2 * NET3 Protocol independent device support routines.
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
9 * Derived from the non IP parts of dev.c 1.0.19
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
15 * Florian la Roche <rzsfl@rz.uni-sb.de>
16 * Alan Cox <gw4pts@gw4pts.ampr.org>
17 * David Hinds <dahinds@users.sourceforge.net>
18 * Alexey Kuznetsov <kuznet@ms2.inr.ac.ru>
19 * Adam Sulmicki <adam@cfar.umd.edu>
20 * Pekka Riikonen <priikone@poesidon.pspt.fi>
23 * D.J. Barrow : Fixed bug where dev->refcnt gets set
24 * to 2 if register_netdev gets called
25 * before net_dev_init & also removed a
26 * few lines of code in the process.
27 * Alan Cox : device private ioctl copies fields back.
28 * Alan Cox : Transmit queue code does relevant
29 * stunts to keep the queue safe.
30 * Alan Cox : Fixed double lock.
31 * Alan Cox : Fixed promisc NULL pointer trap
32 * ???????? : Support the full private ioctl range
33 * Alan Cox : Moved ioctl permission check into
35 * Tim Kordas : SIOCADDMULTI/SIOCDELMULTI
36 * Alan Cox : 100 backlog just doesn't cut it when
37 * you start doing multicast video 8)
38 * Alan Cox : Rewrote net_bh and list manager.
39 * Alan Cox : Fix ETH_P_ALL echoback lengths.
40 * Alan Cox : Took out transmit every packet pass
41 * Saved a few bytes in the ioctl handler
42 * Alan Cox : Network driver sets packet type before
43 * calling netif_rx. Saves a function
45 * Alan Cox : Hashed net_bh()
46 * Richard Kooijman: Timestamp fixes.
47 * Alan Cox : Wrong field in SIOCGIFDSTADDR
48 * Alan Cox : Device lock protection.
49 * Alan Cox : Fixed nasty side effect of device close
51 * Rudi Cilibrasi : Pass the right thing to
53 * Dave Miller : 32bit quantity for the device lock to
54 * make it work out on a Sparc.
55 * Bjorn Ekwall : Added KERNELD hack.
56 * Alan Cox : Cleaned up the backlog initialise.
57 * Craig Metz : SIOCGIFCONF fix if space for under
59 * Thomas Bogendoerfer : Return ENODEV for dev_open, if there
60 * is no device open function.
61 * Andi Kleen : Fix error reporting for SIOCGIFCONF
62 * Michael Chastain : Fix signed/unsigned for SIOCGIFCONF
63 * Cyrus Durgin : Cleaned for KMOD
64 * Adam Sulmicki : Bug Fix : Network Device Unload
65 * A network device unload needs to purge
67 * Paul Rusty Russell : SIOCSIFNAME
68 * Pekka Riikonen : Netdev boot-time settings code
69 * Andrew Morton : Make unregister_netdevice wait
70 * indefinitely on dev->refcnt
71 * J Hadi Salim : - Backlog queue sampling
72 * - netif_rx() feedback
75 #include <linux/uaccess.h>
76 #include <linux/bitops.h>
77 #include <linux/capability.h>
78 #include <linux/cpu.h>
79 #include <linux/types.h>
80 #include <linux/kernel.h>
81 #include <linux/hash.h>
82 #include <linux/slab.h>
83 #include <linux/sched.h>
84 #include <linux/sched/mm.h>
85 #include <linux/mutex.h>
86 #include <linux/string.h>
88 #include <linux/socket.h>
89 #include <linux/sockios.h>
90 #include <linux/errno.h>
91 #include <linux/interrupt.h>
92 #include <linux/if_ether.h>
93 #include <linux/netdevice.h>
94 #include <linux/etherdevice.h>
95 #include <linux/ethtool.h>
96 #include <linux/notifier.h>
97 #include <linux/skbuff.h>
98 #include <linux/bpf.h>
99 #include <linux/bpf_trace.h>
100 #include <net/net_namespace.h>
101 #include <net/sock.h>
102 #include <net/busy_poll.h>
103 #include <linux/rtnetlink.h>
104 #include <linux/stat.h>
106 #include <net/dst_metadata.h>
107 #include <net/pkt_sched.h>
108 #include <net/pkt_cls.h>
109 #include <net/checksum.h>
110 #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 <linux/pci.h>
136 #include <linux/inetdevice.h>
137 #include <linux/cpu_rmap.h>
138 #include <linux/static_key.h>
139 #include <linux/hashtable.h>
140 #include <linux/vmalloc.h>
141 #include <linux/if_macvlan.h>
142 #include <linux/errqueue.h>
143 #include <linux/hrtimer.h>
144 #include <linux/netfilter_ingress.h>
145 #include <linux/crash_dump.h>
146 #include <linux/sctp.h>
147 #include <net/udp_tunnel.h>
148 #include <linux/net_namespace.h>
150 #include "net-sysfs.h"
152 /* Instead of increasing this, you should create a hash table. */
153 #define MAX_GRO_SKBS 8
155 /* This should be increased if a protocol with a bigger head is added. */
156 #define GRO_MAX_HEAD (MAX_HEADER + 128)
158 static DEFINE_SPINLOCK(ptype_lock);
159 static DEFINE_SPINLOCK(offload_lock);
160 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly;
161 struct list_head ptype_all __read_mostly; /* Taps */
162 static struct list_head offload_base __read_mostly;
164 static int netif_rx_internal(struct sk_buff *skb);
165 static int call_netdevice_notifiers_info(unsigned long val,
166 struct netdev_notifier_info *info);
167 static struct napi_struct *napi_by_id(unsigned int napi_id);
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 seqcount_t devnet_rename_seq;
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(struct softnet_data *sd)
222 spin_lock(&sd->input_pkt_queue.lock);
226 static inline void rps_unlock(struct softnet_data *sd)
229 spin_unlock(&sd->input_pkt_queue.lock);
233 /* Device list insertion */
234 static void list_netdevice(struct net_device *dev)
236 struct net *net = dev_net(dev);
240 write_lock_bh(&dev_base_lock);
241 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head);
242 hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
243 hlist_add_head_rcu(&dev->index_hlist,
244 dev_index_hash(net, dev->ifindex));
245 write_unlock_bh(&dev_base_lock);
247 dev_base_seq_inc(net);
250 /* Device list removal
251 * caller must respect a RCU grace period before freeing/reusing dev
253 static void unlist_netdevice(struct net_device *dev)
257 /* Unlink dev from the device chain */
258 write_lock_bh(&dev_base_lock);
259 list_del_rcu(&dev->dev_list);
260 hlist_del_rcu(&dev->name_hlist);
261 hlist_del_rcu(&dev->index_hlist);
262 write_unlock_bh(&dev_base_lock);
264 dev_base_seq_inc(dev_net(dev));
271 static RAW_NOTIFIER_HEAD(netdev_chain);
274 * Device drivers call our routines to queue packets here. We empty the
275 * queue in the local softnet handler.
278 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data);
279 EXPORT_PER_CPU_SYMBOL(softnet_data);
281 #ifdef CONFIG_LOCKDEP
283 * register_netdevice() inits txq->_xmit_lock and sets lockdep class
284 * according to dev->type
286 static const unsigned short netdev_lock_type[] = {
287 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25,
288 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET,
289 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM,
290 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP,
291 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD,
292 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25,
293 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP,
294 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD,
295 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI,
296 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE,
297 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET,
298 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL,
299 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM,
300 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE,
301 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE};
303 static const char *const netdev_lock_name[] = {
304 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25",
305 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET",
306 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM",
307 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP",
308 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD",
309 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25",
310 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP",
311 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD",
312 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI",
313 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE",
314 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET",
315 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL",
316 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM",
317 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE",
318 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"};
320 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)];
321 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)];
323 static inline unsigned short netdev_lock_pos(unsigned short dev_type)
327 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++)
328 if (netdev_lock_type[i] == dev_type)
330 /* the last key is used by default */
331 return ARRAY_SIZE(netdev_lock_type) - 1;
334 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
335 unsigned short dev_type)
339 i = netdev_lock_pos(dev_type);
340 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i],
341 netdev_lock_name[i]);
344 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
348 i = netdev_lock_pos(dev->type);
349 lockdep_set_class_and_name(&dev->addr_list_lock,
350 &netdev_addr_lock_key[i],
351 netdev_lock_name[i]);
354 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
355 unsigned short dev_type)
358 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
363 /*******************************************************************************
365 * Protocol management and registration routines
367 *******************************************************************************/
371 * Add a protocol ID to the list. Now that the input handler is
372 * smarter we can dispense with all the messy stuff that used to be
375 * BEWARE!!! Protocol handlers, mangling input packets,
376 * MUST BE last in hash buckets and checking protocol handlers
377 * MUST start from promiscuous ptype_all chain in net_bh.
378 * It is true now, do not change it.
379 * Explanation follows: if protocol handler, mangling packet, will
380 * be the first on list, it is not able to sense, that packet
381 * is cloned and should be copied-on-write, so that it will
382 * change it and subsequent readers will get broken packet.
386 static inline struct list_head *ptype_head(const struct packet_type *pt)
388 if (pt->type == htons(ETH_P_ALL))
389 return pt->dev ? &pt->dev->ptype_all : &ptype_all;
391 return pt->dev ? &pt->dev->ptype_specific :
392 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK];
396 * dev_add_pack - add packet handler
397 * @pt: packet type declaration
399 * Add a protocol handler to the networking stack. The passed &packet_type
400 * is linked into kernel lists and may not be freed until it has been
401 * removed from the kernel lists.
403 * This call does not sleep therefore it can not
404 * guarantee all CPU's that are in middle of receiving packets
405 * will see the new packet type (until the next received packet).
408 void dev_add_pack(struct packet_type *pt)
410 struct list_head *head = ptype_head(pt);
412 spin_lock(&ptype_lock);
413 list_add_rcu(&pt->list, head);
414 spin_unlock(&ptype_lock);
416 EXPORT_SYMBOL(dev_add_pack);
419 * __dev_remove_pack - remove packet handler
420 * @pt: packet type declaration
422 * Remove a protocol handler that was previously added to the kernel
423 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
424 * from the kernel lists and can be freed or reused once this function
427 * The packet type might still be in use by receivers
428 * and must not be freed until after all the CPU's have gone
429 * through a quiescent state.
431 void __dev_remove_pack(struct packet_type *pt)
433 struct list_head *head = ptype_head(pt);
434 struct packet_type *pt1;
436 spin_lock(&ptype_lock);
438 list_for_each_entry(pt1, head, list) {
440 list_del_rcu(&pt->list);
445 pr_warn("dev_remove_pack: %p not found\n", pt);
447 spin_unlock(&ptype_lock);
449 EXPORT_SYMBOL(__dev_remove_pack);
452 * dev_remove_pack - remove packet handler
453 * @pt: packet type declaration
455 * Remove a protocol handler that was previously added to the kernel
456 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
457 * from the kernel lists and can be freed or reused once this function
460 * This call sleeps to guarantee that no CPU is looking at the packet
463 void dev_remove_pack(struct packet_type *pt)
465 __dev_remove_pack(pt);
469 EXPORT_SYMBOL(dev_remove_pack);
473 * dev_add_offload - register offload handlers
474 * @po: protocol offload declaration
476 * Add protocol offload handlers to the networking stack. The passed
477 * &proto_offload is linked into kernel lists and may not be freed until
478 * it has been removed from the kernel lists.
480 * This call does not sleep therefore it can not
481 * guarantee all CPU's that are in middle of receiving packets
482 * will see the new offload handlers (until the next received packet).
484 void dev_add_offload(struct packet_offload *po)
486 struct packet_offload *elem;
488 spin_lock(&offload_lock);
489 list_for_each_entry(elem, &offload_base, list) {
490 if (po->priority < elem->priority)
493 list_add_rcu(&po->list, elem->list.prev);
494 spin_unlock(&offload_lock);
496 EXPORT_SYMBOL(dev_add_offload);
499 * __dev_remove_offload - remove offload handler
500 * @po: packet offload declaration
502 * Remove a protocol offload handler that was previously added to the
503 * kernel offload handlers by dev_add_offload(). The passed &offload_type
504 * is removed from the kernel lists and can be freed or reused once this
507 * The packet type might still be in use by receivers
508 * and must not be freed until after all the CPU's have gone
509 * through a quiescent state.
511 static void __dev_remove_offload(struct packet_offload *po)
513 struct list_head *head = &offload_base;
514 struct packet_offload *po1;
516 spin_lock(&offload_lock);
518 list_for_each_entry(po1, head, list) {
520 list_del_rcu(&po->list);
525 pr_warn("dev_remove_offload: %p not found\n", po);
527 spin_unlock(&offload_lock);
531 * dev_remove_offload - remove packet offload handler
532 * @po: packet offload declaration
534 * Remove a packet offload handler that was previously added to the kernel
535 * offload handlers by dev_add_offload(). The passed &offload_type is
536 * removed from the kernel lists and can be freed or reused once this
539 * This call sleeps to guarantee that no CPU is looking at the packet
542 void dev_remove_offload(struct packet_offload *po)
544 __dev_remove_offload(po);
548 EXPORT_SYMBOL(dev_remove_offload);
550 /******************************************************************************
552 * Device Boot-time Settings Routines
554 ******************************************************************************/
556 /* Boot time configuration table */
557 static struct netdev_boot_setup dev_boot_setup[NETDEV_BOOT_SETUP_MAX];
560 * netdev_boot_setup_add - add new setup entry
561 * @name: name of the device
562 * @map: configured settings for the device
564 * Adds new setup entry to the dev_boot_setup list. The function
565 * returns 0 on error and 1 on success. This is a generic routine to
568 static int netdev_boot_setup_add(char *name, struct ifmap *map)
570 struct netdev_boot_setup *s;
574 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
575 if (s[i].name[0] == '\0' || s[i].name[0] == ' ') {
576 memset(s[i].name, 0, sizeof(s[i].name));
577 strlcpy(s[i].name, name, IFNAMSIZ);
578 memcpy(&s[i].map, map, sizeof(s[i].map));
583 return i >= NETDEV_BOOT_SETUP_MAX ? 0 : 1;
587 * netdev_boot_setup_check - check boot time settings
588 * @dev: the netdevice
590 * Check boot time settings for the device.
591 * The found settings are set for the device to be used
592 * later in the device probing.
593 * Returns 0 if no settings found, 1 if they are.
595 int netdev_boot_setup_check(struct net_device *dev)
597 struct netdev_boot_setup *s = dev_boot_setup;
600 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
601 if (s[i].name[0] != '\0' && s[i].name[0] != ' ' &&
602 !strcmp(dev->name, s[i].name)) {
603 dev->irq = s[i].map.irq;
604 dev->base_addr = s[i].map.base_addr;
605 dev->mem_start = s[i].map.mem_start;
606 dev->mem_end = s[i].map.mem_end;
612 EXPORT_SYMBOL(netdev_boot_setup_check);
616 * netdev_boot_base - get address from boot time settings
617 * @prefix: prefix for network device
618 * @unit: id for network device
620 * Check boot time settings for the base address of device.
621 * The found settings are set for the device to be used
622 * later in the device probing.
623 * Returns 0 if no settings found.
625 unsigned long netdev_boot_base(const char *prefix, int unit)
627 const struct netdev_boot_setup *s = dev_boot_setup;
631 sprintf(name, "%s%d", prefix, unit);
634 * If device already registered then return base of 1
635 * to indicate not to probe for this interface
637 if (__dev_get_by_name(&init_net, name))
640 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++)
641 if (!strcmp(name, s[i].name))
642 return s[i].map.base_addr;
647 * Saves at boot time configured settings for any netdevice.
649 int __init netdev_boot_setup(char *str)
654 str = get_options(str, ARRAY_SIZE(ints), ints);
659 memset(&map, 0, sizeof(map));
663 map.base_addr = ints[2];
665 map.mem_start = ints[3];
667 map.mem_end = ints[4];
669 /* Add new entry to the list */
670 return netdev_boot_setup_add(str, &map);
673 __setup("netdev=", netdev_boot_setup);
675 /*******************************************************************************
677 * Device Interface Subroutines
679 *******************************************************************************/
682 * dev_get_iflink - get 'iflink' value of a interface
683 * @dev: targeted interface
685 * Indicates the ifindex the interface is linked to.
686 * Physical interfaces have the same 'ifindex' and 'iflink' values.
689 int dev_get_iflink(const struct net_device *dev)
691 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink)
692 return dev->netdev_ops->ndo_get_iflink(dev);
696 EXPORT_SYMBOL(dev_get_iflink);
699 * dev_fill_metadata_dst - Retrieve tunnel egress information.
700 * @dev: targeted interface
703 * For better visibility of tunnel traffic OVS needs to retrieve
704 * egress tunnel information for a packet. Following API allows
705 * user to get this info.
707 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb)
709 struct ip_tunnel_info *info;
711 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst)
714 info = skb_tunnel_info_unclone(skb);
717 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX)))
720 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb);
722 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst);
725 * __dev_get_by_name - find a device by its name
726 * @net: the applicable net namespace
727 * @name: name to find
729 * Find an interface by name. Must be called under RTNL semaphore
730 * or @dev_base_lock. If the name is found a pointer to the device
731 * is returned. If the name is not found then %NULL is returned. The
732 * reference counters are not incremented so the caller must be
733 * careful with locks.
736 struct net_device *__dev_get_by_name(struct net *net, const char *name)
738 struct net_device *dev;
739 struct hlist_head *head = dev_name_hash(net, name);
741 hlist_for_each_entry(dev, head, name_hlist)
742 if (!strncmp(dev->name, name, IFNAMSIZ))
747 EXPORT_SYMBOL(__dev_get_by_name);
750 * dev_get_by_name_rcu - find a device by its name
751 * @net: the applicable net namespace
752 * @name: name to find
754 * Find an interface by name.
755 * If the name is found a pointer to the device is returned.
756 * If the name is not found then %NULL is returned.
757 * The reference counters are not incremented so the caller must be
758 * careful with locks. The caller must hold RCU lock.
761 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name)
763 struct net_device *dev;
764 struct hlist_head *head = dev_name_hash(net, name);
766 hlist_for_each_entry_rcu(dev, head, name_hlist)
767 if (!strncmp(dev->name, name, IFNAMSIZ))
772 EXPORT_SYMBOL(dev_get_by_name_rcu);
775 * dev_get_by_name - find a device by its name
776 * @net: the applicable net namespace
777 * @name: name to find
779 * Find an interface by name. This can be called from any
780 * context and does its own locking. The returned handle has
781 * the usage count incremented and the caller must use dev_put() to
782 * release it when it is no longer needed. %NULL is returned if no
783 * matching device is found.
786 struct net_device *dev_get_by_name(struct net *net, const char *name)
788 struct net_device *dev;
791 dev = dev_get_by_name_rcu(net, name);
797 EXPORT_SYMBOL(dev_get_by_name);
800 * __dev_get_by_index - find a device by its ifindex
801 * @net: the applicable net namespace
802 * @ifindex: index of device
804 * Search for an interface by index. Returns %NULL if the device
805 * is not found or a pointer to the device. The device has not
806 * had its reference counter increased so the caller must be careful
807 * about locking. The caller must hold either the RTNL semaphore
811 struct net_device *__dev_get_by_index(struct net *net, int ifindex)
813 struct net_device *dev;
814 struct hlist_head *head = dev_index_hash(net, ifindex);
816 hlist_for_each_entry(dev, head, index_hlist)
817 if (dev->ifindex == ifindex)
822 EXPORT_SYMBOL(__dev_get_by_index);
825 * dev_get_by_index_rcu - find a device by its ifindex
826 * @net: the applicable net namespace
827 * @ifindex: index of device
829 * Search for an interface by index. Returns %NULL if the device
830 * is not found or a pointer to the device. The device has not
831 * had its reference counter increased so the caller must be careful
832 * about locking. The caller must hold RCU lock.
835 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex)
837 struct net_device *dev;
838 struct hlist_head *head = dev_index_hash(net, ifindex);
840 hlist_for_each_entry_rcu(dev, head, index_hlist)
841 if (dev->ifindex == ifindex)
846 EXPORT_SYMBOL(dev_get_by_index_rcu);
850 * dev_get_by_index - find a device by its ifindex
851 * @net: the applicable net namespace
852 * @ifindex: index of device
854 * Search for an interface by index. Returns NULL if the device
855 * is not found or a pointer to the device. The device returned has
856 * had a reference added and the pointer is safe until the user calls
857 * dev_put to indicate they have finished with it.
860 struct net_device *dev_get_by_index(struct net *net, int ifindex)
862 struct net_device *dev;
865 dev = dev_get_by_index_rcu(net, ifindex);
871 EXPORT_SYMBOL(dev_get_by_index);
874 * dev_get_by_napi_id - find a device by napi_id
875 * @napi_id: ID of the NAPI struct
877 * Search for an interface by NAPI ID. Returns %NULL if the device
878 * is not found or a pointer to the device. The device has not had
879 * its reference counter increased so the caller must be careful
880 * about locking. The caller must hold RCU lock.
883 struct net_device *dev_get_by_napi_id(unsigned int napi_id)
885 struct napi_struct *napi;
887 WARN_ON_ONCE(!rcu_read_lock_held());
889 if (napi_id < MIN_NAPI_ID)
892 napi = napi_by_id(napi_id);
894 return napi ? napi->dev : NULL;
896 EXPORT_SYMBOL(dev_get_by_napi_id);
899 * netdev_get_name - get a netdevice name, knowing its ifindex.
900 * @net: network namespace
901 * @name: a pointer to the buffer where the name will be stored.
902 * @ifindex: the ifindex of the interface to get the name from.
904 * The use of raw_seqcount_begin() and cond_resched() before
905 * retrying is required as we want to give the writers a chance
906 * to complete when CONFIG_PREEMPT is not set.
908 int netdev_get_name(struct net *net, char *name, int ifindex)
910 struct net_device *dev;
914 seq = raw_seqcount_begin(&devnet_rename_seq);
916 dev = dev_get_by_index_rcu(net, ifindex);
922 strcpy(name, dev->name);
924 if (read_seqcount_retry(&devnet_rename_seq, seq)) {
933 * dev_getbyhwaddr_rcu - find a device by its hardware address
934 * @net: the applicable net namespace
935 * @type: media type of device
936 * @ha: hardware address
938 * Search for an interface by MAC address. Returns NULL if the device
939 * is not found or a pointer to the device.
940 * The caller must hold RCU or RTNL.
941 * The returned device has not had its ref count increased
942 * and the caller must therefore be careful about locking
946 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type,
949 struct net_device *dev;
951 for_each_netdev_rcu(net, dev)
952 if (dev->type == type &&
953 !memcmp(dev->dev_addr, ha, dev->addr_len))
958 EXPORT_SYMBOL(dev_getbyhwaddr_rcu);
960 struct net_device *__dev_getfirstbyhwtype(struct net *net, unsigned short type)
962 struct net_device *dev;
965 for_each_netdev(net, dev)
966 if (dev->type == type)
971 EXPORT_SYMBOL(__dev_getfirstbyhwtype);
973 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type)
975 struct net_device *dev, *ret = NULL;
978 for_each_netdev_rcu(net, dev)
979 if (dev->type == type) {
987 EXPORT_SYMBOL(dev_getfirstbyhwtype);
990 * __dev_get_by_flags - find any device with given flags
991 * @net: the applicable net namespace
992 * @if_flags: IFF_* values
993 * @mask: bitmask of bits in if_flags to check
995 * Search for any interface with the given flags. Returns NULL if a device
996 * is not found or a pointer to the device. Must be called inside
997 * rtnl_lock(), and result refcount is unchanged.
1000 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags,
1001 unsigned short mask)
1003 struct net_device *dev, *ret;
1008 for_each_netdev(net, dev) {
1009 if (((dev->flags ^ if_flags) & mask) == 0) {
1016 EXPORT_SYMBOL(__dev_get_by_flags);
1019 * dev_valid_name - check if name is okay for network device
1020 * @name: name string
1022 * Network device names need to be valid file names to
1023 * to allow sysfs to work. We also disallow any kind of
1026 bool dev_valid_name(const char *name)
1030 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ)
1032 if (!strcmp(name, ".") || !strcmp(name, ".."))
1036 if (*name == '/' || *name == ':' || isspace(*name))
1042 EXPORT_SYMBOL(dev_valid_name);
1045 * __dev_alloc_name - allocate a name for a device
1046 * @net: network namespace to allocate the device name in
1047 * @name: name format string
1048 * @buf: scratch buffer and result name string
1050 * Passed a format string - eg "lt%d" it will try and find a suitable
1051 * id. It scans list of devices to build up a free map, then chooses
1052 * the first empty slot. The caller must hold the dev_base or rtnl lock
1053 * while allocating the name and adding the device in order to avoid
1055 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1056 * Returns the number of the unit assigned or a negative errno code.
1059 static int __dev_alloc_name(struct net *net, const char *name, char *buf)
1063 const int max_netdevices = 8*PAGE_SIZE;
1064 unsigned long *inuse;
1065 struct net_device *d;
1067 if (!dev_valid_name(name))
1070 p = strchr(name, '%');
1073 * Verify the string as this thing may have come from
1074 * the user. There must be either one "%d" and no other "%"
1077 if (p[1] != 'd' || strchr(p + 2, '%'))
1080 /* Use one page as a bit array of possible slots */
1081 inuse = (unsigned long *) get_zeroed_page(GFP_ATOMIC);
1085 for_each_netdev(net, d) {
1086 if (!sscanf(d->name, name, &i))
1088 if (i < 0 || i >= max_netdevices)
1091 /* avoid cases where sscanf is not exact inverse of printf */
1092 snprintf(buf, IFNAMSIZ, name, i);
1093 if (!strncmp(buf, d->name, IFNAMSIZ))
1097 i = find_first_zero_bit(inuse, max_netdevices);
1098 free_page((unsigned long) inuse);
1101 snprintf(buf, IFNAMSIZ, name, i);
1102 if (!__dev_get_by_name(net, buf))
1105 /* It is possible to run out of possible slots
1106 * when the name is long and there isn't enough space left
1107 * for the digits, or if all bits are used.
1112 static int dev_alloc_name_ns(struct net *net,
1113 struct net_device *dev,
1120 ret = __dev_alloc_name(net, name, buf);
1122 strlcpy(dev->name, buf, IFNAMSIZ);
1127 * dev_alloc_name - allocate a name for a device
1129 * @name: name format string
1131 * Passed a format string - eg "lt%d" it will try and find a suitable
1132 * id. It scans list of devices to build up a free map, then chooses
1133 * the first empty slot. The caller must hold the dev_base or rtnl lock
1134 * while allocating the name and adding the device in order to avoid
1136 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1137 * Returns the number of the unit assigned or a negative errno code.
1140 int dev_alloc_name(struct net_device *dev, const char *name)
1142 return dev_alloc_name_ns(dev_net(dev), dev, name);
1144 EXPORT_SYMBOL(dev_alloc_name);
1146 int dev_get_valid_name(struct net *net, struct net_device *dev,
1151 if (!dev_valid_name(name))
1154 if (strchr(name, '%'))
1155 return dev_alloc_name_ns(net, dev, name);
1156 else if (__dev_get_by_name(net, name))
1158 else if (dev->name != name)
1159 strlcpy(dev->name, name, IFNAMSIZ);
1163 EXPORT_SYMBOL(dev_get_valid_name);
1166 * dev_change_name - change name of a device
1168 * @newname: name (or format string) must be at least IFNAMSIZ
1170 * Change name of a device, can pass format strings "eth%d".
