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 int 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);
221 /* Device list removal
222 * caller must respect a RCU grace period before freeing/reusing dev
224 static void unlist_netdevice(struct net_device *dev)
228 /* Unlink dev from the device chain */
229 write_lock_bh(&dev_base_lock);
230 list_del_rcu(&dev->dev_list);
231 hlist_del_rcu(&dev->name_hlist);
232 hlist_del_rcu(&dev->index_hlist);
233 write_unlock_bh(&dev_base_lock);
235 dev_base_seq_inc(dev_net(dev));
242 static RAW_NOTIFIER_HEAD(netdev_chain);
245 * Device drivers call our routines to queue packets here. We empty the
246 * queue in the local softnet handler.
249 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data);
250 EXPORT_PER_CPU_SYMBOL(softnet_data);
252 #ifdef CONFIG_LOCKDEP
254 * register_netdevice() inits txq->_xmit_lock and sets lockdep class
255 * according to dev->type
257 static const unsigned short netdev_lock_type[] =
258 {ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25,
259 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET,
260 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM,
261 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP,
262 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD,
263 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25,
264 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP,
265 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD,
266 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI,
267 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE,
268 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET,
269 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL,
270 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM,
271 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE,
272 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE};
274 static const char *const netdev_lock_name[] =
275 {"_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25",
276 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET",
277 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM",
278 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP",
279 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD",
280 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25",
281 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP",
282 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD",
283 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI",
284 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE",
285 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET",
286 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL",
287 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM",
288 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE",
289 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"};
291 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)];
292 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)];
294 static inline unsigned short netdev_lock_pos(unsigned short dev_type)
298 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++)
299 if (netdev_lock_type[i] == dev_type)
301 /* the last key is used by default */
302 return ARRAY_SIZE(netdev_lock_type) - 1;
305 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
306 unsigned short dev_type)
310 i = netdev_lock_pos(dev_type);
311 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i],
312 netdev_lock_name[i]);
315 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
319 i = netdev_lock_pos(dev->type);
320 lockdep_set_class_and_name(&dev->addr_list_lock,
321 &netdev_addr_lock_key[i],
322 netdev_lock_name[i]);
325 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
326 unsigned short dev_type)
329 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
334 /*******************************************************************************
336 Protocol management and registration routines
338 *******************************************************************************/
341 * Add a protocol ID to the list. Now that the input handler is
342 * smarter we can dispense with all the messy stuff that used to be
345 * BEWARE!!! Protocol handlers, mangling input packets,
346 * MUST BE last in hash buckets and checking protocol handlers
347 * MUST start from promiscuous ptype_all chain in net_bh.
348 * It is true now, do not change it.
349 * Explanation follows: if protocol handler, mangling packet, will
350 * be the first on list, it is not able to sense, that packet
351 * is cloned and should be copied-on-write, so that it will
352 * change it and subsequent readers will get broken packet.
356 static inline struct list_head *ptype_head(const struct packet_type *pt)
358 if (pt->type == htons(ETH_P_ALL))
361 return &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK];
365 * dev_add_pack - add packet handler
366 * @pt: packet type declaration
368 * Add a protocol handler to the networking stack. The passed &packet_type
369 * is linked into kernel lists and may not be freed until it has been
370 * removed from the kernel lists.
372 * This call does not sleep therefore it can not
373 * guarantee all CPU's that are in middle of receiving packets
374 * will see the new packet type (until the next received packet).
377 void dev_add_pack(struct packet_type *pt)
379 struct list_head *head = ptype_head(pt);
381 spin_lock(&ptype_lock);
382 list_add_rcu(&pt->list, head);
383 spin_unlock(&ptype_lock);
385 EXPORT_SYMBOL(dev_add_pack);
388 * __dev_remove_pack - remove packet handler
389 * @pt: packet type declaration
391 * Remove a protocol handler that was previously added to the kernel
392 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
393 * from the kernel lists and can be freed or reused once this function
396 * The packet type might still be in use by receivers
397 * and must not be freed until after all the CPU's have gone
398 * through a quiescent state.
400 void __dev_remove_pack(struct packet_type *pt)
402 struct list_head *head = ptype_head(pt);
403 struct packet_type *pt1;
405 spin_lock(&ptype_lock);
407 list_for_each_entry(pt1, head, list) {
409 list_del_rcu(&pt->list);
414 pr_warn("dev_remove_pack: %p not found\n", pt);
416 spin_unlock(&ptype_lock);
418 EXPORT_SYMBOL(__dev_remove_pack);
421 * dev_remove_pack - remove packet handler
422 * @pt: packet type declaration
424 * Remove a protocol handler that was previously added to the kernel
425 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
426 * from the kernel lists and can be freed or reused once this function
429 * This call sleeps to guarantee that no CPU is looking at the packet
432 void dev_remove_pack(struct packet_type *pt)
434 __dev_remove_pack(pt);
438 EXPORT_SYMBOL(dev_remove_pack);
442 * dev_add_offload - register offload handlers
443 * @po: protocol offload declaration
445 * Add protocol offload handlers to the networking stack. The passed
446 * &proto_offload is linked into kernel lists and may not be freed until
447 * it has been removed from the kernel lists.
449 * This call does not sleep therefore it can not
450 * guarantee all CPU's that are in middle of receiving packets
451 * will see the new offload handlers (until the next received packet).
453 void dev_add_offload(struct packet_offload *po)
455 struct list_head *head = &offload_base;
457 spin_lock(&offload_lock);
458 list_add_rcu(&po->list, head);
459 spin_unlock(&offload_lock);
461 EXPORT_SYMBOL(dev_add_offload);
464 * __dev_remove_offload - remove offload handler
465 * @po: packet offload declaration
467 * Remove a protocol offload handler that was previously added to the
468 * kernel offload handlers by dev_add_offload(). The passed &offload_type
469 * is removed from the kernel lists and can be freed or reused once this
472 * The packet type might still be in use by receivers
473 * and must not be freed until after all the CPU's have gone
474 * through a quiescent state.
476 void __dev_remove_offload(struct packet_offload *po)
478 struct list_head *head = &offload_base;
479 struct packet_offload *po1;
481 spin_lock(&offload_lock);
483 list_for_each_entry(po1, head, list) {
485 list_del_rcu(&po->list);
490 pr_warn("dev_remove_offload: %p not found\n", po);
492 spin_unlock(&offload_lock);
494 EXPORT_SYMBOL(__dev_remove_offload);
497 * dev_remove_offload - remove packet offload handler
498 * @po: packet offload declaration
500 * Remove a packet offload handler that was previously added to the kernel
501 * offload handlers by dev_add_offload(). The passed &offload_type is
502 * removed from the kernel lists and can be freed or reused once this
505 * This call sleeps to guarantee that no CPU is looking at the packet
508 void dev_remove_offload(struct packet_offload *po)
510 __dev_remove_offload(po);
514 EXPORT_SYMBOL(dev_remove_offload);
516 /******************************************************************************
518 Device Boot-time Settings Routines
520 *******************************************************************************/
522 /* Boot time configuration table */
523 static struct netdev_boot_setup dev_boot_setup[NETDEV_BOOT_SETUP_MAX];
526 * netdev_boot_setup_add - add new setup entry
527 * @name: name of the device
528 * @map: configured settings for the device
530 * Adds new setup entry to the dev_boot_setup list. The function
531 * returns 0 on error and 1 on success. This is a generic routine to
534 static int netdev_boot_setup_add(char *name, struct ifmap *map)
536 struct netdev_boot_setup *s;
540 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
541 if (s[i].name[0] == '\0' || s[i].name[0] == ' ') {
542 memset(s[i].name, 0, sizeof(s[i].name));
543 strlcpy(s[i].name, name, IFNAMSIZ);
544 memcpy(&s[i].map, map, sizeof(s[i].map));
549 return i >= NETDEV_BOOT_SETUP_MAX ? 0 : 1;
553 * netdev_boot_setup_check - check boot time settings
554 * @dev: the netdevice
556 * Check boot time settings for the device.
557 * The found settings are set for the device to be used
558 * later in the device probing.
559 * Returns 0 if no settings found, 1 if they are.
561 int netdev_boot_setup_check(struct net_device *dev)
563 struct netdev_boot_setup *s = dev_boot_setup;
566 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
567 if (s[i].name[0] != '\0' && s[i].name[0] != ' ' &&
568 !strcmp(dev->name, s[i].name)) {
569 dev->irq = s[i].map.irq;
570 dev->base_addr = s[i].map.base_addr;
571 dev->mem_start = s[i].map.mem_start;
572 dev->mem_end = s[i].map.mem_end;
578 EXPORT_SYMBOL(netdev_boot_setup_check);
582 * netdev_boot_base - get address from boot time settings
583 * @prefix: prefix for network device
584 * @unit: id for network device
586 * Check boot time settings for the base address of device.
587 * The found settings are set for the device to be used
588 * later in the device probing.
589 * Returns 0 if no settings found.
591 unsigned long netdev_boot_base(const char *prefix, int unit)
593 const struct netdev_boot_setup *s = dev_boot_setup;
597 sprintf(name, "%s%d", prefix, unit);
600 * If device already registered then return base of 1
601 * to indicate not to probe for this interface
603 if (__dev_get_by_name(&init_net, name))
606 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++)
607 if (!strcmp(name, s[i].name))
608 return s[i].map.base_addr;
613 * Saves at boot time configured settings for any netdevice.
615 int __init netdev_boot_setup(char *str)
620 str = get_options(str, ARRAY_SIZE(ints), ints);
625 memset(&map, 0, sizeof(map));
629 map.base_addr = ints[2];
631 map.mem_start = ints[3];
633 map.mem_end = ints[4];
635 /* Add new entry to the list */
636 return netdev_boot_setup_add(str, &map);
639 __setup("netdev=", netdev_boot_setup);
641 /*******************************************************************************
643 Device Interface Subroutines
645 *******************************************************************************/
648 * __dev_get_by_name - find a device by its name
649 * @net: the applicable net namespace
650 * @name: name to find
652 * Find an interface by name. Must be called under RTNL semaphore
653 * or @dev_base_lock. If the name is found a pointer to the device
654 * is returned. If the name is not found then %NULL is returned. The
655 * reference counters are not incremented so the caller must be
656 * careful with locks.
659 struct net_device *__dev_get_by_name(struct net *net, const char *name)
661 struct net_device *dev;
662 struct hlist_head *head = dev_name_hash(net, name);
664 hlist_for_each_entry(dev, head, name_hlist)
665 if (!strncmp(dev->name, name, IFNAMSIZ))
670 EXPORT_SYMBOL(__dev_get_by_name);
673 * dev_get_by_name_rcu - find a device by its name
674 * @net: the applicable net namespace
675 * @name: name to find
677 * Find an interface by name.
678 * If the name is found a pointer to the device is returned.
679 * If the name is not found then %NULL is returned.
680 * The reference counters are not incremented so the caller must be
681 * careful with locks. The caller must hold RCU lock.
684 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name)
686 struct net_device *dev;
687 struct hlist_head *head = dev_name_hash(net, name);
689 hlist_for_each_entry_rcu(dev, head, name_hlist)
690 if (!strncmp(dev->name, name, IFNAMSIZ))
695 EXPORT_SYMBOL(dev_get_by_name_rcu);
698 * dev_get_by_name - find a device by its name
699 * @net: the applicable net namespace
700 * @name: name to find
702 * Find an interface by name. This can be called from any
703 * context and does its own locking. The returned handle has
704 * the usage count incremented and the caller must use dev_put() to
705 * release it when it is no longer needed. %NULL is returned if no
706 * matching device is found.
709 struct net_device *dev_get_by_name(struct net *net, const char *name)
711 struct net_device *dev;
714 dev = dev_get_by_name_rcu(net, name);
720 EXPORT_SYMBOL(dev_get_by_name);
723 * __dev_get_by_index - find a device by its ifindex
724 * @net: the applicable net namespace
725 * @ifindex: index of device
727 * Search for an interface by index. Returns %NULL if the device
728 * is not found or a pointer to the device. The device has not
729 * had its reference counter increased so the caller must be careful
730 * about locking. The caller must hold either the RTNL semaphore
734 struct net_device *__dev_get_by_index(struct net *net, int ifindex)
736 struct net_device *dev;
737 struct hlist_head *head = dev_index_hash(net, ifindex);
739 hlist_for_each_entry(dev, head, index_hlist)
740 if (dev->ifindex == ifindex)
745 EXPORT_SYMBOL(__dev_get_by_index);
748 * dev_get_by_index_rcu - find a device by its ifindex
749 * @net: the applicable net namespace
750 * @ifindex: index of device
752 * Search for an interface by index. Returns %NULL if the device
753 * is not found or a pointer to the device. The device has not
754 * had its reference counter increased so the caller must be careful
755 * about locking. The caller must hold RCU lock.
758 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex)
760 struct net_device *dev;
761 struct hlist_head *head = dev_index_hash(net, ifindex);
763 hlist_for_each_entry_rcu(dev, head, index_hlist)
764 if (dev->ifindex == ifindex)
769 EXPORT_SYMBOL(dev_get_by_index_rcu);
773 * dev_get_by_index - find a device by its ifindex
774 * @net: the applicable net namespace
775 * @ifindex: index of device
777 * Search for an interface by index. Returns NULL if the device
778 * is not found or a pointer to the device. The device returned has
779 * had a reference added and the pointer is safe until the user calls
780 * dev_put to indicate they have finished with it.
783 struct net_device *dev_get_by_index(struct net *net, int ifindex)
785 struct net_device *dev;
788 dev = dev_get_by_index_rcu(net, ifindex);
794 EXPORT_SYMBOL(dev_get_by_index);
797 * dev_getbyhwaddr_rcu - find a device by its hardware address
798 * @net: the applicable net namespace
799 * @type: media type of device
800 * @ha: hardware address
802 * Search for an interface by MAC address. Returns NULL if the device
803 * is not found or a pointer to the device.
804 * The caller must hold RCU or RTNL.
805 * The returned device has not had its ref count increased
806 * and the caller must therefore be careful about locking
810 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type,
813 struct net_device *dev;
815 for_each_netdev_rcu(net, dev)
816 if (dev->type == type &&
817 !memcmp(dev->dev_addr, ha, dev->addr_len))
822 EXPORT_SYMBOL(dev_getbyhwaddr_rcu);
824 struct net_device *__dev_getfirstbyhwtype(struct net *net, unsigned short type)
826 struct net_device *dev;
829 for_each_netdev(net, dev)
830 if (dev->type == type)
835 EXPORT_SYMBOL(__dev_getfirstbyhwtype);
837 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type)
839 struct net_device *dev, *ret = NULL;
842 for_each_netdev_rcu(net, dev)
843 if (dev->type == type) {
851 EXPORT_SYMBOL(dev_getfirstbyhwtype);
854 * dev_get_by_flags_rcu - find any device with given flags
855 * @net: the applicable net namespace
856 * @if_flags: IFF_* values
857 * @mask: bitmask of bits in if_flags to check
859 * Search for any interface with the given flags. Returns NULL if a device
860 * is not found or a pointer to the device. Must be called inside
861 * rcu_read_lock(), and result refcount is unchanged.
864 struct net_device *dev_get_by_flags_rcu(struct net *net, unsigned short if_flags,
867 struct net_device *dev, *ret;
870 for_each_netdev_rcu(net, dev) {
871 if (((dev->flags ^ if_flags) & mask) == 0) {
878 EXPORT_SYMBOL(dev_get_by_flags_rcu);
881 * dev_valid_name - check if name is okay for network device
884 * Network device names need to be valid file names to
885 * to allow sysfs to work. We also disallow any kind of
888 bool dev_valid_name(const char *name)
892 if (strlen(name) >= IFNAMSIZ)
894 if (!strcmp(name, ".") || !strcmp(name, ".."))
898 if (*name == '/' || isspace(*name))
904 EXPORT_SYMBOL(dev_valid_name);
907 * __dev_alloc_name - allocate a name for a device
908 * @net: network namespace to allocate the device name in
909 * @name: name format string
910 * @buf: scratch buffer and result name string
912 * Passed a format string - eg "lt%d" it will try and find a suitable
913 * id. It scans list of devices to build up a free map, then chooses
914 * the first empty slot. The caller must hold the dev_base or rtnl lock
915 * while allocating the name and adding the device in order to avoid
917 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
918 * Returns the number of the unit assigned or a negative errno code.
921 static int __dev_alloc_name(struct net *net, const char *name, char *buf)
925 const int max_netdevices = 8*PAGE_SIZE;
926 unsigned long *inuse;
927 struct net_device *d;
929 p = strnchr(name, IFNAMSIZ-1, '%');
932 * Verify the string as this thing may have come from
933 * the user. There must be either one "%d" and no other "%"
936 if (p[1] != 'd' || strchr(p + 2, '%'))
939 /* Use one page as a bit array of possible slots */
940 inuse = (unsigned long *) get_zeroed_page(GFP_ATOMIC);
944 for_each_netdev(net, d) {
945 if (!sscanf(d->name, name, &i))
947 if (i < 0 || i >= max_netdevices)
950 /* avoid cases where sscanf is not exact inverse of printf */
951 snprintf(buf, IFNAMSIZ, name, i);
952 if (!strncmp(buf, d->name, IFNAMSIZ))
956 i = find_first_zero_bit(inuse, max_netdevices);
957 free_page((unsigned long) inuse);
961 snprintf(buf, IFNAMSIZ, name, i);
962 if (!__dev_get_by_name(net, buf))
965 /* It is possible to run out of possible slots
966 * when the name is long and there isn't enough space left
967 * for the digits, or if all bits are used.
973 * dev_alloc_name - allocate a name for a device
975 * @name: name format string
977 * Passed a format string - eg "lt%d" it will try and find a suitable
978 * id. It scans list of devices to build up a free map, then chooses
979 * the first empty slot. The caller must hold the dev_base or rtnl lock
980 * while allocating the name and adding the device in order to avoid
982 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
983 * Returns the number of the unit assigned or a negative errno code.
986 int dev_alloc_name(struct net_device *dev, const char *name)
992 BUG_ON(!dev_net(dev));
994 ret = __dev_alloc_name(net, name, buf);
996 strlcpy(dev->name, buf, IFNAMSIZ);
999 EXPORT_SYMBOL(dev_alloc_name);
1001 static int dev_alloc_name_ns(struct net *net,
1002 struct net_device *dev,
1008 ret = __dev_alloc_name(net, name, buf);
1010 strlcpy(dev->name, buf, IFNAMSIZ);
1014 static int dev_get_valid_name(struct net *net,
1015 struct net_device *dev,
1020 if (!dev_valid_name(name))
1023 if (strchr(name, '%'))
1024 return dev_alloc_name_ns(net, dev, name);
1025 else if (__dev_get_by_name(net, name))
1027 else if (dev->name != name)
1028 strlcpy(dev->name, name, IFNAMSIZ);
1034 * dev_change_name - change name of a device
1036 * @newname: name (or format string) must be at least IFNAMSIZ
1038 * Change name of a device, can pass format strings "eth%d".
