ndo_sk_get_lower_dev)
8277 return ops->ndo_sk_get_lower_dev(dev, sk);
8281 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket
8285 * %NULL is returned if no lower device is found.
8288 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev,
8291 struct net_device *lower;
8293 lower = netdev_sk_get_lower_dev(dev, sk);
8296 lower = netdev_sk_get_lower_dev(dev, sk);
8301 EXPORT_SYMBOL(netdev_sk_get_lowest_dev);
8303 static void netdev_adjacent_add_links(struct net_device *dev)
8305 struct netdev_adjacent *iter;
8307 struct net *net = dev_net(dev);
8309 list_for_each_entry(iter, &dev->adj_list.upper, list) {
8310 if (!net_eq(net, dev_net(iter->dev)))
8312 netdev_adjacent_sysfs_add(iter->dev, dev,
8313 &iter->dev->adj_list.lower);
8314 netdev_adjacent_sysfs_add(dev, iter->dev,
8315 &dev->adj_list.upper);
8318 list_for_each_entry(iter, &dev->adj_list.lower, list) {
8319 if (!net_eq(net, dev_net(iter->dev)))
8321 netdev_adjacent_sysfs_add(iter->dev, dev,
8322 &iter->dev->adj_list.upper);
8323 netdev_adjacent_sysfs_add(dev, iter->dev,
8324 &dev->adj_list.lower);
8328 static void netdev_adjacent_del_links(struct net_device *dev)
8330 struct netdev_adjacent *iter;
8332 struct net *net = dev_net(dev);
8334 list_for_each_entry(iter, &dev->adj_list.upper, list) {
8335 if (!net_eq(net, dev_net(iter->dev)))
8337 netdev_adjacent_sysfs_del(iter->dev, dev->name,
8338 &iter->dev->adj_list.lower);
8339 netdev_adjacent_sysfs_del(dev, iter->dev->name,
8340 &dev->adj_list.upper);
8343 list_for_each_entry(iter, &dev->adj_list.lower, list) {
8344 if (!net_eq(net, dev_net(iter->dev)))
8346 netdev_adjacent_sysfs_del(iter->dev, dev->name,
8347 &iter->dev->adj_list.upper);
8348 netdev_adjacent_sysfs_del(dev, iter->dev->name,
8349 &dev->adj_list.lower);
8353 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname)
8355 struct netdev_adjacent *iter;
8357 struct net *net = dev_net(dev);
8359 list_for_each_entry(iter, &dev->adj_list.upper, list) {
8360 if (!net_eq(net, dev_net(iter->dev)))
8362 netdev_adjacent_sysfs_del(iter->dev, oldname,
8363 &iter->dev->adj_list.lower);
8364 netdev_adjacent_sysfs_add(iter->dev, dev,
8365 &iter->dev->adj_list.lower);
8368 list_for_each_entry(iter, &dev->adj_list.lower, list) {
8369 if (!net_eq(net, dev_net(iter->dev)))
8371 netdev_adjacent_sysfs_del(iter->dev, oldname,
8372 &iter->dev->adj_list.upper);
8373 netdev_adjacent_sysfs_add(iter->dev, dev,
8374 &iter->dev->adj_list.upper);
8378 void *netdev_lower_dev_get_private(struct net_device *dev,
8379 struct net_device *lower_dev)
8381 struct netdev_adjacent *lower;
8385 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower);
8389 return lower->private;
8391 EXPORT_SYMBOL(netdev_lower_dev_get_private);
8395 * netdev_lower_state_changed - Dispatch event about lower device state change
8396 * @lower_dev: device
8397 * @lower_state_info: state to dispatch
8399 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info.
8400 * The caller must hold the RTNL lock.
8402 void netdev_lower_state_changed(struct net_device *lower_dev,
8403 void *lower_state_info)
8405 struct netdev_notifier_changelowerstate_info changelowerstate_info = {
8406 .info.dev = lower_dev,
8410 changelowerstate_info.lower_state_info = lower_state_info;
8411 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE,
8412 &changelowerstate_info.info);
8414 EXPORT_SYMBOL(netdev_lower_state_changed);
8416 static void dev_change_rx_flags(struct net_device *dev, int flags)
8418 const struct net_device_ops *ops = dev->netdev_ops;
8420 if (ops->ndo_change_rx_flags)
8421 ops->ndo_change_rx_flags(dev, flags);
8424 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify)
8426 unsigned int old_flags = dev->flags;
8432 dev->flags |= IFF_PROMISC;
8433 dev->promiscuity += inc;
8434 if (dev->promiscuity == 0) {
8437 * If inc causes overflow, untouch promisc and return error.
8440 dev->flags &= ~IFF_PROMISC;
8442 dev->promiscuity -= inc;
8443 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n");
8447 if (dev->flags != old_flags) {
8448 netdev_info(dev, "%s promiscuous mode\n",
8449 dev->flags & IFF_PROMISC ? "entered" : "left");
8450 if (audit_enabled) {
8451 current_uid_gid(&uid, &gid);
8452 audit_log(audit_context(), GFP_ATOMIC,
8453 AUDIT_ANOM_PROMISCUOUS,
8454 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u",
8455 dev->name, (dev->flags & IFF_PROMISC),
8456 (old_flags & IFF_PROMISC),
8457 from_kuid(&init_user_ns, audit_get_loginuid(current)),
8458 from_kuid(&init_user_ns, uid),
8459 from_kgid(&init_user_ns, gid),
8460 audit_get_sessionid(current));
8463 dev_change_rx_flags(dev, IFF_PROMISC);
8466 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL);
8471 * dev_set_promiscuity - update promiscuity count on a device
8475 * Add or remove promiscuity from a device. While the count in the device
8476 * remains above zero the interface remains promiscuous. Once it hits zero
8477 * the device reverts back to normal filtering operation. A negative inc
8478 * value is used to drop promiscuity on the device.
8479 * Return 0 if successful or a negative errno code on error.
8481 int dev_set_promiscuity(struct net_device *dev, int inc)
8483 unsigned int old_flags = dev->flags;
8486 err = __dev_set_promiscuity(dev, inc, true);
8489 if (dev->flags != old_flags)
8490 dev_set_rx_mode(dev);
8493 EXPORT_SYMBOL(dev_set_promiscuity);
8495 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify)
8497 unsigned int old_flags = dev->flags, old_gflags = dev->gflags;
8501 dev->flags |= IFF_ALLMULTI;
8502 dev->allmulti += inc;
8503 if (dev->allmulti == 0) {
8506 * If inc causes overflow, untouch allmulti and return error.
8509 dev->flags &= ~IFF_ALLMULTI;
8511 dev->allmulti -= inc;
8512 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n");
8516 if (dev->flags ^ old_flags) {
8517 netdev_info(dev, "%s allmulticast mode\n",
8518 dev->flags & IFF_ALLMULTI ? "entered" : "left");
8519 dev_change_rx_flags(dev, IFF_ALLMULTI);
8520 dev_set_rx_mode(dev);
8522 __dev_notify_flags(dev, old_flags,
8523 dev->gflags ^ old_gflags, 0, NULL);
8529 * dev_set_allmulti - update allmulti count on a device
8533 * Add or remove reception of all multicast frames to a device. While the
8534 * count in the device remains above zero the interface remains listening
8535 * to all interfaces. Once it hits zero the device reverts back to normal
8536 * filtering operation. A negative @inc value is used to drop the counter
8537 * when releasing a resource needing all multicasts.
8538 * Return 0 if successful or a negative errno code on error.
8541 int dev_set_allmulti(struct net_device *dev, int inc)
8543 return __dev_set_allmulti(dev, inc, true);
8545 EXPORT_SYMBOL(dev_set_allmulti);
8548 * Upload unicast and multicast address lists to device and
8549 * configure RX filtering. When the device doesn't support unicast
8550 * filtering it is put in promiscuous mode while unicast addresses
8553 void __dev_set_rx_mode(struct net_device *dev)
8555 const struct net_device_ops *ops = dev->netdev_ops;
8557 /* dev_open will call this function so the list will stay sane. */
8558 if (!(dev->flags&IFF_UP))
8561 if (!netif_device_present(dev))
8564 if (!(dev->priv_flags & IFF_UNICAST_FLT)) {
8565 /* Unicast addresses changes may only happen under the rtnl,
8566 * therefore calling __dev_set_promiscuity here is safe.
8568 if (!netdev_uc_empty(dev) && !dev->uc_promisc) {
8569 __dev_set_promiscuity(dev, 1, false);
8570 dev->uc_promisc = true;
8571 } else if (netdev_uc_empty(dev) && dev->uc_promisc) {
8572 __dev_set_promiscuity(dev, -1, false);
8573 dev->uc_promisc = false;
8577 if (ops->ndo_set_rx_mode)
8578 ops->ndo_set_rx_mode(dev);
8581 void dev_set_rx_mode(struct net_device *dev)
8583 netif_addr_lock_bh(dev);
8584 __dev_set_rx_mode(dev);
8585 netif_addr_unlock_bh(dev);
8589 * dev_get_flags - get flags reported to userspace
8592 * Get the combination of flag bits exported through APIs to userspace.
8594 unsigned int dev_get_flags(const struct net_device *dev)
8598 flags = (dev->flags & ~(IFF_PROMISC |
8603 (dev->gflags & (IFF_PROMISC |
8606 if (netif_running(dev)) {
8607 if (netif_oper_up(dev))
8608 flags |= IFF_RUNNING;
8609 if (netif_carrier_ok(dev))
8610 flags |= IFF_LOWER_UP;
8611 if (netif_dormant(dev))
8612 flags |= IFF_DORMANT;
8617 EXPORT_SYMBOL(dev_get_flags);
8619 int __dev_change_flags(struct net_device *dev, unsigned int flags,
8620 struct netlink_ext_ack *extack)
8622 unsigned int old_flags = dev->flags;
8628 * Set the flags on our device.
8631 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP |
8632 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL |
8634 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC |
8638 * Load in the correct multicast list now the flags have changed.
8641 if ((old_flags ^ flags) & IFF_MULTICAST)
8642 dev_change_rx_flags(dev, IFF_MULTICAST);
8644 dev_set_rx_mode(dev);
8647 * Have we downed the interface. We handle IFF_UP ourselves
8648 * according to user attempts to set it, rather than blindly
8653 if ((old_flags ^ flags) & IFF_UP) {
8654 if (old_flags & IFF_UP)
8657 ret = __dev_open(dev, extack);
8660 if ((flags ^ dev->gflags) & IFF_PROMISC) {
8661 int inc = (flags & IFF_PROMISC) ? 1 : -1;
8662 unsigned int old_flags = dev->flags;
8664 dev->gflags ^= IFF_PROMISC;
8666 if (__dev_set_promiscuity(dev, inc, false) >= 0)
8667 if (dev->flags != old_flags)
8668 dev_set_rx_mode(dev);
8671 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI
8672 * is important. Some (broken) drivers set IFF_PROMISC, when
8673 * IFF_ALLMULTI is requested not asking us and not reporting.
