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 (strlen(name) >= 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 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val,
1575 struct net_device *dev)
1577 struct netdev_notifier_info info = {
1581 return nb->notifier_call(nb, val, &info);
1584 static int dev_boot_phase = 1;
1587 * register_netdevice_notifier - register a network notifier block
1590 * Register a notifier to be called when network device events occur.
1591 * The notifier passed is linked into the kernel structures and must
1592 * not be reused until it has been unregistered. A negative errno code
1593 * is returned on a failure.
1595 * When registered all registration and up events are replayed
1596 * to the new notifier to allow device to have a race free
1597 * view of the network device list.
1600 int register_netdevice_notifier(struct notifier_block *nb)
1602 struct net_device *dev;
1603 struct net_device *last;
1608 err = raw_notifier_chain_register(&netdev_chain, nb);
1614 for_each_netdev(net, dev) {
1615 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev);
1616 err = notifier_to_errno(err);
1620 if (!(dev->flags & IFF_UP))
1623 call_netdevice_notifier(nb, NETDEV_UP, dev);
1634 for_each_netdev(net, dev) {
1638 if (dev->flags & IFF_UP) {
1639 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1641 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1643 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1648 raw_notifier_chain_unregister(&netdev_chain, nb);
1651 EXPORT_SYMBOL(register_netdevice_notifier);
1654 * unregister_netdevice_notifier - unregister a network notifier block
1657 * Unregister a notifier previously registered by
1658 * register_netdevice_notifier(). The notifier is unlinked into the
1659 * kernel structures and may then be reused. A negative errno code
1660 * is returned on a failure.
1662 * After unregistering unregister and down device events are synthesized
1663 * for all devices on the device list to the removed notifier to remove
1664 * the need for special case cleanup code.
1667 int unregister_netdevice_notifier(struct notifier_block *nb)
1669 struct net_device *dev;
1674 err = raw_notifier_chain_unregister(&netdev_chain, nb);
1679 for_each_netdev(net, dev) {
1680 if (dev->flags & IFF_UP) {
1681 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1683 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1685 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1692 EXPORT_SYMBOL(unregister_netdevice_notifier);
1695 * call_netdevice_notifiers_info - call all network notifier blocks
1696 * @val: value passed unmodified to notifier function
1697 * @info: notifier information data
1699 * Call all network notifier blocks. Parameters and return value
1700 * are as for raw_notifier_call_chain().
1703 static int call_netdevice_notifiers_info(unsigned long val,
1704 struct netdev_notifier_info *info)
1707 return raw_notifier_call_chain(&netdev_chain, val, info);
1711 * call_netdevice_notifiers - call all network notifier blocks
1712 * @val: value passed unmodified to notifier function
1713 * @dev: net_device pointer passed unmodified to notifier function
1715 * Call all network notifier blocks. Parameters and return value
1716 * are as for raw_notifier_call_chain().
1719 int call_netdevice_notifiers(unsigned long val, struct net_device *dev)
1721 struct netdev_notifier_info info = {
1725 return call_netdevice_notifiers_info(val, &info);
1727 EXPORT_SYMBOL(call_netdevice_notifiers);
1729 #ifdef CONFIG_NET_INGRESS
1730 static struct static_key ingress_needed __read_mostly;
1732 void net_inc_ingress_queue(void)
1734 static_key_slow_inc(&ingress_needed);
1736 EXPORT_SYMBOL_GPL(net_inc_ingress_queue);
1738 void net_dec_ingress_queue(void)
1740 static_key_slow_dec(&ingress_needed);
1742 EXPORT_SYMBOL_GPL(net_dec_ingress_queue);
1745 #ifdef CONFIG_NET_EGRESS
1746 static struct static_key egress_needed __read_mostly;
1748 void net_inc_egress_queue(void)
1750 static_key_slow_inc(&egress_needed);
1752 EXPORT_SYMBOL_GPL(net_inc_egress_queue);
1754 void net_dec_egress_queue(void)
1756 static_key_slow_dec(&egress_needed);
1758 EXPORT_SYMBOL_GPL(net_dec_egress_queue);
1761 static struct static_key netstamp_needed __read_mostly;
1762 #ifdef HAVE_JUMP_LABEL
1763 static atomic_t netstamp_needed_deferred;
1764 static atomic_t netstamp_wanted;
1765 static void netstamp_clear(struct work_struct *work)
1767 int deferred = atomic_xchg(&netstamp_needed_deferred, 0);
1770 wanted = atomic_add_return(deferred, &netstamp_wanted);
1772 static_key_enable(&netstamp_needed);
1774 static_key_disable(&netstamp_needed);
1776 static DECLARE_WORK(netstamp_work, netstamp_clear);
1779 void net_enable_timestamp(void)
1781 #ifdef HAVE_JUMP_LABEL
1785 wanted = atomic_read(&netstamp_wanted);
1788 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted + 1) == wanted)
1791 atomic_inc(&netstamp_needed_deferred);
1792 schedule_work(&netstamp_work);
1794 static_key_slow_inc(&netstamp_needed);
1797 EXPORT_SYMBOL(net_enable_timestamp);
1799 void net_disable_timestamp(void)
1801 #ifdef HAVE_JUMP_LABEL
1805 wanted = atomic_read(&netstamp_wanted);
1808 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted - 1) == wanted)
1811 atomic_dec(&netstamp_needed_deferred);
1812 schedule_work(&netstamp_work);
1814 static_key_slow_dec(&netstamp_needed);
1817 EXPORT_SYMBOL(net_disable_timestamp);
1819 static inline void net_timestamp_set(struct sk_buff *skb)
1822 if (static_key_false(&netstamp_needed))
1823 __net_timestamp(skb);
1826 #define net_timestamp_check(COND, SKB) \
1827 if (static_key_false(&netstamp_needed)) { \
1828 if ((COND) && !(SKB)->tstamp) \
1829 __net_timestamp(SKB); \
1832 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb)
1836 if (!(dev->flags & IFF_UP))
1839 len = dev->mtu + dev->hard_header_len + VLAN_HLEN;
1840 if (skb->len <= len)
1843 /* if TSO is enabled, we don't care about the length as the packet
1844 * could be forwarded without being segmented before
1846 if (skb_is_gso(skb))
1851 EXPORT_SYMBOL_GPL(is_skb_forwardable);
1853 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1855 int ret = ____dev_forward_skb(dev, skb);
1858 skb->protocol = eth_type_trans(skb, dev);
1859 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
1864 EXPORT_SYMBOL_GPL(__dev_forward_skb);
1867 * dev_forward_skb - loopback an skb to another netif
1869 * @dev: destination network device
1870 * @skb: buffer to forward
1873 * NET_RX_SUCCESS (no congestion)
1874 * NET_RX_DROP (packet was dropped, but freed)
1876 * dev_forward_skb can be used for injecting an skb from the
1877 * start_xmit function of one device into the receive queue
1878 * of another device.
1880 * The receiving device may be in another namespace, so
1881 * we have to clear all information in the skb that could
1882 * impact namespace isolation.
1884 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1886 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb);
1888 EXPORT_SYMBOL_GPL(dev_forward_skb);
1890 static inline int deliver_skb(struct sk_buff *skb,
1891 struct packet_type *pt_prev,
1892 struct net_device *orig_dev)
1894 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
1896 refcount_inc(&skb->users);
1897 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
1900 static inline void deliver_ptype_list_skb(struct sk_buff *skb,
1901 struct packet_type **pt,
1902 struct net_device *orig_dev,
1904 struct list_head *ptype_list)
1906 struct packet_type *ptype, *pt_prev = *pt;
1908 list_for_each_entry_rcu(ptype, ptype_list, list) {
1909 if (ptype->type != type)
1912 deliver_skb(skb, pt_prev, orig_dev);
1918 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb)
1920 if (!ptype->af_packet_priv || !skb->sk)
1923 if (ptype->id_match)
1924 return ptype->id_match(ptype, skb->sk);
1925 else if ((struct sock *)ptype->af_packet_priv == skb->sk)
1932 * Support routine. Sends outgoing frames to any network
1933 * taps currently in use.
1936 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
1938 struct packet_type *ptype;
1939 struct sk_buff *skb2 = NULL;
1940 struct packet_type *pt_prev = NULL;
1941 struct list_head *ptype_list = &ptype_all;
1945 list_for_each_entry_rcu(ptype, ptype_list, list) {
1946 /* Never send packets back to the socket
1947 * they originated from - MvS (miquels@drinkel.ow.org)
1949 if (skb_loop_sk(ptype, skb))
1953 deliver_skb(skb2, pt_prev, skb->dev);
1958 /* need to clone skb, done only once */
1959 skb2 = skb_clone(skb, GFP_ATOMIC);
1963 net_timestamp_set(skb2);
1965 /* skb->nh should be correctly
1966 * set by sender, so that the second statement is
1967 * just protection against buggy protocols.
1969 skb_reset_mac_header(skb2);
1971 if (skb_network_header(skb2) < skb2->data ||
1972 skb_network_header(skb2) > skb_tail_pointer(skb2)) {
1973 net_crit_ratelimited("protocol %04x is buggy, dev %s\n",
1974 ntohs(skb2->protocol),
1976 skb_reset_network_header(skb2);
1979 skb2->transport_header = skb2->network_header;
1980 skb2->pkt_type = PACKET_OUTGOING;
1984 if (ptype_list == &ptype_all) {
1985 ptype_list = &dev->ptype_all;
1990 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC))
1991 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
1997 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit);
2000 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change
2001 * @dev: Network device
2002 * @txq: number of queues available
2004 * If real_num_tx_queues is changed the tc mappings may no longer be
2005 * valid. To resolve this verify the tc mapping remains valid and if
2006 * not NULL the mapping. With no priorities mapping to this
2007 * offset/count pair it will no longer be used. In the worst case TC0
2008 * is invalid nothing can be done so disable priority mappings. If is
2009 * expected that drivers will fix this mapping if they can before
2010 * calling netif_set_real_num_tx_queues.
2012 static void netif_setup_tc(struct net_device *dev, unsigned int txq)
2015 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2017 /* If TC0 is invalidated disable TC mapping */
2018 if (tc->offset + tc->count > txq) {
2019 pr_warn("Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n");
2024 /* Invalidated prio to tc mappings set to TC0 */
2025 for (i = 1; i < TC_BITMASK + 1; i++) {
2026 int q = netdev_get_prio_tc_map(dev, i);
2028 tc = &dev->tc_to_txq[q];
2029 if (tc->offset + tc->count > txq) {
2030 pr_warn("Number of in use tx queues changed. Priority %i to tc mapping %i is no longer valid. Setting map to 0\n",
2032 netdev_set_prio_tc_map(dev, i, 0);
2037 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq)
2040 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2043 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) {
2044 if ((txq - tc->offset) < tc->count)
2053 EXPORT_SYMBOL(netdev_txq_to_tc);
2056 static DEFINE_MUTEX(xps_map_mutex);
2057 #define xmap_dereference(P) \
2058 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex))
2060 static bool remove_xps_queue(struct xps_dev_maps *dev_maps,
2063 struct xps_map *map = NULL;
2067 map = xmap_dereference(dev_maps->cpu_map[tci]);
2071 for (pos = map->len; pos--;) {
2072 if (map->queues[pos] != index)
2076 map->queues[pos] = map->queues[--map->len];
2080 RCU_INIT_POINTER(dev_maps->cpu_map[tci], NULL);
2081 kfree_rcu(map, rcu);
2088 static bool remove_xps_queue_cpu(struct net_device *dev,
2089 struct xps_dev_maps *dev_maps,
2090 int cpu, u16 offset, u16 count)
2092 int num_tc = dev->num_tc ? : 1;
2093 bool active = false;
2096 for (tci = cpu * num_tc; num_tc--; tci++) {
2099 for (i = count, j = offset; i--; j++) {
2100 if (!remove_xps_queue(dev_maps, cpu, j))
2110 static void netif_reset_xps_queues(struct net_device *dev, u16 offset,
2113 struct xps_dev_maps *dev_maps;
2115 bool active = false;
2117 mutex_lock(&xps_map_mutex);
2118 dev_maps = xmap_dereference(dev->xps_maps);
2123 for_each_possible_cpu(cpu)
2124 active |= remove_xps_queue_cpu(dev, dev_maps, cpu,
2128 RCU_INIT_POINTER(dev->xps_maps, NULL);
2129 kfree_rcu(dev_maps, rcu);
2132 for (i = offset + (count - 1); count--; i--)
2133 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, i),
2137 mutex_unlock(&xps_map_mutex);
2140 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index)
2142 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index);
2145 static struct xps_map *expand_xps_map(struct xps_map *map,
2148 struct xps_map *new_map;
2149 int alloc_len = XPS_MIN_MAP_ALLOC;
2152 for (pos = 0; map && pos < map->len; pos++) {
2153 if (map->queues[pos] != index)
2158 /* Need to add queue to this CPU's existing map */
2160 if (pos < map->alloc_len)
2163 alloc_len = map->alloc_len * 2;
2166 /* Need to allocate new map to store queue on this CPU's map */
2167 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL,
2172 for (i = 0; i < pos; i++)
2173 new_map->queues[i] = map->queues[i];
2174 new_map->alloc_len = alloc_len;
2180 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask,
2183 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL;
2184 int i, cpu, tci, numa_node_id = -2;
2185 int maps_sz, num_tc = 1, tc = 0;
2186 struct xps_map *map, *new_map;
2187 bool active = false;
2190 num_tc = dev->num_tc;
2191 tc = netdev_txq_to_tc(dev, index);
2196 maps_sz = XPS_DEV_MAPS_SIZE(num_tc);
2197 if (maps_sz < L1_CACHE_BYTES)
2198 maps_sz = L1_CACHE_BYTES;
2200 mutex_lock(&xps_map_mutex);
2202 dev_maps = xmap_dereference(dev->xps_maps);
2204 /* allocate memory for queue storage */
2205 for_each_cpu_and(cpu, cpu_online_mask, mask) {
2207 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL);
2208 if (!new_dev_maps) {
2209 mutex_unlock(&xps_map_mutex);
2213 tci = cpu * num_tc + tc;
2214 map = dev_maps ? xmap_dereference(dev_maps->cpu_map[tci]) :
2217 map = expand_xps_map(map, cpu, index);
2221 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2225 goto out_no_new_maps;
2227 for_each_possible_cpu(cpu) {
2228 /* copy maps belonging to foreign traffic classes */
2229 for (i = tc, tci = cpu * num_tc; dev_maps && i--; tci++) {
2230 /* fill in the new device map from the old device map */
2231 map = xmap_dereference(dev_maps->cpu_map[tci]);
2232 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2235 /* We need to explicitly update tci as prevous loop
2236 * could break out early if dev_maps is NULL.
2238 tci = cpu * num_tc + tc;
2240 if (cpumask_test_cpu(cpu, mask) && cpu_online(cpu)) {
2241 /* add queue to CPU maps */
2244 map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2245 while ((pos < map->len) && (map->queues[pos] != index))
2248 if (pos == map->len)
2249 map->queues[map->len++] = index;
2251 if (numa_node_id == -2)
2252 numa_node_id = cpu_to_node(cpu);
2253 else if (numa_node_id != cpu_to_node(cpu))
2256 } else if (dev_maps) {
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 /* copy maps belonging to foreign traffic classes */
2263 for (i = num_tc - tc, tci++; dev_maps && --i; tci++) {
2264 /* fill in the new device map from the old device map */
2265 map = xmap_dereference(dev_maps->cpu_map[tci]);
2266 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2270 rcu_assign_pointer(dev->xps_maps, new_dev_maps);
2272 /* Cleanup old maps */
2274 goto out_no_old_maps;
2276 for_each_possible_cpu(cpu) {
2277 for (i = num_tc, tci = cpu * num_tc; i--; tci++) {
2278 new_map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2279 map = xmap_dereference(dev_maps->cpu_map[tci]);
2280 if (map && map != new_map)
2281 kfree_rcu(map, rcu);
2285 kfree_rcu(dev_maps, rcu);
2288 dev_maps = new_dev_maps;
2292 /* update Tx queue numa node */
2293 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index),
2294 (numa_node_id >= 0) ? numa_node_id :
2300 /* removes queue from unused CPUs */
2301 for_each_possible_cpu(cpu) {
2302 for (i = tc, tci = cpu * num_tc; i--; tci++)
2303 active |= remove_xps_queue(dev_maps, tci, index);
2304 if (!cpumask_test_cpu(cpu, mask) || !cpu_online(cpu))
2305 active |= remove_xps_queue(dev_maps, tci, index);
2306 for (i = num_tc - tc, tci++; --i; tci++)
2307 active |= remove_xps_queue(dev_maps, tci, index);
2310 /* free map if not active */
2312 RCU_INIT_POINTER(dev->xps_maps, NULL);
2313 kfree_rcu(dev_maps, rcu);
2317 mutex_unlock(&xps_map_mutex);
2321 /* remove any maps that we added */
2322 for_each_possible_cpu(cpu) {
2323 for (i = num_tc, tci = cpu * num_tc; i--; tci++) {
2324 new_map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2326 xmap_dereference(dev_maps->cpu_map[tci]) :
2328 if (new_map && new_map != map)
2333 mutex_unlock(&xps_map_mutex);
2335 kfree(new_dev_maps);
2338 EXPORT_SYMBOL(netif_set_xps_queue);
2341 void netdev_reset_tc(struct net_device *dev)
2344 netif_reset_xps_queues_gt(dev, 0);
2347 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq));
2348 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map));
2350 EXPORT_SYMBOL(netdev_reset_tc);
2352 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset)
2354 if (tc >= dev->num_tc)
2358 netif_reset_xps_queues(dev, offset, count);
2360 dev->tc_to_txq[tc].count = count;
2361 dev->tc_to_txq[tc].offset = offset;
2364 EXPORT_SYMBOL(netdev_set_tc_queue);
2366 int netdev_set_num_tc(struct net_device *dev, u8 num_tc)
2368 if (num_tc > TC_MAX_QUEUE)
2372 netif_reset_xps_queues_gt(dev, 0);
2374 dev->num_tc = num_tc;
2377 EXPORT_SYMBOL(netdev_set_num_tc);
2380 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues
2381 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed.
2383 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq)
2388 disabling = txq < dev->real_num_tx_queues;
2390 if (txq < 1 || txq > dev->num_tx_queues)
2393 if (dev->reg_state == NETREG_REGISTERED ||
2394 dev->reg_state == NETREG_UNREGISTERING) {
2397 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues,
2403 netif_setup_tc(dev, txq);
2405 dev->real_num_tx_queues = txq;
2409 qdisc_reset_all_tx_gt(dev, txq);
2411 netif_reset_xps_queues_gt(dev, txq);
2415 dev->real_num_tx_queues = txq;
2420 EXPORT_SYMBOL(netif_set_real_num_tx_queues);
2424 * netif_set_real_num_rx_queues - set actual number of RX queues used
2425 * @dev: Network device
2426 * @rxq: Actual number of RX queues
2428 * This must be called either with the rtnl_lock held or before
2429 * registration of the net device. Returns 0 on success, or a
2430 * negative error code. If called before registration, it always
2433 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq)
2437 if (rxq < 1 || rxq > dev->num_rx_queues)
2440 if (dev->reg_state == NETREG_REGISTERED) {
2443 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues,
2449 dev->real_num_rx_queues = rxq;
2452 EXPORT_SYMBOL(netif_set_real_num_rx_queues);
2456 * netif_get_num_default_rss_queues - default number of RSS queues
2458 * This routine should set an upper limit on the number of RSS queues
2459 * used by default by multiqueue devices.
2461 int netif_get_num_default_rss_queues(void)
2463 return is_kdump_kernel() ?
2464 1 : min_t(int, DEFAULT_MAX_NUM_RSS_QUEUES, num_online_cpus());
2466 EXPORT_SYMBOL(netif_get_num_default_rss_queues);
2468 static void __netif_reschedule(struct Qdisc *q)
2470 struct softnet_data *sd;
2471 unsigned long flags;
2473 local_irq_save(flags);
2474 sd = this_cpu_ptr(&softnet_data);
2475 q->next_sched = NULL;
2476 *sd->output_queue_tailp = q;
2477 sd->output_queue_tailp = &q->next_sched;
2478 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2479 local_irq_restore(flags);
2482 void __netif_schedule(struct Qdisc *q)
2484 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
2485 __netif_reschedule(q);
2487 EXPORT_SYMBOL(__netif_schedule);
2489 struct dev_kfree_skb_cb {
2490 enum skb_free_reason reason;
2493 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb)
2495 return (struct dev_kfree_skb_cb *)skb->cb;
2498 void netif_schedule_queue(struct netdev_queue *txq)
2501 if (!(txq->state & QUEUE_STATE_ANY_XOFF)) {
2502 struct Qdisc *q = rcu_dereference(txq->qdisc);
2504 __netif_schedule(q);
2508 EXPORT_SYMBOL(netif_schedule_queue);
2510 void netif_tx_wake_queue(struct netdev_queue *dev_queue)
2512 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) {
2516 q = rcu_dereference(dev_queue->qdisc);
2517 __netif_schedule(q);
2521 EXPORT_SYMBOL(netif_tx_wake_queue);
2523 void __dev_kfree_skb_irq(struct sk_buff *skb, enum skb_free_reason reason)
2525 unsigned long flags;
2530 if (likely(refcount_read(&skb->users) == 1)) {
2532 refcount_set(&skb->users, 0);
2533 } else if (likely(!refcount_dec_and_test(&skb->users))) {
2536 get_kfree_skb_cb(skb)->reason = reason;
2537 local_irq_save(flags);
2538 skb->next = __this_cpu_read(softnet_data.completion_queue);
2539 __this_cpu_write(softnet_data.completion_queue, skb);
2540 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2541 local_irq_restore(flags);
2543 EXPORT_SYMBOL(__dev_kfree_skb_irq);
2545 void __dev_kfree_skb_any(struct sk_buff *skb, enum skb_free_reason reason)
2547 if (in_irq() || irqs_disabled())
2548 __dev_kfree_skb_irq(skb, reason);
2552 EXPORT_SYMBOL(__dev_kfree_skb_any);
2556 * netif_device_detach - mark device as removed
2557 * @dev: network device
2559 * Mark device as removed from system and therefore no longer available.
2561 void netif_device_detach(struct net_device *dev)
2563 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) &&
2564 netif_running(dev)) {
2565 netif_tx_stop_all_queues(dev);
2568 EXPORT_SYMBOL(netif_device_detach);
2571 * netif_device_attach - mark device as attached
2572 * @dev: network device
2574 * Mark device as attached from system and restart if needed.
2576 void netif_device_attach(struct net_device *dev)
2578 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) &&
2579 netif_running(dev)) {
2580 netif_tx_wake_all_queues(dev);
2581 __netdev_watchdog_up(dev);
2584 EXPORT_SYMBOL(netif_device_attach);
2587 * Returns a Tx hash based on the given packet descriptor a Tx queues' number
2588 * to be used as a distribution range.