1173 int dev_change_name(struct net_device *dev, const char *newname)
1175 unsigned char old_assign_type;
1176 char oldname[IFNAMSIZ];
1182 BUG_ON(!dev_net(dev));
1185 if (dev->flags & IFF_UP)
1188 write_seqcount_begin(&devnet_rename_seq);
1190 if (strncmp(newname, dev->name, IFNAMSIZ) == 0) {
1191 write_seqcount_end(&devnet_rename_seq);
1195 memcpy(oldname, dev->name, IFNAMSIZ);
1197 err = dev_get_valid_name(net, dev, newname);
1199 write_seqcount_end(&devnet_rename_seq);
1203 if (oldname[0] && !strchr(oldname, '%'))
1204 netdev_info(dev, "renamed from %s\n", oldname);
1206 old_assign_type = dev->name_assign_type;
1207 dev->name_assign_type = NET_NAME_RENAMED;
1210 ret = device_rename(&dev->dev, dev->name);
1212 memcpy(dev->name, oldname, IFNAMSIZ);
1213 dev->name_assign_type = old_assign_type;
1214 write_seqcount_end(&devnet_rename_seq);
1218 write_seqcount_end(&devnet_rename_seq);
1220 netdev_adjacent_rename_links(dev, oldname);
1222 write_lock_bh(&dev_base_lock);
1223 hlist_del_rcu(&dev->name_hlist);
1224 write_unlock_bh(&dev_base_lock);
1228 write_lock_bh(&dev_base_lock);
1229 hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
1230 write_unlock_bh(&dev_base_lock);
1232 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev);
1233 ret = notifier_to_errno(ret);
1236 /* err >= 0 after dev_alloc_name() or stores the first errno */
1239 write_seqcount_begin(&devnet_rename_seq);
1240 memcpy(dev->name, oldname, IFNAMSIZ);
1241 memcpy(oldname, newname, IFNAMSIZ);
1242 dev->name_assign_type = old_assign_type;
1243 old_assign_type = NET_NAME_RENAMED;
1246 pr_err("%s: name change rollback failed: %d\n",
1255 * dev_set_alias - change ifalias of a device
1257 * @alias: name up to IFALIASZ
1258 * @len: limit of bytes to copy from info
1260 * Set ifalias for a device,
1262 int dev_set_alias(struct net_device *dev, const char *alias, size_t len)
1264 struct dev_ifalias *new_alias = NULL;
1266 if (len >= IFALIASZ)
1270 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL);
1274 memcpy(new_alias->ifalias, alias, len);
1275 new_alias->ifalias[len] = 0;
1278 mutex_lock(&ifalias_mutex);
1279 rcu_swap_protected(dev->ifalias, new_alias,
1280 mutex_is_locked(&ifalias_mutex));
1281 mutex_unlock(&ifalias_mutex);
1284 kfree_rcu(new_alias, rcuhead);
1288 EXPORT_SYMBOL(dev_set_alias);
1291 * dev_get_alias - get ifalias of a device
1293 * @name: buffer to store name of ifalias
1294 * @len: size of buffer
1296 * get ifalias for a device. Caller must make sure dev cannot go
1297 * away, e.g. rcu read lock or own a reference count to device.
1299 int dev_get_alias(const struct net_device *dev, char *name, size_t len)
1301 const struct dev_ifalias *alias;
1305 alias = rcu_dereference(dev->ifalias);
1307 ret = snprintf(name, len, "%s", alias->ifalias);
1314 * netdev_features_change - device changes features
1315 * @dev: device to cause notification
1317 * Called to indicate a device has changed features.
1319 void netdev_features_change(struct net_device *dev)
1321 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev);
1323 EXPORT_SYMBOL(netdev_features_change);
1326 * netdev_state_change - device changes state
1327 * @dev: device to cause notification
1329 * Called to indicate a device has changed state. This function calls
1330 * the notifier chains for netdev_chain and sends a NEWLINK message
1331 * to the routing socket.
1333 void netdev_state_change(struct net_device *dev)
1335 if (dev->flags & IFF_UP) {
1336 struct netdev_notifier_change_info change_info = {
1340 call_netdevice_notifiers_info(NETDEV_CHANGE,
1342 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL);
1345 EXPORT_SYMBOL(netdev_state_change);
1348 * netdev_notify_peers - notify network peers about existence of @dev
1349 * @dev: network device
1351 * Generate traffic such that interested network peers are aware of
1352 * @dev, such as by generating a gratuitous ARP. This may be used when
1353 * a device wants to inform the rest of the network about some sort of
1354 * reconfiguration such as a failover event or virtual machine
1357 void netdev_notify_peers(struct net_device *dev)
1360 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
1361 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev);
1364 EXPORT_SYMBOL(netdev_notify_peers);
1366 static int __dev_open(struct net_device *dev)
1368 const struct net_device_ops *ops = dev->netdev_ops;
1373 if (!netif_device_present(dev))
1376 /* Block netpoll from trying to do any rx path servicing.
1377 * If we don't do this there is a chance ndo_poll_controller
1378 * or ndo_poll may be running while we open the device
1380 netpoll_poll_disable(dev);
1382 ret = call_netdevice_notifiers(NETDEV_PRE_UP, dev);
1383 ret = notifier_to_errno(ret);
1387 set_bit(__LINK_STATE_START, &dev->state);
1389 if (ops->ndo_validate_addr)
1390 ret = ops->ndo_validate_addr(dev);
1392 if (!ret && ops->ndo_open)
1393 ret = ops->ndo_open(dev);
1395 netpoll_poll_enable(dev);
1398 clear_bit(__LINK_STATE_START, &dev->state);
1400 dev->flags |= IFF_UP;
1401 dev_set_rx_mode(dev);
1403 add_device_randomness(dev->dev_addr, dev->addr_len);
1410 * dev_open - prepare an interface for use.
1411 * @dev: device to open
1413 * Takes a device from down to up state. The device's private open
1414 * function is invoked and then the multicast lists are loaded. Finally
1415 * the device is moved into the up state and a %NETDEV_UP message is
1416 * sent to the netdev notifier chain.
1418 * Calling this function on an active interface is a nop. On a failure
1419 * a negative errno code is returned.
1421 int dev_open(struct net_device *dev)
1425 if (dev->flags & IFF_UP)
1428 ret = __dev_open(dev);
1432 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1433 call_netdevice_notifiers(NETDEV_UP, dev);
1437 EXPORT_SYMBOL(dev_open);
1439 static void __dev_close_many(struct list_head *head)
1441 struct net_device *dev;
1446 list_for_each_entry(dev, head, close_list) {
1447 /* Temporarily disable netpoll until the interface is down */
1448 netpoll_poll_disable(dev);
1450 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev);
1452 clear_bit(__LINK_STATE_START, &dev->state);
1454 /* Synchronize to scheduled poll. We cannot touch poll list, it
1455 * can be even on different cpu. So just clear netif_running().
1457 * dev->stop() will invoke napi_disable() on all of it's
1458 * napi_struct instances on this device.
1460 smp_mb__after_atomic(); /* Commit netif_running(). */
1463 dev_deactivate_many(head);
1465 list_for_each_entry(dev, head, close_list) {
1466 const struct net_device_ops *ops = dev->netdev_ops;
1469 * Call the device specific close. This cannot fail.
1470 * Only if device is UP
1472 * We allow it to be called even after a DETACH hot-plug
1478 dev->flags &= ~IFF_UP;
1479 netpoll_poll_enable(dev);
1483 static void __dev_close(struct net_device *dev)
1487 list_add(&dev->close_list, &single);
1488 __dev_close_many(&single);
1492 void dev_close_many(struct list_head *head, bool unlink)
1494 struct net_device *dev, *tmp;
1496 /* Remove the devices that don't need to be closed */
1497 list_for_each_entry_safe(dev, tmp, head, close_list)
1498 if (!(dev->flags & IFF_UP))
1499 list_del_init(&dev->close_list);
1501 __dev_close_many(head);
1503 list_for_each_entry_safe(dev, tmp, head, close_list) {
1504 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1505 call_netdevice_notifiers(NETDEV_DOWN, dev);
1507 list_del_init(&dev->close_list);
1510 EXPORT_SYMBOL(dev_close_many);
1513 * dev_close - shutdown an interface.
1514 * @dev: device to shutdown
1516 * This function moves an active device into down state. A
1517 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device
1518 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier
1521 void dev_close(struct net_device *dev)
1523 if (dev->flags & IFF_UP) {
1526 list_add(&dev->close_list, &single);
1527 dev_close_many(&single, true);
1531 EXPORT_SYMBOL(dev_close);
1535 * dev_disable_lro - disable Large Receive Offload on a device
1538 * Disable Large Receive Offload (LRO) on a net device. Must be
1539 * called under RTNL. This is needed if received packets may be
1540 * forwarded to another interface.
1542 void dev_disable_lro(struct net_device *dev)
1544 struct net_device *lower_dev;
1545 struct list_head *iter;
1547 dev->wanted_features &= ~NETIF_F_LRO;
1548 netdev_update_features(dev);
1550 if (unlikely(dev->features & NETIF_F_LRO))
1551 netdev_WARN(dev, "failed to disable LRO!\n");
1553 netdev_for_each_lower_dev(dev, lower_dev, iter)
1554 dev_disable_lro(lower_dev);
1556 EXPORT_SYMBOL(dev_disable_lro);
1559 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device
1562 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be
1563 * called under RTNL. This is needed if Generic XDP is installed on
1566 static void dev_disable_gro_hw(struct net_device *dev)
1568 dev->wanted_features &= ~NETIF_F_GRO_HW;
1569 netdev_update_features(dev);
1571 if (unlikely(dev->features & NETIF_F_GRO_HW))
1572 netdev_WARN(dev, "failed to disable GRO_HW!\n");
1575 const char *netdev_cmd_to_name(enum netdev_cmd cmd)
1578 case NETDEV_##val: \
1579 return "NETDEV_" __stringify(val);
1581 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER)
1582 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE)
1583 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE)
1584 N(POST_INIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) N(CHANGEUPPER)
1585 N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) N(BONDING_INFO)
1586 N(PRECHANGEUPPER) N(CHANGELOWERSTATE) N(UDP_TUNNEL_PUSH_INFO)
1587 N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN)
1588 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO)
1589 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO)
1592 return "UNKNOWN_NETDEV_EVENT";
1594 EXPORT_SYMBOL_GPL(netdev_cmd_to_name);
1596 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val,
1597 struct net_device *dev)
1599 struct netdev_notifier_info info = {
1603 return nb->notifier_call(nb, val, &info);
1606 static int dev_boot_phase = 1;
1609 * register_netdevice_notifier - register a network notifier block
1612 * Register a notifier to be called when network device events occur.
1613 * The notifier passed is linked into the kernel structures and must
1614 * not be reused until it has been unregistered. A negative errno code
1615 * is returned on a failure.
1617 * When registered all registration and up events are replayed
1618 * to the new notifier to allow device to have a race free
1619 * view of the network device list.
1622 int register_netdevice_notifier(struct notifier_block *nb)
1624 struct net_device *dev;
1625 struct net_device *last;
1629 /* Close race with setup_net() and cleanup_net() */
1630 down_write(&pernet_ops_rwsem);
1632 err = raw_notifier_chain_register(&netdev_chain, nb);
1638 for_each_netdev(net, dev) {
1639 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev);
1640 err = notifier_to_errno(err);
1644 if (!(dev->flags & IFF_UP))
1647 call_netdevice_notifier(nb, NETDEV_UP, dev);
1653 up_write(&pernet_ops_rwsem);
1659 for_each_netdev(net, dev) {
1663 if (dev->flags & IFF_UP) {
1664 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1666 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1668 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1673 raw_notifier_chain_unregister(&netdev_chain, nb);
1676 EXPORT_SYMBOL(register_netdevice_notifier);
1679 * unregister_netdevice_notifier - unregister a network notifier block
1682 * Unregister a notifier previously registered by
1683 * register_netdevice_notifier(). The notifier is unlinked into the
1684 * kernel structures and may then be reused. A negative errno code
1685 * is returned on a failure.
1687 * After unregistering unregister and down device events are synthesized
1688 * for all devices on the device list to the removed notifier to remove
1689 * the need for special case cleanup code.
1692 int unregister_netdevice_notifier(struct notifier_block *nb)
1694 struct net_device *dev;
1698 /* Close race with setup_net() and cleanup_net() */
1699 down_write(&pernet_ops_rwsem);
1701 err = raw_notifier_chain_unregister(&netdev_chain, nb);
1706 for_each_netdev(net, dev) {
1707 if (dev->flags & IFF_UP) {
1708 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1710 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1712 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1717 up_write(&pernet_ops_rwsem);
1720 EXPORT_SYMBOL(unregister_netdevice_notifier);
1723 * call_netdevice_notifiers_info - call all network notifier blocks
1724 * @val: value passed unmodified to notifier function
1725 * @info: notifier information data
1727 * Call all network notifier blocks. Parameters and return value
1728 * are as for raw_notifier_call_chain().
1731 static int call_netdevice_notifiers_info(unsigned long val,
1732 struct netdev_notifier_info *info)
1735 return raw_notifier_call_chain(&netdev_chain, val, info);
1739 * call_netdevice_notifiers - call all network notifier blocks
1740 * @val: value passed unmodified to notifier function
1741 * @dev: net_device pointer passed unmodified to notifier function
1743 * Call all network notifier blocks. Parameters and return value
1744 * are as for raw_notifier_call_chain().
1747 int call_netdevice_notifiers(unsigned long val, struct net_device *dev)
1749 struct netdev_notifier_info info = {
1753 return call_netdevice_notifiers_info(val, &info);
1755 EXPORT_SYMBOL(call_netdevice_notifiers);
1757 #ifdef CONFIG_NET_INGRESS
1758 static struct static_key ingress_needed __read_mostly;
1760 void net_inc_ingress_queue(void)
1762 static_key_slow_inc(&ingress_needed);
1764 EXPORT_SYMBOL_GPL(net_inc_ingress_queue);
1766 void net_dec_ingress_queue(void)
1768 static_key_slow_dec(&ingress_needed);
1770 EXPORT_SYMBOL_GPL(net_dec_ingress_queue);
1773 #ifdef CONFIG_NET_EGRESS
1774 static struct static_key egress_needed __read_mostly;
1776 void net_inc_egress_queue(void)
1778 static_key_slow_inc(&egress_needed);
1780 EXPORT_SYMBOL_GPL(net_inc_egress_queue);
1782 void net_dec_egress_queue(void)
1784 static_key_slow_dec(&egress_needed);
1786 EXPORT_SYMBOL_GPL(net_dec_egress_queue);
1789 static struct static_key netstamp_needed __read_mostly;
1790 #ifdef HAVE_JUMP_LABEL
1791 static atomic_t netstamp_needed_deferred;
1792 static atomic_t netstamp_wanted;
1793 static void netstamp_clear(struct work_struct *work)
1795 int deferred = atomic_xchg(&netstamp_needed_deferred, 0);
1798 wanted = atomic_add_return(deferred, &netstamp_wanted);
1800 static_key_enable(&netstamp_needed);
1802 static_key_disable(&netstamp_needed);
1804 static DECLARE_WORK(netstamp_work, netstamp_clear);
1807 void net_enable_timestamp(void)
1809 #ifdef HAVE_JUMP_LABEL
1813 wanted = atomic_read(&netstamp_wanted);
1816 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted + 1) == wanted)
1819 atomic_inc(&netstamp_needed_deferred);
1820 schedule_work(&netstamp_work);
1822 static_key_slow_inc(&netstamp_needed);
1825 EXPORT_SYMBOL(net_enable_timestamp);
1827 void net_disable_timestamp(void)
1829 #ifdef HAVE_JUMP_LABEL
1833 wanted = atomic_read(&netstamp_wanted);
1836 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted - 1) == wanted)
1839 atomic_dec(&netstamp_needed_deferred);
1840 schedule_work(&netstamp_work);
1842 static_key_slow_dec(&netstamp_needed);
1845 EXPORT_SYMBOL(net_disable_timestamp);
1847 static inline void net_timestamp_set(struct sk_buff *skb)
1850 if (static_key_false(&netstamp_needed))
1851 __net_timestamp(skb);
1854 #define net_timestamp_check(COND, SKB) \
1855 if (static_key_false(&netstamp_needed)) { \
1856 if ((COND) && !(SKB)->tstamp) \
1857 __net_timestamp(SKB); \
1860 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb)
1864 if (!(dev->flags & IFF_UP))
1867 len = dev->mtu + dev->hard_header_len + VLAN_HLEN;
1868 if (skb->len <= len)
1871 /* if TSO is enabled, we don't care about the length as the packet
1872 * could be forwarded without being segmented before
1874 if (skb_is_gso(skb))
1879 EXPORT_SYMBOL_GPL(is_skb_forwardable);
1881 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1883 int ret = ____dev_forward_skb(dev, skb);
1886 skb->protocol = eth_type_trans(skb, dev);
1887 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
1892 EXPORT_SYMBOL_GPL(__dev_forward_skb);
1895 * dev_forward_skb - loopback an skb to another netif
1897 * @dev: destination network device
1898 * @skb: buffer to forward
1901 * NET_RX_SUCCESS (no congestion)
1902 * NET_RX_DROP (packet was dropped, but freed)
1904 * dev_forward_skb can be used for injecting an skb from the
1905 * start_xmit function of one device into the receive queue
1906 * of another device.
1908 * The receiving device may be in another namespace, so
1909 * we have to clear all information in the skb that could
1910 * impact namespace isolation.
1912 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1914 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb);
1916 EXPORT_SYMBOL_GPL(dev_forward_skb);
1918 static inline int deliver_skb(struct sk_buff *skb,
1919 struct packet_type *pt_prev,
1920 struct net_device *orig_dev)
1922 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
1924 refcount_inc(&skb->users);
1925 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
1928 static inline void deliver_ptype_list_skb(struct sk_buff *skb,
1929 struct packet_type **pt,
1930 struct net_device *orig_dev,
1932 struct list_head *ptype_list)
1934 struct packet_type *ptype, *pt_prev = *pt;
1936 list_for_each_entry_rcu(ptype, ptype_list, list) {
1937 if (ptype->type != type)
1940 deliver_skb(skb, pt_prev, orig_dev);
1946 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb)
1948 if (!ptype->af_packet_priv || !skb->sk)
1951 if (ptype->id_match)
1952 return ptype->id_match(ptype, skb->sk);
1953 else if ((struct sock *)ptype->af_packet_priv == skb->sk)
1960 * Support routine. Sends outgoing frames to any network
1961 * taps currently in use.
1964 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
1966 struct packet_type *ptype;
1967 struct sk_buff *skb2 = NULL;
1968 struct packet_type *pt_prev = NULL;
1969 struct list_head *ptype_list = &ptype_all;
1973 list_for_each_entry_rcu(ptype, ptype_list, list) {
1974 /* Never send packets back to the socket
1975 * they originated from - MvS (miquels@drinkel.ow.org)
1977 if (skb_loop_sk(ptype, skb))
1981 deliver_skb(skb2, pt_prev, skb->dev);
1986 /* need to clone skb, done only once */
1987 skb2 = skb_clone(skb, GFP_ATOMIC);
1991 net_timestamp_set(skb2);
1993 /* skb->nh should be correctly
1994 * set by sender, so that the second statement is
1995 * just protection against buggy protocols.
1997 skb_reset_mac_header(skb2);
1999 if (skb_network_header(skb2) < skb2->data ||
2000 skb_network_header(skb2) > skb_tail_pointer(skb2)) {
2001 net_crit_ratelimited("protocol %04x is buggy, dev %s\n",
2002 ntohs(skb2->protocol),
2004 skb_reset_network_header(skb2);
2007 skb2->transport_header = skb2->network_header;
2008 skb2->pkt_type = PACKET_OUTGOING;
2012 if (ptype_list == &ptype_all) {
2013 ptype_list = &dev->ptype_all;
2018 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC))
2019 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
2025 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit);
2028 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change
2029 * @dev: Network device
2030 * @txq: number of queues available
2032 * If real_num_tx_queues is changed the tc mappings may no longer be
2033 * valid. To resolve this verify the tc mapping remains valid and if
2034 * not NULL the mapping. With no priorities mapping to this
2035 * offset/count pair it will no longer be used. In the worst case TC0
2036 * is invalid nothing can be done so disable priority mappings. If is
2037 * expected that drivers will fix this mapping if they can before
2038 * calling netif_set_real_num_tx_queues.
2040 static void netif_setup_tc(struct net_device *dev, unsigned int txq)
2043 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2045 /* If TC0 is invalidated disable TC mapping */
2046 if (tc->offset + tc->count > txq) {
2047 pr_warn("Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n");
2052 /* Invalidated prio to tc mappings set to TC0 */
2053 for (i = 1; i < TC_BITMASK + 1; i++) {
2054 int q = netdev_get_prio_tc_map(dev, i);
2056 tc = &dev->tc_to_txq[q];
2057 if (tc->offset + tc->count > txq) {
2058 pr_warn("Number of in use tx queues changed. Priority %i to tc mapping %i is no longer valid. Setting map to 0\n",
2060 netdev_set_prio_tc_map(dev, i, 0);
2065 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq)
2068 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2071 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) {
2072 if ((txq - tc->offset) < tc->count)
2081 EXPORT_SYMBOL(netdev_txq_to_tc);
2084 static DEFINE_MUTEX(xps_map_mutex);
2085 #define xmap_dereference(P) \
2086 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex))
2088 static bool remove_xps_queue(struct xps_dev_maps *dev_maps,
2091 struct xps_map *map = NULL;
2095 map = xmap_dereference(dev_maps->cpu_map[tci]);
2099 for (pos = map->len; pos--;) {
2100 if (map->queues[pos] != index)
2104 map->queues[pos] = map->queues[--map->len];
2108 RCU_INIT_POINTER(dev_maps->cpu_map[tci], NULL);
2109 kfree_rcu(map, rcu);
2116 static bool remove_xps_queue_cpu(struct net_device *dev,
2117 struct xps_dev_maps *dev_maps,
2118 int cpu, u16 offset, u16 count)
2120 int num_tc = dev->num_tc ? : 1;
2121 bool active = false;
2124 for (tci = cpu * num_tc; num_tc--; tci++) {
2127 for (i = count, j = offset; i--; j++) {
2128 if (!remove_xps_queue(dev_maps, cpu, j))
2138 static void netif_reset_xps_queues(struct net_device *dev, u16 offset,
2141 struct xps_dev_maps *dev_maps;
2143 bool active = false;
2145 mutex_lock(&xps_map_mutex);
2146 dev_maps = xmap_dereference(dev->xps_maps);
2151 for_each_possible_cpu(cpu)
2152 active |= remove_xps_queue_cpu(dev, dev_maps, cpu,
2156 RCU_INIT_POINTER(dev->xps_maps, NULL);
2157 kfree_rcu(dev_maps, rcu);
2160 for (i = offset + (count - 1); count--; i--)
2161 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, i),
2165 mutex_unlock(&xps_map_mutex);
2168 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index)
2170 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index);
2173 static struct xps_map *expand_xps_map(struct xps_map *map,
2176 struct xps_map *new_map;
2177 int alloc_len = XPS_MIN_MAP_ALLOC;
2180 for (pos = 0; map && pos < map->len; pos++) {
2181 if (map->queues[pos] != index)
2186 /* Need to add queue to this CPU's existing map */
2188 if (pos < map->alloc_len)
2191 alloc_len = map->alloc_len * 2;
2194 /* Need to allocate new map to store queue on this CPU's map */
2195 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL,
2200 for (i = 0; i < pos; i++)
2201 new_map->queues[i] = map->queues[i];
2202 new_map->alloc_len = alloc_len;
2208 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask,
2211 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL;
2212 int i, cpu, tci, numa_node_id = -2;
2213 int maps_sz, num_tc = 1, tc = 0;
2214 struct xps_map *map, *new_map;
2215 bool active = false;
2218 num_tc = dev->num_tc;
2219 tc = netdev_txq_to_tc(dev, index);
2224 maps_sz = XPS_DEV_MAPS_SIZE(num_tc);
2225 if (maps_sz < L1_CACHE_BYTES)
2226 maps_sz = L1_CACHE_BYTES;
2228 mutex_lock(&xps_map_mutex);
2230 dev_maps = xmap_dereference(dev->xps_maps);
2232 /* allocate memory for queue storage */
2233 for_each_cpu_and(cpu, cpu_online_mask, mask) {
2235 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL);
2236 if (!new_dev_maps) {
2237 mutex_unlock(&xps_map_mutex);
2241 tci = cpu * num_tc + tc;
2242 map = dev_maps ? xmap_dereference(dev_maps->cpu_map[tci]) :
2245 map = expand_xps_map(map, cpu, index);
2249 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2253 goto out_no_new_maps;
2255 for_each_possible_cpu(cpu) {
2256 /* copy maps belonging to foreign traffic classes */
2257 for (i = tc, tci = cpu * num_tc; dev_maps && i--; tci++) {
2258 /* fill in the new device map from the old device map */
2259 map = xmap_dereference(dev_maps->cpu_map[tci]);
2260 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2263 /* We need to explicitly update tci as prevous loop
2264 * could break out early if dev_maps is NULL.
2266 tci = cpu * num_tc + tc;
2268 if (cpumask_test_cpu(cpu, mask) && cpu_online(cpu)) {
2269 /* add queue to CPU maps */
2272 map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2273 while ((pos < map->len) && (map->queues[pos] != index))
2276 if (pos == map->len)
2277 map->queues[map->len++] = index;
2279 if (numa_node_id == -2)
2280 numa_node_id = cpu_to_node(cpu);
2281 else if (numa_node_id != cpu_to_node(cpu))
2284 } else if (dev_maps) {
2285 /* fill in the new device map from the old device map */
2286 map = xmap_dereference(dev_maps->cpu_map[tci]);
2287 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2290 /* copy maps belonging to foreign traffic classes */
2291 for (i = num_tc - tc, tci++; dev_maps && --i; tci++) {
2292 /* fill in the new device map from the old device map */
2293 map = xmap_dereference(dev_maps->cpu_map[tci]);
2294 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2298 rcu_assign_pointer(dev->xps_maps, new_dev_maps);
2300 /* Cleanup old maps */
2302 goto out_no_old_maps;
2304 for_each_possible_cpu(cpu) {
2305 for (i = num_tc, tci = cpu * num_tc; i--; tci++) {
2306 new_map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2307 map = xmap_dereference(dev_maps->cpu_map[tci]);
2308 if (map && map != new_map)
2309 kfree_rcu(map, rcu);
2313 kfree_rcu(dev_maps, rcu);
2316 dev_maps = new_dev_maps;
2320 /* update Tx queue numa node */
2321 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index),
2322 (numa_node_id >= 0) ? numa_node_id :
2328 /* removes queue from unused CPUs */
2329 for_each_possible_cpu(cpu) {
2330 for (i = tc, tci = cpu * num_tc; i--; tci++)
2331 active |= remove_xps_queue(dev_maps, tci, index);
2332 if (!cpumask_test_cpu(cpu, mask) || !cpu_online(cpu))
2333 active |= remove_xps_queue(dev_maps, tci, index);
2334 for (i = num_tc - tc, tci++; --i; tci++)
2335 active |= remove_xps_queue(dev_maps, tci, index);
2338 /* free map if not active */
2340 RCU_INIT_POINTER(dev->xps_maps, NULL);
2341 kfree_rcu(dev_maps, rcu);
2345 mutex_unlock(&xps_map_mutex);
2349 /* remove any maps that we added */
2350 for_each_possible_cpu(cpu) {
2351 for (i = num_tc, tci = cpu * num_tc; i--; tci++) {
2352 new_map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2354 xmap_dereference(dev_maps->cpu_map[tci]) :
2356 if (new_map && new_map != map)
2361 mutex_unlock(&xps_map_mutex);
2363 kfree(new_dev_maps);
2366 EXPORT_SYMBOL(netif_set_xps_queue);
2369 void netdev_reset_tc(struct net_device *dev)
2372 netif_reset_xps_queues_gt(dev, 0);
2375 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq));
2376 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map));
2378 EXPORT_SYMBOL(netdev_reset_tc);
2380 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset)
2382 if (tc >= dev->num_tc)
2386 netif_reset_xps_queues(dev, offset, count);
2388 dev->tc_to_txq[tc].count = count;
2389 dev->tc_to_txq[tc].offset = offset;
2392 EXPORT_SYMBOL(netdev_set_tc_queue);
2394 int netdev_set_num_tc(struct net_device *dev, u8 num_tc)
2396 if (num_tc > TC_MAX_QUEUE)
2400 netif_reset_xps_queues_gt(dev, 0);
2402 dev->num_tc = num_tc;
2405 EXPORT_SYMBOL(netdev_set_num_tc);
2408 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues
2409 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed.
