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 u16 __skb_tx_hash(const struct net_device *dev, struct sk_buff *skb,
2619 unsigned int num_tx_queues)
2623 u16 qcount = num_tx_queues;
2625 if (skb_rx_queue_recorded(skb)) {
2626 hash = skb_get_rx_queue(skb);
2627 while (unlikely(hash >= num_tx_queues))
2628 hash -= num_tx_queues;
2633 u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
2635 qoffset = dev->tc_to_txq[tc].offset;
2636 qcount = dev->tc_to_txq[tc].count;
2639 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset;
2641 EXPORT_SYMBOL(__skb_tx_hash);
2643 static void skb_warn_bad_offload(const struct sk_buff *skb)
2645 static const netdev_features_t null_features;
2646 struct net_device *dev = skb->dev;
2647 const char *name = "";
2649 if (!net_ratelimit())
2653 if (dev->dev.parent)
2654 name = dev_driver_string(dev->dev.parent);
2656 name = netdev_name(dev);
2658 WARN(1, "%s: caps=(%pNF, %pNF) len=%d data_len=%d gso_size=%d "
2659 "gso_type=%d ip_summed=%d\n",
2660 name, dev ? &dev->features : &null_features,
2661 skb->sk ? &skb->sk->sk_route_caps : &null_features,
2662 skb->len, skb->data_len, skb_shinfo(skb)->gso_size,
2663 skb_shinfo(skb)->gso_type, skb->ip_summed);
2667 * Invalidate hardware checksum when packet is to be mangled, and
2668 * complete checksum manually on outgoing path.
2670 int skb_checksum_help(struct sk_buff *skb)
2673 int ret = 0, offset;
2675 if (skb->ip_summed == CHECKSUM_COMPLETE)
2676 goto out_set_summed;
2678 if (unlikely(skb_shinfo(skb)->gso_size)) {
2679 skb_warn_bad_offload(skb);
2683 /* Before computing a checksum, we should make sure no frag could
2684 * be modified by an external entity : checksum could be wrong.
2686 if (skb_has_shared_frag(skb)) {
2687 ret = __skb_linearize(skb);
2692 offset = skb_checksum_start_offset(skb);
2693 BUG_ON(offset >= skb_headlen(skb));
2694 csum = skb_checksum(skb, offset, skb->len - offset, 0);
2696 offset += skb->csum_offset;
2697 BUG_ON(offset + sizeof(__sum16) > skb_headlen(skb));
2699 if (skb_cloned(skb) &&
2700 !skb_clone_writable(skb, offset + sizeof(__sum16))) {
2701 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2706 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0;
2708 skb->ip_summed = CHECKSUM_NONE;
2712 EXPORT_SYMBOL(skb_checksum_help);
2714 int skb_crc32c_csum_help(struct sk_buff *skb)
2717 int ret = 0, offset, start;
2719 if (skb->ip_summed != CHECKSUM_PARTIAL)
2722 if (unlikely(skb_is_gso(skb)))
2725 /* Before computing a checksum, we should make sure no frag could
2726 * be modified by an external entity : checksum could be wrong.
2728 if (unlikely(skb_has_shared_frag(skb))) {
2729 ret = __skb_linearize(skb);
2733 start = skb_checksum_start_offset(skb);
2734 offset = start + offsetof(struct sctphdr, checksum);
2735 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) {
2739 if (skb_cloned(skb) &&
2740 !skb_clone_writable(skb, offset + sizeof(__le32))) {
2741 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2745 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start,
2746 skb->len - start, ~(__u32)0,
2748 *(__le32 *)(skb->data + offset) = crc32c_csum;
2749 skb->ip_summed = CHECKSUM_NONE;
2750 skb->csum_not_inet = 0;
2755 __be16 skb_network_protocol(struct sk_buff *skb, int *depth)
2757 __be16 type = skb->protocol;
2759 /* Tunnel gso handlers can set protocol to ethernet. */
2760 if (type == htons(ETH_P_TEB)) {
2763 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr))))
2766 eth = (struct ethhdr *)skb->data;
2767 type = eth->h_proto;
2770 return __vlan_get_protocol(skb, type, depth);
2774 * skb_mac_gso_segment - mac layer segmentation handler.
2775 * @skb: buffer to segment
2776 * @features: features for the output path (see dev->features)
2778 struct sk_buff *skb_mac_gso_segment(struct sk_buff *skb,
2779 netdev_features_t features)
2781 struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT);
2782 struct packet_offload *ptype;
2783 int vlan_depth = skb->mac_len;
2784 __be16 type = skb_network_protocol(skb, &vlan_depth);
2786 if (unlikely(!type))
2787 return ERR_PTR(-EINVAL);
2789 __skb_pull(skb, vlan_depth);
2792 list_for_each_entry_rcu(ptype, &offload_base, list) {
2793 if (ptype->type == type && ptype->callbacks.gso_segment) {
2794 segs = ptype->callbacks.gso_segment(skb, features);
2800 __skb_push(skb, skb->data - skb_mac_header(skb));
2804 EXPORT_SYMBOL(skb_mac_gso_segment);
2807 /* openvswitch calls this on rx path, so we need a different check.
2809 static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
2812 return skb->ip_summed != CHECKSUM_PARTIAL &&
2813 skb->ip_summed != CHECKSUM_UNNECESSARY;
2815 return skb->ip_summed == CHECKSUM_NONE;
2819 * __skb_gso_segment - Perform segmentation on skb.
2820 * @skb: buffer to segment
2821 * @features: features for the output path (see dev->features)
2822 * @tx_path: whether it is called in TX path
2824 * This function segments the given skb and returns a list of segments.
2826 * It may return NULL if the skb requires no segmentation. This is
2827 * only possible when GSO is used for verifying header integrity.
2829 * Segmentation preserves SKB_SGO_CB_OFFSET bytes of previous skb cb.
2831 struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
2832 netdev_features_t features, bool tx_path)
2834 struct sk_buff *segs;
2836 if (unlikely(skb_needs_check(skb, tx_path))) {
2839 /* We're going to init ->check field in TCP or UDP header */
2840 err = skb_cow_head(skb, 0);
2842 return ERR_PTR(err);
2845 /* Only report GSO partial support if it will enable us to
2846 * support segmentation on this frame without needing additional
2849 if (features & NETIF_F_GSO_PARTIAL) {
2850 netdev_features_t partial_features = NETIF_F_GSO_ROBUST;
2851 struct net_device *dev = skb->dev;
2853 partial_features |= dev->features & dev->gso_partial_features;
2854 if (!skb_gso_ok(skb, features | partial_features))
2855 features &= ~NETIF_F_GSO_PARTIAL;
2858 BUILD_BUG_ON(SKB_SGO_CB_OFFSET +
2859 sizeof(*SKB_GSO_CB(skb)) > sizeof(skb->cb));
2861 SKB_GSO_CB(skb)->mac_offset = skb_headroom(skb);
2862 SKB_GSO_CB(skb)->encap_level = 0;
2864 skb_reset_mac_header(skb);
2865 skb_reset_mac_len(skb);
2867 segs = skb_mac_gso_segment(skb, features);
2869 if (unlikely(skb_needs_check(skb, tx_path) && !IS_ERR(segs)))
2870 skb_warn_bad_offload(skb);
2874 EXPORT_SYMBOL(__skb_gso_segment);
2876 /* Take action when hardware reception checksum errors are detected. */
2878 void netdev_rx_csum_fault(struct net_device *dev)
2880 if (net_ratelimit()) {
2881 pr_err("%s: hw csum failure\n", dev ? dev->name : "<unknown>");
2885 EXPORT_SYMBOL(netdev_rx_csum_fault);
2888 /* Actually, we should eliminate this check as soon as we know, that:
2889 * 1. IOMMU is present and allows to map all the memory.
2890 * 2. No high memory really exists on this machine.
2893 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
2895 #ifdef CONFIG_HIGHMEM
2898 if (!(dev->features & NETIF_F_HIGHDMA)) {
2899 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2900 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2902 if (PageHighMem(skb_frag_page(frag)))
2907 if (PCI_DMA_BUS_IS_PHYS) {
2908 struct device *pdev = dev->dev.parent;
2912 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2913 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2914 dma_addr_t addr = page_to_phys(skb_frag_page(frag));
2916 if (!pdev->dma_mask || addr + PAGE_SIZE - 1 > *pdev->dma_mask)
2924 /* If MPLS offload request, verify we are testing hardware MPLS features
2925 * instead of standard features for the netdev.
2927 #if IS_ENABLED(CONFIG_NET_MPLS_GSO)
2928 static netdev_features_t net_mpls_features(struct sk_buff *skb,
2929 netdev_features_t features,
2932 if (eth_p_mpls(type))
2933 features &= skb->dev->mpls_features;
2938 static netdev_features_t net_mpls_features(struct sk_buff *skb,
2939 netdev_features_t features,
2946 static netdev_features_t harmonize_features(struct sk_buff *skb,
2947 netdev_features_t features)
2952 type = skb_network_protocol(skb, &tmp);
2953 features = net_mpls_features(skb, features, type);
2955 if (skb->ip_summed != CHECKSUM_NONE &&
2956 !can_checksum_protocol(features, type)) {
2957 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
2959 if (illegal_highdma(skb->dev, skb))
2960 features &= ~NETIF_F_SG;
2965 netdev_features_t passthru_features_check(struct sk_buff *skb,
2966 struct net_device *dev,
2967 netdev_features_t features)
2971 EXPORT_SYMBOL(passthru_features_check);
2973 static netdev_features_t dflt_features_check(struct sk_buff *skb,
2974 struct net_device *dev,
2975 netdev_features_t features)
2977 return vlan_features_check(skb, features);
2980 static netdev_features_t gso_features_check(const struct sk_buff *skb,
2981 struct net_device *dev,
2982 netdev_features_t features)
2984 u16 gso_segs = skb_shinfo(skb)->gso_segs;
2986 if (gso_segs > dev->gso_max_segs)
2987 return features & ~NETIF_F_GSO_MASK;
2989 /* Support for GSO partial features requires software
2990 * intervention before we can actually process the packets
2991 * so we need to strip support for any partial features now
2992 * and we can pull them back in after we have partially
2993 * segmented the frame.
2995 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL))
2996 features &= ~dev->gso_partial_features;
2998 /* Make sure to clear the IPv4 ID mangling feature if the
2999 * IPv4 header has the potential to be fragmented.
3001 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
3002 struct iphdr *iph = skb->encapsulation ?
3003 inner_ip_hdr(skb) : ip_hdr(skb);
3005 if (!(iph->frag_off & htons(IP_DF)))
3006 features &= ~NETIF_F_TSO_MANGLEID;
3012 netdev_features_t netif_skb_features(struct sk_buff *skb)
3014 struct net_device *dev = skb->dev;
3015 netdev_features_t features = dev->features;
3017 if (skb_is_gso(skb))
3018 features = gso_features_check(skb, dev, features);
3020 /* If encapsulation offload request, verify we are testing
3021 * hardware encapsulation features instead of standard
3022 * features for the netdev
3024 if (skb->encapsulation)
3025 features &= dev->hw_enc_features;
3027 if (skb_vlan_tagged(skb))
3028 features = netdev_intersect_features(features,
3029 dev->vlan_features |
3030 NETIF_F_HW_VLAN_CTAG_TX |
3031 NETIF_F_HW_VLAN_STAG_TX);
3033 if (dev->netdev_ops->ndo_features_check)
3034 features &= dev->netdev_ops->ndo_features_check(skb, dev,
3037 features &= dflt_features_check(skb, dev, features);
3039 return harmonize_features(skb, features);
3041 EXPORT_SYMBOL(netif_skb_features);
3043 static int xmit_one(struct sk_buff *skb, struct net_device *dev,
3044 struct netdev_queue *txq, bool more)
3049 if (!list_empty(&ptype_all) || !list_empty(&dev->ptype_all))
3050 dev_queue_xmit_nit(skb, dev);
3053 trace_net_dev_start_xmit(skb, dev);
3054 rc = netdev_start_xmit(skb, dev, txq, more);
3055 trace_net_dev_xmit(skb, rc, dev, len);
3060 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev,
3061 struct netdev_queue *txq, int *ret)
3063 struct sk_buff *skb = first;
3064 int rc = NETDEV_TX_OK;
3067 struct sk_buff *next = skb->next;
3070 rc = xmit_one(skb, dev, txq, next != NULL);
3071 if (unlikely(!dev_xmit_complete(rc))) {
3077 if (netif_xmit_stopped(txq) && skb) {
3078 rc = NETDEV_TX_BUSY;
3088 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb,
3089 netdev_features_t features)
3091 if (skb_vlan_tag_present(skb) &&
3092 !vlan_hw_offload_capable(features, skb->vlan_proto))
3093 skb = __vlan_hwaccel_push_inside(skb);
3097 int skb_csum_hwoffload_help(struct sk_buff *skb,
3098 const netdev_features_t features)
3100 if (unlikely(skb->csum_not_inet))
3101 return !!(features & NETIF_F_SCTP_CRC) ? 0 :
3102 skb_crc32c_csum_help(skb);
3104 return !!(features & NETIF_F_CSUM_MASK) ? 0 : skb_checksum_help(skb);
3106 EXPORT_SYMBOL(skb_csum_hwoffload_help);
3108 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again)
3110 netdev_features_t features;
3112 features = netif_skb_features(skb);
3113 skb = validate_xmit_vlan(skb, features);
3117 if (netif_needs_gso(skb, features)) {
3118 struct sk_buff *segs;
3120 segs = skb_gso_segment(skb, features);
3128 if (skb_needs_linearize(skb, features) &&
3129 __skb_linearize(skb))
3132 /* If packet is not checksummed and device does not
3133 * support checksumming for this protocol, complete
3134 * checksumming here.
3136 if (skb->ip_summed == CHECKSUM_PARTIAL) {
3137 if (skb->encapsulation)
3138 skb_set_inner_transport_header(skb,
3139 skb_checksum_start_offset(skb));
3141 skb_set_transport_header(skb,
3142 skb_checksum_start_offset(skb));
3143 if (skb_csum_hwoffload_help(skb, features))
3148 skb = validate_xmit_xfrm(skb, features, again);
3155 atomic_long_inc(&dev->tx_dropped);
3159 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again)
3161 struct sk_buff *next, *head = NULL, *tail;
3163 for (; skb != NULL; skb = next) {
3167 /* in case skb wont be segmented, point to itself */
3170 skb = validate_xmit_skb(skb, dev, again);
3178 /* If skb was segmented, skb->prev points to
3179 * the last segment. If not, it still contains skb.
3185 EXPORT_SYMBOL_GPL(validate_xmit_skb_list);
3187 static void qdisc_pkt_len_init(struct sk_buff *skb)
3189 const struct skb_shared_info *shinfo = skb_shinfo(skb);
3191 qdisc_skb_cb(skb)->pkt_len = skb->len;
3193 /* To get more precise estimation of bytes sent on wire,
3194 * we add to pkt_len the headers size of all segments
3196 if (shinfo->gso_size) {
3197 unsigned int hdr_len;
3198 u16 gso_segs = shinfo->gso_segs;
3200 /* mac layer + network layer */
3201 hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
3203 /* + transport layer */
3204 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) {
3205 const struct tcphdr *th;
3206 struct tcphdr _tcphdr;
3208 th = skb_header_pointer(skb, skb_transport_offset(skb),
3209 sizeof(_tcphdr), &_tcphdr);
3211 hdr_len += __tcp_hdrlen(th);
3213 struct udphdr _udphdr;
3215 if (skb_header_pointer(skb, skb_transport_offset(skb),
3216 sizeof(_udphdr), &_udphdr))
3217 hdr_len += sizeof(struct udphdr);
3220 if (shinfo->gso_type & SKB_GSO_DODGY)
3221 gso_segs = DIV_ROUND_UP(skb->len - hdr_len,
3224 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len;
3228 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
3229 struct net_device *dev,
3230 struct netdev_queue *txq)
3232 spinlock_t *root_lock = qdisc_lock(q);
3233 struct sk_buff *to_free = NULL;
3237 qdisc_calculate_pkt_len(skb, q);
3239 if (q->flags & TCQ_F_NOLOCK) {
3240 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3241 __qdisc_drop(skb, &to_free);
3244 rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
3248 if (unlikely(to_free))
3249 kfree_skb_list(to_free);
3254 * Heuristic to force contended enqueues to serialize on a
3255 * separate lock before trying to get qdisc main lock.
3256 * This permits qdisc->running owner to get the lock more
3257 * often and dequeue packets faster.
3259 contended = qdisc_is_running(q);
3260 if (unlikely(contended))
3261 spin_lock(&q->busylock);
3263 spin_lock(root_lock);
3264 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3265 __qdisc_drop(skb, &to_free);
3267 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
3268 qdisc_run_begin(q)) {
3270 * This is a work-conserving queue; there are no old skbs
3271 * waiting to be sent out; and the qdisc is not running -
3272 * xmit the skb directly.
3275 qdisc_bstats_update(q, skb);
3277 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) {
3278 if (unlikely(contended)) {
3279 spin_unlock(&q->busylock);
3286 rc = NET_XMIT_SUCCESS;
3288 rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
3289 if (qdisc_run_begin(q)) {
3290 if (unlikely(contended)) {
3291 spin_unlock(&q->busylock);
3298 spin_unlock(root_lock);
3299 if (unlikely(to_free))
3300 kfree_skb_list(to_free);
3301 if (unlikely(contended))
3302 spin_unlock(&q->busylock);
3306 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO)
3307 static void skb_update_prio(struct sk_buff *skb)
3309 const struct netprio_map *map;
3310 const struct sock *sk;
3311 unsigned int prioidx;
3315 map = rcu_dereference_bh(skb->dev->priomap);
3318 sk = skb_to_full_sk(skb);
3322 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data);
3324 if (prioidx < map->priomap_len)
3325 skb->priority = map->priomap[prioidx];
3328 #define skb_update_prio(skb)
3331 DEFINE_PER_CPU(int, xmit_recursion);
3332 EXPORT_SYMBOL(xmit_recursion);
3335 * dev_loopback_xmit - loop back @skb
3336 * @net: network namespace this loopback is happening in
3337 * @sk: sk needed to be a netfilter okfn
3338 * @skb: buffer to transmit
3340 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb)
3342 skb_reset_mac_header(skb);
3343 __skb_pull(skb, skb_network_offset(skb));
3344 skb->pkt_type = PACKET_LOOPBACK;
3345 skb->ip_summed = CHECKSUM_UNNECESSARY;
3346 WARN_ON(!skb_dst(skb));
3351 EXPORT_SYMBOL(dev_loopback_xmit);
3353 #ifdef CONFIG_NET_EGRESS
3354 static struct sk_buff *
3355 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
3357 struct mini_Qdisc *miniq = rcu_dereference_bh(dev->miniq_egress);
3358 struct tcf_result cl_res;
3363 /* qdisc_skb_cb(skb)->pkt_len was already set by the caller. */
3364 mini_qdisc_bstats_cpu_update(miniq, skb);
3366 switch (tcf_classify(skb, miniq->filter_list, &cl_res, false)) {
3368 case TC_ACT_RECLASSIFY:
3369 skb->tc_index = TC_H_MIN(cl_res.classid);
3372 mini_qdisc_qstats_cpu_drop(miniq);
3373 *ret = NET_XMIT_DROP;
3379 *ret = NET_XMIT_SUCCESS;
3382 case TC_ACT_REDIRECT:
3383 /* No need to push/pop skb's mac_header here on egress! */
3384 skb_do_redirect(skb);
3385 *ret = NET_XMIT_SUCCESS;
3393 #endif /* CONFIG_NET_EGRESS */
3395 static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb)
3398 struct xps_dev_maps *dev_maps;
3399 struct xps_map *map;
3400 int queue_index = -1;
3403 dev_maps = rcu_dereference(dev->xps_maps);
3405 unsigned int tci = skb->sender_cpu - 1;
3409 tci += netdev_get_prio_tc_map(dev, skb->priority);
3412 map = rcu_dereference(dev_maps->cpu_map[tci]);
3415 queue_index = map->queues[0];
3417 queue_index = map->queues[reciprocal_scale(skb_get_hash(skb),
3419 if (unlikely(queue_index >= dev->real_num_tx_queues))
3431 static u16 __netdev_pick_tx(struct net_device *dev, struct sk_buff *skb)
3433 struct sock *sk = skb->sk;
3434 int queue_index = sk_tx_queue_get(sk);
3436 if (queue_index < 0 || skb->ooo_okay ||
3437 queue_index >= dev->real_num_tx_queues) {
3438 int new_index = get_xps_queue(dev, skb);
3441 new_index = skb_tx_hash(dev, skb);
3443 if (queue_index != new_index && sk &&
3445 rcu_access_pointer(sk->sk_dst_cache))
3446 sk_tx_queue_set(sk, new_index);
3448 queue_index = new_index;
3454 struct netdev_queue *netdev_pick_tx(struct net_device *dev,
3455 struct sk_buff *skb,
3458 int queue_index = 0;
3461 u32 sender_cpu = skb->sender_cpu - 1;
3463 if (sender_cpu >= (u32)NR_CPUS)
3464 skb->sender_cpu = raw_smp_processor_id() + 1;
3467 if (dev->real_num_tx_queues != 1) {
3468 const struct net_device_ops *ops = dev->netdev_ops;
3470 if (ops->ndo_select_queue)
3471 queue_index = ops->ndo_select_queue(dev, skb, accel_priv,
3474 queue_index = __netdev_pick_tx(dev, skb);
3476 queue_index = netdev_cap_txqueue(dev, queue_index);
3479 skb_set_queue_mapping(skb, queue_index);
3480 return netdev_get_tx_queue(dev, queue_index);
3484 * __dev_queue_xmit - transmit a buffer
3485 * @skb: buffer to transmit
3486 * @accel_priv: private data used for L2 forwarding offload
3488 * Queue a buffer for transmission to a network device. The caller must
3489 * have set the device and priority and built the buffer before calling
3490 * this function. The function can be called from an interrupt.