2590 u16 __skb_tx_hash(const struct net_device *dev, struct sk_buff *skb,
2591 unsigned int num_tx_queues)
2595 u16 qcount = num_tx_queues;
2597 if (skb_rx_queue_recorded(skb)) {
2598 hash = skb_get_rx_queue(skb);
2599 while (unlikely(hash >= num_tx_queues))
2600 hash -= num_tx_queues;
2605 u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
2607 qoffset = dev->tc_to_txq[tc].offset;
2608 qcount = dev->tc_to_txq[tc].count;
2611 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset;
2613 EXPORT_SYMBOL(__skb_tx_hash);
2615 static void skb_warn_bad_offload(const struct sk_buff *skb)
2617 static const netdev_features_t null_features;
2618 struct net_device *dev = skb->dev;
2619 const char *name = "";
2621 if (!net_ratelimit())
2625 if (dev->dev.parent)
2626 name = dev_driver_string(dev->dev.parent);
2628 name = netdev_name(dev);
2630 WARN(1, "%s: caps=(%pNF, %pNF) len=%d data_len=%d gso_size=%d "
2631 "gso_type=%d ip_summed=%d\n",
2632 name, dev ? &dev->features : &null_features,
2633 skb->sk ? &skb->sk->sk_route_caps : &null_features,
2634 skb->len, skb->data_len, skb_shinfo(skb)->gso_size,
2635 skb_shinfo(skb)->gso_type, skb->ip_summed);
2639 * Invalidate hardware checksum when packet is to be mangled, and
2640 * complete checksum manually on outgoing path.
2642 int skb_checksum_help(struct sk_buff *skb)
2645 int ret = 0, offset;
2647 if (skb->ip_summed == CHECKSUM_COMPLETE)
2648 goto out_set_summed;
2650 if (unlikely(skb_shinfo(skb)->gso_size)) {
2651 skb_warn_bad_offload(skb);
2655 /* Before computing a checksum, we should make sure no frag could
2656 * be modified by an external entity : checksum could be wrong.
2658 if (skb_has_shared_frag(skb)) {
2659 ret = __skb_linearize(skb);
2664 offset = skb_checksum_start_offset(skb);
2665 BUG_ON(offset >= skb_headlen(skb));
2666 csum = skb_checksum(skb, offset, skb->len - offset, 0);
2668 offset += skb->csum_offset;
2669 BUG_ON(offset + sizeof(__sum16) > skb_headlen(skb));
2671 if (skb_cloned(skb) &&
2672 !skb_clone_writable(skb, offset + sizeof(__sum16))) {
2673 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2678 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0;
2680 skb->ip_summed = CHECKSUM_NONE;
2684 EXPORT_SYMBOL(skb_checksum_help);
2686 int skb_crc32c_csum_help(struct sk_buff *skb)
2689 int ret = 0, offset, start;
2691 if (skb->ip_summed != CHECKSUM_PARTIAL)
2694 if (unlikely(skb_is_gso(skb)))
2697 /* Before computing a checksum, we should make sure no frag could
2698 * be modified by an external entity : checksum could be wrong.
2700 if (unlikely(skb_has_shared_frag(skb))) {
2701 ret = __skb_linearize(skb);
2705 start = skb_checksum_start_offset(skb);
2706 offset = start + offsetof(struct sctphdr, checksum);
2707 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) {
2711 if (skb_cloned(skb) &&
2712 !skb_clone_writable(skb, offset + sizeof(__le32))) {
2713 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2717 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start,
2718 skb->len - start, ~(__u32)0,
2720 *(__le32 *)(skb->data + offset) = crc32c_csum;
2721 skb->ip_summed = CHECKSUM_NONE;
2722 skb->csum_not_inet = 0;
2727 __be16 skb_network_protocol(struct sk_buff *skb, int *depth)
2729 __be16 type = skb->protocol;
2731 /* Tunnel gso handlers can set protocol to ethernet. */
2732 if (type == htons(ETH_P_TEB)) {
2735 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr))))
2738 eth = (struct ethhdr *)skb_mac_header(skb);
2739 type = eth->h_proto;
2742 return __vlan_get_protocol(skb, type, depth);
2746 * skb_mac_gso_segment - mac layer segmentation handler.
2747 * @skb: buffer to segment
2748 * @features: features for the output path (see dev->features)
2750 struct sk_buff *skb_mac_gso_segment(struct sk_buff *skb,
2751 netdev_features_t features)
2753 struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT);
2754 struct packet_offload *ptype;
2755 int vlan_depth = skb->mac_len;
2756 __be16 type = skb_network_protocol(skb, &vlan_depth);
2758 if (unlikely(!type))
2759 return ERR_PTR(-EINVAL);
2761 __skb_pull(skb, vlan_depth);
2764 list_for_each_entry_rcu(ptype, &offload_base, list) {
2765 if (ptype->type == type && ptype->callbacks.gso_segment) {
2766 segs = ptype->callbacks.gso_segment(skb, features);
2772 __skb_push(skb, skb->data - skb_mac_header(skb));
2776 EXPORT_SYMBOL(skb_mac_gso_segment);
2779 /* openvswitch calls this on rx path, so we need a different check.
2781 static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
2784 return skb->ip_summed != CHECKSUM_PARTIAL &&
2785 skb->ip_summed != CHECKSUM_UNNECESSARY;
2787 return skb->ip_summed == CHECKSUM_NONE;
2791 * __skb_gso_segment - Perform segmentation on skb.
2792 * @skb: buffer to segment
2793 * @features: features for the output path (see dev->features)
2794 * @tx_path: whether it is called in TX path
2796 * This function segments the given skb and returns a list of segments.
2798 * It may return NULL if the skb requires no segmentation. This is
2799 * only possible when GSO is used for verifying header integrity.
2801 * Segmentation preserves SKB_SGO_CB_OFFSET bytes of previous skb cb.
2803 struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
2804 netdev_features_t features, bool tx_path)
2806 struct sk_buff *segs;
2808 if (unlikely(skb_needs_check(skb, tx_path))) {
2811 /* We're going to init ->check field in TCP or UDP header */
2812 err = skb_cow_head(skb, 0);
2814 return ERR_PTR(err);
2817 /* Only report GSO partial support if it will enable us to
2818 * support segmentation on this frame without needing additional
2821 if (features & NETIF_F_GSO_PARTIAL) {
2822 netdev_features_t partial_features = NETIF_F_GSO_ROBUST;
2823 struct net_device *dev = skb->dev;
2825 partial_features |= dev->features & dev->gso_partial_features;
2826 if (!skb_gso_ok(skb, features | partial_features))
2827 features &= ~NETIF_F_GSO_PARTIAL;
2830 BUILD_BUG_ON(SKB_SGO_CB_OFFSET +
2831 sizeof(*SKB_GSO_CB(skb)) > sizeof(skb->cb));
2833 SKB_GSO_CB(skb)->mac_offset = skb_headroom(skb);
2834 SKB_GSO_CB(skb)->encap_level = 0;
2836 skb_reset_mac_header(skb);
2837 skb_reset_mac_len(skb);
2839 segs = skb_mac_gso_segment(skb, features);
2841 if (unlikely(skb_needs_check(skb, tx_path) && !IS_ERR(segs)))
2842 skb_warn_bad_offload(skb);
2846 EXPORT_SYMBOL(__skb_gso_segment);
2848 /* Take action when hardware reception checksum errors are detected. */
2850 void netdev_rx_csum_fault(struct net_device *dev)
2852 if (net_ratelimit()) {
2853 pr_err("%s: hw csum failure\n", dev ? dev->name : "<unknown>");
2857 EXPORT_SYMBOL(netdev_rx_csum_fault);
2860 /* Actually, we should eliminate this check as soon as we know, that:
2861 * 1. IOMMU is present and allows to map all the memory.
2862 * 2. No high memory really exists on this machine.
2865 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
2867 #ifdef CONFIG_HIGHMEM
2870 if (!(dev->features & NETIF_F_HIGHDMA)) {
2871 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2872 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2874 if (PageHighMem(skb_frag_page(frag)))
2879 if (PCI_DMA_BUS_IS_PHYS) {
2880 struct device *pdev = dev->dev.parent;
2884 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2885 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2886 dma_addr_t addr = page_to_phys(skb_frag_page(frag));
2888 if (!pdev->dma_mask || addr + PAGE_SIZE - 1 > *pdev->dma_mask)
2896 /* If MPLS offload request, verify we are testing hardware MPLS features
2897 * instead of standard features for the netdev.
2899 #if IS_ENABLED(CONFIG_NET_MPLS_GSO)
2900 static netdev_features_t net_mpls_features(struct sk_buff *skb,
2901 netdev_features_t features,
2904 if (eth_p_mpls(type))
2905 features &= skb->dev->mpls_features;
2910 static netdev_features_t net_mpls_features(struct sk_buff *skb,
2911 netdev_features_t features,
2918 static netdev_features_t harmonize_features(struct sk_buff *skb,
2919 netdev_features_t features)
2924 type = skb_network_protocol(skb, &tmp);
2925 features = net_mpls_features(skb, features, type);
2927 if (skb->ip_summed != CHECKSUM_NONE &&
2928 !can_checksum_protocol(features, type)) {
2929 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
2931 if (illegal_highdma(skb->dev, skb))
2932 features &= ~NETIF_F_SG;
2937 netdev_features_t passthru_features_check(struct sk_buff *skb,
2938 struct net_device *dev,
2939 netdev_features_t features)
2943 EXPORT_SYMBOL(passthru_features_check);
2945 static netdev_features_t dflt_features_check(const struct sk_buff *skb,
2946 struct net_device *dev,
2947 netdev_features_t features)
2949 return vlan_features_check(skb, features);
2952 static netdev_features_t gso_features_check(const struct sk_buff *skb,
2953 struct net_device *dev,
2954 netdev_features_t features)
2956 u16 gso_segs = skb_shinfo(skb)->gso_segs;
2958 if (gso_segs > dev->gso_max_segs)
2959 return features & ~NETIF_F_GSO_MASK;
2961 /* Support for GSO partial features requires software
2962 * intervention before we can actually process the packets
2963 * so we need to strip support for any partial features now
2964 * and we can pull them back in after we have partially
2965 * segmented the frame.
2967 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL))
2968 features &= ~dev->gso_partial_features;
2970 /* Make sure to clear the IPv4 ID mangling feature if the
2971 * IPv4 header has the potential to be fragmented.
2973 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
2974 struct iphdr *iph = skb->encapsulation ?
2975 inner_ip_hdr(skb) : ip_hdr(skb);
2977 if (!(iph->frag_off & htons(IP_DF)))
2978 features &= ~NETIF_F_TSO_MANGLEID;
2984 netdev_features_t netif_skb_features(struct sk_buff *skb)
2986 struct net_device *dev = skb->dev;
2987 netdev_features_t features = dev->features;
2989 if (skb_is_gso(skb))
2990 features = gso_features_check(skb, dev, features);
2992 /* If encapsulation offload request, verify we are testing
2993 * hardware encapsulation features instead of standard
2994 * features for the netdev
2996 if (skb->encapsulation)
2997 features &= dev->hw_enc_features;
2999 if (skb_vlan_tagged(skb))
3000 features = netdev_intersect_features(features,
3001 dev->vlan_features |
3002 NETIF_F_HW_VLAN_CTAG_TX |
3003 NETIF_F_HW_VLAN_STAG_TX);
3005 if (dev->netdev_ops->ndo_features_check)
3006 features &= dev->netdev_ops->ndo_features_check(skb, dev,
3009 features &= dflt_features_check(skb, dev, features);
3011 return harmonize_features(skb, features);
3013 EXPORT_SYMBOL(netif_skb_features);
3015 static int xmit_one(struct sk_buff *skb, struct net_device *dev,
3016 struct netdev_queue *txq, bool more)
3021 if (!list_empty(&ptype_all) || !list_empty(&dev->ptype_all))
3022 dev_queue_xmit_nit(skb, dev);
3025 trace_net_dev_start_xmit(skb, dev);
3026 rc = netdev_start_xmit(skb, dev, txq, more);
3027 trace_net_dev_xmit(skb, rc, dev, len);
3032 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev,
3033 struct netdev_queue *txq, int *ret)
3035 struct sk_buff *skb = first;
3036 int rc = NETDEV_TX_OK;
3039 struct sk_buff *next = skb->next;
3042 rc = xmit_one(skb, dev, txq, next != NULL);
3043 if (unlikely(!dev_xmit_complete(rc))) {
3049 if (netif_xmit_stopped(txq) && skb) {
3050 rc = NETDEV_TX_BUSY;
3060 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb,
3061 netdev_features_t features)
3063 if (skb_vlan_tag_present(skb) &&
3064 !vlan_hw_offload_capable(features, skb->vlan_proto))
3065 skb = __vlan_hwaccel_push_inside(skb);
3069 int skb_csum_hwoffload_help(struct sk_buff *skb,
3070 const netdev_features_t features)
3072 if (unlikely(skb->csum_not_inet))
3073 return !!(features & NETIF_F_SCTP_CRC) ? 0 :
3074 skb_crc32c_csum_help(skb);
3076 return !!(features & NETIF_F_CSUM_MASK) ? 0 : skb_checksum_help(skb);
3078 EXPORT_SYMBOL(skb_csum_hwoffload_help);
3080 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again)
3082 netdev_features_t features;
3084 features = netif_skb_features(skb);
3085 skb = validate_xmit_vlan(skb, features);
3089 if (netif_needs_gso(skb, features)) {
3090 struct sk_buff *segs;
3092 segs = skb_gso_segment(skb, features);
3100 if (skb_needs_linearize(skb, features) &&
3101 __skb_linearize(skb))
3104 /* If packet is not checksummed and device does not
3105 * support checksumming for this protocol, complete
3106 * checksumming here.
3108 if (skb->ip_summed == CHECKSUM_PARTIAL) {
3109 if (skb->encapsulation)
3110 skb_set_inner_transport_header(skb,
3111 skb_checksum_start_offset(skb));
3113 skb_set_transport_header(skb,
3114 skb_checksum_start_offset(skb));
3115 if (skb_csum_hwoffload_help(skb, features))
3120 skb = validate_xmit_xfrm(skb, features, again);
3127 atomic_long_inc(&dev->tx_dropped);
3131 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again)
3133 struct sk_buff *next, *head = NULL, *tail;
3135 for (; skb != NULL; skb = next) {
3139 /* in case skb wont be segmented, point to itself */
3142 skb = validate_xmit_skb(skb, dev, again);
3150 /* If skb was segmented, skb->prev points to
3151 * the last segment. If not, it still contains skb.
3157 EXPORT_SYMBOL_GPL(validate_xmit_skb_list);
3159 static void qdisc_pkt_len_init(struct sk_buff *skb)
3161 const struct skb_shared_info *shinfo = skb_shinfo(skb);
3163 qdisc_skb_cb(skb)->pkt_len = skb->len;
3165 /* To get more precise estimation of bytes sent on wire,
3166 * we add to pkt_len the headers size of all segments
3168 if (shinfo->gso_size) {
3169 unsigned int hdr_len;
3170 u16 gso_segs = shinfo->gso_segs;
3172 /* mac layer + network layer */
3173 hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
3175 /* + transport layer */
3176 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) {
3177 const struct tcphdr *th;
3178 struct tcphdr _tcphdr;
3180 th = skb_header_pointer(skb, skb_transport_offset(skb),
3181 sizeof(_tcphdr), &_tcphdr);
3183 hdr_len += __tcp_hdrlen(th);
3185 struct udphdr _udphdr;
3187 if (skb_header_pointer(skb, skb_transport_offset(skb),
3188 sizeof(_udphdr), &_udphdr))
3189 hdr_len += sizeof(struct udphdr);
3192 if (shinfo->gso_type & SKB_GSO_DODGY)
3193 gso_segs = DIV_ROUND_UP(skb->len - hdr_len,
3196 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len;
3200 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
3201 struct net_device *dev,
3202 struct netdev_queue *txq)
3204 spinlock_t *root_lock = qdisc_lock(q);
3205 struct sk_buff *to_free = NULL;
3209 qdisc_calculate_pkt_len(skb, q);
3211 if (q->flags & TCQ_F_NOLOCK) {
3212 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3213 __qdisc_drop(skb, &to_free);
3216 rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
3220 if (unlikely(to_free))
3221 kfree_skb_list(to_free);
3226 * Heuristic to force contended enqueues to serialize on a
3227 * separate lock before trying to get qdisc main lock.
3228 * This permits qdisc->running owner to get the lock more
3229 * often and dequeue packets faster.
3231 contended = qdisc_is_running(q);
3232 if (unlikely(contended))
3233 spin_lock(&q->busylock);
3235 spin_lock(root_lock);
3236 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3237 __qdisc_drop(skb, &to_free);
3239 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
3240 qdisc_run_begin(q)) {
3242 * This is a work-conserving queue; there are no old skbs
3243 * waiting to be sent out; and the qdisc is not running -
3244 * xmit the skb directly.
3247 qdisc_bstats_update(q, skb);
3249 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) {
3250 if (unlikely(contended)) {
3251 spin_unlock(&q->busylock);
3258 rc = NET_XMIT_SUCCESS;
3260 rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
3261 if (qdisc_run_begin(q)) {
3262 if (unlikely(contended)) {
3263 spin_unlock(&q->busylock);
3270 spin_unlock(root_lock);
3271 if (unlikely(to_free))
3272 kfree_skb_list(to_free);
3273 if (unlikely(contended))
3274 spin_unlock(&q->busylock);
3278 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO)
3279 static void skb_update_prio(struct sk_buff *skb)
3281 const struct netprio_map *map;
3282 const struct sock *sk;
3283 unsigned int prioidx;
3287 map = rcu_dereference_bh(skb->dev->priomap);
3290 sk = skb_to_full_sk(skb);
3294 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data);
3296 if (prioidx < map->priomap_len)
3297 skb->priority = map->priomap[prioidx];
3300 #define skb_update_prio(skb)
3303 DEFINE_PER_CPU(int, xmit_recursion);
3304 EXPORT_SYMBOL(xmit_recursion);
3307 * dev_loopback_xmit - loop back @skb
3308 * @net: network namespace this loopback is happening in
3309 * @sk: sk needed to be a netfilter okfn
3310 * @skb: buffer to transmit
3312 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb)
3314 skb_reset_mac_header(skb);
3315 __skb_pull(skb, skb_network_offset(skb));
3316 skb->pkt_type = PACKET_LOOPBACK;
3317 skb->ip_summed = CHECKSUM_UNNECESSARY;
3318 WARN_ON(!skb_dst(skb));
3323 EXPORT_SYMBOL(dev_loopback_xmit);
3325 #ifdef CONFIG_NET_EGRESS
3326 static struct sk_buff *
3327 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
3329 struct mini_Qdisc *miniq = rcu_dereference_bh(dev->miniq_egress);
3330 struct tcf_result cl_res;
3335 /* qdisc_skb_cb(skb)->pkt_len was already set by the caller. */
3336 mini_qdisc_bstats_cpu_update(miniq, skb);
3338 switch (tcf_classify(skb, miniq->filter_list, &cl_res, false)) {
3340 case TC_ACT_RECLASSIFY:
3341 skb->tc_index = TC_H_MIN(cl_res.classid);
3344 mini_qdisc_qstats_cpu_drop(miniq);
3345 *ret = NET_XMIT_DROP;
3351 *ret = NET_XMIT_SUCCESS;
3354 case TC_ACT_REDIRECT:
3355 /* No need to push/pop skb's mac_header here on egress! */
3356 skb_do_redirect(skb);
3357 *ret = NET_XMIT_SUCCESS;
3365 #endif /* CONFIG_NET_EGRESS */
3367 static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb)
3370 struct xps_dev_maps *dev_maps;
3371 struct xps_map *map;
3372 int queue_index = -1;
3375 dev_maps = rcu_dereference(dev->xps_maps);
3377 unsigned int tci = skb->sender_cpu - 1;
3381 tci += netdev_get_prio_tc_map(dev, skb->priority);
3384 map = rcu_dereference(dev_maps->cpu_map[tci]);
3387 queue_index = map->queues[0];
3389 queue_index = map->queues[reciprocal_scale(skb_get_hash(skb),
3391 if (unlikely(queue_index >= dev->real_num_tx_queues))
3403 static u16 __netdev_pick_tx(struct net_device *dev, struct sk_buff *skb)
3405 struct sock *sk = skb->sk;
3406 int queue_index = sk_tx_queue_get(sk);
3408 if (queue_index < 0 || skb->ooo_okay ||
3409 queue_index >= dev->real_num_tx_queues) {
3410 int new_index = get_xps_queue(dev, skb);
3413 new_index = skb_tx_hash(dev, skb);
3415 if (queue_index != new_index && sk &&
3417 rcu_access_pointer(sk->sk_dst_cache))
3418 sk_tx_queue_set(sk, new_index);
3420 queue_index = new_index;
3426 struct netdev_queue *netdev_pick_tx(struct net_device *dev,
3427 struct sk_buff *skb,
3430 int queue_index = 0;
3433 u32 sender_cpu = skb->sender_cpu - 1;
3435 if (sender_cpu >= (u32)NR_CPUS)
3436 skb->sender_cpu = raw_smp_processor_id() + 1;
3439 if (dev->real_num_tx_queues != 1) {
3440 const struct net_device_ops *ops = dev->netdev_ops;
3442 if (ops->ndo_select_queue)
3443 queue_index = ops->ndo_select_queue(dev, skb, accel_priv,
3446 queue_index = __netdev_pick_tx(dev, skb);
3448 queue_index = netdev_cap_txqueue(dev, queue_index);
3451 skb_set_queue_mapping(skb, queue_index);
3452 return netdev_get_tx_queue(dev, queue_index);
3456 * __dev_queue_xmit - transmit a buffer
3457 * @skb: buffer to transmit
3458 * @accel_priv: private data used for L2 forwarding offload
3460 * Queue a buffer for transmission to a network device. The caller must
3461 * have set the device and priority and built the buffer before calling
3462 * this function. The function can be called from an interrupt.
3464 * A negative errno code is returned on a failure. A success does not
3465 * guarantee the frame will be transmitted as it may be dropped due
3466 * to congestion or traffic shaping.
3468 * -----------------------------------------------------------------------------------
3469 * I notice this method can also return errors from the queue disciplines,
3470 * including NET_XMIT_DROP, which is a positive value. So, errors can also
3473 * Regardless of the return value, the skb is consumed, so it is currently
3474 * difficult to retry a send to this method. (You can bump the ref count
3475 * before sending to hold a reference for retry if you are careful.)