2411 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq)
2416 disabling = txq < dev->real_num_tx_queues;
2418 if (txq < 1 || txq > dev->num_tx_queues)
2421 if (dev->reg_state == NETREG_REGISTERED ||
2422 dev->reg_state == NETREG_UNREGISTERING) {
2425 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues,
2431 netif_setup_tc(dev, txq);
2433 dev->real_num_tx_queues = txq;
2437 qdisc_reset_all_tx_gt(dev, txq);
2439 netif_reset_xps_queues_gt(dev, txq);
2443 dev->real_num_tx_queues = txq;
2448 EXPORT_SYMBOL(netif_set_real_num_tx_queues);
2452 * netif_set_real_num_rx_queues - set actual number of RX queues used
2453 * @dev: Network device
2454 * @rxq: Actual number of RX queues
2456 * This must be called either with the rtnl_lock held or before
2457 * registration of the net device. Returns 0 on success, or a
2458 * negative error code. If called before registration, it always
2461 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq)
2465 if (rxq < 1 || rxq > dev->num_rx_queues)
2468 if (dev->reg_state == NETREG_REGISTERED) {
2471 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues,
2477 dev->real_num_rx_queues = rxq;
2480 EXPORT_SYMBOL(netif_set_real_num_rx_queues);
2484 * netif_get_num_default_rss_queues - default number of RSS queues
2486 * This routine should set an upper limit on the number of RSS queues
2487 * used by default by multiqueue devices.
2489 int netif_get_num_default_rss_queues(void)
2491 return is_kdump_kernel() ?
2492 1 : min_t(int, DEFAULT_MAX_NUM_RSS_QUEUES, num_online_cpus());
2494 EXPORT_SYMBOL(netif_get_num_default_rss_queues);
2496 static void __netif_reschedule(struct Qdisc *q)
2498 struct softnet_data *sd;
2499 unsigned long flags;
2501 local_irq_save(flags);
2502 sd = this_cpu_ptr(&softnet_data);
2503 q->next_sched = NULL;
2504 *sd->output_queue_tailp = q;
2505 sd->output_queue_tailp = &q->next_sched;
2506 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2507 local_irq_restore(flags);
2510 void __netif_schedule(struct Qdisc *q)
2512 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
2513 __netif_reschedule(q);
2515 EXPORT_SYMBOL(__netif_schedule);
2517 struct dev_kfree_skb_cb {
2518 enum skb_free_reason reason;
2521 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb)
2523 return (struct dev_kfree_skb_cb *)skb->cb;
2526 void netif_schedule_queue(struct netdev_queue *txq)
2529 if (!(txq->state & QUEUE_STATE_ANY_XOFF)) {
2530 struct Qdisc *q = rcu_dereference(txq->qdisc);
2532 __netif_schedule(q);
2536 EXPORT_SYMBOL(netif_schedule_queue);
2538 void netif_tx_wake_queue(struct netdev_queue *dev_queue)
2540 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) {
2544 q = rcu_dereference(dev_queue->qdisc);
2545 __netif_schedule(q);
2549 EXPORT_SYMBOL(netif_tx_wake_queue);
2551 void __dev_kfree_skb_irq(struct sk_buff *skb, enum skb_free_reason reason)
2553 unsigned long flags;
2558 if (likely(refcount_read(&skb->users) == 1)) {
2560 refcount_set(&skb->users, 0);
2561 } else if (likely(!refcount_dec_and_test(&skb->users))) {
2564 get_kfree_skb_cb(skb)->reason = reason;
2565 local_irq_save(flags);
2566 skb->next = __this_cpu_read(softnet_data.completion_queue);
2567 __this_cpu_write(softnet_data.completion_queue, skb);
2568 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2569 local_irq_restore(flags);
2571 EXPORT_SYMBOL(__dev_kfree_skb_irq);
2573 void __dev_kfree_skb_any(struct sk_buff *skb, enum skb_free_reason reason)
2575 if (in_irq() || irqs_disabled())
2576 __dev_kfree_skb_irq(skb, reason);
2580 EXPORT_SYMBOL(__dev_kfree_skb_any);
2584 * netif_device_detach - mark device as removed
2585 * @dev: network device
2587 * Mark device as removed from system and therefore no longer available.
2589 void netif_device_detach(struct net_device *dev)
2591 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) &&
2592 netif_running(dev)) {
2593 netif_tx_stop_all_queues(dev);
2596 EXPORT_SYMBOL(netif_device_detach);
2599 * netif_device_attach - mark device as attached
2600 * @dev: network device
2602 * Mark device as attached from system and restart if needed.
2604 void netif_device_attach(struct net_device *dev)
2606 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) &&
2607 netif_running(dev)) {
2608 netif_tx_wake_all_queues(dev);
2609 __netdev_watchdog_up(dev);
2612 EXPORT_SYMBOL(netif_device_attach);
2615 * Returns a Tx hash based on the given packet descriptor a Tx queues' number
2616 * to be used as a distribution range.
2618 static u16 skb_tx_hash(const struct net_device *dev, struct sk_buff *skb)
2622 u16 qcount = dev->real_num_tx_queues;
2624 if (skb_rx_queue_recorded(skb)) {
2625 hash = skb_get_rx_queue(skb);
2626 while (unlikely(hash >= qcount))
2632 u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
2634 qoffset = dev->tc_to_txq[tc].offset;
2635 qcount = dev->tc_to_txq[tc].count;
2638 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset;
2641 static void skb_warn_bad_offload(const struct sk_buff *skb)
2643 static const netdev_features_t null_features;
2644 struct net_device *dev = skb->dev;
2645 const char *name = "";
2647 if (!net_ratelimit())
2651 if (dev->dev.parent)
2652 name = dev_driver_string(dev->dev.parent);
2654 name = netdev_name(dev);
2656 WARN(1, "%s: caps=(%pNF, %pNF) len=%d data_len=%d gso_size=%d "
2657 "gso_type=%d ip_summed=%d\n",
2658 name, dev ? &dev->features : &null_features,
2659 skb->sk ? &skb->sk->sk_route_caps : &null_features,
2660 skb->len, skb->data_len, skb_shinfo(skb)->gso_size,
2661 skb_shinfo(skb)->gso_type, skb->ip_summed);
2665 * Invalidate hardware checksum when packet is to be mangled, and
2666 * complete checksum manually on outgoing path.
2668 int skb_checksum_help(struct sk_buff *skb)
2671 int ret = 0, offset;
2673 if (skb->ip_summed == CHECKSUM_COMPLETE)
2674 goto out_set_summed;
2676 if (unlikely(skb_shinfo(skb)->gso_size)) {
2677 skb_warn_bad_offload(skb);
2681 /* Before computing a checksum, we should make sure no frag could
2682 * be modified by an external entity : checksum could be wrong.
2684 if (skb_has_shared_frag(skb)) {
2685 ret = __skb_linearize(skb);
2690 offset = skb_checksum_start_offset(skb);
2691 BUG_ON(offset >= skb_headlen(skb));
2692 csum = skb_checksum(skb, offset, skb->len - offset, 0);
2694 offset += skb->csum_offset;
2695 BUG_ON(offset + sizeof(__sum16) > skb_headlen(skb));
2697 if (skb_cloned(skb) &&
2698 !skb_clone_writable(skb, offset + sizeof(__sum16))) {
2699 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2704 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0;
2706 skb->ip_summed = CHECKSUM_NONE;
2710 EXPORT_SYMBOL(skb_checksum_help);
2712 int skb_crc32c_csum_help(struct sk_buff *skb)
2715 int ret = 0, offset, start;
2717 if (skb->ip_summed != CHECKSUM_PARTIAL)
2720 if (unlikely(skb_is_gso(skb)))
2723 /* Before computing a checksum, we should make sure no frag could
2724 * be modified by an external entity : checksum could be wrong.
2726 if (unlikely(skb_has_shared_frag(skb))) {
2727 ret = __skb_linearize(skb);
2731 start = skb_checksum_start_offset(skb);
2732 offset = start + offsetof(struct sctphdr, checksum);
2733 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) {
2737 if (skb_cloned(skb) &&
2738 !skb_clone_writable(skb, offset + sizeof(__le32))) {
2739 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2743 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start,
2744 skb->len - start, ~(__u32)0,
2746 *(__le32 *)(skb->data + offset) = crc32c_csum;
2747 skb->ip_summed = CHECKSUM_NONE;
2748 skb->csum_not_inet = 0;
2753 __be16 skb_network_protocol(struct sk_buff *skb, int *depth)
2755 __be16 type = skb->protocol;
2757 /* Tunnel gso handlers can set protocol to ethernet. */
2758 if (type == htons(ETH_P_TEB)) {
2761 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr))))
2764 eth = (struct ethhdr *)skb->data;
2765 type = eth->h_proto;
2768 return __vlan_get_protocol(skb, type, depth);
2772 * skb_mac_gso_segment - mac layer segmentation handler.
2773 * @skb: buffer to segment
2774 * @features: features for the output path (see dev->features)
2776 struct sk_buff *skb_mac_gso_segment(struct sk_buff *skb,
2777 netdev_features_t features)
2779 struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT);
2780 struct packet_offload *ptype;
2781 int vlan_depth = skb->mac_len;
2782 __be16 type = skb_network_protocol(skb, &vlan_depth);
2784 if (unlikely(!type))
2785 return ERR_PTR(-EINVAL);
2787 __skb_pull(skb, vlan_depth);
2790 list_for_each_entry_rcu(ptype, &offload_base, list) {
2791 if (ptype->type == type && ptype->callbacks.gso_segment) {
2792 segs = ptype->callbacks.gso_segment(skb, features);
2798 __skb_push(skb, skb->data - skb_mac_header(skb));
2802 EXPORT_SYMBOL(skb_mac_gso_segment);
2805 /* openvswitch calls this on rx path, so we need a different check.
2807 static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
2810 return skb->ip_summed != CHECKSUM_PARTIAL &&
2811 skb->ip_summed != CHECKSUM_UNNECESSARY;
2813 return skb->ip_summed == CHECKSUM_NONE;
2817 * __skb_gso_segment - Perform segmentation on skb.
2818 * @skb: buffer to segment
2819 * @features: features for the output path (see dev->features)
2820 * @tx_path: whether it is called in TX path
2822 * This function segments the given skb and returns a list of segments.
2824 * It may return NULL if the skb requires no segmentation. This is
2825 * only possible when GSO is used for verifying header integrity.
2827 * Segmentation preserves SKB_SGO_CB_OFFSET bytes of previous skb cb.
2829 struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
2830 netdev_features_t features, bool tx_path)
2832 struct sk_buff *segs;
2834 if (unlikely(skb_needs_check(skb, tx_path))) {
2837 /* We're going to init ->check field in TCP or UDP header */
2838 err = skb_cow_head(skb, 0);
2840 return ERR_PTR(err);
2843 /* Only report GSO partial support if it will enable us to
2844 * support segmentation on this frame without needing additional
2847 if (features & NETIF_F_GSO_PARTIAL) {
2848 netdev_features_t partial_features = NETIF_F_GSO_ROBUST;
2849 struct net_device *dev = skb->dev;
2851 partial_features |= dev->features & dev->gso_partial_features;
2852 if (!skb_gso_ok(skb, features | partial_features))
2853 features &= ~NETIF_F_GSO_PARTIAL;
2856 BUILD_BUG_ON(SKB_SGO_CB_OFFSET +
2857 sizeof(*SKB_GSO_CB(skb)) > sizeof(skb->cb));
2859 SKB_GSO_CB(skb)->mac_offset = skb_headroom(skb);
2860 SKB_GSO_CB(skb)->encap_level = 0;
2862 skb_reset_mac_header(skb);
2863 skb_reset_mac_len(skb);
2865 segs = skb_mac_gso_segment(skb, features);
2867 if (unlikely(skb_needs_check(skb, tx_path) && !IS_ERR(segs)))
2868 skb_warn_bad_offload(skb);
2872 EXPORT_SYMBOL(__skb_gso_segment);
2874 /* Take action when hardware reception checksum errors are detected. */
2876 void netdev_rx_csum_fault(struct net_device *dev)
2878 if (net_ratelimit()) {
2879 pr_err("%s: hw csum failure\n", dev ? dev->name : "<unknown>");
2883 EXPORT_SYMBOL(netdev_rx_csum_fault);
2886 /* Actually, we should eliminate this check as soon as we know, that:
2887 * 1. IOMMU is present and allows to map all the memory.
2888 * 2. No high memory really exists on this machine.
2891 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
2893 #ifdef CONFIG_HIGHMEM
2896 if (!(dev->features & NETIF_F_HIGHDMA)) {
2897 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2898 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2900 if (PageHighMem(skb_frag_page(frag)))
2905 if (PCI_DMA_BUS_IS_PHYS) {
2906 struct device *pdev = dev->dev.parent;
2910 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2911 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2912 dma_addr_t addr = page_to_phys(skb_frag_page(frag));
2914 if (!pdev->dma_mask || addr + PAGE_SIZE - 1 > *pdev->dma_mask)
2922 /* If MPLS offload request, verify we are testing hardware MPLS features
2923 * instead of standard features for the netdev.
2925 #if IS_ENABLED(CONFIG_NET_MPLS_GSO)
2926 static netdev_features_t net_mpls_features(struct sk_buff *skb,
2927 netdev_features_t features,
2930 if (eth_p_mpls(type))
2931 features &= skb->dev->mpls_features;
2936 static netdev_features_t net_mpls_features(struct sk_buff *skb,
2937 netdev_features_t features,
2944 static netdev_features_t harmonize_features(struct sk_buff *skb,
2945 netdev_features_t features)
2950 type = skb_network_protocol(skb, &tmp);
2951 features = net_mpls_features(skb, features, type);
2953 if (skb->ip_summed != CHECKSUM_NONE &&
2954 !can_checksum_protocol(features, type)) {
2955 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
2957 if (illegal_highdma(skb->dev, skb))
2958 features &= ~NETIF_F_SG;
2963 netdev_features_t passthru_features_check(struct sk_buff *skb,
2964 struct net_device *dev,
2965 netdev_features_t features)
2969 EXPORT_SYMBOL(passthru_features_check);
2971 static netdev_features_t dflt_features_check(struct sk_buff *skb,
2972 struct net_device *dev,
2973 netdev_features_t features)
2975 return vlan_features_check(skb, features);
2978 static netdev_features_t gso_features_check(const struct sk_buff *skb,
2979 struct net_device *dev,
2980 netdev_features_t features)
2982 u16 gso_segs = skb_shinfo(skb)->gso_segs;
2984 if (gso_segs > dev->gso_max_segs)
2985 return features & ~NETIF_F_GSO_MASK;
2987 /* Support for GSO partial features requires software
2988 * intervention before we can actually process the packets
2989 * so we need to strip support for any partial features now
2990 * and we can pull them back in after we have partially
2991 * segmented the frame.
2993 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL))
2994 features &= ~dev->gso_partial_features;
2996 /* Make sure to clear the IPv4 ID mangling feature if the
2997 * IPv4 header has the potential to be fragmented.
2999 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
3000 struct iphdr *iph = skb->encapsulation ?
3001 inner_ip_hdr(skb) : ip_hdr(skb);
3003 if (!(iph->frag_off & htons(IP_DF)))
3004 features &= ~NETIF_F_TSO_MANGLEID;
3010 netdev_features_t netif_skb_features(struct sk_buff *skb)
3012 struct net_device *dev = skb->dev;
3013 netdev_features_t features = dev->features;
3015 if (skb_is_gso(skb))
3016 features = gso_features_check(skb, dev, features);
3018 /* If encapsulation offload request, verify we are testing
3019 * hardware encapsulation features instead of standard
3020 * features for the netdev
3022 if (skb->encapsulation)
3023 features &= dev->hw_enc_features;
3025 if (skb_vlan_tagged(skb))
3026 features = netdev_intersect_features(features,
3027 dev->vlan_features |
3028 NETIF_F_HW_VLAN_CTAG_TX |
3029 NETIF_F_HW_VLAN_STAG_TX);
3031 if (dev->netdev_ops->ndo_features_check)
3032 features &= dev->netdev_ops->ndo_features_check(skb, dev,
3035 features &= dflt_features_check(skb, dev, features);
3037 return harmonize_features(skb, features);
3039 EXPORT_SYMBOL(netif_skb_features);
3041 static int xmit_one(struct sk_buff *skb, struct net_device *dev,
3042 struct netdev_queue *txq, bool more)
3047 if (!list_empty(&ptype_all) || !list_empty(&dev->ptype_all))
3048 dev_queue_xmit_nit(skb, dev);
3051 trace_net_dev_start_xmit(skb, dev);
3052 rc = netdev_start_xmit(skb, dev, txq, more);
3053 trace_net_dev_xmit(skb, rc, dev, len);
3058 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev,
3059 struct netdev_queue *txq, int *ret)
3061 struct sk_buff *skb = first;
3062 int rc = NETDEV_TX_OK;
3065 struct sk_buff *next = skb->next;
3068 rc = xmit_one(skb, dev, txq, next != NULL);
3069 if (unlikely(!dev_xmit_complete(rc))) {
3075 if (netif_xmit_stopped(txq) && skb) {
3076 rc = NETDEV_TX_BUSY;
3086 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb,
3087 netdev_features_t features)
3089 if (skb_vlan_tag_present(skb) &&
3090 !vlan_hw_offload_capable(features, skb->vlan_proto))
3091 skb = __vlan_hwaccel_push_inside(skb);
3095 int skb_csum_hwoffload_help(struct sk_buff *skb,
3096 const netdev_features_t features)
3098 if (unlikely(skb->csum_not_inet))
3099 return !!(features & NETIF_F_SCTP_CRC) ? 0 :
3100 skb_crc32c_csum_help(skb);
3102 return !!(features & NETIF_F_CSUM_MASK) ? 0 : skb_checksum_help(skb);
3104 EXPORT_SYMBOL(skb_csum_hwoffload_help);
3106 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again)
3108 netdev_features_t features;
3110 features = netif_skb_features(skb);
3111 skb = validate_xmit_vlan(skb, features);
3115 skb = sk_validate_xmit_skb(skb, dev);
3119 if (netif_needs_gso(skb, features)) {
3120 struct sk_buff *segs;
3122 segs = skb_gso_segment(skb, features);
3130 if (skb_needs_linearize(skb, features) &&
3131 __skb_linearize(skb))
3134 /* If packet is not checksummed and device does not
3135 * support checksumming for this protocol, complete
3136 * checksumming here.
3138 if (skb->ip_summed == CHECKSUM_PARTIAL) {
3139 if (skb->encapsulation)
3140 skb_set_inner_transport_header(skb,
3141 skb_checksum_start_offset(skb));
3143 skb_set_transport_header(skb,
3144 skb_checksum_start_offset(skb));
3145 if (skb_csum_hwoffload_help(skb, features))
3150 skb = validate_xmit_xfrm(skb, features, again);
3157 atomic_long_inc(&dev->tx_dropped);
3161 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again)
3163 struct sk_buff *next, *head = NULL, *tail;
3165 for (; skb != NULL; skb = next) {
3169 /* in case skb wont be segmented, point to itself */
3172 skb = validate_xmit_skb(skb, dev, again);
3180 /* If skb was segmented, skb->prev points to
3181 * the last segment. If not, it still contains skb.
3187 EXPORT_SYMBOL_GPL(validate_xmit_skb_list);
3189 static void qdisc_pkt_len_init(struct sk_buff *skb)
3191 const struct skb_shared_info *shinfo = skb_shinfo(skb);
3193 qdisc_skb_cb(skb)->pkt_len = skb->len;
3195 /* To get more precise estimation of bytes sent on wire,
3196 * we add to pkt_len the headers size of all segments
3198 if (shinfo->gso_size) {
3199 unsigned int hdr_len;
3200 u16 gso_segs = shinfo->gso_segs;
3202 /* mac layer + network layer */
3203 hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
3205 /* + transport layer */
3206 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) {
3207 const struct tcphdr *th;
3208 struct tcphdr _tcphdr;
3210 th = skb_header_pointer(skb, skb_transport_offset(skb),
3211 sizeof(_tcphdr), &_tcphdr);
3213 hdr_len += __tcp_hdrlen(th);
3215 struct udphdr _udphdr;
3217 if (skb_header_pointer(skb, skb_transport_offset(skb),
3218 sizeof(_udphdr), &_udphdr))
3219 hdr_len += sizeof(struct udphdr);
3222 if (shinfo->gso_type & SKB_GSO_DODGY)
3223 gso_segs = DIV_ROUND_UP(skb->len - hdr_len,
3226 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len;
3230 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
3231 struct net_device *dev,
3232 struct netdev_queue *txq)
3234 spinlock_t *root_lock = qdisc_lock(q);
3235 struct sk_buff *to_free = NULL;
3239 qdisc_calculate_pkt_len(skb, q);
3241 if (q->flags & TCQ_F_NOLOCK) {
3242 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3243 __qdisc_drop(skb, &to_free);
3246 rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
3250 if (unlikely(to_free))
3251 kfree_skb_list(to_free);
3256 * Heuristic to force contended enqueues to serialize on a
3257 * separate lock before trying to get qdisc main lock.
3258 * This permits qdisc->running owner to get the lock more
3259 * often and dequeue packets faster.
3261 contended = qdisc_is_running(q);
3262 if (unlikely(contended))
3263 spin_lock(&q->busylock);
3265 spin_lock(root_lock);
3266 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3267 __qdisc_drop(skb, &to_free);
3269 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
3270 qdisc_run_begin(q)) {
3272 * This is a work-conserving queue; there are no old skbs
3273 * waiting to be sent out; and the qdisc is not running -
3274 * xmit the skb directly.
3277 qdisc_bstats_update(q, skb);
3279 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) {
3280 if (unlikely(contended)) {
3281 spin_unlock(&q->busylock);
3288 rc = NET_XMIT_SUCCESS;
3290 rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
3291 if (qdisc_run_begin(q)) {
3292 if (unlikely(contended)) {
3293 spin_unlock(&q->busylock);
3300 spin_unlock(root_lock);
3301 if (unlikely(to_free))
3302 kfree_skb_list(to_free);
3303 if (unlikely(contended))
3304 spin_unlock(&q->busylock);
3308 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO)
3309 static void skb_update_prio(struct sk_buff *skb)
3311 const struct netprio_map *map;
3312 const struct sock *sk;
3313 unsigned int prioidx;
3317 map = rcu_dereference_bh(skb->dev->priomap);
3320 sk = skb_to_full_sk(skb);
3324 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data);
3326 if (prioidx < map->priomap_len)
3327 skb->priority = map->priomap[prioidx];
3330 #define skb_update_prio(skb)
3333 DEFINE_PER_CPU(int, xmit_recursion);
3334 EXPORT_SYMBOL(xmit_recursion);
3337 * dev_loopback_xmit - loop back @skb
3338 * @net: network namespace this loopback is happening in
3339 * @sk: sk needed to be a netfilter okfn
3340 * @skb: buffer to transmit
3342 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb)
3344 skb_reset_mac_header(skb);
3345 __skb_pull(skb, skb_network_offset(skb));
3346 skb->pkt_type = PACKET_LOOPBACK;
3347 skb->ip_summed = CHECKSUM_UNNECESSARY;
3348 WARN_ON(!skb_dst(skb));
3353 EXPORT_SYMBOL(dev_loopback_xmit);
3355 #ifdef CONFIG_NET_EGRESS
3356 static struct sk_buff *
3357 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
3359 struct mini_Qdisc *miniq = rcu_dereference_bh(dev->miniq_egress);
3360 struct tcf_result cl_res;
3365 /* qdisc_skb_cb(skb)->pkt_len was already set by the caller. */
3366 mini_qdisc_bstats_cpu_update(miniq, skb);
3368 switch (tcf_classify(skb, miniq->filter_list, &cl_res, false)) {
3370 case TC_ACT_RECLASSIFY:
3371 skb->tc_index = TC_H_MIN(cl_res.classid);
3374 mini_qdisc_qstats_cpu_drop(miniq);
3375 *ret = NET_XMIT_DROP;
3381 *ret = NET_XMIT_SUCCESS;
3384 case TC_ACT_REDIRECT:
3385 /* No need to push/pop skb's mac_header here on egress! */
3386 skb_do_redirect(skb);
3387 *ret = NET_XMIT_SUCCESS;
3395 #endif /* CONFIG_NET_EGRESS */
3397 static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb)
3400 struct xps_dev_maps *dev_maps;
3401 struct xps_map *map;
3402 int queue_index = -1;
3405 dev_maps = rcu_dereference(dev->xps_maps);
3407 unsigned int tci = skb->sender_cpu - 1;
3411 tci += netdev_get_prio_tc_map(dev, skb->priority);
3414 map = rcu_dereference(dev_maps->cpu_map[tci]);
3417 queue_index = map->queues[0];
3419 queue_index = map->queues[reciprocal_scale(skb_get_hash(skb),
3421 if (unlikely(queue_index >= dev->real_num_tx_queues))
3433 static u16 __netdev_pick_tx(struct net_device *dev, struct sk_buff *skb)
3435 struct sock *sk = skb->sk;
3436 int queue_index = sk_tx_queue_get(sk);
3438 if (queue_index < 0 || skb->ooo_okay ||
3439 queue_index >= dev->real_num_tx_queues) {
3440 int new_index = get_xps_queue(dev, skb);
3443 new_index = skb_tx_hash(dev, skb);
3445 if (queue_index != new_index && sk &&
3447 rcu_access_pointer(sk->sk_dst_cache))
3448 sk_tx_queue_set(sk, new_index);
3450 queue_index = new_index;
3456 struct netdev_queue *netdev_pick_tx(struct net_device *dev,
3457 struct sk_buff *skb,
3460 int queue_index = 0;
3463 u32 sender_cpu = skb->sender_cpu - 1;
3465 if (sender_cpu >= (u32)NR_CPUS)
3466 skb->sender_cpu = raw_smp_processor_id() + 1;
3469 if (dev->real_num_tx_queues != 1) {
3470 const struct net_device_ops *ops = dev->netdev_ops;
3472 if (ops->ndo_select_queue)
3473 queue_index = ops->ndo_select_queue(dev, skb, accel_priv,
3476 queue_index = __netdev_pick_tx(dev, skb);
3478 queue_index = netdev_cap_txqueue(dev, queue_index);
3481 skb_set_queue_mapping(skb, queue_index);
3482 return netdev_get_tx_queue(dev, queue_index);
3486 * __dev_queue_xmit - transmit a buffer
3487 * @skb: buffer to transmit
3488 * @accel_priv: private data used for L2 forwarding offload
3490 * Queue a buffer for transmission to a network device. The caller must
3491 * have set the device and priority and built the buffer before calling
3492 * this function. The function can be called from an interrupt.
