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);
1290 * dev_get_alias - get ifalias of a device
1292 * @name: buffer to store name of ifalias
1293 * @len: size of buffer
1295 * get ifalias for a device. Caller must make sure dev cannot go
1296 * away, e.g. rcu read lock or own a reference count to device.
1298 int dev_get_alias(const struct net_device *dev, char *name, size_t len)
1300 const struct dev_ifalias *alias;
1304 alias = rcu_dereference(dev->ifalias);
1306 ret = snprintf(name, len, "%s", alias->ifalias);
1313 * netdev_features_change - device changes features
1314 * @dev: device to cause notification
1316 * Called to indicate a device has changed features.
1318 void netdev_features_change(struct net_device *dev)
1320 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev);
1322 EXPORT_SYMBOL(netdev_features_change);
1325 * netdev_state_change - device changes state
1326 * @dev: device to cause notification
1328 * Called to indicate a device has changed state. This function calls
1329 * the notifier chains for netdev_chain and sends a NEWLINK message
1330 * to the routing socket.
1332 void netdev_state_change(struct net_device *dev)
1334 if (dev->flags & IFF_UP) {
1335 struct netdev_notifier_change_info change_info = {
1339 call_netdevice_notifiers_info(NETDEV_CHANGE,
1341 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL);
1344 EXPORT_SYMBOL(netdev_state_change);
1347 * netdev_notify_peers - notify network peers about existence of @dev
1348 * @dev: network device
1350 * Generate traffic such that interested network peers are aware of
1351 * @dev, such as by generating a gratuitous ARP. This may be used when
1352 * a device wants to inform the rest of the network about some sort of
1353 * reconfiguration such as a failover event or virtual machine
1356 void netdev_notify_peers(struct net_device *dev)
1359 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
1360 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev);
1363 EXPORT_SYMBOL(netdev_notify_peers);
1365 static int __dev_open(struct net_device *dev)
1367 const struct net_device_ops *ops = dev->netdev_ops;
1372 if (!netif_device_present(dev))
1375 /* Block netpoll from trying to do any rx path servicing.
1376 * If we don't do this there is a chance ndo_poll_controller
1377 * or ndo_poll may be running while we open the device
1379 netpoll_poll_disable(dev);
1381 ret = call_netdevice_notifiers(NETDEV_PRE_UP, dev);
1382 ret = notifier_to_errno(ret);
1386 set_bit(__LINK_STATE_START, &dev->state);
1388 if (ops->ndo_validate_addr)
1389 ret = ops->ndo_validate_addr(dev);
1391 if (!ret && ops->ndo_open)
1392 ret = ops->ndo_open(dev);
1394 netpoll_poll_enable(dev);
1397 clear_bit(__LINK_STATE_START, &dev->state);
1399 dev->flags |= IFF_UP;
1400 dev_set_rx_mode(dev);
1402 add_device_randomness(dev->dev_addr, dev->addr_len);
1409 * dev_open - prepare an interface for use.
1410 * @dev: device to open
1412 * Takes a device from down to up state. The device's private open
1413 * function is invoked and then the multicast lists are loaded. Finally
1414 * the device is moved into the up state and a %NETDEV_UP message is
1415 * sent to the netdev notifier chain.
1417 * Calling this function on an active interface is a nop. On a failure
1418 * a negative errno code is returned.
1420 int dev_open(struct net_device *dev)
1424 if (dev->flags & IFF_UP)
1427 ret = __dev_open(dev);
1431 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1432 call_netdevice_notifiers(NETDEV_UP, dev);
1436 EXPORT_SYMBOL(dev_open);
1438 static void __dev_close_many(struct list_head *head)
1440 struct net_device *dev;
1445 list_for_each_entry(dev, head, close_list) {
1446 /* Temporarily disable netpoll until the interface is down */
1447 netpoll_poll_disable(dev);
1449 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev);
1451 clear_bit(__LINK_STATE_START, &dev->state);
1453 /* Synchronize to scheduled poll. We cannot touch poll list, it
1454 * can be even on different cpu. So just clear netif_running().
1456 * dev->stop() will invoke napi_disable() on all of it's
1457 * napi_struct instances on this device.
1459 smp_mb__after_atomic(); /* Commit netif_running(). */
1462 dev_deactivate_many(head);
1464 list_for_each_entry(dev, head, close_list) {
1465 const struct net_device_ops *ops = dev->netdev_ops;
1468 * Call the device specific close. This cannot fail.
1469 * Only if device is UP
1471 * We allow it to be called even after a DETACH hot-plug
1477 dev->flags &= ~IFF_UP;
1478 netpoll_poll_enable(dev);
1482 static void __dev_close(struct net_device *dev)
1486 list_add(&dev->close_list, &single);
1487 __dev_close_many(&single);
1491 void dev_close_many(struct list_head *head, bool unlink)
1493 struct net_device *dev, *tmp;
1495 /* Remove the devices that don't need to be closed */
1496 list_for_each_entry_safe(dev, tmp, head, close_list)
1497 if (!(dev->flags & IFF_UP))
1498 list_del_init(&dev->close_list);
1500 __dev_close_many(head);
1502 list_for_each_entry_safe(dev, tmp, head, close_list) {
1503 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1504 call_netdevice_notifiers(NETDEV_DOWN, dev);
1506 list_del_init(&dev->close_list);
1509 EXPORT_SYMBOL(dev_close_many);
1512 * dev_close - shutdown an interface.
1513 * @dev: device to shutdown
1515 * This function moves an active device into down state. A
1516 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device
1517 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier
1520 void dev_close(struct net_device *dev)
1522 if (dev->flags & IFF_UP) {
1525 list_add(&dev->close_list, &single);
1526 dev_close_many(&single, true);
1530 EXPORT_SYMBOL(dev_close);
1534 * dev_disable_lro - disable Large Receive Offload on a device
1537 * Disable Large Receive Offload (LRO) on a net device. Must be
1538 * called under RTNL. This is needed if received packets may be
1539 * forwarded to another interface.
1541 void dev_disable_lro(struct net_device *dev)
1543 struct net_device *lower_dev;
1544 struct list_head *iter;
1546 dev->wanted_features &= ~NETIF_F_LRO;
1547 netdev_update_features(dev);
1549 if (unlikely(dev->features & NETIF_F_LRO))
1550 netdev_WARN(dev, "failed to disable LRO!\n");
1552 netdev_for_each_lower_dev(dev, lower_dev, iter)
1553 dev_disable_lro(lower_dev);
1555 EXPORT_SYMBOL(dev_disable_lro);
1558 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device
1561 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be
1562 * called under RTNL. This is needed if Generic XDP is installed on
1565 static void dev_disable_gro_hw(struct net_device *dev)
1567 dev->wanted_features &= ~NETIF_F_GRO_HW;
1568 netdev_update_features(dev);
1570 if (unlikely(dev->features & NETIF_F_GRO_HW))
1571 netdev_WARN(dev, "failed to disable GRO_HW!\n");
1574 const char *netdev_cmd_to_name(enum netdev_cmd cmd)
1577 case NETDEV_##val: \
1578 return "NETDEV_" __stringify(val);
1580 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER)
1581 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE)
1582 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE)
1583 N(POST_INIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) N(CHANGEUPPER)
1584 N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) N(BONDING_INFO)
1585 N(PRECHANGEUPPER) N(CHANGELOWERSTATE) N(UDP_TUNNEL_PUSH_INFO)
1586 N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN)
1587 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO)
1588 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO)
1591 return "UNKNOWN_NETDEV_EVENT";
1593 EXPORT_SYMBOL_GPL(netdev_cmd_to_name);
1595 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val,
1596 struct net_device *dev)
1598 struct netdev_notifier_info info = {
1602 return nb->notifier_call(nb, val, &info);
1605 static int dev_boot_phase = 1;
1608 * register_netdevice_notifier - register a network notifier block
1611 * Register a notifier to be called when network device events occur.
1612 * The notifier passed is linked into the kernel structures and must
1613 * not be reused until it has been unregistered. A negative errno code
1614 * is returned on a failure.
1616 * When registered all registration and up events are replayed
1617 * to the new notifier to allow device to have a race free
1618 * view of the network device list.
1621 int register_netdevice_notifier(struct notifier_block *nb)
1623 struct net_device *dev;
1624 struct net_device *last;
1628 /* Close race with setup_net() and cleanup_net() */
1629 down_write(&pernet_ops_rwsem);
1631 err = raw_notifier_chain_register(&netdev_chain, nb);
1637 for_each_netdev(net, dev) {
1638 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev);
1639 err = notifier_to_errno(err);
1643 if (!(dev->flags & IFF_UP))
1646 call_netdevice_notifier(nb, NETDEV_UP, dev);
1652 up_write(&pernet_ops_rwsem);
1658 for_each_netdev(net, dev) {
1662 if (dev->flags & IFF_UP) {
1663 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1665 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1667 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1672 raw_notifier_chain_unregister(&netdev_chain, nb);
1675 EXPORT_SYMBOL(register_netdevice_notifier);
1678 * unregister_netdevice_notifier - unregister a network notifier block
1681 * Unregister a notifier previously registered by
1682 * register_netdevice_notifier(). The notifier is unlinked into the
1683 * kernel structures and may then be reused. A negative errno code
1684 * is returned on a failure.
1686 * After unregistering unregister and down device events are synthesized
1687 * for all devices on the device list to the removed notifier to remove
1688 * the need for special case cleanup code.
1691 int unregister_netdevice_notifier(struct notifier_block *nb)
1693 struct net_device *dev;
1697 /* Close race with setup_net() and cleanup_net() */
1698 down_write(&pernet_ops_rwsem);
1700 err = raw_notifier_chain_unregister(&netdev_chain, nb);
1705 for_each_netdev(net, dev) {
1706 if (dev->flags & IFF_UP) {
1707 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1709 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1711 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1716 up_write(&pernet_ops_rwsem);
1719 EXPORT_SYMBOL(unregister_netdevice_notifier);
1722 * call_netdevice_notifiers_info - call all network notifier blocks
1723 * @val: value passed unmodified to notifier function
1724 * @info: notifier information data
1726 * Call all network notifier blocks. Parameters and return value
1727 * are as for raw_notifier_call_chain().
1730 static int call_netdevice_notifiers_info(unsigned long val,
1731 struct netdev_notifier_info *info)
1734 return raw_notifier_call_chain(&netdev_chain, val, info);
1738 * call_netdevice_notifiers - call all network notifier blocks
1739 * @val: value passed unmodified to notifier function
1740 * @dev: net_device pointer passed unmodified to notifier function
1742 * Call all network notifier blocks. Parameters and return value
1743 * are as for raw_notifier_call_chain().
1746 int call_netdevice_notifiers(unsigned long val, struct net_device *dev)
1748 struct netdev_notifier_info info = {
1752 return call_netdevice_notifiers_info(val, &info);
1754 EXPORT_SYMBOL(call_netdevice_notifiers);
1756 #ifdef CONFIG_NET_INGRESS
1757 static struct static_key ingress_needed __read_mostly;
1759 void net_inc_ingress_queue(void)
1761 static_key_slow_inc(&ingress_needed);
1763 EXPORT_SYMBOL_GPL(net_inc_ingress_queue);
1765 void net_dec_ingress_queue(void)
1767 static_key_slow_dec(&ingress_needed);
1769 EXPORT_SYMBOL_GPL(net_dec_ingress_queue);
1772 #ifdef CONFIG_NET_EGRESS
1773 static struct static_key egress_needed __read_mostly;
1775 void net_inc_egress_queue(void)
1777 static_key_slow_inc(&egress_needed);
1779 EXPORT_SYMBOL_GPL(net_inc_egress_queue);
1781 void net_dec_egress_queue(void)
1783 static_key_slow_dec(&egress_needed);
1785 EXPORT_SYMBOL_GPL(net_dec_egress_queue);
1788 static struct static_key netstamp_needed __read_mostly;
1789 #ifdef HAVE_JUMP_LABEL
1790 static atomic_t netstamp_needed_deferred;
1791 static atomic_t netstamp_wanted;
1792 static void netstamp_clear(struct work_struct *work)
1794 int deferred = atomic_xchg(&netstamp_needed_deferred, 0);
1797 wanted = atomic_add_return(deferred, &netstamp_wanted);
1799 static_key_enable(&netstamp_needed);
1801 static_key_disable(&netstamp_needed);
1803 static DECLARE_WORK(netstamp_work, netstamp_clear);
1806 void net_enable_timestamp(void)
1808 #ifdef HAVE_JUMP_LABEL
1812 wanted = atomic_read(&netstamp_wanted);
1815 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted + 1) == wanted)
1818 atomic_inc(&netstamp_needed_deferred);
1819 schedule_work(&netstamp_work);
1821 static_key_slow_inc(&netstamp_needed);
1824 EXPORT_SYMBOL(net_enable_timestamp);
1826 void net_disable_timestamp(void)
1828 #ifdef HAVE_JUMP_LABEL
1832 wanted = atomic_read(&netstamp_wanted);
1835 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted - 1) == wanted)
1838 atomic_dec(&netstamp_needed_deferred);
1839 schedule_work(&netstamp_work);
1841 static_key_slow_dec(&netstamp_needed);
1844 EXPORT_SYMBOL(net_disable_timestamp);
1846 static inline void net_timestamp_set(struct sk_buff *skb)
1849 if (static_key_false(&netstamp_needed))
1850 __net_timestamp(skb);
1853 #define net_timestamp_check(COND, SKB) \
1854 if (static_key_false(&netstamp_needed)) { \
1855 if ((COND) && !(SKB)->tstamp) \
1856 __net_timestamp(SKB); \
1859 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb)
1863 if (!(dev->flags & IFF_UP))
1866 len = dev->mtu + dev->hard_header_len + VLAN_HLEN;
1867 if (skb->len <= len)
1870 /* if TSO is enabled, we don't care about the length as the packet
1871 * could be forwarded without being segmented before
1873 if (skb_is_gso(skb))
1878 EXPORT_SYMBOL_GPL(is_skb_forwardable);
1880 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1882 int ret = ____dev_forward_skb(dev, skb);
1885 skb->protocol = eth_type_trans(skb, dev);
1886 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
1891 EXPORT_SYMBOL_GPL(__dev_forward_skb);
1894 * dev_forward_skb - loopback an skb to another netif
1896 * @dev: destination network device
1897 * @skb: buffer to forward
1900 * NET_RX_SUCCESS (no congestion)
1901 * NET_RX_DROP (packet was dropped, but freed)
1903 * dev_forward_skb can be used for injecting an skb from the
1904 * start_xmit function of one device into the receive queue
1905 * of another device.
1907 * The receiving device may be in another namespace, so
1908 * we have to clear all information in the skb that could
1909 * impact namespace isolation.
1911 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1913 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb);
1915 EXPORT_SYMBOL_GPL(dev_forward_skb);
1917 static inline int deliver_skb(struct sk_buff *skb,
1918 struct packet_type *pt_prev,
1919 struct net_device *orig_dev)
1921 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
1923 refcount_inc(&skb->users);
1924 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
1927 static inline void deliver_ptype_list_skb(struct sk_buff *skb,
1928 struct packet_type **pt,
1929 struct net_device *orig_dev,
1931 struct list_head *ptype_list)
1933 struct packet_type *ptype, *pt_prev = *pt;
1935 list_for_each_entry_rcu(ptype, ptype_list, list) {
1936 if (ptype->type != type)
1939 deliver_skb(skb, pt_prev, orig_dev);
1945 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb)
1947 if (!ptype->af_packet_priv || !skb->sk)
1950 if (ptype->id_match)
1951 return ptype->id_match(ptype, skb->sk);
1952 else if ((struct sock *)ptype->af_packet_priv == skb->sk)
1959 * Support routine. Sends outgoing frames to any network
1960 * taps currently in use.
1963 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
1965 struct packet_type *ptype;
1966 struct sk_buff *skb2 = NULL;
1967 struct packet_type *pt_prev = NULL;
1968 struct list_head *ptype_list = &ptype_all;
1972 list_for_each_entry_rcu(ptype, ptype_list, list) {
1973 /* Never send packets back to the socket
1974 * they originated from - MvS (miquels@drinkel.ow.org)
1976 if (skb_loop_sk(ptype, skb))
1980 deliver_skb(skb2, pt_prev, skb->dev);
1985 /* need to clone skb, done only once */
1986 skb2 = skb_clone(skb, GFP_ATOMIC);
1990 net_timestamp_set(skb2);
1992 /* skb->nh should be correctly
1993 * set by sender, so that the second statement is
1994 * just protection against buggy protocols.
1996 skb_reset_mac_header(skb2);
1998 if (skb_network_header(skb2) < skb2->data ||
1999 skb_network_header(skb2) > skb_tail_pointer(skb2)) {
2000 net_crit_ratelimited("protocol %04x is buggy, dev %s\n",
2001 ntohs(skb2->protocol),
2003 skb_reset_network_header(skb2);
2006 skb2->transport_header = skb2->network_header;
2007 skb2->pkt_type = PACKET_OUTGOING;
2011 if (ptype_list == &ptype_all) {
2012 ptype_list = &dev->ptype_all;
2017 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC))
2018 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
2024 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit);
2027 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change
2028 * @dev: Network device
2029 * @txq: number of queues available
2031 * If real_num_tx_queues is changed the tc mappings may no longer be
2032 * valid. To resolve this verify the tc mapping remains valid and if
2033 * not NULL the mapping. With no priorities mapping to this
2034 * offset/count pair it will no longer be used. In the worst case TC0
2035 * is invalid nothing can be done so disable priority mappings. If is
2036 * expected that drivers will fix this mapping if they can before
2037 * calling netif_set_real_num_tx_queues.
2039 static void netif_setup_tc(struct net_device *dev, unsigned int txq)
2042 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2044 /* If TC0 is invalidated disable TC mapping */
2045 if (tc->offset + tc->count > txq) {
2046 pr_warn("Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n");
2051 /* Invalidated prio to tc mappings set to TC0 */
2052 for (i = 1; i < TC_BITMASK + 1; i++) {
2053 int q = netdev_get_prio_tc_map(dev, i);
2055 tc = &dev->tc_to_txq[q];
2056 if (tc->offset + tc->count > txq) {
2057 pr_warn("Number of in use tx queues changed. Priority %i to tc mapping %i is no longer valid. Setting map to 0\n",
2059 netdev_set_prio_tc_map(dev, i, 0);
2064 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq)
2067 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2070 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) {
2071 if ((txq - tc->offset) < tc->count)
2080 EXPORT_SYMBOL(netdev_txq_to_tc);
2083 static DEFINE_MUTEX(xps_map_mutex);
2084 #define xmap_dereference(P) \
2085 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex))
2087 static bool remove_xps_queue(struct xps_dev_maps *dev_maps,
2090 struct xps_map *map = NULL;
2094 map = xmap_dereference(dev_maps->cpu_map[tci]);
2098 for (pos = map->len; pos--;) {
2099 if (map->queues[pos] != index)
2103 map->queues[pos] = map->queues[--map->len];
2107 RCU_INIT_POINTER(dev_maps->cpu_map[tci], NULL);
2108 kfree_rcu(map, rcu);
2115 static bool remove_xps_queue_cpu(struct net_device *dev,
2116 struct xps_dev_maps *dev_maps,
2117 int cpu, u16 offset, u16 count)
2119 int num_tc = dev->num_tc ? : 1;
2120 bool active = false;
2123 for (tci = cpu * num_tc; num_tc--; tci++) {
2126 for (i = count, j = offset; i--; j++) {
2127 if (!remove_xps_queue(dev_maps, tci, j))
2137 static void netif_reset_xps_queues(struct net_device *dev, u16 offset,
2140 struct xps_dev_maps *dev_maps;
2142 bool active = false;
2144 mutex_lock(&xps_map_mutex);
2145 dev_maps = xmap_dereference(dev->xps_maps);
2150 for_each_possible_cpu(cpu)
2151 active |= remove_xps_queue_cpu(dev, dev_maps, cpu,
2155 RCU_INIT_POINTER(dev->xps_maps, NULL);
2156 kfree_rcu(dev_maps, rcu);
2159 for (i = offset + (count - 1); count--; i--)
2160 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, i),
2164 mutex_unlock(&xps_map_mutex);
2167 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index)
2169 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index);
2172 static struct xps_map *expand_xps_map(struct xps_map *map,
2175 struct xps_map *new_map;
2176 int alloc_len = XPS_MIN_MAP_ALLOC;
2179 for (pos = 0; map && pos < map->len; pos++) {
2180 if (map->queues[pos] != index)
2185 /* Need to add queue to this CPU's existing map */
2187 if (pos < map->alloc_len)
2190 alloc_len = map->alloc_len * 2;
2193 /* Need to allocate new map to store queue on this CPU's map */
2194 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL,
2199 for (i = 0; i < pos; i++)
2200 new_map->queues[i] = map->queues[i];
2201 new_map->alloc_len = alloc_len;
2207 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask,
2210 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL;
2211 int i, cpu, tci, numa_node_id = -2;
2212 int maps_sz, num_tc = 1, tc = 0;
2213 struct xps_map *map, *new_map;
2214 bool active = false;
2217 num_tc = dev->num_tc;
2218 tc = netdev_txq_to_tc(dev, index);
2223 maps_sz = XPS_DEV_MAPS_SIZE(num_tc);
2224 if (maps_sz < L1_CACHE_BYTES)
2225 maps_sz = L1_CACHE_BYTES;
2227 mutex_lock(&xps_map_mutex);
2229 dev_maps = xmap_dereference(dev->xps_maps);
2231 /* allocate memory for queue storage */
2232 for_each_cpu_and(cpu, cpu_online_mask, mask) {
2234 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL);
2235 if (!new_dev_maps) {
2236 mutex_unlock(&xps_map_mutex);
2240 tci = cpu * num_tc + tc;
2241 map = dev_maps ? xmap_dereference(dev_maps->cpu_map[tci]) :
2244 map = expand_xps_map(map, cpu, index);
2248 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2252 goto out_no_new_maps;
2254 for_each_possible_cpu(cpu) {
2255 /* copy maps belonging to foreign traffic classes */
2256 for (i = tc, tci = cpu * num_tc; dev_maps && i--; tci++) {
2257 /* fill in the new device map from the old device map */
2258 map = xmap_dereference(dev_maps->cpu_map[tci]);
2259 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2262 /* We need to explicitly update tci as prevous loop
2263 * could break out early if dev_maps is NULL.
