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 <asm/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/mutex.h>
85 #include <linux/string.h>
87 #include <linux/socket.h>
88 #include <linux/sockios.h>
89 #include <linux/errno.h>
90 #include <linux/interrupt.h>
91 #include <linux/if_ether.h>
92 #include <linux/netdevice.h>
93 #include <linux/etherdevice.h>
94 #include <linux/ethtool.h>
95 #include <linux/notifier.h>
96 #include <linux/skbuff.h>
97 #include <net/net_namespace.h>
99 #include <linux/rtnetlink.h>
100 #include <linux/stat.h>
102 #include <net/pkt_sched.h>
103 #include <net/checksum.h>
104 #include <net/xfrm.h>
105 #include <linux/highmem.h>
106 #include <linux/init.h>
107 #include <linux/module.h>
108 #include <linux/netpoll.h>
109 #include <linux/rcupdate.h>
110 #include <linux/delay.h>
111 #include <net/iw_handler.h>
112 #include <asm/current.h>
113 #include <linux/audit.h>
114 #include <linux/dmaengine.h>
115 #include <linux/err.h>
116 #include <linux/ctype.h>
117 #include <linux/if_arp.h>
118 #include <linux/if_vlan.h>
119 #include <linux/ip.h>
121 #include <linux/ipv6.h>
122 #include <linux/in.h>
123 #include <linux/jhash.h>
124 #include <linux/random.h>
125 #include <trace/events/napi.h>
126 #include <trace/events/net.h>
127 #include <trace/events/skb.h>
128 #include <linux/pci.h>
129 #include <linux/inetdevice.h>
130 #include <linux/cpu_rmap.h>
131 #include <linux/static_key.h>
133 #include "net-sysfs.h"
135 /* Instead of increasing this, you should create a hash table. */
136 #define MAX_GRO_SKBS 8
138 /* This should be increased if a protocol with a bigger head is added. */
139 #define GRO_MAX_HEAD (MAX_HEADER + 128)
141 static DEFINE_SPINLOCK(ptype_lock);
142 static DEFINE_SPINLOCK(offload_lock);
143 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly;
144 struct list_head ptype_all __read_mostly; /* Taps */
145 static struct list_head offload_base __read_mostly;
148 * The @dev_base_head list is protected by @dev_base_lock and the rtnl
151 * Pure readers hold dev_base_lock for reading, or rcu_read_lock()
153 * Writers must hold the rtnl semaphore while they loop through the
154 * dev_base_head list, and hold dev_base_lock for writing when they do the
155 * actual updates. This allows pure readers to access the list even
156 * while a writer is preparing to update it.
158 * To put it another way, dev_base_lock is held for writing only to
159 * protect against pure readers; the rtnl semaphore provides the
160 * protection against other writers.
162 * See, for example usages, register_netdevice() and
163 * unregister_netdevice(), which must be called with the rtnl
166 DEFINE_RWLOCK(dev_base_lock);
167 EXPORT_SYMBOL(dev_base_lock);
169 seqcount_t devnet_rename_seq;
171 static inline void dev_base_seq_inc(struct net *net)
173 while (++net->dev_base_seq == 0);
176 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name)
178 unsigned int hash = full_name_hash(name, strnlen(name, IFNAMSIZ));
180 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)];
183 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex)
185 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)];
188 static inline void rps_lock(struct softnet_data *sd)
191 spin_lock(&sd->input_pkt_queue.lock);
195 static inline void rps_unlock(struct softnet_data *sd)
198 spin_unlock(&sd->input_pkt_queue.lock);
202 /* Device list insertion */
203 static void list_netdevice(struct net_device *dev)
205 struct net *net = dev_net(dev);
209 write_lock_bh(&dev_base_lock);
210 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head);
211 hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
212 hlist_add_head_rcu(&dev->index_hlist,
213 dev_index_hash(net, dev->ifindex));
214 write_unlock_bh(&dev_base_lock);
216 dev_base_seq_inc(net);
219 /* Device list removal
220 * caller must respect a RCU grace period before freeing/reusing dev
222 static void unlist_netdevice(struct net_device *dev)
226 /* Unlink dev from the device chain */
227 write_lock_bh(&dev_base_lock);
228 list_del_rcu(&dev->dev_list);
229 hlist_del_rcu(&dev->name_hlist);
230 hlist_del_rcu(&dev->index_hlist);
231 write_unlock_bh(&dev_base_lock);
233 dev_base_seq_inc(dev_net(dev));
240 static RAW_NOTIFIER_HEAD(netdev_chain);
243 * Device drivers call our routines to queue packets here. We empty the
244 * queue in the local softnet handler.
247 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data);
248 EXPORT_PER_CPU_SYMBOL(softnet_data);
250 #ifdef CONFIG_LOCKDEP
252 * register_netdevice() inits txq->_xmit_lock and sets lockdep class
253 * according to dev->type
255 static const unsigned short netdev_lock_type[] =
256 {ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25,
257 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET,
258 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM,
259 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP,
260 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD,
261 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25,
262 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP,
263 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD,
264 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI,
265 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE,
266 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET,
267 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL,
268 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM,
269 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE,
270 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE};
272 static const char *const netdev_lock_name[] =
273 {"_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25",
274 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET",
275 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM",
276 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP",
277 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD",
278 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25",
279 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP",
280 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD",
281 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI",
282 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE",
283 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET",
284 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL",
285 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM",
286 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE",
287 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"};
289 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)];
290 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)];
292 static inline unsigned short netdev_lock_pos(unsigned short dev_type)
296 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++)
297 if (netdev_lock_type[i] == dev_type)
299 /* the last key is used by default */
300 return ARRAY_SIZE(netdev_lock_type) - 1;
303 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
304 unsigned short dev_type)
308 i = netdev_lock_pos(dev_type);
309 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i],
310 netdev_lock_name[i]);
313 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
317 i = netdev_lock_pos(dev->type);
318 lockdep_set_class_and_name(&dev->addr_list_lock,
319 &netdev_addr_lock_key[i],
320 netdev_lock_name[i]);
323 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
324 unsigned short dev_type)
327 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
332 /*******************************************************************************
334 Protocol management and registration routines
336 *******************************************************************************/
339 * Add a protocol ID to the list. Now that the input handler is
340 * smarter we can dispense with all the messy stuff that used to be
343 * BEWARE!!! Protocol handlers, mangling input packets,
344 * MUST BE last in hash buckets and checking protocol handlers
345 * MUST start from promiscuous ptype_all chain in net_bh.
346 * It is true now, do not change it.
347 * Explanation follows: if protocol handler, mangling packet, will
348 * be the first on list, it is not able to sense, that packet
349 * is cloned and should be copied-on-write, so that it will
350 * change it and subsequent readers will get broken packet.
354 static inline struct list_head *ptype_head(const struct packet_type *pt)
356 if (pt->type == htons(ETH_P_ALL))
359 return &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK];
363 * dev_add_pack - add packet handler
364 * @pt: packet type declaration
366 * Add a protocol handler to the networking stack. The passed &packet_type
367 * is linked into kernel lists and may not be freed until it has been
368 * removed from the kernel lists.
370 * This call does not sleep therefore it can not
371 * guarantee all CPU's that are in middle of receiving packets
372 * will see the new packet type (until the next received packet).
375 void dev_add_pack(struct packet_type *pt)
377 struct list_head *head = ptype_head(pt);
379 spin_lock(&ptype_lock);
380 list_add_rcu(&pt->list, head);
381 spin_unlock(&ptype_lock);
383 EXPORT_SYMBOL(dev_add_pack);
386 * __dev_remove_pack - remove packet handler
387 * @pt: packet type declaration
389 * Remove a protocol handler that was previously added to the kernel
390 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
391 * from the kernel lists and can be freed or reused once this function
394 * The packet type might still be in use by receivers
395 * and must not be freed until after all the CPU's have gone
396 * through a quiescent state.
398 void __dev_remove_pack(struct packet_type *pt)
400 struct list_head *head = ptype_head(pt);
401 struct packet_type *pt1;
403 spin_lock(&ptype_lock);
405 list_for_each_entry(pt1, head, list) {
407 list_del_rcu(&pt->list);
412 pr_warn("dev_remove_pack: %p not found\n", pt);
414 spin_unlock(&ptype_lock);
416 EXPORT_SYMBOL(__dev_remove_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 * This call sleeps to guarantee that no CPU is looking at the packet
430 void dev_remove_pack(struct packet_type *pt)
432 __dev_remove_pack(pt);
436 EXPORT_SYMBOL(dev_remove_pack);
440 * dev_add_offload - register offload handlers
441 * @po: protocol offload declaration
443 * Add protocol offload handlers to the networking stack. The passed
444 * &proto_offload is linked into kernel lists and may not be freed until
445 * it has been removed from the kernel lists.
447 * This call does not sleep therefore it can not
448 * guarantee all CPU's that are in middle of receiving packets
449 * will see the new offload handlers (until the next received packet).
451 void dev_add_offload(struct packet_offload *po)
453 struct list_head *head = &offload_base;
455 spin_lock(&offload_lock);
456 list_add_rcu(&po->list, head);
457 spin_unlock(&offload_lock);
459 EXPORT_SYMBOL(dev_add_offload);
462 * __dev_remove_offload - remove offload handler
463 * @po: packet offload declaration
465 * Remove a protocol offload handler that was previously added to the
466 * kernel offload handlers by dev_add_offload(). The passed &offload_type
467 * is removed from the kernel lists and can be freed or reused once this
470 * The packet type might still be in use by receivers
471 * and must not be freed until after all the CPU's have gone
472 * through a quiescent state.
474 void __dev_remove_offload(struct packet_offload *po)
476 struct list_head *head = &offload_base;
477 struct packet_offload *po1;
479 spin_lock(&offload_lock);
481 list_for_each_entry(po1, head, list) {
483 list_del_rcu(&po->list);
488 pr_warn("dev_remove_offload: %p not found\n", po);
490 spin_unlock(&offload_lock);
492 EXPORT_SYMBOL(__dev_remove_offload);
495 * dev_remove_offload - remove packet offload handler
496 * @po: packet offload declaration
498 * Remove a packet offload handler that was previously added to the kernel
499 * offload handlers by dev_add_offload(). The passed &offload_type is
500 * removed from the kernel lists and can be freed or reused once this
503 * This call sleeps to guarantee that no CPU is looking at the packet
506 void dev_remove_offload(struct packet_offload *po)
508 __dev_remove_offload(po);
512 EXPORT_SYMBOL(dev_remove_offload);
514 /******************************************************************************
516 Device Boot-time Settings Routines
518 *******************************************************************************/
520 /* Boot time configuration table */
521 static struct netdev_boot_setup dev_boot_setup[NETDEV_BOOT_SETUP_MAX];
524 * netdev_boot_setup_add - add new setup entry
525 * @name: name of the device
526 * @map: configured settings for the device
528 * Adds new setup entry to the dev_boot_setup list. The function
529 * returns 0 on error and 1 on success. This is a generic routine to
532 static int netdev_boot_setup_add(char *name, struct ifmap *map)
534 struct netdev_boot_setup *s;
538 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
539 if (s[i].name[0] == '\0' || s[i].name[0] == ' ') {
540 memset(s[i].name, 0, sizeof(s[i].name));
541 strlcpy(s[i].name, name, IFNAMSIZ);
542 memcpy(&s[i].map, map, sizeof(s[i].map));
547 return i >= NETDEV_BOOT_SETUP_MAX ? 0 : 1;
551 * netdev_boot_setup_check - check boot time settings
552 * @dev: the netdevice
554 * Check boot time settings for the device.
555 * The found settings are set for the device to be used
556 * later in the device probing.
557 * Returns 0 if no settings found, 1 if they are.
559 int netdev_boot_setup_check(struct net_device *dev)
561 struct netdev_boot_setup *s = dev_boot_setup;
564 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
565 if (s[i].name[0] != '\0' && s[i].name[0] != ' ' &&
566 !strcmp(dev->name, s[i].name)) {
567 dev->irq = s[i].map.irq;
568 dev->base_addr = s[i].map.base_addr;
569 dev->mem_start = s[i].map.mem_start;
570 dev->mem_end = s[i].map.mem_end;
576 EXPORT_SYMBOL(netdev_boot_setup_check);
580 * netdev_boot_base - get address from boot time settings
581 * @prefix: prefix for network device
582 * @unit: id for network device
584 * Check boot time settings for the base address of device.
585 * The found settings are set for the device to be used
586 * later in the device probing.
587 * Returns 0 if no settings found.
589 unsigned long netdev_boot_base(const char *prefix, int unit)
591 const struct netdev_boot_setup *s = dev_boot_setup;
595 sprintf(name, "%s%d", prefix, unit);
598 * If device already registered then return base of 1
599 * to indicate not to probe for this interface
601 if (__dev_get_by_name(&init_net, name))
604 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++)
605 if (!strcmp(name, s[i].name))
606 return s[i].map.base_addr;
611 * Saves at boot time configured settings for any netdevice.
613 int __init netdev_boot_setup(char *str)
618 str = get_options(str, ARRAY_SIZE(ints), ints);
623 memset(&map, 0, sizeof(map));
627 map.base_addr = ints[2];
629 map.mem_start = ints[3];
631 map.mem_end = ints[4];
633 /* Add new entry to the list */
634 return netdev_boot_setup_add(str, &map);
637 __setup("netdev=", netdev_boot_setup);
639 /*******************************************************************************
641 Device Interface Subroutines
643 *******************************************************************************/
646 * __dev_get_by_name - find a device by its name
647 * @net: the applicable net namespace
648 * @name: name to find
650 * Find an interface by name. Must be called under RTNL semaphore
651 * or @dev_base_lock. If the name is found a pointer to the device
652 * is returned. If the name is not found then %NULL is returned. The
653 * reference counters are not incremented so the caller must be
654 * careful with locks.
657 struct net_device *__dev_get_by_name(struct net *net, const char *name)
659 struct net_device *dev;
660 struct hlist_head *head = dev_name_hash(net, name);
662 hlist_for_each_entry(dev, head, name_hlist)
663 if (!strncmp(dev->name, name, IFNAMSIZ))
668 EXPORT_SYMBOL(__dev_get_by_name);
671 * dev_get_by_name_rcu - find a device by its name
672 * @net: the applicable net namespace
673 * @name: name to find
675 * Find an interface by name.
676 * If the name is found a pointer to the device is returned.
677 * If the name is not found then %NULL is returned.
678 * The reference counters are not incremented so the caller must be
679 * careful with locks. The caller must hold RCU lock.
682 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name)
684 struct net_device *dev;
685 struct hlist_head *head = dev_name_hash(net, name);
687 hlist_for_each_entry_rcu(dev, head, name_hlist)
688 if (!strncmp(dev->name, name, IFNAMSIZ))
693 EXPORT_SYMBOL(dev_get_by_name_rcu);
696 * dev_get_by_name - find a device by its name
697 * @net: the applicable net namespace
698 * @name: name to find
700 * Find an interface by name. This can be called from any
701 * context and does its own locking. The returned handle has
702 * the usage count incremented and the caller must use dev_put() to
703 * release it when it is no longer needed. %NULL is returned if no
704 * matching device is found.
707 struct net_device *dev_get_by_name(struct net *net, const char *name)
709 struct net_device *dev;
712 dev = dev_get_by_name_rcu(net, name);
718 EXPORT_SYMBOL(dev_get_by_name);
721 * __dev_get_by_index - find a device by its ifindex
722 * @net: the applicable net namespace
723 * @ifindex: index of device
725 * Search for an interface by index. Returns %NULL if the device
726 * is not found or a pointer to the device. The device has not
727 * had its reference counter increased so the caller must be careful
728 * about locking. The caller must hold either the RTNL semaphore
732 struct net_device *__dev_get_by_index(struct net *net, int ifindex)
734 struct net_device *dev;
735 struct hlist_head *head = dev_index_hash(net, ifindex);
737 hlist_for_each_entry(dev, head, index_hlist)
738 if (dev->ifindex == ifindex)
743 EXPORT_SYMBOL(__dev_get_by_index);
746 * dev_get_by_index_rcu - find a device by its ifindex
747 * @net: the applicable net namespace
748 * @ifindex: index of device
750 * Search for an interface by index. Returns %NULL if the device
751 * is not found or a pointer to the device. The device has not
752 * had its reference counter increased so the caller must be careful
753 * about locking. The caller must hold RCU lock.
756 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex)
758 struct net_device *dev;
759 struct hlist_head *head = dev_index_hash(net, ifindex);
761 hlist_for_each_entry_rcu(dev, head, index_hlist)
762 if (dev->ifindex == ifindex)
767 EXPORT_SYMBOL(dev_get_by_index_rcu);
771 * dev_get_by_index - find a device by its ifindex
772 * @net: the applicable net namespace
773 * @ifindex: index of device
775 * Search for an interface by index. Returns NULL if the device
776 * is not found or a pointer to the device. The device returned has
777 * had a reference added and the pointer is safe until the user calls
778 * dev_put to indicate they have finished with it.
781 struct net_device *dev_get_by_index(struct net *net, int ifindex)
783 struct net_device *dev;
786 dev = dev_get_by_index_rcu(net, ifindex);
792 EXPORT_SYMBOL(dev_get_by_index);
795 * dev_getbyhwaddr_rcu - find a device by its hardware address
796 * @net: the applicable net namespace
797 * @type: media type of device
798 * @ha: hardware address
800 * Search for an interface by MAC address. Returns NULL if the device
801 * is not found or a pointer to the device.
802 * The caller must hold RCU or RTNL.
803 * The returned device has not had its ref count increased
804 * and the caller must therefore be careful about locking
808 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type,
811 struct net_device *dev;
813 for_each_netdev_rcu(net, dev)
814 if (dev->type == type &&
815 !memcmp(dev->dev_addr, ha, dev->addr_len))
820 EXPORT_SYMBOL(dev_getbyhwaddr_rcu);
822 struct net_device *__dev_getfirstbyhwtype(struct net *net, unsigned short type)
824 struct net_device *dev;
827 for_each_netdev(net, dev)
828 if (dev->type == type)
833 EXPORT_SYMBOL(__dev_getfirstbyhwtype);
835 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type)
837 struct net_device *dev, *ret = NULL;
840 for_each_netdev_rcu(net, dev)
841 if (dev->type == type) {
849 EXPORT_SYMBOL(dev_getfirstbyhwtype);
852 * dev_get_by_flags_rcu - find any device with given flags
853 * @net: the applicable net namespace
854 * @if_flags: IFF_* values
855 * @mask: bitmask of bits in if_flags to check
857 * Search for any interface with the given flags. Returns NULL if a device
858 * is not found or a pointer to the device. Must be called inside
859 * rcu_read_lock(), and result refcount is unchanged.
862 struct net_device *dev_get_by_flags_rcu(struct net *net, unsigned short if_flags,
865 struct net_device *dev, *ret;
868 for_each_netdev_rcu(net, dev) {
869 if (((dev->flags ^ if_flags) & mask) == 0) {
876 EXPORT_SYMBOL(dev_get_by_flags_rcu);
879 * dev_valid_name - check if name is okay for network device
882 * Network device names need to be valid file names to
883 * to allow sysfs to work. We also disallow any kind of
886 bool dev_valid_name(const char *name)
890 if (strlen(name) >= IFNAMSIZ)
892 if (!strcmp(name, ".") || !strcmp(name, ".."))
