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/skbuff.h>
97 #include <linux/bpf.h>
98 #include <linux/bpf_trace.h>
99 #include <net/net_namespace.h>
100 #include <net/sock.h>
101 #include <net/busy_poll.h>
102 #include <linux/rtnetlink.h>
103 #include <linux/stat.h>
105 #include <net/dst_metadata.h>
106 #include <net/pkt_sched.h>
107 #include <net/pkt_cls.h>
108 #include <net/checksum.h>
109 #include <net/xfrm.h>
110 #include <linux/highmem.h>
111 #include <linux/init.h>
112 #include <linux/module.h>
113 #include <linux/netpoll.h>
114 #include <linux/rcupdate.h>
115 #include <linux/delay.h>
116 #include <net/iw_handler.h>
117 #include <asm/current.h>
118 #include <linux/audit.h>
119 #include <linux/dmaengine.h>
120 #include <linux/err.h>
121 #include <linux/ctype.h>
122 #include <linux/if_arp.h>
123 #include <linux/if_vlan.h>
124 #include <linux/ip.h>
126 #include <net/mpls.h>
127 #include <linux/ipv6.h>
128 #include <linux/in.h>
129 #include <linux/jhash.h>
130 #include <linux/random.h>
131 #include <trace/events/napi.h>
132 #include <trace/events/net.h>
133 #include <trace/events/skb.h>
134 #include <linux/pci.h>
135 #include <linux/inetdevice.h>
136 #include <linux/cpu_rmap.h>
137 #include <linux/static_key.h>
138 #include <linux/hashtable.h>
139 #include <linux/vmalloc.h>
140 #include <linux/if_macvlan.h>
141 #include <linux/errqueue.h>
142 #include <linux/hrtimer.h>
143 #include <linux/netfilter_ingress.h>
144 #include <linux/crash_dump.h>
145 #include <linux/sctp.h>
146 #include <net/udp_tunnel.h>
147 #include <linux/net_namespace.h>
149 #include "net-sysfs.h"
151 #define MAX_GRO_SKBS 8
153 /* This should be increased if a protocol with a bigger head is added. */
154 #define GRO_MAX_HEAD (MAX_HEADER + 128)
156 static DEFINE_SPINLOCK(ptype_lock);
157 static DEFINE_SPINLOCK(offload_lock);
158 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly;
159 struct list_head ptype_all __read_mostly; /* Taps */
160 static struct list_head offload_base __read_mostly;
162 static int netif_rx_internal(struct sk_buff *skb);
163 static int call_netdevice_notifiers_info(unsigned long val,
164 struct netdev_notifier_info *info);
165 static struct napi_struct *napi_by_id(unsigned int napi_id);
168 * The @dev_base_head list is protected by @dev_base_lock and the rtnl
171 * Pure readers hold dev_base_lock for reading, or rcu_read_lock()
173 * Writers must hold the rtnl semaphore while they loop through the
174 * dev_base_head list, and hold dev_base_lock for writing when they do the
175 * actual updates. This allows pure readers to access the list even
176 * while a writer is preparing to update it.
178 * To put it another way, dev_base_lock is held for writing only to
179 * protect against pure readers; the rtnl semaphore provides the
180 * protection against other writers.
182 * See, for example usages, register_netdevice() and
183 * unregister_netdevice(), which must be called with the rtnl
186 DEFINE_RWLOCK(dev_base_lock);
187 EXPORT_SYMBOL(dev_base_lock);
189 static DEFINE_MUTEX(ifalias_mutex);
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 (strnlen(name, IFNAMSIZ) == 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 if (!dev_valid_name(name))
1068 p = strchr(name, '%');
1071 * Verify the string as this thing may have come from
1072 * the user. There must be either one "%d" and no other "%"
1075 if (p[1] != 'd' || strchr(p + 2, '%'))
1078 /* Use one page as a bit array of possible slots */
1079 inuse = (unsigned long *) get_zeroed_page(GFP_ATOMIC);
1083 for_each_netdev(net, d) {
1084 if (!sscanf(d->name, name, &i))
1086 if (i < 0 || i >= max_netdevices)
1089 /* avoid cases where sscanf is not exact inverse of printf */
1090 snprintf(buf, IFNAMSIZ, name, i);
1091 if (!strncmp(buf, d->name, IFNAMSIZ))
1095 i = find_first_zero_bit(inuse, max_netdevices);
1096 free_page((unsigned long) inuse);
1099 snprintf(buf, IFNAMSIZ, name, i);
1100 if (!__dev_get_by_name(net, buf))
1103 /* It is possible to run out of possible slots
1104 * when the name is long and there isn't enough space left
1105 * for the digits, or if all bits are used.