2265 tci = cpu * num_tc + tc;
2267 if (cpumask_test_cpu(cpu, mask) && cpu_online(cpu)) {
2268 /* add queue to CPU maps */
2271 map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2272 while ((pos < map->len) && (map->queues[pos] != index))
2275 if (pos == map->len)
2276 map->queues[map->len++] = index;
2278 if (numa_node_id == -2)
2279 numa_node_id = cpu_to_node(cpu);
2280 else if (numa_node_id != cpu_to_node(cpu))
2283 } else if (dev_maps) {
2284 /* fill in the new device map from the old device map */
2285 map = xmap_dereference(dev_maps->cpu_map[tci]);
2286 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2289 /* copy maps belonging to foreign traffic classes */
2290 for (i = num_tc - tc, tci++; dev_maps && --i; tci++) {
2291 /* fill in the new device map from the old device map */
2292 map = xmap_dereference(dev_maps->cpu_map[tci]);
2293 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2297 rcu_assign_pointer(dev->xps_maps, new_dev_maps);
2299 /* Cleanup old maps */
2301 goto out_no_old_maps;
2303 for_each_possible_cpu(cpu) {
2304 for (i = num_tc, tci = cpu * num_tc; i--; tci++) {
2305 new_map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2306 map = xmap_dereference(dev_maps->cpu_map[tci]);
2307 if (map && map != new_map)
2308 kfree_rcu(map, rcu);
2312 kfree_rcu(dev_maps, rcu);
2315 dev_maps = new_dev_maps;
2319 /* update Tx queue numa node */
2320 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index),
2321 (numa_node_id >= 0) ? numa_node_id :
2327 /* removes queue from unused CPUs */
2328 for_each_possible_cpu(cpu) {
2329 for (i = tc, tci = cpu * num_tc; i--; tci++)
2330 active |= remove_xps_queue(dev_maps, tci, index);
2331 if (!cpumask_test_cpu(cpu, mask) || !cpu_online(cpu))
2332 active |= remove_xps_queue(dev_maps, tci, index);
2333 for (i = num_tc - tc, tci++; --i; tci++)
2334 active |= remove_xps_queue(dev_maps, tci, index);
2337 /* free map if not active */
2339 RCU_INIT_POINTER(dev->xps_maps, NULL);
2340 kfree_rcu(dev_maps, rcu);
2344 mutex_unlock(&xps_map_mutex);
2348 /* remove any maps that we added */
2349 for_each_possible_cpu(cpu) {
2350 for (i = num_tc, tci = cpu * num_tc; i--; tci++) {
2351 new_map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2353 xmap_dereference(dev_maps->cpu_map[tci]) :
2355 if (new_map && new_map != map)
2360 mutex_unlock(&xps_map_mutex);
2362 kfree(new_dev_maps);
2365 EXPORT_SYMBOL(netif_set_xps_queue);
2368 void netdev_reset_tc(struct net_device *dev)
2371 netif_reset_xps_queues_gt(dev, 0);
2374 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq));
2375 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map));
2377 EXPORT_SYMBOL(netdev_reset_tc);
2379 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset)
2381 if (tc >= dev->num_tc)
2385 netif_reset_xps_queues(dev, offset, count);
2387 dev->tc_to_txq[tc].count = count;
2388 dev->tc_to_txq[tc].offset = offset;
2391 EXPORT_SYMBOL(netdev_set_tc_queue);
2393 int netdev_set_num_tc(struct net_device *dev, u8 num_tc)
2395 if (num_tc > TC_MAX_QUEUE)
2399 netif_reset_xps_queues_gt(dev, 0);
2401 dev->num_tc = num_tc;
2404 EXPORT_SYMBOL(netdev_set_num_tc);
2407 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues
2408 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed.
2410 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq)
2415 disabling = txq < dev->real_num_tx_queues;
2417 if (txq < 1 || txq > dev->num_tx_queues)
2420 if (dev->reg_state == NETREG_REGISTERED ||
2421 dev->reg_state == NETREG_UNREGISTERING) {
2424 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues,
2430 netif_setup_tc(dev, txq);
2432 dev->real_num_tx_queues = txq;
2436 qdisc_reset_all_tx_gt(dev, txq);
2438 netif_reset_xps_queues_gt(dev, txq);
2442 dev->real_num_tx_queues = txq;
2447 EXPORT_SYMBOL(netif_set_real_num_tx_queues);
2451 * netif_set_real_num_rx_queues - set actual number of RX queues used
2452 * @dev: Network device
2453 * @rxq: Actual number of RX queues
2455 * This must be called either with the rtnl_lock held or before
2456 * registration of the net device. Returns 0 on success, or a
2457 * negative error code. If called before registration, it always
2460 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq)
2464 if (rxq < 1 || rxq > dev->num_rx_queues)
2467 if (dev->reg_state == NETREG_REGISTERED) {
2470 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues,
2476 dev->real_num_rx_queues = rxq;
2479 EXPORT_SYMBOL(netif_set_real_num_rx_queues);
2483 * netif_get_num_default_rss_queues - default number of RSS queues
2485 * This routine should set an upper limit on the number of RSS queues
2486 * used by default by multiqueue devices.
2488 int netif_get_num_default_rss_queues(void)
2490 return is_kdump_kernel() ?
2491 1 : min_t(int, DEFAULT_MAX_NUM_RSS_QUEUES, num_online_cpus());
2493 EXPORT_SYMBOL(netif_get_num_default_rss_queues);
2495 static void __netif_reschedule(struct Qdisc *q)
2497 struct softnet_data *sd;
2498 unsigned long flags;
2500 local_irq_save(flags);
2501 sd = this_cpu_ptr(&softnet_data);
2502 q->next_sched = NULL;
2503 *sd->output_queue_tailp = q;
2504 sd->output_queue_tailp = &q->next_sched;
2505 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2506 local_irq_restore(flags);
2509 void __netif_schedule(struct Qdisc *q)
2511 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
2512 __netif_reschedule(q);
2514 EXPORT_SYMBOL(__netif_schedule);
2516 struct dev_kfree_skb_cb {
2517 enum skb_free_reason reason;
2520 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb)
2522 return (struct dev_kfree_skb_cb *)skb->cb;
2525 void netif_schedule_queue(struct netdev_queue *txq)
2528 if (!(txq->state & QUEUE_STATE_ANY_XOFF)) {
2529 struct Qdisc *q = rcu_dereference(txq->qdisc);
2531 __netif_schedule(q);
2535 EXPORT_SYMBOL(netif_schedule_queue);
2537 void netif_tx_wake_queue(struct netdev_queue *dev_queue)
2539 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) {
2543 q = rcu_dereference(dev_queue->qdisc);
2544 __netif_schedule(q);
2548 EXPORT_SYMBOL(netif_tx_wake_queue);
2550 void __dev_kfree_skb_irq(struct sk_buff *skb, enum skb_free_reason reason)
2552 unsigned long flags;
2557 if (likely(refcount_read(&skb->users) == 1)) {
2559 refcount_set(&skb->users, 0);
2560 } else if (likely(!refcount_dec_and_test(&skb->users))) {
2563 get_kfree_skb_cb(skb)->reason = reason;
2564 local_irq_save(flags);
2565 skb->next = __this_cpu_read(softnet_data.completion_queue);
2566 __this_cpu_write(softnet_data.completion_queue, skb);
2567 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2568 local_irq_restore(flags);
2570 EXPORT_SYMBOL(__dev_kfree_skb_irq);
2572 void __dev_kfree_skb_any(struct sk_buff *skb, enum skb_free_reason reason)
2574 if (in_irq() || irqs_disabled())
2575 __dev_kfree_skb_irq(skb, reason);
2579 EXPORT_SYMBOL(__dev_kfree_skb_any);
2583 * netif_device_detach - mark device as removed
2584 * @dev: network device
2586 * Mark device as removed from system and therefore no longer available.
2588 void netif_device_detach(struct net_device *dev)
2590 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) &&
2591 netif_running(dev)) {
2592 netif_tx_stop_all_queues(dev);
2595 EXPORT_SYMBOL(netif_device_detach);
2598 * netif_device_attach - mark device as attached
2599 * @dev: network device
2601 * Mark device as attached from system and restart if needed.
2603 void netif_device_attach(struct net_device *dev)
2605 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) &&
2606 netif_running(dev)) {
2607 netif_tx_wake_all_queues(dev);
2608 __netdev_watchdog_up(dev);
2611 EXPORT_SYMBOL(netif_device_attach);
2614 * Returns a Tx hash based on the given packet descriptor a Tx queues' number
2615 * to be used as a distribution range.
2617 u16 __skb_tx_hash(const struct net_device *dev, struct sk_buff *skb,
2618 unsigned int num_tx_queues)
2622 u16 qcount = num_tx_queues;
2624 if (skb_rx_queue_recorded(skb)) {
2625 hash = skb_get_rx_queue(skb);
2626 while (unlikely(hash >= num_tx_queues))
2627 hash -= num_tx_queues;
2632 u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
2634 qoffset = dev->tc_to_txq[tc].offset;
2635 qcount = dev->tc_to_txq[tc].count;
2638 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset;
2640 EXPORT_SYMBOL(__skb_tx_hash);
2642 static void skb_warn_bad_offload(const struct sk_buff *skb)
2644 static const netdev_features_t null_features;
2645 struct net_device *dev = skb->dev;
2646 const char *name = "";
2648 if (!net_ratelimit())
2652 if (dev->dev.parent)
2653 name = dev_driver_string(dev->dev.parent);
2655 name = netdev_name(dev);
2657 WARN(1, "%s: caps=(%pNF, %pNF) len=%d data_len=%d gso_size=%d "
2658 "gso_type=%d ip_summed=%d\n",
2659 name, dev ? &dev->features : &null_features,
2660 skb->sk ? &skb->sk->sk_route_caps : &null_features,
2661 skb->len, skb->data_len, skb_shinfo(skb)->gso_size,
2662 skb_shinfo(skb)->gso_type, skb->ip_summed);
2666 * Invalidate hardware checksum when packet is to be mangled, and
2667 * complete checksum manually on outgoing path.
2669 int skb_checksum_help(struct sk_buff *skb)
2672 int ret = 0, offset;
2674 if (skb->ip_summed == CHECKSUM_COMPLETE)
2675 goto out_set_summed;
2677 if (unlikely(skb_shinfo(skb)->gso_size)) {
2678 skb_warn_bad_offload(skb);
2682 /* Before computing a checksum, we should make sure no frag could
2683 * be modified by an external entity : checksum could be wrong.
2685 if (skb_has_shared_frag(skb)) {
2686 ret = __skb_linearize(skb);
2691 offset = skb_checksum_start_offset(skb);
2692 BUG_ON(offset >= skb_headlen(skb));
2693 csum = skb_checksum(skb, offset, skb->len - offset, 0);
2695 offset += skb->csum_offset;
2696 BUG_ON(offset + sizeof(__sum16) > skb_headlen(skb));
2698 if (skb_cloned(skb) &&
2699 !skb_clone_writable(skb, offset + sizeof(__sum16))) {
2700 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2705 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0;
2707 skb->ip_summed = CHECKSUM_NONE;
2711 EXPORT_SYMBOL(skb_checksum_help);
2713 int skb_crc32c_csum_help(struct sk_buff *skb)
2716 int ret = 0, offset, start;
2718 if (skb->ip_summed != CHECKSUM_PARTIAL)
2721 if (unlikely(skb_is_gso(skb)))
2724 /* Before computing a checksum, we should make sure no frag could
2725 * be modified by an external entity : checksum could be wrong.
2727 if (unlikely(skb_has_shared_frag(skb))) {
2728 ret = __skb_linearize(skb);
2732 start = skb_checksum_start_offset(skb);
2733 offset = start + offsetof(struct sctphdr, checksum);
2734 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) {
2738 if (skb_cloned(skb) &&
2739 !skb_clone_writable(skb, offset + sizeof(__le32))) {
2740 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2744 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start,
2745 skb->len - start, ~(__u32)0,
2747 *(__le32 *)(skb->data + offset) = crc32c_csum;
2748 skb->ip_summed = CHECKSUM_NONE;
2749 skb->csum_not_inet = 0;
2754 __be16 skb_network_protocol(struct sk_buff *skb, int *depth)
2756 __be16 type = skb->protocol;
2758 /* Tunnel gso handlers can set protocol to ethernet. */
2759 if (type == htons(ETH_P_TEB)) {
2762 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr))))
2765 eth = (struct ethhdr *)skb->data;
2766 type = eth->h_proto;
2769 return __vlan_get_protocol(skb, type, depth);
2773 * skb_mac_gso_segment - mac layer segmentation handler.
2774 * @skb: buffer to segment
2775 * @features: features for the output path (see dev->features)
2777 struct sk_buff *skb_mac_gso_segment(struct sk_buff *skb,
2778 netdev_features_t features)
2780 struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT);
2781 struct packet_offload *ptype;
2782 int vlan_depth = skb->mac_len;
2783 __be16 type = skb_network_protocol(skb, &vlan_depth);
2785 if (unlikely(!type))
2786 return ERR_PTR(-EINVAL);
2788 __skb_pull(skb, vlan_depth);
2791 list_for_each_entry_rcu(ptype, &offload_base, list) {
2792 if (ptype->type == type && ptype->callbacks.gso_segment) {
2793 segs = ptype->callbacks.gso_segment(skb, features);
2799 __skb_push(skb, skb->data - skb_mac_header(skb));
2803 EXPORT_SYMBOL(skb_mac_gso_segment);
2806 /* openvswitch calls this on rx path, so we need a different check.
2808 static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
2811 return skb->ip_summed != CHECKSUM_PARTIAL &&
2812 skb->ip_summed != CHECKSUM_UNNECESSARY;
2814 return skb->ip_summed == CHECKSUM_NONE;
2818 * __skb_gso_segment - Perform segmentation on skb.
2819 * @skb: buffer to segment
2820 * @features: features for the output path (see dev->features)
2821 * @tx_path: whether it is called in TX path
2823 * This function segments the given skb and returns a list of segments.
2825 * It may return NULL if the skb requires no segmentation. This is
2826 * only possible when GSO is used for verifying header integrity.
2828 * Segmentation preserves SKB_SGO_CB_OFFSET bytes of previous skb cb.
2830 struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
2831 netdev_features_t features, bool tx_path)
2833 struct sk_buff *segs;
2835 if (unlikely(skb_needs_check(skb, tx_path))) {
2838 /* We're going to init ->check field in TCP or UDP header */
2839 err = skb_cow_head(skb, 0);
2841 return ERR_PTR(err);
2844 /* Only report GSO partial support if it will enable us to
2845 * support segmentation on this frame without needing additional
2848 if (features & NETIF_F_GSO_PARTIAL) {
2849 netdev_features_t partial_features = NETIF_F_GSO_ROBUST;
2850 struct net_device *dev = skb->dev;
2852 partial_features |= dev->features & dev->gso_partial_features;
2853 if (!skb_gso_ok(skb, features | partial_features))
2854 features &= ~NETIF_F_GSO_PARTIAL;
2857 BUILD_BUG_ON(SKB_SGO_CB_OFFSET +
2858 sizeof(*SKB_GSO_CB(skb)) > sizeof(skb->cb));
2860 SKB_GSO_CB(skb)->mac_offset = skb_headroom(skb);
2861 SKB_GSO_CB(skb)->encap_level = 0;
2863 skb_reset_mac_header(skb);
2864 skb_reset_mac_len(skb);
2866 segs = skb_mac_gso_segment(skb, features);
2868 if (unlikely(skb_needs_check(skb, tx_path) && !IS_ERR(segs)))
2869 skb_warn_bad_offload(skb);
2873 EXPORT_SYMBOL(__skb_gso_segment);
2875 /* Take action when hardware reception checksum errors are detected. */
2877 void netdev_rx_csum_fault(struct net_device *dev)
2879 if (net_ratelimit()) {
2880 pr_err("%s: hw csum failure\n", dev ? dev->name : "<unknown>");
2884 EXPORT_SYMBOL(netdev_rx_csum_fault);
2887 /* XXX: check that highmem exists at all on the given machine. */
2888 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
2890 #ifdef CONFIG_HIGHMEM
2893 if (!(dev->features & NETIF_F_HIGHDMA)) {
2894 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2895 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2897 if (PageHighMem(skb_frag_page(frag)))
2905 /* If MPLS offload request, verify we are testing hardware MPLS features
2906 * instead of standard features for the netdev.
2908 #if IS_ENABLED(CONFIG_NET_MPLS_GSO)
2909 static netdev_features_t net_mpls_features(struct sk_buff *skb,
2910 netdev_features_t features,
2913 if (eth_p_mpls(type))
2914 features &= skb->dev->mpls_features;
2919 static netdev_features_t net_mpls_features(struct sk_buff *skb,
2920 netdev_features_t features,
2927 static netdev_features_t harmonize_features(struct sk_buff *skb,
2928 netdev_features_t features)
2933 type = skb_network_protocol(skb, &tmp);
2934 features = net_mpls_features(skb, features, type);
2936 if (skb->ip_summed != CHECKSUM_NONE &&
2937 !can_checksum_protocol(features, type)) {
2938 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
2940 if (illegal_highdma(skb->dev, skb))
2941 features &= ~NETIF_F_SG;
2946 netdev_features_t passthru_features_check(struct sk_buff *skb,
2947 struct net_device *dev,
2948 netdev_features_t features)
2952 EXPORT_SYMBOL(passthru_features_check);
2954 static netdev_features_t dflt_features_check(struct sk_buff *skb,
2955 struct net_device *dev,
2956 netdev_features_t features)
2958 return vlan_features_check(skb, features);
2961 static netdev_features_t gso_features_check(const struct sk_buff *skb,
2962 struct net_device *dev,
2963 netdev_features_t features)
2965 u16 gso_segs = skb_shinfo(skb)->gso_segs;
2967 if (gso_segs > dev->gso_max_segs)
2968 return features & ~NETIF_F_GSO_MASK;
2970 /* Support for GSO partial features requires software
2971 * intervention before we can actually process the packets
2972 * so we need to strip support for any partial features now
2973 * and we can pull them back in after we have partially
2974 * segmented the frame.
2976 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL))
2977 features &= ~dev->gso_partial_features;
2979 /* Make sure to clear the IPv4 ID mangling feature if the
2980 * IPv4 header has the potential to be fragmented.
2982 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
2983 struct iphdr *iph = skb->encapsulation ?
2984 inner_ip_hdr(skb) : ip_hdr(skb);
2986 if (!(iph->frag_off & htons(IP_DF)))
2987 features &= ~NETIF_F_TSO_MANGLEID;
2993 netdev_features_t netif_skb_features(struct sk_buff *skb)
2995 struct net_device *dev = skb->dev;
2996 netdev_features_t features = dev->features;
2998 if (skb_is_gso(skb))
2999 features = gso_features_check(skb, dev, features);
3001 /* If encapsulation offload request, verify we are testing
3002 * hardware encapsulation features instead of standard
3003 * features for the netdev
3005 if (skb->encapsulation)
3006 features &= dev->hw_enc_features;
3008 if (skb_vlan_tagged(skb))
3009 features = netdev_intersect_features(features,
3010 dev->vlan_features |
3011 NETIF_F_HW_VLAN_CTAG_TX |
3012 NETIF_F_HW_VLAN_STAG_TX);
3014 if (dev->netdev_ops->ndo_features_check)
3015 features &= dev->netdev_ops->ndo_features_check(skb, dev,
3018 features &= dflt_features_check(skb, dev, features);
3020 return harmonize_features(skb, features);
3022 EXPORT_SYMBOL(netif_skb_features);
3024 static int xmit_one(struct sk_buff *skb, struct net_device *dev,
3025 struct netdev_queue *txq, bool more)
3030 if (!list_empty(&ptype_all) || !list_empty(&dev->ptype_all))
3031 dev_queue_xmit_nit(skb, dev);
3034 trace_net_dev_start_xmit(skb, dev);
3035 rc = netdev_start_xmit(skb, dev, txq, more);
3036 trace_net_dev_xmit(skb, rc, dev, len);
3041 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev,
3042 struct netdev_queue *txq, int *ret)
3044 struct sk_buff *skb = first;
3045 int rc = NETDEV_TX_OK;
3048 struct sk_buff *next = skb->next;
3051 rc = xmit_one(skb, dev, txq, next != NULL);
3052 if (unlikely(!dev_xmit_complete(rc))) {
3058 if (netif_xmit_stopped(txq) && skb) {
3059 rc = NETDEV_TX_BUSY;
3069 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb,
3070 netdev_features_t features)
3072 if (skb_vlan_tag_present(skb) &&
3073 !vlan_hw_offload_capable(features, skb->vlan_proto))
3074 skb = __vlan_hwaccel_push_inside(skb);
3078 int skb_csum_hwoffload_help(struct sk_buff *skb,
3079 const netdev_features_t features)
3081 if (unlikely(skb->csum_not_inet))
3082 return !!(features & NETIF_F_SCTP_CRC) ? 0 :
3083 skb_crc32c_csum_help(skb);
3085 return !!(features & NETIF_F_CSUM_MASK) ? 0 : skb_checksum_help(skb);
3087 EXPORT_SYMBOL(skb_csum_hwoffload_help);
3089 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again)
3091 netdev_features_t features;
3093 features = netif_skb_features(skb);
3094 skb = validate_xmit_vlan(skb, features);
3098 if (netif_needs_gso(skb, features)) {
3099 struct sk_buff *segs;
3101 segs = skb_gso_segment(skb, features);
3109 if (skb_needs_linearize(skb, features) &&
3110 __skb_linearize(skb))
3113 /* If packet is not checksummed and device does not
3114 * support checksumming for this protocol, complete
3115 * checksumming here.
3117 if (skb->ip_summed == CHECKSUM_PARTIAL) {
3118 if (skb->encapsulation)
3119 skb_set_inner_transport_header(skb,
3120 skb_checksum_start_offset(skb));
3122 skb_set_transport_header(skb,
3123 skb_checksum_start_offset(skb));
3124 if (skb_csum_hwoffload_help(skb, features))
3129 skb = validate_xmit_xfrm(skb, features, again);
3136 atomic_long_inc(&dev->tx_dropped);
3140 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again)
3142 struct sk_buff *next, *head = NULL, *tail;
3144 for (; skb != NULL; skb = next) {
3148 /* in case skb wont be segmented, point to itself */
3151 skb = validate_xmit_skb(skb, dev, again);
3159 /* If skb was segmented, skb->prev points to
3160 * the last segment. If not, it still contains skb.
3166 EXPORT_SYMBOL_GPL(validate_xmit_skb_list);
3168 static void qdisc_pkt_len_init(struct sk_buff *skb)
3170 const struct skb_shared_info *shinfo = skb_shinfo(skb);
3172 qdisc_skb_cb(skb)->pkt_len = skb->len;
3174 /* To get more precise estimation of bytes sent on wire,
3175 * we add to pkt_len the headers size of all segments
3177 if (shinfo->gso_size) {
3178 unsigned int hdr_len;
3179 u16 gso_segs = shinfo->gso_segs;
3181 /* mac layer + network layer */
3182 hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
3184 /* + transport layer */
3185 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) {
3186 const struct tcphdr *th;
3187 struct tcphdr _tcphdr;
3189 th = skb_header_pointer(skb, skb_transport_offset(skb),
3190 sizeof(_tcphdr), &_tcphdr);
3192 hdr_len += __tcp_hdrlen(th);
3194 struct udphdr _udphdr;
3196 if (skb_header_pointer(skb, skb_transport_offset(skb),
3197 sizeof(_udphdr), &_udphdr))
3198 hdr_len += sizeof(struct udphdr);
3201 if (shinfo->gso_type & SKB_GSO_DODGY)
3202 gso_segs = DIV_ROUND_UP(skb->len - hdr_len,
3205 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len;
3209 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
3210 struct net_device *dev,
3211 struct netdev_queue *txq)
3213 spinlock_t *root_lock = qdisc_lock(q);
3214 struct sk_buff *to_free = NULL;
3218 qdisc_calculate_pkt_len(skb, q);
3220 if (q->flags & TCQ_F_NOLOCK) {
3221 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3222 __qdisc_drop(skb, &to_free);
3225 rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
3229 if (unlikely(to_free))
3230 kfree_skb_list(to_free);
3235 * Heuristic to force contended enqueues to serialize on a
3236 * separate lock before trying to get qdisc main lock.
3237 * This permits qdisc->running owner to get the lock more
3238 * often and dequeue packets faster.
3240 contended = qdisc_is_running(q);
3241 if (unlikely(contended))
3242 spin_lock(&q->busylock);
3244 spin_lock(root_lock);
3245 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3246 __qdisc_drop(skb, &to_free);
3248 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
3249 qdisc_run_begin(q)) {
3251 * This is a work-conserving queue; there are no old skbs
3252 * waiting to be sent out; and the qdisc is not running -
3253 * xmit the skb directly.