3494 * A negative errno code is returned on a failure. A success does not
3495 * guarantee the frame will be transmitted as it may be dropped due
3496 * to congestion or traffic shaping.
3498 * -----------------------------------------------------------------------------------
3499 * I notice this method can also return errors from the queue disciplines,
3500 * including NET_XMIT_DROP, which is a positive value. So, errors can also
3503 * Regardless of the return value, the skb is consumed, so it is currently
3504 * difficult to retry a send to this method. (You can bump the ref count
3505 * before sending to hold a reference for retry if you are careful.)
3507 * When calling this method, interrupts MUST be enabled. This is because
3508 * the BH enable code must have IRQs enabled so that it will not deadlock.
3511 static int __dev_queue_xmit(struct sk_buff *skb, void *accel_priv)
3513 struct net_device *dev = skb->dev;
3514 struct netdev_queue *txq;
3519 skb_reset_mac_header(skb);
3521 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP))
3522 __skb_tstamp_tx(skb, NULL, skb->sk, SCM_TSTAMP_SCHED);
3524 /* Disable soft irqs for various locks below. Also
3525 * stops preemption for RCU.
3529 skb_update_prio(skb);
3531 qdisc_pkt_len_init(skb);
3532 #ifdef CONFIG_NET_CLS_ACT
3533 skb->tc_at_ingress = 0;
3534 # ifdef CONFIG_NET_EGRESS
3535 if (static_key_false(&egress_needed)) {
3536 skb = sch_handle_egress(skb, &rc, dev);
3542 /* If device/qdisc don't need skb->dst, release it right now while
3543 * its hot in this cpu cache.
3545 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
3550 txq = netdev_pick_tx(dev, skb, accel_priv);
3551 q = rcu_dereference_bh(txq->qdisc);
3553 trace_net_dev_queue(skb);
3555 rc = __dev_xmit_skb(skb, q, dev, txq);
3559 /* The device has no queue. Common case for software devices:
3560 * loopback, all the sorts of tunnels...
3562 * Really, it is unlikely that netif_tx_lock protection is necessary
3563 * here. (f.e. loopback and IP tunnels are clean ignoring statistics
3565 * However, it is possible, that they rely on protection
3568 * Check this and shot the lock. It is not prone from deadlocks.
3569 *Either shot noqueue qdisc, it is even simpler 8)
3571 if (dev->flags & IFF_UP) {
3572 int cpu = smp_processor_id(); /* ok because BHs are off */
3574 if (txq->xmit_lock_owner != cpu) {
3575 if (unlikely(__this_cpu_read(xmit_recursion) >
3576 XMIT_RECURSION_LIMIT))
3577 goto recursion_alert;
3579 skb = validate_xmit_skb(skb, dev, &again);
3583 HARD_TX_LOCK(dev, txq, cpu);
3585 if (!netif_xmit_stopped(txq)) {
3586 __this_cpu_inc(xmit_recursion);
3587 skb = dev_hard_start_xmit(skb, dev, txq, &rc);
3588 __this_cpu_dec(xmit_recursion);
3589 if (dev_xmit_complete(rc)) {
3590 HARD_TX_UNLOCK(dev, txq);
3594 HARD_TX_UNLOCK(dev, txq);
3595 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
3598 /* Recursion is detected! It is possible,
3602 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
3608 rcu_read_unlock_bh();
3610 atomic_long_inc(&dev->tx_dropped);
3611 kfree_skb_list(skb);
3614 rcu_read_unlock_bh();
3618 int dev_queue_xmit(struct sk_buff *skb)
3620 return __dev_queue_xmit(skb, NULL);
3622 EXPORT_SYMBOL(dev_queue_xmit);
3624 int dev_queue_xmit_accel(struct sk_buff *skb, void *accel_priv)
3626 return __dev_queue_xmit(skb, accel_priv);
3628 EXPORT_SYMBOL(dev_queue_xmit_accel);
3631 /*************************************************************************
3633 *************************************************************************/
3635 int netdev_max_backlog __read_mostly = 1000;
3636 EXPORT_SYMBOL(netdev_max_backlog);
3638 int netdev_tstamp_prequeue __read_mostly = 1;
3639 int netdev_budget __read_mostly = 300;
3640 unsigned int __read_mostly netdev_budget_usecs = 2000;
3641 int weight_p __read_mostly = 64; /* old backlog weight */
3642 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */
3643 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */
3644 int dev_rx_weight __read_mostly = 64;
3645 int dev_tx_weight __read_mostly = 64;
3647 /* Called with irq disabled */
3648 static inline void ____napi_schedule(struct softnet_data *sd,
3649 struct napi_struct *napi)
3651 list_add_tail(&napi->poll_list, &sd->poll_list);
3652 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3657 /* One global table that all flow-based protocols share. */
3658 struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
3659 EXPORT_SYMBOL(rps_sock_flow_table);
3660 u32 rps_cpu_mask __read_mostly;
3661 EXPORT_SYMBOL(rps_cpu_mask);
3663 struct static_key rps_needed __read_mostly;
3664 EXPORT_SYMBOL(rps_needed);
3665 struct static_key rfs_needed __read_mostly;
3666 EXPORT_SYMBOL(rfs_needed);
3668 static struct rps_dev_flow *
3669 set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3670 struct rps_dev_flow *rflow, u16 next_cpu)
3672 if (next_cpu < nr_cpu_ids) {
3673 #ifdef CONFIG_RFS_ACCEL
3674 struct netdev_rx_queue *rxqueue;
3675 struct rps_dev_flow_table *flow_table;
3676 struct rps_dev_flow *old_rflow;
3681 /* Should we steer this flow to a different hardware queue? */
3682 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
3683 !(dev->features & NETIF_F_NTUPLE))
3685 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
3686 if (rxq_index == skb_get_rx_queue(skb))
3689 rxqueue = dev->_rx + rxq_index;
3690 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3693 flow_id = skb_get_hash(skb) & flow_table->mask;
3694 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
3695 rxq_index, flow_id);
3699 rflow = &flow_table->flows[flow_id];
3701 if (old_rflow->filter == rflow->filter)
3702 old_rflow->filter = RPS_NO_FILTER;
3706 per_cpu(softnet_data, next_cpu).input_queue_head;
3709 rflow->cpu = next_cpu;
3714 * get_rps_cpu is called from netif_receive_skb and returns the target
3715 * CPU from the RPS map of the receiving queue for a given skb.
3716 * rcu_read_lock must be held on entry.
3718 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3719 struct rps_dev_flow **rflowp)
3721 const struct rps_sock_flow_table *sock_flow_table;
3722 struct netdev_rx_queue *rxqueue = dev->_rx;
3723 struct rps_dev_flow_table *flow_table;
3724 struct rps_map *map;
3729 if (skb_rx_queue_recorded(skb)) {
3730 u16 index = skb_get_rx_queue(skb);
3732 if (unlikely(index >= dev->real_num_rx_queues)) {
3733 WARN_ONCE(dev->real_num_rx_queues > 1,
3734 "%s received packet on queue %u, but number "
3735 "of RX queues is %u\n",
3736 dev->name, index, dev->real_num_rx_queues);
3742 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */
3744 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3745 map = rcu_dereference(rxqueue->rps_map);
3746 if (!flow_table && !map)
3749 skb_reset_network_header(skb);
3750 hash = skb_get_hash(skb);
3754 sock_flow_table = rcu_dereference(rps_sock_flow_table);
3755 if (flow_table && sock_flow_table) {
3756 struct rps_dev_flow *rflow;
3760 /* First check into global flow table if there is a match */
3761 ident = sock_flow_table->ents[hash & sock_flow_table->mask];
3762 if ((ident ^ hash) & ~rps_cpu_mask)
3765 next_cpu = ident & rps_cpu_mask;
3767 /* OK, now we know there is a match,
3768 * we can look at the local (per receive queue) flow table
3770 rflow = &flow_table->flows[hash & flow_table->mask];
3774 * If the desired CPU (where last recvmsg was done) is
3775 * different from current CPU (one in the rx-queue flow
3776 * table entry), switch if one of the following holds:
3777 * - Current CPU is unset (>= nr_cpu_ids).
3778 * - Current CPU is offline.
3779 * - The current CPU's queue tail has advanced beyond the
3780 * last packet that was enqueued using this table entry.
3781 * This guarantees that all previous packets for the flow
3782 * have been dequeued, thus preserving in order delivery.
3784 if (unlikely(tcpu != next_cpu) &&
3785 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) ||
3786 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
3787 rflow->last_qtail)) >= 0)) {
3789 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
3792 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) {
3802 tcpu = map->cpus[reciprocal_scale(hash, map->len)];
3803 if (cpu_online(tcpu)) {
3813 #ifdef CONFIG_RFS_ACCEL
3816 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
3817 * @dev: Device on which the filter was set
3818 * @rxq_index: RX queue index
3819 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
3820 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
3822 * Drivers that implement ndo_rx_flow_steer() should periodically call
3823 * this function for each installed filter and remove the filters for
3824 * which it returns %true.
3826 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
3827 u32 flow_id, u16 filter_id)
3829 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
3830 struct rps_dev_flow_table *flow_table;
3831 struct rps_dev_flow *rflow;
3836 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3837 if (flow_table && flow_id <= flow_table->mask) {
3838 rflow = &flow_table->flows[flow_id];
3839 cpu = READ_ONCE(rflow->cpu);
3840 if (rflow->filter == filter_id && cpu < nr_cpu_ids &&
3841 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
3842 rflow->last_qtail) <
3843 (int)(10 * flow_table->mask)))
3849 EXPORT_SYMBOL(rps_may_expire_flow);
3851 #endif /* CONFIG_RFS_ACCEL */
3853 /* Called from hardirq (IPI) context */
3854 static void rps_trigger_softirq(void *data)
3856 struct softnet_data *sd = data;
3858 ____napi_schedule(sd, &sd->backlog);
3862 #endif /* CONFIG_RPS */
3865 * Check if this softnet_data structure is another cpu one
3866 * If yes, queue it to our IPI list and return 1
3869 static int rps_ipi_queued(struct softnet_data *sd)
3872 struct softnet_data *mysd = this_cpu_ptr(&softnet_data);
3875 sd->rps_ipi_next = mysd->rps_ipi_list;
3876 mysd->rps_ipi_list = sd;
3878 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3881 #endif /* CONFIG_RPS */
3885 #ifdef CONFIG_NET_FLOW_LIMIT
3886 int netdev_flow_limit_table_len __read_mostly = (1 << 12);
3889 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
3891 #ifdef CONFIG_NET_FLOW_LIMIT
3892 struct sd_flow_limit *fl;
3893 struct softnet_data *sd;
3894 unsigned int old_flow, new_flow;
3896 if (qlen < (netdev_max_backlog >> 1))
3899 sd = this_cpu_ptr(&softnet_data);
3902 fl = rcu_dereference(sd->flow_limit);
3904 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1);
3905 old_flow = fl->history[fl->history_head];
3906 fl->history[fl->history_head] = new_flow;
3909 fl->history_head &= FLOW_LIMIT_HISTORY - 1;
3911 if (likely(fl->buckets[old_flow]))
3912 fl->buckets[old_flow]--;
3914 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) {
3926 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
3927 * queue (may be a remote CPU queue).
3929 static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
3930 unsigned int *qtail)
3932 struct softnet_data *sd;
3933 unsigned long flags;
3936 sd = &per_cpu(softnet_data, cpu);
3938 local_irq_save(flags);
3941 if (!netif_running(skb->dev))
3943 qlen = skb_queue_len(&sd->input_pkt_queue);
3944 if (qlen <= netdev_max_backlog && !skb_flow_limit(skb, qlen)) {
3947 __skb_queue_tail(&sd->input_pkt_queue, skb);
3948 input_queue_tail_incr_save(sd, qtail);
3950 local_irq_restore(flags);
3951 return NET_RX_SUCCESS;
3954 /* Schedule NAPI for backlog device
3955 * We can use non atomic operation since we own the queue lock
3957 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
3958 if (!rps_ipi_queued(sd))
3959 ____napi_schedule(sd, &sd->backlog);
3968 local_irq_restore(flags);
3970 atomic_long_inc(&skb->dev->rx_dropped);
3975 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb)
3977 struct net_device *dev = skb->dev;
3978 struct netdev_rx_queue *rxqueue;
3982 if (skb_rx_queue_recorded(skb)) {
3983 u16 index = skb_get_rx_queue(skb);
3985 if (unlikely(index >= dev->real_num_rx_queues)) {
3986 WARN_ONCE(dev->real_num_rx_queues > 1,
3987 "%s received packet on queue %u, but number "
3988 "of RX queues is %u\n",
3989 dev->name, index, dev->real_num_rx_queues);
3991 return rxqueue; /* Return first rxqueue */
3998 static u32 netif_receive_generic_xdp(struct sk_buff *skb,
3999 struct bpf_prog *xdp_prog)
4001 struct netdev_rx_queue *rxqueue;
4002 void *orig_data, *orig_data_end;
4003 u32 metalen, act = XDP_DROP;
4004 struct xdp_buff xdp;
4008 /* Reinjected packets coming from act_mirred or similar should
4009 * not get XDP generic processing.
4011 if (skb_cloned(skb))
4014 /* XDP packets must be linear and must have sufficient headroom
4015 * of XDP_PACKET_HEADROOM bytes. This is the guarantee that also
4016 * native XDP provides, thus we need to do it here as well.
4018 if (skb_is_nonlinear(skb) ||
4019 skb_headroom(skb) < XDP_PACKET_HEADROOM) {
4020 int hroom = XDP_PACKET_HEADROOM - skb_headroom(skb);
4021 int troom = skb->tail + skb->data_len - skb->end;
4023 /* In case we have to go down the path and also linearize,
4024 * then lets do the pskb_expand_head() work just once here.
4026 if (pskb_expand_head(skb,
4027 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0,
4028 troom > 0 ? troom + 128 : 0, GFP_ATOMIC))
4030 if (skb_linearize(skb))
4034 /* The XDP program wants to see the packet starting at the MAC
4037 mac_len = skb->data - skb_mac_header(skb);
4038 hlen = skb_headlen(skb) + mac_len;
4039 xdp.data = skb->data - mac_len;
4040 xdp.data_meta = xdp.data;
4041 xdp.data_end = xdp.data + hlen;
4042 xdp.data_hard_start = skb->data - skb_headroom(skb);
4043 orig_data_end = xdp.data_end;
4044 orig_data = xdp.data;
4046 rxqueue = netif_get_rxqueue(skb);
4047 xdp.rxq = &rxqueue->xdp_rxq;
4049 act = bpf_prog_run_xdp(xdp_prog, &xdp);
4051 off = xdp.data - orig_data;
4053 __skb_pull(skb, off);
4055 __skb_push(skb, -off);
4056 skb->mac_header += off;
4058 /* check if bpf_xdp_adjust_tail was used. it can only "shrink"
4061 off = orig_data_end - xdp.data_end;
4063 skb_set_tail_pointer(skb, xdp.data_end - xdp.data);
4070 __skb_push(skb, mac_len);
4073 metalen = xdp.data - xdp.data_meta;
4075 skb_metadata_set(skb, metalen);
4078 bpf_warn_invalid_xdp_action(act);
4081 trace_xdp_exception(skb->dev, xdp_prog, act);
4092 /* When doing generic XDP we have to bypass the qdisc layer and the
4093 * network taps in order to match in-driver-XDP behavior.
4095 void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog)
4097 struct net_device *dev = skb->dev;
4098 struct netdev_queue *txq;
4099 bool free_skb = true;
4102 txq = netdev_pick_tx(dev, skb, NULL);
4103 cpu = smp_processor_id();
4104 HARD_TX_LOCK(dev, txq, cpu);
4105 if (!netif_xmit_stopped(txq)) {
4106 rc = netdev_start_xmit(skb, dev, txq, 0);
4107 if (dev_xmit_complete(rc))
4110 HARD_TX_UNLOCK(dev, txq);
4112 trace_xdp_exception(dev, xdp_prog, XDP_TX);
4116 EXPORT_SYMBOL_GPL(generic_xdp_tx);
4118 static struct static_key generic_xdp_needed __read_mostly;
4120 int do_xdp_generic(struct bpf_prog *xdp_prog, struct sk_buff *skb)
4123 u32 act = netif_receive_generic_xdp(skb, xdp_prog);
4126 if (act != XDP_PASS) {
4129 err = xdp_do_generic_redirect(skb->dev, skb,
4133 /* fallthru to submit skb */
4135 generic_xdp_tx(skb, xdp_prog);
4146 EXPORT_SYMBOL_GPL(do_xdp_generic);
4148 static int netif_rx_internal(struct sk_buff *skb)
4152 net_timestamp_check(netdev_tstamp_prequeue, skb);
4154 trace_netif_rx(skb);
4156 if (static_key_false(&generic_xdp_needed)) {
4161 ret = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
4165 /* Consider XDP consuming the packet a success from
4166 * the netdev point of view we do not want to count
4169 if (ret != XDP_PASS)
4170 return NET_RX_SUCCESS;
4174 if (static_key_false(&rps_needed)) {
4175 struct rps_dev_flow voidflow, *rflow = &voidflow;
4181 cpu = get_rps_cpu(skb->dev, skb, &rflow);
4183 cpu = smp_processor_id();
4185 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4194 ret = enqueue_to_backlog(skb, get_cpu(), &qtail);
4201 * netif_rx - post buffer to the network code
4202 * @skb: buffer to post
4204 * This function receives a packet from a device driver and queues it for
4205 * the upper (protocol) levels to process. It always succeeds. The buffer
4206 * may be dropped during processing for congestion control or by the
4210 * NET_RX_SUCCESS (no congestion)
4211 * NET_RX_DROP (packet was dropped)
4215 int netif_rx(struct sk_buff *skb)
4217 trace_netif_rx_entry(skb);
4219 return netif_rx_internal(skb);
4221 EXPORT_SYMBOL(netif_rx);
4223 int netif_rx_ni(struct sk_buff *skb)
4227 trace_netif_rx_ni_entry(skb);
4230 err = netif_rx_internal(skb);
4231 if (local_softirq_pending())
4237 EXPORT_SYMBOL(netif_rx_ni);
4239 static __latent_entropy void net_tx_action(struct softirq_action *h)
4241 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
4243 if (sd->completion_queue) {
4244 struct sk_buff *clist;
4246 local_irq_disable();
4247 clist = sd->completion_queue;
4248 sd->completion_queue = NULL;
4252 struct sk_buff *skb = clist;
4254 clist = clist->next;
4256 WARN_ON(refcount_read(&skb->users));
4257 if (likely(get_kfree_skb_cb(skb)->reason == SKB_REASON_CONSUMED))
4258 trace_consume_skb(skb);
4260 trace_kfree_skb(skb, net_tx_action);
4262 if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
4265 __kfree_skb_defer(skb);
4268 __kfree_skb_flush();
4271 if (sd->output_queue) {
4274 local_irq_disable();
4275 head = sd->output_queue;
4276 sd->output_queue = NULL;
4277 sd->output_queue_tailp = &sd->output_queue;
4281 struct Qdisc *q = head;
4282 spinlock_t *root_lock = NULL;
4284 head = head->next_sched;
4286 if (!(q->flags & TCQ_F_NOLOCK)) {
4287 root_lock = qdisc_lock(q);
4288 spin_lock(root_lock);
4290 /* We need to make sure head->next_sched is read
4291 * before clearing __QDISC_STATE_SCHED
4293 smp_mb__before_atomic();
4294 clear_bit(__QDISC_STATE_SCHED, &q->state);
4297 spin_unlock(root_lock);
4301 xfrm_dev_backlog(sd);
4304 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE)
4305 /* This hook is defined here for ATM LANE */
4306 int (*br_fdb_test_addr_hook)(struct net_device *dev,
4307 unsigned char *addr) __read_mostly;
4308 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook);
4311 static inline struct sk_buff *
4312 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
4313 struct net_device *orig_dev)
4315 #ifdef CONFIG_NET_CLS_ACT
4316 struct mini_Qdisc *miniq = rcu_dereference_bh(skb->dev->miniq_ingress);
4317 struct tcf_result cl_res;
4319 /* If there's at least one ingress present somewhere (so
4320 * we get here via enabled static key), remaining devices
4321 * that are not configured with an ingress qdisc will bail
4328 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4332 qdisc_skb_cb(skb)->pkt_len = skb->len;
4333 skb->tc_at_ingress = 1;
4334 mini_qdisc_bstats_cpu_update(miniq, skb);
4336 switch (tcf_classify(skb, miniq->filter_list, &cl_res, false)) {
4338 case TC_ACT_RECLASSIFY:
4339 skb->tc_index = TC_H_MIN(cl_res.classid);
4342 mini_qdisc_qstats_cpu_drop(miniq);
4350 case TC_ACT_REDIRECT:
4351 /* skb_mac_header check was done by cls/act_bpf, so
4352 * we can safely push the L2 header back before
4353 * redirecting to another netdev
4355 __skb_push(skb, skb->mac_len);
4356 skb_do_redirect(skb);
4361 #endif /* CONFIG_NET_CLS_ACT */
4366 * netdev_is_rx_handler_busy - check if receive handler is registered
4367 * @dev: device to check
4369 * Check if a receive handler is already registered for a given device.
4370 * Return true if there one.
4372 * The caller must hold the rtnl_mutex.
4374 bool netdev_is_rx_handler_busy(struct net_device *dev)
4377 return dev && rtnl_dereference(dev->rx_handler);
4379 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy);
4382 * netdev_rx_handler_register - register receive handler
4383 * @dev: device to register a handler for
4384 * @rx_handler: receive handler to register
4385 * @rx_handler_data: data pointer that is used by rx handler
4387 * Register a receive handler for a device. This handler will then be
4388 * called from __netif_receive_skb. A negative errno code is returned
4391 * The caller must hold the rtnl_mutex.
4393 * For a general description of rx_handler, see enum rx_handler_result.
4395 int netdev_rx_handler_register(struct net_device *dev,
4396 rx_handler_func_t *rx_handler,
4397 void *rx_handler_data)
4399 if (netdev_is_rx_handler_busy(dev))
4402 if (dev->priv_flags & IFF_NO_RX_HANDLER)
4405 /* Note: rx_handler_data must be set before rx_handler */
4406 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data);
4407 rcu_assign_pointer(dev->rx_handler, rx_handler);
4411 EXPORT_SYMBOL_GPL(netdev_rx_handler_register);
4414 * netdev_rx_handler_unregister - unregister receive handler
4415 * @dev: device to unregister a handler from
4417 * Unregister a receive handler from a device.
4419 * The caller must hold the rtnl_mutex.
4421 void netdev_rx_handler_unregister(struct net_device *dev)
4425 RCU_INIT_POINTER(dev->rx_handler, NULL);
4426 /* a reader seeing a non NULL rx_handler in a rcu_read_lock()
4427 * section has a guarantee to see a non NULL rx_handler_data
4431 RCU_INIT_POINTER(dev->rx_handler_data, NULL);
4433 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister);
4436 * Limit the use of PFMEMALLOC reserves to those protocols that implement
4437 * the special handling of PFMEMALLOC skbs.
4439 static bool skb_pfmemalloc_protocol(struct sk_buff *skb)
4441 switch (skb->protocol) {
4442 case htons(ETH_P_ARP):
4443 case htons(ETH_P_IP):
4444 case htons(ETH_P_IPV6):
4445 case htons(ETH_P_8021Q):
4446 case htons(ETH_P_8021AD):
4453 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev,
4454 int *ret, struct net_device *orig_dev)
4456 #ifdef CONFIG_NETFILTER_INGRESS
4457 if (nf_hook_ingress_active(skb)) {
4461 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4466 ingress_retval = nf_hook_ingress(skb);
4468 return ingress_retval;
4470 #endif /* CONFIG_NETFILTER_INGRESS */
4474 static int __netif_receive_skb_core(struct sk_buff *skb, bool pfmemalloc)
4476 struct packet_type *ptype, *pt_prev;
4477 rx_handler_func_t *rx_handler;
4478 struct net_device *orig_dev;
4479 bool deliver_exact = false;
4480 int ret = NET_RX_DROP;
4483 net_timestamp_check(!netdev_tstamp_prequeue, skb);
4485 trace_netif_receive_skb(skb);
4487 orig_dev = skb->dev;
4489 skb_reset_network_header(skb);
4490 if (!skb_transport_header_was_set(skb))
4491 skb_reset_transport_header(skb);
4492 skb_reset_mac_len(skb);
4497 skb->skb_iif = skb->dev->ifindex;
4499 __this_cpu_inc(softnet_data.processed);
4501 if (skb->protocol == cpu_to_be16(ETH_P_8021Q) ||
4502 skb->protocol == cpu_to_be16(ETH_P_8021AD)) {
4503 skb = skb_vlan_untag(skb);
4508 if (skb_skip_tc_classify(skb))
4514 list_for_each_entry_rcu(ptype, &ptype_all, list) {
4516 ret = deliver_skb(skb, pt_prev, orig_dev);
4520 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) {
4522 ret = deliver_skb(skb, pt_prev, orig_dev);
4527 #ifdef CONFIG_NET_INGRESS
4528 if (static_key_false(&ingress_needed)) {
4529 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev);
4533 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0)
4539 if (pfmemalloc && !skb_pfmemalloc_protocol(skb))
4542 if (skb_vlan_tag_present(skb)) {
4544 ret = deliver_skb(skb, pt_prev, orig_dev);
4547 if (vlan_do_receive(&skb))
4549 else if (unlikely(!skb))
4553 rx_handler = rcu_dereference(skb->dev->rx_handler);
4556 ret = deliver_skb(skb, pt_prev, orig_dev);
4559 switch (rx_handler(&skb)) {
4560 case RX_HANDLER_CONSUMED:
4561 ret = NET_RX_SUCCESS;
4563 case RX_HANDLER_ANOTHER:
4565 case RX_HANDLER_EXACT:
4566 deliver_exact = true;
4567 case RX_HANDLER_PASS:
4574 if (unlikely(skb_vlan_tag_present(skb))) {
4575 if (skb_vlan_tag_get_id(skb))
4576 skb->pkt_type = PACKET_OTHERHOST;
4577 /* Note: we might in the future use prio bits
4578 * and set skb->priority like in vlan_do_receive()
4579 * For the time being, just ignore Priority Code Point
4584 type = skb->protocol;
4586 /* deliver only exact match when indicated */
4587 if (likely(!deliver_exact)) {
4588 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4589 &ptype_base[ntohs(type) &
4593 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4594 &orig_dev->ptype_specific);
4596 if (unlikely(skb->dev != orig_dev)) {
4597 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4598 &skb->dev->ptype_specific);
4602 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
4605 ret = pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
4609 atomic_long_inc(&skb->dev->rx_dropped);
4611 atomic_long_inc(&skb->dev->rx_nohandler);
4613 /* Jamal, now you will not able to escape explaining
4614 * me how you were going to use this. :-)
4624 * netif_receive_skb_core - special purpose version of netif_receive_skb
4625 * @skb: buffer to process
4627 * More direct receive version of netif_receive_skb(). It should
4628 * only be used by callers that have a need to skip RPS and Generic XDP.
4629 * Caller must also take care of handling if (page_is_)pfmemalloc.
4631 * This function may only be called from softirq context and interrupts
4632 * should be enabled.
