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 static_key_slow_inc(&netstamp_needed);
1550 EXPORT_SYMBOL(net_enable_timestamp);
1552 void net_disable_timestamp(void)
1554 #ifdef HAVE_JUMP_LABEL
1555 if (in_interrupt()) {
1556 atomic_inc(&netstamp_needed_deferred);
1560 static_key_slow_dec(&netstamp_needed);
1562 EXPORT_SYMBOL(net_disable_timestamp);
1564 static inline void net_timestamp_set(struct sk_buff *skb)
1566 skb->tstamp.tv64 = 0;
1567 if (static_key_false(&netstamp_needed))
1568 __net_timestamp(skb);
1571 #define net_timestamp_check(COND, SKB) \
1572 if (static_key_false(&netstamp_needed)) { \
1573 if ((COND) && !(SKB)->tstamp.tv64) \
1574 __net_timestamp(SKB); \
1577 static inline bool is_skb_forwardable(struct net_device *dev,
1578 struct sk_buff *skb)
1582 if (!(dev->flags & IFF_UP))
1585 len = dev->mtu + dev->hard_header_len + VLAN_HLEN;
1586 if (skb->len <= len)
1589 /* if TSO is enabled, we don't care about the length as the packet
1590 * could be forwarded without being segmented before
1592 if (skb_is_gso(skb))
1599 * dev_forward_skb - loopback an skb to another netif
1601 * @dev: destination network device
1602 * @skb: buffer to forward
1605 * NET_RX_SUCCESS (no congestion)
1606 * NET_RX_DROP (packet was dropped, but freed)
1608 * dev_forward_skb can be used for injecting an skb from the
1609 * start_xmit function of one device into the receive queue
1610 * of another device.
1612 * The receiving device may be in another namespace, so
1613 * we have to clear all information in the skb that could
1614 * impact namespace isolation.
1616 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1618 if (skb_shinfo(skb)->tx_flags & SKBTX_DEV_ZEROCOPY) {
1619 if (skb_copy_ubufs(skb, GFP_ATOMIC)) {
1620 atomic_long_inc(&dev->rx_dropped);
1628 if (unlikely(!is_skb_forwardable(dev, skb))) {
1629 atomic_long_inc(&dev->rx_dropped);
1636 skb->tstamp.tv64 = 0;
1637 skb->pkt_type = PACKET_HOST;
1638 skb->protocol = eth_type_trans(skb, dev);
1642 nf_reset_trace(skb);
1643 return netif_rx(skb);
1645 EXPORT_SYMBOL_GPL(dev_forward_skb);
1647 static inline int deliver_skb(struct sk_buff *skb,
1648 struct packet_type *pt_prev,
1649 struct net_device *orig_dev)
1651 if (unlikely(skb_orphan_frags(skb, GFP_ATOMIC)))
1653 atomic_inc(&skb->users);
1654 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
1657 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb)
1659 if (!ptype->af_packet_priv || !skb->sk)
1662 if (ptype->id_match)
1663 return ptype->id_match(ptype, skb->sk);
1664 else if ((struct sock *)ptype->af_packet_priv == skb->sk)
1671 * Support routine. Sends outgoing frames to any network
1672 * taps currently in use.
1675 static void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
1677 struct packet_type *ptype;
1678 struct sk_buff *skb2 = NULL;
1679 struct packet_type *pt_prev = NULL;
1682 list_for_each_entry_rcu(ptype, &ptype_all, list) {
1683 /* Never send packets back to the socket
1684 * they originated from - MvS (miquels@drinkel.ow.org)
1686 if ((ptype->dev == dev || !ptype->dev) &&
1687 (!skb_loop_sk(ptype, skb))) {
1689 deliver_skb(skb2, pt_prev, skb->dev);
1694 skb2 = skb_clone(skb, GFP_ATOMIC);
1698 net_timestamp_set(skb2);
1700 /* skb->nh should be correctly
1701 set by sender, so that the second statement is
1702 just protection against buggy protocols.
1704 skb_reset_mac_header(skb2);
1706 if (skb_network_header(skb2) < skb2->data ||
1707 skb2->network_header > skb2->tail) {
1708 net_crit_ratelimited("protocol %04x is buggy, dev %s\n",
1709 ntohs(skb2->protocol),
1711 skb_reset_network_header(skb2);
1714 skb2->transport_header = skb2->network_header;
1715 skb2->pkt_type = PACKET_OUTGOING;
1720 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
1725 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change
1726 * @dev: Network device
1727 * @txq: number of queues available
1729 * If real_num_tx_queues is changed the tc mappings may no longer be
1730 * valid. To resolve this verify the tc mapping remains valid and if
1731 * not NULL the mapping. With no priorities mapping to this
1732 * offset/count pair it will no longer be used. In the worst case TC0
1733 * is invalid nothing can be done so disable priority mappings. If is
1734 * expected that drivers will fix this mapping if they can before
1735 * calling netif_set_real_num_tx_queues.
1737 static void netif_setup_tc(struct net_device *dev, unsigned int txq)
1740 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
1742 /* If TC0 is invalidated disable TC mapping */
1743 if (tc->offset + tc->count > txq) {
1744 pr_warn("Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n");
1749 /* Invalidated prio to tc mappings set to TC0 */
1750 for (i = 1; i < TC_BITMASK + 1; i++) {
1751 int q = netdev_get_prio_tc_map(dev, i);
1753 tc = &dev->tc_to_txq[q];
1754 if (tc->offset + tc->count > txq) {
1755 pr_warn("Number of in use tx queues changed. Priority %i to tc mapping %i is no longer valid. Setting map to 0\n",
1757 netdev_set_prio_tc_map(dev, i, 0);
1763 static DEFINE_MUTEX(xps_map_mutex);
1764 #define xmap_dereference(P) \
1765 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex))
1767 static struct xps_map *remove_xps_queue(struct xps_dev_maps *dev_maps,
1770 struct xps_map *map = NULL;
1774 map = xmap_dereference(dev_maps->cpu_map[cpu]);
1776 for (pos = 0; map && pos < map->len; pos++) {
1777 if (map->queues[pos] == index) {
1779 map->queues[pos] = map->queues[--map->len];
1781 RCU_INIT_POINTER(dev_maps->cpu_map[cpu], NULL);
1782 kfree_rcu(map, rcu);
1792 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index)
1794 struct xps_dev_maps *dev_maps;
1796 bool active = false;
1798 mutex_lock(&xps_map_mutex);
1799 dev_maps = xmap_dereference(dev->xps_maps);
1804 for_each_possible_cpu(cpu) {
1805 for (i = index; i < dev->num_tx_queues; i++) {
1806 if (!remove_xps_queue(dev_maps, cpu, i))
1809 if (i == dev->num_tx_queues)
1814 RCU_INIT_POINTER(dev->xps_maps, NULL);
1815 kfree_rcu(dev_maps, rcu);
1818 for (i = index; i < dev->num_tx_queues; i++)
1819 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, i),
1823 mutex_unlock(&xps_map_mutex);
1826 static struct xps_map *expand_xps_map(struct xps_map *map,
1829 struct xps_map *new_map;
1830 int alloc_len = XPS_MIN_MAP_ALLOC;
1833 for (pos = 0; map && pos < map->len; pos++) {
1834 if (map->queues[pos] != index)
1839 /* Need to add queue to this CPU's existing map */
1841 if (pos < map->alloc_len)
1844 alloc_len = map->alloc_len * 2;
1847 /* Need to allocate new map to store queue on this CPU's map */
1848 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL,
1853 for (i = 0; i < pos; i++)
1854 new_map->queues[i] = map->queues[i];
1855 new_map->alloc_len = alloc_len;
1861 int netif_set_xps_queue(struct net_device *dev, struct cpumask *mask, u16 index)
1863 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL;
1864 struct xps_map *map, *new_map;
1865 int maps_sz = max_t(unsigned int, XPS_DEV_MAPS_SIZE, L1_CACHE_BYTES);
1866 int cpu, numa_node_id = -2;
1867 bool active = false;
1869 mutex_lock(&xps_map_mutex);
1871 dev_maps = xmap_dereference(dev->xps_maps);
1873 /* allocate memory for queue storage */
1874 for_each_online_cpu(cpu) {
1875 if (!cpumask_test_cpu(cpu, mask))
1879 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL);
1880 if (!new_dev_maps) {
1881 mutex_unlock(&xps_map_mutex);
1885 map = dev_maps ? xmap_dereference(dev_maps->cpu_map[cpu]) :
1888 map = expand_xps_map(map, cpu, index);
1892 RCU_INIT_POINTER(new_dev_maps->cpu_map[cpu], map);
1896 goto out_no_new_maps;
1898 for_each_possible_cpu(cpu) {
1899 if (cpumask_test_cpu(cpu, mask) && cpu_online(cpu)) {
1900 /* add queue to CPU maps */
1903 map = xmap_dereference(new_dev_maps->cpu_map[cpu]);
1904 while ((pos < map->len) && (map->queues[pos] != index))
1907 if (pos == map->len)
1908 map->queues[map->len++] = index;
1910 if (numa_node_id == -2)
1911 numa_node_id = cpu_to_node(cpu);
1912 else if (numa_node_id != cpu_to_node(cpu))
1915 } else if (dev_maps) {
1916 /* fill in the new device map from the old device map */
1917 map = xmap_dereference(dev_maps->cpu_map[cpu]);
1918 RCU_INIT_POINTER(new_dev_maps->cpu_map[cpu], map);
1923 rcu_assign_pointer(dev->xps_maps, new_dev_maps);
1925 /* Cleanup old maps */
1927 for_each_possible_cpu(cpu) {
1928 new_map = xmap_dereference(new_dev_maps->cpu_map[cpu]);
1929 map = xmap_dereference(dev_maps->cpu_map[cpu]);
1930 if (map && map != new_map)
1931 kfree_rcu(map, rcu);
1934 kfree_rcu(dev_maps, rcu);
1937 dev_maps = new_dev_maps;
1941 /* update Tx queue numa node */
1942 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index),
1943 (numa_node_id >= 0) ? numa_node_id :
1949 /* removes queue from unused CPUs */
1950 for_each_possible_cpu(cpu) {
1951 if (cpumask_test_cpu(cpu, mask) && cpu_online(cpu))
1954 if (remove_xps_queue(dev_maps, cpu, index))
1958 /* free map if not active */
1960 RCU_INIT_POINTER(dev->xps_maps, NULL);
1961 kfree_rcu(dev_maps, rcu);
1965 mutex_unlock(&xps_map_mutex);
1969 /* remove any maps that we added */
1970 for_each_possible_cpu(cpu) {
1971 new_map = xmap_dereference(new_dev_maps->cpu_map[cpu]);
1972 map = dev_maps ? xmap_dereference(dev_maps->cpu_map[cpu]) :
1974 if (new_map && new_map != map)
1978 mutex_unlock(&xps_map_mutex);
1980 kfree(new_dev_maps);
1983 EXPORT_SYMBOL(netif_set_xps_queue);
1987 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues
1988 * greater then real_num_tx_queues stale skbs on the qdisc must be flushed.
1990 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq)
1994 if (txq < 1 || txq > dev->num_tx_queues)
1997 if (dev->reg_state == NETREG_REGISTERED ||
1998 dev->reg_state == NETREG_UNREGISTERING) {
2001 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues,
2007 netif_setup_tc(dev, txq);
2009 if (txq < dev->real_num_tx_queues) {
2010 qdisc_reset_all_tx_gt(dev, txq);
2012 netif_reset_xps_queues_gt(dev, txq);
2017 dev->real_num_tx_queues = txq;
2020 EXPORT_SYMBOL(netif_set_real_num_tx_queues);
2024 * netif_set_real_num_rx_queues - set actual number of RX queues used
2025 * @dev: Network device
2026 * @rxq: Actual number of RX queues
2028 * This must be called either with the rtnl_lock held or before
2029 * registration of the net device. Returns 0 on success, or a
2030 * negative error code. If called before registration, it always
2033 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq)
2037 if (rxq < 1 || rxq > dev->num_rx_queues)
2040 if (dev->reg_state == NETREG_REGISTERED) {
2043 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues,
2049 dev->real_num_rx_queues = rxq;
2052 EXPORT_SYMBOL(netif_set_real_num_rx_queues);
2056 * netif_get_num_default_rss_queues - default number of RSS queues
2058 * This routine should set an upper limit on the number of RSS queues
2059 * used by default by multiqueue devices.
2061 int netif_get_num_default_rss_queues(void)
2063 return min_t(int, DEFAULT_MAX_NUM_RSS_QUEUES, num_online_cpus());
2065 EXPORT_SYMBOL(netif_get_num_default_rss_queues);
2067 static inline void __netif_reschedule(struct Qdisc *q)
2069 struct softnet_data *sd;
2070 unsigned long flags;
2072 local_irq_save(flags);
2073 sd = &__get_cpu_var(softnet_data);
2074 q->next_sched = NULL;
2075 *sd->output_queue_tailp = q;
2076 sd->output_queue_tailp = &q->next_sched;
2077 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2078 local_irq_restore(flags);
2081 void __netif_schedule(struct Qdisc *q)
2083 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
2084 __netif_reschedule(q);
2086 EXPORT_SYMBOL(__netif_schedule);
2088 void dev_kfree_skb_irq(struct sk_buff *skb)
2090 if (atomic_dec_and_test(&skb->users)) {
2091 struct softnet_data *sd;
2092 unsigned long flags;
2094 local_irq_save(flags);
2095 sd = &__get_cpu_var(softnet_data);
2096 skb->next = sd->completion_queue;
2097 sd->completion_queue = skb;
2098 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2099 local_irq_restore(flags);
2102 EXPORT_SYMBOL(dev_kfree_skb_irq);
2104 void dev_kfree_skb_any(struct sk_buff *skb)
2106 if (in_irq() || irqs_disabled())
2107 dev_kfree_skb_irq(skb);
2111 EXPORT_SYMBOL(dev_kfree_skb_any);
2115 * netif_device_detach - mark device as removed
2116 * @dev: network device
2118 * Mark device as removed from system and therefore no longer available.
2120 void netif_device_detach(struct net_device *dev)
2122 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) &&
2123 netif_running(dev)) {
2124 netif_tx_stop_all_queues(dev);
2127 EXPORT_SYMBOL(netif_device_detach);
2130 * netif_device_attach - mark device as attached
2131 * @dev: network device
2133 * Mark device as attached from system and restart if needed.