896 if (*name == '/' || isspace(*name))
902 EXPORT_SYMBOL(dev_valid_name);
905 * __dev_alloc_name - allocate a name for a device
906 * @net: network namespace to allocate the device name in
907 * @name: name format string
908 * @buf: scratch buffer and result name string
910 * Passed a format string - eg "lt%d" it will try and find a suitable
911 * id. It scans list of devices to build up a free map, then chooses
912 * the first empty slot. The caller must hold the dev_base or rtnl lock
913 * while allocating the name and adding the device in order to avoid
915 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
916 * Returns the number of the unit assigned or a negative errno code.
919 static int __dev_alloc_name(struct net *net, const char *name, char *buf)
923 const int max_netdevices = 8*PAGE_SIZE;
924 unsigned long *inuse;
925 struct net_device *d;
927 p = strnchr(name, IFNAMSIZ-1, '%');
930 * Verify the string as this thing may have come from
931 * the user. There must be either one "%d" and no other "%"
934 if (p[1] != 'd' || strchr(p + 2, '%'))
937 /* Use one page as a bit array of possible slots */
938 inuse = (unsigned long *) get_zeroed_page(GFP_ATOMIC);
942 for_each_netdev(net, d) {
943 if (!sscanf(d->name, name, &i))
945 if (i < 0 || i >= max_netdevices)
948 /* avoid cases where sscanf is not exact inverse of printf */
949 snprintf(buf, IFNAMSIZ, name, i);
950 if (!strncmp(buf, d->name, IFNAMSIZ))
954 i = find_first_zero_bit(inuse, max_netdevices);
955 free_page((unsigned long) inuse);
959 snprintf(buf, IFNAMSIZ, name, i);
960 if (!__dev_get_by_name(net, buf))
963 /* It is possible to run out of possible slots
964 * when the name is long and there isn't enough space left
965 * for the digits, or if all bits are used.
971 * dev_alloc_name - allocate a name for a device
973 * @name: name format string
975 * Passed a format string - eg "lt%d" it will try and find a suitable
976 * id. It scans list of devices to build up a free map, then chooses
977 * the first empty slot. The caller must hold the dev_base or rtnl lock
978 * while allocating the name and adding the device in order to avoid
980 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
981 * Returns the number of the unit assigned or a negative errno code.
984 int dev_alloc_name(struct net_device *dev, const char *name)
990 BUG_ON(!dev_net(dev));
992 ret = __dev_alloc_name(net, name, buf);
994 strlcpy(dev->name, buf, IFNAMSIZ);
997 EXPORT_SYMBOL(dev_alloc_name);
999 static int dev_alloc_name_ns(struct net *net,
1000 struct net_device *dev,
1006 ret = __dev_alloc_name(net, name, buf);
1008 strlcpy(dev->name, buf, IFNAMSIZ);
1012 static int dev_get_valid_name(struct net *net,
1013 struct net_device *dev,
1018 if (!dev_valid_name(name))
1021 if (strchr(name, '%'))
1022 return dev_alloc_name_ns(net, dev, name);
1023 else if (__dev_get_by_name(net, name))
1025 else if (dev->name != name)
1026 strlcpy(dev->name, name, IFNAMSIZ);
1032 * dev_change_name - change name of a device
1034 * @newname: name (or format string) must be at least IFNAMSIZ
1036 * Change name of a device, can pass format strings "eth%d".
1039 int dev_change_name(struct net_device *dev, const char *newname)
1041 char oldname[IFNAMSIZ];
1047 BUG_ON(!dev_net(dev));
1050 if (dev->flags & IFF_UP)
1053 write_seqcount_begin(&devnet_rename_seq);
1055 if (strncmp(newname, dev->name, IFNAMSIZ) == 0) {
1056 write_seqcount_end(&devnet_rename_seq);
1060 memcpy(oldname, dev->name, IFNAMSIZ);
1062 err = dev_get_valid_name(net, dev, newname);
1064 write_seqcount_end(&devnet_rename_seq);
1069 ret = device_rename(&dev->dev, dev->name);
1071 memcpy(dev->name, oldname, IFNAMSIZ);
1072 write_seqcount_end(&devnet_rename_seq);
1076 write_seqcount_end(&devnet_rename_seq);
1078 write_lock_bh(&dev_base_lock);
1079 hlist_del_rcu(&dev->name_hlist);
1080 write_unlock_bh(&dev_base_lock);
1084 write_lock_bh(&dev_base_lock);
1085 hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
1086 write_unlock_bh(&dev_base_lock);
1088 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev);
1089 ret = notifier_to_errno(ret);
1092 /* err >= 0 after dev_alloc_name() or stores the first errno */
1095 write_seqcount_begin(&devnet_rename_seq);
1096 memcpy(dev->name, oldname, IFNAMSIZ);
1099 pr_err("%s: name change rollback failed: %d\n",
1108 * dev_set_alias - change ifalias of a device
1110 * @alias: name up to IFALIASZ
1111 * @len: limit of bytes to copy from info
1113 * Set ifalias for a device,
1115 int dev_set_alias(struct net_device *dev, const char *alias, size_t len)
1121 if (len >= IFALIASZ)
1125 kfree(dev->ifalias);
1126 dev->ifalias = NULL;
1130 new_ifalias = krealloc(dev->ifalias, len + 1, GFP_KERNEL);
1133 dev->ifalias = new_ifalias;
1135 strlcpy(dev->ifalias, alias, len+1);
1141 * netdev_features_change - device changes features
1142 * @dev: device to cause notification
1144 * Called to indicate a device has changed features.
1146 void netdev_features_change(struct net_device *dev)
1148 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev);
1150 EXPORT_SYMBOL(netdev_features_change);
1153 * netdev_state_change - device changes state
1154 * @dev: device to cause notification
1156 * Called to indicate a device has changed state. This function calls
1157 * the notifier chains for netdev_chain and sends a NEWLINK message
1158 * to the routing socket.
1160 void netdev_state_change(struct net_device *dev)
1162 if (dev->flags & IFF_UP) {
1163 call_netdevice_notifiers(NETDEV_CHANGE, dev);
1164 rtmsg_ifinfo(RTM_NEWLINK, dev, 0);
1167 EXPORT_SYMBOL(netdev_state_change);
1170 * netdev_notify_peers - notify network peers about existence of @dev
1171 * @dev: network device
1173 * Generate traffic such that interested network peers are aware of
1174 * @dev, such as by generating a gratuitous ARP. This may be used when
1175 * a device wants to inform the rest of the network about some sort of
1176 * reconfiguration such as a failover event or virtual machine
1179 void netdev_notify_peers(struct net_device *dev)
1182 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
1185 EXPORT_SYMBOL(netdev_notify_peers);
1187 static int __dev_open(struct net_device *dev)
1189 const struct net_device_ops *ops = dev->netdev_ops;
1194 if (!netif_device_present(dev))
1197 /* Block netpoll from trying to do any rx path servicing.
1198 * If we don't do this there is a chance ndo_poll_controller
1199 * or ndo_poll may be running while we open the device
1201 ret = netpoll_rx_disable(dev);
1205 ret = call_netdevice_notifiers(NETDEV_PRE_UP, dev);
1206 ret = notifier_to_errno(ret);
1210 set_bit(__LINK_STATE_START, &dev->state);
1212 if (ops->ndo_validate_addr)
1213 ret = ops->ndo_validate_addr(dev);
1215 if (!ret && ops->ndo_open)
1216 ret = ops->ndo_open(dev);
1218 netpoll_rx_enable(dev);
1221 clear_bit(__LINK_STATE_START, &dev->state);
1223 dev->flags |= IFF_UP;
1224 net_dmaengine_get();
1225 dev_set_rx_mode(dev);
1227 add_device_randomness(dev->dev_addr, dev->addr_len);
1234 * dev_open - prepare an interface for use.
1235 * @dev: device to open
1237 * Takes a device from down to up state. The device's private open
1238 * function is invoked and then the multicast lists are loaded. Finally
1239 * the device is moved into the up state and a %NETDEV_UP message is
1240 * sent to the netdev notifier chain.
1242 * Calling this function on an active interface is a nop. On a failure
1243 * a negative errno code is returned.
1245 int dev_open(struct net_device *dev)
1249 if (dev->flags & IFF_UP)
1252 ret = __dev_open(dev);
1256 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING);
1257 call_netdevice_notifiers(NETDEV_UP, dev);
1261 EXPORT_SYMBOL(dev_open);
1263 static int __dev_close_many(struct list_head *head)
1265 struct net_device *dev;
1270 list_for_each_entry(dev, head, unreg_list) {
1271 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev);
1273 clear_bit(__LINK_STATE_START, &dev->state);
1275 /* Synchronize to scheduled poll. We cannot touch poll list, it
1276 * can be even on different cpu. So just clear netif_running().
1278 * dev->stop() will invoke napi_disable() on all of it's
1279 * napi_struct instances on this device.
1281 smp_mb__after_clear_bit(); /* Commit netif_running(). */
1284 dev_deactivate_many(head);
1286 list_for_each_entry(dev, head, unreg_list) {
1287 const struct net_device_ops *ops = dev->netdev_ops;
1290 * Call the device specific close. This cannot fail.
1291 * Only if device is UP
1293 * We allow it to be called even after a DETACH hot-plug
1299 dev->flags &= ~IFF_UP;
1300 net_dmaengine_put();
1306 static int __dev_close(struct net_device *dev)
1311 /* Temporarily disable netpoll until the interface is down */
1312 retval = netpoll_rx_disable(dev);
1316 list_add(&dev->unreg_list, &single);
1317 retval = __dev_close_many(&single);
1320 netpoll_rx_enable(dev);
1324 static int dev_close_many(struct list_head *head)
1326 struct net_device *dev, *tmp;
1327 LIST_HEAD(tmp_list);
1329 list_for_each_entry_safe(dev, tmp, head, unreg_list)
1330 if (!(dev->flags & IFF_UP))
1331 list_move(&dev->unreg_list, &tmp_list);
1333 __dev_close_many(head);
1335 list_for_each_entry(dev, head, unreg_list) {
1336 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING);
1337 call_netdevice_notifiers(NETDEV_DOWN, dev);
1340 /* rollback_registered_many needs the complete original list */
1341 list_splice(&tmp_list, head);
1346 * dev_close - shutdown an interface.
1347 * @dev: device to shutdown
1349 * This function moves an active device into down state. A
1350 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device
1351 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier
1354 int dev_close(struct net_device *dev)
1357 if (dev->flags & IFF_UP) {
1360 /* Block netpoll rx while the interface is going down */
1361 ret = netpoll_rx_disable(dev);
1365 list_add(&dev->unreg_list, &single);
1366 dev_close_many(&single);
1369 netpoll_rx_enable(dev);
1373 EXPORT_SYMBOL(dev_close);
1377 * dev_disable_lro - disable Large Receive Offload on a device
1380 * Disable Large Receive Offload (LRO) on a net device. Must be
1381 * called under RTNL. This is needed if received packets may be
1382 * forwarded to another interface.
1384 void dev_disable_lro(struct net_device *dev)
1387 * If we're trying to disable lro on a vlan device
1388 * use the underlying physical device instead
1390 if (is_vlan_dev(dev))
1391 dev = vlan_dev_real_dev(dev);
1393 dev->wanted_features &= ~NETIF_F_LRO;
1394 netdev_update_features(dev);
1396 if (unlikely(dev->features & NETIF_F_LRO))
1397 netdev_WARN(dev, "failed to disable LRO!\n");
1399 EXPORT_SYMBOL(dev_disable_lro);
1402 static int dev_boot_phase = 1;
1405 * register_netdevice_notifier - register a network notifier block
1408 * Register a notifier to be called when network device events occur.
1409 * The notifier passed is linked into the kernel structures and must
1410 * not be reused until it has been unregistered. A negative errno code
1411 * is returned on a failure.
1413 * When registered all registration and up events are replayed
1414 * to the new notifier to allow device to have a race free
1415 * view of the network device list.
1418 int register_netdevice_notifier(struct notifier_block *nb)
1420 struct net_device *dev;
1421 struct net_device *last;
1426 err = raw_notifier_chain_register(&netdev_chain, nb);
1432 for_each_netdev(net, dev) {
1433 err = nb->notifier_call(nb, NETDEV_REGISTER, dev);
1434 err = notifier_to_errno(err);
1438 if (!(dev->flags & IFF_UP))
1441 nb->notifier_call(nb, NETDEV_UP, dev);
1452 for_each_netdev(net, dev) {
1456 if (dev->flags & IFF_UP) {
1457 nb->notifier_call(nb, NETDEV_GOING_DOWN, dev);
1458 nb->notifier_call(nb, NETDEV_DOWN, dev);
1460 nb->notifier_call(nb, NETDEV_UNREGISTER, dev);
1465 raw_notifier_chain_unregister(&netdev_chain, nb);
1468 EXPORT_SYMBOL(register_netdevice_notifier);
1471 * unregister_netdevice_notifier - unregister a network notifier block
1474 * Unregister a notifier previously registered by
1475 * register_netdevice_notifier(). The notifier is unlinked into the
1476 * kernel structures and may then be reused. A negative errno code
1477 * is returned on a failure.
1479 * After unregistering unregister and down device events are synthesized
1480 * for all devices on the device list to the removed notifier to remove
1481 * the need for special case cleanup code.
1484 int unregister_netdevice_notifier(struct notifier_block *nb)
1486 struct net_device *dev;
1491 err = raw_notifier_chain_unregister(&netdev_chain, nb);
1496 for_each_netdev(net, dev) {
1497 if (dev->flags & IFF_UP) {
1498 nb->notifier_call(nb, NETDEV_GOING_DOWN, dev);
1499 nb->notifier_call(nb, NETDEV_DOWN, dev);
1501 nb->notifier_call(nb, NETDEV_UNREGISTER, dev);
1508 EXPORT_SYMBOL(unregister_netdevice_notifier);
1511 * call_netdevice_notifiers - call all network notifier blocks
1512 * @val: value passed unmodified to notifier function
1513 * @dev: net_device pointer passed unmodified to notifier function
1515 * Call all network notifier blocks. Parameters and return value
1516 * are as for raw_notifier_call_chain().
1519 int call_netdevice_notifiers(unsigned long val, struct net_device *dev)
1522 return raw_notifier_call_chain(&netdev_chain, val, dev);
1524 EXPORT_SYMBOL(call_netdevice_notifiers);
1526 static struct static_key netstamp_needed __read_mostly;
1527 #ifdef HAVE_JUMP_LABEL
1528 /* We are not allowed to call static_key_slow_dec() from irq context
1529 * If net_disable_timestamp() is called from irq context, defer the
1530 * static_key_slow_dec() calls.
1532 static atomic_t netstamp_needed_deferred;
1535 void net_enable_timestamp(void)
1537 #ifdef HAVE_JUMP_LABEL
1538 int deferred = atomic_xchg(&netstamp_needed_deferred, 0);
1542 static_key_slow_dec(&netstamp_needed);
1546 static_key_slow_inc(&netstamp_needed);
1548 EXPORT_SYMBOL(net_enable_timestamp);
1550 void net_disable_timestamp(void)
1552 #ifdef HAVE_JUMP_LABEL
1553 if (in_interrupt()) {
1554 atomic_inc(&netstamp_needed_deferred);
1558 static_key_slow_dec(&netstamp_needed);
1560 EXPORT_SYMBOL(net_disable_timestamp);
1562 static inline void net_timestamp_set(struct sk_buff *skb)
1564 skb->tstamp.tv64 = 0;
1565 if (static_key_false(&netstamp_needed))
1566 __net_timestamp(skb);
1569 #define net_timestamp_check(COND, SKB) \
1570 if (static_key_false(&netstamp_needed)) { \
1571 if ((COND) && !(SKB)->tstamp.tv64) \
1572 __net_timestamp(SKB); \
1575 static inline bool is_skb_forwardable(struct net_device *dev,
1576 struct sk_buff *skb)
1580 if (!(dev->flags & IFF_UP))
1583 len = dev->mtu + dev->hard_header_len + VLAN_HLEN;
1584 if (skb->len <= len)
1587 /* if TSO is enabled, we don't care about the length as the packet
1588 * could be forwarded without being segmented before
1590 if (skb_is_gso(skb))
1597 * dev_forward_skb - loopback an skb to another netif
1599 * @dev: destination network device
1600 * @skb: buffer to forward
1603 * NET_RX_SUCCESS (no congestion)
1604 * NET_RX_DROP (packet was dropped, but freed)
1606 * dev_forward_skb can be used for injecting an skb from the
1607 * start_xmit function of one device into the receive queue
1608 * of another device.
1610 * The receiving device may be in another namespace, so
1611 * we have to clear all information in the skb that could
1612 * impact namespace isolation.
1614 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1616 if (skb_shinfo(skb)->tx_flags & SKBTX_DEV_ZEROCOPY) {
1617 if (skb_copy_ubufs(skb, GFP_ATOMIC)) {
1618 atomic_long_inc(&dev->rx_dropped);
1626 if (unlikely(!is_skb_forwardable(dev, skb))) {
1627 atomic_long_inc(&dev->rx_dropped);
1633 skb->tstamp.tv64 = 0;
1634 skb->pkt_type = PACKET_HOST;
1635 skb->protocol = eth_type_trans(skb, dev);
1639 nf_reset_trace(skb);
1640 return netif_rx(skb);
1642 EXPORT_SYMBOL_GPL(dev_forward_skb);
1644 static inline int deliver_skb(struct sk_buff *skb,
1645 struct packet_type *pt_prev,
1646 struct net_device *orig_dev)
1648 if (unlikely(skb_orphan_frags(skb, GFP_ATOMIC)))
1650 atomic_inc(&skb->users);
1651 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
1654 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb)
1656 if (!ptype->af_packet_priv || !skb->sk)
1659 if (ptype->id_match)
1660 return ptype->id_match(ptype, skb->sk);
1661 else if ((struct sock *)ptype->af_packet_priv == skb->sk)
1668 * Support routine. Sends outgoing frames to any network
1669 * taps currently in use.
1672 static void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
1674 struct packet_type *ptype;
1675 struct sk_buff *skb2 = NULL;
1676 struct packet_type *pt_prev = NULL;
1679 list_for_each_entry_rcu(ptype, &ptype_all, list) {
1680 /* Never send packets back to the socket
1681 * they originated from - MvS (miquels@drinkel.ow.org)
1683 if ((ptype->dev == dev || !ptype->dev) &&
1684 (!skb_loop_sk(ptype, skb))) {
1686 deliver_skb(skb2, pt_prev, skb->dev);
1691 skb2 = skb_clone(skb, GFP_ATOMIC);
1695 net_timestamp_set(skb2);
1697 /* skb->nh should be correctly
1698 set by sender, so that the second statement is
1699 just protection against buggy protocols.
1701 skb_reset_mac_header(skb2);
1703 if (skb_network_header(skb2) < skb2->data ||
1704 skb2->network_header > skb2->tail) {
1705 net_crit_ratelimited("protocol %04x is buggy, dev %s\n",
1706 ntohs(skb2->protocol),
1708 skb_reset_network_header(skb2);
1711 skb2->transport_header = skb2->network_header;
1712 skb2->pkt_type = PACKET_OUTGOING;
1717 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
1722 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change
1723 * @dev: Network device
1724 * @txq: number of queues available
1726 * If real_num_tx_queues is changed the tc mappings may no longer be
1727 * valid. To resolve this verify the tc mapping remains valid and if
1728 * not NULL the mapping. With no priorities mapping to this
1729 * offset/count pair it will no longer be used. In the worst case TC0
1730 * is invalid nothing can be done so disable priority mappings. If is
1731 * expected that drivers will fix this mapping if they can before
1732 * calling netif_set_real_num_tx_queues.