1110 static int dev_alloc_name_ns(struct net *net,
1111 struct net_device *dev,
1118 ret = __dev_alloc_name(net, name, buf);
1120 strlcpy(dev->name, buf, IFNAMSIZ);
1125 * dev_alloc_name - allocate a name for a device
1127 * @name: name format string
1129 * Passed a format string - eg "lt%d" it will try and find a suitable
1130 * id. It scans list of devices to build up a free map, then chooses
1131 * the first empty slot. The caller must hold the dev_base or rtnl lock
1132 * while allocating the name and adding the device in order to avoid
1134 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1135 * Returns the number of the unit assigned or a negative errno code.
1138 int dev_alloc_name(struct net_device *dev, const char *name)
1140 return dev_alloc_name_ns(dev_net(dev), dev, name);
1142 EXPORT_SYMBOL(dev_alloc_name);
1144 int dev_get_valid_name(struct net *net, struct net_device *dev,
1149 if (!dev_valid_name(name))
1152 if (strchr(name, '%'))
1153 return dev_alloc_name_ns(net, dev, name);
1154 else if (__dev_get_by_name(net, name))
1156 else if (dev->name != name)
1157 strlcpy(dev->name, name, IFNAMSIZ);
1161 EXPORT_SYMBOL(dev_get_valid_name);
1164 * dev_change_name - change name of a device
1166 * @newname: name (or format string) must be at least IFNAMSIZ
1168 * Change name of a device, can pass format strings "eth%d".
1171 int dev_change_name(struct net_device *dev, const char *newname)
1173 unsigned char old_assign_type;
1174 char oldname[IFNAMSIZ];
1180 BUG_ON(!dev_net(dev));
1183 if (dev->flags & IFF_UP)
1186 write_seqcount_begin(&devnet_rename_seq);
1188 if (strncmp(newname, dev->name, IFNAMSIZ) == 0) {
1189 write_seqcount_end(&devnet_rename_seq);
1193 memcpy(oldname, dev->name, IFNAMSIZ);
1195 err = dev_get_valid_name(net, dev, newname);
1197 write_seqcount_end(&devnet_rename_seq);
1201 if (oldname[0] && !strchr(oldname, '%'))
1202 netdev_info(dev, "renamed from %s\n", oldname);
1204 old_assign_type = dev->name_assign_type;
1205 dev->name_assign_type = NET_NAME_RENAMED;
1208 ret = device_rename(&dev->dev, dev->name);
1210 memcpy(dev->name, oldname, IFNAMSIZ);
1211 dev->name_assign_type = old_assign_type;
1212 write_seqcount_end(&devnet_rename_seq);
1216 write_seqcount_end(&devnet_rename_seq);
1218 netdev_adjacent_rename_links(dev, oldname);
1220 write_lock_bh(&dev_base_lock);
1221 hlist_del_rcu(&dev->name_hlist);
1222 write_unlock_bh(&dev_base_lock);
1226 write_lock_bh(&dev_base_lock);
1227 hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
1228 write_unlock_bh(&dev_base_lock);
1230 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev);
1231 ret = notifier_to_errno(ret);
1234 /* err >= 0 after dev_alloc_name() or stores the first errno */
1237 write_seqcount_begin(&devnet_rename_seq);
1238 memcpy(dev->name, oldname, IFNAMSIZ);
1239 memcpy(oldname, newname, IFNAMSIZ);
1240 dev->name_assign_type = old_assign_type;
1241 old_assign_type = NET_NAME_RENAMED;
1244 pr_err("%s: name change rollback failed: %d\n",
1253 * dev_set_alias - change ifalias of a device
1255 * @alias: name up to IFALIASZ
1256 * @len: limit of bytes to copy from info
1258 * Set ifalias for a device,
1260 int dev_set_alias(struct net_device *dev, const char *alias, size_t len)
1262 struct dev_ifalias *new_alias = NULL;