3477 * When calling this method, interrupts MUST be enabled. This is because
3478 * the BH enable code must have IRQs enabled so that it will not deadlock.
3481 static int __dev_queue_xmit(struct sk_buff *skb, void *accel_priv)
3483 struct net_device *dev = skb->dev;
3484 struct netdev_queue *txq;
3489 skb_reset_mac_header(skb);
3491 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP))
3492 __skb_tstamp_tx(skb, NULL, skb->sk, SCM_TSTAMP_SCHED);
3494 /* Disable soft irqs for various locks below. Also
3495 * stops preemption for RCU.
3499 skb_update_prio(skb);
3501 qdisc_pkt_len_init(skb);
3502 #ifdef CONFIG_NET_CLS_ACT
3503 skb->tc_at_ingress = 0;
3504 # ifdef CONFIG_NET_EGRESS
3505 if (static_key_false(&egress_needed)) {
3506 skb = sch_handle_egress(skb, &rc, dev);
3512 /* If device/qdisc don't need skb->dst, release it right now while
3513 * its hot in this cpu cache.
3515 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
3520 txq = netdev_pick_tx(dev, skb, accel_priv);
3521 q = rcu_dereference_bh(txq->qdisc);
3523 trace_net_dev_queue(skb);
3525 rc = __dev_xmit_skb(skb, q, dev, txq);
3529 /* The device has no queue. Common case for software devices:
3530 * loopback, all the sorts of tunnels...
3532 * Really, it is unlikely that netif_tx_lock protection is necessary
3533 * here. (f.e. loopback and IP tunnels are clean ignoring statistics
3535 * However, it is possible, that they rely on protection
3538 * Check this and shot the lock. It is not prone from deadlocks.
3539 *Either shot noqueue qdisc, it is even simpler 8)
3541 if (dev->flags & IFF_UP) {
3542 int cpu = smp_processor_id(); /* ok because BHs are off */
3544 if (txq->xmit_lock_owner != cpu) {
3545 if (unlikely(__this_cpu_read(xmit_recursion) >
3546 XMIT_RECURSION_LIMIT))
3547 goto recursion_alert;
3549 skb = validate_xmit_skb(skb, dev, &again);
3553 HARD_TX_LOCK(dev, txq, cpu);
3555 if (!netif_xmit_stopped(txq)) {
3556 __this_cpu_inc(xmit_recursion);
3557 skb = dev_hard_start_xmit(skb, dev, txq, &rc);
3558 __this_cpu_dec(xmit_recursion);
3559 if (dev_xmit_complete(rc)) {
3560 HARD_TX_UNLOCK(dev, txq);
3564 HARD_TX_UNLOCK(dev, txq);
3565 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
3568 /* Recursion is detected! It is possible,
3572 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
3578 rcu_read_unlock_bh();
3580 atomic_long_inc(&dev->tx_dropped);
3581 kfree_skb_list(skb);
3584 rcu_read_unlock_bh();
3588 int dev_queue_xmit(struct sk_buff *skb)
3590 return __dev_queue_xmit(skb, NULL);
3592 EXPORT_SYMBOL(dev_queue_xmit);
3594 int dev_queue_xmit_accel(struct sk_buff *skb, void *accel_priv)
3596 return __dev_queue_xmit(skb, accel_priv);
3598 EXPORT_SYMBOL(dev_queue_xmit_accel);
3601 /*************************************************************************
3603 *************************************************************************/
3605 int netdev_max_backlog __read_mostly = 1000;
3606 EXPORT_SYMBOL(netdev_max_backlog);
3608 int netdev_tstamp_prequeue __read_mostly = 1;
3609 int netdev_budget __read_mostly = 300;
3610 unsigned int __read_mostly netdev_budget_usecs = 2000;
3611 int weight_p __read_mostly = 64; /* old backlog weight */
3612 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */
3613 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */
3614 int dev_rx_weight __read_mostly = 64;
3615 int dev_tx_weight __read_mostly = 64;
3617 /* Called with irq disabled */
3618 static inline void ____napi_schedule(struct softnet_data *sd,
3619 struct napi_struct *napi)
3621 list_add_tail(&napi->poll_list, &sd->poll_list);
3622 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3627 /* One global table that all flow-based protocols share. */
3628 struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
3629 EXPORT_SYMBOL(rps_sock_flow_table);
3630 u32 rps_cpu_mask __read_mostly;
3631 EXPORT_SYMBOL(rps_cpu_mask);
3633 struct static_key rps_needed __read_mostly;
3634 EXPORT_SYMBOL(rps_needed);
3635 struct static_key rfs_needed __read_mostly;
3636 EXPORT_SYMBOL(rfs_needed);
3638 static struct rps_dev_flow *
3639 set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3640 struct rps_dev_flow *rflow, u16 next_cpu)
3642 if (next_cpu < nr_cpu_ids) {
3643 #ifdef CONFIG_RFS_ACCEL
3644 struct netdev_rx_queue *rxqueue;
3645 struct rps_dev_flow_table *flow_table;
3646 struct rps_dev_flow *old_rflow;
3651 /* Should we steer this flow to a different hardware queue? */
3652 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
3653 !(dev->features & NETIF_F_NTUPLE))
3655 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
3656 if (rxq_index == skb_get_rx_queue(skb))
3659 rxqueue = dev->_rx + rxq_index;
3660 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3663 flow_id = skb_get_hash(skb) & flow_table->mask;
3664 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
3665 rxq_index, flow_id);
3669 rflow = &flow_table->flows[flow_id];
3671 if (old_rflow->filter == rflow->filter)
3672 old_rflow->filter = RPS_NO_FILTER;
3676 per_cpu(softnet_data, next_cpu).input_queue_head;
3679 rflow->cpu = next_cpu;
3684 * get_rps_cpu is called from netif_receive_skb and returns the target
3685 * CPU from the RPS map of the receiving queue for a given skb.
3686 * rcu_read_lock must be held on entry.
3688 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3689 struct rps_dev_flow **rflowp)
3691 const struct rps_sock_flow_table *sock_flow_table;
3692 struct netdev_rx_queue *rxqueue = dev->_rx;
3693 struct rps_dev_flow_table *flow_table;
3694 struct rps_map *map;
3699 if (skb_rx_queue_recorded(skb)) {
3700 u16 index = skb_get_rx_queue(skb);
3702 if (unlikely(index >= dev->real_num_rx_queues)) {
3703 WARN_ONCE(dev->real_num_rx_queues > 1,
3704 "%s received packet on queue %u, but number "
3705 "of RX queues is %u\n",
3706 dev->name, index, dev->real_num_rx_queues);
3712 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */
3714 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3715 map = rcu_dereference(rxqueue->rps_map);
3716 if (!flow_table && !map)
3719 skb_reset_network_header(skb);
3720 hash = skb_get_hash(skb);
3724 sock_flow_table = rcu_dereference(rps_sock_flow_table);
3725 if (flow_table && sock_flow_table) {
3726 struct rps_dev_flow *rflow;
3730 /* First check into global flow table if there is a match */
3731 ident = sock_flow_table->ents[hash & sock_flow_table->mask];
3732 if ((ident ^ hash) & ~rps_cpu_mask)
3735 next_cpu = ident & rps_cpu_mask;
3737 /* OK, now we know there is a match,
3738 * we can look at the local (per receive queue) flow table
3740 rflow = &flow_table->flows[hash & flow_table->mask];
3744 * If the desired CPU (where last recvmsg was done) is
3745 * different from current CPU (one in the rx-queue flow
3746 * table entry), switch if one of the following holds:
3747 * - Current CPU is unset (>= nr_cpu_ids).
3748 * - Current CPU is offline.
3749 * - The current CPU's queue tail has advanced beyond the
3750 * last packet that was enqueued using this table entry.
3751 * This guarantees that all previous packets for the flow
3752 * have been dequeued, thus preserving in order delivery.
3754 if (unlikely(tcpu != next_cpu) &&
3755 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) ||
3756 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
3757 rflow->last_qtail)) >= 0)) {
3759 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
3762 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) {
3772 tcpu = map->cpus[reciprocal_scale(hash, map->len)];
3773 if (cpu_online(tcpu)) {
3783 #ifdef CONFIG_RFS_ACCEL
3786 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
3787 * @dev: Device on which the filter was set
3788 * @rxq_index: RX queue index
3789 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
3790 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
3792 * Drivers that implement ndo_rx_flow_steer() should periodically call
3793 * this function for each installed filter and remove the filters for
3794 * which it returns %true.
3796 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
3797 u32 flow_id, u16 filter_id)
3799 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
3800 struct rps_dev_flow_table *flow_table;
3801 struct rps_dev_flow *rflow;
3806 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3807 if (flow_table && flow_id <= flow_table->mask) {
3808 rflow = &flow_table->flows[flow_id];
3809 cpu = READ_ONCE(rflow->cpu);
3810 if (rflow->filter == filter_id && cpu < nr_cpu_ids &&
3811 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
3812 rflow->last_qtail) <
3813 (int)(10 * flow_table->mask)))
3819 EXPORT_SYMBOL(rps_may_expire_flow);
3821 #endif /* CONFIG_RFS_ACCEL */
3823 /* Called from hardirq (IPI) context */
3824 static void rps_trigger_softirq(void *data)
3826 struct softnet_data *sd = data;
3828 ____napi_schedule(sd, &sd->backlog);
3832 #endif /* CONFIG_RPS */
3835 * Check if this softnet_data structure is another cpu one
3836 * If yes, queue it to our IPI list and return 1
3839 static int rps_ipi_queued(struct softnet_data *sd)
3842 struct softnet_data *mysd = this_cpu_ptr(&softnet_data);
3845 sd->rps_ipi_next = mysd->rps_ipi_list;
3846 mysd->rps_ipi_list = sd;
3848 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3851 #endif /* CONFIG_RPS */
3855 #ifdef CONFIG_NET_FLOW_LIMIT
3856 int netdev_flow_limit_table_len __read_mostly = (1 << 12);
3859 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
3861 #ifdef CONFIG_NET_FLOW_LIMIT
3862 struct sd_flow_limit *fl;
3863 struct softnet_data *sd;
3864 unsigned int old_flow, new_flow;
3866 if (qlen < (netdev_max_backlog >> 1))
3869 sd = this_cpu_ptr(&softnet_data);
3872 fl = rcu_dereference(sd->flow_limit);
3874 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1);
3875 old_flow = fl->history[fl->history_head];
3876 fl->history[fl->history_head] = new_flow;
3879 fl->history_head &= FLOW_LIMIT_HISTORY - 1;
3881 if (likely(fl->buckets[old_flow]))
3882 fl->buckets[old_flow]--;
3884 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) {
3896 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
3897 * queue (may be a remote CPU queue).
3899 static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
3900 unsigned int *qtail)
3902 struct softnet_data *sd;
3903 unsigned long flags;
3906 sd = &per_cpu(softnet_data, cpu);
3908 local_irq_save(flags);
3911 if (!netif_running(skb->dev))
3913 qlen = skb_queue_len(&sd->input_pkt_queue);
3914 if (qlen <= netdev_max_backlog && !skb_flow_limit(skb, qlen)) {
3917 __skb_queue_tail(&sd->input_pkt_queue, skb);
3918 input_queue_tail_incr_save(sd, qtail);
3920 local_irq_restore(flags);
3921 return NET_RX_SUCCESS;
3924 /* Schedule NAPI for backlog device
3925 * We can use non atomic operation since we own the queue lock
3927 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
3928 if (!rps_ipi_queued(sd))
3929 ____napi_schedule(sd, &sd->backlog);
3938 local_irq_restore(flags);
3940 atomic_long_inc(&skb->dev->rx_dropped);
3945 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb)
3947 struct net_device *dev = skb->dev;
3948 struct netdev_rx_queue *rxqueue;
3952 if (skb_rx_queue_recorded(skb)) {
3953 u16 index = skb_get_rx_queue(skb);
3955 if (unlikely(index >= dev->real_num_rx_queues)) {
3956 WARN_ONCE(dev->real_num_rx_queues > 1,
3957 "%s received packet on queue %u, but number "
3958 "of RX queues is %u\n",
3959 dev->name, index, dev->real_num_rx_queues);
3961 return rxqueue; /* Return first rxqueue */
3968 static u32 netif_receive_generic_xdp(struct sk_buff *skb,
3969 struct bpf_prog *xdp_prog)
3971 struct netdev_rx_queue *rxqueue;
3972 u32 metalen, act = XDP_DROP;
3973 struct xdp_buff xdp;
3978 /* Reinjected packets coming from act_mirred or similar should
3979 * not get XDP generic processing.
3981 if (skb_cloned(skb))
3984 /* XDP packets must be linear and must have sufficient headroom
3985 * of XDP_PACKET_HEADROOM bytes. This is the guarantee that also
3986 * native XDP provides, thus we need to do it here as well.
3988 if (skb_is_nonlinear(skb) ||
3989 skb_headroom(skb) < XDP_PACKET_HEADROOM) {
3990 int hroom = XDP_PACKET_HEADROOM - skb_headroom(skb);
3991 int troom = skb->tail + skb->data_len - skb->end;
3993 /* In case we have to go down the path and also linearize,
3994 * then lets do the pskb_expand_head() work just once here.
3996 if (pskb_expand_head(skb,
3997 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0,
3998 troom > 0 ? troom + 128 : 0, GFP_ATOMIC))
4000 if (skb_linearize(skb))
4004 /* The XDP program wants to see the packet starting at the MAC
4007 mac_len = skb->data - skb_mac_header(skb);
4008 hlen = skb_headlen(skb) + mac_len;
4009 xdp.data = skb->data - mac_len;
4010 xdp.data_meta = xdp.data;
4011 xdp.data_end = xdp.data + hlen;
4012 xdp.data_hard_start = skb->data - skb_headroom(skb);
4013 orig_data = xdp.data;
4015 rxqueue = netif_get_rxqueue(skb);
4016 xdp.rxq = &rxqueue->xdp_rxq;
4018 act = bpf_prog_run_xdp(xdp_prog, &xdp);
4020 off = xdp.data - orig_data;
4022 __skb_pull(skb, off);
4024 __skb_push(skb, -off);
4025 skb->mac_header += off;
4030 __skb_push(skb, mac_len);
4033 metalen = xdp.data - xdp.data_meta;
4035 skb_metadata_set(skb, metalen);
4038 bpf_warn_invalid_xdp_action(act);
4041 trace_xdp_exception(skb->dev, xdp_prog, act);
4052 /* When doing generic XDP we have to bypass the qdisc layer and the
4053 * network taps in order to match in-driver-XDP behavior.
4055 void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog)
4057 struct net_device *dev = skb->dev;
4058 struct netdev_queue *txq;
4059 bool free_skb = true;
4062 txq = netdev_pick_tx(dev, skb, NULL);
4063 cpu = smp_processor_id();
4064 HARD_TX_LOCK(dev, txq, cpu);
4065 if (!netif_xmit_stopped(txq)) {
4066 rc = netdev_start_xmit(skb, dev, txq, 0);
4067 if (dev_xmit_complete(rc))
4070 HARD_TX_UNLOCK(dev, txq);
4072 trace_xdp_exception(dev, xdp_prog, XDP_TX);
4076 EXPORT_SYMBOL_GPL(generic_xdp_tx);
4078 static struct static_key generic_xdp_needed __read_mostly;
4080 int do_xdp_generic(struct bpf_prog *xdp_prog, struct sk_buff *skb)
4083 u32 act = netif_receive_generic_xdp(skb, xdp_prog);
4086 if (act != XDP_PASS) {
4089 err = xdp_do_generic_redirect(skb->dev, skb,
4093 /* fallthru to submit skb */
4095 generic_xdp_tx(skb, xdp_prog);
4106 EXPORT_SYMBOL_GPL(do_xdp_generic);
4108 static int netif_rx_internal(struct sk_buff *skb)
4112 net_timestamp_check(netdev_tstamp_prequeue, skb);
4114 trace_netif_rx(skb);
4116 if (static_key_false(&generic_xdp_needed)) {
4121 ret = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
4125 /* Consider XDP consuming the packet a success from
4126 * the netdev point of view we do not want to count
4129 if (ret != XDP_PASS)
4130 return NET_RX_SUCCESS;
4134 if (static_key_false(&rps_needed)) {
4135 struct rps_dev_flow voidflow, *rflow = &voidflow;
4141 cpu = get_rps_cpu(skb->dev, skb, &rflow);
4143 cpu = smp_processor_id();
4145 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4154 ret = enqueue_to_backlog(skb, get_cpu(), &qtail);
4161 * netif_rx - post buffer to the network code
4162 * @skb: buffer to post
4164 * This function receives a packet from a device driver and queues it for
4165 * the upper (protocol) levels to process. It always succeeds. The buffer
4166 * may be dropped during processing for congestion control or by the
4170 * NET_RX_SUCCESS (no congestion)
4171 * NET_RX_DROP (packet was dropped)
4175 int netif_rx(struct sk_buff *skb)
4177 trace_netif_rx_entry(skb);
4179 return netif_rx_internal(skb);
4181 EXPORT_SYMBOL(netif_rx);
4183 int netif_rx_ni(struct sk_buff *skb)
4187 trace_netif_rx_ni_entry(skb);
4190 err = netif_rx_internal(skb);
4191 if (local_softirq_pending())
4197 EXPORT_SYMBOL(netif_rx_ni);
4199 static __latent_entropy void net_tx_action(struct softirq_action *h)
4201 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
4203 if (sd->completion_queue) {
4204 struct sk_buff *clist;
4206 local_irq_disable();
4207 clist = sd->completion_queue;
4208 sd->completion_queue = NULL;
4212 struct sk_buff *skb = clist;
4214 clist = clist->next;
4216 WARN_ON(refcount_read(&skb->users));
4217 if (likely(get_kfree_skb_cb(skb)->reason == SKB_REASON_CONSUMED))
4218 trace_consume_skb(skb);
4220 trace_kfree_skb(skb, net_tx_action);
4222 if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
4225 __kfree_skb_defer(skb);
4228 __kfree_skb_flush();
4231 if (sd->output_queue) {
4234 local_irq_disable();
4235 head = sd->output_queue;
4236 sd->output_queue = NULL;
4237 sd->output_queue_tailp = &sd->output_queue;
4241 struct Qdisc *q = head;
4242 spinlock_t *root_lock = NULL;
4244 head = head->next_sched;
4246 if (!(q->flags & TCQ_F_NOLOCK)) {
4247 root_lock = qdisc_lock(q);
4248 spin_lock(root_lock);
4250 /* We need to make sure head->next_sched is read
4251 * before clearing __QDISC_STATE_SCHED
4253 smp_mb__before_atomic();
4254 clear_bit(__QDISC_STATE_SCHED, &q->state);
4257 spin_unlock(root_lock);
4261 xfrm_dev_backlog(sd);
4264 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE)
4265 /* This hook is defined here for ATM LANE */
4266 int (*br_fdb_test_addr_hook)(struct net_device *dev,
4267 unsigned char *addr) __read_mostly;
4268 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook);
4271 static inline struct sk_buff *
4272 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
4273 struct net_device *orig_dev)
4275 #ifdef CONFIG_NET_CLS_ACT
4276 struct mini_Qdisc *miniq = rcu_dereference_bh(skb->dev->miniq_ingress);
4277 struct tcf_result cl_res;
4279 /* If there's at least one ingress present somewhere (so
4280 * we get here via enabled static key), remaining devices
4281 * that are not configured with an ingress qdisc will bail
4288 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4292 qdisc_skb_cb(skb)->pkt_len = skb->len;
4293 skb->tc_at_ingress = 1;
4294 mini_qdisc_bstats_cpu_update(miniq, skb);
4296 switch (tcf_classify(skb, miniq->filter_list, &cl_res, false)) {
4298 case TC_ACT_RECLASSIFY:
4299 skb->tc_index = TC_H_MIN(cl_res.classid);
4302 mini_qdisc_qstats_cpu_drop(miniq);
4310 case TC_ACT_REDIRECT:
4311 /* skb_mac_header check was done by cls/act_bpf, so
4312 * we can safely push the L2 header back before
4313 * redirecting to another netdev
4315 __skb_push(skb, skb->mac_len);
4316 skb_do_redirect(skb);
4321 #endif /* CONFIG_NET_CLS_ACT */
4326 * netdev_is_rx_handler_busy - check if receive handler is registered
4327 * @dev: device to check
4329 * Check if a receive handler is already registered for a given device.
4330 * Return true if there one.
4332 * The caller must hold the rtnl_mutex.
4334 bool netdev_is_rx_handler_busy(struct net_device *dev)
4337 return dev && rtnl_dereference(dev->rx_handler);
4339 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy);
4342 * netdev_rx_handler_register - register receive handler
4343 * @dev: device to register a handler for
4344 * @rx_handler: receive handler to register
4345 * @rx_handler_data: data pointer that is used by rx handler
4347 * Register a receive handler for a device. This handler will then be
4348 * called from __netif_receive_skb. A negative errno code is returned
4351 * The caller must hold the rtnl_mutex.
4353 * For a general description of rx_handler, see enum rx_handler_result.
4355 int netdev_rx_handler_register(struct net_device *dev,
4356 rx_handler_func_t *rx_handler,
4357 void *rx_handler_data)
4359 if (netdev_is_rx_handler_busy(dev))
4362 if (dev->priv_flags & IFF_NO_RX_HANDLER)
4365 /* Note: rx_handler_data must be set before rx_handler */
4366 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data);
4367 rcu_assign_pointer(dev->rx_handler, rx_handler);
4371 EXPORT_SYMBOL_GPL(netdev_rx_handler_register);
4374 * netdev_rx_handler_unregister - unregister receive handler
4375 * @dev: device to unregister a handler from
4377 * Unregister a receive handler from a device.
4379 * The caller must hold the rtnl_mutex.
4381 void netdev_rx_handler_unregister(struct net_device *dev)
4385 RCU_INIT_POINTER(dev->rx_handler, NULL);
4386 /* a reader seeing a non NULL rx_handler in a rcu_read_lock()
4387 * section has a guarantee to see a non NULL rx_handler_data
4391 RCU_INIT_POINTER(dev->rx_handler_data, NULL);
4393 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister);
4396 * Limit the use of PFMEMALLOC reserves to those protocols that implement
4397 * the special handling of PFMEMALLOC skbs.