3256 qdisc_bstats_update(q, skb);
3258 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) {
3259 if (unlikely(contended)) {
3260 spin_unlock(&q->busylock);
3267 rc = NET_XMIT_SUCCESS;
3269 rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
3270 if (qdisc_run_begin(q)) {
3271 if (unlikely(contended)) {
3272 spin_unlock(&q->busylock);
3279 spin_unlock(root_lock);
3280 if (unlikely(to_free))
3281 kfree_skb_list(to_free);
3282 if (unlikely(contended))
3283 spin_unlock(&q->busylock);
3287 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO)
3288 static void skb_update_prio(struct sk_buff *skb)
3290 const struct netprio_map *map;
3291 const struct sock *sk;
3292 unsigned int prioidx;
3296 map = rcu_dereference_bh(skb->dev->priomap);
3299 sk = skb_to_full_sk(skb);
3303 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data);
3305 if (prioidx < map->priomap_len)
3306 skb->priority = map->priomap[prioidx];
3309 #define skb_update_prio(skb)
3312 DEFINE_PER_CPU(int, xmit_recursion);
3313 EXPORT_SYMBOL(xmit_recursion);
3316 * dev_loopback_xmit - loop back @skb
3317 * @net: network namespace this loopback is happening in
3318 * @sk: sk needed to be a netfilter okfn
3319 * @skb: buffer to transmit
3321 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb)
3323 skb_reset_mac_header(skb);
3324 __skb_pull(skb, skb_network_offset(skb));
3325 skb->pkt_type = PACKET_LOOPBACK;
3326 skb->ip_summed = CHECKSUM_UNNECESSARY;
3327 WARN_ON(!skb_dst(skb));
3332 EXPORT_SYMBOL(dev_loopback_xmit);
3334 #ifdef CONFIG_NET_EGRESS
3335 static struct sk_buff *
3336 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
3338 struct mini_Qdisc *miniq = rcu_dereference_bh(dev->miniq_egress);
3339 struct tcf_result cl_res;
3344 /* qdisc_skb_cb(skb)->pkt_len was already set by the caller. */
3345 mini_qdisc_bstats_cpu_update(miniq, skb);
3347 switch (tcf_classify(skb, miniq->filter_list, &cl_res, false)) {
3349 case TC_ACT_RECLASSIFY:
3350 skb->tc_index = TC_H_MIN(cl_res.classid);
3353 mini_qdisc_qstats_cpu_drop(miniq);
3354 *ret = NET_XMIT_DROP;
3360 *ret = NET_XMIT_SUCCESS;
3363 case TC_ACT_REDIRECT:
3364 /* No need to push/pop skb's mac_header here on egress! */
3365 skb_do_redirect(skb);
3366 *ret = NET_XMIT_SUCCESS;
3374 #endif /* CONFIG_NET_EGRESS */
3376 static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb)
3379 struct xps_dev_maps *dev_maps;
3380 struct xps_map *map;
3381 int queue_index = -1;
3384 dev_maps = rcu_dereference(dev->xps_maps);
3386 unsigned int tci = skb->sender_cpu - 1;
3390 tci += netdev_get_prio_tc_map(dev, skb->priority);
3393 map = rcu_dereference(dev_maps->cpu_map[tci]);
3396 queue_index = map->queues[0];
3398 queue_index = map->queues[reciprocal_scale(skb_get_hash(skb),
3400 if (unlikely(queue_index >= dev->real_num_tx_queues))
3412 static u16 __netdev_pick_tx(struct net_device *dev, struct sk_buff *skb)
3414 struct sock *sk = skb->sk;
3415 int queue_index = sk_tx_queue_get(sk);
3417 if (queue_index < 0 || skb->ooo_okay ||
3418 queue_index >= dev->real_num_tx_queues) {
3419 int new_index = get_xps_queue(dev, skb);
3422 new_index = skb_tx_hash(dev, skb);
3424 if (queue_index != new_index && sk &&
3426 rcu_access_pointer(sk->sk_dst_cache))
3427 sk_tx_queue_set(sk, new_index);
3429 queue_index = new_index;
3435 struct netdev_queue *netdev_pick_tx(struct net_device *dev,
3436 struct sk_buff *skb,
3439 int queue_index = 0;
3442 u32 sender_cpu = skb->sender_cpu - 1;
3444 if (sender_cpu >= (u32)NR_CPUS)
3445 skb->sender_cpu = raw_smp_processor_id() + 1;
3448 if (dev->real_num_tx_queues != 1) {
3449 const struct net_device_ops *ops = dev->netdev_ops;
3451 if (ops->ndo_select_queue)
3452 queue_index = ops->ndo_select_queue(dev, skb, accel_priv,
3455 queue_index = __netdev_pick_tx(dev, skb);
3457 queue_index = netdev_cap_txqueue(dev, queue_index);
3460 skb_set_queue_mapping(skb, queue_index);
3461 return netdev_get_tx_queue(dev, queue_index);
3465 * __dev_queue_xmit - transmit a buffer
3466 * @skb: buffer to transmit
3467 * @accel_priv: private data used for L2 forwarding offload
3469 * Queue a buffer for transmission to a network device. The caller must
3470 * have set the device and priority and built the buffer before calling
3471 * this function. The function can be called from an interrupt.
3473 * A negative errno code is returned on a failure. A success does not
3474 * guarantee the frame will be transmitted as it may be dropped due
3475 * to congestion or traffic shaping.
3477 * -----------------------------------------------------------------------------------
3478 * I notice this method can also return errors from the queue disciplines,
3479 * including NET_XMIT_DROP, which is a positive value. So, errors can also
3482 * Regardless of the return value, the skb is consumed, so it is currently
3483 * difficult to retry a send to this method. (You can bump the ref count
3484 * before sending to hold a reference for retry if you are careful.)
3486 * When calling this method, interrupts MUST be enabled. This is because
3487 * the BH enable code must have IRQs enabled so that it will not deadlock.
3490 static int __dev_queue_xmit(struct sk_buff *skb, void *accel_priv)
3492 struct net_device *dev = skb->dev;
3493 struct netdev_queue *txq;
3498 skb_reset_mac_header(skb);
3500 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP))
3501 __skb_tstamp_tx(skb, NULL, skb->sk, SCM_TSTAMP_SCHED);
3503 /* Disable soft irqs for various locks below. Also
3504 * stops preemption for RCU.
3508 skb_update_prio(skb);
3510 qdisc_pkt_len_init(skb);
3511 #ifdef CONFIG_NET_CLS_ACT
3512 skb->tc_at_ingress = 0;
3513 # ifdef CONFIG_NET_EGRESS
3514 if (static_key_false(&egress_needed)) {
3515 skb = sch_handle_egress(skb, &rc, dev);
3521 /* If device/qdisc don't need skb->dst, release it right now while
3522 * its hot in this cpu cache.
3524 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
3529 txq = netdev_pick_tx(dev, skb, accel_priv);
3530 q = rcu_dereference_bh(txq->qdisc);
3532 trace_net_dev_queue(skb);
3534 rc = __dev_xmit_skb(skb, q, dev, txq);
3538 /* The device has no queue. Common case for software devices:
3539 * loopback, all the sorts of tunnels...
3541 * Really, it is unlikely that netif_tx_lock protection is necessary
3542 * here. (f.e. loopback and IP tunnels are clean ignoring statistics
3544 * However, it is possible, that they rely on protection
3547 * Check this and shot the lock. It is not prone from deadlocks.
3548 *Either shot noqueue qdisc, it is even simpler 8)
3550 if (dev->flags & IFF_UP) {
3551 int cpu = smp_processor_id(); /* ok because BHs are off */
3553 if (txq->xmit_lock_owner != cpu) {
3554 if (unlikely(__this_cpu_read(xmit_recursion) >
3555 XMIT_RECURSION_LIMIT))
3556 goto recursion_alert;
3558 skb = validate_xmit_skb(skb, dev, &again);
3562 HARD_TX_LOCK(dev, txq, cpu);
3564 if (!netif_xmit_stopped(txq)) {
3565 __this_cpu_inc(xmit_recursion);
3566 skb = dev_hard_start_xmit(skb, dev, txq, &rc);
3567 __this_cpu_dec(xmit_recursion);
3568 if (dev_xmit_complete(rc)) {
3569 HARD_TX_UNLOCK(dev, txq);
3573 HARD_TX_UNLOCK(dev, txq);
3574 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
3577 /* Recursion is detected! It is possible,
3581 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
3587 rcu_read_unlock_bh();
3589 atomic_long_inc(&dev->tx_dropped);
3590 kfree_skb_list(skb);
3593 rcu_read_unlock_bh();
3597 int dev_queue_xmit(struct sk_buff *skb)
3599 return __dev_queue_xmit(skb, NULL);
3601 EXPORT_SYMBOL(dev_queue_xmit);
3603 int dev_queue_xmit_accel(struct sk_buff *skb, void *accel_priv)
3605 return __dev_queue_xmit(skb, accel_priv);
3607 EXPORT_SYMBOL(dev_queue_xmit_accel);
3610 /*************************************************************************
3612 *************************************************************************/
3614 int netdev_max_backlog __read_mostly = 1000;
3615 EXPORT_SYMBOL(netdev_max_backlog);
3617 int netdev_tstamp_prequeue __read_mostly = 1;
3618 int netdev_budget __read_mostly = 300;
3619 unsigned int __read_mostly netdev_budget_usecs = 2000;
3620 int weight_p __read_mostly = 64; /* old backlog weight */
3621 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */
3622 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */
3623 int dev_rx_weight __read_mostly = 64;
3624 int dev_tx_weight __read_mostly = 64;
3626 /* Called with irq disabled */
3627 static inline void ____napi_schedule(struct softnet_data *sd,
3628 struct napi_struct *napi)
3630 list_add_tail(&napi->poll_list, &sd->poll_list);
3631 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3636 /* One global table that all flow-based protocols share. */
3637 struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
3638 EXPORT_SYMBOL(rps_sock_flow_table);
3639 u32 rps_cpu_mask __read_mostly;
3640 EXPORT_SYMBOL(rps_cpu_mask);
3642 struct static_key rps_needed __read_mostly;
3643 EXPORT_SYMBOL(rps_needed);
3644 struct static_key rfs_needed __read_mostly;
3645 EXPORT_SYMBOL(rfs_needed);
3647 static struct rps_dev_flow *
3648 set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3649 struct rps_dev_flow *rflow, u16 next_cpu)
3651 if (next_cpu < nr_cpu_ids) {
3652 #ifdef CONFIG_RFS_ACCEL
3653 struct netdev_rx_queue *rxqueue;
3654 struct rps_dev_flow_table *flow_table;
3655 struct rps_dev_flow *old_rflow;
3660 /* Should we steer this flow to a different hardware queue? */
3661 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
3662 !(dev->features & NETIF_F_NTUPLE))
3664 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
3665 if (rxq_index == skb_get_rx_queue(skb))
3668 rxqueue = dev->_rx + rxq_index;
3669 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3672 flow_id = skb_get_hash(skb) & flow_table->mask;
3673 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
3674 rxq_index, flow_id);
3678 rflow = &flow_table->flows[flow_id];
3680 if (old_rflow->filter == rflow->filter)
3681 old_rflow->filter = RPS_NO_FILTER;
3685 per_cpu(softnet_data, next_cpu).input_queue_head;
3688 rflow->cpu = next_cpu;
3693 * get_rps_cpu is called from netif_receive_skb and returns the target
3694 * CPU from the RPS map of the receiving queue for a given skb.
3695 * rcu_read_lock must be held on entry.
3697 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3698 struct rps_dev_flow **rflowp)
3700 const struct rps_sock_flow_table *sock_flow_table;
3701 struct netdev_rx_queue *rxqueue = dev->_rx;
3702 struct rps_dev_flow_table *flow_table;
3703 struct rps_map *map;
3708 if (skb_rx_queue_recorded(skb)) {
3709 u16 index = skb_get_rx_queue(skb);
3711 if (unlikely(index >= dev->real_num_rx_queues)) {
3712 WARN_ONCE(dev->real_num_rx_queues > 1,
3713 "%s received packet on queue %u, but number "
3714 "of RX queues is %u\n",
3715 dev->name, index, dev->real_num_rx_queues);
3721 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */
3723 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3724 map = rcu_dereference(rxqueue->rps_map);
3725 if (!flow_table && !map)
3728 skb_reset_network_header(skb);
3729 hash = skb_get_hash(skb);
3733 sock_flow_table = rcu_dereference(rps_sock_flow_table);
3734 if (flow_table && sock_flow_table) {
3735 struct rps_dev_flow *rflow;
3739 /* First check into global flow table if there is a match */
3740 ident = sock_flow_table->ents[hash & sock_flow_table->mask];
3741 if ((ident ^ hash) & ~rps_cpu_mask)
3744 next_cpu = ident & rps_cpu_mask;
3746 /* OK, now we know there is a match,
3747 * we can look at the local (per receive queue) flow table
3749 rflow = &flow_table->flows[hash & flow_table->mask];
3753 * If the desired CPU (where last recvmsg was done) is
3754 * different from current CPU (one in the rx-queue flow
3755 * table entry), switch if one of the following holds:
3756 * - Current CPU is unset (>= nr_cpu_ids).
3757 * - Current CPU is offline.
3758 * - The current CPU's queue tail has advanced beyond the
3759 * last packet that was enqueued using this table entry.
3760 * This guarantees that all previous packets for the flow
3761 * have been dequeued, thus preserving in order delivery.
3763 if (unlikely(tcpu != next_cpu) &&
3764 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) ||
3765 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
3766 rflow->last_qtail)) >= 0)) {
3768 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
3771 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) {
3781 tcpu = map->cpus[reciprocal_scale(hash, map->len)];
3782 if (cpu_online(tcpu)) {
3792 #ifdef CONFIG_RFS_ACCEL
3795 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
3796 * @dev: Device on which the filter was set
3797 * @rxq_index: RX queue index
3798 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
3799 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
3801 * Drivers that implement ndo_rx_flow_steer() should periodically call
3802 * this function for each installed filter and remove the filters for
3803 * which it returns %true.
3805 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
3806 u32 flow_id, u16 filter_id)
3808 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
3809 struct rps_dev_flow_table *flow_table;
3810 struct rps_dev_flow *rflow;
3815 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3816 if (flow_table && flow_id <= flow_table->mask) {
3817 rflow = &flow_table->flows[flow_id];
3818 cpu = READ_ONCE(rflow->cpu);
3819 if (rflow->filter == filter_id && cpu < nr_cpu_ids &&
3820 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
3821 rflow->last_qtail) <
3822 (int)(10 * flow_table->mask)))
3828 EXPORT_SYMBOL(rps_may_expire_flow);
3830 #endif /* CONFIG_RFS_ACCEL */
3832 /* Called from hardirq (IPI) context */
3833 static void rps_trigger_softirq(void *data)
3835 struct softnet_data *sd = data;
3837 ____napi_schedule(sd, &sd->backlog);
3841 #endif /* CONFIG_RPS */
3844 * Check if this softnet_data structure is another cpu one
3845 * If yes, queue it to our IPI list and return 1
3848 static int rps_ipi_queued(struct softnet_data *sd)
3851 struct softnet_data *mysd = this_cpu_ptr(&softnet_data);
3854 sd->rps_ipi_next = mysd->rps_ipi_list;
3855 mysd->rps_ipi_list = sd;
3857 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3860 #endif /* CONFIG_RPS */
3864 #ifdef CONFIG_NET_FLOW_LIMIT
3865 int netdev_flow_limit_table_len __read_mostly = (1 << 12);
3868 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
3870 #ifdef CONFIG_NET_FLOW_LIMIT
3871 struct sd_flow_limit *fl;
3872 struct softnet_data *sd;
3873 unsigned int old_flow, new_flow;
3875 if (qlen < (netdev_max_backlog >> 1))
3878 sd = this_cpu_ptr(&softnet_data);
3881 fl = rcu_dereference(sd->flow_limit);
3883 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1);
3884 old_flow = fl->history[fl->history_head];
3885 fl->history[fl->history_head] = new_flow;
3888 fl->history_head &= FLOW_LIMIT_HISTORY - 1;
3890 if (likely(fl->buckets[old_flow]))
3891 fl->buckets[old_flow]--;
3893 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) {
3905 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
3906 * queue (may be a remote CPU queue).
3908 static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
3909 unsigned int *qtail)
3911 struct softnet_data *sd;
3912 unsigned long flags;
3915 sd = &per_cpu(softnet_data, cpu);
3917 local_irq_save(flags);
3920 if (!netif_running(skb->dev))
3922 qlen = skb_queue_len(&sd->input_pkt_queue);
3923 if (qlen <= netdev_max_backlog && !skb_flow_limit(skb, qlen)) {
3926 __skb_queue_tail(&sd->input_pkt_queue, skb);
3927 input_queue_tail_incr_save(sd, qtail);
3929 local_irq_restore(flags);
3930 return NET_RX_SUCCESS;
3933 /* Schedule NAPI for backlog device
3934 * We can use non atomic operation since we own the queue lock
3936 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
3937 if (!rps_ipi_queued(sd))
3938 ____napi_schedule(sd, &sd->backlog);
3947 local_irq_restore(flags);
3949 atomic_long_inc(&skb->dev->rx_dropped);
3954 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb)
3956 struct net_device *dev = skb->dev;
3957 struct netdev_rx_queue *rxqueue;
3961 if (skb_rx_queue_recorded(skb)) {
3962 u16 index = skb_get_rx_queue(skb);
3964 if (unlikely(index >= dev->real_num_rx_queues)) {
3965 WARN_ONCE(dev->real_num_rx_queues > 1,
3966 "%s received packet on queue %u, but number "
3967 "of RX queues is %u\n",
3968 dev->name, index, dev->real_num_rx_queues);
3970 return rxqueue; /* Return first rxqueue */
3977 static u32 netif_receive_generic_xdp(struct sk_buff *skb,
3978 struct bpf_prog *xdp_prog)
3980 struct netdev_rx_queue *rxqueue;
3981 u32 metalen, act = XDP_DROP;
3982 struct xdp_buff xdp;
3987 /* Reinjected packets coming from act_mirred or similar should
3988 * not get XDP generic processing.
3990 if (skb_cloned(skb))
3993 /* XDP packets must be linear and must have sufficient headroom
3994 * of XDP_PACKET_HEADROOM bytes. This is the guarantee that also
3995 * native XDP provides, thus we need to do it here as well.
3997 if (skb_is_nonlinear(skb) ||
3998 skb_headroom(skb) < XDP_PACKET_HEADROOM) {
3999 int hroom = XDP_PACKET_HEADROOM - skb_headroom(skb);
4000 int troom = skb->tail + skb->data_len - skb->end;
4002 /* In case we have to go down the path and also linearize,
4003 * then lets do the pskb_expand_head() work just once here.
4005 if (pskb_expand_head(skb,
4006 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0,
4007 troom > 0 ? troom + 128 : 0, GFP_ATOMIC))
4009 if (skb_linearize(skb))
4013 /* The XDP program wants to see the packet starting at the MAC
4016 mac_len = skb->data - skb_mac_header(skb);
4017 hlen = skb_headlen(skb) + mac_len;
4018 xdp.data = skb->data - mac_len;
4019 xdp.data_meta = xdp.data;
4020 xdp.data_end = xdp.data + hlen;
4021 xdp.data_hard_start = skb->data - skb_headroom(skb);
4022 orig_data = xdp.data;
4024 rxqueue = netif_get_rxqueue(skb);
4025 xdp.rxq = &rxqueue->xdp_rxq;
4027 act = bpf_prog_run_xdp(xdp_prog, &xdp);
4029 off = xdp.data - orig_data;
4031 __skb_pull(skb, off);
4033 __skb_push(skb, -off);
4034 skb->mac_header += off;
4039 __skb_push(skb, mac_len);
4042 metalen = xdp.data - xdp.data_meta;
4044 skb_metadata_set(skb, metalen);
4047 bpf_warn_invalid_xdp_action(act);
4050 trace_xdp_exception(skb->dev, xdp_prog, act);
4061 /* When doing generic XDP we have to bypass the qdisc layer and the
4062 * network taps in order to match in-driver-XDP behavior.
4064 void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog)
4066 struct net_device *dev = skb->dev;
4067 struct netdev_queue *txq;
4068 bool free_skb = true;
4071 txq = netdev_pick_tx(dev, skb, NULL);
4072 cpu = smp_processor_id();
4073 HARD_TX_LOCK(dev, txq, cpu);
4074 if (!netif_xmit_stopped(txq)) {
4075 rc = netdev_start_xmit(skb, dev, txq, 0);
4076 if (dev_xmit_complete(rc))
4079 HARD_TX_UNLOCK(dev, txq);
4081 trace_xdp_exception(dev, xdp_prog, XDP_TX);
4085 EXPORT_SYMBOL_GPL(generic_xdp_tx);
4087 static struct static_key generic_xdp_needed __read_mostly;
4089 int do_xdp_generic(struct bpf_prog *xdp_prog, struct sk_buff *skb)
4092 u32 act = netif_receive_generic_xdp(skb, xdp_prog);
4095 if (act != XDP_PASS) {
4098 err = xdp_do_generic_redirect(skb->dev, skb,
4102 /* fallthru to submit skb */
4104 generic_xdp_tx(skb, xdp_prog);
4115 EXPORT_SYMBOL_GPL(do_xdp_generic);
4117 static int netif_rx_internal(struct sk_buff *skb)
4121 net_timestamp_check(netdev_tstamp_prequeue, skb);
4123 trace_netif_rx(skb);
4125 if (static_key_false(&generic_xdp_needed)) {
4130 ret = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
4134 /* Consider XDP consuming the packet a success from
4135 * the netdev point of view we do not want to count
4138 if (ret != XDP_PASS)
4139 return NET_RX_SUCCESS;
4143 if (static_key_false(&rps_needed)) {
4144 struct rps_dev_flow voidflow, *rflow = &voidflow;
4150 cpu = get_rps_cpu(skb->dev, skb, &rflow);
4152 cpu = smp_processor_id();
4154 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4163 ret = enqueue_to_backlog(skb, get_cpu(), &qtail);
4170 * netif_rx - post buffer to the network code
4171 * @skb: buffer to post
4173 * This function receives a packet from a device driver and queues it for
4174 * the upper (protocol) levels to process. It always succeeds. The buffer
4175 * may be dropped during processing for congestion control or by the
4179 * NET_RX_SUCCESS (no congestion)
4180 * NET_RX_DROP (packet was dropped)
4184 int netif_rx(struct sk_buff *skb)
4186 trace_netif_rx_entry(skb);
4188 return netif_rx_internal(skb);
4190 EXPORT_SYMBOL(netif_rx);
4192 int netif_rx_ni(struct sk_buff *skb)
4196 trace_netif_rx_ni_entry(skb);
4199 err = netif_rx_internal(skb);
4200 if (local_softirq_pending())
4206 EXPORT_SYMBOL(netif_rx_ni);
4208 static __latent_entropy void net_tx_action(struct softirq_action *h)
4210 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
4212 if (sd->completion_queue) {
4213 struct sk_buff *clist;
4215 local_irq_disable();
4216 clist = sd->completion_queue;
4217 sd->completion_queue = NULL;
4221 struct sk_buff *skb = clist;
4223 clist = clist->next;
4225 WARN_ON(refcount_read(&skb->users));
4226 if (likely(get_kfree_skb_cb(skb)->reason == SKB_REASON_CONSUMED))
4227 trace_consume_skb(skb);
4229 trace_kfree_skb(skb, net_tx_action);
4231 if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
4234 __kfree_skb_defer(skb);
4237 __kfree_skb_flush();
4240 if (sd->output_queue) {
4243 local_irq_disable();
4244 head = sd->output_queue;
4245 sd->output_queue = NULL;
4246 sd->output_queue_tailp = &sd->output_queue;
4250 struct Qdisc *q = head;
4251 spinlock_t *root_lock = NULL;
4253 head = head->next_sched;
4255 if (!(q->flags & TCQ_F_NOLOCK)) {
4256 root_lock = qdisc_lock(q);
4257 spin_lock(root_lock);
4259 /* We need to make sure head->next_sched is read
4260 * before clearing __QDISC_STATE_SCHED
4262 smp_mb__before_atomic();
4263 clear_bit(__QDISC_STATE_SCHED, &q->state);
4266 spin_unlock(root_lock);
4270 xfrm_dev_backlog(sd);
4273 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE)
4274 /* This hook is defined here for ATM LANE */
4275 int (*br_fdb_test_addr_hook)(struct net_device *dev,
4276 unsigned char *addr) __read_mostly;
4277 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook);
4280 static inline struct sk_buff *
4281 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
4282 struct net_device *orig_dev)
4284 #ifdef CONFIG_NET_CLS_ACT
4285 struct mini_Qdisc *miniq = rcu_dereference_bh(skb->dev->miniq_ingress);
4286 struct tcf_result cl_res;
4288 /* If there's at least one ingress present somewhere (so
4289 * we get here via enabled static key), remaining devices
4290 * that are not configured with an ingress qdisc will bail
4297 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4301 qdisc_skb_cb(skb)->pkt_len = skb->len;
4302 skb->tc_at_ingress = 1;
4303 mini_qdisc_bstats_cpu_update(miniq, skb);
4305 switch (tcf_classify(skb, miniq->filter_list, &cl_res, false)) {
4307 case TC_ACT_RECLASSIFY:
4308 skb->tc_index = TC_H_MIN(cl_res.classid);
4311 mini_qdisc_qstats_cpu_drop(miniq);
4319 case TC_ACT_REDIRECT:
4320 /* skb_mac_header check was done by cls/act_bpf, so
4321 * we can safely push the L2 header back before
4322 * redirecting to another netdev
4324 __skb_push(skb, skb->mac_len);
4325 skb_do_redirect(skb);
4330 #endif /* CONFIG_NET_CLS_ACT */
4335 * netdev_is_rx_handler_busy - check if receive handler is registered
4336 * @dev: device to check
4338 * Check if a receive handler is already registered for a given device.
