2 * NET3 Protocol independent device support routines.
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
9 * Derived from the non IP parts of dev.c 1.0.19
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
15 * Florian la Roche <rzsfl@rz.uni-sb.de>
16 * Alan Cox <gw4pts@gw4pts.ampr.org>
17 * David Hinds <dahinds@users.sourceforge.net>
18 * Alexey Kuznetsov <kuznet@ms2.inr.ac.ru>
19 * Adam Sulmicki <adam@cfar.umd.edu>
20 * Pekka Riikonen <priikone@poesidon.pspt.fi>
23 * D.J. Barrow : Fixed bug where dev->refcnt gets set
24 * to 2 if register_netdev gets called
25 * before net_dev_init & also removed a
26 * few lines of code in the process.
27 * Alan Cox : device private ioctl copies fields back.
28 * Alan Cox : Transmit queue code does relevant
29 * stunts to keep the queue safe.
30 * Alan Cox : Fixed double lock.
31 * Alan Cox : Fixed promisc NULL pointer trap
32 * ???????? : Support the full private ioctl range
33 * Alan Cox : Moved ioctl permission check into
35 * Tim Kordas : SIOCADDMULTI/SIOCDELMULTI
36 * Alan Cox : 100 backlog just doesn't cut it when
37 * you start doing multicast video 8)
38 * Alan Cox : Rewrote net_bh and list manager.
39 * Alan Cox : Fix ETH_P_ALL echoback lengths.
40 * Alan Cox : Took out transmit every packet pass
41 * Saved a few bytes in the ioctl handler
42 * Alan Cox : Network driver sets packet type before
43 * calling netif_rx. Saves a function
45 * Alan Cox : Hashed net_bh()
46 * Richard Kooijman: Timestamp fixes.
47 * Alan Cox : Wrong field in SIOCGIFDSTADDR
48 * Alan Cox : Device lock protection.
49 * Alan Cox : Fixed nasty side effect of device close
51 * Rudi Cilibrasi : Pass the right thing to
53 * Dave Miller : 32bit quantity for the device lock to
54 * make it work out on a Sparc.
55 * Bjorn Ekwall : Added KERNELD hack.
56 * Alan Cox : Cleaned up the backlog initialise.
57 * Craig Metz : SIOCGIFCONF fix if space for under
59 * Thomas Bogendoerfer : Return ENODEV for dev_open, if there
60 * is no device open function.
61 * Andi Kleen : Fix error reporting for SIOCGIFCONF
62 * Michael Chastain : Fix signed/unsigned for SIOCGIFCONF
63 * Cyrus Durgin : Cleaned for KMOD
64 * Adam Sulmicki : Bug Fix : Network Device Unload
65 * A network device unload needs to purge
67 * Paul Rusty Russell : SIOCSIFNAME
68 * Pekka Riikonen : Netdev boot-time settings code
69 * Andrew Morton : Make unregister_netdevice wait
70 * indefinitely on dev->refcnt
71 * J Hadi Salim : - Backlog queue sampling
72 * - netif_rx() feedback
75 #include <linux/uaccess.h>
76 #include <linux/bitops.h>
77 #include <linux/capability.h>
78 #include <linux/cpu.h>
79 #include <linux/types.h>
80 #include <linux/kernel.h>
81 #include <linux/hash.h>
82 #include <linux/slab.h>
83 #include <linux/sched.h>
84 #include <linux/sched/mm.h>
85 #include <linux/mutex.h>
86 #include <linux/string.h>
88 #include <linux/socket.h>
89 #include <linux/sockios.h>
90 #include <linux/errno.h>
91 #include <linux/interrupt.h>
92 #include <linux/if_ether.h>
93 #include <linux/netdevice.h>
94 #include <linux/etherdevice.h>
95 #include <linux/ethtool.h>
96 #include <linux/notifier.h>
97 #include <linux/skbuff.h>
98 #include <linux/bpf.h>
99 #include <linux/bpf_trace.h>
100 #include <net/net_namespace.h>
101 #include <net/sock.h>
102 #include <net/busy_poll.h>
103 #include <linux/rtnetlink.h>
104 #include <linux/stat.h>
106 #include <net/dst_metadata.h>
107 #include <net/pkt_sched.h>
108 #include <net/pkt_cls.h>
109 #include <net/checksum.h>
110 #include <net/xfrm.h>
111 #include <linux/highmem.h>
112 #include <linux/init.h>
113 #include <linux/module.h>
114 #include <linux/netpoll.h>
115 #include <linux/rcupdate.h>
116 #include <linux/delay.h>
117 #include <net/iw_handler.h>
118 #include <asm/current.h>
119 #include <linux/audit.h>
120 #include <linux/dmaengine.h>
121 #include <linux/err.h>
122 #include <linux/ctype.h>
123 #include <linux/if_arp.h>
124 #include <linux/if_vlan.h>
125 #include <linux/ip.h>
127 #include <net/mpls.h>
128 #include <linux/ipv6.h>
129 #include <linux/in.h>
130 #include <linux/jhash.h>
131 #include <linux/random.h>
132 #include <trace/events/napi.h>
133 #include <trace/events/net.h>
134 #include <trace/events/skb.h>
135 #include <linux/pci.h>
136 #include <linux/inetdevice.h>
137 #include <linux/cpu_rmap.h>
138 #include <linux/static_key.h>
139 #include <linux/hashtable.h>
140 #include <linux/vmalloc.h>
141 #include <linux/if_macvlan.h>
142 #include <linux/errqueue.h>
143 #include <linux/hrtimer.h>
144 #include <linux/netfilter_ingress.h>
145 #include <linux/crash_dump.h>
146 #include <linux/sctp.h>
147 #include <net/udp_tunnel.h>
149 #include "net-sysfs.h"
151 /* Instead of increasing this, you should create a hash table. */
152 #define MAX_GRO_SKBS 8
154 /* This should be increased if a protocol with a bigger head is added. */
155 #define GRO_MAX_HEAD (MAX_HEADER + 128)
157 static DEFINE_SPINLOCK(ptype_lock);
158 static DEFINE_SPINLOCK(offload_lock);
159 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly;
160 struct list_head ptype_all __read_mostly; /* Taps */
161 static struct list_head offload_base __read_mostly;
163 static int netif_rx_internal(struct sk_buff *skb);
164 static int call_netdevice_notifiers_info(unsigned long val,
165 struct net_device *dev,
166 struct netdev_notifier_info *info);
167 static struct napi_struct *napi_by_id(unsigned int napi_id);
170 * The @dev_base_head list is protected by @dev_base_lock and the rtnl
173 * Pure readers hold dev_base_lock for reading, or rcu_read_lock()
175 * Writers must hold the rtnl semaphore while they loop through the
176 * dev_base_head list, and hold dev_base_lock for writing when they do the
177 * actual updates. This allows pure readers to access the list even
178 * while a writer is preparing to update it.
180 * To put it another way, dev_base_lock is held for writing only to
181 * protect against pure readers; the rtnl semaphore provides the
182 * protection against other writers.
184 * See, for example usages, register_netdevice() and
185 * unregister_netdevice(), which must be called with the rtnl
188 DEFINE_RWLOCK(dev_base_lock);
189 EXPORT_SYMBOL(dev_base_lock);
191 /* protects napi_hash addition/deletion and napi_gen_id */
192 static DEFINE_SPINLOCK(napi_hash_lock);
194 static unsigned int napi_gen_id = NR_CPUS;
195 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8);
197 static seqcount_t devnet_rename_seq;
199 static inline void dev_base_seq_inc(struct net *net)
201 while (++net->dev_base_seq == 0)
205 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name)
207 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ));
209 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)];
212 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex)
214 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)];
217 static inline void rps_lock(struct softnet_data *sd)
220 spin_lock(&sd->input_pkt_queue.lock);
224 static inline void rps_unlock(struct softnet_data *sd)
227 spin_unlock(&sd->input_pkt_queue.lock);
231 /* Device list insertion */
232 static void list_netdevice(struct net_device *dev)
234 struct net *net = dev_net(dev);
238 write_lock_bh(&dev_base_lock);
239 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head);
240 hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
241 hlist_add_head_rcu(&dev->index_hlist,
242 dev_index_hash(net, dev->ifindex));
243 write_unlock_bh(&dev_base_lock);
245 dev_base_seq_inc(net);
248 /* Device list removal
249 * caller must respect a RCU grace period before freeing/reusing dev
251 static void unlist_netdevice(struct net_device *dev)
255 /* Unlink dev from the device chain */
256 write_lock_bh(&dev_base_lock);
257 list_del_rcu(&dev->dev_list);
258 hlist_del_rcu(&dev->name_hlist);
259 hlist_del_rcu(&dev->index_hlist);
260 write_unlock_bh(&dev_base_lock);
262 dev_base_seq_inc(dev_net(dev));
269 static RAW_NOTIFIER_HEAD(netdev_chain);
272 * Device drivers call our routines to queue packets here. We empty the
273 * queue in the local softnet handler.
276 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data);
277 EXPORT_PER_CPU_SYMBOL(softnet_data);
279 #ifdef CONFIG_LOCKDEP
281 * register_netdevice() inits txq->_xmit_lock and sets lockdep class
282 * according to dev->type
284 static const unsigned short netdev_lock_type[] = {
285 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25,
286 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET,
287 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM,
288 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP,
289 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD,
290 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25,
291 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP,
292 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD,
293 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI,
294 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE,
295 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET,
296 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL,
297 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM,
298 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE,
299 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE};
301 static const char *const netdev_lock_name[] = {
302 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25",
303 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET",
304 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM",
305 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP",
306 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD",
307 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25",
308 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP",
309 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD",
310 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI",
311 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE",
312 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET",
313 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL",
314 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM",
315 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE",
316 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"};
318 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)];
319 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)];
321 static inline unsigned short netdev_lock_pos(unsigned short dev_type)
325 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++)
326 if (netdev_lock_type[i] == dev_type)
328 /* the last key is used by default */
329 return ARRAY_SIZE(netdev_lock_type) - 1;
332 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
333 unsigned short dev_type)
337 i = netdev_lock_pos(dev_type);
338 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i],
339 netdev_lock_name[i]);
342 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
346 i = netdev_lock_pos(dev->type);
347 lockdep_set_class_and_name(&dev->addr_list_lock,
348 &netdev_addr_lock_key[i],
349 netdev_lock_name[i]);
352 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
353 unsigned short dev_type)
356 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
361 /*******************************************************************************
363 * Protocol management and registration routines
365 *******************************************************************************/
369 * Add a protocol ID to the list. Now that the input handler is
370 * smarter we can dispense with all the messy stuff that used to be
373 * BEWARE!!! Protocol handlers, mangling input packets,
374 * MUST BE last in hash buckets and checking protocol handlers
375 * MUST start from promiscuous ptype_all chain in net_bh.
376 * It is true now, do not change it.
377 * Explanation follows: if protocol handler, mangling packet, will
378 * be the first on list, it is not able to sense, that packet
379 * is cloned and should be copied-on-write, so that it will
380 * change it and subsequent readers will get broken packet.
384 static inline struct list_head *ptype_head(const struct packet_type *pt)
386 if (pt->type == htons(ETH_P_ALL))
387 return pt->dev ? &pt->dev->ptype_all : &ptype_all;
389 return pt->dev ? &pt->dev->ptype_specific :
390 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK];
394 * dev_add_pack - add packet handler
395 * @pt: packet type declaration
397 * Add a protocol handler to the networking stack. The passed &packet_type
398 * is linked into kernel lists and may not be freed until it has been
399 * removed from the kernel lists.
401 * This call does not sleep therefore it can not
402 * guarantee all CPU's that are in middle of receiving packets
403 * will see the new packet type (until the next received packet).
406 void dev_add_pack(struct packet_type *pt)
408 struct list_head *head = ptype_head(pt);
410 spin_lock(&ptype_lock);
411 list_add_rcu(&pt->list, head);
412 spin_unlock(&ptype_lock);
414 EXPORT_SYMBOL(dev_add_pack);
417 * __dev_remove_pack - remove packet handler
418 * @pt: packet type declaration
420 * Remove a protocol handler that was previously added to the kernel
421 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
422 * from the kernel lists and can be freed or reused once this function
425 * The packet type might still be in use by receivers
426 * and must not be freed until after all the CPU's have gone
427 * through a quiescent state.
429 void __dev_remove_pack(struct packet_type *pt)
431 struct list_head *head = ptype_head(pt);
432 struct packet_type *pt1;
434 spin_lock(&ptype_lock);
436 list_for_each_entry(pt1, head, list) {
438 list_del_rcu(&pt->list);
443 pr_warn("dev_remove_pack: %p not found\n", pt);
445 spin_unlock(&ptype_lock);
447 EXPORT_SYMBOL(__dev_remove_pack);
450 * dev_remove_pack - remove packet handler
451 * @pt: packet type declaration
453 * Remove a protocol handler that was previously added to the kernel
454 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
455 * from the kernel lists and can be freed or reused once this function
458 * This call sleeps to guarantee that no CPU is looking at the packet
461 void dev_remove_pack(struct packet_type *pt)
463 __dev_remove_pack(pt);
467 EXPORT_SYMBOL(dev_remove_pack);
471 * dev_add_offload - register offload handlers
472 * @po: protocol offload declaration
474 * Add protocol offload handlers to the networking stack. The passed
475 * &proto_offload is linked into kernel lists and may not be freed until
476 * it has been removed from the kernel lists.
478 * This call does not sleep therefore it can not
479 * guarantee all CPU's that are in middle of receiving packets
480 * will see the new offload handlers (until the next received packet).
482 void dev_add_offload(struct packet_offload *po)
484 struct packet_offload *elem;
486 spin_lock(&offload_lock);
487 list_for_each_entry(elem, &offload_base, list) {
488 if (po->priority < elem->priority)
491 list_add_rcu(&po->list, elem->list.prev);
492 spin_unlock(&offload_lock);
494 EXPORT_SYMBOL(dev_add_offload);
497 * __dev_remove_offload - remove offload handler
498 * @po: packet offload declaration
500 * Remove a protocol offload handler that was previously added to the
501 * kernel offload handlers by dev_add_offload(). The passed &offload_type
502 * is removed from the kernel lists and can be freed or reused once this
505 * The packet type might still be in use by receivers
506 * and must not be freed until after all the CPU's have gone
507 * through a quiescent state.
509 static void __dev_remove_offload(struct packet_offload *po)
511 struct list_head *head = &offload_base;
512 struct packet_offload *po1;
514 spin_lock(&offload_lock);
516 list_for_each_entry(po1, head, list) {
518 list_del_rcu(&po->list);
523 pr_warn("dev_remove_offload: %p not found\n", po);
525 spin_unlock(&offload_lock);
529 * dev_remove_offload - remove packet offload handler
530 * @po: packet offload declaration
532 * Remove a packet offload handler that was previously added to the kernel
533 * offload handlers by dev_add_offload(). The passed &offload_type is
534 * removed from the kernel lists and can be freed or reused once this
537 * This call sleeps to guarantee that no CPU is looking at the packet
540 void dev_remove_offload(struct packet_offload *po)
542 __dev_remove_offload(po);
546 EXPORT_SYMBOL(dev_remove_offload);
548 /******************************************************************************
550 * Device Boot-time Settings Routines
552 ******************************************************************************/
554 /* Boot time configuration table */
555 static struct netdev_boot_setup dev_boot_setup[NETDEV_BOOT_SETUP_MAX];
558 * netdev_boot_setup_add - add new setup entry
559 * @name: name of the device
560 * @map: configured settings for the device
562 * Adds new setup entry to the dev_boot_setup list. The function
563 * returns 0 on error and 1 on success. This is a generic routine to
566 static int netdev_boot_setup_add(char *name, struct ifmap *map)
568 struct netdev_boot_setup *s;
572 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
573 if (s[i].name[0] == '\0' || s[i].name[0] == ' ') {
574 memset(s[i].name, 0, sizeof(s[i].name));
575 strlcpy(s[i].name, name, IFNAMSIZ);
576 memcpy(&s[i].map, map, sizeof(s[i].map));
581 return i >= NETDEV_BOOT_SETUP_MAX ? 0 : 1;
585 * netdev_boot_setup_check - check boot time settings
586 * @dev: the netdevice
588 * Check boot time settings for the device.
589 * The found settings are set for the device to be used
590 * later in the device probing.
591 * Returns 0 if no settings found, 1 if they are.
593 int netdev_boot_setup_check(struct net_device *dev)
595 struct netdev_boot_setup *s = dev_boot_setup;
598 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
599 if (s[i].name[0] != '\0' && s[i].name[0] != ' ' &&
600 !strcmp(dev->name, s[i].name)) {
601 dev->irq = s[i].map.irq;
602 dev->base_addr = s[i].map.base_addr;
603 dev->mem_start = s[i].map.mem_start;
604 dev->mem_end = s[i].map.mem_end;
610 EXPORT_SYMBOL(netdev_boot_setup_check);
614 * netdev_boot_base - get address from boot time settings
615 * @prefix: prefix for network device
616 * @unit: id for network device
618 * Check boot time settings for the base address of device.
619 * The found settings are set for the device to be used
620 * later in the device probing.
621 * Returns 0 if no settings found.
623 unsigned long netdev_boot_base(const char *prefix, int unit)
625 const struct netdev_boot_setup *s = dev_boot_setup;
629 sprintf(name, "%s%d", prefix, unit);
632 * If device already registered then return base of 1
633 * to indicate not to probe for this interface
635 if (__dev_get_by_name(&init_net, name))
638 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++)
639 if (!strcmp(name, s[i].name))
640 return s[i].map.base_addr;
645 * Saves at boot time configured settings for any netdevice.
647 int __init netdev_boot_setup(char *str)
652 str = get_options(str, ARRAY_SIZE(ints), ints);
657 memset(&map, 0, sizeof(map));
661 map.base_addr = ints[2];
663 map.mem_start = ints[3];
665 map.mem_end = ints[4];
667 /* Add new entry to the list */
668 return netdev_boot_setup_add(str, &map);
671 __setup("netdev=", netdev_boot_setup);
673 /*******************************************************************************
675 * Device Interface Subroutines
677 *******************************************************************************/
680 * dev_get_iflink - get 'iflink' value of a interface
681 * @dev: targeted interface
683 * Indicates the ifindex the interface is linked to.
684 * Physical interfaces have the same 'ifindex' and 'iflink' values.
687 int dev_get_iflink(const struct net_device *dev)
689 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink)
690 return dev->netdev_ops->ndo_get_iflink(dev);
694 EXPORT_SYMBOL(dev_get_iflink);
697 * dev_fill_metadata_dst - Retrieve tunnel egress information.
698 * @dev: targeted interface
701 * For better visibility of tunnel traffic OVS needs to retrieve
702 * egress tunnel information for a packet. Following API allows
703 * user to get this info.
705 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb)
707 struct ip_tunnel_info *info;
709 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst)
712 info = skb_tunnel_info_unclone(skb);
715 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX)))
718 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb);
720 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst);
723 * __dev_get_by_name - find a device by its name
724 * @net: the applicable net namespace
725 * @name: name to find
727 * Find an interface by name. Must be called under RTNL semaphore
728 * or @dev_base_lock. If the name is found a pointer to the device
729 * is returned. If the name is not found then %NULL is returned. The
730 * reference counters are not incremented so the caller must be
731 * careful with locks.
734 struct net_device *__dev_get_by_name(struct net *net, const char *name)
736 struct net_device *dev;
737 struct hlist_head *head = dev_name_hash(net, name);
739 hlist_for_each_entry(dev, head, name_hlist)
740 if (!strncmp(dev->name, name, IFNAMSIZ))
745 EXPORT_SYMBOL(__dev_get_by_name);
748 * dev_get_by_name_rcu - find a device by its name
749 * @net: the applicable net namespace
750 * @name: name to find
752 * Find an interface by name.
753 * If the name is found a pointer to the device is returned.
754 * If the name is not found then %NULL is returned.
755 * The reference counters are not incremented so the caller must be
756 * careful with locks. The caller must hold RCU lock.
759 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name)
761 struct net_device *dev;
762 struct hlist_head *head = dev_name_hash(net, name);
764 hlist_for_each_entry_rcu(dev, head, name_hlist)
765 if (!strncmp(dev->name, name, IFNAMSIZ))
770 EXPORT_SYMBOL(dev_get_by_name_rcu);
773 * dev_get_by_name - find a device by its name
774 * @net: the applicable net namespace
775 * @name: name to find
777 * Find an interface by name. This can be called from any
778 * context and does its own locking. The returned handle has
779 * the usage count incremented and the caller must use dev_put() to
780 * release it when it is no longer needed. %NULL is returned if no
781 * matching device is found.
784 struct net_device *dev_get_by_name(struct net *net, const char *name)
786 struct net_device *dev;
789 dev = dev_get_by_name_rcu(net, name);
795 EXPORT_SYMBOL(dev_get_by_name);
798 * __dev_get_by_index - find a device by its ifindex
799 * @net: the applicable net namespace
800 * @ifindex: index of device
802 * Search for an interface by index. Returns %NULL if the device
803 * is not found or a pointer to the device. The device has not
804 * had its reference counter increased so the caller must be careful
805 * about locking. The caller must hold either the RTNL semaphore
809 struct net_device *__dev_get_by_index(struct net *net, int ifindex)
811 struct net_device *dev;
812 struct hlist_head *head = dev_index_hash(net, ifindex);
814 hlist_for_each_entry(dev, head, index_hlist)
815 if (dev->ifindex == ifindex)
820 EXPORT_SYMBOL(__dev_get_by_index);
823 * dev_get_by_index_rcu - find a device by its ifindex
824 * @net: the applicable net namespace
825 * @ifindex: index of device
827 * Search for an interface by index. Returns %NULL if the device
828 * is not found or a pointer to the device. The device has not
829 * had its reference counter increased so the caller must be careful
830 * about locking. The caller must hold RCU lock.
833 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex)
835 struct net_device *dev;
836 struct hlist_head *head = dev_index_hash(net, ifindex);
838 hlist_for_each_entry_rcu(dev, head, index_hlist)
839 if (dev->ifindex == ifindex)
844 EXPORT_SYMBOL(dev_get_by_index_rcu);
848 * dev_get_by_index - find a device by its ifindex
849 * @net: the applicable net namespace
850 * @ifindex: index of device
852 * Search for an interface by index. Returns NULL if the device
853 * is not found or a pointer to the device. The device returned has
854 * had a reference added and the pointer is safe until the user calls
855 * dev_put to indicate they have finished with it.
858 struct net_device *dev_get_by_index(struct net *net, int ifindex)
860 struct net_device *dev;
863 dev = dev_get_by_index_rcu(net, ifindex);
869 EXPORT_SYMBOL(dev_get_by_index);
872 * dev_get_by_napi_id - find a device by napi_id
873 * @napi_id: ID of the NAPI struct
875 * Search for an interface by NAPI ID. Returns %NULL if the device
876 * is not found or a pointer to the device. The device has not had
877 * its reference counter increased so the caller must be careful
878 * about locking. The caller must hold RCU lock.
881 struct net_device *dev_get_by_napi_id(unsigned int napi_id)
883 struct napi_struct *napi;
885 WARN_ON_ONCE(!rcu_read_lock_held());
887 if (napi_id < MIN_NAPI_ID)
890 napi = napi_by_id(napi_id);
892 return napi ? napi->dev : NULL;
894 EXPORT_SYMBOL(dev_get_by_napi_id);
897 * netdev_get_name - get a netdevice name, knowing its ifindex.
898 * @net: network namespace
899 * @name: a pointer to the buffer where the name will be stored.
900 * @ifindex: the ifindex of the interface to get the name from.
902 * The use of raw_seqcount_begin() and cond_resched() before
903 * retrying is required as we want to give the writers a chance
904 * to complete when CONFIG_PREEMPT is not set.
906 int netdev_get_name(struct net *net, char *name, int ifindex)
908 struct net_device *dev;
912 seq = raw_seqcount_begin(&devnet_rename_seq);
914 dev = dev_get_by_index_rcu(net, ifindex);
920 strcpy(name, dev->name);
922 if (read_seqcount_retry(&devnet_rename_seq, seq)) {
931 * dev_getbyhwaddr_rcu - find a device by its hardware address
932 * @net: the applicable net namespace
933 * @type: media type of device
934 * @ha: hardware address
936 * Search for an interface by MAC address. Returns NULL if the device
937 * is not found or a pointer to the device.
938 * The caller must hold RCU or RTNL.
939 * The returned device has not had its ref count increased
940 * and the caller must therefore be careful about locking
944 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type,
947 struct net_device *dev;
949 for_each_netdev_rcu(net, dev)
950 if (dev->type == type &&
951 !memcmp(dev->dev_addr, ha, dev->addr_len))
956 EXPORT_SYMBOL(dev_getbyhwaddr_rcu);
958 struct net_device *__dev_getfirstbyhwtype(struct net *net, unsigned short type)
960 struct net_device *dev;
963 for_each_netdev(net, dev)
964 if (dev->type == type)
969 EXPORT_SYMBOL(__dev_getfirstbyhwtype);
971 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type)
973 struct net_device *dev, *ret = NULL;
976 for_each_netdev_rcu(net, dev)
977 if (dev->type == type) {
985 EXPORT_SYMBOL(dev_getfirstbyhwtype);
988 * __dev_get_by_flags - find any device with given flags
989 * @net: the applicable net namespace
990 * @if_flags: IFF_* values
991 * @mask: bitmask of bits in if_flags to check
993 * Search for any interface with the given flags. Returns NULL if a device
994 * is not found or a pointer to the device. Must be called inside
995 * rtnl_lock(), and result refcount is unchanged.
998 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags,
1001 struct net_device *dev, *ret;
1006 for_each_netdev(net, dev) {
1007 if (((dev->flags ^ if_flags) & mask) == 0) {
1014 EXPORT_SYMBOL(__dev_get_by_flags);
1017 * dev_valid_name - check if name is okay for network device
1018 * @name: name string
1020 * Network device names need to be valid file names to
1021 * to allow sysfs to work. We also disallow any kind of
1024 bool dev_valid_name(const char *name)
1028 if (strlen(name) >= IFNAMSIZ)
1030 if (!strcmp(name, ".") || !strcmp(name, ".."))
1034 if (*name == '/' || *name == ':' || isspace(*name))
1040 EXPORT_SYMBOL(dev_valid_name);
1043 * __dev_alloc_name - allocate a name for a device
1044 * @net: network namespace to allocate the device name in
1045 * @name: name format string
1046 * @buf: scratch buffer and result name string
1048 * Passed a format string - eg "lt%d" it will try and find a suitable
1049 * id. It scans list of devices to build up a free map, then chooses
1050 * the first empty slot. The caller must hold the dev_base or rtnl lock
1051 * while allocating the name and adding the device in order to avoid
1053 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1054 * Returns the number of the unit assigned or a negative errno code.