1041 int dev_change_name(struct net_device *dev, const char *newname)
1043 char oldname[IFNAMSIZ];
1049 BUG_ON(!dev_net(dev));
1052 if (dev->flags & IFF_UP)
1055 write_seqcount_begin(&devnet_rename_seq);
1057 if (strncmp(newname, dev->name, IFNAMSIZ) == 0) {
1058 write_seqcount_end(&devnet_rename_seq);
1062 memcpy(oldname, dev->name, IFNAMSIZ);
1064 err = dev_get_valid_name(net, dev, newname);
1066 write_seqcount_end(&devnet_rename_seq);
1071 ret = device_rename(&dev->dev, dev->name);
1073 memcpy(dev->name, oldname, IFNAMSIZ);
1074 write_seqcount_end(&devnet_rename_seq);
1078 write_seqcount_end(&devnet_rename_seq);
1080 write_lock_bh(&dev_base_lock);
1081 hlist_del_rcu(&dev->name_hlist);
1082 write_unlock_bh(&dev_base_lock);
1086 write_lock_bh(&dev_base_lock);
1087 hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
1088 write_unlock_bh(&dev_base_lock);
1090 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev);
1091 ret = notifier_to_errno(ret);
1094 /* err >= 0 after dev_alloc_name() or stores the first errno */
1097 write_seqcount_begin(&devnet_rename_seq);
1098 memcpy(dev->name, oldname, IFNAMSIZ);
1101 pr_err("%s: name change rollback failed: %d\n",
1110 * dev_set_alias - change ifalias of a device
1112 * @alias: name up to IFALIASZ
1113 * @len: limit of bytes to copy from info
1115 * Set ifalias for a device,
1117 int dev_set_alias(struct net_device *dev, const char *alias, size_t len)
1123 if (len >= IFALIASZ)
1127 kfree(dev->ifalias);
1128 dev->ifalias = NULL;
1132 new_ifalias = krealloc(dev->ifalias, len + 1, GFP_KERNEL);
1135 dev->ifalias = new_ifalias;
1137 strlcpy(dev->ifalias, alias, len+1);
1143 * netdev_features_change - device changes features
1144 * @dev: device to cause notification
1146 * Called to indicate a device has changed features.
1148 void netdev_features_change(struct net_device *dev)
1150 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev);
1152 EXPORT_SYMBOL(netdev_features_change);
1155 * netdev_state_change - device changes state
1156 * @dev: device to cause notification
1158 * Called to indicate a device has changed state. This function calls
1159 * the notifier chains for netdev_chain and sends a NEWLINK message
1160 * to the routing socket.
1162 void netdev_state_change(struct net_device *dev)
1164 if (dev->flags & IFF_UP) {
1165 call_netdevice_notifiers(NETDEV_CHANGE, dev);
1166 rtmsg_ifinfo(RTM_NEWLINK, dev, 0);
1169 EXPORT_SYMBOL(netdev_state_change);
1172 * netdev_notify_peers - notify network peers about existence of @dev
1173 * @dev: network device
1175 * Generate traffic such that interested network peers are aware of
1176 * @dev, such as by generating a gratuitous ARP. This may be used when
1177 * a device wants to inform the rest of the network about some sort of
1178 * reconfiguration such as a failover event or virtual machine
1181 void netdev_notify_peers(struct net_device *dev)
1184 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
1187 EXPORT_SYMBOL(netdev_notify_peers);
1189 static int __dev_open(struct net_device *dev)
1191 const struct net_device_ops *ops = dev->netdev_ops;
1196 if (!netif_device_present(dev))
1199 /* Block netpoll from trying to do any rx path servicing.
1200 * If we don't do this there is a chance ndo_poll_controller
1201 * or ndo_poll may be running while we open the device
1203 ret = netpoll_rx_disable(dev);
1207 ret = call_netdevice_notifiers(NETDEV_PRE_UP, dev);
1208 ret = notifier_to_errno(ret);
1212 set_bit(__LINK_STATE_START, &dev->state);
1214 if (ops->ndo_validate_addr)
1215 ret = ops->ndo_validate_addr(dev);
1217 if (!ret && ops->ndo_open)
1218 ret = ops->ndo_open(dev);
1220 netpoll_rx_enable(dev);
1223 clear_bit(__LINK_STATE_START, &dev->state);
1225 dev->flags |= IFF_UP;
1226 net_dmaengine_get();
1227 dev_set_rx_mode(dev);
1229 add_device_randomness(dev->dev_addr, dev->addr_len);
1236 * dev_open - prepare an interface for use.
1237 * @dev: device to open
1239 * Takes a device from down to up state. The device's private open
1240 * function is invoked and then the multicast lists are loaded. Finally
1241 * the device is moved into the up state and a %NETDEV_UP message is
1242 * sent to the netdev notifier chain.
1244 * Calling this function on an active interface is a nop. On a failure
1245 * a negative errno code is returned.
1247 int dev_open(struct net_device *dev)
1251 if (dev->flags & IFF_UP)
1254 ret = __dev_open(dev);
1258 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING);
1259 call_netdevice_notifiers(NETDEV_UP, dev);
1263 EXPORT_SYMBOL(dev_open);
1265 static int __dev_close_many(struct list_head *head)
1267 struct net_device *dev;
1272 list_for_each_entry(dev, head, unreg_list) {
1273 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev);
1275 clear_bit(__LINK_STATE_START, &dev->state);
1277 /* Synchronize to scheduled poll. We cannot touch poll list, it
1278 * can be even on different cpu. So just clear netif_running().
1280 * dev->stop() will invoke napi_disable() on all of it's
1281 * napi_struct instances on this device.
1283 smp_mb__after_clear_bit(); /* Commit netif_running(). */
1286 dev_deactivate_many(head);
1288 list_for_each_entry(dev, head, unreg_list) {
1289 const struct net_device_ops *ops = dev->netdev_ops;
1292 * Call the device specific close. This cannot fail.
1293 * Only if device is UP
1295 * We allow it to be called even after a DETACH hot-plug
1301 dev->flags &= ~IFF_UP;
1302 net_dmaengine_put();
1308 static int __dev_close(struct net_device *dev)
1313 /* Temporarily disable netpoll until the interface is down */
1314 retval = netpoll_rx_disable(dev);
1318 list_add(&dev->unreg_list, &single);
1319 retval = __dev_close_many(&single);
1322 netpoll_rx_enable(dev);
1326 static int dev_close_many(struct list_head *head)
1328 struct net_device *dev, *tmp;
1329 LIST_HEAD(tmp_list);
1331 list_for_each_entry_safe(dev, tmp, head, unreg_list)
1332 if (!(dev->flags & IFF_UP))
1333 list_move(&dev->unreg_list, &tmp_list);
1335 __dev_close_many(head);
1337 list_for_each_entry(dev, head, unreg_list) {
1338 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING);
1339 call_netdevice_notifiers(NETDEV_DOWN, dev);
1342 /* rollback_registered_many needs the complete original list */
1343 list_splice(&tmp_list, head);
1348 * dev_close - shutdown an interface.
1349 * @dev: device to shutdown
1351 * This function moves an active device into down state. A
1352 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device
1353 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier
1356 int dev_close(struct net_device *dev)
1359 if (dev->flags & IFF_UP) {
1362 /* Block netpoll rx while the interface is going down */
1363 ret = netpoll_rx_disable(dev);
1367 list_add(&dev->unreg_list, &single);
1368 dev_close_many(&single);
1371 netpoll_rx_enable(dev);
1375 EXPORT_SYMBOL(dev_close);
1379 * dev_disable_lro - disable Large Receive Offload on a device
1382 * Disable Large Receive Offload (LRO) on a net device. Must be
1383 * called under RTNL. This is needed if received packets may be
1384 * forwarded to another interface.
1386 void dev_disable_lro(struct net_device *dev)
1389 * If we're trying to disable lro on a vlan device
1390 * use the underlying physical device instead
1392 if (is_vlan_dev(dev))
1393 dev = vlan_dev_real_dev(dev);
1395 dev->wanted_features &= ~NETIF_F_LRO;
1396 netdev_update_features(dev);
1398 if (unlikely(dev->features & NETIF_F_LRO))
1399 netdev_WARN(dev, "failed to disable LRO!\n");
1401 EXPORT_SYMBOL(dev_disable_lro);
1404 static int dev_boot_phase = 1;
1407 * register_netdevice_notifier - register a network notifier block
1410 * Register a notifier to be called when network device events occur.
1411 * The notifier passed is linked into the kernel structures and must
1412 * not be reused until it has been unregistered. A negative errno code
1413 * is returned on a failure.
1415 * When registered all registration and up events are replayed
1416 * to the new notifier to allow device to have a race free
1417 * view of the network device list.
1420 int register_netdevice_notifier(struct notifier_block *nb)
1422 struct net_device *dev;
1423 struct net_device *last;
1428 err = raw_notifier_chain_register(&netdev_chain, nb);
1434 for_each_netdev(net, dev) {
1435 err = nb->notifier_call(nb, NETDEV_REGISTER, dev);
1436 err = notifier_to_errno(err);
1440 if (!(dev->flags & IFF_UP))
1443 nb->notifier_call(nb, NETDEV_UP, dev);
1454 for_each_netdev(net, dev) {
1458 if (dev->flags & IFF_UP) {
1459 nb->notifier_call(nb, NETDEV_GOING_DOWN, dev);
1460 nb->notifier_call(nb, NETDEV_DOWN, dev);
1462 nb->notifier_call(nb, NETDEV_UNREGISTER, dev);
1467 raw_notifier_chain_unregister(&netdev_chain, nb);
1470 EXPORT_SYMBOL(register_netdevice_notifier);
1473 * unregister_netdevice_notifier - unregister a network notifier block
1476 * Unregister a notifier previously registered by
1477 * register_netdevice_notifier(). The notifier is unlinked into the
1478 * kernel structures and may then be reused. A negative errno code
1479 * is returned on a failure.
1481 * After unregistering unregister and down device events are synthesized
1482 * for all devices on the device list to the removed notifier to remove
1483 * the need for special case cleanup code.
1486 int unregister_netdevice_notifier(struct notifier_block *nb)
1488 struct net_device *dev;
1493 err = raw_notifier_chain_unregister(&netdev_chain, nb);
1498 for_each_netdev(net, dev) {
1499 if (dev->flags & IFF_UP) {
1500 nb->notifier_call(nb, NETDEV_GOING_DOWN, dev);
1501 nb->notifier_call(nb, NETDEV_DOWN, dev);
1503 nb->notifier_call(nb, NETDEV_UNREGISTER, dev);
1510 EXPORT_SYMBOL(unregister_netdevice_notifier);
1513 * call_netdevice_notifiers - call all network notifier blocks
1514 * @val: value passed unmodified to notifier function
1515 * @dev: net_device pointer passed unmodified to notifier function
1517 * Call all network notifier blocks. Parameters and return value
1518 * are as for raw_notifier_call_chain().
1521 int call_netdevice_notifiers(unsigned long val, struct net_device *dev)
1524 return raw_notifier_call_chain(&netdev_chain, val, dev);
1526 EXPORT_SYMBOL(call_netdevice_notifiers);
1528 static struct static_key netstamp_needed __read_mostly;
1529 #ifdef HAVE_JUMP_LABEL
1530 /* We are not allowed to call static_key_slow_dec() from irq context
1531 * If net_disable_timestamp() is called from irq context, defer the
1532 * static_key_slow_dec() calls.
1534 static atomic_t netstamp_needed_deferred;
1537 void net_enable_timestamp(void)
1539 #ifdef HAVE_JUMP_LABEL
1540 int deferred = atomic_xchg(&netstamp_needed_deferred, 0);
1544 static_key_slow_dec(&netstamp_needed);
1548 WARN_ON(in_interrupt());
1549 static_key_slow_inc(&netstamp_needed);
1551 EXPORT_SYMBOL(net_enable_timestamp);
1553 void net_disable_timestamp(void)
1555 #ifdef HAVE_JUMP_LABEL
1556 if (in_interrupt()) {
1557 atomic_inc(&netstamp_needed_deferred);
1561 static_key_slow_dec(&netstamp_needed);
1563 EXPORT_SYMBOL(net_disable_timestamp);
1565 static inline void net_timestamp_set(struct sk_buff *skb)
1567 skb->tstamp.tv64 = 0;
1568 if (static_key_false(&netstamp_needed))
1569 __net_timestamp(skb);
1572 #define net_timestamp_check(COND, SKB) \
1573 if (static_key_false(&netstamp_needed)) { \
1574 if ((COND) && !(SKB)->tstamp.tv64) \
1575 __net_timestamp(SKB); \
1578 static inline bool is_skb_forwardable(struct net_device *dev,
1579 struct sk_buff *skb)
1583 if (!(dev->flags & IFF_UP))
1586 len = dev->mtu + dev->hard_header_len + VLAN_HLEN;
1587 if (skb->len <= len)
1590 /* if TSO is enabled, we don't care about the length as the packet
1591 * could be forwarded without being segmented before
1593 if (skb_is_gso(skb))
1600 * dev_forward_skb - loopback an skb to another netif
1602 * @dev: destination network device
1603 * @skb: buffer to forward
1606 * NET_RX_SUCCESS (no congestion)
1607 * NET_RX_DROP (packet was dropped, but freed)
1609 * dev_forward_skb can be used for injecting an skb from the
1610 * start_xmit function of one device into the receive queue
1611 * of another device.
1613 * The receiving device may be in another namespace, so
1614 * we have to clear all information in the skb that could
1615 * impact namespace isolation.
1617 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1619 if (skb_shinfo(skb)->tx_flags & SKBTX_DEV_ZEROCOPY) {
1620 if (skb_copy_ubufs(skb, GFP_ATOMIC)) {
1621 atomic_long_inc(&dev->rx_dropped);
1630 if (unlikely(!is_skb_forwardable(dev, skb))) {
1631 atomic_long_inc(&dev->rx_dropped);
1638 skb->tstamp.tv64 = 0;
1639 skb->pkt_type = PACKET_HOST;
1640 skb->protocol = eth_type_trans(skb, dev);
1644 return netif_rx(skb);
1646 EXPORT_SYMBOL_GPL(dev_forward_skb);
1648 static inline int deliver_skb(struct sk_buff *skb,
1649 struct packet_type *pt_prev,
1650 struct net_device *orig_dev)
1652 if (unlikely(skb_orphan_frags(skb, GFP_ATOMIC)))
1654 atomic_inc(&skb->users);
1655 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
1658 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb)
1660 if (!ptype->af_packet_priv || !skb->sk)
1663 if (ptype->id_match)
1664 return ptype->id_match(ptype, skb->sk);
1665 else if ((struct sock *)ptype->af_packet_priv == skb->sk)
1672 * Support routine. Sends outgoing frames to any network
1673 * taps currently in use.
1676 static void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
1678 struct packet_type *ptype;
1679 struct sk_buff *skb2 = NULL;
1680 struct packet_type *pt_prev = NULL;
1683 list_for_each_entry_rcu(ptype, &ptype_all, list) {
1684 /* Never send packets back to the socket
1685 * they originated from - MvS (miquels@drinkel.ow.org)
1687 if ((ptype->dev == dev || !ptype->dev) &&
1688 (!skb_loop_sk(ptype, skb))) {
1690 deliver_skb(skb2, pt_prev, skb->dev);
1695 skb2 = skb_clone(skb, GFP_ATOMIC);
1699 net_timestamp_set(skb2);
1701 /* skb->nh should be correctly
1702 set by sender, so that the second statement is
1703 just protection against buggy protocols.
1705 skb_reset_mac_header(skb2);
1707 if (skb_network_header(skb2) < skb2->data ||
1708 skb2->network_header > skb2->tail) {
1709 net_crit_ratelimited("protocol %04x is buggy, dev %s\n",
1710 ntohs(skb2->protocol),
1712 skb_reset_network_header(skb2);
1715 skb2->transport_header = skb2->network_header;
1716 skb2->pkt_type = PACKET_OUTGOING;
1721 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
1726 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change
1727 * @dev: Network device
1728 * @txq: number of queues available
1730 * If real_num_tx_queues is changed the tc mappings may no longer be
1731 * valid. To resolve this verify the tc mapping remains valid and if
1732 * not NULL the mapping. With no priorities mapping to this
1733 * offset/count pair it will no longer be used. In the worst case TC0
1734 * is invalid nothing can be done so disable priority mappings. If is
1735 * expected that drivers will fix this mapping if they can before
1736 * calling netif_set_real_num_tx_queues.