3492 * A negative errno code is returned on a failure. A success does not
3493 * guarantee the frame will be transmitted as it may be dropped due
3494 * to congestion or traffic shaping.
3496 * -----------------------------------------------------------------------------------
3497 * I notice this method can also return errors from the queue disciplines,
3498 * including NET_XMIT_DROP, which is a positive value. So, errors can also
3501 * Regardless of the return value, the skb is consumed, so it is currently
3502 * difficult to retry a send to this method. (You can bump the ref count
3503 * before sending to hold a reference for retry if you are careful.)
3505 * When calling this method, interrupts MUST be enabled. This is because
3506 * the BH enable code must have IRQs enabled so that it will not deadlock.
3509 static int __dev_queue_xmit(struct sk_buff *skb, void *accel_priv)
3511 struct net_device *dev = skb->dev;
3512 struct netdev_queue *txq;
3517 skb_reset_mac_header(skb);
3519 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP))
3520 __skb_tstamp_tx(skb, NULL, skb->sk, SCM_TSTAMP_SCHED);
3522 /* Disable soft irqs for various locks below. Also
3523 * stops preemption for RCU.
3527 skb_update_prio(skb);
3529 qdisc_pkt_len_init(skb);
3530 #ifdef CONFIG_NET_CLS_ACT
3531 skb->tc_at_ingress = 0;
3532 # ifdef CONFIG_NET_EGRESS
3533 if (static_key_false(&egress_needed)) {
3534 skb = sch_handle_egress(skb, &rc, dev);
3540 /* If device/qdisc don't need skb->dst, release it right now while
3541 * its hot in this cpu cache.
3543 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
3548 txq = netdev_pick_tx(dev, skb, accel_priv);
3549 q = rcu_dereference_bh(txq->qdisc);
3551 trace_net_dev_queue(skb);
3553 rc = __dev_xmit_skb(skb, q, dev, txq);
3557 /* The device has no queue. Common case for software devices:
3558 * loopback, all the sorts of tunnels...
3560 * Really, it is unlikely that netif_tx_lock protection is necessary
3561 * here. (f.e. loopback and IP tunnels are clean ignoring statistics
3563 * However, it is possible, that they rely on protection
3566 * Check this and shot the lock. It is not prone from deadlocks.
3567 *Either shot noqueue qdisc, it is even simpler 8)
3569 if (dev->flags & IFF_UP) {
3570 int cpu = smp_processor_id(); /* ok because BHs are off */
3572 if (txq->xmit_lock_owner != cpu) {
3573 if (unlikely(__this_cpu_read(xmit_recursion) >
3574 XMIT_RECURSION_LIMIT))
3575 goto recursion_alert;
3577 skb = validate_xmit_skb(skb, dev, &again);
3581 HARD_TX_LOCK(dev, txq, cpu);
3583 if (!netif_xmit_stopped(txq)) {
3584 __this_cpu_inc(xmit_recursion);
3585 skb = dev_hard_start_xmit(skb, dev, txq, &rc);
3586 __this_cpu_dec(xmit_recursion);
3587 if (dev_xmit_complete(rc)) {
3588 HARD_TX_UNLOCK(dev, txq);
3592 HARD_TX_UNLOCK(dev, txq);
3593 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
3596 /* Recursion is detected! It is possible,
3600 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
3606 rcu_read_unlock_bh();
3608 atomic_long_inc(&dev->tx_dropped);
3609 kfree_skb_list(skb);
3612 rcu_read_unlock_bh();
3616 int dev_queue_xmit(struct sk_buff *skb)
3618 return __dev_queue_xmit(skb, NULL);
3620 EXPORT_SYMBOL(dev_queue_xmit);
3622 int dev_queue_xmit_accel(struct sk_buff *skb, void *accel_priv)
3624 return __dev_queue_xmit(skb, accel_priv);
3626 EXPORT_SYMBOL(dev_queue_xmit_accel);
3629 /*************************************************************************
3631 *************************************************************************/
3633 int netdev_max_backlog __read_mostly = 1000;
3634 EXPORT_SYMBOL(netdev_max_backlog);
3636 int netdev_tstamp_prequeue __read_mostly = 1;
3637 int netdev_budget __read_mostly = 300;
3638 unsigned int __read_mostly netdev_budget_usecs = 2000;
3639 int weight_p __read_mostly = 64; /* old backlog weight */
3640 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */
3641 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */
3642 int dev_rx_weight __read_mostly = 64;
3643 int dev_tx_weight __read_mostly = 64;
3645 /* Called with irq disabled */
3646 static inline void ____napi_schedule(struct softnet_data *sd,
3647 struct napi_struct *napi)
3649 list_add_tail(&napi->poll_list, &sd->poll_list);
3650 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3655 /* One global table that all flow-based protocols share. */
3656 struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
3657 EXPORT_SYMBOL(rps_sock_flow_table);
3658 u32 rps_cpu_mask __read_mostly;
3659 EXPORT_SYMBOL(rps_cpu_mask);
3661 struct static_key rps_needed __read_mostly;
3662 EXPORT_SYMBOL(rps_needed);
3663 struct static_key rfs_needed __read_mostly;
3664 EXPORT_SYMBOL(rfs_needed);
3666 static struct rps_dev_flow *
3667 set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3668 struct rps_dev_flow *rflow, u16 next_cpu)
3670 if (next_cpu < nr_cpu_ids) {
3671 #ifdef CONFIG_RFS_ACCEL
3672 struct netdev_rx_queue *rxqueue;
3673 struct rps_dev_flow_table *flow_table;
3674 struct rps_dev_flow *old_rflow;
3679 /* Should we steer this flow to a different hardware queue? */
3680 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
3681 !(dev->features & NETIF_F_NTUPLE))
3683 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
3684 if (rxq_index == skb_get_rx_queue(skb))
3687 rxqueue = dev->_rx + rxq_index;
3688 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3691 flow_id = skb_get_hash(skb) & flow_table->mask;
3692 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
3693 rxq_index, flow_id);
3697 rflow = &flow_table->flows[flow_id];
3699 if (old_rflow->filter == rflow->filter)
3700 old_rflow->filter = RPS_NO_FILTER;
3704 per_cpu(softnet_data, next_cpu).input_queue_head;
3707 rflow->cpu = next_cpu;
3712 * get_rps_cpu is called from netif_receive_skb and returns the target
3713 * CPU from the RPS map of the receiving queue for a given skb.
3714 * rcu_read_lock must be held on entry.
3716 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3717 struct rps_dev_flow **rflowp)
3719 const struct rps_sock_flow_table *sock_flow_table;
3720 struct netdev_rx_queue *rxqueue = dev->_rx;
3721 struct rps_dev_flow_table *flow_table;
3722 struct rps_map *map;
3727 if (skb_rx_queue_recorded(skb)) {
3728 u16 index = skb_get_rx_queue(skb);
3730 if (unlikely(index >= dev->real_num_rx_queues)) {
3731 WARN_ONCE(dev->real_num_rx_queues > 1,
3732 "%s received packet on queue %u, but number "
3733 "of RX queues is %u\n",
3734 dev->name, index, dev->real_num_rx_queues);
3740 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */
3742 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3743 map = rcu_dereference(rxqueue->rps_map);
3744 if (!flow_table && !map)
3747 skb_reset_network_header(skb);
3748 hash = skb_get_hash(skb);
3752 sock_flow_table = rcu_dereference(rps_sock_flow_table);
3753 if (flow_table && sock_flow_table) {
3754 struct rps_dev_flow *rflow;
3758 /* First check into global flow table if there is a match */
3759 ident = sock_flow_table->ents[hash & sock_flow_table->mask];
3760 if ((ident ^ hash) & ~rps_cpu_mask)
3763 next_cpu = ident & rps_cpu_mask;
3765 /* OK, now we know there is a match,
3766 * we can look at the local (per receive queue) flow table
3768 rflow = &flow_table->flows[hash & flow_table->mask];
3772 * If the desired CPU (where last recvmsg was done) is
3773 * different from current CPU (one in the rx-queue flow
3774 * table entry), switch if one of the following holds:
3775 * - Current CPU is unset (>= nr_cpu_ids).
3776 * - Current CPU is offline.
3777 * - The current CPU's queue tail has advanced beyond the
3778 * last packet that was enqueued using this table entry.
3779 * This guarantees that all previous packets for the flow
3780 * have been dequeued, thus preserving in order delivery.
3782 if (unlikely(tcpu != next_cpu) &&
3783 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) ||
3784 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
3785 rflow->last_qtail)) >= 0)) {
3787 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
3790 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) {
3800 tcpu = map->cpus[reciprocal_scale(hash, map->len)];
3801 if (cpu_online(tcpu)) {
3811 #ifdef CONFIG_RFS_ACCEL
3814 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
3815 * @dev: Device on which the filter was set
3816 * @rxq_index: RX queue index
3817 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
3818 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
3820 * Drivers that implement ndo_rx_flow_steer() should periodically call
3821 * this function for each installed filter and remove the filters for
3822 * which it returns %true.
3824 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
3825 u32 flow_id, u16 filter_id)
3827 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
3828 struct rps_dev_flow_table *flow_table;
3829 struct rps_dev_flow *rflow;
3834 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3835 if (flow_table && flow_id <= flow_table->mask) {
3836 rflow = &flow_table->flows[flow_id];
3837 cpu = READ_ONCE(rflow->cpu);
3838 if (rflow->filter == filter_id && cpu < nr_cpu_ids &&
3839 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
3840 rflow->last_qtail) <
3841 (int)(10 * flow_table->mask)))
3847 EXPORT_SYMBOL(rps_may_expire_flow);
3849 #endif /* CONFIG_RFS_ACCEL */
3851 /* Called from hardirq (IPI) context */
3852 static void rps_trigger_softirq(void *data)
3854 struct softnet_data *sd = data;
3856 ____napi_schedule(sd, &sd->backlog);
3860 #endif /* CONFIG_RPS */
3863 * Check if this softnet_data structure is another cpu one
3864 * If yes, queue it to our IPI list and return 1
3867 static int rps_ipi_queued(struct softnet_data *sd)
3870 struct softnet_data *mysd = this_cpu_ptr(&softnet_data);
3873 sd->rps_ipi_next = mysd->rps_ipi_list;
3874 mysd->rps_ipi_list = sd;
3876 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3879 #endif /* CONFIG_RPS */
3883 #ifdef CONFIG_NET_FLOW_LIMIT
3884 int netdev_flow_limit_table_len __read_mostly = (1 << 12);
3887 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
3889 #ifdef CONFIG_NET_FLOW_LIMIT
3890 struct sd_flow_limit *fl;
3891 struct softnet_data *sd;
3892 unsigned int old_flow, new_flow;
3894 if (qlen < (netdev_max_backlog >> 1))
3897 sd = this_cpu_ptr(&softnet_data);
3900 fl = rcu_dereference(sd->flow_limit);
3902 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1);
3903 old_flow = fl->history[fl->history_head];
3904 fl->history[fl->history_head] = new_flow;
3907 fl->history_head &= FLOW_LIMIT_HISTORY - 1;
3909 if (likely(fl->buckets[old_flow]))
3910 fl->buckets[old_flow]--;
3912 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) {
3924 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
3925 * queue (may be a remote CPU queue).
3927 static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
3928 unsigned int *qtail)
3930 struct softnet_data *sd;
3931 unsigned long flags;
3934 sd = &per_cpu(softnet_data, cpu);
3936 local_irq_save(flags);
3939 if (!netif_running(skb->dev))
3941 qlen = skb_queue_len(&sd->input_pkt_queue);
3942 if (qlen <= netdev_max_backlog && !skb_flow_limit(skb, qlen)) {
3945 __skb_queue_tail(&sd->input_pkt_queue, skb);
3946 input_queue_tail_incr_save(sd, qtail);
3948 local_irq_restore(flags);
3949 return NET_RX_SUCCESS;
3952 /* Schedule NAPI for backlog device
3953 * We can use non atomic operation since we own the queue lock
3955 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
3956 if (!rps_ipi_queued(sd))
3957 ____napi_schedule(sd, &sd->backlog);
3966 local_irq_restore(flags);
3968 atomic_long_inc(&skb->dev->rx_dropped);
3973 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb)
3975 struct net_device *dev = skb->dev;
3976 struct netdev_rx_queue *rxqueue;
3980 if (skb_rx_queue_recorded(skb)) {
3981 u16 index = skb_get_rx_queue(skb);
3983 if (unlikely(index >= dev->real_num_rx_queues)) {
3984 WARN_ONCE(dev->real_num_rx_queues > 1,
3985 "%s received packet on queue %u, but number "
3986 "of RX queues is %u\n",
3987 dev->name, index, dev->real_num_rx_queues);
3989 return rxqueue; /* Return first rxqueue */
3996 static u32 netif_receive_generic_xdp(struct sk_buff *skb,
3997 struct bpf_prog *xdp_prog)
3999 struct netdev_rx_queue *rxqueue;
4000 void *orig_data, *orig_data_end;
4001 u32 metalen, act = XDP_DROP;
4002 struct xdp_buff xdp;
4006 /* Reinjected packets coming from act_mirred or similar should
4007 * not get XDP generic processing.
4009 if (skb_cloned(skb))
4012 /* XDP packets must be linear and must have sufficient headroom
4013 * of XDP_PACKET_HEADROOM bytes. This is the guarantee that also
4014 * native XDP provides, thus we need to do it here as well.
4016 if (skb_is_nonlinear(skb) ||
4017 skb_headroom(skb) < XDP_PACKET_HEADROOM) {
4018 int hroom = XDP_PACKET_HEADROOM - skb_headroom(skb);
4019 int troom = skb->tail + skb->data_len - skb->end;
4021 /* In case we have to go down the path and also linearize,
4022 * then lets do the pskb_expand_head() work just once here.
4024 if (pskb_expand_head(skb,
4025 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0,
4026 troom > 0 ? troom + 128 : 0, GFP_ATOMIC))
4028 if (skb_linearize(skb))
4032 /* The XDP program wants to see the packet starting at the MAC
4035 mac_len = skb->data - skb_mac_header(skb);
4036 hlen = skb_headlen(skb) + mac_len;
4037 xdp.data = skb->data - mac_len;
4038 xdp.data_meta = xdp.data;
4039 xdp.data_end = xdp.data + hlen;
4040 xdp.data_hard_start = skb->data - skb_headroom(skb);
4041 orig_data_end = xdp.data_end;
4042 orig_data = xdp.data;
4044 rxqueue = netif_get_rxqueue(skb);
4045 xdp.rxq = &rxqueue->xdp_rxq;
4047 act = bpf_prog_run_xdp(xdp_prog, &xdp);
4049 off = xdp.data - orig_data;
4051 __skb_pull(skb, off);
4053 __skb_push(skb, -off);
4054 skb->mac_header += off;
4056 /* check if bpf_xdp_adjust_tail was used. it can only "shrink"
4059 off = orig_data_end - xdp.data_end;
4061 skb_set_tail_pointer(skb, xdp.data_end - xdp.data);
4066 __skb_push(skb, mac_len);
4069 metalen = xdp.data - xdp.data_meta;
4071 skb_metadata_set(skb, metalen);
4074 bpf_warn_invalid_xdp_action(act);
4077 trace_xdp_exception(skb->dev, xdp_prog, act);
4088 /* When doing generic XDP we have to bypass the qdisc layer and the
4089 * network taps in order to match in-driver-XDP behavior.