1057 static int __dev_alloc_name(struct net *net, const char *name, char *buf)
1061 const int max_netdevices = 8*PAGE_SIZE;
1062 unsigned long *inuse;
1063 struct net_device *d;
1065 p = strnchr(name, IFNAMSIZ-1, '%');
1068 * Verify the string as this thing may have come from
1069 * the user. There must be either one "%d" and no other "%"
1072 if (p[1] != 'd' || strchr(p + 2, '%'))
1075 /* Use one page as a bit array of possible slots */
1076 inuse = (unsigned long *) get_zeroed_page(GFP_ATOMIC);
1080 for_each_netdev(net, d) {
1081 if (!sscanf(d->name, name, &i))
1083 if (i < 0 || i >= max_netdevices)
1086 /* avoid cases where sscanf is not exact inverse of printf */
1087 snprintf(buf, IFNAMSIZ, name, i);
1088 if (!strncmp(buf, d->name, IFNAMSIZ))
1092 i = find_first_zero_bit(inuse, max_netdevices);
1093 free_page((unsigned long) inuse);
1097 snprintf(buf, IFNAMSIZ, name, i);
1098 if (!__dev_get_by_name(net, buf))
1101 /* It is possible to run out of possible slots
1102 * when the name is long and there isn't enough space left
1103 * for the digits, or if all bits are used.
1109 * dev_alloc_name - allocate a name for a device
1111 * @name: name format string
1113 * Passed a format string - eg "lt%d" it will try and find a suitable
1114 * id. It scans list of devices to build up a free map, then chooses
1115 * the first empty slot. The caller must hold the dev_base or rtnl lock
1116 * while allocating the name and adding the device in order to avoid
1118 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1119 * Returns the number of the unit assigned or a negative errno code.
1122 int dev_alloc_name(struct net_device *dev, const char *name)
1128 BUG_ON(!dev_net(dev));
1130 ret = __dev_alloc_name(net, name, buf);
1132 strlcpy(dev->name, buf, IFNAMSIZ);
1135 EXPORT_SYMBOL(dev_alloc_name);
1137 static int dev_alloc_name_ns(struct net *net,
1138 struct net_device *dev,
1144 ret = __dev_alloc_name(net, name, buf);
1146 strlcpy(dev->name, buf, IFNAMSIZ);
1150 static int dev_get_valid_name(struct net *net,
1151 struct net_device *dev,
1156 if (!dev_valid_name(name))
1159 if (strchr(name, '%'))
1160 return dev_alloc_name_ns(net, dev, name);
1161 else if (__dev_get_by_name(net, name))
1163 else if (dev->name != name)
1164 strlcpy(dev->name, name, IFNAMSIZ);
1170 * dev_change_name - change name of a device
1172 * @newname: name (or format string) must be at least IFNAMSIZ
1174 * Change name of a device, can pass format strings "eth%d".
1177 int dev_change_name(struct net_device *dev, const char *newname)
1179 unsigned char old_assign_type;
1180 char oldname[IFNAMSIZ];
1186 BUG_ON(!dev_net(dev));
1189 if (dev->flags & IFF_UP)
1192 write_seqcount_begin(&devnet_rename_seq);
1194 if (strncmp(newname, dev->name, IFNAMSIZ) == 0) {
1195 write_seqcount_end(&devnet_rename_seq);
1199 memcpy(oldname, dev->name, IFNAMSIZ);
1201 err = dev_get_valid_name(net, dev, newname);
1203 write_seqcount_end(&devnet_rename_seq);
1207 if (oldname[0] && !strchr(oldname, '%'))
1208 netdev_info(dev, "renamed from %s\n", oldname);
1210 old_assign_type = dev->name_assign_type;
1211 dev->name_assign_type = NET_NAME_RENAMED;
1214 ret = device_rename(&dev->dev, dev->name);
1216 memcpy(dev->name, oldname, IFNAMSIZ);
1217 dev->name_assign_type = old_assign_type;
1218 write_seqcount_end(&devnet_rename_seq);
1222 write_seqcount_end(&devnet_rename_seq);
1224 netdev_adjacent_rename_links(dev, oldname);
1226 write_lock_bh(&dev_base_lock);
1227 hlist_del_rcu(&dev->name_hlist);
1228 write_unlock_bh(&dev_base_lock);
1232 write_lock_bh(&dev_base_lock);
1233 hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
1234 write_unlock_bh(&dev_base_lock);
1236 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev);
1237 ret = notifier_to_errno(ret);
1240 /* err >= 0 after dev_alloc_name() or stores the first errno */
1243 write_seqcount_begin(&devnet_rename_seq);
1244 memcpy(dev->name, oldname, IFNAMSIZ);
1245 memcpy(oldname, newname, IFNAMSIZ);
1246 dev->name_assign_type = old_assign_type;
1247 old_assign_type = NET_NAME_RENAMED;
1250 pr_err("%s: name change rollback failed: %d\n",
1259 * dev_set_alias - change ifalias of a device
1261 * @alias: name up to IFALIASZ
1262 * @len: limit of bytes to copy from info
1264 * Set ifalias for a device,
1266 int dev_set_alias(struct net_device *dev, const char *alias, size_t len)
1272 if (len >= IFALIASZ)
1276 kfree(dev->ifalias);
1277 dev->ifalias = NULL;
1281 new_ifalias = krealloc(dev->ifalias, len + 1, GFP_KERNEL);
1284 dev->ifalias = new_ifalias;
1285 memcpy(dev->ifalias, alias, len);
1286 dev->ifalias[len] = 0;
1293 * netdev_features_change - device changes features
1294 * @dev: device to cause notification
1296 * Called to indicate a device has changed features.
1298 void netdev_features_change(struct net_device *dev)
1300 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev);
1302 EXPORT_SYMBOL(netdev_features_change);
1305 * netdev_state_change - device changes state
1306 * @dev: device to cause notification
1308 * Called to indicate a device has changed state. This function calls
1309 * the notifier chains for netdev_chain and sends a NEWLINK message
1310 * to the routing socket.
1312 void netdev_state_change(struct net_device *dev)
1314 if (dev->flags & IFF_UP) {
1315 struct netdev_notifier_change_info change_info;
1317 change_info.flags_changed = 0;
1318 call_netdevice_notifiers_info(NETDEV_CHANGE, dev,
1320 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL);
1323 EXPORT_SYMBOL(netdev_state_change);
1326 * netdev_notify_peers - notify network peers about existence of @dev
1327 * @dev: network device
1329 * Generate traffic such that interested network peers are aware of
1330 * @dev, such as by generating a gratuitous ARP. This may be used when
1331 * a device wants to inform the rest of the network about some sort of
1332 * reconfiguration such as a failover event or virtual machine
1335 void netdev_notify_peers(struct net_device *dev)
1338 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
1339 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev);
1342 EXPORT_SYMBOL(netdev_notify_peers);
1344 static int __dev_open(struct net_device *dev)
1346 const struct net_device_ops *ops = dev->netdev_ops;
1351 if (!netif_device_present(dev))
1354 /* Block netpoll from trying to do any rx path servicing.
1355 * If we don't do this there is a chance ndo_poll_controller
1356 * or ndo_poll may be running while we open the device
1358 netpoll_poll_disable(dev);
1360 ret = call_netdevice_notifiers(NETDEV_PRE_UP, dev);
1361 ret = notifier_to_errno(ret);
1365 set_bit(__LINK_STATE_START, &dev->state);
1367 if (ops->ndo_validate_addr)
1368 ret = ops->ndo_validate_addr(dev);
1370 if (!ret && ops->ndo_open)
1371 ret = ops->ndo_open(dev);
1373 netpoll_poll_enable(dev);
1376 clear_bit(__LINK_STATE_START, &dev->state);
1378 dev->flags |= IFF_UP;
1379 dev_set_rx_mode(dev);
1381 add_device_randomness(dev->dev_addr, dev->addr_len);
1388 * dev_open - prepare an interface for use.
1389 * @dev: device to open
1391 * Takes a device from down to up state. The device's private open
1392 * function is invoked and then the multicast lists are loaded. Finally
1393 * the device is moved into the up state and a %NETDEV_UP message is
1394 * sent to the netdev notifier chain.
1396 * Calling this function on an active interface is a nop. On a failure
1397 * a negative errno code is returned.
1399 int dev_open(struct net_device *dev)
1403 if (dev->flags & IFF_UP)
1406 ret = __dev_open(dev);
1410 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1411 call_netdevice_notifiers(NETDEV_UP, dev);
1415 EXPORT_SYMBOL(dev_open);
1417 static void __dev_close_many(struct list_head *head)
1419 struct net_device *dev;
1424 list_for_each_entry(dev, head, close_list) {
1425 /* Temporarily disable netpoll until the interface is down */
1426 netpoll_poll_disable(dev);
1428 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev);
1430 clear_bit(__LINK_STATE_START, &dev->state);
1432 /* Synchronize to scheduled poll. We cannot touch poll list, it
1433 * can be even on different cpu. So just clear netif_running().
1435 * dev->stop() will invoke napi_disable() on all of it's
1436 * napi_struct instances on this device.
1438 smp_mb__after_atomic(); /* Commit netif_running(). */
1441 dev_deactivate_many(head);
1443 list_for_each_entry(dev, head, close_list) {
1444 const struct net_device_ops *ops = dev->netdev_ops;
1447 * Call the device specific close. This cannot fail.
1448 * Only if device is UP
1450 * We allow it to be called even after a DETACH hot-plug
1456 dev->flags &= ~IFF_UP;
1457 netpoll_poll_enable(dev);
1461 static void __dev_close(struct net_device *dev)
1465 list_add(&dev->close_list, &single);
1466 __dev_close_many(&single);
1470 void dev_close_many(struct list_head *head, bool unlink)
1472 struct net_device *dev, *tmp;
1474 /* Remove the devices that don't need to be closed */
1475 list_for_each_entry_safe(dev, tmp, head, close_list)
1476 if (!(dev->flags & IFF_UP))
1477 list_del_init(&dev->close_list);
1479 __dev_close_many(head);
1481 list_for_each_entry_safe(dev, tmp, head, close_list) {
1482 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1483 call_netdevice_notifiers(NETDEV_DOWN, dev);
1485 list_del_init(&dev->close_list);
1488 EXPORT_SYMBOL(dev_close_many);
1491 * dev_close - shutdown an interface.
1492 * @dev: device to shutdown
1494 * This function moves an active device into down state. A
1495 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device
1496 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier
1499 void dev_close(struct net_device *dev)
1501 if (dev->flags & IFF_UP) {
1504 list_add(&dev->close_list, &single);
1505 dev_close_many(&single, true);
1509 EXPORT_SYMBOL(dev_close);
1513 * dev_disable_lro - disable Large Receive Offload on a device
1516 * Disable Large Receive Offload (LRO) on a net device. Must be
1517 * called under RTNL. This is needed if received packets may be
1518 * forwarded to another interface.
1520 void dev_disable_lro(struct net_device *dev)
1522 struct net_device *lower_dev;
1523 struct list_head *iter;
1525 dev->wanted_features &= ~NETIF_F_LRO;
1526 netdev_update_features(dev);
1528 if (unlikely(dev->features & NETIF_F_LRO))
1529 netdev_WARN(dev, "failed to disable LRO!\n");
1531 netdev_for_each_lower_dev(dev, lower_dev, iter)
1532 dev_disable_lro(lower_dev);
1534 EXPORT_SYMBOL(dev_disable_lro);
1536 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val,
1537 struct net_device *dev)
1539 struct netdev_notifier_info info;
1541 netdev_notifier_info_init(&info, dev);
1542 return nb->notifier_call(nb, val, &info);
1545 static int dev_boot_phase = 1;
1548 * register_netdevice_notifier - register a network notifier block
1551 * Register a notifier to be called when network device events occur.
1552 * The notifier passed is linked into the kernel structures and must
1553 * not be reused until it has been unregistered. A negative errno code
1554 * is returned on a failure.
1556 * When registered all registration and up events are replayed
1557 * to the new notifier to allow device to have a race free
1558 * view of the network device list.
1561 int register_netdevice_notifier(struct notifier_block *nb)
1563 struct net_device *dev;
1564 struct net_device *last;
1569 err = raw_notifier_chain_register(&netdev_chain, nb);
1575 for_each_netdev(net, dev) {
1576 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev);
1577 err = notifier_to_errno(err);
1581 if (!(dev->flags & IFF_UP))
1584 call_netdevice_notifier(nb, NETDEV_UP, dev);
1595 for_each_netdev(net, dev) {
1599 if (dev->flags & IFF_UP) {
1600 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1602 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1604 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1609 raw_notifier_chain_unregister(&netdev_chain, nb);
1612 EXPORT_SYMBOL(register_netdevice_notifier);
1615 * unregister_netdevice_notifier - unregister a network notifier block
1618 * Unregister a notifier previously registered by
1619 * register_netdevice_notifier(). The notifier is unlinked into the
1620 * kernel structures and may then be reused. A negative errno code
1621 * is returned on a failure.
1623 * After unregistering unregister and down device events are synthesized
1624 * for all devices on the device list to the removed notifier to remove
1625 * the need for special case cleanup code.
1628 int unregister_netdevice_notifier(struct notifier_block *nb)
1630 struct net_device *dev;
1635 err = raw_notifier_chain_unregister(&netdev_chain, nb);
1640 for_each_netdev(net, dev) {
1641 if (dev->flags & IFF_UP) {
1642 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1644 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1646 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1653 EXPORT_SYMBOL(unregister_netdevice_notifier);
1656 * call_netdevice_notifiers_info - call all network notifier blocks
1657 * @val: value passed unmodified to notifier function
1658 * @dev: net_device pointer passed unmodified to notifier function
1659 * @info: notifier information data
1661 * Call all network notifier blocks. Parameters and return value
1662 * are as for raw_notifier_call_chain().
1665 static int call_netdevice_notifiers_info(unsigned long val,
1666 struct net_device *dev,
1667 struct netdev_notifier_info *info)
1670 netdev_notifier_info_init(info, dev);
1671 return raw_notifier_call_chain(&netdev_chain, val, info);
1675 * call_netdevice_notifiers - call all network notifier blocks
1676 * @val: value passed unmodified to notifier function
1677 * @dev: net_device pointer passed unmodified to notifier function
1679 * Call all network notifier blocks. Parameters and return value
1680 * are as for raw_notifier_call_chain().
1683 int call_netdevice_notifiers(unsigned long val, struct net_device *dev)
1685 struct netdev_notifier_info info;
1687 return call_netdevice_notifiers_info(val, dev, &info);
1689 EXPORT_SYMBOL(call_netdevice_notifiers);
1691 #ifdef CONFIG_NET_INGRESS
1692 static struct static_key ingress_needed __read_mostly;
1694 void net_inc_ingress_queue(void)
1696 static_key_slow_inc(&ingress_needed);
1698 EXPORT_SYMBOL_GPL(net_inc_ingress_queue);
1700 void net_dec_ingress_queue(void)
1702 static_key_slow_dec(&ingress_needed);
1704 EXPORT_SYMBOL_GPL(net_dec_ingress_queue);
1707 #ifdef CONFIG_NET_EGRESS
1708 static struct static_key egress_needed __read_mostly;
1710 void net_inc_egress_queue(void)
1712 static_key_slow_inc(&egress_needed);
1714 EXPORT_SYMBOL_GPL(net_inc_egress_queue);
1716 void net_dec_egress_queue(void)
1718 static_key_slow_dec(&egress_needed);
1720 EXPORT_SYMBOL_GPL(net_dec_egress_queue);
1723 static struct static_key netstamp_needed __read_mostly;
1724 #ifdef HAVE_JUMP_LABEL
1725 static atomic_t netstamp_needed_deferred;
1726 static atomic_t netstamp_wanted;
1727 static void netstamp_clear(struct work_struct *work)
1729 int deferred = atomic_xchg(&netstamp_needed_deferred, 0);
1732 wanted = atomic_add_return(deferred, &netstamp_wanted);
1734 static_key_enable(&netstamp_needed);
1736 static_key_disable(&netstamp_needed);
1738 static DECLARE_WORK(netstamp_work, netstamp_clear);
1741 void net_enable_timestamp(void)
1743 #ifdef HAVE_JUMP_LABEL
1747 wanted = atomic_read(&netstamp_wanted);
1750 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted + 1) == wanted)
1753 atomic_inc(&netstamp_needed_deferred);
1754 schedule_work(&netstamp_work);
1756 static_key_slow_inc(&netstamp_needed);
1759 EXPORT_SYMBOL(net_enable_timestamp);
1761 void net_disable_timestamp(void)
1763 #ifdef HAVE_JUMP_LABEL
1767 wanted = atomic_read(&netstamp_wanted);
1770 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted - 1) == wanted)
1773 atomic_dec(&netstamp_needed_deferred);
1774 schedule_work(&netstamp_work);
1776 static_key_slow_dec(&netstamp_needed);
1779 EXPORT_SYMBOL(net_disable_timestamp);
1781 static inline void net_timestamp_set(struct sk_buff *skb)
1784 if (static_key_false(&netstamp_needed))
1785 __net_timestamp(skb);
1788 #define net_timestamp_check(COND, SKB) \
1789 if (static_key_false(&netstamp_needed)) { \
1790 if ((COND) && !(SKB)->tstamp) \
1791 __net_timestamp(SKB); \
1794 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb)
1798 if (!(dev->flags & IFF_UP))
1801 len = dev->mtu + dev->hard_header_len + VLAN_HLEN;
1802 if (skb->len <= len)
1805 /* if TSO is enabled, we don't care about the length as the packet
1806 * could be forwarded without being segmented before
1808 if (skb_is_gso(skb))
1813 EXPORT_SYMBOL_GPL(is_skb_forwardable);
1815 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1817 int ret = ____dev_forward_skb(dev, skb);
1820 skb->protocol = eth_type_trans(skb, dev);
1821 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
1826 EXPORT_SYMBOL_GPL(__dev_forward_skb);
1829 * dev_forward_skb - loopback an skb to another netif
1831 * @dev: destination network device
1832 * @skb: buffer to forward
1835 * NET_RX_SUCCESS (no congestion)
1836 * NET_RX_DROP (packet was dropped, but freed)
1838 * dev_forward_skb can be used for injecting an skb from the
1839 * start_xmit function of one device into the receive queue
1840 * of another device.
1842 * The receiving device may be in another namespace, so
1843 * we have to clear all information in the skb that could
1844 * impact namespace isolation.
1846 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1848 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb);
1850 EXPORT_SYMBOL_GPL(dev_forward_skb);
1852 static inline int deliver_skb(struct sk_buff *skb,
1853 struct packet_type *pt_prev,
1854 struct net_device *orig_dev)
1856 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
1858 refcount_inc(&skb->users);
1859 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
1862 static inline void deliver_ptype_list_skb(struct sk_buff *skb,
1863 struct packet_type **pt,
1864 struct net_device *orig_dev,
1866 struct list_head *ptype_list)
1868 struct packet_type *ptype, *pt_prev = *pt;
1870 list_for_each_entry_rcu(ptype, ptype_list, list) {
1871 if (ptype->type != type)
1874 deliver_skb(skb, pt_prev, orig_dev);
1880 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb)
1882 if (!ptype->af_packet_priv || !skb->sk)
1885 if (ptype->id_match)
1886 return ptype->id_match(ptype, skb->sk);
1887 else if ((struct sock *)ptype->af_packet_priv == skb->sk)
1894 * Support routine. Sends outgoing frames to any network
1895 * taps currently in use.
1898 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
1900 struct packet_type *ptype;
1901 struct sk_buff *skb2 = NULL;
1902 struct packet_type *pt_prev = NULL;
1903 struct list_head *ptype_list = &ptype_all;
1907 list_for_each_entry_rcu(ptype, ptype_list, list) {
1908 /* Never send packets back to the socket
1909 * they originated from - MvS (miquels@drinkel.ow.org)
1911 if (skb_loop_sk(ptype, skb))
1915 deliver_skb(skb2, pt_prev, skb->dev);
1920 /* need to clone skb, done only once */
1921 skb2 = skb_clone(skb, GFP_ATOMIC);
1925 net_timestamp_set(skb2);
1927 /* skb->nh should be correctly
1928 * set by sender, so that the second statement is
1929 * just protection against buggy protocols.
1931 skb_reset_mac_header(skb2);
1933 if (skb_network_header(skb2) < skb2->data ||
1934 skb_network_header(skb2) > skb_tail_pointer(skb2)) {
1935 net_crit_ratelimited("protocol %04x is buggy, dev %s\n",
1936 ntohs(skb2->protocol),
1938 skb_reset_network_header(skb2);
1941 skb2->transport_header = skb2->network_header;
1942 skb2->pkt_type = PACKET_OUTGOING;
1946 if (ptype_list == &ptype_all) {
1947 ptype_list = &dev->ptype_all;
1952 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
1955 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit);
1958 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change
1959 * @dev: Network device
1960 * @txq: number of queues available
1962 * If real_num_tx_queues is changed the tc mappings may no longer be
1963 * valid. To resolve this verify the tc mapping remains valid and if
1964 * not NULL the mapping. With no priorities mapping to this
1965 * offset/count pair it will no longer be used. In the worst case TC0
1966 * is invalid nothing can be done so disable priority mappings. If is
1967 * expected that drivers will fix this mapping if they can before
1968 * calling netif_set_real_num_tx_queues.
1970 static void netif_setup_tc(struct net_device *dev, unsigned int txq)
1973 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
1975 /* If TC0 is invalidated disable TC mapping */
1976 if (tc->offset + tc->count > txq) {
1977 pr_warn("Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n");
1982 /* Invalidated prio to tc mappings set to TC0 */
1983 for (i = 1; i < TC_BITMASK + 1; i++) {
1984 int q = netdev_get_prio_tc_map(dev, i);
1986 tc = &dev->tc_to_txq[q];
1987 if (tc->offset + tc->count > txq) {
1988 pr_warn("Number of in use tx queues changed. Priority %i to tc mapping %i is no longer valid. Setting map to 0\n",
1990 netdev_set_prio_tc_map(dev, i, 0);
1995 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq)
1998 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2001 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) {
2002 if ((txq - tc->offset) < tc->count)
2013 static DEFINE_MUTEX(xps_map_mutex);
2014 #define xmap_dereference(P) \
2015 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex))
2017 static bool remove_xps_queue(struct xps_dev_maps *dev_maps,
2020 struct xps_map *map = NULL;
2024 map = xmap_dereference(dev_maps->cpu_map[tci]);
2028 for (pos = map->len; pos--;) {
2029 if (map->queues[pos] != index)
2033 map->queues[pos] = map->queues[--map->len];
2037 RCU_INIT_POINTER(dev_maps->cpu_map[tci], NULL);
2038 kfree_rcu(map, rcu);
2045 static bool remove_xps_queue_cpu(struct net_device *dev,
2046 struct xps_dev_maps *dev_maps,
2047 int cpu, u16 offset, u16 count)
2049 int num_tc = dev->num_tc ? : 1;
2050 bool active = false;
2053 for (tci = cpu * num_tc; num_tc--; tci++) {
2056 for (i = count, j = offset; i--; j++) {
2057 if (!remove_xps_queue(dev_maps, cpu, j))
2067 static void netif_reset_xps_queues(struct net_device *dev, u16 offset,
2070 struct xps_dev_maps *dev_maps;
2072 bool active = false;
2074 mutex_lock(&xps_map_mutex);
2075 dev_maps = xmap_dereference(dev->xps_maps);
2080 for_each_possible_cpu(cpu)
2081 active |= remove_xps_queue_cpu(dev, dev_maps, cpu,
2085 RCU_INIT_POINTER(dev->xps_maps, NULL);
2086 kfree_rcu(dev_maps, rcu);
2089 for (i = offset + (count - 1); count--; i--)
2090 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, i),
2094 mutex_unlock(&xps_map_mutex);
2097 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index)
2099 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index);
2102 static struct xps_map *expand_xps_map(struct xps_map *map,
2105 struct xps_map *new_map;
2106 int alloc_len = XPS_MIN_MAP_ALLOC;
2109 for (pos = 0; map && pos < map->len; pos++) {
2110 if (map->queues[pos] != index)
2115 /* Need to add queue to this CPU's existing map */
2117 if (pos < map->alloc_len)
2120 alloc_len = map->alloc_len * 2;
2123 /* Need to allocate new map to store queue on this CPU's map */
2124 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL,
2129 for (i = 0; i < pos; i++)
2130 new_map->queues[i] = map->queues[i];
2131 new_map->alloc_len = alloc_len;
2137 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask,
2140 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL;
2141 int i, cpu, tci, numa_node_id = -2;
2142 int maps_sz, num_tc = 1, tc = 0;
2143 struct xps_map *map, *new_map;
2144 bool active = false;
2147 num_tc = dev->num_tc;
2148 tc = netdev_txq_to_tc(dev, index);
2153 maps_sz = XPS_DEV_MAPS_SIZE(num_tc);
2154 if (maps_sz < L1_CACHE_BYTES)
2155 maps_sz = L1_CACHE_BYTES;
2157 mutex_lock(&xps_map_mutex);
2159 dev_maps = xmap_dereference(dev->xps_maps);
2161 /* allocate memory for queue storage */
2162 for_each_cpu_and(cpu, cpu_online_mask, mask) {
2164 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL);
2165 if (!new_dev_maps) {
2166 mutex_unlock(&xps_map_mutex);
2170 tci = cpu * num_tc + tc;
2171 map = dev_maps ? xmap_dereference(dev_maps->cpu_map[tci]) :
2174 map = expand_xps_map(map, cpu, index);
2178 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2182 goto out_no_new_maps;
2184 for_each_possible_cpu(cpu) {
2185 /* copy maps belonging to foreign traffic classes */
2186 for (i = tc, tci = cpu * num_tc; dev_maps && i--; tci++) {
2187 /* fill in the new device map from the old device map */
2188 map = xmap_dereference(dev_maps->cpu_map[tci]);
2189 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2192 /* We need to explicitly update tci as prevous loop
2193 * could break out early if dev_maps is NULL.