1738 static void netif_setup_tc(struct net_device *dev, unsigned int txq)
1741 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
1743 /* If TC0 is invalidated disable TC mapping */
1744 if (tc->offset + tc->count > txq) {
1745 pr_warn("Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n");
1750 /* Invalidated prio to tc mappings set to TC0 */
1751 for (i = 1; i < TC_BITMASK + 1; i++) {
1752 int q = netdev_get_prio_tc_map(dev, i);
1754 tc = &dev->tc_to_txq[q];
1755 if (tc->offset + tc->count > txq) {
1756 pr_warn("Number of in use tx queues changed. Priority %i to tc mapping %i is no longer valid. Setting map to 0\n",
1758 netdev_set_prio_tc_map(dev, i, 0);
1764 static DEFINE_MUTEX(xps_map_mutex);
1765 #define xmap_dereference(P) \
1766 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex))
1768 static struct xps_map *remove_xps_queue(struct xps_dev_maps *dev_maps,
1771 struct xps_map *map = NULL;
1775 map = xmap_dereference(dev_maps->cpu_map[cpu]);
1777 for (pos = 0; map && pos < map->len; pos++) {
1778 if (map->queues[pos] == index) {
1780 map->queues[pos] = map->queues[--map->len];
1782 RCU_INIT_POINTER(dev_maps->cpu_map[cpu], NULL);
1783 kfree_rcu(map, rcu);
1793 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index)
1795 struct xps_dev_maps *dev_maps;
1797 bool active = false;
1799 mutex_lock(&xps_map_mutex);
1800 dev_maps = xmap_dereference(dev->xps_maps);
1805 for_each_possible_cpu(cpu) {
1806 for (i = index; i < dev->num_tx_queues; i++) {
1807 if (!remove_xps_queue(dev_maps, cpu, i))
1810 if (i == dev->num_tx_queues)
1815 RCU_INIT_POINTER(dev->xps_maps, NULL);
1816 kfree_rcu(dev_maps, rcu);
1819 for (i = index; i < dev->num_tx_queues; i++)
1820 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, i),
1824 mutex_unlock(&xps_map_mutex);
1827 static struct xps_map *expand_xps_map(struct xps_map *map,
1830 struct xps_map *new_map;
1831 int alloc_len = XPS_MIN_MAP_ALLOC;
1834 for (pos = 0; map && pos < map->len; pos++) {
1835 if (map->queues[pos] != index)
1840 /* Need to add queue to this CPU's existing map */
1842 if (pos < map->alloc_len)
1845 alloc_len = map->alloc_len * 2;
1848 /* Need to allocate new map to store queue on this CPU's map */
1849 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL,
1854 for (i = 0; i < pos; i++)
1855 new_map->queues[i] = map->queues[i];
1856 new_map->alloc_len = alloc_len;
1862 int netif_set_xps_queue(struct net_device *dev, struct cpumask *mask, u16 index)
1864 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL;
1865 struct xps_map *map, *new_map;
1866 int maps_sz = max_t(unsigned int, XPS_DEV_MAPS_SIZE, L1_CACHE_BYTES);
1867 int cpu, numa_node_id = -2;
1868 bool active = false;
1870 mutex_lock(&xps_map_mutex);
1872 dev_maps = xmap_dereference(dev->xps_maps);
1874 /* allocate memory for queue storage */
1875 for_each_online_cpu(cpu) {
1876 if (!cpumask_test_cpu(cpu, mask))
1880 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL);
1881 if (!new_dev_maps) {
1882 mutex_unlock(&xps_map_mutex);
1886 map = dev_maps ? xmap_dereference(dev_maps->cpu_map[cpu]) :
1889 map = expand_xps_map(map, cpu, index);
1893 RCU_INIT_POINTER(new_dev_maps->cpu_map[cpu], map);
1897 goto out_no_new_maps;
1899 for_each_possible_cpu(cpu) {
1900 if (cpumask_test_cpu(cpu, mask) && cpu_online(cpu)) {
1901 /* add queue to CPU maps */
1904 map = xmap_dereference(new_dev_maps->cpu_map[cpu]);
1905 while ((pos < map->len) && (map->queues[pos] != index))
1908 if (pos == map->len)
1909 map->queues[map->len++] = index;
1911 if (numa_node_id == -2)
1912 numa_node_id = cpu_to_node(cpu);
1913 else if (numa_node_id != cpu_to_node(cpu))
1916 } else if (dev_maps) {
1917 /* fill in the new device map from the old device map */
1918 map = xmap_dereference(dev_maps->cpu_map[cpu]);
1919 RCU_INIT_POINTER(new_dev_maps->cpu_map[cpu], map);
1924 rcu_assign_pointer(dev->xps_maps, new_dev_maps);
1926 /* Cleanup old maps */
1928 for_each_possible_cpu(cpu) {
1929 new_map = xmap_dereference(new_dev_maps->cpu_map[cpu]);
1930 map = xmap_dereference(dev_maps->cpu_map[cpu]);
1931 if (map && map != new_map)
1932 kfree_rcu(map, rcu);
1935 kfree_rcu(dev_maps, rcu);
1938 dev_maps = new_dev_maps;
1942 /* update Tx queue numa node */
1943 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index),
1944 (numa_node_id >= 0) ? numa_node_id :
1950 /* removes queue from unused CPUs */
1951 for_each_possible_cpu(cpu) {
1952 if (cpumask_test_cpu(cpu, mask) && cpu_online(cpu))
1955 if (remove_xps_queue(dev_maps, cpu, index))
1959 /* free map if not active */
1961 RCU_INIT_POINTER(dev->xps_maps, NULL);
1962 kfree_rcu(dev_maps, rcu);
1966 mutex_unlock(&xps_map_mutex);
1970 /* remove any maps that we added */
1971 for_each_possible_cpu(cpu) {
1972 new_map = xmap_dereference(new_dev_maps->cpu_map[cpu]);
1973 map = dev_maps ? xmap_dereference(dev_maps->cpu_map[cpu]) :
1975 if (new_map && new_map != map)
1979 mutex_unlock(&xps_map_mutex);
1981 kfree(new_dev_maps);
1984 EXPORT_SYMBOL(netif_set_xps_queue);
1988 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues
1989 * greater then real_num_tx_queues stale skbs on the qdisc must be flushed.
1991 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq)
1995 if (txq < 1 || txq > dev->num_tx_queues)
1998 if (dev->reg_state == NETREG_REGISTERED ||
1999 dev->reg_state == NETREG_UNREGISTERING) {
2002 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues,
2008 netif_setup_tc(dev, txq);
2010 if (txq < dev->real_num_tx_queues) {
2011 qdisc_reset_all_tx_gt(dev, txq);
2013 netif_reset_xps_queues_gt(dev, txq);
2018 dev->real_num_tx_queues = txq;
2021 EXPORT_SYMBOL(netif_set_real_num_tx_queues);
2025 * netif_set_real_num_rx_queues - set actual number of RX queues used
2026 * @dev: Network device
2027 * @rxq: Actual number of RX queues
2029 * This must be called either with the rtnl_lock held or before
2030 * registration of the net device. Returns 0 on success, or a
2031 * negative error code. If called before registration, it always
2034 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq)
2038 if (rxq < 1 || rxq > dev->num_rx_queues)
2041 if (dev->reg_state == NETREG_REGISTERED) {
2044 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues,
2050 dev->real_num_rx_queues = rxq;
2053 EXPORT_SYMBOL(netif_set_real_num_rx_queues);
2057 * netif_get_num_default_rss_queues - default number of RSS queues
2059 * This routine should set an upper limit on the number of RSS queues
2060 * used by default by multiqueue devices.
2062 int netif_get_num_default_rss_queues(void)
2064 return min_t(int, DEFAULT_MAX_NUM_RSS_QUEUES, num_online_cpus());
2066 EXPORT_SYMBOL(netif_get_num_default_rss_queues);
2068 static inline void __netif_reschedule(struct Qdisc *q)
2070 struct softnet_data *sd;
2071 unsigned long flags;
2073 local_irq_save(flags);
2074 sd = &__get_cpu_var(softnet_data);
2075 q->next_sched = NULL;
2076 *sd->output_queue_tailp = q;
2077 sd->output_queue_tailp = &q->next_sched;
2078 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2079 local_irq_restore(flags);
2082 void __netif_schedule(struct Qdisc *q)
2084 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
2085 __netif_reschedule(q);
2087 EXPORT_SYMBOL(__netif_schedule);
2089 void dev_kfree_skb_irq(struct sk_buff *skb)
2091 if (atomic_dec_and_test(&skb->users)) {
2092 struct softnet_data *sd;
2093 unsigned long flags;
2095 local_irq_save(flags);
2096 sd = &__get_cpu_var(softnet_data);
2097 skb->next = sd->completion_queue;
2098 sd->completion_queue = skb;
2099 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2100 local_irq_restore(flags);
2103 EXPORT_SYMBOL(dev_kfree_skb_irq);
2105 void dev_kfree_skb_any(struct sk_buff *skb)
2107 if (in_irq() || irqs_disabled())
2108 dev_kfree_skb_irq(skb);
2112 EXPORT_SYMBOL(dev_kfree_skb_any);
2116 * netif_device_detach - mark device as removed
2117 * @dev: network device
2119 * Mark device as removed from system and therefore no longer available.
2121 void netif_device_detach(struct net_device *dev)
2123 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) &&
2124 netif_running(dev)) {
2125 netif_tx_stop_all_queues(dev);
2128 EXPORT_SYMBOL(netif_device_detach);
2131 * netif_device_attach - mark device as attached
2132 * @dev: network device
2134 * Mark device as attached from system and restart if needed.
2136 void netif_device_attach(struct net_device *dev)
2138 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) &&
2139 netif_running(dev)) {
2140 netif_tx_wake_all_queues(dev);
2141 __netdev_watchdog_up(dev);
2144 EXPORT_SYMBOL(netif_device_attach);
2146 static void skb_warn_bad_offload(const struct sk_buff *skb)
2148 static const netdev_features_t null_features = 0;
2149 struct net_device *dev = skb->dev;
2150 const char *driver = "";
2152 if (dev && dev->dev.parent)
2153 driver = dev_driver_string(dev->dev.parent);
2155 WARN(1, "%s: caps=(%pNF, %pNF) len=%d data_len=%d gso_size=%d "
2156 "gso_type=%d ip_summed=%d\n",
2157 driver, dev ? &dev->features : &null_features,
2158 skb->sk ? &skb->sk->sk_route_caps : &null_features,
2159 skb->len, skb->data_len, skb_shinfo(skb)->gso_size,
2160 skb_shinfo(skb)->gso_type, skb->ip_summed);
2164 * Invalidate hardware checksum when packet is to be mangled, and
2165 * complete checksum manually on outgoing path.
2167 int skb_checksum_help(struct sk_buff *skb)
2170 int ret = 0, offset;
2172 if (skb->ip_summed == CHECKSUM_COMPLETE)
2173 goto out_set_summed;
2175 if (unlikely(skb_shinfo(skb)->gso_size)) {
2176 skb_warn_bad_offload(skb);
2180 /* Before computing a checksum, we should make sure no frag could
2181 * be modified by an external entity : checksum could be wrong.
2183 if (skb_has_shared_frag(skb)) {
2184 ret = __skb_linearize(skb);
2189 offset = skb_checksum_start_offset(skb);
2190 BUG_ON(offset >= skb_headlen(skb));
2191 csum = skb_checksum(skb, offset, skb->len - offset, 0);
2193 offset += skb->csum_offset;
2194 BUG_ON(offset + sizeof(__sum16) > skb_headlen(skb));
2196 if (skb_cloned(skb) &&
2197 !skb_clone_writable(skb, offset + sizeof(__sum16))) {
2198 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2203 *(__sum16 *)(skb->data + offset) = csum_fold(csum);
2205 skb->ip_summed = CHECKSUM_NONE;
2209 EXPORT_SYMBOL(skb_checksum_help);
2211 __be16 skb_network_protocol(struct sk_buff *skb)
2213 __be16 type = skb->protocol;
2215 while (type == htons(ETH_P_8021Q)) {
2216 int vlan_depth = ETH_HLEN;
2217 struct vlan_hdr *vh;
2219 if (unlikely(!pskb_may_pull(skb, vlan_depth + VLAN_HLEN)))
2222 vh = (struct vlan_hdr *)(skb->data + vlan_depth);
2223 type = vh->h_vlan_encapsulated_proto;
2224 vlan_depth += VLAN_HLEN;
2231 * skb_mac_gso_segment - mac layer segmentation handler.
2232 * @skb: buffer to segment
2233 * @features: features for the output path (see dev->features)
2235 struct sk_buff *skb_mac_gso_segment(struct sk_buff *skb,
2236 netdev_features_t features)
2238 struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT);
2239 struct packet_offload *ptype;
2240 __be16 type = skb_network_protocol(skb);
2242 if (unlikely(!type))
2243 return ERR_PTR(-EINVAL);
2245 __skb_pull(skb, skb->mac_len);
2248 list_for_each_entry_rcu(ptype, &offload_base, list) {
2249 if (ptype->type == type && ptype->callbacks.gso_segment) {
2250 if (unlikely(skb->ip_summed != CHECKSUM_PARTIAL)) {
2253 err = ptype->callbacks.gso_send_check(skb);
2254 segs = ERR_PTR(err);
2255 if (err || skb_gso_ok(skb, features))
2257 __skb_push(skb, (skb->data -
2258 skb_network_header(skb)));
2260 segs = ptype->callbacks.gso_segment(skb, features);
2266 __skb_push(skb, skb->data - skb_mac_header(skb));
2270 EXPORT_SYMBOL(skb_mac_gso_segment);
2273 /* openvswitch calls this on rx path, so we need a different check.
2275 static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
2278 return skb->ip_summed != CHECKSUM_PARTIAL;
2280 return skb->ip_summed == CHECKSUM_NONE;
2284 * __skb_gso_segment - Perform segmentation on skb.
2285 * @skb: buffer to segment
2286 * @features: features for the output path (see dev->features)
2287 * @tx_path: whether it is called in TX path
2289 * This function segments the given skb and returns a list of segments.
2291 * It may return NULL if the skb requires no segmentation. This is
2292 * only possible when GSO is used for verifying header integrity.
2294 struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
2295 netdev_features_t features, bool tx_path)
2297 if (unlikely(skb_needs_check(skb, tx_path))) {
2300 skb_warn_bad_offload(skb);
2302 if (skb_header_cloned(skb) &&
2303 (err = pskb_expand_head(skb, 0, 0, GFP_ATOMIC)))
2304 return ERR_PTR(err);
2307 SKB_GSO_CB(skb)->mac_offset = skb_headroom(skb);
2308 skb_reset_mac_header(skb);
2309 skb_reset_mac_len(skb);
2311 return skb_mac_gso_segment(skb, features);
2313 EXPORT_SYMBOL(__skb_gso_segment);
2315 /* Take action when hardware reception checksum errors are detected. */
2317 void netdev_rx_csum_fault(struct net_device *dev)
2319 if (net_ratelimit()) {
2320 pr_err("%s: hw csum failure\n", dev ? dev->name : "<unknown>");
2324 EXPORT_SYMBOL(netdev_rx_csum_fault);
2327 /* Actually, we should eliminate this check as soon as we know, that:
2328 * 1. IOMMU is present and allows to map all the memory.
2329 * 2. No high memory really exists on this machine.
2332 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
2334 #ifdef CONFIG_HIGHMEM
2336 if (!(dev->features & NETIF_F_HIGHDMA)) {
2337 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2338 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2339 if (PageHighMem(skb_frag_page(frag)))
2344 if (PCI_DMA_BUS_IS_PHYS) {
2345 struct device *pdev = dev->dev.parent;
2349 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2350 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2351 dma_addr_t addr = page_to_phys(skb_frag_page(frag));
2352 if (!pdev->dma_mask || addr + PAGE_SIZE - 1 > *pdev->dma_mask)
2361 void (*destructor)(struct sk_buff *skb);
2364 #define DEV_GSO_CB(skb) ((struct dev_gso_cb *)(skb)->cb)
2366 static void dev_gso_skb_destructor(struct sk_buff *skb)
2368 struct dev_gso_cb *cb;
2371 struct sk_buff *nskb = skb->next;
2373 skb->next = nskb->next;
2376 } while (skb->next);
2378 cb = DEV_GSO_CB(skb);
2380 cb->destructor(skb);
2384 * dev_gso_segment - Perform emulated hardware segmentation on skb.
2385 * @skb: buffer to segment
2386 * @features: device features as applicable to this skb
2388 * This function segments the given skb and stores the list of segments
2391 static int dev_gso_segment(struct sk_buff *skb, netdev_features_t features)
2393 struct sk_buff *segs;
2395 segs = skb_gso_segment(skb, features);
2397 /* Verifying header integrity only. */
2402 return PTR_ERR(segs);
2405 DEV_GSO_CB(skb)->destructor = skb->destructor;
2406 skb->destructor = dev_gso_skb_destructor;
2411 static netdev_features_t harmonize_features(struct sk_buff *skb,
2412 __be16 protocol, netdev_features_t features)
2414 if (skb->ip_summed != CHECKSUM_NONE &&
2415 !can_checksum_protocol(features, protocol)) {
2416 features &= ~NETIF_F_ALL_CSUM;
2417 } else if (illegal_highdma(skb->dev, skb)) {
2418 features &= ~NETIF_F_SG;
2424 netdev_features_t netif_skb_features(struct sk_buff *skb)
2426 __be16 protocol = skb->protocol;
2427 netdev_features_t features = skb->dev->features;
2429 if (skb_shinfo(skb)->gso_segs > skb->dev->gso_max_segs)
2430 features &= ~NETIF_F_GSO_MASK;
2432 if (protocol == htons(ETH_P_8021Q)) {
2433 struct vlan_ethhdr *veh = (struct vlan_ethhdr *)skb->data;
2434 protocol = veh->h_vlan_encapsulated_proto;
2435 } else if (!vlan_tx_tag_present(skb)) {
2436 return harmonize_features(skb, protocol, features);
2439 features &= (skb->dev->vlan_features | NETIF_F_HW_VLAN_TX);
2441 if (protocol != htons(ETH_P_8021Q)) {
2442 return harmonize_features(skb, protocol, features);
2444 features &= NETIF_F_SG | NETIF_F_HIGHDMA | NETIF_F_FRAGLIST |
2445 NETIF_F_GEN_CSUM | NETIF_F_HW_VLAN_TX;
2446 return harmonize_features(skb, protocol, features);
2449 EXPORT_SYMBOL(netif_skb_features);
2452 * Returns true if either:
2453 * 1. skb has frag_list and the device doesn't support FRAGLIST, or
2454 * 2. skb is fragmented and the device does not support SG.
2456 static inline int skb_needs_linearize(struct sk_buff *skb,
2459 return skb_is_nonlinear(skb) &&
2460 ((skb_has_frag_list(skb) &&
2461 !(features & NETIF_F_FRAGLIST)) ||
2462 (skb_shinfo(skb)->nr_frags &&
2463 !(features & NETIF_F_SG)));
2466 int dev_hard_start_xmit(struct sk_buff *skb, struct net_device *dev,
2467 struct netdev_queue *txq)
2469 const struct net_device_ops *ops = dev->netdev_ops;
2470 int rc = NETDEV_TX_OK;
2471 unsigned int skb_len;
2473 if (likely(!skb->next)) {
2474 netdev_features_t features;
2477 * If device doesn't need skb->dst, release it right now while
2478 * its hot in this cpu cache
2480 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
2483 features = netif_skb_features(skb);
2485 if (vlan_tx_tag_present(skb) &&
2486 !(features & NETIF_F_HW_VLAN_TX)) {
2487 skb = __vlan_put_tag(skb, vlan_tx_tag_get(skb));
2494 /* If encapsulation offload request, verify we are testing
2495 * hardware encapsulation features instead of standard
2496 * features for the netdev
2498 if (skb->encapsulation)
2499 features &= dev->hw_enc_features;
2501 if (netif_needs_gso(skb, features)) {
2502 if (unlikely(dev_gso_segment(skb, features)))
2507 if (skb_needs_linearize(skb, features) &&
2508 __skb_linearize(skb))
2511 /* If packet is not checksummed and device does not
2512 * support checksumming for this protocol, complete
2513 * checksumming here.