8675 if ((flags ^ dev->gflags) & IFF_ALLMULTI) {
8676 int inc = (flags & IFF_ALLMULTI) ? 1 : -1;
8678 dev->gflags ^= IFF_ALLMULTI;
8679 __dev_set_allmulti(dev, inc, false);
8685 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags,
8686 unsigned int gchanges, u32 portid,
8687 const struct nlmsghdr *nlh)
8689 unsigned int changes = dev->flags ^ old_flags;
8692 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh);
8694 if (changes & IFF_UP) {
8695 if (dev->flags & IFF_UP)
8696 call_netdevice_notifiers(NETDEV_UP, dev);
8698 call_netdevice_notifiers(NETDEV_DOWN, dev);
8701 if (dev->flags & IFF_UP &&
8702 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) {
8703 struct netdev_notifier_change_info change_info = {
8707 .flags_changed = changes,
8710 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info);
8715 * dev_change_flags - change device settings
8717 * @flags: device state flags
8718 * @extack: netlink extended ack
8720 * Change settings on device based state flags. The flags are
8721 * in the userspace exported format.
8723 int dev_change_flags(struct net_device *dev, unsigned int flags,
8724 struct netlink_ext_ack *extack)
8727 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags;
8729 ret = __dev_change_flags(dev, flags, extack);
8733 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags);
8734 __dev_notify_flags(dev, old_flags, changes, 0, NULL);
8737 EXPORT_SYMBOL(dev_change_flags);
8739 int __dev_set_mtu(struct net_device *dev, int new_mtu)
8741 const struct net_device_ops *ops = dev->netdev_ops;
8743 if (ops->ndo_change_mtu)
8744 return ops->ndo_change_mtu(dev, new_mtu);
8746 /* Pairs with all the lockless reads of dev->mtu in the stack */
8747 WRITE_ONCE(dev->mtu, new_mtu);
8750 EXPORT_SYMBOL(__dev_set_mtu);
8752 int dev_validate_mtu(struct net_device *dev, int new_mtu,
8753 struct netlink_ext_ack *extack)
8755 /* MTU must be positive, and in range */
8756 if (new_mtu < 0 || new_mtu < dev->min_mtu) {
8757 NL_SET_ERR_MSG(extack, "mtu less than device minimum");
8761 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) {
8762 NL_SET_ERR_MSG(extack, "mtu greater than device maximum");
8769 * dev_set_mtu_ext - Change maximum transfer unit
8771 * @new_mtu: new transfer unit
8772 * @extack: netlink extended ack
8774 * Change the maximum transfer size of the network device.
8776 int dev_set_mtu_ext(struct net_device *dev, int new_mtu,
8777 struct netlink_ext_ack *extack)
8781 if (new_mtu == dev->mtu)
8784 err = dev_validate_mtu(dev, new_mtu, extack);
8788 if (!netif_device_present(dev))
8791 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev);
8792 err = notifier_to_errno(err);
8796 orig_mtu = dev->mtu;
8797 err = __dev_set_mtu(dev, new_mtu);
8800 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev,
8802 err = notifier_to_errno(err);
8804 /* setting mtu back and notifying everyone again,
8805 * so that they have a chance to revert changes.
8807 __dev_set_mtu(dev, orig_mtu);
8808 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev,
8815 int dev_set_mtu(struct net_device *dev, int new_mtu)
8817 struct netlink_ext_ack extack;
8820 memset(&extack, 0, sizeof(extack));
8821 err = dev_set_mtu_ext(dev, new_mtu, &extack);
8822 if (err && extack._msg)
8823 net_err_ratelimited("%s: %s\n", dev->name, extack._msg);
8826 EXPORT_SYMBOL(dev_set_mtu);
8829 * dev_change_tx_queue_len - Change TX queue length of a netdevice
8831 * @new_len: new tx queue length
8833 int dev_change_tx_queue_len(struct net_device *dev, unsigned long new_len)
8835 unsigned int orig_len = dev->tx_queue_len;
8838 if (new_len != (unsigned int)new_len)
8841 if (new_len != orig_len) {
8842 dev->tx_queue_len = new_len;
8843 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev);
8844 res = notifier_to_errno(res);
8847 res = dev_qdisc_change_tx_queue_len(dev);
8855 netdev_err(dev, "refused to change device tx_queue_len\n");
8856 dev->tx_queue_len = orig_len;
8861 * dev_set_group - Change group this device belongs to
8863 * @new_group: group this device should belong to
8865 void dev_set_group(struct net_device *dev, int new_group)
8867 dev->group = new_group;
8871 * dev_pre_changeaddr_notify - Call NETDEV_PRE_CHANGEADDR.
8873 * @addr: new address
8874 * @extack: netlink extended ack
8876 int dev_pre_changeaddr_notify(struct net_device *dev, const char *addr,
8877 struct netlink_ext_ack *extack)
8879 struct netdev_notifier_pre_changeaddr_info info = {
8881 .info.extack = extack,
8886 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info);
8887 return notifier_to_errno(rc);
8889 EXPORT_SYMBOL(dev_pre_changeaddr_notify);
8892 * dev_set_mac_address - Change Media Access Control Address
8895 * @extack: netlink extended ack
8897 * Change the hardware (MAC) address of the device
8899 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa,
8900 struct netlink_ext_ack *extack)
8902 const struct net_device_ops *ops = dev->netdev_ops;
8905 if (!ops->ndo_set_mac_address)
8907 if (sa->sa_family != dev->type)
8909 if (!netif_device_present(dev))
8911 err = dev_pre_changeaddr_notify(dev, sa->sa_data, extack);
8914 if (memcmp(dev->dev_addr, sa->sa_data, dev->addr_len)) {
8915 err = ops->ndo_set_mac_address(dev, sa);
8919 dev->addr_assign_type = NET_ADDR_SET;
8920 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
8921 add_device_randomness(dev->dev_addr, dev->addr_len);
8924 EXPORT_SYMBOL(dev_set_mac_address);
8926 static DECLARE_RWSEM(dev_addr_sem);
8928 int dev_set_mac_address_user(struct net_device *dev, struct sockaddr *sa,
8929 struct netlink_ext_ack *extack)
8933 down_write(&dev_addr_sem);
8934 ret = dev_set_mac_address(dev, sa, extack);
8935 up_write(&dev_addr_sem);
8938 EXPORT_SYMBOL(dev_set_mac_address_user);
8940 int dev_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name)
8942 size_t size = sizeof(sa->sa_data_min);
8943 struct net_device *dev;
8946 down_read(&dev_addr_sem);
8949 dev = dev_get_by_name_rcu(net, dev_name);
8955 memset(sa->sa_data, 0, size);
8957 memcpy(sa->sa_data, dev->dev_addr,
8958 min_t(size_t, size, dev->addr_len));
8959 sa->sa_family = dev->type;
8963 up_read(&dev_addr_sem);
8966 EXPORT_SYMBOL(dev_get_mac_address);
8969 * dev_change_carrier - Change device carrier
8971 * @new_carrier: new value
8973 * Change device carrier
8975 int dev_change_carrier(struct net_device *dev, bool new_carrier)
8977 const struct net_device_ops *ops = dev->netdev_ops;
8979 if (!ops->ndo_change_carrier)
8981 if (!netif_device_present(dev))
8983 return ops->ndo_change_carrier(dev, new_carrier);
8987 * dev_get_phys_port_id - Get device physical port ID
8991 * Get device physical port ID
8993 int dev_get_phys_port_id(struct net_device *dev,
8994 struct netdev_phys_item_id *ppid)
8996 const struct net_device_ops *ops = dev->netdev_ops;
8998 if (!ops->ndo_get_phys_port_id)
9000 return ops->ndo_get_phys_port_id(dev, ppid);
9004 * dev_get_phys_port_name - Get device physical port name
9007 * @len: limit of bytes to copy to name
9009 * Get device physical port name
9011 int dev_get_phys_port_name(struct net_device *dev,
9012 char *name, size_t len)
9014 const struct net_device_ops *ops = dev->netdev_ops;
9017 if (ops->ndo_get_phys_port_name) {
9018 err = ops->ndo_get_phys_port_name(dev, name, len);
9019 if (err != -EOPNOTSUPP)
9022 return devlink_compat_phys_port_name_get(dev, name, len);
9026 * dev_get_port_parent_id - Get the device's port parent identifier
9027 * @dev: network device
9028 * @ppid: pointer to a storage for the port's parent identifier
9029 * @recurse: allow/disallow recursion to lower devices
9031 * Get the devices's port parent identifier
9033 int dev_get_port_parent_id(struct net_device *dev,
9034 struct netdev_phys_item_id *ppid,
9037 const struct net_device_ops *ops = dev->netdev_ops;
9038 struct netdev_phys_item_id first = { };
9039 struct net_device *lower_dev;
9040 struct list_head *iter;
9043 if (ops->ndo_get_port_parent_id) {
9044 err = ops->ndo_get_port_parent_id(dev, ppid);
9045 if (err != -EOPNOTSUPP)
9049 err = devlink_compat_switch_id_get(dev, ppid);
9050 if (!recurse || err != -EOPNOTSUPP)
9053 netdev_for_each_lower_dev(dev, lower_dev, iter) {
9054 err = dev_get_port_parent_id(lower_dev, ppid, true);
9059 else if (memcmp(&first, ppid, sizeof(*ppid)))
9065 EXPORT_SYMBOL(dev_get_port_parent_id);
9068 * netdev_port_same_parent_id - Indicate if two network devices have
9069 * the same port parent identifier
9070 * @a: first network device
9071 * @b: second network device
9073 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b)
9075 struct netdev_phys_item_id a_id = { };
9076 struct netdev_phys_item_id b_id = { };
9078 if (dev_get_port_parent_id(a, &a_id, true) ||
9079 dev_get_port_parent_id(b, &b_id, true))
9082 return netdev_phys_item_id_same(&a_id, &b_id);
9084 EXPORT_SYMBOL(netdev_port_same_parent_id);
9087 * dev_change_proto_down - set carrier according to proto_down.