2195 tci = cpu * num_tc + tc;
2197 if (cpumask_test_cpu(cpu, mask) && cpu_online(cpu)) {
2198 /* add queue to CPU maps */
2201 map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2202 while ((pos < map->len) && (map->queues[pos] != index))
2205 if (pos == map->len)
2206 map->queues[map->len++] = index;
2208 if (numa_node_id == -2)
2209 numa_node_id = cpu_to_node(cpu);
2210 else if (numa_node_id != cpu_to_node(cpu))
2213 } else if (dev_maps) {
2214 /* fill in the new device map from the old device map */
2215 map = xmap_dereference(dev_maps->cpu_map[tci]);
2216 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2219 /* copy maps belonging to foreign traffic classes */
2220 for (i = num_tc - tc, tci++; dev_maps && --i; tci++) {
2221 /* fill in the new device map from the old device map */
2222 map = xmap_dereference(dev_maps->cpu_map[tci]);
2223 RCU_INIT_POINTER(new_dev_maps->cpu_map[tci], map);
2227 rcu_assign_pointer(dev->xps_maps, new_dev_maps);
2229 /* Cleanup old maps */
2231 goto out_no_old_maps;
2233 for_each_possible_cpu(cpu) {
2234 for (i = num_tc, tci = cpu * num_tc; i--; tci++) {
2235 new_map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2236 map = xmap_dereference(dev_maps->cpu_map[tci]);
2237 if (map && map != new_map)
2238 kfree_rcu(map, rcu);
2242 kfree_rcu(dev_maps, rcu);
2245 dev_maps = new_dev_maps;
2249 /* update Tx queue numa node */
2250 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index),
2251 (numa_node_id >= 0) ? numa_node_id :
2257 /* removes queue from unused CPUs */
2258 for_each_possible_cpu(cpu) {
2259 for (i = tc, tci = cpu * num_tc; i--; tci++)
2260 active |= remove_xps_queue(dev_maps, tci, index);
2261 if (!cpumask_test_cpu(cpu, mask) || !cpu_online(cpu))
2262 active |= remove_xps_queue(dev_maps, tci, index);
2263 for (i = num_tc - tc, tci++; --i; tci++)
2264 active |= remove_xps_queue(dev_maps, tci, index);
2267 /* free map if not active */
2269 RCU_INIT_POINTER(dev->xps_maps, NULL);
2270 kfree_rcu(dev_maps, rcu);
2274 mutex_unlock(&xps_map_mutex);
2278 /* remove any maps that we added */
2279 for_each_possible_cpu(cpu) {
2280 for (i = num_tc, tci = cpu * num_tc; i--; tci++) {
2281 new_map = xmap_dereference(new_dev_maps->cpu_map[tci]);
2283 xmap_dereference(dev_maps->cpu_map[tci]) :
2285 if (new_map && new_map != map)
2290 mutex_unlock(&xps_map_mutex);
2292 kfree(new_dev_maps);
2295 EXPORT_SYMBOL(netif_set_xps_queue);
2298 void netdev_reset_tc(struct net_device *dev)
2301 netif_reset_xps_queues_gt(dev, 0);
2304 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq));
2305 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map));
2307 EXPORT_SYMBOL(netdev_reset_tc);
2309 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset)
2311 if (tc >= dev->num_tc)
2315 netif_reset_xps_queues(dev, offset, count);
2317 dev->tc_to_txq[tc].count = count;
2318 dev->tc_to_txq[tc].offset = offset;
2321 EXPORT_SYMBOL(netdev_set_tc_queue);
2323 int netdev_set_num_tc(struct net_device *dev, u8 num_tc)
2325 if (num_tc > TC_MAX_QUEUE)
2329 netif_reset_xps_queues_gt(dev, 0);
2331 dev->num_tc = num_tc;
2334 EXPORT_SYMBOL(netdev_set_num_tc);
2337 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues
2338 * greater then real_num_tx_queues stale skbs on the qdisc must be flushed.
2340 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq)
2344 if (txq < 1 || txq > dev->num_tx_queues)
2347 if (dev->reg_state == NETREG_REGISTERED ||
2348 dev->reg_state == NETREG_UNREGISTERING) {
2351 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues,
2357 netif_setup_tc(dev, txq);
2359 if (txq < dev->real_num_tx_queues) {
2360 qdisc_reset_all_tx_gt(dev, txq);
2362 netif_reset_xps_queues_gt(dev, txq);
2367 dev->real_num_tx_queues = txq;
2370 EXPORT_SYMBOL(netif_set_real_num_tx_queues);
2374 * netif_set_real_num_rx_queues - set actual number of RX queues used
2375 * @dev: Network device
2376 * @rxq: Actual number of RX queues
2378 * This must be called either with the rtnl_lock held or before
2379 * registration of the net device. Returns 0 on success, or a
2380 * negative error code. If called before registration, it always
2383 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq)
2387 if (rxq < 1 || rxq > dev->num_rx_queues)
2390 if (dev->reg_state == NETREG_REGISTERED) {
2393 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues,
2399 dev->real_num_rx_queues = rxq;
2402 EXPORT_SYMBOL(netif_set_real_num_rx_queues);
2406 * netif_get_num_default_rss_queues - default number of RSS queues
2408 * This routine should set an upper limit on the number of RSS queues
2409 * used by default by multiqueue devices.
2411 int netif_get_num_default_rss_queues(void)
2413 return is_kdump_kernel() ?
2414 1 : min_t(int, DEFAULT_MAX_NUM_RSS_QUEUES, num_online_cpus());
2416 EXPORT_SYMBOL(netif_get_num_default_rss_queues);
2418 static void __netif_reschedule(struct Qdisc *q)
2420 struct softnet_data *sd;
2421 unsigned long flags;
2423 local_irq_save(flags);
2424 sd = this_cpu_ptr(&softnet_data);
2425 q->next_sched = NULL;
2426 *sd->output_queue_tailp = q;
2427 sd->output_queue_tailp = &q->next_sched;
2428 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2429 local_irq_restore(flags);
2432 void __netif_schedule(struct Qdisc *q)
2434 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
2435 __netif_reschedule(q);
2437 EXPORT_SYMBOL(__netif_schedule);
2439 struct dev_kfree_skb_cb {
2440 enum skb_free_reason reason;
2443 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb)
2445 return (struct dev_kfree_skb_cb *)skb->cb;
2448 void netif_schedule_queue(struct netdev_queue *txq)
2451 if (!(txq->state & QUEUE_STATE_ANY_XOFF)) {
2452 struct Qdisc *q = rcu_dereference(txq->qdisc);
2454 __netif_schedule(q);
2458 EXPORT_SYMBOL(netif_schedule_queue);
2460 void netif_tx_wake_queue(struct netdev_queue *dev_queue)
2462 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) {
2466 q = rcu_dereference(dev_queue->qdisc);
2467 __netif_schedule(q);
2471 EXPORT_SYMBOL(netif_tx_wake_queue);
2473 void __dev_kfree_skb_irq(struct sk_buff *skb, enum skb_free_reason reason)
2475 unsigned long flags;
2480 if (likely(refcount_read(&skb->users) == 1)) {
2482 refcount_set(&skb->users, 0);
2483 } else if (likely(!refcount_dec_and_test(&skb->users))) {
2486 get_kfree_skb_cb(skb)->reason = reason;
2487 local_irq_save(flags);
2488 skb->next = __this_cpu_read(softnet_data.completion_queue);
2489 __this_cpu_write(softnet_data.completion_queue, skb);
2490 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2491 local_irq_restore(flags);
2493 EXPORT_SYMBOL(__dev_kfree_skb_irq);
2495 void __dev_kfree_skb_any(struct sk_buff *skb, enum skb_free_reason reason)
2497 if (in_irq() || irqs_disabled())
2498 __dev_kfree_skb_irq(skb, reason);
2502 EXPORT_SYMBOL(__dev_kfree_skb_any);
2506 * netif_device_detach - mark device as removed
2507 * @dev: network device
2509 * Mark device as removed from system and therefore no longer available.
2511 void netif_device_detach(struct net_device *dev)
2513 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) &&
2514 netif_running(dev)) {
2515 netif_tx_stop_all_queues(dev);
2518 EXPORT_SYMBOL(netif_device_detach);
2521 * netif_device_attach - mark device as attached
2522 * @dev: network device
2524 * Mark device as attached from system and restart if needed.
2526 void netif_device_attach(struct net_device *dev)
2528 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) &&
2529 netif_running(dev)) {
2530 netif_tx_wake_all_queues(dev);
2531 __netdev_watchdog_up(dev);
2534 EXPORT_SYMBOL(netif_device_attach);
2537 * Returns a Tx hash based on the given packet descriptor a Tx queues' number
2538 * to be used as a distribution range.
2540 u16 __skb_tx_hash(const struct net_device *dev, struct sk_buff *skb,
2541 unsigned int num_tx_queues)
2545 u16 qcount = num_tx_queues;
2547 if (skb_rx_queue_recorded(skb)) {
2548 hash = skb_get_rx_queue(skb);
2549 while (unlikely(hash >= num_tx_queues))
2550 hash -= num_tx_queues;
2555 u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
2557 qoffset = dev->tc_to_txq[tc].offset;
2558 qcount = dev->tc_to_txq[tc].count;
2561 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset;
2563 EXPORT_SYMBOL(__skb_tx_hash);
2565 static void skb_warn_bad_offload(const struct sk_buff *skb)
2567 static const netdev_features_t null_features;
2568 struct net_device *dev = skb->dev;
2569 const char *name = "";
2571 if (!net_ratelimit())
2575 if (dev->dev.parent)
2576 name = dev_driver_string(dev->dev.parent);
2578 name = netdev_name(dev);
2580 WARN(1, "%s: caps=(%pNF, %pNF) len=%d data_len=%d gso_size=%d "
2581 "gso_type=%d ip_summed=%d\n",
2582 name, dev ? &dev->features : &null_features,
2583 skb->sk ? &skb->sk->sk_route_caps : &null_features,
2584 skb->len, skb->data_len, skb_shinfo(skb)->gso_size,
2585 skb_shinfo(skb)->gso_type, skb->ip_summed);
2589 * Invalidate hardware checksum when packet is to be mangled, and
2590 * complete checksum manually on outgoing path.
2592 int skb_checksum_help(struct sk_buff *skb)
2595 int ret = 0, offset;
2597 if (skb->ip_summed == CHECKSUM_COMPLETE)
2598 goto out_set_summed;
2600 if (unlikely(skb_shinfo(skb)->gso_size)) {
2601 skb_warn_bad_offload(skb);
2605 /* Before computing a checksum, we should make sure no frag could
2606 * be modified by an external entity : checksum could be wrong.
2608 if (skb_has_shared_frag(skb)) {
2609 ret = __skb_linearize(skb);
2614 offset = skb_checksum_start_offset(skb);
2615 BUG_ON(offset >= skb_headlen(skb));
2616 csum = skb_checksum(skb, offset, skb->len - offset, 0);
2618 offset += skb->csum_offset;
2619 BUG_ON(offset + sizeof(__sum16) > skb_headlen(skb));
2621 if (skb_cloned(skb) &&
2622 !skb_clone_writable(skb, offset + sizeof(__sum16))) {
2623 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2628 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0;
2630 skb->ip_summed = CHECKSUM_NONE;
2634 EXPORT_SYMBOL(skb_checksum_help);
2636 int skb_crc32c_csum_help(struct sk_buff *skb)
2639 int ret = 0, offset, start;
2641 if (skb->ip_summed != CHECKSUM_PARTIAL)
2644 if (unlikely(skb_is_gso(skb)))
2647 /* Before computing a checksum, we should make sure no frag could
2648 * be modified by an external entity : checksum could be wrong.
2650 if (unlikely(skb_has_shared_frag(skb))) {
2651 ret = __skb_linearize(skb);
2655 start = skb_checksum_start_offset(skb);
2656 offset = start + offsetof(struct sctphdr, checksum);
2657 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) {
2661 if (skb_cloned(skb) &&
2662 !skb_clone_writable(skb, offset + sizeof(__le32))) {
2663 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2667 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start,
2668 skb->len - start, ~(__u32)0,
2670 *(__le32 *)(skb->data + offset) = crc32c_csum;
2671 skb->ip_summed = CHECKSUM_NONE;
2672 skb->csum_not_inet = 0;
2677 __be16 skb_network_protocol(struct sk_buff *skb, int *depth)
2679 __be16 type = skb->protocol;
2681 /* Tunnel gso handlers can set protocol to ethernet. */
2682 if (type == htons(ETH_P_TEB)) {
2685 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr))))
2688 eth = (struct ethhdr *)skb_mac_header(skb);
2689 type = eth->h_proto;
2692 return __vlan_get_protocol(skb, type, depth);
2696 * skb_mac_gso_segment - mac layer segmentation handler.
2697 * @skb: buffer to segment
2698 * @features: features for the output path (see dev->features)
2700 struct sk_buff *skb_mac_gso_segment(struct sk_buff *skb,
2701 netdev_features_t features)
2703 struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT);
2704 struct packet_offload *ptype;
2705 int vlan_depth = skb->mac_len;
2706 __be16 type = skb_network_protocol(skb, &vlan_depth);
2708 if (unlikely(!type))
2709 return ERR_PTR(-EINVAL);
2711 __skb_pull(skb, vlan_depth);
2714 list_for_each_entry_rcu(ptype, &offload_base, list) {
2715 if (ptype->type == type && ptype->callbacks.gso_segment) {
2716 segs = ptype->callbacks.gso_segment(skb, features);
2722 __skb_push(skb, skb->data - skb_mac_header(skb));
2726 EXPORT_SYMBOL(skb_mac_gso_segment);
2729 /* openvswitch calls this on rx path, so we need a different check.
2731 static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
2734 return skb->ip_summed != CHECKSUM_PARTIAL &&
2735 skb->ip_summed != CHECKSUM_UNNECESSARY;
2737 return skb->ip_summed == CHECKSUM_NONE;
2741 * __skb_gso_segment - Perform segmentation on skb.
2742 * @skb: buffer to segment
2743 * @features: features for the output path (see dev->features)
2744 * @tx_path: whether it is called in TX path
2746 * This function segments the given skb and returns a list of segments.
2748 * It may return NULL if the skb requires no segmentation. This is
2749 * only possible when GSO is used for verifying header integrity.
2751 * Segmentation preserves SKB_SGO_CB_OFFSET bytes of previous skb cb.
2753 struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
2754 netdev_features_t features, bool tx_path)
2756 struct sk_buff *segs;
2758 if (unlikely(skb_needs_check(skb, tx_path))) {
2761 /* We're going to init ->check field in TCP or UDP header */
2762 err = skb_cow_head(skb, 0);
2764 return ERR_PTR(err);
2767 /* Only report GSO partial support if it will enable us to
2768 * support segmentation on this frame without needing additional
2771 if (features & NETIF_F_GSO_PARTIAL) {
2772 netdev_features_t partial_features = NETIF_F_GSO_ROBUST;
2773 struct net_device *dev = skb->dev;
2775 partial_features |= dev->features & dev->gso_partial_features;
2776 if (!skb_gso_ok(skb, features | partial_features))
2777 features &= ~NETIF_F_GSO_PARTIAL;
2780 BUILD_BUG_ON(SKB_SGO_CB_OFFSET +
2781 sizeof(*SKB_GSO_CB(skb)) > sizeof(skb->cb));
2783 SKB_GSO_CB(skb)->mac_offset = skb_headroom(skb);
2784 SKB_GSO_CB(skb)->encap_level = 0;
2786 skb_reset_mac_header(skb);
2787 skb_reset_mac_len(skb);
2789 segs = skb_mac_gso_segment(skb, features);
2791 if (unlikely(skb_needs_check(skb, tx_path)))
2792 skb_warn_bad_offload(skb);
2796 EXPORT_SYMBOL(__skb_gso_segment);
2798 /* Take action when hardware reception checksum errors are detected. */
2800 void netdev_rx_csum_fault(struct net_device *dev)
2802 if (net_ratelimit()) {
2803 pr_err("%s: hw csum failure\n", dev ? dev->name : "<unknown>");
2807 EXPORT_SYMBOL(netdev_rx_csum_fault);
2810 /* Actually, we should eliminate this check as soon as we know, that:
2811 * 1. IOMMU is present and allows to map all the memory.
2812 * 2. No high memory really exists on this machine.
2815 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
2817 #ifdef CONFIG_HIGHMEM
2820 if (!(dev->features & NETIF_F_HIGHDMA)) {
2821 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2822 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2824 if (PageHighMem(skb_frag_page(frag)))
2829 if (PCI_DMA_BUS_IS_PHYS) {
2830 struct device *pdev = dev->dev.parent;
2834 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
2835 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
2836 dma_addr_t addr = page_to_phys(skb_frag_page(frag));
2838 if (!pdev->dma_mask || addr + PAGE_SIZE - 1 > *pdev->dma_mask)
2846 /* If MPLS offload request, verify we are testing hardware MPLS features
2847 * instead of standard features for the netdev.
2849 #if IS_ENABLED(CONFIG_NET_MPLS_GSO)
2850 static netdev_features_t net_mpls_features(struct sk_buff *skb,
2851 netdev_features_t features,
2854 if (eth_p_mpls(type))
2855 features &= skb->dev->mpls_features;
2860 static netdev_features_t net_mpls_features(struct sk_buff *skb,
2861 netdev_features_t features,
2868 static netdev_features_t harmonize_features(struct sk_buff *skb,
2869 netdev_features_t features)
2874 type = skb_network_protocol(skb, &tmp);
2875 features = net_mpls_features(skb, features, type);
2877 if (skb->ip_summed != CHECKSUM_NONE &&
2878 !can_checksum_protocol(features, type)) {
2879 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
2881 if (illegal_highdma(skb->dev, skb))
2882 features &= ~NETIF_F_SG;
2887 netdev_features_t passthru_features_check(struct sk_buff *skb,
2888 struct net_device *dev,
2889 netdev_features_t features)
2893 EXPORT_SYMBOL(passthru_features_check);
2895 static netdev_features_t dflt_features_check(const struct sk_buff *skb,
2896 struct net_device *dev,
2897 netdev_features_t features)
2899 return vlan_features_check(skb, features);
2902 static netdev_features_t gso_features_check(const struct sk_buff *skb,
2903 struct net_device *dev,
2904 netdev_features_t features)
2906 u16 gso_segs = skb_shinfo(skb)->gso_segs;
2908 if (gso_segs > dev->gso_max_segs)
2909 return features & ~NETIF_F_GSO_MASK;
2911 /* Support for GSO partial features requires software
2912 * intervention before we can actually process the packets
2913 * so we need to strip support for any partial features now
2914 * and we can pull them back in after we have partially
2915 * segmented the frame.
2917 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL))
2918 features &= ~dev->gso_partial_features;
2920 /* Make sure to clear the IPv4 ID mangling feature if the
2921 * IPv4 header has the potential to be fragmented.
2923 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
2924 struct iphdr *iph = skb->encapsulation ?
2925 inner_ip_hdr(skb) : ip_hdr(skb);
2927 if (!(iph->frag_off & htons(IP_DF)))
2928 features &= ~NETIF_F_TSO_MANGLEID;
2934 netdev_features_t netif_skb_features(struct sk_buff *skb)
2936 struct net_device *dev = skb->dev;
2937 netdev_features_t features = dev->features;
2939 if (skb_is_gso(skb))
2940 features = gso_features_check(skb, dev, features);
2942 /* If encapsulation offload request, verify we are testing
2943 * hardware encapsulation features instead of standard
2944 * features for the netdev
2946 if (skb->encapsulation)
2947 features &= dev->hw_enc_features;
2949 if (skb_vlan_tagged(skb))
2950 features = netdev_intersect_features(features,
2951 dev->vlan_features |
2952 NETIF_F_HW_VLAN_CTAG_TX |
2953 NETIF_F_HW_VLAN_STAG_TX);
2955 if (dev->netdev_ops->ndo_features_check)
2956 features &= dev->netdev_ops->ndo_features_check(skb, dev,
2959 features &= dflt_features_check(skb, dev, features);
2961 return harmonize_features(skb, features);
2963 EXPORT_SYMBOL(netif_skb_features);
2965 static int xmit_one(struct sk_buff *skb, struct net_device *dev,
2966 struct netdev_queue *txq, bool more)
2971 if (!list_empty(&ptype_all) || !list_empty(&dev->ptype_all))
2972 dev_queue_xmit_nit(skb, dev);
2975 trace_net_dev_start_xmit(skb, dev);
2976 rc = netdev_start_xmit(skb, dev, txq, more);
2977 trace_net_dev_xmit(skb, rc, dev, len);
2982 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev,
2983 struct netdev_queue *txq, int *ret)
2985 struct sk_buff *skb = first;
2986 int rc = NETDEV_TX_OK;
2989 struct sk_buff *next = skb->next;
2992 rc = xmit_one(skb, dev, txq, next != NULL);
2993 if (unlikely(!dev_xmit_complete(rc))) {
2999 if (netif_xmit_stopped(txq) && skb) {
3000 rc = NETDEV_TX_BUSY;
3010 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb,
3011 netdev_features_t features)
3013 if (skb_vlan_tag_present(skb) &&
3014 !vlan_hw_offload_capable(features, skb->vlan_proto))
3015 skb = __vlan_hwaccel_push_inside(skb);
3019 int skb_csum_hwoffload_help(struct sk_buff *skb,
3020 const netdev_features_t features)
3022 if (unlikely(skb->csum_not_inet))
3023 return !!(features & NETIF_F_SCTP_CRC) ? 0 :
3024 skb_crc32c_csum_help(skb);
3026 return !!(features & NETIF_F_CSUM_MASK) ? 0 : skb_checksum_help(skb);
3028 EXPORT_SYMBOL(skb_csum_hwoffload_help);
3030 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev)
3032 netdev_features_t features;
3034 features = netif_skb_features(skb);
3035 skb = validate_xmit_vlan(skb, features);
3039 if (netif_needs_gso(skb, features)) {
3040 struct sk_buff *segs;
3042 segs = skb_gso_segment(skb, features);
3050 if (skb_needs_linearize(skb, features) &&
3051 __skb_linearize(skb))
3054 if (validate_xmit_xfrm(skb, features))
3057 /* If packet is not checksummed and device does not
3058 * support checksumming for this protocol, complete
3059 * checksumming here.
3061 if (skb->ip_summed == CHECKSUM_PARTIAL) {
3062 if (skb->encapsulation)
3063 skb_set_inner_transport_header(skb,
3064 skb_checksum_start_offset(skb));
3066 skb_set_transport_header(skb,
3067 skb_checksum_start_offset(skb));
3068 if (skb_csum_hwoffload_help(skb, features))
3078 atomic_long_inc(&dev->tx_dropped);
3082 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev)
3084 struct sk_buff *next, *head = NULL, *tail;
3086 for (; skb != NULL; skb = next) {
3090 /* in case skb wont be segmented, point to itself */
3093 skb = validate_xmit_skb(skb, dev);
3101 /* If skb was segmented, skb->prev points to
3102 * the last segment. If not, it still contains skb.
3108 EXPORT_SYMBOL_GPL(validate_xmit_skb_list);
3110 static void qdisc_pkt_len_init(struct sk_buff *skb)
3112 const struct skb_shared_info *shinfo = skb_shinfo(skb);
3114 qdisc_skb_cb(skb)->pkt_len = skb->len;
3116 /* To get more precise estimation of bytes sent on wire,
3117 * we add to pkt_len the headers size of all segments
3119 if (shinfo->gso_size) {
3120 unsigned int hdr_len;
3121 u16 gso_segs = shinfo->gso_segs;
3123 /* mac layer + network layer */
3124 hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
3126 /* + transport layer */
3127 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6)))
3128 hdr_len += tcp_hdrlen(skb);
3130 hdr_len += sizeof(struct udphdr);
3132 if (shinfo->gso_type & SKB_GSO_DODGY)
3133 gso_segs = DIV_ROUND_UP(skb->len - hdr_len,
3136 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len;
3140 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
3141 struct net_device *dev,
3142 struct netdev_queue *txq)
3144 spinlock_t *root_lock = qdisc_lock(q);
3145 struct sk_buff *to_free = NULL;
3149 qdisc_calculate_pkt_len(skb, q);
3151 * Heuristic to force contended enqueues to serialize on a
3152 * separate lock before trying to get qdisc main lock.
3153 * This permits qdisc->running owner to get the lock more
3154 * often and dequeue packets faster.
3156 contended = qdisc_is_running(q);
3157 if (unlikely(contended))
3158 spin_lock(&q->busylock);
3160 spin_lock(root_lock);
3161 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3162 __qdisc_drop(skb, &to_free);
3164 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
3165 qdisc_run_begin(q)) {
3167 * This is a work-conserving queue; there are no old skbs
3168 * waiting to be sent out; and the qdisc is not running -
3169 * xmit the skb directly.
