2 * NET3 Protocol independent device support routines.
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
9 * Derived from the non IP parts of dev.c 1.0.19
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
15 * Florian la Roche <rzsfl@rz.uni-sb.de>
16 * Alan Cox <gw4pts@gw4pts.ampr.org>
17 * David Hinds <dahinds@users.sourceforge.net>
18 * Alexey Kuznetsov <kuznet@ms2.inr.ac.ru>
19 * Adam Sulmicki <adam@cfar.umd.edu>
20 * Pekka Riikonen <priikone@poesidon.pspt.fi>
23 * D.J. Barrow : Fixed bug where dev->refcnt gets set
24 * to 2 if register_netdev gets called
25 * before net_dev_init & also removed a
26 * few lines of code in the process.
27 * Alan Cox : device private ioctl copies fields back.
28 * Alan Cox : Transmit queue code does relevant
29 * stunts to keep the queue safe.
30 * Alan Cox : Fixed double lock.
31 * Alan Cox : Fixed promisc NULL pointer trap
32 * ???????? : Support the full private ioctl range
33 * Alan Cox : Moved ioctl permission check into
35 * Tim Kordas : SIOCADDMULTI/SIOCDELMULTI
36 * Alan Cox : 100 backlog just doesn't cut it when
37 * you start doing multicast video 8)
38 * Alan Cox : Rewrote net_bh and list manager.
39 * Alan Cox : Fix ETH_P_ALL echoback lengths.
40 * Alan Cox : Took out transmit every packet pass
41 * Saved a few bytes in the ioctl handler
42 * Alan Cox : Network driver sets packet type before
43 * calling netif_rx. Saves a function
45 * Alan Cox : Hashed net_bh()
46 * Richard Kooijman: Timestamp fixes.
47 * Alan Cox : Wrong field in SIOCGIFDSTADDR
48 * Alan Cox : Device lock protection.
49 * Alan Cox : Fixed nasty side effect of device close
51 * Rudi Cilibrasi : Pass the right thing to
53 * Dave Miller : 32bit quantity for the device lock to
54 * make it work out on a Sparc.
55 * Bjorn Ekwall : Added KERNELD hack.
56 * Alan Cox : Cleaned up the backlog initialise.
57 * Craig Metz : SIOCGIFCONF fix if space for under
59 * Thomas Bogendoerfer : Return ENODEV for dev_open, if there
60 * is no device open function.
61 * Andi Kleen : Fix error reporting for SIOCGIFCONF
62 * Michael Chastain : Fix signed/unsigned for SIOCGIFCONF
63 * Cyrus Durgin : Cleaned for KMOD
64 * Adam Sulmicki : Bug Fix : Network Device Unload
65 * A network device unload needs to purge
67 * Paul Rusty Russell : SIOCSIFNAME
68 * Pekka Riikonen : Netdev boot-time settings code
69 * Andrew Morton : Make unregister_netdevice wait
70 * indefinitely on dev->refcnt
71 * J Hadi Salim : - Backlog queue sampling
72 * - netif_rx() feedback
75 #include <linux/uaccess.h>
76 #include <linux/bitops.h>
77 #include <linux/capability.h>
78 #include <linux/cpu.h>
79 #include <linux/types.h>
80 #include <linux/kernel.h>
81 #include <linux/hash.h>
82 #include <linux/slab.h>
83 #include <linux/sched.h>
84 #include <linux/sched/mm.h>
85 #include <linux/mutex.h>
86 #include <linux/string.h>
88 #include <linux/socket.h>
89 #include <linux/sockios.h>
90 #include <linux/errno.h>
91 #include <linux/interrupt.h>
92 #include <linux/if_ether.h>
93 #include <linux/netdevice.h>
94 #include <linux/etherdevice.h>
95 #include <linux/ethtool.h>
96 #include <linux/notifier.h>
97 #include <linux/skbuff.h>
98 #include <linux/bpf.h>
99 #include <linux/bpf_trace.h>
100 #include <net/net_namespace.h>
101 #include <net/sock.h>
102 #include <net/busy_poll.h>
103 #include <linux/rtnetlink.h>
104 #include <linux/stat.h>
106 #include <net/dst_metadata.h>
107 #include <net/pkt_sched.h>
108 #include <net/pkt_cls.h>
109 #include <net/checksum.h>
110 #include <net/xfrm.h>
111 #include <linux/highmem.h>
112 #include <linux/init.h>
113 #include <linux/module.h>
114 #include <linux/netpoll.h>
115 #include <linux/rcupdate.h>
116 #include <linux/delay.h>
117 #include <net/iw_handler.h>
118 #include <asm/current.h>
119 #include <linux/audit.h>
120 #include <linux/dmaengine.h>
121 #include <linux/err.h>
122 #include <linux/ctype.h>
123 #include <linux/if_arp.h>
124 #include <linux/if_vlan.h>
125 #include <linux/ip.h>
127 #include <net/mpls.h>
128 #include <linux/ipv6.h>
129 #include <linux/in.h>
130 #include <linux/jhash.h>
131 #include <linux/random.h>
132 #include <trace/events/napi.h>
133 #include <trace/events/net.h>
134 #include <trace/events/skb.h>
135 #include <linux/pci.h>
136 #include <linux/inetdevice.h>
137 #include <linux/cpu_rmap.h>
138 #include <linux/static_key.h>
139 #include <linux/hashtable.h>
140 #include <linux/vmalloc.h>
141 #include <linux/if_macvlan.h>
142 #include <linux/errqueue.h>
143 #include <linux/hrtimer.h>
144 #include <linux/netfilter_ingress.h>
145 #include <linux/crash_dump.h>
146 #include <linux/sctp.h>
148 #include "net-sysfs.h"
150 /* Instead of increasing this, you should create a hash table. */
151 #define MAX_GRO_SKBS 8
153 /* This should be increased if a protocol with a bigger head is added. */
154 #define GRO_MAX_HEAD (MAX_HEADER + 128)
156 static DEFINE_SPINLOCK(ptype_lock);
157 static DEFINE_SPINLOCK(offload_lock);
158 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly;
159 struct list_head ptype_all __read_mostly; /* Taps */
160 static struct list_head offload_base __read_mostly;
162 static int netif_rx_internal(struct sk_buff *skb);
163 static int call_netdevice_notifiers_info(unsigned long val,
164 struct net_device *dev,
165 struct netdev_notifier_info *info);
166 static struct napi_struct *napi_by_id(unsigned int napi_id);
169 * The @dev_base_head list is protected by @dev_base_lock and the rtnl
172 * Pure readers hold dev_base_lock for reading, or rcu_read_lock()
174 * Writers must hold the rtnl semaphore while they loop through the
175 * dev_base_head list, and hold dev_base_lock for writing when they do the
176 * actual updates. This allows pure readers to access the list even
177 * while a writer is preparing to update it.
179 * To put it another way, dev_base_lock is held for writing only to
180 * protect against pure readers; the rtnl semaphore provides the
181 * protection against other writers.
183 * See, for example usages, register_netdevice() and
184 * unregister_netdevice(), which must be called with the rtnl
187 DEFINE_RWLOCK(dev_base_lock);
188 EXPORT_SYMBOL(dev_base_lock);
190 /* protects napi_hash addition/deletion and napi_gen_id */
191 static DEFINE_SPINLOCK(napi_hash_lock);
193 static unsigned int napi_gen_id = NR_CPUS;
194 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8);
196 static seqcount_t devnet_rename_seq;
198 static inline void dev_base_seq_inc(struct net *net)
200 while (++net->dev_base_seq == 0)
204 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name)
206 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ));
208 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)];
211 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex)
213 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)];
216 static inline void rps_lock(struct softnet_data *sd)
219 spin_lock(&sd->input_pkt_queue.lock);
223 static inline void rps_unlock(struct softnet_data *sd)
226 spin_unlock(&sd->input_pkt_queue.lock);
230 /* Device list insertion */
231 static void list_netdevice(struct net_device *dev)
233 struct net *net = dev_net(dev);
237 write_lock_bh(&dev_base_lock);
238 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head);
239 hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
240 hlist_add_head_rcu(&dev->index_hlist,
241 dev_index_hash(net, dev->ifindex));
242 write_unlock_bh(&dev_base_lock);
244 dev_base_seq_inc(net);
247 /* Device list removal
248 * caller must respect a RCU grace period before freeing/reusing dev
250 static void unlist_netdevice(struct net_device *dev)
254 /* Unlink dev from the device chain */
255 write_lock_bh(&dev_base_lock);
256 list_del_rcu(&dev->dev_list);
257 hlist_del_rcu(&dev->name_hlist);
258 hlist_del_rcu(&dev->index_hlist);
259 write_unlock_bh(&dev_base_lock);
261 dev_base_seq_inc(dev_net(dev));
268 static RAW_NOTIFIER_HEAD(netdev_chain);
271 * Device drivers call our routines to queue packets here. We empty the
272 * queue in the local softnet handler.
275 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data);
276 EXPORT_PER_CPU_SYMBOL(softnet_data);
278 #ifdef CONFIG_LOCKDEP
280 * register_netdevice() inits txq->_xmit_lock and sets lockdep class
281 * according to dev->type
283 static const unsigned short netdev_lock_type[] = {
284 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25,
285 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET,
286 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM,
287 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP,
288 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD,
289 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25,
290 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP,
291 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD,
292 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI,
293 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE,
294 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET,
295 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL,
296 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM,
297 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE,
298 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE};
300 static const char *const netdev_lock_name[] = {
301 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25",
302 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET",
303 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM",
304 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP",
305 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD",
306 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25",
307 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP",
308 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD",
309 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI",
310 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE",
311 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET",
312 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL",
313 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM",
314 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE",
315 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"};
317 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)];
318 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)];
320 static inline unsigned short netdev_lock_pos(unsigned short dev_type)
324 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++)
325 if (netdev_lock_type[i] == dev_type)
327 /* the last key is used by default */
328 return ARRAY_SIZE(netdev_lock_type) - 1;
331 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
332 unsigned short dev_type)
336 i = netdev_lock_pos(dev_type);
337 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i],
338 netdev_lock_name[i]);
341 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
345 i = netdev_lock_pos(dev->type);
346 lockdep_set_class_and_name(&dev->addr_list_lock,
347 &netdev_addr_lock_key[i],
348 netdev_lock_name[i]);
351 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
352 unsigned short dev_type)
355 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
360 /*******************************************************************************
362 * Protocol management and registration routines
364 *******************************************************************************/
368 * Add a protocol ID to the list. Now that the input handler is
369 * smarter we can dispense with all the messy stuff that used to be
372 * BEWARE!!! Protocol handlers, mangling input packets,
373 * MUST BE last in hash buckets and checking protocol handlers
374 * MUST start from promiscuous ptype_all chain in net_bh.
375 * It is true now, do not change it.
376 * Explanation follows: if protocol handler, mangling packet, will
377 * be the first on list, it is not able to sense, that packet
378 * is cloned and should be copied-on-write, so that it will
379 * change it and subsequent readers will get broken packet.
383 static inline struct list_head *ptype_head(const struct packet_type *pt)
385 if (pt->type == htons(ETH_P_ALL))
386 return pt->dev ? &pt->dev->ptype_all : &ptype_all;
388 return pt->dev ? &pt->dev->ptype_specific :
389 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK];
393 * dev_add_pack - add packet handler
394 * @pt: packet type declaration
396 * Add a protocol handler to the networking stack. The passed &packet_type
397 * is linked into kernel lists and may not be freed until it has been
398 * removed from the kernel lists.
400 * This call does not sleep therefore it can not
401 * guarantee all CPU's that are in middle of receiving packets
402 * will see the new packet type (until the next received packet).
405 void dev_add_pack(struct packet_type *pt)
407 struct list_head *head = ptype_head(pt);
409 spin_lock(&ptype_lock);
410 list_add_rcu(&pt->list, head);
411 spin_unlock(&ptype_lock);
413 EXPORT_SYMBOL(dev_add_pack);
416 * __dev_remove_pack - remove packet handler
417 * @pt: packet type declaration
419 * Remove a protocol handler that was previously added to the kernel
420 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
421 * from the kernel lists and can be freed or reused once this function
424 * The packet type might still be in use by receivers
425 * and must not be freed until after all the CPU's have gone
426 * through a quiescent state.
428 void __dev_remove_pack(struct packet_type *pt)
430 struct list_head *head = ptype_head(pt);
431 struct packet_type *pt1;
433 spin_lock(&ptype_lock);
435 list_for_each_entry(pt1, head, list) {
437 list_del_rcu(&pt->list);
442 pr_warn("dev_remove_pack: %p not found\n", pt);
444 spin_unlock(&ptype_lock);
446 EXPORT_SYMBOL(__dev_remove_pack);
449 * dev_remove_pack - remove packet handler
450 * @pt: packet type declaration
452 * Remove a protocol handler that was previously added to the kernel
453 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
454 * from the kernel lists and can be freed or reused once this function
457 * This call sleeps to guarantee that no CPU is looking at the packet
460 void dev_remove_pack(struct packet_type *pt)
462 __dev_remove_pack(pt);
466 EXPORT_SYMBOL(dev_remove_pack);
470 * dev_add_offload - register offload handlers
471 * @po: protocol offload declaration
473 * Add protocol offload handlers to the networking stack. The passed
474 * &proto_offload is linked into kernel lists and may not be freed until
475 * it has been removed from the kernel lists.
477 * This call does not sleep therefore it can not
478 * guarantee all CPU's that are in middle of receiving packets
479 * will see the new offload handlers (until the next received packet).
481 void dev_add_offload(struct packet_offload *po)
483 struct packet_offload *elem;
485 spin_lock(&offload_lock);
486 list_for_each_entry(elem, &offload_base, list) {
487 if (po->priority < elem->priority)
490 list_add_rcu(&po->list, elem->list.prev);
491 spin_unlock(&offload_lock);
493 EXPORT_SYMBOL(dev_add_offload);
496 * __dev_remove_offload - remove offload handler
497 * @po: packet offload declaration
499 * Remove a protocol offload handler that was previously added to the
500 * kernel offload handlers by dev_add_offload(). The passed &offload_type
501 * is removed from the kernel lists and can be freed or reused once this
504 * The packet type might still be in use by receivers
505 * and must not be freed until after all the CPU's have gone
506 * through a quiescent state.
508 static void __dev_remove_offload(struct packet_offload *po)
510 struct list_head *head = &offload_base;
511 struct packet_offload *po1;
513 spin_lock(&offload_lock);
515 list_for_each_entry(po1, head, list) {
517 list_del_rcu(&po->list);
522 pr_warn("dev_remove_offload: %p not found\n", po);
524 spin_unlock(&offload_lock);
528 * dev_remove_offload - remove packet offload handler
529 * @po: packet offload declaration
531 * Remove a packet offload handler that was previously added to the kernel
532 * offload handlers by dev_add_offload(). The passed &offload_type is
533 * removed from the kernel lists and can be freed or reused once this
536 * This call sleeps to guarantee that no CPU is looking at the packet
539 void dev_remove_offload(struct packet_offload *po)
541 __dev_remove_offload(po);
545 EXPORT_SYMBOL(dev_remove_offload);
547 /******************************************************************************
549 * Device Boot-time Settings Routines
551 ******************************************************************************/
553 /* Boot time configuration table */
554 static struct netdev_boot_setup dev_boot_setup[NETDEV_BOOT_SETUP_MAX];
557 * netdev_boot_setup_add - add new setup entry
558 * @name: name of the device
559 * @map: configured settings for the device
561 * Adds new setup entry to the dev_boot_setup list. The function
562 * returns 0 on error and 1 on success. This is a generic routine to
565 static int netdev_boot_setup_add(char *name, struct ifmap *map)
567 struct netdev_boot_setup *s;
571 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
572 if (s[i].name[0] == '\0' || s[i].name[0] == ' ') {
573 memset(s[i].name, 0, sizeof(s[i].name));
574 strlcpy(s[i].name, name, IFNAMSIZ);
575 memcpy(&s[i].map, map, sizeof(s[i].map));
580 return i >= NETDEV_BOOT_SETUP_MAX ? 0 : 1;
584 * netdev_boot_setup_check - check boot time settings
585 * @dev: the netdevice
587 * Check boot time settings for the device.
588 * The found settings are set for the device to be used
589 * later in the device probing.
590 * Returns 0 if no settings found, 1 if they are.
592 int netdev_boot_setup_check(struct net_device *dev)
594 struct netdev_boot_setup *s = dev_boot_setup;
597 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
598 if (s[i].name[0] != '\0' && s[i].name[0] != ' ' &&
599 !strcmp(dev->name, s[i].name)) {
600 dev->irq = s[i].map.irq;
601 dev->base_addr = s[i].map.base_addr;
602 dev->mem_start = s[i].map.mem_start;
603 dev->mem_end = s[i].map.mem_end;
609 EXPORT_SYMBOL(netdev_boot_setup_check);
613 * netdev_boot_base - get address from boot time settings
614 * @prefix: prefix for network device
615 * @unit: id for network device
617 * Check boot time settings for the base address of device.
618 * The found settings are set for the device to be used
619 * later in the device probing.
620 * Returns 0 if no settings found.
622 unsigned long netdev_boot_base(const char *prefix, int unit)
624 const struct netdev_boot_setup *s = dev_boot_setup;
628 sprintf(name, "%s%d", prefix, unit);
631 * If device already registered then return base of 1
632 * to indicate not to probe for this interface
634 if (__dev_get_by_name(&init_net, name))
637 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++)
638 if (!strcmp(name, s[i].name))
639 return s[i].map.base_addr;
644 * Saves at boot time configured settings for any netdevice.
646 int __init netdev_boot_setup(char *str)
651 str = get_options(str, ARRAY_SIZE(ints), ints);
656 memset(&map, 0, sizeof(map));
660 map.base_addr = ints[2];
662 map.mem_start = ints[3];
664 map.mem_end = ints[4];
666 /* Add new entry to the list */
667 return netdev_boot_setup_add(str, &map);
670 __setup("netdev=", netdev_boot_setup);
672 /*******************************************************************************
674 * Device Interface Subroutines
676 *******************************************************************************/
679 * dev_get_iflink - get 'iflink' value of a interface
680 * @dev: targeted interface
682 * Indicates the ifindex the interface is linked to.
683 * Physical interfaces have the same 'ifindex' and 'iflink' values.
686 int dev_get_iflink(const struct net_device *dev)
688 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink)
689 return dev->netdev_ops->ndo_get_iflink(dev);
693 EXPORT_SYMBOL(dev_get_iflink);
696 * dev_fill_metadata_dst - Retrieve tunnel egress information.
697 * @dev: targeted interface
700 * For better visibility of tunnel traffic OVS needs to retrieve
701 * egress tunnel information for a packet. Following API allows
702 * user to get this info.
704 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb)
706 struct ip_tunnel_info *info;
708 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst)
711 info = skb_tunnel_info_unclone(skb);
714 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX)))
717 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb);
719 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst);
722 * __dev_get_by_name - find a device by its name
723 * @net: the applicable net namespace
724 * @name: name to find
726 * Find an interface by name. Must be called under RTNL semaphore
727 * or @dev_base_lock. If the name is found a pointer to the device
728 * is returned. If the name is not found then %NULL is returned. The
729 * reference counters are not incremented so the caller must be
730 * careful with locks.
733 struct net_device *__dev_get_by_name(struct net *net, const char *name)
735 struct net_device *dev;
736 struct hlist_head *head = dev_name_hash(net, name);
738 hlist_for_each_entry(dev, head, name_hlist)
739 if (!strncmp(dev->name, name, IFNAMSIZ))
744 EXPORT_SYMBOL(__dev_get_by_name);
747 * dev_get_by_name_rcu - find a device by its name
748 * @net: the applicable net namespace
749 * @name: name to find
751 * Find an interface by name.
752 * If the name is found a pointer to the device is returned.
753 * If the name is not found then %NULL is returned.
754 * The reference counters are not incremented so the caller must be
755 * careful with locks. The caller must hold RCU lock.
758 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name)
760 struct net_device *dev;
761 struct hlist_head *head = dev_name_hash(net, name);
763 hlist_for_each_entry_rcu(dev, head, name_hlist)
764 if (!strncmp(dev->name, name, IFNAMSIZ))
769 EXPORT_SYMBOL(dev_get_by_name_rcu);
772 * dev_get_by_name - find a device by its name
773 * @net: the applicable net namespace
774 * @name: name to find
776 * Find an interface by name. This can be called from any
777 * context and does its own locking. The returned handle has
778 * the usage count incremented and the caller must use dev_put() to
779 * release it when it is no longer needed. %NULL is returned if no
780 * matching device is found.
783 struct net_device *dev_get_by_name(struct net *net, const char *name)
785 struct net_device *dev;
788 dev = dev_get_by_name_rcu(net, name);
794 EXPORT_SYMBOL(dev_get_by_name);
797 * __dev_get_by_index - find a device by its ifindex
798 * @net: the applicable net namespace
799 * @ifindex: index of device
801 * Search for an interface by index. Returns %NULL if the device
802 * is not found or a pointer to the device. The device has not
803 * had its reference counter increased so the caller must be careful
804 * about locking. The caller must hold either the RTNL semaphore
808 struct net_device *__dev_get_by_index(struct net *net, int ifindex)
810 struct net_device *dev;
811 struct hlist_head *head = dev_index_hash(net, ifindex);
813 hlist_for_each_entry(dev, head, index_hlist)
814 if (dev->ifindex == ifindex)
819 EXPORT_SYMBOL(__dev_get_by_index);
822 * dev_get_by_index_rcu - find a device by its ifindex
823 * @net: the applicable net namespace
824 * @ifindex: index of device
826 * Search for an interface by index. Returns %NULL if the device
827 * is not found or a pointer to the device. The device has not
828 * had its reference counter increased so the caller must be careful
829 * about locking. The caller must hold RCU lock.
832 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex)
834 struct net_device *dev;
835 struct hlist_head *head = dev_index_hash(net, ifindex);
837 hlist_for_each_entry_rcu(dev, head, index_hlist)
838 if (dev->ifindex == ifindex)
843 EXPORT_SYMBOL(dev_get_by_index_rcu);
847 * dev_get_by_index - find a device by its ifindex
848 * @net: the applicable net namespace
849 * @ifindex: index of device
851 * Search for an interface by index. Returns NULL if the device
852 * is not found or a pointer to the device. The device returned has
853 * had a reference added and the pointer is safe until the user calls
854 * dev_put to indicate they have finished with it.
857 struct net_device *dev_get_by_index(struct net *net, int ifindex)
859 struct net_device *dev;
862 dev = dev_get_by_index_rcu(net, ifindex);
868 EXPORT_SYMBOL(dev_get_by_index);
871 * dev_get_by_napi_id - find a device by napi_id
872 * @napi_id: ID of the NAPI struct
874 * Search for an interface by NAPI ID. Returns %NULL if the device
875 * is not found or a pointer to the device. The device has not had
876 * its reference counter increased so the caller must be careful
877 * about locking. The caller must hold RCU lock.
880 struct net_device *dev_get_by_napi_id(unsigned int napi_id)
882 struct napi_struct *napi;
884 WARN_ON_ONCE(!rcu_read_lock_held());
886 if (napi_id < MIN_NAPI_ID)
889 napi = napi_by_id(napi_id);
891 return napi ? napi->dev : NULL;
893 EXPORT_SYMBOL(dev_get_by_napi_id);
896 * netdev_get_name - get a netdevice name, knowing its ifindex.
897 * @net: network namespace
898 * @name: a pointer to the buffer where the name will be stored.
899 * @ifindex: the ifindex of the interface to get the name from.
901 * The use of raw_seqcount_begin() and cond_resched() before
902 * retrying is required as we want to give the writers a chance
903 * to complete when CONFIG_PREEMPT is not set.
905 int netdev_get_name(struct net *net, char *name, int ifindex)
907 struct net_device *dev;
911 seq = raw_seqcount_begin(&devnet_rename_seq);
913 dev = dev_get_by_index_rcu(net, ifindex);
919 strcpy(name, dev->name);
921 if (read_seqcount_retry(&devnet_rename_seq, seq)) {
930 * dev_getbyhwaddr_rcu - find a device by its hardware address
931 * @net: the applicable net namespace
932 * @type: media type of device
933 * @ha: hardware address
935 * Search for an interface by MAC address. Returns NULL if the device
936 * is not found or a pointer to the device.
937 * The caller must hold RCU or RTNL.
938 * The returned device has not had its ref count increased
939 * and the caller must therefore be careful about locking
943 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type,
946 struct net_device *dev;
948 for_each_netdev_rcu(net, dev)
949 if (dev->type == type &&
950 !memcmp(dev->dev_addr, ha, dev->addr_len))
955 EXPORT_SYMBOL(dev_getbyhwaddr_rcu);
957 struct net_device *__dev_getfirstbyhwtype(struct net *net, unsigned short type)
959 struct net_device *dev;
962 for_each_netdev(net, dev)
963 if (dev->type == type)
968 EXPORT_SYMBOL(__dev_getfirstbyhwtype);
970 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type)
972 struct net_device *dev, *ret = NULL;
975 for_each_netdev_rcu(net, dev)
976 if (dev->type == type) {
984 EXPORT_SYMBOL(dev_getfirstbyhwtype);
987 * __dev_get_by_flags - find any device with given flags
988 * @net: the applicable net namespace
989 * @if_flags: IFF_* values
990 * @mask: bitmask of bits in if_flags to check
992 * Search for any interface with the given flags. Returns NULL if a device
993 * is not found or a pointer to the device. Must be called inside
994 * rtnl_lock(), and result refcount is unchanged.
997 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags,
1000 struct net_device *dev, *ret;
1005 for_each_netdev(net, dev) {
1006 if (((dev->flags ^ if_flags) & mask) == 0) {
1013 EXPORT_SYMBOL(__dev_get_by_flags);
1016 * dev_valid_name - check if name is okay for network device
1017 * @name: name string
1019 * Network device names need to be valid file names to
1020 * to allow sysfs to work. We also disallow any kind of
1023 bool dev_valid_name(const char *name)
1027 if (strlen(name) >= IFNAMSIZ)
1029 if (!strcmp(name, ".") || !strcmp(name, ".."))
1033 if (*name == '/' || *name == ':' || isspace(*name))
1039 EXPORT_SYMBOL(dev_valid_name);
1042 * __dev_alloc_name - allocate a name for a device
1043 * @net: network namespace to allocate the device name in
1044 * @name: name format string
1045 * @buf: scratch buffer and result name string
1047 * Passed a format string - eg "lt%d" it will try and find a suitable
1048 * id. It scans list of devices to build up a free map, then chooses
1049 * the first empty slot. The caller must hold the dev_base or rtnl lock
1050 * while allocating the name and adding the device in order to avoid
1052 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1053 * Returns the number of the unit assigned or a negative errno code.