4634 * Return values (usually ignored):
4635 * NET_RX_SUCCESS: no congestion
4636 * NET_RX_DROP: packet was dropped
4638 int netif_receive_skb_core(struct sk_buff *skb)
4643 ret = __netif_receive_skb_core(skb, false);
4648 EXPORT_SYMBOL(netif_receive_skb_core);
4650 static int __netif_receive_skb(struct sk_buff *skb)
4654 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) {
4655 unsigned int noreclaim_flag;
4658 * PFMEMALLOC skbs are special, they should
4659 * - be delivered to SOCK_MEMALLOC sockets only
4660 * - stay away from userspace
4661 * - have bounded memory usage
4663 * Use PF_MEMALLOC as this saves us from propagating the allocation
4664 * context down to all allocation sites.
4666 noreclaim_flag = memalloc_noreclaim_save();
4667 ret = __netif_receive_skb_core(skb, true);
4668 memalloc_noreclaim_restore(noreclaim_flag);
4670 ret = __netif_receive_skb_core(skb, false);
4675 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp)
4677 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog);
4678 struct bpf_prog *new = xdp->prog;
4681 switch (xdp->command) {
4682 case XDP_SETUP_PROG:
4683 rcu_assign_pointer(dev->xdp_prog, new);
4688 static_key_slow_dec(&generic_xdp_needed);
4689 } else if (new && !old) {
4690 static_key_slow_inc(&generic_xdp_needed);
4691 dev_disable_lro(dev);
4692 dev_disable_gro_hw(dev);
4696 case XDP_QUERY_PROG:
4697 xdp->prog_attached = !!old;
4698 xdp->prog_id = old ? old->aux->id : 0;
4709 static int netif_receive_skb_internal(struct sk_buff *skb)
4713 net_timestamp_check(netdev_tstamp_prequeue, skb);
4715 if (skb_defer_rx_timestamp(skb))
4716 return NET_RX_SUCCESS;
4718 if (static_key_false(&generic_xdp_needed)) {
4723 ret = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
4727 if (ret != XDP_PASS)
4733 if (static_key_false(&rps_needed)) {
4734 struct rps_dev_flow voidflow, *rflow = &voidflow;
4735 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
4738 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4744 ret = __netif_receive_skb(skb);
4750 * netif_receive_skb - process receive buffer from network
4751 * @skb: buffer to process
4753 * netif_receive_skb() is the main receive data processing function.
4754 * It always succeeds. The buffer may be dropped during processing
4755 * for congestion control or by the protocol layers.
4757 * This function may only be called from softirq context and interrupts
4758 * should be enabled.
4760 * Return values (usually ignored):
4761 * NET_RX_SUCCESS: no congestion
4762 * NET_RX_DROP: packet was dropped
4764 int netif_receive_skb(struct sk_buff *skb)
4766 trace_netif_receive_skb_entry(skb);
4768 return netif_receive_skb_internal(skb);
4770 EXPORT_SYMBOL(netif_receive_skb);
4772 DEFINE_PER_CPU(struct work_struct, flush_works);
4774 /* Network device is going away, flush any packets still pending */
4775 static void flush_backlog(struct work_struct *work)
4777 struct sk_buff *skb, *tmp;
4778 struct softnet_data *sd;
4781 sd = this_cpu_ptr(&softnet_data);
4783 local_irq_disable();
4785 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) {
4786 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
4787 __skb_unlink(skb, &sd->input_pkt_queue);
4789 input_queue_head_incr(sd);
4795 skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
4796 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
4797 __skb_unlink(skb, &sd->process_queue);
4799 input_queue_head_incr(sd);
4805 static void flush_all_backlogs(void)
4811 for_each_online_cpu(cpu)
4812 queue_work_on(cpu, system_highpri_wq,
4813 per_cpu_ptr(&flush_works, cpu));
4815 for_each_online_cpu(cpu)
4816 flush_work(per_cpu_ptr(&flush_works, cpu));
4821 static int napi_gro_complete(struct sk_buff *skb)
4823 struct packet_offload *ptype;
4824 __be16 type = skb->protocol;
4825 struct list_head *head = &offload_base;
4828 BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
4830 if (NAPI_GRO_CB(skb)->count == 1) {
4831 skb_shinfo(skb)->gso_size = 0;
4836 list_for_each_entry_rcu(ptype, head, list) {
4837 if (ptype->type != type || !ptype->callbacks.gro_complete)
4840 err = ptype->callbacks.gro_complete(skb, 0);
4846 WARN_ON(&ptype->list == head);
4848 return NET_RX_SUCCESS;
4852 return netif_receive_skb_internal(skb);
4855 /* napi->gro_list contains packets ordered by age.
4856 * youngest packets at the head of it.
4857 * Complete skbs in reverse order to reduce latencies.
4859 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
4861 struct sk_buff *skb, *prev = NULL;
4863 /* scan list and build reverse chain */
4864 for (skb = napi->gro_list; skb != NULL; skb = skb->next) {
4869 for (skb = prev; skb; skb = prev) {
4872 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
4876 napi_gro_complete(skb);
4880 napi->gro_list = NULL;
4882 EXPORT_SYMBOL(napi_gro_flush);
4884 static void gro_list_prepare(struct napi_struct *napi, struct sk_buff *skb)
4887 unsigned int maclen = skb->dev->hard_header_len;
4888 u32 hash = skb_get_hash_raw(skb);
4890 for (p = napi->gro_list; p; p = p->next) {
4891 unsigned long diffs;
4893 NAPI_GRO_CB(p)->flush = 0;
4895 if (hash != skb_get_hash_raw(p)) {
4896 NAPI_GRO_CB(p)->same_flow = 0;
4900 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
4901 diffs |= p->vlan_tci ^ skb->vlan_tci;
4902 diffs |= skb_metadata_dst_cmp(p, skb);
4903 diffs |= skb_metadata_differs(p, skb);
4904 if (maclen == ETH_HLEN)
4905 diffs |= compare_ether_header(skb_mac_header(p),
4906 skb_mac_header(skb));
4908 diffs = memcmp(skb_mac_header(p),
4909 skb_mac_header(skb),
4911 NAPI_GRO_CB(p)->same_flow = !diffs;
4915 static void skb_gro_reset_offset(struct sk_buff *skb)
4917 const struct skb_shared_info *pinfo = skb_shinfo(skb);
4918 const skb_frag_t *frag0 = &pinfo->frags[0];
4920 NAPI_GRO_CB(skb)->data_offset = 0;
4921 NAPI_GRO_CB(skb)->frag0 = NULL;
4922 NAPI_GRO_CB(skb)->frag0_len = 0;
4924 if (skb_mac_header(skb) == skb_tail_pointer(skb) &&
4926 !PageHighMem(skb_frag_page(frag0))) {
4927 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
4928 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
4929 skb_frag_size(frag0),
4930 skb->end - skb->tail);
4934 static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
4936 struct skb_shared_info *pinfo = skb_shinfo(skb);
4938 BUG_ON(skb->end - skb->tail < grow);
4940 memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
4942 skb->data_len -= grow;
4945 pinfo->frags[0].page_offset += grow;
4946 skb_frag_size_sub(&pinfo->frags[0], grow);
4948 if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
4949 skb_frag_unref(skb, 0);
4950 memmove(pinfo->frags, pinfo->frags + 1,
4951 --pinfo->nr_frags * sizeof(pinfo->frags[0]));
4955 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
4957 struct sk_buff **pp = NULL;
4958 struct packet_offload *ptype;
4959 __be16 type = skb->protocol;
4960 struct list_head *head = &offload_base;
4962 enum gro_result ret;
4965 if (netif_elide_gro(skb->dev))
4968 gro_list_prepare(napi, skb);
4971 list_for_each_entry_rcu(ptype, head, list) {
4972 if (ptype->type != type || !ptype->callbacks.gro_receive)
4975 skb_set_network_header(skb, skb_gro_offset(skb));
4976 skb_reset_mac_len(skb);
4977 NAPI_GRO_CB(skb)->same_flow = 0;
4978 NAPI_GRO_CB(skb)->flush = skb_is_gso(skb) || skb_has_frag_list(skb);
4979 NAPI_GRO_CB(skb)->free = 0;
4980 NAPI_GRO_CB(skb)->encap_mark = 0;
4981 NAPI_GRO_CB(skb)->recursion_counter = 0;
4982 NAPI_GRO_CB(skb)->is_fou = 0;
4983 NAPI_GRO_CB(skb)->is_atomic = 1;
4984 NAPI_GRO_CB(skb)->gro_remcsum_start = 0;
4986 /* Setup for GRO checksum validation */
4987 switch (skb->ip_summed) {
4988 case CHECKSUM_COMPLETE:
4989 NAPI_GRO_CB(skb)->csum = skb->csum;
4990 NAPI_GRO_CB(skb)->csum_valid = 1;
4991 NAPI_GRO_CB(skb)->csum_cnt = 0;
4993 case CHECKSUM_UNNECESSARY:
4994 NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
4995 NAPI_GRO_CB(skb)->csum_valid = 0;
4998 NAPI_GRO_CB(skb)->csum_cnt = 0;
4999 NAPI_GRO_CB(skb)->csum_valid = 0;
5002 pp = ptype->callbacks.gro_receive(&napi->gro_list, skb);
5007 if (&ptype->list == head)
5010 if (IS_ERR(pp) && PTR_ERR(pp) == -EINPROGRESS) {
5015 same_flow = NAPI_GRO_CB(skb)->same_flow;
5016 ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
5019 struct sk_buff *nskb = *pp;
5023 napi_gro_complete(nskb);
5030 if (NAPI_GRO_CB(skb)->flush)
5033 if (unlikely(napi->gro_count >= MAX_GRO_SKBS)) {
5034 struct sk_buff *nskb = napi->gro_list;
5036 /* locate the end of the list to select the 'oldest' flow */
5037 while (nskb->next) {
5043 napi_gro_complete(nskb);
5047 NAPI_GRO_CB(skb)->count = 1;
5048 NAPI_GRO_CB(skb)->age = jiffies;
5049 NAPI_GRO_CB(skb)->last = skb;
5050 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
5051 skb->next = napi->gro_list;
5052 napi->gro_list = skb;
5056 grow = skb_gro_offset(skb) - skb_headlen(skb);
5058 gro_pull_from_frag0(skb, grow);
5067 struct packet_offload *gro_find_receive_by_type(__be16 type)
5069 struct list_head *offload_head = &offload_base;
5070 struct packet_offload *ptype;
5072 list_for_each_entry_rcu(ptype, offload_head, list) {
5073 if (ptype->type != type || !ptype->callbacks.gro_receive)
5079 EXPORT_SYMBOL(gro_find_receive_by_type);
5081 struct packet_offload *gro_find_complete_by_type(__be16 type)
5083 struct list_head *offload_head = &offload_base;
5084 struct packet_offload *ptype;
5086 list_for_each_entry_rcu(ptype, offload_head, list) {
5087 if (ptype->type != type || !ptype->callbacks.gro_complete)
5093 EXPORT_SYMBOL(gro_find_complete_by_type);
5095 static void napi_skb_free_stolen_head(struct sk_buff *skb)
5099 kmem_cache_free(skbuff_head_cache, skb);
5102 static gro_result_t napi_skb_finish(gro_result_t ret, struct sk_buff *skb)
5106 if (netif_receive_skb_internal(skb))
5114 case GRO_MERGED_FREE:
5115 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
5116 napi_skb_free_stolen_head(skb);
5130 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
5132 skb_mark_napi_id(skb, napi);
5133 trace_napi_gro_receive_entry(skb);
5135 skb_gro_reset_offset(skb);
5137 return napi_skb_finish(dev_gro_receive(napi, skb), skb);
5139 EXPORT_SYMBOL(napi_gro_receive);
5141 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
5143 if (unlikely(skb->pfmemalloc)) {
5147 __skb_pull(skb, skb_headlen(skb));
5148 /* restore the reserve we had after netdev_alloc_skb_ip_align() */
5149 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
5151 skb->dev = napi->dev;
5153 skb->encapsulation = 0;
5154 skb_shinfo(skb)->gso_type = 0;
5155 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
5161 struct sk_buff *napi_get_frags(struct napi_struct *napi)
5163 struct sk_buff *skb = napi->skb;
5166 skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
5169 skb_mark_napi_id(skb, napi);
5174 EXPORT_SYMBOL(napi_get_frags);
5176 static gro_result_t napi_frags_finish(struct napi_struct *napi,
5177 struct sk_buff *skb,
5183 __skb_push(skb, ETH_HLEN);
5184 skb->protocol = eth_type_trans(skb, skb->dev);
5185 if (ret == GRO_NORMAL && netif_receive_skb_internal(skb))
5190 napi_reuse_skb(napi, skb);
5193 case GRO_MERGED_FREE:
5194 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
5195 napi_skb_free_stolen_head(skb);
5197 napi_reuse_skb(napi, skb);
5208 /* Upper GRO stack assumes network header starts at gro_offset=0
5209 * Drivers could call both napi_gro_frags() and napi_gro_receive()
5210 * We copy ethernet header into skb->data to have a common layout.
5212 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
5214 struct sk_buff *skb = napi->skb;
5215 const struct ethhdr *eth;
5216 unsigned int hlen = sizeof(*eth);
5220 skb_reset_mac_header(skb);
5221 skb_gro_reset_offset(skb);
5223 eth = skb_gro_header_fast(skb, 0);
5224 if (unlikely(skb_gro_header_hard(skb, hlen))) {
5225 eth = skb_gro_header_slow(skb, hlen, 0);
5226 if (unlikely(!eth)) {
5227 net_warn_ratelimited("%s: dropping impossible skb from %s\n",
5228 __func__, napi->dev->name);
5229 napi_reuse_skb(napi, skb);
5233 gro_pull_from_frag0(skb, hlen);
5234 NAPI_GRO_CB(skb)->frag0 += hlen;
5235 NAPI_GRO_CB(skb)->frag0_len -= hlen;
5237 __skb_pull(skb, hlen);
5240 * This works because the only protocols we care about don't require
5242 * We'll fix it up properly in napi_frags_finish()
5244 skb->protocol = eth->h_proto;
5249 gro_result_t napi_gro_frags(struct napi_struct *napi)
5251 struct sk_buff *skb = napi_frags_skb(napi);
5256 trace_napi_gro_frags_entry(skb);
5258 return napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
5260 EXPORT_SYMBOL(napi_gro_frags);
5262 /* Compute the checksum from gro_offset and return the folded value
5263 * after adding in any pseudo checksum.
5265 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
5270 wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
5272 /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
5273 sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
5275 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
5276 !skb->csum_complete_sw)
5277 netdev_rx_csum_fault(skb->dev);
5280 NAPI_GRO_CB(skb)->csum = wsum;
5281 NAPI_GRO_CB(skb)->csum_valid = 1;
5285 EXPORT_SYMBOL(__skb_gro_checksum_complete);
5287 static void net_rps_send_ipi(struct softnet_data *remsd)
5291 struct softnet_data *next = remsd->rps_ipi_next;
5293 if (cpu_online(remsd->cpu))
5294 smp_call_function_single_async(remsd->cpu, &remsd->csd);
5301 * net_rps_action_and_irq_enable sends any pending IPI's for rps.
5302 * Note: called with local irq disabled, but exits with local irq enabled.
5304 static void net_rps_action_and_irq_enable(struct softnet_data *sd)
5307 struct softnet_data *remsd = sd->rps_ipi_list;
5310 sd->rps_ipi_list = NULL;
5314 /* Send pending IPI's to kick RPS processing on remote cpus. */
5315 net_rps_send_ipi(remsd);
5321 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd)
5324 return sd->rps_ipi_list != NULL;
5330 static int process_backlog(struct napi_struct *napi, int quota)
5332 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
5336 /* Check if we have pending ipi, its better to send them now,
5337 * not waiting net_rx_action() end.
5339 if (sd_has_rps_ipi_waiting(sd)) {
5340 local_irq_disable();
5341 net_rps_action_and_irq_enable(sd);
5344 napi->weight = dev_rx_weight;
5346 struct sk_buff *skb;
5348 while ((skb = __skb_dequeue(&sd->process_queue))) {
5350 __netif_receive_skb(skb);
5352 input_queue_head_incr(sd);
5353 if (++work >= quota)
5358 local_irq_disable();
5360 if (skb_queue_empty(&sd->input_pkt_queue)) {
5362 * Inline a custom version of __napi_complete().
5363 * only current cpu owns and manipulates this napi,
5364 * and NAPI_STATE_SCHED is the only possible flag set
5366 * We can use a plain write instead of clear_bit(),
5367 * and we dont need an smp_mb() memory barrier.
5372 skb_queue_splice_tail_init(&sd->input_pkt_queue,
5373 &sd->process_queue);
5383 * __napi_schedule - schedule for receive
5384 * @n: entry to schedule
5386 * The entry's receive function will be scheduled to run.
5387 * Consider using __napi_schedule_irqoff() if hard irqs are masked.
5389 void __napi_schedule(struct napi_struct *n)
5391 unsigned long flags;
5393 local_irq_save(flags);
5394 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
5395 local_irq_restore(flags);
5397 EXPORT_SYMBOL(__napi_schedule);
5400 * napi_schedule_prep - check if napi can be scheduled
5403 * Test if NAPI routine is already running, and if not mark
5404 * it as running. This is used as a condition variable
5405 * insure only one NAPI poll instance runs. We also make
5406 * sure there is no pending NAPI disable.
5408 bool napi_schedule_prep(struct napi_struct *n)
5410 unsigned long val, new;
5413 val = READ_ONCE(n->state);
5414 if (unlikely(val & NAPIF_STATE_DISABLE))
5416 new = val | NAPIF_STATE_SCHED;
5418 /* Sets STATE_MISSED bit if STATE_SCHED was already set
5419 * This was suggested by Alexander Duyck, as compiler
5420 * emits better code than :
5421 * if (val & NAPIF_STATE_SCHED)
5422 * new |= NAPIF_STATE_MISSED;
5424 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED *
5426 } while (cmpxchg(&n->state, val, new) != val);
5428 return !(val & NAPIF_STATE_SCHED);
5430 EXPORT_SYMBOL(napi_schedule_prep);
5433 * __napi_schedule_irqoff - schedule for receive
5434 * @n: entry to schedule
5436 * Variant of __napi_schedule() assuming hard irqs are masked
5438 void __napi_schedule_irqoff(struct napi_struct *n)
5440 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
5442 EXPORT_SYMBOL(__napi_schedule_irqoff);
5444 bool napi_complete_done(struct napi_struct *n, int work_done)
5446 unsigned long flags, val, new;
5449 * 1) Don't let napi dequeue from the cpu poll list
5450 * just in case its running on a different cpu.
5451 * 2) If we are busy polling, do nothing here, we have
5452 * the guarantee we will be called later.
5454 if (unlikely(n->state & (NAPIF_STATE_NPSVC |
5455 NAPIF_STATE_IN_BUSY_POLL)))
5459 unsigned long timeout = 0;
5462 timeout = n->dev->gro_flush_timeout;
5465 hrtimer_start(&n->timer, ns_to_ktime(timeout),
5466 HRTIMER_MODE_REL_PINNED);
5468 napi_gro_flush(n, false);
5470 if (unlikely(!list_empty(&n->poll_list))) {
5471 /* If n->poll_list is not empty, we need to mask irqs */
5472 local_irq_save(flags);
5473 list_del_init(&n->poll_list);
5474 local_irq_restore(flags);
5478 val = READ_ONCE(n->state);
5480 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED));
5482 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED);
5484 /* If STATE_MISSED was set, leave STATE_SCHED set,
5485 * because we will call napi->poll() one more time.
5486 * This C code was suggested by Alexander Duyck to help gcc.
5488 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED *
5490 } while (cmpxchg(&n->state, val, new) != val);
5492 if (unlikely(val & NAPIF_STATE_MISSED)) {
5499 EXPORT_SYMBOL(napi_complete_done);
5501 /* must be called under rcu_read_lock(), as we dont take a reference */
5502 static struct napi_struct *napi_by_id(unsigned int napi_id)
5504 unsigned int hash = napi_id % HASH_SIZE(napi_hash);
5505 struct napi_struct *napi;
5507 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node)
5508 if (napi->napi_id == napi_id)
5514 #if defined(CONFIG_NET_RX_BUSY_POLL)
5516 #define BUSY_POLL_BUDGET 8
5518 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock)
5522 /* Busy polling means there is a high chance device driver hard irq
5523 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was
5524 * set in napi_schedule_prep().
5525 * Since we are about to call napi->poll() once more, we can safely
5526 * clear NAPI_STATE_MISSED.
5528 * Note: x86 could use a single "lock and ..." instruction
5529 * to perform these two clear_bit()
5531 clear_bit(NAPI_STATE_MISSED, &napi->state);
5532 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state);
5536 /* All we really want here is to re-enable device interrupts.
5537 * Ideally, a new ndo_busy_poll_stop() could avoid another round.
5539 rc = napi->poll(napi, BUSY_POLL_BUDGET);
5540 trace_napi_poll(napi, rc, BUSY_POLL_BUDGET);
5541 netpoll_poll_unlock(have_poll_lock);
5542 if (rc == BUSY_POLL_BUDGET)
5543 __napi_schedule(napi);
5547 void napi_busy_loop(unsigned int napi_id,
5548 bool (*loop_end)(void *, unsigned long),
5551 unsigned long start_time = loop_end ? busy_loop_current_time() : 0;
5552 int (*napi_poll)(struct napi_struct *napi, int budget);
5553 void *have_poll_lock = NULL;
5554 struct napi_struct *napi;
5561 napi = napi_by_id(napi_id);
5571 unsigned long val = READ_ONCE(napi->state);
5573 /* If multiple threads are competing for this napi,
5574 * we avoid dirtying napi->state as much as we can.
5576 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED |
5577 NAPIF_STATE_IN_BUSY_POLL))
5579 if (cmpxchg(&napi->state, val,
5580 val | NAPIF_STATE_IN_BUSY_POLL |
5581 NAPIF_STATE_SCHED) != val)
5583 have_poll_lock = netpoll_poll_lock(napi);
5584 napi_poll = napi->poll;
5586 work = napi_poll(napi, BUSY_POLL_BUDGET);
5587 trace_napi_poll(napi, work, BUSY_POLL_BUDGET);
5590 __NET_ADD_STATS(dev_net(napi->dev),
5591 LINUX_MIB_BUSYPOLLRXPACKETS, work);
5594 if (!loop_end || loop_end(loop_end_arg, start_time))
5597 if (unlikely(need_resched())) {
5599 busy_poll_stop(napi, have_poll_lock);
5603 if (loop_end(loop_end_arg, start_time))
5610 busy_poll_stop(napi, have_poll_lock);
5615 EXPORT_SYMBOL(napi_busy_loop);
5617 #endif /* CONFIG_NET_RX_BUSY_POLL */
5619 static void napi_hash_add(struct napi_struct *napi)
5621 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state) ||
5622 test_and_set_bit(NAPI_STATE_HASHED, &napi->state))
5625 spin_lock(&napi_hash_lock);
5627 /* 0..NR_CPUS range is reserved for sender_cpu use */
5629 if (unlikely(++napi_gen_id < MIN_NAPI_ID))
5630 napi_gen_id = MIN_NAPI_ID;
5631 } while (napi_by_id(napi_gen_id));
5632 napi->napi_id = napi_gen_id;
5634 hlist_add_head_rcu(&napi->napi_hash_node,
5635 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]);
5637 spin_unlock(&napi_hash_lock);
5640 /* Warning : caller is responsible to make sure rcu grace period
5641 * is respected before freeing memory containing @napi
5643 bool napi_hash_del(struct napi_struct *napi)
5645 bool rcu_sync_needed = false;
5647 spin_lock(&napi_hash_lock);
5649 if (test_and_clear_bit(NAPI_STATE_HASHED, &napi->state)) {
5650 rcu_sync_needed = true;
5651 hlist_del_rcu(&napi->napi_hash_node);
5653 spin_unlock(&napi_hash_lock);
5654 return rcu_sync_needed;
5656 EXPORT_SYMBOL_GPL(napi_hash_del);
5658 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer)
5660 struct napi_struct *napi;
5662 napi = container_of(timer, struct napi_struct, timer);
5664 /* Note : we use a relaxed variant of napi_schedule_prep() not setting
5665 * NAPI_STATE_MISSED, since we do not react to a device IRQ.
5667 if (napi->gro_list && !napi_disable_pending(napi) &&
5668 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state))
5669 __napi_schedule_irqoff(napi);
5671 return HRTIMER_NORESTART;
5674 void netif_napi_add(struct net_device *dev, struct napi_struct *napi,
5675 int (*poll)(struct napi_struct *, int), int weight)
5677 INIT_LIST_HEAD(&napi->poll_list);
5678 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
5679 napi->timer.function = napi_watchdog;
5680 napi->gro_count = 0;
5681 napi->gro_list = NULL;
5684 if (weight > NAPI_POLL_WEIGHT)
5685 pr_err_once("netif_napi_add() called with weight %d on device %s\n",
5687 napi->weight = weight;
5688 list_add(&napi->dev_list, &dev->napi_list);
5690 #ifdef CONFIG_NETPOLL
5691 napi->poll_owner = -1;
5693 set_bit(NAPI_STATE_SCHED, &napi->state);
5694 napi_hash_add(napi);
5696 EXPORT_SYMBOL(netif_napi_add);
5698 void napi_disable(struct napi_struct *n)
5701 set_bit(NAPI_STATE_DISABLE, &n->state);
5703 while (test_and_set_bit(NAPI_STATE_SCHED, &n->state))
5705 while (test_and_set_bit(NAPI_STATE_NPSVC, &n->state))
5708 hrtimer_cancel(&n->timer);
5710 clear_bit(NAPI_STATE_DISABLE, &n->state);
5712 EXPORT_SYMBOL(napi_disable);
5714 /* Must be called in process context */
5715 void netif_napi_del(struct napi_struct *napi)
5718 if (napi_hash_del(napi))
5720 list_del_init(&napi->dev_list);
5721 napi_free_frags(napi);
5723 kfree_skb_list(napi->gro_list);
5724 napi->gro_list = NULL;
5725 napi->gro_count = 0;
5727 EXPORT_SYMBOL(netif_napi_del);
5729 static int napi_poll(struct napi_struct *n, struct list_head *repoll)
5734 list_del_init(&n->poll_list);
5736 have = netpoll_poll_lock(n);
5740 /* This NAPI_STATE_SCHED test is for avoiding a race
5741 * with netpoll's poll_napi(). Only the entity which
5742 * obtains the lock and sees NAPI_STATE_SCHED set will
5743 * actually make the ->poll() call. Therefore we avoid
5744 * accidentally calling ->poll() when NAPI is not scheduled.
5747 if (test_bit(NAPI_STATE_SCHED, &n->state)) {
5748 work = n->poll(n, weight);
5749 trace_napi_poll(n, work, weight);
5752 WARN_ON_ONCE(work > weight);
5754 if (likely(work < weight))
5757 /* Drivers must not modify the NAPI state if they
5758 * consume the entire weight. In such cases this code
5759 * still "owns" the NAPI instance and therefore can
5760 * move the instance around on the list at-will.
5762 if (unlikely(napi_disable_pending(n))) {
5768 /* flush too old packets
5769 * If HZ < 1000, flush all packets.
5771 napi_gro_flush(n, HZ >= 1000);
5774 /* Some drivers may have called napi_schedule
5775 * prior to exhausting their budget.