2515 if (skb->ip_summed == CHECKSUM_PARTIAL) {
2516 if (skb->encapsulation)
2517 skb_set_inner_transport_header(skb,
2518 skb_checksum_start_offset(skb));
2520 skb_set_transport_header(skb,
2521 skb_checksum_start_offset(skb));
2522 if (!(features & NETIF_F_ALL_CSUM) &&
2523 skb_checksum_help(skb))
2528 if (!list_empty(&ptype_all))
2529 dev_queue_xmit_nit(skb, dev);
2532 rc = ops->ndo_start_xmit(skb, dev);
2533 trace_net_dev_xmit(skb, rc, dev, skb_len);
2534 if (rc == NETDEV_TX_OK)
2535 txq_trans_update(txq);
2541 struct sk_buff *nskb = skb->next;
2543 skb->next = nskb->next;
2547 * If device doesn't need nskb->dst, release it right now while
2548 * its hot in this cpu cache
2550 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
2553 if (!list_empty(&ptype_all))
2554 dev_queue_xmit_nit(nskb, dev);
2556 skb_len = nskb->len;
2557 rc = ops->ndo_start_xmit(nskb, dev);
2558 trace_net_dev_xmit(nskb, rc, dev, skb_len);
2559 if (unlikely(rc != NETDEV_TX_OK)) {
2560 if (rc & ~NETDEV_TX_MASK)
2561 goto out_kfree_gso_skb;
2562 nskb->next = skb->next;
2566 txq_trans_update(txq);
2567 if (unlikely(netif_xmit_stopped(txq) && skb->next))
2568 return NETDEV_TX_BUSY;
2569 } while (skb->next);
2572 if (likely(skb->next == NULL))
2573 skb->destructor = DEV_GSO_CB(skb)->destructor;
2580 static void qdisc_pkt_len_init(struct sk_buff *skb)
2582 const struct skb_shared_info *shinfo = skb_shinfo(skb);
2584 qdisc_skb_cb(skb)->pkt_len = skb->len;
2586 /* To get more precise estimation of bytes sent on wire,
2587 * we add to pkt_len the headers size of all segments
2589 if (shinfo->gso_size) {
2590 unsigned int hdr_len;
2592 /* mac layer + network layer */
2593 hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
2595 /* + transport layer */
2596 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))
2597 hdr_len += tcp_hdrlen(skb);
2599 hdr_len += sizeof(struct udphdr);
2600 qdisc_skb_cb(skb)->pkt_len += (shinfo->gso_segs - 1) * hdr_len;
2604 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
2605 struct net_device *dev,
2606 struct netdev_queue *txq)
2608 spinlock_t *root_lock = qdisc_lock(q);
2612 qdisc_pkt_len_init(skb);
2613 qdisc_calculate_pkt_len(skb, q);
2615 * Heuristic to force contended enqueues to serialize on a
2616 * separate lock before trying to get qdisc main lock.
2617 * This permits __QDISC_STATE_RUNNING owner to get the lock more often
2618 * and dequeue packets faster.
2620 contended = qdisc_is_running(q);
2621 if (unlikely(contended))
2622 spin_lock(&q->busylock);
2624 spin_lock(root_lock);
2625 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
2628 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
2629 qdisc_run_begin(q)) {
2631 * This is a work-conserving queue; there are no old skbs
2632 * waiting to be sent out; and the qdisc is not running -
2633 * xmit the skb directly.
2635 if (!(dev->priv_flags & IFF_XMIT_DST_RELEASE))
2638 qdisc_bstats_update(q, skb);
2640 if (sch_direct_xmit(skb, q, dev, txq, root_lock)) {
2641 if (unlikely(contended)) {
2642 spin_unlock(&q->busylock);
2649 rc = NET_XMIT_SUCCESS;
2652 rc = q->enqueue(skb, q) & NET_XMIT_MASK;
2653 if (qdisc_run_begin(q)) {
2654 if (unlikely(contended)) {
2655 spin_unlock(&q->busylock);
2661 spin_unlock(root_lock);
2662 if (unlikely(contended))
2663 spin_unlock(&q->busylock);
2667 #if IS_ENABLED(CONFIG_NETPRIO_CGROUP)
2668 static void skb_update_prio(struct sk_buff *skb)
2670 struct netprio_map *map = rcu_dereference_bh(skb->dev->priomap);
2672 if (!skb->priority && skb->sk && map) {
2673 unsigned int prioidx = skb->sk->sk_cgrp_prioidx;
2675 if (prioidx < map->priomap_len)
2676 skb->priority = map->priomap[prioidx];
2680 #define skb_update_prio(skb)
2683 static DEFINE_PER_CPU(int, xmit_recursion);
2684 #define RECURSION_LIMIT 10
2687 * dev_loopback_xmit - loop back @skb
2688 * @skb: buffer to transmit
2690 int dev_loopback_xmit(struct sk_buff *skb)
2692 skb_reset_mac_header(skb);
2693 __skb_pull(skb, skb_network_offset(skb));
2694 skb->pkt_type = PACKET_LOOPBACK;
2695 skb->ip_summed = CHECKSUM_UNNECESSARY;
2696 WARN_ON(!skb_dst(skb));
2701 EXPORT_SYMBOL(dev_loopback_xmit);
2704 * dev_queue_xmit - transmit a buffer
2705 * @skb: buffer to transmit
2707 * Queue a buffer for transmission to a network device. The caller must
2708 * have set the device and priority and built the buffer before calling
2709 * this function. The function can be called from an interrupt.
2711 * A negative errno code is returned on a failure. A success does not
2712 * guarantee the frame will be transmitted as it may be dropped due
2713 * to congestion or traffic shaping.
2715 * -----------------------------------------------------------------------------------
2716 * I notice this method can also return errors from the queue disciplines,
2717 * including NET_XMIT_DROP, which is a positive value. So, errors can also
2720 * Regardless of the return value, the skb is consumed, so it is currently
2721 * difficult to retry a send to this method. (You can bump the ref count
2722 * before sending to hold a reference for retry if you are careful.)
2724 * When calling this method, interrupts MUST be enabled. This is because
2725 * the BH enable code must have IRQs enabled so that it will not deadlock.
2728 int dev_queue_xmit(struct sk_buff *skb)
2730 struct net_device *dev = skb->dev;
2731 struct netdev_queue *txq;
2735 skb_reset_mac_header(skb);
2737 /* Disable soft irqs for various locks below. Also
2738 * stops preemption for RCU.
2742 skb_update_prio(skb);
2744 txq = netdev_pick_tx(dev, skb);
2745 q = rcu_dereference_bh(txq->qdisc);
2747 #ifdef CONFIG_NET_CLS_ACT
2748 skb->tc_verd = SET_TC_AT(skb->tc_verd, AT_EGRESS);
2750 trace_net_dev_queue(skb);
2752 rc = __dev_xmit_skb(skb, q, dev, txq);
2756 /* The device has no queue. Common case for software devices:
2757 loopback, all the sorts of tunnels...
2759 Really, it is unlikely that netif_tx_lock protection is necessary
2760 here. (f.e. loopback and IP tunnels are clean ignoring statistics
2762 However, it is possible, that they rely on protection
2765 Check this and shot the lock. It is not prone from deadlocks.
2766 Either shot noqueue qdisc, it is even simpler 8)
2768 if (dev->flags & IFF_UP) {
2769 int cpu = smp_processor_id(); /* ok because BHs are off */
2771 if (txq->xmit_lock_owner != cpu) {
2773 if (__this_cpu_read(xmit_recursion) > RECURSION_LIMIT)
2774 goto recursion_alert;
2776 HARD_TX_LOCK(dev, txq, cpu);
2778 if (!netif_xmit_stopped(txq)) {
2779 __this_cpu_inc(xmit_recursion);
2780 rc = dev_hard_start_xmit(skb, dev, txq);
2781 __this_cpu_dec(xmit_recursion);
2782 if (dev_xmit_complete(rc)) {
2783 HARD_TX_UNLOCK(dev, txq);
2787 HARD_TX_UNLOCK(dev, txq);
2788 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
2791 /* Recursion is detected! It is possible,
2795 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
2801 rcu_read_unlock_bh();
2806 rcu_read_unlock_bh();
2809 EXPORT_SYMBOL(dev_queue_xmit);
2812 /*=======================================================================
2814 =======================================================================*/
2816 int netdev_max_backlog __read_mostly = 1000;
2817 EXPORT_SYMBOL(netdev_max_backlog);
2819 int netdev_tstamp_prequeue __read_mostly = 1;
2820 int netdev_budget __read_mostly = 300;
2821 int weight_p __read_mostly = 64; /* old backlog weight */
2823 /* Called with irq disabled */
2824 static inline void ____napi_schedule(struct softnet_data *sd,
2825 struct napi_struct *napi)
2827 list_add_tail(&napi->poll_list, &sd->poll_list);
2828 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
2833 /* One global table that all flow-based protocols share. */
2834 struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
2835 EXPORT_SYMBOL(rps_sock_flow_table);
2837 struct static_key rps_needed __read_mostly;
2839 static struct rps_dev_flow *
2840 set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
2841 struct rps_dev_flow *rflow, u16 next_cpu)
2843 if (next_cpu != RPS_NO_CPU) {
2844 #ifdef CONFIG_RFS_ACCEL
2845 struct netdev_rx_queue *rxqueue;
2846 struct rps_dev_flow_table *flow_table;
2847 struct rps_dev_flow *old_rflow;
2852 /* Should we steer this flow to a different hardware queue? */
2853 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
2854 !(dev->features & NETIF_F_NTUPLE))
2856 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
2857 if (rxq_index == skb_get_rx_queue(skb))
2860 rxqueue = dev->_rx + rxq_index;
2861 flow_table = rcu_dereference(rxqueue->rps_flow_table);
2864 flow_id = skb->rxhash & flow_table->mask;
2865 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
2866 rxq_index, flow_id);
2870 rflow = &flow_table->flows[flow_id];
2872 if (old_rflow->filter == rflow->filter)
2873 old_rflow->filter = RPS_NO_FILTER;
2877 per_cpu(softnet_data, next_cpu).input_queue_head;
2880 rflow->cpu = next_cpu;
2885 * get_rps_cpu is called from netif_receive_skb and returns the target
2886 * CPU from the RPS map of the receiving queue for a given skb.
2887 * rcu_read_lock must be held on entry.
2889 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
2890 struct rps_dev_flow **rflowp)
2892 struct netdev_rx_queue *rxqueue;
2893 struct rps_map *map;
2894 struct rps_dev_flow_table *flow_table;
2895 struct rps_sock_flow_table *sock_flow_table;
2899 if (skb_rx_queue_recorded(skb)) {
2900 u16 index = skb_get_rx_queue(skb);
2901 if (unlikely(index >= dev->real_num_rx_queues)) {
2902 WARN_ONCE(dev->real_num_rx_queues > 1,
2903 "%s received packet on queue %u, but number "
2904 "of RX queues is %u\n",
2905 dev->name, index, dev->real_num_rx_queues);
2908 rxqueue = dev->_rx + index;
2912 map = rcu_dereference(rxqueue->rps_map);
2914 if (map->len == 1 &&
2915 !rcu_access_pointer(rxqueue->rps_flow_table)) {
2916 tcpu = map->cpus[0];
2917 if (cpu_online(tcpu))
2921 } else if (!rcu_access_pointer(rxqueue->rps_flow_table)) {
2925 skb_reset_network_header(skb);
2926 if (!skb_get_rxhash(skb))
2929 flow_table = rcu_dereference(rxqueue->rps_flow_table);
2930 sock_flow_table = rcu_dereference(rps_sock_flow_table);
2931 if (flow_table && sock_flow_table) {
2933 struct rps_dev_flow *rflow;
2935 rflow = &flow_table->flows[skb->rxhash & flow_table->mask];
2938 next_cpu = sock_flow_table->ents[skb->rxhash &
2939 sock_flow_table->mask];
2942 * If the desired CPU (where last recvmsg was done) is
2943 * different from current CPU (one in the rx-queue flow
2944 * table entry), switch if one of the following holds:
2945 * - Current CPU is unset (equal to RPS_NO_CPU).
2946 * - Current CPU is offline.
2947 * - The current CPU's queue tail has advanced beyond the
2948 * last packet that was enqueued using this table entry.
2949 * This guarantees that all previous packets for the flow
2950 * have been dequeued, thus preserving in order delivery.
2952 if (unlikely(tcpu != next_cpu) &&
2953 (tcpu == RPS_NO_CPU || !cpu_online(tcpu) ||
2954 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
2955 rflow->last_qtail)) >= 0)) {
2957 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
2960 if (tcpu != RPS_NO_CPU && cpu_online(tcpu)) {
2968 tcpu = map->cpus[((u64) skb->rxhash * map->len) >> 32];
2970 if (cpu_online(tcpu)) {
2980 #ifdef CONFIG_RFS_ACCEL
2983 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
2984 * @dev: Device on which the filter was set
2985 * @rxq_index: RX queue index
2986 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
2987 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
2989 * Drivers that implement ndo_rx_flow_steer() should periodically call
2990 * this function for each installed filter and remove the filters for
2991 * which it returns %true.
2993 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
2994 u32 flow_id, u16 filter_id)
2996 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
2997 struct rps_dev_flow_table *flow_table;
2998 struct rps_dev_flow *rflow;
3003 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3004 if (flow_table && flow_id <= flow_table->mask) {
3005 rflow = &flow_table->flows[flow_id];
3006 cpu = ACCESS_ONCE(rflow->cpu);
3007 if (rflow->filter == filter_id && cpu != RPS_NO_CPU &&
3008 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
3009 rflow->last_qtail) <
3010 (int)(10 * flow_table->mask)))
3016 EXPORT_SYMBOL(rps_may_expire_flow);
3018 #endif /* CONFIG_RFS_ACCEL */
3020 /* Called from hardirq (IPI) context */
3021 static void rps_trigger_softirq(void *data)
3023 struct softnet_data *sd = data;
3025 ____napi_schedule(sd, &sd->backlog);
3029 #endif /* CONFIG_RPS */
3032 * Check if this softnet_data structure is another cpu one
3033 * If yes, queue it to our IPI list and return 1
3036 static int rps_ipi_queued(struct softnet_data *sd)
3039 struct softnet_data *mysd = &__get_cpu_var(softnet_data);
3042 sd->rps_ipi_next = mysd->rps_ipi_list;
3043 mysd->rps_ipi_list = sd;
3045 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3048 #endif /* CONFIG_RPS */
3053 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
3054 * queue (may be a remote CPU queue).
3056 static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
3057 unsigned int *qtail)
3059 struct softnet_data *sd;
3060 unsigned long flags;
3062 sd = &per_cpu(softnet_data, cpu);
3064 local_irq_save(flags);
3067 if (skb_queue_len(&sd->input_pkt_queue) <= netdev_max_backlog) {
3068 if (skb_queue_len(&sd->input_pkt_queue)) {
3070 __skb_queue_tail(&sd->input_pkt_queue, skb);
3071 input_queue_tail_incr_save(sd, qtail);
3073 local_irq_restore(flags);
3074 return NET_RX_SUCCESS;
3077 /* Schedule NAPI for backlog device
3078 * We can use non atomic operation since we own the queue lock
3080 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
3081 if (!rps_ipi_queued(sd))
3082 ____napi_schedule(sd, &sd->backlog);
3090 local_irq_restore(flags);
3092 atomic_long_inc(&skb->dev->rx_dropped);
3098 * netif_rx - post buffer to the network code
3099 * @skb: buffer to post
3101 * This function receives a packet from a device driver and queues it for
3102 * the upper (protocol) levels to process. It always succeeds. The buffer
3103 * may be dropped during processing for congestion control or by the
3107 * NET_RX_SUCCESS (no congestion)
3108 * NET_RX_DROP (packet was dropped)
3112 int netif_rx(struct sk_buff *skb)
3116 /* if netpoll wants it, pretend we never saw it */
3117 if (netpoll_rx(skb))
3120 net_timestamp_check(netdev_tstamp_prequeue, skb);
3122 trace_netif_rx(skb);
3124 if (static_key_false(&rps_needed)) {
3125 struct rps_dev_flow voidflow, *rflow = &voidflow;
3131 cpu = get_rps_cpu(skb->dev, skb, &rflow);
3133 cpu = smp_processor_id();
3135 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
3143 ret = enqueue_to_backlog(skb, get_cpu(), &qtail);
3148 EXPORT_SYMBOL(netif_rx);
3150 int netif_rx_ni(struct sk_buff *skb)
3155 err = netif_rx(skb);
3156 if (local_softirq_pending())
3162 EXPORT_SYMBOL(netif_rx_ni);
3164 static void net_tx_action(struct softirq_action *h)
3166 struct softnet_data *sd = &__get_cpu_var(softnet_data);
3168 if (sd->completion_queue) {
3169 struct sk_buff *clist;
3171 local_irq_disable();
3172 clist = sd->completion_queue;
3173 sd->completion_queue = NULL;
3177 struct sk_buff *skb = clist;
3178 clist = clist->next;
3180 WARN_ON(atomic_read(&skb->users));
3181 trace_kfree_skb(skb, net_tx_action);
3186 if (sd->output_queue) {
3189 local_irq_disable();
3190 head = sd->output_queue;
3191 sd->output_queue = NULL;
3192 sd->output_queue_tailp = &sd->output_queue;
3196 struct Qdisc *q = head;
3197 spinlock_t *root_lock;
3199 head = head->next_sched;
3201 root_lock = qdisc_lock(q);
3202 if (spin_trylock(root_lock)) {
3203 smp_mb__before_clear_bit();
3204 clear_bit(__QDISC_STATE_SCHED,
3207 spin_unlock(root_lock);
3209 if (!test_bit(__QDISC_STATE_DEACTIVATED,
3211 __netif_reschedule(q);
3213 smp_mb__before_clear_bit();
3214 clear_bit(__QDISC_STATE_SCHED,
3222 #if (defined(CONFIG_BRIDGE) || defined(CONFIG_BRIDGE_MODULE)) && \
3223 (defined(CONFIG_ATM_LANE) || defined(CONFIG_ATM_LANE_MODULE))
3224 /* This hook is defined here for ATM LANE */
3225 int (*br_fdb_test_addr_hook)(struct net_device *dev,
3226 unsigned char *addr) __read_mostly;
3227 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook);
3230 #ifdef CONFIG_NET_CLS_ACT
3231 /* TODO: Maybe we should just force sch_ingress to be compiled in
3232 * when CONFIG_NET_CLS_ACT is? otherwise some useless instructions
3233 * a compare and 2 stores extra right now if we dont have it on
3234 * but have CONFIG_NET_CLS_ACT
3235 * NOTE: This doesn't stop any functionality; if you dont have
3236 * the ingress scheduler, you just can't add policies on ingress.