4339 * Return true if there one.
4341 * The caller must hold the rtnl_mutex.
4343 bool netdev_is_rx_handler_busy(struct net_device *dev)
4346 return dev && rtnl_dereference(dev->rx_handler);
4348 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy);
4351 * netdev_rx_handler_register - register receive handler
4352 * @dev: device to register a handler for
4353 * @rx_handler: receive handler to register
4354 * @rx_handler_data: data pointer that is used by rx handler
4356 * Register a receive handler for a device. This handler will then be
4357 * called from __netif_receive_skb. A negative errno code is returned
4360 * The caller must hold the rtnl_mutex.
4362 * For a general description of rx_handler, see enum rx_handler_result.
4364 int netdev_rx_handler_register(struct net_device *dev,
4365 rx_handler_func_t *rx_handler,
4366 void *rx_handler_data)
4368 if (netdev_is_rx_handler_busy(dev))
4371 if (dev->priv_flags & IFF_NO_RX_HANDLER)
4374 /* Note: rx_handler_data must be set before rx_handler */
4375 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data);
4376 rcu_assign_pointer(dev->rx_handler, rx_handler);
4380 EXPORT_SYMBOL_GPL(netdev_rx_handler_register);
4383 * netdev_rx_handler_unregister - unregister receive handler
4384 * @dev: device to unregister a handler from
4386 * Unregister a receive handler from a device.
4388 * The caller must hold the rtnl_mutex.
4390 void netdev_rx_handler_unregister(struct net_device *dev)
4394 RCU_INIT_POINTER(dev->rx_handler, NULL);
4395 /* a reader seeing a non NULL rx_handler in a rcu_read_lock()
4396 * section has a guarantee to see a non NULL rx_handler_data
4400 RCU_INIT_POINTER(dev->rx_handler_data, NULL);
4402 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister);
4405 * Limit the use of PFMEMALLOC reserves to those protocols that implement
4406 * the special handling of PFMEMALLOC skbs.
4408 static bool skb_pfmemalloc_protocol(struct sk_buff *skb)
4410 switch (skb->protocol) {
4411 case htons(ETH_P_ARP):
4412 case htons(ETH_P_IP):
4413 case htons(ETH_P_IPV6):
4414 case htons(ETH_P_8021Q):
4415 case htons(ETH_P_8021AD):
4422 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev,
4423 int *ret, struct net_device *orig_dev)
4425 #ifdef CONFIG_NETFILTER_INGRESS
4426 if (nf_hook_ingress_active(skb)) {
4430 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4435 ingress_retval = nf_hook_ingress(skb);
4437 return ingress_retval;
4439 #endif /* CONFIG_NETFILTER_INGRESS */
4443 static int __netif_receive_skb_core(struct sk_buff *skb, bool pfmemalloc)
4445 struct packet_type *ptype, *pt_prev;
4446 rx_handler_func_t *rx_handler;
4447 struct net_device *orig_dev;
4448 bool deliver_exact = false;
4449 int ret = NET_RX_DROP;
4452 net_timestamp_check(!netdev_tstamp_prequeue, skb);
4454 trace_netif_receive_skb(skb);
4456 orig_dev = skb->dev;
4458 skb_reset_network_header(skb);
4459 if (!skb_transport_header_was_set(skb))
4460 skb_reset_transport_header(skb);
4461 skb_reset_mac_len(skb);
4466 skb->skb_iif = skb->dev->ifindex;
4468 __this_cpu_inc(softnet_data.processed);
4470 if (skb->protocol == cpu_to_be16(ETH_P_8021Q) ||
4471 skb->protocol == cpu_to_be16(ETH_P_8021AD)) {
4472 skb = skb_vlan_untag(skb);
4477 if (skb_skip_tc_classify(skb))
4483 list_for_each_entry_rcu(ptype, &ptype_all, list) {
4485 ret = deliver_skb(skb, pt_prev, orig_dev);
4489 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) {
4491 ret = deliver_skb(skb, pt_prev, orig_dev);
4496 #ifdef CONFIG_NET_INGRESS
4497 if (static_key_false(&ingress_needed)) {
4498 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev);
4502 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0)
4508 if (pfmemalloc && !skb_pfmemalloc_protocol(skb))
4511 if (skb_vlan_tag_present(skb)) {
4513 ret = deliver_skb(skb, pt_prev, orig_dev);
4516 if (vlan_do_receive(&skb))
4518 else if (unlikely(!skb))
4522 rx_handler = rcu_dereference(skb->dev->rx_handler);
4525 ret = deliver_skb(skb, pt_prev, orig_dev);
4528 switch (rx_handler(&skb)) {
4529 case RX_HANDLER_CONSUMED:
4530 ret = NET_RX_SUCCESS;
4532 case RX_HANDLER_ANOTHER:
4534 case RX_HANDLER_EXACT:
4535 deliver_exact = true;
4536 case RX_HANDLER_PASS:
4543 if (unlikely(skb_vlan_tag_present(skb))) {
4544 if (skb_vlan_tag_get_id(skb))
4545 skb->pkt_type = PACKET_OTHERHOST;
4546 /* Note: we might in the future use prio bits
4547 * and set skb->priority like in vlan_do_receive()
4548 * For the time being, just ignore Priority Code Point
4553 type = skb->protocol;
4555 /* deliver only exact match when indicated */
4556 if (likely(!deliver_exact)) {
4557 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4558 &ptype_base[ntohs(type) &
4562 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4563 &orig_dev->ptype_specific);
4565 if (unlikely(skb->dev != orig_dev)) {
4566 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4567 &skb->dev->ptype_specific);
4571 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
4574 ret = pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
4578 atomic_long_inc(&skb->dev->rx_dropped);
4580 atomic_long_inc(&skb->dev->rx_nohandler);
4582 /* Jamal, now you will not able to escape explaining
4583 * me how you were going to use this. :-)
4593 * netif_receive_skb_core - special purpose version of netif_receive_skb
4594 * @skb: buffer to process
4596 * More direct receive version of netif_receive_skb(). It should
4597 * only be used by callers that have a need to skip RPS and Generic XDP.
4598 * Caller must also take care of handling if (page_is_)pfmemalloc.
4600 * This function may only be called from softirq context and interrupts
4601 * should be enabled.
4603 * Return values (usually ignored):
4604 * NET_RX_SUCCESS: no congestion
4605 * NET_RX_DROP: packet was dropped
4607 int netif_receive_skb_core(struct sk_buff *skb)
4612 ret = __netif_receive_skb_core(skb, false);
4617 EXPORT_SYMBOL(netif_receive_skb_core);
4619 static int __netif_receive_skb(struct sk_buff *skb)
4623 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) {
4624 unsigned int noreclaim_flag;
4627 * PFMEMALLOC skbs are special, they should
4628 * - be delivered to SOCK_MEMALLOC sockets only
4629 * - stay away from userspace
4630 * - have bounded memory usage
4632 * Use PF_MEMALLOC as this saves us from propagating the allocation
4633 * context down to all allocation sites.
4635 noreclaim_flag = memalloc_noreclaim_save();
4636 ret = __netif_receive_skb_core(skb, true);
4637 memalloc_noreclaim_restore(noreclaim_flag);
4639 ret = __netif_receive_skb_core(skb, false);
4644 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp)
4646 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog);
4647 struct bpf_prog *new = xdp->prog;
4650 switch (xdp->command) {
4651 case XDP_SETUP_PROG:
4652 rcu_assign_pointer(dev->xdp_prog, new);
4657 static_key_slow_dec(&generic_xdp_needed);
4658 } else if (new && !old) {
4659 static_key_slow_inc(&generic_xdp_needed);
4660 dev_disable_lro(dev);
4661 dev_disable_gro_hw(dev);
4665 case XDP_QUERY_PROG:
4666 xdp->prog_attached = !!old;
4667 xdp->prog_id = old ? old->aux->id : 0;
4678 static int netif_receive_skb_internal(struct sk_buff *skb)
4682 net_timestamp_check(netdev_tstamp_prequeue, skb);
4684 if (skb_defer_rx_timestamp(skb))
4685 return NET_RX_SUCCESS;
4687 if (static_key_false(&generic_xdp_needed)) {
4692 ret = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
4696 if (ret != XDP_PASS)
4702 if (static_key_false(&rps_needed)) {
4703 struct rps_dev_flow voidflow, *rflow = &voidflow;
4704 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
4707 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4713 ret = __netif_receive_skb(skb);
4719 * netif_receive_skb - process receive buffer from network
4720 * @skb: buffer to process
4722 * netif_receive_skb() is the main receive data processing function.
4723 * It always succeeds. The buffer may be dropped during processing
4724 * for congestion control or by the protocol layers.
4726 * This function may only be called from softirq context and interrupts
4727 * should be enabled.
4729 * Return values (usually ignored):
4730 * NET_RX_SUCCESS: no congestion
4731 * NET_RX_DROP: packet was dropped
4733 int netif_receive_skb(struct sk_buff *skb)
4735 trace_netif_receive_skb_entry(skb);
4737 return netif_receive_skb_internal(skb);
4739 EXPORT_SYMBOL(netif_receive_skb);
4741 DEFINE_PER_CPU(struct work_struct, flush_works);
4743 /* Network device is going away, flush any packets still pending */
4744 static void flush_backlog(struct work_struct *work)
4746 struct sk_buff *skb, *tmp;
4747 struct softnet_data *sd;
4750 sd = this_cpu_ptr(&softnet_data);
4752 local_irq_disable();
4754 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) {
4755 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
4756 __skb_unlink(skb, &sd->input_pkt_queue);
4758 input_queue_head_incr(sd);
4764 skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
4765 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
4766 __skb_unlink(skb, &sd->process_queue);
4768 input_queue_head_incr(sd);
4774 static void flush_all_backlogs(void)
4780 for_each_online_cpu(cpu)
4781 queue_work_on(cpu, system_highpri_wq,
4782 per_cpu_ptr(&flush_works, cpu));
4784 for_each_online_cpu(cpu)
4785 flush_work(per_cpu_ptr(&flush_works, cpu));
4790 static int napi_gro_complete(struct sk_buff *skb)
4792 struct packet_offload *ptype;
4793 __be16 type = skb->protocol;
4794 struct list_head *head = &offload_base;
4797 BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
4799 if (NAPI_GRO_CB(skb)->count == 1) {
4800 skb_shinfo(skb)->gso_size = 0;
4805 list_for_each_entry_rcu(ptype, head, list) {
4806 if (ptype->type != type || !ptype->callbacks.gro_complete)
4809 err = ptype->callbacks.gro_complete(skb, 0);
4815 WARN_ON(&ptype->list == head);
4817 return NET_RX_SUCCESS;
4821 return netif_receive_skb_internal(skb);
4824 /* napi->gro_list contains packets ordered by age.
4825 * youngest packets at the head of it.
4826 * Complete skbs in reverse order to reduce latencies.
4828 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
4830 struct sk_buff *skb, *prev = NULL;
4832 /* scan list and build reverse chain */
4833 for (skb = napi->gro_list; skb != NULL; skb = skb->next) {
4838 for (skb = prev; skb; skb = prev) {
4841 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
4845 napi_gro_complete(skb);
4849 napi->gro_list = NULL;
4851 EXPORT_SYMBOL(napi_gro_flush);
4853 static void gro_list_prepare(struct napi_struct *napi, struct sk_buff *skb)
4856 unsigned int maclen = skb->dev->hard_header_len;
4857 u32 hash = skb_get_hash_raw(skb);
4859 for (p = napi->gro_list; p; p = p->next) {
4860 unsigned long diffs;
4862 NAPI_GRO_CB(p)->flush = 0;
4864 if (hash != skb_get_hash_raw(p)) {
4865 NAPI_GRO_CB(p)->same_flow = 0;
4869 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
4870 diffs |= p->vlan_tci ^ skb->vlan_tci;
4871 diffs |= skb_metadata_dst_cmp(p, skb);
4872 diffs |= skb_metadata_differs(p, skb);
4873 if (maclen == ETH_HLEN)
4874 diffs |= compare_ether_header(skb_mac_header(p),
4875 skb_mac_header(skb));
4877 diffs = memcmp(skb_mac_header(p),
4878 skb_mac_header(skb),
4880 NAPI_GRO_CB(p)->same_flow = !diffs;
4884 static void skb_gro_reset_offset(struct sk_buff *skb)
4886 const struct skb_shared_info *pinfo = skb_shinfo(skb);
4887 const skb_frag_t *frag0 = &pinfo->frags[0];
4889 NAPI_GRO_CB(skb)->data_offset = 0;
4890 NAPI_GRO_CB(skb)->frag0 = NULL;
4891 NAPI_GRO_CB(skb)->frag0_len = 0;
4893 if (skb_mac_header(skb) == skb_tail_pointer(skb) &&
4895 !PageHighMem(skb_frag_page(frag0))) {
4896 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
4897 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
4898 skb_frag_size(frag0),
4899 skb->end - skb->tail);
4903 static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
4905 struct skb_shared_info *pinfo = skb_shinfo(skb);
4907 BUG_ON(skb->end - skb->tail < grow);
4909 memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
4911 skb->data_len -= grow;
4914 pinfo->frags[0].page_offset += grow;
4915 skb_frag_size_sub(&pinfo->frags[0], grow);
4917 if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
4918 skb_frag_unref(skb, 0);
4919 memmove(pinfo->frags, pinfo->frags + 1,
4920 --pinfo->nr_frags * sizeof(pinfo->frags[0]));
4924 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
4926 struct sk_buff **pp = NULL;
4927 struct packet_offload *ptype;
4928 __be16 type = skb->protocol;
4929 struct list_head *head = &offload_base;
4931 enum gro_result ret;
4934 if (netif_elide_gro(skb->dev))
4937 gro_list_prepare(napi, skb);
4940 list_for_each_entry_rcu(ptype, head, list) {
4941 if (ptype->type != type || !ptype->callbacks.gro_receive)
4944 skb_set_network_header(skb, skb_gro_offset(skb));
4945 skb_reset_mac_len(skb);
4946 NAPI_GRO_CB(skb)->same_flow = 0;
4947 NAPI_GRO_CB(skb)->flush = skb_is_gso(skb) || skb_has_frag_list(skb);
4948 NAPI_GRO_CB(skb)->free = 0;
4949 NAPI_GRO_CB(skb)->encap_mark = 0;
4950 NAPI_GRO_CB(skb)->recursion_counter = 0;
4951 NAPI_GRO_CB(skb)->is_fou = 0;
4952 NAPI_GRO_CB(skb)->is_atomic = 1;
4953 NAPI_GRO_CB(skb)->gro_remcsum_start = 0;
4955 /* Setup for GRO checksum validation */
4956 switch (skb->ip_summed) {
4957 case CHECKSUM_COMPLETE:
4958 NAPI_GRO_CB(skb)->csum = skb->csum;
4959 NAPI_GRO_CB(skb)->csum_valid = 1;
4960 NAPI_GRO_CB(skb)->csum_cnt = 0;
4962 case CHECKSUM_UNNECESSARY:
4963 NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
4964 NAPI_GRO_CB(skb)->csum_valid = 0;
4967 NAPI_GRO_CB(skb)->csum_cnt = 0;
4968 NAPI_GRO_CB(skb)->csum_valid = 0;
4971 pp = ptype->callbacks.gro_receive(&napi->gro_list, skb);
4976 if (&ptype->list == head)
4979 if (IS_ERR(pp) && PTR_ERR(pp) == -EINPROGRESS) {
4984 same_flow = NAPI_GRO_CB(skb)->same_flow;
4985 ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
4988 struct sk_buff *nskb = *pp;
4992 napi_gro_complete(nskb);
4999 if (NAPI_GRO_CB(skb)->flush)
5002 if (unlikely(napi->gro_count >= MAX_GRO_SKBS)) {
5003 struct sk_buff *nskb = napi->gro_list;
5005 /* locate the end of the list to select the 'oldest' flow */
5006 while (nskb->next) {
5012 napi_gro_complete(nskb);
5016 NAPI_GRO_CB(skb)->count = 1;
5017 NAPI_GRO_CB(skb)->age = jiffies;
5018 NAPI_GRO_CB(skb)->last = skb;
5019 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
5020 skb->next = napi->gro_list;
5021 napi->gro_list = skb;
5025 grow = skb_gro_offset(skb) - skb_headlen(skb);
5027 gro_pull_from_frag0(skb, grow);
5036 struct packet_offload *gro_find_receive_by_type(__be16 type)
5038 struct list_head *offload_head = &offload_base;
5039 struct packet_offload *ptype;
5041 list_for_each_entry_rcu(ptype, offload_head, list) {
5042 if (ptype->type != type || !ptype->callbacks.gro_receive)
5048 EXPORT_SYMBOL(gro_find_receive_by_type);
5050 struct packet_offload *gro_find_complete_by_type(__be16 type)
5052 struct list_head *offload_head = &offload_base;
5053 struct packet_offload *ptype;
5055 list_for_each_entry_rcu(ptype, offload_head, list) {
5056 if (ptype->type != type || !ptype->callbacks.gro_complete)
5062 EXPORT_SYMBOL(gro_find_complete_by_type);
5064 static void napi_skb_free_stolen_head(struct sk_buff *skb)
5068 kmem_cache_free(skbuff_head_cache, skb);
5071 static gro_result_t napi_skb_finish(gro_result_t ret, struct sk_buff *skb)
5075 if (netif_receive_skb_internal(skb))
5083 case GRO_MERGED_FREE:
5084 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
5085 napi_skb_free_stolen_head(skb);
5099 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
5101 skb_mark_napi_id(skb, napi);
5102 trace_napi_gro_receive_entry(skb);
5104 skb_gro_reset_offset(skb);
5106 return napi_skb_finish(dev_gro_receive(napi, skb), skb);
5108 EXPORT_SYMBOL(napi_gro_receive);
5110 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
5112 if (unlikely(skb->pfmemalloc)) {
5116 __skb_pull(skb, skb_headlen(skb));
5117 /* restore the reserve we had after netdev_alloc_skb_ip_align() */
5118 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
5120 skb->dev = napi->dev;
5122 skb->encapsulation = 0;
5123 skb_shinfo(skb)->gso_type = 0;
5124 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
5130 struct sk_buff *napi_get_frags(struct napi_struct *napi)
5132 struct sk_buff *skb = napi->skb;
5135 skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
5138 skb_mark_napi_id(skb, napi);
5143 EXPORT_SYMBOL(napi_get_frags);
5145 static gro_result_t napi_frags_finish(struct napi_struct *napi,
5146 struct sk_buff *skb,
5152 __skb_push(skb, ETH_HLEN);
5153 skb->protocol = eth_type_trans(skb, skb->dev);
5154 if (ret == GRO_NORMAL && netif_receive_skb_internal(skb))
5159 napi_reuse_skb(napi, skb);
5162 case GRO_MERGED_FREE:
5163 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
5164 napi_skb_free_stolen_head(skb);
5166 napi_reuse_skb(napi, skb);
5177 /* Upper GRO stack assumes network header starts at gro_offset=0
5178 * Drivers could call both napi_gro_frags() and napi_gro_receive()
5179 * We copy ethernet header into skb->data to have a common layout.
5181 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
5183 struct sk_buff *skb = napi->skb;
5184 const struct ethhdr *eth;
5185 unsigned int hlen = sizeof(*eth);
5189 skb_reset_mac_header(skb);
5190 skb_gro_reset_offset(skb);
5192 eth = skb_gro_header_fast(skb, 0);
5193 if (unlikely(skb_gro_header_hard(skb, hlen))) {
5194 eth = skb_gro_header_slow(skb, hlen, 0);
5195 if (unlikely(!eth)) {
5196 net_warn_ratelimited("%s: dropping impossible skb from %s\n",
5197 __func__, napi->dev->name);
5198 napi_reuse_skb(napi, skb);
5202 gro_pull_from_frag0(skb, hlen);
5203 NAPI_GRO_CB(skb)->frag0 += hlen;
5204 NAPI_GRO_CB(skb)->frag0_len -= hlen;
5206 __skb_pull(skb, hlen);
5209 * This works because the only protocols we care about don't require
5211 * We'll fix it up properly in napi_frags_finish()
5213 skb->protocol = eth->h_proto;
5218 gro_result_t napi_gro_frags(struct napi_struct *napi)
5220 struct sk_buff *skb = napi_frags_skb(napi);
5225 trace_napi_gro_frags_entry(skb);
5227 return napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
5229 EXPORT_SYMBOL(napi_gro_frags);
5231 /* Compute the checksum from gro_offset and return the folded value
5232 * after adding in any pseudo checksum.
5234 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
5239 wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
5241 /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
5242 sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
5244 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
5245 !skb->csum_complete_sw)
5246 netdev_rx_csum_fault(skb->dev);
5249 NAPI_GRO_CB(skb)->csum = wsum;
5250 NAPI_GRO_CB(skb)->csum_valid = 1;
5254 EXPORT_SYMBOL(__skb_gro_checksum_complete);
5256 static void net_rps_send_ipi(struct softnet_data *remsd)
5260 struct softnet_data *next = remsd->rps_ipi_next;
5262 if (cpu_online(remsd->cpu))
5263 smp_call_function_single_async(remsd->cpu, &remsd->csd);
5270 * net_rps_action_and_irq_enable sends any pending IPI's for rps.
5271 * Note: called with local irq disabled, but exits with local irq enabled.
5273 static void net_rps_action_and_irq_enable(struct softnet_data *sd)
5276 struct softnet_data *remsd = sd->rps_ipi_list;
5279 sd->rps_ipi_list = NULL;
5283 /* Send pending IPI's to kick RPS processing on remote cpus. */
5284 net_rps_send_ipi(remsd);
5290 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd)
5293 return sd->rps_ipi_list != NULL;
5299 static int process_backlog(struct napi_struct *napi, int quota)
5301 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
5305 /* Check if we have pending ipi, its better to send them now,
5306 * not waiting net_rx_action() end.