1056 static int __dev_alloc_name(struct net *net, const char *name, char *buf)
1060 const int max_netdevices = 8*PAGE_SIZE;
1061 unsigned long *inuse;
1062 struct net_device *d;
1064 p = strnchr(name, IFNAMSIZ-1, '%');
1067 * Verify the string as this thing may have come from
1068 * the user. There must be either one "%d" and no other "%"
1071 if (p[1] != 'd' || strchr(p + 2, '%'))
1074 /* Use one page as a bit array of possible slots */
1075 inuse = (unsigned long *) get_zeroed_page(GFP_ATOMIC);
1079 for_each_netdev(net, d) {
1080 if (!sscanf(d->name, name, &i))
1082 if (i < 0 || i >= max_netdevices)
1085 /* avoid cases where sscanf is not exact inverse of printf */
1086 snprintf(buf, IFNAMSIZ, name, i);
1087 if (!strncmp(buf, d->name, IFNAMSIZ))
1091 i = find_first_zero_bit(inuse, max_netdevices);
1092 free_page((unsigned long) inuse);
1096 snprintf(buf, IFNAMSIZ, name, i);
1097 if (!__dev_get_by_name(net, buf))
1100 /* It is possible to run out of possible slots
1101 * when the name is long and there isn't enough space left
1102 * for the digits, or if all bits are used.
1108 * dev_alloc_name - allocate a name for a device
1110 * @name: name format string
1112 * Passed a format string - eg "lt%d" it will try and find a suitable
1113 * id. It scans list of devices to build up a free map, then chooses
1114 * the first empty slot. The caller must hold the dev_base or rtnl lock
1115 * while allocating the name and adding the device in order to avoid
1117 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1118 * Returns the number of the unit assigned or a negative errno code.
1121 int dev_alloc_name(struct net_device *dev, const char *name)
1127 BUG_ON(!dev_net(dev));
1129 ret = __dev_alloc_name(net, name, buf);
1131 strlcpy(dev->name, buf, IFNAMSIZ);
1134 EXPORT_SYMBOL(dev_alloc_name);
1136 static int dev_alloc_name_ns(struct net *net,
1137 struct net_device *dev,
1143 ret = __dev_alloc_name(net, name, buf);
1145 strlcpy(dev->name, buf, IFNAMSIZ);
1149 static int dev_get_valid_name(struct net *net,
1150 struct net_device *dev,
1155 if (!dev_valid_name(name))
1158 if (strchr(name, '%'))
1159 return dev_alloc_name_ns(net, dev, name);
1160 else if (__dev_get_by_name(net, name))
1162 else if (dev->name != name)
1163 strlcpy(dev->name, name, IFNAMSIZ);
1169 * dev_change_name - change name of a device
1171 * @newname: name (or format string) must be at least IFNAMSIZ
1173 * Change name of a device, can pass format strings "eth%d".
1176 int dev_change_name(struct net_device *dev, const char *newname)
1178 unsigned char old_assign_type;
1179 char oldname[IFNAMSIZ];
1185 BUG_ON(!dev_net(dev));
1188 if (dev->flags & IFF_UP)
1191 write_seqcount_begin(&devnet_rename_seq);
1193 if (strncmp(newname, dev->name, IFNAMSIZ) == 0) {
1194 write_seqcount_end(&devnet_rename_seq);
1198 memcpy(oldname, dev->name, IFNAMSIZ);
1200 err = dev_get_valid_name(net, dev, newname);
1202 write_seqcount_end(&devnet_rename_seq);
1206 if (oldname[0] && !strchr(oldname, '%'))
1207 netdev_info(dev, "renamed from %s\n", oldname);
1209 old_assign_type = dev->name_assign_type;
1210 dev->name_assign_type = NET_NAME_RENAMED;
1213 ret = device_rename(&dev->dev, dev->name);
1215 memcpy(dev->name, oldname, IFNAMSIZ);
1216 dev->name_assign_type = old_assign_type;
1217 write_seqcount_end(&devnet_rename_seq);
1221 write_seqcount_end(&devnet_rename_seq);
1223 netdev_adjacent_rename_links(dev, oldname);
1225 write_lock_bh(&dev_base_lock);
1226 hlist_del_rcu(&dev->name_hlist);
1227 write_unlock_bh(&dev_base_lock);
1231 write_lock_bh(&dev_base_lock);
1232 hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
1233 write_unlock_bh(&dev_base_lock);
1235 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev);
1236 ret = notifier_to_errno(ret);
1239 /* err >= 0 after dev_alloc_name() or stores the first errno */
1242 write_seqcount_begin(&devnet_rename_seq);
1243 memcpy(dev->name, oldname, IFNAMSIZ);
1244 memcpy(oldname, newname, IFNAMSIZ);
1245 dev->name_assign_type = old_assign_type;
1246 old_assign_type = NET_NAME_RENAMED;
1249 pr_err("%s: name change rollback failed: %d\n",
1258 * dev_set_alias - change ifalias of a device
1260 * @alias: name up to IFALIASZ
1261 * @len: limit of bytes to copy from info
1263 * Set ifalias for a device,
1265 int dev_set_alias(struct net_device *dev, const char *alias, size_t len)
1271 if (len >= IFALIASZ)
1275 kfree(dev->ifalias);
1276 dev->ifalias = NULL;
1280 new_ifalias = krealloc(dev->ifalias, len + 1, GFP_KERNEL);
1283 dev->ifalias = new_ifalias;
1284 memcpy(dev->ifalias, alias, len);
1285 dev->ifalias[len] = 0;
1292 * netdev_features_change - device changes features
1293 * @dev: device to cause notification
1295 * Called to indicate a device has changed features.
1297 void netdev_features_change(struct net_device *dev)
1299 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev);
1301 EXPORT_SYMBOL(netdev_features_change);
1304 * netdev_state_change - device changes state
1305 * @dev: device to cause notification
1307 * Called to indicate a device has changed state. This function calls
1308 * the notifier chains for netdev_chain and sends a NEWLINK message
1309 * to the routing socket.
1311 void netdev_state_change(struct net_device *dev)
1313 if (dev->flags & IFF_UP) {
1314 struct netdev_notifier_change_info change_info;
1316 change_info.flags_changed = 0;
1317 call_netdevice_notifiers_info(NETDEV_CHANGE, dev,
1319 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL);
1322 EXPORT_SYMBOL(netdev_state_change);
1325 * netdev_notify_peers - notify network peers about existence of @dev
1326 * @dev: network device
1328 * Generate traffic such that interested network peers are aware of
1329 * @dev, such as by generating a gratuitous ARP. This may be used when
1330 * a device wants to inform the rest of the network about some sort of
1331 * reconfiguration such as a failover event or virtual machine
1334 void netdev_notify_peers(struct net_device *dev)
1337 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
1338 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev);
1341 EXPORT_SYMBOL(netdev_notify_peers);
1343 static int __dev_open(struct net_device *dev)
1345 const struct net_device_ops *ops = dev->netdev_ops;
1350 if (!netif_device_present(dev))
1353 /* Block netpoll from trying to do any rx path servicing.
1354 * If we don't do this there is a chance ndo_poll_controller
1355 * or ndo_poll may be running while we open the device
1357 netpoll_poll_disable(dev);
1359 ret = call_netdevice_notifiers(NETDEV_PRE_UP, dev);
1360 ret = notifier_to_errno(ret);
1364 set_bit(__LINK_STATE_START, &dev->state);
1366 if (ops->ndo_validate_addr)
1367 ret = ops->ndo_validate_addr(dev);
1369 if (!ret && ops->ndo_open)
1370 ret = ops->ndo_open(dev);
1372 netpoll_poll_enable(dev);
1375 clear_bit(__LINK_STATE_START, &dev->state);
1377 dev->flags |= IFF_UP;
1378 dev_set_rx_mode(dev);
1380 add_device_randomness(dev->dev_addr, dev->addr_len);
1387 * dev_open - prepare an interface for use.
1388 * @dev: device to open
1390 * Takes a device from down to up state. The device's private open
1391 * function is invoked and then the multicast lists are loaded. Finally
1392 * the device is moved into the up state and a %NETDEV_UP message is
1393 * sent to the netdev notifier chain.
1395 * Calling this function on an active interface is a nop. On a failure
1396 * a negative errno code is returned.
1398 int dev_open(struct net_device *dev)
1402 if (dev->flags & IFF_UP)
1405 ret = __dev_open(dev);
1409 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1410 call_netdevice_notifiers(NETDEV_UP, dev);
1414 EXPORT_SYMBOL(dev_open);
1416 static void __dev_close_many(struct list_head *head)
1418 struct net_device *dev;
1423 list_for_each_entry(dev, head, close_list) {
1424 /* Temporarily disable netpoll until the interface is down */
1425 netpoll_poll_disable(dev);
1427 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev);
1429 clear_bit(__LINK_STATE_START, &dev->state);
1431 /* Synchronize to scheduled poll. We cannot touch poll list, it
1432 * can be even on different cpu. So just clear netif_running().
1434 * dev->stop() will invoke napi_disable() on all of it's
1435 * napi_struct instances on this device.
1437 smp_mb__after_atomic(); /* Commit netif_running(). */
1440 dev_deactivate_many(head);
1442 list_for_each_entry(dev, head, close_list) {
1443 const struct net_device_ops *ops = dev->netdev_ops;
1446 * Call the device specific close. This cannot fail.
1447 * Only if device is UP
1449 * We allow it to be called even after a DETACH hot-plug
1455 dev->flags &= ~IFF_UP;
1456 netpoll_poll_enable(dev);
1460 static void __dev_close(struct net_device *dev)
1464 list_add(&dev->close_list, &single);
1465 __dev_close_many(&single);
1469 void dev_close_many(struct list_head *head, bool unlink)
1471 struct net_device *dev, *tmp;
1473 /* Remove the devices that don't need to be closed */
1474 list_for_each_entry_safe(dev, tmp, head, close_list)
1475 if (!(dev->flags & IFF_UP))
1476 list_del_init(&dev->close_list);
1478 __dev_close_many(head);
1480 list_for_each_entry_safe(dev, tmp, head, close_list) {
1481 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1482 call_netdevice_notifiers(NETDEV_DOWN, dev);
1484 list_del_init(&dev->close_list);
1487 EXPORT_SYMBOL(dev_close_many);
1490 * dev_close - shutdown an interface.
1491 * @dev: device to shutdown
1493 * This function moves an active device into down state. A
1494 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device
1495 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier
1498 void dev_close(struct net_device *dev)
1500 if (dev->flags & IFF_UP) {
1503 list_add(&dev->close_list, &single);
1504 dev_close_many(&single, true);
1508 EXPORT_SYMBOL(dev_close);
1512 * dev_disable_lro - disable Large Receive Offload on a device
1515 * Disable Large Receive Offload (LRO) on a net device. Must be
1516 * called under RTNL. This is needed if received packets may be
1517 * forwarded to another interface.
1519 void dev_disable_lro(struct net_device *dev)
1521 struct net_device *lower_dev;
1522 struct list_head *iter;
1524 dev->wanted_features &= ~NETIF_F_LRO;
1525 netdev_update_features(dev);
1527 if (unlikely(dev->features & NETIF_F_LRO))
1528 netdev_WARN(dev, "failed to disable LRO!\n");
1530 netdev_for_each_lower_dev(dev, lower_dev, iter)
1531 dev_disable_lro(lower_dev);
1533 EXPORT_SYMBOL(dev_disable_lro);
1535 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val,
1536 struct net_device *dev)
1538 struct netdev_notifier_info info;
1540 netdev_notifier_info_init(&info, dev);
1541 return nb->notifier_call(nb, val, &info);
1544 static int dev_boot_phase = 1;
1547 * register_netdevice_notifier - register a network notifier block
1550 * Register a notifier to be called when network device events occur.
1551 * The notifier passed is linked into the kernel structures and must
1552 * not be reused until it has been unregistered. A negative errno code
1553 * is returned on a failure.
1555 * When registered all registration and up events are replayed
1556 * to the new notifier to allow device to have a race free
1557 * view of the network device list.
1560 int register_netdevice_notifier(struct notifier_block *nb)
1562 struct net_device *dev;
1563 struct net_device *last;
1568 err = raw_notifier_chain_register(&netdev_chain, nb);
1574 for_each_netdev(net, dev) {
1575 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev);
1576 err = notifier_to_errno(err);
1580 if (!(dev->flags & IFF_UP))
1583 call_netdevice_notifier(nb, NETDEV_UP, dev);
1594 for_each_netdev(net, dev) {
1598 if (dev->flags & IFF_UP) {
1599 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1601 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1603 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1608 raw_notifier_chain_unregister(&netdev_chain, nb);
1611 EXPORT_SYMBOL(register_netdevice_notifier);
1614 * unregister_netdevice_notifier - unregister a network notifier block
1617 * Unregister a notifier previously registered by
1618 * register_netdevice_notifier(). The notifier is unlinked into the
1619 * kernel structures and may then be reused. A negative errno code
1620 * is returned on a failure.
1622 * After unregistering unregister and down device events are synthesized
1623 * for all devices on the device list to the removed notifier to remove
1624 * the need for special case cleanup code.
1627 int unregister_netdevice_notifier(struct notifier_block *nb)
1629 struct net_device *dev;
1634 err = raw_notifier_chain_unregister(&netdev_chain, nb);
1639 for_each_netdev(net, dev) {
1640 if (dev->flags & IFF_UP) {
1641 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1643 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1645 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1652 EXPORT_SYMBOL(unregister_netdevice_notifier);
1655 * call_netdevice_notifiers_info - call all network notifier blocks
1656 * @val: value passed unmodified to notifier function
1657 * @dev: net_device pointer passed unmodified to notifier function
1658 * @info: notifier information data
1660 * Call all network notifier blocks. Parameters and return value
1661 * are as for raw_notifier_call_chain().
1664 static int call_netdevice_notifiers_info(unsigned long val,
1665 struct net_device *dev,
1666 struct netdev_notifier_info *info)
1669 netdev_notifier_info_init(info, dev);
1670 return raw_notifier_call_chain(&netdev_chain, val, info);
1674 * call_netdevice_notifiers - call all network notifier blocks
1675 * @val: value passed unmodified to notifier function
1676 * @dev: net_device pointer passed unmodified to notifier function
1678 * Call all network notifier blocks. Parameters and return value
1679 * are as for raw_notifier_call_chain().
1682 int call_netdevice_notifiers(unsigned long val, struct net_device *dev)
1684 struct netdev_notifier_info info;
1686 return call_netdevice_notifiers_info(val, dev, &info);
1688 EXPORT_SYMBOL(call_netdevice_notifiers);
1690 #ifdef CONFIG_NET_INGRESS
1691 static struct static_key ingress_needed __read_mostly;
1693 void net_inc_ingress_queue(void)
1695 static_key_slow_inc(&ingress_needed);
1697 EXPORT_SYMBOL_GPL(net_inc_ingress_queue);
1699 void net_dec_ingress_queue(void)
1701 static_key_slow_dec(&ingress_needed);
1703 EXPORT_SYMBOL_GPL(net_dec_ingress_queue);
1706 #ifdef CONFIG_NET_EGRESS
1707 static struct static_key egress_needed __read_mostly;
1709 void net_inc_egress_queue(void)
1711 static_key_slow_inc(&egress_needed);
1713 EXPORT_SYMBOL_GPL(net_inc_egress_queue);
1715 void net_dec_egress_queue(void)
1717 static_key_slow_dec(&egress_needed);
1719 EXPORT_SYMBOL_GPL(net_dec_egress_queue);
1722 static struct static_key netstamp_needed __read_mostly;
1723 #ifdef HAVE_JUMP_LABEL
1724 static atomic_t netstamp_needed_deferred;
1725 static atomic_t netstamp_wanted;
1726 static void netstamp_clear(struct work_struct *work)
1728 int deferred = atomic_xchg(&netstamp_needed_deferred, 0);
1731 wanted = atomic_add_return(deferred, &netstamp_wanted);
1733 static_key_enable(&netstamp_needed);
1735 static_key_disable(&netstamp_needed);
1737 static DECLARE_WORK(netstamp_work, netstamp_clear);
1740 void net_enable_timestamp(void)
1742 #ifdef HAVE_JUMP_LABEL
1746 wanted = atomic_read(&netstamp_wanted);
1749 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted + 1) == wanted)
1752 atomic_inc(&netstamp_needed_deferred);
1753 schedule_work(&netstamp_work);
1755 static_key_slow_inc(&netstamp_needed);
1758 EXPORT_SYMBOL(net_enable_timestamp);
1760 void net_disable_timestamp(void)
1762 #ifdef HAVE_JUMP_LABEL
1766 wanted = atomic_read(&netstamp_wanted);
1769 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted - 1) == wanted)
1772 atomic_dec(&netstamp_needed_deferred);
1773 schedule_work(&netstamp_work);
1775 static_key_slow_dec(&netstamp_needed);
1778 EXPORT_SYMBOL(net_disable_timestamp);
1780 static inline void net_timestamp_set(struct sk_buff *skb)
1783 if (static_key_false(&netstamp_needed))
1784 __net_timestamp(skb);
1787 #define net_timestamp_check(COND, SKB) \
1788 if (static_key_false(&netstamp_needed)) { \
1789 if ((COND) && !(SKB)->tstamp) \
1790 __net_timestamp(SKB); \
1793 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb)
1797 if (!(dev->flags & IFF_UP))
1800 len = dev->mtu + dev->hard_header_len + VLAN_HLEN;
1801 if (skb->len <= len)
1804 /* if TSO is enabled, we don't care about the length as the packet
1805 * could be forwarded without being segmented before
1807 if (skb_is_gso(skb))
1812 EXPORT_SYMBOL_GPL(is_skb_forwardable);
1814 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1816 int ret = ____dev_forward_skb(dev, skb);
1819 skb->protocol = eth_type_trans(skb, dev);
1820 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
1825 EXPORT_SYMBOL_GPL(__dev_forward_skb);
1828 * dev_forward_skb - loopback an skb to another netif
1830 * @dev: destination network device
1831 * @skb: buffer to forward
1834 * NET_RX_SUCCESS (no congestion)
1835 * NET_RX_DROP (packet was dropped, but freed)
1837 * dev_forward_skb can be used for injecting an skb from the
1838 * start_xmit function of one device into the receive queue
1839 * of another device.
1841 * The receiving device may be in another namespace, so
1842 * we have to clear all information in the skb that could
1843 * impact namespace isolation.
1845 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1847 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb);
1849 EXPORT_SYMBOL_GPL(dev_forward_skb);
1851 static inline int deliver_skb(struct sk_buff *skb,
1852 struct packet_type *pt_prev,
1853 struct net_device *orig_dev)
1855 if (unlikely(skb_orphan_frags(skb, GFP_ATOMIC)))
1857 refcount_inc(&skb->users);
1858 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
1861 static inline void deliver_ptype_list_skb(struct sk_buff *skb,
1862 struct packet_type **pt,
1863 struct net_device *orig_dev,
1865 struct list_head *ptype_list)
1867 struct packet_type *ptype, *pt_prev = *pt;
1869 list_for_each_entry_rcu(ptype, ptype_list, list) {
1870 if (ptype->type != type)
1873 deliver_skb(skb, pt_prev, orig_dev);
1879 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb)
1881 if (!ptype->af_packet_priv || !skb->sk)
1884 if (ptype->id_match)
1885 return ptype->id_match(ptype, skb->sk);
1886 else if ((struct sock *)ptype->af_packet_priv == skb->sk)
1893 * Support routine. Sends outgoing frames to any network
1894 * taps currently in use.
1897 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
1899 struct packet_type *ptype;
1900 struct sk_buff *skb2 = NULL;
1901 struct packet_type *pt_prev = NULL;
1902 struct list_head *ptype_list = &ptype_all;
1906 list_for_each_entry_rcu(ptype, ptype_list, list) {
1907 /* Never send packets back to the socket
1908 * they originated from - MvS (miquels@drinkel.ow.org)
1910 if (skb_loop_sk(ptype, skb))
1914 deliver_skb(skb2, pt_prev, skb->dev);
1919 /* need to clone skb, done only once */
1920 skb2 = skb_clone(skb, GFP_ATOMIC);
1924 net_timestamp_set(skb2);
1926 /* skb->nh should be correctly
1927 * set by sender, so that the second statement is
1928 * just protection against buggy protocols.
1930 skb_reset_mac_header(skb2);
1932 if (skb_network_header(skb2) < skb2->data ||
1933 skb_network_header(skb2) > skb_tail_pointer(skb2)) {
1934 net_crit_ratelimited("protocol %04x is buggy, dev %s\n",
1935 ntohs(skb2->protocol),
1937 skb_reset_network_header(skb2);
1940 skb2->transport_header = skb2->network_header;
1941 skb2->pkt_type = PACKET_OUTGOING;
1945 if (ptype_list == &ptype_all) {
1946 ptype_list = &dev->ptype_all;
1951 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
1954 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit);
1957 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change
1958 * @dev: Network device
1959 * @txq: number of queues available
1961 * If real_num_tx_queues is changed the tc mappings may no longer be
1962 * valid. To resolve this verify the tc mapping remains valid and if
1963 * not NULL the mapping. With no priorities mapping to this
1964 * offset/count pair it will no longer be used. In the worst case TC0
1965 * is invalid nothing can be done so disable priority mappings. If is
1966 * expected that drivers will fix this mapping if they can before
1967 * calling netif_set_real_num_tx_queues.
1969 static void netif_setup_tc(struct net_device *dev, unsigned int txq)
1972 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
1974 /* If TC0 is invalidated disable TC mapping */
1975 if (tc->offset + tc->count > txq) {
1976 pr_warn("Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n");
1981 /* Invalidated prio to tc mappings set to TC0 */
1982 for (i = 1; i < TC_BITMASK + 1; i++) {
1983 int q = netdev_get_prio_tc_map(dev, i);
1985 tc = &dev->tc_to_txq[q];
1986 if (tc->offset + tc->count > txq) {
1987 pr_warn("Number of in use tx queues changed. Priority %i to tc mapping %i is no longer valid. Setting map to 0\n",
1989 netdev_set_prio_tc_map(dev, i, 0);
1994 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq)
1997 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2000 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) {
2001 if ((txq - tc->offset) < tc->count)
2012 static DEFINE_MUTEX(xps_map_mutex);
2013 #define xmap_dereference(P) \
2014 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex))
2016 static bool remove_xps_queue(struct xps_dev_maps *dev_maps,
2019 struct xps_map *map = NULL;
2023 map = xmap_dereference(dev_maps->cpu_map[tci]);
2027 for (pos = map->len; pos--;) {
2028 if (map->queues[pos] != index)
2032 map->queues[pos] = map->queues[--map->len];
2036 RCU_INIT_POINTER(dev_maps->cpu_map[tci], NULL);
2037 kfree_rcu(map, rcu);
2044 static bool remove_xps_queue_cpu(struct net_device *dev,
2045 struct xps_dev_maps *dev_maps,
2046 int cpu, u16 offset, u16 count)
2048 int num_tc = dev->num_tc ? : 1;
2049 bool active = false;
2052 for (tci = cpu * num_tc; num_tc--; tci++) {
2055 for (i = count, j = offset; i--; j++) {
2056 if (!remove_xps_queue(dev_maps, cpu, j))
2066 static void netif_reset_xps_queues(struct net_device *dev, u16 offset,
2069 struct xps_dev_maps *dev_maps;
2071 bool active = false;
2073 mutex_lock(&xps_map_mutex);
2074 dev_maps = xmap_dereference(dev->xps_maps);
2079 for_each_possible_cpu(cpu)
2080 active |= remove_xps_queue_cpu(dev, dev_maps, cpu,
2084 RCU_INIT_POINTER(dev->xps_maps, NULL);
2085 kfree_rcu(dev_maps, rcu);
2088 for (i = offset + (count - 1); count--; i--)
2089 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, i),
2093 mutex_unlock(&xps_map_mutex);
2096 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index)
2098 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index);
2101 static struct xps_map *expand_xps_map(struct xps_map *map,
2104 struct xps_map *new_map;
2105 int alloc_len = XPS_MIN_MAP_ALLOC;
2108 for (pos = 0; map && pos < map->len; pos++) {
2109 if (map->queues[pos] != index)
2114 /* Need to add queue to this CPU's existing map */
2116 if (pos < map->alloc_len)
2119 alloc_len = map->alloc_len * 2;
2122 /* Need to allocate new map to store queue on this CPU's map */
2123 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL,
2128 for (i = 0; i < pos; i++)
2129 new_map->queues[i] = map->queues[i];
2130 new_map->alloc_len = alloc_len;
2136 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask,
2139 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL;
2140 int i, cpu, tci, numa_node_id = -2;
2141 int maps_sz, num_tc = 1, tc = 0;
2142 struct xps_map *map, *new_map;
2143 bool active = false;
2146 num_tc = dev->num_tc;
2147 tc = netdev_txq_to_tc(dev, index);
2152 maps_sz = XPS_DEV_MAPS_SIZE(num_tc);
2153 if (maps_sz < L1_CACHE_BYTES)
2154 maps_sz = L1_CACHE_BYTES;
2156 mutex_lock(&xps_map_mutex);
2158 dev_maps = xmap_dereference(dev->xps_maps);
2160 /* allocate memory for queue storage */
2161 for_each_cpu_and(cpu, cpu_online_mask, mask) {
2163 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL);
2164 if (!new_dev_maps) {
2165 mutex_unlock(&xps_map_mutex);
2169 tci = cpu * num_tc + tc;
2170 map = dev_maps ? xmap_dereference(dev_maps->cpu_map[tci]) :
2173 map = expand_xps_map(map, cpu, index);
2177 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2181 goto out_no_new_maps;
2183 for_each_possible_cpu(cpu) {
2184 /* copy maps belonging to foreign traffic classes */
2185 for (i = tc, tci = cpu * num_tc; dev_maps && i--; tci++) {
2186 /* fill in the new device map from the old device map */
2187 map = xmap_dereference(dev_maps->cpu_map[tci]);
2188 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2191 /* We need to explicitly update tci as prevous loop
2192 * could break out early if dev_maps is NULL.