3239 static int ing_filter(struct sk_buff *skb, struct netdev_queue *rxq)
3241 struct net_device *dev = skb->dev;
3242 u32 ttl = G_TC_RTTL(skb->tc_verd);
3243 int result = TC_ACT_OK;
3246 if (unlikely(MAX_RED_LOOP < ttl++)) {
3247 net_warn_ratelimited("Redir loop detected Dropping packet (%d->%d)\n",
3248 skb->skb_iif, dev->ifindex);
3252 skb->tc_verd = SET_TC_RTTL(skb->tc_verd, ttl);
3253 skb->tc_verd = SET_TC_AT(skb->tc_verd, AT_INGRESS);
3256 if (q != &noop_qdisc) {
3257 spin_lock(qdisc_lock(q));
3258 if (likely(!test_bit(__QDISC_STATE_DEACTIVATED, &q->state)))
3259 result = qdisc_enqueue_root(skb, q);
3260 spin_unlock(qdisc_lock(q));
3266 static inline struct sk_buff *handle_ing(struct sk_buff *skb,
3267 struct packet_type **pt_prev,
3268 int *ret, struct net_device *orig_dev)
3270 struct netdev_queue *rxq = rcu_dereference(skb->dev->ingress_queue);
3272 if (!rxq || rxq->qdisc == &noop_qdisc)
3276 *ret = deliver_skb(skb, *pt_prev, orig_dev);
3280 switch (ing_filter(skb, rxq)) {
3294 * netdev_rx_handler_register - register receive handler
3295 * @dev: device to register a handler for
3296 * @rx_handler: receive handler to register
3297 * @rx_handler_data: data pointer that is used by rx handler
3299 * Register a receive hander for a device. This handler will then be
3300 * called from __netif_receive_skb. A negative errno code is returned
3303 * The caller must hold the rtnl_mutex.
3305 * For a general description of rx_handler, see enum rx_handler_result.
3307 int netdev_rx_handler_register(struct net_device *dev,
3308 rx_handler_func_t *rx_handler,
3309 void *rx_handler_data)
3313 if (dev->rx_handler)
3316 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data);
3317 rcu_assign_pointer(dev->rx_handler, rx_handler);
3321 EXPORT_SYMBOL_GPL(netdev_rx_handler_register);
3324 * netdev_rx_handler_unregister - unregister receive handler
3325 * @dev: device to unregister a handler from
3327 * Unregister a receive handler from a device.
3329 * The caller must hold the rtnl_mutex.
3331 void netdev_rx_handler_unregister(struct net_device *dev)
3335 RCU_INIT_POINTER(dev->rx_handler, NULL);
3336 RCU_INIT_POINTER(dev->rx_handler_data, NULL);
3338 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister);
3341 * Limit the use of PFMEMALLOC reserves to those protocols that implement
3342 * the special handling of PFMEMALLOC skbs.
3344 static bool skb_pfmemalloc_protocol(struct sk_buff *skb)
3346 switch (skb->protocol) {
3347 case __constant_htons(ETH_P_ARP):
3348 case __constant_htons(ETH_P_IP):
3349 case __constant_htons(ETH_P_IPV6):
3350 case __constant_htons(ETH_P_8021Q):
3357 static int __netif_receive_skb_core(struct sk_buff *skb, bool pfmemalloc)
3359 struct packet_type *ptype, *pt_prev;
3360 rx_handler_func_t *rx_handler;
3361 struct net_device *orig_dev;
3362 struct net_device *null_or_dev;
3363 bool deliver_exact = false;
3364 int ret = NET_RX_DROP;
3367 net_timestamp_check(!netdev_tstamp_prequeue, skb);
3369 trace_netif_receive_skb(skb);
3371 /* if we've gotten here through NAPI, check netpoll */
3372 if (netpoll_receive_skb(skb))
3375 orig_dev = skb->dev;
3377 skb_reset_network_header(skb);
3378 if (!skb_transport_header_was_set(skb))
3379 skb_reset_transport_header(skb);
3380 skb_reset_mac_len(skb);
3387 skb->skb_iif = skb->dev->ifindex;
3389 __this_cpu_inc(softnet_data.processed);
3391 if (skb->protocol == cpu_to_be16(ETH_P_8021Q)) {
3392 skb = vlan_untag(skb);
3397 #ifdef CONFIG_NET_CLS_ACT
3398 if (skb->tc_verd & TC_NCLS) {
3399 skb->tc_verd = CLR_TC_NCLS(skb->tc_verd);
3407 list_for_each_entry_rcu(ptype, &ptype_all, list) {
3408 if (!ptype->dev || ptype->dev == skb->dev) {
3410 ret = deliver_skb(skb, pt_prev, orig_dev);
3416 #ifdef CONFIG_NET_CLS_ACT
3417 skb = handle_ing(skb, &pt_prev, &ret, orig_dev);
3423 if (pfmemalloc && !skb_pfmemalloc_protocol(skb))
3426 if (vlan_tx_tag_present(skb)) {
3428 ret = deliver_skb(skb, pt_prev, orig_dev);
3431 if (vlan_do_receive(&skb))
3433 else if (unlikely(!skb))
3437 rx_handler = rcu_dereference(skb->dev->rx_handler);
3440 ret = deliver_skb(skb, pt_prev, orig_dev);
3443 switch (rx_handler(&skb)) {
3444 case RX_HANDLER_CONSUMED:
3445 ret = NET_RX_SUCCESS;
3447 case RX_HANDLER_ANOTHER:
3449 case RX_HANDLER_EXACT:
3450 deliver_exact = true;
3451 case RX_HANDLER_PASS:
3458 if (vlan_tx_nonzero_tag_present(skb))
3459 skb->pkt_type = PACKET_OTHERHOST;
3461 /* deliver only exact match when indicated */
3462 null_or_dev = deliver_exact ? skb->dev : NULL;
3464 type = skb->protocol;
3465 list_for_each_entry_rcu(ptype,
3466 &ptype_base[ntohs(type) & PTYPE_HASH_MASK], list) {
3467 if (ptype->type == type &&
3468 (ptype->dev == null_or_dev || ptype->dev == skb->dev ||
3469 ptype->dev == orig_dev)) {
3471 ret = deliver_skb(skb, pt_prev, orig_dev);
3477 if (unlikely(skb_orphan_frags(skb, GFP_ATOMIC)))
3480 ret = pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
3483 atomic_long_inc(&skb->dev->rx_dropped);
3485 /* Jamal, now you will not able to escape explaining
3486 * me how you were going to use this. :-)
3497 static int __netif_receive_skb(struct sk_buff *skb)
3501 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) {
3502 unsigned long pflags = current->flags;
3505 * PFMEMALLOC skbs are special, they should
3506 * - be delivered to SOCK_MEMALLOC sockets only
3507 * - stay away from userspace
3508 * - have bounded memory usage
3510 * Use PF_MEMALLOC as this saves us from propagating the allocation
3511 * context down to all allocation sites.
3513 current->flags |= PF_MEMALLOC;
3514 ret = __netif_receive_skb_core(skb, true);
3515 tsk_restore_flags(current, pflags, PF_MEMALLOC);
3517 ret = __netif_receive_skb_core(skb, false);
3523 * netif_receive_skb - process receive buffer from network
3524 * @skb: buffer to process
3526 * netif_receive_skb() is the main receive data processing function.
3527 * It always succeeds. The buffer may be dropped during processing
3528 * for congestion control or by the protocol layers.
3530 * This function may only be called from softirq context and interrupts
3531 * should be enabled.
3533 * Return values (usually ignored):
3534 * NET_RX_SUCCESS: no congestion
3535 * NET_RX_DROP: packet was dropped
3537 int netif_receive_skb(struct sk_buff *skb)
3539 net_timestamp_check(netdev_tstamp_prequeue, skb);
3541 if (skb_defer_rx_timestamp(skb))
3542 return NET_RX_SUCCESS;
3545 if (static_key_false(&rps_needed)) {
3546 struct rps_dev_flow voidflow, *rflow = &voidflow;
3551 cpu = get_rps_cpu(skb->dev, skb, &rflow);
3554 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
3561 return __netif_receive_skb(skb);
3563 EXPORT_SYMBOL(netif_receive_skb);
3565 /* Network device is going away, flush any packets still pending
3566 * Called with irqs disabled.
3568 static void flush_backlog(void *arg)
3570 struct net_device *dev = arg;
3571 struct softnet_data *sd = &__get_cpu_var(softnet_data);
3572 struct sk_buff *skb, *tmp;
3575 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) {
3576 if (skb->dev == dev) {
3577 __skb_unlink(skb, &sd->input_pkt_queue);
3579 input_queue_head_incr(sd);
3584 skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
3585 if (skb->dev == dev) {
3586 __skb_unlink(skb, &sd->process_queue);
3588 input_queue_head_incr(sd);
3593 static int napi_gro_complete(struct sk_buff *skb)
3595 struct packet_offload *ptype;
3596 __be16 type = skb->protocol;
3597 struct list_head *head = &offload_base;
3600 BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
3602 if (NAPI_GRO_CB(skb)->count == 1) {
3603 skb_shinfo(skb)->gso_size = 0;
3608 list_for_each_entry_rcu(ptype, head, list) {
3609 if (ptype->type != type || !ptype->callbacks.gro_complete)
3612 err = ptype->callbacks.gro_complete(skb);
3618 WARN_ON(&ptype->list == head);
3620 return NET_RX_SUCCESS;
3624 return netif_receive_skb(skb);
3627 /* napi->gro_list contains packets ordered by age.
3628 * youngest packets at the head of it.
3629 * Complete skbs in reverse order to reduce latencies.
3631 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
3633 struct sk_buff *skb, *prev = NULL;
3635 /* scan list and build reverse chain */
3636 for (skb = napi->gro_list; skb != NULL; skb = skb->next) {
3641 for (skb = prev; skb; skb = prev) {
3644 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
3648 napi_gro_complete(skb);
3652 napi->gro_list = NULL;
3654 EXPORT_SYMBOL(napi_gro_flush);
3656 static void gro_list_prepare(struct napi_struct *napi, struct sk_buff *skb)
3659 unsigned int maclen = skb->dev->hard_header_len;
3661 for (p = napi->gro_list; p; p = p->next) {
3662 unsigned long diffs;
3664 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
3665 diffs |= p->vlan_tci ^ skb->vlan_tci;
3666 if (maclen == ETH_HLEN)
3667 diffs |= compare_ether_header(skb_mac_header(p),
3668 skb_gro_mac_header(skb));
3670 diffs = memcmp(skb_mac_header(p),
3671 skb_gro_mac_header(skb),
3673 NAPI_GRO_CB(p)->same_flow = !diffs;
3674 NAPI_GRO_CB(p)->flush = 0;
3678 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
3680 struct sk_buff **pp = NULL;
3681 struct packet_offload *ptype;
3682 __be16 type = skb->protocol;
3683 struct list_head *head = &offload_base;
3685 enum gro_result ret;
3687 if (!(skb->dev->features & NETIF_F_GRO) || netpoll_rx_on(skb))
3690 if (skb_is_gso(skb) || skb_has_frag_list(skb))
3693 gro_list_prepare(napi, skb);
3696 list_for_each_entry_rcu(ptype, head, list) {
3697 if (ptype->type != type || !ptype->callbacks.gro_receive)
3700 skb_set_network_header(skb, skb_gro_offset(skb));
3701 skb_reset_mac_len(skb);
3702 NAPI_GRO_CB(skb)->same_flow = 0;
3703 NAPI_GRO_CB(skb)->flush = 0;
3704 NAPI_GRO_CB(skb)->free = 0;
3706 pp = ptype->callbacks.gro_receive(&napi->gro_list, skb);
3711 if (&ptype->list == head)
3714 same_flow = NAPI_GRO_CB(skb)->same_flow;
3715 ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
3718 struct sk_buff *nskb = *pp;
3722 napi_gro_complete(nskb);
3729 if (NAPI_GRO_CB(skb)->flush || napi->gro_count >= MAX_GRO_SKBS)
3733 NAPI_GRO_CB(skb)->count = 1;
3734 NAPI_GRO_CB(skb)->age = jiffies;
3735 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
3736 skb->next = napi->gro_list;
3737 napi->gro_list = skb;
3741 if (skb_headlen(skb) < skb_gro_offset(skb)) {
3742 int grow = skb_gro_offset(skb) - skb_headlen(skb);
3744 BUG_ON(skb->end - skb->tail < grow);
3746 memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
3749 skb->data_len -= grow;
3751 skb_shinfo(skb)->frags[0].page_offset += grow;
3752 skb_frag_size_sub(&skb_shinfo(skb)->frags[0], grow);
3754 if (unlikely(!skb_frag_size(&skb_shinfo(skb)->frags[0]))) {
3755 skb_frag_unref(skb, 0);
3756 memmove(skb_shinfo(skb)->frags,
3757 skb_shinfo(skb)->frags + 1,
3758 --skb_shinfo(skb)->nr_frags * sizeof(skb_frag_t));
3771 static gro_result_t napi_skb_finish(gro_result_t ret, struct sk_buff *skb)
3775 if (netif_receive_skb(skb))
3783 case GRO_MERGED_FREE:
3784 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
3785 kmem_cache_free(skbuff_head_cache, skb);
3798 static void skb_gro_reset_offset(struct sk_buff *skb)
3800 const struct skb_shared_info *pinfo = skb_shinfo(skb);
3801 const skb_frag_t *frag0 = &pinfo->frags[0];
3803 NAPI_GRO_CB(skb)->data_offset = 0;
3804 NAPI_GRO_CB(skb)->frag0 = NULL;
3805 NAPI_GRO_CB(skb)->frag0_len = 0;
3807 if (skb->mac_header == skb->tail &&
3809 !PageHighMem(skb_frag_page(frag0))) {
3810 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
3811 NAPI_GRO_CB(skb)->frag0_len = skb_frag_size(frag0);
3815 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
3817 skb_gro_reset_offset(skb);
3819 return napi_skb_finish(dev_gro_receive(napi, skb), skb);
3821 EXPORT_SYMBOL(napi_gro_receive);
3823 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
3825 __skb_pull(skb, skb_headlen(skb));
3826 /* restore the reserve we had after netdev_alloc_skb_ip_align() */
3827 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
3829 skb->dev = napi->dev;
3835 struct sk_buff *napi_get_frags(struct napi_struct *napi)
3837 struct sk_buff *skb = napi->skb;
3840 skb = netdev_alloc_skb_ip_align(napi->dev, GRO_MAX_HEAD);
3846 EXPORT_SYMBOL(napi_get_frags);
3848 static gro_result_t napi_frags_finish(struct napi_struct *napi, struct sk_buff *skb,
3854 skb->protocol = eth_type_trans(skb, skb->dev);
3856 if (ret == GRO_HELD)
3857 skb_gro_pull(skb, -ETH_HLEN);
3858 else if (netif_receive_skb(skb))
3863 case GRO_MERGED_FREE:
3864 napi_reuse_skb(napi, skb);
3874 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
3876 struct sk_buff *skb = napi->skb;
3883 skb_reset_mac_header(skb);
3884 skb_gro_reset_offset(skb);
3886 off = skb_gro_offset(skb);
3887 hlen = off + sizeof(*eth);
3888 eth = skb_gro_header_fast(skb, off);
3889 if (skb_gro_header_hard(skb, hlen)) {
3890 eth = skb_gro_header_slow(skb, hlen, off);
3891 if (unlikely(!eth)) {
3892 napi_reuse_skb(napi, skb);
3898 skb_gro_pull(skb, sizeof(*eth));
3901 * This works because the only protocols we care about don't require
3902 * special handling. We'll fix it up properly at the end.
3904 skb->protocol = eth->h_proto;
3910 gro_result_t napi_gro_frags(struct napi_struct *napi)
3912 struct sk_buff *skb = napi_frags_skb(napi);
3917 return napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
3919 EXPORT_SYMBOL(napi_gro_frags);
3922 * net_rps_action sends any pending IPI's for rps.
3923 * Note: called with local irq disabled, but exits with local irq enabled.
3925 static void net_rps_action_and_irq_enable(struct softnet_data *sd)
3928 struct softnet_data *remsd = sd->rps_ipi_list;
3931 sd->rps_ipi_list = NULL;
3935 /* Send pending IPI's to kick RPS processing on remote cpus. */
3937 struct softnet_data *next = remsd->rps_ipi_next;
3939 if (cpu_online(remsd->cpu))
3940 __smp_call_function_single(remsd->cpu,
3949 static int process_backlog(struct napi_struct *napi, int quota)
3952 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
3955 /* Check if we have pending ipi, its better to send them now,
3956 * not waiting net_rx_action() end.
3958 if (sd->rps_ipi_list) {
3959 local_irq_disable();
3960 net_rps_action_and_irq_enable(sd);
3963 napi->weight = weight_p;
3964 local_irq_disable();
3965 while (work < quota) {
3966 struct sk_buff *skb;
3969 while ((skb = __skb_dequeue(&sd->process_queue))) {
3971 __netif_receive_skb(skb);
3972 local_irq_disable();
3973 input_queue_head_incr(sd);
3974 if (++work >= quota) {
3981 qlen = skb_queue_len(&sd->input_pkt_queue);
3983 skb_queue_splice_tail_init(&sd->input_pkt_queue,
3984 &sd->process_queue);
3986 if (qlen < quota - work) {
3988 * Inline a custom version of __napi_complete().
3989 * only current cpu owns and manipulates this napi,
3990 * and NAPI_STATE_SCHED is the only possible flag set on backlog.
3991 * we can use a plain write instead of clear_bit(),
3992 * and we dont need an smp_mb() memory barrier.