5308 if (sd_has_rps_ipi_waiting(sd)) {
5309 local_irq_disable();
5310 net_rps_action_and_irq_enable(sd);
5313 napi->weight = dev_rx_weight;
5315 struct sk_buff *skb;
5317 while ((skb = __skb_dequeue(&sd->process_queue))) {
5319 __netif_receive_skb(skb);
5321 input_queue_head_incr(sd);
5322 if (++work >= quota)
5327 local_irq_disable();
5329 if (skb_queue_empty(&sd->input_pkt_queue)) {
5331 * Inline a custom version of __napi_complete().
5332 * only current cpu owns and manipulates this napi,
5333 * and NAPI_STATE_SCHED is the only possible flag set
5335 * We can use a plain write instead of clear_bit(),
5336 * and we dont need an smp_mb() memory barrier.
5341 skb_queue_splice_tail_init(&sd->input_pkt_queue,
5342 &sd->process_queue);
5352 * __napi_schedule - schedule for receive
5353 * @n: entry to schedule
5355 * The entry's receive function will be scheduled to run.
5356 * Consider using __napi_schedule_irqoff() if hard irqs are masked.
5358 void __napi_schedule(struct napi_struct *n)
5360 unsigned long flags;
5362 local_irq_save(flags);
5363 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
5364 local_irq_restore(flags);
5366 EXPORT_SYMBOL(__napi_schedule);
5369 * napi_schedule_prep - check if napi can be scheduled
5372 * Test if NAPI routine is already running, and if not mark
5373 * it as running. This is used as a condition variable
5374 * insure only one NAPI poll instance runs. We also make
5375 * sure there is no pending NAPI disable.
5377 bool napi_schedule_prep(struct napi_struct *n)
5379 unsigned long val, new;
5382 val = READ_ONCE(n->state);
5383 if (unlikely(val & NAPIF_STATE_DISABLE))
5385 new = val | NAPIF_STATE_SCHED;
5387 /* Sets STATE_MISSED bit if STATE_SCHED was already set
5388 * This was suggested by Alexander Duyck, as compiler
5389 * emits better code than :
5390 * if (val & NAPIF_STATE_SCHED)
5391 * new |= NAPIF_STATE_MISSED;
5393 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED *
5395 } while (cmpxchg(&n->state, val, new) != val);
5397 return !(val & NAPIF_STATE_SCHED);
5399 EXPORT_SYMBOL(napi_schedule_prep);
5402 * __napi_schedule_irqoff - schedule for receive
5403 * @n: entry to schedule
5405 * Variant of __napi_schedule() assuming hard irqs are masked
5407 void __napi_schedule_irqoff(struct napi_struct *n)
5409 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
5411 EXPORT_SYMBOL(__napi_schedule_irqoff);
5413 bool napi_complete_done(struct napi_struct *n, int work_done)
5415 unsigned long flags, val, new;
5418 * 1) Don't let napi dequeue from the cpu poll list
5419 * just in case its running on a different cpu.
5420 * 2) If we are busy polling, do nothing here, we have
5421 * the guarantee we will be called later.
5423 if (unlikely(n->state & (NAPIF_STATE_NPSVC |
5424 NAPIF_STATE_IN_BUSY_POLL)))
5428 unsigned long timeout = 0;
5431 timeout = n->dev->gro_flush_timeout;
5434 hrtimer_start(&n->timer, ns_to_ktime(timeout),
5435 HRTIMER_MODE_REL_PINNED);
5437 napi_gro_flush(n, false);
5439 if (unlikely(!list_empty(&n->poll_list))) {
5440 /* If n->poll_list is not empty, we need to mask irqs */
5441 local_irq_save(flags);
5442 list_del_init(&n->poll_list);
5443 local_irq_restore(flags);
5447 val = READ_ONCE(n->state);
5449 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED));
5451 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED);
5453 /* If STATE_MISSED was set, leave STATE_SCHED set,
5454 * because we will call napi->poll() one more time.
5455 * This C code was suggested by Alexander Duyck to help gcc.
5457 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED *
5459 } while (cmpxchg(&n->state, val, new) != val);
5461 if (unlikely(val & NAPIF_STATE_MISSED)) {
5468 EXPORT_SYMBOL(napi_complete_done);
5470 /* must be called under rcu_read_lock(), as we dont take a reference */
5471 static struct napi_struct *napi_by_id(unsigned int napi_id)
5473 unsigned int hash = napi_id % HASH_SIZE(napi_hash);
5474 struct napi_struct *napi;
5476 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node)
5477 if (napi->napi_id == napi_id)
5483 #if defined(CONFIG_NET_RX_BUSY_POLL)
5485 #define BUSY_POLL_BUDGET 8
5487 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock)
5491 /* Busy polling means there is a high chance device driver hard irq
5492 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was
5493 * set in napi_schedule_prep().
5494 * Since we are about to call napi->poll() once more, we can safely
5495 * clear NAPI_STATE_MISSED.
5497 * Note: x86 could use a single "lock and ..." instruction
5498 * to perform these two clear_bit()
5500 clear_bit(NAPI_STATE_MISSED, &napi->state);
5501 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state);
5505 /* All we really want here is to re-enable device interrupts.
5506 * Ideally, a new ndo_busy_poll_stop() could avoid another round.
5508 rc = napi->poll(napi, BUSY_POLL_BUDGET);
5509 trace_napi_poll(napi, rc, BUSY_POLL_BUDGET);
5510 netpoll_poll_unlock(have_poll_lock);
5511 if (rc == BUSY_POLL_BUDGET)
5512 __napi_schedule(napi);
5516 void napi_busy_loop(unsigned int napi_id,
5517 bool (*loop_end)(void *, unsigned long),
5520 unsigned long start_time = loop_end ? busy_loop_current_time() : 0;
5521 int (*napi_poll)(struct napi_struct *napi, int budget);
5522 void *have_poll_lock = NULL;
5523 struct napi_struct *napi;
5530 napi = napi_by_id(napi_id);
5540 unsigned long val = READ_ONCE(napi->state);
5542 /* If multiple threads are competing for this napi,
5543 * we avoid dirtying napi->state as much as we can.
5545 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED |
5546 NAPIF_STATE_IN_BUSY_POLL))
5548 if (cmpxchg(&napi->state, val,
5549 val | NAPIF_STATE_IN_BUSY_POLL |
5550 NAPIF_STATE_SCHED) != val)
5552 have_poll_lock = netpoll_poll_lock(napi);
5553 napi_poll = napi->poll;
5555 work = napi_poll(napi, BUSY_POLL_BUDGET);
5556 trace_napi_poll(napi, work, BUSY_POLL_BUDGET);
5559 __NET_ADD_STATS(dev_net(napi->dev),
5560 LINUX_MIB_BUSYPOLLRXPACKETS, work);
5563 if (!loop_end || loop_end(loop_end_arg, start_time))
5566 if (unlikely(need_resched())) {
5568 busy_poll_stop(napi, have_poll_lock);
5572 if (loop_end(loop_end_arg, start_time))
5579 busy_poll_stop(napi, have_poll_lock);
5584 EXPORT_SYMBOL(napi_busy_loop);
5586 #endif /* CONFIG_NET_RX_BUSY_POLL */
5588 static void napi_hash_add(struct napi_struct *napi)
5590 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state) ||
5591 test_and_set_bit(NAPI_STATE_HASHED, &napi->state))
5594 spin_lock(&napi_hash_lock);
5596 /* 0..NR_CPUS range is reserved for sender_cpu use */
5598 if (unlikely(++napi_gen_id < MIN_NAPI_ID))
5599 napi_gen_id = MIN_NAPI_ID;
5600 } while (napi_by_id(napi_gen_id));
5601 napi->napi_id = napi_gen_id;
5603 hlist_add_head_rcu(&napi->napi_hash_node,
5604 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]);
5606 spin_unlock(&napi_hash_lock);
5609 /* Warning : caller is responsible to make sure rcu grace period
5610 * is respected before freeing memory containing @napi
5612 bool napi_hash_del(struct napi_struct *napi)
5614 bool rcu_sync_needed = false;
5616 spin_lock(&napi_hash_lock);
5618 if (test_and_clear_bit(NAPI_STATE_HASHED, &napi->state)) {
5619 rcu_sync_needed = true;
5620 hlist_del_rcu(&napi->napi_hash_node);
5622 spin_unlock(&napi_hash_lock);
5623 return rcu_sync_needed;
5625 EXPORT_SYMBOL_GPL(napi_hash_del);
5627 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer)
5629 struct napi_struct *napi;
5631 napi = container_of(timer, struct napi_struct, timer);
5633 /* Note : we use a relaxed variant of napi_schedule_prep() not setting
5634 * NAPI_STATE_MISSED, since we do not react to a device IRQ.
5636 if (napi->gro_list && !napi_disable_pending(napi) &&
5637 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state))
5638 __napi_schedule_irqoff(napi);
5640 return HRTIMER_NORESTART;
5643 void netif_napi_add(struct net_device *dev, struct napi_struct *napi,
5644 int (*poll)(struct napi_struct *, int), int weight)
5646 INIT_LIST_HEAD(&napi->poll_list);
5647 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
5648 napi->timer.function = napi_watchdog;
5649 napi->gro_count = 0;
5650 napi->gro_list = NULL;
5653 if (weight > NAPI_POLL_WEIGHT)
5654 pr_err_once("netif_napi_add() called with weight %d on device %s\n",
5656 napi->weight = weight;
5657 list_add(&napi->dev_list, &dev->napi_list);
5659 #ifdef CONFIG_NETPOLL
5660 napi->poll_owner = -1;
5662 set_bit(NAPI_STATE_SCHED, &napi->state);
5663 napi_hash_add(napi);
5665 EXPORT_SYMBOL(netif_napi_add);
5667 void napi_disable(struct napi_struct *n)
5670 set_bit(NAPI_STATE_DISABLE, &n->state);
5672 while (test_and_set_bit(NAPI_STATE_SCHED, &n->state))
5674 while (test_and_set_bit(NAPI_STATE_NPSVC, &n->state))
5677 hrtimer_cancel(&n->timer);
5679 clear_bit(NAPI_STATE_DISABLE, &n->state);
5681 EXPORT_SYMBOL(napi_disable);
5683 /* Must be called in process context */
5684 void netif_napi_del(struct napi_struct *napi)
5687 if (napi_hash_del(napi))
5689 list_del_init(&napi->dev_list);
5690 napi_free_frags(napi);
5692 kfree_skb_list(napi->gro_list);
5693 napi->gro_list = NULL;
5694 napi->gro_count = 0;
5696 EXPORT_SYMBOL(netif_napi_del);
5698 static int napi_poll(struct napi_struct *n, struct list_head *repoll)
5703 list_del_init(&n->poll_list);
5705 have = netpoll_poll_lock(n);
5709 /* This NAPI_STATE_SCHED test is for avoiding a race
5710 * with netpoll's poll_napi(). Only the entity which
5711 * obtains the lock and sees NAPI_STATE_SCHED set will
5712 * actually make the ->poll() call. Therefore we avoid
5713 * accidentally calling ->poll() when NAPI is not scheduled.
5716 if (test_bit(NAPI_STATE_SCHED, &n->state)) {
5717 work = n->poll(n, weight);
5718 trace_napi_poll(n, work, weight);
5721 WARN_ON_ONCE(work > weight);
5723 if (likely(work < weight))
5726 /* Drivers must not modify the NAPI state if they
5727 * consume the entire weight. In such cases this code
5728 * still "owns" the NAPI instance and therefore can
5729 * move the instance around on the list at-will.
5731 if (unlikely(napi_disable_pending(n))) {
5737 /* flush too old packets
5738 * If HZ < 1000, flush all packets.
5740 napi_gro_flush(n, HZ >= 1000);
5743 /* Some drivers may have called napi_schedule
5744 * prior to exhausting their budget.
5746 if (unlikely(!list_empty(&n->poll_list))) {
5747 pr_warn_once("%s: Budget exhausted after napi rescheduled\n",
5748 n->dev ? n->dev->name : "backlog");
5752 list_add_tail(&n->poll_list, repoll);
5755 netpoll_poll_unlock(have);
5760 static __latent_entropy void net_rx_action(struct softirq_action *h)
5762 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
5763 unsigned long time_limit = jiffies +
5764 usecs_to_jiffies(netdev_budget_usecs);
5765 int budget = netdev_budget;
5769 local_irq_disable();
5770 list_splice_init(&sd->poll_list, &list);
5774 struct napi_struct *n;
5776 if (list_empty(&list)) {
5777 if (!sd_has_rps_ipi_waiting(sd) && list_empty(&repoll))
5782 n = list_first_entry(&list, struct napi_struct, poll_list);
5783 budget -= napi_poll(n, &repoll);
5785 /* If softirq window is exhausted then punt.
5786 * Allow this to run for 2 jiffies since which will allow
5787 * an average latency of 1.5/HZ.
5789 if (unlikely(budget <= 0 ||
5790 time_after_eq(jiffies, time_limit))) {
5796 local_irq_disable();
5798 list_splice_tail_init(&sd->poll_list, &list);
5799 list_splice_tail(&repoll, &list);
5800 list_splice(&list, &sd->poll_list);
5801 if (!list_empty(&sd->poll_list))
5802 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
5804 net_rps_action_and_irq_enable(sd);
5806 __kfree_skb_flush();
5809 struct netdev_adjacent {
5810 struct net_device *dev;
5812 /* upper master flag, there can only be one master device per list */
5815 /* counter for the number of times this device was added to us */
5818 /* private field for the users */
5821 struct list_head list;
5822 struct rcu_head rcu;
5825 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev,
5826 struct list_head *adj_list)
5828 struct netdev_adjacent *adj;
5830 list_for_each_entry(adj, adj_list, list) {
5831 if (adj->dev == adj_dev)
5837 static int __netdev_has_upper_dev(struct net_device *upper_dev, void *data)
5839 struct net_device *dev = data;
5841 return upper_dev == dev;
5845 * netdev_has_upper_dev - Check if device is linked to an upper device
5847 * @upper_dev: upper device to check
5849 * Find out if a device is linked to specified upper device and return true
5850 * in case it is. Note that this checks only immediate upper device,
5851 * not through a complete stack of devices. The caller must hold the RTNL lock.
5853 bool netdev_has_upper_dev(struct net_device *dev,
5854 struct net_device *upper_dev)
5858 return netdev_walk_all_upper_dev_rcu(dev, __netdev_has_upper_dev,
5861 EXPORT_SYMBOL(netdev_has_upper_dev);
5864 * netdev_has_upper_dev_all - Check if device is linked to an upper device
5866 * @upper_dev: upper device to check
5868 * Find out if a device is linked to specified upper device and return true
5869 * in case it is. Note that this checks the entire upper device chain.
5870 * The caller must hold rcu lock.
5873 bool netdev_has_upper_dev_all_rcu(struct net_device *dev,
5874 struct net_device *upper_dev)
5876 return !!netdev_walk_all_upper_dev_rcu(dev, __netdev_has_upper_dev,
5879 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu);
5882 * netdev_has_any_upper_dev - Check if device is linked to some device
5885 * Find out if a device is linked to an upper device and return true in case
5886 * it is. The caller must hold the RTNL lock.
5888 bool netdev_has_any_upper_dev(struct net_device *dev)
5892 return !list_empty(&dev->adj_list.upper);
5894 EXPORT_SYMBOL(netdev_has_any_upper_dev);
5897 * netdev_master_upper_dev_get - Get master upper device
5900 * Find a master upper device and return pointer to it or NULL in case
5901 * it's not there. The caller must hold the RTNL lock.
5903 struct net_device *netdev_master_upper_dev_get(struct net_device *dev)
5905 struct netdev_adjacent *upper;
5909 if (list_empty(&dev->adj_list.upper))
5912 upper = list_first_entry(&dev->adj_list.upper,
5913 struct netdev_adjacent, list);
5914 if (likely(upper->master))
5918 EXPORT_SYMBOL(netdev_master_upper_dev_get);
5921 * netdev_has_any_lower_dev - Check if device is linked to some device
5924 * Find out if a device is linked to a lower device and return true in case
5925 * it is. The caller must hold the RTNL lock.
5927 static bool netdev_has_any_lower_dev(struct net_device *dev)
5931 return !list_empty(&dev->adj_list.lower);
5934 void *netdev_adjacent_get_private(struct list_head *adj_list)
5936 struct netdev_adjacent *adj;
5938 adj = list_entry(adj_list, struct netdev_adjacent, list);
5940 return adj->private;
5942 EXPORT_SYMBOL(netdev_adjacent_get_private);
5945 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list
5947 * @iter: list_head ** of the current position
5949 * Gets the next device from the dev's upper list, starting from iter
5950 * position. The caller must hold RCU read lock.
5952 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev,
5953 struct list_head **iter)
5955 struct netdev_adjacent *upper;
5957 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
5959 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
5961 if (&upper->list == &dev->adj_list.upper)
5964 *iter = &upper->list;
5968 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu);
5970 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev,
5971 struct list_head **iter)
5973 struct netdev_adjacent *upper;
5975 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
5977 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
5979 if (&upper->list == &dev->adj_list.upper)
5982 *iter = &upper->list;
5987 int netdev_walk_all_upper_dev_rcu(struct net_device *dev,
5988 int (*fn)(struct net_device *dev,
5992 struct net_device *udev;
5993 struct list_head *iter;
5996 for (iter = &dev->adj_list.upper,
5997 udev = netdev_next_upper_dev_rcu(dev, &iter);
5999 udev = netdev_next_upper_dev_rcu(dev, &iter)) {
6000 /* first is the upper device itself */
6001 ret = fn(udev, data);
6005 /* then look at all of its upper devices */
6006 ret = netdev_walk_all_upper_dev_rcu(udev, fn, data);
6013 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu);
6016 * netdev_lower_get_next_private - Get the next ->private from the
6017 * lower neighbour list
6019 * @iter: list_head ** of the current position
6021 * Gets the next netdev_adjacent->private from the dev's lower neighbour
6022 * list, starting from iter position. The caller must hold either hold the
6023 * RTNL lock or its own locking that guarantees that the neighbour lower
6024 * list will remain unchanged.
6026 void *netdev_lower_get_next_private(struct net_device *dev,
6027 struct list_head **iter)
6029 struct netdev_adjacent *lower;
6031 lower = list_entry(*iter, struct netdev_adjacent, list);
6033 if (&lower->list == &dev->adj_list.lower)
6036 *iter = lower->list.next;
6038 return lower->private;
6040 EXPORT_SYMBOL(netdev_lower_get_next_private);
6043 * netdev_lower_get_next_private_rcu - Get the next ->private from the
6044 * lower neighbour list, RCU
6047 * @iter: list_head ** of the current position
6049 * Gets the next netdev_adjacent->private from the dev's lower neighbour
6050 * list, starting from iter position. The caller must hold RCU read lock.
6052 void *netdev_lower_get_next_private_rcu(struct net_device *dev,
6053 struct list_head **iter)
6055 struct netdev_adjacent *lower;
6057 WARN_ON_ONCE(!rcu_read_lock_held());
6059 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6061 if (&lower->list == &dev->adj_list.lower)
6064 *iter = &lower->list;
6066 return lower->private;
6068 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu);
6071 * netdev_lower_get_next - Get the next device from the lower neighbour
6074 * @iter: list_head ** of the current position
6076 * Gets the next netdev_adjacent from the dev's lower neighbour
6077 * list, starting from iter position. The caller must hold RTNL lock or
6078 * its own locking that guarantees that the neighbour lower
6079 * list will remain unchanged.
6081 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter)
6083 struct netdev_adjacent *lower;
6085 lower = list_entry(*iter, struct netdev_adjacent, list);
6087 if (&lower->list == &dev->adj_list.lower)
6090 *iter = lower->list.next;
6094 EXPORT_SYMBOL(netdev_lower_get_next);
6096 static struct net_device *netdev_next_lower_dev(struct net_device *dev,
6097 struct list_head **iter)
6099 struct netdev_adjacent *lower;
6101 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
6103 if (&lower->list == &dev->adj_list.lower)
6106 *iter = &lower->list;
6111 int netdev_walk_all_lower_dev(struct net_device *dev,
6112 int (*fn)(struct net_device *dev,
6116 struct net_device *ldev;
6117 struct list_head *iter;
6120 for (iter = &dev->adj_list.lower,
6121 ldev = netdev_next_lower_dev(dev, &iter);
6123 ldev = netdev_next_lower_dev(dev, &iter)) {
6124 /* first is the lower device itself */
6125 ret = fn(ldev, data);
6129 /* then look at all of its lower devices */
6130 ret = netdev_walk_all_lower_dev(ldev, fn, data);
6137 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev);
6139 static struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev,
6140 struct list_head **iter)
6142 struct netdev_adjacent *lower;
6144 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6145 if (&lower->list == &dev->adj_list.lower)
6148 *iter = &lower->list;
6153 int netdev_walk_all_lower_dev_rcu(struct net_device *dev,
6154 int (*fn)(struct net_device *dev,
6158 struct net_device *ldev;
6159 struct list_head *iter;
6162 for (iter = &dev->adj_list.lower,
6163 ldev = netdev_next_lower_dev_rcu(dev, &iter);
6165 ldev = netdev_next_lower_dev_rcu(dev, &iter)) {
6166 /* first is the lower device itself */
6167 ret = fn(ldev, data);
6171 /* then look at all of its lower devices */
6172 ret = netdev_walk_all_lower_dev_rcu(ldev, fn, data);
6179 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu);
6182 * netdev_lower_get_first_private_rcu - Get the first ->private from the
6183 * lower neighbour list, RCU
6187 * Gets the first netdev_adjacent->private from the dev's lower neighbour
6188 * list. The caller must hold RCU read lock.
6190 void *netdev_lower_get_first_private_rcu(struct net_device *dev)
6192 struct netdev_adjacent *lower;
6194 lower = list_first_or_null_rcu(&dev->adj_list.lower,
6195 struct netdev_adjacent, list);
6197 return lower->private;
6200 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu);
6203 * netdev_master_upper_dev_get_rcu - Get master upper device
6206 * Find a master upper device and return pointer to it or NULL in case
6207 * it's not there. The caller must hold the RCU read lock.
6209 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev)
6211 struct netdev_adjacent *upper;
6213 upper = list_first_or_null_rcu(&dev->adj_list.upper,
6214 struct netdev_adjacent, list);
6215 if (upper && likely(upper->master))
6219 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu);
6221 static int netdev_adjacent_sysfs_add(struct net_device *dev,
6222 struct net_device *adj_dev,
6223 struct list_head *dev_list)
6225 char linkname[IFNAMSIZ+7];
6227 sprintf(linkname, dev_list == &dev->adj_list.upper ?
6228 "upper_%s" : "lower_%s", adj_dev->name);
6229 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj),
6232 static void netdev_adjacent_sysfs_del(struct net_device *dev,
6234 struct list_head *dev_list)
6236 char linkname[IFNAMSIZ+7];
6238 sprintf(linkname, dev_list == &dev->adj_list.upper ?