4399 static bool skb_pfmemalloc_protocol(struct sk_buff *skb)
4401 switch (skb->protocol) {
4402 case htons(ETH_P_ARP):
4403 case htons(ETH_P_IP):
4404 case htons(ETH_P_IPV6):
4405 case htons(ETH_P_8021Q):
4406 case htons(ETH_P_8021AD):
4413 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev,
4414 int *ret, struct net_device *orig_dev)
4416 #ifdef CONFIG_NETFILTER_INGRESS
4417 if (nf_hook_ingress_active(skb)) {
4421 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4426 ingress_retval = nf_hook_ingress(skb);
4428 return ingress_retval;
4430 #endif /* CONFIG_NETFILTER_INGRESS */
4434 static int __netif_receive_skb_core(struct sk_buff *skb, bool pfmemalloc)
4436 struct packet_type *ptype, *pt_prev;
4437 rx_handler_func_t *rx_handler;
4438 struct net_device *orig_dev;
4439 bool deliver_exact = false;
4440 int ret = NET_RX_DROP;
4443 net_timestamp_check(!netdev_tstamp_prequeue, skb);
4445 trace_netif_receive_skb(skb);
4447 orig_dev = skb->dev;
4449 skb_reset_network_header(skb);
4450 if (!skb_transport_header_was_set(skb))
4451 skb_reset_transport_header(skb);
4452 skb_reset_mac_len(skb);
4457 skb->skb_iif = skb->dev->ifindex;
4459 __this_cpu_inc(softnet_data.processed);
4461 if (skb->protocol == cpu_to_be16(ETH_P_8021Q) ||
4462 skb->protocol == cpu_to_be16(ETH_P_8021AD)) {
4463 skb = skb_vlan_untag(skb);
4468 if (skb_skip_tc_classify(skb))
4474 list_for_each_entry_rcu(ptype, &ptype_all, list) {
4476 ret = deliver_skb(skb, pt_prev, orig_dev);
4480 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) {
4482 ret = deliver_skb(skb, pt_prev, orig_dev);
4487 #ifdef CONFIG_NET_INGRESS
4488 if (static_key_false(&ingress_needed)) {
4489 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev);
4493 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0)
4499 if (pfmemalloc && !skb_pfmemalloc_protocol(skb))
4502 if (skb_vlan_tag_present(skb)) {
4504 ret = deliver_skb(skb, pt_prev, orig_dev);
4507 if (vlan_do_receive(&skb))
4509 else if (unlikely(!skb))
4513 rx_handler = rcu_dereference(skb->dev->rx_handler);
4516 ret = deliver_skb(skb, pt_prev, orig_dev);
4519 switch (rx_handler(&skb)) {
4520 case RX_HANDLER_CONSUMED:
4521 ret = NET_RX_SUCCESS;
4523 case RX_HANDLER_ANOTHER:
4525 case RX_HANDLER_EXACT:
4526 deliver_exact = true;
4527 case RX_HANDLER_PASS:
4534 if (unlikely(skb_vlan_tag_present(skb))) {
4535 if (skb_vlan_tag_get_id(skb))
4536 skb->pkt_type = PACKET_OTHERHOST;
4537 /* Note: we might in the future use prio bits
4538 * and set skb->priority like in vlan_do_receive()
4539 * For the time being, just ignore Priority Code Point
4544 type = skb->protocol;
4546 /* deliver only exact match when indicated */
4547 if (likely(!deliver_exact)) {
4548 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4549 &ptype_base[ntohs(type) &
4553 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4554 &orig_dev->ptype_specific);
4556 if (unlikely(skb->dev != orig_dev)) {
4557 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4558 &skb->dev->ptype_specific);
4562 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
4565 ret = pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
4569 atomic_long_inc(&skb->dev->rx_dropped);
4571 atomic_long_inc(&skb->dev->rx_nohandler);
4573 /* Jamal, now you will not able to escape explaining
4574 * me how you were going to use this. :-)
4584 * netif_receive_skb_core - special purpose version of netif_receive_skb
4585 * @skb: buffer to process
4587 * More direct receive version of netif_receive_skb(). It should
4588 * only be used by callers that have a need to skip RPS and Generic XDP.
4589 * Caller must also take care of handling if (page_is_)pfmemalloc.
4591 * This function may only be called from softirq context and interrupts
4592 * should be enabled.
4594 * Return values (usually ignored):
4595 * NET_RX_SUCCESS: no congestion
4596 * NET_RX_DROP: packet was dropped
4598 int netif_receive_skb_core(struct sk_buff *skb)
4603 ret = __netif_receive_skb_core(skb, false);
4608 EXPORT_SYMBOL(netif_receive_skb_core);
4610 static int __netif_receive_skb(struct sk_buff *skb)
4614 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) {
4615 unsigned int noreclaim_flag;
4618 * PFMEMALLOC skbs are special, they should
4619 * - be delivered to SOCK_MEMALLOC sockets only
4620 * - stay away from userspace
4621 * - have bounded memory usage
4623 * Use PF_MEMALLOC as this saves us from propagating the allocation
4624 * context down to all allocation sites.
4626 noreclaim_flag = memalloc_noreclaim_save();
4627 ret = __netif_receive_skb_core(skb, true);
4628 memalloc_noreclaim_restore(noreclaim_flag);
4630 ret = __netif_receive_skb_core(skb, false);
4635 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp)
4637 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog);
4638 struct bpf_prog *new = xdp->prog;
4641 switch (xdp->command) {
4642 case XDP_SETUP_PROG:
4643 rcu_assign_pointer(dev->xdp_prog, new);
4648 static_key_slow_dec(&generic_xdp_needed);
4649 } else if (new && !old) {
4650 static_key_slow_inc(&generic_xdp_needed);
4651 dev_disable_lro(dev);
4652 dev_disable_gro_hw(dev);
4656 case XDP_QUERY_PROG:
4657 xdp->prog_attached = !!old;
4658 xdp->prog_id = old ? old->aux->id : 0;
4669 static int netif_receive_skb_internal(struct sk_buff *skb)
4673 net_timestamp_check(netdev_tstamp_prequeue, skb);
4675 if (skb_defer_rx_timestamp(skb))
4676 return NET_RX_SUCCESS;
4678 if (static_key_false(&generic_xdp_needed)) {
4683 ret = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
4687 if (ret != XDP_PASS)
4693 if (static_key_false(&rps_needed)) {
4694 struct rps_dev_flow voidflow, *rflow = &voidflow;
4695 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
4698 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4704 ret = __netif_receive_skb(skb);
4710 * netif_receive_skb - process receive buffer from network
4711 * @skb: buffer to process
4713 * netif_receive_skb() is the main receive data processing function.
4714 * It always succeeds. The buffer may be dropped during processing
4715 * for congestion control or by the protocol layers.
4717 * This function may only be called from softirq context and interrupts
4718 * should be enabled.
4720 * Return values (usually ignored):
4721 * NET_RX_SUCCESS: no congestion
4722 * NET_RX_DROP: packet was dropped
4724 int netif_receive_skb(struct sk_buff *skb)
4726 trace_netif_receive_skb_entry(skb);
4728 return netif_receive_skb_internal(skb);
4730 EXPORT_SYMBOL(netif_receive_skb);
4732 DEFINE_PER_CPU(struct work_struct, flush_works);
4734 /* Network device is going away, flush any packets still pending */
4735 static void flush_backlog(struct work_struct *work)
4737 struct sk_buff *skb, *tmp;
4738 struct softnet_data *sd;
4741 sd = this_cpu_ptr(&softnet_data);
4743 local_irq_disable();
4745 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) {
4746 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
4747 __skb_unlink(skb, &sd->input_pkt_queue);
4749 input_queue_head_incr(sd);
4755 skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
4756 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
4757 __skb_unlink(skb, &sd->process_queue);
4759 input_queue_head_incr(sd);
4765 static void flush_all_backlogs(void)
4771 for_each_online_cpu(cpu)
4772 queue_work_on(cpu, system_highpri_wq,
4773 per_cpu_ptr(&flush_works, cpu));
4775 for_each_online_cpu(cpu)
4776 flush_work(per_cpu_ptr(&flush_works, cpu));
4781 static int napi_gro_complete(struct sk_buff *skb)
4783 struct packet_offload *ptype;
4784 __be16 type = skb->protocol;
4785 struct list_head *head = &offload_base;
4788 BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
4790 if (NAPI_GRO_CB(skb)->count == 1) {
4791 skb_shinfo(skb)->gso_size = 0;
4796 list_for_each_entry_rcu(ptype, head, list) {
4797 if (ptype->type != type || !ptype->callbacks.gro_complete)
4800 err = ptype->callbacks.gro_complete(skb, 0);
4806 WARN_ON(&ptype->list == head);
4808 return NET_RX_SUCCESS;
4812 return netif_receive_skb_internal(skb);
4815 /* napi->gro_list contains packets ordered by age.
4816 * youngest packets at the head of it.
4817 * Complete skbs in reverse order to reduce latencies.
4819 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
4821 struct sk_buff *skb, *prev = NULL;
4823 /* scan list and build reverse chain */
4824 for (skb = napi->gro_list; skb != NULL; skb = skb->next) {
4829 for (skb = prev; skb; skb = prev) {
4832 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
4836 napi_gro_complete(skb);
4840 napi->gro_list = NULL;
4842 EXPORT_SYMBOL(napi_gro_flush);
4844 static void gro_list_prepare(struct napi_struct *napi, struct sk_buff *skb)
4847 unsigned int maclen = skb->dev->hard_header_len;
4848 u32 hash = skb_get_hash_raw(skb);
4850 for (p = napi->gro_list; p; p = p->next) {
4851 unsigned long diffs;
4853 NAPI_GRO_CB(p)->flush = 0;
4855 if (hash != skb_get_hash_raw(p)) {
4856 NAPI_GRO_CB(p)->same_flow = 0;
4860 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
4861 diffs |= p->vlan_tci ^ skb->vlan_tci;
4862 diffs |= skb_metadata_dst_cmp(p, skb);
4863 diffs |= skb_metadata_differs(p, skb);
4864 if (maclen == ETH_HLEN)
4865 diffs |= compare_ether_header(skb_mac_header(p),
4866 skb_mac_header(skb));
4868 diffs = memcmp(skb_mac_header(p),
4869 skb_mac_header(skb),
4871 NAPI_GRO_CB(p)->same_flow = !diffs;
4875 static void skb_gro_reset_offset(struct sk_buff *skb)
4877 const struct skb_shared_info *pinfo = skb_shinfo(skb);
4878 const skb_frag_t *frag0 = &pinfo->frags[0];
4880 NAPI_GRO_CB(skb)->data_offset = 0;
4881 NAPI_GRO_CB(skb)->frag0 = NULL;
4882 NAPI_GRO_CB(skb)->frag0_len = 0;
4884 if (skb_mac_header(skb) == skb_tail_pointer(skb) &&
4886 !PageHighMem(skb_frag_page(frag0))) {
4887 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
4888 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
4889 skb_frag_size(frag0),
4890 skb->end - skb->tail);
4894 static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
4896 struct skb_shared_info *pinfo = skb_shinfo(skb);
4898 BUG_ON(skb->end - skb->tail < grow);
4900 memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
4902 skb->data_len -= grow;
4905 pinfo->frags[0].page_offset += grow;
4906 skb_frag_size_sub(&pinfo->frags[0], grow);
4908 if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
4909 skb_frag_unref(skb, 0);
4910 memmove(pinfo->frags, pinfo->frags + 1,
4911 --pinfo->nr_frags * sizeof(pinfo->frags[0]));
4915 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
4917 struct sk_buff **pp = NULL;
4918 struct packet_offload *ptype;
4919 __be16 type = skb->protocol;
4920 struct list_head *head = &offload_base;
4922 enum gro_result ret;
4925 if (netif_elide_gro(skb->dev))
4928 gro_list_prepare(napi, skb);
4931 list_for_each_entry_rcu(ptype, head, list) {
4932 if (ptype->type != type || !ptype->callbacks.gro_receive)
4935 skb_set_network_header(skb, skb_gro_offset(skb));
4936 skb_reset_mac_len(skb);
4937 NAPI_GRO_CB(skb)->same_flow = 0;
4938 NAPI_GRO_CB(skb)->flush = skb_is_gso(skb) || skb_has_frag_list(skb);
4939 NAPI_GRO_CB(skb)->free = 0;
4940 NAPI_GRO_CB(skb)->encap_mark = 0;
4941 NAPI_GRO_CB(skb)->recursion_counter = 0;
4942 NAPI_GRO_CB(skb)->is_fou = 0;
4943 NAPI_GRO_CB(skb)->is_atomic = 1;
4944 NAPI_GRO_CB(skb)->gro_remcsum_start = 0;
4946 /* Setup for GRO checksum validation */
4947 switch (skb->ip_summed) {
4948 case CHECKSUM_COMPLETE:
4949 NAPI_GRO_CB(skb)->csum = skb->csum;
4950 NAPI_GRO_CB(skb)->csum_valid = 1;
4951 NAPI_GRO_CB(skb)->csum_cnt = 0;
4953 case CHECKSUM_UNNECESSARY:
4954 NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
4955 NAPI_GRO_CB(skb)->csum_valid = 0;
4958 NAPI_GRO_CB(skb)->csum_cnt = 0;
4959 NAPI_GRO_CB(skb)->csum_valid = 0;
4962 pp = ptype->callbacks.gro_receive(&napi->gro_list, skb);
4967 if (&ptype->list == head)
4970 if (IS_ERR(pp) && PTR_ERR(pp) == -EINPROGRESS) {
4975 same_flow = NAPI_GRO_CB(skb)->same_flow;
4976 ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
4979 struct sk_buff *nskb = *pp;
4983 napi_gro_complete(nskb);
4990 if (NAPI_GRO_CB(skb)->flush)
4993 if (unlikely(napi->gro_count >= MAX_GRO_SKBS)) {
4994 struct sk_buff *nskb = napi->gro_list;
4996 /* locate the end of the list to select the 'oldest' flow */
4997 while (nskb->next) {
5003 napi_gro_complete(nskb);
5007 NAPI_GRO_CB(skb)->count = 1;
5008 NAPI_GRO_CB(skb)->age = jiffies;
5009 NAPI_GRO_CB(skb)->last = skb;
5010 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
5011 skb->next = napi->gro_list;
5012 napi->gro_list = skb;
5016 grow = skb_gro_offset(skb) - skb_headlen(skb);
5018 gro_pull_from_frag0(skb, grow);
5027 struct packet_offload *gro_find_receive_by_type(__be16 type)
5029 struct list_head *offload_head = &offload_base;
5030 struct packet_offload *ptype;
5032 list_for_each_entry_rcu(ptype, offload_head, list) {
5033 if (ptype->type != type || !ptype->callbacks.gro_receive)
5039 EXPORT_SYMBOL(gro_find_receive_by_type);
5041 struct packet_offload *gro_find_complete_by_type(__be16 type)
5043 struct list_head *offload_head = &offload_base;
5044 struct packet_offload *ptype;
5046 list_for_each_entry_rcu(ptype, offload_head, list) {
5047 if (ptype->type != type || !ptype->callbacks.gro_complete)
5053 EXPORT_SYMBOL(gro_find_complete_by_type);
5055 static void napi_skb_free_stolen_head(struct sk_buff *skb)
5059 kmem_cache_free(skbuff_head_cache, skb);
5062 static gro_result_t napi_skb_finish(gro_result_t ret, struct sk_buff *skb)
5066 if (netif_receive_skb_internal(skb))
5074 case GRO_MERGED_FREE:
5075 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
5076 napi_skb_free_stolen_head(skb);
5090 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
5092 skb_mark_napi_id(skb, napi);
5093 trace_napi_gro_receive_entry(skb);
5095 skb_gro_reset_offset(skb);
5097 return napi_skb_finish(dev_gro_receive(napi, skb), skb);
5099 EXPORT_SYMBOL(napi_gro_receive);
5101 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
5103 if (unlikely(skb->pfmemalloc)) {
5107 __skb_pull(skb, skb_headlen(skb));
5108 /* restore the reserve we had after netdev_alloc_skb_ip_align() */
5109 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
5111 skb->dev = napi->dev;
5113 skb->encapsulation = 0;
5114 skb_shinfo(skb)->gso_type = 0;
5115 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
5121 struct sk_buff *napi_get_frags(struct napi_struct *napi)
5123 struct sk_buff *skb = napi->skb;
5126 skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
5129 skb_mark_napi_id(skb, napi);
5134 EXPORT_SYMBOL(napi_get_frags);
5136 static gro_result_t napi_frags_finish(struct napi_struct *napi,
5137 struct sk_buff *skb,
5143 __skb_push(skb, ETH_HLEN);
5144 skb->protocol = eth_type_trans(skb, skb->dev);
5145 if (ret == GRO_NORMAL && netif_receive_skb_internal(skb))
5150 napi_reuse_skb(napi, skb);
5153 case GRO_MERGED_FREE:
5154 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
5155 napi_skb_free_stolen_head(skb);
5157 napi_reuse_skb(napi, skb);
5168 /* Upper GRO stack assumes network header starts at gro_offset=0
5169 * Drivers could call both napi_gro_frags() and napi_gro_receive()
5170 * We copy ethernet header into skb->data to have a common layout.
5172 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
5174 struct sk_buff *skb = napi->skb;
5175 const struct ethhdr *eth;
5176 unsigned int hlen = sizeof(*eth);
5180 skb_reset_mac_header(skb);
5181 skb_gro_reset_offset(skb);
5183 eth = skb_gro_header_fast(skb, 0);
5184 if (unlikely(skb_gro_header_hard(skb, hlen))) {
5185 eth = skb_gro_header_slow(skb, hlen, 0);
5186 if (unlikely(!eth)) {
5187 net_warn_ratelimited("%s: dropping impossible skb from %s\n",
5188 __func__, napi->dev->name);
5189 napi_reuse_skb(napi, skb);
5193 gro_pull_from_frag0(skb, hlen);
5194 NAPI_GRO_CB(skb)->frag0 += hlen;
5195 NAPI_GRO_CB(skb)->frag0_len -= hlen;
5197 __skb_pull(skb, hlen);
5200 * This works because the only protocols we care about don't require
5202 * We'll fix it up properly in napi_frags_finish()
5204 skb->protocol = eth->h_proto;
5209 gro_result_t napi_gro_frags(struct napi_struct *napi)
5211 struct sk_buff *skb = napi_frags_skb(napi);
5216 trace_napi_gro_frags_entry(skb);
5218 return napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
5220 EXPORT_SYMBOL(napi_gro_frags);
5222 /* Compute the checksum from gro_offset and return the folded value
5223 * after adding in any pseudo checksum.
5225 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
5230 wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
5232 /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
5233 sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
5235 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
5236 !skb->csum_complete_sw)
5237 netdev_rx_csum_fault(skb->dev);
5240 NAPI_GRO_CB(skb)->csum = wsum;
5241 NAPI_GRO_CB(skb)->csum_valid = 1;
5245 EXPORT_SYMBOL(__skb_gro_checksum_complete);
5247 static void net_rps_send_ipi(struct softnet_data *remsd)
5251 struct softnet_data *next = remsd->rps_ipi_next;
5253 if (cpu_online(remsd->cpu))
5254 smp_call_function_single_async(remsd->cpu, &remsd->csd);
5261 * net_rps_action_and_irq_enable sends any pending IPI's for rps.
5262 * Note: called with local irq disabled, but exits with local irq enabled.
5264 static void net_rps_action_and_irq_enable(struct softnet_data *sd)
5267 struct softnet_data *remsd = sd->rps_ipi_list;
5270 sd->rps_ipi_list = NULL;
5274 /* Send pending IPI's to kick RPS processing on remote cpus. */
5275 net_rps_send_ipi(remsd);
5281 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd)
5284 return sd->rps_ipi_list != NULL;
5290 static int process_backlog(struct napi_struct *napi, int quota)
5292 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
5296 /* Check if we have pending ipi, its better to send them now,
5297 * not waiting net_rx_action() end.
5299 if (sd_has_rps_ipi_waiting(sd)) {
5300 local_irq_disable();
5301 net_rps_action_and_irq_enable(sd);
5304 napi->weight = dev_rx_weight;
5306 struct sk_buff *skb;
5308 while ((skb = __skb_dequeue(&sd->process_queue))) {
5310 __netif_receive_skb(skb);
5312 input_queue_head_incr(sd);
5313 if (++work >= quota)
5318 local_irq_disable();
5320 if (skb_queue_empty(&sd->input_pkt_queue)) {
5322 * Inline a custom version of __napi_complete().
5323 * only current cpu owns and manipulates this napi,
5324 * and NAPI_STATE_SCHED is the only possible flag set
5326 * We can use a plain write instead of clear_bit(),
5327 * and we dont need an smp_mb() memory barrier.
5332 skb_queue_splice_tail_init(&sd->input_pkt_queue,
5333 &sd->process_queue);
5343 * __napi_schedule - schedule for receive
5344 * @n: entry to schedule
5346 * The entry's receive function will be scheduled to run.
5347 * Consider using __napi_schedule_irqoff() if hard irqs are masked.
5349 void __napi_schedule(struct napi_struct *n)
5351 unsigned long flags;
5353 local_irq_save(flags);
5354 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
5355 local_irq_restore(flags);
5357 EXPORT_SYMBOL(__napi_schedule);
5360 * napi_schedule_prep - check if napi can be scheduled
5363 * Test if NAPI routine is already running, and if not mark
5364 * it as running. This is used as a condition variable
5365 * insure only one NAPI poll instance runs. We also make
5366 * sure there is no pending NAPI disable.
5368 bool napi_schedule_prep(struct napi_struct *n)
5370 unsigned long val, new;
5373 val = READ_ONCE(n->state);
5374 if (unlikely(val & NAPIF_STATE_DISABLE))
5376 new = val | NAPIF_STATE_SCHED;
5378 /* Sets STATE_MISSED bit if STATE_SCHED was already set
5379 * This was suggested by Alexander Duyck, as compiler
5380 * emits better code than :
5381 * if (val & NAPIF_STATE_SCHED)
5382 * new |= NAPIF_STATE_MISSED;
5384 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED *
5386 } while (cmpxchg(&n->state, val, new) != val);
5388 return !(val & NAPIF_STATE_SCHED);
5390 EXPORT_SYMBOL(napi_schedule_prep);
5393 * __napi_schedule_irqoff - schedule for receive
5394 * @n: entry to schedule
5396 * Variant of __napi_schedule() assuming hard irqs are masked
5398 void __napi_schedule_irqoff(struct napi_struct *n)
5400 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
5402 EXPORT_SYMBOL(__napi_schedule_irqoff);
5404 bool napi_complete_done(struct napi_struct *n, int work_done)
5406 unsigned long flags, val, new;
5409 * 1) Don't let napi dequeue from the cpu poll list
5410 * just in case its running on a different cpu.
5411 * 2) If we are busy polling, do nothing here, we have
5412 * the guarantee we will be called later.