5777 if (unlikely(!list_empty(&n->poll_list))) {
5778 pr_warn_once("%s: Budget exhausted after napi rescheduled\n",
5779 n->dev ? n->dev->name : "backlog");
5783 list_add_tail(&n->poll_list, repoll);
5786 netpoll_poll_unlock(have);
5791 static __latent_entropy void net_rx_action(struct softirq_action *h)
5793 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
5794 unsigned long time_limit = jiffies +
5795 usecs_to_jiffies(netdev_budget_usecs);
5796 int budget = netdev_budget;
5800 local_irq_disable();
5801 list_splice_init(&sd->poll_list, &list);
5805 struct napi_struct *n;
5807 if (list_empty(&list)) {
5808 if (!sd_has_rps_ipi_waiting(sd) && list_empty(&repoll))
5813 n = list_first_entry(&list, struct napi_struct, poll_list);
5814 budget -= napi_poll(n, &repoll);
5816 /* If softirq window is exhausted then punt.
5817 * Allow this to run for 2 jiffies since which will allow
5818 * an average latency of 1.5/HZ.
5820 if (unlikely(budget <= 0 ||
5821 time_after_eq(jiffies, time_limit))) {
5827 local_irq_disable();
5829 list_splice_tail_init(&sd->poll_list, &list);
5830 list_splice_tail(&repoll, &list);
5831 list_splice(&list, &sd->poll_list);
5832 if (!list_empty(&sd->poll_list))
5833 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
5835 net_rps_action_and_irq_enable(sd);
5837 __kfree_skb_flush();
5840 struct netdev_adjacent {
5841 struct net_device *dev;
5843 /* upper master flag, there can only be one master device per list */
5846 /* counter for the number of times this device was added to us */
5849 /* private field for the users */
5852 struct list_head list;
5853 struct rcu_head rcu;
5856 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev,
5857 struct list_head *adj_list)
5859 struct netdev_adjacent *adj;
5861 list_for_each_entry(adj, adj_list, list) {
5862 if (adj->dev == adj_dev)
5868 static int __netdev_has_upper_dev(struct net_device *upper_dev, void *data)
5870 struct net_device *dev = data;
5872 return upper_dev == dev;
5876 * netdev_has_upper_dev - Check if device is linked to an upper device
5878 * @upper_dev: upper device to check
5880 * Find out if a device is linked to specified upper device and return true
5881 * in case it is. Note that this checks only immediate upper device,
5882 * not through a complete stack of devices. The caller must hold the RTNL lock.
5884 bool netdev_has_upper_dev(struct net_device *dev,
5885 struct net_device *upper_dev)
5889 return netdev_walk_all_upper_dev_rcu(dev, __netdev_has_upper_dev,
5892 EXPORT_SYMBOL(netdev_has_upper_dev);
5895 * netdev_has_upper_dev_all - Check if device is linked to an upper device
5897 * @upper_dev: upper device to check
5899 * Find out if a device is linked to specified upper device and return true
5900 * in case it is. Note that this checks the entire upper device chain.
5901 * The caller must hold rcu lock.
5904 bool netdev_has_upper_dev_all_rcu(struct net_device *dev,
5905 struct net_device *upper_dev)
5907 return !!netdev_walk_all_upper_dev_rcu(dev, __netdev_has_upper_dev,
5910 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu);
5913 * netdev_has_any_upper_dev - Check if device is linked to some device
5916 * Find out if a device is linked to an upper device and return true in case
5917 * it is. The caller must hold the RTNL lock.
5919 bool netdev_has_any_upper_dev(struct net_device *dev)
5923 return !list_empty(&dev->adj_list.upper);
5925 EXPORT_SYMBOL(netdev_has_any_upper_dev);
5928 * netdev_master_upper_dev_get - Get master upper device
5931 * Find a master upper device and return pointer to it or NULL in case
5932 * it's not there. The caller must hold the RTNL lock.
5934 struct net_device *netdev_master_upper_dev_get(struct net_device *dev)
5936 struct netdev_adjacent *upper;
5940 if (list_empty(&dev->adj_list.upper))
5943 upper = list_first_entry(&dev->adj_list.upper,
5944 struct netdev_adjacent, list);
5945 if (likely(upper->master))
5949 EXPORT_SYMBOL(netdev_master_upper_dev_get);
5952 * netdev_has_any_lower_dev - Check if device is linked to some device
5955 * Find out if a device is linked to a lower device and return true in case
5956 * it is. The caller must hold the RTNL lock.
5958 static bool netdev_has_any_lower_dev(struct net_device *dev)
5962 return !list_empty(&dev->adj_list.lower);
5965 void *netdev_adjacent_get_private(struct list_head *adj_list)
5967 struct netdev_adjacent *adj;
5969 adj = list_entry(adj_list, struct netdev_adjacent, list);
5971 return adj->private;
5973 EXPORT_SYMBOL(netdev_adjacent_get_private);
5976 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list
5978 * @iter: list_head ** of the current position
5980 * Gets the next device from the dev's upper list, starting from iter
5981 * position. The caller must hold RCU read lock.
5983 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev,
5984 struct list_head **iter)
5986 struct netdev_adjacent *upper;
5988 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
5990 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
5992 if (&upper->list == &dev->adj_list.upper)
5995 *iter = &upper->list;
5999 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu);
6001 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev,
6002 struct list_head **iter)
6004 struct netdev_adjacent *upper;
6006 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
6008 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6010 if (&upper->list == &dev->adj_list.upper)
6013 *iter = &upper->list;
6018 int netdev_walk_all_upper_dev_rcu(struct net_device *dev,
6019 int (*fn)(struct net_device *dev,
6023 struct net_device *udev;
6024 struct list_head *iter;
6027 for (iter = &dev->adj_list.upper,
6028 udev = netdev_next_upper_dev_rcu(dev, &iter);
6030 udev = netdev_next_upper_dev_rcu(dev, &iter)) {
6031 /* first is the upper device itself */
6032 ret = fn(udev, data);
6036 /* then look at all of its upper devices */
6037 ret = netdev_walk_all_upper_dev_rcu(udev, fn, data);
6044 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu);
6047 * netdev_lower_get_next_private - Get the next ->private from the
6048 * lower neighbour list
6050 * @iter: list_head ** of the current position
6052 * Gets the next netdev_adjacent->private from the dev's lower neighbour
6053 * list, starting from iter position. The caller must hold either hold the
6054 * RTNL lock or its own locking that guarantees that the neighbour lower
6055 * list will remain unchanged.
6057 void *netdev_lower_get_next_private(struct net_device *dev,
6058 struct list_head **iter)
6060 struct netdev_adjacent *lower;
6062 lower = list_entry(*iter, struct netdev_adjacent, list);
6064 if (&lower->list == &dev->adj_list.lower)
6067 *iter = lower->list.next;
6069 return lower->private;
6071 EXPORT_SYMBOL(netdev_lower_get_next_private);
6074 * netdev_lower_get_next_private_rcu - Get the next ->private from the
6075 * lower neighbour list, RCU
6078 * @iter: list_head ** of the current position
6080 * Gets the next netdev_adjacent->private from the dev's lower neighbour
6081 * list, starting from iter position. The caller must hold RCU read lock.
6083 void *netdev_lower_get_next_private_rcu(struct net_device *dev,
6084 struct list_head **iter)
6086 struct netdev_adjacent *lower;
6088 WARN_ON_ONCE(!rcu_read_lock_held());
6090 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6092 if (&lower->list == &dev->adj_list.lower)
6095 *iter = &lower->list;
6097 return lower->private;
6099 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu);
6102 * netdev_lower_get_next - Get the next device from the lower neighbour
6105 * @iter: list_head ** of the current position
6107 * Gets the next netdev_adjacent from the dev's lower neighbour
6108 * list, starting from iter position. The caller must hold RTNL lock or
6109 * its own locking that guarantees that the neighbour lower
6110 * list will remain unchanged.
6112 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter)
6114 struct netdev_adjacent *lower;
6116 lower = list_entry(*iter, struct netdev_adjacent, list);
6118 if (&lower->list == &dev->adj_list.lower)
6121 *iter = lower->list.next;
6125 EXPORT_SYMBOL(netdev_lower_get_next);
6127 static struct net_device *netdev_next_lower_dev(struct net_device *dev,
6128 struct list_head **iter)
6130 struct netdev_adjacent *lower;
6132 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
6134 if (&lower->list == &dev->adj_list.lower)
6137 *iter = &lower->list;
6142 int netdev_walk_all_lower_dev(struct net_device *dev,
6143 int (*fn)(struct net_device *dev,
6147 struct net_device *ldev;
6148 struct list_head *iter;
6151 for (iter = &dev->adj_list.lower,
6152 ldev = netdev_next_lower_dev(dev, &iter);
6154 ldev = netdev_next_lower_dev(dev, &iter)) {
6155 /* first is the lower device itself */
6156 ret = fn(ldev, data);
6160 /* then look at all of its lower devices */
6161 ret = netdev_walk_all_lower_dev(ldev, fn, data);
6168 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev);
6170 static struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev,
6171 struct list_head **iter)
6173 struct netdev_adjacent *lower;
6175 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6176 if (&lower->list == &dev->adj_list.lower)
6179 *iter = &lower->list;
6184 int netdev_walk_all_lower_dev_rcu(struct net_device *dev,
6185 int (*fn)(struct net_device *dev,
6189 struct net_device *ldev;
6190 struct list_head *iter;
6193 for (iter = &dev->adj_list.lower,
6194 ldev = netdev_next_lower_dev_rcu(dev, &iter);
6196 ldev = netdev_next_lower_dev_rcu(dev, &iter)) {
6197 /* first is the lower device itself */
6198 ret = fn(ldev, data);
6202 /* then look at all of its lower devices */
6203 ret = netdev_walk_all_lower_dev_rcu(ldev, fn, data);
6210 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu);
6213 * netdev_lower_get_first_private_rcu - Get the first ->private from the
6214 * lower neighbour list, RCU
6218 * Gets the first netdev_adjacent->private from the dev's lower neighbour
6219 * list. The caller must hold RCU read lock.
6221 void *netdev_lower_get_first_private_rcu(struct net_device *dev)
6223 struct netdev_adjacent *lower;
6225 lower = list_first_or_null_rcu(&dev->adj_list.lower,
6226 struct netdev_adjacent, list);
6228 return lower->private;
6231 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu);
6234 * netdev_master_upper_dev_get_rcu - Get master upper device
6237 * Find a master upper device and return pointer to it or NULL in case
6238 * it's not there. The caller must hold the RCU read lock.
6240 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev)
6242 struct netdev_adjacent *upper;
6244 upper = list_first_or_null_rcu(&dev->adj_list.upper,
6245 struct netdev_adjacent, list);
6246 if (upper && likely(upper->master))
6250 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu);
6252 static int netdev_adjacent_sysfs_add(struct net_device *dev,
6253 struct net_device *adj_dev,
6254 struct list_head *dev_list)
6256 char linkname[IFNAMSIZ+7];
6258 sprintf(linkname, dev_list == &dev->adj_list.upper ?
6259 "upper_%s" : "lower_%s", adj_dev->name);
6260 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj),
6263 static void netdev_adjacent_sysfs_del(struct net_device *dev,
6265 struct list_head *dev_list)
6267 char linkname[IFNAMSIZ+7];
6269 sprintf(linkname, dev_list == &dev->adj_list.upper ?
6270 "upper_%s" : "lower_%s", name);
6271 sysfs_remove_link(&(dev->dev.kobj), linkname);
6274 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev,
6275 struct net_device *adj_dev,
6276 struct list_head *dev_list)
6278 return (dev_list == &dev->adj_list.upper ||
6279 dev_list == &dev->adj_list.lower) &&
6280 net_eq(dev_net(dev), dev_net(adj_dev));
6283 static int __netdev_adjacent_dev_insert(struct net_device *dev,
6284 struct net_device *adj_dev,
6285 struct list_head *dev_list,
6286 void *private, bool master)
6288 struct netdev_adjacent *adj;
6291 adj = __netdev_find_adj(adj_dev, dev_list);
6295 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n",
6296 dev->name, adj_dev->name, adj->ref_nr);
6301 adj = kmalloc(sizeof(*adj), GFP_KERNEL);
6306 adj->master = master;
6308 adj->private = private;
6311 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n",
6312 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name);
6314 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) {
6315 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list);
6320 /* Ensure that master link is always the first item in list. */
6322 ret = sysfs_create_link(&(dev->dev.kobj),
6323 &(adj_dev->dev.kobj), "master");
6325 goto remove_symlinks;
6327 list_add_rcu(&adj->list, dev_list);
6329 list_add_tail_rcu(&adj->list, dev_list);
6335 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
6336 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
6344 static void __netdev_adjacent_dev_remove(struct net_device *dev,
6345 struct net_device *adj_dev,
6347 struct list_head *dev_list)
6349 struct netdev_adjacent *adj;
6351 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n",
6352 dev->name, adj_dev->name, ref_nr);
6354 adj = __netdev_find_adj(adj_dev, dev_list);
6357 pr_err("Adjacency does not exist for device %s from %s\n",
6358 dev->name, adj_dev->name);
6363 if (adj->ref_nr > ref_nr) {
6364 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n",
6365 dev->name, adj_dev->name, ref_nr,
6366 adj->ref_nr - ref_nr);
6367 adj->ref_nr -= ref_nr;
6372 sysfs_remove_link(&(dev->dev.kobj), "master");
6374 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
6375 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
6377 list_del_rcu(&adj->list);
6378 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n",
6379 adj_dev->name, dev->name, adj_dev->name);
6381 kfree_rcu(adj, rcu);
6384 static int __netdev_adjacent_dev_link_lists(struct net_device *dev,
6385 struct net_device *upper_dev,
6386 struct list_head *up_list,
6387 struct list_head *down_list,
6388 void *private, bool master)
6392 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list,
6397 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list,
6400 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list);
6407 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev,
6408 struct net_device *upper_dev,
6410 struct list_head *up_list,
6411 struct list_head *down_list)
6413 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list);
6414 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list);
6417 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev,
6418 struct net_device *upper_dev,
6419 void *private, bool master)
6421 return __netdev_adjacent_dev_link_lists(dev, upper_dev,
6422 &dev->adj_list.upper,
6423 &upper_dev->adj_list.lower,
6427 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev,
6428 struct net_device *upper_dev)
6430 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1,
6431 &dev->adj_list.upper,
6432 &upper_dev->adj_list.lower);
6435 static int __netdev_upper_dev_link(struct net_device *dev,
6436 struct net_device *upper_dev, bool master,
6437 void *upper_priv, void *upper_info,
6438 struct netlink_ext_ack *extack)
6440 struct netdev_notifier_changeupper_info changeupper_info = {
6445 .upper_dev = upper_dev,
6448 .upper_info = upper_info,
6450 struct net_device *master_dev;
6455 if (dev == upper_dev)
6458 /* To prevent loops, check if dev is not upper device to upper_dev. */
6459 if (netdev_has_upper_dev(upper_dev, dev))
6463 if (netdev_has_upper_dev(dev, upper_dev))
6466 master_dev = netdev_master_upper_dev_get(dev);
6468 return master_dev == upper_dev ? -EEXIST : -EBUSY;
6471 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
6472 &changeupper_info.info);
6473 ret = notifier_to_errno(ret);
6477 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv,
6482 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
6483 &changeupper_info.info);
6484 ret = notifier_to_errno(ret);
6491 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
6497 * netdev_upper_dev_link - Add a link to the upper device
6499 * @upper_dev: new upper device
6500 * @extack: netlink extended ack
6502 * Adds a link to device which is upper to this one. The caller must hold
6503 * the RTNL lock. On a failure a negative errno code is returned.
6504 * On success the reference counts are adjusted and the function
6507 int netdev_upper_dev_link(struct net_device *dev,
6508 struct net_device *upper_dev,
6509 struct netlink_ext_ack *extack)
6511 return __netdev_upper_dev_link(dev, upper_dev, false,
6512 NULL, NULL, extack);
6514 EXPORT_SYMBOL(netdev_upper_dev_link);
6517 * netdev_master_upper_dev_link - Add a master link to the upper device
6519 * @upper_dev: new upper device
6520 * @upper_priv: upper device private
6521 * @upper_info: upper info to be passed down via notifier
6522 * @extack: netlink extended ack
6524 * Adds a link to device which is upper to this one. In this case, only
6525 * one master upper device can be linked, although other non-master devices
6526 * might be linked as well. The caller must hold the RTNL lock.
6527 * On a failure a negative errno code is returned. On success the reference
6528 * counts are adjusted and the function returns zero.
6530 int netdev_master_upper_dev_link(struct net_device *dev,
6531 struct net_device *upper_dev,
6532 void *upper_priv, void *upper_info,
6533 struct netlink_ext_ack *extack)
6535 return __netdev_upper_dev_link(dev, upper_dev, true,
6536 upper_priv, upper_info, extack);
6538 EXPORT_SYMBOL(netdev_master_upper_dev_link);
6541 * netdev_upper_dev_unlink - Removes a link to upper device
6543 * @upper_dev: new upper device
6545 * Removes a link to device which is upper to this one. The caller must hold
6548 void netdev_upper_dev_unlink(struct net_device *dev,
6549 struct net_device *upper_dev)
6551 struct netdev_notifier_changeupper_info changeupper_info = {
6555 .upper_dev = upper_dev,
6561 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev;
6563 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
6564 &changeupper_info.info);
6566 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
6568 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
6569 &changeupper_info.info);
6571 EXPORT_SYMBOL(netdev_upper_dev_unlink);
6574 * netdev_bonding_info_change - Dispatch event about slave change
6576 * @bonding_info: info to dispatch
6578 * Send NETDEV_BONDING_INFO to netdev notifiers with info.
6579 * The caller must hold the RTNL lock.
6581 void netdev_bonding_info_change(struct net_device *dev,
6582 struct netdev_bonding_info *bonding_info)
6584 struct netdev_notifier_bonding_info info = {
6588 memcpy(&info.bonding_info, bonding_info,
6589 sizeof(struct netdev_bonding_info));
6590 call_netdevice_notifiers_info(NETDEV_BONDING_INFO,
6593 EXPORT_SYMBOL(netdev_bonding_info_change);
6595 static void netdev_adjacent_add_links(struct net_device *dev)
6597 struct netdev_adjacent *iter;
6599 struct net *net = dev_net(dev);
6601 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6602 if (!net_eq(net, dev_net(iter->dev)))
6604 netdev_adjacent_sysfs_add(iter->dev, dev,
6605 &iter->dev->adj_list.lower);
6606 netdev_adjacent_sysfs_add(dev, iter->dev,
6607 &dev->adj_list.upper);
6610 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6611 if (!net_eq(net, dev_net(iter->dev)))
6613 netdev_adjacent_sysfs_add(iter->dev, dev,
6614 &iter->dev->adj_list.upper);
6615 netdev_adjacent_sysfs_add(dev, iter->dev,
6616 &dev->adj_list.lower);
6620 static void netdev_adjacent_del_links(struct net_device *dev)
6622 struct netdev_adjacent *iter;
6624 struct net *net = dev_net(dev);
6626 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6627 if (!net_eq(net, dev_net(iter->dev)))
6629 netdev_adjacent_sysfs_del(iter->dev, dev->name,
6630 &iter->dev->adj_list.lower);
6631 netdev_adjacent_sysfs_del(dev, iter->dev->name,
6632 &dev->adj_list.upper);
6635 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6636 if (!net_eq(net, dev_net(iter->dev)))
6638 netdev_adjacent_sysfs_del(iter->dev, dev->name,
6639 &iter->dev->adj_list.upper);
6640 netdev_adjacent_sysfs_del(dev, iter->dev->name,
6641 &dev->adj_list.lower);
6645 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname)
6647 struct netdev_adjacent *iter;
6649 struct net *net = dev_net(dev);
6651 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6652 if (!net_eq(net, dev_net(iter->dev)))
6654 netdev_adjacent_sysfs_del(iter->dev, oldname,
6655 &iter->dev->adj_list.lower);
6656 netdev_adjacent_sysfs_add(iter->dev, dev,
6657 &iter->dev->adj_list.lower);
6660 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6661 if (!net_eq(net, dev_net(iter->dev)))
6663 netdev_adjacent_sysfs_del(iter->dev, oldname,
6664 &iter->dev->adj_list.upper);
6665 netdev_adjacent_sysfs_add(iter->dev, dev,
6666 &iter->dev->adj_list.upper);
6670 void *netdev_lower_dev_get_private(struct net_device *dev,
6671 struct net_device *lower_dev)
6673 struct netdev_adjacent *lower;
6677 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower);
6681 return lower->private;
6683 EXPORT_SYMBOL(netdev_lower_dev_get_private);
6686 int dev_get_nest_level(struct net_device *dev)
6688 struct net_device *lower = NULL;
6689 struct list_head *iter;
6695 netdev_for_each_lower_dev(dev, lower, iter) {
6696 nest = dev_get_nest_level(lower);
6697 if (max_nest < nest)
6701 return max_nest + 1;
6703 EXPORT_SYMBOL(dev_get_nest_level);
6706 * netdev_lower_change - Dispatch event about lower device state change
6707 * @lower_dev: device
6708 * @lower_state_info: state to dispatch
6710 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info.
6711 * The caller must hold the RTNL lock.
6713 void netdev_lower_state_changed(struct net_device *lower_dev,
6714 void *lower_state_info)
6716 struct netdev_notifier_changelowerstate_info changelowerstate_info = {
6717 .info.dev = lower_dev,
6721 changelowerstate_info.lower_state_info = lower_state_info;
6722 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE,
6723 &changelowerstate_info.info);
6725 EXPORT_SYMBOL(netdev_lower_state_changed);
6727 static void dev_change_rx_flags(struct net_device *dev, int flags)
6729 const struct net_device_ops *ops = dev->netdev_ops;
6731 if (ops->ndo_change_rx_flags)
6732 ops->ndo_change_rx_flags(dev, flags);
6735 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify)
6737 unsigned int old_flags = dev->flags;
6743 dev->flags |= IFF_PROMISC;
6744 dev->promiscuity += inc;
6745 if (dev->promiscuity == 0) {
6748 * If inc causes overflow, untouch promisc and return error.
6751 dev->flags &= ~IFF_PROMISC;
6753 dev->promiscuity -= inc;
6754 pr_warn("%s: promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n",
6759 if (dev->flags != old_flags) {
6760 pr_info("device %s %s promiscuous mode\n",
6762 dev->flags & IFF_PROMISC ? "entered" : "left");
6763 if (audit_enabled) {
6764 current_uid_gid(&uid, &gid);
6765 audit_log(current->audit_context, GFP_ATOMIC,
6766 AUDIT_ANOM_PROMISCUOUS,
6767 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u",
6768 dev->name, (dev->flags & IFF_PROMISC),
6769 (old_flags & IFF_PROMISC),
6770 from_kuid(&init_user_ns, audit_get_loginuid(current)),
6771 from_kuid(&init_user_ns, uid),
6772 from_kgid(&init_user_ns, gid),
6773 audit_get_sessionid(current));
6776 dev_change_rx_flags(dev, IFF_PROMISC);
6779 __dev_notify_flags(dev, old_flags, IFF_PROMISC);
6784 * dev_set_promiscuity - update promiscuity count on a device
6788 * Add or remove promiscuity from a device. While the count in the device
6789 * remains above zero the interface remains promiscuous. Once it hits zero
6790 * the device reverts back to normal filtering operation. A negative inc
6791 * value is used to drop promiscuity on the device.
6792 * Return 0 if successful or a negative errno code on error.
6794 int dev_set_promiscuity(struct net_device *dev, int inc)
6796 unsigned int old_flags = dev->flags;
6799 err = __dev_set_promiscuity(dev, inc, true);
6802 if (dev->flags != old_flags)
6803 dev_set_rx_mode(dev);
6806 EXPORT_SYMBOL(dev_set_promiscuity);
6808 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify)
6810 unsigned int old_flags = dev->flags, old_gflags = dev->gflags;
6814 dev->flags |= IFF_ALLMULTI;
6815 dev->allmulti += inc;
6816 if (dev->allmulti == 0) {
6819 * If inc causes overflow, untouch allmulti and return error.
6822 dev->flags &= ~IFF_ALLMULTI;
6824 dev->allmulti -= inc;
6825 pr_warn("%s: allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n",
6830 if (dev->flags ^ old_flags) {
6831 dev_change_rx_flags(dev, IFF_ALLMULTI);
6832 dev_set_rx_mode(dev);
6834 __dev_notify_flags(dev, old_flags,
6835 dev->gflags ^ old_gflags);
6841 * dev_set_allmulti - update allmulti count on a device
6845 * Add or remove reception of all multicast frames to a device. While the
6846 * count in the device remains above zero the interface remains listening
6847 * to all interfaces. Once it hits zero the device reverts back to normal
6848 * filtering operation. A negative @inc value is used to drop the counter
6849 * when releasing a resource needing all multicasts.
6850 * Return 0 if successful or a negative errno code on error.
6853 int dev_set_allmulti(struct net_device *dev, int inc)
6855 return __dev_set_allmulti(dev, inc, true);
6857 EXPORT_SYMBOL(dev_set_allmulti);
6860 * Upload unicast and multicast address lists to device and
6861 * configure RX filtering. When the device doesn't support unicast
6862 * filtering it is put in promiscuous mode while unicast addresses
6865 void __dev_set_rx_mode(struct net_device *dev)
6867 const struct net_device_ops *ops = dev->netdev_ops;
6869 /* dev_open will call this function so the list will stay sane. */
6870 if (!(dev->flags&IFF_UP))
6873 if (!netif_device_present(dev))
6876 if (!(dev->priv_flags & IFF_UNICAST_FLT)) {
6877 /* Unicast addresses changes may only happen under the rtnl,
6878 * therefore calling __dev_set_promiscuity here is safe.
6880 if (!netdev_uc_empty(dev) && !dev->uc_promisc) {
6881 __dev_set_promiscuity(dev, 1, false);
6882 dev->uc_promisc = true;
6883 } else if (netdev_uc_empty(dev) && dev->uc_promisc) {
6884 __dev_set_promiscuity(dev, -1, false);
6885 dev->uc_promisc = false;
6889 if (ops->ndo_set_rx_mode)
6890 ops->ndo_set_rx_mode(dev);
6893 void dev_set_rx_mode(struct net_device *dev)
6895 netif_addr_lock_bh(dev);
6896 __dev_set_rx_mode(dev);
6897 netif_addr_unlock_bh(dev);
6901 * dev_get_flags - get flags reported to userspace
6904 * Get the combination of flag bits exported through APIs to userspace.
6906 unsigned int dev_get_flags(const struct net_device *dev)
6910 flags = (dev->flags & ~(IFF_PROMISC |
6915 (dev->gflags & (IFF_PROMISC |
6918 if (netif_running(dev)) {
6919 if (netif_oper_up(dev))
6920 flags |= IFF_RUNNING;
6921 if (netif_carrier_ok(dev))
6922 flags |= IFF_LOWER_UP;
6923 if (netif_dormant(dev))
6924 flags |= IFF_DORMANT;
6929 EXPORT_SYMBOL(dev_get_flags);
6931 int __dev_change_flags(struct net_device *dev, unsigned int flags)
6933 unsigned int old_flags = dev->flags;
6939 * Set the flags on our device.
6942 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP |
6943 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL |
6945 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC |
6949 * Load in the correct multicast list now the flags have changed.