1734 static void netif_setup_tc(struct net_device *dev, unsigned int txq)
1737 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
1739 /* If TC0 is invalidated disable TC mapping */
1740 if (tc->offset + tc->count > txq) {
1741 pr_warn("Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n");
1746 /* Invalidated prio to tc mappings set to TC0 */
1747 for (i = 1; i < TC_BITMASK + 1; i++) {
1748 int q = netdev_get_prio_tc_map(dev, i);
1750 tc = &dev->tc_to_txq[q];
1751 if (tc->offset + tc->count > txq) {
1752 pr_warn("Number of in use tx queues changed. Priority %i to tc mapping %i is no longer valid. Setting map to 0\n",
1754 netdev_set_prio_tc_map(dev, i, 0);
1760 static DEFINE_MUTEX(xps_map_mutex);
1761 #define xmap_dereference(P) \
1762 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex))
1764 static struct xps_map *remove_xps_queue(struct xps_dev_maps *dev_maps,
1767 struct xps_map *map = NULL;
1771 map = xmap_dereference(dev_maps->cpu_map[cpu]);
1773 for (pos = 0; map && pos < map->len; pos++) {
1774 if (map->queues[pos] == index) {
1776 map->queues[pos] = map->queues[--map->len];
1778 RCU_INIT_POINTER(dev_maps->cpu_map[cpu], NULL);
1779 kfree_rcu(map, rcu);
1789 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index)
1791 struct xps_dev_maps *dev_maps;
1793 bool active = false;
1795 mutex_lock(&xps_map_mutex);
1796 dev_maps = xmap_dereference(dev->xps_maps);
1801 for_each_possible_cpu(cpu) {
1802 for (i = index; i < dev->num_tx_queues; i++) {
1803 if (!remove_xps_queue(dev_maps, cpu, i))
1806 if (i == dev->num_tx_queues)
1811 RCU_INIT_POINTER(dev->xps_maps, NULL);
1812 kfree_rcu(dev_maps, rcu);
1815 for (i = index; i < dev->num_tx_queues; i++)
1816 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, i),
1820 mutex_unlock(&xps_map_mutex);
1823 static struct xps_map *expand_xps_map(struct xps_map *map,
1826 struct xps_map *new_map;
1827 int alloc_len = XPS_MIN_MAP_ALLOC;
1830 for (pos = 0; map && pos < map->len; pos++) {
1831 if (map->queues[pos] != index)
1836 /* Need to add queue to this CPU's existing map */
1838 if (pos < map->alloc_len)
1841 alloc_len = map->alloc_len * 2;
1844 /* Need to allocate new map to store queue on this CPU's map */
1845 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL,
1850 for (i = 0; i < pos; i++)
1851 new_map->queues[i] = map->queues[i];
1852 new_map->alloc_len = alloc_len;
1858 int netif_set_xps_queue(struct net_device *dev, struct cpumask *mask, u16 index)
1860 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL;
1861 struct xps_map *map, *new_map;
1862 int maps_sz = max_t(unsigned int, XPS_DEV_MAPS_SIZE, L1_CACHE_BYTES);
1863 int cpu, numa_node_id = -2;
1864 bool active = false;
1866 mutex_lock(&xps_map_mutex);
1868 dev_maps = xmap_dereference(dev->xps_maps);
1870 /* allocate memory for queue storage */
1871 for_each_online_cpu(cpu) {
1872 if (!cpumask_test_cpu(cpu, mask))
1876 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL);
1877 if (!new_dev_maps) {
1878 mutex_unlock(&xps_map_mutex);
1882 map = dev_maps ? xmap_dereference(dev_maps->cpu_map[cpu]) :
1885 map = expand_xps_map(map, cpu, index);
1889 RCU_INIT_POINTER(new_dev_maps->cpu_map[cpu], map);
1893 goto out_no_new_maps;
1895 for_each_possible_cpu(cpu) {
1896 if (cpumask_test_cpu(cpu, mask) && cpu_online(cpu)) {
1897 /* add queue to CPU maps */
1900 map = xmap_dereference(new_dev_maps->cpu_map[cpu]);
1901 while ((pos < map->len) && (map->queues[pos] != index))
1904 if (pos == map->len)
1905 map->queues[map->len++] = index;
1907 if (numa_node_id == -2)
1908 numa_node_id = cpu_to_node(cpu);
1909 else if (numa_node_id != cpu_to_node(cpu))
1912 } else if (dev_maps) {
1913 /* fill in the new device map from the old device map */
1914 map = xmap_dereference(dev_maps->cpu_map[cpu]);
1915 RCU_INIT_POINTER(new_dev_maps->cpu_map[cpu], map);
1920 rcu_assign_pointer(dev->xps_maps, new_dev_maps);
1922 /* Cleanup old maps */
1924 for_each_possible_cpu(cpu) {
1925 new_map = xmap_dereference(new_dev_maps->cpu_map[cpu]);
1926 map = xmap_dereference(dev_maps->cpu_map[cpu]);
1927 if (map && map != new_map)
1928 kfree_rcu(map, rcu);
1931 kfree_rcu(dev_maps, rcu);
1934 dev_maps = new_dev_maps;
1938 /* update Tx queue numa node */
1939 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index),
1940 (numa_node_id >= 0) ? numa_node_id :
1946 /* removes queue from unused CPUs */
1947 for_each_possible_cpu(cpu) {
1948 if (cpumask_test_cpu(cpu, mask) && cpu_online(cpu))
1951 if (remove_xps_queue(dev_maps, cpu, index))
1955 /* free map if not active */
1957 RCU_INIT_POINTER(dev->xps_maps, NULL);
1958 kfree_rcu(dev_maps, rcu);
1962 mutex_unlock(&xps_map_mutex);
1966 /* remove any maps that we added */
1967 for_each_possible_cpu(cpu) {
1968 new_map = xmap_dereference(new_dev_maps->cpu_map[cpu]);
1969 map = dev_maps ? xmap_dereference(dev_maps->cpu_map[cpu]) :
1971 if (new_map && new_map != map)
1975 mutex_unlock(&xps_map_mutex);
1977 kfree(new_dev_maps);
1980 EXPORT_SYMBOL(netif_set_xps_queue);
1984 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues
1985 * greater then real_num_tx_queues stale skbs on the qdisc must be flushed.
1987 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq)
1991 if (txq < 1 || txq > dev->num_tx_queues)
1994 if (dev->reg_state == NETREG_REGISTERED ||
1995 dev->reg_state == NETREG_UNREGISTERING) {
1998 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues,
2004 netif_setup_tc(dev, txq);
2006 if (txq < dev->real_num_tx_queues) {
2007 qdisc_reset_all_tx_gt(dev, txq);
2009 netif_reset_xps_queues_gt(dev, txq);
2014 dev->real_num_tx_queues = txq;
2017 EXPORT_SYMBOL(netif_set_real_num_tx_queues);
2021 * netif_set_real_num_rx_queues - set actual number of RX queues used
2022 * @dev: Network device
2023 * @rxq: Actual number of RX queues
2025 * This must be called either with the rtnl_lock held or before
2026 * registration of the net device. Returns 0 on success, or a
2027 * negative error code. If called before registration, it always
2030 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq)
2034 if (rxq < 1 || rxq > dev->num_rx_queues)
2037 if (dev->reg_state == NETREG_REGISTERED) {
2040 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues,
2046 dev->real_num_rx_queues = rxq;
2049 EXPORT_SYMBOL(netif_set_real_num_rx_queues);
2053 * netif_get_num_default_rss_queues - default number of RSS queues
2055 * This routine should set an upper limit on the number of RSS queues
2056 * used by default by multiqueue devices.
2058 int netif_get_num_default_rss_queues(void)
2060 return min_t(int, DEFAULT_MAX_NUM_RSS_QUEUES, num_online_cpus());
2062 EXPORT_SYMBOL(netif_get_num_default_rss_queues);
2064 static inline void __netif_reschedule(struct Qdisc *q)
2066 struct softnet_data *sd;
2067 unsigned long flags;
2069 local_irq_save(flags);
2070 sd = &__get_cpu_var(softnet_data);
2071 q->next_sched = NULL;
2072 *sd->output_queue_tailp = q;
2073 sd->output_queue_tailp = &q->next_sched;
2074 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2075 local_irq_restore(flags);
2078 void __netif_schedule(struct Qdisc *q)
2080 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
2081 __netif_reschedule(q);
2083 EXPORT_SYMBOL(__netif_schedule);
2085 void dev_kfree_skb_irq(struct sk_buff *skb)
2087 if (atomic_dec_and_test(&skb->users)) {
2088 struct softnet_data *sd;
2089 unsigned long flags;
2091 local_irq_save(flags);
2092 sd = &__get_cpu_var(softnet_data);
2093 skb->next = sd->completion_queue;
2094 sd->completion_queue = skb;
2095 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2096 local_irq_restore(flags);
2099 EXPORT_SYMBOL(dev_kfree_skb_irq);
2101 void dev_kfree_skb_any(struct sk_buff *skb)
2103 if (in_irq() || irqs_disabled())
2104 dev_kfree_skb_irq(skb);
2108 EXPORT_SYMBOL(dev_kfree_skb_any);
2112 * netif_device_detach - mark device as removed
2113 * @dev: network device
2115 * Mark device as removed from system and therefore no longer available.
2117 void netif_device_detach(struct net_device *dev)
2119 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) &&
2120 netif_running(dev)) {
2121 netif_tx_stop_all_queues(dev);
2124 EXPORT_SYMBOL(netif_device_detach);
2127 * netif_device_attach - mark device as attached
2128 * @dev: network device
2130 * Mark device as attached from system and restart if needed.
2132 void netif_device_attach(struct net_device *dev)
2134 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) &&
2135 netif_running(dev)) {
2136 netif_tx_wake_all_queues(dev);
2137 __netdev_watchdog_up(dev);
2140 EXPORT_SYMBOL(netif_device_attach);
2142 static void skb_warn_bad_offload(const struct sk_buff *skb)
2144 static const netdev_features_t null_features = 0;
2145 struct net_device *dev = skb->dev;
2146 const char *driver = "";
2148 if (!net_ratelimit())
2151 if (dev && dev->dev.parent)
2152 driver = dev_driver_string(dev->dev.parent);
2154 WARN(1, "%s: caps=(%pNF, %pNF) len=%d data_len=%d gso_size=%d "
2155 "gso_type=%d ip_summed=%d\n",
2156 driver, dev ? &dev->features : &null_features,
2157 skb->sk ? &skb->sk->sk_route_caps : &null_features,
2158 skb->len, skb->data_len, skb_shinfo(skb)->gso_size,
2159 skb_shinfo(skb)->gso_type, skb->ip_summed);
2163 * Invalidate hardware checksum when packet is to be mangled, and
2164 * complete checksum manually on outgoing path.
2166 int skb_checksum_help(struct sk_buff *skb)
2169 int ret = 0, offset;
2171 if (skb->ip_summed == CHECKSUM_COMPLETE)
2172 goto out_set_summed;
2174 if (unlikely(skb_shinfo(skb)->gso_size)) {
2175 skb_warn_bad_offload(skb);
2179 /* Before computing a checksum, we should make sure no frag could
2180 * be modified by an external entity : checksum could be wrong.
2182 if (skb_has_shared_frag(skb)) {
2183 ret = __skb_linearize(skb);
2188 offset = skb_checksum_start_offset(skb);
2189 BUG_ON(offset >= skb_headlen(skb));
2190 csum = skb_checksum(skb, offset, skb->len - offset, 0);
2192 offset += skb->csum_offset;
2193 BUG_ON(offset + sizeof(__sum16) > skb_headlen(skb));
2195 if (skb_cloned(skb) &&
2196 !skb_clone_writable(skb, offset + sizeof(__sum16))) {
2197 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2202 *(__sum16 *)(skb->data + offset) = csum_fold(csum);
2204 skb->ip_summed = CHECKSUM_NONE;
2208 EXPORT_SYMBOL(skb_checksum_help);
2210 __be16 skb_network_protocol(struct sk_buff *skb)
2212 __be16 type = skb->protocol;
2213 int vlan_depth = ETH_HLEN;
2215 /* Tunnel gso handlers can set protocol to ethernet. */
2216 if (type == htons(ETH_P_TEB)) {
2219 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr))))
2222 eth = (struct ethhdr *)skb_mac_header(skb);
2223 type = eth->h_proto;
2226 while (type == htons(ETH_P_8021Q) || type == htons(ETH_P_8021AD)) {
2227 struct vlan_hdr *vh;
2229 if (unlikely(!pskb_may_pull(skb, vlan_depth + VLAN_HLEN)))
2232 vh = (struct vlan_hdr *)(skb->data + vlan_depth);
2233 type = vh->h_vlan_encapsulated_proto;
2234 vlan_depth += VLAN_HLEN;
2241 * skb_mac_gso_segment - mac layer segmentation handler.
2242 * @skb: buffer to segment
2243 * @features: features for the output path (see dev->features)
2245 struct sk_buff *skb_mac_gso_segment(struct sk_buff *skb,
2246 netdev_features_t features)
2248 struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT);
2249 struct packet_offload *ptype;
2250 __be16 type = skb_network_protocol(skb);
2252 if (unlikely(!type))
2253 return ERR_PTR(-EINVAL);
2255 __skb_pull(skb, skb->mac_len);
2258 list_for_each_entry_rcu(ptype, &offload_base, list) {
2259 if (ptype->type == type && ptype->callbacks.gso_segment) {
2260 if (unlikely(skb->ip_summed != CHECKSUM_PARTIAL)) {
2263 err = ptype->callbacks.gso_send_check(skb);
2264 segs = ERR_PTR(err);
2265 if (err || skb_gso_ok(skb, features))
2267 __skb_push(skb, (skb->data -
2268 skb_network_header(skb)));
2270 segs = ptype->callbacks.gso_segment(skb, features);
2276 __skb_push(skb, skb->data - skb_mac_header(skb));
2280 EXPORT_SYMBOL(skb_mac_gso_segment);
2283 /* openvswitch calls this on rx path, so we need a different check.
2285 static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
2288 return skb->ip_summed != CHECKSUM_PARTIAL;
2290 return skb->ip_summed == CHECKSUM_NONE;
2294 * __skb_gso_segment - Perform segmentation on skb.
2295 * @skb: buffer to segment
2296 * @features: features for the output path (see dev->features)
2297 * @tx_path: whether it is called in TX path
2299 * This function segments the given skb and returns a list of segments.
2301 * It may return NULL if the skb requires no segmentation. This is
2302 * only possible when GSO is used for verifying header integrity.
2304 struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
2305 netdev_features_t features, bool tx_path)
2307 if (unlikely(skb_needs_check(skb, tx_path))) {
2310 skb_warn_bad_offload(skb);
2312 if (skb_header_cloned(skb) &&
2313 (err = pskb_expand_head(skb, 0, 0, GFP_ATOMIC)))
2314 return ERR_PTR(err);
2317 SKB_GSO_CB(skb)->mac_offset = skb_headroom(skb);
2318 skb_reset_mac_header(skb);
2319 skb_reset_mac_len(skb);
2321 return skb_mac_gso_segment(skb, features);
2323 EXPORT_SYMBOL(__skb_gso_segment);
2325 /* Take action when hardware reception checksum errors are detected. */
2327 void netdev_rx_csum_fault(struct net_device *dev)
2329 if (net_ratelimit()) {
2330 pr_err("%s: hw csum failure\n", dev ? dev->name : "<unknown>");
2334 EXPORT_SYMBOL(netdev_rx_csum_fault);
2337 /* Actually, we should eliminate this check as soon as we know, that:
2338 * 1. IOMMU is present and allows to map all the memory.
2339 * 2. No high memory really exists on this machine.
2342 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
2344 #ifdef CONFIG_HIGHMEM
2346 if (!(dev->features & NETIF_F_HIGHDMA)) {
2347 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2348 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2349 if (PageHighMem(skb_frag_page(frag)))
2354 if (PCI_DMA_BUS_IS_PHYS) {
2355 struct device *pdev = dev->dev.parent;
2359 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2360 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2361 dma_addr_t addr = page_to_phys(skb_frag_page(frag));
2362 if (!pdev->dma_mask || addr + PAGE_SIZE - 1 > *pdev->dma_mask)
2371 void (*destructor)(struct sk_buff *skb);
2374 #define DEV_GSO_CB(skb) ((struct dev_gso_cb *)(skb)->cb)
2376 static void dev_gso_skb_destructor(struct sk_buff *skb)
2378 struct dev_gso_cb *cb;
2381 struct sk_buff *nskb = skb->next;
2383 skb->next = nskb->next;
2386 } while (skb->next);
2388 cb = DEV_GSO_CB(skb);
2390 cb->destructor(skb);
2394 * dev_gso_segment - Perform emulated hardware segmentation on skb.
2395 * @skb: buffer to segment
2396 * @features: device features as applicable to this skb
2398 * This function segments the given skb and stores the list of segments
2401 static int dev_gso_segment(struct sk_buff *skb, netdev_features_t features)
2403 struct sk_buff *segs;
2405 segs = skb_gso_segment(skb, features);
2407 /* Verifying header integrity only. */
2412 return PTR_ERR(segs);
2415 DEV_GSO_CB(skb)->destructor = skb->destructor;
2416 skb->destructor = dev_gso_skb_destructor;
2421 static netdev_features_t harmonize_features(struct sk_buff *skb,
2422 __be16 protocol, netdev_features_t features)
2424 if (skb->ip_summed != CHECKSUM_NONE &&
2425 !can_checksum_protocol(features, protocol)) {
2426 features &= ~NETIF_F_ALL_CSUM;
2427 } else if (illegal_highdma(skb->dev, skb)) {
2428 features &= ~NETIF_F_SG;
2434 netdev_features_t netif_skb_features(struct sk_buff *skb)
2436 __be16 protocol = skb->protocol;
2437 netdev_features_t features = skb->dev->features;
2439 if (skb_shinfo(skb)->gso_segs > skb->dev->gso_max_segs)
2440 features &= ~NETIF_F_GSO_MASK;
2442 if (protocol == htons(ETH_P_8021Q) || protocol == htons(ETH_P_8021AD)) {
2443 struct vlan_ethhdr *veh = (struct vlan_ethhdr *)skb->data;
2444 protocol = veh->h_vlan_encapsulated_proto;
2445 } else if (!vlan_tx_tag_present(skb)) {
2446 return harmonize_features(skb, protocol, features);
2449 features &= (skb->dev->vlan_features | NETIF_F_HW_VLAN_CTAG_TX |
2450 NETIF_F_HW_VLAN_STAG_TX);
2452 if (protocol != htons(ETH_P_8021Q) && protocol != htons(ETH_P_8021AD)) {
2453 return harmonize_features(skb, protocol, features);
2455 features &= NETIF_F_SG | NETIF_F_HIGHDMA | NETIF_F_FRAGLIST |
2456 NETIF_F_GEN_CSUM | NETIF_F_HW_VLAN_CTAG_TX |
2457 NETIF_F_HW_VLAN_STAG_TX;
2458 return harmonize_features(skb, protocol, features);
2461 EXPORT_SYMBOL(netif_skb_features);
2464 * Returns true if either:
2465 * 1. skb has frag_list and the device doesn't support FRAGLIST, or
2466 * 2. skb is fragmented and the device does not support SG.