2135 void netif_device_attach(struct net_device *dev)
2137 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) &&
2138 netif_running(dev)) {
2139 netif_tx_wake_all_queues(dev);
2140 __netdev_watchdog_up(dev);
2143 EXPORT_SYMBOL(netif_device_attach);
2145 static void skb_warn_bad_offload(const struct sk_buff *skb)
2147 static const netdev_features_t null_features = 0;
2148 struct net_device *dev = skb->dev;
2149 const char *driver = "";
2151 if (dev && dev->dev.parent)
2152 driver = dev_driver_string(dev->dev.parent);
2154 WARN(1, "%s: caps=(%pNF, %pNF) len=%d data_len=%d gso_size=%d "
2155 "gso_type=%d ip_summed=%d\n",
2156 driver, dev ? &dev->features : &null_features,
2157 skb->sk ? &skb->sk->sk_route_caps : &null_features,
2158 skb->len, skb->data_len, skb_shinfo(skb)->gso_size,
2159 skb_shinfo(skb)->gso_type, skb->ip_summed);
2163 * Invalidate hardware checksum when packet is to be mangled, and
2164 * complete checksum manually on outgoing path.
2166 int skb_checksum_help(struct sk_buff *skb)
2169 int ret = 0, offset;
2171 if (skb->ip_summed == CHECKSUM_COMPLETE)
2172 goto out_set_summed;
2174 if (unlikely(skb_shinfo(skb)->gso_size)) {
2175 skb_warn_bad_offload(skb);
2179 /* Before computing a checksum, we should make sure no frag could
2180 * be modified by an external entity : checksum could be wrong.
2182 if (skb_has_shared_frag(skb)) {
2183 ret = __skb_linearize(skb);
2188 offset = skb_checksum_start_offset(skb);
2189 BUG_ON(offset >= skb_headlen(skb));
2190 csum = skb_checksum(skb, offset, skb->len - offset, 0);
2192 offset += skb->csum_offset;
2193 BUG_ON(offset + sizeof(__sum16) > skb_headlen(skb));
2195 if (skb_cloned(skb) &&
2196 !skb_clone_writable(skb, offset + sizeof(__sum16))) {
2197 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2202 *(__sum16 *)(skb->data + offset) = csum_fold(csum);
2204 skb->ip_summed = CHECKSUM_NONE;
2208 EXPORT_SYMBOL(skb_checksum_help);
2211 * skb_mac_gso_segment - mac layer segmentation handler.
2212 * @skb: buffer to segment
2213 * @features: features for the output path (see dev->features)
2215 struct sk_buff *skb_mac_gso_segment(struct sk_buff *skb,
2216 netdev_features_t features)
2218 struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT);
2219 struct packet_offload *ptype;
2220 __be16 type = skb->protocol;
2221 int vlan_depth = ETH_HLEN;
2223 while (type == htons(ETH_P_8021Q)) {
2224 struct vlan_hdr *vh;
2226 if (unlikely(!pskb_may_pull(skb, vlan_depth + VLAN_HLEN)))
2227 return ERR_PTR(-EINVAL);
2229 vh = (struct vlan_hdr *)(skb->data + vlan_depth);
2230 type = vh->h_vlan_encapsulated_proto;
2231 vlan_depth += VLAN_HLEN;
2234 __skb_pull(skb, skb->mac_len);
2237 list_for_each_entry_rcu(ptype, &offload_base, list) {
2238 if (ptype->type == type && ptype->callbacks.gso_segment) {
2239 if (unlikely(skb->ip_summed != CHECKSUM_PARTIAL)) {
2242 err = ptype->callbacks.gso_send_check(skb);
2243 segs = ERR_PTR(err);
2244 if (err || skb_gso_ok(skb, features))
2246 __skb_push(skb, (skb->data -
2247 skb_network_header(skb)));
2249 segs = ptype->callbacks.gso_segment(skb, features);
2255 __skb_push(skb, skb->data - skb_mac_header(skb));
2259 EXPORT_SYMBOL(skb_mac_gso_segment);
2262 /* openvswitch calls this on rx path, so we need a different check.
2264 static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
2267 return skb->ip_summed != CHECKSUM_PARTIAL;
2269 return skb->ip_summed == CHECKSUM_NONE;
2273 * __skb_gso_segment - Perform segmentation on skb.
2274 * @skb: buffer to segment
2275 * @features: features for the output path (see dev->features)
2276 * @tx_path: whether it is called in TX path
2278 * This function segments the given skb and returns a list of segments.
2280 * It may return NULL if the skb requires no segmentation. This is
2281 * only possible when GSO is used for verifying header integrity.
2283 struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
2284 netdev_features_t features, bool tx_path)
2286 if (unlikely(skb_needs_check(skb, tx_path))) {
2289 skb_warn_bad_offload(skb);
2291 if (skb_header_cloned(skb) &&
2292 (err = pskb_expand_head(skb, 0, 0, GFP_ATOMIC)))
2293 return ERR_PTR(err);
2296 SKB_GSO_CB(skb)->mac_offset = skb_headroom(skb);
2297 skb_reset_mac_header(skb);
2298 skb_reset_mac_len(skb);
2300 return skb_mac_gso_segment(skb, features);
2302 EXPORT_SYMBOL(__skb_gso_segment);
2304 /* Take action when hardware reception checksum errors are detected. */
2306 void netdev_rx_csum_fault(struct net_device *dev)
2308 if (net_ratelimit()) {
2309 pr_err("%s: hw csum failure\n", dev ? dev->name : "<unknown>");
2313 EXPORT_SYMBOL(netdev_rx_csum_fault);
2316 /* Actually, we should eliminate this check as soon as we know, that:
2317 * 1. IOMMU is present and allows to map all the memory.
2318 * 2. No high memory really exists on this machine.
2321 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
2323 #ifdef CONFIG_HIGHMEM
2325 if (!(dev->features & NETIF_F_HIGHDMA)) {
2326 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2327 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2328 if (PageHighMem(skb_frag_page(frag)))
2333 if (PCI_DMA_BUS_IS_PHYS) {
2334 struct device *pdev = dev->dev.parent;
2338 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2339 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2340 dma_addr_t addr = page_to_phys(skb_frag_page(frag));
2341 if (!pdev->dma_mask || addr + PAGE_SIZE - 1 > *pdev->dma_mask)
2350 void (*destructor)(struct sk_buff *skb);
2353 #define DEV_GSO_CB(skb) ((struct dev_gso_cb *)(skb)->cb)
2355 static void dev_gso_skb_destructor(struct sk_buff *skb)
2357 struct dev_gso_cb *cb;
2360 struct sk_buff *nskb = skb->next;
2362 skb->next = nskb->next;
2365 } while (skb->next);
2367 cb = DEV_GSO_CB(skb);
2369 cb->destructor(skb);
2373 * dev_gso_segment - Perform emulated hardware segmentation on skb.
2374 * @skb: buffer to segment
2375 * @features: device features as applicable to this skb
2377 * This function segments the given skb and stores the list of segments
2380 static int dev_gso_segment(struct sk_buff *skb, netdev_features_t features)
2382 struct sk_buff *segs;
2384 segs = skb_gso_segment(skb, features);
2386 /* Verifying header integrity only. */
2391 return PTR_ERR(segs);
2394 DEV_GSO_CB(skb)->destructor = skb->destructor;
2395 skb->destructor = dev_gso_skb_destructor;
2400 static bool can_checksum_protocol(netdev_features_t features, __be16 protocol)
2402 return ((features & NETIF_F_GEN_CSUM) ||
2403 ((features & NETIF_F_V4_CSUM) &&
2404 protocol == htons(ETH_P_IP)) ||
2405 ((features & NETIF_F_V6_CSUM) &&
2406 protocol == htons(ETH_P_IPV6)) ||
2407 ((features & NETIF_F_FCOE_CRC) &&
2408 protocol == htons(ETH_P_FCOE)));
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 features &= ~NETIF_F_SG;
2418 } else if (illegal_highdma(skb->dev, skb)) {
2419 features &= ~NETIF_F_SG;
2425 netdev_features_t netif_skb_features(struct sk_buff *skb)
2427 __be16 protocol = skb->protocol;
2428 netdev_features_t features = skb->dev->features;
2430 if (skb_shinfo(skb)->gso_segs > skb->dev->gso_max_segs)
2431 features &= ~NETIF_F_GSO_MASK;
2433 if (protocol == htons(ETH_P_8021Q)) {
2434 struct vlan_ethhdr *veh = (struct vlan_ethhdr *)skb->data;
2435 protocol = veh->h_vlan_encapsulated_proto;
2436 } else if (!vlan_tx_tag_present(skb)) {
2437 return harmonize_features(skb, protocol, features);
2440 features &= (skb->dev->vlan_features | NETIF_F_HW_VLAN_TX);
2442 if (protocol != htons(ETH_P_8021Q)) {
2443 return harmonize_features(skb, protocol, features);
2445 features &= NETIF_F_SG | NETIF_F_HIGHDMA | NETIF_F_FRAGLIST |
2446 NETIF_F_GEN_CSUM | NETIF_F_HW_VLAN_TX;
2447 return harmonize_features(skb, protocol, features);
2450 EXPORT_SYMBOL(netif_skb_features);
2453 * Returns true if either:
2454 * 1. skb has frag_list and the device doesn't support FRAGLIST, or
2455 * 2. skb is fragmented and the device does not support SG.
2457 static inline int skb_needs_linearize(struct sk_buff *skb,
2460 return skb_is_nonlinear(skb) &&
2461 ((skb_has_frag_list(skb) &&
2462 !(features & NETIF_F_FRAGLIST)) ||
2463 (skb_shinfo(skb)->nr_frags &&
2464 !(features & NETIF_F_SG)));
2467 int dev_hard_start_xmit(struct sk_buff *skb, struct net_device *dev,
2468 struct netdev_queue *txq)
2470 const struct net_device_ops *ops = dev->netdev_ops;
2471 int rc = NETDEV_TX_OK;
2472 unsigned int skb_len;
2474 if (likely(!skb->next)) {
2475 netdev_features_t features;
2478 * If device doesn't need skb->dst, release it right now while
2479 * its hot in this cpu cache
2481 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
2484 features = netif_skb_features(skb);
2486 if (vlan_tx_tag_present(skb) &&
2487 !(features & NETIF_F_HW_VLAN_TX)) {
2488 skb = __vlan_put_tag(skb, vlan_tx_tag_get(skb));
2495 /* If encapsulation offload request, verify we are testing
2496 * hardware encapsulation features instead of standard
2497 * features for the netdev
2499 if (skb->encapsulation)
2500 features &= dev->hw_enc_features;
2502 if (netif_needs_gso(skb, features)) {
2503 if (unlikely(dev_gso_segment(skb, features)))
2508 if (skb_needs_linearize(skb, features) &&
2509 __skb_linearize(skb))
2512 /* If packet is not checksummed and device does not
2513 * support checksumming for this protocol, complete
2514 * checksumming here.
2516 if (skb->ip_summed == CHECKSUM_PARTIAL) {
2517 if (skb->encapsulation)
2518 skb_set_inner_transport_header(skb,
2519 skb_checksum_start_offset(skb));
2521 skb_set_transport_header(skb,
2522 skb_checksum_start_offset(skb));
2523 if (!(features & NETIF_F_ALL_CSUM) &&
2524 skb_checksum_help(skb))
2529 if (!list_empty(&ptype_all))
2530 dev_queue_xmit_nit(skb, dev);
2533 rc = ops->ndo_start_xmit(skb, dev);
2534 trace_net_dev_xmit(skb, rc, dev, skb_len);
2535 if (rc == NETDEV_TX_OK)
2536 txq_trans_update(txq);
2542 struct sk_buff *nskb = skb->next;
2544 skb->next = nskb->next;
2548 * If device doesn't need nskb->dst, release it right now while
2549 * its hot in this cpu cache
2551 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
2554 if (!list_empty(&ptype_all))
2555 dev_queue_xmit_nit(nskb, dev);
2557 skb_len = nskb->len;
2558 rc = ops->ndo_start_xmit(nskb, dev);
2559 trace_net_dev_xmit(nskb, rc, dev, skb_len);
2560 if (unlikely(rc != NETDEV_TX_OK)) {
2561 if (rc & ~NETDEV_TX_MASK)
2562 goto out_kfree_gso_skb;
2563 nskb->next = skb->next;
2567 txq_trans_update(txq);
2568 if (unlikely(netif_xmit_stopped(txq) && skb->next))
2569 return NETDEV_TX_BUSY;
2570 } while (skb->next);
2573 if (likely(skb->next == NULL))
2574 skb->destructor = DEV_GSO_CB(skb)->destructor;
2581 static void qdisc_pkt_len_init(struct sk_buff *skb)
2583 const struct skb_shared_info *shinfo = skb_shinfo(skb);
2585 qdisc_skb_cb(skb)->pkt_len = skb->len;
2587 /* To get more precise estimation of bytes sent on wire,
2588 * we add to pkt_len the headers size of all segments
2590 if (shinfo->gso_size) {
2591 unsigned int hdr_len;
2593 /* mac layer + network layer */
2594 hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
2596 /* + transport layer */
2597 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))
2598 hdr_len += tcp_hdrlen(skb);
2600 hdr_len += sizeof(struct udphdr);
2601 qdisc_skb_cb(skb)->pkt_len += (shinfo->gso_segs - 1) * hdr_len;
2605 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
2606 struct net_device *dev,
2607 struct netdev_queue *txq)
2609 spinlock_t *root_lock = qdisc_lock(q);
2613 qdisc_pkt_len_init(skb);
2614 qdisc_calculate_pkt_len(skb, q);
2616 * Heuristic to force contended enqueues to serialize on a
2617 * separate lock before trying to get qdisc main lock.
2618 * This permits __QDISC_STATE_RUNNING owner to get the lock more often
2619 * and dequeue packets faster.
2621 contended = qdisc_is_running(q);
2622 if (unlikely(contended))
2623 spin_lock(&q->busylock);
2625 spin_lock(root_lock);
2626 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
2629 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
2630 qdisc_run_begin(q)) {
2632 * This is a work-conserving queue; there are no old skbs
2633 * waiting to be sent out; and the qdisc is not running -
2634 * xmit the skb directly.
2636 if (!(dev->priv_flags & IFF_XMIT_DST_RELEASE))
2639 qdisc_bstats_update(q, skb);
2641 if (sch_direct_xmit(skb, q, dev, txq, root_lock)) {
2642 if (unlikely(contended)) {
2643 spin_unlock(&q->busylock);
2650 rc = NET_XMIT_SUCCESS;
2653 rc = q->enqueue(skb, q) & NET_XMIT_MASK;
2654 if (qdisc_run_begin(q)) {
2655 if (unlikely(contended)) {
2656 spin_unlock(&q->busylock);
2662 spin_unlock(root_lock);
2663 if (unlikely(contended))
2664 spin_unlock(&q->busylock);
2668 #if IS_ENABLED(CONFIG_NETPRIO_CGROUP)
2669 static void skb_update_prio(struct sk_buff *skb)
2671 struct netprio_map *map = rcu_dereference_bh(skb->dev->priomap);
2673 if (!skb->priority && skb->sk && map) {
2674 unsigned int prioidx = skb->sk->sk_cgrp_prioidx;
2676 if (prioidx < map->priomap_len)
2677 skb->priority = map->priomap[prioidx];
2681 #define skb_update_prio(skb)
2684 static DEFINE_PER_CPU(int, xmit_recursion);
2685 #define RECURSION_LIMIT 10
2688 * dev_loopback_xmit - loop back @skb
2689 * @skb: buffer to transmit
2691 int dev_loopback_xmit(struct sk_buff *skb)
2693 skb_reset_mac_header(skb);
2694 __skb_pull(skb, skb_network_offset(skb));
2695 skb->pkt_type = PACKET_LOOPBACK;
2696 skb->ip_summed = CHECKSUM_UNNECESSARY;
2697 WARN_ON(!skb_dst(skb));
2702 EXPORT_SYMBOL(dev_loopback_xmit);
2705 * dev_queue_xmit - transmit a buffer
2706 * @skb: buffer to transmit
2708 * Queue a buffer for transmission to a network device. The caller must
2709 * have set the device and priority and built the buffer before calling
2710 * this function. The function can be called from an interrupt.