6239 "upper_%s" : "lower_%s", name);
6240 sysfs_remove_link(&(dev->dev.kobj), linkname);
6243 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev,
6244 struct net_device *adj_dev,
6245 struct list_head *dev_list)
6247 return (dev_list == &dev->adj_list.upper ||
6248 dev_list == &dev->adj_list.lower) &&
6249 net_eq(dev_net(dev), dev_net(adj_dev));
6252 static int __netdev_adjacent_dev_insert(struct net_device *dev,
6253 struct net_device *adj_dev,
6254 struct list_head *dev_list,
6255 void *private, bool master)
6257 struct netdev_adjacent *adj;
6260 adj = __netdev_find_adj(adj_dev, dev_list);
6264 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n",
6265 dev->name, adj_dev->name, adj->ref_nr);
6270 adj = kmalloc(sizeof(*adj), GFP_KERNEL);
6275 adj->master = master;
6277 adj->private = private;
6280 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n",
6281 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name);
6283 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) {
6284 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list);
6289 /* Ensure that master link is always the first item in list. */
6291 ret = sysfs_create_link(&(dev->dev.kobj),
6292 &(adj_dev->dev.kobj), "master");
6294 goto remove_symlinks;
6296 list_add_rcu(&adj->list, dev_list);
6298 list_add_tail_rcu(&adj->list, dev_list);
6304 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
6305 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
6313 static void __netdev_adjacent_dev_remove(struct net_device *dev,
6314 struct net_device *adj_dev,
6316 struct list_head *dev_list)
6318 struct netdev_adjacent *adj;
6320 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n",
6321 dev->name, adj_dev->name, ref_nr);
6323 adj = __netdev_find_adj(adj_dev, dev_list);
6326 pr_err("Adjacency does not exist for device %s from %s\n",
6327 dev->name, adj_dev->name);
6332 if (adj->ref_nr > ref_nr) {
6333 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n",
6334 dev->name, adj_dev->name, ref_nr,
6335 adj->ref_nr - ref_nr);
6336 adj->ref_nr -= ref_nr;
6341 sysfs_remove_link(&(dev->dev.kobj), "master");
6343 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
6344 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
6346 list_del_rcu(&adj->list);
6347 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n",
6348 adj_dev->name, dev->name, adj_dev->name);
6350 kfree_rcu(adj, rcu);
6353 static int __netdev_adjacent_dev_link_lists(struct net_device *dev,
6354 struct net_device *upper_dev,
6355 struct list_head *up_list,
6356 struct list_head *down_list,
6357 void *private, bool master)
6361 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list,
6366 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list,
6369 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list);
6376 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev,
6377 struct net_device *upper_dev,
6379 struct list_head *up_list,
6380 struct list_head *down_list)
6382 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list);
6383 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list);
6386 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev,
6387 struct net_device *upper_dev,
6388 void *private, bool master)
6390 return __netdev_adjacent_dev_link_lists(dev, upper_dev,
6391 &dev->adj_list.upper,
6392 &upper_dev->adj_list.lower,
6396 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev,
6397 struct net_device *upper_dev)
6399 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1,
6400 &dev->adj_list.upper,
6401 &upper_dev->adj_list.lower);
6404 static int __netdev_upper_dev_link(struct net_device *dev,
6405 struct net_device *upper_dev, bool master,
6406 void *upper_priv, void *upper_info,
6407 struct netlink_ext_ack *extack)
6409 struct netdev_notifier_changeupper_info changeupper_info = {
6414 .upper_dev = upper_dev,
6417 .upper_info = upper_info,
6419 struct net_device *master_dev;
6424 if (dev == upper_dev)
6427 /* To prevent loops, check if dev is not upper device to upper_dev. */
6428 if (netdev_has_upper_dev(upper_dev, dev))
6432 if (netdev_has_upper_dev(dev, upper_dev))
6435 master_dev = netdev_master_upper_dev_get(dev);
6437 return master_dev == upper_dev ? -EEXIST : -EBUSY;
6440 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
6441 &changeupper_info.info);
6442 ret = notifier_to_errno(ret);
6446 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv,
6451 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
6452 &changeupper_info.info);
6453 ret = notifier_to_errno(ret);
6460 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
6466 * netdev_upper_dev_link - Add a link to the upper device
6468 * @upper_dev: new upper device
6469 * @extack: netlink extended ack
6471 * Adds a link to device which is upper to this one. The caller must hold
6472 * the RTNL lock. On a failure a negative errno code is returned.
6473 * On success the reference counts are adjusted and the function
6476 int netdev_upper_dev_link(struct net_device *dev,
6477 struct net_device *upper_dev,
6478 struct netlink_ext_ack *extack)
6480 return __netdev_upper_dev_link(dev, upper_dev, false,
6481 NULL, NULL, extack);
6483 EXPORT_SYMBOL(netdev_upper_dev_link);
6486 * netdev_master_upper_dev_link - Add a master link to the upper device
6488 * @upper_dev: new upper device
6489 * @upper_priv: upper device private
6490 * @upper_info: upper info to be passed down via notifier
6491 * @extack: netlink extended ack
6493 * Adds a link to device which is upper to this one. In this case, only
6494 * one master upper device can be linked, although other non-master devices
6495 * might be linked as well. The caller must hold the RTNL lock.
6496 * On a failure a negative errno code is returned. On success the reference
6497 * counts are adjusted and the function returns zero.
6499 int netdev_master_upper_dev_link(struct net_device *dev,
6500 struct net_device *upper_dev,
6501 void *upper_priv, void *upper_info,
6502 struct netlink_ext_ack *extack)
6504 return __netdev_upper_dev_link(dev, upper_dev, true,
6505 upper_priv, upper_info, extack);
6507 EXPORT_SYMBOL(netdev_master_upper_dev_link);
6510 * netdev_upper_dev_unlink - Removes a link to upper device
6512 * @upper_dev: new upper device
6514 * Removes a link to device which is upper to this one. The caller must hold
6517 void netdev_upper_dev_unlink(struct net_device *dev,
6518 struct net_device *upper_dev)
6520 struct netdev_notifier_changeupper_info changeupper_info = {
6524 .upper_dev = upper_dev,
6530 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev;
6532 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
6533 &changeupper_info.info);
6535 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
6537 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
6538 &changeupper_info.info);
6540 EXPORT_SYMBOL(netdev_upper_dev_unlink);
6543 * netdev_bonding_info_change - Dispatch event about slave change
6545 * @bonding_info: info to dispatch
6547 * Send NETDEV_BONDING_INFO to netdev notifiers with info.
6548 * The caller must hold the RTNL lock.
6550 void netdev_bonding_info_change(struct net_device *dev,
6551 struct netdev_bonding_info *bonding_info)
6553 struct netdev_notifier_bonding_info info = {
6557 memcpy(&info.bonding_info, bonding_info,
6558 sizeof(struct netdev_bonding_info));
6559 call_netdevice_notifiers_info(NETDEV_BONDING_INFO,
6562 EXPORT_SYMBOL(netdev_bonding_info_change);
6564 static void netdev_adjacent_add_links(struct net_device *dev)
6566 struct netdev_adjacent *iter;
6568 struct net *net = dev_net(dev);
6570 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6571 if (!net_eq(net, dev_net(iter->dev)))
6573 netdev_adjacent_sysfs_add(iter->dev, dev,
6574 &iter->dev->adj_list.lower);
6575 netdev_adjacent_sysfs_add(dev, iter->dev,
6576 &dev->adj_list.upper);
6579 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6580 if (!net_eq(net, dev_net(iter->dev)))
6582 netdev_adjacent_sysfs_add(iter->dev, dev,
6583 &iter->dev->adj_list.upper);
6584 netdev_adjacent_sysfs_add(dev, iter->dev,
6585 &dev->adj_list.lower);
6589 static void netdev_adjacent_del_links(struct net_device *dev)
6591 struct netdev_adjacent *iter;
6593 struct net *net = dev_net(dev);
6595 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6596 if (!net_eq(net, dev_net(iter->dev)))
6598 netdev_adjacent_sysfs_del(iter->dev, dev->name,
6599 &iter->dev->adj_list.lower);
6600 netdev_adjacent_sysfs_del(dev, iter->dev->name,
6601 &dev->adj_list.upper);
6604 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6605 if (!net_eq(net, dev_net(iter->dev)))
6607 netdev_adjacent_sysfs_del(iter->dev, dev->name,
6608 &iter->dev->adj_list.upper);
6609 netdev_adjacent_sysfs_del(dev, iter->dev->name,
6610 &dev->adj_list.lower);
6614 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname)
6616 struct netdev_adjacent *iter;
6618 struct net *net = dev_net(dev);
6620 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6621 if (!net_eq(net, dev_net(iter->dev)))
6623 netdev_adjacent_sysfs_del(iter->dev, oldname,
6624 &iter->dev->adj_list.lower);
6625 netdev_adjacent_sysfs_add(iter->dev, dev,
6626 &iter->dev->adj_list.lower);
6629 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6630 if (!net_eq(net, dev_net(iter->dev)))
6632 netdev_adjacent_sysfs_del(iter->dev, oldname,
6633 &iter->dev->adj_list.upper);
6634 netdev_adjacent_sysfs_add(iter->dev, dev,
6635 &iter->dev->adj_list.upper);
6639 void *netdev_lower_dev_get_private(struct net_device *dev,
6640 struct net_device *lower_dev)
6642 struct netdev_adjacent *lower;
6646 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower);
6650 return lower->private;
6652 EXPORT_SYMBOL(netdev_lower_dev_get_private);
6655 int dev_get_nest_level(struct net_device *dev)
6657 struct net_device *lower = NULL;
6658 struct list_head *iter;
6664 netdev_for_each_lower_dev(dev, lower, iter) {
6665 nest = dev_get_nest_level(lower);
6666 if (max_nest < nest)
6670 return max_nest + 1;
6672 EXPORT_SYMBOL(dev_get_nest_level);
6675 * netdev_lower_change - Dispatch event about lower device state change
6676 * @lower_dev: device
6677 * @lower_state_info: state to dispatch
6679 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info.
6680 * The caller must hold the RTNL lock.
6682 void netdev_lower_state_changed(struct net_device *lower_dev,
6683 void *lower_state_info)
6685 struct netdev_notifier_changelowerstate_info changelowerstate_info = {
6686 .info.dev = lower_dev,
6690 changelowerstate_info.lower_state_info = lower_state_info;
6691 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE,
6692 &changelowerstate_info.info);
6694 EXPORT_SYMBOL(netdev_lower_state_changed);
6696 static void dev_change_rx_flags(struct net_device *dev, int flags)
6698 const struct net_device_ops *ops = dev->netdev_ops;
6700 if (ops->ndo_change_rx_flags)
6701 ops->ndo_change_rx_flags(dev, flags);
6704 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify)
6706 unsigned int old_flags = dev->flags;
6712 dev->flags |= IFF_PROMISC;
6713 dev->promiscuity += inc;
6714 if (dev->promiscuity == 0) {
6717 * If inc causes overflow, untouch promisc and return error.
6720 dev->flags &= ~IFF_PROMISC;
6722 dev->promiscuity -= inc;
6723 pr_warn("%s: promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n",
6728 if (dev->flags != old_flags) {
6729 pr_info("device %s %s promiscuous mode\n",
6731 dev->flags & IFF_PROMISC ? "entered" : "left");
6732 if (audit_enabled) {
6733 current_uid_gid(&uid, &gid);
6734 audit_log(audit_context(), GFP_ATOMIC,
6735 AUDIT_ANOM_PROMISCUOUS,
6736 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u",
6737 dev->name, (dev->flags & IFF_PROMISC),
6738 (old_flags & IFF_PROMISC),
6739 from_kuid(&init_user_ns, audit_get_loginuid(current)),
6740 from_kuid(&init_user_ns, uid),
6741 from_kgid(&init_user_ns, gid),
6742 audit_get_sessionid(current));
6745 dev_change_rx_flags(dev, IFF_PROMISC);
6748 __dev_notify_flags(dev, old_flags, IFF_PROMISC);
6753 * dev_set_promiscuity - update promiscuity count on a device
6757 * Add or remove promiscuity from a device. While the count in the device
6758 * remains above zero the interface remains promiscuous. Once it hits zero
6759 * the device reverts back to normal filtering operation. A negative inc
6760 * value is used to drop promiscuity on the device.
6761 * Return 0 if successful or a negative errno code on error.
6763 int dev_set_promiscuity(struct net_device *dev, int inc)
6765 unsigned int old_flags = dev->flags;
6768 err = __dev_set_promiscuity(dev, inc, true);
6771 if (dev->flags != old_flags)
6772 dev_set_rx_mode(dev);
6775 EXPORT_SYMBOL(dev_set_promiscuity);
6777 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify)
6779 unsigned int old_flags = dev->flags, old_gflags = dev->gflags;
6783 dev->flags |= IFF_ALLMULTI;
6784 dev->allmulti += inc;
6785 if (dev->allmulti == 0) {
6788 * If inc causes overflow, untouch allmulti and return error.
6791 dev->flags &= ~IFF_ALLMULTI;
6793 dev->allmulti -= inc;
6794 pr_warn("%s: allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n",
6799 if (dev->flags ^ old_flags) {
6800 dev_change_rx_flags(dev, IFF_ALLMULTI);
6801 dev_set_rx_mode(dev);
6803 __dev_notify_flags(dev, old_flags,
6804 dev->gflags ^ old_gflags);
6810 * dev_set_allmulti - update allmulti count on a device
6814 * Add or remove reception of all multicast frames to a device. While the
6815 * count in the device remains above zero the interface remains listening
6816 * to all interfaces. Once it hits zero the device reverts back to normal
6817 * filtering operation. A negative @inc value is used to drop the counter
6818 * when releasing a resource needing all multicasts.
6819 * Return 0 if successful or a negative errno code on error.
6822 int dev_set_allmulti(struct net_device *dev, int inc)
6824 return __dev_set_allmulti(dev, inc, true);
6826 EXPORT_SYMBOL(dev_set_allmulti);
6829 * Upload unicast and multicast address lists to device and
6830 * configure RX filtering. When the device doesn't support unicast
6831 * filtering it is put in promiscuous mode while unicast addresses
6834 void __dev_set_rx_mode(struct net_device *dev)
6836 const struct net_device_ops *ops = dev->netdev_ops;
6838 /* dev_open will call this function so the list will stay sane. */
6839 if (!(dev->flags&IFF_UP))
6842 if (!netif_device_present(dev))
6845 if (!(dev->priv_flags & IFF_UNICAST_FLT)) {
6846 /* Unicast addresses changes may only happen under the rtnl,
6847 * therefore calling __dev_set_promiscuity here is safe.
6849 if (!netdev_uc_empty(dev) && !dev->uc_promisc) {
6850 __dev_set_promiscuity(dev, 1, false);
6851 dev->uc_promisc = true;
6852 } else if (netdev_uc_empty(dev) && dev->uc_promisc) {
6853 __dev_set_promiscuity(dev, -1, false);
6854 dev->uc_promisc = false;
6858 if (ops->ndo_set_rx_mode)
6859 ops->ndo_set_rx_mode(dev);
6862 void dev_set_rx_mode(struct net_device *dev)
6864 netif_addr_lock_bh(dev);
6865 __dev_set_rx_mode(dev);
6866 netif_addr_unlock_bh(dev);
6870 * dev_get_flags - get flags reported to userspace
6873 * Get the combination of flag bits exported through APIs to userspace.
6875 unsigned int dev_get_flags(const struct net_device *dev)
6879 flags = (dev->flags & ~(IFF_PROMISC |
6884 (dev->gflags & (IFF_PROMISC |
6887 if (netif_running(dev)) {
6888 if (netif_oper_up(dev))
6889 flags |= IFF_RUNNING;
6890 if (netif_carrier_ok(dev))
6891 flags |= IFF_LOWER_UP;
6892 if (netif_dormant(dev))
6893 flags |= IFF_DORMANT;
6898 EXPORT_SYMBOL(dev_get_flags);
6900 int __dev_change_flags(struct net_device *dev, unsigned int flags)
6902 unsigned int old_flags = dev->flags;
6908 * Set the flags on our device.
6911 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP |
6912 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL |
6914 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC |
6918 * Load in the correct multicast list now the flags have changed.
6921 if ((old_flags ^ flags) & IFF_MULTICAST)
6922 dev_change_rx_flags(dev, IFF_MULTICAST);
6924 dev_set_rx_mode(dev);
6927 * Have we downed the interface. We handle IFF_UP ourselves
6928 * according to user attempts to set it, rather than blindly
6933 if ((old_flags ^ flags) & IFF_UP) {
6934 if (old_flags & IFF_UP)
6937 ret = __dev_open(dev);
6940 if ((flags ^ dev->gflags) & IFF_PROMISC) {
6941 int inc = (flags & IFF_PROMISC) ? 1 : -1;
6942 unsigned int old_flags = dev->flags;
6944 dev->gflags ^= IFF_PROMISC;
6946 if (__dev_set_promiscuity(dev, inc, false) >= 0)
6947 if (dev->flags != old_flags)
6948 dev_set_rx_mode(dev);
6951 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI
6952 * is important. Some (broken) drivers set IFF_PROMISC, when
6953 * IFF_ALLMULTI is requested not asking us and not reporting.
6955 if ((flags ^ dev->gflags) & IFF_ALLMULTI) {
6956 int inc = (flags & IFF_ALLMULTI) ? 1 : -1;
6958 dev->gflags ^= IFF_ALLMULTI;
6959 __dev_set_allmulti(dev, inc, false);
6965 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags,
6966 unsigned int gchanges)
6968 unsigned int changes = dev->flags ^ old_flags;
6971 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC);
6973 if (changes & IFF_UP) {
6974 if (dev->flags & IFF_UP)
6975 call_netdevice_notifiers(NETDEV_UP, dev);
6977 call_netdevice_notifiers(NETDEV_DOWN, dev);
6980 if (dev->flags & IFF_UP &&
6981 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) {
6982 struct netdev_notifier_change_info change_info = {
6986 .flags_changed = changes,
6989 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info);
6994 * dev_change_flags - change device settings
6996 * @flags: device state flags
6998 * Change settings on device based state flags. The flags are
6999 * in the userspace exported format.
7001 int dev_change_flags(struct net_device *dev, unsigned int flags)
7004 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags;
7006 ret = __dev_change_flags(dev, flags);
7010 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags);
7011 __dev_notify_flags(dev, old_flags, changes);
7014 EXPORT_SYMBOL(dev_change_flags);
7016 int __dev_set_mtu(struct net_device *dev, int new_mtu)
7018 const struct net_device_ops *ops = dev->netdev_ops;
7020 if (ops->ndo_change_mtu)
7021 return ops->ndo_change_mtu(dev, new_mtu);
7026 EXPORT_SYMBOL(__dev_set_mtu);
7029 * dev_set_mtu - Change maximum transfer unit
7031 * @new_mtu: new transfer unit
7033 * Change the maximum transfer size of the network device.
7035 int dev_set_mtu(struct net_device *dev, int new_mtu)
7039 if (new_mtu == dev->mtu)
7042 /* MTU must be positive, and in range */
7043 if (new_mtu < 0 || new_mtu < dev->min_mtu) {
7044 net_err_ratelimited("%s: Invalid MTU %d requested, hw min %d\n",
7045 dev->name, new_mtu, dev->min_mtu);
7049 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) {
7050 net_err_ratelimited("%s: Invalid MTU %d requested, hw max %d\n",
7051 dev->name, new_mtu, dev->max_mtu);
7055 if (!netif_device_present(dev))
7058 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev);
7059 err = notifier_to_errno(err);
7063 orig_mtu = dev->mtu;
7064 err = __dev_set_mtu(dev, new_mtu);
7067 err = call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
7068 err = notifier_to_errno(err);
7070 /* setting mtu back and notifying everyone again,
7071 * so that they have a chance to revert changes.
7073 __dev_set_mtu(dev, orig_mtu);
7074 call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
7079 EXPORT_SYMBOL(dev_set_mtu);
7082 * dev_change_tx_queue_len - Change TX queue length of a netdevice
7084 * @new_len: new tx queue length
7086 int dev_change_tx_queue_len(struct net_device *dev, unsigned long new_len)
7088 unsigned int orig_len = dev->tx_queue_len;
7091 if (new_len != (unsigned int)new_len)
7094 if (new_len != orig_len) {
7095 dev->tx_queue_len = new_len;
7096 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev);
7097 res = notifier_to_errno(res);
7100 "refused to change device tx_queue_len\n");
7101 dev->tx_queue_len = orig_len;
7104 return dev_qdisc_change_tx_queue_len(dev);
7111 * dev_set_group - Change group this device belongs to
7113 * @new_group: group this device should belong to
7115 void dev_set_group(struct net_device *dev, int new_group)
7117 dev->group = new_group;
7119 EXPORT_SYMBOL(dev_set_group);
7122 * dev_set_mac_address - Change Media Access Control Address
7126 * Change the hardware (MAC) address of the device
7128 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa)
7130 const struct net_device_ops *ops = dev->netdev_ops;
7133 if (!ops->ndo_set_mac_address)
7135 if (sa->sa_family != dev->type)
7137 if (!netif_device_present(dev))
7139 err = ops->ndo_set_mac_address(dev, sa);
7142 dev->addr_assign_type = NET_ADDR_SET;
7143 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
7144 add_device_randomness(dev->dev_addr, dev->addr_len);
7147 EXPORT_SYMBOL(dev_set_mac_address);
7150 * dev_change_carrier - Change device carrier
7152 * @new_carrier: new value
7154 * Change device carrier
7156 int dev_change_carrier(struct net_device *dev, bool new_carrier)
7158 const struct net_device_ops *ops = dev->netdev_ops;
7160 if (!ops->ndo_change_carrier)
7162 if (!netif_device_present(dev))
7164 return ops->ndo_change_carrier(dev, new_carrier);
7166 EXPORT_SYMBOL(dev_change_carrier);
7169 * dev_get_phys_port_id - Get device physical port ID
7173 * Get device physical port ID
7175 int dev_get_phys_port_id(struct net_device *dev,
7176 struct netdev_phys_item_id *ppid)
7178 const struct net_device_ops *ops = dev->netdev_ops;
7180 if (!ops->ndo_get_phys_port_id)
7182 return ops->ndo_get_phys_port_id(dev, ppid);
7184 EXPORT_SYMBOL(dev_get_phys_port_id);
7187 * dev_get_phys_port_name - Get device physical port name
7190 * @len: limit of bytes to copy to name
7192 * Get device physical port name
7194 int dev_get_phys_port_name(struct net_device *dev,
7195 char *name, size_t len)
7197 const struct net_device_ops *ops = dev->netdev_ops;
7199 if (!ops->ndo_get_phys_port_name)
7201 return ops->ndo_get_phys_port_name(dev, name, len);
7203 EXPORT_SYMBOL(dev_get_phys_port_name);
7206 * dev_change_proto_down - update protocol port state information
7208 * @proto_down: new value
7210 * This info can be used by switch drivers to set the phys state of the
7213 int dev_change_proto_down(struct net_device *dev, bool proto_down)
7215 const struct net_device_ops *ops = dev->netdev_ops;
7217 if (!ops->ndo_change_proto_down)
7219 if (!netif_device_present(dev))
7221 return ops->ndo_change_proto_down(dev, proto_down);
7223 EXPORT_SYMBOL(dev_change_proto_down);
7225 void __dev_xdp_query(struct net_device *dev, bpf_op_t bpf_op,
7226 struct netdev_bpf *xdp)
7228 memset(xdp, 0, sizeof(*xdp));
7229 xdp->command = XDP_QUERY_PROG;
7231 /* Query must always succeed. */
7232 WARN_ON(bpf_op(dev, xdp) < 0);
7235 static u8 __dev_xdp_attached(struct net_device *dev, bpf_op_t bpf_op)
7237 struct netdev_bpf xdp;
7239 __dev_xdp_query(dev, bpf_op, &xdp);
7241 return xdp.prog_attached;
7244 static int dev_xdp_install(struct net_device *dev, bpf_op_t bpf_op,
7245 struct netlink_ext_ack *extack, u32 flags,
7246 struct bpf_prog *prog)
7248 struct netdev_bpf xdp;
7250 memset(&xdp, 0, sizeof(xdp));
7251 if (flags & XDP_FLAGS_HW_MODE)
7252 xdp.command = XDP_SETUP_PROG_HW;
7254 xdp.command = XDP_SETUP_PROG;
7255 xdp.extack = extack;
7259 return bpf_op(dev, &xdp);
7262 static void dev_xdp_uninstall(struct net_device *dev)
7264 struct netdev_bpf xdp;
7267 /* Remove generic XDP */
7268 WARN_ON(dev_xdp_install(dev, generic_xdp_install, NULL, 0, NULL));
7270 /* Remove from the driver */
7271 ndo_bpf = dev->netdev_ops->ndo_bpf;
7275 __dev_xdp_query(dev, ndo_bpf, &xdp);
7276 if (xdp.prog_attached == XDP_ATTACHED_NONE)
7279 /* Program removal should always succeed */
7280 WARN_ON(dev_xdp_install(dev, ndo_bpf, NULL, xdp.prog_flags, NULL));
7284 * dev_change_xdp_fd - set or clear a bpf program for a device rx path
7286 * @extack: netlink extended ack
7287 * @fd: new program fd or negative value to clear
7288 * @flags: xdp-related flags
7290 * Set or clear a bpf program for a device
7292 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack,
7295 const struct net_device_ops *ops = dev->netdev_ops;
7296 struct bpf_prog *prog = NULL;
7297 bpf_op_t bpf_op, bpf_chk;
7302 bpf_op = bpf_chk = ops->ndo_bpf;
7303 if (!bpf_op && (flags & (XDP_FLAGS_DRV_MODE | XDP_FLAGS_HW_MODE)))
7305 if (!bpf_op || (flags & XDP_FLAGS_SKB_MODE))
7306 bpf_op = generic_xdp_install;
7307 if (bpf_op == bpf_chk)
7308 bpf_chk = generic_xdp_install;
7311 if (bpf_chk && __dev_xdp_attached(dev, bpf_chk))
7313 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) &&
7314 __dev_xdp_attached(dev, bpf_op))
7317 prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP,
7318 bpf_op == ops->ndo_bpf);
7320 return PTR_ERR(prog);
7322 if (!(flags & XDP_FLAGS_HW_MODE) &&
7323 bpf_prog_is_dev_bound(prog->aux)) {
7324 NL_SET_ERR_MSG(extack, "using device-bound program without HW_MODE flag is not supported");
7330 err = dev_xdp_install(dev, bpf_op, extack, flags, prog);
7331 if (err < 0 && prog)
7338 * dev_new_index - allocate an ifindex
7339 * @net: the applicable net namespace
7341 * Returns a suitable unique value for a new device interface
7342 * number. The caller must hold the rtnl semaphore or the
7343 * dev_base_lock to be sure it remains unique.