2468 static inline int skb_needs_linearize(struct sk_buff *skb,
2469 netdev_features_t features)
2471 return skb_is_nonlinear(skb) &&
2472 ((skb_has_frag_list(skb) &&
2473 !(features & NETIF_F_FRAGLIST)) ||
2474 (skb_shinfo(skb)->nr_frags &&
2475 !(features & NETIF_F_SG)));
2478 int dev_hard_start_xmit(struct sk_buff *skb, struct net_device *dev,
2479 struct netdev_queue *txq)
2481 const struct net_device_ops *ops = dev->netdev_ops;
2482 int rc = NETDEV_TX_OK;
2483 unsigned int skb_len;
2485 if (likely(!skb->next)) {
2486 netdev_features_t features;
2489 * If device doesn't need skb->dst, release it right now while
2490 * its hot in this cpu cache
2492 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
2495 features = netif_skb_features(skb);
2497 if (vlan_tx_tag_present(skb) &&
2498 !vlan_hw_offload_capable(features, skb->vlan_proto)) {
2499 skb = __vlan_put_tag(skb, skb->vlan_proto,
2500 vlan_tx_tag_get(skb));
2507 /* If encapsulation offload request, verify we are testing
2508 * hardware encapsulation features instead of standard
2509 * features for the netdev
2511 if (skb->encapsulation)
2512 features &= dev->hw_enc_features;
2514 if (netif_needs_gso(skb, features)) {
2515 if (unlikely(dev_gso_segment(skb, features)))
2520 if (skb_needs_linearize(skb, features) &&
2521 __skb_linearize(skb))
2524 /* If packet is not checksummed and device does not
2525 * support checksumming for this protocol, complete
2526 * checksumming here.
2528 if (skb->ip_summed == CHECKSUM_PARTIAL) {
2529 if (skb->encapsulation)
2530 skb_set_inner_transport_header(skb,
2531 skb_checksum_start_offset(skb));
2533 skb_set_transport_header(skb,
2534 skb_checksum_start_offset(skb));
2535 if (!(features & NETIF_F_ALL_CSUM) &&
2536 skb_checksum_help(skb))
2541 if (!list_empty(&ptype_all))
2542 dev_queue_xmit_nit(skb, dev);
2545 rc = ops->ndo_start_xmit(skb, dev);
2546 trace_net_dev_xmit(skb, rc, dev, skb_len);
2547 if (rc == NETDEV_TX_OK)
2548 txq_trans_update(txq);
2554 struct sk_buff *nskb = skb->next;
2556 skb->next = nskb->next;
2559 if (!list_empty(&ptype_all))
2560 dev_queue_xmit_nit(nskb, dev);
2562 skb_len = nskb->len;
2563 rc = ops->ndo_start_xmit(nskb, dev);
2564 trace_net_dev_xmit(nskb, rc, dev, skb_len);
2565 if (unlikely(rc != NETDEV_TX_OK)) {
2566 if (rc & ~NETDEV_TX_MASK)
2567 goto out_kfree_gso_skb;
2568 nskb->next = skb->next;
2572 txq_trans_update(txq);
2573 if (unlikely(netif_xmit_stopped(txq) && skb->next))
2574 return NETDEV_TX_BUSY;
2575 } while (skb->next);
2578 if (likely(skb->next == NULL)) {
2579 skb->destructor = DEV_GSO_CB(skb)->destructor;
2589 static void qdisc_pkt_len_init(struct sk_buff *skb)
2591 const struct skb_shared_info *shinfo = skb_shinfo(skb);
2593 qdisc_skb_cb(skb)->pkt_len = skb->len;
2595 /* To get more precise estimation of bytes sent on wire,
2596 * we add to pkt_len the headers size of all segments
2598 if (shinfo->gso_size) {
2599 unsigned int hdr_len;
2600 u16 gso_segs = shinfo->gso_segs;
2602 /* mac layer + network layer */
2603 hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
2605 /* + transport layer */
2606 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))
2607 hdr_len += tcp_hdrlen(skb);
2609 hdr_len += sizeof(struct udphdr);
2611 if (shinfo->gso_type & SKB_GSO_DODGY)
2612 gso_segs = DIV_ROUND_UP(skb->len - hdr_len,
2615 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len;
2619 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
2620 struct net_device *dev,
2621 struct netdev_queue *txq)
2623 spinlock_t *root_lock = qdisc_lock(q);
2627 qdisc_pkt_len_init(skb);
2628 qdisc_calculate_pkt_len(skb, q);
2630 * Heuristic to force contended enqueues to serialize on a
2631 * separate lock before trying to get qdisc main lock.
2632 * This permits __QDISC_STATE_RUNNING owner to get the lock more often
2633 * and dequeue packets faster.
2635 contended = qdisc_is_running(q);
2636 if (unlikely(contended))
2637 spin_lock(&q->busylock);
2639 spin_lock(root_lock);
2640 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
2643 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
2644 qdisc_run_begin(q)) {
2646 * This is a work-conserving queue; there are no old skbs
2647 * waiting to be sent out; and the qdisc is not running -
2648 * xmit the skb directly.
2650 if (!(dev->priv_flags & IFF_XMIT_DST_RELEASE))
2653 qdisc_bstats_update(q, skb);
2655 if (sch_direct_xmit(skb, q, dev, txq, root_lock)) {
2656 if (unlikely(contended)) {
2657 spin_unlock(&q->busylock);
2664 rc = NET_XMIT_SUCCESS;
2667 rc = q->enqueue(skb, q) & NET_XMIT_MASK;
2668 if (qdisc_run_begin(q)) {
2669 if (unlikely(contended)) {
2670 spin_unlock(&q->busylock);
2676 spin_unlock(root_lock);
2677 if (unlikely(contended))
2678 spin_unlock(&q->busylock);
2682 #if IS_ENABLED(CONFIG_NETPRIO_CGROUP)
2683 static void skb_update_prio(struct sk_buff *skb)
2685 struct netprio_map *map = rcu_dereference_bh(skb->dev->priomap);
2687 if (!skb->priority && skb->sk && map) {
2688 unsigned int prioidx = skb->sk->sk_cgrp_prioidx;
2690 if (prioidx < map->priomap_len)
2691 skb->priority = map->priomap[prioidx];
2695 #define skb_update_prio(skb)
2698 static DEFINE_PER_CPU(int, xmit_recursion);
2699 #define RECURSION_LIMIT 10
2702 * dev_loopback_xmit - loop back @skb
2703 * @skb: buffer to transmit
2705 int dev_loopback_xmit(struct sk_buff *skb)
2707 skb_reset_mac_header(skb);
2708 __skb_pull(skb, skb_network_offset(skb));
2709 skb->pkt_type = PACKET_LOOPBACK;
2710 skb->ip_summed = CHECKSUM_UNNECESSARY;
2711 WARN_ON(!skb_dst(skb));
2716 EXPORT_SYMBOL(dev_loopback_xmit);
2719 * dev_queue_xmit - transmit a buffer
2720 * @skb: buffer to transmit
2722 * Queue a buffer for transmission to a network device. The caller must
2723 * have set the device and priority and built the buffer before calling
2724 * this function. The function can be called from an interrupt.
2726 * A negative errno code is returned on a failure. A success does not
2727 * guarantee the frame will be transmitted as it may be dropped due
2728 * to congestion or traffic shaping.
2730 * -----------------------------------------------------------------------------------
2731 * I notice this method can also return errors from the queue disciplines,
2732 * including NET_XMIT_DROP, which is a positive value. So, errors can also
2735 * Regardless of the return value, the skb is consumed, so it is currently
2736 * difficult to retry a send to this method. (You can bump the ref count
2737 * before sending to hold a reference for retry if you are careful.)
2739 * When calling this method, interrupts MUST be enabled. This is because
2740 * the BH enable code must have IRQs enabled so that it will not deadlock.
2743 int dev_queue_xmit(struct sk_buff *skb)
2745 struct net_device *dev = skb->dev;
2746 struct netdev_queue *txq;
2750 skb_reset_mac_header(skb);
2752 /* Disable soft irqs for various locks below. Also
2753 * stops preemption for RCU.
2757 skb_update_prio(skb);
2759 txq = netdev_pick_tx(dev, skb);
2760 q = rcu_dereference_bh(txq->qdisc);
2762 #ifdef CONFIG_NET_CLS_ACT
2763 skb->tc_verd = SET_TC_AT(skb->tc_verd, AT_EGRESS);
2765 trace_net_dev_queue(skb);
2767 rc = __dev_xmit_skb(skb, q, dev, txq);
2771 /* The device has no queue. Common case for software devices:
2772 loopback, all the sorts of tunnels...
2774 Really, it is unlikely that netif_tx_lock protection is necessary
2775 here. (f.e. loopback and IP tunnels are clean ignoring statistics
2777 However, it is possible, that they rely on protection
2780 Check this and shot the lock. It is not prone from deadlocks.
2781 Either shot noqueue qdisc, it is even simpler 8)
2783 if (dev->flags & IFF_UP) {
2784 int cpu = smp_processor_id(); /* ok because BHs are off */
2786 if (txq->xmit_lock_owner != cpu) {
2788 if (__this_cpu_read(xmit_recursion) > RECURSION_LIMIT)
2789 goto recursion_alert;
2791 HARD_TX_LOCK(dev, txq, cpu);
2793 if (!netif_xmit_stopped(txq)) {
2794 __this_cpu_inc(xmit_recursion);
2795 rc = dev_hard_start_xmit(skb, dev, txq);
2796 __this_cpu_dec(xmit_recursion);
2797 if (dev_xmit_complete(rc)) {
2798 HARD_TX_UNLOCK(dev, txq);
2802 HARD_TX_UNLOCK(dev, txq);
2803 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
2806 /* Recursion is detected! It is possible,
2810 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
2816 rcu_read_unlock_bh();
2821 rcu_read_unlock_bh();
2824 EXPORT_SYMBOL(dev_queue_xmit);
2827 /*=======================================================================
2829 =======================================================================*/
2831 int netdev_max_backlog __read_mostly = 1000;
2832 EXPORT_SYMBOL(netdev_max_backlog);
2834 int netdev_tstamp_prequeue __read_mostly = 1;
2835 int netdev_budget __read_mostly = 300;
2836 int weight_p __read_mostly = 64; /* old backlog weight */
2838 /* Called with irq disabled */
2839 static inline void ____napi_schedule(struct softnet_data *sd,
2840 struct napi_struct *napi)
2842 list_add_tail(&napi->poll_list, &sd->poll_list);
2843 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
2848 /* One global table that all flow-based protocols share. */
2849 struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
2850 EXPORT_SYMBOL(rps_sock_flow_table);
2852 struct static_key rps_needed __read_mostly;
2854 static struct rps_dev_flow *
2855 set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
2856 struct rps_dev_flow *rflow, u16 next_cpu)
2858 if (next_cpu != RPS_NO_CPU) {
2859 #ifdef CONFIG_RFS_ACCEL
2860 struct netdev_rx_queue *rxqueue;
2861 struct rps_dev_flow_table *flow_table;
2862 struct rps_dev_flow *old_rflow;
2867 /* Should we steer this flow to a different hardware queue? */
2868 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
2869 !(dev->features & NETIF_F_NTUPLE))
2871 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
2872 if (rxq_index == skb_get_rx_queue(skb))
2875 rxqueue = dev->_rx + rxq_index;
2876 flow_table = rcu_dereference(rxqueue->rps_flow_table);
2879 flow_id = skb->rxhash & flow_table->mask;
2880 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
2881 rxq_index, flow_id);
2885 rflow = &flow_table->flows[flow_id];
2887 if (old_rflow->filter == rflow->filter)
2888 old_rflow->filter = RPS_NO_FILTER;
2892 per_cpu(softnet_data, next_cpu).input_queue_head;
2895 rflow->cpu = next_cpu;
2900 * get_rps_cpu is called from netif_receive_skb and returns the target
2901 * CPU from the RPS map of the receiving queue for a given skb.
2902 * rcu_read_lock must be held on entry.
2904 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
2905 struct rps_dev_flow **rflowp)
2907 struct netdev_rx_queue *rxqueue;
2908 struct rps_map *map;
2909 struct rps_dev_flow_table *flow_table;
2910 struct rps_sock_flow_table *sock_flow_table;
2914 if (skb_rx_queue_recorded(skb)) {
2915 u16 index = skb_get_rx_queue(skb);
2916 if (unlikely(index >= dev->real_num_rx_queues)) {
2917 WARN_ONCE(dev->real_num_rx_queues > 1,
2918 "%s received packet on queue %u, but number "
2919 "of RX queues is %u\n",
2920 dev->name, index, dev->real_num_rx_queues);
2923 rxqueue = dev->_rx + index;
2927 map = rcu_dereference(rxqueue->rps_map);
2929 if (map->len == 1 &&
2930 !rcu_access_pointer(rxqueue->rps_flow_table)) {
2931 tcpu = map->cpus[0];
2932 if (cpu_online(tcpu))
2936 } else if (!rcu_access_pointer(rxqueue->rps_flow_table)) {
2940 skb_reset_network_header(skb);
2941 if (!skb_get_rxhash(skb))
2944 flow_table = rcu_dereference(rxqueue->rps_flow_table);
2945 sock_flow_table = rcu_dereference(rps_sock_flow_table);
2946 if (flow_table && sock_flow_table) {
2948 struct rps_dev_flow *rflow;
2950 rflow = &flow_table->flows[skb->rxhash & flow_table->mask];
2953 next_cpu = sock_flow_table->ents[skb->rxhash &
2954 sock_flow_table->mask];
2957 * If the desired CPU (where last recvmsg was done) is
2958 * different from current CPU (one in the rx-queue flow
2959 * table entry), switch if one of the following holds:
2960 * - Current CPU is unset (equal to RPS_NO_CPU).
2961 * - Current CPU is offline.
2962 * - The current CPU's queue tail has advanced beyond the
2963 * last packet that was enqueued using this table entry.
2964 * This guarantees that all previous packets for the flow
2965 * have been dequeued, thus preserving in order delivery.
2967 if (unlikely(tcpu != next_cpu) &&
2968 (tcpu == RPS_NO_CPU || !cpu_online(tcpu) ||
2969 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
2970 rflow->last_qtail)) >= 0)) {
2972 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
2975 if (tcpu != RPS_NO_CPU && cpu_online(tcpu)) {
2983 tcpu = map->cpus[((u64) skb->rxhash * map->len) >> 32];
2985 if (cpu_online(tcpu)) {
2995 #ifdef CONFIG_RFS_ACCEL
2998 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
2999 * @dev: Device on which the filter was set
3000 * @rxq_index: RX queue index
3001 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
3002 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
3004 * Drivers that implement ndo_rx_flow_steer() should periodically call
3005 * this function for each installed filter and remove the filters for
3006 * which it returns %true.
3008 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
3009 u32 flow_id, u16 filter_id)
3011 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
3012 struct rps_dev_flow_table *flow_table;
3013 struct rps_dev_flow *rflow;
3018 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3019 if (flow_table && flow_id <= flow_table->mask) {
3020 rflow = &flow_table->flows[flow_id];
3021 cpu = ACCESS_ONCE(rflow->cpu);
3022 if (rflow->filter == filter_id && cpu != RPS_NO_CPU &&
3023 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
3024 rflow->last_qtail) <
3025 (int)(10 * flow_table->mask)))
3031 EXPORT_SYMBOL(rps_may_expire_flow);
3033 #endif /* CONFIG_RFS_ACCEL */
3035 /* Called from hardirq (IPI) context */
3036 static void rps_trigger_softirq(void *data)
3038 struct softnet_data *sd = data;
3040 ____napi_schedule(sd, &sd->backlog);
3044 #endif /* CONFIG_RPS */
3047 * Check if this softnet_data structure is another cpu one
3048 * If yes, queue it to our IPI list and return 1
3051 static int rps_ipi_queued(struct softnet_data *sd)
3054 struct softnet_data *mysd = &__get_cpu_var(softnet_data);
3057 sd->rps_ipi_next = mysd->rps_ipi_list;
3058 mysd->rps_ipi_list = sd;
3060 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3063 #endif /* CONFIG_RPS */
3067 #ifdef CONFIG_NET_FLOW_LIMIT
3068 int netdev_flow_limit_table_len __read_mostly = (1 << 12);
3071 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
3073 #ifdef CONFIG_NET_FLOW_LIMIT
3074 struct sd_flow_limit *fl;
3075 struct softnet_data *sd;
3076 unsigned int old_flow, new_flow;
3078 if (qlen < (netdev_max_backlog >> 1))
3081 sd = &__get_cpu_var(softnet_data);
3084 fl = rcu_dereference(sd->flow_limit);
3086 new_flow = skb_get_rxhash(skb) & (fl->num_buckets - 1);
3087 old_flow = fl->history[fl->history_head];
3088 fl->history[fl->history_head] = new_flow;
3091 fl->history_head &= FLOW_LIMIT_HISTORY - 1;
3093 if (likely(fl->buckets[old_flow]))
3094 fl->buckets[old_flow]--;
3096 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) {
3108 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
3109 * queue (may be a remote CPU queue).
3111 static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
3112 unsigned int *qtail)
3114 struct softnet_data *sd;
3115 unsigned long flags;
3118 sd = &per_cpu(softnet_data, cpu);
3120 local_irq_save(flags);
3123 qlen = skb_queue_len(&sd->input_pkt_queue);
3124 if (qlen <= netdev_max_backlog && !skb_flow_limit(skb, qlen)) {
3125 if (skb_queue_len(&sd->input_pkt_queue)) {
3127 __skb_queue_tail(&sd->input_pkt_queue, skb);
3128 input_queue_tail_incr_save(sd, qtail);
3130 local_irq_restore(flags);
3131 return NET_RX_SUCCESS;
3134 /* Schedule NAPI for backlog device
3135 * We can use non atomic operation since we own the queue lock
3137 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
3138 if (!rps_ipi_queued(sd))
3139 ____napi_schedule(sd, &sd->backlog);
3147 local_irq_restore(flags);
3149 atomic_long_inc(&skb->dev->rx_dropped);
3155 * netif_rx - post buffer to the network code
3156 * @skb: buffer to post
3158 * This function receives a packet from a device driver and queues it for
3159 * the upper (protocol) levels to process. It always succeeds. The buffer
3160 * may be dropped during processing for congestion control or by the
3164 * NET_RX_SUCCESS (no congestion)
3165 * NET_RX_DROP (packet was dropped)
3169 int netif_rx(struct sk_buff *skb)
3173 /* if netpoll wants it, pretend we never saw it */
3174 if (netpoll_rx(skb))
3177 net_timestamp_check(netdev_tstamp_prequeue, skb);
3179 trace_netif_rx(skb);
3181 if (static_key_false(&rps_needed)) {
3182 struct rps_dev_flow voidflow, *rflow = &voidflow;
3188 cpu = get_rps_cpu(skb->dev, skb, &rflow);
3190 cpu = smp_processor_id();
3192 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
3200 ret = enqueue_to_backlog(skb, get_cpu(), &qtail);
3205 EXPORT_SYMBOL(netif_rx);
3207 int netif_rx_ni(struct sk_buff *skb)
3212 err = netif_rx(skb);
3213 if (local_softirq_pending())
3219 EXPORT_SYMBOL(netif_rx_ni);
3221 static void net_tx_action(struct softirq_action *h)
3223 struct softnet_data *sd = &__get_cpu_var(softnet_data);
3225 if (sd->completion_queue) {
3226 struct sk_buff *clist;
3228 local_irq_disable();
3229 clist = sd->completion_queue;
3230 sd->completion_queue = NULL;
3234 struct sk_buff *skb = clist;
3235 clist = clist->next;
3237 WARN_ON(atomic_read(&skb->users));
3238 trace_kfree_skb(skb, net_tx_action);
3243 if (sd->output_queue) {
3246 local_irq_disable();
3247 head = sd->output_queue;
3248 sd->output_queue = NULL;
3249 sd->output_queue_tailp = &sd->output_queue;
3253 struct Qdisc *q = head;
3254 spinlock_t *root_lock;
3256 head = head->next_sched;
3258 root_lock = qdisc_lock(q);
3259 if (spin_trylock(root_lock)) {
3260 smp_mb__before_clear_bit();
3261 clear_bit(__QDISC_STATE_SCHED,
3264 spin_unlock(root_lock);
3266 if (!test_bit(__QDISC_STATE_DEACTIVATED,
3268 __netif_reschedule(q);
3270 smp_mb__before_clear_bit();
3271 clear_bit(__QDISC_STATE_SCHED,
3279 #if (defined(CONFIG_BRIDGE) || defined(CONFIG_BRIDGE_MODULE)) && \
3280 (defined(CONFIG_ATM_LANE) || defined(CONFIG_ATM_LANE_MODULE))
3281 /* This hook is defined here for ATM LANE */
3282 int (*br_fdb_test_addr_hook)(struct net_device *dev,
3283 unsigned char *addr) __read_mostly;
3284 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook);
3287 #ifdef CONFIG_NET_CLS_ACT
3288 /* TODO: Maybe we should just force sch_ingress to be compiled in
3289 * when CONFIG_NET_CLS_ACT is? otherwise some useless instructions
3290 * a compare and 2 stores extra right now if we dont have it on
3291 * but have CONFIG_NET_CLS_ACT
3292 * NOTE: This doesn't stop any functionality; if you dont have
3293 * the ingress scheduler, you just can't add policies on ingress.