9090 * @proto_down: new value
9092 int dev_change_proto_down(struct net_device *dev, bool proto_down)
9094 if (!(dev->priv_flags & IFF_CHANGE_PROTO_DOWN))
9096 if (!netif_device_present(dev))
9099 netif_carrier_off(dev);
9101 netif_carrier_on(dev);
9102 dev->proto_down = proto_down;
9107 * dev_change_proto_down_reason - proto down reason
9110 * @mask: proto down mask
9111 * @value: proto down value
9113 void dev_change_proto_down_reason(struct net_device *dev, unsigned long mask,
9119 dev->proto_down_reason = value;
9121 for_each_set_bit(b, &mask, 32) {
9122 if (value & (1 << b))
9123 dev->proto_down_reason |= BIT(b);
9125 dev->proto_down_reason &= ~BIT(b);
9130 struct bpf_xdp_link {
9131 struct bpf_link link;
9132 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */
9136 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags)
9138 if (flags & XDP_FLAGS_HW_MODE)
9140 if (flags & XDP_FLAGS_DRV_MODE)
9141 return XDP_MODE_DRV;
9142 if (flags & XDP_FLAGS_SKB_MODE)
9143 return XDP_MODE_SKB;
9144 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB;
9147 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode)
9151 return generic_xdp_install;
9154 return dev->netdev_ops->ndo_bpf;
9160 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev,
9161 enum bpf_xdp_mode mode)
9163 return dev->xdp_state[mode].link;
9166 static struct bpf_prog *dev_xdp_prog(struct net_device *dev,
9167 enum bpf_xdp_mode mode)
9169 struct bpf_xdp_link *link = dev_xdp_link(dev, mode);
9172 return link->link.prog;
9173 return dev->xdp_state[mode].prog;
9176 u8 dev_xdp_prog_count(struct net_device *dev)
9181 for (i = 0; i < __MAX_XDP_MODE; i++)
9182 if (dev->xdp_state[i].prog || dev->xdp_state[i].link)
9186 EXPORT_SYMBOL_GPL(dev_xdp_prog_count);
9188 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode)
9190 struct bpf_prog *prog = dev_xdp_prog(dev, mode);
9192 return prog ? prog->aux->id : 0;
9195 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode,
9196 struct bpf_xdp_link *link)
9198 dev->xdp_state[mode].link = link;
9199 dev->xdp_state[mode].prog = NULL;
9202 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode,
9203 struct bpf_prog *prog)
9205 dev->xdp_state[mode].link = NULL;
9206 dev->xdp_state[mode].prog = prog;
9209 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode,
9210 bpf_op_t bpf_op, struct netlink_ext_ack *extack,
9211 u32 flags, struct bpf_prog *prog)
9213 struct netdev_bpf xdp;
9216 memset(&xdp, 0, sizeof(xdp));
9217 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG;
9218 xdp.extack = extack;
9222 /* Drivers assume refcnt is already incremented (i.e, prog pointer is
9223 * "moved" into driver), so they don't increment it on their own, but
9224 * they do decrement refcnt when program is detached or replaced.
9225 * Given net_device also owns link/prog, we need to bump refcnt here
9226 * to prevent drivers from underflowing it.
9230 err = bpf_op(dev, &xdp);
9237 if (mode != XDP_MODE_HW)
9238 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog);
9243 static void dev_xdp_uninstall(struct net_device *dev)
9245 struct bpf_xdp_link *link;
9246 struct bpf_prog *prog;
9247 enum bpf_xdp_mode mode;
9252 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) {
9253 prog = dev_xdp_prog(dev, mode);
9257 bpf_op = dev_xdp_bpf_op(dev, mode);
9261 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL));
9263 /* auto-detach link from net device */
9264 link = dev_xdp_link(dev, mode);
9270 dev_xdp_set_link(dev, mode, NULL);
9274 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack,
9275 struct bpf_xdp_link *link, struct bpf_prog *new_prog,
9276 struct bpf_prog *old_prog, u32 flags)
9278 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES);
9279 struct bpf_prog *cur_prog;
9280 struct net_device *upper;
9281 struct list_head *iter;
9282 enum bpf_xdp_mode mode;
9288 /* either link or prog attachment, never both */
9289 if (link && (new_prog || old_prog))
9291 /* link supports only XDP mode flags */
9292 if (link && (flags & ~XDP_FLAGS_MODES)) {
9293 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment");
9296 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */
9297 if (num_modes > 1) {
9298 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set");
9301 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */
9302 if (!num_modes && dev_xdp_prog_count(dev) > 1) {
9303 NL_SET_ERR_MSG(extack,
9304 "More than one program loaded, unset mode is ambiguous");
9307 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */
9308 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) {
9309 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified");
9313 mode = dev_xdp_mode(dev, flags);
9314 /* can't replace attached link */
9315 if (dev_xdp_link(dev, mode)) {
9316 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link");
9320 /* don't allow if an upper device already has a program */
9321 netdev_for_each_upper_dev_rcu(dev, upper, iter) {
9322 if (dev_xdp_prog_count(upper) > 0) {
9323 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program");
9328 cur_prog = dev_xdp_prog(dev, mode);
9329 /* can't replace attached prog with link */
9330 if (link && cur_prog) {
9331 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link");
9334 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) {
9335 NL_SET_ERR_MSG(extack, "Active program does not match expected");
9339 /* put effective new program into new_prog */
9341 new_prog = link->link.prog;
9344 bool offload = mode == XDP_MODE_HW;
9345 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB
9346 ? XDP_MODE_DRV : XDP_MODE_SKB;
9348 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) {
9349 NL_SET_ERR_MSG(extack, "XDP program already attached");
9352 if (!offload && dev_xdp_prog(dev, other_mode)) {
9353 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time");
9356 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) {
9357 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported");
9360 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) {
9361 NL_SET_ERR_MSG(extack, "Program bound to different device");
9364 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) {
9365 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device");
9368 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) {
9369 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device");
9374 /* don't call drivers if the effective program didn't change */
9375 if (new_prog != cur_prog) {
9376 bpf_op = dev_xdp_bpf_op(dev, mode);
9378 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode");
9382 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog);
9388 dev_xdp_set_link(dev, mode, link);
9390 dev_xdp_set_prog(dev, mode, new_prog);
9392 bpf_prog_put(cur_prog);
9397 static int dev_xdp_attach_link(struct net_device *dev,
9398 struct netlink_ext_ack *extack,
9399 struct bpf_xdp_link *link)
9401 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags);
9404 static int dev_xdp_detach_link(struct net_device *dev,
9405 struct netlink_ext_ack *extack,
9406 struct bpf_xdp_link *link)
9408 enum bpf_xdp_mode mode;
9413 mode = dev_xdp_mode(dev, link->flags);
9414 if (dev_xdp_link(dev, mode) != link)
9417 bpf_op = dev_xdp_bpf_op(dev, mode);
9418 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL));
9419 dev_xdp_set_link(dev, mode, NULL);
9423 static void bpf_xdp_link_release(struct bpf_link *link)
9425 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9429 /* if racing with net_device's tear down, xdp_link->dev might be
9430 * already NULL, in which case link was already auto-detached
9432 if (xdp_link->dev) {
9433 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link));
9434 xdp_link->dev = NULL;
9440 static int bpf_xdp_link_detach(struct bpf_link *link)
9442 bpf_xdp_link_release(link);
9446 static void bpf_xdp_link_dealloc(struct bpf_link *link)
9448 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9453 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link,
9454 struct seq_file *seq)
9456 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9461 ifindex = xdp_link->dev->ifindex;
9464 seq_printf(seq, "ifindex:\t%u\n", ifindex);
9467 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link,
9468 struct bpf_link_info *info)
9470 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9475 ifindex = xdp_link->dev->ifindex;
9478 info->xdp.ifindex = ifindex;
9482 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog,
9483 struct bpf_prog *old_prog)
9485 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9486 enum bpf_xdp_mode mode;
9492 /* link might have been auto-released already, so fail */
9493 if (!xdp_link->dev) {
9498 if (old_prog && link->prog != old_prog) {
9502 old_prog = link->prog;
9503 if (old_prog->type != new_prog->type ||
9504 old_prog->expected_attach_type != new_prog->expected_attach_type) {
9509 if (old_prog == new_prog) {
9510 /* no-op, don't disturb drivers */
9511 bpf_prog_put(new_prog);
9515 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags);
9516 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode);
9517 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL,
9518 xdp_link->flags, new_prog);
9522 old_prog = xchg(&link->prog, new_prog);
9523 bpf_prog_put(old_prog);
9530 static const struct bpf_link_ops bpf_xdp_link_lops = {
9531 .release = bpf_xdp_link_release,
9532 .dealloc = bpf_xdp_link_dealloc,
9533 .detach = bpf_xdp_link_detach,
9534 .show_fdinfo = bpf_xdp_link_show_fdinfo,
9535 .fill_link_info = bpf_xdp_link_fill_link_info,
9536 .update_prog = bpf_xdp_link_update,
9539 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog)
9541 struct net *net = current->nsproxy->net_ns;
9542 struct bpf_link_primer link_primer;
9543 struct netlink_ext_ack extack = {};
9544 struct bpf_xdp_link *link;
9545 struct net_device *dev;
9549 dev = dev_get_by_index(net, attr->link_create.target_ifindex);
9555 link = kzalloc(sizeof(*link), GFP_USER);
9561 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog);
9563 link->flags = attr->link_create.flags;
9565 err = bpf_link_prime(&link->link, &link_primer);
9571 err = dev_xdp_attach_link(dev, &extack, link);
9576 bpf_link_cleanup(&link_primer);
9577 trace_bpf_xdp_link_attach_failed(extack._msg);
9581 fd = bpf_link_settle(&link_primer);
9582 /* link itself doesn't hold dev's refcnt to not complicate shutdown */
9595 * dev_change_xdp_fd - set or clear a bpf program for a device rx path
9597 * @extack: netlink extended ack
9598 * @fd: new program fd or negative value to clear
9599 * @expected_fd: old program fd that userspace expects to replace or clear
9600 * @flags: xdp-related flags
9602 * Set or clear a bpf program for a device
9604 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack,
9605 int fd, int expected_fd, u32 flags)
9607 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags);
9608 struct bpf_prog *new_prog = NULL, *old_prog = NULL;
9614 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP,
9615 mode != XDP_MODE_SKB);
9616 if (IS_ERR(new_prog))
9617 return PTR_ERR(new_prog);
9620 if (expected_fd >= 0) {
9621 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP,
9622 mode != XDP_MODE_SKB);
9623 if (IS_ERR(old_prog)) {
9624 err = PTR_ERR(old_prog);
9630 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags);
9633 if (err && new_prog)
9634 bpf_prog_put(new_prog);
9636 bpf_prog_put(old_prog);
9641 * dev_index_reserve() - allocate an ifindex in a namespace
9642 * @net: the applicable net namespace
9643 * @ifindex: requested ifindex, pass %0 to get one allocated
9645 * Allocate a ifindex for a new device. Caller must either use the ifindex
9646 * to store the device (via list_netdevice()) or call dev_index_release()
9647 * to give the index up.
9649 * Return: a suitable unique value for a new device interface number or -errno.