2194 tci = cpu * num_tc + tc;
2196 if (cpumask_test_cpu(cpu, mask) && cpu_online(cpu)) {
2197 /* add queue to CPU maps */
2200 map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2201 while ((pos < map->len) && (map->queues[pos] != index))
2204 if (pos == map->len)
2205 map->queues[map->len++] = index;
2207 if (numa_node_id == -2)
2208 numa_node_id = cpu_to_node(cpu);
2209 else if (numa_node_id != cpu_to_node(cpu))
2212 } else if (dev_maps) {
2213 /* fill in the new device map from the old device map */
2214 map = xmap_dereference(dev_maps->cpu_map[tci]);
2215 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2218 /* copy maps belonging to foreign traffic classes */
2219 for (i = num_tc - tc, tci++; dev_maps && --i; tci++) {
2220 /* fill in the new device map from the old device map */
2221 map = xmap_dereference(dev_maps->cpu_map[tci]);
2222 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2226 rcu_assign_pointer(dev->xps_maps, new_dev_maps);
2228 /* Cleanup old maps */
2230 goto out_no_old_maps;
2232 for_each_possible_cpu(cpu) {
2233 for (i = num_tc, tci = cpu * num_tc; i--; tci++) {
2234 new_map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2235 map = xmap_dereference(dev_maps->cpu_map[tci]);
2236 if (map && map != new_map)
2237 kfree_rcu(map, rcu);
2241 kfree_rcu(dev_maps, rcu);
2244 dev_maps = new_dev_maps;
2248 /* update Tx queue numa node */
2249 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index),
2250 (numa_node_id >= 0) ? numa_node_id :
2256 /* removes queue from unused CPUs */
2257 for_each_possible_cpu(cpu) {
2258 for (i = tc, tci = cpu * num_tc; i--; tci++)
2259 active |= remove_xps_queue(dev_maps, tci, index);
2260 if (!cpumask_test_cpu(cpu, mask) || !cpu_online(cpu))
2261 active |= remove_xps_queue(dev_maps, tci, index);
2262 for (i = num_tc - tc, tci++; --i; tci++)
2263 active |= remove_xps_queue(dev_maps, tci, index);
2266 /* free map if not active */
2268 RCU_INIT_POINTER(dev->xps_maps, NULL);
2269 kfree_rcu(dev_maps, rcu);
2273 mutex_unlock(&xps_map_mutex);
2277 /* remove any maps that we added */
2278 for_each_possible_cpu(cpu) {
2279 for (i = num_tc, tci = cpu * num_tc; i--; tci++) {
2280 new_map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2282 xmap_dereference(dev_maps->cpu_map[tci]) :
2284 if (new_map && new_map != map)
2289 mutex_unlock(&xps_map_mutex);
2291 kfree(new_dev_maps);
2294 EXPORT_SYMBOL(netif_set_xps_queue);
2297 void netdev_reset_tc(struct net_device *dev)
2300 netif_reset_xps_queues_gt(dev, 0);
2303 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq));
2304 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map));
2306 EXPORT_SYMBOL(netdev_reset_tc);
2308 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset)
2310 if (tc >= dev->num_tc)
2314 netif_reset_xps_queues(dev, offset, count);
2316 dev->tc_to_txq[tc].count = count;
2317 dev->tc_to_txq[tc].offset = offset;
2320 EXPORT_SYMBOL(netdev_set_tc_queue);
2322 int netdev_set_num_tc(struct net_device *dev, u8 num_tc)
2324 if (num_tc > TC_MAX_QUEUE)
2328 netif_reset_xps_queues_gt(dev, 0);
2330 dev->num_tc = num_tc;
2333 EXPORT_SYMBOL(netdev_set_num_tc);
2336 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues
2337 * greater then real_num_tx_queues stale skbs on the qdisc must be flushed.
2339 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq)
2343 if (txq < 1 || txq > dev->num_tx_queues)
2346 if (dev->reg_state == NETREG_REGISTERED ||
2347 dev->reg_state == NETREG_UNREGISTERING) {
2350 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues,
2356 netif_setup_tc(dev, txq);
2358 if (txq < dev->real_num_tx_queues) {
2359 qdisc_reset_all_tx_gt(dev, txq);
2361 netif_reset_xps_queues_gt(dev, txq);
2366 dev->real_num_tx_queues = txq;
2369 EXPORT_SYMBOL(netif_set_real_num_tx_queues);
2373 * netif_set_real_num_rx_queues - set actual number of RX queues used
2374 * @dev: Network device
2375 * @rxq: Actual number of RX queues
2377 * This must be called either with the rtnl_lock held or before
2378 * registration of the net device. Returns 0 on success, or a
2379 * negative error code. If called before registration, it always
2382 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq)
2386 if (rxq < 1 || rxq > dev->num_rx_queues)
2389 if (dev->reg_state == NETREG_REGISTERED) {
2392 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues,
2398 dev->real_num_rx_queues = rxq;
2401 EXPORT_SYMBOL(netif_set_real_num_rx_queues);
2405 * netif_get_num_default_rss_queues - default number of RSS queues
2407 * This routine should set an upper limit on the number of RSS queues
2408 * used by default by multiqueue devices.
2410 int netif_get_num_default_rss_queues(void)
2412 return is_kdump_kernel() ?
2413 1 : min_t(int, DEFAULT_MAX_NUM_RSS_QUEUES, num_online_cpus());
2415 EXPORT_SYMBOL(netif_get_num_default_rss_queues);
2417 static void __netif_reschedule(struct Qdisc *q)
2419 struct softnet_data *sd;
2420 unsigned long flags;
2422 local_irq_save(flags);
2423 sd = this_cpu_ptr(&softnet_data);
2424 q->next_sched = NULL;
2425 *sd->output_queue_tailp = q;
2426 sd->output_queue_tailp = &q->next_sched;
2427 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2428 local_irq_restore(flags);
2431 void __netif_schedule(struct Qdisc *q)
2433 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
2434 __netif_reschedule(q);
2436 EXPORT_SYMBOL(__netif_schedule);
2438 struct dev_kfree_skb_cb {
2439 enum skb_free_reason reason;
2442 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb)
2444 return (struct dev_kfree_skb_cb *)skb->cb;
2447 void netif_schedule_queue(struct netdev_queue *txq)
2450 if (!(txq->state & QUEUE_STATE_ANY_XOFF)) {
2451 struct Qdisc *q = rcu_dereference(txq->qdisc);
2453 __netif_schedule(q);
2457 EXPORT_SYMBOL(netif_schedule_queue);
2459 void netif_tx_wake_queue(struct netdev_queue *dev_queue)
2461 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) {
2465 q = rcu_dereference(dev_queue->qdisc);
2466 __netif_schedule(q);
2470 EXPORT_SYMBOL(netif_tx_wake_queue);
2472 void __dev_kfree_skb_irq(struct sk_buff *skb, enum skb_free_reason reason)
2474 unsigned long flags;
2479 if (likely(refcount_read(&skb->users) == 1)) {
2481 refcount_set(&skb->users, 0);
2482 } else if (likely(!refcount_dec_and_test(&skb->users))) {
2485 get_kfree_skb_cb(skb)->reason = reason;
2486 local_irq_save(flags);
2487 skb->next = __this_cpu_read(softnet_data.completion_queue);
2488 __this_cpu_write(softnet_data.completion_queue, skb);
2489 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2490 local_irq_restore(flags);
2492 EXPORT_SYMBOL(__dev_kfree_skb_irq);
2494 void __dev_kfree_skb_any(struct sk_buff *skb, enum skb_free_reason reason)
2496 if (in_irq() || irqs_disabled())
2497 __dev_kfree_skb_irq(skb, reason);
2501 EXPORT_SYMBOL(__dev_kfree_skb_any);
2505 * netif_device_detach - mark device as removed
2506 * @dev: network device
2508 * Mark device as removed from system and therefore no longer available.
2510 void netif_device_detach(struct net_device *dev)
2512 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) &&
2513 netif_running(dev)) {
2514 netif_tx_stop_all_queues(dev);
2517 EXPORT_SYMBOL(netif_device_detach);
2520 * netif_device_attach - mark device as attached
2521 * @dev: network device
2523 * Mark device as attached from system and restart if needed.
2525 void netif_device_attach(struct net_device *dev)
2527 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) &&
2528 netif_running(dev)) {
2529 netif_tx_wake_all_queues(dev);
2530 __netdev_watchdog_up(dev);
2533 EXPORT_SYMBOL(netif_device_attach);
2536 * Returns a Tx hash based on the given packet descriptor a Tx queues' number
2537 * to be used as a distribution range.
2539 u16 __skb_tx_hash(const struct net_device *dev, struct sk_buff *skb,
2540 unsigned int num_tx_queues)
2544 u16 qcount = num_tx_queues;
2546 if (skb_rx_queue_recorded(skb)) {
2547 hash = skb_get_rx_queue(skb);
2548 while (unlikely(hash >= num_tx_queues))
2549 hash -= num_tx_queues;
2554 u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
2556 qoffset = dev->tc_to_txq[tc].offset;
2557 qcount = dev->tc_to_txq[tc].count;
2560 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset;
2562 EXPORT_SYMBOL(__skb_tx_hash);
2564 static void skb_warn_bad_offload(const struct sk_buff *skb)
2566 static const netdev_features_t null_features;
2567 struct net_device *dev = skb->dev;
2568 const char *name = "";
2570 if (!net_ratelimit())
2574 if (dev->dev.parent)
2575 name = dev_driver_string(dev->dev.parent);
2577 name = netdev_name(dev);
2579 WARN(1, "%s: caps=(%pNF, %pNF) len=%d data_len=%d gso_size=%d "
2580 "gso_type=%d ip_summed=%d\n",
2581 name, dev ? &dev->features : &null_features,
2582 skb->sk ? &skb->sk->sk_route_caps : &null_features,
2583 skb->len, skb->data_len, skb_shinfo(skb)->gso_size,
2584 skb_shinfo(skb)->gso_type, skb->ip_summed);
2588 * Invalidate hardware checksum when packet is to be mangled, and
2589 * complete checksum manually on outgoing path.
2591 int skb_checksum_help(struct sk_buff *skb)
2594 int ret = 0, offset;
2596 if (skb->ip_summed == CHECKSUM_COMPLETE)
2597 goto out_set_summed;
2599 if (unlikely(skb_shinfo(skb)->gso_size)) {
2600 skb_warn_bad_offload(skb);
2604 /* Before computing a checksum, we should make sure no frag could
2605 * be modified by an external entity : checksum could be wrong.
2607 if (skb_has_shared_frag(skb)) {
2608 ret = __skb_linearize(skb);
2613 offset = skb_checksum_start_offset(skb);
2614 BUG_ON(offset >= skb_headlen(skb));
2615 csum = skb_checksum(skb, offset, skb->len - offset, 0);
2617 offset += skb->csum_offset;
2618 BUG_ON(offset + sizeof(__sum16) > skb_headlen(skb));
2620 if (skb_cloned(skb) &&
2621 !skb_clone_writable(skb, offset + sizeof(__sum16))) {
2622 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2627 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0;
2629 skb->ip_summed = CHECKSUM_NONE;
2633 EXPORT_SYMBOL(skb_checksum_help);
2635 int skb_crc32c_csum_help(struct sk_buff *skb)
2638 int ret = 0, offset, start;
2640 if (skb->ip_summed != CHECKSUM_PARTIAL)
2643 if (unlikely(skb_is_gso(skb)))
2646 /* Before computing a checksum, we should make sure no frag could
2647 * be modified by an external entity : checksum could be wrong.
2649 if (unlikely(skb_has_shared_frag(skb))) {
2650 ret = __skb_linearize(skb);
2654 start = skb_checksum_start_offset(skb);
2655 offset = start + offsetof(struct sctphdr, checksum);
2656 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) {
2660 if (skb_cloned(skb) &&
2661 !skb_clone_writable(skb, offset + sizeof(__le32))) {
2662 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2666 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start,
2667 skb->len - start, ~(__u32)0,
2669 *(__le32 *)(skb->data + offset) = crc32c_csum;
2670 skb->ip_summed = CHECKSUM_NONE;
2671 skb->csum_not_inet = 0;
2676 __be16 skb_network_protocol(struct sk_buff *skb, int *depth)
2678 __be16 type = skb->protocol;
2680 /* Tunnel gso handlers can set protocol to ethernet. */
2681 if (type == htons(ETH_P_TEB)) {
2684 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr))))
2687 eth = (struct ethhdr *)skb_mac_header(skb);
2688 type = eth->h_proto;
2691 return __vlan_get_protocol(skb, type, depth);
2695 * skb_mac_gso_segment - mac layer segmentation handler.
2696 * @skb: buffer to segment
2697 * @features: features for the output path (see dev->features)
2699 struct sk_buff *skb_mac_gso_segment(struct sk_buff *skb,
2700 netdev_features_t features)
2702 struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT);
2703 struct packet_offload *ptype;
2704 int vlan_depth = skb->mac_len;
2705 __be16 type = skb_network_protocol(skb, &vlan_depth);
2707 if (unlikely(!type))
2708 return ERR_PTR(-EINVAL);
2710 __skb_pull(skb, vlan_depth);
2713 list_for_each_entry_rcu(ptype, &offload_base, list) {
2714 if (ptype->type == type && ptype->callbacks.gso_segment) {
2715 segs = ptype->callbacks.gso_segment(skb, features);
2721 __skb_push(skb, skb->data - skb_mac_header(skb));
2725 EXPORT_SYMBOL(skb_mac_gso_segment);
2728 /* openvswitch calls this on rx path, so we need a different check.
2730 static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
2733 return skb->ip_summed != CHECKSUM_PARTIAL &&
2734 skb->ip_summed != CHECKSUM_NONE;
2736 return skb->ip_summed == CHECKSUM_NONE;
2740 * __skb_gso_segment - Perform segmentation on skb.
2741 * @skb: buffer to segment
2742 * @features: features for the output path (see dev->features)
2743 * @tx_path: whether it is called in TX path
2745 * This function segments the given skb and returns a list of segments.
2747 * It may return NULL if the skb requires no segmentation. This is
2748 * only possible when GSO is used for verifying header integrity.
2750 * Segmentation preserves SKB_SGO_CB_OFFSET bytes of previous skb cb.
2752 struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
2753 netdev_features_t features, bool tx_path)
2755 struct sk_buff *segs;
2757 if (unlikely(skb_needs_check(skb, tx_path))) {
2760 /* We're going to init ->check field in TCP or UDP header */
2761 err = skb_cow_head(skb, 0);
2763 return ERR_PTR(err);
2766 /* Only report GSO partial support if it will enable us to
2767 * support segmentation on this frame without needing additional
2770 if (features & NETIF_F_GSO_PARTIAL) {
2771 netdev_features_t partial_features = NETIF_F_GSO_ROBUST;
2772 struct net_device *dev = skb->dev;
2774 partial_features |= dev->features & dev->gso_partial_features;
2775 if (!skb_gso_ok(skb, features | partial_features))
2776 features &= ~NETIF_F_GSO_PARTIAL;
2779 BUILD_BUG_ON(SKB_SGO_CB_OFFSET +
2780 sizeof(*SKB_GSO_CB(skb)) > sizeof(skb->cb));
2782 SKB_GSO_CB(skb)->mac_offset = skb_headroom(skb);
2783 SKB_GSO_CB(skb)->encap_level = 0;
2785 skb_reset_mac_header(skb);
2786 skb_reset_mac_len(skb);
2788 segs = skb_mac_gso_segment(skb, features);
2790 if (unlikely(skb_needs_check(skb, tx_path)))
2791 skb_warn_bad_offload(skb);
2795 EXPORT_SYMBOL(__skb_gso_segment);
2797 /* Take action when hardware reception checksum errors are detected. */
2799 void netdev_rx_csum_fault(struct net_device *dev)
2801 if (net_ratelimit()) {
2802 pr_err("%s: hw csum failure\n", dev ? dev->name : "<unknown>");
2806 EXPORT_SYMBOL(netdev_rx_csum_fault);
2809 /* Actually, we should eliminate this check as soon as we know, that:
2810 * 1. IOMMU is present and allows to map all the memory.
2811 * 2. No high memory really exists on this machine.
2814 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
2816 #ifdef CONFIG_HIGHMEM
2819 if (!(dev->features & NETIF_F_HIGHDMA)) {
2820 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2821 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2823 if (PageHighMem(skb_frag_page(frag)))
2828 if (PCI_DMA_BUS_IS_PHYS) {
2829 struct device *pdev = dev->dev.parent;
2833 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2834 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2835 dma_addr_t addr = page_to_phys(skb_frag_page(frag));
2837 if (!pdev->dma_mask || addr + PAGE_SIZE - 1 > *pdev->dma_mask)
2845 /* If MPLS offload request, verify we are testing hardware MPLS features
2846 * instead of standard features for the netdev.
2848 #if IS_ENABLED(CONFIG_NET_MPLS_GSO)
2849 static netdev_features_t net_mpls_features(struct sk_buff *skb,
2850 netdev_features_t features,
2853 if (eth_p_mpls(type))
2854 features &= skb->dev->mpls_features;
2859 static netdev_features_t net_mpls_features(struct sk_buff *skb,
2860 netdev_features_t features,
2867 static netdev_features_t harmonize_features(struct sk_buff *skb,
2868 netdev_features_t features)
2873 type = skb_network_protocol(skb, &tmp);
2874 features = net_mpls_features(skb, features, type);
2876 if (skb->ip_summed != CHECKSUM_NONE &&
2877 !can_checksum_protocol(features, type)) {
2878 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
2880 if (illegal_highdma(skb->dev, skb))
2881 features &= ~NETIF_F_SG;
2886 netdev_features_t passthru_features_check(struct sk_buff *skb,
2887 struct net_device *dev,
2888 netdev_features_t features)
2892 EXPORT_SYMBOL(passthru_features_check);
2894 static netdev_features_t dflt_features_check(const struct sk_buff *skb,
2895 struct net_device *dev,
2896 netdev_features_t features)
2898 return vlan_features_check(skb, features);
2901 static netdev_features_t gso_features_check(const struct sk_buff *skb,
2902 struct net_device *dev,
2903 netdev_features_t features)
2905 u16 gso_segs = skb_shinfo(skb)->gso_segs;
2907 if (gso_segs > dev->gso_max_segs)
2908 return features & ~NETIF_F_GSO_MASK;
2910 /* Support for GSO partial features requires software
2911 * intervention before we can actually process the packets
2912 * so we need to strip support for any partial features now
2913 * and we can pull them back in after we have partially
2914 * segmented the frame.
2916 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL))
2917 features &= ~dev->gso_partial_features;
2919 /* Make sure to clear the IPv4 ID mangling feature if the
2920 * IPv4 header has the potential to be fragmented.
2922 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
2923 struct iphdr *iph = skb->encapsulation ?
2924 inner_ip_hdr(skb) : ip_hdr(skb);
2926 if (!(iph->frag_off & htons(IP_DF)))
2927 features &= ~NETIF_F_TSO_MANGLEID;
2933 netdev_features_t netif_skb_features(struct sk_buff *skb)
2935 struct net_device *dev = skb->dev;
2936 netdev_features_t features = dev->features;
2938 if (skb_is_gso(skb))
2939 features = gso_features_check(skb, dev, features);
2941 /* If encapsulation offload request, verify we are testing
2942 * hardware encapsulation features instead of standard
2943 * features for the netdev
2945 if (skb->encapsulation)
2946 features &= dev->hw_enc_features;
2948 if (skb_vlan_tagged(skb))
2949 features = netdev_intersect_features(features,
2950 dev->vlan_features |
2951 NETIF_F_HW_VLAN_CTAG_TX |
2952 NETIF_F_HW_VLAN_STAG_TX);
2954 if (dev->netdev_ops->ndo_features_check)
2955 features &= dev->netdev_ops->ndo_features_check(skb, dev,
2958 features &= dflt_features_check(skb, dev, features);
2960 return harmonize_features(skb, features);
2962 EXPORT_SYMBOL(netif_skb_features);
2964 static int xmit_one(struct sk_buff *skb, struct net_device *dev,
2965 struct netdev_queue *txq, bool more)
2970 if (!list_empty(&ptype_all) || !list_empty(&dev->ptype_all))
2971 dev_queue_xmit_nit(skb, dev);
2974 trace_net_dev_start_xmit(skb, dev);
2975 rc = netdev_start_xmit(skb, dev, txq, more);
2976 trace_net_dev_xmit(skb, rc, dev, len);
2981 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev,
2982 struct netdev_queue *txq, int *ret)
2984 struct sk_buff *skb = first;
2985 int rc = NETDEV_TX_OK;
2988 struct sk_buff *next = skb->next;
2991 rc = xmit_one(skb, dev, txq, next != NULL);
2992 if (unlikely(!dev_xmit_complete(rc))) {
2998 if (netif_xmit_stopped(txq) && skb) {
2999 rc = NETDEV_TX_BUSY;
3009 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb,
3010 netdev_features_t features)
3012 if (skb_vlan_tag_present(skb) &&
3013 !vlan_hw_offload_capable(features, skb->vlan_proto))
3014 skb = __vlan_hwaccel_push_inside(skb);
3018 int skb_csum_hwoffload_help(struct sk_buff *skb,
3019 const netdev_features_t features)
3021 if (unlikely(skb->csum_not_inet))
3022 return !!(features & NETIF_F_SCTP_CRC) ? 0 :
3023 skb_crc32c_csum_help(skb);
3025 return !!(features & NETIF_F_CSUM_MASK) ? 0 : skb_checksum_help(skb);
3027 EXPORT_SYMBOL(skb_csum_hwoffload_help);
3029 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev)
3031 netdev_features_t features;
3033 features = netif_skb_features(skb);
3034 skb = validate_xmit_vlan(skb, features);
3038 if (netif_needs_gso(skb, features)) {
3039 struct sk_buff *segs;
3041 segs = skb_gso_segment(skb, features);
3049 if (skb_needs_linearize(skb, features) &&
3050 __skb_linearize(skb))
3053 if (validate_xmit_xfrm(skb, features))
3056 /* If packet is not checksummed and device does not
3057 * support checksumming for this protocol, complete
3058 * checksumming here.
3060 if (skb->ip_summed == CHECKSUM_PARTIAL) {
3061 if (skb->encapsulation)
3062 skb_set_inner_transport_header(skb,
3063 skb_checksum_start_offset(skb));
3065 skb_set_transport_header(skb,
3066 skb_checksum_start_offset(skb));
3067 if (skb_csum_hwoffload_help(skb, features))
3077 atomic_long_inc(&dev->tx_dropped);
3081 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev)
3083 struct sk_buff *next, *head = NULL, *tail;
3085 for (; skb != NULL; skb = next) {
3089 /* in case skb wont be segmented, point to itself */
3092 skb = validate_xmit_skb(skb, dev);
3100 /* If skb was segmented, skb->prev points to
3101 * the last segment. If not, it still contains skb.
3107 EXPORT_SYMBOL_GPL(validate_xmit_skb_list);
3109 static void qdisc_pkt_len_init(struct sk_buff *skb)
3111 const struct skb_shared_info *shinfo = skb_shinfo(skb);
3113 qdisc_skb_cb(skb)->pkt_len = skb->len;
3115 /* To get more precise estimation of bytes sent on wire,
3116 * we add to pkt_len the headers size of all segments
3118 if (shinfo->gso_size) {
3119 unsigned int hdr_len;
3120 u16 gso_segs = shinfo->gso_segs;
3122 /* mac layer + network layer */
3123 hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
3125 /* + transport layer */
3126 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))
3127 hdr_len += tcp_hdrlen(skb);
3129 hdr_len += sizeof(struct udphdr);
3131 if (shinfo->gso_type & SKB_GSO_DODGY)
3132 gso_segs = DIV_ROUND_UP(skb->len - hdr_len,
3135 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len;
3139 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
3140 struct net_device *dev,
3141 struct netdev_queue *txq)
3143 spinlock_t *root_lock = qdisc_lock(q);
3144 struct sk_buff *to_free = NULL;
3148 qdisc_calculate_pkt_len(skb, q);
3150 * Heuristic to force contended enqueues to serialize on a
3151 * separate lock before trying to get qdisc main lock.
3152 * This permits qdisc->running owner to get the lock more
3153 * often and dequeue packets faster.
3155 contended = qdisc_is_running(q);
3156 if (unlikely(contended))
3157 spin_lock(&q->busylock);
3159 spin_lock(root_lock);
3160 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3161 __qdisc_drop(skb, &to_free);
3163 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
3164 qdisc_run_begin(q)) {
3166 * This is a work-conserving queue; there are no old skbs
3167 * waiting to be sent out; and the qdisc is not running -
3168 * xmit the skb directly.