2712 * A negative errno code is returned on a failure. A success does not
2713 * guarantee the frame will be transmitted as it may be dropped due
2714 * to congestion or traffic shaping.
2716 * -----------------------------------------------------------------------------------
2717 * I notice this method can also return errors from the queue disciplines,
2718 * including NET_XMIT_DROP, which is a positive value. So, errors can also
2721 * Regardless of the return value, the skb is consumed, so it is currently
2722 * difficult to retry a send to this method. (You can bump the ref count
2723 * before sending to hold a reference for retry if you are careful.)
2725 * When calling this method, interrupts MUST be enabled. This is because
2726 * the BH enable code must have IRQs enabled so that it will not deadlock.
2729 int dev_queue_xmit(struct sk_buff *skb)
2731 struct net_device *dev = skb->dev;
2732 struct netdev_queue *txq;
2736 skb_reset_mac_header(skb);
2738 /* Disable soft irqs for various locks below. Also
2739 * stops preemption for RCU.
2743 skb_update_prio(skb);
2745 txq = netdev_pick_tx(dev, skb);
2746 q = rcu_dereference_bh(txq->qdisc);
2748 #ifdef CONFIG_NET_CLS_ACT
2749 skb->tc_verd = SET_TC_AT(skb->tc_verd, AT_EGRESS);
2751 trace_net_dev_queue(skb);
2753 rc = __dev_xmit_skb(skb, q, dev, txq);
2757 /* The device has no queue. Common case for software devices:
2758 loopback, all the sorts of tunnels...
2760 Really, it is unlikely that netif_tx_lock protection is necessary
2761 here. (f.e. loopback and IP tunnels are clean ignoring statistics
2763 However, it is possible, that they rely on protection
2766 Check this and shot the lock. It is not prone from deadlocks.
2767 Either shot noqueue qdisc, it is even simpler 8)
2769 if (dev->flags & IFF_UP) {
2770 int cpu = smp_processor_id(); /* ok because BHs are off */
2772 if (txq->xmit_lock_owner != cpu) {
2774 if (__this_cpu_read(xmit_recursion) > RECURSION_LIMIT)
2775 goto recursion_alert;
2777 HARD_TX_LOCK(dev, txq, cpu);
2779 if (!netif_xmit_stopped(txq)) {
2780 __this_cpu_inc(xmit_recursion);
2781 rc = dev_hard_start_xmit(skb, dev, txq);
2782 __this_cpu_dec(xmit_recursion);
2783 if (dev_xmit_complete(rc)) {
2784 HARD_TX_UNLOCK(dev, txq);
2788 HARD_TX_UNLOCK(dev, txq);
2789 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
2792 /* Recursion is detected! It is possible,
2796 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
2802 rcu_read_unlock_bh();
2807 rcu_read_unlock_bh();
2810 EXPORT_SYMBOL(dev_queue_xmit);
2813 /*=======================================================================
2815 =======================================================================*/
2817 int netdev_max_backlog __read_mostly = 1000;
2818 EXPORT_SYMBOL(netdev_max_backlog);
2820 int netdev_tstamp_prequeue __read_mostly = 1;
2821 int netdev_budget __read_mostly = 300;
2822 int weight_p __read_mostly = 64; /* old backlog weight */
2824 /* Called with irq disabled */
2825 static inline void ____napi_schedule(struct softnet_data *sd,
2826 struct napi_struct *napi)
2828 list_add_tail(&napi->poll_list, &sd->poll_list);
2829 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
2834 /* One global table that all flow-based protocols share. */
2835 struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
2836 EXPORT_SYMBOL(rps_sock_flow_table);
2838 struct static_key rps_needed __read_mostly;
2840 static struct rps_dev_flow *
2841 set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
2842 struct rps_dev_flow *rflow, u16 next_cpu)
2844 if (next_cpu != RPS_NO_CPU) {
2845 #ifdef CONFIG_RFS_ACCEL
2846 struct netdev_rx_queue *rxqueue;
2847 struct rps_dev_flow_table *flow_table;
2848 struct rps_dev_flow *old_rflow;
2853 /* Should we steer this flow to a different hardware queue? */
2854 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
2855 !(dev->features & NETIF_F_NTUPLE))
2857 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
2858 if (rxq_index == skb_get_rx_queue(skb))
2861 rxqueue = dev->_rx + rxq_index;
2862 flow_table = rcu_dereference(rxqueue->rps_flow_table);
2865 flow_id = skb->rxhash & flow_table->mask;
2866 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
2867 rxq_index, flow_id);
2871 rflow = &flow_table->flows[flow_id];
2873 if (old_rflow->filter == rflow->filter)
2874 old_rflow->filter = RPS_NO_FILTER;
2878 per_cpu(softnet_data, next_cpu).input_queue_head;
2881 rflow->cpu = next_cpu;
2886 * get_rps_cpu is called from netif_receive_skb and returns the target
2887 * CPU from the RPS map of the receiving queue for a given skb.
2888 * rcu_read_lock must be held on entry.
2890 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
2891 struct rps_dev_flow **rflowp)
2893 struct netdev_rx_queue *rxqueue;
2894 struct rps_map *map;
2895 struct rps_dev_flow_table *flow_table;
2896 struct rps_sock_flow_table *sock_flow_table;
2900 if (skb_rx_queue_recorded(skb)) {
2901 u16 index = skb_get_rx_queue(skb);
2902 if (unlikely(index >= dev->real_num_rx_queues)) {
2903 WARN_ONCE(dev->real_num_rx_queues > 1,
2904 "%s received packet on queue %u, but number "
2905 "of RX queues is %u\n",
2906 dev->name, index, dev->real_num_rx_queues);
2909 rxqueue = dev->_rx + index;
2913 map = rcu_dereference(rxqueue->rps_map);
2915 if (map->len == 1 &&
2916 !rcu_access_pointer(rxqueue->rps_flow_table)) {
2917 tcpu = map->cpus[0];
2918 if (cpu_online(tcpu))
2922 } else if (!rcu_access_pointer(rxqueue->rps_flow_table)) {
2926 skb_reset_network_header(skb);
2927 if (!skb_get_rxhash(skb))
2930 flow_table = rcu_dereference(rxqueue->rps_flow_table);
2931 sock_flow_table = rcu_dereference(rps_sock_flow_table);
2932 if (flow_table && sock_flow_table) {
2934 struct rps_dev_flow *rflow;
2936 rflow = &flow_table->flows[skb->rxhash & flow_table->mask];
2939 next_cpu = sock_flow_table->ents[skb->rxhash &
2940 sock_flow_table->mask];
2943 * If the desired CPU (where last recvmsg was done) is
2944 * different from current CPU (one in the rx-queue flow
2945 * table entry), switch if one of the following holds:
2946 * - Current CPU is unset (equal to RPS_NO_CPU).
2947 * - Current CPU is offline.
2948 * - The current CPU's queue tail has advanced beyond the
2949 * last packet that was enqueued using this table entry.
2950 * This guarantees that all previous packets for the flow
2951 * have been dequeued, thus preserving in order delivery.
2953 if (unlikely(tcpu != next_cpu) &&
2954 (tcpu == RPS_NO_CPU || !cpu_online(tcpu) ||
2955 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
2956 rflow->last_qtail)) >= 0)) {
2958 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
2961 if (tcpu != RPS_NO_CPU && cpu_online(tcpu)) {
2969 tcpu = map->cpus[((u64) skb->rxhash * map->len) >> 32];
2971 if (cpu_online(tcpu)) {
2981 #ifdef CONFIG_RFS_ACCEL
2984 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
2985 * @dev: Device on which the filter was set
2986 * @rxq_index: RX queue index
2987 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
2988 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
2990 * Drivers that implement ndo_rx_flow_steer() should periodically call
2991 * this function for each installed filter and remove the filters for
2992 * which it returns %true.
2994 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
2995 u32 flow_id, u16 filter_id)
2997 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
2998 struct rps_dev_flow_table *flow_table;
2999 struct rps_dev_flow *rflow;
3004 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3005 if (flow_table && flow_id <= flow_table->mask) {
3006 rflow = &flow_table->flows[flow_id];
3007 cpu = ACCESS_ONCE(rflow->cpu);
3008 if (rflow->filter == filter_id && cpu != RPS_NO_CPU &&
3009 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
3010 rflow->last_qtail) <
3011 (int)(10 * flow_table->mask)))
3017 EXPORT_SYMBOL(rps_may_expire_flow);
3019 #endif /* CONFIG_RFS_ACCEL */
3021 /* Called from hardirq (IPI) context */
3022 static void rps_trigger_softirq(void *data)
3024 struct softnet_data *sd = data;
3026 ____napi_schedule(sd, &sd->backlog);
3030 #endif /* CONFIG_RPS */
3033 * Check if this softnet_data structure is another cpu one
3034 * If yes, queue it to our IPI list and return 1
3037 static int rps_ipi_queued(struct softnet_data *sd)
3040 struct softnet_data *mysd = &__get_cpu_var(softnet_data);
3043 sd->rps_ipi_next = mysd->rps_ipi_list;
3044 mysd->rps_ipi_list = sd;
3046 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3049 #endif /* CONFIG_RPS */
3054 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
3055 * queue (may be a remote CPU queue).
3057 static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
3058 unsigned int *qtail)
3060 struct softnet_data *sd;
3061 unsigned long flags;
3063 sd = &per_cpu(softnet_data, cpu);
3065 local_irq_save(flags);
3068 if (skb_queue_len(&sd->input_pkt_queue) <= netdev_max_backlog) {
3069 if (skb_queue_len(&sd->input_pkt_queue)) {
3071 __skb_queue_tail(&sd->input_pkt_queue, skb);
3072 input_queue_tail_incr_save(sd, qtail);
3074 local_irq_restore(flags);
3075 return NET_RX_SUCCESS;
3078 /* Schedule NAPI for backlog device
3079 * We can use non atomic operation since we own the queue lock
3081 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
3082 if (!rps_ipi_queued(sd))
3083 ____napi_schedule(sd, &sd->backlog);
3091 local_irq_restore(flags);
3093 atomic_long_inc(&skb->dev->rx_dropped);
3099 * netif_rx - post buffer to the network code
3100 * @skb: buffer to post
3102 * This function receives a packet from a device driver and queues it for
3103 * the upper (protocol) levels to process. It always succeeds. The buffer
3104 * may be dropped during processing for congestion control or by the
3108 * NET_RX_SUCCESS (no congestion)
3109 * NET_RX_DROP (packet was dropped)
3113 int netif_rx(struct sk_buff *skb)
3117 /* if netpoll wants it, pretend we never saw it */
3118 if (netpoll_rx(skb))
3121 net_timestamp_check(netdev_tstamp_prequeue, skb);
3123 trace_netif_rx(skb);
3125 if (static_key_false(&rps_needed)) {
3126 struct rps_dev_flow voidflow, *rflow = &voidflow;
3132 cpu = get_rps_cpu(skb->dev, skb, &rflow);
3134 cpu = smp_processor_id();
3136 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
3144 ret = enqueue_to_backlog(skb, get_cpu(), &qtail);
3149 EXPORT_SYMBOL(netif_rx);
3151 int netif_rx_ni(struct sk_buff *skb)
3156 err = netif_rx(skb);
3157 if (local_softirq_pending())
3163 EXPORT_SYMBOL(netif_rx_ni);
3165 static void net_tx_action(struct softirq_action *h)
3167 struct softnet_data *sd = &__get_cpu_var(softnet_data);
3169 if (sd->completion_queue) {
3170 struct sk_buff *clist;
3172 local_irq_disable();
3173 clist = sd->completion_queue;
3174 sd->completion_queue = NULL;
3178 struct sk_buff *skb = clist;
3179 clist = clist->next;
3181 WARN_ON(atomic_read(&skb->users));
3182 trace_kfree_skb(skb, net_tx_action);
3187 if (sd->output_queue) {
3190 local_irq_disable();
3191 head = sd->output_queue;
3192 sd->output_queue = NULL;
3193 sd->output_queue_tailp = &sd->output_queue;
3197 struct Qdisc *q = head;
3198 spinlock_t *root_lock;
3200 head = head->next_sched;
3202 root_lock = qdisc_lock(q);
3203 if (spin_trylock(root_lock)) {
3204 smp_mb__before_clear_bit();
3205 clear_bit(__QDISC_STATE_SCHED,
3208 spin_unlock(root_lock);
3210 if (!test_bit(__QDISC_STATE_DEACTIVATED,
3212 __netif_reschedule(q);
3214 smp_mb__before_clear_bit();
3215 clear_bit(__QDISC_STATE_SCHED,
3223 #if (defined(CONFIG_BRIDGE) || defined(CONFIG_BRIDGE_MODULE)) && \
3224 (defined(CONFIG_ATM_LANE) || defined(CONFIG_ATM_LANE_MODULE))
3225 /* This hook is defined here for ATM LANE */
3226 int (*br_fdb_test_addr_hook)(struct net_device *dev,
3227 unsigned char *addr) __read_mostly;
3228 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook);
3231 #ifdef CONFIG_NET_CLS_ACT
3232 /* TODO: Maybe we should just force sch_ingress to be compiled in
3233 * when CONFIG_NET_CLS_ACT is? otherwise some useless instructions
3234 * a compare and 2 stores extra right now if we dont have it on
3235 * but have CONFIG_NET_CLS_ACT
3236 * NOTE: This doesn't stop any functionality; if you dont have
3237 * the ingress scheduler, you just can't add policies on ingress.