7345 static int dev_new_index(struct net *net)
7347 int ifindex = net->ifindex;
7352 if (!__dev_get_by_index(net, ifindex))
7353 return net->ifindex = ifindex;
7357 /* Delayed registration/unregisteration */
7358 static LIST_HEAD(net_todo_list);
7359 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq);
7361 static void net_set_todo(struct net_device *dev)
7363 list_add_tail(&dev->todo_list, &net_todo_list);
7364 dev_net(dev)->dev_unreg_count++;
7367 static void rollback_registered_many(struct list_head *head)
7369 struct net_device *dev, *tmp;
7370 LIST_HEAD(close_head);
7372 BUG_ON(dev_boot_phase);
7375 list_for_each_entry_safe(dev, tmp, head, unreg_list) {
7376 /* Some devices call without registering
7377 * for initialization unwind. Remove those
7378 * devices and proceed with the remaining.
7380 if (dev->reg_state == NETREG_UNINITIALIZED) {
7381 pr_debug("unregister_netdevice: device %s/%p never was registered\n",
7385 list_del(&dev->unreg_list);
7388 dev->dismantle = true;
7389 BUG_ON(dev->reg_state != NETREG_REGISTERED);
7392 /* If device is running, close it first. */
7393 list_for_each_entry(dev, head, unreg_list)
7394 list_add_tail(&dev->close_list, &close_head);
7395 dev_close_many(&close_head, true);
7397 list_for_each_entry(dev, head, unreg_list) {
7398 /* And unlink it from device chain. */
7399 unlist_netdevice(dev);
7401 dev->reg_state = NETREG_UNREGISTERING;
7403 flush_all_backlogs();
7407 list_for_each_entry(dev, head, unreg_list) {
7408 struct sk_buff *skb = NULL;
7410 /* Shutdown queueing discipline. */
7413 dev_xdp_uninstall(dev);
7415 /* Notify protocols, that we are about to destroy
7416 * this device. They should clean all the things.
7418 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
7420 if (!dev->rtnl_link_ops ||
7421 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
7422 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0,
7423 GFP_KERNEL, NULL, 0);
7426 * Flush the unicast and multicast chains
7431 if (dev->netdev_ops->ndo_uninit)
7432 dev->netdev_ops->ndo_uninit(dev);
7435 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL);
7437 /* Notifier chain MUST detach us all upper devices. */
7438 WARN_ON(netdev_has_any_upper_dev(dev));
7439 WARN_ON(netdev_has_any_lower_dev(dev));
7441 /* Remove entries from kobject tree */
7442 netdev_unregister_kobject(dev);
7444 /* Remove XPS queueing entries */
7445 netif_reset_xps_queues_gt(dev, 0);
7451 list_for_each_entry(dev, head, unreg_list)
7455 static void rollback_registered(struct net_device *dev)
7459 list_add(&dev->unreg_list, &single);
7460 rollback_registered_many(&single);
7464 static netdev_features_t netdev_sync_upper_features(struct net_device *lower,
7465 struct net_device *upper, netdev_features_t features)
7467 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
7468 netdev_features_t feature;
7471 for_each_netdev_feature(&upper_disables, feature_bit) {
7472 feature = __NETIF_F_BIT(feature_bit);
7473 if (!(upper->wanted_features & feature)
7474 && (features & feature)) {
7475 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n",
7476 &feature, upper->name);
7477 features &= ~feature;
7484 static void netdev_sync_lower_features(struct net_device *upper,
7485 struct net_device *lower, netdev_features_t features)
7487 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
7488 netdev_features_t feature;
7491 for_each_netdev_feature(&upper_disables, feature_bit) {
7492 feature = __NETIF_F_BIT(feature_bit);
7493 if (!(features & feature) && (lower->features & feature)) {
7494 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n",
7495 &feature, lower->name);
7496 lower->wanted_features &= ~feature;
7497 netdev_update_features(lower);
7499 if (unlikely(lower->features & feature))
7500 netdev_WARN(upper, "failed to disable %pNF on %s!\n",
7501 &feature, lower->name);
7506 static netdev_features_t netdev_fix_features(struct net_device *dev,
7507 netdev_features_t features)
7509 /* Fix illegal checksum combinations */
7510 if ((features & NETIF_F_HW_CSUM) &&
7511 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
7512 netdev_warn(dev, "mixed HW and IP checksum settings.\n");
7513 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
7516 /* TSO requires that SG is present as well. */
7517 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) {
7518 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n");
7519 features &= ~NETIF_F_ALL_TSO;
7522 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) &&
7523 !(features & NETIF_F_IP_CSUM)) {
7524 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n");
7525 features &= ~NETIF_F_TSO;
7526 features &= ~NETIF_F_TSO_ECN;
7529 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) &&
7530 !(features & NETIF_F_IPV6_CSUM)) {
7531 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n");
7532 features &= ~NETIF_F_TSO6;
7535 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */
7536 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO))
7537 features &= ~NETIF_F_TSO_MANGLEID;
7539 /* TSO ECN requires that TSO is present as well. */
7540 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN)
7541 features &= ~NETIF_F_TSO_ECN;
7543 /* Software GSO depends on SG. */
7544 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) {
7545 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n");
7546 features &= ~NETIF_F_GSO;
7549 /* GSO partial features require GSO partial be set */
7550 if ((features & dev->gso_partial_features) &&
7551 !(features & NETIF_F_GSO_PARTIAL)) {
7553 "Dropping partially supported GSO features since no GSO partial.\n");
7554 features &= ~dev->gso_partial_features;
7557 if (!(features & NETIF_F_RXCSUM)) {
7558 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet
7559 * successfully merged by hardware must also have the
7560 * checksum verified by hardware. If the user does not
7561 * want to enable RXCSUM, logically, we should disable GRO_HW.
7563 if (features & NETIF_F_GRO_HW) {
7564 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n");
7565 features &= ~NETIF_F_GRO_HW;
7569 /* LRO/HW-GRO features cannot be combined with RX-FCS */
7570 if (features & NETIF_F_RXFCS) {
7571 if (features & NETIF_F_LRO) {
7572 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n");
7573 features &= ~NETIF_F_LRO;
7576 if (features & NETIF_F_GRO_HW) {
7577 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n");
7578 features &= ~NETIF_F_GRO_HW;
7585 int __netdev_update_features(struct net_device *dev)
7587 struct net_device *upper, *lower;
7588 netdev_features_t features;
7589 struct list_head *iter;
7594 features = netdev_get_wanted_features(dev);
7596 if (dev->netdev_ops->ndo_fix_features)
7597 features = dev->netdev_ops->ndo_fix_features(dev, features);
7599 /* driver might be less strict about feature dependencies */
7600 features = netdev_fix_features(dev, features);
7602 /* some features can't be enabled if they're off an an upper device */
7603 netdev_for_each_upper_dev_rcu(dev, upper, iter)
7604 features = netdev_sync_upper_features(dev, upper, features);
7606 if (dev->features == features)
7609 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n",
7610 &dev->features, &features);
7612 if (dev->netdev_ops->ndo_set_features)
7613 err = dev->netdev_ops->ndo_set_features(dev, features);
7617 if (unlikely(err < 0)) {
7619 "set_features() failed (%d); wanted %pNF, left %pNF\n",
7620 err, &features, &dev->features);
7621 /* return non-0 since some features might have changed and
7622 * it's better to fire a spurious notification than miss it
7628 /* some features must be disabled on lower devices when disabled
7629 * on an upper device (think: bonding master or bridge)
7631 netdev_for_each_lower_dev(dev, lower, iter)
7632 netdev_sync_lower_features(dev, lower, features);
7635 netdev_features_t diff = features ^ dev->features;
7637 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) {
7638 /* udp_tunnel_{get,drop}_rx_info both need
7639 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the
7640 * device, or they won't do anything.
7641 * Thus we need to update dev->features
7642 * *before* calling udp_tunnel_get_rx_info,
7643 * but *after* calling udp_tunnel_drop_rx_info.
7645 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) {
7646 dev->features = features;
7647 udp_tunnel_get_rx_info(dev);
7649 udp_tunnel_drop_rx_info(dev);
7653 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) {
7654 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) {
7655 dev->features = features;
7656 err |= vlan_get_rx_ctag_filter_info(dev);
7658 vlan_drop_rx_ctag_filter_info(dev);
7662 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) {
7663 if (features & NETIF_F_HW_VLAN_STAG_FILTER) {
7664 dev->features = features;
7665 err |= vlan_get_rx_stag_filter_info(dev);
7667 vlan_drop_rx_stag_filter_info(dev);
7671 dev->features = features;
7674 return err < 0 ? 0 : 1;
7678 * netdev_update_features - recalculate device features
7679 * @dev: the device to check
7681 * Recalculate dev->features set and send notifications if it
7682 * has changed. Should be called after driver or hardware dependent
7683 * conditions might have changed that influence the features.
7685 void netdev_update_features(struct net_device *dev)
7687 if (__netdev_update_features(dev))
7688 netdev_features_change(dev);
7690 EXPORT_SYMBOL(netdev_update_features);
7693 * netdev_change_features - recalculate device features
7694 * @dev: the device to check
7696 * Recalculate dev->features set and send notifications even
7697 * if they have not changed. Should be called instead of
7698 * netdev_update_features() if also dev->vlan_features might
7699 * have changed to allow the changes to be propagated to stacked
7702 void netdev_change_features(struct net_device *dev)
7704 __netdev_update_features(dev);
7705 netdev_features_change(dev);
7707 EXPORT_SYMBOL(netdev_change_features);
7710 * netif_stacked_transfer_operstate - transfer operstate
7711 * @rootdev: the root or lower level device to transfer state from
7712 * @dev: the device to transfer operstate to
7714 * Transfer operational state from root to device. This is normally
7715 * called when a stacking relationship exists between the root
7716 * device and the device(a leaf device).
7718 void netif_stacked_transfer_operstate(const struct net_device *rootdev,
7719 struct net_device *dev)
7721 if (rootdev->operstate == IF_OPER_DORMANT)
7722 netif_dormant_on(dev);
7724 netif_dormant_off(dev);
7726 if (netif_carrier_ok(rootdev))
7727 netif_carrier_on(dev);
7729 netif_carrier_off(dev);
7731 EXPORT_SYMBOL(netif_stacked_transfer_operstate);
7733 static int netif_alloc_rx_queues(struct net_device *dev)
7735 unsigned int i, count = dev->num_rx_queues;
7736 struct netdev_rx_queue *rx;
7737 size_t sz = count * sizeof(*rx);
7742 rx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
7748 for (i = 0; i < count; i++) {
7751 /* XDP RX-queue setup */
7752 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i);
7759 /* Rollback successful reg's and free other resources */
7761 xdp_rxq_info_unreg(&rx[i].xdp_rxq);
7767 static void netif_free_rx_queues(struct net_device *dev)
7769 unsigned int i, count = dev->num_rx_queues;
7771 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */
7775 for (i = 0; i < count; i++)
7776 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq);
7781 static void netdev_init_one_queue(struct net_device *dev,
7782 struct netdev_queue *queue, void *_unused)
7784 /* Initialize queue lock */
7785 spin_lock_init(&queue->_xmit_lock);
7786 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type);
7787 queue->xmit_lock_owner = -1;
7788 netdev_queue_numa_node_write(queue, NUMA_NO_NODE);
7791 dql_init(&queue->dql, HZ);
7795 static void netif_free_tx_queues(struct net_device *dev)
7800 static int netif_alloc_netdev_queues(struct net_device *dev)
7802 unsigned int count = dev->num_tx_queues;
7803 struct netdev_queue *tx;
7804 size_t sz = count * sizeof(*tx);
7806 if (count < 1 || count > 0xffff)
7809 tx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
7815 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL);
7816 spin_lock_init(&dev->tx_global_lock);
7821 void netif_tx_stop_all_queues(struct net_device *dev)
7825 for (i = 0; i < dev->num_tx_queues; i++) {
7826 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
7828 netif_tx_stop_queue(txq);
7831 EXPORT_SYMBOL(netif_tx_stop_all_queues);
7834 * register_netdevice - register a network device
7835 * @dev: device to register
7837 * Take a completed network device structure and add it to the kernel
7838 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
7839 * chain. 0 is returned on success. A negative errno code is returned
7840 * on a failure to set up the device, or if the name is a duplicate.
7842 * Callers must hold the rtnl semaphore. You may want
7843 * register_netdev() instead of this.
7846 * The locking appears insufficient to guarantee two parallel registers
7847 * will not get the same name.
7850 int register_netdevice(struct net_device *dev)
7853 struct net *net = dev_net(dev);
7855 BUG_ON(dev_boot_phase);
7860 /* When net_device's are persistent, this will be fatal. */
7861 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
7864 spin_lock_init(&dev->addr_list_lock);
7865 netdev_set_addr_lockdep_class(dev);
7867 ret = dev_get_valid_name(net, dev, dev->name);
7871 /* Init, if this function is available */
7872 if (dev->netdev_ops->ndo_init) {
7873 ret = dev->netdev_ops->ndo_init(dev);
7881 if (((dev->hw_features | dev->features) &
7882 NETIF_F_HW_VLAN_CTAG_FILTER) &&
7883 (!dev->netdev_ops->ndo_vlan_rx_add_vid ||
7884 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) {
7885 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n");
7892 dev->ifindex = dev_new_index(net);
7893 else if (__dev_get_by_index(net, dev->ifindex))
7896 /* Transfer changeable features to wanted_features and enable
7897 * software offloads (GSO and GRO).
7899 dev->hw_features |= NETIF_F_SOFT_FEATURES;
7900 dev->features |= NETIF_F_SOFT_FEATURES;
7902 if (dev->netdev_ops->ndo_udp_tunnel_add) {
7903 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT;
7904 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT;
7907 dev->wanted_features = dev->features & dev->hw_features;
7909 if (!(dev->flags & IFF_LOOPBACK))
7910 dev->hw_features |= NETIF_F_NOCACHE_COPY;
7912 /* If IPv4 TCP segmentation offload is supported we should also
7913 * allow the device to enable segmenting the frame with the option
7914 * of ignoring a static IP ID value. This doesn't enable the
7915 * feature itself but allows the user to enable it later.
7917 if (dev->hw_features & NETIF_F_TSO)
7918 dev->hw_features |= NETIF_F_TSO_MANGLEID;
7919 if (dev->vlan_features & NETIF_F_TSO)
7920 dev->vlan_features |= NETIF_F_TSO_MANGLEID;
7921 if (dev->mpls_features & NETIF_F_TSO)
7922 dev->mpls_features |= NETIF_F_TSO_MANGLEID;
7923 if (dev->hw_enc_features & NETIF_F_TSO)
7924 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID;
7926 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices.
7928 dev->vlan_features |= NETIF_F_HIGHDMA;
7930 /* Make NETIF_F_SG inheritable to tunnel devices.
7932 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL;
7934 /* Make NETIF_F_SG inheritable to MPLS.
7936 dev->mpls_features |= NETIF_F_SG;
7938 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev);
7939 ret = notifier_to_errno(ret);
7943 ret = netdev_register_kobject(dev);
7946 dev->reg_state = NETREG_REGISTERED;
7948 __netdev_update_features(dev);
7951 * Default initial state at registry is that the
7952 * device is present.
7955 set_bit(__LINK_STATE_PRESENT, &dev->state);
7957 linkwatch_init_dev(dev);
7959 dev_init_scheduler(dev);
7961 list_netdevice(dev);
7962 add_device_randomness(dev->dev_addr, dev->addr_len);
7964 /* If the device has permanent device address, driver should
7965 * set dev_addr and also addr_assign_type should be set to
7966 * NET_ADDR_PERM (default value).
7968 if (dev->addr_assign_type == NET_ADDR_PERM)
7969 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len);
7971 /* Notify protocols, that a new device appeared. */
7972 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
7973 ret = notifier_to_errno(ret);
7975 rollback_registered(dev);
7976 dev->reg_state = NETREG_UNREGISTERED;
7979 * Prevent userspace races by waiting until the network
7980 * device is fully setup before sending notifications.
7982 if (!dev->rtnl_link_ops ||
7983 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
7984 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
7990 if (dev->netdev_ops->ndo_uninit)
7991 dev->netdev_ops->ndo_uninit(dev);
7992 if (dev->priv_destructor)
7993 dev->priv_destructor(dev);
7996 EXPORT_SYMBOL(register_netdevice);
7999 * init_dummy_netdev - init a dummy network device for NAPI
8000 * @dev: device to init
8002 * This takes a network device structure and initialize the minimum
8003 * amount of fields so it can be used to schedule NAPI polls without
8004 * registering a full blown interface. This is to be used by drivers
8005 * that need to tie several hardware interfaces to a single NAPI
8006 * poll scheduler due to HW limitations.
8008 int init_dummy_netdev(struct net_device *dev)
8010 /* Clear everything. Note we don't initialize spinlocks
8011 * are they aren't supposed to be taken by any of the
8012 * NAPI code and this dummy netdev is supposed to be
8013 * only ever used for NAPI polls
8015 memset(dev, 0, sizeof(struct net_device));
8017 /* make sure we BUG if trying to hit standard
8018 * register/unregister code path
8020 dev->reg_state = NETREG_DUMMY;
8022 /* NAPI wants this */
8023 INIT_LIST_HEAD(&dev->napi_list);
8025 /* a dummy interface is started by default */
8026 set_bit(__LINK_STATE_PRESENT, &dev->state);
8027 set_bit(__LINK_STATE_START, &dev->state);
8029 /* Note : We dont allocate pcpu_refcnt for dummy devices,
8030 * because users of this 'device' dont need to change
8036 EXPORT_SYMBOL_GPL(init_dummy_netdev);
8040 * register_netdev - register a network device
8041 * @dev: device to register
8043 * Take a completed network device structure and add it to the kernel
8044 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
8045 * chain. 0 is returned on success. A negative errno code is returned
8046 * on a failure to set up the device, or if the name is a duplicate.
8048 * This is a wrapper around register_netdevice that takes the rtnl semaphore
8049 * and expands the device name if you passed a format string to
8052 int register_netdev(struct net_device *dev)
8056 if (rtnl_lock_killable())
8058 err = register_netdevice(dev);
8062 EXPORT_SYMBOL(register_netdev);
8064 int netdev_refcnt_read(const struct net_device *dev)
8068 for_each_possible_cpu(i)
8069 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
8072 EXPORT_SYMBOL(netdev_refcnt_read);
8075 * netdev_wait_allrefs - wait until all references are gone.
8076 * @dev: target net_device
8078 * This is called when unregistering network devices.
8080 * Any protocol or device that holds a reference should register
8081 * for netdevice notification, and cleanup and put back the
8082 * reference if they receive an UNREGISTER event.
8083 * We can get stuck here if buggy protocols don't correctly
8086 static void netdev_wait_allrefs(struct net_device *dev)
8088 unsigned long rebroadcast_time, warning_time;
8091 linkwatch_forget_dev(dev);
8093 rebroadcast_time = warning_time = jiffies;
8094 refcnt = netdev_refcnt_read(dev);
8096 while (refcnt != 0) {
8097 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
8100 /* Rebroadcast unregister notification */
8101 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
8107 if (test_bit(__LINK_STATE_LINKWATCH_PENDING,
8109 /* We must not have linkwatch events
8110 * pending on unregister. If this
8111 * happens, we simply run the queue
8112 * unscheduled, resulting in a noop
8115 linkwatch_run_queue();
8120 rebroadcast_time = jiffies;
8125 refcnt = netdev_refcnt_read(dev);
8127 if (time_after(jiffies, warning_time + 10 * HZ)) {
8128 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n",
8130 warning_time = jiffies;
8139 * register_netdevice(x1);
8140 * register_netdevice(x2);
8142 * unregister_netdevice(y1);
8143 * unregister_netdevice(y2);
8149 * We are invoked by rtnl_unlock().
8150 * This allows us to deal with problems:
8151 * 1) We can delete sysfs objects which invoke hotplug
8152 * without deadlocking with linkwatch via keventd.
8153 * 2) Since we run with the RTNL semaphore not held, we can sleep
8154 * safely in order to wait for the netdev refcnt to drop to zero.
8156 * We must not return until all unregister events added during
8157 * the interval the lock was held have been completed.
8159 void netdev_run_todo(void)
8161 struct list_head list;
8163 /* Snapshot list, allow later requests */
8164 list_replace_init(&net_todo_list, &list);
8169 /* Wait for rcu callbacks to finish before next phase */
8170 if (!list_empty(&list))
8173 while (!list_empty(&list)) {
8174 struct net_device *dev
8175 = list_first_entry(&list, struct net_device, todo_list);
8176 list_del(&dev->todo_list);
8178 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) {
8179 pr_err("network todo '%s' but state %d\n",
8180 dev->name, dev->reg_state);
8185 dev->reg_state = NETREG_UNREGISTERED;
8187 netdev_wait_allrefs(dev);
8190 BUG_ON(netdev_refcnt_read(dev));
8191 BUG_ON(!list_empty(&dev->ptype_all));
8192 BUG_ON(!list_empty(&dev->ptype_specific));
8193 WARN_ON(rcu_access_pointer(dev->ip_ptr));
8194 WARN_ON(rcu_access_pointer(dev->ip6_ptr));
8195 #if IS_ENABLED(CONFIG_DECNET)
8196 WARN_ON(dev->dn_ptr);
8198 if (dev->priv_destructor)
8199 dev->priv_destructor(dev);
8200 if (dev->needs_free_netdev)
8203 /* Report a network device has been unregistered */
8205 dev_net(dev)->dev_unreg_count--;
8207 wake_up(&netdev_unregistering_wq);
8209 /* Free network device */
8210 kobject_put(&dev->dev.kobj);
8214 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has
8215 * all the same fields in the same order as net_device_stats, with only
8216 * the type differing, but rtnl_link_stats64 may have additional fields
8217 * at the end for newer counters.