3172 qdisc_bstats_update(q, skb);
3174 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) {
3175 if (unlikely(contended)) {
3176 spin_unlock(&q->busylock);
3183 rc = NET_XMIT_SUCCESS;
3185 rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
3186 if (qdisc_run_begin(q)) {
3187 if (unlikely(contended)) {
3188 spin_unlock(&q->busylock);
3194 spin_unlock(root_lock);
3195 if (unlikely(to_free))
3196 kfree_skb_list(to_free);
3197 if (unlikely(contended))
3198 spin_unlock(&q->busylock);
3202 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO)
3203 static void skb_update_prio(struct sk_buff *skb)
3205 struct netprio_map *map = rcu_dereference_bh(skb->dev->priomap);
3207 if (!skb->priority && skb->sk && map) {
3208 unsigned int prioidx =
3209 sock_cgroup_prioidx(&skb->sk->sk_cgrp_data);
3211 if (prioidx < map->priomap_len)
3212 skb->priority = map->priomap[prioidx];
3216 #define skb_update_prio(skb)
3219 DEFINE_PER_CPU(int, xmit_recursion);
3220 EXPORT_SYMBOL(xmit_recursion);
3223 * dev_loopback_xmit - loop back @skb
3224 * @net: network namespace this loopback is happening in
3225 * @sk: sk needed to be a netfilter okfn
3226 * @skb: buffer to transmit
3228 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb)
3230 skb_reset_mac_header(skb);
3231 __skb_pull(skb, skb_network_offset(skb));
3232 skb->pkt_type = PACKET_LOOPBACK;
3233 skb->ip_summed = CHECKSUM_UNNECESSARY;
3234 WARN_ON(!skb_dst(skb));
3239 EXPORT_SYMBOL(dev_loopback_xmit);
3241 #ifdef CONFIG_NET_EGRESS
3242 static struct sk_buff *
3243 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
3245 struct tcf_proto *cl = rcu_dereference_bh(dev->egress_cl_list);
3246 struct tcf_result cl_res;
3251 /* qdisc_skb_cb(skb)->pkt_len was already set by the caller. */
3252 qdisc_bstats_cpu_update(cl->q, skb);
3254 switch (tcf_classify(skb, cl, &cl_res, false)) {
3256 case TC_ACT_RECLASSIFY:
3257 skb->tc_index = TC_H_MIN(cl_res.classid);
3260 qdisc_qstats_cpu_drop(cl->q);
3261 *ret = NET_XMIT_DROP;
3267 *ret = NET_XMIT_SUCCESS;
3270 case TC_ACT_REDIRECT:
3271 /* No need to push/pop skb's mac_header here on egress! */
3272 skb_do_redirect(skb);
3273 *ret = NET_XMIT_SUCCESS;
3281 #endif /* CONFIG_NET_EGRESS */
3283 static inline int get_xps_queue(struct net_device *dev, struct sk_buff *skb)
3286 struct xps_dev_maps *dev_maps;
3287 struct xps_map *map;
3288 int queue_index = -1;
3291 dev_maps = rcu_dereference(dev->xps_maps);
3293 unsigned int tci = skb->sender_cpu - 1;
3297 tci += netdev_get_prio_tc_map(dev, skb->priority);
3300 map = rcu_dereference(dev_maps->cpu_map[tci]);
3303 queue_index = map->queues[0];
3305 queue_index = map->queues[reciprocal_scale(skb_get_hash(skb),
3307 if (unlikely(queue_index >= dev->real_num_tx_queues))
3319 static u16 __netdev_pick_tx(struct net_device *dev, struct sk_buff *skb)
3321 struct sock *sk = skb->sk;
3322 int queue_index = sk_tx_queue_get(sk);
3324 if (queue_index < 0 || skb->ooo_okay ||
3325 queue_index >= dev->real_num_tx_queues) {
3326 int new_index = get_xps_queue(dev, skb);
3329 new_index = skb_tx_hash(dev, skb);
3331 if (queue_index != new_index && sk &&
3333 rcu_access_pointer(sk->sk_dst_cache))
3334 sk_tx_queue_set(sk, new_index);
3336 queue_index = new_index;
3342 struct netdev_queue *netdev_pick_tx(struct net_device *dev,
3343 struct sk_buff *skb,
3346 int queue_index = 0;
3349 u32 sender_cpu = skb->sender_cpu - 1;
3351 if (sender_cpu >= (u32)NR_CPUS)
3352 skb->sender_cpu = raw_smp_processor_id() + 1;
3355 if (dev->real_num_tx_queues != 1) {
3356 const struct net_device_ops *ops = dev->netdev_ops;
3358 if (ops->ndo_select_queue)
3359 queue_index = ops->ndo_select_queue(dev, skb, accel_priv,
3362 queue_index = __netdev_pick_tx(dev, skb);
3365 queue_index = netdev_cap_txqueue(dev, queue_index);
3368 skb_set_queue_mapping(skb, queue_index);
3369 return netdev_get_tx_queue(dev, queue_index);
3373 * __dev_queue_xmit - transmit a buffer
3374 * @skb: buffer to transmit
3375 * @accel_priv: private data used for L2 forwarding offload
3377 * Queue a buffer for transmission to a network device. The caller must
3378 * have set the device and priority and built the buffer before calling
3379 * this function. The function can be called from an interrupt.
3381 * A negative errno code is returned on a failure. A success does not
3382 * guarantee the frame will be transmitted as it may be dropped due
3383 * to congestion or traffic shaping.
3385 * -----------------------------------------------------------------------------------
3386 * I notice this method can also return errors from the queue disciplines,
3387 * including NET_XMIT_DROP, which is a positive value. So, errors can also
3390 * Regardless of the return value, the skb is consumed, so it is currently
3391 * difficult to retry a send to this method. (You can bump the ref count
3392 * before sending to hold a reference for retry if you are careful.)
3394 * When calling this method, interrupts MUST be enabled. This is because
3395 * the BH enable code must have IRQs enabled so that it will not deadlock.
3398 static int __dev_queue_xmit(struct sk_buff *skb, void *accel_priv)
3400 struct net_device *dev = skb->dev;
3401 struct netdev_queue *txq;
3405 skb_reset_mac_header(skb);
3407 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP))
3408 __skb_tstamp_tx(skb, NULL, skb->sk, SCM_TSTAMP_SCHED);
3410 /* Disable soft irqs for various locks below. Also
3411 * stops preemption for RCU.
3415 skb_update_prio(skb);
3417 qdisc_pkt_len_init(skb);
3418 #ifdef CONFIG_NET_CLS_ACT
3419 skb->tc_at_ingress = 0;
3420 # ifdef CONFIG_NET_EGRESS
3421 if (static_key_false(&egress_needed)) {
3422 skb = sch_handle_egress(skb, &rc, dev);
3428 /* If device/qdisc don't need skb->dst, release it right now while
3429 * its hot in this cpu cache.
3431 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
3436 txq = netdev_pick_tx(dev, skb, accel_priv);
3437 q = rcu_dereference_bh(txq->qdisc);
3439 trace_net_dev_queue(skb);
3441 rc = __dev_xmit_skb(skb, q, dev, txq);
3445 /* The device has no queue. Common case for software devices:
3446 * loopback, all the sorts of tunnels...
3448 * Really, it is unlikely that netif_tx_lock protection is necessary
3449 * here. (f.e. loopback and IP tunnels are clean ignoring statistics
3451 * However, it is possible, that they rely on protection
3454 * Check this and shot the lock. It is not prone from deadlocks.
3455 *Either shot noqueue qdisc, it is even simpler 8)
3457 if (dev->flags & IFF_UP) {
3458 int cpu = smp_processor_id(); /* ok because BHs are off */
3460 if (txq->xmit_lock_owner != cpu) {
3461 if (unlikely(__this_cpu_read(xmit_recursion) >
3462 XMIT_RECURSION_LIMIT))
3463 goto recursion_alert;
3465 skb = validate_xmit_skb(skb, dev);
3469 HARD_TX_LOCK(dev, txq, cpu);
3471 if (!netif_xmit_stopped(txq)) {
3472 __this_cpu_inc(xmit_recursion);
3473 skb = dev_hard_start_xmit(skb, dev, txq, &rc);
3474 __this_cpu_dec(xmit_recursion);
3475 if (dev_xmit_complete(rc)) {
3476 HARD_TX_UNLOCK(dev, txq);
3480 HARD_TX_UNLOCK(dev, txq);
3481 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
3484 /* Recursion is detected! It is possible,
3488 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
3494 rcu_read_unlock_bh();
3496 atomic_long_inc(&dev->tx_dropped);
3497 kfree_skb_list(skb);
3500 rcu_read_unlock_bh();
3504 int dev_queue_xmit(struct sk_buff *skb)
3506 return __dev_queue_xmit(skb, NULL);
3508 EXPORT_SYMBOL(dev_queue_xmit);
3510 int dev_queue_xmit_accel(struct sk_buff *skb, void *accel_priv)
3512 return __dev_queue_xmit(skb, accel_priv);
3514 EXPORT_SYMBOL(dev_queue_xmit_accel);
3517 /*************************************************************************
3519 *************************************************************************/
3521 int netdev_max_backlog __read_mostly = 1000;
3522 EXPORT_SYMBOL(netdev_max_backlog);
3524 int netdev_tstamp_prequeue __read_mostly = 1;
3525 int netdev_budget __read_mostly = 300;
3526 unsigned int __read_mostly netdev_budget_usecs = 2000;
3527 int weight_p __read_mostly = 64; /* old backlog weight */
3528 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */
3529 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */
3530 int dev_rx_weight __read_mostly = 64;
3531 int dev_tx_weight __read_mostly = 64;
3533 /* Called with irq disabled */
3534 static inline void ____napi_schedule(struct softnet_data *sd,
3535 struct napi_struct *napi)
3537 list_add_tail(&napi->poll_list, &sd->poll_list);
3538 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3543 /* One global table that all flow-based protocols share. */
3544 struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
3545 EXPORT_SYMBOL(rps_sock_flow_table);
3546 u32 rps_cpu_mask __read_mostly;
3547 EXPORT_SYMBOL(rps_cpu_mask);
3549 struct static_key rps_needed __read_mostly;
3550 EXPORT_SYMBOL(rps_needed);
3551 struct static_key rfs_needed __read_mostly;
3552 EXPORT_SYMBOL(rfs_needed);
3554 static struct rps_dev_flow *
3555 set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3556 struct rps_dev_flow *rflow, u16 next_cpu)
3558 if (next_cpu < nr_cpu_ids) {
3559 #ifdef CONFIG_RFS_ACCEL
3560 struct netdev_rx_queue *rxqueue;
3561 struct rps_dev_flow_table *flow_table;
3562 struct rps_dev_flow *old_rflow;
3567 /* Should we steer this flow to a different hardware queue? */
3568 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
3569 !(dev->features & NETIF_F_NTUPLE))
3571 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
3572 if (rxq_index == skb_get_rx_queue(skb))
3575 rxqueue = dev->_rx + rxq_index;
3576 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3579 flow_id = skb_get_hash(skb) & flow_table->mask;
3580 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
3581 rxq_index, flow_id);
3585 rflow = &flow_table->flows[flow_id];
3587 if (old_rflow->filter == rflow->filter)
3588 old_rflow->filter = RPS_NO_FILTER;
3592 per_cpu(softnet_data, next_cpu).input_queue_head;
3595 rflow->cpu = next_cpu;
3600 * get_rps_cpu is called from netif_receive_skb and returns the target
3601 * CPU from the RPS map of the receiving queue for a given skb.
3602 * rcu_read_lock must be held on entry.
3604 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3605 struct rps_dev_flow **rflowp)
3607 const struct rps_sock_flow_table *sock_flow_table;
3608 struct netdev_rx_queue *rxqueue = dev->_rx;
3609 struct rps_dev_flow_table *flow_table;
3610 struct rps_map *map;
3615 if (skb_rx_queue_recorded(skb)) {
3616 u16 index = skb_get_rx_queue(skb);
3618 if (unlikely(index >= dev->real_num_rx_queues)) {
3619 WARN_ONCE(dev->real_num_rx_queues > 1,
3620 "%s received packet on queue %u, but number "
3621 "of RX queues is %u\n",
3622 dev->name, index, dev->real_num_rx_queues);
3628 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */
3630 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3631 map = rcu_dereference(rxqueue->rps_map);
3632 if (!flow_table && !map)
3635 skb_reset_network_header(skb);
3636 hash = skb_get_hash(skb);
3640 sock_flow_table = rcu_dereference(rps_sock_flow_table);
3641 if (flow_table && sock_flow_table) {
3642 struct rps_dev_flow *rflow;
3646 /* First check into global flow table if there is a match */
3647 ident = sock_flow_table->ents[hash & sock_flow_table->mask];
3648 if ((ident ^ hash) & ~rps_cpu_mask)
3651 next_cpu = ident & rps_cpu_mask;
3653 /* OK, now we know there is a match,
3654 * we can look at the local (per receive queue) flow table
3656 rflow = &flow_table->flows[hash & flow_table->mask];
3660 * If the desired CPU (where last recvmsg was done) is
3661 * different from current CPU (one in the rx-queue flow
3662 * table entry), switch if one of the following holds:
3663 * - Current CPU is unset (>= nr_cpu_ids).
3664 * - Current CPU is offline.
3665 * - The current CPU's queue tail has advanced beyond the
3666 * last packet that was enqueued using this table entry.
3667 * This guarantees that all previous packets for the flow
3668 * have been dequeued, thus preserving in order delivery.
3670 if (unlikely(tcpu != next_cpu) &&
3671 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) ||
3672 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
3673 rflow->last_qtail)) >= 0)) {
3675 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
3678 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) {
3688 tcpu = map->cpus[reciprocal_scale(hash, map->len)];
3689 if (cpu_online(tcpu)) {
3699 #ifdef CONFIG_RFS_ACCEL
3702 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
3703 * @dev: Device on which the filter was set
3704 * @rxq_index: RX queue index
3705 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
3706 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
3708 * Drivers that implement ndo_rx_flow_steer() should periodically call
3709 * this function for each installed filter and remove the filters for
3710 * which it returns %true.
3712 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
3713 u32 flow_id, u16 filter_id)
3715 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
3716 struct rps_dev_flow_table *flow_table;
3717 struct rps_dev_flow *rflow;
3722 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3723 if (flow_table && flow_id <= flow_table->mask) {
3724 rflow = &flow_table->flows[flow_id];
3725 cpu = ACCESS_ONCE(rflow->cpu);
3726 if (rflow->filter == filter_id && cpu < nr_cpu_ids &&
3727 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
3728 rflow->last_qtail) <
3729 (int)(10 * flow_table->mask)))
3735 EXPORT_SYMBOL(rps_may_expire_flow);
3737 #endif /* CONFIG_RFS_ACCEL */
3739 /* Called from hardirq (IPI) context */
3740 static void rps_trigger_softirq(void *data)
3742 struct softnet_data *sd = data;
3744 ____napi_schedule(sd, &sd->backlog);
3748 #endif /* CONFIG_RPS */
3751 * Check if this softnet_data structure is another cpu one
3752 * If yes, queue it to our IPI list and return 1
3755 static int rps_ipi_queued(struct softnet_data *sd)
3758 struct softnet_data *mysd = this_cpu_ptr(&softnet_data);
3761 sd->rps_ipi_next = mysd->rps_ipi_list;
3762 mysd->rps_ipi_list = sd;
3764 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3767 #endif /* CONFIG_RPS */
3771 #ifdef CONFIG_NET_FLOW_LIMIT
3772 int netdev_flow_limit_table_len __read_mostly = (1 << 12);
3775 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
3777 #ifdef CONFIG_NET_FLOW_LIMIT
3778 struct sd_flow_limit *fl;
3779 struct softnet_data *sd;
3780 unsigned int old_flow, new_flow;
3782 if (qlen < (netdev_max_backlog >> 1))
3785 sd = this_cpu_ptr(&softnet_data);
3788 fl = rcu_dereference(sd->flow_limit);
3790 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1);
3791 old_flow = fl->history[fl->history_head];
3792 fl->history[fl->history_head] = new_flow;
3795 fl->history_head &= FLOW_LIMIT_HISTORY - 1;
3797 if (likely(fl->buckets[old_flow]))
3798 fl->buckets[old_flow]--;
3800 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) {
3812 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
3813 * queue (may be a remote CPU queue).
3815 static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
3816 unsigned int *qtail)
3818 struct softnet_data *sd;
3819 unsigned long flags;
3822 sd = &per_cpu(softnet_data, cpu);
3824 local_irq_save(flags);
3827 if (!netif_running(skb->dev))
3829 qlen = skb_queue_len(&sd->input_pkt_queue);
3830 if (qlen <= netdev_max_backlog && !skb_flow_limit(skb, qlen)) {
3833 __skb_queue_tail(&sd->input_pkt_queue, skb);
3834 input_queue_tail_incr_save(sd, qtail);
3836 local_irq_restore(flags);
3837 return NET_RX_SUCCESS;
3840 /* Schedule NAPI for backlog device
3841 * We can use non atomic operation since we own the queue lock
3843 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
3844 if (!rps_ipi_queued(sd))
3845 ____napi_schedule(sd, &sd->backlog);
3854 local_irq_restore(flags);
3856 atomic_long_inc(&skb->dev->rx_dropped);
3861 static u32 netif_receive_generic_xdp(struct sk_buff *skb,
3862 struct bpf_prog *xdp_prog)
3864 struct xdp_buff xdp;
3870 /* Reinjected packets coming from act_mirred or similar should
3871 * not get XDP generic processing.
3873 if (skb_cloned(skb))
3876 if (skb_linearize(skb))
3879 /* The XDP program wants to see the packet starting at the MAC
3882 mac_len = skb->data - skb_mac_header(skb);
3883 hlen = skb_headlen(skb) + mac_len;
3884 xdp.data = skb->data - mac_len;
3885 xdp.data_end = xdp.data + hlen;
3886 xdp.data_hard_start = skb->data - skb_headroom(skb);
3887 orig_data = xdp.data;
3889 act = bpf_prog_run_xdp(xdp_prog, &xdp);
3891 off = xdp.data - orig_data;
3893 __skb_pull(skb, off);
3895 __skb_push(skb, -off);
3900 __skb_push(skb, mac_len);
3906 bpf_warn_invalid_xdp_action(act);
3909 trace_xdp_exception(skb->dev, xdp_prog, act);
3920 /* When doing generic XDP we have to bypass the qdisc layer and the
3921 * network taps in order to match in-driver-XDP behavior.
3923 static void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog)
3925 struct net_device *dev = skb->dev;
3926 struct netdev_queue *txq;
3927 bool free_skb = true;
3930 txq = netdev_pick_tx(dev, skb, NULL);
3931 cpu = smp_processor_id();
3932 HARD_TX_LOCK(dev, txq, cpu);
3933 if (!netif_xmit_stopped(txq)) {
3934 rc = netdev_start_xmit(skb, dev, txq, 0);
3935 if (dev_xmit_complete(rc))
3938 HARD_TX_UNLOCK(dev, txq);
3940 trace_xdp_exception(dev, xdp_prog, XDP_TX);
3945 static struct static_key generic_xdp_needed __read_mostly;
3947 static int do_xdp_generic(struct sk_buff *skb)
3949 struct bpf_prog *xdp_prog = rcu_dereference(skb->dev->xdp_prog);
3952 u32 act = netif_receive_generic_xdp(skb, xdp_prog);
3955 if (act != XDP_PASS) {
3958 err = xdp_do_generic_redirect(skb->dev, skb);
3961 /* fallthru to submit skb */
3963 generic_xdp_tx(skb, xdp_prog);
3971 trace_xdp_exception(skb->dev, xdp_prog, XDP_REDIRECT);
3976 static int netif_rx_internal(struct sk_buff *skb)
3980 net_timestamp_check(netdev_tstamp_prequeue, skb);
3982 trace_netif_rx(skb);
3984 if (static_key_false(&generic_xdp_needed)) {
3985 int ret = do_xdp_generic(skb);
3987 /* Consider XDP consuming the packet a success from
3988 * the netdev point of view we do not want to count
3991 if (ret != XDP_PASS)
3992 return NET_RX_SUCCESS;
3996 if (static_key_false(&rps_needed)) {
3997 struct rps_dev_flow voidflow, *rflow = &voidflow;
4003 cpu = get_rps_cpu(skb->dev, skb, &rflow);
4005 cpu = smp_processor_id();
4007 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4016 ret = enqueue_to_backlog(skb, get_cpu(), &qtail);
4023 * netif_rx - post buffer to the network code
4024 * @skb: buffer to post
4026 * This function receives a packet from a device driver and queues it for
4027 * the upper (protocol) levels to process. It always succeeds. The buffer
4028 * may be dropped during processing for congestion control or by the
4032 * NET_RX_SUCCESS (no congestion)
4033 * NET_RX_DROP (packet was dropped)
4037 int netif_rx(struct sk_buff *skb)
4039 trace_netif_rx_entry(skb);
4041 return netif_rx_internal(skb);
4043 EXPORT_SYMBOL(netif_rx);
4045 int netif_rx_ni(struct sk_buff *skb)
4049 trace_netif_rx_ni_entry(skb);
4052 err = netif_rx_internal(skb);
4053 if (local_softirq_pending())
4059 EXPORT_SYMBOL(netif_rx_ni);
4061 static __latent_entropy void net_tx_action(struct softirq_action *h)
4063 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
4065 if (sd->completion_queue) {
4066 struct sk_buff *clist;
4068 local_irq_disable();
4069 clist = sd->completion_queue;
4070 sd->completion_queue = NULL;
4074 struct sk_buff *skb = clist;
4076 clist = clist->next;
4078 WARN_ON(refcount_read(&skb->users));
4079 if (likely(get_kfree_skb_cb(skb)->reason == SKB_REASON_CONSUMED))
4080 trace_consume_skb(skb);
4082 trace_kfree_skb(skb, net_tx_action);
4084 if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
4087 __kfree_skb_defer(skb);
4090 __kfree_skb_flush();
4093 if (sd->output_queue) {
4096 local_irq_disable();
4097 head = sd->output_queue;
4098 sd->output_queue = NULL;
4099 sd->output_queue_tailp = &sd->output_queue;
4103 struct Qdisc *q = head;
4104 spinlock_t *root_lock;
4106 head = head->next_sched;
4108 root_lock = qdisc_lock(q);
4109 spin_lock(root_lock);
4110 /* We need to make sure head->next_sched is read
4111 * before clearing __QDISC_STATE_SCHED
4113 smp_mb__before_atomic();
4114 clear_bit(__QDISC_STATE_SCHED, &q->state);
4116 spin_unlock(root_lock);
4121 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE)
4122 /* This hook is defined here for ATM LANE */
4123 int (*br_fdb_test_addr_hook)(struct net_device *dev,
4124 unsigned char *addr) __read_mostly;
4125 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook);
4128 static inline struct sk_buff *
4129 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
4130 struct net_device *orig_dev)
4132 #ifdef CONFIG_NET_CLS_ACT
4133 struct tcf_proto *cl = rcu_dereference_bh(skb->dev->ingress_cl_list);
4134 struct tcf_result cl_res;
4136 /* If there's at least one ingress present somewhere (so
4137 * we get here via enabled static key), remaining devices
4138 * that are not configured with an ingress qdisc will bail
4144 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4148 qdisc_skb_cb(skb)->pkt_len = skb->len;
4149 skb->tc_at_ingress = 1;
4150 qdisc_bstats_cpu_update(cl->q, skb);
4152 switch (tcf_classify(skb, cl, &cl_res, false)) {
4154 case TC_ACT_RECLASSIFY:
4155 skb->tc_index = TC_H_MIN(cl_res.classid);
4158 qdisc_qstats_cpu_drop(cl->q);
4166 case TC_ACT_REDIRECT:
4167 /* skb_mac_header check was done by cls/act_bpf, so
4168 * we can safely push the L2 header back before
4169 * redirecting to another netdev
4171 __skb_push(skb, skb->mac_len);
4172 skb_do_redirect(skb);
4177 #endif /* CONFIG_NET_CLS_ACT */
4182 * netdev_is_rx_handler_busy - check if receive handler is registered
4183 * @dev: device to check
4185 * Check if a receive handler is already registered for a given device.
4186 * Return true if there one.
4188 * The caller must hold the rtnl_mutex.
4190 bool netdev_is_rx_handler_busy(struct net_device *dev)
4193 return dev && rtnl_dereference(dev->rx_handler);
4195 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy);
4198 * netdev_rx_handler_register - register receive handler
4199 * @dev: device to register a handler for
4200 * @rx_handler: receive handler to register
4201 * @rx_handler_data: data pointer that is used by rx handler
4203 * Register a receive handler for a device. This handler will then be
4204 * called from __netif_receive_skb. A negative errno code is returned
4207 * The caller must hold the rtnl_mutex.
4209 * For a general description of rx_handler, see enum rx_handler_result.
4211 int netdev_rx_handler_register(struct net_device *dev,
4212 rx_handler_func_t *rx_handler,
4213 void *rx_handler_data)
4215 if (netdev_is_rx_handler_busy(dev))
4218 /* Note: rx_handler_data must be set before rx_handler */
4219 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data);
4220 rcu_assign_pointer(dev->rx_handler, rx_handler);
4224 EXPORT_SYMBOL_GPL(netdev_rx_handler_register);
4227 * netdev_rx_handler_unregister - unregister receive handler
4228 * @dev: device to unregister a handler from
4230 * Unregister a receive handler from a device.
4232 * The caller must hold the rtnl_mutex.
4234 void netdev_rx_handler_unregister(struct net_device *dev)
4238 RCU_INIT_POINTER(dev->rx_handler, NULL);
4239 /* a reader seeing a non NULL rx_handler in a rcu_read_lock()
4240 * section has a guarantee to see a non NULL rx_handler_data
4244 RCU_INIT_POINTER(dev->rx_handler_data, NULL);
4246 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister);
4249 * Limit the use of PFMEMALLOC reserves to those protocols that implement
4250 * the special handling of PFMEMALLOC skbs.
4252 static bool skb_pfmemalloc_protocol(struct sk_buff *skb)
4254 switch (skb->protocol) {
4255 case htons(ETH_P_ARP):
4256 case htons(ETH_P_IP):
4257 case htons(ETH_P_IPV6):
4258 case htons(ETH_P_8021Q):
4259 case htons(ETH_P_8021AD):
4266 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev,
4267 int *ret, struct net_device *orig_dev)
4269 #ifdef CONFIG_NETFILTER_INGRESS
4270 if (nf_hook_ingress_active(skb)) {
4274 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4279 ingress_retval = nf_hook_ingress(skb);
4281 return ingress_retval;
4283 #endif /* CONFIG_NETFILTER_INGRESS */
4287 static int __netif_receive_skb_core(struct sk_buff *skb, bool pfmemalloc)
4289 struct packet_type *ptype, *pt_prev;
4290 rx_handler_func_t *rx_handler;
4291 struct net_device *orig_dev;
4292 bool deliver_exact = false;
4293 int ret = NET_RX_DROP;
4296 net_timestamp_check(!netdev_tstamp_prequeue, skb);
4298 trace_netif_receive_skb(skb);
4300 orig_dev = skb->dev;
4302 skb_reset_network_header(skb);
4303 if (!skb_transport_header_was_set(skb))
4304 skb_reset_transport_header(skb);
4305 skb_reset_mac_len(skb);
4310 skb->skb_iif = skb->dev->ifindex;
4312 __this_cpu_inc(softnet_data.processed);
4314 if (skb->protocol == cpu_to_be16(ETH_P_8021Q) ||
4315 skb->protocol == cpu_to_be16(ETH_P_8021AD)) {
4316 skb = skb_vlan_untag(skb);
4321 if (skb_skip_tc_classify(skb))
4327 list_for_each_entry_rcu(ptype, &ptype_all, list) {
4329 ret = deliver_skb(skb, pt_prev, orig_dev);
4333 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) {
4335 ret = deliver_skb(skb, pt_prev, orig_dev);
4340 #ifdef CONFIG_NET_INGRESS
4341 if (static_key_false(&ingress_needed)) {
4342 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev);
4346 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0)
4352 if (pfmemalloc && !skb_pfmemalloc_protocol(skb))
4355 if (skb_vlan_tag_present(skb)) {
4357 ret = deliver_skb(skb, pt_prev, orig_dev);
4360 if (vlan_do_receive(&skb))
4362 else if (unlikely(!skb))
4366 rx_handler = rcu_dereference(skb->dev->rx_handler);
4369 ret = deliver_skb(skb, pt_prev, orig_dev);
4372 switch (rx_handler(&skb)) {
4373 case RX_HANDLER_CONSUMED:
4374 ret = NET_RX_SUCCESS;
4376 case RX_HANDLER_ANOTHER:
4378 case RX_HANDLER_EXACT:
4379 deliver_exact = true;
4380 case RX_HANDLER_PASS:
4387 if (unlikely(skb_vlan_tag_present(skb))) {
4388 if (skb_vlan_tag_get_id(skb))
4389 skb->pkt_type = PACKET_OTHERHOST;
4390 /* Note: we might in the future use prio bits
4391 * and set skb->priority like in vlan_do_receive()
4392 * For the time being, just ignore Priority Code Point
4397 type = skb->protocol;
4399 /* deliver only exact match when indicated */
4400 if (likely(!deliver_exact)) {
4401 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4402 &ptype_base[ntohs(type) &
4406 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4407 &orig_dev->ptype_specific);
4409 if (unlikely(skb->dev != orig_dev)) {
4410 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4411 &skb->dev->ptype_specific);
4415 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
4418 ret = pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
4422 atomic_long_inc(&skb->dev->rx_dropped);
4424 atomic_long_inc(&skb->dev->rx_nohandler);
4426 /* Jamal, now you will not able to escape explaining
4427 * me how you were going to use this. :-)
4436 static int __netif_receive_skb(struct sk_buff *skb)
4440 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) {
4441 unsigned int noreclaim_flag;
4444 * PFMEMALLOC skbs are special, they should
4445 * - be delivered to SOCK_MEMALLOC sockets only
4446 * - stay away from userspace
4447 * - have bounded memory usage
4449 * Use PF_MEMALLOC as this saves us from propagating the allocation
4450 * context down to all allocation sites.