4091 void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog)
4093 struct net_device *dev = skb->dev;
4094 struct netdev_queue *txq;
4095 bool free_skb = true;
4098 txq = netdev_pick_tx(dev, skb, NULL);
4099 cpu = smp_processor_id();
4100 HARD_TX_LOCK(dev, txq, cpu);
4101 if (!netif_xmit_stopped(txq)) {
4102 rc = netdev_start_xmit(skb, dev, txq, 0);
4103 if (dev_xmit_complete(rc))
4106 HARD_TX_UNLOCK(dev, txq);
4108 trace_xdp_exception(dev, xdp_prog, XDP_TX);
4112 EXPORT_SYMBOL_GPL(generic_xdp_tx);
4114 static struct static_key generic_xdp_needed __read_mostly;
4116 int do_xdp_generic(struct bpf_prog *xdp_prog, struct sk_buff *skb)
4119 u32 act = netif_receive_generic_xdp(skb, xdp_prog);
4122 if (act != XDP_PASS) {
4125 err = xdp_do_generic_redirect(skb->dev, skb,
4129 /* fallthru to submit skb */
4131 generic_xdp_tx(skb, xdp_prog);
4142 EXPORT_SYMBOL_GPL(do_xdp_generic);
4144 static int netif_rx_internal(struct sk_buff *skb)
4148 net_timestamp_check(netdev_tstamp_prequeue, skb);
4150 trace_netif_rx(skb);
4152 if (static_key_false(&generic_xdp_needed)) {
4157 ret = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
4161 /* Consider XDP consuming the packet a success from
4162 * the netdev point of view we do not want to count
4165 if (ret != XDP_PASS)
4166 return NET_RX_SUCCESS;
4170 if (static_key_false(&rps_needed)) {
4171 struct rps_dev_flow voidflow, *rflow = &voidflow;
4177 cpu = get_rps_cpu(skb->dev, skb, &rflow);
4179 cpu = smp_processor_id();
4181 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4190 ret = enqueue_to_backlog(skb, get_cpu(), &qtail);
4197 * netif_rx - post buffer to the network code
4198 * @skb: buffer to post
4200 * This function receives a packet from a device driver and queues it for
4201 * the upper (protocol) levels to process. It always succeeds. The buffer
4202 * may be dropped during processing for congestion control or by the
4206 * NET_RX_SUCCESS (no congestion)
4207 * NET_RX_DROP (packet was dropped)
4211 int netif_rx(struct sk_buff *skb)
4213 trace_netif_rx_entry(skb);
4215 return netif_rx_internal(skb);
4217 EXPORT_SYMBOL(netif_rx);
4219 int netif_rx_ni(struct sk_buff *skb)
4223 trace_netif_rx_ni_entry(skb);
4226 err = netif_rx_internal(skb);
4227 if (local_softirq_pending())
4233 EXPORT_SYMBOL(netif_rx_ni);
4235 static __latent_entropy void net_tx_action(struct softirq_action *h)
4237 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
4239 if (sd->completion_queue) {
4240 struct sk_buff *clist;
4242 local_irq_disable();
4243 clist = sd->completion_queue;
4244 sd->completion_queue = NULL;
4248 struct sk_buff *skb = clist;
4250 clist = clist->next;
4252 WARN_ON(refcount_read(&skb->users));
4253 if (likely(get_kfree_skb_cb(skb)->reason == SKB_REASON_CONSUMED))
4254 trace_consume_skb(skb);
4256 trace_kfree_skb(skb, net_tx_action);
4258 if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
4261 __kfree_skb_defer(skb);
4264 __kfree_skb_flush();
4267 if (sd->output_queue) {
4270 local_irq_disable();
4271 head = sd->output_queue;
4272 sd->output_queue = NULL;
4273 sd->output_queue_tailp = &sd->output_queue;
4277 struct Qdisc *q = head;
4278 spinlock_t *root_lock = NULL;
4280 head = head->next_sched;
4282 if (!(q->flags & TCQ_F_NOLOCK)) {
4283 root_lock = qdisc_lock(q);
4284 spin_lock(root_lock);
4286 /* We need to make sure head->next_sched is read
4287 * before clearing __QDISC_STATE_SCHED
4289 smp_mb__before_atomic();
4290 clear_bit(__QDISC_STATE_SCHED, &q->state);
4293 spin_unlock(root_lock);
4297 xfrm_dev_backlog(sd);
4300 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE)
4301 /* This hook is defined here for ATM LANE */
4302 int (*br_fdb_test_addr_hook)(struct net_device *dev,
4303 unsigned char *addr) __read_mostly;
4304 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook);
4307 static inline struct sk_buff *
4308 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
4309 struct net_device *orig_dev)
4311 #ifdef CONFIG_NET_CLS_ACT
4312 struct mini_Qdisc *miniq = rcu_dereference_bh(skb->dev->miniq_ingress);
4313 struct tcf_result cl_res;
4315 /* If there's at least one ingress present somewhere (so
4316 * we get here via enabled static key), remaining devices
4317 * that are not configured with an ingress qdisc will bail
4324 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4328 qdisc_skb_cb(skb)->pkt_len = skb->len;
4329 skb->tc_at_ingress = 1;
4330 mini_qdisc_bstats_cpu_update(miniq, skb);
4332 switch (tcf_classify(skb, miniq->filter_list, &cl_res, false)) {
4334 case TC_ACT_RECLASSIFY:
4335 skb->tc_index = TC_H_MIN(cl_res.classid);
4338 mini_qdisc_qstats_cpu_drop(miniq);
4346 case TC_ACT_REDIRECT:
4347 /* skb_mac_header check was done by cls/act_bpf, so
4348 * we can safely push the L2 header back before
4349 * redirecting to another netdev
4351 __skb_push(skb, skb->mac_len);
4352 skb_do_redirect(skb);
4357 #endif /* CONFIG_NET_CLS_ACT */
4362 * netdev_is_rx_handler_busy - check if receive handler is registered
4363 * @dev: device to check
4365 * Check if a receive handler is already registered for a given device.
4366 * Return true if there one.
4368 * The caller must hold the rtnl_mutex.
4370 bool netdev_is_rx_handler_busy(struct net_device *dev)
4373 return dev && rtnl_dereference(dev->rx_handler);
4375 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy);
4378 * netdev_rx_handler_register - register receive handler
4379 * @dev: device to register a handler for
4380 * @rx_handler: receive handler to register
4381 * @rx_handler_data: data pointer that is used by rx handler
4383 * Register a receive handler for a device. This handler will then be
4384 * called from __netif_receive_skb. A negative errno code is returned
4387 * The caller must hold the rtnl_mutex.
4389 * For a general description of rx_handler, see enum rx_handler_result.
4391 int netdev_rx_handler_register(struct net_device *dev,
4392 rx_handler_func_t *rx_handler,
4393 void *rx_handler_data)
4395 if (netdev_is_rx_handler_busy(dev))
4398 if (dev->priv_flags & IFF_NO_RX_HANDLER)
4401 /* Note: rx_handler_data must be set before rx_handler */
4402 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data);
4403 rcu_assign_pointer(dev->rx_handler, rx_handler);
4407 EXPORT_SYMBOL_GPL(netdev_rx_handler_register);
4410 * netdev_rx_handler_unregister - unregister receive handler
4411 * @dev: device to unregister a handler from
4413 * Unregister a receive handler from a device.
4415 * The caller must hold the rtnl_mutex.
4417 void netdev_rx_handler_unregister(struct net_device *dev)
4421 RCU_INIT_POINTER(dev->rx_handler, NULL);
4422 /* a reader seeing a non NULL rx_handler in a rcu_read_lock()
4423 * section has a guarantee to see a non NULL rx_handler_data
4427 RCU_INIT_POINTER(dev->rx_handler_data, NULL);
4429 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister);
4432 * Limit the use of PFMEMALLOC reserves to those protocols that implement
4433 * the special handling of PFMEMALLOC skbs.
4435 static bool skb_pfmemalloc_protocol(struct sk_buff *skb)
4437 switch (skb->protocol) {
4438 case htons(ETH_P_ARP):
4439 case htons(ETH_P_IP):
4440 case htons(ETH_P_IPV6):
4441 case htons(ETH_P_8021Q):
4442 case htons(ETH_P_8021AD):
4449 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev,
4450 int *ret, struct net_device *orig_dev)
4452 #ifdef CONFIG_NETFILTER_INGRESS
4453 if (nf_hook_ingress_active(skb)) {
4457 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4462 ingress_retval = nf_hook_ingress(skb);
4464 return ingress_retval;
4466 #endif /* CONFIG_NETFILTER_INGRESS */
4470 static int __netif_receive_skb_core(struct sk_buff *skb, bool pfmemalloc)
4472 struct packet_type *ptype, *pt_prev;
4473 rx_handler_func_t *rx_handler;
4474 struct net_device *orig_dev;
4475 bool deliver_exact = false;
4476 int ret = NET_RX_DROP;
4479 net_timestamp_check(!netdev_tstamp_prequeue, skb);
4481 trace_netif_receive_skb(skb);
4483 orig_dev = skb->dev;
4485 skb_reset_network_header(skb);
4486 if (!skb_transport_header_was_set(skb))
4487 skb_reset_transport_header(skb);
4488 skb_reset_mac_len(skb);
4493 skb->skb_iif = skb->dev->ifindex;
4495 __this_cpu_inc(softnet_data.processed);
4497 if (skb->protocol == cpu_to_be16(ETH_P_8021Q) ||
4498 skb->protocol == cpu_to_be16(ETH_P_8021AD)) {
4499 skb = skb_vlan_untag(skb);
4504 if (skb_skip_tc_classify(skb))
4510 list_for_each_entry_rcu(ptype, &ptype_all, list) {
4512 ret = deliver_skb(skb, pt_prev, orig_dev);
4516 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) {
4518 ret = deliver_skb(skb, pt_prev, orig_dev);
4523 #ifdef CONFIG_NET_INGRESS
4524 if (static_key_false(&ingress_needed)) {
4525 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev);
4529 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0)
4535 if (pfmemalloc && !skb_pfmemalloc_protocol(skb))
4538 if (skb_vlan_tag_present(skb)) {
4540 ret = deliver_skb(skb, pt_prev, orig_dev);
4543 if (vlan_do_receive(&skb))
4545 else if (unlikely(!skb))
4549 rx_handler = rcu_dereference(skb->dev->rx_handler);
4552 ret = deliver_skb(skb, pt_prev, orig_dev);
4555 switch (rx_handler(&skb)) {
4556 case RX_HANDLER_CONSUMED:
4557 ret = NET_RX_SUCCESS;
4559 case RX_HANDLER_ANOTHER:
4561 case RX_HANDLER_EXACT:
4562 deliver_exact = true;
4563 case RX_HANDLER_PASS:
4570 if (unlikely(skb_vlan_tag_present(skb))) {
4571 if (skb_vlan_tag_get_id(skb))
4572 skb->pkt_type = PACKET_OTHERHOST;
4573 /* Note: we might in the future use prio bits
4574 * and set skb->priority like in vlan_do_receive()
4575 * For the time being, just ignore Priority Code Point
4580 type = skb->protocol;
4582 /* deliver only exact match when indicated */
4583 if (likely(!deliver_exact)) {
4584 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4585 &ptype_base[ntohs(type) &
4589 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4590 &orig_dev->ptype_specific);
4592 if (unlikely(skb->dev != orig_dev)) {
4593 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4594 &skb->dev->ptype_specific);
4598 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
4601 ret = pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
4605 atomic_long_inc(&skb->dev->rx_dropped);
4607 atomic_long_inc(&skb->dev->rx_nohandler);
4609 /* Jamal, now you will not able to escape explaining
4610 * me how you were going to use this. :-)
4620 * netif_receive_skb_core - special purpose version of netif_receive_skb
4621 * @skb: buffer to process
4623 * More direct receive version of netif_receive_skb(). It should
4624 * only be used by callers that have a need to skip RPS and Generic XDP.
4625 * Caller must also take care of handling if (page_is_)pfmemalloc.
4627 * This function may only be called from softirq context and interrupts
4628 * should be enabled.
4630 * Return values (usually ignored):
4631 * NET_RX_SUCCESS: no congestion
4632 * NET_RX_DROP: packet was dropped
4634 int netif_receive_skb_core(struct sk_buff *skb)
4639 ret = __netif_receive_skb_core(skb, false);
4644 EXPORT_SYMBOL(netif_receive_skb_core);
4646 static int __netif_receive_skb(struct sk_buff *skb)
4650 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) {
4651 unsigned int noreclaim_flag;
4654 * PFMEMALLOC skbs are special, they should
4655 * - be delivered to SOCK_MEMALLOC sockets only
4656 * - stay away from userspace
4657 * - have bounded memory usage
4659 * Use PF_MEMALLOC as this saves us from propagating the allocation
4660 * context down to all allocation sites.
4662 noreclaim_flag = memalloc_noreclaim_save();
4663 ret = __netif_receive_skb_core(skb, true);
4664 memalloc_noreclaim_restore(noreclaim_flag);
4666 ret = __netif_receive_skb_core(skb, false);
4671 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp)
4673 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog);
4674 struct bpf_prog *new = xdp->prog;
4677 switch (xdp->command) {
4678 case XDP_SETUP_PROG:
4679 rcu_assign_pointer(dev->xdp_prog, new);
4684 static_key_slow_dec(&generic_xdp_needed);
4685 } else if (new && !old) {
4686 static_key_slow_inc(&generic_xdp_needed);
4687 dev_disable_lro(dev);
4688 dev_disable_gro_hw(dev);
4692 case XDP_QUERY_PROG:
4693 xdp->prog_attached = !!old;
4694 xdp->prog_id = old ? old->aux->id : 0;
4705 static int netif_receive_skb_internal(struct sk_buff *skb)
4709 net_timestamp_check(netdev_tstamp_prequeue, skb);
4711 if (skb_defer_rx_timestamp(skb))
4712 return NET_RX_SUCCESS;
4714 if (static_key_false(&generic_xdp_needed)) {
4719 ret = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
4723 if (ret != XDP_PASS)
4729 if (static_key_false(&rps_needed)) {
4730 struct rps_dev_flow voidflow, *rflow = &voidflow;
4731 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
4734 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4740 ret = __netif_receive_skb(skb);
4746 * netif_receive_skb - process receive buffer from network
4747 * @skb: buffer to process
4749 * netif_receive_skb() is the main receive data processing function.
4750 * It always succeeds. The buffer may be dropped during processing
4751 * for congestion control or by the protocol layers.
4753 * This function may only be called from softirq context and interrupts
4754 * should be enabled.
4756 * Return values (usually ignored):
4757 * NET_RX_SUCCESS: no congestion
4758 * NET_RX_DROP: packet was dropped
4760 int netif_receive_skb(struct sk_buff *skb)
4762 trace_netif_receive_skb_entry(skb);
4764 return netif_receive_skb_internal(skb);
4766 EXPORT_SYMBOL(netif_receive_skb);
4768 DEFINE_PER_CPU(struct work_struct, flush_works);
4770 /* Network device is going away, flush any packets still pending */
4771 static void flush_backlog(struct work_struct *work)
4773 struct sk_buff *skb, *tmp;
4774 struct softnet_data *sd;
4777 sd = this_cpu_ptr(&softnet_data);
4779 local_irq_disable();
4781 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) {
4782 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
4783 __skb_unlink(skb, &sd->input_pkt_queue);
4785 input_queue_head_incr(sd);
4791 skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
4792 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
4793 __skb_unlink(skb, &sd->process_queue);
4795 input_queue_head_incr(sd);
4801 static void flush_all_backlogs(void)
4807 for_each_online_cpu(cpu)
4808 queue_work_on(cpu, system_highpri_wq,
4809 per_cpu_ptr(&flush_works, cpu));
4811 for_each_online_cpu(cpu)
4812 flush_work(per_cpu_ptr(&flush_works, cpu));
4817 static int napi_gro_complete(struct sk_buff *skb)
4819 struct packet_offload *ptype;
4820 __be16 type = skb->protocol;
4821 struct list_head *head = &offload_base;
4824 BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
4826 if (NAPI_GRO_CB(skb)->count == 1) {
4827 skb_shinfo(skb)->gso_size = 0;
4832 list_for_each_entry_rcu(ptype, head, list) {
4833 if (ptype->type != type || !ptype->callbacks.gro_complete)
4836 err = ptype->callbacks.gro_complete(skb, 0);
4842 WARN_ON(&ptype->list == head);
4844 return NET_RX_SUCCESS;
4848 return netif_receive_skb_internal(skb);
4851 /* napi->gro_list contains packets ordered by age.
4852 * youngest packets at the head of it.
4853 * Complete skbs in reverse order to reduce latencies.
4855 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
4857 struct sk_buff *skb, *prev = NULL;
4859 /* scan list and build reverse chain */
4860 for (skb = napi->gro_list; skb != NULL; skb = skb->next) {
4865 for (skb = prev; skb; skb = prev) {
4868 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
4872 napi_gro_complete(skb);
4876 napi->gro_list = NULL;
4878 EXPORT_SYMBOL(napi_gro_flush);
4880 static void gro_list_prepare(struct napi_struct *napi, struct sk_buff *skb)
4883 unsigned int maclen = skb->dev->hard_header_len;
4884 u32 hash = skb_get_hash_raw(skb);
4886 for (p = napi->gro_list; p; p = p->next) {
4887 unsigned long diffs;
4889 NAPI_GRO_CB(p)->flush = 0;
4891 if (hash != skb_get_hash_raw(p)) {
4892 NAPI_GRO_CB(p)->same_flow = 0;
4896 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
4897 diffs |= p->vlan_tci ^ skb->vlan_tci;
4898 diffs |= skb_metadata_dst_cmp(p, skb);
4899 diffs |= skb_metadata_differs(p, skb);
4900 if (maclen == ETH_HLEN)
4901 diffs |= compare_ether_header(skb_mac_header(p),
4902 skb_mac_header(skb));
4904 diffs = memcmp(skb_mac_header(p),
4905 skb_mac_header(skb),
4907 NAPI_GRO_CB(p)->same_flow = !diffs;
4911 static void skb_gro_reset_offset(struct sk_buff *skb)
4913 const struct skb_shared_info *pinfo = skb_shinfo(skb);
4914 const skb_frag_t *frag0 = &pinfo->frags[0];
4916 NAPI_GRO_CB(skb)->data_offset = 0;
4917 NAPI_GRO_CB(skb)->frag0 = NULL;
4918 NAPI_GRO_CB(skb)->frag0_len = 0;
4920 if (skb_mac_header(skb) == skb_tail_pointer(skb) &&
4922 !PageHighMem(skb_frag_page(frag0))) {
4923 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
4924 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
4925 skb_frag_size(frag0),
4926 skb->end - skb->tail);
4930 static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
4932 struct skb_shared_info *pinfo = skb_shinfo(skb);
4934 BUG_ON(skb->end - skb->tail < grow);
4936 memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
4938 skb->data_len -= grow;
4941 pinfo->frags[0].page_offset += grow;
4942 skb_frag_size_sub(&pinfo->frags[0], grow);
4944 if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
4945 skb_frag_unref(skb, 0);
4946 memmove(pinfo->frags, pinfo->frags + 1,
4947 --pinfo->nr_frags * sizeof(pinfo->frags[0]));
4951 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
4953 struct sk_buff **pp = NULL;
4954 struct packet_offload *ptype;
4955 __be16 type = skb->protocol;
4956 struct list_head *head = &offload_base;
4958 enum gro_result ret;
4961 if (netif_elide_gro(skb->dev))
4964 gro_list_prepare(napi, skb);
4967 list_for_each_entry_rcu(ptype, head, list) {
4968 if (ptype->type != type || !ptype->callbacks.gro_receive)
4971 skb_set_network_header(skb, skb_gro_offset(skb));
4972 skb_reset_mac_len(skb);
4973 NAPI_GRO_CB(skb)->same_flow = 0;
4974 NAPI_GRO_CB(skb)->flush = skb_is_gso(skb) || skb_has_frag_list(skb);
4975 NAPI_GRO_CB(skb)->free = 0;
4976 NAPI_GRO_CB(skb)->encap_mark = 0;
4977 NAPI_GRO_CB(skb)->recursion_counter = 0;
4978 NAPI_GRO_CB(skb)->is_fou = 0;
4979 NAPI_GRO_CB(skb)->is_atomic = 1;
4980 NAPI_GRO_CB(skb)->gro_remcsum_start = 0;
4982 /* Setup for GRO checksum validation */
4983 switch (skb->ip_summed) {
4984 case CHECKSUM_COMPLETE:
4985 NAPI_GRO_CB(skb)->csum = skb->csum;
4986 NAPI_GRO_CB(skb)->csum_valid = 1;
4987 NAPI_GRO_CB(skb)->csum_cnt = 0;
4989 case CHECKSUM_UNNECESSARY:
4990 NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
4991 NAPI_GRO_CB(skb)->csum_valid = 0;
4994 NAPI_GRO_CB(skb)->csum_cnt = 0;
4995 NAPI_GRO_CB(skb)->csum_valid = 0;
4998 pp = ptype->callbacks.gro_receive(&napi->gro_list, skb);
5003 if (&ptype->list == head)
5006 if (IS_ERR(pp) && PTR_ERR(pp) == -EINPROGRESS) {
5011 same_flow = NAPI_GRO_CB(skb)->same_flow;
5012 ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
5015 struct sk_buff *nskb = *pp;
5019 napi_gro_complete(nskb);
5026 if (NAPI_GRO_CB(skb)->flush)
5029 if (unlikely(napi->gro_count >= MAX_GRO_SKBS)) {
5030 struct sk_buff *nskb = napi->gro_list;
5032 /* locate the end of the list to select the 'oldest' flow */
5033 while (nskb->next) {
5039 napi_gro_complete(nskb);
5043 NAPI_GRO_CB(skb)->count = 1;
5044 NAPI_GRO_CB(skb)->age = jiffies;
5045 NAPI_GRO_CB(skb)->last = skb;
5046 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
5047 skb->next = napi->gro_list;
5048 napi->gro_list = skb;
5052 grow = skb_gro_offset(skb) - skb_headlen(skb);
5054 gro_pull_from_frag0(skb, grow);
5063 struct packet_offload *gro_find_receive_by_type(__be16 type)
5065 struct list_head *offload_head = &offload_base;
5066 struct packet_offload *ptype;
5068 list_for_each_entry_rcu(ptype, offload_head, list) {
5069 if (ptype->type != type || !ptype->callbacks.gro_receive)
5075 EXPORT_SYMBOL(gro_find_receive_by_type);
5077 struct packet_offload *gro_find_complete_by_type(__be16 type)
5079 struct list_head *offload_head = &offload_base;
5080 struct packet_offload *ptype;
5082 list_for_each_entry_rcu(ptype, offload_head, list) {
5083 if (ptype->type != type || !ptype->callbacks.gro_complete)
5089 EXPORT_SYMBOL(gro_find_complete_by_type);
5091 static void napi_skb_free_stolen_head(struct sk_buff *skb)
5095 kmem_cache_free(skbuff_head_cache, skb);
5098 static gro_result_t napi_skb_finish(gro_result_t ret, struct sk_buff *skb)
5102 if (netif_receive_skb_internal(skb))
5110 case GRO_MERGED_FREE:
5111 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
5112 napi_skb_free_stolen_head(skb);
5126 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
5128 skb_mark_napi_id(skb, napi);
5129 trace_napi_gro_receive_entry(skb);
5131 skb_gro_reset_offset(skb);
5133 return napi_skb_finish(dev_gro_receive(napi, skb), skb);
5135 EXPORT_SYMBOL(napi_gro_receive);
5137 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
5139 if (unlikely(skb->pfmemalloc)) {
5143 __skb_pull(skb, skb_headlen(skb));
5144 /* restore the reserve we had after netdev_alloc_skb_ip_align() */
5145 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
5147 skb->dev = napi->dev;
5149 skb->encapsulation = 0;
5150 skb_shinfo(skb)->gso_type = 0;
5151 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
5157 struct sk_buff *napi_get_frags(struct napi_struct *napi)
5159 struct sk_buff *skb = napi->skb;
5162 skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
5165 skb_mark_napi_id(skb, napi);
5170 EXPORT_SYMBOL(napi_get_frags);
5172 static gro_result_t napi_frags_finish(struct napi_struct *napi,
5173 struct sk_buff *skb,
5179 __skb_push(skb, ETH_HLEN);
5180 skb->protocol = eth_type_trans(skb, skb->dev);
5181 if (ret == GRO_NORMAL && netif_receive_skb_internal(skb))
5186 napi_reuse_skb(napi, skb);
5189 case GRO_MERGED_FREE:
5190 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
5191 napi_skb_free_stolen_head(skb);
5193 napi_reuse_skb(napi, skb);
5204 /* Upper GRO stack assumes network header starts at gro_offset=0
5205 * Drivers could call both napi_gro_frags() and napi_gro_receive()
5206 * We copy ethernet header into skb->data to have a common layout.