3240 static int ing_filter(struct sk_buff *skb, struct netdev_queue *rxq)
3242 struct net_device *dev = skb->dev;
3243 u32 ttl = G_TC_RTTL(skb->tc_verd);
3244 int result = TC_ACT_OK;
3247 if (unlikely(MAX_RED_LOOP < ttl++)) {
3248 net_warn_ratelimited("Redir loop detected Dropping packet (%d->%d)\n",
3249 skb->skb_iif, dev->ifindex);
3253 skb->tc_verd = SET_TC_RTTL(skb->tc_verd, ttl);
3254 skb->tc_verd = SET_TC_AT(skb->tc_verd, AT_INGRESS);
3257 if (q != &noop_qdisc) {
3258 spin_lock(qdisc_lock(q));
3259 if (likely(!test_bit(__QDISC_STATE_DEACTIVATED, &q->state)))
3260 result = qdisc_enqueue_root(skb, q);
3261 spin_unlock(qdisc_lock(q));
3267 static inline struct sk_buff *handle_ing(struct sk_buff *skb,
3268 struct packet_type **pt_prev,
3269 int *ret, struct net_device *orig_dev)
3271 struct netdev_queue *rxq = rcu_dereference(skb->dev->ingress_queue);
3273 if (!rxq || rxq->qdisc == &noop_qdisc)
3277 *ret = deliver_skb(skb, *pt_prev, orig_dev);
3281 switch (ing_filter(skb, rxq)) {
3295 * netdev_rx_handler_register - register receive handler
3296 * @dev: device to register a handler for
3297 * @rx_handler: receive handler to register
3298 * @rx_handler_data: data pointer that is used by rx handler
3300 * Register a receive hander for a device. This handler will then be
3301 * called from __netif_receive_skb. A negative errno code is returned
3304 * The caller must hold the rtnl_mutex.
3306 * For a general description of rx_handler, see enum rx_handler_result.
3308 int netdev_rx_handler_register(struct net_device *dev,
3309 rx_handler_func_t *rx_handler,
3310 void *rx_handler_data)
3314 if (dev->rx_handler)
3317 /* Note: rx_handler_data must be set before rx_handler */
3318 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data);
3319 rcu_assign_pointer(dev->rx_handler, rx_handler);
3323 EXPORT_SYMBOL_GPL(netdev_rx_handler_register);
3326 * netdev_rx_handler_unregister - unregister receive handler
3327 * @dev: device to unregister a handler from
3329 * Unregister a receive hander from a device.
3331 * The caller must hold the rtnl_mutex.
3333 void netdev_rx_handler_unregister(struct net_device *dev)
3337 RCU_INIT_POINTER(dev->rx_handler, NULL);
3338 /* a reader seeing a non NULL rx_handler in a rcu_read_lock()
3339 * section has a guarantee to see a non NULL rx_handler_data
3343 RCU_INIT_POINTER(dev->rx_handler_data, NULL);
3345 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister);
3348 * Limit the use of PFMEMALLOC reserves to those protocols that implement
3349 * the special handling of PFMEMALLOC skbs.
3351 static bool skb_pfmemalloc_protocol(struct sk_buff *skb)
3353 switch (skb->protocol) {
3354 case __constant_htons(ETH_P_ARP):
3355 case __constant_htons(ETH_P_IP):
3356 case __constant_htons(ETH_P_IPV6):
3357 case __constant_htons(ETH_P_8021Q):
3364 static int __netif_receive_skb_core(struct sk_buff *skb, bool pfmemalloc)
3366 struct packet_type *ptype, *pt_prev;
3367 rx_handler_func_t *rx_handler;
3368 struct net_device *orig_dev;
3369 struct net_device *null_or_dev;
3370 bool deliver_exact = false;
3371 int ret = NET_RX_DROP;
3374 net_timestamp_check(!netdev_tstamp_prequeue, skb);
3376 trace_netif_receive_skb(skb);
3378 /* if we've gotten here through NAPI, check netpoll */
3379 if (netpoll_receive_skb(skb))
3382 orig_dev = skb->dev;
3384 skb_reset_network_header(skb);
3385 if (!skb_transport_header_was_set(skb))
3386 skb_reset_transport_header(skb);
3387 skb_reset_mac_len(skb);
3394 skb->skb_iif = skb->dev->ifindex;
3396 __this_cpu_inc(softnet_data.processed);
3398 if (skb->protocol == cpu_to_be16(ETH_P_8021Q)) {
3399 skb = vlan_untag(skb);
3404 #ifdef CONFIG_NET_CLS_ACT
3405 if (skb->tc_verd & TC_NCLS) {
3406 skb->tc_verd = CLR_TC_NCLS(skb->tc_verd);
3414 list_for_each_entry_rcu(ptype, &ptype_all, list) {
3415 if (!ptype->dev || ptype->dev == skb->dev) {
3417 ret = deliver_skb(skb, pt_prev, orig_dev);
3423 #ifdef CONFIG_NET_CLS_ACT
3424 skb = handle_ing(skb, &pt_prev, &ret, orig_dev);
3430 if (pfmemalloc && !skb_pfmemalloc_protocol(skb))
3433 if (vlan_tx_tag_present(skb)) {
3435 ret = deliver_skb(skb, pt_prev, orig_dev);
3438 if (vlan_do_receive(&skb))
3440 else if (unlikely(!skb))
3444 rx_handler = rcu_dereference(skb->dev->rx_handler);
3447 ret = deliver_skb(skb, pt_prev, orig_dev);
3450 switch (rx_handler(&skb)) {
3451 case RX_HANDLER_CONSUMED:
3452 ret = NET_RX_SUCCESS;
3454 case RX_HANDLER_ANOTHER:
3456 case RX_HANDLER_EXACT:
3457 deliver_exact = true;
3458 case RX_HANDLER_PASS:
3465 if (vlan_tx_nonzero_tag_present(skb))
3466 skb->pkt_type = PACKET_OTHERHOST;
3468 /* deliver only exact match when indicated */
3469 null_or_dev = deliver_exact ? skb->dev : NULL;
3471 type = skb->protocol;
3472 list_for_each_entry_rcu(ptype,
3473 &ptype_base[ntohs(type) & PTYPE_HASH_MASK], list) {
3474 if (ptype->type == type &&
3475 (ptype->dev == null_or_dev || ptype->dev == skb->dev ||
3476 ptype->dev == orig_dev)) {
3478 ret = deliver_skb(skb, pt_prev, orig_dev);
3484 if (unlikely(skb_orphan_frags(skb, GFP_ATOMIC)))
3487 ret = pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
3490 atomic_long_inc(&skb->dev->rx_dropped);
3492 /* Jamal, now you will not able to escape explaining
3493 * me how you were going to use this. :-)
3504 static int __netif_receive_skb(struct sk_buff *skb)
3508 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) {
3509 unsigned long pflags = current->flags;
3512 * PFMEMALLOC skbs are special, they should
3513 * - be delivered to SOCK_MEMALLOC sockets only
3514 * - stay away from userspace
3515 * - have bounded memory usage
3517 * Use PF_MEMALLOC as this saves us from propagating the allocation
3518 * context down to all allocation sites.
3520 current->flags |= PF_MEMALLOC;
3521 ret = __netif_receive_skb_core(skb, true);
3522 tsk_restore_flags(current, pflags, PF_MEMALLOC);
3524 ret = __netif_receive_skb_core(skb, false);
3530 * netif_receive_skb - process receive buffer from network
3531 * @skb: buffer to process
3533 * netif_receive_skb() is the main receive data processing function.
3534 * It always succeeds. The buffer may be dropped during processing
3535 * for congestion control or by the protocol layers.
3537 * This function may only be called from softirq context and interrupts
3538 * should be enabled.
3540 * Return values (usually ignored):
3541 * NET_RX_SUCCESS: no congestion
3542 * NET_RX_DROP: packet was dropped
3544 int netif_receive_skb(struct sk_buff *skb)
3546 net_timestamp_check(netdev_tstamp_prequeue, skb);
3548 if (skb_defer_rx_timestamp(skb))
3549 return NET_RX_SUCCESS;
3552 if (static_key_false(&rps_needed)) {
3553 struct rps_dev_flow voidflow, *rflow = &voidflow;
3558 cpu = get_rps_cpu(skb->dev, skb, &rflow);
3561 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
3568 return __netif_receive_skb(skb);
3570 EXPORT_SYMBOL(netif_receive_skb);
3572 /* Network device is going away, flush any packets still pending
3573 * Called with irqs disabled.
3575 static void flush_backlog(void *arg)
3577 struct net_device *dev = arg;
3578 struct softnet_data *sd = &__get_cpu_var(softnet_data);
3579 struct sk_buff *skb, *tmp;
3582 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) {
3583 if (skb->dev == dev) {
3584 __skb_unlink(skb, &sd->input_pkt_queue);
3586 input_queue_head_incr(sd);
3591 skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
3592 if (skb->dev == dev) {
3593 __skb_unlink(skb, &sd->process_queue);
3595 input_queue_head_incr(sd);
3600 static int napi_gro_complete(struct sk_buff *skb)
3602 struct packet_offload *ptype;
3603 __be16 type = skb->protocol;
3604 struct list_head *head = &offload_base;
3607 BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
3609 if (NAPI_GRO_CB(skb)->count == 1) {
3610 skb_shinfo(skb)->gso_size = 0;
3615 list_for_each_entry_rcu(ptype, head, list) {
3616 if (ptype->type != type || !ptype->callbacks.gro_complete)
3619 err = ptype->callbacks.gro_complete(skb);
3625 WARN_ON(&ptype->list == head);
3627 return NET_RX_SUCCESS;
3631 return netif_receive_skb(skb);
3634 /* napi->gro_list contains packets ordered by age.
3635 * youngest packets at the head of it.
3636 * Complete skbs in reverse order to reduce latencies.
3638 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
3640 struct sk_buff *skb, *prev = NULL;
3642 /* scan list and build reverse chain */
3643 for (skb = napi->gro_list; skb != NULL; skb = skb->next) {
3648 for (skb = prev; skb; skb = prev) {
3651 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
3655 napi_gro_complete(skb);
3659 napi->gro_list = NULL;
3661 EXPORT_SYMBOL(napi_gro_flush);
3663 static void gro_list_prepare(struct napi_struct *napi, struct sk_buff *skb)
3666 unsigned int maclen = skb->dev->hard_header_len;
3668 for (p = napi->gro_list; p; p = p->next) {
3669 unsigned long diffs;
3671 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
3672 diffs |= p->vlan_tci ^ skb->vlan_tci;
3673 if (maclen == ETH_HLEN)
3674 diffs |= compare_ether_header(skb_mac_header(p),
3675 skb_gro_mac_header(skb));
3677 diffs = memcmp(skb_mac_header(p),
3678 skb_gro_mac_header(skb),
3680 NAPI_GRO_CB(p)->same_flow = !diffs;
3681 NAPI_GRO_CB(p)->flush = 0;
3685 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
3687 struct sk_buff **pp = NULL;
3688 struct packet_offload *ptype;
3689 __be16 type = skb->protocol;
3690 struct list_head *head = &offload_base;
3692 enum gro_result ret;
3694 if (!(skb->dev->features & NETIF_F_GRO) || netpoll_rx_on(skb))
3697 if (skb_is_gso(skb) || skb_has_frag_list(skb))
3700 gro_list_prepare(napi, skb);
3703 list_for_each_entry_rcu(ptype, head, list) {
3704 if (ptype->type != type || !ptype->callbacks.gro_receive)
3707 skb_set_network_header(skb, skb_gro_offset(skb));
3708 skb_reset_mac_len(skb);
3709 NAPI_GRO_CB(skb)->same_flow = 0;
3710 NAPI_GRO_CB(skb)->flush = 0;
3711 NAPI_GRO_CB(skb)->free = 0;
3713 pp = ptype->callbacks.gro_receive(&napi->gro_list, skb);
3718 if (&ptype->list == head)
3721 same_flow = NAPI_GRO_CB(skb)->same_flow;
3722 ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
3725 struct sk_buff *nskb = *pp;
3729 napi_gro_complete(nskb);
3736 if (NAPI_GRO_CB(skb)->flush || napi->gro_count >= MAX_GRO_SKBS)
3740 NAPI_GRO_CB(skb)->count = 1;
3741 NAPI_GRO_CB(skb)->age = jiffies;
3742 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
3743 skb->next = napi->gro_list;
3744 napi->gro_list = skb;
3748 if (skb_headlen(skb) < skb_gro_offset(skb)) {
3749 int grow = skb_gro_offset(skb) - skb_headlen(skb);
3751 BUG_ON(skb->end - skb->tail < grow);
3753 memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
3756 skb->data_len -= grow;
3758 skb_shinfo(skb)->frags[0].page_offset += grow;
3759 skb_frag_size_sub(&skb_shinfo(skb)->frags[0], grow);
3761 if (unlikely(!skb_frag_size(&skb_shinfo(skb)->frags[0]))) {
3762 skb_frag_unref(skb, 0);
3763 memmove(skb_shinfo(skb)->frags,
3764 skb_shinfo(skb)->frags + 1,
3765 --skb_shinfo(skb)->nr_frags * sizeof(skb_frag_t));
3778 static gro_result_t napi_skb_finish(gro_result_t ret, struct sk_buff *skb)
3782 if (netif_receive_skb(skb))
3790 case GRO_MERGED_FREE:
3791 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
3792 kmem_cache_free(skbuff_head_cache, skb);
3805 static void skb_gro_reset_offset(struct sk_buff *skb)
3807 const struct skb_shared_info *pinfo = skb_shinfo(skb);
3808 const skb_frag_t *frag0 = &pinfo->frags[0];
3810 NAPI_GRO_CB(skb)->data_offset = 0;
3811 NAPI_GRO_CB(skb)->frag0 = NULL;
3812 NAPI_GRO_CB(skb)->frag0_len = 0;
3814 if (skb->mac_header == skb->tail &&
3816 !PageHighMem(skb_frag_page(frag0))) {
3817 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
3818 NAPI_GRO_CB(skb)->frag0_len = skb_frag_size(frag0);
3822 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
3824 skb_gro_reset_offset(skb);
3826 return napi_skb_finish(dev_gro_receive(napi, skb), skb);
3828 EXPORT_SYMBOL(napi_gro_receive);
3830 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
3832 __skb_pull(skb, skb_headlen(skb));
3833 /* restore the reserve we had after netdev_alloc_skb_ip_align() */
3834 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
3836 skb->dev = napi->dev;
3842 struct sk_buff *napi_get_frags(struct napi_struct *napi)
3844 struct sk_buff *skb = napi->skb;
3847 skb = netdev_alloc_skb_ip_align(napi->dev, GRO_MAX_HEAD);
3853 EXPORT_SYMBOL(napi_get_frags);
3855 static gro_result_t napi_frags_finish(struct napi_struct *napi, struct sk_buff *skb,
3861 skb->protocol = eth_type_trans(skb, skb->dev);
3863 if (ret == GRO_HELD)
3864 skb_gro_pull(skb, -ETH_HLEN);
3865 else if (netif_receive_skb(skb))
3870 case GRO_MERGED_FREE:
3871 napi_reuse_skb(napi, skb);
3881 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
3883 struct sk_buff *skb = napi->skb;
3890 skb_reset_mac_header(skb);
3891 skb_gro_reset_offset(skb);
3893 off = skb_gro_offset(skb);
3894 hlen = off + sizeof(*eth);
3895 eth = skb_gro_header_fast(skb, off);
3896 if (skb_gro_header_hard(skb, hlen)) {
3897 eth = skb_gro_header_slow(skb, hlen, off);
3898 if (unlikely(!eth)) {
3899 napi_reuse_skb(napi, skb);
3905 skb_gro_pull(skb, sizeof(*eth));
3908 * This works because the only protocols we care about don't require
3909 * special handling. We'll fix it up properly at the end.