1264 if (len >= IFALIASZ)
1268 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL);
1272 memcpy(new_alias->ifalias, alias, len);
1273 new_alias->ifalias[len] = 0;
1276 mutex_lock(&ifalias_mutex);
1277 rcu_swap_protected(dev->ifalias, new_alias,
1278 mutex_is_locked(&ifalias_mutex));
1279 mutex_unlock(&ifalias_mutex);
1282 kfree_rcu(new_alias, rcuhead);
1286 EXPORT_SYMBOL(dev_set_alias);
1289 * dev_get_alias - get ifalias of a device
1291 * @name: buffer to store name of ifalias
1292 * @len: size of buffer
1294 * get ifalias for a device. Caller must make sure dev cannot go
1295 * away, e.g. rcu read lock or own a reference count to device.
1297 int dev_get_alias(const struct net_device *dev, char *name, size_t len)
1299 const struct dev_ifalias *alias;
1303 alias = rcu_dereference(dev->ifalias);
1305 ret = snprintf(name, len, "%s", alias->ifalias);
1312 * netdev_features_change - device changes features
1313 * @dev: device to cause notification
1315 * Called to indicate a device has changed features.
1317 void netdev_features_change(struct net_device *dev)
1319 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev);
1321 EXPORT_SYMBOL(netdev_features_change);
1324 * netdev_state_change - device changes state
1325 * @dev: device to cause notification
1327 * Called to indicate a device has changed state. This function calls
1328 * the notifier chains for netdev_chain and sends a NEWLINK message
1329 * to the routing socket.
1331 void netdev_state_change(struct net_device *dev)
1333 if (dev->flags & IFF_UP) {
1334 struct netdev_notifier_change_info change_info = {
1338 call_netdevice_notifiers_info(NETDEV_CHANGE,
1340 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL);
1343 EXPORT_SYMBOL(netdev_state_change);
1346 * netdev_notify_peers - notify network peers about existence of @dev
1347 * @dev: network device
1349 * Generate traffic such that interested network peers are aware of
1350 * @dev, such as by generating a gratuitous ARP. This may be used when
1351 * a device wants to inform the rest of the network about some sort of
1352 * reconfiguration such as a failover event or virtual machine
1355 void netdev_notify_peers(struct net_device *dev)
1358 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
1359 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev);
1362 EXPORT_SYMBOL(netdev_notify_peers);
1364 static int __dev_open(struct net_device *dev)
1366 const struct net_device_ops *ops = dev->netdev_ops;
1371 if (!netif_device_present(dev))
1374 /* Block netpoll from trying to do any rx path servicing.
1375 * If we don't do this there is a chance ndo_poll_controller
1376 * or ndo_poll may be running while we open the device
1378 netpoll_poll_disable(dev);
1380 ret = call_netdevice_notifiers(NETDEV_PRE_UP, dev);
1381 ret = notifier_to_errno(ret);
1385 set_bit(__LINK_STATE_START, &dev->state);
1387 if (ops->ndo_validate_addr)
1388 ret = ops->ndo_validate_addr(dev);
1390 if (!ret && ops->ndo_open)
1391 ret = ops->ndo_open(dev);
1393 netpoll_poll_enable(dev);
1396 clear_bit(__LINK_STATE_START, &dev->state);
1398 dev->flags |= IFF_UP;
1399 dev_set_rx_mode(dev);
1401 add_device_randomness(dev->dev_addr, dev->addr_len);
1408 * dev_open - prepare an interface for use.