4452 noreclaim_flag = memalloc_noreclaim_save();
4453 ret = __netif_receive_skb_core(skb, true);
4454 memalloc_noreclaim_restore(noreclaim_flag);
4456 ret = __netif_receive_skb_core(skb, false);
4461 static int generic_xdp_install(struct net_device *dev, struct netdev_xdp *xdp)
4463 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog);
4464 struct bpf_prog *new = xdp->prog;
4467 switch (xdp->command) {
4468 case XDP_SETUP_PROG:
4469 rcu_assign_pointer(dev->xdp_prog, new);
4474 static_key_slow_dec(&generic_xdp_needed);
4475 } else if (new && !old) {
4476 static_key_slow_inc(&generic_xdp_needed);
4477 dev_disable_lro(dev);
4481 case XDP_QUERY_PROG:
4482 xdp->prog_attached = !!old;
4483 xdp->prog_id = old ? old->aux->id : 0;
4494 static int netif_receive_skb_internal(struct sk_buff *skb)
4498 net_timestamp_check(netdev_tstamp_prequeue, skb);
4500 if (skb_defer_rx_timestamp(skb))
4501 return NET_RX_SUCCESS;
4505 if (static_key_false(&generic_xdp_needed)) {
4506 int ret = do_xdp_generic(skb);
4508 if (ret != XDP_PASS) {
4515 if (static_key_false(&rps_needed)) {
4516 struct rps_dev_flow voidflow, *rflow = &voidflow;
4517 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
4520 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4526 ret = __netif_receive_skb(skb);
4532 * netif_receive_skb - process receive buffer from network
4533 * @skb: buffer to process
4535 * netif_receive_skb() is the main receive data processing function.
4536 * It always succeeds. The buffer may be dropped during processing
4537 * for congestion control or by the protocol layers.
4539 * This function may only be called from softirq context and interrupts
4540 * should be enabled.
4542 * Return values (usually ignored):
4543 * NET_RX_SUCCESS: no congestion
4544 * NET_RX_DROP: packet was dropped
4546 int netif_receive_skb(struct sk_buff *skb)
4548 trace_netif_receive_skb_entry(skb);
4550 return netif_receive_skb_internal(skb);
4552 EXPORT_SYMBOL(netif_receive_skb);
4554 DEFINE_PER_CPU(struct work_struct, flush_works);
4556 /* Network device is going away, flush any packets still pending */
4557 static void flush_backlog(struct work_struct *work)
4559 struct sk_buff *skb, *tmp;
4560 struct softnet_data *sd;
4563 sd = this_cpu_ptr(&softnet_data);
4565 local_irq_disable();
4567 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) {
4568 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
4569 __skb_unlink(skb, &sd->input_pkt_queue);
4571 input_queue_head_incr(sd);
4577 skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
4578 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
4579 __skb_unlink(skb, &sd->process_queue);
4581 input_queue_head_incr(sd);
4587 static void flush_all_backlogs(void)
4593 for_each_online_cpu(cpu)
4594 queue_work_on(cpu, system_highpri_wq,
4595 per_cpu_ptr(&flush_works, cpu));
4597 for_each_online_cpu(cpu)
4598 flush_work(per_cpu_ptr(&flush_works, cpu));
4603 static int napi_gro_complete(struct sk_buff *skb)
4605 struct packet_offload *ptype;
4606 __be16 type = skb->protocol;
4607 struct list_head *head = &offload_base;
4610 BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
4612 if (NAPI_GRO_CB(skb)->count == 1) {
4613 skb_shinfo(skb)->gso_size = 0;
4618 list_for_each_entry_rcu(ptype, head, list) {
4619 if (ptype->type != type || !ptype->callbacks.gro_complete)
4622 err = ptype->callbacks.gro_complete(skb, 0);
4628 WARN_ON(&ptype->list == head);
4630 return NET_RX_SUCCESS;
4634 return netif_receive_skb_internal(skb);
4637 /* napi->gro_list contains packets ordered by age.
4638 * youngest packets at the head of it.
4639 * Complete skbs in reverse order to reduce latencies.
4641 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
4643 struct sk_buff *skb, *prev = NULL;
4645 /* scan list and build reverse chain */
4646 for (skb = napi->gro_list; skb != NULL; skb = skb->next) {
4651 for (skb = prev; skb; skb = prev) {
4654 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
4658 napi_gro_complete(skb);
4662 napi->gro_list = NULL;
4664 EXPORT_SYMBOL(napi_gro_flush);
4666 static void gro_list_prepare(struct napi_struct *napi, struct sk_buff *skb)
4669 unsigned int maclen = skb->dev->hard_header_len;
4670 u32 hash = skb_get_hash_raw(skb);
4672 for (p = napi->gro_list; p; p = p->next) {
4673 unsigned long diffs;
4675 NAPI_GRO_CB(p)->flush = 0;
4677 if (hash != skb_get_hash_raw(p)) {
4678 NAPI_GRO_CB(p)->same_flow = 0;
4682 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
4683 diffs |= p->vlan_tci ^ skb->vlan_tci;
4684 diffs |= skb_metadata_dst_cmp(p, skb);
4685 if (maclen == ETH_HLEN)
4686 diffs |= compare_ether_header(skb_mac_header(p),
4687 skb_mac_header(skb));
4689 diffs = memcmp(skb_mac_header(p),
4690 skb_mac_header(skb),
4692 NAPI_GRO_CB(p)->same_flow = !diffs;
4696 static void skb_gro_reset_offset(struct sk_buff *skb)
4698 const struct skb_shared_info *pinfo = skb_shinfo(skb);
4699 const skb_frag_t *frag0 = &pinfo->frags[0];
4701 NAPI_GRO_CB(skb)->data_offset = 0;
4702 NAPI_GRO_CB(skb)->frag0 = NULL;
4703 NAPI_GRO_CB(skb)->frag0_len = 0;
4705 if (skb_mac_header(skb) == skb_tail_pointer(skb) &&
4707 !PageHighMem(skb_frag_page(frag0))) {
4708 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
4709 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
4710 skb_frag_size(frag0),
4711 skb->end - skb->tail);
4715 static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
4717 struct skb_shared_info *pinfo = skb_shinfo(skb);
4719 BUG_ON(skb->end - skb->tail < grow);
4721 memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
4723 skb->data_len -= grow;
4726 pinfo->frags[0].page_offset += grow;
4727 skb_frag_size_sub(&pinfo->frags[0], grow);
4729 if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
4730 skb_frag_unref(skb, 0);
4731 memmove(pinfo->frags, pinfo->frags + 1,
4732 --pinfo->nr_frags * sizeof(pinfo->frags[0]));
4736 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
4738 struct sk_buff **pp = NULL;
4739 struct packet_offload *ptype;
4740 __be16 type = skb->protocol;
4741 struct list_head *head = &offload_base;
4743 enum gro_result ret;
4746 if (netif_elide_gro(skb->dev))
4749 gro_list_prepare(napi, skb);
4752 list_for_each_entry_rcu(ptype, head, list) {
4753 if (ptype->type != type || !ptype->callbacks.gro_receive)
4756 skb_set_network_header(skb, skb_gro_offset(skb));
4757 skb_reset_mac_len(skb);
4758 NAPI_GRO_CB(skb)->same_flow = 0;
4759 NAPI_GRO_CB(skb)->flush = skb_is_gso(skb) || skb_has_frag_list(skb);
4760 NAPI_GRO_CB(skb)->free = 0;
4761 NAPI_GRO_CB(skb)->encap_mark = 0;
4762 NAPI_GRO_CB(skb)->recursion_counter = 0;
4763 NAPI_GRO_CB(skb)->is_fou = 0;
4764 NAPI_GRO_CB(skb)->is_atomic = 1;
4765 NAPI_GRO_CB(skb)->gro_remcsum_start = 0;
4767 /* Setup for GRO checksum validation */
4768 switch (skb->ip_summed) {
4769 case CHECKSUM_COMPLETE:
4770 NAPI_GRO_CB(skb)->csum = skb->csum;
4771 NAPI_GRO_CB(skb)->csum_valid = 1;
4772 NAPI_GRO_CB(skb)->csum_cnt = 0;
4774 case CHECKSUM_UNNECESSARY:
4775 NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
4776 NAPI_GRO_CB(skb)->csum_valid = 0;
4779 NAPI_GRO_CB(skb)->csum_cnt = 0;
4780 NAPI_GRO_CB(skb)->csum_valid = 0;
4783 pp = ptype->callbacks.gro_receive(&napi->gro_list, skb);
4788 if (&ptype->list == head)
4791 if (IS_ERR(pp) && PTR_ERR(pp) == -EINPROGRESS) {
4796 same_flow = NAPI_GRO_CB(skb)->same_flow;
4797 ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
4800 struct sk_buff *nskb = *pp;
4804 napi_gro_complete(nskb);
4811 if (NAPI_GRO_CB(skb)->flush)
4814 if (unlikely(napi->gro_count >= MAX_GRO_SKBS)) {
4815 struct sk_buff *nskb = napi->gro_list;
4817 /* locate the end of the list to select the 'oldest' flow */
4818 while (nskb->next) {
4824 napi_gro_complete(nskb);
4828 NAPI_GRO_CB(skb)->count = 1;
4829 NAPI_GRO_CB(skb)->age = jiffies;
4830 NAPI_GRO_CB(skb)->last = skb;
4831 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
4832 skb->next = napi->gro_list;
4833 napi->gro_list = skb;
4837 grow = skb_gro_offset(skb) - skb_headlen(skb);
4839 gro_pull_from_frag0(skb, grow);
4848 struct packet_offload *gro_find_receive_by_type(__be16 type)
4850 struct list_head *offload_head = &offload_base;
4851 struct packet_offload *ptype;
4853 list_for_each_entry_rcu(ptype, offload_head, list) {
4854 if (ptype->type != type || !ptype->callbacks.gro_receive)
4860 EXPORT_SYMBOL(gro_find_receive_by_type);
4862 struct packet_offload *gro_find_complete_by_type(__be16 type)
4864 struct list_head *offload_head = &offload_base;
4865 struct packet_offload *ptype;
4867 list_for_each_entry_rcu(ptype, offload_head, list) {
4868 if (ptype->type != type || !ptype->callbacks.gro_complete)
4874 EXPORT_SYMBOL(gro_find_complete_by_type);
4876 static void napi_skb_free_stolen_head(struct sk_buff *skb)
4880 kmem_cache_free(skbuff_head_cache, skb);
4883 static gro_result_t napi_skb_finish(gro_result_t ret, struct sk_buff *skb)
4887 if (netif_receive_skb_internal(skb))
4895 case GRO_MERGED_FREE:
4896 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
4897 napi_skb_free_stolen_head(skb);
4911 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
4913 skb_mark_napi_id(skb, napi);
4914 trace_napi_gro_receive_entry(skb);
4916 skb_gro_reset_offset(skb);
4918 return napi_skb_finish(dev_gro_receive(napi, skb), skb);
4920 EXPORT_SYMBOL(napi_gro_receive);
4922 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
4924 if (unlikely(skb->pfmemalloc)) {
4928 __skb_pull(skb, skb_headlen(skb));
4929 /* restore the reserve we had after netdev_alloc_skb_ip_align() */
4930 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
4932 skb->dev = napi->dev;
4934 skb->encapsulation = 0;
4935 skb_shinfo(skb)->gso_type = 0;
4936 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
4942 struct sk_buff *napi_get_frags(struct napi_struct *napi)
4944 struct sk_buff *skb = napi->skb;
4947 skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
4950 skb_mark_napi_id(skb, napi);
4955 EXPORT_SYMBOL(napi_get_frags);
4957 static gro_result_t napi_frags_finish(struct napi_struct *napi,
4958 struct sk_buff *skb,
4964 __skb_push(skb, ETH_HLEN);
4965 skb->protocol = eth_type_trans(skb, skb->dev);
4966 if (ret == GRO_NORMAL && netif_receive_skb_internal(skb))
4971 napi_reuse_skb(napi, skb);
4974 case GRO_MERGED_FREE:
4975 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
4976 napi_skb_free_stolen_head(skb);
4978 napi_reuse_skb(napi, skb);
4989 /* Upper GRO stack assumes network header starts at gro_offset=0
4990 * Drivers could call both napi_gro_frags() and napi_gro_receive()
4991 * We copy ethernet header into skb->data to have a common layout.
4993 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
4995 struct sk_buff *skb = napi->skb;
4996 const struct ethhdr *eth;
4997 unsigned int hlen = sizeof(*eth);
5001 skb_reset_mac_header(skb);
5002 skb_gro_reset_offset(skb);
5004 eth = skb_gro_header_fast(skb, 0);
5005 if (unlikely(skb_gro_header_hard(skb, hlen))) {
5006 eth = skb_gro_header_slow(skb, hlen, 0);
5007 if (unlikely(!eth)) {
5008 net_warn_ratelimited("%s: dropping impossible skb from %s\n",
5009 __func__, napi->dev->name);
5010 napi_reuse_skb(napi, skb);
5014 gro_pull_from_frag0(skb, hlen);
5015 NAPI_GRO_CB(skb)->frag0 += hlen;
5016 NAPI_GRO_CB(skb)->frag0_len -= hlen;
5018 __skb_pull(skb, hlen);
5021 * This works because the only protocols we care about don't require
5023 * We'll fix it up properly in napi_frags_finish()
5025 skb->protocol = eth->h_proto;
5030 gro_result_t napi_gro_frags(struct napi_struct *napi)
5032 struct sk_buff *skb = napi_frags_skb(napi);
5037 trace_napi_gro_frags_entry(skb);
5039 return napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
5041 EXPORT_SYMBOL(napi_gro_frags);
5043 /* Compute the checksum from gro_offset and return the folded value
5044 * after adding in any pseudo checksum.
5046 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
5051 wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
5053 /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
5054 sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
5056 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
5057 !skb->csum_complete_sw)
5058 netdev_rx_csum_fault(skb->dev);
5061 NAPI_GRO_CB(skb)->csum = wsum;
5062 NAPI_GRO_CB(skb)->csum_valid = 1;
5066 EXPORT_SYMBOL(__skb_gro_checksum_complete);
5068 static void net_rps_send_ipi(struct softnet_data *remsd)
5072 struct softnet_data *next = remsd->rps_ipi_next;
5074 if (cpu_online(remsd->cpu))
5075 smp_call_function_single_async(remsd->cpu, &remsd->csd);
5082 * net_rps_action_and_irq_enable sends any pending IPI's for rps.
5083 * Note: called with local irq disabled, but exits with local irq enabled.
5085 static void net_rps_action_and_irq_enable(struct softnet_data *sd)
5088 struct softnet_data *remsd = sd->rps_ipi_list;
5091 sd->rps_ipi_list = NULL;
5095 /* Send pending IPI's to kick RPS processing on remote cpus. */
5096 net_rps_send_ipi(remsd);
5102 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd)
5105 return sd->rps_ipi_list != NULL;
5111 static int process_backlog(struct napi_struct *napi, int quota)
5113 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
5117 /* Check if we have pending ipi, its better to send them now,
5118 * not waiting net_rx_action() end.
5120 if (sd_has_rps_ipi_waiting(sd)) {
5121 local_irq_disable();
5122 net_rps_action_and_irq_enable(sd);
5125 napi->weight = dev_rx_weight;
5127 struct sk_buff *skb;
5129 while ((skb = __skb_dequeue(&sd->process_queue))) {
5131 __netif_receive_skb(skb);
5133 input_queue_head_incr(sd);
5134 if (++work >= quota)
5139 local_irq_disable();
5141 if (skb_queue_empty(&sd->input_pkt_queue)) {
5143 * Inline a custom version of __napi_complete().
5144 * only current cpu owns and manipulates this napi,
5145 * and NAPI_STATE_SCHED is the only possible flag set
5147 * We can use a plain write instead of clear_bit(),
5148 * and we dont need an smp_mb() memory barrier.
5153 skb_queue_splice_tail_init(&sd->input_pkt_queue,
5154 &sd->process_queue);
5164 * __napi_schedule - schedule for receive
5165 * @n: entry to schedule
5167 * The entry's receive function will be scheduled to run.
5168 * Consider using __napi_schedule_irqoff() if hard irqs are masked.
5170 void __napi_schedule(struct napi_struct *n)
5172 unsigned long flags;
5174 local_irq_save(flags);
5175 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
5176 local_irq_restore(flags);
5178 EXPORT_SYMBOL(__napi_schedule);
5181 * napi_schedule_prep - check if napi can be scheduled
5184 * Test if NAPI routine is already running, and if not mark
5185 * it as running. This is used as a condition variable
5186 * insure only one NAPI poll instance runs. We also make
5187 * sure there is no pending NAPI disable.
5189 bool napi_schedule_prep(struct napi_struct *n)
5191 unsigned long val, new;
5194 val = READ_ONCE(n->state);
5195 if (unlikely(val & NAPIF_STATE_DISABLE))
5197 new = val | NAPIF_STATE_SCHED;
5199 /* Sets STATE_MISSED bit if STATE_SCHED was already set
5200 * This was suggested by Alexander Duyck, as compiler
5201 * emits better code than :
5202 * if (val & NAPIF_STATE_SCHED)
5203 * new |= NAPIF_STATE_MISSED;
5205 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED *
5207 } while (cmpxchg(&n->state, val, new) != val);
5209 return !(val & NAPIF_STATE_SCHED);
5211 EXPORT_SYMBOL(napi_schedule_prep);
5214 * __napi_schedule_irqoff - schedule for receive
5215 * @n: entry to schedule
5217 * Variant of __napi_schedule() assuming hard irqs are masked
5219 void __napi_schedule_irqoff(struct napi_struct *n)
5221 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
5223 EXPORT_SYMBOL(__napi_schedule_irqoff);
5225 bool napi_complete_done(struct napi_struct *n, int work_done)
5227 unsigned long flags, val, new;
5230 * 1) Don't let napi dequeue from the cpu poll list
5231 * just in case its running on a different cpu.
5232 * 2) If we are busy polling, do nothing here, we have
5233 * the guarantee we will be called later.
5235 if (unlikely(n->state & (NAPIF_STATE_NPSVC |
5236 NAPIF_STATE_IN_BUSY_POLL)))
5240 unsigned long timeout = 0;
5243 timeout = n->dev->gro_flush_timeout;
5246 hrtimer_start(&n->timer, ns_to_ktime(timeout),
5247 HRTIMER_MODE_REL_PINNED);
5249 napi_gro_flush(n, false);
5251 if (unlikely(!list_empty(&n->poll_list))) {
5252 /* If n->poll_list is not empty, we need to mask irqs */
5253 local_irq_save(flags);
5254 list_del_init(&n->poll_list);
5255 local_irq_restore(flags);
5259 val = READ_ONCE(n->state);
5261 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED));
5263 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED);
5265 /* If STATE_MISSED was set, leave STATE_SCHED set,
5266 * because we will call napi->poll() one more time.
5267 * This C code was suggested by Alexander Duyck to help gcc.
5269 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED *
5271 } while (cmpxchg(&n->state, val, new) != val);
5273 if (unlikely(val & NAPIF_STATE_MISSED)) {
5280 EXPORT_SYMBOL(napi_complete_done);
5282 /* must be called under rcu_read_lock(), as we dont take a reference */
5283 static struct napi_struct *napi_by_id(unsigned int napi_id)
5285 unsigned int hash = napi_id % HASH_SIZE(napi_hash);
5286 struct napi_struct *napi;
5288 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node)
5289 if (napi->napi_id == napi_id)
5295 #if defined(CONFIG_NET_RX_BUSY_POLL)
5297 #define BUSY_POLL_BUDGET 8
5299 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock)
5303 /* Busy polling means there is a high chance device driver hard irq
5304 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was
5305 * set in napi_schedule_prep().
5306 * Since we are about to call napi->poll() once more, we can safely
5307 * clear NAPI_STATE_MISSED.
5309 * Note: x86 could use a single "lock and ..." instruction
5310 * to perform these two clear_bit()
5312 clear_bit(NAPI_STATE_MISSED, &napi->state);
5313 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state);
5317 /* All we really want here is to re-enable device interrupts.
5318 * Ideally, a new ndo_busy_poll_stop() could avoid another round.
5320 rc = napi->poll(napi, BUSY_POLL_BUDGET);
5321 netpoll_poll_unlock(have_poll_lock);
5322 if (rc == BUSY_POLL_BUDGET)
5323 __napi_schedule(napi);
5327 void napi_busy_loop(unsigned int napi_id,
5328 bool (*loop_end)(void *, unsigned long),
5331 unsigned long start_time = loop_end ? busy_loop_current_time() : 0;
5332 int (*napi_poll)(struct napi_struct *napi, int budget);
5333 void *have_poll_lock = NULL;
5334 struct napi_struct *napi;
5341 napi = napi_by_id(napi_id);
5351 unsigned long val = READ_ONCE(napi->state);
5353 /* If multiple threads are competing for this napi,
5354 * we avoid dirtying napi->state as much as we can.
5356 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED |
5357 NAPIF_STATE_IN_BUSY_POLL))
5359 if (cmpxchg(&napi->state, val,
5360 val | NAPIF_STATE_IN_BUSY_POLL |
5361 NAPIF_STATE_SCHED) != val)
5363 have_poll_lock = netpoll_poll_lock(napi);
5364 napi_poll = napi->poll;
5366 work = napi_poll(napi, BUSY_POLL_BUDGET);
5367 trace_napi_poll(napi, work, BUSY_POLL_BUDGET);
5370 __NET_ADD_STATS(dev_net(napi->dev),
5371 LINUX_MIB_BUSYPOLLRXPACKETS, work);
5374 if (!loop_end || loop_end(loop_end_arg, start_time))
5377 if (unlikely(need_resched())) {
5379 busy_poll_stop(napi, have_poll_lock);
5383 if (loop_end(loop_end_arg, start_time))
5390 busy_poll_stop(napi, have_poll_lock);
5395 EXPORT_SYMBOL(napi_busy_loop);
5397 #endif /* CONFIG_NET_RX_BUSY_POLL */
5399 static void napi_hash_add(struct napi_struct *napi)
5401 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state) ||
5402 test_and_set_bit(NAPI_STATE_HASHED, &napi->state))
5405 spin_lock(&napi_hash_lock);
5407 /* 0..NR_CPUS range is reserved for sender_cpu use */
5409 if (unlikely(++napi_gen_id < MIN_NAPI_ID))
5410 napi_gen_id = MIN_NAPI_ID;
5411 } while (napi_by_id(napi_gen_id));
5412 napi->napi_id = napi_gen_id;
5414 hlist_add_head_rcu(&napi->napi_hash_node,
5415 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]);
5417 spin_unlock(&napi_hash_lock);
5420 /* Warning : caller is responsible to make sure rcu grace period
5421 * is respected before freeing memory containing @napi
5423 bool napi_hash_del(struct napi_struct *napi)
5425 bool rcu_sync_needed = false;
5427 spin_lock(&napi_hash_lock);
5429 if (test_and_clear_bit(NAPI_STATE_HASHED, &napi->state)) {
5430 rcu_sync_needed = true;
5431 hlist_del_rcu(&napi->napi_hash_node);
5433 spin_unlock(&napi_hash_lock);
5434 return rcu_sync_needed;
5436 EXPORT_SYMBOL_GPL(napi_hash_del);
5438 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer)
5440 struct napi_struct *napi;
5442 napi = container_of(timer, struct napi_struct, timer);
5444 /* Note : we use a relaxed variant of napi_schedule_prep() not setting
5445 * NAPI_STATE_MISSED, since we do not react to a device IRQ.
5447 if (napi->gro_list && !napi_disable_pending(napi) &&
5448 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state))
5449 __napi_schedule_irqoff(napi);
5451 return HRTIMER_NORESTART;
5454 void netif_napi_add(struct net_device *dev, struct napi_struct *napi,
5455 int (*poll)(struct napi_struct *, int), int weight)
5457 INIT_LIST_HEAD(&napi->poll_list);
5458 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
5459 napi->timer.function = napi_watchdog;
5460 napi->gro_count = 0;
5461 napi->gro_list = NULL;
5464 if (weight > NAPI_POLL_WEIGHT)
5465 pr_err_once("netif_napi_add() called with weight %d on device %s\n",
5467 napi->weight = weight;
5468 list_add(&napi->dev_list, &dev->napi_list);
5470 #ifdef CONFIG_NETPOLL
5471 napi->poll_owner = -1;
5473 set_bit(NAPI_STATE_SCHED, &napi->state);
5474 napi_hash_add(napi);
5476 EXPORT_SYMBOL(netif_napi_add);
5478 void napi_disable(struct napi_struct *n)
5481 set_bit(NAPI_STATE_DISABLE, &n->state);
5483 while (test_and_set_bit(NAPI_STATE_SCHED, &n->state))
5485 while (test_and_set_bit(NAPI_STATE_NPSVC, &n->state))
5488 hrtimer_cancel(&n->timer);
5490 clear_bit(NAPI_STATE_DISABLE, &n->state);
5492 EXPORT_SYMBOL(napi_disable);
5494 /* Must be called in process context */
5495 void netif_napi_del(struct napi_struct *napi)
5498 if (napi_hash_del(napi))
5500 list_del_init(&napi->dev_list);
5501 napi_free_frags(napi);
5503 kfree_skb_list(napi->gro_list);
5504 napi->gro_list = NULL;
5505 napi->gro_count = 0;
5507 EXPORT_SYMBOL(netif_napi_del);
5509 static int napi_poll(struct napi_struct *n, struct list_head *repoll)
5514 list_del_init(&n->poll_list);
5516 have = netpoll_poll_lock(n);
5520 /* This NAPI_STATE_SCHED test is for avoiding a race
5521 * with netpoll's poll_napi(). Only the entity which
5522 * obtains the lock and sees NAPI_STATE_SCHED set will
5523 * actually make the ->poll() call. Therefore we avoid
5524 * accidentally calling ->poll() when NAPI is not scheduled.