5208 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
5210 struct sk_buff *skb = napi->skb;
5211 const struct ethhdr *eth;
5212 unsigned int hlen = sizeof(*eth);
5216 skb_reset_mac_header(skb);
5217 skb_gro_reset_offset(skb);
5219 eth = skb_gro_header_fast(skb, 0);
5220 if (unlikely(skb_gro_header_hard(skb, hlen))) {
5221 eth = skb_gro_header_slow(skb, hlen, 0);
5222 if (unlikely(!eth)) {
5223 net_warn_ratelimited("%s: dropping impossible skb from %s\n",
5224 __func__, napi->dev->name);
5225 napi_reuse_skb(napi, skb);
5229 gro_pull_from_frag0(skb, hlen);
5230 NAPI_GRO_CB(skb)->frag0 += hlen;
5231 NAPI_GRO_CB(skb)->frag0_len -= hlen;
5233 __skb_pull(skb, hlen);
5236 * This works because the only protocols we care about don't require
5238 * We'll fix it up properly in napi_frags_finish()
5240 skb->protocol = eth->h_proto;
5245 gro_result_t napi_gro_frags(struct napi_struct *napi)
5247 struct sk_buff *skb = napi_frags_skb(napi);
5252 trace_napi_gro_frags_entry(skb);
5254 return napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
5256 EXPORT_SYMBOL(napi_gro_frags);
5258 /* Compute the checksum from gro_offset and return the folded value
5259 * after adding in any pseudo checksum.
5261 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
5266 wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
5268 /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
5269 sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
5271 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
5272 !skb->csum_complete_sw)
5273 netdev_rx_csum_fault(skb->dev);
5276 NAPI_GRO_CB(skb)->csum = wsum;
5277 NAPI_GRO_CB(skb)->csum_valid = 1;
5281 EXPORT_SYMBOL(__skb_gro_checksum_complete);
5283 static void net_rps_send_ipi(struct softnet_data *remsd)
5287 struct softnet_data *next = remsd->rps_ipi_next;
5289 if (cpu_online(remsd->cpu))
5290 smp_call_function_single_async(remsd->cpu, &remsd->csd);
5297 * net_rps_action_and_irq_enable sends any pending IPI's for rps.
5298 * Note: called with local irq disabled, but exits with local irq enabled.
5300 static void net_rps_action_and_irq_enable(struct softnet_data *sd)
5303 struct softnet_data *remsd = sd->rps_ipi_list;
5306 sd->rps_ipi_list = NULL;
5310 /* Send pending IPI's to kick RPS processing on remote cpus. */
5311 net_rps_send_ipi(remsd);
5317 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd)
5320 return sd->rps_ipi_list != NULL;
5326 static int process_backlog(struct napi_struct *napi, int quota)
5328 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
5332 /* Check if we have pending ipi, its better to send them now,
5333 * not waiting net_rx_action() end.
5335 if (sd_has_rps_ipi_waiting(sd)) {
5336 local_irq_disable();
5337 net_rps_action_and_irq_enable(sd);
5340 napi->weight = dev_rx_weight;
5342 struct sk_buff *skb;
5344 while ((skb = __skb_dequeue(&sd->process_queue))) {
5346 __netif_receive_skb(skb);
5348 input_queue_head_incr(sd);
5349 if (++work >= quota)
5354 local_irq_disable();
5356 if (skb_queue_empty(&sd->input_pkt_queue)) {
5358 * Inline a custom version of __napi_complete().
5359 * only current cpu owns and manipulates this napi,
5360 * and NAPI_STATE_SCHED is the only possible flag set
5362 * We can use a plain write instead of clear_bit(),
5363 * and we dont need an smp_mb() memory barrier.
5368 skb_queue_splice_tail_init(&sd->input_pkt_queue,
5369 &sd->process_queue);
5379 * __napi_schedule - schedule for receive
5380 * @n: entry to schedule
5382 * The entry's receive function will be scheduled to run.
5383 * Consider using __napi_schedule_irqoff() if hard irqs are masked.
5385 void __napi_schedule(struct napi_struct *n)
5387 unsigned long flags;
5389 local_irq_save(flags);
5390 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
5391 local_irq_restore(flags);
5393 EXPORT_SYMBOL(__napi_schedule);
5396 * napi_schedule_prep - check if napi can be scheduled
5399 * Test if NAPI routine is already running, and if not mark
5400 * it as running. This is used as a condition variable
5401 * insure only one NAPI poll instance runs. We also make
5402 * sure there is no pending NAPI disable.
5404 bool napi_schedule_prep(struct napi_struct *n)
5406 unsigned long val, new;
5409 val = READ_ONCE(n->state);
5410 if (unlikely(val & NAPIF_STATE_DISABLE))
5412 new = val | NAPIF_STATE_SCHED;
5414 /* Sets STATE_MISSED bit if STATE_SCHED was already set
5415 * This was suggested by Alexander Duyck, as compiler
5416 * emits better code than :
5417 * if (val & NAPIF_STATE_SCHED)
5418 * new |= NAPIF_STATE_MISSED;
5420 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED *
5422 } while (cmpxchg(&n->state, val, new) != val);
5424 return !(val & NAPIF_STATE_SCHED);
5426 EXPORT_SYMBOL(napi_schedule_prep);
5429 * __napi_schedule_irqoff - schedule for receive
5430 * @n: entry to schedule
5432 * Variant of __napi_schedule() assuming hard irqs are masked
5434 void __napi_schedule_irqoff(struct napi_struct *n)
5436 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
5438 EXPORT_SYMBOL(__napi_schedule_irqoff);
5440 bool napi_complete_done(struct napi_struct *n, int work_done)
5442 unsigned long flags, val, new;
5445 * 1) Don't let napi dequeue from the cpu poll list
5446 * just in case its running on a different cpu.
5447 * 2) If we are busy polling, do nothing here, we have
5448 * the guarantee we will be called later.
5450 if (unlikely(n->state & (NAPIF_STATE_NPSVC |
5451 NAPIF_STATE_IN_BUSY_POLL)))
5455 unsigned long timeout = 0;
5458 timeout = n->dev->gro_flush_timeout;
5461 hrtimer_start(&n->timer, ns_to_ktime(timeout),
5462 HRTIMER_MODE_REL_PINNED);
5464 napi_gro_flush(n, false);
5466 if (unlikely(!list_empty(&n->poll_list))) {
5467 /* If n->poll_list is not empty, we need to mask irqs */
5468 local_irq_save(flags);
5469 list_del_init(&n->poll_list);
5470 local_irq_restore(flags);
5474 val = READ_ONCE(n->state);
5476 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED));
5478 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED);
5480 /* If STATE_MISSED was set, leave STATE_SCHED set,
5481 * because we will call napi->poll() one more time.
5482 * This C code was suggested by Alexander Duyck to help gcc.
5484 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED *
5486 } while (cmpxchg(&n->state, val, new) != val);
5488 if (unlikely(val & NAPIF_STATE_MISSED)) {
5495 EXPORT_SYMBOL(napi_complete_done);
5497 /* must be called under rcu_read_lock(), as we dont take a reference */
5498 static struct napi_struct *napi_by_id(unsigned int napi_id)
5500 unsigned int hash = napi_id % HASH_SIZE(napi_hash);
5501 struct napi_struct *napi;
5503 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node)
5504 if (napi->napi_id == napi_id)
5510 #if defined(CONFIG_NET_RX_BUSY_POLL)
5512 #define BUSY_POLL_BUDGET 8
5514 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock)
5518 /* Busy polling means there is a high chance device driver hard irq
5519 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was
5520 * set in napi_schedule_prep().
5521 * Since we are about to call napi->poll() once more, we can safely
5522 * clear NAPI_STATE_MISSED.
5524 * Note: x86 could use a single "lock and ..." instruction
5525 * to perform these two clear_bit()
5527 clear_bit(NAPI_STATE_MISSED, &napi->state);
5528 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state);
5532 /* All we really want here is to re-enable device interrupts.
5533 * Ideally, a new ndo_busy_poll_stop() could avoid another round.
5535 rc = napi->poll(napi, BUSY_POLL_BUDGET);
5536 trace_napi_poll(napi, rc, BUSY_POLL_BUDGET);
5537 netpoll_poll_unlock(have_poll_lock);
5538 if (rc == BUSY_POLL_BUDGET)
5539 __napi_schedule(napi);
5543 void napi_busy_loop(unsigned int napi_id,
5544 bool (*loop_end)(void *, unsigned long),
5547 unsigned long start_time = loop_end ? busy_loop_current_time() : 0;
5548 int (*napi_poll)(struct napi_struct *napi, int budget);
5549 void *have_poll_lock = NULL;
5550 struct napi_struct *napi;
5557 napi = napi_by_id(napi_id);
5567 unsigned long val = READ_ONCE(napi->state);
5569 /* If multiple threads are competing for this napi,
5570 * we avoid dirtying napi->state as much as we can.
5572 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED |
5573 NAPIF_STATE_IN_BUSY_POLL))
5575 if (cmpxchg(&napi->state, val,
5576 val | NAPIF_STATE_IN_BUSY_POLL |
5577 NAPIF_STATE_SCHED) != val)
5579 have_poll_lock = netpoll_poll_lock(napi);
5580 napi_poll = napi->poll;
5582 work = napi_poll(napi, BUSY_POLL_BUDGET);
5583 trace_napi_poll(napi, work, BUSY_POLL_BUDGET);
5586 __NET_ADD_STATS(dev_net(napi->dev),
5587 LINUX_MIB_BUSYPOLLRXPACKETS, work);
5590 if (!loop_end || loop_end(loop_end_arg, start_time))
5593 if (unlikely(need_resched())) {
5595 busy_poll_stop(napi, have_poll_lock);
5599 if (loop_end(loop_end_arg, start_time))
5606 busy_poll_stop(napi, have_poll_lock);
5611 EXPORT_SYMBOL(napi_busy_loop);
5613 #endif /* CONFIG_NET_RX_BUSY_POLL */
5615 static void napi_hash_add(struct napi_struct *napi)
5617 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state) ||
5618 test_and_set_bit(NAPI_STATE_HASHED, &napi->state))
5621 spin_lock(&napi_hash_lock);
5623 /* 0..NR_CPUS range is reserved for sender_cpu use */
5625 if (unlikely(++napi_gen_id < MIN_NAPI_ID))
5626 napi_gen_id = MIN_NAPI_ID;
5627 } while (napi_by_id(napi_gen_id));
5628 napi->napi_id = napi_gen_id;
5630 hlist_add_head_rcu(&napi->napi_hash_node,
5631 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]);
5633 spin_unlock(&napi_hash_lock);
5636 /* Warning : caller is responsible to make sure rcu grace period
5637 * is respected before freeing memory containing @napi
5639 bool napi_hash_del(struct napi_struct *napi)
5641 bool rcu_sync_needed = false;
5643 spin_lock(&napi_hash_lock);
5645 if (test_and_clear_bit(NAPI_STATE_HASHED, &napi->state)) {
5646 rcu_sync_needed = true;
5647 hlist_del_rcu(&napi->napi_hash_node);
5649 spin_unlock(&napi_hash_lock);
5650 return rcu_sync_needed;
5652 EXPORT_SYMBOL_GPL(napi_hash_del);
5654 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer)
5656 struct napi_struct *napi;
5658 napi = container_of(timer, struct napi_struct, timer);
5660 /* Note : we use a relaxed variant of napi_schedule_prep() not setting
5661 * NAPI_STATE_MISSED, since we do not react to a device IRQ.
5663 if (napi->gro_list && !napi_disable_pending(napi) &&
5664 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state))
5665 __napi_schedule_irqoff(napi);
5667 return HRTIMER_NORESTART;
5670 void netif_napi_add(struct net_device *dev, struct napi_struct *napi,
5671 int (*poll)(struct napi_struct *, int), int weight)
5673 INIT_LIST_HEAD(&napi->poll_list);
5674 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
5675 napi->timer.function = napi_watchdog;
5676 napi->gro_count = 0;
5677 napi->gro_list = NULL;
5680 if (weight > NAPI_POLL_WEIGHT)
5681 pr_err_once("netif_napi_add() called with weight %d on device %s\n",
5683 napi->weight = weight;
5684 list_add(&napi->dev_list, &dev->napi_list);
5686 #ifdef CONFIG_NETPOLL
5687 napi->poll_owner = -1;
5689 set_bit(NAPI_STATE_SCHED, &napi->state);
5690 napi_hash_add(napi);
5692 EXPORT_SYMBOL(netif_napi_add);
5694 void napi_disable(struct napi_struct *n)
5697 set_bit(NAPI_STATE_DISABLE, &n->state);
5699 while (test_and_set_bit(NAPI_STATE_SCHED, &n->state))
5701 while (test_and_set_bit(NAPI_STATE_NPSVC, &n->state))
5704 hrtimer_cancel(&n->timer);
5706 clear_bit(NAPI_STATE_DISABLE, &n->state);
5708 EXPORT_SYMBOL(napi_disable);
5710 /* Must be called in process context */
5711 void netif_napi_del(struct napi_struct *napi)
5714 if (napi_hash_del(napi))
5716 list_del_init(&napi->dev_list);
5717 napi_free_frags(napi);
5719 kfree_skb_list(napi->gro_list);
5720 napi->gro_list = NULL;
5721 napi->gro_count = 0;
5723 EXPORT_SYMBOL(netif_napi_del);
5725 static int napi_poll(struct napi_struct *n, struct list_head *repoll)
5730 list_del_init(&n->poll_list);
5732 have = netpoll_poll_lock(n);
5736 /* This NAPI_STATE_SCHED test is for avoiding a race
5737 * with netpoll's poll_napi(). Only the entity which
5738 * obtains the lock and sees NAPI_STATE_SCHED set will
5739 * actually make the ->poll() call. Therefore we avoid
5740 * accidentally calling ->poll() when NAPI is not scheduled.
5743 if (test_bit(NAPI_STATE_SCHED, &n->state)) {
5744 work = n->poll(n, weight);
5745 trace_napi_poll(n, work, weight);
5748 WARN_ON_ONCE(work > weight);
5750 if (likely(work < weight))
5753 /* Drivers must not modify the NAPI state if they
5754 * consume the entire weight. In such cases this code
5755 * still "owns" the NAPI instance and therefore can
5756 * move the instance around on the list at-will.
5758 if (unlikely(napi_disable_pending(n))) {
5764 /* flush too old packets
5765 * If HZ < 1000, flush all packets.
5767 napi_gro_flush(n, HZ >= 1000);
5770 /* Some drivers may have called napi_schedule
5771 * prior to exhausting their budget.
5773 if (unlikely(!list_empty(&n->poll_list))) {
5774 pr_warn_once("%s: Budget exhausted after napi rescheduled\n",
5775 n->dev ? n->dev->name : "backlog");
5779 list_add_tail(&n->poll_list, repoll);
5782 netpoll_poll_unlock(have);
5787 static __latent_entropy void net_rx_action(struct softirq_action *h)
5789 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
5790 unsigned long time_limit = jiffies +
5791 usecs_to_jiffies(netdev_budget_usecs);
5792 int budget = netdev_budget;
5796 local_irq_disable();
5797 list_splice_init(&sd->poll_list, &list);
5801 struct napi_struct *n;
5803 if (list_empty(&list)) {
5804 if (!sd_has_rps_ipi_waiting(sd) && list_empty(&repoll))
5809 n = list_first_entry(&list, struct napi_struct, poll_list);
5810 budget -= napi_poll(n, &repoll);
5812 /* If softirq window is exhausted then punt.
5813 * Allow this to run for 2 jiffies since which will allow
5814 * an average latency of 1.5/HZ.
5816 if (unlikely(budget <= 0 ||
5817 time_after_eq(jiffies, time_limit))) {
5823 local_irq_disable();
5825 list_splice_tail_init(&sd->poll_list, &list);
5826 list_splice_tail(&repoll, &list);
5827 list_splice(&list, &sd->poll_list);
5828 if (!list_empty(&sd->poll_list))
5829 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
5831 net_rps_action_and_irq_enable(sd);
5833 __kfree_skb_flush();
5836 struct netdev_adjacent {
5837 struct net_device *dev;
5839 /* upper master flag, there can only be one master device per list */
5842 /* counter for the number of times this device was added to us */
5845 /* private field for the users */
5848 struct list_head list;
5849 struct rcu_head rcu;
5852 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev,
5853 struct list_head *adj_list)
5855 struct netdev_adjacent *adj;
5857 list_for_each_entry(adj, adj_list, list) {
5858 if (adj->dev == adj_dev)
5864 static int __netdev_has_upper_dev(struct net_device *upper_dev, void *data)
5866 struct net_device *dev = data;
5868 return upper_dev == dev;
5872 * netdev_has_upper_dev - Check if device is linked to an upper device
5874 * @upper_dev: upper device to check
5876 * Find out if a device is linked to specified upper device and return true
5877 * in case it is. Note that this checks only immediate upper device,
5878 * not through a complete stack of devices. The caller must hold the RTNL lock.
5880 bool netdev_has_upper_dev(struct net_device *dev,
5881 struct net_device *upper_dev)
5885 return netdev_walk_all_upper_dev_rcu(dev, __netdev_has_upper_dev,
5888 EXPORT_SYMBOL(netdev_has_upper_dev);
5891 * netdev_has_upper_dev_all - Check if device is linked to an upper device
5893 * @upper_dev: upper device to check
5895 * Find out if a device is linked to specified upper device and return true
5896 * in case it is. Note that this checks the entire upper device chain.
5897 * The caller must hold rcu lock.