1409 * @dev: device to open
1411 * Takes a device from down to up state. The device's private open
1412 * function is invoked and then the multicast lists are loaded. Finally
1413 * the device is moved into the up state and a %NETDEV_UP message is
1414 * sent to the netdev notifier chain.
1416 * Calling this function on an active interface is a nop. On a failure
1417 * a negative errno code is returned.
1419 int dev_open(struct net_device *dev)
1423 if (dev->flags & IFF_UP)
1426 ret = __dev_open(dev);
1430 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1431 call_netdevice_notifiers(NETDEV_UP, dev);
1435 EXPORT_SYMBOL(dev_open);
1437 static void __dev_close_many(struct list_head *head)
1439 struct net_device *dev;
1444 list_for_each_entry(dev, head, close_list) {
1445 /* Temporarily disable netpoll until the interface is down */
1446 netpoll_poll_disable(dev);
1448 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev);
1450 clear_bit(__LINK_STATE_START, &dev->state);
1452 /* Synchronize to scheduled poll. We cannot touch poll list, it
1453 * can be even on different cpu. So just clear netif_running().
1455 * dev->stop() will invoke napi_disable() on all of it's
1456 * napi_struct instances on this device.
1458 smp_mb__after_atomic(); /* Commit netif_running(). */
1461 dev_deactivate_many(head);
1463 list_for_each_entry(dev, head, close_list) {
1464 const struct net_device_ops *ops = dev->netdev_ops;
1467 * Call the device specific close. This cannot fail.
1468 * Only if device is UP
1470 * We allow it to be called even after a DETACH hot-plug
1476 dev->flags &= ~IFF_UP;
1477 netpoll_poll_enable(dev);
1481 static void __dev_close(struct net_device *dev)
1485 list_add(&dev->close_list, &single);
1486 __dev_close_many(&single);
1490 void dev_close_many(struct list_head *head, bool unlink)
1492 struct net_device *dev, *tmp;
1494 /* Remove the devices that don't need to be closed */
1495 list_for_each_entry_safe(dev, tmp, head, close_list)
1496 if (!(dev->flags & IFF_UP))
1497 list_del_init(&dev->close_list);
1499 __dev_close_many(head);
1501 list_for_each_entry_safe(dev, tmp, head, close_list) {
1502 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1503 call_netdevice_notifiers(NETDEV_DOWN, dev);
1505 list_del_init(&dev->close_list);
1508 EXPORT_SYMBOL(dev_close_many);
1511 * dev_close - shutdown an interface.
1512 * @dev: device to shutdown
1514 * This function moves an active device into down state. A
1515 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device
1516 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier
1519 void dev_close(struct net_device *dev)
1521 if (dev->flags & IFF_UP) {
1524 list_add(&dev->close_list, &single);
1525 dev_close_many(&single, true);
1529 EXPORT_SYMBOL(dev_close);
1533 * dev_disable_lro - disable Large Receive Offload on a device
1536 * Disable Large Receive Offload (LRO) on a net device. Must be
1537 * called under RTNL. This is needed if received packets may be
1538 * forwarded to another interface.