5527 if (test_bit(NAPI_STATE_SCHED, &n->state)) {
5528 work = n->poll(n, weight);
5529 trace_napi_poll(n, work, weight);
5532 WARN_ON_ONCE(work > weight);
5534 if (likely(work < weight))
5537 /* Drivers must not modify the NAPI state if they
5538 * consume the entire weight. In such cases this code
5539 * still "owns" the NAPI instance and therefore can
5540 * move the instance around on the list at-will.
5542 if (unlikely(napi_disable_pending(n))) {
5548 /* flush too old packets
5549 * If HZ < 1000, flush all packets.
5551 napi_gro_flush(n, HZ >= 1000);
5554 /* Some drivers may have called napi_schedule
5555 * prior to exhausting their budget.
5557 if (unlikely(!list_empty(&n->poll_list))) {
5558 pr_warn_once("%s: Budget exhausted after napi rescheduled\n",
5559 n->dev ? n->dev->name : "backlog");
5563 list_add_tail(&n->poll_list, repoll);
5566 netpoll_poll_unlock(have);
5571 static __latent_entropy void net_rx_action(struct softirq_action *h)
5573 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
5574 unsigned long time_limit = jiffies +
5575 usecs_to_jiffies(netdev_budget_usecs);
5576 int budget = netdev_budget;
5580 local_irq_disable();
5581 list_splice_init(&sd->poll_list, &list);
5585 struct napi_struct *n;
5587 if (list_empty(&list)) {
5588 if (!sd_has_rps_ipi_waiting(sd) && list_empty(&repoll))
5593 n = list_first_entry(&list, struct napi_struct, poll_list);
5594 budget -= napi_poll(n, &repoll);
5596 /* If softirq window is exhausted then punt.
5597 * Allow this to run for 2 jiffies since which will allow
5598 * an average latency of 1.5/HZ.
5600 if (unlikely(budget <= 0 ||
5601 time_after_eq(jiffies, time_limit))) {
5607 local_irq_disable();
5609 list_splice_tail_init(&sd->poll_list, &list);
5610 list_splice_tail(&repoll, &list);
5611 list_splice(&list, &sd->poll_list);
5612 if (!list_empty(&sd->poll_list))
5613 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
5615 net_rps_action_and_irq_enable(sd);
5617 __kfree_skb_flush();
5620 struct netdev_adjacent {
5621 struct net_device *dev;
5623 /* upper master flag, there can only be one master device per list */
5626 /* counter for the number of times this device was added to us */
5629 /* private field for the users */
5632 struct list_head list;
5633 struct rcu_head rcu;
5636 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev,
5637 struct list_head *adj_list)
5639 struct netdev_adjacent *adj;
5641 list_for_each_entry(adj, adj_list, list) {
5642 if (adj->dev == adj_dev)
5648 static int __netdev_has_upper_dev(struct net_device *upper_dev, void *data)
5650 struct net_device *dev = data;
5652 return upper_dev == dev;
5656 * netdev_has_upper_dev - Check if device is linked to an upper device
5658 * @upper_dev: upper device to check
5660 * Find out if a device is linked to specified upper device and return true
5661 * in case it is. Note that this checks only immediate upper device,
5662 * not through a complete stack of devices. The caller must hold the RTNL lock.
5664 bool netdev_has_upper_dev(struct net_device *dev,
5665 struct net_device *upper_dev)
5669 return netdev_walk_all_upper_dev_rcu(dev, __netdev_has_upper_dev,
5672 EXPORT_SYMBOL(netdev_has_upper_dev);
5675 * netdev_has_upper_dev_all - Check if device is linked to an upper device
5677 * @upper_dev: upper device to check
5679 * Find out if a device is linked to specified upper device and return true
5680 * in case it is. Note that this checks the entire upper device chain.
5681 * The caller must hold rcu lock.
5684 bool netdev_has_upper_dev_all_rcu(struct net_device *dev,
5685 struct net_device *upper_dev)
5687 return !!netdev_walk_all_upper_dev_rcu(dev, __netdev_has_upper_dev,
5690 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu);
5693 * netdev_has_any_upper_dev - Check if device is linked to some device
5696 * Find out if a device is linked to an upper device and return true in case
5697 * it is. The caller must hold the RTNL lock.
5699 static bool netdev_has_any_upper_dev(struct net_device *dev)
5703 return !list_empty(&dev->adj_list.upper);
5707 * netdev_master_upper_dev_get - Get master upper device
5710 * Find a master upper device and return pointer to it or NULL in case
5711 * it's not there. The caller must hold the RTNL lock.
5713 struct net_device *netdev_master_upper_dev_get(struct net_device *dev)
5715 struct netdev_adjacent *upper;
5719 if (list_empty(&dev->adj_list.upper))
5722 upper = list_first_entry(&dev->adj_list.upper,
5723 struct netdev_adjacent, list);
5724 if (likely(upper->master))
5728 EXPORT_SYMBOL(netdev_master_upper_dev_get);
5731 * netdev_has_any_lower_dev - Check if device is linked to some device
5734 * Find out if a device is linked to a lower device and return true in case
5735 * it is. The caller must hold the RTNL lock.
5737 static bool netdev_has_any_lower_dev(struct net_device *dev)
5741 return !list_empty(&dev->adj_list.lower);
5744 void *netdev_adjacent_get_private(struct list_head *adj_list)
5746 struct netdev_adjacent *adj;
5748 adj = list_entry(adj_list, struct netdev_adjacent, list);
5750 return adj->private;
5752 EXPORT_SYMBOL(netdev_adjacent_get_private);
5755 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list
5757 * @iter: list_head ** of the current position
5759 * Gets the next device from the dev's upper list, starting from iter
5760 * position. The caller must hold RCU read lock.
5762 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev,
5763 struct list_head **iter)
5765 struct netdev_adjacent *upper;
5767 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
5769 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
5771 if (&upper->list == &dev->adj_list.upper)
5774 *iter = &upper->list;
5778 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu);
5780 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev,
5781 struct list_head **iter)
5783 struct netdev_adjacent *upper;
5785 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
5787 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
5789 if (&upper->list == &dev->adj_list.upper)
5792 *iter = &upper->list;
5797 int netdev_walk_all_upper_dev_rcu(struct net_device *dev,
5798 int (*fn)(struct net_device *dev,
5802 struct net_device *udev;
5803 struct list_head *iter;
5806 for (iter = &dev->adj_list.upper,
5807 udev = netdev_next_upper_dev_rcu(dev, &iter);
5809 udev = netdev_next_upper_dev_rcu(dev, &iter)) {
5810 /* first is the upper device itself */
5811 ret = fn(udev, data);
5815 /* then look at all of its upper devices */
5816 ret = netdev_walk_all_upper_dev_rcu(udev, fn, data);
5823 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu);
5826 * netdev_lower_get_next_private - Get the next ->private from the
5827 * lower neighbour list
5829 * @iter: list_head ** of the current position
5831 * Gets the next netdev_adjacent->private from the dev's lower neighbour
5832 * list, starting from iter position. The caller must hold either hold the
5833 * RTNL lock or its own locking that guarantees that the neighbour lower
5834 * list will remain unchanged.
5836 void *netdev_lower_get_next_private(struct net_device *dev,
5837 struct list_head **iter)
5839 struct netdev_adjacent *lower;
5841 lower = list_entry(*iter, struct netdev_adjacent, list);
5843 if (&lower->list == &dev->adj_list.lower)
5846 *iter = lower->list.next;
5848 return lower->private;
5850 EXPORT_SYMBOL(netdev_lower_get_next_private);
5853 * netdev_lower_get_next_private_rcu - Get the next ->private from the
5854 * lower neighbour list, RCU
5857 * @iter: list_head ** of the current position
5859 * Gets the next netdev_adjacent->private from the dev's lower neighbour
5860 * list, starting from iter position. The caller must hold RCU read lock.
5862 void *netdev_lower_get_next_private_rcu(struct net_device *dev,
5863 struct list_head **iter)
5865 struct netdev_adjacent *lower;
5867 WARN_ON_ONCE(!rcu_read_lock_held());
5869 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
5871 if (&lower->list == &dev->adj_list.lower)
5874 *iter = &lower->list;
5876 return lower->private;
5878 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu);
5881 * netdev_lower_get_next - Get the next device from the lower neighbour
5884 * @iter: list_head ** of the current position
5886 * Gets the next netdev_adjacent from the dev's lower neighbour
5887 * list, starting from iter position. The caller must hold RTNL lock or
5888 * its own locking that guarantees that the neighbour lower
5889 * list will remain unchanged.
5891 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter)
5893 struct netdev_adjacent *lower;
5895 lower = list_entry(*iter, struct netdev_adjacent, list);
5897 if (&lower->list == &dev->adj_list.lower)
5900 *iter = lower->list.next;
5904 EXPORT_SYMBOL(netdev_lower_get_next);
5906 static struct net_device *netdev_next_lower_dev(struct net_device *dev,
5907 struct list_head **iter)
5909 struct netdev_adjacent *lower;
5911 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
5913 if (&lower->list == &dev->adj_list.lower)
5916 *iter = &lower->list;
5921 int netdev_walk_all_lower_dev(struct net_device *dev,
5922 int (*fn)(struct net_device *dev,
5926 struct net_device *ldev;
5927 struct list_head *iter;
5930 for (iter = &dev->adj_list.lower,
5931 ldev = netdev_next_lower_dev(dev, &iter);
5933 ldev = netdev_next_lower_dev(dev, &iter)) {
5934 /* first is the lower device itself */
5935 ret = fn(ldev, data);
5939 /* then look at all of its lower devices */
5940 ret = netdev_walk_all_lower_dev(ldev, fn, data);
5947 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev);
5949 static struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev,
5950 struct list_head **iter)
5952 struct netdev_adjacent *lower;
5954 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
5955 if (&lower->list == &dev->adj_list.lower)
5958 *iter = &lower->list;
5963 int netdev_walk_all_lower_dev_rcu(struct net_device *dev,
5964 int (*fn)(struct net_device *dev,
5968 struct net_device *ldev;
5969 struct list_head *iter;
5972 for (iter = &dev->adj_list.lower,
5973 ldev = netdev_next_lower_dev_rcu(dev, &iter);
5975 ldev = netdev_next_lower_dev_rcu(dev, &iter)) {
5976 /* first is the lower device itself */
5977 ret = fn(ldev, data);
5981 /* then look at all of its lower devices */
5982 ret = netdev_walk_all_lower_dev_rcu(ldev, fn, data);
5989 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu);
5992 * netdev_lower_get_first_private_rcu - Get the first ->private from the
5993 * lower neighbour list, RCU
5997 * Gets the first netdev_adjacent->private from the dev's lower neighbour
5998 * list. The caller must hold RCU read lock.
6000 void *netdev_lower_get_first_private_rcu(struct net_device *dev)
6002 struct netdev_adjacent *lower;
6004 lower = list_first_or_null_rcu(&dev->adj_list.lower,
6005 struct netdev_adjacent, list);
6007 return lower->private;
6010 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu);
6013 * netdev_master_upper_dev_get_rcu - Get master upper device
6016 * Find a master upper device and return pointer to it or NULL in case
6017 * it's not there. The caller must hold the RCU read lock.
6019 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev)
6021 struct netdev_adjacent *upper;
6023 upper = list_first_or_null_rcu(&dev->adj_list.upper,
6024 struct netdev_adjacent, list);
6025 if (upper && likely(upper->master))
6029 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu);
6031 static int netdev_adjacent_sysfs_add(struct net_device *dev,
6032 struct net_device *adj_dev,
6033 struct list_head *dev_list)
6035 char linkname[IFNAMSIZ+7];
6037 sprintf(linkname, dev_list == &dev->adj_list.upper ?
6038 "upper_%s" : "lower_%s", adj_dev->name);
6039 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj),
6042 static void netdev_adjacent_sysfs_del(struct net_device *dev,
6044 struct list_head *dev_list)
6046 char linkname[IFNAMSIZ+7];
6048 sprintf(linkname, dev_list == &dev->adj_list.upper ?
6049 "upper_%s" : "lower_%s", name);
6050 sysfs_remove_link(&(dev->dev.kobj), linkname);
6053 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev,
6054 struct net_device *adj_dev,
6055 struct list_head *dev_list)
6057 return (dev_list == &dev->adj_list.upper ||
6058 dev_list == &dev->adj_list.lower) &&
6059 net_eq(dev_net(dev), dev_net(adj_dev));
6062 static int __netdev_adjacent_dev_insert(struct net_device *dev,
6063 struct net_device *adj_dev,
6064 struct list_head *dev_list,
6065 void *private, bool master)
6067 struct netdev_adjacent *adj;
6070 adj = __netdev_find_adj(adj_dev, dev_list);
6074 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n",
6075 dev->name, adj_dev->name, adj->ref_nr);
6080 adj = kmalloc(sizeof(*adj), GFP_KERNEL);
6085 adj->master = master;
6087 adj->private = private;
6090 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n",
6091 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name);
6093 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) {
6094 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list);
6099 /* Ensure that master link is always the first item in list. */
6101 ret = sysfs_create_link(&(dev->dev.kobj),
6102 &(adj_dev->dev.kobj), "master");
6104 goto remove_symlinks;
6106 list_add_rcu(&adj->list, dev_list);
6108 list_add_tail_rcu(&adj->list, dev_list);
6114 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
6115 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
6123 static void __netdev_adjacent_dev_remove(struct net_device *dev,
6124 struct net_device *adj_dev,
6126 struct list_head *dev_list)
6128 struct netdev_adjacent *adj;
6130 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n",
6131 dev->name, adj_dev->name, ref_nr);
6133 adj = __netdev_find_adj(adj_dev, dev_list);
6136 pr_err("Adjacency does not exist for device %s from %s\n",
6137 dev->name, adj_dev->name);
6142 if (adj->ref_nr > ref_nr) {
6143 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n",
6144 dev->name, adj_dev->name, ref_nr,
6145 adj->ref_nr - ref_nr);
6146 adj->ref_nr -= ref_nr;
6151 sysfs_remove_link(&(dev->dev.kobj), "master");
6153 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
6154 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
6156 list_del_rcu(&adj->list);
6157 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n",
6158 adj_dev->name, dev->name, adj_dev->name);
6160 kfree_rcu(adj, rcu);
6163 static int __netdev_adjacent_dev_link_lists(struct net_device *dev,
6164 struct net_device *upper_dev,
6165 struct list_head *up_list,
6166 struct list_head *down_list,
6167 void *private, bool master)
6171 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list,
6176 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list,
6179 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list);
6186 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev,
6187 struct net_device *upper_dev,
6189 struct list_head *up_list,
6190 struct list_head *down_list)
6192 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list);
6193 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list);
6196 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev,
6197 struct net_device *upper_dev,
6198 void *private, bool master)
6200 return __netdev_adjacent_dev_link_lists(dev, upper_dev,
6201 &dev->adj_list.upper,
6202 &upper_dev->adj_list.lower,
6206 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev,
6207 struct net_device *upper_dev)
6209 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1,
6210 &dev->adj_list.upper,
6211 &upper_dev->adj_list.lower);
6214 static int __netdev_upper_dev_link(struct net_device *dev,
6215 struct net_device *upper_dev, bool master,
6216 void *upper_priv, void *upper_info)
6218 struct netdev_notifier_changeupper_info changeupper_info;
6223 if (dev == upper_dev)
6226 /* To prevent loops, check if dev is not upper device to upper_dev. */
6227 if (netdev_has_upper_dev(upper_dev, dev))
6230 if (netdev_has_upper_dev(dev, upper_dev))
6233 if (master && netdev_master_upper_dev_get(dev))
6236 changeupper_info.upper_dev = upper_dev;
6237 changeupper_info.master = master;
6238 changeupper_info.linking = true;
6239 changeupper_info.upper_info = upper_info;
6241 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, dev,
6242 &changeupper_info.info);
6243 ret = notifier_to_errno(ret);
6247 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv,
6252 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, dev,
6253 &changeupper_info.info);
6254 ret = notifier_to_errno(ret);
6261 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
6267 * netdev_upper_dev_link - Add a link to the upper device
6269 * @upper_dev: new upper device
6271 * Adds a link to device which is upper to this one. The caller must hold
6272 * the RTNL lock. On a failure a negative errno code is returned.
6273 * On success the reference counts are adjusted and the function
6276 int netdev_upper_dev_link(struct net_device *dev,
6277 struct net_device *upper_dev)
6279 return __netdev_upper_dev_link(dev, upper_dev, false, NULL, NULL);
6281 EXPORT_SYMBOL(netdev_upper_dev_link);
6284 * netdev_master_upper_dev_link - Add a master link to the upper device
6286 * @upper_dev: new upper device
6287 * @upper_priv: upper device private
6288 * @upper_info: upper info to be passed down via notifier
6290 * Adds a link to device which is upper to this one. In this case, only
6291 * one master upper device can be linked, although other non-master devices
6292 * might be linked as well. The caller must hold the RTNL lock.
6293 * On a failure a negative errno code is returned. On success the reference
6294 * counts are adjusted and the function returns zero.
6296 int netdev_master_upper_dev_link(struct net_device *dev,
6297 struct net_device *upper_dev,
6298 void *upper_priv, void *upper_info)
6300 return __netdev_upper_dev_link(dev, upper_dev, true,
6301 upper_priv, upper_info);
6303 EXPORT_SYMBOL(netdev_master_upper_dev_link);
6306 * netdev_upper_dev_unlink - Removes a link to upper device
6308 * @upper_dev: new upper device
6310 * Removes a link to device which is upper to this one. The caller must hold
6313 void netdev_upper_dev_unlink(struct net_device *dev,
6314 struct net_device *upper_dev)
6316 struct netdev_notifier_changeupper_info changeupper_info;
6320 changeupper_info.upper_dev = upper_dev;
6321 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev;
6322 changeupper_info.linking = false;
6324 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, dev,
6325 &changeupper_info.info);
6327 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
6329 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, dev,
6330 &changeupper_info.info);
6332 EXPORT_SYMBOL(netdev_upper_dev_unlink);
6335 * netdev_bonding_info_change - Dispatch event about slave change
6337 * @bonding_info: info to dispatch
6339 * Send NETDEV_BONDING_INFO to netdev notifiers with info.
6340 * The caller must hold the RTNL lock.
6342 void netdev_bonding_info_change(struct net_device *dev,
6343 struct netdev_bonding_info *bonding_info)
6345 struct netdev_notifier_bonding_info info;
6347 memcpy(&info.bonding_info, bonding_info,
6348 sizeof(struct netdev_bonding_info));
6349 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, dev,
6352 EXPORT_SYMBOL(netdev_bonding_info_change);
6354 static void netdev_adjacent_add_links(struct net_device *dev)
6356 struct netdev_adjacent *iter;
6358 struct net *net = dev_net(dev);
6360 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6361 if (!net_eq(net, dev_net(iter->dev)))
6363 netdev_adjacent_sysfs_add(iter->dev, dev,
6364 &iter->dev->adj_list.lower);
6365 netdev_adjacent_sysfs_add(dev, iter->dev,
6366 &dev->adj_list.upper);
6369 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6370 if (!net_eq(net, dev_net(iter->dev)))
6372 netdev_adjacent_sysfs_add(iter->dev, dev,
6373 &iter->dev->adj_list.upper);
6374 netdev_adjacent_sysfs_add(dev, iter->dev,
6375 &dev->adj_list.lower);
6379 static void netdev_adjacent_del_links(struct net_device *dev)
6381 struct netdev_adjacent *iter;
6383 struct net *net = dev_net(dev);
6385 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6386 if (!net_eq(net, dev_net(iter->dev)))
6388 netdev_adjacent_sysfs_del(iter->dev, dev->name,
6389 &iter->dev->adj_list.lower);
6390 netdev_adjacent_sysfs_del(dev, iter->dev->name,
6391 &dev->adj_list.upper);
6394 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6395 if (!net_eq(net, dev_net(iter->dev)))
6397 netdev_adjacent_sysfs_del(iter->dev, dev->name,
6398 &iter->dev->adj_list.upper);
6399 netdev_adjacent_sysfs_del(dev, iter->dev->name,
6400 &dev->adj_list.lower);
6404 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname)
6406 struct netdev_adjacent *iter;
6408 struct net *net = dev_net(dev);
6410 list_for_each_entry(iter, &dev->adj_list.upper, list) {
6411 if (!net_eq(net, dev_net(iter->dev)))
6413 netdev_adjacent_sysfs_del(iter->dev, oldname,
6414 &iter->dev->adj_list.lower);
6415 netdev_adjacent_sysfs_add(iter->dev, dev,
6416 &iter->dev->adj_list.lower);
6419 list_for_each_entry(iter, &dev->adj_list.lower, list) {
6420 if (!net_eq(net, dev_net(iter->dev)))
6422 netdev_adjacent_sysfs_del(iter->dev, oldname,
6423 &iter->dev->adj_list.upper);
6424 netdev_adjacent_sysfs_add(iter->dev, dev,
6425 &iter->dev->adj_list.upper);
6429 void *netdev_lower_dev_get_private(struct net_device *dev,
6430 struct net_device *lower_dev)
6432 struct netdev_adjacent *lower;
6436 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower);
6440 return lower->private;
6442 EXPORT_SYMBOL(netdev_lower_dev_get_private);
6445 int dev_get_nest_level(struct net_device *dev)
6447 struct net_device *lower = NULL;
6448 struct list_head *iter;
6454 netdev_for_each_lower_dev(dev, lower, iter) {
6455 nest = dev_get_nest_level(lower);
6456 if (max_nest < nest)
6460 return max_nest + 1;
6462 EXPORT_SYMBOL(dev_get_nest_level);
6465 * netdev_lower_change - Dispatch event about lower device state change
6466 * @lower_dev: device
6467 * @lower_state_info: state to dispatch
6469 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info.
6470 * The caller must hold the RTNL lock.
6472 void netdev_lower_state_changed(struct net_device *lower_dev,
6473 void *lower_state_info)
6475 struct netdev_notifier_changelowerstate_info changelowerstate_info;
6478 changelowerstate_info.lower_state_info = lower_state_info;
6479 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, lower_dev,
6480 &changelowerstate_info.info);
6482 EXPORT_SYMBOL(netdev_lower_state_changed);
6484 static void dev_change_rx_flags(struct net_device *dev, int flags)
6486 const struct net_device_ops *ops = dev->netdev_ops;
6488 if (ops->ndo_change_rx_flags)
6489 ops->ndo_change_rx_flags(dev, flags);
6492 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify)
6494 unsigned int old_flags = dev->flags;
6500 dev->flags |= IFF_PROMISC;
6501 dev->promiscuity += inc;
6502 if (dev->promiscuity == 0) {
6505 * If inc causes overflow, untouch promisc and return error.
6508 dev->flags &= ~IFF_PROMISC;
6510 dev->promiscuity -= inc;
6511 pr_warn("%s: promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n",
6516 if (dev->flags != old_flags) {
6517 pr_info("device %s %s promiscuous mode\n",
6519 dev->flags & IFF_PROMISC ? "entered" : "left");
6520 if (audit_enabled) {
6521 current_uid_gid(&uid, &gid);
6522 audit_log(current->audit_context, GFP_ATOMIC,
6523 AUDIT_ANOM_PROMISCUOUS,
6524 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u",
6525 dev->name, (dev->flags & IFF_PROMISC),
6526 (old_flags & IFF_PROMISC),
6527 from_kuid(&init_user_ns, audit_get_loginuid(current)),
6528 from_kuid(&init_user_ns, uid),
6529 from_kgid(&init_user_ns, gid),
6530 audit_get_sessionid(current));
6533 dev_change_rx_flags(dev, IFF_PROMISC);
6536 __dev_notify_flags(dev, old_flags, IFF_PROMISC);
6541 * dev_set_promiscuity - update promiscuity count on a device
6545 * Add or remove promiscuity from a device. While the count in the device
6546 * remains above zero the interface remains promiscuous. Once it hits zero
6547 * the device reverts back to normal filtering operation. A negative inc
6548 * value is used to drop promiscuity on the device.
6549 * Return 0 if successful or a negative errno code on error.
6551 int dev_set_promiscuity(struct net_device *dev, int inc)
6553 unsigned int old_flags = dev->flags;
6556 err = __dev_set_promiscuity(dev, inc, true);
6559 if (dev->flags != old_flags)
6560 dev_set_rx_mode(dev);
6563 EXPORT_SYMBOL(dev_set_promiscuity);
6565 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify)
6567 unsigned int old_flags = dev->flags, old_gflags = dev->gflags;
6571 dev->flags |= IFF_ALLMULTI;
6572 dev->allmulti += inc;
6573 if (dev->allmulti == 0) {
6576 * If inc causes overflow, untouch allmulti and return error.
6579 dev->flags &= ~IFF_ALLMULTI;
6581 dev->allmulti -= inc;
6582 pr_warn("%s: allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n",
6587 if (dev->flags ^ old_flags) {
6588 dev_change_rx_flags(dev, IFF_ALLMULTI);
6589 dev_set_rx_mode(dev);
6591 __dev_notify_flags(dev, old_flags,
6592 dev->gflags ^ old_gflags);
6598 * dev_set_allmulti - update allmulti count on a device
6602 * Add or remove reception of all multicast frames to a device. While the
6603 * count in the device remains above zero the interface remains listening
6604 * to all interfaces. Once it hits zero the device reverts back to normal
6605 * filtering operation. A negative @inc value is used to drop the counter
6606 * when releasing a resource needing all multicasts.
6607 * Return 0 if successful or a negative errno code on error.
6610 int dev_set_allmulti(struct net_device *dev, int inc)
6612 return __dev_set_allmulti(dev, inc, true);
6614 EXPORT_SYMBOL(dev_set_allmulti);
6617 * Upload unicast and multicast address lists to device and
6618 * configure RX filtering. When the device doesn't support unicast
6619 * filtering it is put in promiscuous mode while unicast addresses
6622 void __dev_set_rx_mode(struct net_device *dev)
6624 const struct net_device_ops *ops = dev->netdev_ops;
6626 /* dev_open will call this function so the list will stay sane. */
6627 if (!(dev->flags&IFF_UP))
6630 if (!netif_device_present(dev))
6633 if (!(dev->priv_flags & IFF_UNICAST_FLT)) {
6634 /* Unicast addresses changes may only happen under the rtnl,
6635 * therefore calling __dev_set_promiscuity here is safe.