5414 if (unlikely(n->state & (NAPIF_STATE_NPSVC |
5415 NAPIF_STATE_IN_BUSY_POLL)))
5419 unsigned long timeout = 0;
5422 timeout = n->dev->gro_flush_timeout;
5425 hrtimer_start(&n->timer, ns_to_ktime(timeout),
5426 HRTIMER_MODE_REL_PINNED);
5428 napi_gro_flush(n, false);
5430 if (unlikely(!list_empty(&n->poll_list))) {
5431 /* If n->poll_list is not empty, we need to mask irqs */
5432 local_irq_save(flags);
5433 list_del_init(&n->poll_list);
5434 local_irq_restore(flags);
5438 val = READ_ONCE(n->state);
5440 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED));
5442 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED);
5444 /* If STATE_MISSED was set, leave STATE_SCHED set,
5445 * because we will call napi->poll() one more time.
5446 * This C code was suggested by Alexander Duyck to help gcc.
5448 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED *
5450 } while (cmpxchg(&n->state, val, new) != val);
5452 if (unlikely(val & NAPIF_STATE_MISSED)) {
5459 EXPORT_SYMBOL(napi_complete_done);
5461 /* must be called under rcu_read_lock(), as we dont take a reference */
5462 static struct napi_struct *napi_by_id(unsigned int napi_id)
5464 unsigned int hash = napi_id % HASH_SIZE(napi_hash);
5465 struct napi_struct *napi;
5467 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node)
5468 if (napi->napi_id == napi_id)
5474 #if defined(CONFIG_NET_RX_BUSY_POLL)
5476 #define BUSY_POLL_BUDGET 8
5478 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock)
5482 /* Busy polling means there is a high chance device driver hard irq
5483 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was
5484 * set in napi_schedule_prep().
5485 * Since we are about to call napi->poll() once more, we can safely
5486 * clear NAPI_STATE_MISSED.
5488 * Note: x86 could use a single "lock and ..." instruction
5489 * to perform these two clear_bit()
5491 clear_bit(NAPI_STATE_MISSED, &napi->state);
5492 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state);
5496 /* All we really want here is to re-enable device interrupts.
5497 * Ideally, a new ndo_busy_poll_stop() could avoid another round.
5499 rc = napi->poll(napi, BUSY_POLL_BUDGET);
5500 trace_napi_poll(napi, rc, BUSY_POLL_BUDGET);
5501 netpoll_poll_unlock(have_poll_lock);
5502 if (rc == BUSY_POLL_BUDGET)
5503 __napi_schedule(napi);
5507 void napi_busy_loop(unsigned int napi_id,
5508 bool (*loop_end)(void *, unsigned long),
5511 unsigned long start_time = loop_end ? busy_loop_current_time() : 0;
5512 int (*napi_poll)(struct napi_struct *napi, int budget);
5513 void *have_poll_lock = NULL;
5514 struct napi_struct *napi;
5521 napi = napi_by_id(napi_id);
5531 unsigned long val = READ_ONCE(napi->state);
5533 /* If multiple threads are competing for this napi,
5534 * we avoid dirtying napi->state as much as we can.
5536 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED |
5537 NAPIF_STATE_IN_BUSY_POLL))
5539 if (cmpxchg(&napi->state, val,
5540 val | NAPIF_STATE_IN_BUSY_POLL |
5541 NAPIF_STATE_SCHED) != val)
5543 have_poll_lock = netpoll_poll_lock(napi);
5544 napi_poll = napi->poll;
5546 work = napi_poll(napi, BUSY_POLL_BUDGET);
5547 trace_napi_poll(napi, work, BUSY_POLL_BUDGET);
5550 __NET_ADD_STATS(dev_net(napi->dev),
5551 LINUX_MIB_BUSYPOLLRXPACKETS, work);
5554 if (!loop_end || loop_end(loop_end_arg, start_time))
5557 if (unlikely(need_resched())) {
5559 busy_poll_stop(napi, have_poll_lock);
5563 if (loop_end(loop_end_arg, start_time))
5570 busy_poll_stop(napi, have_poll_lock);
5575 EXPORT_SYMBOL(napi_busy_loop);
5577 #endif /* CONFIG_NET_RX_BUSY_POLL */
5579 static void napi_hash_add(struct napi_struct *napi)
5581 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state) ||
5582 test_and_set_bit(NAPI_STATE_HASHED, &napi->state))
5585 spin_lock(&napi_hash_lock);
5587 /* 0..NR_CPUS range is reserved for sender_cpu use */
5589 if (unlikely(++napi_gen_id < MIN_NAPI_ID))
5590 napi_gen_id = MIN_NAPI_ID;
5591 } while (napi_by_id(napi_gen_id));
5592 napi->napi_id = napi_gen_id;
5594 hlist_add_head_rcu(&napi->napi_hash_node,
5595 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]);
5597 spin_unlock(&napi_hash_lock);
5600 /* Warning : caller is responsible to make sure rcu grace period
5601 * is respected before freeing memory containing @napi
5603 bool napi_hash_del(struct napi_struct *napi)
5605 bool rcu_sync_needed = false;
5607 spin_lock(&napi_hash_lock);
5609 if (test_and_clear_bit(NAPI_STATE_HASHED, &napi->state)) {
5610 rcu_sync_needed = true;
5611 hlist_del_rcu(&napi->napi_hash_node);
5613 spin_unlock(&napi_hash_lock);
5614 return rcu_sync_needed;
5616 EXPORT_SYMBOL_GPL(napi_hash_del);
5618 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer)
5620 struct napi_struct *napi;
5622 napi = container_of(timer, struct napi_struct, timer);
5624 /* Note : we use a relaxed variant of napi_schedule_prep() not setting
5625 * NAPI_STATE_MISSED, since we do not react to a device IRQ.
5627 if (napi->gro_list && !napi_disable_pending(napi) &&
5628 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state))
5629 __napi_schedule_irqoff(napi);
5631 return HRTIMER_NORESTART;
5634 void netif_napi_add(struct net_device *dev, struct napi_struct *napi,
5635 int (*poll)(struct napi_struct *, int), int weight)
5637 INIT_LIST_HEAD(&napi->poll_list);
5638 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
5639 napi->timer.function = napi_watchdog;
5640 napi->gro_count = 0;
5641 napi->gro_list = NULL;
5644 if (weight > NAPI_POLL_WEIGHT)
5645 pr_err_once("netif_napi_add() called with weight %d on device %s\n",
5647 napi->weight = weight;
5648 list_add(&napi->dev_list, &dev->napi_list);
5650 #ifdef CONFIG_NETPOLL
5651 napi->poll_owner = -1;
5653 set_bit(NAPI_STATE_SCHED, &napi->state);
5654 napi_hash_add(napi);
5656 EXPORT_SYMBOL(netif_napi_add);
5658 void napi_disable(struct napi_struct *n)
5661 set_bit(NAPI_STATE_DISABLE, &n->state);
5663 while (test_and_set_bit(NAPI_STATE_SCHED, &n->state))
5665 while (test_and_set_bit(NAPI_STATE_NPSVC, &n->state))
5668 hrtimer_cancel(&n->timer);
5670 clear_bit(NAPI_STATE_DISABLE, &n->state);
5672 EXPORT_SYMBOL(napi_disable);
5674 /* Must be called in process context */
5675 void netif_napi_del(struct napi_struct *napi)
5678 if (napi_hash_del(napi))
5680 list_del_init(&napi->dev_list);
5681 napi_free_frags(napi);
5683 kfree_skb_list(napi->gro_list);
5684 napi->gro_list = NULL;
5685 napi->gro_count = 0;
5687 EXPORT_SYMBOL(netif_napi_del);
5689 static int napi_poll(struct napi_struct *n, struct list_head *repoll)
5694 list_del_init(&n->poll_list);
5696 have = netpoll_poll_lock(n);
5700 /* This NAPI_STATE_SCHED test is for avoiding a race
5701 * with netpoll's poll_napi(). Only the entity which
5702 * obtains the lock and sees NAPI_STATE_SCHED set will
5703 * actually make the ->poll() call. Therefore we avoid
5704 * accidentally calling ->poll() when NAPI is not scheduled.
5707 if (test_bit(NAPI_STATE_SCHED, &n->state)) {
5708 work = n->poll(n, weight);
5709 trace_napi_poll(n, work, weight);
5712 WARN_ON_ONCE(work > weight);
5714 if (likely(work < weight))
5717 /* Drivers must not modify the NAPI state if they
5718 * consume the entire weight. In such cases this code
5719 * still "owns" the NAPI instance and therefore can
5720 * move the instance around on the list at-will.
5722 if (unlikely(napi_disable_pending(n))) {
5728 /* flush too old packets
5729 * If HZ < 1000, flush all packets.
5731 napi_gro_flush(n, HZ >= 1000);
5734 /* Some drivers may have called napi_schedule
5735 * prior to exhausting their budget.
5737 if (unlikely(!list_empty(&n->poll_list))) {
5738 pr_warn_once("%s: Budget exhausted after napi rescheduled\n",
5739 n->dev ? n->dev->name : "backlog");
5743 list_add_tail(&n->poll_list, repoll);
5746 netpoll_poll_unlock(have);
5751 static __latent_entropy void net_rx_action(struct softirq_action *h)
5753 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
5754 unsigned long time_limit = jiffies +
5755 usecs_to_jiffies(netdev_budget_usecs);
5756 int budget = netdev_budget;
5760 local_irq_disable();
5761 list_splice_init(&sd->poll_list, &list);
5765 struct napi_struct *n;
5767 if (list_empty(&list)) {
5768 if (!sd_has_rps_ipi_waiting(sd) && list_empty(&repoll))
5773 n = list_first_entry(&list, struct napi_struct, poll_list);
5774 budget -= napi_poll(n, &repoll);
5776 /* If softirq window is exhausted then punt.
5777 * Allow this to run for 2 jiffies since which will allow
5778 * an average latency of 1.5/HZ.
5780 if (unlikely(budget <= 0 ||
5781 time_after_eq(jiffies, time_limit))) {
5787 local_irq_disable();
5789 list_splice_tail_init(&sd->poll_list, &list);
5790 list_splice_tail(&repoll, &list);
5791 list_splice(&list, &sd->poll_list);
5792 if (!list_empty(&sd->poll_list))
5793 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
5795 net_rps_action_and_irq_enable(sd);
5797 __kfree_skb_flush();
5800 struct netdev_adjacent {
5801 struct net_device *dev;
5803 /* upper master flag, there can only be one master device per list */
5806 /* counter for the number of times this device was added to us */
5809 /* private field for the users */
5812 struct list_head list;
5813 struct rcu_head rcu;
5816 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev,
5817 struct list_head *adj_list)
5819 struct netdev_adjacent *adj;
5821 list_for_each_entry(adj, adj_list, list) {
5822 if (adj->dev == adj_dev)
5828 static int __netdev_has_upper_dev(struct net_device *upper_dev, void *data)
5830 struct net_device *dev = data;
5832 return upper_dev == dev;
5836 * netdev_has_upper_dev - Check if device is linked to an upper device
5838 * @upper_dev: upper device to check
5840 * Find out if a device is linked to specified upper device and return true
5841 * in case it is. Note that this checks only immediate upper device,
5842 * not through a complete stack of devices. The caller must hold the RTNL lock.
5844 bool netdev_has_upper_dev(struct net_device *dev,
5845 struct net_device *upper_dev)
5849 return netdev_walk_all_upper_dev_rcu(dev, __netdev_has_upper_dev,
5852 EXPORT_SYMBOL(netdev_has_upper_dev);
5855 * netdev_has_upper_dev_all - Check if device is linked to an upper device
5857 * @upper_dev: upper device to check
5859 * Find out if a device is linked to specified upper device and return true
5860 * in case it is. Note that this checks the entire upper device chain.
5861 * The caller must hold rcu lock.
5864 bool netdev_has_upper_dev_all_rcu(struct net_device *dev,
5865 struct net_device *upper_dev)
5867 return !!netdev_walk_all_upper_dev_rcu(dev, __netdev_has_upper_dev,
5870 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu);
5873 * netdev_has_any_upper_dev - Check if device is linked to some device
5876 * Find out if a device is linked to an upper device and return true in case
5877 * it is. The caller must hold the RTNL lock.
5879 bool netdev_has_any_upper_dev(struct net_device *dev)
5883 return !list_empty(&dev->adj_list.upper);
5885 EXPORT_SYMBOL(netdev_has_any_upper_dev);
5888 * netdev_master_upper_dev_get - Get master upper device
5891 * Find a master upper device and return pointer to it or NULL in case
5892 * it's not there. The caller must hold the RTNL lock.
5894 struct net_device *netdev_master_upper_dev_get(struct net_device *dev)
5896 struct netdev_adjacent *upper;
5900 if (list_empty(&dev->adj_list.upper))
5903 upper = list_first_entry(&dev->adj_list.upper,
5904 struct netdev_adjacent, list);
5905 if (likely(upper->master))
5909 EXPORT_SYMBOL(netdev_master_upper_dev_get);
5912 * netdev_has_any_lower_dev - Check if device is linked to some device
5915 * Find out if a device is linked to a lower device and return true in case
5916 * it is. The caller must hold the RTNL lock.
5918 static bool netdev_has_any_lower_dev(struct net_device *dev)
5922 return !list_empty(&dev->adj_list.lower);
5925 void *netdev_adjacent_get_private(struct list_head *adj_list)
5927 struct netdev_adjacent *adj;
5929 adj = list_entry(adj_list, struct netdev_adjacent, list);
5931 return adj->private;
5933 EXPORT_SYMBOL(netdev_adjacent_get_private);
5936 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list
5938 * @iter: list_head ** of the current position
5940 * Gets the next device from the dev's upper list, starting from iter
5941 * position. The caller must hold RCU read lock.
5943 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev,
5944 struct list_head **iter)
5946 struct netdev_adjacent *upper;
5948 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
5950 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
5952 if (&upper->list == &dev->adj_list.upper)
5955 *iter = &upper->list;
5959 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu);
5961 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev,
5962 struct list_head **iter)
5964 struct netdev_adjacent *upper;
5966 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
5968 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
5970 if (&upper->list == &dev->adj_list.upper)
5973 *iter = &upper->list;
5978 int netdev_walk_all_upper_dev_rcu(struct net_device *dev,
5979 int (*fn)(struct net_device *dev,
5983 struct net_device *udev;
5984 struct list_head *iter;
5987 for (iter = &dev->adj_list.upper,
5988 udev = netdev_next_upper_dev_rcu(dev, &iter);
5990 udev = netdev_next_upper_dev_rcu(dev, &iter)) {
5991 /* first is the upper device itself */
5992 ret = fn(udev, data);
5996 /* then look at all of its upper devices */
5997 ret = netdev_walk_all_upper_dev_rcu(udev, fn, data);
6004 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu);
6007 * netdev_lower_get_next_private - Get the next ->private from the
6008 * lower neighbour list
6010 * @iter: list_head ** of the current position
6012 * Gets the next netdev_adjacent->private from the dev's lower neighbour
6013 * list, starting from iter position. The caller must hold either hold the
6014 * RTNL lock or its own locking that guarantees that the neighbour lower
6015 * list will remain unchanged.
6017 void *netdev_lower_get_next_private(struct net_device *dev,
6018 struct list_head **iter)
6020 struct netdev_adjacent *lower;
6022 lower = list_entry(*iter, struct netdev_adjacent, list);
6024 if (&lower->list == &dev->adj_list.lower)
6027 *iter = lower->list.next;
6029 return lower->private;
6031 EXPORT_SYMBOL(netdev_lower_get_next_private);
6034 * netdev_lower_get_next_private_rcu - Get the next ->private from the
6035 * lower neighbour list, RCU
6038 * @iter: list_head ** of the current position
6040 * Gets the next netdev_adjacent->private from the dev's lower neighbour
6041 * list, starting from iter position. The caller must hold RCU read lock.
6043 void *netdev_lower_get_next_private_rcu(struct net_device *dev,
6044 struct list_head **iter)
6046 struct netdev_adjacent *lower;
6048 WARN_ON_ONCE(!rcu_read_lock_held());
6050 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6052 if (&lower->list == &dev->adj_list.lower)
6055 *iter = &lower->list;
6057 return lower->private;
6059 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu);
6062 * netdev_lower_get_next - Get the next device from the lower neighbour
6065 * @iter: list_head ** of the current position
6067 * Gets the next netdev_adjacent from the dev's lower neighbour
6068 * list, starting from iter position. The caller must hold RTNL lock or
6069 * its own locking that guarantees that the neighbour lower
6070 * list will remain unchanged.
6072 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter)
6074 struct netdev_adjacent *lower;
6076 lower = list_entry(*iter, struct netdev_adjacent, list);
6078 if (&lower->list == &dev->adj_list.lower)
6081 *iter = lower->list.next;
6085 EXPORT_SYMBOL(netdev_lower_get_next);
6087 static struct net_device *netdev_next_lower_dev(struct net_device *dev,
6088 struct list_head **iter)
6090 struct netdev_adjacent *lower;
6092 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
6094 if (&lower->list == &dev->adj_list.lower)
6097 *iter = &lower->list;
6102 int netdev_walk_all_lower_dev(struct net_device *dev,
6103 int (*fn)(struct net_device *dev,
6107 struct net_device *ldev;
6108 struct list_head *iter;
6111 for (iter = &dev->adj_list.lower,
6112 ldev = netdev_next_lower_dev(dev, &iter);
6114 ldev = netdev_next_lower_dev(dev, &iter)) {
6115 /* first is the lower device itself */
6116 ret = fn(ldev, data);
6120 /* then look at all of its lower devices */
6121 ret = netdev_walk_all_lower_dev(ldev, fn, data);
6128 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev);
6130 static struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev,
6131 struct list_head **iter)
6133 struct netdev_adjacent *lower;
6135 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6136 if (&lower->list == &dev->adj_list.lower)
6139 *iter = &lower->list;
6144 int netdev_walk_all_lower_dev_rcu(struct net_device *dev,
6145 int (*fn)(struct net_device *dev,
6149 struct net_device *ldev;
6150 struct list_head *iter;
6153 for (iter = &dev->adj_list.lower,
6154 ldev = netdev_next_lower_dev_rcu(dev, &iter);
6156 ldev = netdev_next_lower_dev_rcu(dev, &iter)) {
6157 /* first is the lower device itself */
6158 ret = fn(ldev, data);
6162 /* then look at all of its lower devices */
6163 ret = netdev_walk_all_lower_dev_rcu(ldev, fn, data);
6170 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu);
6173 * netdev_lower_get_first_private_rcu - Get the first ->private from the
6174 * lower neighbour list, RCU
6178 * Gets the first netdev_adjacent->private from the dev's lower neighbour
6179 * list. The caller must hold RCU read lock.
6181 void *netdev_lower_get_first_private_rcu(struct net_device *dev)
6183 struct netdev_adjacent *lower;
6185 lower = list_first_or_null_rcu(&dev->adj_list.lower,
6186 struct netdev_adjacent, list);
6188 return lower->private;
6191 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu);
6194 * netdev_master_upper_dev_get_rcu - Get master upper device
6197 * Find a master upper device and return pointer to it or NULL in case
6198 * it's not there. The caller must hold the RCU read lock.
6200 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev)
6202 struct netdev_adjacent *upper;
6204 upper = list_first_or_null_rcu(&dev->adj_list.upper,
6205 struct netdev_adjacent, list);
6206 if (upper && likely(upper->master))
6210 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu);
6212 static int netdev_adjacent_sysfs_add(struct net_device *dev,
6213 struct net_device *adj_dev,
6214 struct list_head *dev_list)
6216 char linkname[IFNAMSIZ+7];
6218 sprintf(linkname, dev_list == &dev->adj_list.upper ?
6219 "upper_%s" : "lower_%s", adj_dev->name);
6220 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj),
6223 static void netdev_adjacent_sysfs_del(struct net_device *dev,
6225 struct list_head *dev_list)
6227 char linkname[IFNAMSIZ+7];
6229 sprintf(linkname, dev_list == &dev->adj_list.upper ?
6230 "upper_%s" : "lower_%s", name);
6231 sysfs_remove_link(&(dev->dev.kobj), linkname);
6234 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev,
6235 struct net_device *adj_dev,
6236 struct list_head *dev_list)
6238 return (dev_list == &dev->adj_list.upper ||
6239 dev_list == &dev->adj_list.lower) &&
6240 net_eq(dev_net(dev), dev_net(adj_dev));
6243 static int __netdev_adjacent_dev_insert(struct net_device *dev,
6244 struct net_device *adj_dev,
6245 struct list_head *dev_list,
6246 void *private, bool master)
6248 struct netdev_adjacent *adj;
6251 adj = __netdev_find_adj(adj_dev, dev_list);
6255 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n",
6256 dev->name, adj_dev->name, adj->ref_nr);
6261 adj = kmalloc(sizeof(*adj), GFP_KERNEL);
6266 adj->master = master;
6268 adj->private = private;
6271 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n",
6272 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name);
6274 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) {
6275 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list);
6280 /* Ensure that master link is always the first item in list. */
6282 ret = sysfs_create_link(&(dev->dev.kobj),
6283 &(adj_dev->dev.kobj), "master");
6285 goto remove_symlinks;
6287 list_add_rcu(&adj->list, dev_list);
6289 list_add_tail_rcu(&adj->list, dev_list);
6295 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
6296 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
6304 static void __netdev_adjacent_dev_remove(struct net_device *dev,
6305 struct net_device *adj_dev,
6307 struct list_head *dev_list)
6309 struct netdev_adjacent *adj;
6311 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n",
6312 dev->name, adj_dev->name, ref_nr);
6314 adj = __netdev_find_adj(adj_dev, dev_list);
6317 pr_err("Adjacency does not exist for device %s from %s\n",
6318 dev->name, adj_dev->name);
6323 if (adj->ref_nr > ref_nr) {
6324 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n",
6325 dev->name, adj_dev->name, ref_nr,
6326 adj->ref_nr - ref_nr);
6327 adj->ref_nr -= ref_nr;
6332 sysfs_remove_link(&(dev->dev.kobj), "master");
6334 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
6335 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
6337 list_del_rcu(&adj->list);
6338 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n",
6339 adj_dev->name, dev->name, adj_dev->name);
6341 kfree_rcu(adj, rcu);
6344 static int __netdev_adjacent_dev_link_lists(struct net_device *dev,
6345 struct net_device *upper_dev,
6346 struct list_head *up_list,
6347 struct list_head *down_list,
6348 void *private, bool master)
6352 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list,
6357 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list,
6360 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list);
6367 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev,
6368 struct net_device *upper_dev,
6370 struct list_head *up_list,
6371 struct list_head *down_list)
6373 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list);
6374 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list);
6377 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev,
6378 struct net_device *upper_dev,
6379 void *private, bool master)
6381 return __netdev_adjacent_dev_link_lists(dev, upper_dev,
6382 &dev->adj_list.upper,
6383 &upper_dev->adj_list.lower,
6387 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev,
6388 struct net_device *upper_dev)
6390 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1,
6391 &dev->adj_list.upper,
6392 &upper_dev->adj_list.lower);
6395 static int __netdev_upper_dev_link(struct net_device *dev,
6396 struct net_device *upper_dev, bool master,
6397 void *upper_priv, void *upper_info,
6398 struct netlink_ext_ack *extack)
6400 struct netdev_notifier_changeupper_info changeupper_info = {
6405 .upper_dev = upper_dev,
6408 .upper_info = upper_info,
6410 struct net_device *master_dev;
6415 if (dev == upper_dev)
6418 /* To prevent loops, check if dev is not upper device to upper_dev. */
6419 if (netdev_has_upper_dev(upper_dev, dev))
6423 if (netdev_has_upper_dev(dev, upper_dev))
6426 master_dev = netdev_master_upper_dev_get(dev);
6428 return master_dev == upper_dev ? -EEXIST : -EBUSY;
6431 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
6432 &changeupper_info.info);
6433 ret = notifier_to_errno(ret);
6437 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv,
6442 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
6443 &changeupper_info.info);
6444 ret = notifier_to_errno(ret);
6451 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
6457 * netdev_upper_dev_link - Add a link to the upper device
6459 * @upper_dev: new upper device
6460 * @extack: netlink extended ack
6462 * Adds a link to device which is upper to this one. The caller must hold
6463 * the RTNL lock. On a failure a negative errno code is returned.