5900 bool netdev_has_upper_dev_all_rcu(struct net_device *dev,
5901 struct net_device *upper_dev)
5903 return !!netdev_walk_all_upper_dev_rcu(dev, __netdev_has_upper_dev,
5906 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu);
5909 * netdev_has_any_upper_dev - Check if device is linked to some device
5912 * Find out if a device is linked to an upper device and return true in case
5913 * it is. The caller must hold the RTNL lock.
5915 bool netdev_has_any_upper_dev(struct net_device *dev)
5919 return !list_empty(&dev->adj_list.upper);
5921 EXPORT_SYMBOL(netdev_has_any_upper_dev);
5924 * netdev_master_upper_dev_get - Get master upper device
5927 * Find a master upper device and return pointer to it or NULL in case
5928 * it's not there. The caller must hold the RTNL lock.
5930 struct net_device *netdev_master_upper_dev_get(struct net_device *dev)
5932 struct netdev_adjacent *upper;
5936 if (list_empty(&dev->adj_list.upper))
5939 upper = list_first_entry(&dev->adj_list.upper,
5940 struct netdev_adjacent, list);
5941 if (likely(upper->master))
5945 EXPORT_SYMBOL(netdev_master_upper_dev_get);
5948 * netdev_has_any_lower_dev - Check if device is linked to some device
5951 * Find out if a device is linked to a lower device and return true in case
5952 * it is. The caller must hold the RTNL lock.
5954 static bool netdev_has_any_lower_dev(struct net_device *dev)
5958 return !list_empty(&dev->adj_list.lower);
5961 void *netdev_adjacent_get_private(struct list_head *adj_list)
5963 struct netdev_adjacent *adj;
5965 adj = list_entry(adj_list, struct netdev_adjacent, list);
5967 return adj->private;
5969 EXPORT_SYMBOL(netdev_adjacent_get_private);
5972 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list
5974 * @iter: list_head ** of the current position
5976 * Gets the next device from the dev's upper list, starting from iter
5977 * position. The caller must hold RCU read lock.
5979 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev,
5980 struct list_head **iter)
5982 struct netdev_adjacent *upper;
5984 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
5986 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
5988 if (&upper->list == &dev->adj_list.upper)
5991 *iter = &upper->list;
5995 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu);
5997 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev,
5998 struct list_head **iter)
6000 struct netdev_adjacent *upper;
6002 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
6004 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6006 if (&upper->list == &dev->adj_list.upper)
6009 *iter = &upper->list;
6014 int netdev_walk_all_upper_dev_rcu(struct net_device *dev,
6015 int (*fn)(struct net_device *dev,
6019 struct net_device *udev;
6020 struct list_head *iter;
6023 for (iter = &dev->adj_list.upper,
6024 udev = netdev_next_upper_dev_rcu(dev, &iter);
6026 udev = netdev_next_upper_dev_rcu(dev, &iter)) {
6027 /* first is the upper device itself */
6028 ret = fn(udev, data);
6032 /* then look at all of its upper devices */
6033 ret = netdev_walk_all_upper_dev_rcu(udev, fn, data);
6040 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu);
6043 * netdev_lower_get_next_private - Get the next ->private from the
6044 * lower neighbour list
6046 * @iter: list_head ** of the current position
6048 * Gets the next netdev_adjacent->private from the dev's lower neighbour
6049 * list, starting from iter position. The caller must hold either hold the
6050 * RTNL lock or its own locking that guarantees that the neighbour lower
6051 * list will remain unchanged.
6053 void *netdev_lower_get_next_private(struct net_device *dev,
6054 struct list_head **iter)
6056 struct netdev_adjacent *lower;
6058 lower = list_entry(*iter, struct netdev_adjacent, list);
6060 if (&lower->list == &dev->adj_list.lower)
6063 *iter = lower->list.next;
6065 return lower->private;
6067 EXPORT_SYMBOL(netdev_lower_get_next_private);
6070 * netdev_lower_get_next_private_rcu - Get the next ->private from the
6071 * lower neighbour list, RCU
6074 * @iter: list_head ** of the current position
6076 * Gets the next netdev_adjacent->private from the dev's lower neighbour
6077 * list, starting from iter position. The caller must hold RCU read lock.
6079 void *netdev_lower_get_next_private_rcu(struct net_device *dev,
6080 struct list_head **iter)
6082 struct netdev_adjacent *lower;
6084 WARN_ON_ONCE(!rcu_read_lock_held());
6086 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6088 if (&lower->list == &dev->adj_list.lower)
6091 *iter = &lower->list;
6093 return lower->private;
6095 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu);
6098 * netdev_lower_get_next - Get the next device from the lower neighbour
6101 * @iter: list_head ** of the current position
6103 * Gets the next netdev_adjacent from the dev's lower neighbour
6104 * list, starting from iter position. The caller must hold RTNL lock or
6105 * its own locking that guarantees that the neighbour lower
6106 * list will remain unchanged.
6108 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter)
6110 struct netdev_adjacent *lower;
6112 lower = list_entry(*iter, struct netdev_adjacent, list);
6114 if (&lower->list == &dev->adj_list.lower)
6117 *iter = lower->list.next;
6121 EXPORT_SYMBOL(netdev_lower_get_next);
6123 static struct net_device *netdev_next_lower_dev(struct net_device *dev,
6124 struct list_head **iter)
6126 struct netdev_adjacent *lower;
6128 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
6130 if (&lower->list == &dev->adj_list.lower)
6133 *iter = &lower->list;
6138 int netdev_walk_all_lower_dev(struct net_device *dev,
6139 int (*fn)(struct net_device *dev,
6143 struct net_device *ldev;
6144 struct list_head *iter;
6147 for (iter = &dev->adj_list.lower,
6148 ldev = netdev_next_lower_dev(dev, &iter);
6150 ldev = netdev_next_lower_dev(dev, &iter)) {
6151 /* first is the lower device itself */
6152 ret = fn(ldev, data);
6156 /* then look at all of its lower devices */
6157 ret = netdev_walk_all_lower_dev(ldev, fn, data);
6164 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev);
6166 static struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev,
6167 struct list_head **iter)
6169 struct netdev_adjacent *lower;
6171 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6172 if (&lower->list == &dev->adj_list.lower)
6175 *iter = &lower->list;
6180 int netdev_walk_all_lower_dev_rcu(struct net_device *dev,
6181 int (*fn)(struct net_device *dev,
6185 struct net_device *ldev;
6186 struct list_head *iter;
6189 for (iter = &dev->adj_list.lower,
6190 ldev = netdev_next_lower_dev_rcu(dev, &iter);
6192 ldev = netdev_next_lower_dev_rcu(dev, &iter)) {
6193 /* first is the lower device itself */
6194 ret = fn(ldev, data);
6198 /* then look at all of its lower devices */
6199 ret = netdev_walk_all_lower_dev_rcu(ldev, fn, data);
6206 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu);
6209 * netdev_lower_get_first_private_rcu - Get the first ->private from the
6210 * lower neighbour list, RCU
6214 * Gets the first netdev_adjacent->private from the dev's lower neighbour
6215 * list. The caller must hold RCU read lock.
6217 void *netdev_lower_get_first_private_rcu(struct net_device *dev)
6219 struct netdev_adjacent *lower;
6221 lower = list_first_or_null_rcu(&dev->adj_list.lower,
6222 struct netdev_adjacent, list);
6224 return lower->private;
6227 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu);
6230 * netdev_master_upper_dev_get_rcu - Get master upper device
6233 * Find a master upper device and return pointer to it or NULL in case
6234 * it's not there. The caller must hold the RCU read lock.
6236 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev)
6238 struct netdev_adjacent *upper;
6240 upper = list_first_or_null_rcu(&dev->adj_list.upper,
6241 struct netdev_adjacent, list);
6242 if (upper && likely(upper->master))
6246 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu);
6248 static int netdev_adjacent_sysfs_add(struct net_device *dev,
6249 struct net_device *adj_dev,
6250 struct list_head *dev_list)
6252 char linkname[IFNAMSIZ+7];
6254 sprintf(linkname, dev_list == &dev->adj_list.upper ?
6255 "upper_%s" : "lower_%s", adj_dev->name);
6256 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj),
6259 static void netdev_adjacent_sysfs_del(struct net_device *dev,
6261 struct list_head *dev_list)
6263 char linkname[IFNAMSIZ+7];
6265 sprintf(linkname, dev_list == &dev->adj_list.upper ?
6266 "upper_%s" : "lower_%s", name);
6267 sysfs_remove_link(&(dev->dev.kobj), linkname);
6270 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev,
6271 struct net_device *adj_dev,
6272 struct list_head *dev_list)
6274 return (dev_list == &dev->adj_list.upper ||
6275 dev_list == &dev->adj_list.lower) &&
6276 net_eq(dev_net(dev), dev_net(adj_dev));
6279 static int __netdev_adjacent_dev_insert(struct net_device *dev,
6280 struct net_device *adj_dev,
6281 struct list_head *dev_list,
6282 void *private, bool master)
6284 struct netdev_adjacent *adj;
6287 adj = __netdev_find_adj(adj_dev, dev_list);
6291 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n",
6292 dev->name, adj_dev->name, adj->ref_nr);
6297 adj = kmalloc(sizeof(*adj), GFP_KERNEL);
6302 adj->master = master;
6304 adj->private = private;
6307 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n",
6308 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name);
6310 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) {
6311 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list);
6316 /* Ensure that master link is always the first item in list. */
6318 ret = sysfs_create_link(&(dev->dev.kobj),
6319 &(adj_dev->dev.kobj), "master");
6321 goto remove_symlinks;
6323 list_add_rcu(&adj->list, dev_list);
6325 list_add_tail_rcu(&adj->list, dev_list);
6331 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
6332 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
6340 static void __netdev_adjacent_dev_remove(struct net_device *dev,
6341 struct net_device *adj_dev,
6343 struct list_head *dev_list)
6345 struct netdev_adjacent *adj;
6347 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n",
6348 dev->name, adj_dev->name, ref_nr);
6350 adj = __netdev_find_adj(adj_dev, dev_list);
6353 pr_err("Adjacency does not exist for device %s from %s\n",
6354 dev->name, adj_dev->name);
6359 if (adj->ref_nr > ref_nr) {
6360 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n",
6361 dev->name, adj_dev->name, ref_nr,
6362 adj->ref_nr - ref_nr);
6363 adj->ref_nr -= ref_nr;
6368 sysfs_remove_link(&(dev->dev.kobj), "master");
6370 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
6371 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
6373 list_del_rcu(&adj->list);
6374 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n",
6375 adj_dev->name, dev->name, adj_dev->name);
6377 kfree_rcu(adj, rcu);
6380 static int __netdev_adjacent_dev_link_lists(struct net_device *dev,
6381 struct net_device *upper_dev,
6382 struct list_head *up_list,
6383 struct list_head *down_list,
6384 void *private, bool master)
6388 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list,
6393 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list,
6396 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list);
6403 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev,
6404 struct net_device *upper_dev,
6406 struct list_head *up_list,
6407 struct list_head *down_list)
6409 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list);
6410 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list);
6413 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev,
6414 struct net_device *upper_dev,
6415 void *private, bool master)
6417 return __netdev_adjacent_dev_link_lists(dev, upper_dev,
6418 &dev->adj_list.upper,
6419 &upper_dev->adj_list.lower,
6423 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev,
6424 struct net_device *upper_dev)
6426 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1,
6427 &dev->adj_list.upper,
6428 &upper_dev->adj_list.lower);
6431 static int __netdev_upper_dev_link(struct net_device *dev,
6432 struct net_device *upper_dev, bool master,
6433 void *upper_priv, void *upper_info,
6434 struct netlink_ext_ack *extack)
6436 struct netdev_notifier_changeupper_info changeupper_info = {
6441 .upper_dev = upper_dev,
6444 .upper_info = upper_info,
6446 struct net_device *master_dev;
6451 if (dev == upper_dev)
6454 /* To prevent loops, check if dev is not upper device to upper_dev. */
6455 if (netdev_has_upper_dev(upper_dev, dev))
6459 if (netdev_has_upper_dev(dev, upper_dev))
6462 master_dev = netdev_master_upper_dev_get(dev);
6464 return master_dev == upper_dev ? -EEXIST : -EBUSY;
6467 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
6468 &changeupper_info.info);
6469 ret = notifier_to_errno(ret);
6473 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv,
6478 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
6479 &changeupper_info.info);
6480 ret = notifier_to_errno(ret);
6487 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
6493 * netdev_upper_dev_link - Add a link to the upper device
6495 * @upper_dev: new upper device
6496 * @extack: netlink extended ack
6498 * Adds a link to device which is upper to this one. The caller must hold
6499 * the RTNL lock. On a failure a negative errno code is returned.
6500 * On success the reference counts are adjusted and the function
6503 int netdev_upper_dev_link(struct net_device *dev,
6504 struct net_device *upper_dev,
6505 struct netlink_ext_ack *extack)
6507 return __netdev_upper_dev_link(dev, upper_dev, false,
6508 NULL, NULL, extack);
6510 EXPORT_SYMBOL(netdev_upper_dev_link);
6513 * netdev_master_upper_dev_link - Add a master link to the upper device
6515 * @upper_dev: new upper device
6516 * @upper_priv: upper device private
6517 * @upper_info: upper info to be passed down via notifier
6518 * @extack: netlink extended ack
6520 * Adds a link to device which is upper to this one. In this case, only
6521 * one master upper device can be linked, although other non-master devices
6522 * might be linked as well. The caller must hold the RTNL lock.
6523 * On a failure a negative errno code is returned. On success the reference
6524 * counts are adjusted and the function returns zero.
6526 int netdev_master_upper_dev_link(struct net_device *dev,
6527 struct net_device *upper_dev,
6528 void *upper_priv, void *upper_info,
6529 struct netlink_ext_ack *extack)
6531 return __netdev_upper_dev_link(dev, upper_dev, true,
6532 upper_priv, upper_info, extack);
6534 EXPORT_SYMBOL(netdev_master_upper_dev_link);
6537 * netdev_upper_dev_unlink - Removes a link to upper device
6539 * @upper_dev: new upper device
6541 * Removes a link to device which is upper to this one. The caller must hold
6544 void netdev_upper_dev_unlink(struct net_device *dev,
6545 struct net_device *upper_dev)
6547 struct netdev_notifier_changeupper_info changeupper_info = {
6551 .upper_dev = upper_dev,
6557 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev;
6559 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
6560 &changeupper_info.info);
6562 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
6564 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
6565 &changeupper_info.info);
6567 EXPORT_SYMBOL(netdev_upper_dev_unlink);
6570 * netdev_bonding_info_change - Dispatch event about slave change
6572 * @bonding_info: info to dispatch
6574 * Send NETDEV_BONDING_INFO to netdev notifiers with info.
6575 * The caller must hold the RTNL lock.
6577 void netdev_bonding_info_change(struct net_device *dev,
6578 struct netdev_bonding_info *bonding_info)
6580 struct netdev_notifier_bonding_info info = {
6584 memcpy(&info.bonding_info, bonding_info,
6585 sizeof(struct netdev_bonding_info));
6586 call_netdevice_notifiers_info(NETDEV_BONDING_INFO,
6589 EXPORT_SYMBOL(netdev_bonding_info_change);
6591 static void netdev_adjacent_add_links(struct net_device *dev)
6593 struct netdev_adjacent *iter;
6595 struct net *net = dev_net(dev);
6597 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6598 if (!net_eq(net, dev_net(iter->dev)))
6600 netdev_adjacent_sysfs_add(iter->dev, dev,
6601 &iter->dev->adj_list.lower);
6602 netdev_adjacent_sysfs_add(dev, iter->dev,
6603 &dev->adj_list.upper);
6606 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6607 if (!net_eq(net, dev_net(iter->dev)))
6609 netdev_adjacent_sysfs_add(iter->dev, dev,
6610 &iter->dev->adj_list.upper);
6611 netdev_adjacent_sysfs_add(dev, iter->dev,
6612 &dev->adj_list.lower);
6616 static void netdev_adjacent_del_links(struct net_device *dev)
6618 struct netdev_adjacent *iter;
6620 struct net *net = dev_net(dev);
6622 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6623 if (!net_eq(net, dev_net(iter->dev)))
6625 netdev_adjacent_sysfs_del(iter->dev, dev->name,
6626 &iter->dev->adj_list.lower);
6627 netdev_adjacent_sysfs_del(dev, iter->dev->name,
6628 &dev->adj_list.upper);
6631 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6632 if (!net_eq(net, dev_net(iter->dev)))
6634 netdev_adjacent_sysfs_del(iter->dev, dev->name,
6635 &iter->dev->adj_list.upper);
6636 netdev_adjacent_sysfs_del(dev, iter->dev->name,
6637 &dev->adj_list.lower);
6641 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname)
6643 struct netdev_adjacent *iter;
6645 struct net *net = dev_net(dev);
6647 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6648 if (!net_eq(net, dev_net(iter->dev)))
6650 netdev_adjacent_sysfs_del(iter->dev, oldname,
6651 &iter->dev->adj_list.lower);
6652 netdev_adjacent_sysfs_add(iter->dev, dev,
6653 &iter->dev->adj_list.lower);
6656 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6657 if (!net_eq(net, dev_net(iter->dev)))
6659 netdev_adjacent_sysfs_del(iter->dev, oldname,
6660 &iter->dev->adj_list.upper);
6661 netdev_adjacent_sysfs_add(iter->dev, dev,
6662 &iter->dev->adj_list.upper);
6666 void *netdev_lower_dev_get_private(struct net_device *dev,
6667 struct net_device *lower_dev)
6669 struct netdev_adjacent *lower;
6673 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower);
6677 return lower->private;
6679 EXPORT_SYMBOL(netdev_lower_dev_get_private);
6682 int dev_get_nest_level(struct net_device *dev)
6684 struct net_device *lower = NULL;
6685 struct list_head *iter;
6691 netdev_for_each_lower_dev(dev, lower, iter) {
6692 nest = dev_get_nest_level(lower);
6693 if (max_nest < nest)
6697 return max_nest + 1;
6699 EXPORT_SYMBOL(dev_get_nest_level);
6702 * netdev_lower_change - Dispatch event about lower device state change
6703 * @lower_dev: device
6704 * @lower_state_info: state to dispatch
6706 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info.
6707 * The caller must hold the RTNL lock.
6709 void netdev_lower_state_changed(struct net_device *lower_dev,
6710 void *lower_state_info)
6712 struct netdev_notifier_changelowerstate_info changelowerstate_info = {
6713 .info.dev = lower_dev,
6717 changelowerstate_info.lower_state_info = lower_state_info;
6718 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE,
6719 &changelowerstate_info.info);
6721 EXPORT_SYMBOL(netdev_lower_state_changed);
6723 static void dev_change_rx_flags(struct net_device *dev, int flags)
6725 const struct net_device_ops *ops = dev->netdev_ops;
6727 if (ops->ndo_change_rx_flags)
6728 ops->ndo_change_rx_flags(dev, flags);
6731 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify)
6733 unsigned int old_flags = dev->flags;
6739 dev->flags |= IFF_PROMISC;
6740 dev->promiscuity += inc;
6741 if (dev->promiscuity == 0) {
6744 * If inc causes overflow, untouch promisc and return error.
6747 dev->flags &= ~IFF_PROMISC;
6749 dev->promiscuity -= inc;
6750 pr_warn("%s: promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n",
6755 if (dev->flags != old_flags) {
6756 pr_info("device %s %s promiscuous mode\n",
6758 dev->flags & IFF_PROMISC ? "entered" : "left");
6759 if (audit_enabled) {
6760 current_uid_gid(&uid, &gid);
6761 audit_log(current->audit_context, GFP_ATOMIC,
6762 AUDIT_ANOM_PROMISCUOUS,
6763 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u",
6764 dev->name, (dev->flags & IFF_PROMISC),
6765 (old_flags & IFF_PROMISC),
6766 from_kuid(&init_user_ns, audit_get_loginuid(current)),
6767 from_kuid(&init_user_ns, uid),
6768 from_kgid(&init_user_ns, gid),
6769 audit_get_sessionid(current));
6772 dev_change_rx_flags(dev, IFF_PROMISC);
6775 __dev_notify_flags(dev, old_flags, IFF_PROMISC);
6780 * dev_set_promiscuity - update promiscuity count on a device
6784 * Add or remove promiscuity from a device. While the count in the device
6785 * remains above zero the interface remains promiscuous. Once it hits zero
6786 * the device reverts back to normal filtering operation. A negative inc
6787 * value is used to drop promiscuity on the device.