3171 qdisc_bstats_update(q, skb);
3173 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) {
3174 if (unlikely(contended)) {
3175 spin_unlock(&q->busylock);
3182 rc = NET_XMIT_SUCCESS;
3184 rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
3185 if (qdisc_run_begin(q)) {
3186 if (unlikely(contended)) {
3187 spin_unlock(&q->busylock);
3193 spin_unlock(root_lock);
3194 if (unlikely(to_free))
3195 kfree_skb_list(to_free);
3196 if (unlikely(contended))
3197 spin_unlock(&q->busylock);
3201 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO)
3202 static void skb_update_prio(struct sk_buff *skb)
3204 struct netprio_map *map = rcu_dereference_bh(skb->dev->priomap);
3206 if (!skb->priority && skb->sk && map) {
3207 unsigned int prioidx =
3208 sock_cgroup_prioidx(&skb->sk->sk_cgrp_data);
3210 if (prioidx < map->priomap_len)
3211 skb->priority = map->priomap[prioidx];
3215 #define skb_update_prio(skb)
3218 DEFINE_PER_CPU(int, xmit_recursion);
3219 EXPORT_SYMBOL(xmit_recursion);
3222 * dev_loopback_xmit - loop back @skb
3223 * @net: network namespace this loopback is happening in
3224 * @sk: sk needed to be a netfilter okfn
3225 * @skb: buffer to transmit
3227 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb)
3229 skb_reset_mac_header(skb);
3230 __skb_pull(skb, skb_network_offset(skb));
3231 skb->pkt_type = PACKET_LOOPBACK;
3232 skb->ip_summed = CHECKSUM_UNNECESSARY;
3233 WARN_ON(!skb_dst(skb));
3238 EXPORT_SYMBOL(dev_loopback_xmit);
3240 #ifdef CONFIG_NET_EGRESS
3241 static struct sk_buff *
3242 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
3244 struct tcf_proto *cl = rcu_dereference_bh(dev->egress_cl_list);
3245 struct tcf_result cl_res;
3250 /* qdisc_skb_cb(skb)->pkt_len was already set by the caller. */
3251 qdisc_bstats_cpu_update(cl->q, skb);
3253 switch (tcf_classify(skb, cl, &cl_res, false)) {
3255 case TC_ACT_RECLASSIFY:
3256 skb->tc_index = TC_H_MIN(cl_res.classid);
3259 qdisc_qstats_cpu_drop(cl->q);
3260 *ret = NET_XMIT_DROP;
3266 *ret = NET_XMIT_SUCCESS;
3269 case TC_ACT_REDIRECT:
3270 /* No need to push/pop skb's mac_header here on egress! */
3271 skb_do_redirect(skb);
3272 *ret = NET_XMIT_SUCCESS;
3280 #endif /* CONFIG_NET_EGRESS */
3282 static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb)
3285 struct xps_dev_maps *dev_maps;
3286 struct xps_map *map;
3287 int queue_index = -1;
3290 dev_maps = rcu_dereference(dev->xps_maps);
3292 unsigned int tci = skb->sender_cpu - 1;
3296 tci += netdev_get_prio_tc_map(dev, skb->priority);
3299 map = rcu_dereference(dev_maps->cpu_map[tci]);
3302 queue_index = map->queues[0];
3304 queue_index = map->queues[reciprocal_scale(skb_get_hash(skb),
3306 if (unlikely(queue_index >= dev->real_num_tx_queues))
3318 static u16 __netdev_pick_tx(struct net_device *dev, struct sk_buff *skb)
3320 struct sock *sk = skb->sk;
3321 int queue_index = sk_tx_queue_get(sk);
3323 if (queue_index < 0 || skb->ooo_okay ||
3324 queue_index >= dev->real_num_tx_queues) {
3325 int new_index = get_xps_queue(dev, skb);
3328 new_index = skb_tx_hash(dev, skb);
3330 if (queue_index != new_index && sk &&
3332 rcu_access_pointer(sk->sk_dst_cache))
3333 sk_tx_queue_set(sk, new_index);
3335 queue_index = new_index;
3341 struct netdev_queue *netdev_pick_tx(struct net_device *dev,
3342 struct sk_buff *skb,
3345 int queue_index = 0;
3348 u32 sender_cpu = skb->sender_cpu - 1;
3350 if (sender_cpu >= (u32)NR_CPUS)
3351 skb->sender_cpu = raw_smp_processor_id() + 1;
3354 if (dev->real_num_tx_queues != 1) {
3355 const struct net_device_ops *ops = dev->netdev_ops;
3357 if (ops->ndo_select_queue)
3358 queue_index = ops->ndo_select_queue(dev, skb, accel_priv,
3361 queue_index = __netdev_pick_tx(dev, skb);
3364 queue_index = netdev_cap_txqueue(dev, queue_index);
3367 skb_set_queue_mapping(skb, queue_index);
3368 return netdev_get_tx_queue(dev, queue_index);
3372 * __dev_queue_xmit - transmit a buffer
3373 * @skb: buffer to transmit
3374 * @accel_priv: private data used for L2 forwarding offload
3376 * Queue a buffer for transmission to a network device. The caller must
3377 * have set the device and priority and built the buffer before calling
3378 * this function. The function can be called from an interrupt.
3380 * A negative errno code is returned on a failure. A success does not
3381 * guarantee the frame will be transmitted as it may be dropped due
3382 * to congestion or traffic shaping.
3384 * -----------------------------------------------------------------------------------
3385 * I notice this method can also return errors from the queue disciplines,
3386 * including NET_XMIT_DROP, which is a positive value. So, errors can also
3389 * Regardless of the return value, the skb is consumed, so it is currently
3390 * difficult to retry a send to this method. (You can bump the ref count
3391 * before sending to hold a reference for retry if you are careful.)
3393 * When calling this method, interrupts MUST be enabled. This is because
3394 * the BH enable code must have IRQs enabled so that it will not deadlock.
3397 static int __dev_queue_xmit(struct sk_buff *skb, void *accel_priv)
3399 struct net_device *dev = skb->dev;
3400 struct netdev_queue *txq;
3404 skb_reset_mac_header(skb);
3406 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP))
3407 __skb_tstamp_tx(skb, NULL, skb->sk, SCM_TSTAMP_SCHED);
3409 /* Disable soft irqs for various locks below. Also
3410 * stops preemption for RCU.
3414 skb_update_prio(skb);
3416 qdisc_pkt_len_init(skb);
3417 #ifdef CONFIG_NET_CLS_ACT
3418 skb->tc_at_ingress = 0;
3419 # ifdef CONFIG_NET_EGRESS
3420 if (static_key_false(&egress_needed)) {
3421 skb = sch_handle_egress(skb, &rc, dev);
3427 /* If device/qdisc don't need skb->dst, release it right now while
3428 * its hot in this cpu cache.
3430 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
3435 txq = netdev_pick_tx(dev, skb, accel_priv);
3436 q = rcu_dereference_bh(txq->qdisc);
3438 trace_net_dev_queue(skb);
3440 rc = __dev_xmit_skb(skb, q, dev, txq);
3444 /* The device has no queue. Common case for software devices:
3445 * loopback, all the sorts of tunnels...
3447 * Really, it is unlikely that netif_tx_lock protection is necessary
3448 * here. (f.e. loopback and IP tunnels are clean ignoring statistics
3450 * However, it is possible, that they rely on protection
3453 * Check this and shot the lock. It is not prone from deadlocks.
3454 *Either shot noqueue qdisc, it is even simpler 8)
3456 if (dev->flags & IFF_UP) {
3457 int cpu = smp_processor_id(); /* ok because BHs are off */
3459 if (txq->xmit_lock_owner != cpu) {
3460 if (unlikely(__this_cpu_read(xmit_recursion) >
3461 XMIT_RECURSION_LIMIT))
3462 goto recursion_alert;
3464 skb = validate_xmit_skb(skb, dev);
3468 HARD_TX_LOCK(dev, txq, cpu);
3470 if (!netif_xmit_stopped(txq)) {
3471 __this_cpu_inc(xmit_recursion);
3472 skb = dev_hard_start_xmit(skb, dev, txq, &rc);
3473 __this_cpu_dec(xmit_recursion);
3474 if (dev_xmit_complete(rc)) {
3475 HARD_TX_UNLOCK(dev, txq);
3479 HARD_TX_UNLOCK(dev, txq);
3480 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
3483 /* Recursion is detected! It is possible,
3487 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
3493 rcu_read_unlock_bh();
3495 atomic_long_inc(&dev->tx_dropped);
3496 kfree_skb_list(skb);
3499 rcu_read_unlock_bh();
3503 int dev_queue_xmit(struct sk_buff *skb)
3505 return __dev_queue_xmit(skb, NULL);
3507 EXPORT_SYMBOL(dev_queue_xmit);
3509 int dev_queue_xmit_accel(struct sk_buff *skb, void *accel_priv)
3511 return __dev_queue_xmit(skb, accel_priv);
3513 EXPORT_SYMBOL(dev_queue_xmit_accel);
3516 /*************************************************************************
3518 *************************************************************************/
3520 int netdev_max_backlog __read_mostly = 1000;
3521 EXPORT_SYMBOL(netdev_max_backlog);
3523 int netdev_tstamp_prequeue __read_mostly = 1;
3524 int netdev_budget __read_mostly = 300;
3525 unsigned int __read_mostly netdev_budget_usecs = 2000;
3526 int weight_p __read_mostly = 64; /* old backlog weight */
3527 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */
3528 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */
3529 int dev_rx_weight __read_mostly = 64;
3530 int dev_tx_weight __read_mostly = 64;
3532 /* Called with irq disabled */
3533 static inline void ____napi_schedule(struct softnet_data *sd,
3534 struct napi_struct *napi)
3536 list_add_tail(&napi->poll_list, &sd->poll_list);
3537 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3542 /* One global table that all flow-based protocols share. */
3543 struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
3544 EXPORT_SYMBOL(rps_sock_flow_table);
3545 u32 rps_cpu_mask __read_mostly;
3546 EXPORT_SYMBOL(rps_cpu_mask);
3548 struct static_key rps_needed __read_mostly;
3549 EXPORT_SYMBOL(rps_needed);
3550 struct static_key rfs_needed __read_mostly;
3551 EXPORT_SYMBOL(rfs_needed);
3553 static struct rps_dev_flow *
3554 set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3555 struct rps_dev_flow *rflow, u16 next_cpu)
3557 if (next_cpu < nr_cpu_ids) {
3558 #ifdef CONFIG_RFS_ACCEL
3559 struct netdev_rx_queue *rxqueue;
3560 struct rps_dev_flow_table *flow_table;
3561 struct rps_dev_flow *old_rflow;
3566 /* Should we steer this flow to a different hardware queue? */
3567 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
3568 !(dev->features & NETIF_F_NTUPLE))
3570 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
3571 if (rxq_index == skb_get_rx_queue(skb))
3574 rxqueue = dev->_rx + rxq_index;
3575 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3578 flow_id = skb_get_hash(skb) & flow_table->mask;
3579 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
3580 rxq_index, flow_id);
3584 rflow = &flow_table->flows[flow_id];
3586 if (old_rflow->filter == rflow->filter)
3587 old_rflow->filter = RPS_NO_FILTER;
3591 per_cpu(softnet_data, next_cpu).input_queue_head;
3594 rflow->cpu = next_cpu;
3599 * get_rps_cpu is called from netif_receive_skb and returns the target
3600 * CPU from the RPS map of the receiving queue for a given skb.
3601 * rcu_read_lock must be held on entry.
3603 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3604 struct rps_dev_flow **rflowp)
3606 const struct rps_sock_flow_table *sock_flow_table;
3607 struct netdev_rx_queue *rxqueue = dev->_rx;
3608 struct rps_dev_flow_table *flow_table;
3609 struct rps_map *map;
3614 if (skb_rx_queue_recorded(skb)) {
3615 u16 index = skb_get_rx_queue(skb);
3617 if (unlikely(index >= dev->real_num_rx_queues)) {
3618 WARN_ONCE(dev->real_num_rx_queues > 1,
3619 "%s received packet on queue %u, but number "
3620 "of RX queues is %u\n",
3621 dev->name, index, dev->real_num_rx_queues);
3627 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */
3629 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3630 map = rcu_dereference(rxqueue->rps_map);
3631 if (!flow_table && !map)
3634 skb_reset_network_header(skb);
3635 hash = skb_get_hash(skb);
3639 sock_flow_table = rcu_dereference(rps_sock_flow_table);
3640 if (flow_table && sock_flow_table) {
3641 struct rps_dev_flow *rflow;
3645 /* First check into global flow table if there is a match */
3646 ident = sock_flow_table->ents[hash & sock_flow_table->mask];
3647 if ((ident ^ hash) & ~rps_cpu_mask)
3650 next_cpu = ident & rps_cpu_mask;
3652 /* OK, now we know there is a match,
3653 * we can look at the local (per receive queue) flow table
3655 rflow = &flow_table->flows[hash & flow_table->mask];
3659 * If the desired CPU (where last recvmsg was done) is
3660 * different from current CPU (one in the rx-queue flow
3661 * table entry), switch if one of the following holds:
3662 * - Current CPU is unset (>= nr_cpu_ids).
3663 * - Current CPU is offline.
3664 * - The current CPU's queue tail has advanced beyond the
3665 * last packet that was enqueued using this table entry.
3666 * This guarantees that all previous packets for the flow
3667 * have been dequeued, thus preserving in order delivery.
3669 if (unlikely(tcpu != next_cpu) &&
3670 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) ||
3671 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
3672 rflow->last_qtail)) >= 0)) {
3674 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
3677 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) {
3687 tcpu = map->cpus[reciprocal_scale(hash, map->len)];
3688 if (cpu_online(tcpu)) {
3698 #ifdef CONFIG_RFS_ACCEL
3701 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
3702 * @dev: Device on which the filter was set
3703 * @rxq_index: RX queue index
3704 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
3705 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
3707 * Drivers that implement ndo_rx_flow_steer() should periodically call
3708 * this function for each installed filter and remove the filters for
3709 * which it returns %true.
3711 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
3712 u32 flow_id, u16 filter_id)
3714 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
3715 struct rps_dev_flow_table *flow_table;
3716 struct rps_dev_flow *rflow;
3721 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3722 if (flow_table && flow_id <= flow_table->mask) {
3723 rflow = &flow_table->flows[flow_id];
3724 cpu = ACCESS_ONCE(rflow->cpu);
3725 if (rflow->filter == filter_id && cpu < nr_cpu_ids &&
3726 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
3727 rflow->last_qtail) <
3728 (int)(10 * flow_table->mask)))
3734 EXPORT_SYMBOL(rps_may_expire_flow);
3736 #endif /* CONFIG_RFS_ACCEL */
3738 /* Called from hardirq (IPI) context */
3739 static void rps_trigger_softirq(void *data)
3741 struct softnet_data *sd = data;
3743 ____napi_schedule(sd, &sd->backlog);
3747 #endif /* CONFIG_RPS */
3750 * Check if this softnet_data structure is another cpu one
3751 * If yes, queue it to our IPI list and return 1
3754 static int rps_ipi_queued(struct softnet_data *sd)
3757 struct softnet_data *mysd = this_cpu_ptr(&softnet_data);
3760 sd->rps_ipi_next = mysd->rps_ipi_list;
3761 mysd->rps_ipi_list = sd;
3763 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3766 #endif /* CONFIG_RPS */
3770 #ifdef CONFIG_NET_FLOW_LIMIT
3771 int netdev_flow_limit_table_len __read_mostly = (1 << 12);
3774 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
3776 #ifdef CONFIG_NET_FLOW_LIMIT
3777 struct sd_flow_limit *fl;
3778 struct softnet_data *sd;
3779 unsigned int old_flow, new_flow;
3781 if (qlen < (netdev_max_backlog >> 1))
3784 sd = this_cpu_ptr(&softnet_data);
3787 fl = rcu_dereference(sd->flow_limit);
3789 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1);
3790 old_flow = fl->history[fl->history_head];
3791 fl->history[fl->history_head] = new_flow;
3794 fl->history_head &= FLOW_LIMIT_HISTORY - 1;
3796 if (likely(fl->buckets[old_flow]))
3797 fl->buckets[old_flow]--;
3799 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) {
3811 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
3812 * queue (may be a remote CPU queue).
3814 static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
3815 unsigned int *qtail)
3817 struct softnet_data *sd;
3818 unsigned long flags;
3821 sd = &per_cpu(softnet_data, cpu);
3823 local_irq_save(flags);
3826 if (!netif_running(skb->dev))
3828 qlen = skb_queue_len(&sd->input_pkt_queue);
3829 if (qlen <= netdev_max_backlog && !skb_flow_limit(skb, qlen)) {
3832 __skb_queue_tail(&sd->input_pkt_queue, skb);
3833 input_queue_tail_incr_save(sd, qtail);
3835 local_irq_restore(flags);
3836 return NET_RX_SUCCESS;
3839 /* Schedule NAPI for backlog device
3840 * We can use non atomic operation since we own the queue lock
3842 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
3843 if (!rps_ipi_queued(sd))
3844 ____napi_schedule(sd, &sd->backlog);
3853 local_irq_restore(flags);
3855 atomic_long_inc(&skb->dev->rx_dropped);
3860 static u32 netif_receive_generic_xdp(struct sk_buff *skb,
3861 struct bpf_prog *xdp_prog)
3863 struct xdp_buff xdp;
3869 /* Reinjected packets coming from act_mirred or similar should
3870 * not get XDP generic processing.
3872 if (skb_cloned(skb))
3875 if (skb_linearize(skb))
3878 /* The XDP program wants to see the packet starting at the MAC
3881 mac_len = skb->data - skb_mac_header(skb);
3882 hlen = skb_headlen(skb) + mac_len;
3883 xdp.data = skb->data - mac_len;
3884 xdp.data_end = xdp.data + hlen;
3885 xdp.data_hard_start = skb->data - skb_headroom(skb);
3886 orig_data = xdp.data;
3888 act = bpf_prog_run_xdp(xdp_prog, &xdp);
3890 off = xdp.data - orig_data;
3892 __skb_pull(skb, off);
3894 __skb_push(skb, -off);
3899 __skb_push(skb, mac_len);
3905 bpf_warn_invalid_xdp_action(act);
3908 trace_xdp_exception(skb->dev, xdp_prog, act);
3919 /* When doing generic XDP we have to bypass the qdisc layer and the
3920 * network taps in order to match in-driver-XDP behavior.
3922 static void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog)
3924 struct net_device *dev = skb->dev;
3925 struct netdev_queue *txq;
3926 bool free_skb = true;
3929 txq = netdev_pick_tx(dev, skb, NULL);
3930 cpu = smp_processor_id();
3931 HARD_TX_LOCK(dev, txq, cpu);
3932 if (!netif_xmit_stopped(txq)) {
3933 rc = netdev_start_xmit(skb, dev, txq, 0);
3934 if (dev_xmit_complete(rc))
3937 HARD_TX_UNLOCK(dev, txq);
3939 trace_xdp_exception(dev, xdp_prog, XDP_TX);
3944 static struct static_key generic_xdp_needed __read_mostly;
3946 static int do_xdp_generic(struct sk_buff *skb)
3948 struct bpf_prog *xdp_prog = rcu_dereference(skb->dev->xdp_prog);
3951 u32 act = netif_receive_generic_xdp(skb, xdp_prog);
3954 if (act != XDP_PASS) {
3957 err = xdp_do_generic_redirect(skb->dev, skb);
3960 /* fallthru to submit skb */
3962 generic_xdp_tx(skb, xdp_prog);
3970 trace_xdp_exception(skb->dev, xdp_prog, XDP_REDIRECT);
3975 static int netif_rx_internal(struct sk_buff *skb)
3979 net_timestamp_check(netdev_tstamp_prequeue, skb);
3981 trace_netif_rx(skb);
3983 if (static_key_false(&generic_xdp_needed)) {
3984 int ret = do_xdp_generic(skb);
3986 /* Consider XDP consuming the packet a success from
3987 * the netdev point of view we do not want to count
3990 if (ret != XDP_PASS)
3991 return NET_RX_SUCCESS;
3995 if (static_key_false(&rps_needed)) {
3996 struct rps_dev_flow voidflow, *rflow = &voidflow;
4002 cpu = get_rps_cpu(skb->dev, skb, &rflow);
4004 cpu = smp_processor_id();
4006 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4015 ret = enqueue_to_backlog(skb, get_cpu(), &qtail);
4022 * netif_rx - post buffer to the network code
4023 * @skb: buffer to post
4025 * This function receives a packet from a device driver and queues it for
4026 * the upper (protocol) levels to process. It always succeeds. The buffer
4027 * may be dropped during processing for congestion control or by the
4031 * NET_RX_SUCCESS (no congestion)
4032 * NET_RX_DROP (packet was dropped)
4036 int netif_rx(struct sk_buff *skb)
4038 trace_netif_rx_entry(skb);
4040 return netif_rx_internal(skb);
4042 EXPORT_SYMBOL(netif_rx);
4044 int netif_rx_ni(struct sk_buff *skb)
4048 trace_netif_rx_ni_entry(skb);
4051 err = netif_rx_internal(skb);
4052 if (local_softirq_pending())
4058 EXPORT_SYMBOL(netif_rx_ni);
4060 static __latent_entropy void net_tx_action(struct softirq_action *h)
4062 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
4064 if (sd->completion_queue) {
4065 struct sk_buff *clist;
4067 local_irq_disable();
4068 clist = sd->completion_queue;
4069 sd->completion_queue = NULL;
4073 struct sk_buff *skb = clist;
4075 clist = clist->next;
4077 WARN_ON(refcount_read(&skb->users));
4078 if (likely(get_kfree_skb_cb(skb)->reason == SKB_REASON_CONSUMED))
4079 trace_consume_skb(skb);
4081 trace_kfree_skb(skb, net_tx_action);
4083 if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
4086 __kfree_skb_defer(skb);
4089 __kfree_skb_flush();
4092 if (sd->output_queue) {
4095 local_irq_disable();
4096 head = sd->output_queue;
4097 sd->output_queue = NULL;
4098 sd->output_queue_tailp = &sd->output_queue;
4102 struct Qdisc *q = head;
4103 spinlock_t *root_lock;
4105 head = head->next_sched;
4107 root_lock = qdisc_lock(q);
4108 spin_lock(root_lock);
4109 /* We need to make sure head->next_sched is read
4110 * before clearing __QDISC_STATE_SCHED
4112 smp_mb__before_atomic();
4113 clear_bit(__QDISC_STATE_SCHED, &q->state);
4115 spin_unlock(root_lock);
4120 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE)
4121 /* This hook is defined here for ATM LANE */
4122 int (*br_fdb_test_addr_hook)(struct net_device *dev,
4123 unsigned char *addr) __read_mostly;
4124 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook);
4127 static inline struct sk_buff *
4128 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
4129 struct net_device *orig_dev)
4131 #ifdef CONFIG_NET_CLS_ACT
4132 struct tcf_proto *cl = rcu_dereference_bh(skb->dev->ingress_cl_list);
4133 struct tcf_result cl_res;
4135 /* If there's at least one ingress present somewhere (so
4136 * we get here via enabled static key), remaining devices
4137 * that are not configured with an ingress qdisc will bail
4143 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4147 qdisc_skb_cb(skb)->pkt_len = skb->len;
4148 skb->tc_at_ingress = 1;
4149 qdisc_bstats_cpu_update(cl->q, skb);
4151 switch (tcf_classify(skb, cl, &cl_res, false)) {
4153 case TC_ACT_RECLASSIFY:
4154 skb->tc_index = TC_H_MIN(cl_res.classid);
4157 qdisc_qstats_cpu_drop(cl->q);
4165 case TC_ACT_REDIRECT:
4166 /* skb_mac_header check was done by cls/act_bpf, so
4167 * we can safely push the L2 header back before
4168 * redirecting to another netdev
4170 __skb_push(skb, skb->mac_len);
4171 skb_do_redirect(skb);
4176 #endif /* CONFIG_NET_CLS_ACT */
4181 * netdev_is_rx_handler_busy - check if receive handler is registered
4182 * @dev: device to check
4184 * Check if a receive handler is already registered for a given device.
4185 * Return true if there one.
4187 * The caller must hold the rtnl_mutex.
4189 bool netdev_is_rx_handler_busy(struct net_device *dev)
4192 return dev && rtnl_dereference(dev->rx_handler);
4194 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy);
4197 * netdev_rx_handler_register - register receive handler
4198 * @dev: device to register a handler for
4199 * @rx_handler: receive handler to register
4200 * @rx_handler_data: data pointer that is used by rx handler
4202 * Register a receive handler for a device. This handler will then be
4203 * called from __netif_receive_skb. A negative errno code is returned
4206 * The caller must hold the rtnl_mutex.
4208 * For a general description of rx_handler, see enum rx_handler_result.
4210 int netdev_rx_handler_register(struct net_device *dev,
4211 rx_handler_func_t *rx_handler,
4212 void *rx_handler_data)
4214 if (netdev_is_rx_handler_busy(dev))
4217 /* Note: rx_handler_data must be set before rx_handler */
4218 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data);
4219 rcu_assign_pointer(dev->rx_handler, rx_handler);
4223 EXPORT_SYMBOL_GPL(netdev_rx_handler_register);
4226 * netdev_rx_handler_unregister - unregister receive handler
4227 * @dev: device to unregister a handler from
4229 * Unregister a receive handler from a device.
4231 * The caller must hold the rtnl_mutex.
4233 void netdev_rx_handler_unregister(struct net_device *dev)
4237 RCU_INIT_POINTER(dev->rx_handler, NULL);
4238 /* a reader seeing a non NULL rx_handler in a rcu_read_lock()
4239 * section has a guarantee to see a non NULL rx_handler_data
4243 RCU_INIT_POINTER(dev->rx_handler_data, NULL);
4245 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister);
4248 * Limit the use of PFMEMALLOC reserves to those protocols that implement
4249 * the special handling of PFMEMALLOC skbs.
4251 static bool skb_pfmemalloc_protocol(struct sk_buff *skb)
4253 switch (skb->protocol) {
4254 case htons(ETH_P_ARP):
4255 case htons(ETH_P_IP):
4256 case htons(ETH_P_IPV6):
4257 case htons(ETH_P_8021Q):
4258 case htons(ETH_P_8021AD):
4265 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev,
4266 int *ret, struct net_device *orig_dev)
4268 #ifdef CONFIG_NETFILTER_INGRESS
4269 if (nf_hook_ingress_active(skb)) {
4273 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4278 ingress_retval = nf_hook_ingress(skb);
4280 return ingress_retval;
4282 #endif /* CONFIG_NETFILTER_INGRESS */
4286 static int __netif_receive_skb_core(struct sk_buff *skb, bool pfmemalloc)
4288 struct packet_type *ptype, *pt_prev;
4289 rx_handler_func_t *rx_handler;
4290 struct net_device *orig_dev;
4291 bool deliver_exact = false;
4292 int ret = NET_RX_DROP;
4295 net_timestamp_check(!netdev_tstamp_prequeue, skb);
4297 trace_netif_receive_skb(skb);
4299 orig_dev = skb->dev;
4301 skb_reset_network_header(skb);
4302 if (!skb_transport_header_was_set(skb))
4303 skb_reset_transport_header(skb);
4304 skb_reset_mac_len(skb);
4309 skb->skb_iif = skb->dev->ifindex;
4311 __this_cpu_inc(softnet_data.processed);
4313 if (skb->protocol == cpu_to_be16(ETH_P_8021Q) ||
4314 skb->protocol == cpu_to_be16(ETH_P_8021AD)) {
4315 skb = skb_vlan_untag(skb);
4320 if (skb_skip_tc_classify(skb))
4326 list_for_each_entry_rcu(ptype, &ptype_all, list) {
4328 ret = deliver_skb(skb, pt_prev, orig_dev);
4332 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) {
4334 ret = deliver_skb(skb, pt_prev, orig_dev);
4339 #ifdef CONFIG_NET_INGRESS
4340 if (static_key_false(&ingress_needed)) {
4341 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev);
4345 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0)
4351 if (pfmemalloc && !skb_pfmemalloc_protocol(skb))
4354 if (skb_vlan_tag_present(skb)) {
4356 ret = deliver_skb(skb, pt_prev, orig_dev);
4359 if (vlan_do_receive(&skb))
4361 else if (unlikely(!skb))
4365 rx_handler = rcu_dereference(skb->dev->rx_handler);
4368 ret = deliver_skb(skb, pt_prev, orig_dev);
4371 switch (rx_handler(&skb)) {
4372 case RX_HANDLER_CONSUMED:
4373 ret = NET_RX_SUCCESS;
4375 case RX_HANDLER_ANOTHER:
4377 case RX_HANDLER_EXACT:
4378 deliver_exact = true;
4379 case RX_HANDLER_PASS:
4386 if (unlikely(skb_vlan_tag_present(skb))) {
4387 if (skb_vlan_tag_get_id(skb))
4388 skb->pkt_type = PACKET_OTHERHOST;
4389 /* Note: we might in the future use prio bits
4390 * and set skb->priority like in vlan_do_receive()
4391 * For the time being, just ignore Priority Code Point
4396 type = skb->protocol;
4398 /* deliver only exact match when indicated */
4399 if (likely(!deliver_exact)) {
4400 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4401 &ptype_base[ntohs(type) &
4405 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4406 &orig_dev->ptype_specific);
4408 if (unlikely(skb->dev != orig_dev)) {
4409 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4410 &skb->dev->ptype_specific);
4414 if (unlikely(skb_orphan_frags(skb, GFP_ATOMIC)))
4417 ret = pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
4421 atomic_long_inc(&skb->dev->rx_dropped);
4423 atomic_long_inc(&skb->dev->rx_nohandler);
4425 /* Jamal, now you will not able to escape explaining
4426 * me how you were going to use this. :-)
4435 static int __netif_receive_skb(struct sk_buff *skb)
4439 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) {
4440 unsigned int noreclaim_flag;
4443 * PFMEMALLOC skbs are special, they should
4444 * - be delivered to SOCK_MEMALLOC sockets only
4445 * - stay away from userspace
4446 * - have bounded memory usage
4448 * Use PF_MEMALLOC as this saves us from propagating the allocation
4449 * context down to all allocation sites.