9651 static int dev_index_reserve(struct net *net, u32 ifindex)
9655 if (ifindex > INT_MAX) {
9656 DEBUG_NET_WARN_ON_ONCE(1);
9661 err = xa_alloc_cyclic(&net->dev_by_index, &ifindex, NULL,
9662 xa_limit_31b, &net->ifindex, GFP_KERNEL);
9664 err = xa_insert(&net->dev_by_index, ifindex, NULL, GFP_KERNEL);
9671 static void dev_index_release(struct net *net, int ifindex)
9673 /* Expect only unused indexes, unlist_netdevice() removes the used */
9674 WARN_ON(xa_erase(&net->dev_by_index, ifindex));
9677 /* Delayed registration/unregisteration */
9678 LIST_HEAD(net_todo_list);
9679 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq);
9681 static void net_set_todo(struct net_device *dev)
9683 list_add_tail(&dev->todo_list, &net_todo_list);
9684 atomic_inc(&dev_net(dev)->dev_unreg_count);
9687 static netdev_features_t netdev_sync_upper_features(struct net_device *lower,
9688 struct net_device *upper, netdev_features_t features)
9690 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
9691 netdev_features_t feature;
9694 for_each_netdev_feature(upper_disables, feature_bit) {
9695 feature = __NETIF_F_BIT(feature_bit);
9696 if (!(upper->wanted_features & feature)
9697 && (features & feature)) {
9698 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n",
9699 &feature, upper->name);
9700 features &= ~feature;
9707 static void netdev_sync_lower_features(struct net_device *upper,
9708 struct net_device *lower, netdev_features_t features)
9710 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
9711 netdev_features_t feature;
9714 for_each_netdev_feature(upper_disables, feature_bit) {
9715 feature = __NETIF_F_BIT(feature_bit);
9716 if (!(features & feature) && (lower->features & feature)) {
9717 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n",
9718 &feature, lower->name);
9719 lower->wanted_features &= ~feature;
9720 __netdev_update_features(lower);
9722 if (unlikely(lower->features & feature))
9723 netdev_WARN(upper, "failed to disable %pNF on %s!\n",
9724 &feature, lower->name);
9726 netdev_features_change(lower);
9731 static netdev_features_t netdev_fix_features(struct net_device *dev,
9732 netdev_features_t features)
9734 /* Fix illegal checksum combinations */
9735 if ((features & NETIF_F_HW_CSUM) &&
9736 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
9737 netdev_warn(dev, "mixed HW and IP checksum settings.\n");
9738 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
9741 /* TSO requires that SG is present as well. */
9742 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) {
9743 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n");
9744 features &= ~NETIF_F_ALL_TSO;
9747 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) &&
9748 !(features & NETIF_F_IP_CSUM)) {
9749 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n");
9750 features &= ~NETIF_F_TSO;
9751 features &= ~NETIF_F_TSO_ECN;
9754 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) &&
9755 !(features & NETIF_F_IPV6_CSUM)) {
9756 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n");
9757 features &= ~NETIF_F_TSO6;
9760 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */
9761 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO))
9762 features &= ~NETIF_F_TSO_MANGLEID;
9764 /* TSO ECN requires that TSO is present as well. */
9765 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN)
9766 features &= ~NETIF_F_TSO_ECN;
9768 /* Software GSO depends on SG. */
9769 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) {
9770 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n");
9771 features &= ~NETIF_F_GSO;
9774 /* GSO partial features require GSO partial be set */
9775 if ((features & dev->gso_partial_features) &&
9776 !(features & NETIF_F_GSO_PARTIAL)) {
9778 "Dropping partially supported GSO features since no GSO partial.\n");
9779 features &= ~dev->gso_partial_features;
9782 if (!(features & NETIF_F_RXCSUM)) {
9783 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet
9784 * successfully merged by hardware must also have the
9785 * checksum verified by hardware. If the user does not
9786 * want to enable RXCSUM, logically, we should disable GRO_HW.
9788 if (features & NETIF_F_GRO_HW) {
9789 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n");
9790 features &= ~NETIF_F_GRO_HW;
9794 /* LRO/HW-GRO features cannot be combined with RX-FCS */
9795 if (features & NETIF_F_RXFCS) {
9796 if (features & NETIF_F_LRO) {
9797 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n");
9798 features &= ~NETIF_F_LRO;
9801 if (features & NETIF_F_GRO_HW) {
9802 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n");
9803 features &= ~NETIF_F_GRO_HW;
9807 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) {
9808 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n");
9809 features &= ~NETIF_F_LRO;
9812 if (features & NETIF_F_HW_TLS_TX) {
9813 bool ip_csum = (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) ==
9814 (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM);
9815 bool hw_csum = features & NETIF_F_HW_CSUM;
9817 if (!ip_csum && !hw_csum) {
9818 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n");
9819 features &= ~NETIF_F_HW_TLS_TX;
9823 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) {
9824 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n");
9825 features &= ~NETIF_F_HW_TLS_RX;
9831 int __netdev_update_features(struct net_device *dev)
9833 struct net_device *upper, *lower;
9834 netdev_features_t features;
9835 struct list_head *iter;
9840 features = netdev_get_wanted_features(dev);
9842 if (dev->netdev_ops->ndo_fix_features)
9843 features = dev->netdev_ops->ndo_fix_features(dev, features);
9845 /* driver might be less strict about feature dependencies */
9846 features = netdev_fix_features(dev, features);
9848 /* some features can't be enabled if they're off on an upper device */
9849 netdev_for_each_upper_dev_rcu(dev, upper, iter)
9850 features = netdev_sync_upper_features(dev, upper, features);
9852 if (dev->features == features)
9855 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n",
9856 &dev->features, &features);
9858 if (dev->netdev_ops->ndo_set_features)
9859 err = dev->netdev_ops->ndo_set_features(dev, features);
9863 if (unlikely(err < 0)) {
9865 "set_features() failed (%d); wanted %pNF, left %pNF\n",
9866 err, &features, &dev->features);
9867 /* return non-0 since some features might have changed and
9868 * it's better to fire a spurious notification than miss it
9874 /* some features must be disabled on lower devices when disabled
9875 * on an upper device (think: bonding master or bridge)
9877 netdev_for_each_lower_dev(dev, lower, iter)
9878 netdev_sync_lower_features(dev, lower, features);
9881 netdev_features_t diff = features ^ dev->features;
9883 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) {
9884 /* udp_tunnel_{get,drop}_rx_info both need
9885 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the
9886 * device, or they won't do anything.
9887 * Thus we need to update dev->features
9888 * *before* calling udp_tunnel_get_rx_info,
9889 * but *after* calling udp_tunnel_drop_rx_info.
9891 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) {
9892 dev->features = features;
9893 udp_tunnel_get_rx_info(dev);
9895 udp_tunnel_drop_rx_info(dev);
9899 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) {
9900 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) {
9901 dev->features = features;
9902 err |= vlan_get_rx_ctag_filter_info(dev);
9904 vlan_drop_rx_ctag_filter_info(dev);
9908 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) {
9909 if (features & NETIF_F_HW_VLAN_STAG_FILTER) {
9910 dev->features = features;
9911 err |= vlan_get_rx_stag_filter_info(dev);
9913 vlan_drop_rx_stag_filter_info(dev);
9917 dev->features = features;
9920 return err < 0 ? 0 : 1;
9924 * netdev_update_features - recalculate device features
9925 * @dev: the device to check
9927 * Recalculate dev->features set and send notifications if it
9928 * has changed. Should be called after driver or hardware dependent
9929 * conditions might have changed that influence the features.
9931 void netdev_update_features(struct net_device *dev)
9933 if (__netdev_update_features(dev))
9934 netdev_features_change(dev);
9936 EXPORT_SYMBOL(netdev_update_features);
9939 * netdev_change_features - recalculate device features
9940 * @dev: the device to check
9942 * Recalculate dev->features set and send notifications even
9943 * if they have not changed. Should be called instead of
9944 * netdev_update_features() if also dev->vlan_features might
9945 * have changed to allow the changes to be propagated to stacked
9948 void netdev_change_features(struct net_device *dev)
9950 __netdev_update_features(dev);
9951 netdev_features_change(dev);
9953 EXPORT_SYMBOL(netdev_change_features);
9956 * netif_stacked_transfer_operstate - transfer operstate
9957 * @rootdev: the root or lower level device to transfer state from
9958 * @dev: the device to transfer operstate to
9960 * Transfer operational state from root to device. This is normally
9961 * called when a stacking relationship exists between the root
9962 * device and the device(a leaf device).
9964 void netif_stacked_transfer_operstate(const struct net_device *rootdev,
9965 struct net_device *dev)
9967 if (rootdev->operstate == IF_OPER_DORMANT)
9968 netif_dormant_on(dev);
9970 netif_dormant_off(dev);
9972 if (rootdev->operstate == IF_OPER_TESTING)
9973 netif_testing_on(dev);
9975 netif_testing_off(dev);
9977 if (netif_carrier_ok(rootdev))
9978 netif_carrier_on(dev);
9980 netif_carrier_off(dev);
9982 EXPORT_SYMBOL(netif_stacked_transfer_operstate);
9984 static int netif_alloc_rx_queues(struct net_device *dev)
9986 unsigned int i, count = dev->num_rx_queues;
9987 struct netdev_rx_queue *rx;
9988 size_t sz = count * sizeof(*rx);
9993 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL);
9999 for (i = 0; i < count; i++) {
10002 /* XDP RX-queue setup */
10003 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0);
10010 /* Rollback successful reg's and free other resources */
10012 xdp_rxq_info_unreg(&rx[i].xdp_rxq);
10018 static void netif_free_rx_queues(struct net_device *dev)
10020 unsigned int i, count = dev->num_rx_queues;
10022 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */
10026 for (i = 0; i < count; i++)
10027 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq);
10032 static void netdev_init_one_queue(struct net_device *dev,
10033 struct netdev_queue *queue, void *_unused)
10035 /* Initialize queue lock */
10036 spin_lock_init(&queue->_xmit_lock);
10037 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type);
10038 queue->xmit_lock_owner = -1;
10039 netdev_queue_numa_node_write(queue, NUMA_NO_NODE);
10042 dql_init(&queue->dql, HZ);
10046 static void netif_free_tx_queues(struct net_device *dev)
10051 static int netif_alloc_netdev_queues(struct net_device *dev)
10053 unsigned int count = dev->num_tx_queues;
10054 struct netdev_queue *tx;
10055 size_t sz = count * sizeof(*tx);
10057 if (count < 1 || count > 0xffff)
10060 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL);
10066 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL);
10067 spin_lock_init(&dev->tx_global_lock);
10072 void netif_tx_stop_all_queues(struct net_device *dev)
10076 for (i = 0; i < dev->num_tx_queues; i++) {
10077 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
10079 netif_tx_stop_queue(txq);
10082 EXPORT_SYMBOL(netif_tx_stop_all_queues);
10084 static int netdev_do_alloc_pcpu_stats(struct net_device *dev)
10088 /* Drivers implementing ndo_get_peer_dev must support tstat
10089 * accounting, so that skb_do_redirect() can bump the dev's
10090 * RX stats upon network namespace switch.
10092 if (dev->netdev_ops->ndo_get_peer_dev &&
10093 dev->pcpu_stat_type != NETDEV_PCPU_STAT_TSTATS)
10094 return -EOPNOTSUPP;
10096 switch (dev->pcpu_stat_type) {
10097 case NETDEV_PCPU_STAT_NONE:
10099 case NETDEV_PCPU_STAT_LSTATS:
10100 v = dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats);
10102 case NETDEV_PCPU_STAT_TSTATS:
10103 v = dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats);
10105 case NETDEV_PCPU_STAT_DSTATS:
10106 v = dev->dstats = netdev_alloc_pcpu_stats(struct pcpu_dstats);
10112 return v ? 0 : -ENOMEM;
10115 static void netdev_do_free_pcpu_stats(struct net_device *dev)
10117 switch (dev->pcpu_stat_type) {
10118 case NETDEV_PCPU_STAT_NONE:
10120 case NETDEV_PCPU_STAT_LSTATS:
10121 free_percpu(dev->lstats);
10123 case NETDEV_PCPU_STAT_TSTATS:
10124 free_percpu(dev->tstats);
10126 case NETDEV_PCPU_STAT_DSTATS:
10127 free_percpu(dev->dstats);
10133 * register_netdevice() - register a network device
10134 * @dev: device to register
10136 * Take a prepared network device structure and make it externally accessible.