3911 skb->protocol = eth->h_proto;
3917 gro_result_t napi_gro_frags(struct napi_struct *napi)
3919 struct sk_buff *skb = napi_frags_skb(napi);
3924 return napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
3926 EXPORT_SYMBOL(napi_gro_frags);
3929 * net_rps_action sends any pending IPI's for rps.
3930 * Note: called with local irq disabled, but exits with local irq enabled.
3932 static void net_rps_action_and_irq_enable(struct softnet_data *sd)
3935 struct softnet_data *remsd = sd->rps_ipi_list;
3938 sd->rps_ipi_list = NULL;
3942 /* Send pending IPI's to kick RPS processing on remote cpus. */
3944 struct softnet_data *next = remsd->rps_ipi_next;
3946 if (cpu_online(remsd->cpu))
3947 __smp_call_function_single(remsd->cpu,
3956 static int process_backlog(struct napi_struct *napi, int quota)
3959 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
3962 /* Check if we have pending ipi, its better to send them now,
3963 * not waiting net_rx_action() end.
3965 if (sd->rps_ipi_list) {
3966 local_irq_disable();
3967 net_rps_action_and_irq_enable(sd);
3970 napi->weight = weight_p;
3971 local_irq_disable();
3972 while (work < quota) {
3973 struct sk_buff *skb;
3976 while ((skb = __skb_dequeue(&sd->process_queue))) {
3978 __netif_receive_skb(skb);
3979 local_irq_disable();
3980 input_queue_head_incr(sd);
3981 if (++work >= quota) {
3988 qlen = skb_queue_len(&sd->input_pkt_queue);
3990 skb_queue_splice_tail_init(&sd->input_pkt_queue,
3991 &sd->process_queue);
3993 if (qlen < quota - work) {
3995 * Inline a custom version of __napi_complete().
3996 * only current cpu owns and manipulates this napi,
3997 * and NAPI_STATE_SCHED is the only possible flag set on backlog.
3998 * we can use a plain write instead of clear_bit(),
3999 * and we dont need an smp_mb() memory barrier.
4001 list_del(&napi->poll_list);
4004 quota = work + qlen;
4014 * __napi_schedule - schedule for receive
4015 * @n: entry to schedule
4017 * The entry's receive function will be scheduled to run
4019 void __napi_schedule(struct napi_struct *n)
4021 unsigned long flags;
4023 local_irq_save(flags);
4024 ____napi_schedule(&__get_cpu_var(softnet_data), n);
4025 local_irq_restore(flags);
4027 EXPORT_SYMBOL(__napi_schedule);
4029 void __napi_complete(struct napi_struct *n)
4031 BUG_ON(!test_bit(NAPI_STATE_SCHED, &n->state));
4032 BUG_ON(n->gro_list);
4034 list_del(&n->poll_list);
4035 smp_mb__before_clear_bit();
4036 clear_bit(NAPI_STATE_SCHED, &n->state);
4038 EXPORT_SYMBOL(__napi_complete);
4040 void napi_complete(struct napi_struct *n)
4042 unsigned long flags;
4045 * don't let napi dequeue from the cpu poll list
4046 * just in case its running on a different cpu
4048 if (unlikely(test_bit(NAPI_STATE_NPSVC, &n->state)))
4051 napi_gro_flush(n, false);
4052 local_irq_save(flags);
4054 local_irq_restore(flags);
4056 EXPORT_SYMBOL(napi_complete);
4058 void netif_napi_add(struct net_device *dev, struct napi_struct *napi,
4059 int (*poll)(struct napi_struct *, int), int weight)
4061 INIT_LIST_HEAD(&napi->poll_list);
4062 napi->gro_count = 0;
4063 napi->gro_list = NULL;
4066 napi->weight = weight;
4067 list_add(&napi->dev_list, &dev->napi_list);
4069 #ifdef CONFIG_NETPOLL
4070 spin_lock_init(&napi->poll_lock);
4071 napi->poll_owner = -1;
4073 set_bit(NAPI_STATE_SCHED, &napi->state);
4075 EXPORT_SYMBOL(netif_napi_add);
4077 void netif_napi_del(struct napi_struct *napi)
4079 struct sk_buff *skb, *next;
4081 list_del_init(&napi->dev_list);
4082 napi_free_frags(napi);
4084 for (skb = napi->gro_list; skb; skb = next) {
4090 napi->gro_list = NULL;
4091 napi->gro_count = 0;
4093 EXPORT_SYMBOL(netif_napi_del);
4095 static void net_rx_action(struct softirq_action *h)
4097 struct softnet_data *sd = &__get_cpu_var(softnet_data);
4098 unsigned long time_limit = jiffies + 2;
4099 int budget = netdev_budget;
4102 local_irq_disable();
4104 while (!list_empty(&sd->poll_list)) {
4105 struct napi_struct *n;
4108 /* If softirq window is exhuasted then punt.
4109 * Allow this to run for 2 jiffies since which will allow
4110 * an average latency of 1.5/HZ.
4112 if (unlikely(budget <= 0 || time_after_eq(jiffies, time_limit)))
4117 /* Even though interrupts have been re-enabled, this
4118 * access is safe because interrupts can only add new
4119 * entries to the tail of this list, and only ->poll()
4120 * calls can remove this head entry from the list.
4122 n = list_first_entry(&sd->poll_list, struct napi_struct, poll_list);
4124 have = netpoll_poll_lock(n);
4128 /* This NAPI_STATE_SCHED test is for avoiding a race
4129 * with netpoll's poll_napi(). Only the entity which
4130 * obtains the lock and sees NAPI_STATE_SCHED set will
4131 * actually make the ->poll() call. Therefore we avoid
4132 * accidentally calling ->poll() when NAPI is not scheduled.
4135 if (test_bit(NAPI_STATE_SCHED, &n->state)) {
4136 work = n->poll(n, weight);
4140 WARN_ON_ONCE(work > weight);
4144 local_irq_disable();
4146 /* Drivers must not modify the NAPI state if they
4147 * consume the entire weight. In such cases this code
4148 * still "owns" the NAPI instance and therefore can
4149 * move the instance around on the list at-will.
4151 if (unlikely(work == weight)) {
4152 if (unlikely(napi_disable_pending(n))) {
4155 local_irq_disable();
4158 /* flush too old packets
4159 * If HZ < 1000, flush all packets.
4162 napi_gro_flush(n, HZ >= 1000);
4163 local_irq_disable();
4165 list_move_tail(&n->poll_list, &sd->poll_list);
4169 netpoll_poll_unlock(have);
4172 net_rps_action_and_irq_enable(sd);
4174 #ifdef CONFIG_NET_DMA
4176 * There may not be any more sk_buffs coming right now, so push
4177 * any pending DMA copies to hardware
4179 dma_issue_pending_all();
4186 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
4190 struct netdev_upper {
4191 struct net_device *dev;
4193 struct list_head list;
4194 struct rcu_head rcu;
4195 struct list_head search_list;
4198 static void __append_search_uppers(struct list_head *search_list,
4199 struct net_device *dev)
4201 struct netdev_upper *upper;
4203 list_for_each_entry(upper, &dev->upper_dev_list, list) {
4204 /* check if this upper is not already in search list */
4205 if (list_empty(&upper->search_list))
4206 list_add_tail(&upper->search_list, search_list);
4210 static bool __netdev_search_upper_dev(struct net_device *dev,
4211 struct net_device *upper_dev)
4213 LIST_HEAD(search_list);
4214 struct netdev_upper *upper;
4215 struct netdev_upper *tmp;
4218 __append_search_uppers(&search_list, dev);
4219 list_for_each_entry(upper, &search_list, search_list) {
4220 if (upper->dev == upper_dev) {
4224 __append_search_uppers(&search_list, upper->dev);
4226 list_for_each_entry_safe(upper, tmp, &search_list, search_list)
4227 INIT_LIST_HEAD(&upper->search_list);
4231 static struct netdev_upper *__netdev_find_upper(struct net_device *dev,
4232 struct net_device *upper_dev)
4234 struct netdev_upper *upper;
4236 list_for_each_entry(upper, &dev->upper_dev_list, list) {
4237 if (upper->dev == upper_dev)
4244 * netdev_has_upper_dev - Check if device is linked to an upper device
4246 * @upper_dev: upper device to check
4248 * Find out if a device is linked to specified upper device and return true
4249 * in case it is. Note that this checks only immediate upper device,
4250 * not through a complete stack of devices. The caller must hold the RTNL lock.
4252 bool netdev_has_upper_dev(struct net_device *dev,
4253 struct net_device *upper_dev)
4257 return __netdev_find_upper(dev, upper_dev);
4259 EXPORT_SYMBOL(netdev_has_upper_dev);
4262 * netdev_has_any_upper_dev - Check if device is linked to some device
4265 * Find out if a device is linked to an upper device and return true in case
4266 * it is. The caller must hold the RTNL lock.
4268 bool netdev_has_any_upper_dev(struct net_device *dev)
4272 return !list_empty(&dev->upper_dev_list);
4274 EXPORT_SYMBOL(netdev_has_any_upper_dev);
4277 * netdev_master_upper_dev_get - Get master upper device
4280 * Find a master upper device and return pointer to it or NULL in case
4281 * it's not there. The caller must hold the RTNL lock.
4283 struct net_device *netdev_master_upper_dev_get(struct net_device *dev)
4285 struct netdev_upper *upper;
4289 if (list_empty(&dev->upper_dev_list))
4292 upper = list_first_entry(&dev->upper_dev_list,
4293 struct netdev_upper, list);
4294 if (likely(upper->master))
4298 EXPORT_SYMBOL(netdev_master_upper_dev_get);
4301 * netdev_master_upper_dev_get_rcu - Get master upper device
4304 * Find a master upper device and return pointer to it or NULL in case
4305 * it's not there. The caller must hold the RCU read lock.
4307 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev)
4309 struct netdev_upper *upper;
4311 upper = list_first_or_null_rcu(&dev->upper_dev_list,
4312 struct netdev_upper, list);
4313 if (upper && likely(upper->master))
4317 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu);
4319 static int __netdev_upper_dev_link(struct net_device *dev,
4320 struct net_device *upper_dev, bool master)
4322 struct netdev_upper *upper;
4326 if (dev == upper_dev)
4329 /* To prevent loops, check if dev is not upper device to upper_dev. */
4330 if (__netdev_search_upper_dev(upper_dev, dev))
4333 if (__netdev_find_upper(dev, upper_dev))
4336 if (master && netdev_master_upper_dev_get(dev))
4339 upper = kmalloc(sizeof(*upper), GFP_KERNEL);
4343 upper->dev = upper_dev;
4344 upper->master = master;
4345 INIT_LIST_HEAD(&upper->search_list);
4347 /* Ensure that master upper link is always the first item in list. */
4349 list_add_rcu(&upper->list, &dev->upper_dev_list);
4351 list_add_tail_rcu(&upper->list, &dev->upper_dev_list);
4352 dev_hold(upper_dev);
4358 * netdev_upper_dev_link - Add a link to the upper device
4360 * @upper_dev: new upper device
4362 * Adds a link to device which is upper to this one. The caller must hold
4363 * the RTNL lock. On a failure a negative errno code is returned.
4364 * On success the reference counts are adjusted and the function
4367 int netdev_upper_dev_link(struct net_device *dev,
4368 struct net_device *upper_dev)
4370 return __netdev_upper_dev_link(dev, upper_dev, false);
4372 EXPORT_SYMBOL(netdev_upper_dev_link);
4375 * netdev_master_upper_dev_link - Add a master link to the upper device
4377 * @upper_dev: new upper device
4379 * Adds a link to device which is upper to this one. In this case, only
4380 * one master upper device can be linked, although other non-master devices
4381 * might be linked as well. The caller must hold the RTNL lock.
4382 * On a failure a negative errno code is returned. On success the reference
4383 * counts are adjusted and the function returns zero.
4385 int netdev_master_upper_dev_link(struct net_device *dev,
4386 struct net_device *upper_dev)
4388 return __netdev_upper_dev_link(dev, upper_dev, true);
4390 EXPORT_SYMBOL(netdev_master_upper_dev_link);
4393 * netdev_upper_dev_unlink - Removes a link to upper device
4395 * @upper_dev: new upper device
4397 * Removes a link to device which is upper to this one. The caller must hold
4400 void netdev_upper_dev_unlink(struct net_device *dev,
4401 struct net_device *upper_dev)
4403 struct netdev_upper *upper;
4407 upper = __netdev_find_upper(dev, upper_dev);
4410 list_del_rcu(&upper->list);
4412 kfree_rcu(upper, rcu);
4414 EXPORT_SYMBOL(netdev_upper_dev_unlink);
4416 static void dev_change_rx_flags(struct net_device *dev, int flags)
4418 const struct net_device_ops *ops = dev->netdev_ops;
4420 if ((dev->flags & IFF_UP) && ops->ndo_change_rx_flags)
4421 ops->ndo_change_rx_flags(dev, flags);
4424 static int __dev_set_promiscuity(struct net_device *dev, int inc)
4426 unsigned int old_flags = dev->flags;
4432 dev->flags |= IFF_PROMISC;
4433 dev->promiscuity += inc;
4434 if (dev->promiscuity == 0) {
4437 * If inc causes overflow, untouch promisc and return error.
4440 dev->flags &= ~IFF_PROMISC;
4442 dev->promiscuity -= inc;
4443 pr_warn("%s: promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n",
4448 if (dev->flags != old_flags) {
4449 pr_info("device %s %s promiscuous mode\n",
4451 dev->flags & IFF_PROMISC ? "entered" : "left");
4452 if (audit_enabled) {
4453 current_uid_gid(&uid, &gid);
4454 audit_log(current->audit_context, GFP_ATOMIC,
4455 AUDIT_ANOM_PROMISCUOUS,
4456 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u",
4457 dev->name, (dev->flags & IFF_PROMISC),
4458 (old_flags & IFF_PROMISC),
4459 from_kuid(&init_user_ns, audit_get_loginuid(current)),
4460 from_kuid(&init_user_ns, uid),
4461 from_kgid(&init_user_ns, gid),
4462 audit_get_sessionid(current));
4465 dev_change_rx_flags(dev, IFF_PROMISC);
4471 * dev_set_promiscuity - update promiscuity count on a device
4475 * Add or remove promiscuity from a device. While the count in the device
4476 * remains above zero the interface remains promiscuous. Once it hits zero
4477 * the device reverts back to normal filtering operation. A negative inc
4478 * value is used to drop promiscuity on the device.
4479 * Return 0 if successful or a negative errno code on error.