4451 noreclaim_flag = memalloc_noreclaim_save();
4452 ret = __netif_receive_skb_core(skb, true);
4453 memalloc_noreclaim_restore(noreclaim_flag);
4455 ret = __netif_receive_skb_core(skb, false);
4460 static int generic_xdp_install(struct net_device *dev, struct netdev_xdp *xdp)
4462 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog);
4463 struct bpf_prog *new = xdp->prog;
4466 switch (xdp->command) {
4467 case XDP_SETUP_PROG:
4468 rcu_assign_pointer(dev->xdp_prog, new);
4473 static_key_slow_dec(&generic_xdp_needed);
4474 } else if (new && !old) {
4475 static_key_slow_inc(&generic_xdp_needed);
4476 dev_disable_lro(dev);
4480 case XDP_QUERY_PROG:
4481 xdp->prog_attached = !!old;
4482 xdp->prog_id = old ? old->aux->id : 0;
4493 static int netif_receive_skb_internal(struct sk_buff *skb)
4497 net_timestamp_check(netdev_tstamp_prequeue, skb);
4499 if (skb_defer_rx_timestamp(skb))
4500 return NET_RX_SUCCESS;
4504 if (static_key_false(&generic_xdp_needed)) {
4505 int ret = do_xdp_generic(skb);
4507 if (ret != XDP_PASS) {
4514 if (static_key_false(&rps_needed)) {
4515 struct rps_dev_flow voidflow, *rflow = &voidflow;
4516 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
4519 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4525 ret = __netif_receive_skb(skb);
4531 * netif_receive_skb - process receive buffer from network
4532 * @skb: buffer to process
4534 * netif_receive_skb() is the main receive data processing function.
4535 * It always succeeds. The buffer may be dropped during processing
4536 * for congestion control or by the protocol layers.
4538 * This function may only be called from softirq context and interrupts
4539 * should be enabled.
4541 * Return values (usually ignored):
4542 * NET_RX_SUCCESS: no congestion
4543 * NET_RX_DROP: packet was dropped
4545 int netif_receive_skb(struct sk_buff *skb)
4547 trace_netif_receive_skb_entry(skb);
4549 return netif_receive_skb_internal(skb);
4551 EXPORT_SYMBOL(netif_receive_skb);
4553 DEFINE_PER_CPU(struct work_struct, flush_works);
4555 /* Network device is going away, flush any packets still pending */
4556 static void flush_backlog(struct work_struct *work)
4558 struct sk_buff *skb, *tmp;
4559 struct softnet_data *sd;
4562 sd = this_cpu_ptr(&softnet_data);
4564 local_irq_disable();
4566 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) {
4567 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
4568 __skb_unlink(skb, &sd->input_pkt_queue);
4570 input_queue_head_incr(sd);
4576 skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
4577 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
4578 __skb_unlink(skb, &sd->process_queue);
4580 input_queue_head_incr(sd);
4586 static void flush_all_backlogs(void)
4592 for_each_online_cpu(cpu)
4593 queue_work_on(cpu, system_highpri_wq,
4594 per_cpu_ptr(&flush_works, cpu));
4596 for_each_online_cpu(cpu)
4597 flush_work(per_cpu_ptr(&flush_works, cpu));
4602 static int napi_gro_complete(struct sk_buff *skb)
4604 struct packet_offload *ptype;
4605 __be16 type = skb->protocol;
4606 struct list_head *head = &offload_base;
4609 BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
4611 if (NAPI_GRO_CB(skb)->count == 1) {
4612 skb_shinfo(skb)->gso_size = 0;
4617 list_for_each_entry_rcu(ptype, head, list) {
4618 if (ptype->type != type || !ptype->callbacks.gro_complete)
4621 err = ptype->callbacks.gro_complete(skb, 0);
4627 WARN_ON(&ptype->list == head);
4629 return NET_RX_SUCCESS;
4633 return netif_receive_skb_internal(skb);
4636 /* napi->gro_list contains packets ordered by age.
4637 * youngest packets at the head of it.
4638 * Complete skbs in reverse order to reduce latencies.
4640 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
4642 struct sk_buff *skb, *prev = NULL;
4644 /* scan list and build reverse chain */
4645 for (skb = napi->gro_list; skb != NULL; skb = skb->next) {
4650 for (skb = prev; skb; skb = prev) {
4653 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
4657 napi_gro_complete(skb);
4661 napi->gro_list = NULL;
4663 EXPORT_SYMBOL(napi_gro_flush);
4665 static void gro_list_prepare(struct napi_struct *napi, struct sk_buff *skb)
4668 unsigned int maclen = skb->dev->hard_header_len;
4669 u32 hash = skb_get_hash_raw(skb);
4671 for (p = napi->gro_list; p; p = p->next) {
4672 unsigned long diffs;
4674 NAPI_GRO_CB(p)->flush = 0;
4676 if (hash != skb_get_hash_raw(p)) {
4677 NAPI_GRO_CB(p)->same_flow = 0;
4681 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
4682 diffs |= p->vlan_tci ^ skb->vlan_tci;
4683 diffs |= skb_metadata_dst_cmp(p, skb);
4684 if (maclen == ETH_HLEN)
4685 diffs |= compare_ether_header(skb_mac_header(p),
4686 skb_mac_header(skb));
4688 diffs = memcmp(skb_mac_header(p),
4689 skb_mac_header(skb),
4691 NAPI_GRO_CB(p)->same_flow = !diffs;
4695 static void skb_gro_reset_offset(struct sk_buff *skb)
4697 const struct skb_shared_info *pinfo = skb_shinfo(skb);
4698 const skb_frag_t *frag0 = &pinfo->frags[0];
4700 NAPI_GRO_CB(skb)->data_offset = 0;
4701 NAPI_GRO_CB(skb)->frag0 = NULL;
4702 NAPI_GRO_CB(skb)->frag0_len = 0;
4704 if (skb_mac_header(skb) == skb_tail_pointer(skb) &&
4706 !PageHighMem(skb_frag_page(frag0))) {
4707 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
4708 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
4709 skb_frag_size(frag0),
4710 skb->end - skb->tail);
4714 static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
4716 struct skb_shared_info *pinfo = skb_shinfo(skb);
4718 BUG_ON(skb->end - skb->tail < grow);
4720 memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
4722 skb->data_len -= grow;
4725 pinfo->frags[0].page_offset += grow;
4726 skb_frag_size_sub(&pinfo->frags[0], grow);
4728 if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
4729 skb_frag_unref(skb, 0);
4730 memmove(pinfo->frags, pinfo->frags + 1,
4731 --pinfo->nr_frags * sizeof(pinfo->frags[0]));
4735 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
4737 struct sk_buff **pp = NULL;
4738 struct packet_offload *ptype;
4739 __be16 type = skb->protocol;
4740 struct list_head *head = &offload_base;
4742 enum gro_result ret;
4745 if (netif_elide_gro(skb->dev))
4748 gro_list_prepare(napi, skb);
4751 list_for_each_entry_rcu(ptype, head, list) {
4752 if (ptype->type != type || !ptype->callbacks.gro_receive)
4755 skb_set_network_header(skb, skb_gro_offset(skb));
4756 skb_reset_mac_len(skb);
4757 NAPI_GRO_CB(skb)->same_flow = 0;
4758 NAPI_GRO_CB(skb)->flush = skb_is_gso(skb) || skb_has_frag_list(skb);
4759 NAPI_GRO_CB(skb)->free = 0;
4760 NAPI_GRO_CB(skb)->encap_mark = 0;
4761 NAPI_GRO_CB(skb)->recursion_counter = 0;
4762 NAPI_GRO_CB(skb)->is_fou = 0;
4763 NAPI_GRO_CB(skb)->is_atomic = 1;
4764 NAPI_GRO_CB(skb)->gro_remcsum_start = 0;
4766 /* Setup for GRO checksum validation */
4767 switch (skb->ip_summed) {
4768 case CHECKSUM_COMPLETE:
4769 NAPI_GRO_CB(skb)->csum = skb->csum;
4770 NAPI_GRO_CB(skb)->csum_valid = 1;
4771 NAPI_GRO_CB(skb)->csum_cnt = 0;
4773 case CHECKSUM_UNNECESSARY:
4774 NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
4775 NAPI_GRO_CB(skb)->csum_valid = 0;
4778 NAPI_GRO_CB(skb)->csum_cnt = 0;
4779 NAPI_GRO_CB(skb)->csum_valid = 0;
4782 pp = ptype->callbacks.gro_receive(&napi->gro_list, skb);
4787 if (&ptype->list == head)
4790 if (IS_ERR(pp) && PTR_ERR(pp) == -EINPROGRESS) {
4795 same_flow = NAPI_GRO_CB(skb)->same_flow;
4796 ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
4799 struct sk_buff *nskb = *pp;
4803 napi_gro_complete(nskb);
4810 if (NAPI_GRO_CB(skb)->flush)
4813 if (unlikely(napi->gro_count >= MAX_GRO_SKBS)) {
4814 struct sk_buff *nskb = napi->gro_list;
4816 /* locate the end of the list to select the 'oldest' flow */
4817 while (nskb->next) {
4823 napi_gro_complete(nskb);
4827 NAPI_GRO_CB(skb)->count = 1;
4828 NAPI_GRO_CB(skb)->age = jiffies;
4829 NAPI_GRO_CB(skb)->last = skb;
4830 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
4831 skb->next = napi->gro_list;
4832 napi->gro_list = skb;
4836 grow = skb_gro_offset(skb) - skb_headlen(skb);
4838 gro_pull_from_frag0(skb, grow);
4847 struct packet_offload *gro_find_receive_by_type(__be16 type)
4849 struct list_head *offload_head = &offload_base;
4850 struct packet_offload *ptype;
4852 list_for_each_entry_rcu(ptype, offload_head, list) {
4853 if (ptype->type != type || !ptype->callbacks.gro_receive)
4859 EXPORT_SYMBOL(gro_find_receive_by_type);
4861 struct packet_offload *gro_find_complete_by_type(__be16 type)
4863 struct list_head *offload_head = &offload_base;
4864 struct packet_offload *ptype;
4866 list_for_each_entry_rcu(ptype, offload_head, list) {
4867 if (ptype->type != type || !ptype->callbacks.gro_complete)
4873 EXPORT_SYMBOL(gro_find_complete_by_type);
4875 static void napi_skb_free_stolen_head(struct sk_buff *skb)
4879 kmem_cache_free(skbuff_head_cache, skb);
4882 static gro_result_t napi_skb_finish(gro_result_t ret, struct sk_buff *skb)
4886 if (netif_receive_skb_internal(skb))
4894 case GRO_MERGED_FREE:
4895 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
4896 napi_skb_free_stolen_head(skb);
4910 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
4912 skb_mark_napi_id(skb, napi);
4913 trace_napi_gro_receive_entry(skb);
4915 skb_gro_reset_offset(skb);
4917 return napi_skb_finish(dev_gro_receive(napi, skb), skb);
4919 EXPORT_SYMBOL(napi_gro_receive);
4921 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
4923 if (unlikely(skb->pfmemalloc)) {
4927 __skb_pull(skb, skb_headlen(skb));
4928 /* restore the reserve we had after netdev_alloc_skb_ip_align() */
4929 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
4931 skb->dev = napi->dev;
4933 skb->encapsulation = 0;
4934 skb_shinfo(skb)->gso_type = 0;
4935 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
4941 struct sk_buff *napi_get_frags(struct napi_struct *napi)
4943 struct sk_buff *skb = napi->skb;
4946 skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
4949 skb_mark_napi_id(skb, napi);
4954 EXPORT_SYMBOL(napi_get_frags);
4956 static gro_result_t napi_frags_finish(struct napi_struct *napi,
4957 struct sk_buff *skb,
4963 __skb_push(skb, ETH_HLEN);
4964 skb->protocol = eth_type_trans(skb, skb->dev);
4965 if (ret == GRO_NORMAL && netif_receive_skb_internal(skb))
4970 napi_reuse_skb(napi, skb);
4973 case GRO_MERGED_FREE:
4974 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
4975 napi_skb_free_stolen_head(skb);
4977 napi_reuse_skb(napi, skb);
4988 /* Upper GRO stack assumes network header starts at gro_offset=0
4989 * Drivers could call both napi_gro_frags() and napi_gro_receive()
4990 * We copy ethernet header into skb->data to have a common layout.
4992 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
4994 struct sk_buff *skb = napi->skb;
4995 const struct ethhdr *eth;
4996 unsigned int hlen = sizeof(*eth);
5000 skb_reset_mac_header(skb);
5001 skb_gro_reset_offset(skb);
5003 eth = skb_gro_header_fast(skb, 0);
5004 if (unlikely(skb_gro_header_hard(skb, hlen))) {
5005 eth = skb_gro_header_slow(skb, hlen, 0);
5006 if (unlikely(!eth)) {
5007 net_warn_ratelimited("%s: dropping impossible skb from %s\n",
5008 __func__, napi->dev->name);
5009 napi_reuse_skb(napi, skb);
5013 gro_pull_from_frag0(skb, hlen);
5014 NAPI_GRO_CB(skb)->frag0 += hlen;
5015 NAPI_GRO_CB(skb)->frag0_len -= hlen;
5017 __skb_pull(skb, hlen);
5020 * This works because the only protocols we care about don't require
5022 * We'll fix it up properly in napi_frags_finish()
5024 skb->protocol = eth->h_proto;
5029 gro_result_t napi_gro_frags(struct napi_struct *napi)
5031 struct sk_buff *skb = napi_frags_skb(napi);
5036 trace_napi_gro_frags_entry(skb);
5038 return napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
5040 EXPORT_SYMBOL(napi_gro_frags);
5042 /* Compute the checksum from gro_offset and return the folded value
5043 * after adding in any pseudo checksum.
5045 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
5050 wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
5052 /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
5053 sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
5055 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
5056 !skb->csum_complete_sw)
5057 netdev_rx_csum_fault(skb->dev);
5060 NAPI_GRO_CB(skb)->csum = wsum;
5061 NAPI_GRO_CB(skb)->csum_valid = 1;
5065 EXPORT_SYMBOL(__skb_gro_checksum_complete);
5067 static void net_rps_send_ipi(struct softnet_data *remsd)
5071 struct softnet_data *next = remsd->rps_ipi_next;
5073 if (cpu_online(remsd->cpu))
5074 smp_call_function_single_async(remsd->cpu, &remsd->csd);
5081 * net_rps_action_and_irq_enable sends any pending IPI's for rps.
5082 * Note: called with local irq disabled, but exits with local irq enabled.
5084 static void net_rps_action_and_irq_enable(struct softnet_data *sd)
5087 struct softnet_data *remsd = sd->rps_ipi_list;
5090 sd->rps_ipi_list = NULL;
5094 /* Send pending IPI's to kick RPS processing on remote cpus. */
5095 net_rps_send_ipi(remsd);
5101 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd)
5104 return sd->rps_ipi_list != NULL;
5110 static int process_backlog(struct napi_struct *napi, int quota)
5112 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
5116 /* Check if we have pending ipi, its better to send them now,
5117 * not waiting net_rx_action() end.
5119 if (sd_has_rps_ipi_waiting(sd)) {
5120 local_irq_disable();
5121 net_rps_action_and_irq_enable(sd);
5124 napi->weight = dev_rx_weight;
5126 struct sk_buff *skb;
5128 while ((skb = __skb_dequeue(&sd->process_queue))) {
5130 __netif_receive_skb(skb);
5132 input_queue_head_incr(sd);
5133 if (++work >= quota)
5138 local_irq_disable();
5140 if (skb_queue_empty(&sd->input_pkt_queue)) {
5142 * Inline a custom version of __napi_complete().
5143 * only current cpu owns and manipulates this napi,
5144 * and NAPI_STATE_SCHED is the only possible flag set
5146 * We can use a plain write instead of clear_bit(),
5147 * and we dont need an smp_mb() memory barrier.
5152 skb_queue_splice_tail_init(&sd->input_pkt_queue,
5153 &sd->process_queue);
5163 * __napi_schedule - schedule for receive
5164 * @n: entry to schedule
5166 * The entry's receive function will be scheduled to run.
5167 * Consider using __napi_schedule_irqoff() if hard irqs are masked.
5169 void __napi_schedule(struct napi_struct *n)
5171 unsigned long flags;
5173 local_irq_save(flags);
5174 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
5175 local_irq_restore(flags);
5177 EXPORT_SYMBOL(__napi_schedule);
5180 * napi_schedule_prep - check if napi can be scheduled
5183 * Test if NAPI routine is already running, and if not mark
5184 * it as running. This is used as a condition variable
5185 * insure only one NAPI poll instance runs. We also make
5186 * sure there is no pending NAPI disable.
5188 bool napi_schedule_prep(struct napi_struct *n)
5190 unsigned long val, new;
5193 val = READ_ONCE(n->state);
5194 if (unlikely(val & NAPIF_STATE_DISABLE))
5196 new = val | NAPIF_STATE_SCHED;
5198 /* Sets STATE_MISSED bit if STATE_SCHED was already set
5199 * This was suggested by Alexander Duyck, as compiler
5200 * emits better code than :
5201 * if (val & NAPIF_STATE_SCHED)
5202 * new |= NAPIF_STATE_MISSED;
5204 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED *
5206 } while (cmpxchg(&n->state, val, new) != val);
5208 return !(val & NAPIF_STATE_SCHED);
5210 EXPORT_SYMBOL(napi_schedule_prep);
5213 * __napi_schedule_irqoff - schedule for receive
5214 * @n: entry to schedule
5216 * Variant of __napi_schedule() assuming hard irqs are masked
5218 void __napi_schedule_irqoff(struct napi_struct *n)
5220 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
5222 EXPORT_SYMBOL(__napi_schedule_irqoff);
5224 bool napi_complete_done(struct napi_struct *n, int work_done)
5226 unsigned long flags, val, new;
5229 * 1) Don't let napi dequeue from the cpu poll list
5230 * just in case its running on a different cpu.
5231 * 2) If we are busy polling, do nothing here, we have
5232 * the guarantee we will be called later.
5234 if (unlikely(n->state & (NAPIF_STATE_NPSVC |
5235 NAPIF_STATE_IN_BUSY_POLL)))
5239 unsigned long timeout = 0;
5242 timeout = n->dev->gro_flush_timeout;
5245 hrtimer_start(&n->timer, ns_to_ktime(timeout),
5246 HRTIMER_MODE_REL_PINNED);
5248 napi_gro_flush(n, false);
5250 if (unlikely(!list_empty(&n->poll_list))) {
5251 /* If n->poll_list is not empty, we need to mask irqs */
5252 local_irq_save(flags);
5253 list_del_init(&n->poll_list);
5254 local_irq_restore(flags);
5258 val = READ_ONCE(n->state);
5260 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED));
5262 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED);
5264 /* If STATE_MISSED was set, leave STATE_SCHED set,
5265 * because we will call napi->poll() one more time.
5266 * This C code was suggested by Alexander Duyck to help gcc.
5268 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED *
5270 } while (cmpxchg(&n->state, val, new) != val);
5272 if (unlikely(val & NAPIF_STATE_MISSED)) {
5279 EXPORT_SYMBOL(napi_complete_done);
5281 /* must be called under rcu_read_lock(), as we dont take a reference */
5282 static struct napi_struct *napi_by_id(unsigned int napi_id)
5284 unsigned int hash = napi_id % HASH_SIZE(napi_hash);
5285 struct napi_struct *napi;
5287 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node)
5288 if (napi->napi_id == napi_id)
5294 #if defined(CONFIG_NET_RX_BUSY_POLL)
5296 #define BUSY_POLL_BUDGET 8
5298 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock)
5302 /* Busy polling means there is a high chance device driver hard irq
5303 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was
5304 * set in napi_schedule_prep().
5305 * Since we are about to call napi->poll() once more, we can safely
5306 * clear NAPI_STATE_MISSED.
5308 * Note: x86 could use a single "lock and ..." instruction
5309 * to perform these two clear_bit()
5311 clear_bit(NAPI_STATE_MISSED, &napi->state);
5312 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state);
5316 /* All we really want here is to re-enable device interrupts.
5317 * Ideally, a new ndo_busy_poll_stop() could avoid another round.
5319 rc = napi->poll(napi, BUSY_POLL_BUDGET);
5320 netpoll_poll_unlock(have_poll_lock);
5321 if (rc == BUSY_POLL_BUDGET)
5322 __napi_schedule(napi);
5326 void napi_busy_loop(unsigned int napi_id,
5327 bool (*loop_end)(void *, unsigned long),
5330 unsigned long start_time = loop_end ? busy_loop_current_time() : 0;
5331 int (*napi_poll)(struct napi_struct *napi, int budget);
5332 void *have_poll_lock = NULL;
5333 struct napi_struct *napi;
5340 napi = napi_by_id(napi_id);
5350 unsigned long val = READ_ONCE(napi->state);
5352 /* If multiple threads are competing for this napi,
5353 * we avoid dirtying napi->state as much as we can.
5355 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED |
5356 NAPIF_STATE_IN_BUSY_POLL))
5358 if (cmpxchg(&napi->state, val,
5359 val | NAPIF_STATE_IN_BUSY_POLL |
5360 NAPIF_STATE_SCHED) != val)
5362 have_poll_lock = netpoll_poll_lock(napi);
5363 napi_poll = napi->poll;
5365 work = napi_poll(napi, BUSY_POLL_BUDGET);
5366 trace_napi_poll(napi, work, BUSY_POLL_BUDGET);
5369 __NET_ADD_STATS(dev_net(napi->dev),
5370 LINUX_MIB_BUSYPOLLRXPACKETS, work);
5373 if (!loop_end || loop_end(loop_end_arg, start_time))
5376 if (unlikely(need_resched())) {
5378 busy_poll_stop(napi, have_poll_lock);
5382 if (loop_end(loop_end_arg, start_time))
5389 busy_poll_stop(napi, have_poll_lock);
5394 EXPORT_SYMBOL(napi_busy_loop);
5396 #endif /* CONFIG_NET_RX_BUSY_POLL */
5398 static void napi_hash_add(struct napi_struct *napi)
5400 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state) ||
5401 test_and_set_bit(NAPI_STATE_HASHED, &napi->state))
5404 spin_lock(&napi_hash_lock);
5406 /* 0..NR_CPUS range is reserved for sender_cpu use */
5408 if (unlikely(++napi_gen_id < MIN_NAPI_ID))
5409 napi_gen_id = MIN_NAPI_ID;
5410 } while (napi_by_id(napi_gen_id));
5411 napi->napi_id = napi_gen_id;
5413 hlist_add_head_rcu(&napi->napi_hash_node,
5414 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]);
5416 spin_unlock(&napi_hash_lock);
5419 /* Warning : caller is responsible to make sure rcu grace period
5420 * is respected before freeing memory containing @napi
5422 bool napi_hash_del(struct napi_struct *napi)
5424 bool rcu_sync_needed = false;
5426 spin_lock(&napi_hash_lock);
5428 if (test_and_clear_bit(NAPI_STATE_HASHED, &napi->state)) {
5429 rcu_sync_needed = true;
5430 hlist_del_rcu(&napi->napi_hash_node);
5432 spin_unlock(&napi_hash_lock);
5433 return rcu_sync_needed;
5435 EXPORT_SYMBOL_GPL(napi_hash_del);
5437 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer)
5439 struct napi_struct *napi;
5441 napi = container_of(timer, struct napi_struct, timer);
5443 /* Note : we use a relaxed variant of napi_schedule_prep() not setting
5444 * NAPI_STATE_MISSED, since we do not react to a device IRQ.
5446 if (napi->gro_list && !napi_disable_pending(napi) &&
5447 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state))
5448 __napi_schedule_irqoff(napi);
5450 return HRTIMER_NORESTART;
5453 void netif_napi_add(struct net_device *dev, struct napi_struct *napi,
5454 int (*poll)(struct napi_struct *, int), int weight)
5456 INIT_LIST_HEAD(&napi->poll_list);
5457 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
5458 napi->timer.function = napi_watchdog;
5459 napi->gro_count = 0;
5460 napi->gro_list = NULL;
5463 if (weight > NAPI_POLL_WEIGHT)
5464 pr_err_once("netif_napi_add() called with weight %d on device %s\n",
5466 napi->weight = weight;
5467 list_add(&napi->dev_list, &dev->napi_list);
5469 #ifdef CONFIG_NETPOLL
5470 napi->poll_owner = -1;
5472 set_bit(NAPI_STATE_SCHED, &napi->state);
5473 napi_hash_add(napi);
5475 EXPORT_SYMBOL(netif_napi_add);
5477 void napi_disable(struct napi_struct *n)
5480 set_bit(NAPI_STATE_DISABLE, &n->state);
5482 while (test_and_set_bit(NAPI_STATE_SCHED, &n->state))
5484 while (test_and_set_bit(NAPI_STATE_NPSVC, &n->state))
5487 hrtimer_cancel(&n->timer);
5489 clear_bit(NAPI_STATE_DISABLE, &n->state);
5491 EXPORT_SYMBOL(napi_disable);
5493 /* Must be called in process context */
5494 void netif_napi_del(struct napi_struct *napi)
5497 if (napi_hash_del(napi))
5499 list_del_init(&napi->dev_list);
5500 napi_free_frags(napi);
5502 kfree_skb_list(napi->gro_list);
5503 napi->gro_list = NULL;
5504 napi->gro_count = 0;
5506 EXPORT_SYMBOL(netif_napi_del);
5508 static int napi_poll(struct napi_struct *n, struct list_head *repoll)
5513 list_del_init(&n->poll_list);
5515 have = netpoll_poll_lock(n);
5519 /* This NAPI_STATE_SCHED test is for avoiding a race
5520 * with netpoll's poll_napi(). Only the entity which
5521 * obtains the lock and sees NAPI_STATE_SCHED set will
5522 * actually make the ->poll() call. Therefore we avoid
5523 * accidentally calling ->poll() when NAPI is not scheduled.