10137 * A %NETDEV_REGISTER message is sent to the netdev notifier chain.
10138 * Callers must hold the rtnl lock - you may want register_netdev()
10141 int register_netdevice(struct net_device *dev)
10144 struct net *net = dev_net(dev);
10146 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE <
10147 NETDEV_FEATURE_COUNT);
10148 BUG_ON(dev_boot_phase);
10153 /* When net_device's are persistent, this will be fatal. */
10154 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
10157 ret = ethtool_check_ops(dev->ethtool_ops);
10161 spin_lock_init(&dev->addr_list_lock);
10162 netdev_set_addr_lockdep_class(dev);
10164 ret = dev_get_valid_name(net, dev, dev->name);
10169 dev->name_node = netdev_name_node_head_alloc(dev);
10170 if (!dev->name_node)
10173 /* Init, if this function is available */
10174 if (dev->netdev_ops->ndo_init) {
10175 ret = dev->netdev_ops->ndo_init(dev);
10179 goto err_free_name;
10183 if (((dev->hw_features | dev->features) &
10184 NETIF_F_HW_VLAN_CTAG_FILTER) &&
10185 (!dev->netdev_ops->ndo_vlan_rx_add_vid ||
10186 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) {
10187 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n");
10192 ret = netdev_do_alloc_pcpu_stats(dev);
10196 ret = dev_index_reserve(net, dev->ifindex);
10198 goto err_free_pcpu;
10199 dev->ifindex = ret;
10201 /* Transfer changeable features to wanted_features and enable
10202 * software offloads (GSO and GRO).
10204 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF);
10205 dev->features |= NETIF_F_SOFT_FEATURES;
10207 if (dev->udp_tunnel_nic_info) {
10208 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT;
10209 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT;
10212 dev->wanted_features = dev->features & dev->hw_features;
10214 if (!(dev->flags & IFF_LOOPBACK))
10215 dev->hw_features |= NETIF_F_NOCACHE_COPY;
10217 /* If IPv4 TCP segmentation offload is supported we should also
10218 * allow the device to enable segmenting the frame with the option
10219 * of ignoring a static IP ID value. This doesn't enable the
10220 * feature itself but allows the user to enable it later.
10222 if (dev->hw_features & NETIF_F_TSO)
10223 dev->hw_features |= NETIF_F_TSO_MANGLEID;
10224 if (dev->vlan_features & NETIF_F_TSO)
10225 dev->vlan_features |= NETIF_F_TSO_MANGLEID;
10226 if (dev->mpls_features & NETIF_F_TSO)
10227 dev->mpls_features |= NETIF_F_TSO_MANGLEID;
10228 if (dev->hw_enc_features & NETIF_F_TSO)
10229 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID;
10231 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices.
10233 dev->vlan_features |= NETIF_F_HIGHDMA;
10235 /* Make NETIF_F_SG inheritable to tunnel devices.
10237 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL;
10239 /* Make NETIF_F_SG inheritable to MPLS.
10241 dev->mpls_features |= NETIF_F_SG;
10243 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev);
10244 ret = notifier_to_errno(ret);
10246 goto err_ifindex_release;
10248 ret = netdev_register_kobject(dev);
10249 write_lock(&dev_base_lock);
10250 dev->reg_state = ret ? NETREG_UNREGISTERED : NETREG_REGISTERED;
10251 write_unlock(&dev_base_lock);
10253 goto err_uninit_notify;
10255 __netdev_update_features(dev);
10258 * Default initial state at registry is that the
10259 * device is present.
10262 set_bit(__LINK_STATE_PRESENT, &dev->state);
10264 linkwatch_init_dev(dev);
10266 dev_init_scheduler(dev);
10268 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL);
10269 list_netdevice(dev);
10271 add_device_randomness(dev->dev_addr, dev->addr_len);
10273 /* If the device has permanent device address, driver should
10274 * set dev_addr and also addr_assign_type should be set to
10275 * NET_ADDR_PERM (default value).
10277 if (dev->addr_assign_type == NET_ADDR_PERM)
10278 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len);
10280 /* Notify protocols, that a new device appeared. */
10281 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
10282 ret = notifier_to_errno(ret);
10284 /* Expect explicit free_netdev() on failure */
10285 dev->needs_free_netdev = false;
10286 unregister_netdevice_queue(dev, NULL);
10290 * Prevent userspace races by waiting until the network
10291 * device is fully setup before sending notifications.
10293 if (!dev->rtnl_link_ops ||
10294 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
10295 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL);
10301 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev);
10302 err_ifindex_release:
10303 dev_index_release(net, dev->ifindex);
10305 netdev_do_free_pcpu_stats(dev);
10307 if (dev->netdev_ops->ndo_uninit)
10308 dev->netdev_ops->ndo_uninit(dev);
10309 if (dev->priv_destructor)
10310 dev->priv_destructor(dev);
10312 netdev_name_node_free(dev->name_node);
10315 EXPORT_SYMBOL(register_netdevice);
10318 * init_dummy_netdev - init a dummy network device for NAPI
10319 * @dev: device to init
10321 * This takes a network device structure and initialize the minimum
10322 * amount of fields so it can be used to schedule NAPI polls without
10323 * registering a full blown interface. This is to be used by drivers
10324 * that need to tie several hardware interfaces to a single NAPI
10325 * poll scheduler due to HW limitations.
10327 int init_dummy_netdev(struct net_device *dev)
10329 /* Clear everything. Note we don't initialize spinlocks
10330 * are they aren't supposed to be taken by any of the
10331 * NAPI code and this dummy netdev is supposed to be
10332 * only ever used for NAPI polls
10334 memset(dev, 0, sizeof(struct net_device));
10336 /* make sure we BUG if trying to hit standard
10337 * register/unregister code path
10339 dev->reg_state = NETREG_DUMMY;
10341 /* NAPI wants this */
10342 INIT_LIST_HEAD(&dev->napi_list);
10344 /* a dummy interface is started by default */
10345 set_bit(__LINK_STATE_PRESENT, &dev->state);
10346 set_bit(__LINK_STATE_START, &dev->state);
10348 /* napi_busy_loop stats accounting wants this */
10349 dev_net_set(dev, &init_net);
10351 /* Note : We dont allocate pcpu_refcnt for dummy devices,
10352 * because users of this 'device' dont need to change
10358 EXPORT_SYMBOL_GPL(init_dummy_netdev);
10362 * register_netdev - register a network device
10363 * @dev: device to register
10365 * Take a completed network device structure and add it to the kernel
10366 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
10367 * chain. 0 is returned on success. A negative errno code is returned
10368 * on a failure to set up the device, or if the name is a duplicate.
10370 * This is a wrapper around register_netdevice that takes the rtnl semaphore
10371 * and expands the device name if you passed a format string to
10374 int register_netdev(struct net_device *dev)
10378 if (rtnl_lock_killable())
10380 err = register_netdevice(dev);
10384 EXPORT_SYMBOL(register_netdev);
10386 int netdev_refcnt_read(const struct net_device *dev)
10388 #ifdef CONFIG_PCPU_DEV_REFCNT
10391 for_each_possible_cpu(i)
10392 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
10395 return refcount_read(&dev->dev_refcnt);
10398 EXPORT_SYMBOL(netdev_refcnt_read);
10400 int netdev_unregister_timeout_secs __read_mostly = 10;
10402 #define WAIT_REFS_MIN_MSECS 1
10403 #define WAIT_REFS_MAX_MSECS 250
10405 * netdev_wait_allrefs_any - wait until all references are gone.
10406 * @list: list of net_devices to wait on
10408 * This is called when unregistering network devices.
10410 * Any protocol or device that holds a reference should register
10411 * for netdevice notification, and cleanup and put back the
10412 * reference if they receive an UNREGISTER event.
10413 * We can get stuck here if buggy protocols don't correctly
10416 static struct net_device *netdev_wait_allrefs_any(struct list_head *list)
10418 unsigned long rebroadcast_time, warning_time;
10419 struct net_device *dev;
10422 rebroadcast_time = warning_time = jiffies;
10424 list_for_each_entry(dev, list, todo_list)
10425 if (netdev_refcnt_read(dev) == 1)
10429 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
10432 /* Rebroadcast unregister notification */
10433 list_for_each_entry(dev, list, todo_list)
10434 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
10440 list_for_each_entry(dev, list, todo_list)
10441 if (test_bit(__LINK_STATE_LINKWATCH_PENDING,
10443 /* We must not have linkwatch events
10444 * pending on unregister. If this
10445 * happens, we simply run the queue
10446 * unscheduled, resulting in a noop
10449 linkwatch_run_queue();
10455 rebroadcast_time = jiffies;
10460 wait = WAIT_REFS_MIN_MSECS;
10463 wait = min(wait << 1, WAIT_REFS_MAX_MSECS);
10466 list_for_each_entry(dev, list, todo_list)
10467 if (netdev_refcnt_read(dev) == 1)
10470 if (time_after(jiffies, warning_time +
10471 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) {
10472 list_for_each_entry(dev, list, todo_list) {
10473 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n",
10474 dev->name, netdev_refcnt_read(dev));
10475 ref_tracker_dir_print(&dev->refcnt_tracker, 10);
10478 warning_time = jiffies;
10483 /* The sequence is:
10487 * register_netdevice(x1);
10488 * register_netdevice(x2);
10490 * unregister_netdevice(y1);
10491 * unregister_netdevice(y2);
10497 * We are invoked by rtnl_unlock().
10498 * This allows us to deal with problems:
10499 * 1) We can delete sysfs objects which invoke hotplug
10500 * without deadlocking with linkwatch via keventd.
10501 * 2) Since we run with the RTNL semaphore not held, we can sleep
10502 * safely in order to wait for the netdev refcnt to drop to zero.
10504 * We must not return until all unregister events added during
10505 * the interval the lock was held have been completed.
10507 void netdev_run_todo(void)
10509 struct net_device *dev, *tmp;
10510 struct list_head list;
10511 #ifdef CONFIG_LOCKDEP
10512 struct list_head unlink_list;
10514 list_replace_init(&net_unlink_list, &unlink_list);
10516 while (!list_empty(&unlink_list)) {
10517 struct net_device *dev = list_first_entry(&unlink_list,
10520 list_del_init(&dev->unlink_list);
10521 dev->nested_level = dev->lower_level - 1;
10525 /* Snapshot list, allow later requests */
10526 list_replace_init(&net_todo_list, &list);
10530 /* Wait for rcu callbacks to finish before next phase */
10531 if (!list_empty(&list))
10534 list_for_each_entry_safe(dev, tmp, &list, todo_list) {
10535 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) {
10536 netdev_WARN(dev, "run_todo but not unregistering\n");
10537 list_del(&dev->todo_list);
10541 write_lock(&dev_base_lock);
10542 dev->reg_state = NETREG_UNREGISTERED;
10543 write_unlock(&dev_base_lock);
10544 linkwatch_sync_dev(dev);
10547 while (!list_empty(&list)) {
10548 dev = netdev_wait_allrefs_any(&list);
10549 list_del(&dev->todo_list);
10552 BUG_ON(netdev_refcnt_read(dev) != 1);
10553 BUG_ON(!list_empty(&dev->ptype_all));
10554 BUG_ON(!list_empty(&dev->ptype_specific));
10555 WARN_ON(rcu_access_pointer(dev->ip_ptr));
10556 WARN_ON(rcu_access_pointer(dev->ip6_ptr));
10558 netdev_do_free_pcpu_stats(dev);
10559 if (dev->priv_destructor)
10560 dev->priv_destructor(dev);
10561 if (dev->needs_free_netdev)
10564 if (atomic_dec_and_test(&dev_net(dev)->dev_unreg_count))
10565 wake_up(&netdev_unregistering_wq);
10567 /* Free network device */
10568 kobject_put(&dev->dev.kobj);
10572 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has
10573 * all the same fields in the same order as net_device_stats, with only
10574 * the type differing, but rtnl_link_stats64 may have additional fields
10575 * at the end for newer counters.