6788 * Return 0 if successful or a negative errno code on error.
6790 int dev_set_promiscuity(struct net_device *dev, int inc)
6792 unsigned int old_flags = dev->flags;
6795 err = __dev_set_promiscuity(dev, inc, true);
6798 if (dev->flags != old_flags)
6799 dev_set_rx_mode(dev);
6802 EXPORT_SYMBOL(dev_set_promiscuity);
6804 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify)
6806 unsigned int old_flags = dev->flags, old_gflags = dev->gflags;
6810 dev->flags |= IFF_ALLMULTI;
6811 dev->allmulti += inc;
6812 if (dev->allmulti == 0) {
6815 * If inc causes overflow, untouch allmulti and return error.
6818 dev->flags &= ~IFF_ALLMULTI;
6820 dev->allmulti -= inc;
6821 pr_warn("%s: allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n",
6826 if (dev->flags ^ old_flags) {
6827 dev_change_rx_flags(dev, IFF_ALLMULTI);
6828 dev_set_rx_mode(dev);
6830 __dev_notify_flags(dev, old_flags,
6831 dev->gflags ^ old_gflags);
6837 * dev_set_allmulti - update allmulti count on a device
6841 * Add or remove reception of all multicast frames to a device. While the
6842 * count in the device remains above zero the interface remains listening
6843 * to all interfaces. Once it hits zero the device reverts back to normal
6844 * filtering operation. A negative @inc value is used to drop the counter
6845 * when releasing a resource needing all multicasts.
6846 * Return 0 if successful or a negative errno code on error.
6849 int dev_set_allmulti(struct net_device *dev, int inc)
6851 return __dev_set_allmulti(dev, inc, true);
6853 EXPORT_SYMBOL(dev_set_allmulti);
6856 * Upload unicast and multicast address lists to device and
6857 * configure RX filtering. When the device doesn't support unicast
6858 * filtering it is put in promiscuous mode while unicast addresses
6861 void __dev_set_rx_mode(struct net_device *dev)
6863 const struct net_device_ops *ops = dev->netdev_ops;
6865 /* dev_open will call this function so the list will stay sane. */
6866 if (!(dev->flags&IFF_UP))
6869 if (!netif_device_present(dev))
6872 if (!(dev->priv_flags & IFF_UNICAST_FLT)) {
6873 /* Unicast addresses changes may only happen under the rtnl,
6874 * therefore calling __dev_set_promiscuity here is safe.
6876 if (!netdev_uc_empty(dev) && !dev->uc_promisc) {
6877 __dev_set_promiscuity(dev, 1, false);
6878 dev->uc_promisc = true;
6879 } else if (netdev_uc_empty(dev) && dev->uc_promisc) {
6880 __dev_set_promiscuity(dev, -1, false);
6881 dev->uc_promisc = false;
6885 if (ops->ndo_set_rx_mode)
6886 ops->ndo_set_rx_mode(dev);
6889 void dev_set_rx_mode(struct net_device *dev)
6891 netif_addr_lock_bh(dev);
6892 __dev_set_rx_mode(dev);
6893 netif_addr_unlock_bh(dev);
6897 * dev_get_flags - get flags reported to userspace
6900 * Get the combination of flag bits exported through APIs to userspace.
6902 unsigned int dev_get_flags(const struct net_device *dev)
6906 flags = (dev->flags & ~(IFF_PROMISC |
6911 (dev->gflags & (IFF_PROMISC |
6914 if (netif_running(dev)) {
6915 if (netif_oper_up(dev))
6916 flags |= IFF_RUNNING;
6917 if (netif_carrier_ok(dev))
6918 flags |= IFF_LOWER_UP;
6919 if (netif_dormant(dev))
6920 flags |= IFF_DORMANT;
6925 EXPORT_SYMBOL(dev_get_flags);
6927 int __dev_change_flags(struct net_device *dev, unsigned int flags)
6929 unsigned int old_flags = dev->flags;
6935 * Set the flags on our device.
6938 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP |
6939 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL |
6941 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC |
6945 * Load in the correct multicast list now the flags have changed.
6948 if ((old_flags ^ flags) & IFF_MULTICAST)
6949 dev_change_rx_flags(dev, IFF_MULTICAST);
6951 dev_set_rx_mode(dev);
6954 * Have we downed the interface. We handle IFF_UP ourselves
6955 * according to user attempts to set it, rather than blindly
6960 if ((old_flags ^ flags) & IFF_UP) {
6961 if (old_flags & IFF_UP)
6964 ret = __dev_open(dev);
6967 if ((flags ^ dev->gflags) & IFF_PROMISC) {
6968 int inc = (flags & IFF_PROMISC) ? 1 : -1;
6969 unsigned int old_flags = dev->flags;
6971 dev->gflags ^= IFF_PROMISC;
6973 if (__dev_set_promiscuity(dev, inc, false) >= 0)
6974 if (dev->flags != old_flags)
6975 dev_set_rx_mode(dev);
6978 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI
6979 * is important. Some (broken) drivers set IFF_PROMISC, when
6980 * IFF_ALLMULTI is requested not asking us and not reporting.
6982 if ((flags ^ dev->gflags) & IFF_ALLMULTI) {
6983 int inc = (flags & IFF_ALLMULTI) ? 1 : -1;
6985 dev->gflags ^= IFF_ALLMULTI;
6986 __dev_set_allmulti(dev, inc, false);
6992 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags,
6993 unsigned int gchanges)
6995 unsigned int changes = dev->flags ^ old_flags;
6998 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC);
7000 if (changes & IFF_UP) {
7001 if (dev->flags & IFF_UP)
7002 call_netdevice_notifiers(NETDEV_UP, dev);
7004 call_netdevice_notifiers(NETDEV_DOWN, dev);
7007 if (dev->flags & IFF_UP &&
7008 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) {
7009 struct netdev_notifier_change_info change_info = {
7013 .flags_changed = changes,
7016 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info);
7021 * dev_change_flags - change device settings
7023 * @flags: device state flags
7025 * Change settings on device based state flags. The flags are
7026 * in the userspace exported format.
7028 int dev_change_flags(struct net_device *dev, unsigned int flags)
7031 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags;
7033 ret = __dev_change_flags(dev, flags);
7037 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags);
7038 __dev_notify_flags(dev, old_flags, changes);
7041 EXPORT_SYMBOL(dev_change_flags);
7043 int __dev_set_mtu(struct net_device *dev, int new_mtu)
7045 const struct net_device_ops *ops = dev->netdev_ops;
7047 if (ops->ndo_change_mtu)
7048 return ops->ndo_change_mtu(dev, new_mtu);
7053 EXPORT_SYMBOL(__dev_set_mtu);
7056 * dev_set_mtu - Change maximum transfer unit
7058 * @new_mtu: new transfer unit
7060 * Change the maximum transfer size of the network device.
7062 int dev_set_mtu(struct net_device *dev, int new_mtu)
7066 if (new_mtu == dev->mtu)
7069 /* MTU must be positive, and in range */
7070 if (new_mtu < 0 || new_mtu < dev->min_mtu) {
7071 net_err_ratelimited("%s: Invalid MTU %d requested, hw min %d\n",
7072 dev->name, new_mtu, dev->min_mtu);
7076 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) {
7077 net_err_ratelimited("%s: Invalid MTU %d requested, hw max %d\n",
7078 dev->name, new_mtu, dev->max_mtu);
7082 if (!netif_device_present(dev))
7085 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev);
7086 err = notifier_to_errno(err);
7090 orig_mtu = dev->mtu;
7091 err = __dev_set_mtu(dev, new_mtu);
7094 err = call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
7095 err = notifier_to_errno(err);
7097 /* setting mtu back and notifying everyone again,
7098 * so that they have a chance to revert changes.
7100 __dev_set_mtu(dev, orig_mtu);
7101 call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
7106 EXPORT_SYMBOL(dev_set_mtu);
7109 * dev_change_tx_queue_len - Change TX queue length of a netdevice
7111 * @new_len: new tx queue length
7113 int dev_change_tx_queue_len(struct net_device *dev, unsigned long new_len)
7115 unsigned int orig_len = dev->tx_queue_len;
7118 if (new_len != (unsigned int)new_len)
7121 if (new_len != orig_len) {
7122 dev->tx_queue_len = new_len;
7123 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev);
7124 res = notifier_to_errno(res);
7127 "refused to change device tx_queue_len\n");
7128 dev->tx_queue_len = orig_len;
7131 return dev_qdisc_change_tx_queue_len(dev);
7138 * dev_set_group - Change group this device belongs to
7140 * @new_group: group this device should belong to
7142 void dev_set_group(struct net_device *dev, int new_group)
7144 dev->group = new_group;
7146 EXPORT_SYMBOL(dev_set_group);
7149 * dev_set_mac_address - Change Media Access Control Address
7153 * Change the hardware (MAC) address of the device
7155 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa)
7157 const struct net_device_ops *ops = dev->netdev_ops;
7160 if (!ops->ndo_set_mac_address)
7162 if (sa->sa_family != dev->type)
7164 if (!netif_device_present(dev))
7166 err = ops->ndo_set_mac_address(dev, sa);
7169 dev->addr_assign_type = NET_ADDR_SET;
7170 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
7171 add_device_randomness(dev->dev_addr, dev->addr_len);
7174 EXPORT_SYMBOL(dev_set_mac_address);
7177 * dev_change_carrier - Change device carrier
7179 * @new_carrier: new value
7181 * Change device carrier
7183 int dev_change_carrier(struct net_device *dev, bool new_carrier)
7185 const struct net_device_ops *ops = dev->netdev_ops;
7187 if (!ops->ndo_change_carrier)
7189 if (!netif_device_present(dev))
7191 return ops->ndo_change_carrier(dev, new_carrier);
7193 EXPORT_SYMBOL(dev_change_carrier);
7196 * dev_get_phys_port_id - Get device physical port ID
7200 * Get device physical port ID
7202 int dev_get_phys_port_id(struct net_device *dev,
7203 struct netdev_phys_item_id *ppid)
7205 const struct net_device_ops *ops = dev->netdev_ops;
7207 if (!ops->ndo_get_phys_port_id)
7209 return ops->ndo_get_phys_port_id(dev, ppid);
7211 EXPORT_SYMBOL(dev_get_phys_port_id);
7214 * dev_get_phys_port_name - Get device physical port name
7217 * @len: limit of bytes to copy to name
7219 * Get device physical port name
7221 int dev_get_phys_port_name(struct net_device *dev,
7222 char *name, size_t len)
7224 const struct net_device_ops *ops = dev->netdev_ops;
7226 if (!ops->ndo_get_phys_port_name)
7228 return ops->ndo_get_phys_port_name(dev, name, len);
7230 EXPORT_SYMBOL(dev_get_phys_port_name);
7233 * dev_change_proto_down - update protocol port state information
7235 * @proto_down: new value
7237 * This info can be used by switch drivers to set the phys state of the
7240 int dev_change_proto_down(struct net_device *dev, bool proto_down)
7242 const struct net_device_ops *ops = dev->netdev_ops;
7244 if (!ops->ndo_change_proto_down)
7246 if (!netif_device_present(dev))
7248 return ops->ndo_change_proto_down(dev, proto_down);
7250 EXPORT_SYMBOL(dev_change_proto_down);
7252 void __dev_xdp_query(struct net_device *dev, bpf_op_t bpf_op,
7253 struct netdev_bpf *xdp)
7255 memset(xdp, 0, sizeof(*xdp));
7256 xdp->command = XDP_QUERY_PROG;
7258 /* Query must always succeed. */
7259 WARN_ON(bpf_op(dev, xdp) < 0);
7262 static u8 __dev_xdp_attached(struct net_device *dev, bpf_op_t bpf_op)
7264 struct netdev_bpf xdp;
7266 __dev_xdp_query(dev, bpf_op, &xdp);
7268 return xdp.prog_attached;
7271 static int dev_xdp_install(struct net_device *dev, bpf_op_t bpf_op,
7272 struct netlink_ext_ack *extack, u32 flags,
7273 struct bpf_prog *prog)
7275 struct netdev_bpf xdp;
7277 memset(&xdp, 0, sizeof(xdp));
7278 if (flags & XDP_FLAGS_HW_MODE)
7279 xdp.command = XDP_SETUP_PROG_HW;
7281 xdp.command = XDP_SETUP_PROG;
7282 xdp.extack = extack;
7286 return bpf_op(dev, &xdp);
7289 static void dev_xdp_uninstall(struct net_device *dev)
7291 struct netdev_bpf xdp;
7294 /* Remove generic XDP */
7295 WARN_ON(dev_xdp_install(dev, generic_xdp_install, NULL, 0, NULL));
7297 /* Remove from the driver */
7298 ndo_bpf = dev->netdev_ops->ndo_bpf;
7302 __dev_xdp_query(dev, ndo_bpf, &xdp);
7303 if (xdp.prog_attached == XDP_ATTACHED_NONE)
7306 /* Program removal should always succeed */
7307 WARN_ON(dev_xdp_install(dev, ndo_bpf, NULL, xdp.prog_flags, NULL));
7311 * dev_change_xdp_fd - set or clear a bpf program for a device rx path
7313 * @extack: netlink extended ack
7314 * @fd: new program fd or negative value to clear
7315 * @flags: xdp-related flags
7317 * Set or clear a bpf program for a device
7319 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack,
7322 const struct net_device_ops *ops = dev->netdev_ops;
7323 struct bpf_prog *prog = NULL;
7324 bpf_op_t bpf_op, bpf_chk;
7329 bpf_op = bpf_chk = ops->ndo_bpf;
7330 if (!bpf_op && (flags & (XDP_FLAGS_DRV_MODE | XDP_FLAGS_HW_MODE)))
7332 if (!bpf_op || (flags & XDP_FLAGS_SKB_MODE))
7333 bpf_op = generic_xdp_install;
7334 if (bpf_op == bpf_chk)
7335 bpf_chk = generic_xdp_install;
7338 if (bpf_chk && __dev_xdp_attached(dev, bpf_chk))
7340 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) &&
7341 __dev_xdp_attached(dev, bpf_op))
7344 prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP,
7345 bpf_op == ops->ndo_bpf);
7347 return PTR_ERR(prog);
7349 if (!(flags & XDP_FLAGS_HW_MODE) &&
7350 bpf_prog_is_dev_bound(prog->aux)) {
7351 NL_SET_ERR_MSG(extack, "using device-bound program without HW_MODE flag is not supported");
7357 err = dev_xdp_install(dev, bpf_op, extack, flags, prog);
7358 if (err < 0 && prog)
7365 * dev_new_index - allocate an ifindex
7366 * @net: the applicable net namespace
7368 * Returns a suitable unique value for a new device interface
7369 * number. The caller must hold the rtnl semaphore or the
7370 * dev_base_lock to be sure it remains unique.
7372 static int dev_new_index(struct net *net)
7374 int ifindex = net->ifindex;
7379 if (!__dev_get_by_index(net, ifindex))
7380 return net->ifindex = ifindex;
7384 /* Delayed registration/unregisteration */
7385 static LIST_HEAD(net_todo_list);
7386 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq);
7388 static void net_set_todo(struct net_device *dev)
7390 list_add_tail(&dev->todo_list, &net_todo_list);
7391 dev_net(dev)->dev_unreg_count++;
7394 static void rollback_registered_many(struct list_head *head)
7396 struct net_device *dev, *tmp;
7397 LIST_HEAD(close_head);
7399 BUG_ON(dev_boot_phase);
7402 list_for_each_entry_safe(dev, tmp, head, unreg_list) {
7403 /* Some devices call without registering
7404 * for initialization unwind. Remove those
7405 * devices and proceed with the remaining.
7407 if (dev->reg_state == NETREG_UNINITIALIZED) {
7408 pr_debug("unregister_netdevice: device %s/%p never was registered\n",
7412 list_del(&dev->unreg_list);
7415 dev->dismantle = true;
7416 BUG_ON(dev->reg_state != NETREG_REGISTERED);
7419 /* If device is running, close it first. */
7420 list_for_each_entry(dev, head, unreg_list)
7421 list_add_tail(&dev->close_list, &close_head);
7422 dev_close_many(&close_head, true);
7424 list_for_each_entry(dev, head, unreg_list) {
7425 /* And unlink it from device chain. */
7426 unlist_netdevice(dev);
7428 dev->reg_state = NETREG_UNREGISTERING;
7430 flush_all_backlogs();
7434 list_for_each_entry(dev, head, unreg_list) {
7435 struct sk_buff *skb = NULL;
7437 /* Shutdown queueing discipline. */
7440 dev_xdp_uninstall(dev);
7442 /* Notify protocols, that we are about to destroy
7443 * this device. They should clean all the things.
7445 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
7447 if (!dev->rtnl_link_ops ||
7448 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
7449 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0,
7450 GFP_KERNEL, NULL, 0);
7453 * Flush the unicast and multicast chains
7458 if (dev->netdev_ops->ndo_uninit)
7459 dev->netdev_ops->ndo_uninit(dev);
7462 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL);
7464 /* Notifier chain MUST detach us all upper devices. */
7465 WARN_ON(netdev_has_any_upper_dev(dev));
7466 WARN_ON(netdev_has_any_lower_dev(dev));
7468 /* Remove entries from kobject tree */
7469 netdev_unregister_kobject(dev);
7471 /* Remove XPS queueing entries */
7472 netif_reset_xps_queues_gt(dev, 0);
7478 list_for_each_entry(dev, head, unreg_list)
7482 static void rollback_registered(struct net_device *dev)
7486 list_add(&dev->unreg_list, &single);
7487 rollback_registered_many(&single);
7491 static netdev_features_t netdev_sync_upper_features(struct net_device *lower,
7492 struct net_device *upper, netdev_features_t features)
7494 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
7495 netdev_features_t feature;
7498 for_each_netdev_feature(&upper_disables, feature_bit) {
7499 feature = __NETIF_F_BIT(feature_bit);
7500 if (!(upper->wanted_features & feature)
7501 && (features & feature)) {
7502 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n",
7503 &feature, upper->name);
7504 features &= ~feature;
7511 static void netdev_sync_lower_features(struct net_device *upper,
7512 struct net_device *lower, netdev_features_t features)
7514 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
7515 netdev_features_t feature;
7518 for_each_netdev_feature(&upper_disables, feature_bit) {
7519 feature = __NETIF_F_BIT(feature_bit);
7520 if (!(features & feature) && (lower->features & feature)) {
7521 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n",
7522 &feature, lower->name);
7523 lower->wanted_features &= ~feature;
7524 netdev_update_features(lower);
7526 if (unlikely(lower->features & feature))
7527 netdev_WARN(upper, "failed to disable %pNF on %s!\n",
7528 &feature, lower->name);
7533 static netdev_features_t netdev_fix_features(struct net_device *dev,
7534 netdev_features_t features)
7536 /* Fix illegal checksum combinations */
7537 if ((features & NETIF_F_HW_CSUM) &&
7538 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
7539 netdev_warn(dev, "mixed HW and IP checksum settings.\n");
7540 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
7543 /* TSO requires that SG is present as well. */
7544 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) {
7545 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n");
7546 features &= ~NETIF_F_ALL_TSO;
7549 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) &&
7550 !(features & NETIF_F_IP_CSUM)) {
7551 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n");
7552 features &= ~NETIF_F_TSO;
7553 features &= ~NETIF_F_TSO_ECN;
7556 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) &&
7557 !(features & NETIF_F_IPV6_CSUM)) {
7558 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n");
7559 features &= ~NETIF_F_TSO6;
7562 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */
7563 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO))
7564 features &= ~NETIF_F_TSO_MANGLEID;
7566 /* TSO ECN requires that TSO is present as well. */
7567 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN)
7568 features &= ~NETIF_F_TSO_ECN;
7570 /* Software GSO depends on SG. */
7571 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) {
7572 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n");
7573 features &= ~NETIF_F_GSO;
7576 /* GSO partial features require GSO partial be set */
7577 if ((features & dev->gso_partial_features) &&
7578 !(features & NETIF_F_GSO_PARTIAL)) {
7580 "Dropping partially supported GSO features since no GSO partial.\n");
7581 features &= ~dev->gso_partial_features;
7584 if (!(features & NETIF_F_RXCSUM)) {
7585 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet
7586 * successfully merged by hardware must also have the
7587 * checksum verified by hardware. If the user does not
7588 * want to enable RXCSUM, logically, we should disable GRO_HW.