5526 if (test_bit(NAPI_STATE_SCHED, &n->state)) {
5527 work = n->poll(n, weight);
5528 trace_napi_poll(n, work, weight);
5531 WARN_ON_ONCE(work > weight);
5533 if (likely(work < weight))
5536 /* Drivers must not modify the NAPI state if they
5537 * consume the entire weight. In such cases this code
5538 * still "owns" the NAPI instance and therefore can
5539 * move the instance around on the list at-will.
5541 if (unlikely(napi_disable_pending(n))) {
5547 /* flush too old packets
5548 * If HZ < 1000, flush all packets.
5550 napi_gro_flush(n, HZ >= 1000);
5553 /* Some drivers may have called napi_schedule
5554 * prior to exhausting their budget.
5556 if (unlikely(!list_empty(&n->poll_list))) {
5557 pr_warn_once("%s: Budget exhausted after napi rescheduled\n",
5558 n->dev ? n->dev->name : "backlog");
5562 list_add_tail(&n->poll_list, repoll);
5565 netpoll_poll_unlock(have);
5570 static __latent_entropy void net_rx_action(struct softirq_action *h)
5572 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
5573 unsigned long time_limit = jiffies +
5574 usecs_to_jiffies(netdev_budget_usecs);
5575 int budget = netdev_budget;
5579 local_irq_disable();
5580 list_splice_init(&sd->poll_list, &list);
5584 struct napi_struct *n;
5586 if (list_empty(&list)) {
5587 if (!sd_has_rps_ipi_waiting(sd) && list_empty(&repoll))
5592 n = list_first_entry(&list, struct napi_struct, poll_list);
5593 budget -= napi_poll(n, &repoll);
5595 /* If softirq window is exhausted then punt.
5596 * Allow this to run for 2 jiffies since which will allow
5597 * an average latency of 1.5/HZ.
5599 if (unlikely(budget <= 0 ||
5600 time_after_eq(jiffies, time_limit))) {
5606 local_irq_disable();
5608 list_splice_tail_init(&sd->poll_list, &list);
5609 list_splice_tail(&repoll, &list);
5610 list_splice(&list, &sd->poll_list);
5611 if (!list_empty(&sd->poll_list))
5612 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
5614 net_rps_action_and_irq_enable(sd);
5616 __kfree_skb_flush();
5619 struct netdev_adjacent {
5620 struct net_device *dev;
5622 /* upper master flag, there can only be one master device per list */
5625 /* counter for the number of times this device was added to us */
5628 /* private field for the users */
5631 struct list_head list;
5632 struct rcu_head rcu;
5635 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev,
5636 struct list_head *adj_list)
5638 struct netdev_adjacent *adj;
5640 list_for_each_entry(adj, adj_list, list) {
5641 if (adj->dev == adj_dev)
5647 static int __netdev_has_upper_dev(struct net_device *upper_dev, void *data)
5649 struct net_device *dev = data;
5651 return upper_dev == dev;
5655 * netdev_has_upper_dev - Check if device is linked to an upper device
5657 * @upper_dev: upper device to check
5659 * Find out if a device is linked to specified upper device and return true
5660 * in case it is. Note that this checks only immediate upper device,
5661 * not through a complete stack of devices. The caller must hold the RTNL lock.
5663 bool netdev_has_upper_dev(struct net_device *dev,
5664 struct net_device *upper_dev)
5668 return netdev_walk_all_upper_dev_rcu(dev, __netdev_has_upper_dev,
5671 EXPORT_SYMBOL(netdev_has_upper_dev);
5674 * netdev_has_upper_dev_all - Check if device is linked to an upper device
5676 * @upper_dev: upper device to check
5678 * Find out if a device is linked to specified upper device and return true
5679 * in case it is. Note that this checks the entire upper device chain.
5680 * The caller must hold rcu lock.
5683 bool netdev_has_upper_dev_all_rcu(struct net_device *dev,
5684 struct net_device *upper_dev)
5686 return !!netdev_walk_all_upper_dev_rcu(dev, __netdev_has_upper_dev,
5689 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu);
5692 * netdev_has_any_upper_dev - Check if device is linked to some device
5695 * Find out if a device is linked to an upper device and return true in case
5696 * it is. The caller must hold the RTNL lock.
5698 static bool netdev_has_any_upper_dev(struct net_device *dev)
5702 return !list_empty(&dev->adj_list.upper);
5706 * netdev_master_upper_dev_get - Get master upper device
5709 * Find a master upper device and return pointer to it or NULL in case
5710 * it's not there. The caller must hold the RTNL lock.
5712 struct net_device *netdev_master_upper_dev_get(struct net_device *dev)
5714 struct netdev_adjacent *upper;
5718 if (list_empty(&dev->adj_list.upper))
5721 upper = list_first_entry(&dev->adj_list.upper,
5722 struct netdev_adjacent, list);
5723 if (likely(upper->master))
5727 EXPORT_SYMBOL(netdev_master_upper_dev_get);
5730 * netdev_has_any_lower_dev - Check if device is linked to some device
5733 * Find out if a device is linked to a lower device and return true in case
5734 * it is. The caller must hold the RTNL lock.
5736 static bool netdev_has_any_lower_dev(struct net_device *dev)
5740 return !list_empty(&dev->adj_list.lower);
5743 void *netdev_adjacent_get_private(struct list_head *adj_list)
5745 struct netdev_adjacent *adj;
5747 adj = list_entry(adj_list, struct netdev_adjacent, list);
5749 return adj->private;
5751 EXPORT_SYMBOL(netdev_adjacent_get_private);
5754 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list
5756 * @iter: list_head ** of the current position
5758 * Gets the next device from the dev's upper list, starting from iter
5759 * position. The caller must hold RCU read lock.
5761 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev,
5762 struct list_head **iter)
5764 struct netdev_adjacent *upper;
5766 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
5768 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
5770 if (&upper->list == &dev->adj_list.upper)
5773 *iter = &upper->list;
5777 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu);
5779 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev,
5780 struct list_head **iter)
5782 struct netdev_adjacent *upper;
5784 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
5786 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
5788 if (&upper->list == &dev->adj_list.upper)
5791 *iter = &upper->list;
5796 int netdev_walk_all_upper_dev_rcu(struct net_device *dev,
5797 int (*fn)(struct net_device *dev,
5801 struct net_device *udev;
5802 struct list_head *iter;
5805 for (iter = &dev->adj_list.upper,
5806 udev = netdev_next_upper_dev_rcu(dev, &iter);
5808 udev = netdev_next_upper_dev_rcu(dev, &iter)) {
5809 /* first is the upper device itself */
5810 ret = fn(udev, data);
5814 /* then look at all of its upper devices */
5815 ret = netdev_walk_all_upper_dev_rcu(udev, fn, data);
5822 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu);
5825 * netdev_lower_get_next_private - Get the next ->private from the
5826 * lower neighbour list
5828 * @iter: list_head ** of the current position
5830 * Gets the next netdev_adjacent->private from the dev's lower neighbour
5831 * list, starting from iter position. The caller must hold either hold the
5832 * RTNL lock or its own locking that guarantees that the neighbour lower
5833 * list will remain unchanged.
5835 void *netdev_lower_get_next_private(struct net_device *dev,
5836 struct list_head **iter)
5838 struct netdev_adjacent *lower;
5840 lower = list_entry(*iter, struct netdev_adjacent, list);
5842 if (&lower->list == &dev->adj_list.lower)
5845 *iter = lower->list.next;
5847 return lower->private;
5849 EXPORT_SYMBOL(netdev_lower_get_next_private);
5852 * netdev_lower_get_next_private_rcu - Get the next ->private from the
5853 * lower neighbour list, RCU
5856 * @iter: list_head ** of the current position
5858 * Gets the next netdev_adjacent->private from the dev's lower neighbour
5859 * list, starting from iter position. The caller must hold RCU read lock.
5861 void *netdev_lower_get_next_private_rcu(struct net_device *dev,
5862 struct list_head **iter)
5864 struct netdev_adjacent *lower;
5866 WARN_ON_ONCE(!rcu_read_lock_held());
5868 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
5870 if (&lower->list == &dev->adj_list.lower)
5873 *iter = &lower->list;
5875 return lower->private;
5877 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu);
5880 * netdev_lower_get_next - Get the next device from the lower neighbour
5883 * @iter: list_head ** of the current position
5885 * Gets the next netdev_adjacent from the dev's lower neighbour
5886 * list, starting from iter position. The caller must hold RTNL lock or
5887 * its own locking that guarantees that the neighbour lower
5888 * list will remain unchanged.
5890 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter)
5892 struct netdev_adjacent *lower;
5894 lower = list_entry(*iter, struct netdev_adjacent, list);
5896 if (&lower->list == &dev->adj_list.lower)
5899 *iter = lower->list.next;
5903 EXPORT_SYMBOL(netdev_lower_get_next);
5905 static struct net_device *netdev_next_lower_dev(struct net_device *dev,
5906 struct list_head **iter)
5908 struct netdev_adjacent *lower;
5910 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
5912 if (&lower->list == &dev->adj_list.lower)
5915 *iter = &lower->list;
5920 int netdev_walk_all_lower_dev(struct net_device *dev,
5921 int (*fn)(struct net_device *dev,
5925 struct net_device *ldev;
5926 struct list_head *iter;
5929 for (iter = &dev->adj_list.lower,
5930 ldev = netdev_next_lower_dev(dev, &iter);
5932 ldev = netdev_next_lower_dev(dev, &iter)) {
5933 /* first is the lower device itself */
5934 ret = fn(ldev, data);
5938 /* then look at all of its lower devices */
5939 ret = netdev_walk_all_lower_dev(ldev, fn, data);
5946 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev);
5948 static struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev,
5949 struct list_head **iter)
5951 struct netdev_adjacent *lower;
5953 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
5954 if (&lower->list == &dev->adj_list.lower)
5957 *iter = &lower->list;
5962 int netdev_walk_all_lower_dev_rcu(struct net_device *dev,
5963 int (*fn)(struct net_device *dev,
5967 struct net_device *ldev;
5968 struct list_head *iter;
5971 for (iter = &dev->adj_list.lower,
5972 ldev = netdev_next_lower_dev_rcu(dev, &iter);
5974 ldev = netdev_next_lower_dev_rcu(dev, &iter)) {
5975 /* first is the lower device itself */
5976 ret = fn(ldev, data);
5980 /* then look at all of its lower devices */
5981 ret = netdev_walk_all_lower_dev_rcu(ldev, fn, data);
5988 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu);
5991 * netdev_lower_get_first_private_rcu - Get the first ->private from the
5992 * lower neighbour list, RCU
5996 * Gets the first netdev_adjacent->private from the dev's lower neighbour
5997 * list. The caller must hold RCU read lock.
5999 void *netdev_lower_get_first_private_rcu(struct net_device *dev)
6001 struct netdev_adjacent *lower;
6003 lower = list_first_or_null_rcu(&dev->adj_list.lower,
6004 struct netdev_adjacent, list);
6006 return lower->private;
6009 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu);
6012 * netdev_master_upper_dev_get_rcu - Get master upper device
6015 * Find a master upper device and return pointer to it or NULL in case
6016 * it's not there. The caller must hold the RCU read lock.
6018 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev)
6020 struct netdev_adjacent *upper;
6022 upper = list_first_or_null_rcu(&dev->adj_list.upper,
6023 struct netdev_adjacent, list);
6024 if (upper && likely(upper->master))
6028 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu);
6030 static int netdev_adjacent_sysfs_add(struct net_device *dev,
6031 struct net_device *adj_dev,
6032 struct list_head *dev_list)
6034 char linkname[IFNAMSIZ+7];
6036 sprintf(linkname, dev_list == &dev->adj_list.upper ?
6037 "upper_%s" : "lower_%s", adj_dev->name);
6038 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj),
6041 static void netdev_adjacent_sysfs_del(struct net_device *dev,
6043 struct list_head *dev_list)
6045 char linkname[IFNAMSIZ+7];
6047 sprintf(linkname, dev_list == &dev->adj_list.upper ?
6048 "upper_%s" : "lower_%s", name);
6049 sysfs_remove_link(&(dev->dev.kobj), linkname);
6052 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev,
6053 struct net_device *adj_dev,
6054 struct list_head *dev_list)
6056 return (dev_list == &dev->adj_list.upper ||
6057 dev_list == &dev->adj_list.lower) &&
6058 net_eq(dev_net(dev), dev_net(adj_dev));
6061 static int __netdev_adjacent_dev_insert(struct net_device *dev,
6062 struct net_device *adj_dev,
6063 struct list_head *dev_list,
6064 void *private, bool master)
6066 struct netdev_adjacent *adj;
6069 adj = __netdev_find_adj(adj_dev, dev_list);
6073 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n",
6074 dev->name, adj_dev->name, adj->ref_nr);
6079 adj = kmalloc(sizeof(*adj), GFP_KERNEL);
6084 adj->master = master;
6086 adj->private = private;
6089 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n",
6090 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name);
6092 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) {
6093 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list);
6098 /* Ensure that master link is always the first item in list. */
6100 ret = sysfs_create_link(&(dev->dev.kobj),
6101 &(adj_dev->dev.kobj), "master");
6103 goto remove_symlinks;
6105 list_add_rcu(&adj->list, dev_list);
6107 list_add_tail_rcu(&adj->list, dev_list);
6113 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
6114 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
6122 static void __netdev_adjacent_dev_remove(struct net_device *dev,
6123 struct net_device *adj_dev,
6125 struct list_head *dev_list)
6127 struct netdev_adjacent *adj;
6129 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n",
6130 dev->name, adj_dev->name, ref_nr);
6132 adj = __netdev_find_adj(adj_dev, dev_list);
6135 pr_err("Adjacency does not exist for device %s from %s\n",
6136 dev->name, adj_dev->name);
6141 if (adj->ref_nr > ref_nr) {
6142 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n",
6143 dev->name, adj_dev->name, ref_nr,
6144 adj->ref_nr - ref_nr);
6145 adj->ref_nr -= ref_nr;
6150 sysfs_remove_link(&(dev->dev.kobj), "master");
6152 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
6153 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
6155 list_del_rcu(&adj->list);
6156 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n",
6157 adj_dev->name, dev->name, adj_dev->name);
6159 kfree_rcu(adj, rcu);
6162 static int __netdev_adjacent_dev_link_lists(struct net_device *dev,
6163 struct net_device *upper_dev,
6164 struct list_head *up_list,
6165 struct list_head *down_list,
6166 void *private, bool master)
6170 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list,
6175 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list,
6178 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list);
6185 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev,
6186 struct net_device *upper_dev,
6188 struct list_head *up_list,
6189 struct list_head *down_list)
6191 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list);
6192 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list);
6195 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev,
6196 struct net_device *upper_dev,
6197 void *private, bool master)
6199 return __netdev_adjacent_dev_link_lists(dev, upper_dev,
6200 &dev->adj_list.upper,
6201 &upper_dev->adj_list.lower,
6205 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev,
6206 struct net_device *upper_dev)
6208 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1,
6209 &dev->adj_list.upper,
6210 &upper_dev->adj_list.lower);
6213 static int __netdev_upper_dev_link(struct net_device *dev,
6214 struct net_device *upper_dev, bool master,
6215 void *upper_priv, void *upper_info)
6217 struct netdev_notifier_changeupper_info changeupper_info;
6222 if (dev == upper_dev)
6225 /* To prevent loops, check if dev is not upper device to upper_dev. */
6226 if (netdev_has_upper_dev(upper_dev, dev))
6229 if (netdev_has_upper_dev(dev, upper_dev))
6232 if (master && netdev_master_upper_dev_get(dev))
6235 changeupper_info.upper_dev = upper_dev;
6236 changeupper_info.master = master;
6237 changeupper_info.linking = true;
6238 changeupper_info.upper_info = upper_info;
6240 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, dev,
6241 &changeupper_info.info);
6242 ret = notifier_to_errno(ret);
6246 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv,
6251 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, dev,
6252 &changeupper_info.info);
6253 ret = notifier_to_errno(ret);
6260 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
6266 * netdev_upper_dev_link - Add a link to the upper device
6268 * @upper_dev: new upper device
6270 * Adds a link to device which is upper to this one. The caller must hold
6271 * the RTNL lock. On a failure a negative errno code is returned.
6272 * On success the reference counts are adjusted and the function
6275 int netdev_upper_dev_link(struct net_device *dev,
6276 struct net_device *upper_dev)
6278 return __netdev_upper_dev_link(dev, upper_dev, false, NULL, NULL);
6280 EXPORT_SYMBOL(netdev_upper_dev_link);
6283 * netdev_master_upper_dev_link - Add a master link to the upper device
6285 * @upper_dev: new upper device
6286 * @upper_priv: upper device private
6287 * @upper_info: upper info to be passed down via notifier
6289 * Adds a link to device which is upper to this one. In this case, only
6290 * one master upper device can be linked, although other non-master devices
6291 * might be linked as well. The caller must hold the RTNL lock.
6292 * On a failure a negative errno code is returned. On success the reference
6293 * counts are adjusted and the function returns zero.
6295 int netdev_master_upper_dev_link(struct net_device *dev,
6296 struct net_device *upper_dev,
6297 void *upper_priv, void *upper_info)
6299 return __netdev_upper_dev_link(dev, upper_dev, true,
6300 upper_priv, upper_info);
6302 EXPORT_SYMBOL(netdev_master_upper_dev_link);
6305 * netdev_upper_dev_unlink - Removes a link to upper device
6307 * @upper_dev: new upper device
6309 * Removes a link to device which is upper to this one. The caller must hold
6312 void netdev_upper_dev_unlink(struct net_device *dev,
6313 struct net_device *upper_dev)
6315 struct netdev_notifier_changeupper_info changeupper_info;
6319 changeupper_info.upper_dev = upper_dev;
6320 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev;
6321 changeupper_info.linking = false;
6323 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, dev,
6324 &changeupper_info.info);
6326 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
6328 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, dev,
6329 &changeupper_info.info);
6331 EXPORT_SYMBOL(netdev_upper_dev_unlink);
6334 * netdev_bonding_info_change - Dispatch event about slave change
6336 * @bonding_info: info to dispatch
6338 * Send NETDEV_BONDING_INFO to netdev notifiers with info.
6339 * The caller must hold the RTNL lock.
6341 void netdev_bonding_info_change(struct net_device *dev,
6342 struct netdev_bonding_info *bonding_info)
6344 struct netdev_notifier_bonding_info info;
6346 memcpy(&info.bonding_info, bonding_info,
6347 sizeof(struct netdev_bonding_info));
6348 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, dev,
6351 EXPORT_SYMBOL(netdev_bonding_info_change);
6353 static void netdev_adjacent_add_links(struct net_device *dev)
6355 struct netdev_adjacent *iter;
6357 struct net *net = dev_net(dev);
6359 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6360 if (!net_eq(net, dev_net(iter->dev)))
6362 netdev_adjacent_sysfs_add(iter->dev, dev,
6363 &iter->dev->adj_list.lower);
6364 netdev_adjacent_sysfs_add(dev, iter->dev,
6365 &dev->adj_list.upper);
6368 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6369 if (!net_eq(net, dev_net(iter->dev)))
6371 netdev_adjacent_sysfs_add(iter->dev, dev,
6372 &iter->dev->adj_list.upper);
6373 netdev_adjacent_sysfs_add(dev, iter->dev,
6374 &dev->adj_list.lower);
6378 static void netdev_adjacent_del_links(struct net_device *dev)
6380 struct netdev_adjacent *iter;
6382 struct net *net = dev_net(dev);
6384 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6385 if (!net_eq(net, dev_net(iter->dev)))
6387 netdev_adjacent_sysfs_del(iter->dev, dev->name,
6388 &iter->dev->adj_list.lower);
6389 netdev_adjacent_sysfs_del(dev, iter->dev->name,
6390 &dev->adj_list.upper);
6393 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6394 if (!net_eq(net, dev_net(iter->dev)))
6396 netdev_adjacent_sysfs_del(iter->dev, dev->name,
6397 &iter->dev->adj_list.upper);
6398 netdev_adjacent_sysfs_del(dev, iter->dev->name,
6399 &dev->adj_list.lower);
6403 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname)
6405 struct netdev_adjacent *iter;
6407 struct net *net = dev_net(dev);
6409 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6410 if (!net_eq(net, dev_net(iter->dev)))
6412 netdev_adjacent_sysfs_del(iter->dev, oldname,
6413 &iter->dev->adj_list.lower);
6414 netdev_adjacent_sysfs_add(iter->dev, dev,
6415 &iter->dev->adj_list.lower);
6418 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6419 if (!net_eq(net, dev_net(iter->dev)))
6421 netdev_adjacent_sysfs_del(iter->dev, oldname,
6422 &iter->dev->adj_list.upper);
6423 netdev_adjacent_sysfs_add(iter->dev, dev,
6424 &iter->dev->adj_list.upper);
6428 void *netdev_lower_dev_get_private(struct net_device *dev,
6429 struct net_device *lower_dev)
6431 struct netdev_adjacent *lower;
6435 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower);
6439 return lower->private;
6441 EXPORT_SYMBOL(netdev_lower_dev_get_private);
6444 int dev_get_nest_level(struct net_device *dev)
6446 struct net_device *lower = NULL;
6447 struct list_head *iter;
6453 netdev_for_each_lower_dev(dev, lower, iter) {
6454 nest = dev_get_nest_level(lower);
6455 if (max_nest < nest)
6459 return max_nest + 1;
6461 EXPORT_SYMBOL(dev_get_nest_level);
6464 * netdev_lower_change - Dispatch event about lower device state change
6465 * @lower_dev: device
6466 * @lower_state_info: state to dispatch
6468 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info.
6469 * The caller must hold the RTNL lock.
6471 void netdev_lower_state_changed(struct net_device *lower_dev,
6472 void *lower_state_info)
6474 struct netdev_notifier_changelowerstate_info changelowerstate_info;
6477 changelowerstate_info.lower_state_info = lower_state_info;
6478 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, lower_dev,
6479 &changelowerstate_info.info);
6481 EXPORT_SYMBOL(netdev_lower_state_changed);
6483 static void dev_change_rx_flags(struct net_device *dev, int flags)
6485 const struct net_device_ops *ops = dev->netdev_ops;
6487 if (ops->ndo_change_rx_flags)
6488 ops->ndo_change_rx_flags(dev, flags);
6491 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify)
6493 unsigned int old_flags = dev->flags;
6499 dev->flags |= IFF_PROMISC;
6500 dev->promiscuity += inc;
6501 if (dev->promiscuity == 0) {
6504 * If inc causes overflow, untouch promisc and return error.
6507 dev->flags &= ~IFF_PROMISC;
6509 dev->promiscuity -= inc;
6510 pr_warn("%s: promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n",
6515 if (dev->flags != old_flags) {
6516 pr_info("device %s %s promiscuous mode\n",
6518 dev->flags & IFF_PROMISC ? "entered" : "left");
6519 if (audit_enabled) {
6520 current_uid_gid(&uid, &gid);
6521 audit_log(current->audit_context, GFP_ATOMIC,
6522 AUDIT_ANOM_PROMISCUOUS,
6523 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u",
6524 dev->name, (dev->flags & IFF_PROMISC),
6525 (old_flags & IFF_PROMISC),
6526 from_kuid(&init_user_ns, audit_get_loginuid(current)),
6527 from_kuid(&init_user_ns, uid),
6528 from_kgid(&init_user_ns, gid),
6529 audit_get_sessionid(current));
6532 dev_change_rx_flags(dev, IFF_PROMISC);
6535 __dev_notify_flags(dev, old_flags, IFF_PROMISC);
6540 * dev_set_promiscuity - update promiscuity count on a device
6544 * Add or remove promiscuity from a device. While the count in the device
6545 * remains above zero the interface remains promiscuous. Once it hits zero
6546 * the device reverts back to normal filtering operation. A negative inc
6547 * value is used to drop promiscuity on the device.
6548 * Return 0 if successful or a negative errno code on error.
6550 int dev_set_promiscuity(struct net_device *dev, int inc)
6552 unsigned int old_flags = dev->flags;
6555 err = __dev_set_promiscuity(dev, inc, true);
6558 if (dev->flags != old_flags)
6559 dev_set_rx_mode(dev);
6562 EXPORT_SYMBOL(dev_set_promiscuity);
6564 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify)
6566 unsigned int old_flags = dev->flags, old_gflags = dev->gflags;
6570 dev->flags |= IFF_ALLMULTI;
6571 dev->allmulti += inc;
6572 if (dev->allmulti == 0) {
6575 * If inc causes overflow, untouch allmulti and return error.
6578 dev->flags &= ~IFF_ALLMULTI;
6580 dev->allmulti -= inc;
6581 pr_warn("%s: allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n",
6586 if (dev->flags ^ old_flags) {
6587 dev_change_rx_flags(dev, IFF_ALLMULTI);
6588 dev_set_rx_mode(dev);
6590 __dev_notify_flags(dev, old_flags,
6591 dev->gflags ^ old_gflags);
6597 * dev_set_allmulti - update allmulti count on a device
6601 * Add or remove reception of all multicast frames to a device. While the
6602 * count in the device remains above zero the interface remains listening
6603 * to all interfaces. Once it hits zero the device reverts back to normal
6604 * filtering operation. A negative @inc value is used to drop the counter
6605 * when releasing a resource needing all multicasts.
6606 * Return 0 if successful or a negative errno code on error.
6609 int dev_set_allmulti(struct net_device *dev, int inc)
6611 return __dev_set_allmulti(dev, inc, true);
6613 EXPORT_SYMBOL(dev_set_allmulti);
6616 * Upload unicast and multicast address lists to device and
6617 * configure RX filtering. When the device doesn't support unicast
6618 * filtering it is put in promiscuous mode while unicast addresses
6621 void __dev_set_rx_mode(struct net_device *dev)
6623 const struct net_device_ops *ops = dev->netdev_ops;
6625 /* dev_open will call this function so the list will stay sane. */
6626 if (!(dev->flags&IFF_UP))
6629 if (!netif_device_present(dev))
6632 if (!(dev->priv_flags & IFF_UNICAST_FLT)) {
6633 /* Unicast addresses changes may only happen under the rtnl,
6634 * therefore calling __dev_set_promiscuity here is safe.