3296 static int ing_filter(struct sk_buff *skb, struct netdev_queue *rxq)
3298 struct net_device *dev = skb->dev;
3299 u32 ttl = G_TC_RTTL(skb->tc_verd);
3300 int result = TC_ACT_OK;
3303 if (unlikely(MAX_RED_LOOP < ttl++)) {
3304 net_warn_ratelimited("Redir loop detected Dropping packet (%d->%d)\n",
3305 skb->skb_iif, dev->ifindex);
3309 skb->tc_verd = SET_TC_RTTL(skb->tc_verd, ttl);
3310 skb->tc_verd = SET_TC_AT(skb->tc_verd, AT_INGRESS);
3313 if (q != &noop_qdisc) {
3314 spin_lock(qdisc_lock(q));
3315 if (likely(!test_bit(__QDISC_STATE_DEACTIVATED, &q->state)))
3316 result = qdisc_enqueue_root(skb, q);
3317 spin_unlock(qdisc_lock(q));
3323 static inline struct sk_buff *handle_ing(struct sk_buff *skb,
3324 struct packet_type **pt_prev,
3325 int *ret, struct net_device *orig_dev)
3327 struct netdev_queue *rxq = rcu_dereference(skb->dev->ingress_queue);
3329 if (!rxq || rxq->qdisc == &noop_qdisc)
3333 *ret = deliver_skb(skb, *pt_prev, orig_dev);
3337 switch (ing_filter(skb, rxq)) {
3351 * netdev_rx_handler_register - register receive handler
3352 * @dev: device to register a handler for
3353 * @rx_handler: receive handler to register
3354 * @rx_handler_data: data pointer that is used by rx handler
3356 * Register a receive hander for a device. This handler will then be
3357 * called from __netif_receive_skb. A negative errno code is returned
3360 * The caller must hold the rtnl_mutex.
3362 * For a general description of rx_handler, see enum rx_handler_result.
3364 int netdev_rx_handler_register(struct net_device *dev,
3365 rx_handler_func_t *rx_handler,
3366 void *rx_handler_data)
3370 if (dev->rx_handler)
3373 /* Note: rx_handler_data must be set before rx_handler */
3374 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data);
3375 rcu_assign_pointer(dev->rx_handler, rx_handler);
3379 EXPORT_SYMBOL_GPL(netdev_rx_handler_register);
3382 * netdev_rx_handler_unregister - unregister receive handler
3383 * @dev: device to unregister a handler from
3385 * Unregister a receive handler from a device.
3387 * The caller must hold the rtnl_mutex.
3389 void netdev_rx_handler_unregister(struct net_device *dev)
3393 RCU_INIT_POINTER(dev->rx_handler, NULL);
3394 /* a reader seeing a non NULL rx_handler in a rcu_read_lock()
3395 * section has a guarantee to see a non NULL rx_handler_data
3399 RCU_INIT_POINTER(dev->rx_handler_data, NULL);
3401 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister);
3404 * Limit the use of PFMEMALLOC reserves to those protocols that implement
3405 * the special handling of PFMEMALLOC skbs.
3407 static bool skb_pfmemalloc_protocol(struct sk_buff *skb)
3409 switch (skb->protocol) {
3410 case __constant_htons(ETH_P_ARP):
3411 case __constant_htons(ETH_P_IP):
3412 case __constant_htons(ETH_P_IPV6):
3413 case __constant_htons(ETH_P_8021Q):
3414 case __constant_htons(ETH_P_8021AD):
3421 static int __netif_receive_skb_core(struct sk_buff *skb, bool pfmemalloc)
3423 struct packet_type *ptype, *pt_prev;
3424 rx_handler_func_t *rx_handler;
3425 struct net_device *orig_dev;
3426 struct net_device *null_or_dev;
3427 bool deliver_exact = false;
3428 int ret = NET_RX_DROP;
3431 net_timestamp_check(!netdev_tstamp_prequeue, skb);
3433 trace_netif_receive_skb(skb);
3435 /* if we've gotten here through NAPI, check netpoll */
3436 if (netpoll_receive_skb(skb))
3439 orig_dev = skb->dev;
3441 skb_reset_network_header(skb);
3442 if (!skb_transport_header_was_set(skb))
3443 skb_reset_transport_header(skb);
3444 skb_reset_mac_len(skb);
3451 skb->skb_iif = skb->dev->ifindex;
3453 __this_cpu_inc(softnet_data.processed);
3455 if (skb->protocol == cpu_to_be16(ETH_P_8021Q) ||
3456 skb->protocol == cpu_to_be16(ETH_P_8021AD)) {
3457 skb = vlan_untag(skb);
3462 #ifdef CONFIG_NET_CLS_ACT
3463 if (skb->tc_verd & TC_NCLS) {
3464 skb->tc_verd = CLR_TC_NCLS(skb->tc_verd);
3472 list_for_each_entry_rcu(ptype, &ptype_all, list) {
3473 if (!ptype->dev || ptype->dev == skb->dev) {
3475 ret = deliver_skb(skb, pt_prev, orig_dev);
3481 #ifdef CONFIG_NET_CLS_ACT
3482 skb = handle_ing(skb, &pt_prev, &ret, orig_dev);
3488 if (pfmemalloc && !skb_pfmemalloc_protocol(skb))
3491 if (vlan_tx_tag_present(skb)) {
3493 ret = deliver_skb(skb, pt_prev, orig_dev);
3496 if (vlan_do_receive(&skb))
3498 else if (unlikely(!skb))
3502 rx_handler = rcu_dereference(skb->dev->rx_handler);
3505 ret = deliver_skb(skb, pt_prev, orig_dev);
3508 switch (rx_handler(&skb)) {
3509 case RX_HANDLER_CONSUMED:
3510 ret = NET_RX_SUCCESS;
3512 case RX_HANDLER_ANOTHER:
3514 case RX_HANDLER_EXACT:
3515 deliver_exact = true;
3516 case RX_HANDLER_PASS:
3523 if (vlan_tx_nonzero_tag_present(skb))
3524 skb->pkt_type = PACKET_OTHERHOST;
3526 /* deliver only exact match when indicated */
3527 null_or_dev = deliver_exact ? skb->dev : NULL;
3529 type = skb->protocol;
3530 list_for_each_entry_rcu(ptype,
3531 &ptype_base[ntohs(type) & PTYPE_HASH_MASK], list) {
3532 if (ptype->type == type &&
3533 (ptype->dev == null_or_dev || ptype->dev == skb->dev ||
3534 ptype->dev == orig_dev)) {
3536 ret = deliver_skb(skb, pt_prev, orig_dev);
3542 if (unlikely(skb_orphan_frags(skb, GFP_ATOMIC)))
3545 ret = pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
3548 atomic_long_inc(&skb->dev->rx_dropped);
3550 /* Jamal, now you will not able to escape explaining
3551 * me how you were going to use this. :-)
3562 static int __netif_receive_skb(struct sk_buff *skb)
3566 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) {
3567 unsigned long pflags = current->flags;
3570 * PFMEMALLOC skbs are special, they should
3571 * - be delivered to SOCK_MEMALLOC sockets only
3572 * - stay away from userspace
3573 * - have bounded memory usage
3575 * Use PF_MEMALLOC as this saves us from propagating the allocation
3576 * context down to all allocation sites.
3578 current->flags |= PF_MEMALLOC;
3579 ret = __netif_receive_skb_core(skb, true);
3580 tsk_restore_flags(current, pflags, PF_MEMALLOC);
3582 ret = __netif_receive_skb_core(skb, false);
3588 * netif_receive_skb - process receive buffer from network
3589 * @skb: buffer to process
3591 * netif_receive_skb() is the main receive data processing function.
3592 * It always succeeds. The buffer may be dropped during processing
3593 * for congestion control or by the protocol layers.
3595 * This function may only be called from softirq context and interrupts
3596 * should be enabled.
3598 * Return values (usually ignored):
3599 * NET_RX_SUCCESS: no congestion
3600 * NET_RX_DROP: packet was dropped
3602 int netif_receive_skb(struct sk_buff *skb)
3604 net_timestamp_check(netdev_tstamp_prequeue, skb);
3606 if (skb_defer_rx_timestamp(skb))
3607 return NET_RX_SUCCESS;
3610 if (static_key_false(&rps_needed)) {
3611 struct rps_dev_flow voidflow, *rflow = &voidflow;
3616 cpu = get_rps_cpu(skb->dev, skb, &rflow);
3619 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
3626 return __netif_receive_skb(skb);
3628 EXPORT_SYMBOL(netif_receive_skb);
3630 /* Network device is going away, flush any packets still pending
3631 * Called with irqs disabled.
3633 static void flush_backlog(void *arg)
3635 struct net_device *dev = arg;
3636 struct softnet_data *sd = &__get_cpu_var(softnet_data);
3637 struct sk_buff *skb, *tmp;
3640 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) {
3641 if (skb->dev == dev) {
3642 __skb_unlink(skb, &sd->input_pkt_queue);
3644 input_queue_head_incr(sd);
3649 skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
3650 if (skb->dev == dev) {
3651 __skb_unlink(skb, &sd->process_queue);
3653 input_queue_head_incr(sd);
3658 static int napi_gro_complete(struct sk_buff *skb)
3660 struct packet_offload *ptype;
3661 __be16 type = skb->protocol;
3662 struct list_head *head = &offload_base;
3665 BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
3667 if (NAPI_GRO_CB(skb)->count == 1) {
3668 skb_shinfo(skb)->gso_size = 0;
3673 list_for_each_entry_rcu(ptype, head, list) {
3674 if (ptype->type != type || !ptype->callbacks.gro_complete)
3677 err = ptype->callbacks.gro_complete(skb);
3683 WARN_ON(&ptype->list == head);
3685 return NET_RX_SUCCESS;
3689 return netif_receive_skb(skb);
3692 /* napi->gro_list contains packets ordered by age.
3693 * youngest packets at the head of it.
3694 * Complete skbs in reverse order to reduce latencies.
3696 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
3698 struct sk_buff *skb, *prev = NULL;
3700 /* scan list and build reverse chain */
3701 for (skb = napi->gro_list; skb != NULL; skb = skb->next) {
3706 for (skb = prev; skb; skb = prev) {
3709 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
3713 napi_gro_complete(skb);
3717 napi->gro_list = NULL;
3719 EXPORT_SYMBOL(napi_gro_flush);
3721 static void gro_list_prepare(struct napi_struct *napi, struct sk_buff *skb)
3724 unsigned int maclen = skb->dev->hard_header_len;
3726 for (p = napi->gro_list; p; p = p->next) {
3727 unsigned long diffs;
3729 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
3730 diffs |= p->vlan_tci ^ skb->vlan_tci;
3731 if (maclen == ETH_HLEN)
3732 diffs |= compare_ether_header(skb_mac_header(p),
3733 skb_gro_mac_header(skb));
3735 diffs = memcmp(skb_mac_header(p),
3736 skb_gro_mac_header(skb),
3738 NAPI_GRO_CB(p)->same_flow = !diffs;
3739 NAPI_GRO_CB(p)->flush = 0;
3743 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
3745 struct sk_buff **pp = NULL;
3746 struct packet_offload *ptype;
3747 __be16 type = skb->protocol;
3748 struct list_head *head = &offload_base;
3750 enum gro_result ret;
3752 if (!(skb->dev->features & NETIF_F_GRO) || netpoll_rx_on(skb))
3755 if (skb_is_gso(skb) || skb_has_frag_list(skb))
3758 gro_list_prepare(napi, skb);
3761 list_for_each_entry_rcu(ptype, head, list) {
3762 if (ptype->type != type || !ptype->callbacks.gro_receive)
3765 skb_set_network_header(skb, skb_gro_offset(skb));
3766 skb_reset_mac_len(skb);
3767 NAPI_GRO_CB(skb)->same_flow = 0;
3768 NAPI_GRO_CB(skb)->flush = 0;
3769 NAPI_GRO_CB(skb)->free = 0;
3771 pp = ptype->callbacks.gro_receive(&napi->gro_list, skb);
3776 if (&ptype->list == head)
3779 same_flow = NAPI_GRO_CB(skb)->same_flow;
3780 ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
3783 struct sk_buff *nskb = *pp;
3787 napi_gro_complete(nskb);
3794 if (NAPI_GRO_CB(skb)->flush || napi->gro_count >= MAX_GRO_SKBS)
3798 NAPI_GRO_CB(skb)->count = 1;
3799 NAPI_GRO_CB(skb)->age = jiffies;
3800 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
3801 skb->next = napi->gro_list;
3802 napi->gro_list = skb;
3806 if (skb_headlen(skb) < skb_gro_offset(skb)) {
3807 int grow = skb_gro_offset(skb) - skb_headlen(skb);
3809 BUG_ON(skb->end - skb->tail < grow);
3811 memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
3814 skb->data_len -= grow;
3816 skb_shinfo(skb)->frags[0].page_offset += grow;
3817 skb_frag_size_sub(&skb_shinfo(skb)->frags[0], grow);
3819 if (unlikely(!skb_frag_size(&skb_shinfo(skb)->frags[0]))) {
3820 skb_frag_unref(skb, 0);
3821 memmove(skb_shinfo(skb)->frags,
3822 skb_shinfo(skb)->frags + 1,
3823 --skb_shinfo(skb)->nr_frags * sizeof(skb_frag_t));
3836 static gro_result_t napi_skb_finish(gro_result_t ret, struct sk_buff *skb)
3840 if (netif_receive_skb(skb))
3848 case GRO_MERGED_FREE:
3849 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
3850 kmem_cache_free(skbuff_head_cache, skb);
3863 static void skb_gro_reset_offset(struct sk_buff *skb)
3865 const struct skb_shared_info *pinfo = skb_shinfo(skb);
3866 const skb_frag_t *frag0 = &pinfo->frags[0];
3868 NAPI_GRO_CB(skb)->data_offset = 0;
3869 NAPI_GRO_CB(skb)->frag0 = NULL;
3870 NAPI_GRO_CB(skb)->frag0_len = 0;
3872 if (skb->mac_header == skb->tail &&
3874 !PageHighMem(skb_frag_page(frag0))) {
3875 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
3876 NAPI_GRO_CB(skb)->frag0_len = skb_frag_size(frag0);
3880 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
3882 skb_gro_reset_offset(skb);
3884 return napi_skb_finish(dev_gro_receive(napi, skb), skb);
3886 EXPORT_SYMBOL(napi_gro_receive);
3888 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
3890 __skb_pull(skb, skb_headlen(skb));
3891 /* restore the reserve we had after netdev_alloc_skb_ip_align() */
3892 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
3894 skb->dev = napi->dev;
3900 struct sk_buff *napi_get_frags(struct napi_struct *napi)
3902 struct sk_buff *skb = napi->skb;
3905 skb = netdev_alloc_skb_ip_align(napi->dev, GRO_MAX_HEAD);
3911 EXPORT_SYMBOL(napi_get_frags);
3913 static gro_result_t napi_frags_finish(struct napi_struct *napi, struct sk_buff *skb,
3919 skb->protocol = eth_type_trans(skb, skb->dev);
3921 if (ret == GRO_HELD)
3922 skb_gro_pull(skb, -ETH_HLEN);
3923 else if (netif_receive_skb(skb))
3928 case GRO_MERGED_FREE:
3929 napi_reuse_skb(napi, skb);
3939 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
3941 struct sk_buff *skb = napi->skb;
3948 skb_reset_mac_header(skb);
3949 skb_gro_reset_offset(skb);
3951 off = skb_gro_offset(skb);
3952 hlen = off + sizeof(*eth);
3953 eth = skb_gro_header_fast(skb, off);
3954 if (skb_gro_header_hard(skb, hlen)) {
3955 eth = skb_gro_header_slow(skb, hlen, off);
3956 if (unlikely(!eth)) {
3957 napi_reuse_skb(napi, skb);
3963 skb_gro_pull(skb, sizeof(*eth));
3966 * This works because the only protocols we care about don't require
3967 * special handling. We'll fix it up properly at the end.
3969 skb->protocol = eth->h_proto;
3975 gro_result_t napi_gro_frags(struct napi_struct *napi)
3977 struct sk_buff *skb = napi_frags_skb(napi);
3982 return napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
3984 EXPORT_SYMBOL(napi_gro_frags);
3987 * net_rps_action sends any pending IPI's for rps.
3988 * Note: called with local irq disabled, but exits with local irq enabled.
3990 static void net_rps_action_and_irq_enable(struct softnet_data *sd)
3993 struct softnet_data *remsd = sd->rps_ipi_list;
3996 sd->rps_ipi_list = NULL;
4000 /* Send pending IPI's to kick RPS processing on remote cpus. */
4002 struct softnet_data *next = remsd->rps_ipi_next;
4004 if (cpu_online(remsd->cpu))
4005 __smp_call_function_single(remsd->cpu,
4014 static int process_backlog(struct napi_struct *napi, int quota)
4017 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
4020 /* Check if we have pending ipi, its better to send them now,
4021 * not waiting net_rx_action() end.