3994 list_del(&napi->poll_list);
3997 quota = work + qlen;
4007 * __napi_schedule - schedule for receive
4008 * @n: entry to schedule
4010 * The entry's receive function will be scheduled to run
4012 void __napi_schedule(struct napi_struct *n)
4014 unsigned long flags;
4016 local_irq_save(flags);
4017 ____napi_schedule(&__get_cpu_var(softnet_data), n);
4018 local_irq_restore(flags);
4020 EXPORT_SYMBOL(__napi_schedule);
4022 void __napi_complete(struct napi_struct *n)
4024 BUG_ON(!test_bit(NAPI_STATE_SCHED, &n->state));
4025 BUG_ON(n->gro_list);
4027 list_del(&n->poll_list);
4028 smp_mb__before_clear_bit();
4029 clear_bit(NAPI_STATE_SCHED, &n->state);
4031 EXPORT_SYMBOL(__napi_complete);
4033 void napi_complete(struct napi_struct *n)
4035 unsigned long flags;
4038 * don't let napi dequeue from the cpu poll list
4039 * just in case its running on a different cpu
4041 if (unlikely(test_bit(NAPI_STATE_NPSVC, &n->state)))
4044 napi_gro_flush(n, false);
4045 local_irq_save(flags);
4047 local_irq_restore(flags);
4049 EXPORT_SYMBOL(napi_complete);
4051 void netif_napi_add(struct net_device *dev, struct napi_struct *napi,
4052 int (*poll)(struct napi_struct *, int), int weight)
4054 INIT_LIST_HEAD(&napi->poll_list);
4055 napi->gro_count = 0;
4056 napi->gro_list = NULL;
4059 if (weight > NAPI_POLL_WEIGHT)
4060 pr_err_once("netif_napi_add() called with weight %d on device %s\n",
4062 napi->weight = weight;
4063 list_add(&napi->dev_list, &dev->napi_list);
4065 #ifdef CONFIG_NETPOLL
4066 spin_lock_init(&napi->poll_lock);
4067 napi->poll_owner = -1;
4069 set_bit(NAPI_STATE_SCHED, &napi->state);
4071 EXPORT_SYMBOL(netif_napi_add);
4073 void netif_napi_del(struct napi_struct *napi)
4075 struct sk_buff *skb, *next;
4077 list_del_init(&napi->dev_list);
4078 napi_free_frags(napi);
4080 for (skb = napi->gro_list; skb; skb = next) {
4086 napi->gro_list = NULL;
4087 napi->gro_count = 0;
4089 EXPORT_SYMBOL(netif_napi_del);
4091 static void net_rx_action(struct softirq_action *h)
4093 struct softnet_data *sd = &__get_cpu_var(softnet_data);
4094 unsigned long time_limit = jiffies + 2;
4095 int budget = netdev_budget;
4098 local_irq_disable();
4100 while (!list_empty(&sd->poll_list)) {
4101 struct napi_struct *n;
4104 /* If softirq window is exhuasted then punt.
4105 * Allow this to run for 2 jiffies since which will allow
4106 * an average latency of 1.5/HZ.
4108 if (unlikely(budget <= 0 || time_after_eq(jiffies, time_limit)))
4113 /* Even though interrupts have been re-enabled, this
4114 * access is safe because interrupts can only add new
4115 * entries to the tail of this list, and only ->poll()
4116 * calls can remove this head entry from the list.
4118 n = list_first_entry(&sd->poll_list, struct napi_struct, poll_list);
4120 have = netpoll_poll_lock(n);
4124 /* This NAPI_STATE_SCHED test is for avoiding a race
4125 * with netpoll's poll_napi(). Only the entity which
4126 * obtains the lock and sees NAPI_STATE_SCHED set will
4127 * actually make the ->poll() call. Therefore we avoid
4128 * accidentally calling ->poll() when NAPI is not scheduled.
4131 if (test_bit(NAPI_STATE_SCHED, &n->state)) {
4132 work = n->poll(n, weight);
4136 WARN_ON_ONCE(work > weight);
4140 local_irq_disable();
4142 /* Drivers must not modify the NAPI state if they
4143 * consume the entire weight. In such cases this code
4144 * still "owns" the NAPI instance and therefore can
4145 * move the instance around on the list at-will.
4147 if (unlikely(work == weight)) {
4148 if (unlikely(napi_disable_pending(n))) {
4151 local_irq_disable();
4154 /* flush too old packets
4155 * If HZ < 1000, flush all packets.
4158 napi_gro_flush(n, HZ >= 1000);
4159 local_irq_disable();
4161 list_move_tail(&n->poll_list, &sd->poll_list);
4165 netpoll_poll_unlock(have);
4168 net_rps_action_and_irq_enable(sd);
4170 #ifdef CONFIG_NET_DMA
4172 * There may not be any more sk_buffs coming right now, so push
4173 * any pending DMA copies to hardware
4175 dma_issue_pending_all();
4182 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
4186 struct netdev_upper {
4187 struct net_device *dev;
4189 struct list_head list;
4190 struct rcu_head rcu;
4191 struct list_head search_list;
4194 static void __append_search_uppers(struct list_head *search_list,
4195 struct net_device *dev)
4197 struct netdev_upper *upper;
4199 list_for_each_entry(upper, &dev->upper_dev_list, list) {
4200 /* check if this upper is not already in search list */
4201 if (list_empty(&upper->search_list))
4202 list_add_tail(&upper->search_list, search_list);
4206 static bool __netdev_search_upper_dev(struct net_device *dev,
4207 struct net_device *upper_dev)
4209 LIST_HEAD(search_list);
4210 struct netdev_upper *upper;
4211 struct netdev_upper *tmp;
4214 __append_search_uppers(&search_list, dev);
4215 list_for_each_entry(upper, &search_list, search_list) {
4216 if (upper->dev == upper_dev) {
4220 __append_search_uppers(&search_list, upper->dev);
4222 list_for_each_entry_safe(upper, tmp, &search_list, search_list)
4223 INIT_LIST_HEAD(&upper->search_list);
4227 static struct netdev_upper *__netdev_find_upper(struct net_device *dev,
4228 struct net_device *upper_dev)
4230 struct netdev_upper *upper;
4232 list_for_each_entry(upper, &dev->upper_dev_list, list) {
4233 if (upper->dev == upper_dev)
4240 * netdev_has_upper_dev - Check if device is linked to an upper device
4242 * @upper_dev: upper device to check
4244 * Find out if a device is linked to specified upper device and return true
4245 * in case it is. Note that this checks only immediate upper device,
4246 * not through a complete stack of devices. The caller must hold the RTNL lock.
4248 bool netdev_has_upper_dev(struct net_device *dev,
4249 struct net_device *upper_dev)
4253 return __netdev_find_upper(dev, upper_dev);
4255 EXPORT_SYMBOL(netdev_has_upper_dev);
4258 * netdev_has_any_upper_dev - Check if device is linked to some device
4261 * Find out if a device is linked to an upper device and return true in case
4262 * it is. The caller must hold the RTNL lock.
4264 bool netdev_has_any_upper_dev(struct net_device *dev)
4268 return !list_empty(&dev->upper_dev_list);
4270 EXPORT_SYMBOL(netdev_has_any_upper_dev);
4273 * netdev_master_upper_dev_get - Get master upper device
4276 * Find a master upper device and return pointer to it or NULL in case
4277 * it's not there. The caller must hold the RTNL lock.
4279 struct net_device *netdev_master_upper_dev_get(struct net_device *dev)
4281 struct netdev_upper *upper;
4285 if (list_empty(&dev->upper_dev_list))
4288 upper = list_first_entry(&dev->upper_dev_list,
4289 struct netdev_upper, list);
4290 if (likely(upper->master))
4294 EXPORT_SYMBOL(netdev_master_upper_dev_get);
4297 * netdev_master_upper_dev_get_rcu - Get master upper device
4300 * Find a master upper device and return pointer to it or NULL in case
4301 * it's not there. The caller must hold the RCU read lock.
4303 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev)
4305 struct netdev_upper *upper;
4307 upper = list_first_or_null_rcu(&dev->upper_dev_list,
4308 struct netdev_upper, list);
4309 if (upper && likely(upper->master))
4313 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu);
4315 static int __netdev_upper_dev_link(struct net_device *dev,
4316 struct net_device *upper_dev, bool master)
4318 struct netdev_upper *upper;
4322 if (dev == upper_dev)
4325 /* To prevent loops, check if dev is not upper device to upper_dev. */
4326 if (__netdev_search_upper_dev(upper_dev, dev))
4329 if (__netdev_find_upper(dev, upper_dev))
4332 if (master && netdev_master_upper_dev_get(dev))
4335 upper = kmalloc(sizeof(*upper), GFP_KERNEL);
4339 upper->dev = upper_dev;
4340 upper->master = master;
4341 INIT_LIST_HEAD(&upper->search_list);
4343 /* Ensure that master upper link is always the first item in list. */
4345 list_add_rcu(&upper->list, &dev->upper_dev_list);
4347 list_add_tail_rcu(&upper->list, &dev->upper_dev_list);
4348 dev_hold(upper_dev);
4354 * netdev_upper_dev_link - Add a link to the upper device
4356 * @upper_dev: new upper device
4358 * Adds a link to device which is upper to this one. The caller must hold
4359 * the RTNL lock. On a failure a negative errno code is returned.
4360 * On success the reference counts are adjusted and the function
4363 int netdev_upper_dev_link(struct net_device *dev,
4364 struct net_device *upper_dev)
4366 return __netdev_upper_dev_link(dev, upper_dev, false);
4368 EXPORT_SYMBOL(netdev_upper_dev_link);
4371 * netdev_master_upper_dev_link - Add a master link to the upper device
4373 * @upper_dev: new upper device
4375 * Adds a link to device which is upper to this one. In this case, only
4376 * one master upper device can be linked, although other non-master devices
4377 * might be linked as well. The caller must hold the RTNL lock.
4378 * On a failure a negative errno code is returned. On success the reference
4379 * counts are adjusted and the function returns zero.
4381 int netdev_master_upper_dev_link(struct net_device *dev,
4382 struct net_device *upper_dev)
4384 return __netdev_upper_dev_link(dev, upper_dev, true);
4386 EXPORT_SYMBOL(netdev_master_upper_dev_link);
4389 * netdev_upper_dev_unlink - Removes a link to upper device
4391 * @upper_dev: new upper device
4393 * Removes a link to device which is upper to this one. The caller must hold
4396 void netdev_upper_dev_unlink(struct net_device *dev,
4397 struct net_device *upper_dev)
4399 struct netdev_upper *upper;
4403 upper = __netdev_find_upper(dev, upper_dev);
4406 list_del_rcu(&upper->list);
4408 kfree_rcu(upper, rcu);
4410 EXPORT_SYMBOL(netdev_upper_dev_unlink);
4412 static void dev_change_rx_flags(struct net_device *dev, int flags)
4414 const struct net_device_ops *ops = dev->netdev_ops;
4416 if ((dev->flags & IFF_UP) && ops->ndo_change_rx_flags)
4417 ops->ndo_change_rx_flags(dev, flags);
4420 static int __dev_set_promiscuity(struct net_device *dev, int inc)
4422 unsigned int old_flags = dev->flags;
4428 dev->flags |= IFF_PROMISC;
4429 dev->promiscuity += inc;
4430 if (dev->promiscuity == 0) {
4433 * If inc causes overflow, untouch promisc and return error.
4436 dev->flags &= ~IFF_PROMISC;
4438 dev->promiscuity -= inc;
4439 pr_warn("%s: promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n",
4444 if (dev->flags != old_flags) {
4445 pr_info("device %s %s promiscuous mode\n",
4447 dev->flags & IFF_PROMISC ? "entered" : "left");
4448 if (audit_enabled) {
4449 current_uid_gid(&uid, &gid);
4450 audit_log(current->audit_context, GFP_ATOMIC,
4451 AUDIT_ANOM_PROMISCUOUS,
4452 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u",
4453 dev->name, (dev->flags & IFF_PROMISC),
4454 (old_flags & IFF_PROMISC),
4455 from_kuid(&init_user_ns, audit_get_loginuid(current)),
4456 from_kuid(&init_user_ns, uid),
4457 from_kgid(&init_user_ns, gid),
4458 audit_get_sessionid(current));
4461 dev_change_rx_flags(dev, IFF_PROMISC);
4467 * dev_set_promiscuity - update promiscuity count on a device
4471 * Add or remove promiscuity from a device. While the count in the device
4472 * remains above zero the interface remains promiscuous. Once it hits zero
4473 * the device reverts back to normal filtering operation. A negative inc
4474 * value is used to drop promiscuity on the device.
4475 * Return 0 if successful or a negative errno code on error.
4477 int dev_set_promiscuity(struct net_device *dev, int inc)
4479 unsigned int old_flags = dev->flags;
4482 err = __dev_set_promiscuity(dev, inc);
4485 if (dev->flags != old_flags)
4486 dev_set_rx_mode(dev);
4489 EXPORT_SYMBOL(dev_set_promiscuity);
4492 * dev_set_allmulti - update allmulti count on a device
4496 * Add or remove reception of all multicast frames to a device. While the
4497 * count in the device remains above zero the interface remains listening
4498 * to all interfaces. Once it hits zero the device reverts back to normal
4499 * filtering operation. A negative @inc value is used to drop the counter
4500 * when releasing a resource needing all multicasts.
4501 * Return 0 if successful or a negative errno code on error.
4504 int dev_set_allmulti(struct net_device *dev, int inc)
4506 unsigned int old_flags = dev->flags;
4510 dev->flags |= IFF_ALLMULTI;
4511 dev->allmulti += inc;
4512 if (dev->allmulti == 0) {
4515 * If inc causes overflow, untouch allmulti and return error.
4518 dev->flags &= ~IFF_ALLMULTI;
4520 dev->allmulti -= inc;
4521 pr_warn("%s: allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n",
4526 if (dev->flags ^ old_flags) {
4527 dev_change_rx_flags(dev, IFF_ALLMULTI);
4528 dev_set_rx_mode(dev);
4532 EXPORT_SYMBOL(dev_set_allmulti);
4535 * Upload unicast and multicast address lists to device and
4536 * configure RX filtering. When the device doesn't support unicast
4537 * filtering it is put in promiscuous mode while unicast addresses
4540 void __dev_set_rx_mode(struct net_device *dev)
4542 const struct net_device_ops *ops = dev->netdev_ops;
4544 /* dev_open will call this function so the list will stay sane. */
4545 if (!(dev->flags&IFF_UP))
4548 if (!netif_device_present(dev))
4551 if (!(dev->priv_flags & IFF_UNICAST_FLT)) {
4552 /* Unicast addresses changes may only happen under the rtnl,
4553 * therefore calling __dev_set_promiscuity here is safe.
4555 if (!netdev_uc_empty(dev) && !dev->uc_promisc) {
4556 __dev_set_promiscuity(dev, 1);
4557 dev->uc_promisc = true;
4558 } else if (netdev_uc_empty(dev) && dev->uc_promisc) {
4559 __dev_set_promiscuity(dev, -1);
4560 dev->uc_promisc = false;
4564 if (ops->ndo_set_rx_mode)
4565 ops->ndo_set_rx_mode(dev);
4568 void dev_set_rx_mode(struct net_device *dev)
4570 netif_addr_lock_bh(dev);
4571 __dev_set_rx_mode(dev);
4572 netif_addr_unlock_bh(dev);
4576 * dev_get_flags - get flags reported to userspace
4579 * Get the combination of flag bits exported through APIs to userspace.
4581 unsigned int dev_get_flags(const struct net_device *dev)
4585 flags = (dev->flags & ~(IFF_PROMISC |
4590 (dev->gflags & (IFF_PROMISC |
4593 if (netif_running(dev)) {
4594 if (netif_oper_up(dev))
4595 flags |= IFF_RUNNING;
4596 if (netif_carrier_ok(dev))
4597 flags |= IFF_LOWER_UP;
4598 if (netif_dormant(dev))
4599 flags |= IFF_DORMANT;
4604 EXPORT_SYMBOL(dev_get_flags);
4606 int __dev_change_flags(struct net_device *dev, unsigned int flags)
4608 unsigned int old_flags = dev->flags;
4614 * Set the flags on our device.
4617 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP |
4618 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL |
4620 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC |
4624 * Load in the correct multicast list now the flags have changed.
4627 if ((old_flags ^ flags) & IFF_MULTICAST)
4628 dev_change_rx_flags(dev, IFF_MULTICAST);
4630 dev_set_rx_mode(dev);
4633 * Have we downed the interface. We handle IFF_UP ourselves
4634 * according to user attempts to set it, rather than blindly
4639 if ((old_flags ^ flags) & IFF_UP) { /* Bit is different ? */
4640 ret = ((old_flags & IFF_UP) ? __dev_close : __dev_open)(dev);
4643 dev_set_rx_mode(dev);
4646 if ((flags ^ dev->gflags) & IFF_PROMISC) {
4647 int inc = (flags & IFF_PROMISC) ? 1 : -1;
4649 dev->gflags ^= IFF_PROMISC;
4650 dev_set_promiscuity(dev, inc);
4653 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI
4654 is important. Some (broken) drivers set IFF_PROMISC, when
4655 IFF_ALLMULTI is requested not asking us and not reporting.
4657 if ((flags ^ dev->gflags) & IFF_ALLMULTI) {
4658 int inc = (flags & IFF_ALLMULTI) ? 1 : -1;
4660 dev->gflags ^= IFF_ALLMULTI;
4661 dev_set_allmulti(dev, inc);
4667 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags)
4669 unsigned int changes = dev->flags ^ old_flags;
4671 if (changes & IFF_UP) {
4672 if (dev->flags & IFF_UP)
4673 call_netdevice_notifiers(NETDEV_UP, dev);
4675 call_netdevice_notifiers(NETDEV_DOWN, dev);
4678 if (dev->flags & IFF_UP &&
4679 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE)))
4680 call_netdevice_notifiers(NETDEV_CHANGE, dev);
4684 * dev_change_flags - change device settings
4686 * @flags: device state flags
4688 * Change settings on device based state flags. The flags are
4689 * in the userspace exported format.
4691 int dev_change_flags(struct net_device *dev, unsigned int flags)
4694 unsigned int changes, old_flags = dev->flags;
4696 ret = __dev_change_flags(dev, flags);
4700 changes = old_flags ^ dev->flags;
4702 rtmsg_ifinfo(RTM_NEWLINK, dev, changes);
4704 __dev_notify_flags(dev, old_flags);
4707 EXPORT_SYMBOL(dev_change_flags);
4710 * dev_set_mtu - Change maximum transfer unit
4712 * @new_mtu: new transfer unit
4714 * Change the maximum transfer size of the network device.
4716 int dev_set_mtu(struct net_device *dev, int new_mtu)
4718 const struct net_device_ops *ops = dev->netdev_ops;
4721 if (new_mtu == dev->mtu)
4724 /* MTU must be positive. */
4728 if (!netif_device_present(dev))
4732 if (ops->ndo_change_mtu)
4733 err = ops->ndo_change_mtu(dev, new_mtu);
4738 call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
4741 EXPORT_SYMBOL(dev_set_mtu);
4744 * dev_set_group - Change group this device belongs to
4746 * @new_group: group this device should belong to
4748 void dev_set_group(struct net_device *dev, int new_group)
4750 dev->group = new_group;
4752 EXPORT_SYMBOL(dev_set_group);
4755 * dev_set_mac_address - Change Media Access Control Address
4759 * Change the hardware (MAC) address of the device
4761 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa)
4763 const struct net_device_ops *ops = dev->netdev_ops;
4766 if (!ops->ndo_set_mac_address)
4768 if (sa->sa_family != dev->type)
4770 if (!netif_device_present(dev))
4772 err = ops->ndo_set_mac_address(dev, sa);
4775 dev->addr_assign_type = NET_ADDR_SET;
4776 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
4777 add_device_randomness(dev->dev_addr, dev->addr_len);
4780 EXPORT_SYMBOL(dev_set_mac_address);
4783 * dev_change_carrier - Change device carrier
4785 * @new_carrier: new value
4787 * Change device carrier
4789 int dev_change_carrier(struct net_device *dev, bool new_carrier)
4791 const struct net_device_ops *ops = dev->netdev_ops;
4793 if (!ops->ndo_change_carrier)
4795 if (!netif_device_present(dev))
4797 return ops->ndo_change_carrier(dev, new_carrier);
4799 EXPORT_SYMBOL(dev_change_carrier);
4802 * dev_new_index - allocate an ifindex
4803 * @net: the applicable net namespace
4805 * Returns a suitable unique value for a new device interface
4806 * number. The caller must hold the rtnl semaphore or the
4807 * dev_base_lock to be sure it remains unique.