1540 void dev_disable_lro(struct net_device *dev)
1542 struct net_device *lower_dev;
1543 struct list_head *iter;
1545 dev->wanted_features &= ~NETIF_F_LRO;
1546 netdev_update_features(dev);
1548 if (unlikely(dev->features & NETIF_F_LRO))
1549 netdev_WARN(dev, "failed to disable LRO!\n");
1551 netdev_for_each_lower_dev(dev, lower_dev, iter)
1552 dev_disable_lro(lower_dev);
1554 EXPORT_SYMBOL(dev_disable_lro);
1557 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device
1560 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be
1561 * called under RTNL. This is needed if Generic XDP is installed on
1564 static void dev_disable_gro_hw(struct net_device *dev)
1566 dev->wanted_features &= ~NETIF_F_GRO_HW;
1567 netdev_update_features(dev);
1569 if (unlikely(dev->features & NETIF_F_GRO_HW))
1570 netdev_WARN(dev, "failed to disable GRO_HW!\n");
1573 const char *netdev_cmd_to_name(enum netdev_cmd cmd)
1576 case NETDEV_##val: \
1577 return "NETDEV_" __stringify(val);
1579 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER)
1580 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE)
1581 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE)
1582 N(POST_INIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) N(CHANGEUPPER)
1583 N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) N(BONDING_INFO)
1584 N(PRECHANGEUPPER) N(CHANGELOWERSTATE) N(UDP_TUNNEL_PUSH_INFO)
1585 N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN)
1586 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO)
1587 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO)
1590 return "UNKNOWN_NETDEV_EVENT";
1592 EXPORT_SYMBOL_GPL(netdev_cmd_to_name);
1594 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val,
1595 struct net_device *dev)
1597 struct netdev_notifier_info info = {
1601 return nb->notifier_call(nb, val, &info);
1604 static int dev_boot_phase = 1;
1607 * register_netdevice_notifier - register a network notifier block
1610 * Register a notifier to be called when network device events occur.
1611 * The notifier passed is linked into the kernel structures and must
1612 * not be reused until it has been unregistered. A negative errno code
1613 * is returned on a failure.
1615 * When registered all registration and up events are replayed
1616 * to the new notifier to allow device to have a race free
1617 * view of the network device list.
1620 int register_netdevice_notifier(struct notifier_block *nb)
1622 struct net_device *dev;
1623 struct net_device *last;
1627 /* Close race with setup_net() and cleanup_net() */
1628 down_write(&pernet_ops_rwsem);
1630 err = raw_notifier_chain_register(&netdev_chain, nb);
1636 for_each_netdev(net, dev) {
1637 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev);
1638 err = notifier_to_errno(err);
1642 if (!(dev->flags & IFF_UP))
1645 call_netdevice_notifier(nb, NETDEV_UP, dev);
1651 up_write(&pernet_ops_rwsem);
1657 for_each_netdev(net, dev) {
1661 if (dev->flags & IFF_UP) {
1662 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1664 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1666 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1671 raw_notifier_chain_unregister(&netdev_chain, nb);
1674 EXPORT_SYMBOL(register_netdevice_notifier);
1677 * unregister_netdevice_notifier - unregister a network notifier block
1680 * Unregister a notifier previously registered by
1681 * register_netdevice_notifier(). The notifier is unlinked into the
1682 * kernel structures and may then be reused. A negative errno code
1683 * is returned on a failure.
1685 * After unregistering unregister and down device events are synthesized
1686 * for all devices on the device list to the removed notifier to remove
1687 * the need for special case cleanup code.
1690 int unregister_netdevice_notifier(struct notifier_block *nb)
1692 struct net_device *dev;
1696 /* Close race with setup_net() and cleanup_net() */
1697 down_write(&pernet_ops_rwsem);
1699 err = raw_notifier_chain_unregister(&netdev_chain, nb);
1704 for_each_netdev(net, dev) {
1705 if (dev->flags & IFF_UP) {
1706 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1708 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1710 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1715 up_write(&pernet_ops_rwsem);
1718 EXPORT_SYMBOL(unregister_netdevice_notifier);
1721 * call_netdevice_notifiers_info - call all network notifier blocks
1722 * @val: value passed unmodified to notifier function
1723 * @info: notifier information data
1725 * Call all network notifier blocks. Parameters and return value
1726 * are as for raw_notifier_call_chain().
1729 static int call_netdevice_notifiers_info(unsigned long val,
1730 struct netdev_notifier_info *info)
1733 return raw_notifier_call_chain(&netdev_chain, val, info);
1737 * call_netdevice_notifiers - call all network notifier blocks
1738 * @val: value passed unmodified to notifier function
1739 * @dev: net_device pointer passed unmodified to notifier function
1741 * Call all network notifier blocks. Parameters and return value
1742 * are as for raw_notifier_call_chain().