10577 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
10578 const struct net_device_stats *netdev_stats)
10580 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t);
10581 const atomic_long_t *src = (atomic_long_t *)netdev_stats;
10582 u64 *dst = (u64 *)stats64;
10584 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64));
10585 for (i = 0; i < n; i++)
10586 dst[i] = (unsigned long)atomic_long_read(&src[i]);
10587 /* zero out counters that only exist in rtnl_link_stats64 */
10588 memset((char *)stats64 + n * sizeof(u64), 0,
10589 sizeof(*stats64) - n * sizeof(u64));
10591 EXPORT_SYMBOL(netdev_stats_to_stats64);
10593 static __cold struct net_device_core_stats __percpu *netdev_core_stats_alloc(
10594 struct net_device *dev)
10596 struct net_device_core_stats __percpu *p;
10598 p = alloc_percpu_gfp(struct net_device_core_stats,
10599 GFP_ATOMIC | __GFP_NOWARN);
10601 if (p && cmpxchg(&dev->core_stats, NULL, p))
10604 /* This READ_ONCE() pairs with the cmpxchg() above */
10605 return READ_ONCE(dev->core_stats);
10608 noinline void netdev_core_stats_inc(struct net_device *dev, u32 offset)
10610 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */
10611 struct net_device_core_stats __percpu *p = READ_ONCE(dev->core_stats);
10612 unsigned long __percpu *field;
10614 if (unlikely(!p)) {
10615 p = netdev_core_stats_alloc(dev);
10620 field = (__force unsigned long __percpu *)((__force void *)p + offset);
10621 this_cpu_inc(*field);
10623 EXPORT_SYMBOL_GPL(netdev_core_stats_inc);
10626 * dev_get_stats - get network device statistics
10627 * @dev: device to get statistics from
10628 * @storage: place to store stats
10630 * Get network statistics from device. Return @storage.
10631 * The device driver may provide its own method by setting
10632 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats;
10633 * otherwise the internal statistics structure is used.
10635 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev,
10636 struct rtnl_link_stats64 *storage)
10638 const struct net_device_ops *ops = dev->netdev_ops;
10639 const struct net_device_core_stats __percpu *p;
10641 if (ops->ndo_get_stats64) {
10642 memset(storage, 0, sizeof(*storage));
10643 ops->ndo_get_stats64(dev, storage);
10644 } else if (ops->ndo_get_stats) {
10645 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev));
10647 netdev_stats_to_stats64(storage, &dev->stats);
10650 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */
10651 p = READ_ONCE(dev->core_stats);
10653 const struct net_device_core_stats *core_stats;
10656 for_each_possible_cpu(i) {
10657 core_stats = per_cpu_ptr(p, i);
10658 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped);
10659 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped);
10660 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler);
10661 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped);
10666 EXPORT_SYMBOL(dev_get_stats);
10669 * dev_fetch_sw_netstats - get per-cpu network device statistics
10670 * @s: place to store stats
10671 * @netstats: per-cpu network stats to read from
10673 * Read per-cpu network statistics and populate the related fields in @s.
10675 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s,
10676 const struct pcpu_sw_netstats __percpu *netstats)
10680 for_each_possible_cpu(cpu) {
10681 u64 rx_packets, rx_bytes, tx_packets, tx_bytes;
10682 const struct pcpu_sw_netstats *stats;
10683 unsigned int start;
10685 stats = per_cpu_ptr(netstats, cpu);
10687 start = u64_stats_fetch_begin(&stats->syncp);
10688 rx_packets = u64_stats_read(&stats->rx_packets);
10689 rx_bytes = u64_stats_read(&stats->rx_bytes);
10690 tx_packets = u64_stats_read(&stats->tx_packets);
10691 tx_bytes = u64_stats_read(&stats->tx_bytes);
10692 } while (u64_stats_fetch_retry(&stats->syncp, start));
10694 s->rx_packets += rx_packets;
10695 s->rx_bytes += rx_bytes;
10696 s->tx_packets += tx_packets;
10697 s->tx_bytes += tx_bytes;
10700 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats);
10703 * dev_get_tstats64 - ndo_get_stats64 implementation
10704 * @dev: device to get statistics from
10705 * @s: place to store stats
10707 * Populate @s from dev->stats and dev->tstats. Can be used as
10708 * ndo_get_stats64() callback.
10710 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s)
10712 netdev_stats_to_stats64(s, &dev->stats);
10713 dev_fetch_sw_netstats(s, dev->tstats);
10715 EXPORT_SYMBOL_GPL(dev_get_tstats64);
10717 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev)
10719 struct netdev_queue *queue = dev_ingress_queue(dev);
10721 #ifdef CONFIG_NET_CLS_ACT
10724 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
10727 netdev_init_one_queue(dev, queue, NULL);
10728 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc);
10729 RCU_INIT_POINTER(queue->qdisc_sleeping, &noop_qdisc);
10730 rcu_assign_pointer(dev->ingress_queue, queue);
10735 static const struct ethtool_ops default_ethtool_ops;
10737 void netdev_set_default_ethtool_ops(struct net_device *dev,
10738 const struct ethtool_ops *ops)
10740 if (dev->ethtool_ops == &default_ethtool_ops)
10741 dev->ethtool_ops = ops;
10743 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops);
10746 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default
10747 * @dev: netdev to enable the IRQ coalescing on
10749 * Sets a conservative default for SW IRQ coalescing. Users can use
10750 * sysfs attributes to override the default values.
10752 void netdev_sw_irq_coalesce_default_on(struct net_device *dev)
10754 WARN_ON(dev->reg_state == NETREG_REGISTERED);
10756 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) {
10757 dev->gro_flush_timeout = 20000;
10758 dev->napi_defer_hard_irqs = 1;
10761 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on);
10763 void netdev_freemem(struct net_device *dev)
10765 char *addr = (char *)dev - dev->padded;
10771 * alloc_netdev_mqs - allocate network device
10772 * @sizeof_priv: size of private data to allocate space for
10773 * @name: device name format string
10774 * @name_assign_type: origin of device name
10775 * @setup: callback to initialize device
10776 * @txqs: the number of TX subqueues to allocate
10777 * @rxqs: the number of RX subqueues to allocate
10779 * Allocates a struct net_device with private data area for driver use
10780 * and performs basic initialization. Also allocates subqueue structs
10781 * for each queue on the device.
10783 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
10784 unsigned char name_assign_type,
10785 void (*setup)(struct net_device *),
10786 unsigned int txqs, unsigned int rxqs)
10788 struct net_device *dev;
10789 unsigned int alloc_size;
10790 struct net_device *p;
10792 BUG_ON(strlen(name) >= sizeof(dev->name));
10795 pr_err("alloc_netdev: Unable to allocate device with zero queues\n");
10800 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n");
10804 alloc_size = sizeof(struct net_device);
10806 /* ensure 32-byte alignment of private area */
10807 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
10808 alloc_size += sizeof_priv;
10810 /* ensure 32-byte alignment of whole construct */
10811 alloc_size += NETDEV_ALIGN - 1;
10813 p = kvzalloc(alloc_size, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL);
10817 dev = PTR_ALIGN(p, NETDEV_ALIGN);
10818 dev->padded = (char *)dev - (char *)p;
10820 ref_tracker_dir_init(&dev->refcnt_tracker, 128, name);
10821 #ifdef CONFIG_PCPU_DEV_REFCNT
10822 dev->pcpu_refcnt = alloc_percpu(int);
10823 if (!dev->pcpu_refcnt)
10827 refcount_set(&dev->dev_refcnt, 1);
10830 if (dev_addr_init(dev))
10836 dev_net_set(dev, &init_net);
10838 dev->gso_max_size = GSO_LEGACY_MAX_SIZE;
10839 dev->xdp_zc_max_segs = 1;
10840 dev->gso_max_segs = GSO_MAX_SEGS;
10841 dev->gro_max_size = GRO_LEGACY_MAX_SIZE;
10842 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE;
10843 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE;
10844 dev->tso_max_size = TSO_LEGACY_MAX_SIZE;
10845 dev->tso_max_segs = TSO_MAX_SEGS;
10846 dev->upper_level = 1;
10847 dev->lower_level = 1;
10848 #ifdef CONFIG_LOCKDEP
10849 dev->nested_level = 0;
10850 INIT_LIST_HEAD(&dev->unlink_list);
10853 INIT_LIST_HEAD(&dev->napi_list);
10854 INIT_LIST_HEAD(&dev->unreg_list);
10855 INIT_LIST_HEAD(&dev->close_list);
10856 INIT_LIST_HEAD(&dev->link_watch_list);
10857 INIT_LIST_HEAD(&dev->adj_list.upper);
10858 INIT_LIST_HEAD(&dev->adj_list.lower);
10859 INIT_LIST_HEAD(&dev->ptype_all);
10860 INIT_LIST_HEAD(&dev->ptype_specific);
10861 INIT_LIST_HEAD(&dev->net_notifier_list);
10862 #ifdef CONFIG_NET_SCHED
10863 hash_init(dev->qdisc_hash);
10865 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM;
10868 if (!dev->tx_queue_len) {
10869 dev->priv_flags |= IFF_NO_QUEUE;
10870 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN;
10873 dev->num_tx_queues = txqs;
10874 dev->real_num_tx_queues = txqs;
10875 if (netif_alloc_netdev_queues(dev))
10878 dev->num_rx_queues = rxqs;
10879 dev->real_num_rx_queues = rxqs;
10880 if (netif_alloc_rx_queues(dev))
10883 strcpy(dev->name, name);
10884 dev->name_assign_type = name_assign_type;
10885 dev->group = INIT_NETDEV_GROUP;
10886 if (!dev->ethtool_ops)
10887 dev->ethtool_ops = &default_ethtool_ops;
10889 nf_hook_netdev_init(dev);
10898 #ifdef CONFIG_PCPU_DEV_REFCNT
10899 free_percpu(dev->pcpu_refcnt);
10902 netdev_freemem(dev);
10905 EXPORT_SYMBOL(alloc_netdev_mqs);
10908 * free_netdev - free network device
10911 * This function does the last stage of destroying an allocated device
10912 * interface. The reference to the device object is released. If this
10913 * is the last reference then it will be freed.Must be called in process
10916 void free_netdev(struct net_device *dev)
10918 struct napi_struct *p, *n;
10922 /* When called immediately after register_netdevice() failed the unwind
10923 * handling may still be dismantling the device. Handle that case by
10924 * deferring the free.