4809 static int dev_new_index(struct net *net)
4811 int ifindex = net->ifindex;
4815 if (!__dev_get_by_index(net, ifindex))
4816 return net->ifindex = ifindex;
4820 /* Delayed registration/unregisteration */
4821 static LIST_HEAD(net_todo_list);
4823 static void net_set_todo(struct net_device *dev)
4825 list_add_tail(&dev->todo_list, &net_todo_list);
4828 static void rollback_registered_many(struct list_head *head)
4830 struct net_device *dev, *tmp;
4832 BUG_ON(dev_boot_phase);
4835 list_for_each_entry_safe(dev, tmp, head, unreg_list) {
4836 /* Some devices call without registering
4837 * for initialization unwind. Remove those
4838 * devices and proceed with the remaining.
4840 if (dev->reg_state == NETREG_UNINITIALIZED) {
4841 pr_debug("unregister_netdevice: device %s/%p never was registered\n",
4845 list_del(&dev->unreg_list);
4848 dev->dismantle = true;
4849 BUG_ON(dev->reg_state != NETREG_REGISTERED);
4852 /* If device is running, close it first. */
4853 dev_close_many(head);
4855 list_for_each_entry(dev, head, unreg_list) {
4856 /* And unlink it from device chain. */
4857 unlist_netdevice(dev);
4859 dev->reg_state = NETREG_UNREGISTERING;
4864 list_for_each_entry(dev, head, unreg_list) {
4865 /* Shutdown queueing discipline. */
4869 /* Notify protocols, that we are about to destroy
4870 this device. They should clean all the things.
4872 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
4874 if (!dev->rtnl_link_ops ||
4875 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
4876 rtmsg_ifinfo(RTM_DELLINK, dev, ~0U);
4879 * Flush the unicast and multicast chains
4884 if (dev->netdev_ops->ndo_uninit)
4885 dev->netdev_ops->ndo_uninit(dev);
4887 /* Notifier chain MUST detach us all upper devices. */
4888 WARN_ON(netdev_has_any_upper_dev(dev));
4890 /* Remove entries from kobject tree */
4891 netdev_unregister_kobject(dev);
4893 /* Remove XPS queueing entries */
4894 netif_reset_xps_queues_gt(dev, 0);
4900 list_for_each_entry(dev, head, unreg_list)
4904 static void rollback_registered(struct net_device *dev)
4908 list_add(&dev->unreg_list, &single);
4909 rollback_registered_many(&single);
4913 static netdev_features_t netdev_fix_features(struct net_device *dev,
4914 netdev_features_t features)
4916 /* Fix illegal checksum combinations */
4917 if ((features & NETIF_F_HW_CSUM) &&
4918 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
4919 netdev_warn(dev, "mixed HW and IP checksum settings.\n");
4920 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
4923 /* TSO requires that SG is present as well. */
4924 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) {
4925 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n");
4926 features &= ~NETIF_F_ALL_TSO;
4929 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) &&
4930 !(features & NETIF_F_IP_CSUM)) {
4931 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n");
4932 features &= ~NETIF_F_TSO;
4933 features &= ~NETIF_F_TSO_ECN;
4936 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) &&
4937 !(features & NETIF_F_IPV6_CSUM)) {
4938 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n");
4939 features &= ~NETIF_F_TSO6;
4942 /* TSO ECN requires that TSO is present as well. */
4943 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN)
4944 features &= ~NETIF_F_TSO_ECN;
4946 /* Software GSO depends on SG. */
4947 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) {
4948 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n");
4949 features &= ~NETIF_F_GSO;
4952 /* UFO needs SG and checksumming */
4953 if (features & NETIF_F_UFO) {
4954 /* maybe split UFO into V4 and V6? */
4955 if (!((features & NETIF_F_GEN_CSUM) ||
4956 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))
4957 == (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
4959 "Dropping NETIF_F_UFO since no checksum offload features.\n");
4960 features &= ~NETIF_F_UFO;
4963 if (!(features & NETIF_F_SG)) {
4965 "Dropping NETIF_F_UFO since no NETIF_F_SG feature.\n");
4966 features &= ~NETIF_F_UFO;
4973 int __netdev_update_features(struct net_device *dev)
4975 netdev_features_t features;
4980 features = netdev_get_wanted_features(dev);
4982 if (dev->netdev_ops->ndo_fix_features)
4983 features = dev->netdev_ops->ndo_fix_features(dev, features);
4985 /* driver might be less strict about feature dependencies */
4986 features = netdev_fix_features(dev, features);
4988 if (dev->features == features)
4991 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n",
4992 &dev->features, &features);
4994 if (dev->netdev_ops->ndo_set_features)
4995 err = dev->netdev_ops->ndo_set_features(dev, features);
4997 if (unlikely(err < 0)) {
4999 "set_features() failed (%d); wanted %pNF, left %pNF\n",
5000 err, &features, &dev->features);
5005 dev->features = features;
5011 * netdev_update_features - recalculate device features
5012 * @dev: the device to check
5014 * Recalculate dev->features set and send notifications if it
5015 * has changed. Should be called after driver or hardware dependent
5016 * conditions might have changed that influence the features.
5018 void netdev_update_features(struct net_device *dev)
5020 if (__netdev_update_features(dev))
5021 netdev_features_change(dev);
5023 EXPORT_SYMBOL(netdev_update_features);
5026 * netdev_change_features - recalculate device features
5027 * @dev: the device to check
5029 * Recalculate dev->features set and send notifications even
5030 * if they have not changed. Should be called instead of
5031 * netdev_update_features() if also dev->vlan_features might
5032 * have changed to allow the changes to be propagated to stacked
5035 void netdev_change_features(struct net_device *dev)
5037 __netdev_update_features(dev);
5038 netdev_features_change(dev);
5040 EXPORT_SYMBOL(netdev_change_features);
5043 * netif_stacked_transfer_operstate - transfer operstate
5044 * @rootdev: the root or lower level device to transfer state from
5045 * @dev: the device to transfer operstate to
5047 * Transfer operational state from root to device. This is normally
5048 * called when a stacking relationship exists between the root
5049 * device and the device(a leaf device).
5051 void netif_stacked_transfer_operstate(const struct net_device *rootdev,
5052 struct net_device *dev)
5054 if (rootdev->operstate == IF_OPER_DORMANT)
5055 netif_dormant_on(dev);
5057 netif_dormant_off(dev);
5059 if (netif_carrier_ok(rootdev)) {
5060 if (!netif_carrier_ok(dev))
5061 netif_carrier_on(dev);
5063 if (netif_carrier_ok(dev))
5064 netif_carrier_off(dev);
5067 EXPORT_SYMBOL(netif_stacked_transfer_operstate);
5070 static int netif_alloc_rx_queues(struct net_device *dev)
5072 unsigned int i, count = dev->num_rx_queues;
5073 struct netdev_rx_queue *rx;
5077 rx = kcalloc(count, sizeof(struct netdev_rx_queue), GFP_KERNEL);
5083 for (i = 0; i < count; i++)
5089 static void netdev_init_one_queue(struct net_device *dev,
5090 struct netdev_queue *queue, void *_unused)
5092 /* Initialize queue lock */
5093 spin_lock_init(&queue->_xmit_lock);
5094 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type);
5095 queue->xmit_lock_owner = -1;
5096 netdev_queue_numa_node_write(queue, NUMA_NO_NODE);
5099 dql_init(&queue->dql, HZ);
5103 static int netif_alloc_netdev_queues(struct net_device *dev)
5105 unsigned int count = dev->num_tx_queues;
5106 struct netdev_queue *tx;
5110 tx = kcalloc(count, sizeof(struct netdev_queue), GFP_KERNEL);
5116 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL);
5117 spin_lock_init(&dev->tx_global_lock);
5123 * register_netdevice - register a network device
5124 * @dev: device to register
5126 * Take a completed network device structure and add it to the kernel
5127 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
5128 * chain. 0 is returned on success. A negative errno code is returned
5129 * on a failure to set up the device, or if the name is a duplicate.
5131 * Callers must hold the rtnl semaphore. You may want
5132 * register_netdev() instead of this.
5135 * The locking appears insufficient to guarantee two parallel registers
5136 * will not get the same name.
5139 int register_netdevice(struct net_device *dev)
5142 struct net *net = dev_net(dev);
5144 BUG_ON(dev_boot_phase);
5149 /* When net_device's are persistent, this will be fatal. */
5150 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
5153 spin_lock_init(&dev->addr_list_lock);
5154 netdev_set_addr_lockdep_class(dev);
5158 ret = dev_get_valid_name(net, dev, dev->name);
5162 /* Init, if this function is available */
5163 if (dev->netdev_ops->ndo_init) {
5164 ret = dev->netdev_ops->ndo_init(dev);
5172 if (((dev->hw_features | dev->features) & NETIF_F_HW_VLAN_FILTER) &&
5173 (!dev->netdev_ops->ndo_vlan_rx_add_vid ||
5174 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) {
5175 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n");
5182 dev->ifindex = dev_new_index(net);
5183 else if (__dev_get_by_index(net, dev->ifindex))
5186 if (dev->iflink == -1)
5187 dev->iflink = dev->ifindex;
5189 /* Transfer changeable features to wanted_features and enable
5190 * software offloads (GSO and GRO).
5192 dev->hw_features |= NETIF_F_SOFT_FEATURES;
5193 dev->features |= NETIF_F_SOFT_FEATURES;
5194 dev->wanted_features = dev->features & dev->hw_features;
5196 /* Turn on no cache copy if HW is doing checksum */
5197 if (!(dev->flags & IFF_LOOPBACK)) {
5198 dev->hw_features |= NETIF_F_NOCACHE_COPY;
5199 if (dev->features & NETIF_F_ALL_CSUM) {
5200 dev->wanted_features |= NETIF_F_NOCACHE_COPY;
5201 dev->features |= NETIF_F_NOCACHE_COPY;
5205 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices.
5207 dev->vlan_features |= NETIF_F_HIGHDMA;
5209 /* Make NETIF_F_SG inheritable to tunnel devices.
5211 dev->hw_enc_features |= NETIF_F_SG;
5213 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev);
5214 ret = notifier_to_errno(ret);
5218 ret = netdev_register_kobject(dev);
5221 dev->reg_state = NETREG_REGISTERED;
5223 __netdev_update_features(dev);
5226 * Default initial state at registry is that the
5227 * device is present.
5230 set_bit(__LINK_STATE_PRESENT, &dev->state);
5232 linkwatch_init_dev(dev);
5234 dev_init_scheduler(dev);
5236 list_netdevice(dev);
5237 add_device_randomness(dev->dev_addr, dev->addr_len);
5239 /* If the device has permanent device address, driver should
5240 * set dev_addr and also addr_assign_type should be set to
5241 * NET_ADDR_PERM (default value).
5243 if (dev->addr_assign_type == NET_ADDR_PERM)
5244 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len);
5246 /* Notify protocols, that a new device appeared. */
5247 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
5248 ret = notifier_to_errno(ret);
5250 rollback_registered(dev);
5251 dev->reg_state = NETREG_UNREGISTERED;
5254 * Prevent userspace races by waiting until the network
5255 * device is fully setup before sending notifications.
5257 if (!dev->rtnl_link_ops ||
5258 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
5259 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U);
5265 if (dev->netdev_ops->ndo_uninit)
5266 dev->netdev_ops->ndo_uninit(dev);
5269 EXPORT_SYMBOL(register_netdevice);
5272 * init_dummy_netdev - init a dummy network device for NAPI
5273 * @dev: device to init
5275 * This takes a network device structure and initialize the minimum
5276 * amount of fields so it can be used to schedule NAPI polls without
5277 * registering a full blown interface. This is to be used by drivers
5278 * that need to tie several hardware interfaces to a single NAPI
5279 * poll scheduler due to HW limitations.
5281 int init_dummy_netdev(struct net_device *dev)
5283 /* Clear everything. Note we don't initialize spinlocks
5284 * are they aren't supposed to be taken by any of the
5285 * NAPI code and this dummy netdev is supposed to be
5286 * only ever used for NAPI polls
5288 memset(dev, 0, sizeof(struct net_device));
5290 /* make sure we BUG if trying to hit standard
5291 * register/unregister code path
5293 dev->reg_state = NETREG_DUMMY;
5295 /* NAPI wants this */
5296 INIT_LIST_HEAD(&dev->napi_list);
5298 /* a dummy interface is started by default */
5299 set_bit(__LINK_STATE_PRESENT, &dev->state);
5300 set_bit(__LINK_STATE_START, &dev->state);
5302 /* Note : We dont allocate pcpu_refcnt for dummy devices,
5303 * because users of this 'device' dont need to change
5309 EXPORT_SYMBOL_GPL(init_dummy_netdev);
5313 * register_netdev - register a network device
5314 * @dev: device to register
5316 * Take a completed network device structure and add it to the kernel
5317 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
5318 * chain. 0 is returned on success. A negative errno code is returned
5319 * on a failure to set up the device, or if the name is a duplicate.
5321 * This is a wrapper around register_netdevice that takes the rtnl semaphore
5322 * and expands the device name if you passed a format string to
5325 int register_netdev(struct net_device *dev)
5330 err = register_netdevice(dev);
5334 EXPORT_SYMBOL(register_netdev);
5336 int netdev_refcnt_read(const struct net_device *dev)
5340 for_each_possible_cpu(i)
5341 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
5344 EXPORT_SYMBOL(netdev_refcnt_read);
5347 * netdev_wait_allrefs - wait until all references are gone.
5348 * @dev: target net_device
5350 * This is called when unregistering network devices.
5352 * Any protocol or device that holds a reference should register
5353 * for netdevice notification, and cleanup and put back the
5354 * reference if they receive an UNREGISTER event.
5355 * We can get stuck here if buggy protocols don't correctly
5358 static void netdev_wait_allrefs(struct net_device *dev)
5360 unsigned long rebroadcast_time, warning_time;
5363 linkwatch_forget_dev(dev);
5365 rebroadcast_time = warning_time = jiffies;
5366 refcnt = netdev_refcnt_read(dev);
5368 while (refcnt != 0) {
5369 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
5372 /* Rebroadcast unregister notification */
5373 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
5379 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
5380 if (test_bit(__LINK_STATE_LINKWATCH_PENDING,
5382 /* We must not have linkwatch events
5383 * pending on unregister. If this
5384 * happens, we simply run the queue
5385 * unscheduled, resulting in a noop
5388 linkwatch_run_queue();
5393 rebroadcast_time = jiffies;
5398 refcnt = netdev_refcnt_read(dev);
5400 if (time_after(jiffies, warning_time + 10 * HZ)) {
5401 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n",
5403 warning_time = jiffies;
5412 * register_netdevice(x1);
5413 * register_netdevice(x2);
5415 * unregister_netdevice(y1);
5416 * unregister_netdevice(y2);
5422 * We are invoked by rtnl_unlock().
5423 * This allows us to deal with problems:
5424 * 1) We can delete sysfs objects which invoke hotplug
5425 * without deadlocking with linkwatch via keventd.
5426 * 2) Since we run with the RTNL semaphore not held, we can sleep
5427 * safely in order to wait for the netdev refcnt to drop to zero.
5429 * We must not return until all unregister events added during
5430 * the interval the lock was held have been completed.
5432 void netdev_run_todo(void)
5434 struct list_head list;
5436 /* Snapshot list, allow later requests */
5437 list_replace_init(&net_todo_list, &list);
5442 /* Wait for rcu callbacks to finish before next phase */
5443 if (!list_empty(&list))
5446 while (!list_empty(&list)) {
5447 struct net_device *dev
5448 = list_first_entry(&list, struct net_device, todo_list);
5449 list_del(&dev->todo_list);
5452 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
5455 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) {
5456 pr_err("network todo '%s' but state %d\n",
5457 dev->name, dev->reg_state);
5462 dev->reg_state = NETREG_UNREGISTERED;
5464 on_each_cpu(flush_backlog, dev, 1);
5466 netdev_wait_allrefs(dev);
5469 BUG_ON(netdev_refcnt_read(dev));
5470 WARN_ON(rcu_access_pointer(dev->ip_ptr));
5471 WARN_ON(rcu_access_pointer(dev->ip6_ptr));
5472 WARN_ON(dev->dn_ptr);
5474 if (dev->destructor)
5475 dev->destructor(dev);
5477 /* Free network device */
5478 kobject_put(&dev->dev.kobj);
5482 /* Convert net_device_stats to rtnl_link_stats64. They have the same
5483 * fields in the same order, with only the type differing.
5485 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
5486 const struct net_device_stats *netdev_stats)
5488 #if BITS_PER_LONG == 64
5489 BUILD_BUG_ON(sizeof(*stats64) != sizeof(*netdev_stats));
5490 memcpy(stats64, netdev_stats, sizeof(*stats64));
5492 size_t i, n = sizeof(*stats64) / sizeof(u64);
5493 const unsigned long *src = (const unsigned long *)netdev_stats;
5494 u64 *dst = (u64 *)stats64;
5496 BUILD_BUG_ON(sizeof(*netdev_stats) / sizeof(unsigned long) !=
5497 sizeof(*stats64) / sizeof(u64));
5498 for (i = 0; i < n; i++)
5502 EXPORT_SYMBOL(netdev_stats_to_stats64);
5505 * dev_get_stats - get network device statistics
5506 * @dev: device to get statistics from
5507 * @storage: place to store stats
5509 * Get network statistics from device. Return @storage.
5510 * The device driver may provide its own method by setting
5511 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats;
5512 * otherwise the internal statistics structure is used.