6637 if (!netdev_uc_empty(dev) && !dev->uc_promisc) {
6638 __dev_set_promiscuity(dev, 1, false);
6639 dev->uc_promisc = true;
6640 } else if (netdev_uc_empty(dev) && dev->uc_promisc) {
6641 __dev_set_promiscuity(dev, -1, false);
6642 dev->uc_promisc = false;
6646 if (ops->ndo_set_rx_mode)
6647 ops->ndo_set_rx_mode(dev);
6650 void dev_set_rx_mode(struct net_device *dev)
6652 netif_addr_lock_bh(dev);
6653 __dev_set_rx_mode(dev);
6654 netif_addr_unlock_bh(dev);
6658 * dev_get_flags - get flags reported to userspace
6661 * Get the combination of flag bits exported through APIs to userspace.
6663 unsigned int dev_get_flags(const struct net_device *dev)
6667 flags = (dev->flags & ~(IFF_PROMISC |
6672 (dev->gflags & (IFF_PROMISC |
6675 if (netif_running(dev)) {
6676 if (netif_oper_up(dev))
6677 flags |= IFF_RUNNING;
6678 if (netif_carrier_ok(dev))
6679 flags |= IFF_LOWER_UP;
6680 if (netif_dormant(dev))
6681 flags |= IFF_DORMANT;
6686 EXPORT_SYMBOL(dev_get_flags);
6688 int __dev_change_flags(struct net_device *dev, unsigned int flags)
6690 unsigned int old_flags = dev->flags;
6696 * Set the flags on our device.
6699 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP |
6700 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL |
6702 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC |
6706 * Load in the correct multicast list now the flags have changed.
6709 if ((old_flags ^ flags) & IFF_MULTICAST)
6710 dev_change_rx_flags(dev, IFF_MULTICAST);
6712 dev_set_rx_mode(dev);
6715 * Have we downed the interface. We handle IFF_UP ourselves
6716 * according to user attempts to set it, rather than blindly
6721 if ((old_flags ^ flags) & IFF_UP) {
6722 if (old_flags & IFF_UP)
6725 ret = __dev_open(dev);
6728 if ((flags ^ dev->gflags) & IFF_PROMISC) {
6729 int inc = (flags & IFF_PROMISC) ? 1 : -1;
6730 unsigned int old_flags = dev->flags;
6732 dev->gflags ^= IFF_PROMISC;
6734 if (__dev_set_promiscuity(dev, inc, false) >= 0)
6735 if (dev->flags != old_flags)
6736 dev_set_rx_mode(dev);
6739 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI
6740 * is important. Some (broken) drivers set IFF_PROMISC, when
6741 * IFF_ALLMULTI is requested not asking us and not reporting.
6743 if ((flags ^ dev->gflags) & IFF_ALLMULTI) {
6744 int inc = (flags & IFF_ALLMULTI) ? 1 : -1;
6746 dev->gflags ^= IFF_ALLMULTI;
6747 __dev_set_allmulti(dev, inc, false);
6753 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags,
6754 unsigned int gchanges)
6756 unsigned int changes = dev->flags ^ old_flags;
6759 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC);
6761 if (changes & IFF_UP) {
6762 if (dev->flags & IFF_UP)
6763 call_netdevice_notifiers(NETDEV_UP, dev);
6765 call_netdevice_notifiers(NETDEV_DOWN, dev);
6768 if (dev->flags & IFF_UP &&
6769 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) {
6770 struct netdev_notifier_change_info change_info;
6772 change_info.flags_changed = changes;
6773 call_netdevice_notifiers_info(NETDEV_CHANGE, dev,
6779 * dev_change_flags - change device settings
6781 * @flags: device state flags
6783 * Change settings on device based state flags. The flags are
6784 * in the userspace exported format.
6786 int dev_change_flags(struct net_device *dev, unsigned int flags)
6789 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags;
6791 ret = __dev_change_flags(dev, flags);
6795 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags);
6796 __dev_notify_flags(dev, old_flags, changes);
6799 EXPORT_SYMBOL(dev_change_flags);
6801 int __dev_set_mtu(struct net_device *dev, int new_mtu)
6803 const struct net_device_ops *ops = dev->netdev_ops;
6805 if (ops->ndo_change_mtu)
6806 return ops->ndo_change_mtu(dev, new_mtu);
6811 EXPORT_SYMBOL(__dev_set_mtu);
6814 * dev_set_mtu - Change maximum transfer unit
6816 * @new_mtu: new transfer unit
6818 * Change the maximum transfer size of the network device.
6820 int dev_set_mtu(struct net_device *dev, int new_mtu)
6824 if (new_mtu == dev->mtu)
6827 /* MTU must be positive, and in range */
6828 if (new_mtu < 0 || new_mtu < dev->min_mtu) {
6829 net_err_ratelimited("%s: Invalid MTU %d requested, hw min %d\n",
6830 dev->name, new_mtu, dev->min_mtu);
6834 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) {
6835 net_err_ratelimited("%s: Invalid MTU %d requested, hw max %d\n",
6836 dev->name, new_mtu, dev->max_mtu);
6840 if (!netif_device_present(dev))
6843 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev);
6844 err = notifier_to_errno(err);
6848 orig_mtu = dev->mtu;
6849 err = __dev_set_mtu(dev, new_mtu);
6852 err = call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
6853 err = notifier_to_errno(err);
6855 /* setting mtu back and notifying everyone again,
6856 * so that they have a chance to revert changes.
6858 __dev_set_mtu(dev, orig_mtu);
6859 call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
6864 EXPORT_SYMBOL(dev_set_mtu);
6867 * dev_set_group - Change group this device belongs to
6869 * @new_group: group this device should belong to
6871 void dev_set_group(struct net_device *dev, int new_group)
6873 dev->group = new_group;
6875 EXPORT_SYMBOL(dev_set_group);
6878 * dev_set_mac_address - Change Media Access Control Address
6882 * Change the hardware (MAC) address of the device
6884 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa)
6886 const struct net_device_ops *ops = dev->netdev_ops;
6889 if (!ops->ndo_set_mac_address)
6891 if (sa->sa_family != dev->type)
6893 if (!netif_device_present(dev))
6895 err = ops->ndo_set_mac_address(dev, sa);
6898 dev->addr_assign_type = NET_ADDR_SET;
6899 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
6900 add_device_randomness(dev->dev_addr, dev->addr_len);
6903 EXPORT_SYMBOL(dev_set_mac_address);
6906 * dev_change_carrier - Change device carrier
6908 * @new_carrier: new value
6910 * Change device carrier
6912 int dev_change_carrier(struct net_device *dev, bool new_carrier)
6914 const struct net_device_ops *ops = dev->netdev_ops;
6916 if (!ops->ndo_change_carrier)
6918 if (!netif_device_present(dev))
6920 return ops->ndo_change_carrier(dev, new_carrier);
6922 EXPORT_SYMBOL(dev_change_carrier);
6925 * dev_get_phys_port_id - Get device physical port ID
6929 * Get device physical port ID
6931 int dev_get_phys_port_id(struct net_device *dev,
6932 struct netdev_phys_item_id *ppid)
6934 const struct net_device_ops *ops = dev->netdev_ops;
6936 if (!ops->ndo_get_phys_port_id)
6938 return ops->ndo_get_phys_port_id(dev, ppid);
6940 EXPORT_SYMBOL(dev_get_phys_port_id);
6943 * dev_get_phys_port_name - Get device physical port name
6946 * @len: limit of bytes to copy to name
6948 * Get device physical port name
6950 int dev_get_phys_port_name(struct net_device *dev,
6951 char *name, size_t len)
6953 const struct net_device_ops *ops = dev->netdev_ops;
6955 if (!ops->ndo_get_phys_port_name)
6957 return ops->ndo_get_phys_port_name(dev, name, len);
6959 EXPORT_SYMBOL(dev_get_phys_port_name);
6962 * dev_change_proto_down - update protocol port state information
6964 * @proto_down: new value
6966 * This info can be used by switch drivers to set the phys state of the
6969 int dev_change_proto_down(struct net_device *dev, bool proto_down)
6971 const struct net_device_ops *ops = dev->netdev_ops;
6973 if (!ops->ndo_change_proto_down)
6975 if (!netif_device_present(dev))
6977 return ops->ndo_change_proto_down(dev, proto_down);
6979 EXPORT_SYMBOL(dev_change_proto_down);
6981 u8 __dev_xdp_attached(struct net_device *dev, xdp_op_t xdp_op, u32 *prog_id)
6983 struct netdev_xdp xdp;
6985 memset(&xdp, 0, sizeof(xdp));
6986 xdp.command = XDP_QUERY_PROG;
6988 /* Query must always succeed. */
6989 WARN_ON(xdp_op(dev, &xdp) < 0);
6991 *prog_id = xdp.prog_id;
6993 return xdp.prog_attached;
6996 static int dev_xdp_install(struct net_device *dev, xdp_op_t xdp_op,
6997 struct netlink_ext_ack *extack, u32 flags,
6998 struct bpf_prog *prog)
7000 struct netdev_xdp xdp;
7002 memset(&xdp, 0, sizeof(xdp));
7003 if (flags & XDP_FLAGS_HW_MODE)
7004 xdp.command = XDP_SETUP_PROG_HW;
7006 xdp.command = XDP_SETUP_PROG;
7007 xdp.extack = extack;
7011 return xdp_op(dev, &xdp);
7015 * dev_change_xdp_fd - set or clear a bpf program for a device rx path
7017 * @extack: netlink extended ack
7018 * @fd: new program fd or negative value to clear
7019 * @flags: xdp-related flags
7021 * Set or clear a bpf program for a device
7023 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack,
7026 const struct net_device_ops *ops = dev->netdev_ops;
7027 struct bpf_prog *prog = NULL;
7028 xdp_op_t xdp_op, xdp_chk;
7033 xdp_op = xdp_chk = ops->ndo_xdp;
7034 if (!xdp_op && (flags & (XDP_FLAGS_DRV_MODE | XDP_FLAGS_HW_MODE)))
7036 if (!xdp_op || (flags & XDP_FLAGS_SKB_MODE))
7037 xdp_op = generic_xdp_install;
7038 if (xdp_op == xdp_chk)
7039 xdp_chk = generic_xdp_install;
7042 if (xdp_chk && __dev_xdp_attached(dev, xdp_chk, NULL))
7044 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) &&
7045 __dev_xdp_attached(dev, xdp_op, NULL))
7048 prog = bpf_prog_get_type(fd, BPF_PROG_TYPE_XDP);
7050 return PTR_ERR(prog);
7053 err = dev_xdp_install(dev, xdp_op, extack, flags, prog);
7054 if (err < 0 && prog)
7061 * dev_new_index - allocate an ifindex
7062 * @net: the applicable net namespace
7064 * Returns a suitable unique value for a new device interface
7065 * number. The caller must hold the rtnl semaphore or the
7066 * dev_base_lock to be sure it remains unique.
7068 static int dev_new_index(struct net *net)
7070 int ifindex = net->ifindex;
7075 if (!__dev_get_by_index(net, ifindex))
7076 return net->ifindex = ifindex;
7080 /* Delayed registration/unregisteration */
7081 static LIST_HEAD(net_todo_list);
7082 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq);
7084 static void net_set_todo(struct net_device *dev)
7086 list_add_tail(&dev->todo_list, &net_todo_list);
7087 dev_net(dev)->dev_unreg_count++;
7090 static void rollback_registered_many(struct list_head *head)
7092 struct net_device *dev, *tmp;
7093 LIST_HEAD(close_head);
7095 BUG_ON(dev_boot_phase);
7098 list_for_each_entry_safe(dev, tmp, head, unreg_list) {
7099 /* Some devices call without registering
7100 * for initialization unwind. Remove those
7101 * devices and proceed with the remaining.
7103 if (dev->reg_state == NETREG_UNINITIALIZED) {
7104 pr_debug("unregister_netdevice: device %s/%p never was registered\n",
7108 list_del(&dev->unreg_list);
7111 dev->dismantle = true;
7112 BUG_ON(dev->reg_state != NETREG_REGISTERED);
7115 /* If device is running, close it first. */
7116 list_for_each_entry(dev, head, unreg_list)
7117 list_add_tail(&dev->close_list, &close_head);
7118 dev_close_many(&close_head, true);
7120 list_for_each_entry(dev, head, unreg_list) {
7121 /* And unlink it from device chain. */
7122 unlist_netdevice(dev);
7124 dev->reg_state = NETREG_UNREGISTERING;
7126 flush_all_backlogs();
7130 list_for_each_entry(dev, head, unreg_list) {
7131 struct sk_buff *skb = NULL;
7133 /* Shutdown queueing discipline. */
7137 /* Notify protocols, that we are about to destroy
7138 * this device. They should clean all the things.
7140 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
7142 if (!dev->rtnl_link_ops ||
7143 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
7144 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0,
7148 * Flush the unicast and multicast chains
7153 if (dev->netdev_ops->ndo_uninit)
7154 dev->netdev_ops->ndo_uninit(dev);
7157 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL);
7159 /* Notifier chain MUST detach us all upper devices. */
7160 WARN_ON(netdev_has_any_upper_dev(dev));
7161 WARN_ON(netdev_has_any_lower_dev(dev));
7163 /* Remove entries from kobject tree */
7164 netdev_unregister_kobject(dev);
7166 /* Remove XPS queueing entries */
7167 netif_reset_xps_queues_gt(dev, 0);
7173 list_for_each_entry(dev, head, unreg_list)
7177 static void rollback_registered(struct net_device *dev)
7181 list_add(&dev->unreg_list, &single);
7182 rollback_registered_many(&single);
7186 static netdev_features_t netdev_sync_upper_features(struct net_device *lower,
7187 struct net_device *upper, netdev_features_t features)
7189 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
7190 netdev_features_t feature;
7193 for_each_netdev_feature(&upper_disables, feature_bit) {
7194 feature = __NETIF_F_BIT(feature_bit);
7195 if (!(upper->wanted_features & feature)
7196 && (features & feature)) {
7197 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n",
7198 &feature, upper->name);
7199 features &= ~feature;
7206 static void netdev_sync_lower_features(struct net_device *upper,
7207 struct net_device *lower, netdev_features_t features)
7209 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
7210 netdev_features_t feature;
7213 for_each_netdev_feature(&upper_disables, feature_bit) {
7214 feature = __NETIF_F_BIT(feature_bit);
7215 if (!(features & feature) && (lower->features & feature)) {
7216 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n",
7217 &feature, lower->name);
7218 lower->wanted_features &= ~feature;
7219 netdev_update_features(lower);
7221 if (unlikely(lower->features & feature))
7222 netdev_WARN(upper, "failed to disable %pNF on %s!\n",
7223 &feature, lower->name);
7228 static netdev_features_t netdev_fix_features(struct net_device *dev,
7229 netdev_features_t features)
7231 /* Fix illegal checksum combinations */
7232 if ((features & NETIF_F_HW_CSUM) &&
7233 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
7234 netdev_warn(dev, "mixed HW and IP checksum settings.\n");
7235 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
7238 /* TSO requires that SG is present as well. */
7239 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) {
7240 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n");
7241 features &= ~NETIF_F_ALL_TSO;
7244 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) &&
7245 !(features & NETIF_F_IP_CSUM)) {
7246 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n");
7247 features &= ~NETIF_F_TSO;
7248 features &= ~NETIF_F_TSO_ECN;
7251 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) &&
7252 !(features & NETIF_F_IPV6_CSUM)) {
7253 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n");
7254 features &= ~NETIF_F_TSO6;
7257 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */
7258 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO))
7259 features &= ~NETIF_F_TSO_MANGLEID;
7261 /* TSO ECN requires that TSO is present as well. */
7262 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN)
7263 features &= ~NETIF_F_TSO_ECN;
7265 /* Software GSO depends on SG. */
7266 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) {
7267 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n");
7268 features &= ~NETIF_F_GSO;
7271 /* GSO partial features require GSO partial be set */
7272 if ((features & dev->gso_partial_features) &&
7273 !(features & NETIF_F_GSO_PARTIAL)) {
7275 "Dropping partially supported GSO features since no GSO partial.\n");
7276 features &= ~dev->gso_partial_features;
7282 int __netdev_update_features(struct net_device *dev)
7284 struct net_device *upper, *lower;
7285 netdev_features_t features;
7286 struct list_head *iter;
7291 features = netdev_get_wanted_features(dev);
7293 if (dev->netdev_ops->ndo_fix_features)
7294 features = dev->netdev_ops->ndo_fix_features(dev, features);
7296 /* driver might be less strict about feature dependencies */
7297 features = netdev_fix_features(dev, features);
7299 /* some features can't be enabled if they're off an an upper device */
7300 netdev_for_each_upper_dev_rcu(dev, upper, iter)
7301 features = netdev_sync_upper_features(dev, upper, features);
7303 if (dev->features == features)
7306 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n",
7307 &dev->features, &features);
7309 if (dev->netdev_ops->ndo_set_features)
7310 err = dev->netdev_ops->ndo_set_features(dev, features);
7314 if (unlikely(err < 0)) {
7316 "set_features() failed (%d); wanted %pNF, left %pNF\n",
7317 err, &features, &dev->features);
7318 /* return non-0 since some features might have changed and
7319 * it's better to fire a spurious notification than miss it
7325 /* some features must be disabled on lower devices when disabled
7326 * on an upper device (think: bonding master or bridge)
7328 netdev_for_each_lower_dev(dev, lower, iter)
7329 netdev_sync_lower_features(dev, lower, features);
7332 netdev_features_t diff = features ^ dev->features;
7334 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) {
7335 /* udp_tunnel_{get,drop}_rx_info both need
7336 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the
7337 * device, or they won't do anything.
7338 * Thus we need to update dev->features
7339 * *before* calling udp_tunnel_get_rx_info,
7340 * but *after* calling udp_tunnel_drop_rx_info.
7342 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) {
7343 dev->features = features;
7344 udp_tunnel_get_rx_info(dev);
7346 udp_tunnel_drop_rx_info(dev);
7350 dev->features = features;
7353 return err < 0 ? 0 : 1;
7357 * netdev_update_features - recalculate device features
7358 * @dev: the device to check
7360 * Recalculate dev->features set and send notifications if it
7361 * has changed. Should be called after driver or hardware dependent
7362 * conditions might have changed that influence the features.
7364 void netdev_update_features(struct net_device *dev)
7366 if (__netdev_update_features(dev))
7367 netdev_features_change(dev);
7369 EXPORT_SYMBOL(netdev_update_features);
7372 * netdev_change_features - recalculate device features
7373 * @dev: the device to check
7375 * Recalculate dev->features set and send notifications even
7376 * if they have not changed. Should be called instead of
7377 * netdev_update_features() if also dev->vlan_features might
7378 * have changed to allow the changes to be propagated to stacked
7381 void netdev_change_features(struct net_device *dev)
7383 __netdev_update_features(dev);
7384 netdev_features_change(dev);
7386 EXPORT_SYMBOL(netdev_change_features);
7389 * netif_stacked_transfer_operstate - transfer operstate
7390 * @rootdev: the root or lower level device to transfer state from
7391 * @dev: the device to transfer operstate to
7393 * Transfer operational state from root to device. This is normally
7394 * called when a stacking relationship exists between the root
7395 * device and the device(a leaf device).
7397 void netif_stacked_transfer_operstate(const struct net_device *rootdev,
7398 struct net_device *dev)
7400 if (rootdev->operstate == IF_OPER_DORMANT)
7401 netif_dormant_on(dev);
7403 netif_dormant_off(dev);
7405 if (netif_carrier_ok(rootdev))
7406 netif_carrier_on(dev);
7408 netif_carrier_off(dev);
7410 EXPORT_SYMBOL(netif_stacked_transfer_operstate);
7413 static int netif_alloc_rx_queues(struct net_device *dev)
7415 unsigned int i, count = dev->num_rx_queues;
7416 struct netdev_rx_queue *rx;
7417 size_t sz = count * sizeof(*rx);
7421 rx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
7427 for (i = 0; i < count; i++)
7433 static void netdev_init_one_queue(struct net_device *dev,
7434 struct netdev_queue *queue, void *_unused)
7436 /* Initialize queue lock */
7437 spin_lock_init(&queue->_xmit_lock);
7438 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type);
7439 queue->xmit_lock_owner = -1;
7440 netdev_queue_numa_node_write(queue, NUMA_NO_NODE);
7443 dql_init(&queue->dql, HZ);
7447 static void netif_free_tx_queues(struct net_device *dev)
7452 static int netif_alloc_netdev_queues(struct net_device *dev)
7454 unsigned int count = dev->num_tx_queues;
7455 struct netdev_queue *tx;
7456 size_t sz = count * sizeof(*tx);
7458 if (count < 1 || count > 0xffff)
7461 tx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
7467 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL);
7468 spin_lock_init(&dev->tx_global_lock);
7473 void netif_tx_stop_all_queues(struct net_device *dev)
7477 for (i = 0; i < dev->num_tx_queues; i++) {
7478 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
7480 netif_tx_stop_queue(txq);
7483 EXPORT_SYMBOL(netif_tx_stop_all_queues);
7486 * register_netdevice - register a network device
7487 * @dev: device to register
7489 * Take a completed network device structure and add it to the kernel
7490 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
7491 * chain. 0 is returned on success. A negative errno code is returned
7492 * on a failure to set up the device, or if the name is a duplicate.
7494 * Callers must hold the rtnl semaphore. You may want
7495 * register_netdev() instead of this.
7498 * The locking appears insufficient to guarantee two parallel registers
7499 * will not get the same name.
7502 int register_netdevice(struct net_device *dev)
7505 struct net *net = dev_net(dev);
7507 BUG_ON(dev_boot_phase);
7512 /* When net_device's are persistent, this will be fatal. */
7513 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
7516 spin_lock_init(&dev->addr_list_lock);
7517 netdev_set_addr_lockdep_class(dev);
7519 ret = dev_get_valid_name(net, dev, dev->name);
7523 /* Init, if this function is available */
7524 if (dev->netdev_ops->ndo_init) {
7525 ret = dev->netdev_ops->ndo_init(dev);
7533 if (((dev->hw_features | dev->features) &
7534 NETIF_F_HW_VLAN_CTAG_FILTER) &&
7535 (!dev->netdev_ops->ndo_vlan_rx_add_vid ||
7536 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) {
7537 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n");
7544 dev->ifindex = dev_new_index(net);
7545 else if (__dev_get_by_index(net, dev->ifindex))
7548 /* Transfer changeable features to wanted_features and enable
7549 * software offloads (GSO and GRO).
7551 dev->hw_features |= NETIF_F_SOFT_FEATURES;
7552 dev->features |= NETIF_F_SOFT_FEATURES;
7554 if (dev->netdev_ops->ndo_udp_tunnel_add) {
7555 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT;
7556 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT;
7559 dev->wanted_features = dev->features & dev->hw_features;
7561 if (!(dev->flags & IFF_LOOPBACK))
7562 dev->hw_features |= NETIF_F_NOCACHE_COPY;
7564 /* If IPv4 TCP segmentation offload is supported we should also
7565 * allow the device to enable segmenting the frame with the option
7566 * of ignoring a static IP ID value. This doesn't enable the
7567 * feature itself but allows the user to enable it later.
7569 if (dev->hw_features & NETIF_F_TSO)
7570 dev->hw_features |= NETIF_F_TSO_MANGLEID;
7571 if (dev->vlan_features & NETIF_F_TSO)
7572 dev->vlan_features |= NETIF_F_TSO_MANGLEID;
7573 if (dev->mpls_features & NETIF_F_TSO)
7574 dev->mpls_features |= NETIF_F_TSO_MANGLEID;
7575 if (dev->hw_enc_features & NETIF_F_TSO)
7576 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID;
7578 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices.
7580 dev->vlan_features |= NETIF_F_HIGHDMA;
7582 /* Make NETIF_F_SG inheritable to tunnel devices.
7584 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL;
7586 /* Make NETIF_F_SG inheritable to MPLS.
7588 dev->mpls_features |= NETIF_F_SG;
7590 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev);
7591 ret = notifier_to_errno(ret);
7595 ret = netdev_register_kobject(dev);
7598 dev->reg_state = NETREG_REGISTERED;
7600 __netdev_update_features(dev);
7603 * Default initial state at registry is that the
7604 * device is present.
7607 set_bit(__LINK_STATE_PRESENT, &dev->state);
7609 linkwatch_init_dev(dev);
7611 dev_init_scheduler(dev);
7613 list_netdevice(dev);
7614 add_device_randomness(dev->dev_addr, dev->addr_len);
7616 /* If the device has permanent device address, driver should
7617 * set dev_addr and also addr_assign_type should be set to
7618 * NET_ADDR_PERM (default value).
7620 if (dev->addr_assign_type == NET_ADDR_PERM)
7621 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len);
7623 /* Notify protocols, that a new device appeared. */
7624 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
7625 ret = notifier_to_errno(ret);
7627 rollback_registered(dev);
7628 dev->reg_state = NETREG_UNREGISTERED;
7631 * Prevent userspace races by waiting until the network
7632 * device is fully setup before sending notifications.
7634 if (!dev->rtnl_link_ops ||
7635 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
7636 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
7642 if (dev->netdev_ops->ndo_uninit)
7643 dev->netdev_ops->ndo_uninit(dev);
7644 if (dev->priv_destructor)
7645 dev->priv_destructor(dev);
7648 EXPORT_SYMBOL(register_netdevice);
7651 * init_dummy_netdev - init a dummy network device for NAPI
7652 * @dev: device to init
7654 * This takes a network device structure and initialize the minimum
7655 * amount of fields so it can be used to schedule NAPI polls without
7656 * registering a full blown interface. This is to be used by drivers
7657 * that need to tie several hardware interfaces to a single NAPI
7658 * poll scheduler due to HW limitations.
7660 int init_dummy_netdev(struct net_device *dev)
7662 /* Clear everything. Note we don't initialize spinlocks
7663 * are they aren't supposed to be taken by any of the
7664 * NAPI code and this dummy netdev is supposed to be
7665 * only ever used for NAPI polls
7667 memset(dev, 0, sizeof(struct net_device));
7669 /* make sure we BUG if trying to hit standard
7670 * register/unregister code path
7672 dev->reg_state = NETREG_DUMMY;
7674 /* NAPI wants this */
7675 INIT_LIST_HEAD(&dev->napi_list);
7677 /* a dummy interface is started by default */
7678 set_bit(__LINK_STATE_PRESENT, &dev->state);
7679 set_bit(__LINK_STATE_START, &dev->state);
7681 /* Note : We dont allocate pcpu_refcnt for dummy devices,
7682 * because users of this 'device' dont need to change
7688 EXPORT_SYMBOL_GPL(init_dummy_netdev);
7692 * register_netdev - register a network device
7693 * @dev: device to register
7695 * Take a completed network device structure and add it to the kernel
7696 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
7697 * chain. 0 is returned on success. A negative errno code is returned
7698 * on a failure to set up the device, or if the name is a duplicate.