7590 if (features & NETIF_F_GRO_HW) {
7591 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n");
7592 features &= ~NETIF_F_GRO_HW;
7596 /* LRO/HW-GRO features cannot be combined with RX-FCS */
7597 if (features & NETIF_F_RXFCS) {
7598 if (features & NETIF_F_LRO) {
7599 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n");
7600 features &= ~NETIF_F_LRO;
7603 if (features & NETIF_F_GRO_HW) {
7604 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n");
7605 features &= ~NETIF_F_GRO_HW;
7612 int __netdev_update_features(struct net_device *dev)
7614 struct net_device *upper, *lower;
7615 netdev_features_t features;
7616 struct list_head *iter;
7621 features = netdev_get_wanted_features(dev);
7623 if (dev->netdev_ops->ndo_fix_features)
7624 features = dev->netdev_ops->ndo_fix_features(dev, features);
7626 /* driver might be less strict about feature dependencies */
7627 features = netdev_fix_features(dev, features);
7629 /* some features can't be enabled if they're off an an upper device */
7630 netdev_for_each_upper_dev_rcu(dev, upper, iter)
7631 features = netdev_sync_upper_features(dev, upper, features);
7633 if (dev->features == features)
7636 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n",
7637 &dev->features, &features);
7639 if (dev->netdev_ops->ndo_set_features)
7640 err = dev->netdev_ops->ndo_set_features(dev, features);
7644 if (unlikely(err < 0)) {
7646 "set_features() failed (%d); wanted %pNF, left %pNF\n",
7647 err, &features, &dev->features);
7648 /* return non-0 since some features might have changed and
7649 * it's better to fire a spurious notification than miss it
7655 /* some features must be disabled on lower devices when disabled
7656 * on an upper device (think: bonding master or bridge)
7658 netdev_for_each_lower_dev(dev, lower, iter)
7659 netdev_sync_lower_features(dev, lower, features);
7662 netdev_features_t diff = features ^ dev->features;
7664 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) {
7665 /* udp_tunnel_{get,drop}_rx_info both need
7666 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the
7667 * device, or they won't do anything.
7668 * Thus we need to update dev->features
7669 * *before* calling udp_tunnel_get_rx_info,
7670 * but *after* calling udp_tunnel_drop_rx_info.
7672 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) {
7673 dev->features = features;
7674 udp_tunnel_get_rx_info(dev);
7676 udp_tunnel_drop_rx_info(dev);
7680 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) {
7681 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) {
7682 dev->features = features;
7683 err |= vlan_get_rx_ctag_filter_info(dev);
7685 vlan_drop_rx_ctag_filter_info(dev);
7689 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) {
7690 if (features & NETIF_F_HW_VLAN_STAG_FILTER) {
7691 dev->features = features;
7692 err |= vlan_get_rx_stag_filter_info(dev);
7694 vlan_drop_rx_stag_filter_info(dev);
7698 dev->features = features;
7701 return err < 0 ? 0 : 1;
7705 * netdev_update_features - recalculate device features
7706 * @dev: the device to check
7708 * Recalculate dev->features set and send notifications if it
7709 * has changed. Should be called after driver or hardware dependent
7710 * conditions might have changed that influence the features.
7712 void netdev_update_features(struct net_device *dev)
7714 if (__netdev_update_features(dev))
7715 netdev_features_change(dev);
7717 EXPORT_SYMBOL(netdev_update_features);
7720 * netdev_change_features - recalculate device features
7721 * @dev: the device to check
7723 * Recalculate dev->features set and send notifications even
7724 * if they have not changed. Should be called instead of
7725 * netdev_update_features() if also dev->vlan_features might
7726 * have changed to allow the changes to be propagated to stacked
7729 void netdev_change_features(struct net_device *dev)
7731 __netdev_update_features(dev);
7732 netdev_features_change(dev);
7734 EXPORT_SYMBOL(netdev_change_features);
7737 * netif_stacked_transfer_operstate - transfer operstate
7738 * @rootdev: the root or lower level device to transfer state from
7739 * @dev: the device to transfer operstate to
7741 * Transfer operational state from root to device. This is normally
7742 * called when a stacking relationship exists between the root
7743 * device and the device(a leaf device).
7745 void netif_stacked_transfer_operstate(const struct net_device *rootdev,
7746 struct net_device *dev)
7748 if (rootdev->operstate == IF_OPER_DORMANT)
7749 netif_dormant_on(dev);
7751 netif_dormant_off(dev);
7753 if (netif_carrier_ok(rootdev))
7754 netif_carrier_on(dev);
7756 netif_carrier_off(dev);
7758 EXPORT_SYMBOL(netif_stacked_transfer_operstate);
7760 static int netif_alloc_rx_queues(struct net_device *dev)
7762 unsigned int i, count = dev->num_rx_queues;
7763 struct netdev_rx_queue *rx;
7764 size_t sz = count * sizeof(*rx);
7769 rx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
7775 for (i = 0; i < count; i++) {
7778 /* XDP RX-queue setup */
7779 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i);
7786 /* Rollback successful reg's and free other resources */
7788 xdp_rxq_info_unreg(&rx[i].xdp_rxq);
7794 static void netif_free_rx_queues(struct net_device *dev)
7796 unsigned int i, count = dev->num_rx_queues;
7798 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */
7802 for (i = 0; i < count; i++)
7803 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq);
7808 static void netdev_init_one_queue(struct net_device *dev,
7809 struct netdev_queue *queue, void *_unused)
7811 /* Initialize queue lock */
7812 spin_lock_init(&queue->_xmit_lock);
7813 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type);
7814 queue->xmit_lock_owner = -1;
7815 netdev_queue_numa_node_write(queue, NUMA_NO_NODE);
7818 dql_init(&queue->dql, HZ);
7822 static void netif_free_tx_queues(struct net_device *dev)
7827 static int netif_alloc_netdev_queues(struct net_device *dev)
7829 unsigned int count = dev->num_tx_queues;
7830 struct netdev_queue *tx;
7831 size_t sz = count * sizeof(*tx);
7833 if (count < 1 || count > 0xffff)
7836 tx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
7842 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL);
7843 spin_lock_init(&dev->tx_global_lock);
7848 void netif_tx_stop_all_queues(struct net_device *dev)
7852 for (i = 0; i < dev->num_tx_queues; i++) {
7853 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
7855 netif_tx_stop_queue(txq);
7858 EXPORT_SYMBOL(netif_tx_stop_all_queues);
7861 * register_netdevice - register a network device
7862 * @dev: device to register
7864 * Take a completed network device structure and add it to the kernel
7865 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
7866 * chain. 0 is returned on success. A negative errno code is returned
7867 * on a failure to set up the device, or if the name is a duplicate.
7869 * Callers must hold the rtnl semaphore. You may want
7870 * register_netdev() instead of this.
7873 * The locking appears insufficient to guarantee two parallel registers
7874 * will not get the same name.
7877 int register_netdevice(struct net_device *dev)
7880 struct net *net = dev_net(dev);
7882 BUG_ON(dev_boot_phase);
7887 /* When net_device's are persistent, this will be fatal. */
7888 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
7891 spin_lock_init(&dev->addr_list_lock);
7892 netdev_set_addr_lockdep_class(dev);
7894 ret = dev_get_valid_name(net, dev, dev->name);
7898 /* Init, if this function is available */
7899 if (dev->netdev_ops->ndo_init) {
7900 ret = dev->netdev_ops->ndo_init(dev);
7908 if (((dev->hw_features | dev->features) &
7909 NETIF_F_HW_VLAN_CTAG_FILTER) &&
7910 (!dev->netdev_ops->ndo_vlan_rx_add_vid ||
7911 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) {
7912 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n");
7919 dev->ifindex = dev_new_index(net);
7920 else if (__dev_get_by_index(net, dev->ifindex))
7923 /* Transfer changeable features to wanted_features and enable
7924 * software offloads (GSO and GRO).
7926 dev->hw_features |= NETIF_F_SOFT_FEATURES;
7927 dev->features |= NETIF_F_SOFT_FEATURES;
7929 if (dev->netdev_ops->ndo_udp_tunnel_add) {
7930 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT;
7931 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT;
7934 dev->wanted_features = dev->features & dev->hw_features;
7936 if (!(dev->flags & IFF_LOOPBACK))
7937 dev->hw_features |= NETIF_F_NOCACHE_COPY;
7939 /* If IPv4 TCP segmentation offload is supported we should also
7940 * allow the device to enable segmenting the frame with the option
7941 * of ignoring a static IP ID value. This doesn't enable the
7942 * feature itself but allows the user to enable it later.
7944 if (dev->hw_features & NETIF_F_TSO)
7945 dev->hw_features |= NETIF_F_TSO_MANGLEID;
7946 if (dev->vlan_features & NETIF_F_TSO)
7947 dev->vlan_features |= NETIF_F_TSO_MANGLEID;
7948 if (dev->mpls_features & NETIF_F_TSO)
7949 dev->mpls_features |= NETIF_F_TSO_MANGLEID;
7950 if (dev->hw_enc_features & NETIF_F_TSO)
7951 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID;
7953 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices.
7955 dev->vlan_features |= NETIF_F_HIGHDMA;
7957 /* Make NETIF_F_SG inheritable to tunnel devices.
7959 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL;
7961 /* Make NETIF_F_SG inheritable to MPLS.
7963 dev->mpls_features |= NETIF_F_SG;
7965 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev);
7966 ret = notifier_to_errno(ret);
7970 ret = netdev_register_kobject(dev);
7973 dev->reg_state = NETREG_REGISTERED;
7975 __netdev_update_features(dev);
7978 * Default initial state at registry is that the
7979 * device is present.
7982 set_bit(__LINK_STATE_PRESENT, &dev->state);
7984 linkwatch_init_dev(dev);
7986 dev_init_scheduler(dev);
7988 list_netdevice(dev);
7989 add_device_randomness(dev->dev_addr, dev->addr_len);
7991 /* If the device has permanent device address, driver should
7992 * set dev_addr and also addr_assign_type should be set to
7993 * NET_ADDR_PERM (default value).
7995 if (dev->addr_assign_type == NET_ADDR_PERM)
7996 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len);
7998 /* Notify protocols, that a new device appeared. */
7999 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
8000 ret = notifier_to_errno(ret);
8002 rollback_registered(dev);
8003 dev->reg_state = NETREG_UNREGISTERED;
8006 * Prevent userspace races by waiting until the network
8007 * device is fully setup before sending notifications.
8009 if (!dev->rtnl_link_ops ||
8010 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
8011 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
8017 if (dev->netdev_ops->ndo_uninit)
8018 dev->netdev_ops->ndo_uninit(dev);
8019 if (dev->priv_destructor)
8020 dev->priv_destructor(dev);
8023 EXPORT_SYMBOL(register_netdevice);
8026 * init_dummy_netdev - init a dummy network device for NAPI
8027 * @dev: device to init
8029 * This takes a network device structure and initialize the minimum
8030 * amount of fields so it can be used to schedule NAPI polls without
8031 * registering a full blown interface. This is to be used by drivers
8032 * that need to tie several hardware interfaces to a single NAPI
8033 * poll scheduler due to HW limitations.
8035 int init_dummy_netdev(struct net_device *dev)
8037 /* Clear everything. Note we don't initialize spinlocks
8038 * are they aren't supposed to be taken by any of the
8039 * NAPI code and this dummy netdev is supposed to be
8040 * only ever used for NAPI polls
8042 memset(dev, 0, sizeof(struct net_device));
8044 /* make sure we BUG if trying to hit standard
8045 * register/unregister code path
8047 dev->reg_state = NETREG_DUMMY;
8049 /* NAPI wants this */
8050 INIT_LIST_HEAD(&dev->napi_list);
8052 /* a dummy interface is started by default */
8053 set_bit(__LINK_STATE_PRESENT, &dev->state);
8054 set_bit(__LINK_STATE_START, &dev->state);
8056 /* Note : We dont allocate pcpu_refcnt for dummy devices,
8057 * because users of this 'device' dont need to change
8063 EXPORT_SYMBOL_GPL(init_dummy_netdev);
8067 * register_netdev - register a network device
8068 * @dev: device to register
8070 * Take a completed network device structure and add it to the kernel
8071 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
8072 * chain. 0 is returned on success. A negative errno code is returned
8073 * on a failure to set up the device, or if the name is a duplicate.
8075 * This is a wrapper around register_netdevice that takes the rtnl semaphore
8076 * and expands the device name if you passed a format string to
8079 int register_netdev(struct net_device *dev)
8083 if (rtnl_lock_killable())
8085 err = register_netdevice(dev);
8089 EXPORT_SYMBOL(register_netdev);
8091 int netdev_refcnt_read(const struct net_device *dev)
8095 for_each_possible_cpu(i)
8096 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
8099 EXPORT_SYMBOL(netdev_refcnt_read);
8102 * netdev_wait_allrefs - wait until all references are gone.
8103 * @dev: target net_device
8105 * This is called when unregistering network devices.
8107 * Any protocol or device that holds a reference should register
8108 * for netdevice notification, and cleanup and put back the
8109 * reference if they receive an UNREGISTER event.
8110 * We can get stuck here if buggy protocols don't correctly
8113 static void netdev_wait_allrefs(struct net_device *dev)
8115 unsigned long rebroadcast_time, warning_time;
8118 linkwatch_forget_dev(dev);
8120 rebroadcast_time = warning_time = jiffies;
8121 refcnt = netdev_refcnt_read(dev);
8123 while (refcnt != 0) {
8124 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
8127 /* Rebroadcast unregister notification */
8128 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
8134 if (test_bit(__LINK_STATE_LINKWATCH_PENDING,
8136 /* We must not have linkwatch events
8137 * pending on unregister. If this
8138 * happens, we simply run the queue
8139 * unscheduled, resulting in a noop
8142 linkwatch_run_queue();
8147 rebroadcast_time = jiffies;
8152 refcnt = netdev_refcnt_read(dev);
8154 if (time_after(jiffies, warning_time + 10 * HZ)) {
8155 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n",
8157 warning_time = jiffies;
8166 * register_netdevice(x1);
8167 * register_netdevice(x2);
8169 * unregister_netdevice(y1);
8170 * unregister_netdevice(y2);
8176 * We are invoked by rtnl_unlock().
8177 * This allows us to deal with problems:
8178 * 1) We can delete sysfs objects which invoke hotplug
8179 * without deadlocking with linkwatch via keventd.
8180 * 2) Since we run with the RTNL semaphore not held, we can sleep
8181 * safely in order to wait for the netdev refcnt to drop to zero.
8183 * We must not return until all unregister events added during
8184 * the interval the lock was held have been completed.
8186 void netdev_run_todo(void)
8188 struct list_head list;
8190 /* Snapshot list, allow later requests */
8191 list_replace_init(&net_todo_list, &list);
8196 /* Wait for rcu callbacks to finish before next phase */
8197 if (!list_empty(&list))
8200 while (!list_empty(&list)) {
8201 struct net_device *dev
8202 = list_first_entry(&list, struct net_device, todo_list);
8203 list_del(&dev->todo_list);
8205 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) {
8206 pr_err("network todo '%s' but state %d\n",
8207 dev->name, dev->reg_state);
8212 dev->reg_state = NETREG_UNREGISTERED;
8214 netdev_wait_allrefs(dev);
8217 BUG_ON(netdev_refcnt_read(dev));
8218 BUG_ON(!list_empty(&dev->ptype_all));
8219 BUG_ON(!list_empty(&dev->ptype_specific));
8220 WARN_ON(rcu_access_pointer(dev->ip_ptr));
8221 WARN_ON(rcu_access_pointer(dev->ip6_ptr));
8222 #if IS_ENABLED(CONFIG_DECNET)
8223 WARN_ON(dev->dn_ptr);
8225 if (dev->priv_destructor)
8226 dev->priv_destructor(dev);
8227 if (dev->needs_free_netdev)
8230 /* Report a network device has been unregistered */
8232 dev_net(dev)->dev_unreg_count--;
8234 wake_up(&netdev_unregistering_wq);
8236 /* Free network device */
8237 kobject_put(&dev->dev.kobj);
8241 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has
8242 * all the same fields in the same order as net_device_stats, with only
8243 * the type differing, but rtnl_link_stats64 may have additional fields
8244 * at the end for newer counters.
8246 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
8247 const struct net_device_stats *netdev_stats)
8249 #if BITS_PER_LONG == 64
8250 BUILD_BUG_ON(sizeof(*stats64) < sizeof(*netdev_stats));
8251 memcpy(stats64, netdev_stats, sizeof(*netdev_stats));
8252 /* zero out counters that only exist in rtnl_link_stats64 */
8253 memset((char *)stats64 + sizeof(*netdev_stats), 0,
8254 sizeof(*stats64) - sizeof(*netdev_stats));
8256 size_t i, n = sizeof(*netdev_stats) / sizeof(unsigned long);
8257 const unsigned long *src = (const unsigned long *)netdev_stats;
8258 u64 *dst = (u64 *)stats64;
8260 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64));
8261 for (i = 0; i < n; i++)
8263 /* zero out counters that only exist in rtnl_link_stats64 */
8264 memset((char *)stats64 + n * sizeof(u64), 0,
8265 sizeof(*stats64) - n * sizeof(u64));
8268 EXPORT_SYMBOL(netdev_stats_to_stats64);
8271 * dev_get_stats - get network device statistics
8272 * @dev: device to get statistics from
8273 * @storage: place to store stats
8275 * Get network statistics from device. Return @storage.
8276 * The device driver may provide its own method by setting
8277 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats;
8278 * otherwise the internal statistics structure is used.
8280 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev,
8281 struct rtnl_link_stats64 *storage)
8283 const struct net_device_ops *ops = dev->netdev_ops;
8285 if (ops->ndo_get_stats64) {
8286 memset(storage, 0, sizeof(*storage));
8287 ops->ndo_get_stats64(dev, storage);
8288 } else if (ops->ndo_get_stats) {
8289 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev));
8291 netdev_stats_to_stats64(storage, &dev->stats);
8293 storage->rx_dropped += (unsigned long)atomic_long_read(&dev->rx_dropped);
8294 storage->tx_dropped += (unsigned long)atomic_long_read(&dev->tx_dropped);
8295 storage->rx_nohandler += (unsigned long)atomic_long_read(&dev->rx_nohandler);
8298 EXPORT_SYMBOL(dev_get_stats);
8300 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev)
8302 struct netdev_queue *queue = dev_ingress_queue(dev);
8304 #ifdef CONFIG_NET_CLS_ACT
8307 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
8310 netdev_init_one_queue(dev, queue, NULL);
8311 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc);
8312 queue->qdisc_sleeping = &noop_qdisc;
8313 rcu_assign_pointer(dev->ingress_queue, queue);
8318 static const struct ethtool_ops default_ethtool_ops;
8320 void netdev_set_default_ethtool_ops(struct net_device *dev,
8321 const struct ethtool_ops *ops)
8323 if (dev->ethtool_ops == &default_ethtool_ops)
8324 dev->ethtool_ops = ops;
8326 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops);
8328 void netdev_freemem(struct net_device *dev)
8330 char *addr = (char *)dev - dev->padded;
8336 * alloc_netdev_mqs - allocate network device
8337 * @sizeof_priv: size of private data to allocate space for
8338 * @name: device name format string
8339 * @name_assign_type: origin of device name
8340 * @setup: callback to initialize device
8341 * @txqs: the number of TX subqueues to allocate
8342 * @rxqs: the number of RX subqueues to allocate
8344 * Allocates a struct net_device with private data area for driver use
8345 * and performs basic initialization. Also allocates subqueue structs
8346 * for each queue on the device.