6636 if (!netdev_uc_empty(dev) && !dev->uc_promisc) {
6637 __dev_set_promiscuity(dev, 1, false);
6638 dev->uc_promisc = true;
6639 } else if (netdev_uc_empty(dev) && dev->uc_promisc) {
6640 __dev_set_promiscuity(dev, -1, false);
6641 dev->uc_promisc = false;
6645 if (ops->ndo_set_rx_mode)
6646 ops->ndo_set_rx_mode(dev);
6649 void dev_set_rx_mode(struct net_device *dev)
6651 netif_addr_lock_bh(dev);
6652 __dev_set_rx_mode(dev);
6653 netif_addr_unlock_bh(dev);
6657 * dev_get_flags - get flags reported to userspace
6660 * Get the combination of flag bits exported through APIs to userspace.
6662 unsigned int dev_get_flags(const struct net_device *dev)
6666 flags = (dev->flags & ~(IFF_PROMISC |
6671 (dev->gflags & (IFF_PROMISC |
6674 if (netif_running(dev)) {
6675 if (netif_oper_up(dev))
6676 flags |= IFF_RUNNING;
6677 if (netif_carrier_ok(dev))
6678 flags |= IFF_LOWER_UP;
6679 if (netif_dormant(dev))
6680 flags |= IFF_DORMANT;
6685 EXPORT_SYMBOL(dev_get_flags);
6687 int __dev_change_flags(struct net_device *dev, unsigned int flags)
6689 unsigned int old_flags = dev->flags;
6695 * Set the flags on our device.
6698 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP |
6699 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL |
6701 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC |
6705 * Load in the correct multicast list now the flags have changed.
6708 if ((old_flags ^ flags) & IFF_MULTICAST)
6709 dev_change_rx_flags(dev, IFF_MULTICAST);
6711 dev_set_rx_mode(dev);
6714 * Have we downed the interface. We handle IFF_UP ourselves
6715 * according to user attempts to set it, rather than blindly
6720 if ((old_flags ^ flags) & IFF_UP) {
6721 if (old_flags & IFF_UP)
6724 ret = __dev_open(dev);
6727 if ((flags ^ dev->gflags) & IFF_PROMISC) {
6728 int inc = (flags & IFF_PROMISC) ? 1 : -1;
6729 unsigned int old_flags = dev->flags;
6731 dev->gflags ^= IFF_PROMISC;
6733 if (__dev_set_promiscuity(dev, inc, false) >= 0)
6734 if (dev->flags != old_flags)
6735 dev_set_rx_mode(dev);
6738 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI
6739 * is important. Some (broken) drivers set IFF_PROMISC, when
6740 * IFF_ALLMULTI is requested not asking us and not reporting.
6742 if ((flags ^ dev->gflags) & IFF_ALLMULTI) {
6743 int inc = (flags & IFF_ALLMULTI) ? 1 : -1;
6745 dev->gflags ^= IFF_ALLMULTI;
6746 __dev_set_allmulti(dev, inc, false);
6752 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags,
6753 unsigned int gchanges)
6755 unsigned int changes = dev->flags ^ old_flags;
6758 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC);
6760 if (changes & IFF_UP) {
6761 if (dev->flags & IFF_UP)
6762 call_netdevice_notifiers(NETDEV_UP, dev);
6764 call_netdevice_notifiers(NETDEV_DOWN, dev);
6767 if (dev->flags & IFF_UP &&
6768 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) {
6769 struct netdev_notifier_change_info change_info;
6771 change_info.flags_changed = changes;
6772 call_netdevice_notifiers_info(NETDEV_CHANGE, dev,
6778 * dev_change_flags - change device settings
6780 * @flags: device state flags
6782 * Change settings on device based state flags. The flags are
6783 * in the userspace exported format.
6785 int dev_change_flags(struct net_device *dev, unsigned int flags)
6788 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags;
6790 ret = __dev_change_flags(dev, flags);
6794 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags);
6795 __dev_notify_flags(dev, old_flags, changes);
6798 EXPORT_SYMBOL(dev_change_flags);
6800 int __dev_set_mtu(struct net_device *dev, int new_mtu)
6802 const struct net_device_ops *ops = dev->netdev_ops;
6804 if (ops->ndo_change_mtu)
6805 return ops->ndo_change_mtu(dev, new_mtu);
6810 EXPORT_SYMBOL(__dev_set_mtu);
6813 * dev_set_mtu - Change maximum transfer unit
6815 * @new_mtu: new transfer unit
6817 * Change the maximum transfer size of the network device.
6819 int dev_set_mtu(struct net_device *dev, int new_mtu)
6823 if (new_mtu == dev->mtu)
6826 /* MTU must be positive, and in range */
6827 if (new_mtu < 0 || new_mtu < dev->min_mtu) {
6828 net_err_ratelimited("%s: Invalid MTU %d requested, hw min %d\n",
6829 dev->name, new_mtu, dev->min_mtu);
6833 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) {
6834 net_err_ratelimited("%s: Invalid MTU %d requested, hw max %d\n",
6835 dev->name, new_mtu, dev->max_mtu);
6839 if (!netif_device_present(dev))
6842 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev);
6843 err = notifier_to_errno(err);
6847 orig_mtu = dev->mtu;
6848 err = __dev_set_mtu(dev, new_mtu);
6851 err = call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
6852 err = notifier_to_errno(err);
6854 /* setting mtu back and notifying everyone again,
6855 * so that they have a chance to revert changes.
6857 __dev_set_mtu(dev, orig_mtu);
6858 call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
6863 EXPORT_SYMBOL(dev_set_mtu);
6866 * dev_set_group - Change group this device belongs to
6868 * @new_group: group this device should belong to
6870 void dev_set_group(struct net_device *dev, int new_group)
6872 dev->group = new_group;
6874 EXPORT_SYMBOL(dev_set_group);
6877 * dev_set_mac_address - Change Media Access Control Address
6881 * Change the hardware (MAC) address of the device
6883 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa)
6885 const struct net_device_ops *ops = dev->netdev_ops;
6888 if (!ops->ndo_set_mac_address)
6890 if (sa->sa_family != dev->type)
6892 if (!netif_device_present(dev))
6894 err = ops->ndo_set_mac_address(dev, sa);
6897 dev->addr_assign_type = NET_ADDR_SET;
6898 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
6899 add_device_randomness(dev->dev_addr, dev->addr_len);
6902 EXPORT_SYMBOL(dev_set_mac_address);
6905 * dev_change_carrier - Change device carrier
6907 * @new_carrier: new value
6909 * Change device carrier
6911 int dev_change_carrier(struct net_device *dev, bool new_carrier)
6913 const struct net_device_ops *ops = dev->netdev_ops;
6915 if (!ops->ndo_change_carrier)
6917 if (!netif_device_present(dev))
6919 return ops->ndo_change_carrier(dev, new_carrier);
6921 EXPORT_SYMBOL(dev_change_carrier);
6924 * dev_get_phys_port_id - Get device physical port ID
6928 * Get device physical port ID
6930 int dev_get_phys_port_id(struct net_device *dev,
6931 struct netdev_phys_item_id *ppid)
6933 const struct net_device_ops *ops = dev->netdev_ops;
6935 if (!ops->ndo_get_phys_port_id)
6937 return ops->ndo_get_phys_port_id(dev, ppid);
6939 EXPORT_SYMBOL(dev_get_phys_port_id);
6942 * dev_get_phys_port_name - Get device physical port name
6945 * @len: limit of bytes to copy to name
6947 * Get device physical port name
6949 int dev_get_phys_port_name(struct net_device *dev,
6950 char *name, size_t len)
6952 const struct net_device_ops *ops = dev->netdev_ops;
6954 if (!ops->ndo_get_phys_port_name)
6956 return ops->ndo_get_phys_port_name(dev, name, len);
6958 EXPORT_SYMBOL(dev_get_phys_port_name);
6961 * dev_change_proto_down - update protocol port state information
6963 * @proto_down: new value
6965 * This info can be used by switch drivers to set the phys state of the
6968 int dev_change_proto_down(struct net_device *dev, bool proto_down)
6970 const struct net_device_ops *ops = dev->netdev_ops;
6972 if (!ops->ndo_change_proto_down)
6974 if (!netif_device_present(dev))
6976 return ops->ndo_change_proto_down(dev, proto_down);
6978 EXPORT_SYMBOL(dev_change_proto_down);
6980 u8 __dev_xdp_attached(struct net_device *dev, xdp_op_t xdp_op, u32 *prog_id)
6982 struct netdev_xdp xdp;
6984 memset(&xdp, 0, sizeof(xdp));
6985 xdp.command = XDP_QUERY_PROG;
6987 /* Query must always succeed. */
6988 WARN_ON(xdp_op(dev, &xdp) < 0);
6990 *prog_id = xdp.prog_id;
6992 return xdp.prog_attached;
6995 static int dev_xdp_install(struct net_device *dev, xdp_op_t xdp_op,
6996 struct netlink_ext_ack *extack, u32 flags,
6997 struct bpf_prog *prog)
6999 struct netdev_xdp xdp;
7001 memset(&xdp, 0, sizeof(xdp));
7002 if (flags & XDP_FLAGS_HW_MODE)
7003 xdp.command = XDP_SETUP_PROG_HW;
7005 xdp.command = XDP_SETUP_PROG;
7006 xdp.extack = extack;
7010 return xdp_op(dev, &xdp);
7014 * dev_change_xdp_fd - set or clear a bpf program for a device rx path
7016 * @extack: netlink extended ack
7017 * @fd: new program fd or negative value to clear
7018 * @flags: xdp-related flags
7020 * Set or clear a bpf program for a device
7022 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack,
7025 const struct net_device_ops *ops = dev->netdev_ops;
7026 struct bpf_prog *prog = NULL;
7027 xdp_op_t xdp_op, xdp_chk;
7032 xdp_op = xdp_chk = ops->ndo_xdp;
7033 if (!xdp_op && (flags & (XDP_FLAGS_DRV_MODE | XDP_FLAGS_HW_MODE)))
7035 if (!xdp_op || (flags & XDP_FLAGS_SKB_MODE))
7036 xdp_op = generic_xdp_install;
7037 if (xdp_op == xdp_chk)
7038 xdp_chk = generic_xdp_install;
7041 if (xdp_chk && __dev_xdp_attached(dev, xdp_chk, NULL))
7043 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) &&
7044 __dev_xdp_attached(dev, xdp_op, NULL))
7047 prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP);
7049 return PTR_ERR(prog);
7052 err = dev_xdp_install(dev, xdp_op, extack, flags, prog);
7053 if (err < 0 && prog)
7060 * dev_new_index - allocate an ifindex
7061 * @net: the applicable net namespace
7063 * Returns a suitable unique value for a new device interface
7064 * number. The caller must hold the rtnl semaphore or the
7065 * dev_base_lock to be sure it remains unique.
7067 static int dev_new_index(struct net *net)
7069 int ifindex = net->ifindex;
7074 if (!__dev_get_by_index(net, ifindex))
7075 return net->ifindex = ifindex;
7079 /* Delayed registration/unregisteration */
7080 static LIST_HEAD(net_todo_list);
7081 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq);
7083 static void net_set_todo(struct net_device *dev)
7085 list_add_tail(&dev->todo_list, &net_todo_list);
7086 dev_net(dev)->dev_unreg_count++;
7089 static void rollback_registered_many(struct list_head *head)
7091 struct net_device *dev, *tmp;
7092 LIST_HEAD(close_head);
7094 BUG_ON(dev_boot_phase);
7097 list_for_each_entry_safe(dev, tmp, head, unreg_list) {
7098 /* Some devices call without registering
7099 * for initialization unwind. Remove those
7100 * devices and proceed with the remaining.
7102 if (dev->reg_state == NETREG_UNINITIALIZED) {
7103 pr_debug("unregister_netdevice: device %s/%p never was registered\n",
7107 list_del(&dev->unreg_list);
7110 dev->dismantle = true;
7111 BUG_ON(dev->reg_state != NETREG_REGISTERED);
7114 /* If device is running, close it first. */
7115 list_for_each_entry(dev, head, unreg_list)
7116 list_add_tail(&dev->close_list, &close_head);
7117 dev_close_many(&close_head, true);
7119 list_for_each_entry(dev, head, unreg_list) {
7120 /* And unlink it from device chain. */
7121 unlist_netdevice(dev);
7123 dev->reg_state = NETREG_UNREGISTERING;
7125 flush_all_backlogs();
7129 list_for_each_entry(dev, head, unreg_list) {
7130 struct sk_buff *skb = NULL;
7132 /* Shutdown queueing discipline. */
7136 /* Notify protocols, that we are about to destroy
7137 * this device. They should clean all the things.
7139 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
7141 if (!dev->rtnl_link_ops ||
7142 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
7143 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0,
7147 * Flush the unicast and multicast chains
7152 if (dev->netdev_ops->ndo_uninit)
7153 dev->netdev_ops->ndo_uninit(dev);
7156 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL);
7158 /* Notifier chain MUST detach us all upper devices. */
7159 WARN_ON(netdev_has_any_upper_dev(dev));
7160 WARN_ON(netdev_has_any_lower_dev(dev));
7162 /* Remove entries from kobject tree */
7163 netdev_unregister_kobject(dev);
7165 /* Remove XPS queueing entries */
7166 netif_reset_xps_queues_gt(dev, 0);
7172 list_for_each_entry(dev, head, unreg_list)
7176 static void rollback_registered(struct net_device *dev)
7180 list_add(&dev->unreg_list, &single);
7181 rollback_registered_many(&single);
7185 static netdev_features_t netdev_sync_upper_features(struct net_device *lower,
7186 struct net_device *upper, netdev_features_t features)
7188 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
7189 netdev_features_t feature;
7192 for_each_netdev_feature(&upper_disables, feature_bit) {
7193 feature = __NETIF_F_BIT(feature_bit);
7194 if (!(upper->wanted_features & feature)
7195 && (features & feature)) {
7196 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n",
7197 &feature, upper->name);
7198 features &= ~feature;
7205 static void netdev_sync_lower_features(struct net_device *upper,
7206 struct net_device *lower, netdev_features_t features)
7208 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
7209 netdev_features_t feature;
7212 for_each_netdev_feature(&upper_disables, feature_bit) {
7213 feature = __NETIF_F_BIT(feature_bit);
7214 if (!(features & feature) && (lower->features & feature)) {
7215 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n",
7216 &feature, lower->name);
7217 lower->wanted_features &= ~feature;
7218 netdev_update_features(lower);
7220 if (unlikely(lower->features & feature))
7221 netdev_WARN(upper, "failed to disable %pNF on %s!\n",
7222 &feature, lower->name);
7227 static netdev_features_t netdev_fix_features(struct net_device *dev,
7228 netdev_features_t features)
7230 /* Fix illegal checksum combinations */
7231 if ((features & NETIF_F_HW_CSUM) &&
7232 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
7233 netdev_warn(dev, "mixed HW and IP checksum settings.\n");
7234 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
7237 /* TSO requires that SG is present as well. */
7238 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) {
7239 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n");
7240 features &= ~NETIF_F_ALL_TSO;
7243 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) &&
7244 !(features & NETIF_F_IP_CSUM)) {
7245 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n");
7246 features &= ~NETIF_F_TSO;
7247 features &= ~NETIF_F_TSO_ECN;
7250 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) &&
7251 !(features & NETIF_F_IPV6_CSUM)) {
7252 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n");
7253 features &= ~NETIF_F_TSO6;
7256 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */
7257 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO))
7258 features &= ~NETIF_F_TSO_MANGLEID;
7260 /* TSO ECN requires that TSO is present as well. */
7261 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN)
7262 features &= ~NETIF_F_TSO_ECN;
7264 /* Software GSO depends on SG. */
7265 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) {
7266 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n");
7267 features &= ~NETIF_F_GSO;
7270 /* GSO partial features require GSO partial be set */
7271 if ((features & dev->gso_partial_features) &&
7272 !(features & NETIF_F_GSO_PARTIAL)) {
7274 "Dropping partially supported GSO features since no GSO partial.\n");
7275 features &= ~dev->gso_partial_features;
7281 int __netdev_update_features(struct net_device *dev)
7283 struct net_device *upper, *lower;
7284 netdev_features_t features;
7285 struct list_head *iter;
7290 features = netdev_get_wanted_features(dev);
7292 if (dev->netdev_ops->ndo_fix_features)
7293 features = dev->netdev_ops->ndo_fix_features(dev, features);
7295 /* driver might be less strict about feature dependencies */
7296 features = netdev_fix_features(dev, features);
7298 /* some features can't be enabled if they're off an an upper device */
7299 netdev_for_each_upper_dev_rcu(dev, upper, iter)
7300 features = netdev_sync_upper_features(dev, upper, features);
7302 if (dev->features == features)
7305 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n",
7306 &dev->features, &features);
7308 if (dev->netdev_ops->ndo_set_features)
7309 err = dev->netdev_ops->ndo_set_features(dev, features);
7313 if (unlikely(err < 0)) {
7315 "set_features() failed (%d); wanted %pNF, left %pNF\n",
7316 err, &features, &dev->features);
7317 /* return non-0 since some features might have changed and
7318 * it's better to fire a spurious notification than miss it
7324 /* some features must be disabled on lower devices when disabled
7325 * on an upper device (think: bonding master or bridge)
7327 netdev_for_each_lower_dev(dev, lower, iter)
7328 netdev_sync_lower_features(dev, lower, features);
7331 dev->features = features;
7333 return err < 0 ? 0 : 1;
7337 * netdev_update_features - recalculate device features
7338 * @dev: the device to check
7340 * Recalculate dev->features set and send notifications if it
7341 * has changed. Should be called after driver or hardware dependent
7342 * conditions might have changed that influence the features.
7344 void netdev_update_features(struct net_device *dev)
7346 if (__netdev_update_features(dev))
7347 netdev_features_change(dev);
7349 EXPORT_SYMBOL(netdev_update_features);
7352 * netdev_change_features - recalculate device features
7353 * @dev: the device to check
7355 * Recalculate dev->features set and send notifications even
7356 * if they have not changed. Should be called instead of
7357 * netdev_update_features() if also dev->vlan_features might
7358 * have changed to allow the changes to be propagated to stacked
7361 void netdev_change_features(struct net_device *dev)
7363 __netdev_update_features(dev);
7364 netdev_features_change(dev);
7366 EXPORT_SYMBOL(netdev_change_features);
7369 * netif_stacked_transfer_operstate - transfer operstate
7370 * @rootdev: the root or lower level device to transfer state from
7371 * @dev: the device to transfer operstate to
7373 * Transfer operational state from root to device. This is normally
7374 * called when a stacking relationship exists between the root
7375 * device and the device(a leaf device).
7377 void netif_stacked_transfer_operstate(const struct net_device *rootdev,
7378 struct net_device *dev)
7380 if (rootdev->operstate == IF_OPER_DORMANT)
7381 netif_dormant_on(dev);
7383 netif_dormant_off(dev);
7385 if (netif_carrier_ok(rootdev))
7386 netif_carrier_on(dev);
7388 netif_carrier_off(dev);
7390 EXPORT_SYMBOL(netif_stacked_transfer_operstate);
7393 static int netif_alloc_rx_queues(struct net_device *dev)
7395 unsigned int i, count = dev->num_rx_queues;
7396 struct netdev_rx_queue *rx;
7397 size_t sz = count * sizeof(*rx);
7401 rx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
7407 for (i = 0; i < count; i++)
7413 static void netdev_init_one_queue(struct net_device *dev,
7414 struct netdev_queue *queue, void *_unused)
7416 /* Initialize queue lock */
7417 spin_lock_init(&queue->_xmit_lock);
7418 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type);
7419 queue->xmit_lock_owner = -1;
7420 netdev_queue_numa_node_write(queue, NUMA_NO_NODE);
7423 dql_init(&queue->dql, HZ);
7427 static void netif_free_tx_queues(struct net_device *dev)
7432 static int netif_alloc_netdev_queues(struct net_device *dev)
7434 unsigned int count = dev->num_tx_queues;
7435 struct netdev_queue *tx;
7436 size_t sz = count * sizeof(*tx);
7438 if (count < 1 || count > 0xffff)
7441 tx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
7447 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL);
7448 spin_lock_init(&dev->tx_global_lock);
7453 void netif_tx_stop_all_queues(struct net_device *dev)
7457 for (i = 0; i < dev->num_tx_queues; i++) {
7458 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
7460 netif_tx_stop_queue(txq);
7463 EXPORT_SYMBOL(netif_tx_stop_all_queues);
7466 * register_netdevice - register a network device
7467 * @dev: device to register
7469 * Take a completed network device structure and add it to the kernel
7470 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
7471 * chain. 0 is returned on success. A negative errno code is returned
7472 * on a failure to set up the device, or if the name is a duplicate.
7474 * Callers must hold the rtnl semaphore. You may want
7475 * register_netdev() instead of this.
7478 * The locking appears insufficient to guarantee two parallel registers
7479 * will not get the same name.
7482 int register_netdevice(struct net_device *dev)
7485 struct net *net = dev_net(dev);
7487 BUG_ON(dev_boot_phase);
7492 /* When net_device's are persistent, this will be fatal. */
7493 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
7496 spin_lock_init(&dev->addr_list_lock);
7497 netdev_set_addr_lockdep_class(dev);
7499 ret = dev_get_valid_name(net, dev, dev->name);
7503 /* Init, if this function is available */
7504 if (dev->netdev_ops->ndo_init) {
7505 ret = dev->netdev_ops->ndo_init(dev);
7513 if (((dev->hw_features | dev->features) &
7514 NETIF_F_HW_VLAN_CTAG_FILTER) &&
7515 (!dev->netdev_ops->ndo_vlan_rx_add_vid ||
7516 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) {
7517 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n");
7524 dev->ifindex = dev_new_index(net);
7525 else if (__dev_get_by_index(net, dev->ifindex))
7528 /* Transfer changeable features to wanted_features and enable
7529 * software offloads (GSO and GRO).
7531 dev->hw_features |= NETIF_F_SOFT_FEATURES;
7532 dev->features |= NETIF_F_SOFT_FEATURES;
7533 dev->wanted_features = dev->features & dev->hw_features;
7535 if (!(dev->flags & IFF_LOOPBACK))
7536 dev->hw_features |= NETIF_F_NOCACHE_COPY;
7538 /* If IPv4 TCP segmentation offload is supported we should also
7539 * allow the device to enable segmenting the frame with the option
7540 * of ignoring a static IP ID value. This doesn't enable the
7541 * feature itself but allows the user to enable it later.
7543 if (dev->hw_features & NETIF_F_TSO)
7544 dev->hw_features |= NETIF_F_TSO_MANGLEID;
7545 if (dev->vlan_features & NETIF_F_TSO)
7546 dev->vlan_features |= NETIF_F_TSO_MANGLEID;
7547 if (dev->mpls_features & NETIF_F_TSO)
7548 dev->mpls_features |= NETIF_F_TSO_MANGLEID;
7549 if (dev->hw_enc_features & NETIF_F_TSO)
7550 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID;
7552 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices.
7554 dev->vlan_features |= NETIF_F_HIGHDMA;
7556 /* Make NETIF_F_SG inheritable to tunnel devices.
7558 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL;
7560 /* Make NETIF_F_SG inheritable to MPLS.
7562 dev->mpls_features |= NETIF_F_SG;
7564 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev);
7565 ret = notifier_to_errno(ret);
7569 ret = netdev_register_kobject(dev);
7572 dev->reg_state = NETREG_REGISTERED;
7574 __netdev_update_features(dev);
7577 * Default initial state at registry is that the
7578 * device is present.
7581 set_bit(__LINK_STATE_PRESENT, &dev->state);
7583 linkwatch_init_dev(dev);
7585 dev_init_scheduler(dev);
7587 list_netdevice(dev);
7588 add_device_randomness(dev->dev_addr, dev->addr_len);
7590 /* If the device has permanent device address, driver should
7591 * set dev_addr and also addr_assign_type should be set to
7592 * NET_ADDR_PERM (default value).
7594 if (dev->addr_assign_type == NET_ADDR_PERM)
7595 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len);
7597 /* Notify protocols, that a new device appeared. */
7598 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
7599 ret = notifier_to_errno(ret);
7601 rollback_registered(dev);
7602 dev->reg_state = NETREG_UNREGISTERED;
7605 * Prevent userspace races by waiting until the network
7606 * device is fully setup before sending notifications.
7608 if (!dev->rtnl_link_ops ||
7609 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
7610 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
7616 if (dev->netdev_ops->ndo_uninit)
7617 dev->netdev_ops->ndo_uninit(dev);
7618 if (dev->priv_destructor)
7619 dev->priv_destructor(dev);
7622 EXPORT_SYMBOL(register_netdevice);
7625 * init_dummy_netdev - init a dummy network device for NAPI
7626 * @dev: device to init
7628 * This takes a network device structure and initialize the minimum
7629 * amount of fields so it can be used to schedule NAPI polls without
7630 * registering a full blown interface. This is to be used by drivers
7631 * that need to tie several hardware interfaces to a single NAPI
7632 * poll scheduler due to HW limitations.
7634 int init_dummy_netdev(struct net_device *dev)
7636 /* Clear everything. Note we don't initialize spinlocks
7637 * are they aren't supposed to be taken by any of the
7638 * NAPI code and this dummy netdev is supposed to be
7639 * only ever used for NAPI polls
7641 memset(dev, 0, sizeof(struct net_device));
7643 /* make sure we BUG if trying to hit standard
7644 * register/unregister code path
7646 dev->reg_state = NETREG_DUMMY;
7648 /* NAPI wants this */
7649 INIT_LIST_HEAD(&dev->napi_list);
7651 /* a dummy interface is started by default */
7652 set_bit(__LINK_STATE_PRESENT, &dev->state);
7653 set_bit(__LINK_STATE_START, &dev->state);
7655 /* Note : We dont allocate pcpu_refcnt for dummy devices,
7656 * because users of this 'device' dont need to change
7662 EXPORT_SYMBOL_GPL(init_dummy_netdev);
7666 * register_netdev - register a network device
7667 * @dev: device to register
7669 * Take a completed network device structure and add it to the kernel
7670 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
7671 * chain. 0 is returned on success. A negative errno code is returned
7672 * on a failure to set up the device, or if the name is a duplicate.