4481 int dev_set_promiscuity(struct net_device *dev, int inc)
4483 unsigned int old_flags = dev->flags;
4486 err = __dev_set_promiscuity(dev, inc);
4489 if (dev->flags != old_flags)
4490 dev_set_rx_mode(dev);
4493 EXPORT_SYMBOL(dev_set_promiscuity);
4496 * dev_set_allmulti - update allmulti count on a device
4500 * Add or remove reception of all multicast frames to a device. While the
4501 * count in the device remains above zero the interface remains listening
4502 * to all interfaces. Once it hits zero the device reverts back to normal
4503 * filtering operation. A negative @inc value is used to drop the counter
4504 * when releasing a resource needing all multicasts.
4505 * Return 0 if successful or a negative errno code on error.
4508 int dev_set_allmulti(struct net_device *dev, int inc)
4510 unsigned int old_flags = dev->flags;
4514 dev->flags |= IFF_ALLMULTI;
4515 dev->allmulti += inc;
4516 if (dev->allmulti == 0) {
4519 * If inc causes overflow, untouch allmulti and return error.
4522 dev->flags &= ~IFF_ALLMULTI;
4524 dev->allmulti -= inc;
4525 pr_warn("%s: allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n",
4530 if (dev->flags ^ old_flags) {
4531 dev_change_rx_flags(dev, IFF_ALLMULTI);
4532 dev_set_rx_mode(dev);
4536 EXPORT_SYMBOL(dev_set_allmulti);
4539 * Upload unicast and multicast address lists to device and
4540 * configure RX filtering. When the device doesn't support unicast
4541 * filtering it is put in promiscuous mode while unicast addresses
4544 void __dev_set_rx_mode(struct net_device *dev)
4546 const struct net_device_ops *ops = dev->netdev_ops;
4548 /* dev_open will call this function so the list will stay sane. */
4549 if (!(dev->flags&IFF_UP))
4552 if (!netif_device_present(dev))
4555 if (!(dev->priv_flags & IFF_UNICAST_FLT)) {
4556 /* Unicast addresses changes may only happen under the rtnl,
4557 * therefore calling __dev_set_promiscuity here is safe.
4559 if (!netdev_uc_empty(dev) && !dev->uc_promisc) {
4560 __dev_set_promiscuity(dev, 1);
4561 dev->uc_promisc = true;
4562 } else if (netdev_uc_empty(dev) && dev->uc_promisc) {
4563 __dev_set_promiscuity(dev, -1);
4564 dev->uc_promisc = false;
4568 if (ops->ndo_set_rx_mode)
4569 ops->ndo_set_rx_mode(dev);
4572 void dev_set_rx_mode(struct net_device *dev)
4574 netif_addr_lock_bh(dev);
4575 __dev_set_rx_mode(dev);
4576 netif_addr_unlock_bh(dev);
4580 * dev_get_flags - get flags reported to userspace
4583 * Get the combination of flag bits exported through APIs to userspace.
4585 unsigned int dev_get_flags(const struct net_device *dev)
4589 flags = (dev->flags & ~(IFF_PROMISC |
4594 (dev->gflags & (IFF_PROMISC |
4597 if (netif_running(dev)) {
4598 if (netif_oper_up(dev))
4599 flags |= IFF_RUNNING;
4600 if (netif_carrier_ok(dev))
4601 flags |= IFF_LOWER_UP;
4602 if (netif_dormant(dev))
4603 flags |= IFF_DORMANT;
4608 EXPORT_SYMBOL(dev_get_flags);
4610 int __dev_change_flags(struct net_device *dev, unsigned int flags)
4612 unsigned int old_flags = dev->flags;
4618 * Set the flags on our device.
4621 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP |
4622 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL |
4624 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC |
4628 * Load in the correct multicast list now the flags have changed.
4631 if ((old_flags ^ flags) & IFF_MULTICAST)
4632 dev_change_rx_flags(dev, IFF_MULTICAST);
4634 dev_set_rx_mode(dev);
4637 * Have we downed the interface. We handle IFF_UP ourselves
4638 * according to user attempts to set it, rather than blindly
4643 if ((old_flags ^ flags) & IFF_UP) { /* Bit is different ? */
4644 ret = ((old_flags & IFF_UP) ? __dev_close : __dev_open)(dev);
4647 dev_set_rx_mode(dev);
4650 if ((flags ^ dev->gflags) & IFF_PROMISC) {
4651 int inc = (flags & IFF_PROMISC) ? 1 : -1;
4653 dev->gflags ^= IFF_PROMISC;
4654 dev_set_promiscuity(dev, inc);
4657 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI
4658 is important. Some (broken) drivers set IFF_PROMISC, when
4659 IFF_ALLMULTI is requested not asking us and not reporting.
4661 if ((flags ^ dev->gflags) & IFF_ALLMULTI) {
4662 int inc = (flags & IFF_ALLMULTI) ? 1 : -1;
4664 dev->gflags ^= IFF_ALLMULTI;
4665 dev_set_allmulti(dev, inc);
4671 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags)
4673 unsigned int changes = dev->flags ^ old_flags;
4675 if (changes & IFF_UP) {
4676 if (dev->flags & IFF_UP)
4677 call_netdevice_notifiers(NETDEV_UP, dev);
4679 call_netdevice_notifiers(NETDEV_DOWN, dev);
4682 if (dev->flags & IFF_UP &&
4683 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE)))
4684 call_netdevice_notifiers(NETDEV_CHANGE, dev);
4688 * dev_change_flags - change device settings
4690 * @flags: device state flags
4692 * Change settings on device based state flags. The flags are
4693 * in the userspace exported format.
4695 int dev_change_flags(struct net_device *dev, unsigned int flags)
4698 unsigned int changes, old_flags = dev->flags;
4700 ret = __dev_change_flags(dev, flags);
4704 changes = old_flags ^ dev->flags;
4706 rtmsg_ifinfo(RTM_NEWLINK, dev, changes);
4708 __dev_notify_flags(dev, old_flags);
4711 EXPORT_SYMBOL(dev_change_flags);
4714 * dev_set_mtu - Change maximum transfer unit
4716 * @new_mtu: new transfer unit
4718 * Change the maximum transfer size of the network device.
4720 int dev_set_mtu(struct net_device *dev, int new_mtu)
4722 const struct net_device_ops *ops = dev->netdev_ops;
4725 if (new_mtu == dev->mtu)
4728 /* MTU must be positive. */
4732 if (!netif_device_present(dev))
4736 if (ops->ndo_change_mtu)
4737 err = ops->ndo_change_mtu(dev, new_mtu);
4742 call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
4745 EXPORT_SYMBOL(dev_set_mtu);
4748 * dev_set_group - Change group this device belongs to
4750 * @new_group: group this device should belong to
4752 void dev_set_group(struct net_device *dev, int new_group)
4754 dev->group = new_group;
4756 EXPORT_SYMBOL(dev_set_group);
4759 * dev_set_mac_address - Change Media Access Control Address
4763 * Change the hardware (MAC) address of the device
4765 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa)
4767 const struct net_device_ops *ops = dev->netdev_ops;
4770 if (!ops->ndo_set_mac_address)
4772 if (sa->sa_family != dev->type)
4774 if (!netif_device_present(dev))
4776 err = ops->ndo_set_mac_address(dev, sa);
4779 dev->addr_assign_type = NET_ADDR_SET;
4780 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
4781 add_device_randomness(dev->dev_addr, dev->addr_len);
4784 EXPORT_SYMBOL(dev_set_mac_address);
4787 * dev_change_carrier - Change device carrier
4789 * @new_carrier: new value
4791 * Change device carrier
4793 int dev_change_carrier(struct net_device *dev, bool new_carrier)
4795 const struct net_device_ops *ops = dev->netdev_ops;
4797 if (!ops->ndo_change_carrier)
4799 if (!netif_device_present(dev))
4801 return ops->ndo_change_carrier(dev, new_carrier);
4803 EXPORT_SYMBOL(dev_change_carrier);
4806 * dev_new_index - allocate an ifindex
4807 * @net: the applicable net namespace
4809 * Returns a suitable unique value for a new device interface
4810 * number. The caller must hold the rtnl semaphore or the
4811 * dev_base_lock to be sure it remains unique.
4813 static int dev_new_index(struct net *net)
4815 int ifindex = net->ifindex;
4819 if (!__dev_get_by_index(net, ifindex))
4820 return net->ifindex = ifindex;
4824 /* Delayed registration/unregisteration */
4825 static LIST_HEAD(net_todo_list);
4827 static void net_set_todo(struct net_device *dev)
4829 list_add_tail(&dev->todo_list, &net_todo_list);
4832 static void rollback_registered_many(struct list_head *head)
4834 struct net_device *dev, *tmp;
4836 BUG_ON(dev_boot_phase);
4839 list_for_each_entry_safe(dev, tmp, head, unreg_list) {
4840 /* Some devices call without registering
4841 * for initialization unwind. Remove those
4842 * devices and proceed with the remaining.
4844 if (dev->reg_state == NETREG_UNINITIALIZED) {
4845 pr_debug("unregister_netdevice: device %s/%p never was registered\n",
4849 list_del(&dev->unreg_list);
4852 dev->dismantle = true;
4853 BUG_ON(dev->reg_state != NETREG_REGISTERED);
4856 /* If device is running, close it first. */
4857 dev_close_many(head);
4859 list_for_each_entry(dev, head, unreg_list) {
4860 /* And unlink it from device chain. */
4861 unlist_netdevice(dev);
4863 dev->reg_state = NETREG_UNREGISTERING;
4868 list_for_each_entry(dev, head, unreg_list) {
4869 /* Shutdown queueing discipline. */
4873 /* Notify protocols, that we are about to destroy
4874 this device. They should clean all the things.
4876 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
4878 if (!dev->rtnl_link_ops ||
4879 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
4880 rtmsg_ifinfo(RTM_DELLINK, dev, ~0U);
4883 * Flush the unicast and multicast chains
4888 if (dev->netdev_ops->ndo_uninit)
4889 dev->netdev_ops->ndo_uninit(dev);
4891 /* Notifier chain MUST detach us all upper devices. */
4892 WARN_ON(netdev_has_any_upper_dev(dev));
4894 /* Remove entries from kobject tree */
4895 netdev_unregister_kobject(dev);
4897 /* Remove XPS queueing entries */
4898 netif_reset_xps_queues_gt(dev, 0);
4904 list_for_each_entry(dev, head, unreg_list)
4908 static void rollback_registered(struct net_device *dev)
4912 list_add(&dev->unreg_list, &single);
4913 rollback_registered_many(&single);
4917 static netdev_features_t netdev_fix_features(struct net_device *dev,
4918 netdev_features_t features)
4920 /* Fix illegal checksum combinations */
4921 if ((features & NETIF_F_HW_CSUM) &&
4922 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
4923 netdev_warn(dev, "mixed HW and IP checksum settings.\n");
4924 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
4927 /* Fix illegal SG+CSUM combinations. */
4928 if ((features & NETIF_F_SG) &&
4929 !(features & NETIF_F_ALL_CSUM)) {
4931 "Dropping NETIF_F_SG since no checksum feature.\n");
4932 features &= ~NETIF_F_SG;
4935 /* TSO requires that SG is present as well. */
4936 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) {
4937 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n");
4938 features &= ~NETIF_F_ALL_TSO;
4941 /* TSO ECN requires that TSO is present as well. */
4942 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN)
4943 features &= ~NETIF_F_TSO_ECN;
4945 /* Software GSO depends on SG. */
4946 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) {
4947 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n");
4948 features &= ~NETIF_F_GSO;
4951 /* UFO needs SG and checksumming */
4952 if (features & NETIF_F_UFO) {
4953 /* maybe split UFO into V4 and V6? */
4954 if (!((features & NETIF_F_GEN_CSUM) ||
4955 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))
4956 == (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
4958 "Dropping NETIF_F_UFO since no checksum offload features.\n");
4959 features &= ~NETIF_F_UFO;
4962 if (!(features & NETIF_F_SG)) {
4964 "Dropping NETIF_F_UFO since no NETIF_F_SG feature.\n");
4965 features &= ~NETIF_F_UFO;
4972 int __netdev_update_features(struct net_device *dev)
4974 netdev_features_t features;
4979 features = netdev_get_wanted_features(dev);
4981 if (dev->netdev_ops->ndo_fix_features)
4982 features = dev->netdev_ops->ndo_fix_features(dev, features);
4984 /* driver might be less strict about feature dependencies */
4985 features = netdev_fix_features(dev, features);
4987 if (dev->features == features)
4990 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n",
4991 &dev->features, &features);
4993 if (dev->netdev_ops->ndo_set_features)
4994 err = dev->netdev_ops->ndo_set_features(dev, features);
4996 if (unlikely(err < 0)) {
4998 "set_features() failed (%d); wanted %pNF, left %pNF\n",
4999 err, &features, &dev->features);
5004 dev->features = features;
5010 * netdev_update_features - recalculate device features
5011 * @dev: the device to check
5013 * Recalculate dev->features set and send notifications if it
5014 * has changed. Should be called after driver or hardware dependent
5015 * conditions might have changed that influence the features.
5017 void netdev_update_features(struct net_device *dev)
5019 if (__netdev_update_features(dev))
5020 netdev_features_change(dev);
5022 EXPORT_SYMBOL(netdev_update_features);
5025 * netdev_change_features - recalculate device features
5026 * @dev: the device to check
5028 * Recalculate dev->features set and send notifications even
5029 * if they have not changed. Should be called instead of
5030 * netdev_update_features() if also dev->vlan_features might
5031 * have changed to allow the changes to be propagated to stacked
5034 void netdev_change_features(struct net_device *dev)
5036 __netdev_update_features(dev);
5037 netdev_features_change(dev);
5039 EXPORT_SYMBOL(netdev_change_features);
5042 * netif_stacked_transfer_operstate - transfer operstate
5043 * @rootdev: the root or lower level device to transfer state from
5044 * @dev: the device to transfer operstate to
5046 * Transfer operational state from root to device. This is normally
5047 * called when a stacking relationship exists between the root
5048 * device and the device(a leaf device).