4023 if (sd->rps_ipi_list) {
4024 local_irq_disable();
4025 net_rps_action_and_irq_enable(sd);
4028 napi->weight = weight_p;
4029 local_irq_disable();
4030 while (work < quota) {
4031 struct sk_buff *skb;
4034 while ((skb = __skb_dequeue(&sd->process_queue))) {
4036 __netif_receive_skb(skb);
4037 local_irq_disable();
4038 input_queue_head_incr(sd);
4039 if (++work >= quota) {
4046 qlen = skb_queue_len(&sd->input_pkt_queue);
4048 skb_queue_splice_tail_init(&sd->input_pkt_queue,
4049 &sd->process_queue);
4051 if (qlen < quota - work) {
4053 * Inline a custom version of __napi_complete().
4054 * only current cpu owns and manipulates this napi,
4055 * and NAPI_STATE_SCHED is the only possible flag set on backlog.
4056 * we can use a plain write instead of clear_bit(),
4057 * and we dont need an smp_mb() memory barrier.
4059 list_del(&napi->poll_list);
4062 quota = work + qlen;
4072 * __napi_schedule - schedule for receive
4073 * @n: entry to schedule
4075 * The entry's receive function will be scheduled to run
4077 void __napi_schedule(struct napi_struct *n)
4079 unsigned long flags;
4081 local_irq_save(flags);
4082 ____napi_schedule(&__get_cpu_var(softnet_data), n);
4083 local_irq_restore(flags);
4085 EXPORT_SYMBOL(__napi_schedule);
4087 void __napi_complete(struct napi_struct *n)
4089 BUG_ON(!test_bit(NAPI_STATE_SCHED, &n->state));
4090 BUG_ON(n->gro_list);
4092 list_del(&n->poll_list);
4093 smp_mb__before_clear_bit();
4094 clear_bit(NAPI_STATE_SCHED, &n->state);
4096 EXPORT_SYMBOL(__napi_complete);
4098 void napi_complete(struct napi_struct *n)
4100 unsigned long flags;
4103 * don't let napi dequeue from the cpu poll list
4104 * just in case its running on a different cpu
4106 if (unlikely(test_bit(NAPI_STATE_NPSVC, &n->state)))
4109 napi_gro_flush(n, false);
4110 local_irq_save(flags);
4112 local_irq_restore(flags);
4114 EXPORT_SYMBOL(napi_complete);
4116 void netif_napi_add(struct net_device *dev, struct napi_struct *napi,
4117 int (*poll)(struct napi_struct *, int), int weight)
4119 INIT_LIST_HEAD(&napi->poll_list);
4120 napi->gro_count = 0;
4121 napi->gro_list = NULL;
4124 if (weight > NAPI_POLL_WEIGHT)
4125 pr_err_once("netif_napi_add() called with weight %d on device %s\n",
4127 napi->weight = weight;
4128 list_add(&napi->dev_list, &dev->napi_list);
4130 #ifdef CONFIG_NETPOLL
4131 spin_lock_init(&napi->poll_lock);
4132 napi->poll_owner = -1;
4134 set_bit(NAPI_STATE_SCHED, &napi->state);
4136 EXPORT_SYMBOL(netif_napi_add);
4138 void netif_napi_del(struct napi_struct *napi)
4140 struct sk_buff *skb, *next;
4142 list_del_init(&napi->dev_list);
4143 napi_free_frags(napi);
4145 for (skb = napi->gro_list; skb; skb = next) {
4151 napi->gro_list = NULL;
4152 napi->gro_count = 0;
4154 EXPORT_SYMBOL(netif_napi_del);
4156 static void net_rx_action(struct softirq_action *h)
4158 struct softnet_data *sd = &__get_cpu_var(softnet_data);
4159 unsigned long time_limit = jiffies + 2;
4160 int budget = netdev_budget;
4163 local_irq_disable();
4165 while (!list_empty(&sd->poll_list)) {
4166 struct napi_struct *n;
4169 /* If softirq window is exhuasted then punt.
4170 * Allow this to run for 2 jiffies since which will allow
4171 * an average latency of 1.5/HZ.
4173 if (unlikely(budget <= 0 || time_after_eq(jiffies, time_limit)))
4178 /* Even though interrupts have been re-enabled, this
4179 * access is safe because interrupts can only add new
4180 * entries to the tail of this list, and only ->poll()
4181 * calls can remove this head entry from the list.
4183 n = list_first_entry(&sd->poll_list, struct napi_struct, poll_list);
4185 have = netpoll_poll_lock(n);
4189 /* This NAPI_STATE_SCHED test is for avoiding a race
4190 * with netpoll's poll_napi(). Only the entity which
4191 * obtains the lock and sees NAPI_STATE_SCHED set will
4192 * actually make the ->poll() call. Therefore we avoid
4193 * accidentally calling ->poll() when NAPI is not scheduled.
4196 if (test_bit(NAPI_STATE_SCHED, &n->state)) {
4197 work = n->poll(n, weight);
4201 WARN_ON_ONCE(work > weight);
4205 local_irq_disable();
4207 /* Drivers must not modify the NAPI state if they
4208 * consume the entire weight. In such cases this code
4209 * still "owns" the NAPI instance and therefore can
4210 * move the instance around on the list at-will.
4212 if (unlikely(work == weight)) {
4213 if (unlikely(napi_disable_pending(n))) {
4216 local_irq_disable();
4219 /* flush too old packets
4220 * If HZ < 1000, flush all packets.
4223 napi_gro_flush(n, HZ >= 1000);
4224 local_irq_disable();
4226 list_move_tail(&n->poll_list, &sd->poll_list);
4230 netpoll_poll_unlock(have);
4233 net_rps_action_and_irq_enable(sd);
4235 #ifdef CONFIG_NET_DMA
4237 * There may not be any more sk_buffs coming right now, so push
4238 * any pending DMA copies to hardware
4240 dma_issue_pending_all();
4247 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
4251 struct netdev_upper {
4252 struct net_device *dev;
4254 struct list_head list;
4255 struct rcu_head rcu;
4256 struct list_head search_list;
4259 static void __append_search_uppers(struct list_head *search_list,
4260 struct net_device *dev)
4262 struct netdev_upper *upper;
4264 list_for_each_entry(upper, &dev->upper_dev_list, list) {
4265 /* check if this upper is not already in search list */
4266 if (list_empty(&upper->search_list))
4267 list_add_tail(&upper->search_list, search_list);
4271 static bool __netdev_search_upper_dev(struct net_device *dev,
4272 struct net_device *upper_dev)
4274 LIST_HEAD(search_list);
4275 struct netdev_upper *upper;
4276 struct netdev_upper *tmp;
4279 __append_search_uppers(&search_list, dev);
4280 list_for_each_entry(upper, &search_list, search_list) {
4281 if (upper->dev == upper_dev) {
4285 __append_search_uppers(&search_list, upper->dev);
4287 list_for_each_entry_safe(upper, tmp, &search_list, search_list)
4288 INIT_LIST_HEAD(&upper->search_list);
4292 static struct netdev_upper *__netdev_find_upper(struct net_device *dev,
4293 struct net_device *upper_dev)
4295 struct netdev_upper *upper;
4297 list_for_each_entry(upper, &dev->upper_dev_list, list) {
4298 if (upper->dev == upper_dev)
4305 * netdev_has_upper_dev - Check if device is linked to an upper device
4307 * @upper_dev: upper device to check
4309 * Find out if a device is linked to specified upper device and return true
4310 * in case it is. Note that this checks only immediate upper device,
4311 * not through a complete stack of devices. The caller must hold the RTNL lock.
4313 bool netdev_has_upper_dev(struct net_device *dev,
4314 struct net_device *upper_dev)
4318 return __netdev_find_upper(dev, upper_dev);
4320 EXPORT_SYMBOL(netdev_has_upper_dev);
4323 * netdev_has_any_upper_dev - Check if device is linked to some device
4326 * Find out if a device is linked to an upper device and return true in case
4327 * it is. The caller must hold the RTNL lock.
4329 bool netdev_has_any_upper_dev(struct net_device *dev)
4333 return !list_empty(&dev->upper_dev_list);
4335 EXPORT_SYMBOL(netdev_has_any_upper_dev);
4338 * netdev_master_upper_dev_get - Get master upper device
4341 * Find a master upper device and return pointer to it or NULL in case
4342 * it's not there. The caller must hold the RTNL lock.
4344 struct net_device *netdev_master_upper_dev_get(struct net_device *dev)
4346 struct netdev_upper *upper;
4350 if (list_empty(&dev->upper_dev_list))
4353 upper = list_first_entry(&dev->upper_dev_list,
4354 struct netdev_upper, list);
4355 if (likely(upper->master))
4359 EXPORT_SYMBOL(netdev_master_upper_dev_get);
4362 * netdev_master_upper_dev_get_rcu - Get master upper device
4365 * Find a master upper device and return pointer to it or NULL in case
4366 * it's not there. The caller must hold the RCU read lock.
4368 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev)
4370 struct netdev_upper *upper;
4372 upper = list_first_or_null_rcu(&dev->upper_dev_list,
4373 struct netdev_upper, list);
4374 if (upper && likely(upper->master))
4378 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu);
4380 static int __netdev_upper_dev_link(struct net_device *dev,
4381 struct net_device *upper_dev, bool master)
4383 struct netdev_upper *upper;
4387 if (dev == upper_dev)
4390 /* To prevent loops, check if dev is not upper device to upper_dev. */
4391 if (__netdev_search_upper_dev(upper_dev, dev))
4394 if (__netdev_find_upper(dev, upper_dev))
4397 if (master && netdev_master_upper_dev_get(dev))
4400 upper = kmalloc(sizeof(*upper), GFP_KERNEL);
4404 upper->dev = upper_dev;
4405 upper->master = master;
4406 INIT_LIST_HEAD(&upper->search_list);
4408 /* Ensure that master upper link is always the first item in list. */
4410 list_add_rcu(&upper->list, &dev->upper_dev_list);
4412 list_add_tail_rcu(&upper->list, &dev->upper_dev_list);
4413 dev_hold(upper_dev);
4419 * netdev_upper_dev_link - Add a link to the upper device
4421 * @upper_dev: new upper device
4423 * Adds a link to device which is upper to this one. The caller must hold
4424 * the RTNL lock. On a failure a negative errno code is returned.
4425 * On success the reference counts are adjusted and the function
4428 int netdev_upper_dev_link(struct net_device *dev,
4429 struct net_device *upper_dev)
4431 return __netdev_upper_dev_link(dev, upper_dev, false);
4433 EXPORT_SYMBOL(netdev_upper_dev_link);
4436 * netdev_master_upper_dev_link - Add a master link to the upper device
4438 * @upper_dev: new upper device
4440 * Adds a link to device which is upper to this one. In this case, only
4441 * one master upper device can be linked, although other non-master devices
4442 * might be linked as well. The caller must hold the RTNL lock.
4443 * On a failure a negative errno code is returned. On success the reference
4444 * counts are adjusted and the function returns zero.
4446 int netdev_master_upper_dev_link(struct net_device *dev,
4447 struct net_device *upper_dev)
4449 return __netdev_upper_dev_link(dev, upper_dev, true);
4451 EXPORT_SYMBOL(netdev_master_upper_dev_link);
4454 * netdev_upper_dev_unlink - Removes a link to upper device
4456 * @upper_dev: new upper device
4458 * Removes a link to device which is upper to this one. The caller must hold
4461 void netdev_upper_dev_unlink(struct net_device *dev,
4462 struct net_device *upper_dev)
4464 struct netdev_upper *upper;
4468 upper = __netdev_find_upper(dev, upper_dev);
4471 list_del_rcu(&upper->list);
4473 kfree_rcu(upper, rcu);
4475 EXPORT_SYMBOL(netdev_upper_dev_unlink);
4477 static void dev_change_rx_flags(struct net_device *dev, int flags)
4479 const struct net_device_ops *ops = dev->netdev_ops;
4481 if ((dev->flags & IFF_UP) && ops->ndo_change_rx_flags)
4482 ops->ndo_change_rx_flags(dev, flags);
4485 static int __dev_set_promiscuity(struct net_device *dev, int inc)
4487 unsigned int old_flags = dev->flags;
4493 dev->flags |= IFF_PROMISC;
4494 dev->promiscuity += inc;
4495 if (dev->promiscuity == 0) {
4498 * If inc causes overflow, untouch promisc and return error.
4501 dev->flags &= ~IFF_PROMISC;
4503 dev->promiscuity -= inc;
4504 pr_warn("%s: promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n",
4509 if (dev->flags != old_flags) {
4510 pr_info("device %s %s promiscuous mode\n",
4512 dev->flags & IFF_PROMISC ? "entered" : "left");
4513 if (audit_enabled) {
4514 current_uid_gid(&uid, &gid);
4515 audit_log(current->audit_context, GFP_ATOMIC,
4516 AUDIT_ANOM_PROMISCUOUS,
4517 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u",
4518 dev->name, (dev->flags & IFF_PROMISC),
4519 (old_flags & IFF_PROMISC),
4520 from_kuid(&init_user_ns, audit_get_loginuid(current)),
4521 from_kuid(&init_user_ns, uid),
4522 from_kgid(&init_user_ns, gid),
4523 audit_get_sessionid(current));
4526 dev_change_rx_flags(dev, IFF_PROMISC);
4532 * dev_set_promiscuity - update promiscuity count on a device
4536 * Add or remove promiscuity from a device. While the count in the device
4537 * remains above zero the interface remains promiscuous. Once it hits zero
4538 * the device reverts back to normal filtering operation. A negative inc
4539 * value is used to drop promiscuity on the device.
4540 * Return 0 if successful or a negative errno code on error.
4542 int dev_set_promiscuity(struct net_device *dev, int inc)
4544 unsigned int old_flags = dev->flags;
4547 err = __dev_set_promiscuity(dev, inc);
4550 if (dev->flags != old_flags)
4551 dev_set_rx_mode(dev);
4554 EXPORT_SYMBOL(dev_set_promiscuity);
4557 * dev_set_allmulti - update allmulti count on a device
4561 * Add or remove reception of all multicast frames to a device. While the
4562 * count in the device remains above zero the interface remains listening
4563 * to all interfaces. Once it hits zero the device reverts back to normal
4564 * filtering operation. A negative @inc value is used to drop the counter
4565 * when releasing a resource needing all multicasts.
4566 * Return 0 if successful or a negative errno code on error.
4569 int dev_set_allmulti(struct net_device *dev, int inc)
4571 unsigned int old_flags = dev->flags;
4575 dev->flags |= IFF_ALLMULTI;
4576 dev->allmulti += inc;
4577 if (dev->allmulti == 0) {
4580 * If inc causes overflow, untouch allmulti and return error.
4583 dev->flags &= ~IFF_ALLMULTI;
4585 dev->allmulti -= inc;
4586 pr_warn("%s: allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n",
4591 if (dev->flags ^ old_flags) {
4592 dev_change_rx_flags(dev, IFF_ALLMULTI);
4593 dev_set_rx_mode(dev);
4597 EXPORT_SYMBOL(dev_set_allmulti);
4600 * Upload unicast and multicast address lists to device and
4601 * configure RX filtering. When the device doesn't support unicast
4602 * filtering it is put in promiscuous mode while unicast addresses
4605 void __dev_set_rx_mode(struct net_device *dev)
4607 const struct net_device_ops *ops = dev->netdev_ops;
4609 /* dev_open will call this function so the list will stay sane. */
4610 if (!(dev->flags&IFF_UP))
4613 if (!netif_device_present(dev))
4616 if (!(dev->priv_flags & IFF_UNICAST_FLT)) {
4617 /* Unicast addresses changes may only happen under the rtnl,
4618 * therefore calling __dev_set_promiscuity here is safe.
4620 if (!netdev_uc_empty(dev) && !dev->uc_promisc) {
4621 __dev_set_promiscuity(dev, 1);
4622 dev->uc_promisc = true;
4623 } else if (netdev_uc_empty(dev) && dev->uc_promisc) {
4624 __dev_set_promiscuity(dev, -1);
4625 dev->uc_promisc = false;
4629 if (ops->ndo_set_rx_mode)
4630 ops->ndo_set_rx_mode(dev);
4633 void dev_set_rx_mode(struct net_device *dev)
4635 netif_addr_lock_bh(dev);
4636 __dev_set_rx_mode(dev);
4637 netif_addr_unlock_bh(dev);
4641 * dev_get_flags - get flags reported to userspace
4644 * Get the combination of flag bits exported through APIs to userspace.
4646 unsigned int dev_get_flags(const struct net_device *dev)
4650 flags = (dev->flags & ~(IFF_PROMISC |
4655 (dev->gflags & (IFF_PROMISC |
4658 if (netif_running(dev)) {
4659 if (netif_oper_up(dev))
4660 flags |= IFF_RUNNING;
4661 if (netif_carrier_ok(dev))
4662 flags |= IFF_LOWER_UP;
4663 if (netif_dormant(dev))
4664 flags |= IFF_DORMANT;
4669 EXPORT_SYMBOL(dev_get_flags);
4671 int __dev_change_flags(struct net_device *dev, unsigned int flags)
4673 unsigned int old_flags = dev->flags;
4679 * Set the flags on our device.
4682 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP |
4683 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL |
4685 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC |
4689 * Load in the correct multicast list now the flags have changed.
4692 if ((old_flags ^ flags) & IFF_MULTICAST)
4693 dev_change_rx_flags(dev, IFF_MULTICAST);
4695 dev_set_rx_mode(dev);
4698 * Have we downed the interface. We handle IFF_UP ourselves
4699 * according to user attempts to set it, rather than blindly
4704 if ((old_flags ^ flags) & IFF_UP) { /* Bit is different ? */
4705 ret = ((old_flags & IFF_UP) ? __dev_close : __dev_open)(dev);
4708 dev_set_rx_mode(dev);
4711 if ((flags ^ dev->gflags) & IFF_PROMISC) {
4712 int inc = (flags & IFF_PROMISC) ? 1 : -1;
4714 dev->gflags ^= IFF_PROMISC;
4715 dev_set_promiscuity(dev, inc);
4718 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI
4719 is important. Some (broken) drivers set IFF_PROMISC, when
4720 IFF_ALLMULTI is requested not asking us and not reporting.
4722 if ((flags ^ dev->gflags) & IFF_ALLMULTI) {
4723 int inc = (flags & IFF_ALLMULTI) ? 1 : -1;
4725 dev->gflags ^= IFF_ALLMULTI;
4726 dev_set_allmulti(dev, inc);
4732 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags)
4734 unsigned int changes = dev->flags ^ old_flags;
4736 if (changes & IFF_UP) {
4737 if (dev->flags & IFF_UP)
4738 call_netdevice_notifiers(NETDEV_UP, dev);
4740 call_netdevice_notifiers(NETDEV_DOWN, dev);
4743 if (dev->flags & IFF_UP &&
4744 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE)))
4745 call_netdevice_notifiers(NETDEV_CHANGE, dev);
4749 * dev_change_flags - change device settings
4751 * @flags: device state flags
4753 * Change settings on device based state flags. The flags are
4754 * in the userspace exported format.
4756 int dev_change_flags(struct net_device *dev, unsigned int flags)
4759 unsigned int changes, old_flags = dev->flags;
4761 ret = __dev_change_flags(dev, flags);
4765 changes = old_flags ^ dev->flags;
4767 rtmsg_ifinfo(RTM_NEWLINK, dev, changes);
4769 __dev_notify_flags(dev, old_flags);
4772 EXPORT_SYMBOL(dev_change_flags);
4775 * dev_set_mtu - Change maximum transfer unit
4777 * @new_mtu: new transfer unit
4779 * Change the maximum transfer size of the network device.