8219 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
8220 const struct net_device_stats *netdev_stats)
8222 #if BITS_PER_LONG == 64
8223 BUILD_BUG_ON(sizeof(*stats64) < sizeof(*netdev_stats));
8224 memcpy(stats64, netdev_stats, sizeof(*netdev_stats));
8225 /* zero out counters that only exist in rtnl_link_stats64 */
8226 memset((char *)stats64 + sizeof(*netdev_stats), 0,
8227 sizeof(*stats64) - sizeof(*netdev_stats));
8229 size_t i, n = sizeof(*netdev_stats) / sizeof(unsigned long);
8230 const unsigned long *src = (const unsigned long *)netdev_stats;
8231 u64 *dst = (u64 *)stats64;
8233 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64));
8234 for (i = 0; i < n; i++)
8236 /* zero out counters that only exist in rtnl_link_stats64 */
8237 memset((char *)stats64 + n * sizeof(u64), 0,
8238 sizeof(*stats64) - n * sizeof(u64));
8241 EXPORT_SYMBOL(netdev_stats_to_stats64);
8244 * dev_get_stats - get network device statistics
8245 * @dev: device to get statistics from
8246 * @storage: place to store stats
8248 * Get network statistics from device. Return @storage.
8249 * The device driver may provide its own method by setting
8250 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats;
8251 * otherwise the internal statistics structure is used.
8253 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev,
8254 struct rtnl_link_stats64 *storage)
8256 const struct net_device_ops *ops = dev->netdev_ops;
8258 if (ops->ndo_get_stats64) {
8259 memset(storage, 0, sizeof(*storage));
8260 ops->ndo_get_stats64(dev, storage);
8261 } else if (ops->ndo_get_stats) {
8262 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev));
8264 netdev_stats_to_stats64(storage, &dev->stats);
8266 storage->rx_dropped += (unsigned long)atomic_long_read(&dev->rx_dropped);
8267 storage->tx_dropped += (unsigned long)atomic_long_read(&dev->tx_dropped);
8268 storage->rx_nohandler += (unsigned long)atomic_long_read(&dev->rx_nohandler);
8271 EXPORT_SYMBOL(dev_get_stats);
8273 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev)
8275 struct netdev_queue *queue = dev_ingress_queue(dev);
8277 #ifdef CONFIG_NET_CLS_ACT
8280 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
8283 netdev_init_one_queue(dev, queue, NULL);
8284 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc);
8285 queue->qdisc_sleeping = &noop_qdisc;
8286 rcu_assign_pointer(dev->ingress_queue, queue);
8291 static const struct ethtool_ops default_ethtool_ops;
8293 void netdev_set_default_ethtool_ops(struct net_device *dev,
8294 const struct ethtool_ops *ops)
8296 if (dev->ethtool_ops == &default_ethtool_ops)
8297 dev->ethtool_ops = ops;
8299 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops);
8301 void netdev_freemem(struct net_device *dev)
8303 char *addr = (char *)dev - dev->padded;
8309 * alloc_netdev_mqs - allocate network device
8310 * @sizeof_priv: size of private data to allocate space for
8311 * @name: device name format string
8312 * @name_assign_type: origin of device name
8313 * @setup: callback to initialize device
8314 * @txqs: the number of TX subqueues to allocate
8315 * @rxqs: the number of RX subqueues to allocate
8317 * Allocates a struct net_device with private data area for driver use
8318 * and performs basic initialization. Also allocates subqueue structs
8319 * for each queue on the device.
8321 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
8322 unsigned char name_assign_type,
8323 void (*setup)(struct net_device *),
8324 unsigned int txqs, unsigned int rxqs)
8326 struct net_device *dev;
8327 unsigned int alloc_size;
8328 struct net_device *p;
8330 BUG_ON(strlen(name) >= sizeof(dev->name));
8333 pr_err("alloc_netdev: Unable to allocate device with zero queues\n");
8338 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n");
8342 alloc_size = sizeof(struct net_device);
8344 /* ensure 32-byte alignment of private area */
8345 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
8346 alloc_size += sizeof_priv;
8348 /* ensure 32-byte alignment of whole construct */
8349 alloc_size += NETDEV_ALIGN - 1;
8351 p = kvzalloc(alloc_size, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
8355 dev = PTR_ALIGN(p, NETDEV_ALIGN);
8356 dev->padded = (char *)dev - (char *)p;
8358 dev->pcpu_refcnt = alloc_percpu(int);
8359 if (!dev->pcpu_refcnt)
8362 if (dev_addr_init(dev))
8368 dev_net_set(dev, &init_net);
8370 dev->gso_max_size = GSO_MAX_SIZE;
8371 dev->gso_max_segs = GSO_MAX_SEGS;
8373 INIT_LIST_HEAD(&dev->napi_list);
8374 INIT_LIST_HEAD(&dev->unreg_list);
8375 INIT_LIST_HEAD(&dev->close_list);
8376 INIT_LIST_HEAD(&dev->link_watch_list);
8377 INIT_LIST_HEAD(&dev->adj_list.upper);
8378 INIT_LIST_HEAD(&dev->adj_list.lower);
8379 INIT_LIST_HEAD(&dev->ptype_all);
8380 INIT_LIST_HEAD(&dev->ptype_specific);
8381 #ifdef CONFIG_NET_SCHED
8382 hash_init(dev->qdisc_hash);
8384 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM;
8387 if (!dev->tx_queue_len) {
8388 dev->priv_flags |= IFF_NO_QUEUE;
8389 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN;
8392 dev->num_tx_queues = txqs;
8393 dev->real_num_tx_queues = txqs;
8394 if (netif_alloc_netdev_queues(dev))
8397 dev->num_rx_queues = rxqs;
8398 dev->real_num_rx_queues = rxqs;
8399 if (netif_alloc_rx_queues(dev))
8402 strcpy(dev->name, name);
8403 dev->name_assign_type = name_assign_type;
8404 dev->group = INIT_NETDEV_GROUP;
8405 if (!dev->ethtool_ops)
8406 dev->ethtool_ops = &default_ethtool_ops;
8408 nf_hook_ingress_init(dev);
8417 free_percpu(dev->pcpu_refcnt);
8419 netdev_freemem(dev);
8422 EXPORT_SYMBOL(alloc_netdev_mqs);
8425 * free_netdev - free network device
8428 * This function does the last stage of destroying an allocated device
8429 * interface. The reference to the device object is released. If this
8430 * is the last reference then it will be freed.Must be called in process
8433 void free_netdev(struct net_device *dev)
8435 struct napi_struct *p, *n;
8438 netif_free_tx_queues(dev);
8439 netif_free_rx_queues(dev);
8441 kfree(rcu_dereference_protected(dev->ingress_queue, 1));
8443 /* Flush device addresses */
8444 dev_addr_flush(dev);
8446 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
8449 free_percpu(dev->pcpu_refcnt);
8450 dev->pcpu_refcnt = NULL;
8452 /* Compatibility with error handling in drivers */
8453 if (dev->reg_state == NETREG_UNINITIALIZED) {
8454 netdev_freemem(dev);
8458 BUG_ON(dev->reg_state != NETREG_UNREGISTERED);
8459 dev->reg_state = NETREG_RELEASED;
8461 /* will free via device release */
8462 put_device(&dev->dev);
8464 EXPORT_SYMBOL(free_netdev);
8467 * synchronize_net - Synchronize with packet receive processing
8469 * Wait for packets currently being received to be done.
8470 * Does not block later packets from starting.
8472 void synchronize_net(void)
8475 if (rtnl_is_locked())
8476 synchronize_rcu_expedited();
8480 EXPORT_SYMBOL(synchronize_net);
8483 * unregister_netdevice_queue - remove device from the kernel
8487 * This function shuts down a device interface and removes it
8488 * from the kernel tables.
8489 * If head not NULL, device is queued to be unregistered later.
8491 * Callers must hold the rtnl semaphore. You may want
8492 * unregister_netdev() instead of this.
8495 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head)
8500 list_move_tail(&dev->unreg_list, head);
8502 rollback_registered(dev);
8503 /* Finish processing unregister after unlock */
8507 EXPORT_SYMBOL(unregister_netdevice_queue);
8510 * unregister_netdevice_many - unregister many devices
8511 * @head: list of devices
8513 * Note: As most callers use a stack allocated list_head,
8514 * we force a list_del() to make sure stack wont be corrupted later.
8516 void unregister_netdevice_many(struct list_head *head)
8518 struct net_device *dev;
8520 if (!list_empty(head)) {
8521 rollback_registered_many(head);
8522 list_for_each_entry(dev, head, unreg_list)
8527 EXPORT_SYMBOL(unregister_netdevice_many);
8530 * unregister_netdev - remove device from the kernel
8533 * This function shuts down a device interface and removes it
8534 * from the kernel tables.
8536 * This is just a wrapper for unregister_netdevice that takes
8537 * the rtnl semaphore. In general you want to use this and not
8538 * unregister_netdevice.
8540 void unregister_netdev(struct net_device *dev)
8543 unregister_netdevice(dev);
8546 EXPORT_SYMBOL(unregister_netdev);
8549 * dev_change_net_namespace - move device to different nethost namespace
8551 * @net: network namespace
8552 * @pat: If not NULL name pattern to try if the current device name
8553 * is already taken in the destination network namespace.
8555 * This function shuts down a device interface and moves it
8556 * to a new network namespace. On success 0 is returned, on
8557 * a failure a netagive errno code is returned.
8559 * Callers must hold the rtnl semaphore.
8562 int dev_change_net_namespace(struct net_device *dev, struct net *net, const char *pat)
8564 int err, new_nsid, new_ifindex;
8568 /* Don't allow namespace local devices to be moved. */
8570 if (dev->features & NETIF_F_NETNS_LOCAL)
8573 /* Ensure the device has been registrered */
8574 if (dev->reg_state != NETREG_REGISTERED)
8577 /* Get out if there is nothing todo */
8579 if (net_eq(dev_net(dev), net))
8582 /* Pick the destination device name, and ensure
8583 * we can use it in the destination network namespace.
8586 if (__dev_get_by_name(net, dev->name)) {
8587 /* We get here if we can't use the current device name */
8590 if (dev_get_valid_name(net, dev, pat) < 0)
8595 * And now a mini version of register_netdevice unregister_netdevice.
8598 /* If device is running close it first. */
8601 /* And unlink it from device chain */
8603 unlist_netdevice(dev);
8607 /* Shutdown queueing discipline. */
8610 /* Notify protocols, that we are about to destroy
8611 * this device. They should clean all the things.
8613 * Note that dev->reg_state stays at NETREG_REGISTERED.
8614 * This is wanted because this way 8021q and macvlan know
8615 * the device is just moving and can keep their slaves up.
8617 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
8620 new_nsid = peernet2id_alloc(dev_net(dev), net);
8621 /* If there is an ifindex conflict assign a new one */
8622 if (__dev_get_by_index(net, dev->ifindex))
8623 new_ifindex = dev_new_index(net);
8625 new_ifindex = dev->ifindex;
8627 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid,
8631 * Flush the unicast and multicast chains
8636 /* Send a netdev-removed uevent to the old namespace */
8637 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE);
8638 netdev_adjacent_del_links(dev);
8640 /* Actually switch the network namespace */
8641 dev_net_set(dev, net);
8642 dev->ifindex = new_ifindex;
8644 /* Send a netdev-add uevent to the new namespace */
8645 kobject_uevent(&dev->dev.kobj, KOBJ_ADD);
8646 netdev_adjacent_add_links(dev);
8648 /* Fixup kobjects */
8649 err = device_rename(&dev->dev, dev->name);
8652 /* Add the device back in the hashes */
8653 list_netdevice(dev);
8655 /* Notify protocols, that a new device appeared. */
8656 call_netdevice_notifiers(NETDEV_REGISTER, dev);
8659 * Prevent userspace races by waiting until the network
8660 * device is fully setup before sending notifications.
8662 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
8669 EXPORT_SYMBOL_GPL(dev_change_net_namespace);
8671 static int dev_cpu_dead(unsigned int oldcpu)
8673 struct sk_buff **list_skb;
8674 struct sk_buff *skb;
8676 struct softnet_data *sd, *oldsd, *remsd = NULL;
8678 local_irq_disable();
8679 cpu = smp_processor_id();
8680 sd = &per_cpu(softnet_data, cpu);
8681 oldsd = &per_cpu(softnet_data, oldcpu);
8683 /* Find end of our completion_queue. */
8684 list_skb = &sd->completion_queue;
8686 list_skb = &(*list_skb)->next;
8687 /* Append completion queue from offline CPU. */
8688 *list_skb = oldsd->completion_queue;
8689 oldsd->completion_queue = NULL;
8691 /* Append output queue from offline CPU. */
8692 if (oldsd->output_queue) {
8693 *sd->output_queue_tailp = oldsd->output_queue;
8694 sd->output_queue_tailp = oldsd->output_queue_tailp;
8695 oldsd->output_queue = NULL;
8696 oldsd->output_queue_tailp = &oldsd->output_queue;
8698 /* Append NAPI poll list from offline CPU, with one exception :
8699 * process_backlog() must be called by cpu owning percpu backlog.
8700 * We properly handle process_queue & input_pkt_queue later.
8702 while (!list_empty(&oldsd->poll_list)) {
8703 struct napi_struct *napi = list_first_entry(&oldsd->poll_list,
8707 list_del_init(&napi->poll_list);
8708 if (napi->poll == process_backlog)
8711 ____napi_schedule(sd, napi);
8714 raise_softirq_irqoff(NET_TX_SOFTIRQ);
8718 remsd = oldsd->rps_ipi_list;
8719 oldsd->rps_ipi_list = NULL;
8721 /* send out pending IPI's on offline CPU */
8722 net_rps_send_ipi(remsd);
8724 /* Process offline CPU's input_pkt_queue */
8725 while ((skb = __skb_dequeue(&oldsd->process_queue))) {
8727 input_queue_head_incr(oldsd);
8729 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) {
8731 input_queue_head_incr(oldsd);
8738 * netdev_increment_features - increment feature set by one
8739 * @all: current feature set
8740 * @one: new feature set
8741 * @mask: mask feature set
8743 * Computes a new feature set after adding a device with feature set
8744 * @one to the master device with current feature set @all. Will not
8745 * enable anything that is off in @mask. Returns the new feature set.
8747 netdev_features_t netdev_increment_features(netdev_features_t all,
8748 netdev_features_t one, netdev_features_t mask)
8750 if (mask & NETIF_F_HW_CSUM)
8751 mask |= NETIF_F_CSUM_MASK;
8752 mask |= NETIF_F_VLAN_CHALLENGED;
8754 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask;
8755 all &= one | ~NETIF_F_ALL_FOR_ALL;
8757 /* If one device supports hw checksumming, set for all. */
8758 if (all & NETIF_F_HW_CSUM)
8759 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM);
8763 EXPORT_SYMBOL(netdev_increment_features);
8765 static struct hlist_head * __net_init netdev_create_hash(void)
8768 struct hlist_head *hash;
8770 hash = kmalloc(sizeof(*hash) * NETDEV_HASHENTRIES, GFP_KERNEL);
8772 for (i = 0; i < NETDEV_HASHENTRIES; i++)
8773 INIT_HLIST_HEAD(&hash[i]);
8778 /* Initialize per network namespace state */
8779 static int __net_init netdev_init(struct net *net)
8781 if (net != &init_net)
8782 INIT_LIST_HEAD(&net->dev_base_head);
8784 net->dev_name_head = netdev_create_hash();
8785 if (net->dev_name_head == NULL)
8788 net->dev_index_head = netdev_create_hash();
8789 if (net->dev_index_head == NULL)
8795 kfree(net->dev_name_head);
8801 * netdev_drivername - network driver for the device
8802 * @dev: network device
8804 * Determine network driver for device.
8806 const char *netdev_drivername(const struct net_device *dev)
8808 const struct device_driver *driver;
8809 const struct device *parent;
8810 const char *empty = "";
8812 parent = dev->dev.parent;
8816 driver = parent->driver;
8817 if (driver && driver->name)
8818 return driver->name;
8822 static void __netdev_printk(const char *level, const struct net_device *dev,
8823 struct va_format *vaf)
8825 if (dev && dev->dev.parent) {
8826 dev_printk_emit(level[1] - '0',
8829 dev_driver_string(dev->dev.parent),
8830 dev_name(dev->dev.parent),
8831 netdev_name(dev), netdev_reg_state(dev),
8834 printk("%s%s%s: %pV",
8835 level, netdev_name(dev), netdev_reg_state(dev), vaf);
8837 printk("%s(NULL net_device): %pV", level, vaf);
8841 void netdev_printk(const char *level, const struct net_device *dev,
8842 const char *format, ...)
8844 struct va_format vaf;
8847 va_start(args, format);
8852 __netdev_printk(level, dev, &vaf);
8856 EXPORT_SYMBOL(netdev_printk);
8858 #define define_netdev_printk_level(func, level) \
8859 void func(const struct net_device *dev, const char *fmt, ...) \
8861 struct va_format vaf; \
8864 va_start(args, fmt); \
8869 __netdev_printk(level, dev, &vaf); \
8873 EXPORT_SYMBOL(func);
8875 define_netdev_printk_level(netdev_emerg, KERN_EMERG);
8876 define_netdev_printk_level(netdev_alert, KERN_ALERT);
8877 define_netdev_printk_level(netdev_crit, KERN_CRIT);
8878 define_netdev_printk_level(netdev_err, KERN_ERR);
8879 define_netdev_printk_level(netdev_warn, KERN_WARNING);
8880 define_netdev_printk_level(netdev_notice, KERN_NOTICE);
8881 define_netdev_printk_level(netdev_info, KERN_INFO);
8883 static void __net_exit netdev_exit(struct net *net)
8885 kfree(net->dev_name_head);
8886 kfree(net->dev_index_head);
8887 if (net != &init_net)
8888 WARN_ON_ONCE(!list_empty(&net->dev_base_head));
8891 static struct pernet_operations __net_initdata netdev_net_ops = {
8892 .init = netdev_init,
8893 .exit = netdev_exit,
8896 static void __net_exit default_device_exit(struct net *net)
8898 struct net_device *dev, *aux;
8900 * Push all migratable network devices back to the
8901 * initial network namespace
8904 for_each_netdev_safe(net, dev, aux) {
8906 char fb_name[IFNAMSIZ];
8908 /* Ignore unmoveable devices (i.e. loopback) */
8909 if (dev->features & NETIF_F_NETNS_LOCAL)
8912 /* Leave virtual devices for the generic cleanup */
8913 if (dev->rtnl_link_ops)
8916 /* Push remaining network devices to init_net */
8917 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex);
8918 err = dev_change_net_namespace(dev, &init_net, fb_name);
8920 pr_emerg("%s: failed to move %s to init_net: %d\n",
8921 __func__, dev->name, err);
8928 static void __net_exit rtnl_lock_unregistering(struct list_head *net_list)
8930 /* Return with the rtnl_lock held when there are no network
8931 * devices unregistering in any network namespace in net_list.
8935 DEFINE_WAIT_FUNC(wait, woken_wake_function);
8937 add_wait_queue(&netdev_unregistering_wq, &wait);
8939 unregistering = false;
8941 list_for_each_entry(net, net_list, exit_list) {
8942 if (net->dev_unreg_count > 0) {
8943 unregistering = true;
8951 wait_woken(&wait, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
8953 remove_wait_queue(&netdev_unregistering_wq, &wait);
8956 static void __net_exit default_device_exit_batch(struct list_head *net_list)
8958 /* At exit all network devices most be removed from a network
8959 * namespace. Do this in the reverse order of registration.
8960 * Do this across as many network namespaces as possible to
8961 * improve batching efficiency.
8963 struct net_device *dev;
8965 LIST_HEAD(dev_kill_list);
8967 /* To prevent network device cleanup code from dereferencing
8968 * loopback devices or network devices that have been freed
8969 * wait here for all pending unregistrations to complete,
8970 * before unregistring the loopback device and allowing the
8971 * network namespace be freed.
8973 * The netdev todo list containing all network devices
8974 * unregistrations that happen in default_device_exit_batch
8975 * will run in the rtnl_unlock() at the end of
8976 * default_device_exit_batch.
8978 rtnl_lock_unregistering(net_list);
8979 list_for_each_entry(net, net_list, exit_list) {
8980 for_each_netdev_reverse(net, dev) {
8981 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink)
8982 dev->rtnl_link_ops->dellink(dev, &dev_kill_list);
8984 unregister_netdevice_queue(dev, &dev_kill_list);
8987 unregister_netdevice_many(&dev_kill_list);
8991 static struct pernet_operations __net_initdata default_device_ops = {
8992 .exit = default_device_exit,
8993 .exit_batch = default_device_exit_batch,
8997 * Initialize the DEV module. At boot time this walks the device list and
8998 * unhooks any devices that fail to initialise (normally hardware not
8999 * present) and leaves us with a valid list of present and active devices.
9004 * This is called single threaded during boot, so no need
9005 * to take the rtnl semaphore.
9007 static int __init net_dev_init(void)
9009 int i, rc = -ENOMEM;
9011 BUG_ON(!dev_boot_phase);
9013 if (dev_proc_init())
9016 if (netdev_kobject_init())
9019 INIT_LIST_HEAD(&ptype_all);
9020 for (i = 0; i < PTYPE_HASH_SIZE; i++)
9021 INIT_LIST_HEAD(&ptype_base[i]);
9023 INIT_LIST_HEAD(&offload_base);
9025 if (register_pernet_subsys(&netdev_net_ops))
9029 * Initialise the packet receive queues.
9032 for_each_possible_cpu(i) {
9033 struct work_struct *flush = per_cpu_ptr(&flush_works, i);
9034 struct softnet_data *sd = &per_cpu(softnet_data, i);
9036 INIT_WORK(flush, flush_backlog);
9038 skb_queue_head_init(&sd->input_pkt_queue);
9039 skb_queue_head_init(&sd->process_queue);
9040 #ifdef CONFIG_XFRM_OFFLOAD
9041 skb_queue_head_init(&sd->xfrm_backlog);
9043 INIT_LIST_HEAD(&sd->poll_list);
9044 sd->output_queue_tailp = &sd->output_queue;
9046 sd->csd.func = rps_trigger_softirq;
9051 sd->backlog.poll = process_backlog;
9052 sd->backlog.weight = weight_p;
9057 /* The loopback device is special if any other network devices
9058 * is present in a network namespace the loopback device must
9059 * be present. Since we now dynamically allocate and free the
9060 * loopback device ensure this invariant is maintained by
9061 * keeping the loopback device as the first device on the
9062 * list of network devices. Ensuring the loopback devices
9063 * is the first device that appears and the last network device
9066 if (register_pernet_device(&loopback_net_ops))
9069 if (register_pernet_device(&default_device_ops))
9072 open_softirq(NET_TX_SOFTIRQ, net_tx_action);
9073 open_softirq(NET_RX_SOFTIRQ, net_rx_action);
9075 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead",
9076 NULL, dev_cpu_dead);
9083 subsys_initcall(net_dev_init);