5514 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev,
5515 struct rtnl_link_stats64 *storage)
5517 const struct net_device_ops *ops = dev->netdev_ops;
5519 if (ops->ndo_get_stats64) {
5520 memset(storage, 0, sizeof(*storage));
5521 ops->ndo_get_stats64(dev, storage);
5522 } else if (ops->ndo_get_stats) {
5523 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev));
5525 netdev_stats_to_stats64(storage, &dev->stats);
5527 storage->rx_dropped += atomic_long_read(&dev->rx_dropped);
5530 EXPORT_SYMBOL(dev_get_stats);
5532 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev)
5534 struct netdev_queue *queue = dev_ingress_queue(dev);
5536 #ifdef CONFIG_NET_CLS_ACT
5539 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
5542 netdev_init_one_queue(dev, queue, NULL);
5543 queue->qdisc = &noop_qdisc;
5544 queue->qdisc_sleeping = &noop_qdisc;
5545 rcu_assign_pointer(dev->ingress_queue, queue);
5550 static const struct ethtool_ops default_ethtool_ops;
5552 void netdev_set_default_ethtool_ops(struct net_device *dev,
5553 const struct ethtool_ops *ops)
5555 if (dev->ethtool_ops == &default_ethtool_ops)
5556 dev->ethtool_ops = ops;
5558 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops);
5561 * alloc_netdev_mqs - allocate network device
5562 * @sizeof_priv: size of private data to allocate space for
5563 * @name: device name format string
5564 * @setup: callback to initialize device
5565 * @txqs: the number of TX subqueues to allocate
5566 * @rxqs: the number of RX subqueues to allocate
5568 * Allocates a struct net_device with private data area for driver use
5569 * and performs basic initialization. Also allocates subquue structs
5570 * for each queue on the device.
5572 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
5573 void (*setup)(struct net_device *),
5574 unsigned int txqs, unsigned int rxqs)
5576 struct net_device *dev;
5578 struct net_device *p;
5580 BUG_ON(strlen(name) >= sizeof(dev->name));
5583 pr_err("alloc_netdev: Unable to allocate device with zero queues\n");
5589 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n");
5594 alloc_size = sizeof(struct net_device);
5596 /* ensure 32-byte alignment of private area */
5597 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
5598 alloc_size += sizeof_priv;
5600 /* ensure 32-byte alignment of whole construct */
5601 alloc_size += NETDEV_ALIGN - 1;
5603 p = kzalloc(alloc_size, GFP_KERNEL);
5607 dev = PTR_ALIGN(p, NETDEV_ALIGN);
5608 dev->padded = (char *)dev - (char *)p;
5610 dev->pcpu_refcnt = alloc_percpu(int);
5611 if (!dev->pcpu_refcnt)
5614 if (dev_addr_init(dev))
5620 dev_net_set(dev, &init_net);
5622 dev->gso_max_size = GSO_MAX_SIZE;
5623 dev->gso_max_segs = GSO_MAX_SEGS;
5625 INIT_LIST_HEAD(&dev->napi_list);
5626 INIT_LIST_HEAD(&dev->unreg_list);
5627 INIT_LIST_HEAD(&dev->link_watch_list);
5628 INIT_LIST_HEAD(&dev->upper_dev_list);
5629 dev->priv_flags = IFF_XMIT_DST_RELEASE;
5632 dev->num_tx_queues = txqs;
5633 dev->real_num_tx_queues = txqs;
5634 if (netif_alloc_netdev_queues(dev))
5638 dev->num_rx_queues = rxqs;
5639 dev->real_num_rx_queues = rxqs;
5640 if (netif_alloc_rx_queues(dev))
5644 strcpy(dev->name, name);
5645 dev->group = INIT_NETDEV_GROUP;
5646 if (!dev->ethtool_ops)
5647 dev->ethtool_ops = &default_ethtool_ops;
5655 free_percpu(dev->pcpu_refcnt);
5665 EXPORT_SYMBOL(alloc_netdev_mqs);
5668 * free_netdev - free network device
5671 * This function does the last stage of destroying an allocated device
5672 * interface. The reference to the device object is released.
5673 * If this is the last reference then it will be freed.
5675 void free_netdev(struct net_device *dev)
5677 struct napi_struct *p, *n;
5679 release_net(dev_net(dev));
5686 kfree(rcu_dereference_protected(dev->ingress_queue, 1));
5688 /* Flush device addresses */
5689 dev_addr_flush(dev);
5691 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
5694 free_percpu(dev->pcpu_refcnt);
5695 dev->pcpu_refcnt = NULL;
5697 /* Compatibility with error handling in drivers */
5698 if (dev->reg_state == NETREG_UNINITIALIZED) {
5699 kfree((char *)dev - dev->padded);
5703 BUG_ON(dev->reg_state != NETREG_UNREGISTERED);
5704 dev->reg_state = NETREG_RELEASED;
5706 /* will free via device release */
5707 put_device(&dev->dev);
5709 EXPORT_SYMBOL(free_netdev);
5712 * synchronize_net - Synchronize with packet receive processing
5714 * Wait for packets currently being received to be done.
5715 * Does not block later packets from starting.
5717 void synchronize_net(void)
5720 if (rtnl_is_locked())
5721 synchronize_rcu_expedited();
5725 EXPORT_SYMBOL(synchronize_net);
5728 * unregister_netdevice_queue - remove device from the kernel
5732 * This function shuts down a device interface and removes it
5733 * from the kernel tables.
5734 * If head not NULL, device is queued to be unregistered later.
5736 * Callers must hold the rtnl semaphore. You may want
5737 * unregister_netdev() instead of this.
5740 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head)
5745 list_move_tail(&dev->unreg_list, head);
5747 rollback_registered(dev);
5748 /* Finish processing unregister after unlock */
5752 EXPORT_SYMBOL(unregister_netdevice_queue);
5755 * unregister_netdevice_many - unregister many devices
5756 * @head: list of devices
5758 void unregister_netdevice_many(struct list_head *head)
5760 struct net_device *dev;
5762 if (!list_empty(head)) {
5763 rollback_registered_many(head);
5764 list_for_each_entry(dev, head, unreg_list)
5768 EXPORT_SYMBOL(unregister_netdevice_many);
5771 * unregister_netdev - remove device from the kernel
5774 * This function shuts down a device interface and removes it
5775 * from the kernel tables.
5777 * This is just a wrapper for unregister_netdevice that takes
5778 * the rtnl semaphore. In general you want to use this and not
5779 * unregister_netdevice.
5781 void unregister_netdev(struct net_device *dev)
5784 unregister_netdevice(dev);
5787 EXPORT_SYMBOL(unregister_netdev);
5790 * dev_change_net_namespace - move device to different nethost namespace
5792 * @net: network namespace
5793 * @pat: If not NULL name pattern to try if the current device name
5794 * is already taken in the destination network namespace.
5796 * This function shuts down a device interface and moves it
5797 * to a new network namespace. On success 0 is returned, on
5798 * a failure a netagive errno code is returned.
5800 * Callers must hold the rtnl semaphore.
5803 int dev_change_net_namespace(struct net_device *dev, struct net *net, const char *pat)
5809 /* Don't allow namespace local devices to be moved. */
5811 if (dev->features & NETIF_F_NETNS_LOCAL)
5814 /* Ensure the device has been registrered */
5815 if (dev->reg_state != NETREG_REGISTERED)
5818 /* Get out if there is nothing todo */
5820 if (net_eq(dev_net(dev), net))
5823 /* Pick the destination device name, and ensure
5824 * we can use it in the destination network namespace.
5827 if (__dev_get_by_name(net, dev->name)) {
5828 /* We get here if we can't use the current device name */
5831 if (dev_get_valid_name(net, dev, pat) < 0)
5836 * And now a mini version of register_netdevice unregister_netdevice.
5839 /* If device is running close it first. */
5842 /* And unlink it from device chain */
5844 unlist_netdevice(dev);
5848 /* Shutdown queueing discipline. */
5851 /* Notify protocols, that we are about to destroy
5852 this device. They should clean all the things.
5854 Note that dev->reg_state stays at NETREG_REGISTERED.
5855 This is wanted because this way 8021q and macvlan know
5856 the device is just moving and can keep their slaves up.
5858 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
5860 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
5861 rtmsg_ifinfo(RTM_DELLINK, dev, ~0U);
5864 * Flush the unicast and multicast chains
5869 /* Send a netdev-removed uevent to the old namespace */
5870 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE);
5872 /* Actually switch the network namespace */
5873 dev_net_set(dev, net);
5875 /* If there is an ifindex conflict assign a new one */
5876 if (__dev_get_by_index(net, dev->ifindex)) {
5877 int iflink = (dev->iflink == dev->ifindex);
5878 dev->ifindex = dev_new_index(net);
5880 dev->iflink = dev->ifindex;
5883 /* Send a netdev-add uevent to the new namespace */
5884 kobject_uevent(&dev->dev.kobj, KOBJ_ADD);
5886 /* Fixup kobjects */
5887 err = device_rename(&dev->dev, dev->name);
5890 /* Add the device back in the hashes */
5891 list_netdevice(dev);
5893 /* Notify protocols, that a new device appeared. */
5894 call_netdevice_notifiers(NETDEV_REGISTER, dev);
5897 * Prevent userspace races by waiting until the network
5898 * device is fully setup before sending notifications.
5900 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U);
5907 EXPORT_SYMBOL_GPL(dev_change_net_namespace);
5909 static int dev_cpu_callback(struct notifier_block *nfb,
5910 unsigned long action,
5913 struct sk_buff **list_skb;
5914 struct sk_buff *skb;
5915 unsigned int cpu, oldcpu = (unsigned long)ocpu;
5916 struct softnet_data *sd, *oldsd;
5918 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
5921 local_irq_disable();
5922 cpu = smp_processor_id();
5923 sd = &per_cpu(softnet_data, cpu);
5924 oldsd = &per_cpu(softnet_data, oldcpu);
5926 /* Find end of our completion_queue. */
5927 list_skb = &sd->completion_queue;
5929 list_skb = &(*list_skb)->next;
5930 /* Append completion queue from offline CPU. */
5931 *list_skb = oldsd->completion_queue;
5932 oldsd->completion_queue = NULL;
5934 /* Append output queue from offline CPU. */
5935 if (oldsd->output_queue) {
5936 *sd->output_queue_tailp = oldsd->output_queue;
5937 sd->output_queue_tailp = oldsd->output_queue_tailp;
5938 oldsd->output_queue = NULL;
5939 oldsd->output_queue_tailp = &oldsd->output_queue;
5941 /* Append NAPI poll list from offline CPU. */
5942 if (!list_empty(&oldsd->poll_list)) {
5943 list_splice_init(&oldsd->poll_list, &sd->poll_list);
5944 raise_softirq_irqoff(NET_RX_SOFTIRQ);
5947 raise_softirq_irqoff(NET_TX_SOFTIRQ);
5950 /* Process offline CPU's input_pkt_queue */
5951 while ((skb = __skb_dequeue(&oldsd->process_queue))) {
5953 input_queue_head_incr(oldsd);
5955 while ((skb = __skb_dequeue(&oldsd->input_pkt_queue))) {
5957 input_queue_head_incr(oldsd);
5965 * netdev_increment_features - increment feature set by one
5966 * @all: current feature set
5967 * @one: new feature set
5968 * @mask: mask feature set
5970 * Computes a new feature set after adding a device with feature set
5971 * @one to the master device with current feature set @all. Will not
5972 * enable anything that is off in @mask. Returns the new feature set.
5974 netdev_features_t netdev_increment_features(netdev_features_t all,
5975 netdev_features_t one, netdev_features_t mask)
5977 if (mask & NETIF_F_GEN_CSUM)
5978 mask |= NETIF_F_ALL_CSUM;
5979 mask |= NETIF_F_VLAN_CHALLENGED;
5981 all |= one & (NETIF_F_ONE_FOR_ALL|NETIF_F_ALL_CSUM) & mask;
5982 all &= one | ~NETIF_F_ALL_FOR_ALL;
5984 /* If one device supports hw checksumming, set for all. */
5985 if (all & NETIF_F_GEN_CSUM)
5986 all &= ~(NETIF_F_ALL_CSUM & ~NETIF_F_GEN_CSUM);
5990 EXPORT_SYMBOL(netdev_increment_features);
5992 static struct hlist_head *netdev_create_hash(void)
5995 struct hlist_head *hash;
5997 hash = kmalloc(sizeof(*hash) * NETDEV_HASHENTRIES, GFP_KERNEL);
5999 for (i = 0; i < NETDEV_HASHENTRIES; i++)
6000 INIT_HLIST_HEAD(&hash[i]);
6005 /* Initialize per network namespace state */
6006 static int __net_init netdev_init(struct net *net)
6008 if (net != &init_net)
6009 INIT_LIST_HEAD(&net->dev_base_head);
6011 net->dev_name_head = netdev_create_hash();
6012 if (net->dev_name_head == NULL)
6015 net->dev_index_head = netdev_create_hash();
6016 if (net->dev_index_head == NULL)
6022 kfree(net->dev_name_head);
6028 * netdev_drivername - network driver for the device
6029 * @dev: network device
6031 * Determine network driver for device.
6033 const char *netdev_drivername(const struct net_device *dev)
6035 const struct device_driver *driver;
6036 const struct device *parent;
6037 const char *empty = "";
6039 parent = dev->dev.parent;
6043 driver = parent->driver;
6044 if (driver && driver->name)
6045 return driver->name;
6049 static int __netdev_printk(const char *level, const struct net_device *dev,
6050 struct va_format *vaf)
6054 if (dev && dev->dev.parent) {
6055 r = dev_printk_emit(level[1] - '0',
6058 dev_driver_string(dev->dev.parent),
6059 dev_name(dev->dev.parent),
6060 netdev_name(dev), vaf);
6062 r = printk("%s%s: %pV", level, netdev_name(dev), vaf);
6064 r = printk("%s(NULL net_device): %pV", level, vaf);
6070 int netdev_printk(const char *level, const struct net_device *dev,
6071 const char *format, ...)
6073 struct va_format vaf;
6077 va_start(args, format);
6082 r = __netdev_printk(level, dev, &vaf);
6088 EXPORT_SYMBOL(netdev_printk);
6090 #define define_netdev_printk_level(func, level) \
6091 int func(const struct net_device *dev, const char *fmt, ...) \
6094 struct va_format vaf; \
6097 va_start(args, fmt); \
6102 r = __netdev_printk(level, dev, &vaf); \
6108 EXPORT_SYMBOL(func);
6110 define_netdev_printk_level(netdev_emerg, KERN_EMERG);
6111 define_netdev_printk_level(netdev_alert, KERN_ALERT);
6112 define_netdev_printk_level(netdev_crit, KERN_CRIT);
6113 define_netdev_printk_level(netdev_err, KERN_ERR);
6114 define_netdev_printk_level(netdev_warn, KERN_WARNING);
6115 define_netdev_printk_level(netdev_notice, KERN_NOTICE);
6116 define_netdev_printk_level(netdev_info, KERN_INFO);
6118 static void __net_exit netdev_exit(struct net *net)
6120 kfree(net->dev_name_head);
6121 kfree(net->dev_index_head);
6124 static struct pernet_operations __net_initdata netdev_net_ops = {
6125 .init = netdev_init,
6126 .exit = netdev_exit,
6129 static void __net_exit default_device_exit(struct net *net)
6131 struct net_device *dev, *aux;
6133 * Push all migratable network devices back to the
6134 * initial network namespace
6137 for_each_netdev_safe(net, dev, aux) {
6139 char fb_name[IFNAMSIZ];
6141 /* Ignore unmoveable devices (i.e. loopback) */
6142 if (dev->features & NETIF_F_NETNS_LOCAL)
6145 /* Leave virtual devices for the generic cleanup */
6146 if (dev->rtnl_link_ops)
6149 /* Push remaining network devices to init_net */
6150 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex);
6151 err = dev_change_net_namespace(dev, &init_net, fb_name);
6153 pr_emerg("%s: failed to move %s to init_net: %d\n",
6154 __func__, dev->name, err);
6161 static void __net_exit default_device_exit_batch(struct list_head *net_list)
6163 /* At exit all network devices most be removed from a network
6164 * namespace. Do this in the reverse order of registration.
6165 * Do this across as many network namespaces as possible to
6166 * improve batching efficiency.
6168 struct net_device *dev;
6170 LIST_HEAD(dev_kill_list);
6173 list_for_each_entry(net, net_list, exit_list) {
6174 for_each_netdev_reverse(net, dev) {
6175 if (dev->rtnl_link_ops)
6176 dev->rtnl_link_ops->dellink(dev, &dev_kill_list);
6178 unregister_netdevice_queue(dev, &dev_kill_list);
6181 unregister_netdevice_many(&dev_kill_list);
6182 list_del(&dev_kill_list);
6186 static struct pernet_operations __net_initdata default_device_ops = {
6187 .exit = default_device_exit,
6188 .exit_batch = default_device_exit_batch,
6192 * Initialize the DEV module. At boot time this walks the device list and
6193 * unhooks any devices that fail to initialise (normally hardware not
6194 * present) and leaves us with a valid list of present and active devices.
6199 * This is called single threaded during boot, so no need
6200 * to take the rtnl semaphore.
6202 static int __init net_dev_init(void)
6204 int i, rc = -ENOMEM;
6206 BUG_ON(!dev_boot_phase);
6208 if (dev_proc_init())
6211 if (netdev_kobject_init())
6214 INIT_LIST_HEAD(&ptype_all);
6215 for (i = 0; i < PTYPE_HASH_SIZE; i++)
6216 INIT_LIST_HEAD(&ptype_base[i]);
6218 INIT_LIST_HEAD(&offload_base);
6220 if (register_pernet_subsys(&netdev_net_ops))
6224 * Initialise the packet receive queues.
6227 for_each_possible_cpu(i) {
6228 struct softnet_data *sd = &per_cpu(softnet_data, i);
6230 memset(sd, 0, sizeof(*sd));
6231 skb_queue_head_init(&sd->input_pkt_queue);
6232 skb_queue_head_init(&sd->process_queue);
6233 sd->completion_queue = NULL;
6234 INIT_LIST_HEAD(&sd->poll_list);
6235 sd->output_queue = NULL;
6236 sd->output_queue_tailp = &sd->output_queue;
6238 sd->csd.func = rps_trigger_softirq;
6244 sd->backlog.poll = process_backlog;
6245 sd->backlog.weight = weight_p;
6246 sd->backlog.gro_list = NULL;
6247 sd->backlog.gro_count = 0;
6252 /* The loopback device is special if any other network devices
6253 * is present in a network namespace the loopback device must
6254 * be present. Since we now dynamically allocate and free the
6255 * loopback device ensure this invariant is maintained by
6256 * keeping the loopback device as the first device on the
6257 * list of network devices. Ensuring the loopback devices
6258 * is the first device that appears and the last network device
6261 if (register_pernet_device(&loopback_net_ops))
6264 if (register_pernet_device(&default_device_ops))
6267 open_softirq(NET_TX_SOFTIRQ, net_tx_action);
6268 open_softirq(NET_RX_SOFTIRQ, net_rx_action);
6270 hotcpu_notifier(dev_cpu_callback, 0);
6277 subsys_initcall(net_dev_init);