10926 if (dev->reg_state == NETREG_UNREGISTERING) {
10928 dev->needs_free_netdev = true;
10932 netif_free_tx_queues(dev);
10933 netif_free_rx_queues(dev);
10935 kfree(rcu_dereference_protected(dev->ingress_queue, 1));
10937 /* Flush device addresses */
10938 dev_addr_flush(dev);
10940 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
10943 ref_tracker_dir_exit(&dev->refcnt_tracker);
10944 #ifdef CONFIG_PCPU_DEV_REFCNT
10945 free_percpu(dev->pcpu_refcnt);
10946 dev->pcpu_refcnt = NULL;
10948 free_percpu(dev->core_stats);
10949 dev->core_stats = NULL;
10950 free_percpu(dev->xdp_bulkq);
10951 dev->xdp_bulkq = NULL;
10953 /* Compatibility with error handling in drivers */
10954 if (dev->reg_state == NETREG_UNINITIALIZED) {
10955 netdev_freemem(dev);
10959 BUG_ON(dev->reg_state != NETREG_UNREGISTERED);
10960 dev->reg_state = NETREG_RELEASED;
10962 /* will free via device release */
10963 put_device(&dev->dev);
10965 EXPORT_SYMBOL(free_netdev);
10968 * synchronize_net - Synchronize with packet receive processing
10970 * Wait for packets currently being received to be done.
10971 * Does not block later packets from starting.
10973 void synchronize_net(void)
10976 if (rtnl_is_locked())
10977 synchronize_rcu_expedited();
10981 EXPORT_SYMBOL(synchronize_net);
10984 * unregister_netdevice_queue - remove device from the kernel
10988 * This function shuts down a device interface and removes it
10989 * from the kernel tables.
10990 * If head not NULL, device is queued to be unregistered later.
10992 * Callers must hold the rtnl semaphore. You may want
10993 * unregister_netdev() instead of this.
10996 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head)
11001 list_move_tail(&dev->unreg_list, head);
11005 list_add(&dev->unreg_list, &single);
11006 unregister_netdevice_many(&single);
11009 EXPORT_SYMBOL(unregister_netdevice_queue);
11011 void unregister_netdevice_many_notify(struct list_head *head,
11012 u32 portid, const struct nlmsghdr *nlh)
11014 struct net_device *dev, *tmp;
11015 LIST_HEAD(close_head);
11017 BUG_ON(dev_boot_phase);
11020 if (list_empty(head))
11023 list_for_each_entry_safe(dev, tmp, head, unreg_list) {
11024 /* Some devices call without registering
11025 * for initialization unwind. Remove those
11026 * devices and proceed with the remaining.
11028 if (dev->reg_state == NETREG_UNINITIALIZED) {
11029 pr_debug("unregister_netdevice: device %s/%p never was registered\n",
11033 list_del(&dev->unreg_list);
11036 dev->dismantle = true;
11037 BUG_ON(dev->reg_state != NETREG_REGISTERED);
11040 /* If device is running, close it first. */
11041 list_for_each_entry(dev, head, unreg_list)
11042 list_add_tail(&dev->close_list, &close_head);
11043 dev_close_many(&close_head, true);
11045 list_for_each_entry(dev, head, unreg_list) {
11046 /* And unlink it from device chain. */
11047 write_lock(&dev_base_lock);
11048 unlist_netdevice(dev, false);
11049 dev->reg_state = NETREG_UNREGISTERING;
11050 write_unlock(&dev_base_lock);
11052 flush_all_backlogs();
11056 list_for_each_entry(dev, head, unreg_list) {
11057 struct sk_buff *skb = NULL;
11059 /* Shutdown queueing discipline. */
11061 dev_tcx_uninstall(dev);
11062 dev_xdp_uninstall(dev);
11063 bpf_dev_bound_netdev_unregister(dev);
11065 netdev_offload_xstats_disable_all(dev);
11067 /* Notify protocols, that we are about to destroy
11068 * this device. They should clean all the things.
11070 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
11072 if (!dev->rtnl_link_ops ||
11073 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
11074 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0,
11075 GFP_KERNEL, NULL, 0,
11079 * Flush the unicast and multicast chains
11084 netdev_name_node_alt_flush(dev);
11085 netdev_name_node_free(dev->name_node);
11087 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev);
11089 if (dev->netdev_ops->ndo_uninit)
11090 dev->netdev_ops->ndo_uninit(dev);
11093 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh);
11095 /* Notifier chain MUST detach us all upper devices. */
11096 WARN_ON(netdev_has_any_upper_dev(dev));
11097 WARN_ON(netdev_has_any_lower_dev(dev));
11099 /* Remove entries from kobject tree */
11100 netdev_unregister_kobject(dev);
11102 /* Remove XPS queueing entries */
11103 netif_reset_xps_queues_gt(dev, 0);
11109 list_for_each_entry(dev, head, unreg_list) {
11110 netdev_put(dev, &dev->dev_registered_tracker);
11118 * unregister_netdevice_many - unregister many devices
11119 * @head: list of devices
11121 * Note: As most callers use a stack allocated list_head,
11122 * we force a list_del() to make sure stack wont be corrupted later.
11124 void unregister_netdevice_many(struct list_head *head)
11126 unregister_netdevice_many_notify(head, 0, NULL);
11128 EXPORT_SYMBOL(unregister_netdevice_many);
11131 * unregister_netdev - remove device from the kernel
11134 * This function shuts down a device interface and removes it
11135 * from the kernel tables.
11137 * This is just a wrapper for unregister_netdevice that takes
11138 * the rtnl semaphore. In general you want to use this and not
11139 * unregister_netdevice.
11141 void unregister_netdev(struct net_device *dev)
11144 unregister_netdevice(dev);
11147 EXPORT_SYMBOL(unregister_netdev);
11150 * __dev_change_net_namespace - move device to different nethost namespace
11152 * @net: network namespace
11153 * @pat: If not NULL name pattern to try if the current device name
11154 * is already taken in the destination network namespace.
11155 * @new_ifindex: If not zero, specifies device index in the target
11158 * This function shuts down a device interface and moves it
11159 * to a new network namespace. On success 0 is returned, on
11160 * a failure a netagive errno code is returned.
11162 * Callers must hold the rtnl semaphore.
11165 int __dev_change_net_namespace(struct net_device *dev, struct net *net,
11166 const char *pat, int new_ifindex)
11168 struct netdev_name_node *name_node;
11169 struct net *net_old = dev_net(dev);
11170 char new_name[IFNAMSIZ] = {};
11175 /* Don't allow namespace local devices to be moved. */
11177 if (dev->features & NETIF_F_NETNS_LOCAL)
11180 /* Ensure the device has been registrered */
11181 if (dev->reg_state != NETREG_REGISTERED)
11184 /* Get out if there is nothing todo */
11186 if (net_eq(net_old, net))
11189 /* Pick the destination device name, and ensure
11190 * we can use it in the destination network namespace.
11193 if (netdev_name_in_use(net, dev->name)) {
11194 /* We get here if we can't use the current device name */
11197 err = dev_prep_valid_name(net, dev, pat, new_name, EEXIST);
11201 /* Check that none of the altnames conflicts. */
11203 netdev_for_each_altname(dev, name_node)
11204 if (netdev_name_in_use(net, name_node->name))
11207 /* Check that new_ifindex isn't used yet. */
11209 err = dev_index_reserve(net, new_ifindex);
11213 /* If there is an ifindex conflict assign a new one */
11214 err = dev_index_reserve(net, dev->ifindex);
11216 err = dev_index_reserve(net, 0);
11223 * And now a mini version of register_netdevice unregister_netdevice.
11226 /* If device is running close it first. */
11229 /* And unlink it from device chain */
11230 unlist_netdevice(dev, true);
11234 /* Shutdown queueing discipline. */
11237 /* Notify protocols, that we are about to destroy
11238 * this device. They should clean all the things.
11240 * Note that dev->reg_state stays at NETREG_REGISTERED.
11241 * This is wanted because this way 8021q and macvlan know
11242 * the device is just moving and can keep their slaves up.
11244 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
11247 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL);
11249 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid,
11253 * Flush the unicast and multicast chains
11258 /* Send a netdev-removed uevent to the old namespace */
11259 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE);
11260 netdev_adjacent_del_links(dev);
11262 /* Move per-net netdevice notifiers that are following the netdevice */
11263 move_netdevice_notifiers_dev_net(dev, net);
11265 /* Actually switch the network namespace */
11266 dev_net_set(dev, net);
11267 dev->ifindex = new_ifindex;
11269 if (new_name[0]) /* Rename the netdev to prepared name */
11270 strscpy(dev->name, new_name, IFNAMSIZ);
11272 /* Fixup kobjects */
11273 dev_set_uevent_suppress(&dev->dev, 1);
11274 err = device_rename(&dev->dev, dev->name);
11275 dev_set_uevent_suppress(&dev->dev, 0);
11278 /* Send a netdev-add uevent to the new namespace */
11279 kobject_uevent(&dev->dev.kobj, KOBJ_ADD);
11280 netdev_adjacent_add_links(dev);
11282 /* Adapt owner in case owning user namespace of target network
11283 * namespace is different from the original one.
11285 err = netdev_change_owner(dev, net_old, net);
11288 /* Add the device back in the hashes */
11289 list_netdevice(dev);
11291 /* Notify protocols, that a new device appeared. */
11292 call_netdevice_notifiers(NETDEV_REGISTER, dev);
11295 * Prevent userspace races by waiting until the network
11296 * device is fully setup before sending notifications.
11298 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL);
11305 EXPORT_SYMBOL_GPL(__dev_change_net_namespace);
11307 static int dev_cpu_dead(unsigned int oldcpu)
11309 struct sk_buff **list_skb;
11310 struct sk_buff *skb;
11312 struct softnet_data *sd, *oldsd, *remsd = NULL;
11314 local_irq_disable();
11315 cpu = smp_processor_id();
11316 sd = &per_cpu(softnet_data, cpu);
11317 oldsd = &per_cpu(softnet_data, oldcpu);
11319 /* Find end of our completion_queue. */
11320 list_skb = &sd->completion_queue;
11322 list_skb = &(*list_skb)->next;
11323 /* Append completion queue from offline CPU. */
11324 *list_skb = oldsd->completion_queue;
11325 oldsd->completion_queue = NULL;
11327 /* Append output queue from offline CPU. */
11328 if (oldsd->output_queue) {
11329 *sd->output_queue_tailp = oldsd->output_queue;
11330 sd->output_queue_tailp = oldsd->output_queue_tailp;
11331 oldsd->output_queue = NULL;
11332 oldsd->output_queue_tailp = &oldsd->output_queue;
11334 /* Append NAPI poll list from offline CPU, with one exception :
11335 * process_backlog() must be called by cpu owning percpu backlog.
11336 * We properly handle process_queue & input_pkt_queue later.