7674 * This is a wrapper around register_netdevice that takes the rtnl semaphore
7675 * and expands the device name if you passed a format string to
7678 int register_netdev(struct net_device *dev)
7683 err = register_netdevice(dev);
7687 EXPORT_SYMBOL(register_netdev);
7689 int netdev_refcnt_read(const struct net_device *dev)
7693 for_each_possible_cpu(i)
7694 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
7697 EXPORT_SYMBOL(netdev_refcnt_read);
7700 * netdev_wait_allrefs - wait until all references are gone.
7701 * @dev: target net_device
7703 * This is called when unregistering network devices.
7705 * Any protocol or device that holds a reference should register
7706 * for netdevice notification, and cleanup and put back the
7707 * reference if they receive an UNREGISTER event.
7708 * We can get stuck here if buggy protocols don't correctly
7711 static void netdev_wait_allrefs(struct net_device *dev)
7713 unsigned long rebroadcast_time, warning_time;
7716 linkwatch_forget_dev(dev);
7718 rebroadcast_time = warning_time = jiffies;
7719 refcnt = netdev_refcnt_read(dev);
7721 while (refcnt != 0) {
7722 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
7725 /* Rebroadcast unregister notification */
7726 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
7732 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
7733 if (test_bit(__LINK_STATE_LINKWATCH_PENDING,
7735 /* We must not have linkwatch events
7736 * pending on unregister. If this
7737 * happens, we simply run the queue
7738 * unscheduled, resulting in a noop
7741 linkwatch_run_queue();
7746 rebroadcast_time = jiffies;
7751 refcnt = netdev_refcnt_read(dev);
7753 if (time_after(jiffies, warning_time + 10 * HZ)) {
7754 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n",
7756 warning_time = jiffies;
7765 * register_netdevice(x1);
7766 * register_netdevice(x2);
7768 * unregister_netdevice(y1);
7769 * unregister_netdevice(y2);
7775 * We are invoked by rtnl_unlock().
7776 * This allows us to deal with problems:
7777 * 1) We can delete sysfs objects which invoke hotplug
7778 * without deadlocking with linkwatch via keventd.
7779 * 2) Since we run with the RTNL semaphore not held, we can sleep
7780 * safely in order to wait for the netdev refcnt to drop to zero.
7782 * We must not return until all unregister events added during
7783 * the interval the lock was held have been completed.
7785 void netdev_run_todo(void)
7787 struct list_head list;
7789 /* Snapshot list, allow later requests */
7790 list_replace_init(&net_todo_list, &list);
7795 /* Wait for rcu callbacks to finish before next phase */
7796 if (!list_empty(&list))
7799 while (!list_empty(&list)) {
7800 struct net_device *dev
7801 = list_first_entry(&list, struct net_device, todo_list);
7802 list_del(&dev->todo_list);
7805 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
7808 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) {
7809 pr_err("network todo '%s' but state %d\n",
7810 dev->name, dev->reg_state);
7815 dev->reg_state = NETREG_UNREGISTERED;
7817 netdev_wait_allrefs(dev);
7820 BUG_ON(netdev_refcnt_read(dev));
7821 BUG_ON(!list_empty(&dev->ptype_all));
7822 BUG_ON(!list_empty(&dev->ptype_specific));
7823 WARN_ON(rcu_access_pointer(dev->ip_ptr));
7824 WARN_ON(rcu_access_pointer(dev->ip6_ptr));
7825 WARN_ON(dev->dn_ptr);
7827 if (dev->priv_destructor)
7828 dev->priv_destructor(dev);
7829 if (dev->needs_free_netdev)
7832 /* Report a network device has been unregistered */
7834 dev_net(dev)->dev_unreg_count--;
7836 wake_up(&netdev_unregistering_wq);
7838 /* Free network device */
7839 kobject_put(&dev->dev.kobj);
7843 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has
7844 * all the same fields in the same order as net_device_stats, with only
7845 * the type differing, but rtnl_link_stats64 may have additional fields
7846 * at the end for newer counters.
7848 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
7849 const struct net_device_stats *netdev_stats)
7851 #if BITS_PER_LONG == 64
7852 BUILD_BUG_ON(sizeof(*stats64) < sizeof(*netdev_stats));
7853 memcpy(stats64, netdev_stats, sizeof(*netdev_stats));
7854 /* zero out counters that only exist in rtnl_link_stats64 */
7855 memset((char *)stats64 + sizeof(*netdev_stats), 0,
7856 sizeof(*stats64) - sizeof(*netdev_stats));
7858 size_t i, n = sizeof(*netdev_stats) / sizeof(unsigned long);
7859 const unsigned long *src = (const unsigned long *)netdev_stats;
7860 u64 *dst = (u64 *)stats64;
7862 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64));
7863 for (i = 0; i < n; i++)
7865 /* zero out counters that only exist in rtnl_link_stats64 */
7866 memset((char *)stats64 + n * sizeof(u64), 0,
7867 sizeof(*stats64) - n * sizeof(u64));
7870 EXPORT_SYMBOL(netdev_stats_to_stats64);
7873 * dev_get_stats - get network device statistics
7874 * @dev: device to get statistics from
7875 * @storage: place to store stats
7877 * Get network statistics from device. Return @storage.
7878 * The device driver may provide its own method by setting
7879 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats;
7880 * otherwise the internal statistics structure is used.
7882 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev,
7883 struct rtnl_link_stats64 *storage)
7885 const struct net_device_ops *ops = dev->netdev_ops;
7887 if (ops->ndo_get_stats64) {
7888 memset(storage, 0, sizeof(*storage));
7889 ops->ndo_get_stats64(dev, storage);
7890 } else if (ops->ndo_get_stats) {
7891 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev));
7893 netdev_stats_to_stats64(storage, &dev->stats);
7895 storage->rx_dropped += (unsigned long)atomic_long_read(&dev->rx_dropped);
7896 storage->tx_dropped += (unsigned long)atomic_long_read(&dev->tx_dropped);
7897 storage->rx_nohandler += (unsigned long)atomic_long_read(&dev->rx_nohandler);
7900 EXPORT_SYMBOL(dev_get_stats);
7902 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev)
7904 struct netdev_queue *queue = dev_ingress_queue(dev);
7906 #ifdef CONFIG_NET_CLS_ACT
7909 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
7912 netdev_init_one_queue(dev, queue, NULL);
7913 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc);
7914 queue->qdisc_sleeping = &noop_qdisc;
7915 rcu_assign_pointer(dev->ingress_queue, queue);
7920 static const struct ethtool_ops default_ethtool_ops;
7922 void netdev_set_default_ethtool_ops(struct net_device *dev,
7923 const struct ethtool_ops *ops)
7925 if (dev->ethtool_ops == &default_ethtool_ops)
7926 dev->ethtool_ops = ops;
7928 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops);
7930 void netdev_freemem(struct net_device *dev)
7932 char *addr = (char *)dev - dev->padded;
7938 * alloc_netdev_mqs - allocate network device
7939 * @sizeof_priv: size of private data to allocate space for
7940 * @name: device name format string
7941 * @name_assign_type: origin of device name
7942 * @setup: callback to initialize device
7943 * @txqs: the number of TX subqueues to allocate
7944 * @rxqs: the number of RX subqueues to allocate
7946 * Allocates a struct net_device with private data area for driver use
7947 * and performs basic initialization. Also allocates subqueue structs
7948 * for each queue on the device.
7950 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
7951 unsigned char name_assign_type,
7952 void (*setup)(struct net_device *),
7953 unsigned int txqs, unsigned int rxqs)
7955 struct net_device *dev;
7957 struct net_device *p;
7959 BUG_ON(strlen(name) >= sizeof(dev->name));
7962 pr_err("alloc_netdev: Unable to allocate device with zero queues\n");
7968 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n");
7973 alloc_size = sizeof(struct net_device);
7975 /* ensure 32-byte alignment of private area */
7976 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
7977 alloc_size += sizeof_priv;
7979 /* ensure 32-byte alignment of whole construct */
7980 alloc_size += NETDEV_ALIGN - 1;
7982 p = kvzalloc(alloc_size, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
7986 dev = PTR_ALIGN(p, NETDEV_ALIGN);
7987 dev->padded = (char *)dev - (char *)p;
7989 dev->pcpu_refcnt = alloc_percpu(int);
7990 if (!dev->pcpu_refcnt)
7993 if (dev_addr_init(dev))
7999 dev_net_set(dev, &init_net);
8001 dev->gso_max_size = GSO_MAX_SIZE;
8002 dev->gso_max_segs = GSO_MAX_SEGS;
8004 INIT_LIST_HEAD(&dev->napi_list);
8005 INIT_LIST_HEAD(&dev->unreg_list);
8006 INIT_LIST_HEAD(&dev->close_list);
8007 INIT_LIST_HEAD(&dev->link_watch_list);
8008 INIT_LIST_HEAD(&dev->adj_list.upper);
8009 INIT_LIST_HEAD(&dev->adj_list.lower);
8010 INIT_LIST_HEAD(&dev->ptype_all);
8011 INIT_LIST_HEAD(&dev->ptype_specific);
8012 #ifdef CONFIG_NET_SCHED
8013 hash_init(dev->qdisc_hash);
8015 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM;
8018 if (!dev->tx_queue_len) {
8019 dev->priv_flags |= IFF_NO_QUEUE;
8020 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN;
8023 dev->num_tx_queues = txqs;
8024 dev->real_num_tx_queues = txqs;
8025 if (netif_alloc_netdev_queues(dev))
8029 dev->num_rx_queues = rxqs;
8030 dev->real_num_rx_queues = rxqs;
8031 if (netif_alloc_rx_queues(dev))
8035 strcpy(dev->name, name);
8036 dev->name_assign_type = name_assign_type;
8037 dev->group = INIT_NETDEV_GROUP;
8038 if (!dev->ethtool_ops)
8039 dev->ethtool_ops = &default_ethtool_ops;
8041 nf_hook_ingress_init(dev);
8050 free_percpu(dev->pcpu_refcnt);
8052 netdev_freemem(dev);
8055 EXPORT_SYMBOL(alloc_netdev_mqs);
8058 * free_netdev - free network device
8061 * This function does the last stage of destroying an allocated device
8062 * interface. The reference to the device object is released. If this
8063 * is the last reference then it will be freed.Must be called in process
8066 void free_netdev(struct net_device *dev)
8068 struct napi_struct *p, *n;
8069 struct bpf_prog *prog;
8072 netif_free_tx_queues(dev);
8077 kfree(rcu_dereference_protected(dev->ingress_queue, 1));
8079 /* Flush device addresses */
8080 dev_addr_flush(dev);
8082 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
8085 free_percpu(dev->pcpu_refcnt);
8086 dev->pcpu_refcnt = NULL;
8088 prog = rcu_dereference_protected(dev->xdp_prog, 1);
8091 static_key_slow_dec(&generic_xdp_needed);
8094 /* Compatibility with error handling in drivers */
8095 if (dev->reg_state == NETREG_UNINITIALIZED) {
8096 netdev_freemem(dev);
8100 BUG_ON(dev->reg_state != NETREG_UNREGISTERED);
8101 dev->reg_state = NETREG_RELEASED;
8103 /* will free via device release */
8104 put_device(&dev->dev);
8106 EXPORT_SYMBOL(free_netdev);
8109 * synchronize_net - Synchronize with packet receive processing
8111 * Wait for packets currently being received to be done.
8112 * Does not block later packets from starting.
8114 void synchronize_net(void)
8117 if (rtnl_is_locked())
8118 synchronize_rcu_expedited();
8122 EXPORT_SYMBOL(synchronize_net);
8125 * unregister_netdevice_queue - remove device from the kernel
8129 * This function shuts down a device interface and removes it
8130 * from the kernel tables.
8131 * If head not NULL, device is queued to be unregistered later.
8133 * Callers must hold the rtnl semaphore. You may want
8134 * unregister_netdev() instead of this.
8137 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head)
8142 list_move_tail(&dev->unreg_list, head);
8144 rollback_registered(dev);
8145 /* Finish processing unregister after unlock */
8149 EXPORT_SYMBOL(unregister_netdevice_queue);
8152 * unregister_netdevice_many - unregister many devices
8153 * @head: list of devices
8155 * Note: As most callers use a stack allocated list_head,
8156 * we force a list_del() to make sure stack wont be corrupted later.
8158 void unregister_netdevice_many(struct list_head *head)
8160 struct net_device *dev;
8162 if (!list_empty(head)) {
8163 rollback_registered_many(head);
8164 list_for_each_entry(dev, head, unreg_list)
8169 EXPORT_SYMBOL(unregister_netdevice_many);
8172 * unregister_netdev - remove device from the kernel
8175 * This function shuts down a device interface and removes it
8176 * from the kernel tables.
8178 * This is just a wrapper for unregister_netdevice that takes
8179 * the rtnl semaphore. In general you want to use this and not
8180 * unregister_netdevice.
8182 void unregister_netdev(struct net_device *dev)
8185 unregister_netdevice(dev);
8188 EXPORT_SYMBOL(unregister_netdev);
8191 * dev_change_net_namespace - move device to different nethost namespace
8193 * @net: network namespace
8194 * @pat: If not NULL name pattern to try if the current device name
8195 * is already taken in the destination network namespace.
8197 * This function shuts down a device interface and moves it
8198 * to a new network namespace. On success 0 is returned, on
8199 * a failure a netagive errno code is returned.
8201 * Callers must hold the rtnl semaphore.
8204 int dev_change_net_namespace(struct net_device *dev, struct net *net, const char *pat)
8210 /* Don't allow namespace local devices to be moved. */
8212 if (dev->features & NETIF_F_NETNS_LOCAL)
8215 /* Ensure the device has been registrered */
8216 if (dev->reg_state != NETREG_REGISTERED)
8219 /* Get out if there is nothing todo */
8221 if (net_eq(dev_net(dev), net))
8224 /* Pick the destination device name, and ensure
8225 * we can use it in the destination network namespace.
8228 if (__dev_get_by_name(net, dev->name)) {
8229 /* We get here if we can't use the current device name */
8232 if (dev_get_valid_name(net, dev, pat) < 0)
8237 * And now a mini version of register_netdevice unregister_netdevice.
8240 /* If device is running close it first. */
8243 /* And unlink it from device chain */
8245 unlist_netdevice(dev);
8249 /* Shutdown queueing discipline. */
8252 /* Notify protocols, that we are about to destroy
8253 * this device. They should clean all the things.
8255 * Note that dev->reg_state stays at NETREG_REGISTERED.
8256 * This is wanted because this way 8021q and macvlan know
8257 * the device is just moving and can keep their slaves up.
8259 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
8261 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
8262 rtmsg_ifinfo(RTM_DELLINK, dev, ~0U, GFP_KERNEL);
8265 * Flush the unicast and multicast chains
8270 /* Send a netdev-removed uevent to the old namespace */
8271 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE);
8272 netdev_adjacent_del_links(dev);
8274 /* Actually switch the network namespace */
8275 dev_net_set(dev, net);
8277 /* If there is an ifindex conflict assign a new one */
8278 if (__dev_get_by_index(net, dev->ifindex))
8279 dev->ifindex = dev_new_index(net);
8281 /* Send a netdev-add uevent to the new namespace */
8282 kobject_uevent(&dev->dev.kobj, KOBJ_ADD);
8283 netdev_adjacent_add_links(dev);
8285 /* Fixup kobjects */
8286 err = device_rename(&dev->dev, dev->name);
8289 /* Add the device back in the hashes */
8290 list_netdevice(dev);
8292 /* Notify protocols, that a new device appeared. */
8293 call_netdevice_notifiers(NETDEV_REGISTER, dev);
8296 * Prevent userspace races by waiting until the network
8297 * device is fully setup before sending notifications.
8299 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
8306 EXPORT_SYMBOL_GPL(dev_change_net_namespace);
8308 static int dev_cpu_dead(unsigned int oldcpu)
8310 struct sk_buff **list_skb;
8311 struct sk_buff *skb;
8313 struct softnet_data *sd, *oldsd, *remsd = NULL;
8315 local_irq_disable();
8316 cpu = smp_processor_id();
8317 sd = &per_cpu(softnet_data, cpu);
8318 oldsd = &per_cpu(softnet_data, oldcpu);
8320 /* Find end of our completion_queue. */
8321 list_skb = &sd->completion_queue;
8323 list_skb = &(*list_skb)->next;
8324 /* Append completion queue from offline CPU. */
8325 *list_skb = oldsd->completion_queue;
8326 oldsd->completion_queue = NULL;
8328 /* Append output queue from offline CPU. */
8329 if (oldsd->output_queue) {
8330 *sd->output_queue_tailp = oldsd->output_queue;
8331 sd->output_queue_tailp = oldsd->output_queue_tailp;
8332 oldsd->output_queue = NULL;
8333 oldsd->output_queue_tailp = &oldsd->output_queue;
8335 /* Append NAPI poll list from offline CPU, with one exception :
8336 * process_backlog() must be called by cpu owning percpu backlog.
8337 * We properly handle process_queue & input_pkt_queue later.
8339 while (!list_empty(&oldsd->poll_list)) {
8340 struct napi_struct *napi = list_first_entry(&oldsd->poll_list,
8344 list_del_init(&napi->poll_list);
8345 if (napi->poll == process_backlog)
8348 ____napi_schedule(sd, napi);
8351 raise_softirq_irqoff(NET_TX_SOFTIRQ);
8355 remsd = oldsd->rps_ipi_list;
8356 oldsd->rps_ipi_list = NULL;
8358 /* send out pending IPI's on offline CPU */
8359 net_rps_send_ipi(remsd);
8361 /* Process offline CPU's input_pkt_queue */
8362 while ((skb = __skb_dequeue(&oldsd->process_queue))) {
8364 input_queue_head_incr(oldsd);
8366 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) {
8368 input_queue_head_incr(oldsd);
8375 * netdev_increment_features - increment feature set by one
8376 * @all: current feature set
8377 * @one: new feature set
8378 * @mask: mask feature set
8380 * Computes a new feature set after adding a device with feature set
8381 * @one to the master device with current feature set @all. Will not
8382 * enable anything that is off in @mask. Returns the new feature set.
8384 netdev_features_t netdev_increment_features(netdev_features_t all,
8385 netdev_features_t one, netdev_features_t mask)
8387 if (mask & NETIF_F_HW_CSUM)
8388 mask |= NETIF_F_CSUM_MASK;
8389 mask |= NETIF_F_VLAN_CHALLENGED;
8391 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask;
8392 all &= one | ~NETIF_F_ALL_FOR_ALL;
8394 /* If one device supports hw checksumming, set for all. */
8395 if (all & NETIF_F_HW_CSUM)
8396 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM);
8400 EXPORT_SYMBOL(netdev_increment_features);
8402 static struct hlist_head * __net_init netdev_create_hash(void)
8405 struct hlist_head *hash;
8407 hash = kmalloc(sizeof(*hash) * NETDEV_HASHENTRIES, GFP_KERNEL);
8409 for (i = 0; i < NETDEV_HASHENTRIES; i++)
8410 INIT_HLIST_HEAD(&hash[i]);
8415 /* Initialize per network namespace state */
8416 static int __net_init netdev_init(struct net *net)
8418 if (net != &init_net)
8419 INIT_LIST_HEAD(&net->dev_base_head);
8421 net->dev_name_head = netdev_create_hash();
8422 if (net->dev_name_head == NULL)
8425 net->dev_index_head = netdev_create_hash();
8426 if (net->dev_index_head == NULL)
8432 kfree(net->dev_name_head);
8438 * netdev_drivername - network driver for the device
8439 * @dev: network device
8441 * Determine network driver for device.
8443 const char *netdev_drivername(const struct net_device *dev)
8445 const struct device_driver *driver;
8446 const struct device *parent;
8447 const char *empty = "";
8449 parent = dev->dev.parent;
8453 driver = parent->driver;
8454 if (driver && driver->name)
8455 return driver->name;
8459 static void __netdev_printk(const char *level, const struct net_device *dev,
8460 struct va_format *vaf)
8462 if (dev && dev->dev.parent) {
8463 dev_printk_emit(level[1] - '0',
8466 dev_driver_string(dev->dev.parent),
8467 dev_name(dev->dev.parent),
8468 netdev_name(dev), netdev_reg_state(dev),
8471 printk("%s%s%s: %pV",
8472 level, netdev_name(dev), netdev_reg_state(dev), vaf);
8474 printk("%s(NULL net_device): %pV", level, vaf);
8478 void netdev_printk(const char *level, const struct net_device *dev,
8479 const char *format, ...)
8481 struct va_format vaf;
8484 va_start(args, format);
8489 __netdev_printk(level, dev, &vaf);
8493 EXPORT_SYMBOL(netdev_printk);
8495 #define define_netdev_printk_level(func, level) \
8496 void func(const struct net_device *dev, const char *fmt, ...) \
8498 struct va_format vaf; \
8501 va_start(args, fmt); \
8506 __netdev_printk(level, dev, &vaf); \
8510 EXPORT_SYMBOL(func);
8512 define_netdev_printk_level(netdev_emerg, KERN_EMERG);
8513 define_netdev_printk_level(netdev_alert, KERN_ALERT);
8514 define_netdev_printk_level(netdev_crit, KERN_CRIT);
8515 define_netdev_printk_level(netdev_err, KERN_ERR);
8516 define_netdev_printk_level(netdev_warn, KERN_WARNING);
8517 define_netdev_printk_level(netdev_notice, KERN_NOTICE);
8518 define_netdev_printk_level(netdev_info, KERN_INFO);
8520 static void __net_exit netdev_exit(struct net *net)
8522 kfree(net->dev_name_head);
8523 kfree(net->dev_index_head);
8526 static struct pernet_operations __net_initdata netdev_net_ops = {
8527 .init = netdev_init,
8528 .exit = netdev_exit,
8531 static void __net_exit default_device_exit(struct net *net)
8533 struct net_device *dev, *aux;
8535 * Push all migratable network devices back to the
8536 * initial network namespace
8539 for_each_netdev_safe(net, dev, aux) {
8541 char fb_name[IFNAMSIZ];
8543 /* Ignore unmoveable devices (i.e. loopback) */
8544 if (dev->features & NETIF_F_NETNS_LOCAL)
8547 /* Leave virtual devices for the generic cleanup */
8548 if (dev->rtnl_link_ops)
8551 /* Push remaining network devices to init_net */
8552 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex);
8553 err = dev_change_net_namespace(dev, &init_net, fb_name);
8555 pr_emerg("%s: failed to move %s to init_net: %d\n",
8556 __func__, dev->name, err);
8563 static void __net_exit rtnl_lock_unregistering(struct list_head *net_list)
8565 /* Return with the rtnl_lock held when there are no network
8566 * devices unregistering in any network namespace in net_list.
8570 DEFINE_WAIT_FUNC(wait, woken_wake_function);
8572 add_wait_queue(&netdev_unregistering_wq, &wait);
8574 unregistering = false;
8576 list_for_each_entry(net, net_list, exit_list) {
8577 if (net->dev_unreg_count > 0) {
8578 unregistering = true;
8586 wait_woken(&wait, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
8588 remove_wait_queue(&netdev_unregistering_wq, &wait);
8591 static void __net_exit default_device_exit_batch(struct list_head *net_list)
8593 /* At exit all network devices most be removed from a network
8594 * namespace. Do this in the reverse order of registration.
8595 * Do this across as many network namespaces as possible to
8596 * improve batching efficiency.
8598 struct net_device *dev;
8600 LIST_HEAD(dev_kill_list);
8602 /* To prevent network device cleanup code from dereferencing
8603 * loopback devices or network devices that have been freed
8604 * wait here for all pending unregistrations to complete,
8605 * before unregistring the loopback device and allowing the
8606 * network namespace be freed.
8608 * The netdev todo list containing all network devices
8609 * unregistrations that happen in default_device_exit_batch
8610 * will run in the rtnl_unlock() at the end of
8611 * default_device_exit_batch.
8613 rtnl_lock_unregistering(net_list);
8614 list_for_each_entry(net, net_list, exit_list) {
8615 for_each_netdev_reverse(net, dev) {
8616 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink)
8617 dev->rtnl_link_ops->dellink(dev, &dev_kill_list);
8619 unregister_netdevice_queue(dev, &dev_kill_list);
8622 unregister_netdevice_many(&dev_kill_list);
8626 static struct pernet_operations __net_initdata default_device_ops = {
8627 .exit = default_device_exit,
8628 .exit_batch = default_device_exit_batch,
8632 * Initialize the DEV module. At boot time this walks the device list and
8633 * unhooks any devices that fail to initialise (normally hardware not
8634 * present) and leaves us with a valid list of present and active devices.
8639 * This is called single threaded during boot, so no need
8640 * to take the rtnl semaphore.
8642 static int __init net_dev_init(void)
8644 int i, rc = -ENOMEM;
8646 BUG_ON(!dev_boot_phase);
8648 if (dev_proc_init())
8651 if (netdev_kobject_init())
8654 INIT_LIST_HEAD(&ptype_all);
8655 for (i = 0; i < PTYPE_HASH_SIZE; i++)
8656 INIT_LIST_HEAD(&ptype_base[i]);
8658 INIT_LIST_HEAD(&offload_base);
8660 if (register_pernet_subsys(&netdev_net_ops))
8664 * Initialise the packet receive queues.
8667 for_each_possible_cpu(i) {
8668 struct work_struct *flush = per_cpu_ptr(&flush_works, i);
8669 struct softnet_data *sd = &per_cpu(softnet_data, i);
8671 INIT_WORK(flush, flush_backlog);
8673 skb_queue_head_init(&sd->input_pkt_queue);
8674 skb_queue_head_init(&sd->process_queue);
8675 INIT_LIST_HEAD(&sd->poll_list);
8676 sd->output_queue_tailp = &sd->output_queue;
8678 sd->csd.func = rps_trigger_softirq;
8683 sd->backlog.poll = process_backlog;
8684 sd->backlog.weight = weight_p;
8689 /* The loopback device is special if any other network devices
8690 * is present in a network namespace the loopback device must
8691 * be present. Since we now dynamically allocate and free the
8692 * loopback device ensure this invariant is maintained by
8693 * keeping the loopback device as the first device on the
8694 * list of network devices. Ensuring the loopback devices
8695 * is the first device that appears and the last network device
8698 if (register_pernet_device(&loopback_net_ops))
8701 if (register_pernet_device(&default_device_ops))
8704 open_softirq(NET_TX_SOFTIRQ, net_tx_action);
8705 open_softirq(NET_RX_SOFTIRQ, net_rx_action);
8707 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead",
8708 NULL, dev_cpu_dead);
8715 subsys_initcall(net_dev_init);