4781 int dev_set_mtu(struct net_device *dev, int new_mtu)
4783 const struct net_device_ops *ops = dev->netdev_ops;
4786 if (new_mtu == dev->mtu)
4789 /* MTU must be positive. */
4793 if (!netif_device_present(dev))
4797 if (ops->ndo_change_mtu)
4798 err = ops->ndo_change_mtu(dev, new_mtu);
4803 call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
4806 EXPORT_SYMBOL(dev_set_mtu);
4809 * dev_set_group - Change group this device belongs to
4811 * @new_group: group this device should belong to
4813 void dev_set_group(struct net_device *dev, int new_group)
4815 dev->group = new_group;
4817 EXPORT_SYMBOL(dev_set_group);
4820 * dev_set_mac_address - Change Media Access Control Address
4824 * Change the hardware (MAC) address of the device
4826 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa)
4828 const struct net_device_ops *ops = dev->netdev_ops;
4831 if (!ops->ndo_set_mac_address)
4833 if (sa->sa_family != dev->type)
4835 if (!netif_device_present(dev))
4837 err = ops->ndo_set_mac_address(dev, sa);
4840 dev->addr_assign_type = NET_ADDR_SET;
4841 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
4842 add_device_randomness(dev->dev_addr, dev->addr_len);
4845 EXPORT_SYMBOL(dev_set_mac_address);
4848 * dev_change_carrier - Change device carrier
4850 * @new_carrier: new value
4852 * Change device carrier
4854 int dev_change_carrier(struct net_device *dev, bool new_carrier)
4856 const struct net_device_ops *ops = dev->netdev_ops;
4858 if (!ops->ndo_change_carrier)
4860 if (!netif_device_present(dev))
4862 return ops->ndo_change_carrier(dev, new_carrier);
4864 EXPORT_SYMBOL(dev_change_carrier);
4867 * dev_new_index - allocate an ifindex
4868 * @net: the applicable net namespace
4870 * Returns a suitable unique value for a new device interface
4871 * number. The caller must hold the rtnl semaphore or the
4872 * dev_base_lock to be sure it remains unique.
4874 static int dev_new_index(struct net *net)
4876 int ifindex = net->ifindex;
4880 if (!__dev_get_by_index(net, ifindex))
4881 return net->ifindex = ifindex;
4885 /* Delayed registration/unregisteration */
4886 static LIST_HEAD(net_todo_list);
4888 static void net_set_todo(struct net_device *dev)
4890 list_add_tail(&dev->todo_list, &net_todo_list);
4893 static void rollback_registered_many(struct list_head *head)
4895 struct net_device *dev, *tmp;
4897 BUG_ON(dev_boot_phase);
4900 list_for_each_entry_safe(dev, tmp, head, unreg_list) {
4901 /* Some devices call without registering
4902 * for initialization unwind. Remove those
4903 * devices and proceed with the remaining.
4905 if (dev->reg_state == NETREG_UNINITIALIZED) {
4906 pr_debug("unregister_netdevice: device %s/%p never was registered\n",
4910 list_del(&dev->unreg_list);
4913 dev->dismantle = true;
4914 BUG_ON(dev->reg_state != NETREG_REGISTERED);
4917 /* If device is running, close it first. */
4918 dev_close_many(head);
4920 list_for_each_entry(dev, head, unreg_list) {
4921 /* And unlink it from device chain. */
4922 unlist_netdevice(dev);
4924 dev->reg_state = NETREG_UNREGISTERING;
4929 list_for_each_entry(dev, head, unreg_list) {
4930 /* Shutdown queueing discipline. */
4934 /* Notify protocols, that we are about to destroy
4935 this device. They should clean all the things.
4937 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
4939 if (!dev->rtnl_link_ops ||
4940 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
4941 rtmsg_ifinfo(RTM_DELLINK, dev, ~0U);
4944 * Flush the unicast and multicast chains
4949 if (dev->netdev_ops->ndo_uninit)
4950 dev->netdev_ops->ndo_uninit(dev);
4952 /* Notifier chain MUST detach us all upper devices. */
4953 WARN_ON(netdev_has_any_upper_dev(dev));
4955 /* Remove entries from kobject tree */
4956 netdev_unregister_kobject(dev);
4958 /* Remove XPS queueing entries */
4959 netif_reset_xps_queues_gt(dev, 0);
4965 list_for_each_entry(dev, head, unreg_list)
4969 static void rollback_registered(struct net_device *dev)
4973 list_add(&dev->unreg_list, &single);
4974 rollback_registered_many(&single);
4978 static netdev_features_t netdev_fix_features(struct net_device *dev,
4979 netdev_features_t features)
4981 /* Fix illegal checksum combinations */
4982 if ((features & NETIF_F_HW_CSUM) &&
4983 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
4984 netdev_warn(dev, "mixed HW and IP checksum settings.\n");
4985 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
4988 /* TSO requires that SG is present as well. */
4989 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) {
4990 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n");
4991 features &= ~NETIF_F_ALL_TSO;
4994 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) &&
4995 !(features & NETIF_F_IP_CSUM)) {
4996 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n");
4997 features &= ~NETIF_F_TSO;
4998 features &= ~NETIF_F_TSO_ECN;
5001 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) &&
5002 !(features & NETIF_F_IPV6_CSUM)) {
5003 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n");
5004 features &= ~NETIF_F_TSO6;
5007 /* TSO ECN requires that TSO is present as well. */
5008 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN)
5009 features &= ~NETIF_F_TSO_ECN;
5011 /* Software GSO depends on SG. */
5012 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) {
5013 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n");
5014 features &= ~NETIF_F_GSO;
5017 /* UFO needs SG and checksumming */
5018 if (features & NETIF_F_UFO) {
5019 /* maybe split UFO into V4 and V6? */
5020 if (!((features & NETIF_F_GEN_CSUM) ||
5021 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))
5022 == (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
5024 "Dropping NETIF_F_UFO since no checksum offload features.\n");
5025 features &= ~NETIF_F_UFO;
5028 if (!(features & NETIF_F_SG)) {
5030 "Dropping NETIF_F_UFO since no NETIF_F_SG feature.\n");
5031 features &= ~NETIF_F_UFO;
5038 int __netdev_update_features(struct net_device *dev)
5040 netdev_features_t features;
5045 features = netdev_get_wanted_features(dev);
5047 if (dev->netdev_ops->ndo_fix_features)
5048 features = dev->netdev_ops->ndo_fix_features(dev, features);
5050 /* driver might be less strict about feature dependencies */
5051 features = netdev_fix_features(dev, features);
5053 if (dev->features == features)
5056 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n",
5057 &dev->features, &features);
5059 if (dev->netdev_ops->ndo_set_features)
5060 err = dev->netdev_ops->ndo_set_features(dev, features);
5062 if (unlikely(err < 0)) {
5064 "set_features() failed (%d); wanted %pNF, left %pNF\n",
5065 err, &features, &dev->features);
5070 dev->features = features;
5076 * netdev_update_features - recalculate device features
5077 * @dev: the device to check
5079 * Recalculate dev->features set and send notifications if it
5080 * has changed. Should be called after driver or hardware dependent
5081 * conditions might have changed that influence the features.
5083 void netdev_update_features(struct net_device *dev)
5085 if (__netdev_update_features(dev))
5086 netdev_features_change(dev);
5088 EXPORT_SYMBOL(netdev_update_features);
5091 * netdev_change_features - recalculate device features
5092 * @dev: the device to check
5094 * Recalculate dev->features set and send notifications even
5095 * if they have not changed. Should be called instead of
5096 * netdev_update_features() if also dev->vlan_features might
5097 * have changed to allow the changes to be propagated to stacked
5100 void netdev_change_features(struct net_device *dev)
5102 __netdev_update_features(dev);
5103 netdev_features_change(dev);
5105 EXPORT_SYMBOL(netdev_change_features);
5108 * netif_stacked_transfer_operstate - transfer operstate
5109 * @rootdev: the root or lower level device to transfer state from
5110 * @dev: the device to transfer operstate to
5112 * Transfer operational state from root to device. This is normally
5113 * called when a stacking relationship exists between the root
5114 * device and the device(a leaf device).
5116 void netif_stacked_transfer_operstate(const struct net_device *rootdev,
5117 struct net_device *dev)
5119 if (rootdev->operstate == IF_OPER_DORMANT)
5120 netif_dormant_on(dev);
5122 netif_dormant_off(dev);
5124 if (netif_carrier_ok(rootdev)) {
5125 if (!netif_carrier_ok(dev))
5126 netif_carrier_on(dev);
5128 if (netif_carrier_ok(dev))
5129 netif_carrier_off(dev);
5132 EXPORT_SYMBOL(netif_stacked_transfer_operstate);
5135 static int netif_alloc_rx_queues(struct net_device *dev)
5137 unsigned int i, count = dev->num_rx_queues;
5138 struct netdev_rx_queue *rx;
5142 rx = kcalloc(count, sizeof(struct netdev_rx_queue), GFP_KERNEL);
5148 for (i = 0; i < count; i++)
5154 static void netdev_init_one_queue(struct net_device *dev,
5155 struct netdev_queue *queue, void *_unused)
5157 /* Initialize queue lock */
5158 spin_lock_init(&queue->_xmit_lock);
5159 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type);
5160 queue->xmit_lock_owner = -1;
5161 netdev_queue_numa_node_write(queue, NUMA_NO_NODE);
5164 dql_init(&queue->dql, HZ);
5168 static int netif_alloc_netdev_queues(struct net_device *dev)
5170 unsigned int count = dev->num_tx_queues;
5171 struct netdev_queue *tx;
5175 tx = kcalloc(count, sizeof(struct netdev_queue), GFP_KERNEL);
5181 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL);
5182 spin_lock_init(&dev->tx_global_lock);
5188 * register_netdevice - register a network device
5189 * @dev: device to register
5191 * Take a completed network device structure and add it to the kernel
5192 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
5193 * chain. 0 is returned on success. A negative errno code is returned
5194 * on a failure to set up the device, or if the name is a duplicate.
5196 * Callers must hold the rtnl semaphore. You may want
5197 * register_netdev() instead of this.
5200 * The locking appears insufficient to guarantee two parallel registers
5201 * will not get the same name.
5204 int register_netdevice(struct net_device *dev)
5207 struct net *net = dev_net(dev);
5209 BUG_ON(dev_boot_phase);
5214 /* When net_device's are persistent, this will be fatal. */
5215 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
5218 spin_lock_init(&dev->addr_list_lock);
5219 netdev_set_addr_lockdep_class(dev);
5223 ret = dev_get_valid_name(net, dev, dev->name);
5227 /* Init, if this function is available */
5228 if (dev->netdev_ops->ndo_init) {
5229 ret = dev->netdev_ops->ndo_init(dev);
5237 if (((dev->hw_features | dev->features) &
5238 NETIF_F_HW_VLAN_CTAG_FILTER) &&
5239 (!dev->netdev_ops->ndo_vlan_rx_add_vid ||
5240 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) {
5241 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n");
5248 dev->ifindex = dev_new_index(net);
5249 else if (__dev_get_by_index(net, dev->ifindex))
5252 if (dev->iflink == -1)
5253 dev->iflink = dev->ifindex;
5255 /* Transfer changeable features to wanted_features and enable
5256 * software offloads (GSO and GRO).
5258 dev->hw_features |= NETIF_F_SOFT_FEATURES;
5259 dev->features |= NETIF_F_SOFT_FEATURES;
5260 dev->wanted_features = dev->features & dev->hw_features;
5262 /* Turn on no cache copy if HW is doing checksum */
5263 if (!(dev->flags & IFF_LOOPBACK)) {
5264 dev->hw_features |= NETIF_F_NOCACHE_COPY;
5265 if (dev->features & NETIF_F_ALL_CSUM) {
5266 dev->wanted_features |= NETIF_F_NOCACHE_COPY;
5267 dev->features |= NETIF_F_NOCACHE_COPY;
5271 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices.
5273 dev->vlan_features |= NETIF_F_HIGHDMA;
5275 /* Make NETIF_F_SG inheritable to tunnel devices.
5277 dev->hw_enc_features |= NETIF_F_SG;
5279 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev);
5280 ret = notifier_to_errno(ret);
5284 ret = netdev_register_kobject(dev);
5287 dev->reg_state = NETREG_REGISTERED;
5289 __netdev_update_features(dev);
5292 * Default initial state at registry is that the
5293 * device is present.
5296 set_bit(__LINK_STATE_PRESENT, &dev->state);
5298 linkwatch_init_dev(dev);
5300 dev_init_scheduler(dev);
5302 list_netdevice(dev);
5303 add_device_randomness(dev->dev_addr, dev->addr_len);
5305 /* If the device has permanent device address, driver should
5306 * set dev_addr and also addr_assign_type should be set to
5307 * NET_ADDR_PERM (default value).
5309 if (dev->addr_assign_type == NET_ADDR_PERM)
5310 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len);
5312 /* Notify protocols, that a new device appeared. */
5313 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
5314 ret = notifier_to_errno(ret);
5316 rollback_registered(dev);
5317 dev->reg_state = NETREG_UNREGISTERED;
5320 * Prevent userspace races by waiting until the network
5321 * device is fully setup before sending notifications.
5323 if (!dev->rtnl_link_ops ||
5324 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
5325 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U);
5331 if (dev->netdev_ops->ndo_uninit)
5332 dev->netdev_ops->ndo_uninit(dev);
5335 EXPORT_SYMBOL(register_netdevice);
5338 * init_dummy_netdev - init a dummy network device for NAPI
5339 * @dev: device to init
5341 * This takes a network device structure and initialize the minimum
5342 * amount of fields so it can be used to schedule NAPI polls without
5343 * registering a full blown interface. This is to be used by drivers
5344 * that need to tie several hardware interfaces to a single NAPI
5345 * poll scheduler due to HW limitations.
5347 int init_dummy_netdev(struct net_device *dev)
5349 /* Clear everything. Note we don't initialize spinlocks
5350 * are they aren't supposed to be taken by any of the
5351 * NAPI code and this dummy netdev is supposed to be
5352 * only ever used for NAPI polls
5354 memset(dev, 0, sizeof(struct net_device));
5356 /* make sure we BUG if trying to hit standard
5357 * register/unregister code path
5359 dev->reg_state = NETREG_DUMMY;
5361 /* NAPI wants this */
5362 INIT_LIST_HEAD(&dev->napi_list);
5364 /* a dummy interface is started by default */
5365 set_bit(__LINK_STATE_PRESENT, &dev->state);
5366 set_bit(__LINK_STATE_START, &dev->state);
5368 /* Note : We dont allocate pcpu_refcnt for dummy devices,
5369 * because users of this 'device' dont need to change
5375 EXPORT_SYMBOL_GPL(init_dummy_netdev);
5379 * register_netdev - register a network device
5380 * @dev: device to register
5382 * Take a completed network device structure and add it to the kernel
5383 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
5384 * chain. 0 is returned on success. A negative errno code is returned
5385 * on a failure to set up the device, or if the name is a duplicate.
5387 * This is a wrapper around register_netdevice that takes the rtnl semaphore
5388 * and expands the device name if you passed a format string to
5391 int register_netdev(struct net_device *dev)
5396 err = register_netdevice(dev);
5400 EXPORT_SYMBOL(register_netdev);
5402 int netdev_refcnt_read(const struct net_device *dev)
5406 for_each_possible_cpu(i)
5407 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
5410 EXPORT_SYMBOL(netdev_refcnt_read);
5413 * netdev_wait_allrefs - wait until all references are gone.
5414 * @dev: target net_device
5416 * This is called when unregistering network devices.
5418 * Any protocol or device that holds a reference should register
5419 * for netdevice notification, and cleanup and put back the
5420 * reference if they receive an UNREGISTER event.
5421 * We can get stuck here if buggy protocols don't correctly
5424 static void netdev_wait_allrefs(struct net_device *dev)
5426 unsigned long rebroadcast_time, warning_time;
5429 linkwatch_forget_dev(dev);
5431 rebroadcast_time = warning_time = jiffies;
5432 refcnt = netdev_refcnt_read(dev);
5434 while (refcnt != 0) {
5435 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
5438 /* Rebroadcast unregister notification */
5439 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
5445 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
5446 if (test_bit(__LINK_STATE_LINKWATCH_PENDING,
5448 /* We must not have linkwatch events
5449 * pending on unregister. If this
5450 * happens, we simply run the queue
5451 * unscheduled, resulting in a noop
5454 linkwatch_run_queue();
5459 rebroadcast_time = jiffies;
5464 refcnt = netdev_refcnt_read(dev);
5466 if (time_after(jiffies, warning_time + 10 * HZ)) {
5467 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n",
5469 warning_time = jiffies;
5478 * register_netdevice(x1);
5479 * register_netdevice(x2);
5481 * unregister_netdevice(y1);
5482 * unregister_netdevice(y2);
5488 * We are invoked by rtnl_unlock().
5489 * This allows us to deal with problems:
5490 * 1) We can delete sysfs objects which invoke hotplug
5491 * without deadlocking with linkwatch via keventd.
5492 * 2) Since we run with the RTNL semaphore not held, we can sleep
5493 * safely in order to wait for the netdev refcnt to drop to zero.
5495 * We must not return until all unregister events added during
5496 * the interval the lock was held have been completed.
5498 void netdev_run_todo(void)
5500 struct list_head list;
5502 /* Snapshot list, allow later requests */
5503 list_replace_init(&net_todo_list, &list);
5508 /* Wait for rcu callbacks to finish before next phase */
5509 if (!list_empty(&list))
5512 while (!list_empty(&list)) {
5513 struct net_device *dev
5514 = list_first_entry(&list, struct net_device, todo_list);
5515 list_del(&dev->todo_list);
5518 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
5521 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) {
5522 pr_err("network todo '%s' but state %d\n",
5523 dev->name, dev->reg_state);
5528 dev->reg_state = NETREG_UNREGISTERED;
5530 on_each_cpu(flush_backlog, dev, 1);
5532 netdev_wait_allrefs(dev);
5535 BUG_ON(netdev_refcnt_read(dev));
5536 WARN_ON(rcu_access_pointer(dev->ip_ptr));
5537 WARN_ON(rcu_access_pointer(dev->ip6_ptr));
5538 WARN_ON(dev->dn_ptr);
5540 if (dev->destructor)
5541 dev->destructor(dev);
5543 /* Free network device */
5544 kobject_put(&dev->dev.kobj);
5548 /* Convert net_device_stats to rtnl_link_stats64. They have the same
5549 * fields in the same order, with only the type differing.
5551 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
5552 const struct net_device_stats *netdev_stats)
5554 #if BITS_PER_LONG == 64
5555 BUILD_BUG_ON(sizeof(*stats64) != sizeof(*netdev_stats));
5556 memcpy(stats64, netdev_stats, sizeof(*stats64));
5558 size_t i, n = sizeof(*stats64) / sizeof(u64);
5559 const unsigned long *src = (const unsigned long *)netdev_stats;
5560 u64 *dst = (u64 *)stats64;
5562 BUILD_BUG_ON(sizeof(*netdev_stats) / sizeof(unsigned long) !=
5563 sizeof(*stats64) / sizeof(u64));
5564 for (i = 0; i < n; i++)
5568 EXPORT_SYMBOL(netdev_stats_to_stats64);
5571 * dev_get_stats - get network device statistics
5572 * @dev: device to get statistics from
5573 * @storage: place to store stats
5575 * Get network statistics from device. Return @storage.