6952 if ((old_flags ^ flags) & IFF_MULTICAST)
6953 dev_change_rx_flags(dev, IFF_MULTICAST);
6955 dev_set_rx_mode(dev);
6958 * Have we downed the interface. We handle IFF_UP ourselves
6959 * according to user attempts to set it, rather than blindly
6964 if ((old_flags ^ flags) & IFF_UP) {
6965 if (old_flags & IFF_UP)
6968 ret = __dev_open(dev);
6971 if ((flags ^ dev->gflags) & IFF_PROMISC) {
6972 int inc = (flags & IFF_PROMISC) ? 1 : -1;
6973 unsigned int old_flags = dev->flags;
6975 dev->gflags ^= IFF_PROMISC;
6977 if (__dev_set_promiscuity(dev, inc, false) >= 0)
6978 if (dev->flags != old_flags)
6979 dev_set_rx_mode(dev);
6982 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI
6983 * is important. Some (broken) drivers set IFF_PROMISC, when
6984 * IFF_ALLMULTI is requested not asking us and not reporting.
6986 if ((flags ^ dev->gflags) & IFF_ALLMULTI) {
6987 int inc = (flags & IFF_ALLMULTI) ? 1 : -1;
6989 dev->gflags ^= IFF_ALLMULTI;
6990 __dev_set_allmulti(dev, inc, false);
6996 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags,
6997 unsigned int gchanges)
6999 unsigned int changes = dev->flags ^ old_flags;
7002 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC);
7004 if (changes & IFF_UP) {
7005 if (dev->flags & IFF_UP)
7006 call_netdevice_notifiers(NETDEV_UP, dev);
7008 call_netdevice_notifiers(NETDEV_DOWN, dev);
7011 if (dev->flags & IFF_UP &&
7012 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) {
7013 struct netdev_notifier_change_info change_info = {
7017 .flags_changed = changes,
7020 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info);
7025 * dev_change_flags - change device settings
7027 * @flags: device state flags
7029 * Change settings on device based state flags. The flags are
7030 * in the userspace exported format.
7032 int dev_change_flags(struct net_device *dev, unsigned int flags)
7035 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags;
7037 ret = __dev_change_flags(dev, flags);
7041 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags);
7042 __dev_notify_flags(dev, old_flags, changes);
7045 EXPORT_SYMBOL(dev_change_flags);
7047 int __dev_set_mtu(struct net_device *dev, int new_mtu)
7049 const struct net_device_ops *ops = dev->netdev_ops;
7051 if (ops->ndo_change_mtu)
7052 return ops->ndo_change_mtu(dev, new_mtu);
7057 EXPORT_SYMBOL(__dev_set_mtu);
7060 * dev_set_mtu - Change maximum transfer unit
7062 * @new_mtu: new transfer unit
7064 * Change the maximum transfer size of the network device.
7066 int dev_set_mtu(struct net_device *dev, int new_mtu)
7070 if (new_mtu == dev->mtu)
7073 /* MTU must be positive, and in range */
7074 if (new_mtu < 0 || new_mtu < dev->min_mtu) {
7075 net_err_ratelimited("%s: Invalid MTU %d requested, hw min %d\n",
7076 dev->name, new_mtu, dev->min_mtu);
7080 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) {
7081 net_err_ratelimited("%s: Invalid MTU %d requested, hw max %d\n",
7082 dev->name, new_mtu, dev->max_mtu);
7086 if (!netif_device_present(dev))
7089 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev);
7090 err = notifier_to_errno(err);
7094 orig_mtu = dev->mtu;
7095 err = __dev_set_mtu(dev, new_mtu);
7098 err = call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
7099 err = notifier_to_errno(err);
7101 /* setting mtu back and notifying everyone again,
7102 * so that they have a chance to revert changes.
7104 __dev_set_mtu(dev, orig_mtu);
7105 call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
7110 EXPORT_SYMBOL(dev_set_mtu);
7113 * dev_change_tx_queue_len - Change TX queue length of a netdevice
7115 * @new_len: new tx queue length
7117 int dev_change_tx_queue_len(struct net_device *dev, unsigned long new_len)
7119 unsigned int orig_len = dev->tx_queue_len;
7122 if (new_len != (unsigned int)new_len)
7125 if (new_len != orig_len) {
7126 dev->tx_queue_len = new_len;
7127 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev);
7128 res = notifier_to_errno(res);
7131 "refused to change device tx_queue_len\n");
7132 dev->tx_queue_len = orig_len;
7135 return dev_qdisc_change_tx_queue_len(dev);
7142 * dev_set_group - Change group this device belongs to
7144 * @new_group: group this device should belong to
7146 void dev_set_group(struct net_device *dev, int new_group)
7148 dev->group = new_group;
7150 EXPORT_SYMBOL(dev_set_group);
7153 * dev_set_mac_address - Change Media Access Control Address
7157 * Change the hardware (MAC) address of the device
7159 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa)
7161 const struct net_device_ops *ops = dev->netdev_ops;
7164 if (!ops->ndo_set_mac_address)
7166 if (sa->sa_family != dev->type)
7168 if (!netif_device_present(dev))
7170 err = ops->ndo_set_mac_address(dev, sa);
7173 dev->addr_assign_type = NET_ADDR_SET;
7174 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
7175 add_device_randomness(dev->dev_addr, dev->addr_len);
7178 EXPORT_SYMBOL(dev_set_mac_address);
7181 * dev_change_carrier - Change device carrier
7183 * @new_carrier: new value
7185 * Change device carrier
7187 int dev_change_carrier(struct net_device *dev, bool new_carrier)
7189 const struct net_device_ops *ops = dev->netdev_ops;
7191 if (!ops->ndo_change_carrier)
7193 if (!netif_device_present(dev))
7195 return ops->ndo_change_carrier(dev, new_carrier);
7197 EXPORT_SYMBOL(dev_change_carrier);
7200 * dev_get_phys_port_id - Get device physical port ID
7204 * Get device physical port ID
7206 int dev_get_phys_port_id(struct net_device *dev,
7207 struct netdev_phys_item_id *ppid)
7209 const struct net_device_ops *ops = dev->netdev_ops;
7211 if (!ops->ndo_get_phys_port_id)
7213 return ops->ndo_get_phys_port_id(dev, ppid);
7215 EXPORT_SYMBOL(dev_get_phys_port_id);
7218 * dev_get_phys_port_name - Get device physical port name
7221 * @len: limit of bytes to copy to name
7223 * Get device physical port name
7225 int dev_get_phys_port_name(struct net_device *dev,
7226 char *name, size_t len)
7228 const struct net_device_ops *ops = dev->netdev_ops;
7230 if (!ops->ndo_get_phys_port_name)
7232 return ops->ndo_get_phys_port_name(dev, name, len);
7234 EXPORT_SYMBOL(dev_get_phys_port_name);
7237 * dev_change_proto_down - update protocol port state information
7239 * @proto_down: new value
7241 * This info can be used by switch drivers to set the phys state of the
7244 int dev_change_proto_down(struct net_device *dev, bool proto_down)
7246 const struct net_device_ops *ops = dev->netdev_ops;
7248 if (!ops->ndo_change_proto_down)
7250 if (!netif_device_present(dev))
7252 return ops->ndo_change_proto_down(dev, proto_down);
7254 EXPORT_SYMBOL(dev_change_proto_down);
7256 void __dev_xdp_query(struct net_device *dev, bpf_op_t bpf_op,
7257 struct netdev_bpf *xdp)
7259 memset(xdp, 0, sizeof(*xdp));
7260 xdp->command = XDP_QUERY_PROG;
7262 /* Query must always succeed. */
7263 WARN_ON(bpf_op(dev, xdp) < 0);
7266 static u8 __dev_xdp_attached(struct net_device *dev, bpf_op_t bpf_op)
7268 struct netdev_bpf xdp;
7270 __dev_xdp_query(dev, bpf_op, &xdp);
7272 return xdp.prog_attached;
7275 static int dev_xdp_install(struct net_device *dev, bpf_op_t bpf_op,
7276 struct netlink_ext_ack *extack, u32 flags,
7277 struct bpf_prog *prog)
7279 struct netdev_bpf xdp;
7281 memset(&xdp, 0, sizeof(xdp));
7282 if (flags & XDP_FLAGS_HW_MODE)
7283 xdp.command = XDP_SETUP_PROG_HW;
7285 xdp.command = XDP_SETUP_PROG;
7286 xdp.extack = extack;
7290 return bpf_op(dev, &xdp);
7293 static void dev_xdp_uninstall(struct net_device *dev)
7295 struct netdev_bpf xdp;
7298 /* Remove generic XDP */
7299 WARN_ON(dev_xdp_install(dev, generic_xdp_install, NULL, 0, NULL));
7301 /* Remove from the driver */
7302 ndo_bpf = dev->netdev_ops->ndo_bpf;
7306 __dev_xdp_query(dev, ndo_bpf, &xdp);
7307 if (xdp.prog_attached == XDP_ATTACHED_NONE)
7310 /* Program removal should always succeed */
7311 WARN_ON(dev_xdp_install(dev, ndo_bpf, NULL, xdp.prog_flags, NULL));
7315 * dev_change_xdp_fd - set or clear a bpf program for a device rx path
7317 * @extack: netlink extended ack
7318 * @fd: new program fd or negative value to clear
7319 * @flags: xdp-related flags
7321 * Set or clear a bpf program for a device
7323 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack,
7326 const struct net_device_ops *ops = dev->netdev_ops;
7327 struct bpf_prog *prog = NULL;
7328 bpf_op_t bpf_op, bpf_chk;
7333 bpf_op = bpf_chk = ops->ndo_bpf;
7334 if (!bpf_op && (flags & (XDP_FLAGS_DRV_MODE | XDP_FLAGS_HW_MODE)))
7336 if (!bpf_op || (flags & XDP_FLAGS_SKB_MODE))
7337 bpf_op = generic_xdp_install;
7338 if (bpf_op == bpf_chk)
7339 bpf_chk = generic_xdp_install;
7342 if (bpf_chk && __dev_xdp_attached(dev, bpf_chk))
7344 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) &&
7345 __dev_xdp_attached(dev, bpf_op))
7348 prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP,
7349 bpf_op == ops->ndo_bpf);
7351 return PTR_ERR(prog);
7353 if (!(flags & XDP_FLAGS_HW_MODE) &&
7354 bpf_prog_is_dev_bound(prog->aux)) {
7355 NL_SET_ERR_MSG(extack, "using device-bound program without HW_MODE flag is not supported");
7361 err = dev_xdp_install(dev, bpf_op, extack, flags, prog);
7362 if (err < 0 && prog)
7369 * dev_new_index - allocate an ifindex
7370 * @net: the applicable net namespace
7372 * Returns a suitable unique value for a new device interface
7373 * number. The caller must hold the rtnl semaphore or the
7374 * dev_base_lock to be sure it remains unique.
7376 static int dev_new_index(struct net *net)
7378 int ifindex = net->ifindex;
7383 if (!__dev_get_by_index(net, ifindex))
7384 return net->ifindex = ifindex;
7388 /* Delayed registration/unregisteration */
7389 static LIST_HEAD(net_todo_list);
7390 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq);
7392 static void net_set_todo(struct net_device *dev)
7394 list_add_tail(&dev->todo_list, &net_todo_list);
7395 dev_net(dev)->dev_unreg_count++;
7398 static void rollback_registered_many(struct list_head *head)
7400 struct net_device *dev, *tmp;
7401 LIST_HEAD(close_head);
7403 BUG_ON(dev_boot_phase);
7406 list_for_each_entry_safe(dev, tmp, head, unreg_list) {
7407 /* Some devices call without registering
7408 * for initialization unwind. Remove those
7409 * devices and proceed with the remaining.
7411 if (dev->reg_state == NETREG_UNINITIALIZED) {
7412 pr_debug("unregister_netdevice: device %s/%p never was registered\n",
7416 list_del(&dev->unreg_list);
7419 dev->dismantle = true;
7420 BUG_ON(dev->reg_state != NETREG_REGISTERED);
7423 /* If device is running, close it first. */
7424 list_for_each_entry(dev, head, unreg_list)
7425 list_add_tail(&dev->close_list, &close_head);
7426 dev_close_many(&close_head, true);
7428 list_for_each_entry(dev, head, unreg_list) {
7429 /* And unlink it from device chain. */
7430 unlist_netdevice(dev);
7432 dev->reg_state = NETREG_UNREGISTERING;
7434 flush_all_backlogs();
7438 list_for_each_entry(dev, head, unreg_list) {
7439 struct sk_buff *skb = NULL;
7441 /* Shutdown queueing discipline. */
7444 dev_xdp_uninstall(dev);
7446 /* Notify protocols, that we are about to destroy
7447 * this device. They should clean all the things.
7449 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
7451 if (!dev->rtnl_link_ops ||
7452 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
7453 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0,
7454 GFP_KERNEL, NULL, 0);
7457 * Flush the unicast and multicast chains
7462 if (dev->netdev_ops->ndo_uninit)
7463 dev->netdev_ops->ndo_uninit(dev);
7466 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL);
7468 /* Notifier chain MUST detach us all upper devices. */
7469 WARN_ON(netdev_has_any_upper_dev(dev));
7470 WARN_ON(netdev_has_any_lower_dev(dev));
7472 /* Remove entries from kobject tree */
7473 netdev_unregister_kobject(dev);
7475 /* Remove XPS queueing entries */
7476 netif_reset_xps_queues_gt(dev, 0);
7482 list_for_each_entry(dev, head, unreg_list)
7486 static void rollback_registered(struct net_device *dev)
7490 list_add(&dev->unreg_list, &single);
7491 rollback_registered_many(&single);
7495 static netdev_features_t netdev_sync_upper_features(struct net_device *lower,
7496 struct net_device *upper, netdev_features_t features)
7498 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
7499 netdev_features_t feature;
7502 for_each_netdev_feature(&upper_disables, feature_bit) {
7503 feature = __NETIF_F_BIT(feature_bit);
7504 if (!(upper->wanted_features & feature)
7505 && (features & feature)) {
7506 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n",
7507 &feature, upper->name);
7508 features &= ~feature;
7515 static void netdev_sync_lower_features(struct net_device *upper,
7516 struct net_device *lower, netdev_features_t features)
7518 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
7519 netdev_features_t feature;
7522 for_each_netdev_feature(&upper_disables, feature_bit) {
7523 feature = __NETIF_F_BIT(feature_bit);
7524 if (!(features & feature) && (lower->features & feature)) {
7525 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n",
7526 &feature, lower->name);
7527 lower->wanted_features &= ~feature;
7528 netdev_update_features(lower);
7530 if (unlikely(lower->features & feature))
7531 netdev_WARN(upper, "failed to disable %pNF on %s!\n",
7532 &feature, lower->name);
7537 static netdev_features_t netdev_fix_features(struct net_device *dev,
7538 netdev_features_t features)
7540 /* Fix illegal checksum combinations */
7541 if ((features & NETIF_F_HW_CSUM) &&
7542 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
7543 netdev_warn(dev, "mixed HW and IP checksum settings.\n");
7544 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
7547 /* TSO requires that SG is present as well. */
7548 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) {
7549 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n");
7550 features &= ~NETIF_F_ALL_TSO;
7553 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) &&
7554 !(features & NETIF_F_IP_CSUM)) {
7555 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n");
7556 features &= ~NETIF_F_TSO;
7557 features &= ~NETIF_F_TSO_ECN;
7560 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) &&
7561 !(features & NETIF_F_IPV6_CSUM)) {
7562 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n");
7563 features &= ~NETIF_F_TSO6;
7566 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */
7567 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO))
7568 features &= ~NETIF_F_TSO_MANGLEID;
7570 /* TSO ECN requires that TSO is present as well. */
7571 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN)
7572 features &= ~NETIF_F_TSO_ECN;
7574 /* Software GSO depends on SG. */
7575 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) {
7576 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n");
7577 features &= ~NETIF_F_GSO;
7580 /* GSO partial features require GSO partial be set */
7581 if ((features & dev->gso_partial_features) &&
7582 !(features & NETIF_F_GSO_PARTIAL)) {
7584 "Dropping partially supported GSO features since no GSO partial.\n");
7585 features &= ~dev->gso_partial_features;
7588 if (!(features & NETIF_F_RXCSUM)) {
7589 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet
7590 * successfully merged by hardware must also have the
7591 * checksum verified by hardware. If the user does not
7592 * want to enable RXCSUM, logically, we should disable GRO_HW.
7594 if (features & NETIF_F_GRO_HW) {
7595 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n");
7596 features &= ~NETIF_F_GRO_HW;
7600 /* LRO/HW-GRO features cannot be combined with RX-FCS */
7601 if (features & NETIF_F_RXFCS) {
7602 if (features & NETIF_F_LRO) {
7603 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n");
7604 features &= ~NETIF_F_LRO;
7607 if (features & NETIF_F_GRO_HW) {
7608 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n");
7609 features &= ~NETIF_F_GRO_HW;
7616 int __netdev_update_features(struct net_device *dev)
7618 struct net_device *upper, *lower;
7619 netdev_features_t features;
7620 struct list_head *iter;
7625 features = netdev_get_wanted_features(dev);
7627 if (dev->netdev_ops->ndo_fix_features)
7628 features = dev->netdev_ops->ndo_fix_features(dev, features);
7630 /* driver might be less strict about feature dependencies */
7631 features = netdev_fix_features(dev, features);
7633 /* some features can't be enabled if they're off an an upper device */
7634 netdev_for_each_upper_dev_rcu(dev, upper, iter)
7635 features = netdev_sync_upper_features(dev, upper, features);
7637 if (dev->features == features)
7640 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n",
7641 &dev->features, &features);
7643 if (dev->netdev_ops->ndo_set_features)
7644 err = dev->netdev_ops->ndo_set_features(dev, features);
7648 if (unlikely(err < 0)) {
7650 "set_features() failed (%d); wanted %pNF, left %pNF\n",
7651 err, &features, &dev->features);
7652 /* return non-0 since some features might have changed and
7653 * it's better to fire a spurious notification than miss it
7659 /* some features must be disabled on lower devices when disabled
7660 * on an upper device (think: bonding master or bridge)
7662 netdev_for_each_lower_dev(dev, lower, iter)
7663 netdev_sync_lower_features(dev, lower, features);
7666 netdev_features_t diff = features ^ dev->features;
7668 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) {
7669 /* udp_tunnel_{get,drop}_rx_info both need
7670 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the
7671 * device, or they won't do anything.
7672 * Thus we need to update dev->features
7673 * *before* calling udp_tunnel_get_rx_info,
7674 * but *after* calling udp_tunnel_drop_rx_info.
7676 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) {
7677 dev->features = features;
7678 udp_tunnel_get_rx_info(dev);
7680 udp_tunnel_drop_rx_info(dev);
7684 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) {
7685 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) {
7686 dev->features = features;
7687 err |= vlan_get_rx_ctag_filter_info(dev);
7689 vlan_drop_rx_ctag_filter_info(dev);
7693 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) {
7694 if (features & NETIF_F_HW_VLAN_STAG_FILTER) {
7695 dev->features = features;
7696 err |= vlan_get_rx_stag_filter_info(dev);
7698 vlan_drop_rx_stag_filter_info(dev);
7702 dev->features = features;
7705 return err < 0 ? 0 : 1;
7709 * netdev_update_features - recalculate device features
7710 * @dev: the device to check
7712 * Recalculate dev->features set and send notifications if it
7713 * has changed. Should be called after driver or hardware dependent
7714 * conditions might have changed that influence the features.
7716 void netdev_update_features(struct net_device *dev)
7718 if (__netdev_update_features(dev))
7719 netdev_features_change(dev);
7721 EXPORT_SYMBOL(netdev_update_features);
7724 * netdev_change_features - recalculate device features
7725 * @dev: the device to check
7727 * Recalculate dev->features set and send notifications even
7728 * if they have not changed. Should be called instead of
7729 * netdev_update_features() if also dev->vlan_features might
7730 * have changed to allow the changes to be propagated to stacked
7733 void netdev_change_features(struct net_device *dev)
7735 __netdev_update_features(dev);
7736 netdev_features_change(dev);
7738 EXPORT_SYMBOL(netdev_change_features);
7741 * netif_stacked_transfer_operstate - transfer operstate
7742 * @rootdev: the root or lower level device to transfer state from
7743 * @dev: the device to transfer operstate to
7745 * Transfer operational state from root to device. This is normally
7746 * called when a stacking relationship exists between the root
7747 * device and the device(a leaf device).
7749 void netif_stacked_transfer_operstate(const struct net_device *rootdev,
7750 struct net_device *dev)
7752 if (rootdev->operstate == IF_OPER_DORMANT)
7753 netif_dormant_on(dev);
7755 netif_dormant_off(dev);
7757 if (netif_carrier_ok(rootdev))
7758 netif_carrier_on(dev);
7760 netif_carrier_off(dev);
7762 EXPORT_SYMBOL(netif_stacked_transfer_operstate);
7764 static int netif_alloc_rx_queues(struct net_device *dev)
7766 unsigned int i, count = dev->num_rx_queues;
7767 struct netdev_rx_queue *rx;
7768 size_t sz = count * sizeof(*rx);
7773 rx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
7779 for (i = 0; i < count; i++) {
7782 /* XDP RX-queue setup */
7783 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i);
7790 /* Rollback successful reg's and free other resources */
7792 xdp_rxq_info_unreg(&rx[i].xdp_rxq);
7798 static void netif_free_rx_queues(struct net_device *dev)
7800 unsigned int i, count = dev->num_rx_queues;
7802 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */
7806 for (i = 0; i < count; i++)
7807 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq);
7812 static void netdev_init_one_queue(struct net_device *dev,
7813 struct netdev_queue *queue, void *_unused)
7815 /* Initialize queue lock */
7816 spin_lock_init(&queue->_xmit_lock);
7817 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type);
7818 queue->xmit_lock_owner = -1;
7819 netdev_queue_numa_node_write(queue, NUMA_NO_NODE);
7822 dql_init(&queue->dql, HZ);
7826 static void netif_free_tx_queues(struct net_device *dev)
7831 static int netif_alloc_netdev_queues(struct net_device *dev)
7833 unsigned int count = dev->num_tx_queues;
7834 struct netdev_queue *tx;
7835 size_t sz = count * sizeof(*tx);
7837 if (count < 1 || count > 0xffff)
7840 tx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
7846 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL);
7847 spin_lock_init(&dev->tx_global_lock);
7852 void netif_tx_stop_all_queues(struct net_device *dev)
7856 for (i = 0; i < dev->num_tx_queues; i++) {
7857 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
7859 netif_tx_stop_queue(txq);
7862 EXPORT_SYMBOL(netif_tx_stop_all_queues);
7865 * register_netdevice - register a network device
7866 * @dev: device to register
7868 * Take a completed network device structure and add it to the kernel
7869 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
7870 * chain. 0 is returned on success. A negative errno code is returned
7871 * on a failure to set up the device, or if the name is a duplicate.
7873 * Callers must hold the rtnl semaphore. You may want
7874 * register_netdev() instead of this.
7877 * The locking appears insufficient to guarantee two parallel registers
7878 * will not get the same name.
7881 int register_netdevice(struct net_device *dev)
7884 struct net *net = dev_net(dev);
7886 netdev_features_size_check();
7887 BUG_ON(dev_boot_phase);
7892 /* When net_device's are persistent, this will be fatal. */
7893 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
7896 spin_lock_init(&dev->addr_list_lock);
7897 netdev_set_addr_lockdep_class(dev);
7899 ret = dev_get_valid_name(net, dev, dev->name);
7903 /* Init, if this function is available */
7904 if (dev->netdev_ops->ndo_init) {
7905 ret = dev->netdev_ops->ndo_init(dev);
7913 if (((dev->hw_features | dev->features) &
7914 NETIF_F_HW_VLAN_CTAG_FILTER) &&
7915 (!dev->netdev_ops->ndo_vlan_rx_add_vid ||
7916 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) {
7917 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n");
7924 dev->ifindex = dev_new_index(net);
7925 else if (__dev_get_by_index(net, dev->ifindex))
7928 /* Transfer changeable features to wanted_features and enable
7929 * software offloads (GSO and GRO).
7931 dev->hw_features |= NETIF_F_SOFT_FEATURES;
7932 dev->features |= NETIF_F_SOFT_FEATURES;
7934 if (dev->netdev_ops->ndo_udp_tunnel_add) {
7935 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT;
7936 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT;
7939 dev->wanted_features = dev->features & dev->hw_features;
7941 if (!(dev->flags & IFF_LOOPBACK))
7942 dev->hw_features |= NETIF_F_NOCACHE_COPY;
7944 /* If IPv4 TCP segmentation offload is supported we should also
7945 * allow the device to enable segmenting the frame with the option
7946 * of ignoring a static IP ID value. This doesn't enable the
7947 * feature itself but allows the user to enable it later.
7949 if (dev->hw_features & NETIF_F_TSO)
7950 dev->hw_features |= NETIF_F_TSO_MANGLEID;
7951 if (dev->vlan_features & NETIF_F_TSO)
7952 dev->vlan_features |= NETIF_F_TSO_MANGLEID;
7953 if (dev->mpls_features & NETIF_F_TSO)
7954 dev->mpls_features |= NETIF_F_TSO_MANGLEID;
7955 if (dev->hw_enc_features & NETIF_F_TSO)
7956 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID;
7958 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices.
7960 dev->vlan_features |= NETIF_F_HIGHDMA;
7962 /* Make NETIF_F_SG inheritable to tunnel devices.
7964 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL;
7966 /* Make NETIF_F_SG inheritable to MPLS.
7968 dev->mpls_features |= NETIF_F_SG;
7970 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev);
7971 ret = notifier_to_errno(ret);
7975 ret = netdev_register_kobject(dev);
7978 dev->reg_state = NETREG_REGISTERED;
7980 __netdev_update_features(dev);
7983 * Default initial state at registry is that the
7984 * device is present.
7987 set_bit(__LINK_STATE_PRESENT, &dev->state);
7989 linkwatch_init_dev(dev);
7991 dev_init_scheduler(dev);
7993 list_netdevice(dev);
7994 add_device_randomness(dev->dev_addr, dev->addr_len);
7996 /* If the device has permanent device address, driver should
7997 * set dev_addr and also addr_assign_type should be set to
7998 * NET_ADDR_PERM (default value).
8000 if (dev->addr_assign_type == NET_ADDR_PERM)
8001 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len);
8003 /* Notify protocols, that a new device appeared. */
8004 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
8005 ret = notifier_to_errno(ret);
8007 rollback_registered(dev);
8008 dev->reg_state = NETREG_UNREGISTERED;
8011 * Prevent userspace races by waiting until the network
8012 * device is fully setup before sending notifications.
8014 if (!dev->rtnl_link_ops ||
8015 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
8016 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
8022 if (dev->netdev_ops->ndo_uninit)
8023 dev->netdev_ops->ndo_uninit(dev);
8024 if (dev->priv_destructor)
8025 dev->priv_destructor(dev);
8028 EXPORT_SYMBOL(register_netdevice);
8031 * init_dummy_netdev - init a dummy network device for NAPI
8032 * @dev: device to init
8034 * This takes a network device structure and initialize the minimum
8035 * amount of fields so it can be used to schedule NAPI polls without
8036 * registering a full blown interface. This is to be used by drivers
8037 * that need to tie several hardware interfaces to a single NAPI
8038 * poll scheduler due to HW limitations.