6464 * On success the reference counts are adjusted and the function
6467 int netdev_upper_dev_link(struct net_device *dev,
6468 struct net_device *upper_dev,
6469 struct netlink_ext_ack *extack)
6471 return __netdev_upper_dev_link(dev, upper_dev, false,
6472 NULL, NULL, extack);
6474 EXPORT_SYMBOL(netdev_upper_dev_link);
6477 * netdev_master_upper_dev_link - Add a master link to the upper device
6479 * @upper_dev: new upper device
6480 * @upper_priv: upper device private
6481 * @upper_info: upper info to be passed down via notifier
6482 * @extack: netlink extended ack
6484 * Adds a link to device which is upper to this one. In this case, only
6485 * one master upper device can be linked, although other non-master devices
6486 * might be linked as well. The caller must hold the RTNL lock.
6487 * On a failure a negative errno code is returned. On success the reference
6488 * counts are adjusted and the function returns zero.
6490 int netdev_master_upper_dev_link(struct net_device *dev,
6491 struct net_device *upper_dev,
6492 void *upper_priv, void *upper_info,
6493 struct netlink_ext_ack *extack)
6495 return __netdev_upper_dev_link(dev, upper_dev, true,
6496 upper_priv, upper_info, extack);
6498 EXPORT_SYMBOL(netdev_master_upper_dev_link);
6501 * netdev_upper_dev_unlink - Removes a link to upper device
6503 * @upper_dev: new upper device
6505 * Removes a link to device which is upper to this one. The caller must hold
6508 void netdev_upper_dev_unlink(struct net_device *dev,
6509 struct net_device *upper_dev)
6511 struct netdev_notifier_changeupper_info changeupper_info = {
6515 .upper_dev = upper_dev,
6521 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev;
6523 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
6524 &changeupper_info.info);
6526 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
6528 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
6529 &changeupper_info.info);
6531 EXPORT_SYMBOL(netdev_upper_dev_unlink);
6534 * netdev_bonding_info_change - Dispatch event about slave change
6536 * @bonding_info: info to dispatch
6538 * Send NETDEV_BONDING_INFO to netdev notifiers with info.
6539 * The caller must hold the RTNL lock.
6541 void netdev_bonding_info_change(struct net_device *dev,
6542 struct netdev_bonding_info *bonding_info)
6544 struct netdev_notifier_bonding_info info = {
6548 memcpy(&info.bonding_info, bonding_info,
6549 sizeof(struct netdev_bonding_info));
6550 call_netdevice_notifiers_info(NETDEV_BONDING_INFO,
6553 EXPORT_SYMBOL(netdev_bonding_info_change);
6555 static void netdev_adjacent_add_links(struct net_device *dev)
6557 struct netdev_adjacent *iter;
6559 struct net *net = dev_net(dev);
6561 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6562 if (!net_eq(net, dev_net(iter->dev)))
6564 netdev_adjacent_sysfs_add(iter->dev, dev,
6565 &iter->dev->adj_list.lower);
6566 netdev_adjacent_sysfs_add(dev, iter->dev,
6567 &dev->adj_list.upper);
6570 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6571 if (!net_eq(net, dev_net(iter->dev)))
6573 netdev_adjacent_sysfs_add(iter->dev, dev,
6574 &iter->dev->adj_list.upper);
6575 netdev_adjacent_sysfs_add(dev, iter->dev,
6576 &dev->adj_list.lower);
6580 static void netdev_adjacent_del_links(struct net_device *dev)
6582 struct netdev_adjacent *iter;
6584 struct net *net = dev_net(dev);
6586 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6587 if (!net_eq(net, dev_net(iter->dev)))
6589 netdev_adjacent_sysfs_del(iter->dev, dev->name,
6590 &iter->dev->adj_list.lower);
6591 netdev_adjacent_sysfs_del(dev, iter->dev->name,
6592 &dev->adj_list.upper);
6595 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6596 if (!net_eq(net, dev_net(iter->dev)))
6598 netdev_adjacent_sysfs_del(iter->dev, dev->name,
6599 &iter->dev->adj_list.upper);
6600 netdev_adjacent_sysfs_del(dev, iter->dev->name,
6601 &dev->adj_list.lower);
6605 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname)
6607 struct netdev_adjacent *iter;
6609 struct net *net = dev_net(dev);
6611 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6612 if (!net_eq(net, dev_net(iter->dev)))
6614 netdev_adjacent_sysfs_del(iter->dev, oldname,
6615 &iter->dev->adj_list.lower);
6616 netdev_adjacent_sysfs_add(iter->dev, dev,
6617 &iter->dev->adj_list.lower);
6620 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6621 if (!net_eq(net, dev_net(iter->dev)))
6623 netdev_adjacent_sysfs_del(iter->dev, oldname,
6624 &iter->dev->adj_list.upper);
6625 netdev_adjacent_sysfs_add(iter->dev, dev,
6626 &iter->dev->adj_list.upper);
6630 void *netdev_lower_dev_get_private(struct net_device *dev,
6631 struct net_device *lower_dev)
6633 struct netdev_adjacent *lower;
6637 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower);
6641 return lower->private;
6643 EXPORT_SYMBOL(netdev_lower_dev_get_private);
6646 int dev_get_nest_level(struct net_device *dev)
6648 struct net_device *lower = NULL;
6649 struct list_head *iter;
6655 netdev_for_each_lower_dev(dev, lower, iter) {
6656 nest = dev_get_nest_level(lower);
6657 if (max_nest < nest)
6661 return max_nest + 1;
6663 EXPORT_SYMBOL(dev_get_nest_level);
6666 * netdev_lower_change - Dispatch event about lower device state change
6667 * @lower_dev: device
6668 * @lower_state_info: state to dispatch
6670 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info.
6671 * The caller must hold the RTNL lock.
6673 void netdev_lower_state_changed(struct net_device *lower_dev,
6674 void *lower_state_info)
6676 struct netdev_notifier_changelowerstate_info changelowerstate_info = {
6677 .info.dev = lower_dev,
6681 changelowerstate_info.lower_state_info = lower_state_info;
6682 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE,
6683 &changelowerstate_info.info);
6685 EXPORT_SYMBOL(netdev_lower_state_changed);
6687 static void dev_change_rx_flags(struct net_device *dev, int flags)
6689 const struct net_device_ops *ops = dev->netdev_ops;
6691 if (ops->ndo_change_rx_flags)
6692 ops->ndo_change_rx_flags(dev, flags);
6695 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify)
6697 unsigned int old_flags = dev->flags;
6703 dev->flags |= IFF_PROMISC;
6704 dev->promiscuity += inc;
6705 if (dev->promiscuity == 0) {
6708 * If inc causes overflow, untouch promisc and return error.
6711 dev->flags &= ~IFF_PROMISC;
6713 dev->promiscuity -= inc;
6714 pr_warn("%s: promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n",
6719 if (dev->flags != old_flags) {
6720 pr_info("device %s %s promiscuous mode\n",
6722 dev->flags & IFF_PROMISC ? "entered" : "left");
6723 if (audit_enabled) {
6724 current_uid_gid(&uid, &gid);
6725 audit_log(current->audit_context, GFP_ATOMIC,
6726 AUDIT_ANOM_PROMISCUOUS,
6727 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u",
6728 dev->name, (dev->flags & IFF_PROMISC),
6729 (old_flags & IFF_PROMISC),
6730 from_kuid(&init_user_ns, audit_get_loginuid(current)),
6731 from_kuid(&init_user_ns, uid),
6732 from_kgid(&init_user_ns, gid),
6733 audit_get_sessionid(current));
6736 dev_change_rx_flags(dev, IFF_PROMISC);
6739 __dev_notify_flags(dev, old_flags, IFF_PROMISC);
6744 * dev_set_promiscuity - update promiscuity count on a device
6748 * Add or remove promiscuity from a device. While the count in the device
6749 * remains above zero the interface remains promiscuous. Once it hits zero
6750 * the device reverts back to normal filtering operation. A negative inc
6751 * value is used to drop promiscuity on the device.
6752 * Return 0 if successful or a negative errno code on error.
6754 int dev_set_promiscuity(struct net_device *dev, int inc)
6756 unsigned int old_flags = dev->flags;
6759 err = __dev_set_promiscuity(dev, inc, true);
6762 if (dev->flags != old_flags)
6763 dev_set_rx_mode(dev);
6766 EXPORT_SYMBOL(dev_set_promiscuity);
6768 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify)
6770 unsigned int old_flags = dev->flags, old_gflags = dev->gflags;
6774 dev->flags |= IFF_ALLMULTI;
6775 dev->allmulti += inc;
6776 if (dev->allmulti == 0) {
6779 * If inc causes overflow, untouch allmulti and return error.
6782 dev->flags &= ~IFF_ALLMULTI;
6784 dev->allmulti -= inc;
6785 pr_warn("%s: allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n",
6790 if (dev->flags ^ old_flags) {
6791 dev_change_rx_flags(dev, IFF_ALLMULTI);
6792 dev_set_rx_mode(dev);
6794 __dev_notify_flags(dev, old_flags,
6795 dev->gflags ^ old_gflags);
6801 * dev_set_allmulti - update allmulti count on a device
6805 * Add or remove reception of all multicast frames to a device. While the
6806 * count in the device remains above zero the interface remains listening
6807 * to all interfaces. Once it hits zero the device reverts back to normal
6808 * filtering operation. A negative @inc value is used to drop the counter
6809 * when releasing a resource needing all multicasts.
6810 * Return 0 if successful or a negative errno code on error.
6813 int dev_set_allmulti(struct net_device *dev, int inc)
6815 return __dev_set_allmulti(dev, inc, true);
6817 EXPORT_SYMBOL(dev_set_allmulti);
6820 * Upload unicast and multicast address lists to device and
6821 * configure RX filtering. When the device doesn't support unicast
6822 * filtering it is put in promiscuous mode while unicast addresses
6825 void __dev_set_rx_mode(struct net_device *dev)
6827 const struct net_device_ops *ops = dev->netdev_ops;
6829 /* dev_open will call this function so the list will stay sane. */
6830 if (!(dev->flags&IFF_UP))
6833 if (!netif_device_present(dev))
6836 if (!(dev->priv_flags & IFF_UNICAST_FLT)) {
6837 /* Unicast addresses changes may only happen under the rtnl,
6838 * therefore calling __dev_set_promiscuity here is safe.
6840 if (!netdev_uc_empty(dev) && !dev->uc_promisc) {
6841 __dev_set_promiscuity(dev, 1, false);
6842 dev->uc_promisc = true;
6843 } else if (netdev_uc_empty(dev) && dev->uc_promisc) {
6844 __dev_set_promiscuity(dev, -1, false);
6845 dev->uc_promisc = false;
6849 if (ops->ndo_set_rx_mode)
6850 ops->ndo_set_rx_mode(dev);
6853 void dev_set_rx_mode(struct net_device *dev)
6855 netif_addr_lock_bh(dev);
6856 __dev_set_rx_mode(dev);
6857 netif_addr_unlock_bh(dev);
6861 * dev_get_flags - get flags reported to userspace
6864 * Get the combination of flag bits exported through APIs to userspace.
6866 unsigned int dev_get_flags(const struct net_device *dev)
6870 flags = (dev->flags & ~(IFF_PROMISC |
6875 (dev->gflags & (IFF_PROMISC |
6878 if (netif_running(dev)) {
6879 if (netif_oper_up(dev))
6880 flags |= IFF_RUNNING;
6881 if (netif_carrier_ok(dev))
6882 flags |= IFF_LOWER_UP;
6883 if (netif_dormant(dev))
6884 flags |= IFF_DORMANT;
6889 EXPORT_SYMBOL(dev_get_flags);
6891 int __dev_change_flags(struct net_device *dev, unsigned int flags)
6893 unsigned int old_flags = dev->flags;
6899 * Set the flags on our device.
6902 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP |
6903 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL |
6905 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC |
6909 * Load in the correct multicast list now the flags have changed.
6912 if ((old_flags ^ flags) & IFF_MULTICAST)
6913 dev_change_rx_flags(dev, IFF_MULTICAST);
6915 dev_set_rx_mode(dev);
6918 * Have we downed the interface. We handle IFF_UP ourselves
6919 * according to user attempts to set it, rather than blindly
6924 if ((old_flags ^ flags) & IFF_UP) {
6925 if (old_flags & IFF_UP)
6928 ret = __dev_open(dev);
6931 if ((flags ^ dev->gflags) & IFF_PROMISC) {
6932 int inc = (flags & IFF_PROMISC) ? 1 : -1;
6933 unsigned int old_flags = dev->flags;
6935 dev->gflags ^= IFF_PROMISC;
6937 if (__dev_set_promiscuity(dev, inc, false) >= 0)
6938 if (dev->flags != old_flags)
6939 dev_set_rx_mode(dev);
6942 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI
6943 * is important. Some (broken) drivers set IFF_PROMISC, when
6944 * IFF_ALLMULTI is requested not asking us and not reporting.
6946 if ((flags ^ dev->gflags) & IFF_ALLMULTI) {
6947 int inc = (flags & IFF_ALLMULTI) ? 1 : -1;
6949 dev->gflags ^= IFF_ALLMULTI;
6950 __dev_set_allmulti(dev, inc, false);
6956 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags,
6957 unsigned int gchanges)
6959 unsigned int changes = dev->flags ^ old_flags;
6962 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC);
6964 if (changes & IFF_UP) {
6965 if (dev->flags & IFF_UP)
6966 call_netdevice_notifiers(NETDEV_UP, dev);
6968 call_netdevice_notifiers(NETDEV_DOWN, dev);
6971 if (dev->flags & IFF_UP &&
6972 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) {
6973 struct netdev_notifier_change_info change_info = {
6977 .flags_changed = changes,
6980 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info);
6985 * dev_change_flags - change device settings
6987 * @flags: device state flags
6989 * Change settings on device based state flags. The flags are
6990 * in the userspace exported format.
6992 int dev_change_flags(struct net_device *dev, unsigned int flags)
6995 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags;
6997 ret = __dev_change_flags(dev, flags);
7001 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags);
7002 __dev_notify_flags(dev, old_flags, changes);
7005 EXPORT_SYMBOL(dev_change_flags);
7007 int __dev_set_mtu(struct net_device *dev, int new_mtu)
7009 const struct net_device_ops *ops = dev->netdev_ops;
7011 if (ops->ndo_change_mtu)
7012 return ops->ndo_change_mtu(dev, new_mtu);
7017 EXPORT_SYMBOL(__dev_set_mtu);
7020 * dev_set_mtu - Change maximum transfer unit
7022 * @new_mtu: new transfer unit
7024 * Change the maximum transfer size of the network device.
7026 int dev_set_mtu(struct net_device *dev, int new_mtu)
7030 if (new_mtu == dev->mtu)
7033 /* MTU must be positive, and in range */
7034 if (new_mtu < 0 || new_mtu < dev->min_mtu) {
7035 net_err_ratelimited("%s: Invalid MTU %d requested, hw min %d\n",
7036 dev->name, new_mtu, dev->min_mtu);
7040 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) {
7041 net_err_ratelimited("%s: Invalid MTU %d requested, hw max %d\n",
7042 dev->name, new_mtu, dev->max_mtu);
7046 if (!netif_device_present(dev))
7049 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev);
7050 err = notifier_to_errno(err);
7054 orig_mtu = dev->mtu;
7055 err = __dev_set_mtu(dev, new_mtu);
7058 err = call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
7059 err = notifier_to_errno(err);
7061 /* setting mtu back and notifying everyone again,
7062 * so that they have a chance to revert changes.
7064 __dev_set_mtu(dev, orig_mtu);
7065 call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
7070 EXPORT_SYMBOL(dev_set_mtu);
7073 * dev_change_tx_queue_len - Change TX queue length of a netdevice
7075 * @new_len: new tx queue length
7077 int dev_change_tx_queue_len(struct net_device *dev, unsigned long new_len)
7079 unsigned int orig_len = dev->tx_queue_len;
7082 if (new_len != (unsigned int)new_len)
7085 if (new_len != orig_len) {
7086 dev->tx_queue_len = new_len;
7087 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev);
7088 res = notifier_to_errno(res);
7091 "refused to change device tx_queue_len\n");
7092 dev->tx_queue_len = orig_len;
7095 return dev_qdisc_change_tx_queue_len(dev);
7102 * dev_set_group - Change group this device belongs to
7104 * @new_group: group this device should belong to
7106 void dev_set_group(struct net_device *dev, int new_group)
7108 dev->group = new_group;
7110 EXPORT_SYMBOL(dev_set_group);
7113 * dev_set_mac_address - Change Media Access Control Address
7117 * Change the hardware (MAC) address of the device
7119 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa)
7121 const struct net_device_ops *ops = dev->netdev_ops;
7124 if (!ops->ndo_set_mac_address)
7126 if (sa->sa_family != dev->type)
7128 if (!netif_device_present(dev))
7130 err = ops->ndo_set_mac_address(dev, sa);
7133 dev->addr_assign_type = NET_ADDR_SET;
7134 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
7135 add_device_randomness(dev->dev_addr, dev->addr_len);
7138 EXPORT_SYMBOL(dev_set_mac_address);
7141 * dev_change_carrier - Change device carrier
7143 * @new_carrier: new value
7145 * Change device carrier
7147 int dev_change_carrier(struct net_device *dev, bool new_carrier)
7149 const struct net_device_ops *ops = dev->netdev_ops;
7151 if (!ops->ndo_change_carrier)
7153 if (!netif_device_present(dev))
7155 return ops->ndo_change_carrier(dev, new_carrier);
7157 EXPORT_SYMBOL(dev_change_carrier);
7160 * dev_get_phys_port_id - Get device physical port ID
7164 * Get device physical port ID
7166 int dev_get_phys_port_id(struct net_device *dev,
7167 struct netdev_phys_item_id *ppid)
7169 const struct net_device_ops *ops = dev->netdev_ops;
7171 if (!ops->ndo_get_phys_port_id)
7173 return ops->ndo_get_phys_port_id(dev, ppid);
7175 EXPORT_SYMBOL(dev_get_phys_port_id);
7178 * dev_get_phys_port_name - Get device physical port name
7181 * @len: limit of bytes to copy to name
7183 * Get device physical port name
7185 int dev_get_phys_port_name(struct net_device *dev,
7186 char *name, size_t len)
7188 const struct net_device_ops *ops = dev->netdev_ops;
7190 if (!ops->ndo_get_phys_port_name)
7192 return ops->ndo_get_phys_port_name(dev, name, len);
7194 EXPORT_SYMBOL(dev_get_phys_port_name);
7197 * dev_change_proto_down - update protocol port state information
7199 * @proto_down: new value
7201 * This info can be used by switch drivers to set the phys state of the
7204 int dev_change_proto_down(struct net_device *dev, bool proto_down)
7206 const struct net_device_ops *ops = dev->netdev_ops;
7208 if (!ops->ndo_change_proto_down)
7210 if (!netif_device_present(dev))
7212 return ops->ndo_change_proto_down(dev, proto_down);
7214 EXPORT_SYMBOL(dev_change_proto_down);
7216 void __dev_xdp_query(struct net_device *dev, bpf_op_t bpf_op,
7217 struct netdev_bpf *xdp)
7219 memset(xdp, 0, sizeof(*xdp));
7220 xdp->command = XDP_QUERY_PROG;
7222 /* Query must always succeed. */
7223 WARN_ON(bpf_op(dev, xdp) < 0);
7226 static u8 __dev_xdp_attached(struct net_device *dev, bpf_op_t bpf_op)
7228 struct netdev_bpf xdp;
7230 __dev_xdp_query(dev, bpf_op, &xdp);
7232 return xdp.prog_attached;
7235 static int dev_xdp_install(struct net_device *dev, bpf_op_t bpf_op,
7236 struct netlink_ext_ack *extack, u32 flags,
7237 struct bpf_prog *prog)
7239 struct netdev_bpf xdp;
7241 memset(&xdp, 0, sizeof(xdp));
7242 if (flags & XDP_FLAGS_HW_MODE)
7243 xdp.command = XDP_SETUP_PROG_HW;
7245 xdp.command = XDP_SETUP_PROG;
7246 xdp.extack = extack;
7250 return bpf_op(dev, &xdp);
7253 static void dev_xdp_uninstall(struct net_device *dev)
7255 struct netdev_bpf xdp;
7258 /* Remove generic XDP */
7259 WARN_ON(dev_xdp_install(dev, generic_xdp_install, NULL, 0, NULL));
7261 /* Remove from the driver */
7262 ndo_bpf = dev->netdev_ops->ndo_bpf;
7266 __dev_xdp_query(dev, ndo_bpf, &xdp);
7267 if (xdp.prog_attached == XDP_ATTACHED_NONE)
7270 /* Program removal should always succeed */
7271 WARN_ON(dev_xdp_install(dev, ndo_bpf, NULL, xdp.prog_flags, NULL));
7275 * dev_change_xdp_fd - set or clear a bpf program for a device rx path
7277 * @extack: netlink extended ack
7278 * @fd: new program fd or negative value to clear
7279 * @flags: xdp-related flags
7281 * Set or clear a bpf program for a device
7283 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack,
7286 const struct net_device_ops *ops = dev->netdev_ops;
7287 struct bpf_prog *prog = NULL;
7288 bpf_op_t bpf_op, bpf_chk;
7293 bpf_op = bpf_chk = ops->ndo_bpf;
7294 if (!bpf_op && (flags & (XDP_FLAGS_DRV_MODE | XDP_FLAGS_HW_MODE)))
7296 if (!bpf_op || (flags & XDP_FLAGS_SKB_MODE))
7297 bpf_op = generic_xdp_install;
7298 if (bpf_op == bpf_chk)
7299 bpf_chk = generic_xdp_install;
7302 if (bpf_chk && __dev_xdp_attached(dev, bpf_chk))
7304 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) &&
7305 __dev_xdp_attached(dev, bpf_op))
7308 prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP,
7309 bpf_op == ops->ndo_bpf);
7311 return PTR_ERR(prog);
7313 if (!(flags & XDP_FLAGS_HW_MODE) &&
7314 bpf_prog_is_dev_bound(prog->aux)) {
7315 NL_SET_ERR_MSG(extack, "using device-bound program without HW_MODE flag is not supported");
7321 err = dev_xdp_install(dev, bpf_op, extack, flags, prog);
7322 if (err < 0 && prog)
7329 * dev_new_index - allocate an ifindex
7330 * @net: the applicable net namespace
7332 * Returns a suitable unique value for a new device interface
7333 * number. The caller must hold the rtnl semaphore or the
7334 * dev_base_lock to be sure it remains unique.