5576 * The device driver may provide its own method by setting
5577 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats;
5578 * otherwise the internal statistics structure is used.
5580 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev,
5581 struct rtnl_link_stats64 *storage)
5583 const struct net_device_ops *ops = dev->netdev_ops;
5585 if (ops->ndo_get_stats64) {
5586 memset(storage, 0, sizeof(*storage));
5587 ops->ndo_get_stats64(dev, storage);
5588 } else if (ops->ndo_get_stats) {
5589 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev));
5591 netdev_stats_to_stats64(storage, &dev->stats);
5593 storage->rx_dropped += atomic_long_read(&dev->rx_dropped);
5596 EXPORT_SYMBOL(dev_get_stats);
5598 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev)
5600 struct netdev_queue *queue = dev_ingress_queue(dev);
5602 #ifdef CONFIG_NET_CLS_ACT
5605 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
5608 netdev_init_one_queue(dev, queue, NULL);
5609 queue->qdisc = &noop_qdisc;
5610 queue->qdisc_sleeping = &noop_qdisc;
5611 rcu_assign_pointer(dev->ingress_queue, queue);
5616 static const struct ethtool_ops default_ethtool_ops;
5618 void netdev_set_default_ethtool_ops(struct net_device *dev,
5619 const struct ethtool_ops *ops)
5621 if (dev->ethtool_ops == &default_ethtool_ops)
5622 dev->ethtool_ops = ops;
5624 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops);
5627 * alloc_netdev_mqs - allocate network device
5628 * @sizeof_priv: size of private data to allocate space for
5629 * @name: device name format string
5630 * @setup: callback to initialize device
5631 * @txqs: the number of TX subqueues to allocate
5632 * @rxqs: the number of RX subqueues to allocate
5634 * Allocates a struct net_device with private data area for driver use
5635 * and performs basic initialization. Also allocates subquue structs
5636 * for each queue on the device.
5638 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
5639 void (*setup)(struct net_device *),
5640 unsigned int txqs, unsigned int rxqs)
5642 struct net_device *dev;
5644 struct net_device *p;
5646 BUG_ON(strlen(name) >= sizeof(dev->name));
5649 pr_err("alloc_netdev: Unable to allocate device with zero queues\n");
5655 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n");
5660 alloc_size = sizeof(struct net_device);
5662 /* ensure 32-byte alignment of private area */
5663 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
5664 alloc_size += sizeof_priv;
5666 /* ensure 32-byte alignment of whole construct */
5667 alloc_size += NETDEV_ALIGN - 1;
5669 p = kzalloc(alloc_size, GFP_KERNEL);
5673 dev = PTR_ALIGN(p, NETDEV_ALIGN);
5674 dev->padded = (char *)dev - (char *)p;
5676 dev->pcpu_refcnt = alloc_percpu(int);
5677 if (!dev->pcpu_refcnt)
5680 if (dev_addr_init(dev))
5686 dev_net_set(dev, &init_net);
5688 dev->gso_max_size = GSO_MAX_SIZE;
5689 dev->gso_max_segs = GSO_MAX_SEGS;
5691 INIT_LIST_HEAD(&dev->napi_list);
5692 INIT_LIST_HEAD(&dev->unreg_list);
5693 INIT_LIST_HEAD(&dev->link_watch_list);
5694 INIT_LIST_HEAD(&dev->upper_dev_list);
5695 dev->priv_flags = IFF_XMIT_DST_RELEASE;
5698 dev->num_tx_queues = txqs;
5699 dev->real_num_tx_queues = txqs;
5700 if (netif_alloc_netdev_queues(dev))
5704 dev->num_rx_queues = rxqs;
5705 dev->real_num_rx_queues = rxqs;
5706 if (netif_alloc_rx_queues(dev))
5710 strcpy(dev->name, name);
5711 dev->group = INIT_NETDEV_GROUP;
5712 if (!dev->ethtool_ops)
5713 dev->ethtool_ops = &default_ethtool_ops;
5721 free_percpu(dev->pcpu_refcnt);
5731 EXPORT_SYMBOL(alloc_netdev_mqs);
5734 * free_netdev - free network device
5737 * This function does the last stage of destroying an allocated device
5738 * interface. The reference to the device object is released.
5739 * If this is the last reference then it will be freed.
5741 void free_netdev(struct net_device *dev)
5743 struct napi_struct *p, *n;
5745 release_net(dev_net(dev));
5752 kfree(rcu_dereference_protected(dev->ingress_queue, 1));
5754 /* Flush device addresses */
5755 dev_addr_flush(dev);
5757 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
5760 free_percpu(dev->pcpu_refcnt);
5761 dev->pcpu_refcnt = NULL;
5763 /* Compatibility with error handling in drivers */
5764 if (dev->reg_state == NETREG_UNINITIALIZED) {
5765 kfree((char *)dev - dev->padded);
5769 BUG_ON(dev->reg_state != NETREG_UNREGISTERED);
5770 dev->reg_state = NETREG_RELEASED;
5772 /* will free via device release */
5773 put_device(&dev->dev);
5775 EXPORT_SYMBOL(free_netdev);
5778 * synchronize_net - Synchronize with packet receive processing
5780 * Wait for packets currently being received to be done.
5781 * Does not block later packets from starting.
5783 void synchronize_net(void)
5786 if (rtnl_is_locked())
5787 synchronize_rcu_expedited();
5791 EXPORT_SYMBOL(synchronize_net);
5794 * unregister_netdevice_queue - remove device from the kernel
5798 * This function shuts down a device interface and removes it
5799 * from the kernel tables.
5800 * If head not NULL, device is queued to be unregistered later.
5802 * Callers must hold the rtnl semaphore. You may want
5803 * unregister_netdev() instead of this.
5806 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head)
5811 list_move_tail(&dev->unreg_list, head);
5813 rollback_registered(dev);
5814 /* Finish processing unregister after unlock */
5818 EXPORT_SYMBOL(unregister_netdevice_queue);
5821 * unregister_netdevice_many - unregister many devices
5822 * @head: list of devices
5824 void unregister_netdevice_many(struct list_head *head)
5826 struct net_device *dev;
5828 if (!list_empty(head)) {
5829 rollback_registered_many(head);
5830 list_for_each_entry(dev, head, unreg_list)
5834 EXPORT_SYMBOL(unregister_netdevice_many);
5837 * unregister_netdev - remove device from the kernel
5840 * This function shuts down a device interface and removes it
5841 * from the kernel tables.
5843 * This is just a wrapper for unregister_netdevice that takes
5844 * the rtnl semaphore. In general you want to use this and not
5845 * unregister_netdevice.
5847 void unregister_netdev(struct net_device *dev)
5850 unregister_netdevice(dev);
5853 EXPORT_SYMBOL(unregister_netdev);
5856 * dev_change_net_namespace - move device to different nethost namespace
5858 * @net: network namespace
5859 * @pat: If not NULL name pattern to try if the current device name
5860 * is already taken in the destination network namespace.
5862 * This function shuts down a device interface and moves it
5863 * to a new network namespace. On success 0 is returned, on
5864 * a failure a netagive errno code is returned.
5866 * Callers must hold the rtnl semaphore.
5869 int dev_change_net_namespace(struct net_device *dev, struct net *net, const char *pat)
5875 /* Don't allow namespace local devices to be moved. */
5877 if (dev->features & NETIF_F_NETNS_LOCAL)
5880 /* Ensure the device has been registrered */
5881 if (dev->reg_state != NETREG_REGISTERED)
5884 /* Get out if there is nothing todo */
5886 if (net_eq(dev_net(dev), net))
5889 /* Pick the destination device name, and ensure
5890 * we can use it in the destination network namespace.
5893 if (__dev_get_by_name(net, dev->name)) {
5894 /* We get here if we can't use the current device name */
5897 if (dev_get_valid_name(net, dev, pat) < 0)
5902 * And now a mini version of register_netdevice unregister_netdevice.
5905 /* If device is running close it first. */
5908 /* And unlink it from device chain */
5910 unlist_netdevice(dev);
5914 /* Shutdown queueing discipline. */
5917 /* Notify protocols, that we are about to destroy
5918 this device. They should clean all the things.
5920 Note that dev->reg_state stays at NETREG_REGISTERED.
5921 This is wanted because this way 8021q and macvlan know
5922 the device is just moving and can keep their slaves up.
5924 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
5926 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
5927 rtmsg_ifinfo(RTM_DELLINK, dev, ~0U);
5930 * Flush the unicast and multicast chains
5935 /* Send a netdev-removed uevent to the old namespace */
5936 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE);
5938 /* Actually switch the network namespace */
5939 dev_net_set(dev, net);
5941 /* If there is an ifindex conflict assign a new one */
5942 if (__dev_get_by_index(net, dev->ifindex)) {
5943 int iflink = (dev->iflink == dev->ifindex);
5944 dev->ifindex = dev_new_index(net);
5946 dev->iflink = dev->ifindex;
5949 /* Send a netdev-add uevent to the new namespace */
5950 kobject_uevent(&dev->dev.kobj, KOBJ_ADD);
5952 /* Fixup kobjects */
5953 err = device_rename(&dev->dev, dev->name);
5956 /* Add the device back in the hashes */
5957 list_netdevice(dev);
5959 /* Notify protocols, that a new device appeared. */
5960 call_netdevice_notifiers(NETDEV_REGISTER, dev);
5963 * Prevent userspace races by waiting until the network
5964 * device is fully setup before sending notifications.
5966 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U);
5973 EXPORT_SYMBOL_GPL(dev_change_net_namespace);
5975 static int dev_cpu_callback(struct notifier_block *nfb,
5976 unsigned long action,
5979 struct sk_buff **list_skb;
5980 struct sk_buff *skb;
5981 unsigned int cpu, oldcpu = (unsigned long)ocpu;
5982 struct softnet_data *sd, *oldsd;
5984 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
5987 local_irq_disable();
5988 cpu = smp_processor_id();
5989 sd = &per_cpu(softnet_data, cpu);
5990 oldsd = &per_cpu(softnet_data, oldcpu);
5992 /* Find end of our completion_queue. */
5993 list_skb = &sd->completion_queue;
5995 list_skb = &(*list_skb)->next;
5996 /* Append completion queue from offline CPU. */
5997 *list_skb = oldsd->completion_queue;
5998 oldsd->completion_queue = NULL;
6000 /* Append output queue from offline CPU. */
6001 if (oldsd->output_queue) {
6002 *sd->output_queue_tailp = oldsd->output_queue;
6003 sd->output_queue_tailp = oldsd->output_queue_tailp;
6004 oldsd->output_queue = NULL;
6005 oldsd->output_queue_tailp = &oldsd->output_queue;
6007 /* Append NAPI poll list from offline CPU. */
6008 if (!list_empty(&oldsd->poll_list)) {
6009 list_splice_init(&oldsd->poll_list, &sd->poll_list);
6010 raise_softirq_irqoff(NET_RX_SOFTIRQ);
6013 raise_softirq_irqoff(NET_TX_SOFTIRQ);
6016 /* Process offline CPU's input_pkt_queue */
6017 while ((skb = __skb_dequeue(&oldsd->process_queue))) {
6019 input_queue_head_incr(oldsd);
6021 while ((skb = __skb_dequeue(&oldsd->input_pkt_queue))) {
6023 input_queue_head_incr(oldsd);
6031 * netdev_increment_features - increment feature set by one
6032 * @all: current feature set
6033 * @one: new feature set
6034 * @mask: mask feature set
6036 * Computes a new feature set after adding a device with feature set
6037 * @one to the master device with current feature set @all. Will not
6038 * enable anything that is off in @mask. Returns the new feature set.
6040 netdev_features_t netdev_increment_features(netdev_features_t all,
6041 netdev_features_t one, netdev_features_t mask)
6043 if (mask & NETIF_F_GEN_CSUM)
6044 mask |= NETIF_F_ALL_CSUM;
6045 mask |= NETIF_F_VLAN_CHALLENGED;
6047 all |= one & (NETIF_F_ONE_FOR_ALL|NETIF_F_ALL_CSUM) & mask;
6048 all &= one | ~NETIF_F_ALL_FOR_ALL;
6050 /* If one device supports hw checksumming, set for all. */
6051 if (all & NETIF_F_GEN_CSUM)
6052 all &= ~(NETIF_F_ALL_CSUM & ~NETIF_F_GEN_CSUM);
6056 EXPORT_SYMBOL(netdev_increment_features);
6058 static struct hlist_head *netdev_create_hash(void)
6061 struct hlist_head *hash;
6063 hash = kmalloc(sizeof(*hash) * NETDEV_HASHENTRIES, GFP_KERNEL);
6065 for (i = 0; i < NETDEV_HASHENTRIES; i++)
6066 INIT_HLIST_HEAD(&hash[i]);
6071 /* Initialize per network namespace state */
6072 static int __net_init netdev_init(struct net *net)
6074 if (net != &init_net)
6075 INIT_LIST_HEAD(&net->dev_base_head);
6077 net->dev_name_head = netdev_create_hash();
6078 if (net->dev_name_head == NULL)
6081 net->dev_index_head = netdev_create_hash();
6082 if (net->dev_index_head == NULL)
6088 kfree(net->dev_name_head);
6094 * netdev_drivername - network driver for the device
6095 * @dev: network device
6097 * Determine network driver for device.
6099 const char *netdev_drivername(const struct net_device *dev)
6101 const struct device_driver *driver;
6102 const struct device *parent;
6103 const char *empty = "";
6105 parent = dev->dev.parent;
6109 driver = parent->driver;
6110 if (driver && driver->name)
6111 return driver->name;
6115 static int __netdev_printk(const char *level, const struct net_device *dev,
6116 struct va_format *vaf)
6120 if (dev && dev->dev.parent) {
6121 r = dev_printk_emit(level[1] - '0',
6124 dev_driver_string(dev->dev.parent),
6125 dev_name(dev->dev.parent),
6126 netdev_name(dev), vaf);
6128 r = printk("%s%s: %pV", level, netdev_name(dev), vaf);
6130 r = printk("%s(NULL net_device): %pV", level, vaf);
6136 int netdev_printk(const char *level, const struct net_device *dev,
6137 const char *format, ...)
6139 struct va_format vaf;
6143 va_start(args, format);
6148 r = __netdev_printk(level, dev, &vaf);
6154 EXPORT_SYMBOL(netdev_printk);
6156 #define define_netdev_printk_level(func, level) \
6157 int func(const struct net_device *dev, const char *fmt, ...) \
6160 struct va_format vaf; \
6163 va_start(args, fmt); \
6168 r = __netdev_printk(level, dev, &vaf); \
6174 EXPORT_SYMBOL(func);
6176 define_netdev_printk_level(netdev_emerg, KERN_EMERG);
6177 define_netdev_printk_level(netdev_alert, KERN_ALERT);
6178 define_netdev_printk_level(netdev_crit, KERN_CRIT);
6179 define_netdev_printk_level(netdev_err, KERN_ERR);
6180 define_netdev_printk_level(netdev_warn, KERN_WARNING);
6181 define_netdev_printk_level(netdev_notice, KERN_NOTICE);
6182 define_netdev_printk_level(netdev_info, KERN_INFO);
6184 static void __net_exit netdev_exit(struct net *net)
6186 kfree(net->dev_name_head);
6187 kfree(net->dev_index_head);
6190 static struct pernet_operations __net_initdata netdev_net_ops = {
6191 .init = netdev_init,
6192 .exit = netdev_exit,
6195 static void __net_exit default_device_exit(struct net *net)
6197 struct net_device *dev, *aux;
6199 * Push all migratable network devices back to the
6200 * initial network namespace
6203 for_each_netdev_safe(net, dev, aux) {
6205 char fb_name[IFNAMSIZ];
6207 /* Ignore unmoveable devices (i.e. loopback) */
6208 if (dev->features & NETIF_F_NETNS_LOCAL)
6211 /* Leave virtual devices for the generic cleanup */
6212 if (dev->rtnl_link_ops)
6215 /* Push remaining network devices to init_net */
6216 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex);
6217 err = dev_change_net_namespace(dev, &init_net, fb_name);
6219 pr_emerg("%s: failed to move %s to init_net: %d\n",
6220 __func__, dev->name, err);
6227 static void __net_exit default_device_exit_batch(struct list_head *net_list)
6229 /* At exit all network devices most be removed from a network
6230 * namespace. Do this in the reverse order of registration.
6231 * Do this across as many network namespaces as possible to
6232 * improve batching efficiency.
6234 struct net_device *dev;
6236 LIST_HEAD(dev_kill_list);
6239 list_for_each_entry(net, net_list, exit_list) {
6240 for_each_netdev_reverse(net, dev) {
6241 if (dev->rtnl_link_ops)
6242 dev->rtnl_link_ops->dellink(dev, &dev_kill_list);
6244 unregister_netdevice_queue(dev, &dev_kill_list);
6247 unregister_netdevice_many(&dev_kill_list);
6248 list_del(&dev_kill_list);
6252 static struct pernet_operations __net_initdata default_device_ops = {
6253 .exit = default_device_exit,
6254 .exit_batch = default_device_exit_batch,
6258 * Initialize the DEV module. At boot time this walks the device list and
6259 * unhooks any devices that fail to initialise (normally hardware not
6260 * present) and leaves us with a valid list of present and active devices.
6265 * This is called single threaded during boot, so no need
6266 * to take the rtnl semaphore.
6268 static int __init net_dev_init(void)
6270 int i, rc = -ENOMEM;
6272 BUG_ON(!dev_boot_phase);
6274 if (dev_proc_init())
6277 if (netdev_kobject_init())
6280 INIT_LIST_HEAD(&ptype_all);
6281 for (i = 0; i < PTYPE_HASH_SIZE; i++)
6282 INIT_LIST_HEAD(&ptype_base[i]);
6284 INIT_LIST_HEAD(&offload_base);
6286 if (register_pernet_subsys(&netdev_net_ops))
6290 * Initialise the packet receive queues.
6293 for_each_possible_cpu(i) {
6294 struct softnet_data *sd = &per_cpu(softnet_data, i);
6296 memset(sd, 0, sizeof(*sd));
6297 skb_queue_head_init(&sd->input_pkt_queue);
6298 skb_queue_head_init(&sd->process_queue);
6299 sd->completion_queue = NULL;
6300 INIT_LIST_HEAD(&sd->poll_list);
6301 sd->output_queue = NULL;
6302 sd->output_queue_tailp = &sd->output_queue;
6304 sd->csd.func = rps_trigger_softirq;
6310 sd->backlog.poll = process_backlog;
6311 sd->backlog.weight = weight_p;
6312 sd->backlog.gro_list = NULL;
6313 sd->backlog.gro_count = 0;
6315 #ifdef CONFIG_NET_FLOW_LIMIT
6316 sd->flow_limit = NULL;
6322 /* The loopback device is special if any other network devices
6323 * is present in a network namespace the loopback device must
6324 * be present. Since we now dynamically allocate and free the
6325 * loopback device ensure this invariant is maintained by
6326 * keeping the loopback device as the first device on the
6327 * list of network devices. Ensuring the loopback devices
6328 * is the first device that appears and the last network device
6331 if (register_pernet_device(&loopback_net_ops))
6334 if (register_pernet_device(&default_device_ops))
6337 open_softirq(NET_TX_SOFTIRQ, net_tx_action);
6338 open_softirq(NET_RX_SOFTIRQ, net_rx_action);
6340 hotcpu_notifier(dev_cpu_callback, 0);
6347 subsys_initcall(net_dev_init);