1745 int call_netdevice_notifiers(unsigned long val, struct net_device *dev)
1747 struct netdev_notifier_info info = {
1751 return call_netdevice_notifiers_info(val, &info);
1753 EXPORT_SYMBOL(call_netdevice_notifiers);
1755 #ifdef CONFIG_NET_INGRESS
1756 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key);
1758 void net_inc_ingress_queue(void)
1760 static_branch_inc(&ingress_needed_key);
1762 EXPORT_SYMBOL_GPL(net_inc_ingress_queue);
1764 void net_dec_ingress_queue(void)
1766 static_branch_dec(&ingress_needed_key);
1768 EXPORT_SYMBOL_GPL(net_dec_ingress_queue);
1771 #ifdef CONFIG_NET_EGRESS
1772 static DEFINE_STATIC_KEY_FALSE(egress_needed_key);
1774 void net_inc_egress_queue(void)
1776 static_branch_inc(&egress_needed_key);
1778 EXPORT_SYMBOL_GPL(net_inc_egress_queue);
1780 void net_dec_egress_queue(void)
1782 static_branch_dec(&egress_needed_key);
1784 EXPORT_SYMBOL_GPL(net_dec_egress_queue);
1787 static DEFINE_STATIC_KEY_FALSE(netstamp_needed_key);
1788 #ifdef HAVE_JUMP_LABEL
1789 static atomic_t netstamp_needed_deferred;
1790 static atomic_t netstamp_wanted;
1791 static void netstamp_clear(struct work_struct *work)
1793 int deferred = atomic_xchg(&netstamp_needed_deferred, 0);
1796 wanted = atomic_add_return(deferred, &netstamp_wanted);
1798 static_branch_enable(&netstamp_needed_key);
1800 static_branch_disable(&netstamp_needed_key);
1802 static DECLARE_WORK(netstamp_work, netstamp_clear);
1805 void net_enable_timestamp(void)
1807 #ifdef HAVE_JUMP_LABEL
1811 wanted = atomic_read(&netstamp_wanted);
1814 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted + 1) == wanted)
1817 atomic_inc(&netstamp_needed_deferred);
1818 schedule_work(&netstamp_work);
1820 static_branch_inc(&netstamp_needed_key);
1823 EXPORT_SYMBOL(net_enable_timestamp);
1825 void net_disable_timestamp(void)
1827 #ifdef HAVE_JUMP_LABEL
1831 wanted = atomic_read(&netstamp_wanted);
1834 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted - 1) == wanted)
1837 atomic_dec(&netstamp_needed_deferred);
1838 schedule_work(&netstamp_work);
1840 static_branch_dec(&netstamp_needed_key);
1843 EXPORT_SYMBOL(net_disable_timestamp);
1845 static inline void net_timestamp_set(struct sk_buff *skb)
1848 if (static_branch_unlikely(&netstamp_needed_key))
1849 __net_timestamp(skb);
1852 #define net_timestamp_check(COND, SKB) \
1853 if (static_branch_unlikely(&netstamp_needed_key)) { \
1854 if ((COND) && !(SKB)->tstamp) \
1855 __net_timestamp(SKB); \
1858 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb)
1862 if (!(dev->flags & IFF_UP))
1865 len = dev->mtu + dev->hard_header_len + VLAN_HLEN;
1866 if (skb->len <= len)
1869 /* if TSO is enabled, we don't care about the length as the packet
1870 * could be forwarded without being segmented before
1872 if (skb_is_gso(skb))
1877 EXPORT_SYMBOL_GPL(is_skb_forwardable);
1879 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1881 int ret = ____dev_forward_skb(dev, skb);
1884 skb->protocol = eth_type_trans(skb, dev);
1885 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
1890 EXPORT_SYMBOL_GPL(__dev_forward_skb);
1893 * dev_forward_skb - loopback an skb to another netif
1895 * @dev: destination network device
1896 * @skb: buffer to forward
1899 * NET_RX_SUCCESS (no congestion)
1900 * NET_RX_DROP (packet was dropped, but freed)
1902 * dev_forward_skb can be used for injecting an skb from the
1903 * start_xmit function of one device into the receive queue
1904 * of another device.