8348 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
8349 unsigned char name_assign_type,
8350 void (*setup)(struct net_device *),
8351 unsigned int txqs, unsigned int rxqs)
8353 struct net_device *dev;
8354 unsigned int alloc_size;
8355 struct net_device *p;
8357 BUG_ON(strlen(name) >= sizeof(dev->name));
8360 pr_err("alloc_netdev: Unable to allocate device with zero queues\n");
8365 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n");
8369 alloc_size = sizeof(struct net_device);
8371 /* ensure 32-byte alignment of private area */
8372 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
8373 alloc_size += sizeof_priv;
8375 /* ensure 32-byte alignment of whole construct */
8376 alloc_size += NETDEV_ALIGN - 1;
8378 p = kvzalloc(alloc_size, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
8382 dev = PTR_ALIGN(p, NETDEV_ALIGN);
8383 dev->padded = (char *)dev - (char *)p;
8385 dev->pcpu_refcnt = alloc_percpu(int);
8386 if (!dev->pcpu_refcnt)
8389 if (dev_addr_init(dev))
8395 dev_net_set(dev, &init_net);
8397 dev->gso_max_size = GSO_MAX_SIZE;
8398 dev->gso_max_segs = GSO_MAX_SEGS;
8400 INIT_LIST_HEAD(&dev->napi_list);
8401 INIT_LIST_HEAD(&dev->unreg_list);
8402 INIT_LIST_HEAD(&dev->close_list);
8403 INIT_LIST_HEAD(&dev->link_watch_list);
8404 INIT_LIST_HEAD(&dev->adj_list.upper);
8405 INIT_LIST_HEAD(&dev->adj_list.lower);
8406 INIT_LIST_HEAD(&dev->ptype_all);
8407 INIT_LIST_HEAD(&dev->ptype_specific);
8408 #ifdef CONFIG_NET_SCHED
8409 hash_init(dev->qdisc_hash);
8411 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM;
8414 if (!dev->tx_queue_len) {
8415 dev->priv_flags |= IFF_NO_QUEUE;
8416 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN;
8419 dev->num_tx_queues = txqs;
8420 dev->real_num_tx_queues = txqs;
8421 if (netif_alloc_netdev_queues(dev))
8424 dev->num_rx_queues = rxqs;
8425 dev->real_num_rx_queues = rxqs;
8426 if (netif_alloc_rx_queues(dev))
8429 strcpy(dev->name, name);
8430 dev->name_assign_type = name_assign_type;
8431 dev->group = INIT_NETDEV_GROUP;
8432 if (!dev->ethtool_ops)
8433 dev->ethtool_ops = &default_ethtool_ops;
8435 nf_hook_ingress_init(dev);
8444 free_percpu(dev->pcpu_refcnt);
8446 netdev_freemem(dev);
8449 EXPORT_SYMBOL(alloc_netdev_mqs);
8452 * free_netdev - free network device
8455 * This function does the last stage of destroying an allocated device
8456 * interface. The reference to the device object is released. If this
8457 * is the last reference then it will be freed.Must be called in process
8460 void free_netdev(struct net_device *dev)
8462 struct napi_struct *p, *n;
8465 netif_free_tx_queues(dev);
8466 netif_free_rx_queues(dev);
8468 kfree(rcu_dereference_protected(dev->ingress_queue, 1));
8470 /* Flush device addresses */
8471 dev_addr_flush(dev);
8473 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
8476 free_percpu(dev->pcpu_refcnt);
8477 dev->pcpu_refcnt = NULL;
8479 /* Compatibility with error handling in drivers */
8480 if (dev->reg_state == NETREG_UNINITIALIZED) {
8481 netdev_freemem(dev);
8485 BUG_ON(dev->reg_state != NETREG_UNREGISTERED);
8486 dev->reg_state = NETREG_RELEASED;
8488 /* will free via device release */
8489 put_device(&dev->dev);
8491 EXPORT_SYMBOL(free_netdev);
8494 * synchronize_net - Synchronize with packet receive processing
8496 * Wait for packets currently being received to be done.
8497 * Does not block later packets from starting.
8499 void synchronize_net(void)
8502 if (rtnl_is_locked())
8503 synchronize_rcu_expedited();
8507 EXPORT_SYMBOL(synchronize_net);
8510 * unregister_netdevice_queue - remove device from the kernel
8514 * This function shuts down a device interface and removes it
8515 * from the kernel tables.
8516 * If head not NULL, device is queued to be unregistered later.
8518 * Callers must hold the rtnl semaphore. You may want
8519 * unregister_netdev() instead of this.
8522 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head)
8527 list_move_tail(&dev->unreg_list, head);
8529 rollback_registered(dev);
8530 /* Finish processing unregister after unlock */
8534 EXPORT_SYMBOL(unregister_netdevice_queue);
8537 * unregister_netdevice_many - unregister many devices
8538 * @head: list of devices
8540 * Note: As most callers use a stack allocated list_head,
8541 * we force a list_del() to make sure stack wont be corrupted later.
8543 void unregister_netdevice_many(struct list_head *head)
8545 struct net_device *dev;
8547 if (!list_empty(head)) {
8548 rollback_registered_many(head);
8549 list_for_each_entry(dev, head, unreg_list)
8554 EXPORT_SYMBOL(unregister_netdevice_many);
8557 * unregister_netdev - remove device from the kernel
8560 * This function shuts down a device interface and removes it
8561 * from the kernel tables.
8563 * This is just a wrapper for unregister_netdevice that takes
8564 * the rtnl semaphore. In general you want to use this and not
8565 * unregister_netdevice.
8567 void unregister_netdev(struct net_device *dev)
8570 unregister_netdevice(dev);
8573 EXPORT_SYMBOL(unregister_netdev);
8576 * dev_change_net_namespace - move device to different nethost namespace
8578 * @net: network namespace
8579 * @pat: If not NULL name pattern to try if the current device name
8580 * is already taken in the destination network namespace.
8582 * This function shuts down a device interface and moves it
8583 * to a new network namespace. On success 0 is returned, on
8584 * a failure a netagive errno code is returned.
8586 * Callers must hold the rtnl semaphore.
8589 int dev_change_net_namespace(struct net_device *dev, struct net *net, const char *pat)
8591 int err, new_nsid, new_ifindex;
8595 /* Don't allow namespace local devices to be moved. */
8597 if (dev->features & NETIF_F_NETNS_LOCAL)
8600 /* Ensure the device has been registrered */
8601 if (dev->reg_state != NETREG_REGISTERED)
8604 /* Get out if there is nothing todo */
8606 if (net_eq(dev_net(dev), net))
8609 /* Pick the destination device name, and ensure
8610 * we can use it in the destination network namespace.
8613 if (__dev_get_by_name(net, dev->name)) {
8614 /* We get here if we can't use the current device name */
8617 if (dev_get_valid_name(net, dev, pat) < 0)
8622 * And now a mini version of register_netdevice unregister_netdevice.
8625 /* If device is running close it first. */
8628 /* And unlink it from device chain */
8630 unlist_netdevice(dev);
8634 /* Shutdown queueing discipline. */
8637 /* Notify protocols, that we are about to destroy
8638 * this device. They should clean all the things.
8640 * Note that dev->reg_state stays at NETREG_REGISTERED.
8641 * This is wanted because this way 8021q and macvlan know
8642 * the device is just moving and can keep their slaves up.
8644 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
8647 new_nsid = peernet2id_alloc(dev_net(dev), net);
8648 /* If there is an ifindex conflict assign a new one */
8649 if (__dev_get_by_index(net, dev->ifindex))
8650 new_ifindex = dev_new_index(net);
8652 new_ifindex = dev->ifindex;
8654 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid,
8658 * Flush the unicast and multicast chains
8663 /* Send a netdev-removed uevent to the old namespace */
8664 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE);
8665 netdev_adjacent_del_links(dev);
8667 /* Actually switch the network namespace */
8668 dev_net_set(dev, net);
8669 dev->ifindex = new_ifindex;
8671 /* Send a netdev-add uevent to the new namespace */
8672 kobject_uevent(&dev->dev.kobj, KOBJ_ADD);
8673 netdev_adjacent_add_links(dev);
8675 /* Fixup kobjects */
8676 err = device_rename(&dev->dev, dev->name);
8679 /* Add the device back in the hashes */
8680 list_netdevice(dev);
8682 /* Notify protocols, that a new device appeared. */
8683 call_netdevice_notifiers(NETDEV_REGISTER, dev);
8686 * Prevent userspace races by waiting until the network
8687 * device is fully setup before sending notifications.
8689 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
8696 EXPORT_SYMBOL_GPL(dev_change_net_namespace);
8698 static int dev_cpu_dead(unsigned int oldcpu)
8700 struct sk_buff **list_skb;
8701 struct sk_buff *skb;
8703 struct softnet_data *sd, *oldsd, *remsd = NULL;
8705 local_irq_disable();
8706 cpu = smp_processor_id();
8707 sd = &per_cpu(softnet_data, cpu);
8708 oldsd = &per_cpu(softnet_data, oldcpu);
8710 /* Find end of our completion_queue. */
8711 list_skb = &sd->completion_queue;
8713 list_skb = &(*list_skb)->next;
8714 /* Append completion queue from offline CPU. */
8715 *list_skb = oldsd->completion_queue;
8716 oldsd->completion_queue = NULL;
8718 /* Append output queue from offline CPU. */
8719 if (oldsd->output_queue) {
8720 *sd->output_queue_tailp = oldsd->output_queue;
8721 sd->output_queue_tailp = oldsd->output_queue_tailp;
8722 oldsd->output_queue = NULL;
8723 oldsd->output_queue_tailp = &oldsd->output_queue;
8725 /* Append NAPI poll list from offline CPU, with one exception :
8726 * process_backlog() must be called by cpu owning percpu backlog.
8727 * We properly handle process_queue & input_pkt_queue later.
8729 while (!list_empty(&oldsd->poll_list)) {
8730 struct napi_struct *napi = list_first_entry(&oldsd->poll_list,
8734 list_del_init(&napi->poll_list);
8735 if (napi->poll == process_backlog)
8738 ____napi_schedule(sd, napi);
8741 raise_softirq_irqoff(NET_TX_SOFTIRQ);
8745 remsd = oldsd->rps_ipi_list;
8746 oldsd->rps_ipi_list = NULL;
8748 /* send out pending IPI's on offline CPU */
8749 net_rps_send_ipi(remsd);
8751 /* Process offline CPU's input_pkt_queue */
8752 while ((skb = __skb_dequeue(&oldsd->process_queue))) {
8754 input_queue_head_incr(oldsd);
8756 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) {
8758 input_queue_head_incr(oldsd);
8765 * netdev_increment_features - increment feature set by one
8766 * @all: current feature set
8767 * @one: new feature set
8768 * @mask: mask feature set
8770 * Computes a new feature set after adding a device with feature set
8771 * @one to the master device with current feature set @all. Will not
8772 * enable anything that is off in @mask. Returns the new feature set.
8774 netdev_features_t netdev_increment_features(netdev_features_t all,
8775 netdev_features_t one, netdev_features_t mask)
8777 if (mask & NETIF_F_HW_CSUM)
8778 mask |= NETIF_F_CSUM_MASK;
8779 mask |= NETIF_F_VLAN_CHALLENGED;
8781 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask;
8782 all &= one | ~NETIF_F_ALL_FOR_ALL;
8784 /* If one device supports hw checksumming, set for all. */
8785 if (all & NETIF_F_HW_CSUM)
8786 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM);
8790 EXPORT_SYMBOL(netdev_increment_features);
8792 static struct hlist_head * __net_init netdev_create_hash(void)
8795 struct hlist_head *hash;
8797 hash = kmalloc(sizeof(*hash) * NETDEV_HASHENTRIES, GFP_KERNEL);
8799 for (i = 0; i < NETDEV_HASHENTRIES; i++)
8800 INIT_HLIST_HEAD(&hash[i]);
8805 /* Initialize per network namespace state */
8806 static int __net_init netdev_init(struct net *net)
8808 if (net != &init_net)
8809 INIT_LIST_HEAD(&net->dev_base_head);
8811 net->dev_name_head = netdev_create_hash();
8812 if (net->dev_name_head == NULL)
8815 net->dev_index_head = netdev_create_hash();
8816 if (net->dev_index_head == NULL)
8822 kfree(net->dev_name_head);
8828 * netdev_drivername - network driver for the device
8829 * @dev: network device
8831 * Determine network driver for device.
8833 const char *netdev_drivername(const struct net_device *dev)
8835 const struct device_driver *driver;
8836 const struct device *parent;
8837 const char *empty = "";
8839 parent = dev->dev.parent;
8843 driver = parent->driver;
8844 if (driver && driver->name)
8845 return driver->name;
8849 static void __netdev_printk(const char *level, const struct net_device *dev,
8850 struct va_format *vaf)
8852 if (dev && dev->dev.parent) {
8853 dev_printk_emit(level[1] - '0',
8856 dev_driver_string(dev->dev.parent),
8857 dev_name(dev->dev.parent),
8858 netdev_name(dev), netdev_reg_state(dev),
8861 printk("%s%s%s: %pV",
8862 level, netdev_name(dev), netdev_reg_state(dev), vaf);
8864 printk("%s(NULL net_device): %pV", level, vaf);
8868 void netdev_printk(const char *level, const struct net_device *dev,
8869 const char *format, ...)
8871 struct va_format vaf;
8874 va_start(args, format);
8879 __netdev_printk(level, dev, &vaf);
8883 EXPORT_SYMBOL(netdev_printk);
8885 #define define_netdev_printk_level(func, level) \
8886 void func(const struct net_device *dev, const char *fmt, ...) \
8888 struct va_format vaf; \
8891 va_start(args, fmt); \
8896 __netdev_printk(level, dev, &vaf); \
8900 EXPORT_SYMBOL(func);
8902 define_netdev_printk_level(netdev_emerg, KERN_EMERG);
8903 define_netdev_printk_level(netdev_alert, KERN_ALERT);
8904 define_netdev_printk_level(netdev_crit, KERN_CRIT);
8905 define_netdev_printk_level(netdev_err, KERN_ERR);
8906 define_netdev_printk_level(netdev_warn, KERN_WARNING);
8907 define_netdev_printk_level(netdev_notice, KERN_NOTICE);
8908 define_netdev_printk_level(netdev_info, KERN_INFO);
8910 static void __net_exit netdev_exit(struct net *net)
8912 kfree(net->dev_name_head);
8913 kfree(net->dev_index_head);
8914 if (net != &init_net)
8915 WARN_ON_ONCE(!list_empty(&net->dev_base_head));
8918 static struct pernet_operations __net_initdata netdev_net_ops = {
8919 .init = netdev_init,
8920 .exit = netdev_exit,
8923 static void __net_exit default_device_exit(struct net *net)
8925 struct net_device *dev, *aux;
8927 * Push all migratable network devices back to the
8928 * initial network namespace
8931 for_each_netdev_safe(net, dev, aux) {
8933 char fb_name[IFNAMSIZ];
8935 /* Ignore unmoveable devices (i.e. loopback) */
8936 if (dev->features & NETIF_F_NETNS_LOCAL)
8939 /* Leave virtual devices for the generic cleanup */
8940 if (dev->rtnl_link_ops)
8943 /* Push remaining network devices to init_net */
8944 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex);
8945 err = dev_change_net_namespace(dev, &init_net, fb_name);
8947 pr_emerg("%s: failed to move %s to init_net: %d\n",
8948 __func__, dev->name, err);
8955 static void __net_exit rtnl_lock_unregistering(struct list_head *net_list)
8957 /* Return with the rtnl_lock held when there are no network
8958 * devices unregistering in any network namespace in net_list.
8962 DEFINE_WAIT_FUNC(wait, woken_wake_function);
8964 add_wait_queue(&netdev_unregistering_wq, &wait);
8966 unregistering = false;
8968 list_for_each_entry(net, net_list, exit_list) {
8969 if (net->dev_unreg_count > 0) {
8970 unregistering = true;
8978 wait_woken(&wait, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
8980 remove_wait_queue(&netdev_unregistering_wq, &wait);
8983 static void __net_exit default_device_exit_batch(struct list_head *net_list)
8985 /* At exit all network devices most be removed from a network
8986 * namespace. Do this in the reverse order of registration.
8987 * Do this across as many network namespaces as possible to
8988 * improve batching efficiency.
8990 struct net_device *dev;
8992 LIST_HEAD(dev_kill_list);
8994 /* To prevent network device cleanup code from dereferencing
8995 * loopback devices or network devices that have been freed
8996 * wait here for all pending unregistrations to complete,
8997 * before unregistring the loopback device and allowing the
8998 * network namespace be freed.
9000 * The netdev todo list containing all network devices
9001 * unregistrations that happen in default_device_exit_batch
9002 * will run in the rtnl_unlock() at the end of
9003 * default_device_exit_batch.
9005 rtnl_lock_unregistering(net_list);
9006 list_for_each_entry(net, net_list, exit_list) {
9007 for_each_netdev_reverse(net, dev) {
9008 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink)
9009 dev->rtnl_link_ops->dellink(dev, &dev_kill_list);
9011 unregister_netdevice_queue(dev, &dev_kill_list);
9014 unregister_netdevice_many(&dev_kill_list);
9018 static struct pernet_operations __net_initdata default_device_ops = {
9019 .exit = default_device_exit,
9020 .exit_batch = default_device_exit_batch,
9024 * Initialize the DEV module. At boot time this walks the device list and
9025 * unhooks any devices that fail to initialise (normally hardware not
9026 * present) and leaves us with a valid list of present and active devices.
9031 * This is called single threaded during boot, so no need
9032 * to take the rtnl semaphore.
9034 static int __init net_dev_init(void)
9036 int i, rc = -ENOMEM;
9038 BUG_ON(!dev_boot_phase);
9040 if (dev_proc_init())
9043 if (netdev_kobject_init())
9046 INIT_LIST_HEAD(&ptype_all);
9047 for (i = 0; i < PTYPE_HASH_SIZE; i++)
9048 INIT_LIST_HEAD(&ptype_base[i]);
9050 INIT_LIST_HEAD(&offload_base);
9052 if (register_pernet_subsys(&netdev_net_ops))
9056 * Initialise the packet receive queues.
9059 for_each_possible_cpu(i) {
9060 struct work_struct *flush = per_cpu_ptr(&flush_works, i);
9061 struct softnet_data *sd = &per_cpu(softnet_data, i);
9063 INIT_WORK(flush, flush_backlog);
9065 skb_queue_head_init(&sd->input_pkt_queue);
9066 skb_queue_head_init(&sd->process_queue);
9067 #ifdef CONFIG_XFRM_OFFLOAD
9068 skb_queue_head_init(&sd->xfrm_backlog);
9070 INIT_LIST_HEAD(&sd->poll_list);
9071 sd->output_queue_tailp = &sd->output_queue;
9073 sd->csd.func = rps_trigger_softirq;
9078 sd->backlog.poll = process_backlog;
9079 sd->backlog.weight = weight_p;
9084 /* The loopback device is special if any other network devices
9085 * is present in a network namespace the loopback device must
9086 * be present. Since we now dynamically allocate and free the
9087 * loopback device ensure this invariant is maintained by
9088 * keeping the loopback device as the first device on the
9089 * list of network devices. Ensuring the loopback devices
9090 * is the first device that appears and the last network device
9093 if (register_pernet_device(&loopback_net_ops))
9096 if (register_pernet_device(&default_device_ops))
9099 open_softirq(NET_TX_SOFTIRQ, net_tx_action);
9100 open_softirq(NET_RX_SOFTIRQ, net_rx_action);
9102 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead",
9103 NULL, dev_cpu_dead);
9110 subsys_initcall(net_dev_init);