7336 static int dev_new_index(struct net *net)
7338 int ifindex = net->ifindex;
7343 if (!__dev_get_by_index(net, ifindex))
7344 return net->ifindex = ifindex;
7348 /* Delayed registration/unregisteration */
7349 static LIST_HEAD(net_todo_list);
7350 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq);
7352 static void net_set_todo(struct net_device *dev)
7354 list_add_tail(&dev->todo_list, &net_todo_list);
7355 dev_net(dev)->dev_unreg_count++;
7358 static void rollback_registered_many(struct list_head *head)
7360 struct net_device *dev, *tmp;
7361 LIST_HEAD(close_head);
7363 BUG_ON(dev_boot_phase);
7366 list_for_each_entry_safe(dev, tmp, head, unreg_list) {
7367 /* Some devices call without registering
7368 * for initialization unwind. Remove those
7369 * devices and proceed with the remaining.
7371 if (dev->reg_state == NETREG_UNINITIALIZED) {
7372 pr_debug("unregister_netdevice: device %s/%p never was registered\n",
7376 list_del(&dev->unreg_list);
7379 dev->dismantle = true;
7380 BUG_ON(dev->reg_state != NETREG_REGISTERED);
7383 /* If device is running, close it first. */
7384 list_for_each_entry(dev, head, unreg_list)
7385 list_add_tail(&dev->close_list, &close_head);
7386 dev_close_many(&close_head, true);
7388 list_for_each_entry(dev, head, unreg_list) {
7389 /* And unlink it from device chain. */
7390 unlist_netdevice(dev);
7392 dev->reg_state = NETREG_UNREGISTERING;
7394 flush_all_backlogs();
7398 list_for_each_entry(dev, head, unreg_list) {
7399 struct sk_buff *skb = NULL;
7401 /* Shutdown queueing discipline. */
7404 dev_xdp_uninstall(dev);
7406 /* Notify protocols, that we are about to destroy
7407 * this device. They should clean all the things.
7409 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
7411 if (!dev->rtnl_link_ops ||
7412 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
7413 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0,
7414 GFP_KERNEL, NULL, 0);
7417 * Flush the unicast and multicast chains
7422 if (dev->netdev_ops->ndo_uninit)
7423 dev->netdev_ops->ndo_uninit(dev);
7426 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL);
7428 /* Notifier chain MUST detach us all upper devices. */
7429 WARN_ON(netdev_has_any_upper_dev(dev));
7430 WARN_ON(netdev_has_any_lower_dev(dev));
7432 /* Remove entries from kobject tree */
7433 netdev_unregister_kobject(dev);
7435 /* Remove XPS queueing entries */
7436 netif_reset_xps_queues_gt(dev, 0);
7442 list_for_each_entry(dev, head, unreg_list)
7446 static void rollback_registered(struct net_device *dev)
7450 list_add(&dev->unreg_list, &single);
7451 rollback_registered_many(&single);
7455 static netdev_features_t netdev_sync_upper_features(struct net_device *lower,
7456 struct net_device *upper, netdev_features_t features)
7458 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
7459 netdev_features_t feature;
7462 for_each_netdev_feature(&upper_disables, feature_bit) {
7463 feature = __NETIF_F_BIT(feature_bit);
7464 if (!(upper->wanted_features & feature)
7465 && (features & feature)) {
7466 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n",
7467 &feature, upper->name);
7468 features &= ~feature;
7475 static void netdev_sync_lower_features(struct net_device *upper,
7476 struct net_device *lower, netdev_features_t features)
7478 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
7479 netdev_features_t feature;
7482 for_each_netdev_feature(&upper_disables, feature_bit) {
7483 feature = __NETIF_F_BIT(feature_bit);
7484 if (!(features & feature) && (lower->features & feature)) {
7485 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n",
7486 &feature, lower->name);
7487 lower->wanted_features &= ~feature;
7488 netdev_update_features(lower);
7490 if (unlikely(lower->features & feature))
7491 netdev_WARN(upper, "failed to disable %pNF on %s!\n",
7492 &feature, lower->name);
7497 static netdev_features_t netdev_fix_features(struct net_device *dev,
7498 netdev_features_t features)
7500 /* Fix illegal checksum combinations */
7501 if ((features & NETIF_F_HW_CSUM) &&
7502 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
7503 netdev_warn(dev, "mixed HW and IP checksum settings.\n");
7504 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
7507 /* TSO requires that SG is present as well. */
7508 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) {
7509 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n");
7510 features &= ~NETIF_F_ALL_TSO;
7513 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) &&
7514 !(features & NETIF_F_IP_CSUM)) {
7515 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n");
7516 features &= ~NETIF_F_TSO;
7517 features &= ~NETIF_F_TSO_ECN;
7520 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) &&
7521 !(features & NETIF_F_IPV6_CSUM)) {
7522 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n");
7523 features &= ~NETIF_F_TSO6;
7526 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */
7527 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO))
7528 features &= ~NETIF_F_TSO_MANGLEID;
7530 /* TSO ECN requires that TSO is present as well. */
7531 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN)
7532 features &= ~NETIF_F_TSO_ECN;
7534 /* Software GSO depends on SG. */
7535 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) {
7536 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n");
7537 features &= ~NETIF_F_GSO;
7540 /* GSO partial features require GSO partial be set */
7541 if ((features & dev->gso_partial_features) &&
7542 !(features & NETIF_F_GSO_PARTIAL)) {
7544 "Dropping partially supported GSO features since no GSO partial.\n");
7545 features &= ~dev->gso_partial_features;
7548 if (!(features & NETIF_F_RXCSUM)) {
7549 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet
7550 * successfully merged by hardware must also have the
7551 * checksum verified by hardware. If the user does not
7552 * want to enable RXCSUM, logically, we should disable GRO_HW.
7554 if (features & NETIF_F_GRO_HW) {
7555 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n");
7556 features &= ~NETIF_F_GRO_HW;
7560 /* LRO/HW-GRO features cannot be combined with RX-FCS */
7561 if (features & NETIF_F_RXFCS) {
7562 if (features & NETIF_F_LRO) {
7563 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n");
7564 features &= ~NETIF_F_LRO;
7567 if (features & NETIF_F_GRO_HW) {
7568 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n");
7569 features &= ~NETIF_F_GRO_HW;
7576 int __netdev_update_features(struct net_device *dev)
7578 struct net_device *upper, *lower;
7579 netdev_features_t features;
7580 struct list_head *iter;
7585 features = netdev_get_wanted_features(dev);
7587 if (dev->netdev_ops->ndo_fix_features)
7588 features = dev->netdev_ops->ndo_fix_features(dev, features);
7590 /* driver might be less strict about feature dependencies */
7591 features = netdev_fix_features(dev, features);
7593 /* some features can't be enabled if they're off an an upper device */
7594 netdev_for_each_upper_dev_rcu(dev, upper, iter)
7595 features = netdev_sync_upper_features(dev, upper, features);
7597 if (dev->features == features)
7600 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n",
7601 &dev->features, &features);
7603 if (dev->netdev_ops->ndo_set_features)
7604 err = dev->netdev_ops->ndo_set_features(dev, features);
7608 if (unlikely(err < 0)) {
7610 "set_features() failed (%d); wanted %pNF, left %pNF\n",
7611 err, &features, &dev->features);
7612 /* return non-0 since some features might have changed and
7613 * it's better to fire a spurious notification than miss it
7619 /* some features must be disabled on lower devices when disabled
7620 * on an upper device (think: bonding master or bridge)
7622 netdev_for_each_lower_dev(dev, lower, iter)
7623 netdev_sync_lower_features(dev, lower, features);
7626 netdev_features_t diff = features ^ dev->features;
7628 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) {
7629 /* udp_tunnel_{get,drop}_rx_info both need
7630 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the
7631 * device, or they won't do anything.
7632 * Thus we need to update dev->features
7633 * *before* calling udp_tunnel_get_rx_info,
7634 * but *after* calling udp_tunnel_drop_rx_info.
7636 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) {
7637 dev->features = features;
7638 udp_tunnel_get_rx_info(dev);
7640 udp_tunnel_drop_rx_info(dev);
7644 dev->features = features;
7647 return err < 0 ? 0 : 1;
7651 * netdev_update_features - recalculate device features
7652 * @dev: the device to check
7654 * Recalculate dev->features set and send notifications if it
7655 * has changed. Should be called after driver or hardware dependent
7656 * conditions might have changed that influence the features.
7658 void netdev_update_features(struct net_device *dev)
7660 if (__netdev_update_features(dev))
7661 netdev_features_change(dev);
7663 EXPORT_SYMBOL(netdev_update_features);
7666 * netdev_change_features - recalculate device features
7667 * @dev: the device to check
7669 * Recalculate dev->features set and send notifications even
7670 * if they have not changed. Should be called instead of
7671 * netdev_update_features() if also dev->vlan_features might
7672 * have changed to allow the changes to be propagated to stacked
7675 void netdev_change_features(struct net_device *dev)
7677 __netdev_update_features(dev);
7678 netdev_features_change(dev);
7680 EXPORT_SYMBOL(netdev_change_features);
7683 * netif_stacked_transfer_operstate - transfer operstate
7684 * @rootdev: the root or lower level device to transfer state from
7685 * @dev: the device to transfer operstate to
7687 * Transfer operational state from root to device. This is normally
7688 * called when a stacking relationship exists between the root
7689 * device and the device(a leaf device).
7691 void netif_stacked_transfer_operstate(const struct net_device *rootdev,
7692 struct net_device *dev)
7694 if (rootdev->operstate == IF_OPER_DORMANT)
7695 netif_dormant_on(dev);
7697 netif_dormant_off(dev);
7699 if (netif_carrier_ok(rootdev))
7700 netif_carrier_on(dev);
7702 netif_carrier_off(dev);
7704 EXPORT_SYMBOL(netif_stacked_transfer_operstate);
7706 static int netif_alloc_rx_queues(struct net_device *dev)
7708 unsigned int i, count = dev->num_rx_queues;
7709 struct netdev_rx_queue *rx;
7710 size_t sz = count * sizeof(*rx);
7715 rx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
7721 for (i = 0; i < count; i++) {
7724 /* XDP RX-queue setup */
7725 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i);
7732 /* Rollback successful reg's and free other resources */
7734 xdp_rxq_info_unreg(&rx[i].xdp_rxq);
7740 static void netif_free_rx_queues(struct net_device *dev)
7742 unsigned int i, count = dev->num_rx_queues;
7744 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */
7748 for (i = 0; i < count; i++)
7749 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq);
7754 static void netdev_init_one_queue(struct net_device *dev,
7755 struct netdev_queue *queue, void *_unused)
7757 /* Initialize queue lock */
7758 spin_lock_init(&queue->_xmit_lock);
7759 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type);
7760 queue->xmit_lock_owner = -1;
7761 netdev_queue_numa_node_write(queue, NUMA_NO_NODE);
7764 dql_init(&queue->dql, HZ);
7768 static void netif_free_tx_queues(struct net_device *dev)
7773 static int netif_alloc_netdev_queues(struct net_device *dev)
7775 unsigned int count = dev->num_tx_queues;
7776 struct netdev_queue *tx;
7777 size_t sz = count * sizeof(*tx);
7779 if (count < 1 || count > 0xffff)
7782 tx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
7788 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL);
7789 spin_lock_init(&dev->tx_global_lock);
7794 void netif_tx_stop_all_queues(struct net_device *dev)
7798 for (i = 0; i < dev->num_tx_queues; i++) {
7799 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
7801 netif_tx_stop_queue(txq);
7804 EXPORT_SYMBOL(netif_tx_stop_all_queues);
7807 * register_netdevice - register a network device
7808 * @dev: device to register
7810 * Take a completed network device structure and add it to the kernel
7811 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
7812 * chain. 0 is returned on success. A negative errno code is returned
7813 * on a failure to set up the device, or if the name is a duplicate.
7815 * Callers must hold the rtnl semaphore. You may want
7816 * register_netdev() instead of this.
7819 * The locking appears insufficient to guarantee two parallel registers
7820 * will not get the same name.
7823 int register_netdevice(struct net_device *dev)
7826 struct net *net = dev_net(dev);
7828 BUG_ON(dev_boot_phase);
7833 /* When net_device's are persistent, this will be fatal. */
7834 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
7837 spin_lock_init(&dev->addr_list_lock);
7838 netdev_set_addr_lockdep_class(dev);
7840 ret = dev_get_valid_name(net, dev, dev->name);
7844 /* Init, if this function is available */
7845 if (dev->netdev_ops->ndo_init) {
7846 ret = dev->netdev_ops->ndo_init(dev);
7854 if (((dev->hw_features | dev->features) &
7855 NETIF_F_HW_VLAN_CTAG_FILTER) &&
7856 (!dev->netdev_ops->ndo_vlan_rx_add_vid ||
7857 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) {
7858 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n");
7865 dev->ifindex = dev_new_index(net);
7866 else if (__dev_get_by_index(net, dev->ifindex))
7869 /* Transfer changeable features to wanted_features and enable
7870 * software offloads (GSO and GRO).
7872 dev->hw_features |= NETIF_F_SOFT_FEATURES;
7873 dev->features |= NETIF_F_SOFT_FEATURES;
7875 if (dev->netdev_ops->ndo_udp_tunnel_add) {
7876 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT;
7877 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT;
7880 dev->wanted_features = dev->features & dev->hw_features;
7882 if (!(dev->flags & IFF_LOOPBACK))
7883 dev->hw_features |= NETIF_F_NOCACHE_COPY;
7885 /* If IPv4 TCP segmentation offload is supported we should also
7886 * allow the device to enable segmenting the frame with the option
7887 * of ignoring a static IP ID value. This doesn't enable the
7888 * feature itself but allows the user to enable it later.
7890 if (dev->hw_features & NETIF_F_TSO)
7891 dev->hw_features |= NETIF_F_TSO_MANGLEID;
7892 if (dev->vlan_features & NETIF_F_TSO)
7893 dev->vlan_features |= NETIF_F_TSO_MANGLEID;
7894 if (dev->mpls_features & NETIF_F_TSO)
7895 dev->mpls_features |= NETIF_F_TSO_MANGLEID;
7896 if (dev->hw_enc_features & NETIF_F_TSO)
7897 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID;
7899 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices.
7901 dev->vlan_features |= NETIF_F_HIGHDMA;
7903 /* Make NETIF_F_SG inheritable to tunnel devices.
7905 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL;
7907 /* Make NETIF_F_SG inheritable to MPLS.
7909 dev->mpls_features |= NETIF_F_SG;
7911 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev);
7912 ret = notifier_to_errno(ret);
7916 ret = netdev_register_kobject(dev);
7919 dev->reg_state = NETREG_REGISTERED;
7921 __netdev_update_features(dev);
7924 * Default initial state at registry is that the
7925 * device is present.
7928 set_bit(__LINK_STATE_PRESENT, &dev->state);
7930 linkwatch_init_dev(dev);
7932 dev_init_scheduler(dev);
7934 list_netdevice(dev);
7935 add_device_randomness(dev->dev_addr, dev->addr_len);
7937 /* If the device has permanent device address, driver should
7938 * set dev_addr and also addr_assign_type should be set to
7939 * NET_ADDR_PERM (default value).
7941 if (dev->addr_assign_type == NET_ADDR_PERM)
7942 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len);
7944 /* Notify protocols, that a new device appeared. */
7945 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
7946 ret = notifier_to_errno(ret);
7948 rollback_registered(dev);
7949 dev->reg_state = NETREG_UNREGISTERED;
7952 * Prevent userspace races by waiting until the network
7953 * device is fully setup before sending notifications.
7955 if (!dev->rtnl_link_ops ||
7956 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
7957 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
7963 if (dev->netdev_ops->ndo_uninit)
7964 dev->netdev_ops->ndo_uninit(dev);
7965 if (dev->priv_destructor)
7966 dev->priv_destructor(dev);
7969 EXPORT_SYMBOL(register_netdevice);
7972 * init_dummy_netdev - init a dummy network device for NAPI
7973 * @dev: device to init
7975 * This takes a network device structure and initialize the minimum
7976 * amount of fields so it can be used to schedule NAPI polls without
7977 * registering a full blown interface. This is to be used by drivers
7978 * that need to tie several hardware interfaces to a single NAPI
7979 * poll scheduler due to HW limitations.
7981 int init_dummy_netdev(struct net_device *dev)
7983 /* Clear everything. Note we don't initialize spinlocks
7984 * are they aren't supposed to be taken by any of the
7985 * NAPI code and this dummy netdev is supposed to be
7986 * only ever used for NAPI polls
7988 memset(dev, 0, sizeof(struct net_device));
7990 /* make sure we BUG if trying to hit standard
7991 * register/unregister code path
7993 dev->reg_state = NETREG_DUMMY;
7995 /* NAPI wants this */
7996 INIT_LIST_HEAD(&dev->napi_list);
7998 /* a dummy interface is started by default */
7999 set_bit(__LINK_STATE_PRESENT, &dev->state);
8000 set_bit(__LINK_STATE_START, &dev->state);
8002 /* Note : We dont allocate pcpu_refcnt for dummy devices,
8003 * because users of this 'device' dont need to change
8009 EXPORT_SYMBOL_GPL(init_dummy_netdev);
8013 * register_netdev - register a network device
8014 * @dev: device to register
8016 * Take a completed network device structure and add it to the kernel
8017 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
8018 * chain. 0 is returned on success. A negative errno code is returned
8019 * on a failure to set up the device, or if the name is a duplicate.
8021 * This is a wrapper around register_netdevice that takes the rtnl semaphore
8022 * and expands the device name if you passed a format string to
8025 int register_netdev(struct net_device *dev)
8029 if (rtnl_lock_killable())
8031 err = register_netdevice(dev);
8035 EXPORT_SYMBOL(register_netdev);
8037 int netdev_refcnt_read(const struct net_device *dev)
8041 for_each_possible_cpu(i)
8042 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
8045 EXPORT_SYMBOL(netdev_refcnt_read);
8048 * netdev_wait_allrefs - wait until all references are gone.
8049 * @dev: target net_device
8051 * This is called when unregistering network devices.
8053 * Any protocol or device that holds a reference should register
8054 * for netdevice notification, and cleanup and put back the
8055 * reference if they receive an UNREGISTER event.
8056 * We can get stuck here if buggy protocols don't correctly
8059 static void netdev_wait_allrefs(struct net_device *dev)
8061 unsigned long rebroadcast_time, warning_time;
8064 linkwatch_forget_dev(dev);
8066 rebroadcast_time = warning_time = jiffies;
8067 refcnt = netdev_refcnt_read(dev);
8069 while (refcnt != 0) {
8070 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
8073 /* Rebroadcast unregister notification */
8074 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
8080 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
8081 if (test_bit(__LINK_STATE_LINKWATCH_PENDING,
8083 /* We must not have linkwatch events
8084 * pending on unregister. If this
8085 * happens, we simply run the queue
8086 * unscheduled, resulting in a noop
8089 linkwatch_run_queue();
8094 rebroadcast_time = jiffies;
8099 refcnt = netdev_refcnt_read(dev);
8101 if (time_after(jiffies, warning_time + 10 * HZ)) {
8102 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n",
8104 warning_time = jiffies;
8113 * register_netdevice(x1);
8114 * register_netdevice(x2);
8116 * unregister_netdevice(y1);
8117 * unregister_netdevice(y2);
8123 * We are invoked by rtnl_unlock().
8124 * This allows us to deal with problems:
8125 * 1) We can delete sysfs objects which invoke hotplug
8126 * without deadlocking with linkwatch via keventd.
8127 * 2) Since we run with the RTNL semaphore not held, we can sleep
8128 * safely in order to wait for the netdev refcnt to drop to zero.
8130 * We must not return until all unregister events added during
8131 * the interval the lock was held have been completed.
8133 void netdev_run_todo(void)
8135 struct list_head list;
8137 /* Snapshot list, allow later requests */
8138 list_replace_init(&net_todo_list, &list);
8143 /* Wait for rcu callbacks to finish before next phase */
8144 if (!list_empty(&list))
8147 while (!list_empty(&list)) {
8148 struct net_device *dev
8149 = list_first_entry(&list, struct net_device, todo_list);
8150 list_del(&dev->todo_list);
8153 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
8156 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) {
8157 pr_err("network todo '%s' but state %d\n",
8158 dev->name, dev->reg_state);
8163 dev->reg_state = NETREG_UNREGISTERED;
8165 netdev_wait_allrefs(dev);
8168 BUG_ON(netdev_refcnt_read(dev));
8169 BUG_ON(!list_empty(&dev->ptype_all));
8170 BUG_ON(!list_empty(&dev->ptype_specific));
8171 WARN_ON(rcu_access_pointer(dev->ip_ptr));
8172 WARN_ON(rcu_access_pointer(dev->ip6_ptr));
8173 #if IS_ENABLED(CONFIG_DECNET)
8174 WARN_ON(dev->dn_ptr);
8176 if (dev->priv_destructor)
8177 dev->priv_destructor(dev);
8178 if (dev->needs_free_netdev)
8181 /* Report a network device has been unregistered */
8183 dev_net(dev)->dev_unreg_count--;
8185 wake_up(&netdev_unregistering_wq);
8187 /* Free network device */
8188 kobject_put(&dev->dev.kobj);
8192 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has
8193 * all the same fields in the same order as net_device_stats, with only
8194 * the type differing, but rtnl_link_stats64 may have additional fields
8195 * at the end for newer counters.