5050 void netif_stacked_transfer_operstate(const struct net_device *rootdev,
5051 struct net_device *dev)
5053 if (rootdev->operstate == IF_OPER_DORMANT)
5054 netif_dormant_on(dev);
5056 netif_dormant_off(dev);
5058 if (netif_carrier_ok(rootdev)) {
5059 if (!netif_carrier_ok(dev))
5060 netif_carrier_on(dev);
5062 if (netif_carrier_ok(dev))
5063 netif_carrier_off(dev);
5066 EXPORT_SYMBOL(netif_stacked_transfer_operstate);
5069 static int netif_alloc_rx_queues(struct net_device *dev)
5071 unsigned int i, count = dev->num_rx_queues;
5072 struct netdev_rx_queue *rx;
5076 rx = kcalloc(count, sizeof(struct netdev_rx_queue), GFP_KERNEL);
5082 for (i = 0; i < count; i++)
5088 static void netdev_init_one_queue(struct net_device *dev,
5089 struct netdev_queue *queue, void *_unused)
5091 /* Initialize queue lock */
5092 spin_lock_init(&queue->_xmit_lock);
5093 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type);
5094 queue->xmit_lock_owner = -1;
5095 netdev_queue_numa_node_write(queue, NUMA_NO_NODE);
5098 dql_init(&queue->dql, HZ);
5102 static int netif_alloc_netdev_queues(struct net_device *dev)
5104 unsigned int count = dev->num_tx_queues;
5105 struct netdev_queue *tx;
5109 tx = kcalloc(count, sizeof(struct netdev_queue), GFP_KERNEL);
5115 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL);
5116 spin_lock_init(&dev->tx_global_lock);
5122 * register_netdevice - register a network device
5123 * @dev: device to register
5125 * Take a completed network device structure and add it to the kernel
5126 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
5127 * chain. 0 is returned on success. A negative errno code is returned
5128 * on a failure to set up the device, or if the name is a duplicate.
5130 * Callers must hold the rtnl semaphore. You may want
5131 * register_netdev() instead of this.
5134 * The locking appears insufficient to guarantee two parallel registers
5135 * will not get the same name.
5138 int register_netdevice(struct net_device *dev)
5141 struct net *net = dev_net(dev);
5143 BUG_ON(dev_boot_phase);
5148 /* When net_device's are persistent, this will be fatal. */
5149 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
5152 spin_lock_init(&dev->addr_list_lock);
5153 netdev_set_addr_lockdep_class(dev);
5157 ret = dev_get_valid_name(net, dev, dev->name);
5161 /* Init, if this function is available */
5162 if (dev->netdev_ops->ndo_init) {
5163 ret = dev->netdev_ops->ndo_init(dev);
5171 if (((dev->hw_features | dev->features) & NETIF_F_HW_VLAN_FILTER) &&
5172 (!dev->netdev_ops->ndo_vlan_rx_add_vid ||
5173 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) {
5174 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n");
5181 dev->ifindex = dev_new_index(net);
5182 else if (__dev_get_by_index(net, dev->ifindex))
5185 if (dev->iflink == -1)
5186 dev->iflink = dev->ifindex;
5188 /* Transfer changeable features to wanted_features and enable
5189 * software offloads (GSO and GRO).
5191 dev->hw_features |= NETIF_F_SOFT_FEATURES;
5192 dev->features |= NETIF_F_SOFT_FEATURES;
5193 dev->wanted_features = dev->features & dev->hw_features;
5195 /* Turn on no cache copy if HW is doing checksum */
5196 if (!(dev->flags & IFF_LOOPBACK)) {
5197 dev->hw_features |= NETIF_F_NOCACHE_COPY;
5198 if (dev->features & NETIF_F_ALL_CSUM) {
5199 dev->wanted_features |= NETIF_F_NOCACHE_COPY;
5200 dev->features |= NETIF_F_NOCACHE_COPY;
5204 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices.
5206 dev->vlan_features |= NETIF_F_HIGHDMA;
5208 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev);
5209 ret = notifier_to_errno(ret);
5213 ret = netdev_register_kobject(dev);
5216 dev->reg_state = NETREG_REGISTERED;
5218 __netdev_update_features(dev);
5221 * Default initial state at registry is that the
5222 * device is present.
5225 set_bit(__LINK_STATE_PRESENT, &dev->state);
5227 linkwatch_init_dev(dev);
5229 dev_init_scheduler(dev);
5231 list_netdevice(dev);
5232 add_device_randomness(dev->dev_addr, dev->addr_len);
5234 /* If the device has permanent device address, driver should
5235 * set dev_addr and also addr_assign_type should be set to
5236 * NET_ADDR_PERM (default value).
5238 if (dev->addr_assign_type == NET_ADDR_PERM)
5239 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len);
5241 /* Notify protocols, that a new device appeared. */
5242 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
5243 ret = notifier_to_errno(ret);
5245 rollback_registered(dev);
5246 dev->reg_state = NETREG_UNREGISTERED;
5249 * Prevent userspace races by waiting until the network
5250 * device is fully setup before sending notifications.
5252 if (!dev->rtnl_link_ops ||
5253 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
5254 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U);
5260 if (dev->netdev_ops->ndo_uninit)
5261 dev->netdev_ops->ndo_uninit(dev);
5264 EXPORT_SYMBOL(register_netdevice);
5267 * init_dummy_netdev - init a dummy network device for NAPI
5268 * @dev: device to init
5270 * This takes a network device structure and initialize the minimum
5271 * amount of fields so it can be used to schedule NAPI polls without
5272 * registering a full blown interface. This is to be used by drivers
5273 * that need to tie several hardware interfaces to a single NAPI
5274 * poll scheduler due to HW limitations.
5276 int init_dummy_netdev(struct net_device *dev)
5278 /* Clear everything. Note we don't initialize spinlocks
5279 * are they aren't supposed to be taken by any of the
5280 * NAPI code and this dummy netdev is supposed to be
5281 * only ever used for NAPI polls
5283 memset(dev, 0, sizeof(struct net_device));
5285 /* make sure we BUG if trying to hit standard
5286 * register/unregister code path
5288 dev->reg_state = NETREG_DUMMY;
5290 /* NAPI wants this */
5291 INIT_LIST_HEAD(&dev->napi_list);
5293 /* a dummy interface is started by default */
5294 set_bit(__LINK_STATE_PRESENT, &dev->state);
5295 set_bit(__LINK_STATE_START, &dev->state);
5297 /* Note : We dont allocate pcpu_refcnt for dummy devices,
5298 * because users of this 'device' dont need to change
5304 EXPORT_SYMBOL_GPL(init_dummy_netdev);
5308 * register_netdev - register a network device
5309 * @dev: device to register
5311 * Take a completed network device structure and add it to the kernel
5312 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
5313 * chain. 0 is returned on success. A negative errno code is returned
5314 * on a failure to set up the device, or if the name is a duplicate.
5316 * This is a wrapper around register_netdevice that takes the rtnl semaphore
5317 * and expands the device name if you passed a format string to
5320 int register_netdev(struct net_device *dev)
5325 err = register_netdevice(dev);
5329 EXPORT_SYMBOL(register_netdev);
5331 int netdev_refcnt_read(const struct net_device *dev)
5335 for_each_possible_cpu(i)
5336 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
5339 EXPORT_SYMBOL(netdev_refcnt_read);
5342 * netdev_wait_allrefs - wait until all references are gone.
5343 * @dev: target net_device
5345 * This is called when unregistering network devices.
5347 * Any protocol or device that holds a reference should register
5348 * for netdevice notification, and cleanup and put back the
5349 * reference if they receive an UNREGISTER event.
5350 * We can get stuck here if buggy protocols don't correctly
5353 static void netdev_wait_allrefs(struct net_device *dev)
5355 unsigned long rebroadcast_time, warning_time;
5358 linkwatch_forget_dev(dev);
5360 rebroadcast_time = warning_time = jiffies;
5361 refcnt = netdev_refcnt_read(dev);
5363 while (refcnt != 0) {
5364 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
5367 /* Rebroadcast unregister notification */
5368 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
5374 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
5375 if (test_bit(__LINK_STATE_LINKWATCH_PENDING,
5377 /* We must not have linkwatch events
5378 * pending on unregister. If this
5379 * happens, we simply run the queue
5380 * unscheduled, resulting in a noop
5383 linkwatch_run_queue();
5388 rebroadcast_time = jiffies;
5393 refcnt = netdev_refcnt_read(dev);
5395 if (time_after(jiffies, warning_time + 10 * HZ)) {
5396 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n",
5398 warning_time = jiffies;
5407 * register_netdevice(x1);
5408 * register_netdevice(x2);
5410 * unregister_netdevice(y1);
5411 * unregister_netdevice(y2);
5417 * We are invoked by rtnl_unlock().
5418 * This allows us to deal with problems:
5419 * 1) We can delete sysfs objects which invoke hotplug
5420 * without deadlocking with linkwatch via keventd.
5421 * 2) Since we run with the RTNL semaphore not held, we can sleep
5422 * safely in order to wait for the netdev refcnt to drop to zero.
5424 * We must not return until all unregister events added during
5425 * the interval the lock was held have been completed.
5427 void netdev_run_todo(void)
5429 struct list_head list;
5431 /* Snapshot list, allow later requests */
5432 list_replace_init(&net_todo_list, &list);
5437 /* Wait for rcu callbacks to finish before next phase */
5438 if (!list_empty(&list))
5441 while (!list_empty(&list)) {
5442 struct net_device *dev
5443 = list_first_entry(&list, struct net_device, todo_list);
5444 list_del(&dev->todo_list);
5447 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
5450 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) {
5451 pr_err("network todo '%s' but state %d\n",
5452 dev->name, dev->reg_state);
5457 dev->reg_state = NETREG_UNREGISTERED;
5459 on_each_cpu(flush_backlog, dev, 1);
5461 netdev_wait_allrefs(dev);
5464 BUG_ON(netdev_refcnt_read(dev));
5465 WARN_ON(rcu_access_pointer(dev->ip_ptr));
5466 WARN_ON(rcu_access_pointer(dev->ip6_ptr));
5467 WARN_ON(dev->dn_ptr);
5469 if (dev->destructor)
5470 dev->destructor(dev);
5472 /* Free network device */
5473 kobject_put(&dev->dev.kobj);
5477 /* Convert net_device_stats to rtnl_link_stats64. They have the same
5478 * fields in the same order, with only the type differing.
5480 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
5481 const struct net_device_stats *netdev_stats)
5483 #if BITS_PER_LONG == 64
5484 BUILD_BUG_ON(sizeof(*stats64) != sizeof(*netdev_stats));
5485 memcpy(stats64, netdev_stats, sizeof(*stats64));
5487 size_t i, n = sizeof(*stats64) / sizeof(u64);
5488 const unsigned long *src = (const unsigned long *)netdev_stats;
5489 u64 *dst = (u64 *)stats64;
5491 BUILD_BUG_ON(sizeof(*netdev_stats) / sizeof(unsigned long) !=
5492 sizeof(*stats64) / sizeof(u64));
5493 for (i = 0; i < n; i++)
5497 EXPORT_SYMBOL(netdev_stats_to_stats64);
5500 * dev_get_stats - get network device statistics
5501 * @dev: device to get statistics from
5502 * @storage: place to store stats
5504 * Get network statistics from device. Return @storage.
5505 * The device driver may provide its own method by setting
5506 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats;
5507 * otherwise the internal statistics structure is used.
5509 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev,
5510 struct rtnl_link_stats64 *storage)
5512 const struct net_device_ops *ops = dev->netdev_ops;
5514 if (ops->ndo_get_stats64) {
5515 memset(storage, 0, sizeof(*storage));
5516 ops->ndo_get_stats64(dev, storage);
5517 } else if (ops->ndo_get_stats) {
5518 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev));
5520 netdev_stats_to_stats64(storage, &dev->stats);
5522 storage->rx_dropped += atomic_long_read(&dev->rx_dropped);
5525 EXPORT_SYMBOL(dev_get_stats);
5527 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev)
5529 struct netdev_queue *queue = dev_ingress_queue(dev);
5531 #ifdef CONFIG_NET_CLS_ACT
5534 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
5537 netdev_init_one_queue(dev, queue, NULL);
5538 queue->qdisc = &noop_qdisc;
5539 queue->qdisc_sleeping = &noop_qdisc;
5540 rcu_assign_pointer(dev->ingress_queue, queue);
5545 static const struct ethtool_ops default_ethtool_ops;
5547 void netdev_set_default_ethtool_ops(struct net_device *dev,
5548 const struct ethtool_ops *ops)
5550 if (dev->ethtool_ops == &default_ethtool_ops)
5551 dev->ethtool_ops = ops;
5553 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops);
5556 * alloc_netdev_mqs - allocate network device
5557 * @sizeof_priv: size of private data to allocate space for
5558 * @name: device name format string
5559 * @setup: callback to initialize device
5560 * @txqs: the number of TX subqueues to allocate
5561 * @rxqs: the number of RX subqueues to allocate
5563 * Allocates a struct net_device with private data area for driver use
5564 * and performs basic initialization. Also allocates subquue structs
5565 * for each queue on the device.
5567 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
5568 void (*setup)(struct net_device *),
5569 unsigned int txqs, unsigned int rxqs)
5571 struct net_device *dev;
5573 struct net_device *p;
5575 BUG_ON(strlen(name) >= sizeof(dev->name));
5578 pr_err("alloc_netdev: Unable to allocate device with zero queues\n");
5584 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n");
5589 alloc_size = sizeof(struct net_device);
5591 /* ensure 32-byte alignment of private area */
5592 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
5593 alloc_size += sizeof_priv;
5595 /* ensure 32-byte alignment of whole construct */
5596 alloc_size += NETDEV_ALIGN - 1;
5598 p = kzalloc(alloc_size, GFP_KERNEL);
5602 dev = PTR_ALIGN(p, NETDEV_ALIGN);
5603 dev->padded = (char *)dev - (char *)p;
5605 dev->pcpu_refcnt = alloc_percpu(int);
5606 if (!dev->pcpu_refcnt)
5609 if (dev_addr_init(dev))
5615 dev_net_set(dev, &init_net);
5617 dev->gso_max_size = GSO_MAX_SIZE;
5618 dev->gso_max_segs = GSO_MAX_SEGS;
5620 INIT_LIST_HEAD(&dev->napi_list);
5621 INIT_LIST_HEAD(&dev->unreg_list);
5622 INIT_LIST_HEAD(&dev->link_watch_list);
5623 INIT_LIST_HEAD(&dev->upper_dev_list);
5624 dev->priv_flags = IFF_XMIT_DST_RELEASE;
5627 dev->num_tx_queues = txqs;
5628 dev->real_num_tx_queues = txqs;
5629 if (netif_alloc_netdev_queues(dev))
5633 dev->num_rx_queues = rxqs;
5634 dev->real_num_rx_queues = rxqs;
5635 if (netif_alloc_rx_queues(dev))
5639 strcpy(dev->name, name);
5640 dev->group = INIT_NETDEV_GROUP;
5641 if (!dev->ethtool_ops)
5642 dev->ethtool_ops = &default_ethtool_ops;
5650 free_percpu(dev->pcpu_refcnt);
5660 EXPORT_SYMBOL(alloc_netdev_mqs);
5663 * free_netdev - free network device
5666 * This function does the last stage of destroying an allocated device
5667 * interface. The reference to the device object is released.
5668 * If this is the last reference then it will be freed.