8040 int init_dummy_netdev(struct net_device *dev)
8042 /* Clear everything. Note we don't initialize spinlocks
8043 * are they aren't supposed to be taken by any of the
8044 * NAPI code and this dummy netdev is supposed to be
8045 * only ever used for NAPI polls
8047 memset(dev, 0, sizeof(struct net_device));
8049 /* make sure we BUG if trying to hit standard
8050 * register/unregister code path
8052 dev->reg_state = NETREG_DUMMY;
8054 /* NAPI wants this */
8055 INIT_LIST_HEAD(&dev->napi_list);
8057 /* a dummy interface is started by default */
8058 set_bit(__LINK_STATE_PRESENT, &dev->state);
8059 set_bit(__LINK_STATE_START, &dev->state);
8061 /* Note : We dont allocate pcpu_refcnt for dummy devices,
8062 * because users of this 'device' dont need to change
8068 EXPORT_SYMBOL_GPL(init_dummy_netdev);
8072 * register_netdev - register a network device
8073 * @dev: device to register
8075 * Take a completed network device structure and add it to the kernel
8076 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
8077 * chain. 0 is returned on success. A negative errno code is returned
8078 * on a failure to set up the device, or if the name is a duplicate.
8080 * This is a wrapper around register_netdevice that takes the rtnl semaphore
8081 * and expands the device name if you passed a format string to
8084 int register_netdev(struct net_device *dev)
8088 if (rtnl_lock_killable())
8090 err = register_netdevice(dev);
8094 EXPORT_SYMBOL(register_netdev);
8096 int netdev_refcnt_read(const struct net_device *dev)
8100 for_each_possible_cpu(i)
8101 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
8104 EXPORT_SYMBOL(netdev_refcnt_read);
8107 * netdev_wait_allrefs - wait until all references are gone.
8108 * @dev: target net_device
8110 * This is called when unregistering network devices.
8112 * Any protocol or device that holds a reference should register
8113 * for netdevice notification, and cleanup and put back the
8114 * reference if they receive an UNREGISTER event.
8115 * We can get stuck here if buggy protocols don't correctly
8118 static void netdev_wait_allrefs(struct net_device *dev)
8120 unsigned long rebroadcast_time, warning_time;
8123 linkwatch_forget_dev(dev);
8125 rebroadcast_time = warning_time = jiffies;
8126 refcnt = netdev_refcnt_read(dev);
8128 while (refcnt != 0) {
8129 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
8132 /* Rebroadcast unregister notification */
8133 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
8139 if (test_bit(__LINK_STATE_LINKWATCH_PENDING,
8141 /* We must not have linkwatch events
8142 * pending on unregister. If this
8143 * happens, we simply run the queue
8144 * unscheduled, resulting in a noop
8147 linkwatch_run_queue();
8152 rebroadcast_time = jiffies;
8157 refcnt = netdev_refcnt_read(dev);
8159 if (time_after(jiffies, warning_time + 10 * HZ)) {
8160 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n",
8162 warning_time = jiffies;
8171 * register_netdevice(x1);
8172 * register_netdevice(x2);
8174 * unregister_netdevice(y1);
8175 * unregister_netdevice(y2);
8181 * We are invoked by rtnl_unlock().
8182 * This allows us to deal with problems:
8183 * 1) We can delete sysfs objects which invoke hotplug
8184 * without deadlocking with linkwatch via keventd.
8185 * 2) Since we run with the RTNL semaphore not held, we can sleep
8186 * safely in order to wait for the netdev refcnt to drop to zero.
8188 * We must not return until all unregister events added during
8189 * the interval the lock was held have been completed.
8191 void netdev_run_todo(void)
8193 struct list_head list;
8195 /* Snapshot list, allow later requests */
8196 list_replace_init(&net_todo_list, &list);
8201 /* Wait for rcu callbacks to finish before next phase */
8202 if (!list_empty(&list))
8205 while (!list_empty(&list)) {
8206 struct net_device *dev
8207 = list_first_entry(&list, struct net_device, todo_list);
8208 list_del(&dev->todo_list);
8210 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) {
8211 pr_err("network todo '%s' but state %d\n",
8212 dev->name, dev->reg_state);
8217 dev->reg_state = NETREG_UNREGISTERED;
8219 netdev_wait_allrefs(dev);
8222 BUG_ON(netdev_refcnt_read(dev));
8223 BUG_ON(!list_empty(&dev->ptype_all));
8224 BUG_ON(!list_empty(&dev->ptype_specific));
8225 WARN_ON(rcu_access_pointer(dev->ip_ptr));
8226 WARN_ON(rcu_access_pointer(dev->ip6_ptr));
8227 #if IS_ENABLED(CONFIG_DECNET)
8228 WARN_ON(dev->dn_ptr);
8230 if (dev->priv_destructor)
8231 dev->priv_destructor(dev);
8232 if (dev->needs_free_netdev)
8235 /* Report a network device has been unregistered */
8237 dev_net(dev)->dev_unreg_count--;
8239 wake_up(&netdev_unregistering_wq);
8241 /* Free network device */
8242 kobject_put(&dev->dev.kobj);
8246 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has
8247 * all the same fields in the same order as net_device_stats, with only
8248 * the type differing, but rtnl_link_stats64 may have additional fields
8249 * at the end for newer counters.
8251 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
8252 const struct net_device_stats *netdev_stats)
8254 #if BITS_PER_LONG == 64
8255 BUILD_BUG_ON(sizeof(*stats64) < sizeof(*netdev_stats));
8256 memcpy(stats64, netdev_stats, sizeof(*netdev_stats));
8257 /* zero out counters that only exist in rtnl_link_stats64 */
8258 memset((char *)stats64 + sizeof(*netdev_stats), 0,
8259 sizeof(*stats64) - sizeof(*netdev_stats));
8261 size_t i, n = sizeof(*netdev_stats) / sizeof(unsigned long);
8262 const unsigned long *src = (const unsigned long *)netdev_stats;
8263 u64 *dst = (u64 *)stats64;
8265 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64));
8266 for (i = 0; i < n; i++)
8268 /* zero out counters that only exist in rtnl_link_stats64 */
8269 memset((char *)stats64 + n * sizeof(u64), 0,
8270 sizeof(*stats64) - n * sizeof(u64));
8273 EXPORT_SYMBOL(netdev_stats_to_stats64);
8276 * dev_get_stats - get network device statistics
8277 * @dev: device to get statistics from
8278 * @storage: place to store stats
8280 * Get network statistics from device. Return @storage.
8281 * The device driver may provide its own method by setting
8282 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats;
8283 * otherwise the internal statistics structure is used.
8285 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev,
8286 struct rtnl_link_stats64 *storage)
8288 const struct net_device_ops *ops = dev->netdev_ops;
8290 if (ops->ndo_get_stats64) {
8291 memset(storage, 0, sizeof(*storage));
8292 ops->ndo_get_stats64(dev, storage);
8293 } else if (ops->ndo_get_stats) {
8294 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev));
8296 netdev_stats_to_stats64(storage, &dev->stats);
8298 storage->rx_dropped += (unsigned long)atomic_long_read(&dev->rx_dropped);
8299 storage->tx_dropped += (unsigned long)atomic_long_read(&dev->tx_dropped);
8300 storage->rx_nohandler += (unsigned long)atomic_long_read(&dev->rx_nohandler);
8303 EXPORT_SYMBOL(dev_get_stats);
8305 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev)
8307 struct netdev_queue *queue = dev_ingress_queue(dev);
8309 #ifdef CONFIG_NET_CLS_ACT
8312 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
8315 netdev_init_one_queue(dev, queue, NULL);
8316 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc);
8317 queue->qdisc_sleeping = &noop_qdisc;
8318 rcu_assign_pointer(dev->ingress_queue, queue);
8323 static const struct ethtool_ops default_ethtool_ops;
8325 void netdev_set_default_ethtool_ops(struct net_device *dev,
8326 const struct ethtool_ops *ops)
8328 if (dev->ethtool_ops == &default_ethtool_ops)
8329 dev->ethtool_ops = ops;
8331 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops);
8333 void netdev_freemem(struct net_device *dev)
8335 char *addr = (char *)dev - dev->padded;
8341 * alloc_netdev_mqs - allocate network device
8342 * @sizeof_priv: size of private data to allocate space for
8343 * @name: device name format string
8344 * @name_assign_type: origin of device name
8345 * @setup: callback to initialize device
8346 * @txqs: the number of TX subqueues to allocate
8347 * @rxqs: the number of RX subqueues to allocate
8349 * Allocates a struct net_device with private data area for driver use
8350 * and performs basic initialization. Also allocates subqueue structs
8351 * for each queue on the device.
8353 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
8354 unsigned char name_assign_type,
8355 void (*setup)(struct net_device *),
8356 unsigned int txqs, unsigned int rxqs)
8358 struct net_device *dev;
8359 unsigned int alloc_size;
8360 struct net_device *p;
8362 BUG_ON(strlen(name) >= sizeof(dev->name));
8365 pr_err("alloc_netdev: Unable to allocate device with zero queues\n");
8370 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n");
8374 alloc_size = sizeof(struct net_device);
8376 /* ensure 32-byte alignment of private area */
8377 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
8378 alloc_size += sizeof_priv;
8380 /* ensure 32-byte alignment of whole construct */
8381 alloc_size += NETDEV_ALIGN - 1;
8383 p = kvzalloc(alloc_size, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
8387 dev = PTR_ALIGN(p, NETDEV_ALIGN);
8388 dev->padded = (char *)dev - (char *)p;
8390 dev->pcpu_refcnt = alloc_percpu(int);
8391 if (!dev->pcpu_refcnt)
8394 if (dev_addr_init(dev))
8400 dev_net_set(dev, &init_net);
8402 dev->gso_max_size = GSO_MAX_SIZE;
8403 dev->gso_max_segs = GSO_MAX_SEGS;
8405 INIT_LIST_HEAD(&dev->napi_list);
8406 INIT_LIST_HEAD(&dev->unreg_list);
8407 INIT_LIST_HEAD(&dev->close_list);
8408 INIT_LIST_HEAD(&dev->link_watch_list);
8409 INIT_LIST_HEAD(&dev->adj_list.upper);
8410 INIT_LIST_HEAD(&dev->adj_list.lower);
8411 INIT_LIST_HEAD(&dev->ptype_all);
8412 INIT_LIST_HEAD(&dev->ptype_specific);
8413 #ifdef CONFIG_NET_SCHED
8414 hash_init(dev->qdisc_hash);
8416 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM;
8419 if (!dev->tx_queue_len) {
8420 dev->priv_flags |= IFF_NO_QUEUE;
8421 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN;
8424 dev->num_tx_queues = txqs;
8425 dev->real_num_tx_queues = txqs;
8426 if (netif_alloc_netdev_queues(dev))
8429 dev->num_rx_queues = rxqs;
8430 dev->real_num_rx_queues = rxqs;
8431 if (netif_alloc_rx_queues(dev))
8434 strcpy(dev->name, name);
8435 dev->name_assign_type = name_assign_type;
8436 dev->group = INIT_NETDEV_GROUP;
8437 if (!dev->ethtool_ops)
8438 dev->ethtool_ops = &default_ethtool_ops;
8440 nf_hook_ingress_init(dev);
8449 free_percpu(dev->pcpu_refcnt);
8451 netdev_freemem(dev);
8454 EXPORT_SYMBOL(alloc_netdev_mqs);
8457 * free_netdev - free network device
8460 * This function does the last stage of destroying an allocated device
8461 * interface. The reference to the device object is released. If this
8462 * is the last reference then it will be freed.Must be called in process
8465 void free_netdev(struct net_device *dev)
8467 struct napi_struct *p, *n;
8470 netif_free_tx_queues(dev);
8471 netif_free_rx_queues(dev);
8473 kfree(rcu_dereference_protected(dev->ingress_queue, 1));
8475 /* Flush device addresses */
8476 dev_addr_flush(dev);
8478 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
8481 free_percpu(dev->pcpu_refcnt);
8482 dev->pcpu_refcnt = NULL;
8484 /* Compatibility with error handling in drivers */
8485 if (dev->reg_state == NETREG_UNINITIALIZED) {
8486 netdev_freemem(dev);
8490 BUG_ON(dev->reg_state != NETREG_UNREGISTERED);
8491 dev->reg_state = NETREG_RELEASED;
8493 /* will free via device release */
8494 put_device(&dev->dev);
8496 EXPORT_SYMBOL(free_netdev);
8499 * synchronize_net - Synchronize with packet receive processing
8501 * Wait for packets currently being received to be done.
8502 * Does not block later packets from starting.
8504 void synchronize_net(void)
8507 if (rtnl_is_locked())
8508 synchronize_rcu_expedited();
8512 EXPORT_SYMBOL(synchronize_net);
8515 * unregister_netdevice_queue - remove device from the kernel
8519 * This function shuts down a device interface and removes it
8520 * from the kernel tables.
8521 * If head not NULL, device is queued to be unregistered later.
8523 * Callers must hold the rtnl semaphore. You may want
8524 * unregister_netdev() instead of this.
8527 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head)
8532 list_move_tail(&dev->unreg_list, head);
8534 rollback_registered(dev);
8535 /* Finish processing unregister after unlock */
8539 EXPORT_SYMBOL(unregister_netdevice_queue);
8542 * unregister_netdevice_many - unregister many devices
8543 * @head: list of devices
8545 * Note: As most callers use a stack allocated list_head,
8546 * we force a list_del() to make sure stack wont be corrupted later.
8548 void unregister_netdevice_many(struct list_head *head)
8550 struct net_device *dev;
8552 if (!list_empty(head)) {
8553 rollback_registered_many(head);
8554 list_for_each_entry(dev, head, unreg_list)
8559 EXPORT_SYMBOL(unregister_netdevice_many);
8562 * unregister_netdev - remove device from the kernel
8565 * This function shuts down a device interface and removes it
8566 * from the kernel tables.
8568 * This is just a wrapper for unregister_netdevice that takes
8569 * the rtnl semaphore. In general you want to use this and not
8570 * unregister_netdevice.
8572 void unregister_netdev(struct net_device *dev)
8575 unregister_netdevice(dev);
8578 EXPORT_SYMBOL(unregister_netdev);
8581 * dev_change_net_namespace - move device to different nethost namespace
8583 * @net: network namespace
8584 * @pat: If not NULL name pattern to try if the current device name
8585 * is already taken in the destination network namespace.
8587 * This function shuts down a device interface and moves it
8588 * to a new network namespace. On success 0 is returned, on
8589 * a failure a netagive errno code is returned.
8591 * Callers must hold the rtnl semaphore.
8594 int dev_change_net_namespace(struct net_device *dev, struct net *net, const char *pat)
8596 int err, new_nsid, new_ifindex;
8600 /* Don't allow namespace local devices to be moved. */
8602 if (dev->features & NETIF_F_NETNS_LOCAL)
8605 /* Ensure the device has been registrered */
8606 if (dev->reg_state != NETREG_REGISTERED)
8609 /* Get out if there is nothing todo */
8611 if (net_eq(dev_net(dev), net))
8614 /* Pick the destination device name, and ensure
8615 * we can use it in the destination network namespace.
8618 if (__dev_get_by_name(net, dev->name)) {
8619 /* We get here if we can't use the current device name */
8622 if (dev_get_valid_name(net, dev, pat) < 0)
8627 * And now a mini version of register_netdevice unregister_netdevice.
8630 /* If device is running close it first. */
8633 /* And unlink it from device chain */
8635 unlist_netdevice(dev);
8639 /* Shutdown queueing discipline. */
8642 /* Notify protocols, that we are about to destroy
8643 * this device. They should clean all the things.
8645 * Note that dev->reg_state stays at NETREG_REGISTERED.
8646 * This is wanted because this way 8021q and macvlan know
8647 * the device is just moving and can keep their slaves up.
8649 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
8652 new_nsid = peernet2id_alloc(dev_net(dev), net);
8653 /* If there is an ifindex conflict assign a new one */
8654 if (__dev_get_by_index(net, dev->ifindex))
8655 new_ifindex = dev_new_index(net);
8657 new_ifindex = dev->ifindex;
8659 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid,
8663 * Flush the unicast and multicast chains
8668 /* Send a netdev-removed uevent to the old namespace */
8669 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE);
8670 netdev_adjacent_del_links(dev);
8672 /* Actually switch the network namespace */
8673 dev_net_set(dev, net);
8674 dev->ifindex = new_ifindex;
8676 /* Send a netdev-add uevent to the new namespace */
8677 kobject_uevent(&dev->dev.kobj, KOBJ_ADD);
8678 netdev_adjacent_add_links(dev);
8680 /* Fixup kobjects */
8681 err = device_rename(&dev->dev, dev->name);
8684 /* Add the device back in the hashes */
8685 list_netdevice(dev);
8687 /* Notify protocols, that a new device appeared. */
8688 call_netdevice_notifiers(NETDEV_REGISTER, dev);
8691 * Prevent userspace races by waiting until the network
8692 * device is fully setup before sending notifications.
8694 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
8701 EXPORT_SYMBOL_GPL(dev_change_net_namespace);
8703 static int dev_cpu_dead(unsigned int oldcpu)
8705 struct sk_buff **list_skb;
8706 struct sk_buff *skb;
8708 struct softnet_data *sd, *oldsd, *remsd = NULL;
8710 local_irq_disable();
8711 cpu = smp_processor_id();
8712 sd = &per_cpu(softnet_data, cpu);
8713 oldsd = &per_cpu(softnet_data, oldcpu);
8715 /* Find end of our completion_queue. */
8716 list_skb = &sd->completion_queue;
8718 list_skb = &(*list_skb)->next;
8719 /* Append completion queue from offline CPU. */
8720 *list_skb = oldsd->completion_queue;
8721 oldsd->completion_queue = NULL;
8723 /* Append output queue from offline CPU. */
8724 if (oldsd->output_queue) {
8725 *sd->output_queue_tailp = oldsd->output_queue;
8726 sd->output_queue_tailp = oldsd->output_queue_tailp;
8727 oldsd->output_queue = NULL;
8728 oldsd->output_queue_tailp = &oldsd->output_queue;
8730 /* Append NAPI poll list from offline CPU, with one exception :
8731 * process_backlog() must be called by cpu owning percpu backlog.
8732 * We properly handle process_queue & input_pkt_queue later.
8734 while (!list_empty(&oldsd->poll_list)) {
8735 struct napi_struct *napi = list_first_entry(&oldsd->poll_list,
8739 list_del_init(&napi->poll_list);
8740 if (napi->poll == process_backlog)
8743 ____napi_schedule(sd, napi);
8746 raise_softirq_irqoff(NET_TX_SOFTIRQ);
8750 remsd = oldsd->rps_ipi_list;
8751 oldsd->rps_ipi_list = NULL;
8753 /* send out pending IPI's on offline CPU */
8754 net_rps_send_ipi(remsd);
8756 /* Process offline CPU's input_pkt_queue */
8757 while ((skb = __skb_dequeue(&oldsd->process_queue))) {
8759 input_queue_head_incr(oldsd);
8761 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) {
8763 input_queue_head_incr(oldsd);
8770 * netdev_increment_features - increment feature set by one
8771 * @all: current feature set
8772 * @one: new feature set
8773 * @mask: mask feature set
8775 * Computes a new feature set after adding a device with feature set
8776 * @one to the master device with current feature set @all. Will not
8777 * enable anything that is off in @mask. Returns the new feature set.
8779 netdev_features_t netdev_increment_features(netdev_features_t all,
8780 netdev_features_t one, netdev_features_t mask)
8782 if (mask & NETIF_F_HW_CSUM)
8783 mask |= NETIF_F_CSUM_MASK;
8784 mask |= NETIF_F_VLAN_CHALLENGED;
8786 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask;
8787 all &= one | ~NETIF_F_ALL_FOR_ALL;
8789 /* If one device supports hw checksumming, set for all. */
8790 if (all & NETIF_F_HW_CSUM)
8791 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM);
8795 EXPORT_SYMBOL(netdev_increment_features);
8797 static struct hlist_head * __net_init netdev_create_hash(void)
8800 struct hlist_head *hash;
8802 hash = kmalloc(sizeof(*hash) * NETDEV_HASHENTRIES, GFP_KERNEL);
8804 for (i = 0; i < NETDEV_HASHENTRIES; i++)
8805 INIT_HLIST_HEAD(&hash[i]);
8810 /* Initialize per network namespace state */
8811 static int __net_init netdev_init(struct net *net)
8813 if (net != &init_net)
8814 INIT_LIST_HEAD(&net->dev_base_head);
8816 net->dev_name_head = netdev_create_hash();
8817 if (net->dev_name_head == NULL)
8820 net->dev_index_head = netdev_create_hash();
8821 if (net->dev_index_head == NULL)
8827 kfree(net->dev_name_head);
8833 * netdev_drivername - network driver for the device
8834 * @dev: network device
8836 * Determine network driver for device.
8838 const char *netdev_drivername(const struct net_device *dev)
8840 const struct device_driver *driver;
8841 const struct device *parent;
8842 const char *empty = "";
8844 parent = dev->dev.parent;
8848 driver = parent->driver;
8849 if (driver && driver->name)
8850 return driver->name;
8854 static void __netdev_printk(const char *level, const struct net_device *dev,
8855 struct va_format *vaf)
8857 if (dev && dev->dev.parent) {
8858 dev_printk_emit(level[1] - '0',
8861 dev_driver_string(dev->dev.parent),
8862 dev_name(dev->dev.parent),
8863 netdev_name(dev), netdev_reg_state(dev),
8866 printk("%s%s%s: %pV",
8867 level, netdev_name(dev), netdev_reg_state(dev), vaf);
8869 printk("%s(NULL net_device): %pV", level, vaf);
8873 void netdev_printk(const char *level, const struct net_device *dev,
8874 const char *format, ...)
8876 struct va_format vaf;
8879 va_start(args, format);
8884 __netdev_printk(level, dev, &vaf);
8888 EXPORT_SYMBOL(netdev_printk);
8890 #define define_netdev_printk_level(func, level) \
8891 void func(const struct net_device *dev, const char *fmt, ...) \
8893 struct va_format vaf; \
8896 va_start(args, fmt); \
8901 __netdev_printk(level, dev, &vaf); \
8905 EXPORT_SYMBOL(func);
8907 define_netdev_printk_level(netdev_emerg, KERN_EMERG);
8908 define_netdev_printk_level(netdev_alert, KERN_ALERT);
8909 define_netdev_printk_level(netdev_crit, KERN_CRIT);
8910 define_netdev_printk_level(netdev_err, KERN_ERR);
8911 define_netdev_printk_level(netdev_warn, KERN_WARNING);
8912 define_netdev_printk_level(netdev_notice, KERN_NOTICE);
8913 define_netdev_printk_level(netdev_info, KERN_INFO);
8915 static void __net_exit netdev_exit(struct net *net)
8917 kfree(net->dev_name_head);
8918 kfree(net->dev_index_head);
8919 if (net != &init_net)
8920 WARN_ON_ONCE(!list_empty(&net->dev_base_head));
8923 static struct pernet_operations __net_initdata netdev_net_ops = {
8924 .init = netdev_init,
8925 .exit = netdev_exit,
8928 static void __net_exit default_device_exit(struct net *net)
8930 struct net_device *dev, *aux;
8932 * Push all migratable network devices back to the
8933 * initial network namespace
8936 for_each_netdev_safe(net, dev, aux) {
8938 char fb_name[IFNAMSIZ];
8940 /* Ignore unmoveable devices (i.e. loopback) */
8941 if (dev->features & NETIF_F_NETNS_LOCAL)
8944 /* Leave virtual devices for the generic cleanup */
8945 if (dev->rtnl_link_ops)
8948 /* Push remaining network devices to init_net */
8949 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex);
8950 err = dev_change_net_namespace(dev, &init_net, fb_name);
8952 pr_emerg("%s: failed to move %s to init_net: %d\n",
8953 __func__, dev->name, err);
8960 static void __net_exit rtnl_lock_unregistering(struct list_head *net_list)
8962 /* Return with the rtnl_lock held when there are no network
8963 * devices unregistering in any network namespace in net_list.
8967 DEFINE_WAIT_FUNC(wait, woken_wake_function);
8969 add_wait_queue(&netdev_unregistering_wq, &wait);
8971 unregistering = false;
8973 list_for_each_entry(net, net_list, exit_list) {
8974 if (net->dev_unreg_count > 0) {
8975 unregistering = true;
8983 wait_woken(&wait, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
8985 remove_wait_queue(&netdev_unregistering_wq, &wait);
8988 static void __net_exit default_device_exit_batch(struct list_head *net_list)
8990 /* At exit all network devices most be removed from a network
8991 * namespace. Do this in the reverse order of registration.
8992 * Do this across as many network namespaces as possible to
8993 * improve batching efficiency.
8995 struct net_device *dev;
8997 LIST_HEAD(dev_kill_list);
8999 /* To prevent network device cleanup code from dereferencing
9000 * loopback devices or network devices that have been freed
9001 * wait here for all pending unregistrations to complete,
9002 * before unregistring the loopback device and allowing the
9003 * network namespace be freed.
9005 * The netdev todo list containing all network devices
9006 * unregistrations that happen in default_device_exit_batch
9007 * will run in the rtnl_unlock() at the end of
9008 * default_device_exit_batch.
9010 rtnl_lock_unregistering(net_list);
9011 list_for_each_entry(net, net_list, exit_list) {
9012 for_each_netdev_reverse(net, dev) {
9013 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink)
9014 dev->rtnl_link_ops->dellink(dev, &dev_kill_list);
9016 unregister_netdevice_queue(dev, &dev_kill_list);
9019 unregister_netdevice_many(&dev_kill_list);
9023 static struct pernet_operations __net_initdata default_device_ops = {
9024 .exit = default_device_exit,
9025 .exit_batch = default_device_exit_batch,
9029 * Initialize the DEV module. At boot time this walks the device list and
9030 * unhooks any devices that fail to initialise (normally hardware not
9031 * present) and leaves us with a valid list of present and active devices.
9036 * This is called single threaded during boot, so no need
9037 * to take the rtnl semaphore.
9039 static int __init net_dev_init(void)
9041 int i, rc = -ENOMEM;
9043 BUG_ON(!dev_boot_phase);
9045 if (dev_proc_init())
9048 if (netdev_kobject_init())
9051 INIT_LIST_HEAD(&ptype_all);
9052 for (i = 0; i < PTYPE_HASH_SIZE; i++)
9053 INIT_LIST_HEAD(&ptype_base[i]);
9055 INIT_LIST_HEAD(&offload_base);
9057 if (register_pernet_subsys(&netdev_net_ops))
9061 * Initialise the packet receive queues.
9064 for_each_possible_cpu(i) {
9065 struct work_struct *flush = per_cpu_ptr(&flush_works, i);
9066 struct softnet_data *sd = &per_cpu(softnet_data, i);
9068 INIT_WORK(flush, flush_backlog);
9070 skb_queue_head_init(&sd->input_pkt_queue);
9071 skb_queue_head_init(&sd->process_queue);
9072 #ifdef CONFIG_XFRM_OFFLOAD
9073 skb_queue_head_init(&sd->xfrm_backlog);
9075 INIT_LIST_HEAD(&sd->poll_list);
9076 sd->output_queue_tailp = &sd->output_queue;
9078 sd->csd.func = rps_trigger_softirq;
9083 sd->backlog.poll = process_backlog;
9084 sd->backlog.weight = weight_p;
9089 /* The loopback device is special if any other network devices
9090 * is present in a network namespace the loopback device must
9091 * be present. Since we now dynamically allocate and free the
9092 * loopback device ensure this invariant is maintained by
9093 * keeping the loopback device as the first device on the
9094 * list of network devices. Ensuring the loopback devices
9095 * is the first device that appears and the last network device
9098 if (register_pernet_device(&loopback_net_ops))
9101 if (register_pernet_device(&default_device_ops))
9104 open_softirq(NET_TX_SOFTIRQ, net_tx_action);
9105 open_softirq(NET_RX_SOFTIRQ, net_rx_action);
9107 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead",
9108 NULL, dev_cpu_dead);
9115 subsys_initcall(net_dev_init);