11338 while (!list_empty(&oldsd->poll_list)) {
11339 struct napi_struct *napi = list_first_entry(&oldsd->poll_list,
11340 struct napi_struct,
11343 list_del_init(&napi->poll_list);
11344 if (napi->poll == process_backlog)
11347 ____napi_schedule(sd, napi);
11350 raise_softirq_irqoff(NET_TX_SOFTIRQ);
11351 local_irq_enable();
11354 remsd = oldsd->rps_ipi_list;
11355 oldsd->rps_ipi_list = NULL;
11357 /* send out pending IPI's on offline CPU */
11358 net_rps_send_ipi(remsd);
11360 /* Process offline CPU's input_pkt_queue */
11361 while ((skb = __skb_dequeue(&oldsd->process_queue))) {
11363 input_queue_head_incr(oldsd);
11365 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) {
11367 input_queue_head_incr(oldsd);
11374 * netdev_increment_features - increment feature set by one
11375 * @all: current feature set
11376 * @one: new feature set
11377 * @mask: mask feature set
11379 * Computes a new feature set after adding a device with feature set
11380 * @one to the master device with current feature set @all. Will not
11381 * enable anything that is off in @mask. Returns the new feature set.
11383 netdev_features_t netdev_increment_features(netdev_features_t all,
11384 netdev_features_t one, netdev_features_t mask)
11386 if (mask & NETIF_F_HW_CSUM)
11387 mask |= NETIF_F_CSUM_MASK;
11388 mask |= NETIF_F_VLAN_CHALLENGED;
11390 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask;
11391 all &= one | ~NETIF_F_ALL_FOR_ALL;
11393 /* If one device supports hw checksumming, set for all. */
11394 if (all & NETIF_F_HW_CSUM)
11395 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM);
11399 EXPORT_SYMBOL(netdev_increment_features);
11401 static struct hlist_head * __net_init netdev_create_hash(void)
11404 struct hlist_head *hash;
11406 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL);
11408 for (i = 0; i < NETDEV_HASHENTRIES; i++)
11409 INIT_HLIST_HEAD(&hash[i]);
11414 /* Initialize per network namespace state */
11415 static int __net_init netdev_init(struct net *net)
11417 BUILD_BUG_ON(GRO_HASH_BUCKETS >
11418 8 * sizeof_field(struct napi_struct, gro_bitmask));
11420 INIT_LIST_HEAD(&net->dev_base_head);
11422 net->dev_name_head = netdev_create_hash();
11423 if (net->dev_name_head == NULL)
11426 net->dev_index_head = netdev_create_hash();
11427 if (net->dev_index_head == NULL)
11430 xa_init_flags(&net->dev_by_index, XA_FLAGS_ALLOC1);
11432 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain);
11437 kfree(net->dev_name_head);
11443 * netdev_drivername - network driver for the device
11444 * @dev: network device
11446 * Determine network driver for device.
11448 const char *netdev_drivername(const struct net_device *dev)
11450 const struct device_driver *driver;
11451 const struct device *parent;
11452 const char *empty = "";
11454 parent = dev->dev.parent;
11458 driver = parent->driver;
11459 if (driver && driver->name)
11460 return driver->name;
11464 static void __netdev_printk(const char *level, const struct net_device *dev,
11465 struct va_format *vaf)
11467 if (dev && dev->dev.parent) {
11468 dev_printk_emit(level[1] - '0',
11471 dev_driver_string(dev->dev.parent),
11472 dev_name(dev->dev.parent),
11473 netdev_name(dev), netdev_reg_state(dev),
11476 printk("%s%s%s: %pV",
11477 level, netdev_name(dev), netdev_reg_state(dev), vaf);
11479 printk("%s(NULL net_device): %pV", level, vaf);
11483 void netdev_printk(const char *level, const struct net_device *dev,
11484 const char *format, ...)
11486 struct va_format vaf;
11489 va_start(args, format);
11494 __netdev_printk(level, dev, &vaf);
11498 EXPORT_SYMBOL(netdev_printk);
11500 #define define_netdev_printk_level(func, level) \
11501 void func(const struct net_device *dev, const char *fmt, ...) \
11503 struct va_format vaf; \
11506 va_start(args, fmt); \
11511 __netdev_printk(level, dev, &vaf); \
11515 EXPORT_SYMBOL(func);
11517 define_netdev_printk_level(netdev_emerg, KERN_EMERG);
11518 define_netdev_printk_level(netdev_alert, KERN_ALERT);
11519 define_netdev_printk_level(netdev_crit, KERN_CRIT);
11520 define_netdev_printk_level(netdev_err, KERN_ERR);
11521 define_netdev_printk_level(netdev_warn, KERN_WARNING);
11522 define_netdev_printk_level(netdev_notice, KERN_NOTICE);
11523 define_netdev_printk_level(netdev_info, KERN_INFO);
11525 static void __net_exit netdev_exit(struct net *net)
11527 kfree(net->dev_name_head);
11528 kfree(net->dev_index_head);
11529 xa_destroy(&net->dev_by_index);
11530 if (net != &init_net)
11531 WARN_ON_ONCE(!list_empty(&net->dev_base_head));
11534 static struct pernet_operations __net_initdata netdev_net_ops = {
11535 .init = netdev_init,
11536 .exit = netdev_exit,
11539 static void __net_exit default_device_exit_net(struct net *net)
11541 struct netdev_name_node *name_node, *tmp;
11542 struct net_device *dev, *aux;
11544 * Push all migratable network devices back to the
11545 * initial network namespace
11548 for_each_netdev_safe(net, dev, aux) {
11550 char fb_name[IFNAMSIZ];
11552 /* Ignore unmoveable devices (i.e. loopback) */
11553 if (dev->features & NETIF_F_NETNS_LOCAL)
11556 /* Leave virtual devices for the generic cleanup */
11557 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund)
11560 /* Push remaining network devices to init_net */
11561 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex);
11562 if (netdev_name_in_use(&init_net, fb_name))
11563 snprintf(fb_name, IFNAMSIZ, "dev%%d");
11565 netdev_for_each_altname_safe(dev, name_node, tmp)
11566 if (netdev_name_in_use(&init_net, name_node->name)) {
11567 netdev_name_node_del(name_node);
11569 __netdev_name_node_alt_destroy(name_node);
11572 err = dev_change_net_namespace(dev, &init_net, fb_name);
11574 pr_emerg("%s: failed to move %s to init_net: %d\n",
11575 __func__, dev->name, err);
11581 static void __net_exit default_device_exit_batch(struct list_head *net_list)
11583 /* At exit all network devices most be removed from a network
11584 * namespace. Do this in the reverse order of registration.
11585 * Do this across as many network namespaces as possible to
11586 * improve batching efficiency.
11588 struct net_device *dev;
11590 LIST_HEAD(dev_kill_list);
11593 list_for_each_entry(net, net_list, exit_list) {
11594 default_device_exit_net(net);
11598 list_for_each_entry(net, net_list, exit_list) {
11599 for_each_netdev_reverse(net, dev) {
11600 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink)
11601 dev->rtnl_link_ops->dellink(dev, &dev_kill_list);
11603 unregister_netdevice_queue(dev, &dev_kill_list);
11606 unregister_netdevice_many(&dev_kill_list);
11610 static struct pernet_operations __net_initdata default_device_ops = {
11611 .exit_batch = default_device_exit_batch,
11614 static void __init net_dev_struct_check(void)
11616 /* TX read-mostly hotpath */
11617 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, priv_flags);
11618 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, netdev_ops);
11619 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, header_ops);
11620 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, _tx);
11621 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, real_num_tx_queues);
11622 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_size);
11623 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_ipv4_max_size);
11624 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_segs);
11625 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_partial_features);
11626 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, num_tc);
11627 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, mtu);
11628 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, needed_headroom);
11629 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tc_to_txq);
11631 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, xps_maps);
11633 #ifdef CONFIG_NETFILTER_EGRESS
11634 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, nf_hooks_egress);
11636 #ifdef CONFIG_NET_XGRESS
11637 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tcx_egress);
11639 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_tx, 160);
11641 /* TXRX read-mostly hotpath */
11642 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, lstats);
11643 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, state);
11644 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, flags);
11645 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, hard_header_len);
11646 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, features);
11647 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, ip6_ptr);
11648 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_txrx, 46);
11650 /* RX read-mostly hotpath */
11651 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ptype_specific);
11652 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ifindex);
11653 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, real_num_rx_queues);
11654 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, _rx);
11655 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_flush_timeout);
11656 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, napi_defer_hard_irqs);
11657 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_max_size);
11658 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_ipv4_max_size);
11659 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler);
11660 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler_data);
11661 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, nd_net);
11662 #ifdef CONFIG_NETPOLL
11663 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, npinfo);
11665 #ifdef CONFIG_NET_XGRESS
11666 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, tcx_ingress);
11668 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_rx, 104);
11672 * Initialize the DEV module. At boot time this walks the device list and
11673 * unhooks any devices that fail to initialise (normally hardware not
11674 * present) and leaves us with a valid list of present and active devices.
11679 * This is called single threaded during boot, so no need
11680 * to take the rtnl semaphore.
11682 static int __init net_dev_init(void)
11684 int i, rc = -ENOMEM;
11686 BUG_ON(!dev_boot_phase);
11688 net_dev_struct_check();
11690 if (dev_proc_init())
11693 if (netdev_kobject_init())
11696 INIT_LIST_HEAD(&ptype_all);
11697 for (i = 0; i < PTYPE_HASH_SIZE; i++)
11698 INIT_LIST_HEAD(&ptype_base[i]);
11700 if (register_pernet_subsys(&netdev_net_ops))
11704 * Initialise the packet receive queues.
11707 for_each_possible_cpu(i) {
11708 struct work_struct *flush = per_cpu_ptr(&flush_works, i);
11709 struct softnet_data *sd = &per_cpu(softnet_data, i);
11711 INIT_WORK(flush, flush_backlog);
11713 skb_queue_head_init(&sd->input_pkt_queue);
11714 skb_queue_head_init(&sd->process_queue);
11715 #ifdef CONFIG_XFRM_OFFLOAD
11716 skb_queue_head_init(&sd->xfrm_backlog);
11718 INIT_LIST_HEAD(&sd->poll_list);
11719 sd->output_queue_tailp = &sd->output_queue;
11721 INIT_CSD(&sd->csd, rps_trigger_softirq, sd);
11724 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd);
11725 spin_lock_init(&sd->defer_lock);
11727 init_gro_hash(&sd->backlog);
11728 sd->backlog.poll = process_backlog;
11729 sd->backlog.weight = weight_p;
11732 dev_boot_phase = 0;
11734 /* The loopback device is special if any other network devices
11735 * is present in a network namespace the loopback device must
11736 * be present. Since we now dynamically allocate and free the
11737 * loopback device ensure this invariant is maintained by
11738 * keeping the loopback device as the first device on the
11739 * list of network devices. Ensuring the loopback devices
11740 * is the first device that appears and the last network device
11743 if (register_pernet_device(&loopback_net_ops))
11746 if (register_pernet_device(&default_device_ops))
11749 open_softirq(NET_TX_SOFTIRQ, net_tx_action);
11750 open_softirq(NET_RX_SOFTIRQ, net_rx_action);
11752 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead",
11753 NULL, dev_cpu_dead);
11760 subsys_initcall(net_dev_init);