5670 void free_netdev(struct net_device *dev)
5672 struct napi_struct *p, *n;
5674 release_net(dev_net(dev));
5681 kfree(rcu_dereference_protected(dev->ingress_queue, 1));
5683 /* Flush device addresses */
5684 dev_addr_flush(dev);
5686 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
5689 free_percpu(dev->pcpu_refcnt);
5690 dev->pcpu_refcnt = NULL;
5692 /* Compatibility with error handling in drivers */
5693 if (dev->reg_state == NETREG_UNINITIALIZED) {
5694 kfree((char *)dev - dev->padded);
5698 BUG_ON(dev->reg_state != NETREG_UNREGISTERED);
5699 dev->reg_state = NETREG_RELEASED;
5701 /* will free via device release */
5702 put_device(&dev->dev);
5704 EXPORT_SYMBOL(free_netdev);
5707 * synchronize_net - Synchronize with packet receive processing
5709 * Wait for packets currently being received to be done.
5710 * Does not block later packets from starting.
5712 void synchronize_net(void)
5715 if (rtnl_is_locked())
5716 synchronize_rcu_expedited();
5720 EXPORT_SYMBOL(synchronize_net);
5723 * unregister_netdevice_queue - remove device from the kernel
5727 * This function shuts down a device interface and removes it
5728 * from the kernel tables.
5729 * If head not NULL, device is queued to be unregistered later.
5731 * Callers must hold the rtnl semaphore. You may want
5732 * unregister_netdev() instead of this.
5735 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head)
5740 list_move_tail(&dev->unreg_list, head);
5742 rollback_registered(dev);
5743 /* Finish processing unregister after unlock */
5747 EXPORT_SYMBOL(unregister_netdevice_queue);
5750 * unregister_netdevice_many - unregister many devices
5751 * @head: list of devices
5753 void unregister_netdevice_many(struct list_head *head)
5755 struct net_device *dev;
5757 if (!list_empty(head)) {
5758 rollback_registered_many(head);
5759 list_for_each_entry(dev, head, unreg_list)
5763 EXPORT_SYMBOL(unregister_netdevice_many);
5766 * unregister_netdev - remove device from the kernel
5769 * This function shuts down a device interface and removes it
5770 * from the kernel tables.
5772 * This is just a wrapper for unregister_netdevice that takes
5773 * the rtnl semaphore. In general you want to use this and not
5774 * unregister_netdevice.
5776 void unregister_netdev(struct net_device *dev)
5779 unregister_netdevice(dev);
5782 EXPORT_SYMBOL(unregister_netdev);
5785 * dev_change_net_namespace - move device to different nethost namespace
5787 * @net: network namespace
5788 * @pat: If not NULL name pattern to try if the current device name
5789 * is already taken in the destination network namespace.
5791 * This function shuts down a device interface and moves it
5792 * to a new network namespace. On success 0 is returned, on
5793 * a failure a netagive errno code is returned.
5795 * Callers must hold the rtnl semaphore.
5798 int dev_change_net_namespace(struct net_device *dev, struct net *net, const char *pat)
5804 /* Don't allow namespace local devices to be moved. */
5806 if (dev->features & NETIF_F_NETNS_LOCAL)
5809 /* Ensure the device has been registrered */
5810 if (dev->reg_state != NETREG_REGISTERED)
5813 /* Get out if there is nothing todo */
5815 if (net_eq(dev_net(dev), net))
5818 /* Pick the destination device name, and ensure
5819 * we can use it in the destination network namespace.
5822 if (__dev_get_by_name(net, dev->name)) {
5823 /* We get here if we can't use the current device name */
5826 if (dev_get_valid_name(net, dev, pat) < 0)
5831 * And now a mini version of register_netdevice unregister_netdevice.
5834 /* If device is running close it first. */
5837 /* And unlink it from device chain */
5839 unlist_netdevice(dev);
5843 /* Shutdown queueing discipline. */
5846 /* Notify protocols, that we are about to destroy
5847 this device. They should clean all the things.
5849 Note that dev->reg_state stays at NETREG_REGISTERED.
5850 This is wanted because this way 8021q and macvlan know
5851 the device is just moving and can keep their slaves up.
5853 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
5855 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
5856 rtmsg_ifinfo(RTM_DELLINK, dev, ~0U);
5859 * Flush the unicast and multicast chains
5864 /* Send a netdev-removed uevent to the old namespace */
5865 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE);
5867 /* Actually switch the network namespace */
5868 dev_net_set(dev, net);
5870 /* If there is an ifindex conflict assign a new one */
5871 if (__dev_get_by_index(net, dev->ifindex)) {
5872 int iflink = (dev->iflink == dev->ifindex);
5873 dev->ifindex = dev_new_index(net);
5875 dev->iflink = dev->ifindex;
5878 /* Send a netdev-add uevent to the new namespace */
5879 kobject_uevent(&dev->dev.kobj, KOBJ_ADD);
5881 /* Fixup kobjects */
5882 err = device_rename(&dev->dev, dev->name);
5885 /* Add the device back in the hashes */
5886 list_netdevice(dev);
5888 /* Notify protocols, that a new device appeared. */
5889 call_netdevice_notifiers(NETDEV_REGISTER, dev);
5892 * Prevent userspace races by waiting until the network
5893 * device is fully setup before sending notifications.
5895 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U);
5902 EXPORT_SYMBOL_GPL(dev_change_net_namespace);
5904 static int dev_cpu_callback(struct notifier_block *nfb,
5905 unsigned long action,
5908 struct sk_buff **list_skb;
5909 struct sk_buff *skb;
5910 unsigned int cpu, oldcpu = (unsigned long)ocpu;
5911 struct softnet_data *sd, *oldsd;
5913 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
5916 local_irq_disable();
5917 cpu = smp_processor_id();
5918 sd = &per_cpu(softnet_data, cpu);
5919 oldsd = &per_cpu(softnet_data, oldcpu);
5921 /* Find end of our completion_queue. */
5922 list_skb = &sd->completion_queue;
5924 list_skb = &(*list_skb)->next;
5925 /* Append completion queue from offline CPU. */
5926 *list_skb = oldsd->completion_queue;
5927 oldsd->completion_queue = NULL;
5929 /* Append output queue from offline CPU. */
5930 if (oldsd->output_queue) {
5931 *sd->output_queue_tailp = oldsd->output_queue;
5932 sd->output_queue_tailp = oldsd->output_queue_tailp;
5933 oldsd->output_queue = NULL;
5934 oldsd->output_queue_tailp = &oldsd->output_queue;
5936 /* Append NAPI poll list from offline CPU. */
5937 if (!list_empty(&oldsd->poll_list)) {
5938 list_splice_init(&oldsd->poll_list, &sd->poll_list);
5939 raise_softirq_irqoff(NET_RX_SOFTIRQ);
5942 raise_softirq_irqoff(NET_TX_SOFTIRQ);
5945 /* Process offline CPU's input_pkt_queue */
5946 while ((skb = __skb_dequeue(&oldsd->process_queue))) {
5948 input_queue_head_incr(oldsd);
5950 while ((skb = __skb_dequeue(&oldsd->input_pkt_queue))) {
5952 input_queue_head_incr(oldsd);
5960 * netdev_increment_features - increment feature set by one
5961 * @all: current feature set
5962 * @one: new feature set
5963 * @mask: mask feature set
5965 * Computes a new feature set after adding a device with feature set
5966 * @one to the master device with current feature set @all. Will not
5967 * enable anything that is off in @mask. Returns the new feature set.
5969 netdev_features_t netdev_increment_features(netdev_features_t all,
5970 netdev_features_t one, netdev_features_t mask)
5972 if (mask & NETIF_F_GEN_CSUM)
5973 mask |= NETIF_F_ALL_CSUM;
5974 mask |= NETIF_F_VLAN_CHALLENGED;
5976 all |= one & (NETIF_F_ONE_FOR_ALL|NETIF_F_ALL_CSUM) & mask;
5977 all &= one | ~NETIF_F_ALL_FOR_ALL;
5979 /* If one device supports hw checksumming, set for all. */
5980 if (all & NETIF_F_GEN_CSUM)
5981 all &= ~(NETIF_F_ALL_CSUM & ~NETIF_F_GEN_CSUM);
5985 EXPORT_SYMBOL(netdev_increment_features);
5987 static struct hlist_head *netdev_create_hash(void)
5990 struct hlist_head *hash;
5992 hash = kmalloc(sizeof(*hash) * NETDEV_HASHENTRIES, GFP_KERNEL);
5994 for (i = 0; i < NETDEV_HASHENTRIES; i++)
5995 INIT_HLIST_HEAD(&hash[i]);
6000 /* Initialize per network namespace state */
6001 static int __net_init netdev_init(struct net *net)
6003 if (net != &init_net)
6004 INIT_LIST_HEAD(&net->dev_base_head);
6006 net->dev_name_head = netdev_create_hash();
6007 if (net->dev_name_head == NULL)
6010 net->dev_index_head = netdev_create_hash();
6011 if (net->dev_index_head == NULL)
6017 kfree(net->dev_name_head);
6023 * netdev_drivername - network driver for the device
6024 * @dev: network device
6026 * Determine network driver for device.
6028 const char *netdev_drivername(const struct net_device *dev)
6030 const struct device_driver *driver;
6031 const struct device *parent;
6032 const char *empty = "";
6034 parent = dev->dev.parent;
6038 driver = parent->driver;
6039 if (driver && driver->name)
6040 return driver->name;
6044 static int __netdev_printk(const char *level, const struct net_device *dev,
6045 struct va_format *vaf)
6049 if (dev && dev->dev.parent) {
6050 r = dev_printk_emit(level[1] - '0',
6053 dev_driver_string(dev->dev.parent),
6054 dev_name(dev->dev.parent),
6055 netdev_name(dev), vaf);
6057 r = printk("%s%s: %pV", level, netdev_name(dev), vaf);
6059 r = printk("%s(NULL net_device): %pV", level, vaf);
6065 int netdev_printk(const char *level, const struct net_device *dev,
6066 const char *format, ...)
6068 struct va_format vaf;
6072 va_start(args, format);
6077 r = __netdev_printk(level, dev, &vaf);
6083 EXPORT_SYMBOL(netdev_printk);
6085 #define define_netdev_printk_level(func, level) \
6086 int func(const struct net_device *dev, const char *fmt, ...) \
6089 struct va_format vaf; \
6092 va_start(args, fmt); \
6097 r = __netdev_printk(level, dev, &vaf); \
6103 EXPORT_SYMBOL(func);
6105 define_netdev_printk_level(netdev_emerg, KERN_EMERG);
6106 define_netdev_printk_level(netdev_alert, KERN_ALERT);
6107 define_netdev_printk_level(netdev_crit, KERN_CRIT);
6108 define_netdev_printk_level(netdev_err, KERN_ERR);
6109 define_netdev_printk_level(netdev_warn, KERN_WARNING);
6110 define_netdev_printk_level(netdev_notice, KERN_NOTICE);
6111 define_netdev_printk_level(netdev_info, KERN_INFO);
6113 static void __net_exit netdev_exit(struct net *net)
6115 kfree(net->dev_name_head);
6116 kfree(net->dev_index_head);
6119 static struct pernet_operations __net_initdata netdev_net_ops = {
6120 .init = netdev_init,
6121 .exit = netdev_exit,
6124 static void __net_exit default_device_exit(struct net *net)
6126 struct net_device *dev, *aux;
6128 * Push all migratable network devices back to the
6129 * initial network namespace
6132 for_each_netdev_safe(net, dev, aux) {
6134 char fb_name[IFNAMSIZ];
6136 /* Ignore unmoveable devices (i.e. loopback) */
6137 if (dev->features & NETIF_F_NETNS_LOCAL)
6140 /* Leave virtual devices for the generic cleanup */
6141 if (dev->rtnl_link_ops)
6144 /* Push remaining network devices to init_net */
6145 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex);
6146 err = dev_change_net_namespace(dev, &init_net, fb_name);
6148 pr_emerg("%s: failed to move %s to init_net: %d\n",
6149 __func__, dev->name, err);
6156 static void __net_exit default_device_exit_batch(struct list_head *net_list)
6158 /* At exit all network devices most be removed from a network
6159 * namespace. Do this in the reverse order of registration.
6160 * Do this across as many network namespaces as possible to
6161 * improve batching efficiency.
6163 struct net_device *dev;
6165 LIST_HEAD(dev_kill_list);
6168 list_for_each_entry(net, net_list, exit_list) {
6169 for_each_netdev_reverse(net, dev) {
6170 if (dev->rtnl_link_ops)
6171 dev->rtnl_link_ops->dellink(dev, &dev_kill_list);
6173 unregister_netdevice_queue(dev, &dev_kill_list);
6176 unregister_netdevice_many(&dev_kill_list);
6177 list_del(&dev_kill_list);
6181 static struct pernet_operations __net_initdata default_device_ops = {
6182 .exit = default_device_exit,
6183 .exit_batch = default_device_exit_batch,
6187 * Initialize the DEV module. At boot time this walks the device list and
6188 * unhooks any devices that fail to initialise (normally hardware not
6189 * present) and leaves us with a valid list of present and active devices.
6194 * This is called single threaded during boot, so no need
6195 * to take the rtnl semaphore.
6197 static int __init net_dev_init(void)
6199 int i, rc = -ENOMEM;
6201 BUG_ON(!dev_boot_phase);
6203 if (dev_proc_init())
6206 if (netdev_kobject_init())
6209 INIT_LIST_HEAD(&ptype_all);
6210 for (i = 0; i < PTYPE_HASH_SIZE; i++)
6211 INIT_LIST_HEAD(&ptype_base[i]);
6213 INIT_LIST_HEAD(&offload_base);
6215 if (register_pernet_subsys(&netdev_net_ops))
6219 * Initialise the packet receive queues.
6222 for_each_possible_cpu(i) {
6223 struct softnet_data *sd = &per_cpu(softnet_data, i);
6225 memset(sd, 0, sizeof(*sd));
6226 skb_queue_head_init(&sd->input_pkt_queue);
6227 skb_queue_head_init(&sd->process_queue);
6228 sd->completion_queue = NULL;
6229 INIT_LIST_HEAD(&sd->poll_list);
6230 sd->output_queue = NULL;
6231 sd->output_queue_tailp = &sd->output_queue;
6233 sd->csd.func = rps_trigger_softirq;
6239 sd->backlog.poll = process_backlog;
6240 sd->backlog.weight = weight_p;
6241 sd->backlog.gro_list = NULL;
6242 sd->backlog.gro_count = 0;
6247 /* The loopback device is special if any other network devices
6248 * is present in a network namespace the loopback device must
6249 * be present. Since we now dynamically allocate and free the
6250 * loopback device ensure this invariant is maintained by
6251 * keeping the loopback device as the first device on the
6252 * list of network devices. Ensuring the loopback devices
6253 * is the first device that appears and the last network device
6256 if (register_pernet_device(&loopback_net_ops))
6259 if (register_pernet_device(&default_device_ops))
6262 open_softirq(NET_TX_SOFTIRQ, net_tx_action);
6263 open_softirq(NET_RX_SOFTIRQ, net_rx_action);
6265 hotcpu_notifier(dev_cpu_callback, 0);
6272 subsys_initcall(net_dev_init);