7700 * This is a wrapper around register_netdevice that takes the rtnl semaphore
7701 * and expands the device name if you passed a format string to
7704 int register_netdev(struct net_device *dev)
7709 err = register_netdevice(dev);
7713 EXPORT_SYMBOL(register_netdev);
7715 int netdev_refcnt_read(const struct net_device *dev)
7719 for_each_possible_cpu(i)
7720 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
7723 EXPORT_SYMBOL(netdev_refcnt_read);
7726 * netdev_wait_allrefs - wait until all references are gone.
7727 * @dev: target net_device
7729 * This is called when unregistering network devices.
7731 * Any protocol or device that holds a reference should register
7732 * for netdevice notification, and cleanup and put back the
7733 * reference if they receive an UNREGISTER event.
7734 * We can get stuck here if buggy protocols don't correctly
7737 static void netdev_wait_allrefs(struct net_device *dev)
7739 unsigned long rebroadcast_time, warning_time;
7742 linkwatch_forget_dev(dev);
7744 rebroadcast_time = warning_time = jiffies;
7745 refcnt = netdev_refcnt_read(dev);
7747 while (refcnt != 0) {
7748 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
7751 /* Rebroadcast unregister notification */
7752 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
7758 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
7759 if (test_bit(__LINK_STATE_LINKWATCH_PENDING,
7761 /* We must not have linkwatch events
7762 * pending on unregister. If this
7763 * happens, we simply run the queue
7764 * unscheduled, resulting in a noop
7767 linkwatch_run_queue();
7772 rebroadcast_time = jiffies;
7777 refcnt = netdev_refcnt_read(dev);
7779 if (time_after(jiffies, warning_time + 10 * HZ)) {
7780 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n",
7782 warning_time = jiffies;
7791 * register_netdevice(x1);
7792 * register_netdevice(x2);
7794 * unregister_netdevice(y1);
7795 * unregister_netdevice(y2);
7801 * We are invoked by rtnl_unlock().
7802 * This allows us to deal with problems:
7803 * 1) We can delete sysfs objects which invoke hotplug
7804 * without deadlocking with linkwatch via keventd.
7805 * 2) Since we run with the RTNL semaphore not held, we can sleep
7806 * safely in order to wait for the netdev refcnt to drop to zero.
7808 * We must not return until all unregister events added during
7809 * the interval the lock was held have been completed.
7811 void netdev_run_todo(void)
7813 struct list_head list;
7815 /* Snapshot list, allow later requests */
7816 list_replace_init(&net_todo_list, &list);
7821 /* Wait for rcu callbacks to finish before next phase */
7822 if (!list_empty(&list))
7825 while (!list_empty(&list)) {
7826 struct net_device *dev
7827 = list_first_entry(&list, struct net_device, todo_list);
7828 list_del(&dev->todo_list);
7831 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
7834 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) {
7835 pr_err("network todo '%s' but state %d\n",
7836 dev->name, dev->reg_state);
7841 dev->reg_state = NETREG_UNREGISTERED;
7843 netdev_wait_allrefs(dev);
7846 BUG_ON(netdev_refcnt_read(dev));
7847 BUG_ON(!list_empty(&dev->ptype_all));
7848 BUG_ON(!list_empty(&dev->ptype_specific));
7849 WARN_ON(rcu_access_pointer(dev->ip_ptr));
7850 WARN_ON(rcu_access_pointer(dev->ip6_ptr));
7851 WARN_ON(dev->dn_ptr);
7853 if (dev->priv_destructor)
7854 dev->priv_destructor(dev);
7855 if (dev->needs_free_netdev)
7858 /* Report a network device has been unregistered */
7860 dev_net(dev)->dev_unreg_count--;
7862 wake_up(&netdev_unregistering_wq);
7864 /* Free network device */
7865 kobject_put(&dev->dev.kobj);
7869 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has
7870 * all the same fields in the same order as net_device_stats, with only
7871 * the type differing, but rtnl_link_stats64 may have additional fields
7872 * at the end for newer counters.
7874 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
7875 const struct net_device_stats *netdev_stats)
7877 #if BITS_PER_LONG == 64
7878 BUILD_BUG_ON(sizeof(*stats64) < sizeof(*netdev_stats));
7879 memcpy(stats64, netdev_stats, sizeof(*netdev_stats));
7880 /* zero out counters that only exist in rtnl_link_stats64 */
7881 memset((char *)stats64 + sizeof(*netdev_stats), 0,
7882 sizeof(*stats64) - sizeof(*netdev_stats));
7884 size_t i, n = sizeof(*netdev_stats) / sizeof(unsigned long);
7885 const unsigned long *src = (const unsigned long *)netdev_stats;
7886 u64 *dst = (u64 *)stats64;
7888 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64));
7889 for (i = 0; i < n; i++)
7891 /* zero out counters that only exist in rtnl_link_stats64 */
7892 memset((char *)stats64 + n * sizeof(u64), 0,
7893 sizeof(*stats64) - n * sizeof(u64));
7896 EXPORT_SYMBOL(netdev_stats_to_stats64);
7899 * dev_get_stats - get network device statistics
7900 * @dev: device to get statistics from
7901 * @storage: place to store stats
7903 * Get network statistics from device. Return @storage.
7904 * The device driver may provide its own method by setting
7905 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats;
7906 * otherwise the internal statistics structure is used.
7908 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev,
7909 struct rtnl_link_stats64 *storage)
7911 const struct net_device_ops *ops = dev->netdev_ops;
7913 if (ops->ndo_get_stats64) {
7914 memset(storage, 0, sizeof(*storage));
7915 ops->ndo_get_stats64(dev, storage);
7916 } else if (ops->ndo_get_stats) {
7917 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev));
7919 netdev_stats_to_stats64(storage, &dev->stats);
7921 storage->rx_dropped += (unsigned long)atomic_long_read(&dev->rx_dropped);
7922 storage->tx_dropped += (unsigned long)atomic_long_read(&dev->tx_dropped);
7923 storage->rx_nohandler += (unsigned long)atomic_long_read(&dev->rx_nohandler);
7926 EXPORT_SYMBOL(dev_get_stats);
7928 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev)
7930 struct netdev_queue *queue = dev_ingress_queue(dev);
7932 #ifdef CONFIG_NET_CLS_ACT
7935 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
7938 netdev_init_one_queue(dev, queue, NULL);
7939 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc);
7940 queue->qdisc_sleeping = &noop_qdisc;
7941 rcu_assign_pointer(dev->ingress_queue, queue);
7946 static const struct ethtool_ops default_ethtool_ops;
7948 void netdev_set_default_ethtool_ops(struct net_device *dev,
7949 const struct ethtool_ops *ops)
7951 if (dev->ethtool_ops == &default_ethtool_ops)
7952 dev->ethtool_ops = ops;
7954 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops);
7956 void netdev_freemem(struct net_device *dev)
7958 char *addr = (char *)dev - dev->padded;
7964 * alloc_netdev_mqs - allocate network device
7965 * @sizeof_priv: size of private data to allocate space for
7966 * @name: device name format string
7967 * @name_assign_type: origin of device name
7968 * @setup: callback to initialize device
7969 * @txqs: the number of TX subqueues to allocate
7970 * @rxqs: the number of RX subqueues to allocate
7972 * Allocates a struct net_device with private data area for driver use
7973 * and performs basic initialization. Also allocates subqueue structs
7974 * for each queue on the device.
7976 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
7977 unsigned char name_assign_type,
7978 void (*setup)(struct net_device *),
7979 unsigned int txqs, unsigned int rxqs)
7981 struct net_device *dev;
7983 struct net_device *p;
7985 BUG_ON(strlen(name) >= sizeof(dev->name));
7988 pr_err("alloc_netdev: Unable to allocate device with zero queues\n");
7994 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n");
7999 alloc_size = sizeof(struct net_device);
8001 /* ensure 32-byte alignment of private area */
8002 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
8003 alloc_size += sizeof_priv;
8005 /* ensure 32-byte alignment of whole construct */
8006 alloc_size += NETDEV_ALIGN - 1;
8008 p = kvzalloc(alloc_size, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
8012 dev = PTR_ALIGN(p, NETDEV_ALIGN);
8013 dev->padded = (char *)dev - (char *)p;
8015 dev->pcpu_refcnt = alloc_percpu(int);
8016 if (!dev->pcpu_refcnt)
8019 if (dev_addr_init(dev))
8025 dev_net_set(dev, &init_net);
8027 dev->gso_max_size = GSO_MAX_SIZE;
8028 dev->gso_max_segs = GSO_MAX_SEGS;
8030 INIT_LIST_HEAD(&dev->napi_list);
8031 INIT_LIST_HEAD(&dev->unreg_list);
8032 INIT_LIST_HEAD(&dev->close_list);
8033 INIT_LIST_HEAD(&dev->link_watch_list);
8034 INIT_LIST_HEAD(&dev->adj_list.upper);
8035 INIT_LIST_HEAD(&dev->adj_list.lower);
8036 INIT_LIST_HEAD(&dev->ptype_all);
8037 INIT_LIST_HEAD(&dev->ptype_specific);
8038 #ifdef CONFIG_NET_SCHED
8039 hash_init(dev->qdisc_hash);
8041 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM;
8044 if (!dev->tx_queue_len) {
8045 dev->priv_flags |= IFF_NO_QUEUE;
8046 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN;
8049 dev->num_tx_queues = txqs;
8050 dev->real_num_tx_queues = txqs;
8051 if (netif_alloc_netdev_queues(dev))
8055 dev->num_rx_queues = rxqs;
8056 dev->real_num_rx_queues = rxqs;
8057 if (netif_alloc_rx_queues(dev))
8061 strcpy(dev->name, name);
8062 dev->name_assign_type = name_assign_type;
8063 dev->group = INIT_NETDEV_GROUP;
8064 if (!dev->ethtool_ops)
8065 dev->ethtool_ops = &default_ethtool_ops;
8067 nf_hook_ingress_init(dev);
8076 free_percpu(dev->pcpu_refcnt);
8078 netdev_freemem(dev);
8081 EXPORT_SYMBOL(alloc_netdev_mqs);
8084 * free_netdev - free network device
8087 * This function does the last stage of destroying an allocated device
8088 * interface. The reference to the device object is released. If this
8089 * is the last reference then it will be freed.Must be called in process
8092 void free_netdev(struct net_device *dev)
8094 struct napi_struct *p, *n;
8095 struct bpf_prog *prog;
8098 netif_free_tx_queues(dev);
8103 kfree(rcu_dereference_protected(dev->ingress_queue, 1));
8105 /* Flush device addresses */
8106 dev_addr_flush(dev);
8108 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
8111 free_percpu(dev->pcpu_refcnt);
8112 dev->pcpu_refcnt = NULL;
8114 prog = rcu_dereference_protected(dev->xdp_prog, 1);
8117 static_key_slow_dec(&generic_xdp_needed);
8120 /* Compatibility with error handling in drivers */
8121 if (dev->reg_state == NETREG_UNINITIALIZED) {
8122 netdev_freemem(dev);
8126 BUG_ON(dev->reg_state != NETREG_UNREGISTERED);
8127 dev->reg_state = NETREG_RELEASED;
8129 /* will free via device release */
8130 put_device(&dev->dev);
8132 EXPORT_SYMBOL(free_netdev);
8135 * synchronize_net - Synchronize with packet receive processing
8137 * Wait for packets currently being received to be done.
8138 * Does not block later packets from starting.
8140 void synchronize_net(void)
8143 if (rtnl_is_locked())
8144 synchronize_rcu_expedited();
8148 EXPORT_SYMBOL(synchronize_net);
8151 * unregister_netdevice_queue - remove device from the kernel
8155 * This function shuts down a device interface and removes it
8156 * from the kernel tables.
8157 * If head not NULL, device is queued to be unregistered later.
8159 * Callers must hold the rtnl semaphore. You may want
8160 * unregister_netdev() instead of this.
8163 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head)
8168 list_move_tail(&dev->unreg_list, head);
8170 rollback_registered(dev);
8171 /* Finish processing unregister after unlock */
8175 EXPORT_SYMBOL(unregister_netdevice_queue);
8178 * unregister_netdevice_many - unregister many devices
8179 * @head: list of devices
8181 * Note: As most callers use a stack allocated list_head,
8182 * we force a list_del() to make sure stack wont be corrupted later.
8184 void unregister_netdevice_many(struct list_head *head)
8186 struct net_device *dev;
8188 if (!list_empty(head)) {
8189 rollback_registered_many(head);
8190 list_for_each_entry(dev, head, unreg_list)
8195 EXPORT_SYMBOL(unregister_netdevice_many);
8198 * unregister_netdev - remove device from the kernel
8201 * This function shuts down a device interface and removes it
8202 * from the kernel tables.
8204 * This is just a wrapper for unregister_netdevice that takes
8205 * the rtnl semaphore. In general you want to use this and not
8206 * unregister_netdevice.
8208 void unregister_netdev(struct net_device *dev)
8211 unregister_netdevice(dev);
8214 EXPORT_SYMBOL(unregister_netdev);
8217 * dev_change_net_namespace - move device to different nethost namespace
8219 * @net: network namespace
8220 * @pat: If not NULL name pattern to try if the current device name
8221 * is already taken in the destination network namespace.
8223 * This function shuts down a device interface and moves it
8224 * to a new network namespace. On success 0 is returned, on
8225 * a failure a netagive errno code is returned.
8227 * Callers must hold the rtnl semaphore.
8230 int dev_change_net_namespace(struct net_device *dev, struct net *net, const char *pat)
8236 /* Don't allow namespace local devices to be moved. */
8238 if (dev->features & NETIF_F_NETNS_LOCAL)
8241 /* Ensure the device has been registrered */
8242 if (dev->reg_state != NETREG_REGISTERED)
8245 /* Get out if there is nothing todo */
8247 if (net_eq(dev_net(dev), net))
8250 /* Pick the destination device name, and ensure
8251 * we can use it in the destination network namespace.
8254 if (__dev_get_by_name(net, dev->name)) {
8255 /* We get here if we can't use the current device name */
8258 if (dev_get_valid_name(net, dev, pat) < 0)
8263 * And now a mini version of register_netdevice unregister_netdevice.
8266 /* If device is running close it first. */
8269 /* And unlink it from device chain */
8271 unlist_netdevice(dev);
8275 /* Shutdown queueing discipline. */
8278 /* Notify protocols, that we are about to destroy
8279 * this device. They should clean all the things.
8281 * Note that dev->reg_state stays at NETREG_REGISTERED.
8282 * This is wanted because this way 8021q and macvlan know
8283 * the device is just moving and can keep their slaves up.
8285 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
8287 call_netdevice_notifiers(NETDEV_UNREGISTER_FINAL, dev);
8288 rtmsg_ifinfo(RTM_DELLINK, dev, ~0U, GFP_KERNEL);
8291 * Flush the unicast and multicast chains
8296 /* Send a netdev-removed uevent to the old namespace */
8297 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE);
8298 netdev_adjacent_del_links(dev);
8300 /* Actually switch the network namespace */
8301 dev_net_set(dev, net);
8303 /* If there is an ifindex conflict assign a new one */
8304 if (__dev_get_by_index(net, dev->ifindex))
8305 dev->ifindex = dev_new_index(net);
8307 /* Send a netdev-add uevent to the new namespace */
8308 kobject_uevent(&dev->dev.kobj, KOBJ_ADD);
8309 netdev_adjacent_add_links(dev);
8311 /* Fixup kobjects */
8312 err = device_rename(&dev->dev, dev->name);
8315 /* Add the device back in the hashes */
8316 list_netdevice(dev);
8318 /* Notify protocols, that a new device appeared. */
8319 call_netdevice_notifiers(NETDEV_REGISTER, dev);
8322 * Prevent userspace races by waiting until the network
8323 * device is fully setup before sending notifications.
8325 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
8332 EXPORT_SYMBOL_GPL(dev_change_net_namespace);
8334 static int dev_cpu_dead(unsigned int oldcpu)
8336 struct sk_buff **list_skb;
8337 struct sk_buff *skb;
8339 struct softnet_data *sd, *oldsd, *remsd = NULL;
8341 local_irq_disable();
8342 cpu = smp_processor_id();
8343 sd = &per_cpu(softnet_data, cpu);
8344 oldsd = &per_cpu(softnet_data, oldcpu);
8346 /* Find end of our completion_queue. */
8347 list_skb = &sd->completion_queue;
8349 list_skb = &(*list_skb)->next;
8350 /* Append completion queue from offline CPU. */
8351 *list_skb = oldsd->completion_queue;
8352 oldsd->completion_queue = NULL;
8354 /* Append output queue from offline CPU. */
8355 if (oldsd->output_queue) {
8356 *sd->output_queue_tailp = oldsd->output_queue;
8357 sd->output_queue_tailp = oldsd->output_queue_tailp;
8358 oldsd->output_queue = NULL;
8359 oldsd->output_queue_tailp = &oldsd->output_queue;
8361 /* Append NAPI poll list from offline CPU, with one exception :
8362 * process_backlog() must be called by cpu owning percpu backlog.
8363 * We properly handle process_queue & input_pkt_queue later.
8365 while (!list_empty(&oldsd->poll_list)) {
8366 struct napi_struct *napi = list_first_entry(&oldsd->poll_list,
8370 list_del_init(&napi->poll_list);
8371 if (napi->poll == process_backlog)
8374 ____napi_schedule(sd, napi);
8377 raise_softirq_irqoff(NET_TX_SOFTIRQ);
8381 remsd = oldsd->rps_ipi_list;
8382 oldsd->rps_ipi_list = NULL;
8384 /* send out pending IPI's on offline CPU */
8385 net_rps_send_ipi(remsd);
8387 /* Process offline CPU's input_pkt_queue */
8388 while ((skb = __skb_dequeue(&oldsd->process_queue))) {
8390 input_queue_head_incr(oldsd);
8392 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) {
8394 input_queue_head_incr(oldsd);
8401 * netdev_increment_features - increment feature set by one
8402 * @all: current feature set
8403 * @one: new feature set
8404 * @mask: mask feature set
8406 * Computes a new feature set after adding a device with feature set
8407 * @one to the master device with current feature set @all. Will not
8408 * enable anything that is off in @mask. Returns the new feature set.
8410 netdev_features_t netdev_increment_features(netdev_features_t all,
8411 netdev_features_t one, netdev_features_t mask)
8413 if (mask & NETIF_F_HW_CSUM)
8414 mask |= NETIF_F_CSUM_MASK;
8415 mask |= NETIF_F_VLAN_CHALLENGED;
8417 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask;
8418 all &= one | ~NETIF_F_ALL_FOR_ALL;
8420 /* If one device supports hw checksumming, set for all. */
8421 if (all & NETIF_F_HW_CSUM)
8422 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM);
8426 EXPORT_SYMBOL(netdev_increment_features);
8428 static struct hlist_head * __net_init netdev_create_hash(void)
8431 struct hlist_head *hash;
8433 hash = kmalloc(sizeof(*hash) * NETDEV_HASHENTRIES, GFP_KERNEL);
8435 for (i = 0; i < NETDEV_HASHENTRIES; i++)
8436 INIT_HLIST_HEAD(&hash[i]);
8441 /* Initialize per network namespace state */
8442 static int __net_init netdev_init(struct net *net)
8444 if (net != &init_net)
8445 INIT_LIST_HEAD(&net->dev_base_head);
8447 net->dev_name_head = netdev_create_hash();
8448 if (net->dev_name_head == NULL)
8451 net->dev_index_head = netdev_create_hash();
8452 if (net->dev_index_head == NULL)
8458 kfree(net->dev_name_head);
8464 * netdev_drivername - network driver for the device
8465 * @dev: network device
8467 * Determine network driver for device.
8469 const char *netdev_drivername(const struct net_device *dev)
8471 const struct device_driver *driver;
8472 const struct device *parent;
8473 const char *empty = "";
8475 parent = dev->dev.parent;
8479 driver = parent->driver;
8480 if (driver && driver->name)
8481 return driver->name;
8485 static void __netdev_printk(const char *level, const struct net_device *dev,
8486 struct va_format *vaf)
8488 if (dev && dev->dev.parent) {
8489 dev_printk_emit(level[1] - '0',
8492 dev_driver_string(dev->dev.parent),
8493 dev_name(dev->dev.parent),
8494 netdev_name(dev), netdev_reg_state(dev),
8497 printk("%s%s%s: %pV",
8498 level, netdev_name(dev), netdev_reg_state(dev), vaf);
8500 printk("%s(NULL net_device): %pV", level, vaf);
8504 void netdev_printk(const char *level, const struct net_device *dev,
8505 const char *format, ...)
8507 struct va_format vaf;
8510 va_start(args, format);
8515 __netdev_printk(level, dev, &vaf);
8519 EXPORT_SYMBOL(netdev_printk);
8521 #define define_netdev_printk_level(func, level) \
8522 void func(const struct net_device *dev, const char *fmt, ...) \
8524 struct va_format vaf; \
8527 va_start(args, fmt); \
8532 __netdev_printk(level, dev, &vaf); \
8536 EXPORT_SYMBOL(func);
8538 define_netdev_printk_level(netdev_emerg, KERN_EMERG);
8539 define_netdev_printk_level(netdev_alert, KERN_ALERT);
8540 define_netdev_printk_level(netdev_crit, KERN_CRIT);
8541 define_netdev_printk_level(netdev_err, KERN_ERR);
8542 define_netdev_printk_level(netdev_warn, KERN_WARNING);
8543 define_netdev_printk_level(netdev_notice, KERN_NOTICE);
8544 define_netdev_printk_level(netdev_info, KERN_INFO);
8546 static void __net_exit netdev_exit(struct net *net)
8548 kfree(net->dev_name_head);
8549 kfree(net->dev_index_head);
8552 static struct pernet_operations __net_initdata netdev_net_ops = {
8553 .init = netdev_init,
8554 .exit = netdev_exit,
8557 static void __net_exit default_device_exit(struct net *net)
8559 struct net_device *dev, *aux;
8561 * Push all migratable network devices back to the
8562 * initial network namespace
8565 for_each_netdev_safe(net, dev, aux) {
8567 char fb_name[IFNAMSIZ];
8569 /* Ignore unmoveable devices (i.e. loopback) */
8570 if (dev->features & NETIF_F_NETNS_LOCAL)
8573 /* Leave virtual devices for the generic cleanup */
8574 if (dev->rtnl_link_ops)
8577 /* Push remaining network devices to init_net */
8578 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex);
8579 err = dev_change_net_namespace(dev, &init_net, fb_name);
8581 pr_emerg("%s: failed to move %s to init_net: %d\n",
8582 __func__, dev->name, err);
8589 static void __net_exit rtnl_lock_unregistering(struct list_head *net_list)
8591 /* Return with the rtnl_lock held when there are no network
8592 * devices unregistering in any network namespace in net_list.
8596 DEFINE_WAIT_FUNC(wait, woken_wake_function);
8598 add_wait_queue(&netdev_unregistering_wq, &wait);
8600 unregistering = false;
8602 list_for_each_entry(net, net_list, exit_list) {
8603 if (net->dev_unreg_count > 0) {
8604 unregistering = true;
8612 wait_woken(&wait, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
8614 remove_wait_queue(&netdev_unregistering_wq, &wait);
8617 static void __net_exit default_device_exit_batch(struct list_head *net_list)
8619 /* At exit all network devices most be removed from a network
8620 * namespace. Do this in the reverse order of registration.
8621 * Do this across as many network namespaces as possible to
8622 * improve batching efficiency.
8624 struct net_device *dev;
8626 LIST_HEAD(dev_kill_list);
8628 /* To prevent network device cleanup code from dereferencing
8629 * loopback devices or network devices that have been freed
8630 * wait here for all pending unregistrations to complete,
8631 * before unregistring the loopback device and allowing the
8632 * network namespace be freed.
8634 * The netdev todo list containing all network devices
8635 * unregistrations that happen in default_device_exit_batch
8636 * will run in the rtnl_unlock() at the end of
8637 * default_device_exit_batch.
8639 rtnl_lock_unregistering(net_list);
8640 list_for_each_entry(net, net_list, exit_list) {
8641 for_each_netdev_reverse(net, dev) {
8642 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink)
8643 dev->rtnl_link_ops->dellink(dev, &dev_kill_list);
8645 unregister_netdevice_queue(dev, &dev_kill_list);
8648 unregister_netdevice_many(&dev_kill_list);
8652 static struct pernet_operations __net_initdata default_device_ops = {
8653 .exit = default_device_exit,
8654 .exit_batch = default_device_exit_batch,
8658 * Initialize the DEV module. At boot time this walks the device list and
8659 * unhooks any devices that fail to initialise (normally hardware not
8660 * present) and leaves us with a valid list of present and active devices.
8665 * This is called single threaded during boot, so no need
8666 * to take the rtnl semaphore.
8668 static int __init net_dev_init(void)
8670 int i, rc = -ENOMEM;
8672 BUG_ON(!dev_boot_phase);
8674 if (dev_proc_init())
8677 if (netdev_kobject_init())
8680 INIT_LIST_HEAD(&ptype_all);
8681 for (i = 0; i < PTYPE_HASH_SIZE; i++)
8682 INIT_LIST_HEAD(&ptype_base[i]);
8684 INIT_LIST_HEAD(&offload_base);
8686 if (register_pernet_subsys(&netdev_net_ops))
8690 * Initialise the packet receive queues.
8693 for_each_possible_cpu(i) {
8694 struct work_struct *flush = per_cpu_ptr(&flush_works, i);
8695 struct softnet_data *sd = &per_cpu(softnet_data, i);
8697 INIT_WORK(flush, flush_backlog);
8699 skb_queue_head_init(&sd->input_pkt_queue);
8700 skb_queue_head_init(&sd->process_queue);
8701 INIT_LIST_HEAD(&sd->poll_list);
8702 sd->output_queue_tailp = &sd->output_queue;
8704 sd->csd.func = rps_trigger_softirq;
8709 sd->backlog.poll = process_backlog;
8710 sd->backlog.weight = weight_p;
8715 /* The loopback device is special if any other network devices
8716 * is present in a network namespace the loopback device must
8717 * be present. Since we now dynamically allocate and free the
8718 * loopback device ensure this invariant is maintained by
8719 * keeping the loopback device as the first device on the
8720 * list of network devices. Ensuring the loopback devices
8721 * is the first device that appears and the last network device
8724 if (register_pernet_device(&loopback_net_ops))
8727 if (register_pernet_device(&default_device_ops))
8730 open_softirq(NET_TX_SOFTIRQ, net_tx_action);
8731 open_softirq(NET_RX_SOFTIRQ, net_rx_action);
8733 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead",
8734 NULL, dev_cpu_dead);
8741 subsys_initcall(net_dev_init);