8197 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
8198 const struct net_device_stats *netdev_stats)
8200 #if BITS_PER_LONG == 64
8201 BUILD_BUG_ON(sizeof(*stats64) < sizeof(*netdev_stats));
8202 memcpy(stats64, netdev_stats, sizeof(*netdev_stats));
8203 /* zero out counters that only exist in rtnl_link_stats64 */
8204 memset((char *)stats64 + sizeof(*netdev_stats), 0,
8205 sizeof(*stats64) - sizeof(*netdev_stats));
8207 size_t i, n = sizeof(*netdev_stats) / sizeof(unsigned long);
8208 const unsigned long *src = (const unsigned long *)netdev_stats;
8209 u64 *dst = (u64 *)stats64;
8211 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64));
8212 for (i = 0; i < n; i++)
8214 /* zero out counters that only exist in rtnl_link_stats64 */
8215 memset((char *)stats64 + n * sizeof(u64), 0,
8216 sizeof(*stats64) - n * sizeof(u64));
8219 EXPORT_SYMBOL(netdev_stats_to_stats64);
8222 * dev_get_stats - get network device statistics
8223 * @dev: device to get statistics from
8224 * @storage: place to store stats
8226 * Get network statistics from device. Return @storage.
8227 * The device driver may provide its own method by setting
8228 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats;
8229 * otherwise the internal statistics structure is used.
8231 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev,
8232 struct rtnl_link_stats64 *storage)
8234 const struct net_device_ops *ops = dev->netdev_ops;
8236 if (ops->ndo_get_stats64) {
8237 memset(storage, 0, sizeof(*storage));
8238 ops->ndo_get_stats64(dev, storage);
8239 } else if (ops->ndo_get_stats) {
8240 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev));
8242 netdev_stats_to_stats64(storage, &dev->stats);
8244 storage->rx_dropped += (unsigned long)atomic_long_read(&dev->rx_dropped);
8245 storage->tx_dropped += (unsigned long)atomic_long_read(&dev->tx_dropped);
8246 storage->rx_nohandler += (unsigned long)atomic_long_read(&dev->rx_nohandler);
8249 EXPORT_SYMBOL(dev_get_stats);
8251 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev)
8253 struct netdev_queue *queue = dev_ingress_queue(dev);
8255 #ifdef CONFIG_NET_CLS_ACT
8258 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
8261 netdev_init_one_queue(dev, queue, NULL);
8262 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc);
8263 queue->qdisc_sleeping = &noop_qdisc;
8264 rcu_assign_pointer(dev->ingress_queue, queue);
8269 static const struct ethtool_ops default_ethtool_ops;
8271 void netdev_set_default_ethtool_ops(struct net_device *dev,
8272 const struct ethtool_ops *ops)
8274 if (dev->ethtool_ops == &default_ethtool_ops)
8275 dev->ethtool_ops = ops;
8277 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops);
8279 void netdev_freemem(struct net_device *dev)
8281 char *addr = (char *)dev - dev->padded;
8287 * alloc_netdev_mqs - allocate network device
8288 * @sizeof_priv: size of private data to allocate space for
8289 * @name: device name format string
8290 * @name_assign_type: origin of device name
8291 * @setup: callback to initialize device
8292 * @txqs: the number of TX subqueues to allocate
8293 * @rxqs: the number of RX subqueues to allocate
8295 * Allocates a struct net_device with private data area for driver use
8296 * and performs basic initialization. Also allocates subqueue structs
8297 * for each queue on the device.
8299 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
8300 unsigned char name_assign_type,
8301 void (*setup)(struct net_device *),
8302 unsigned int txqs, unsigned int rxqs)
8304 struct net_device *dev;
8305 unsigned int alloc_size;
8306 struct net_device *p;
8308 BUG_ON(strlen(name) >= sizeof(dev->name));
8311 pr_err("alloc_netdev: Unable to allocate device with zero queues\n");
8316 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n");
8320 alloc_size = sizeof(struct net_device);
8322 /* ensure 32-byte alignment of private area */
8323 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
8324 alloc_size += sizeof_priv;
8326 /* ensure 32-byte alignment of whole construct */
8327 alloc_size += NETDEV_ALIGN - 1;
8329 p = kvzalloc(alloc_size, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
8333 dev = PTR_ALIGN(p, NETDEV_ALIGN);
8334 dev->padded = (char *)dev - (char *)p;
8336 dev->pcpu_refcnt = alloc_percpu(int);
8337 if (!dev->pcpu_refcnt)
8340 if (dev_addr_init(dev))
8346 dev_net_set(dev, &init_net);
8348 dev->gso_max_size = GSO_MAX_SIZE;
8349 dev->gso_max_segs = GSO_MAX_SEGS;
8351 INIT_LIST_HEAD(&dev->napi_list);
8352 INIT_LIST_HEAD(&dev->unreg_list);
8353 INIT_LIST_HEAD(&dev->close_list);
8354 INIT_LIST_HEAD(&dev->link_watch_list);
8355 INIT_LIST_HEAD(&dev->adj_list.upper);
8356 INIT_LIST_HEAD(&dev->adj_list.lower);
8357 INIT_LIST_HEAD(&dev->ptype_all);
8358 INIT_LIST_HEAD(&dev->ptype_specific);
8359 #ifdef CONFIG_NET_SCHED
8360 hash_init(dev->qdisc_hash);
8362 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM;
8365 if (!dev->tx_queue_len) {
8366 dev->priv_flags |= IFF_NO_QUEUE;
8367 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN;
8370 dev->num_tx_queues = txqs;
8371 dev->real_num_tx_queues = txqs;
8372 if (netif_alloc_netdev_queues(dev))
8375 dev->num_rx_queues = rxqs;
8376 dev->real_num_rx_queues = rxqs;
8377 if (netif_alloc_rx_queues(dev))
8380 strcpy(dev->name, name);
8381 dev->name_assign_type = name_assign_type;
8382 dev->group = INIT_NETDEV_GROUP;
8383 if (!dev->ethtool_ops)
8384 dev->ethtool_ops = &default_ethtool_ops;
8386 nf_hook_ingress_init(dev);
8395 free_percpu(dev->pcpu_refcnt);
8397 netdev_freemem(dev);
8400 EXPORT_SYMBOL(alloc_netdev_mqs);
8403 * free_netdev - free network device
8406 * This function does the last stage of destroying an allocated device
8407 * interface. The reference to the device object is released. If this
8408 * is the last reference then it will be freed.Must be called in process
8411 void free_netdev(struct net_device *dev)
8413 struct napi_struct *p, *n;
8416 netif_free_tx_queues(dev);
8417 netif_free_rx_queues(dev);
8419 kfree(rcu_dereference_protected(dev->ingress_queue, 1));
8421 /* Flush device addresses */
8422 dev_addr_flush(dev);
8424 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
8427 free_percpu(dev->pcpu_refcnt);
8428 dev->pcpu_refcnt = NULL;
8430 /* Compatibility with error handling in drivers */
8431 if (dev->reg_state == NETREG_UNINITIALIZED) {
8432 netdev_freemem(dev);
8436 BUG_ON(dev->reg_state != NETREG_UNREGISTERED);
8437 dev->reg_state = NETREG_RELEASED;
8439 /* will free via device release */
8440 put_device(&dev->dev);
8442 EXPORT_SYMBOL(free_netdev);
8445 * synchronize_net - Synchronize with packet receive processing
8447 * Wait for packets currently being received to be done.
8448 * Does not block later packets from starting.
8450 void synchronize_net(void)
8453 if (rtnl_is_locked())
8454 synchronize_rcu_expedited();
8458 EXPORT_SYMBOL(synchronize_net);
8461 * unregister_netdevice_queue - remove device from the kernel
8465 * This function shuts down a device interface and removes it
8466 * from the kernel tables.
8467 * If head not NULL, device is queued to be unregistered later.
8469 * Callers must hold the rtnl semaphore. You may want
8470 * unregister_netdev() instead of this.
8473 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head)
8478 list_move_tail(&dev->unreg_list, head);
8480 rollback_registered(dev);
8481 /* Finish processing unregister after unlock */
8485 EXPORT_SYMBOL(unregister_netdevice_queue);
8488 * unregister_netdevice_many - unregister many devices
8489 * @head: list of devices
8491 * Note: As most callers use a stack allocated list_head,
8492 * we force a list_del() to make sure stack wont be corrupted later.
8494 void unregister_netdevice_many(struct list_head *head)
8496 struct net_device *dev;
8498 if (!list_empty(head)) {
8499 rollback_registered_many(head);
8500 list_for_each_entry(dev, head, unreg_list)
8505 EXPORT_SYMBOL(unregister_netdevice_many);
8508 * unregister_netdev - remove device from the kernel
8511 * This function shuts down a device interface and removes it
8512 * from the kernel tables.
8514 * This is just a wrapper for unregister_netdevice that takes
8515 * the rtnl semaphore. In general you want to use this and not
8516 * unregister_netdevice.
8518 void unregister_netdev(struct net_device *dev)
8521 unregister_netdevice(dev);
8524 EXPORT_SYMBOL(unregister_netdev);
8527 * dev_change_net_namespace - move device to different nethost namespace
8529 * @net: network namespace
8530 * @pat: If not NULL name pattern to try if the current device name
8531 * is already taken in the destination network namespace.
8533 * This function shuts down a device interface and moves it
8534 * to a new network namespace. On success 0 is returned, on
8535 * a failure a netagive errno code is returned.
8537 * Callers must hold the rtnl semaphore.
8540 int dev_change_net_namespace(struct net_device *dev, struct net *net, const char *pat)
8542 int err, new_nsid, new_ifindex;
8546 /* Don't allow namespace local devices to be moved. */
8548 if (dev->features & NETIF_F_NETNS_LOCAL)
8551 /* Ensure the device has been registrered */
8552 if (dev->reg_state != NETREG_REGISTERED)
8555 /* Get out if there is nothing todo */
8557 if (net_eq(dev_net(dev), net))
8560 /* Pick the destination device name, and ensure
8561 * we can use it in the destination network namespace.
8564 if (__dev_get_by_name(net, dev->name)) {
8565 /* We get here if we can't use the current device name */
8568 if (dev_get_valid_name(net, dev, pat) < 0)
8573 * And now a mini version of register_netdevice unregister_netdevice.
8576 /* If device is running close it first. */
8579 /* And unlink it from device chain */
8581 unlist_netdevice(dev);
8585 /* Shutdown queueing discipline. */
8588 /* Notify protocols, that we are about to destroy
8589 * this device. They should clean all the things.
8591 * Note that dev->reg_state stays at NETREG_REGISTERED.
8592 * This is wanted because this way 8021q and macvlan know
8593 * the device is just moving and can keep their slaves up.
8595 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
8597 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
8599 new_nsid = peernet2id_alloc(dev_net(dev), net);
8600 /* If there is an ifindex conflict assign a new one */
8601 if (__dev_get_by_index(net, dev->ifindex))
8602 new_ifindex = dev_new_index(net);
8604 new_ifindex = dev->ifindex;
8606 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid,
8610 * Flush the unicast and multicast chains
8615 /* Send a netdev-removed uevent to the old namespace */
8616 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE);
8617 netdev_adjacent_del_links(dev);
8619 /* Actually switch the network namespace */
8620 dev_net_set(dev, net);
8621 dev->ifindex = new_ifindex;
8623 /* Send a netdev-add uevent to the new namespace */
8624 kobject_uevent(&dev->dev.kobj, KOBJ_ADD);
8625 netdev_adjacent_add_links(dev);
8627 /* Fixup kobjects */
8628 err = device_rename(&dev->dev, dev->name);
8631 /* Add the device back in the hashes */
8632 list_netdevice(dev);
8634 /* Notify protocols, that a new device appeared. */
8635 call_netdevice_notifiers(NETDEV_REGISTER, dev);
8638 * Prevent userspace races by waiting until the network
8639 * device is fully setup before sending notifications.
8641 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
8648 EXPORT_SYMBOL_GPL(dev_change_net_namespace);
8650 static int dev_cpu_dead(unsigned int oldcpu)
8652 struct sk_buff **list_skb;
8653 struct sk_buff *skb;
8655 struct softnet_data *sd, *oldsd, *remsd = NULL;
8657 local_irq_disable();
8658 cpu = smp_processor_id();
8659 sd = &per_cpu(softnet_data, cpu);
8660 oldsd = &per_cpu(softnet_data, oldcpu);
8662 /* Find end of our completion_queue. */
8663 list_skb = &sd->completion_queue;
8665 list_skb = &(*list_skb)->next;
8666 /* Append completion queue from offline CPU. */
8667 *list_skb = oldsd->completion_queue;
8668 oldsd->completion_queue = NULL;
8670 /* Append output queue from offline CPU. */
8671 if (oldsd->output_queue) {
8672 *sd->output_queue_tailp = oldsd->output_queue;
8673 sd->output_queue_tailp = oldsd->output_queue_tailp;
8674 oldsd->output_queue = NULL;
8675 oldsd->output_queue_tailp = &oldsd->output_queue;
8677 /* Append NAPI poll list from offline CPU, with one exception :
8678 * process_backlog() must be called by cpu owning percpu backlog.
8679 * We properly handle process_queue & input_pkt_queue later.
8681 while (!list_empty(&oldsd->poll_list)) {
8682 struct napi_struct *napi = list_first_entry(&oldsd->poll_list,
8686 list_del_init(&napi->poll_list);
8687 if (napi->poll == process_backlog)
8690 ____napi_schedule(sd, napi);
8693 raise_softirq_irqoff(NET_TX_SOFTIRQ);
8697 remsd = oldsd->rps_ipi_list;
8698 oldsd->rps_ipi_list = NULL;
8700 /* send out pending IPI's on offline CPU */
8701 net_rps_send_ipi(remsd);
8703 /* Process offline CPU's input_pkt_queue */
8704 while ((skb = __skb_dequeue(&oldsd->process_queue))) {
8706 input_queue_head_incr(oldsd);
8708 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) {
8710 input_queue_head_incr(oldsd);
8717 * netdev_increment_features - increment feature set by one
8718 * @all: current feature set
8719 * @one: new feature set
8720 * @mask: mask feature set
8722 * Computes a new feature set after adding a device with feature set
8723 * @one to the master device with current feature set @all. Will not
8724 * enable anything that is off in @mask. Returns the new feature set.
8726 netdev_features_t netdev_increment_features(netdev_features_t all,
8727 netdev_features_t one, netdev_features_t mask)
8729 if (mask & NETIF_F_HW_CSUM)
8730 mask |= NETIF_F_CSUM_MASK;
8731 mask |= NETIF_F_VLAN_CHALLENGED;
8733 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask;
8734 all &= one | ~NETIF_F_ALL_FOR_ALL;
8736 /* If one device supports hw checksumming, set for all. */
8737 if (all & NETIF_F_HW_CSUM)
8738 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM);
8742 EXPORT_SYMBOL(netdev_increment_features);
8744 static struct hlist_head * __net_init netdev_create_hash(void)
8747 struct hlist_head *hash;
8749 hash = kmalloc(sizeof(*hash) * NETDEV_HASHENTRIES, GFP_KERNEL);
8751 for (i = 0; i < NETDEV_HASHENTRIES; i++)
8752 INIT_HLIST_HEAD(&hash[i]);
8757 /* Initialize per network namespace state */
8758 static int __net_init netdev_init(struct net *net)
8760 if (net != &init_net)
8761 INIT_LIST_HEAD(&net->dev_base_head);
8763 net->dev_name_head = netdev_create_hash();
8764 if (net->dev_name_head == NULL)
8767 net->dev_index_head = netdev_create_hash();
8768 if (net->dev_index_head == NULL)
8774 kfree(net->dev_name_head);
8780 * netdev_drivername - network driver for the device
8781 * @dev: network device
8783 * Determine network driver for device.
8785 const char *netdev_drivername(const struct net_device *dev)
8787 const struct device_driver *driver;
8788 const struct device *parent;
8789 const char *empty = "";
8791 parent = dev->dev.parent;
8795 driver = parent->driver;
8796 if (driver && driver->name)
8797 return driver->name;
8801 static void __netdev_printk(const char *level, const struct net_device *dev,
8802 struct va_format *vaf)
8804 if (dev && dev->dev.parent) {
8805 dev_printk_emit(level[1] - '0',
8808 dev_driver_string(dev->dev.parent),
8809 dev_name(dev->dev.parent),
8810 netdev_name(dev), netdev_reg_state(dev),
8813 printk("%s%s%s: %pV",
8814 level, netdev_name(dev), netdev_reg_state(dev), vaf);
8816 printk("%s(NULL net_device): %pV", level, vaf);
8820 void netdev_printk(const char *level, const struct net_device *dev,
8821 const char *format, ...)
8823 struct va_format vaf;
8826 va_start(args, format);
8831 __netdev_printk(level, dev, &vaf);
8835 EXPORT_SYMBOL(netdev_printk);
8837 #define define_netdev_printk_level(func, level) \
8838 void func(const struct net_device *dev, const char *fmt, ...) \
8840 struct va_format vaf; \
8843 va_start(args, fmt); \
8848 __netdev_printk(level, dev, &vaf); \
8852 EXPORT_SYMBOL(func);
8854 define_netdev_printk_level(netdev_emerg, KERN_EMERG);
8855 define_netdev_printk_level(netdev_alert, KERN_ALERT);
8856 define_netdev_printk_level(netdev_crit, KERN_CRIT);
8857 define_netdev_printk_level(netdev_err, KERN_ERR);
8858 define_netdev_printk_level(netdev_warn, KERN_WARNING);
8859 define_netdev_printk_level(netdev_notice, KERN_NOTICE);
8860 define_netdev_printk_level(netdev_info, KERN_INFO);
8862 static void __net_exit netdev_exit(struct net *net)
8864 kfree(net->dev_name_head);
8865 kfree(net->dev_index_head);
8866 if (net != &init_net)
8867 WARN_ON_ONCE(!list_empty(&net->dev_base_head));
8870 static struct pernet_operations __net_initdata netdev_net_ops = {
8871 .init = netdev_init,
8872 .exit = netdev_exit,
8876 static void __net_exit default_device_exit(struct net *net)
8878 struct net_device *dev, *aux;
8880 * Push all migratable network devices back to the
8881 * initial network namespace
8884 for_each_netdev_safe(net, dev, aux) {
8886 char fb_name[IFNAMSIZ];
8888 /* Ignore unmoveable devices (i.e. loopback) */
8889 if (dev->features & NETIF_F_NETNS_LOCAL)
8892 /* Leave virtual devices for the generic cleanup */
8893 if (dev->rtnl_link_ops)
8896 /* Push remaining network devices to init_net */
8897 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex);
8898 err = dev_change_net_namespace(dev, &init_net, fb_name);
8900 pr_emerg("%s: failed to move %s to init_net: %d\n",
8901 __func__, dev->name, err);
8908 static void __net_exit rtnl_lock_unregistering(struct list_head *net_list)
8910 /* Return with the rtnl_lock held when there are no network
8911 * devices unregistering in any network namespace in net_list.
8915 DEFINE_WAIT_FUNC(wait, woken_wake_function);
8917 add_wait_queue(&netdev_unregistering_wq, &wait);
8919 unregistering = false;
8921 list_for_each_entry(net, net_list, exit_list) {
8922 if (net->dev_unreg_count > 0) {
8923 unregistering = true;
8931 wait_woken(&wait, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
8933 remove_wait_queue(&netdev_unregistering_wq, &wait);
8936 static void __net_exit default_device_exit_batch(struct list_head *net_list)
8938 /* At exit all network devices most be removed from a network
8939 * namespace. Do this in the reverse order of registration.
8940 * Do this across as many network namespaces as possible to
8941 * improve batching efficiency.
8943 struct net_device *dev;
8945 LIST_HEAD(dev_kill_list);
8947 /* To prevent network device cleanup code from dereferencing
8948 * loopback devices or network devices that have been freed
8949 * wait here for all pending unregistrations to complete,
8950 * before unregistring the loopback device and allowing the
8951 * network namespace be freed.
8953 * The netdev todo list containing all network devices
8954 * unregistrations that happen in default_device_exit_batch
8955 * will run in the rtnl_unlock() at the end of
8956 * default_device_exit_batch.
8958 rtnl_lock_unregistering(net_list);
8959 list_for_each_entry(net, net_list, exit_list) {
8960 for_each_netdev_reverse(net, dev) {
8961 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink)
8962 dev->rtnl_link_ops->dellink(dev, &dev_kill_list);
8964 unregister_netdevice_queue(dev, &dev_kill_list);
8967 unregister_netdevice_many(&dev_kill_list);
8971 static struct pernet_operations __net_initdata default_device_ops = {
8972 .exit = default_device_exit,
8973 .exit_batch = default_device_exit_batch,
8978 * Initialize the DEV module. At boot time this walks the device list and
8979 * unhooks any devices that fail to initialise (normally hardware not
8980 * present) and leaves us with a valid list of present and active devices.
8985 * This is called single threaded during boot, so no need
8986 * to take the rtnl semaphore.
8988 static int __init net_dev_init(void)
8990 int i, rc = -ENOMEM;
8992 BUG_ON(!dev_boot_phase);
8994 if (dev_proc_init())
8997 if (netdev_kobject_init())
9000 INIT_LIST_HEAD(&ptype_all);
9001 for (i = 0; i < PTYPE_HASH_SIZE; i++)
9002 INIT_LIST_HEAD(&ptype_base[i]);
9004 INIT_LIST_HEAD(&offload_base);
9006 if (register_pernet_subsys(&netdev_net_ops))
9010 * Initialise the packet receive queues.
9013 for_each_possible_cpu(i) {
9014 struct work_struct *flush = per_cpu_ptr(&flush_works, i);
9015 struct softnet_data *sd = &per_cpu(softnet_data, i);
9017 INIT_WORK(flush, flush_backlog);
9019 skb_queue_head_init(&sd->input_pkt_queue);
9020 skb_queue_head_init(&sd->process_queue);
9021 #ifdef CONFIG_XFRM_OFFLOAD
9022 skb_queue_head_init(&sd->xfrm_backlog);
9024 INIT_LIST_HEAD(&sd->poll_list);
9025 sd->output_queue_tailp = &sd->output_queue;
9027 sd->csd.func = rps_trigger_softirq;
9032 sd->backlog.poll = process_backlog;
9033 sd->backlog.weight = weight_p;
9038 /* The loopback device is special if any other network devices
9039 * is present in a network namespace the loopback device must
9040 * be present. Since we now dynamically allocate and free the
9041 * loopback device ensure this invariant is maintained by
9042 * keeping the loopback device as the first device on the
9043 * list of network devices. Ensuring the loopback devices
9044 * is the first device that appears and the last network device
9047 if (register_pernet_device(&loopback_net_ops))
9050 if (register_pernet_device(&default_device_ops))
9053 open_softirq(NET_TX_SOFTIRQ, net_tx_action);
9054 open_softirq(NET_RX_SOFTIRQ, net_rx_action);
9056 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead",
9057 NULL, dev_cpu_dead);
9064 subsys_initcall(net_dev_init);