1906 * The receiving device may be in another namespace, so
1907 * we have to clear all information in the skb that could
1908 * impact namespace isolation.
1910 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1912 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb);
1914 EXPORT_SYMBOL_GPL(dev_forward_skb);
1916 static inline int deliver_skb(struct sk_buff *skb,
1917 struct packet_type *pt_prev,
1918 struct net_device *orig_dev)
1920 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
1922 refcount_inc(&skb->users);
1923 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
1926 static inline void deliver_ptype_list_skb(struct sk_buff *skb,
1927 struct packet_type **pt,
1928 struct net_device *orig_dev,
1930 struct list_head *ptype_list)
1932 struct packet_type *ptype, *pt_prev = *pt;
1934 list_for_each_entry_rcu(ptype, ptype_list, list) {
1935 if (ptype->type != type)
1938 deliver_skb(skb, pt_prev, orig_dev);
1944 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb)
1946 if (!ptype->af_packet_priv || !skb->sk)
1949 if (ptype->id_match)
1950 return ptype->id_match(ptype, skb->sk);
1951 else if ((struct sock *)ptype->af_packet_priv == skb->sk)
1958 * Support routine. Sends outgoing frames to any network
1959 * taps currently in use.
1962 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
1964 struct packet_type *ptype;
1965 struct sk_buff *skb2 = NULL;
1966 struct packet_type *pt_prev = NULL;
1967 struct list_head *ptype_list = &ptype_all;
1971 list_for_each_entry_rcu(ptype, ptype_list, list) {
1972 if (ptype->ignore_outgoing)
1975 /* Never send packets back to the socket
1976 * they originated from - MvS (miquels@drinkel.ow.org)
1978 if (skb_loop_sk(ptype, skb))
1982 deliver_skb(skb2, pt_prev, skb->dev);
1987 /* need to clone skb, done only once */
1988 skb2 = skb_clone(skb, GFP_ATOMIC);
1992 net_timestamp_set(skb2);
1994 /* skb->nh should be correctly
1995 * set by sender, so that the second statement is
1996 * just protection against buggy protocols.
1998 skb_reset_mac_header(skb2);
2000 if (skb_network_header(skb2) < skb2->data ||
2001 skb_network_header(skb2) > skb_tail_pointer(skb2)) {
2002 net_crit_ratelimited("protocol %04x is buggy, dev %s\n",
2003 ntohs(skb2->protocol),
2005 skb_reset_network_header(skb2);
2008 skb2->transport_header = skb2->network_header;
2009 skb2->pkt_type = PACKET_OUTGOING;
2013 if (ptype_list == &ptype_all) {
2014 ptype_list = &dev->ptype_all;
2019 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC))
2020 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
2026 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit);
2029 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change
2030 * @dev: Network device
2031 * @txq: number of queues available
2033 * If real_num_tx_queues is changed the tc mappings may no longer be
2034 * valid. To resolve this verify the tc mapping remains valid and if
2035 * not NULL the mapping. With no priorities mapping to this
2036 * offset/count pair it will no longer be used. In the worst case TC0
2037 * is invalid nothing can be done so disable priority mappings. If is
2038 * expected that drivers will fix this mapping if they can before
2039 * calling netif_set_real_num_tx_queues.
2041 static void netif_setup_tc(struct net_device *dev, unsigned int txq)
2044 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2046 /* If TC0 is invalidated disable TC mapping */
2047 if (tc->offset + tc->count > txq) {
2048 pr_warn("Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n");
2053 /* Invalidated prio to tc mappings set to TC0 */
2054 for (i = 1; i < TC_BITMASK + 1; i++) {
2055 int q = netdev_get_prio_tc_map(dev, i);
2057 tc = &dev->tc_to_txq[q];
2058 if (tc->offset + tc->count > txq) {
2059 pr_warn("Number of in use tx queues changed. Priority %i to tc mapping %i is no longer valid. Setting map to 0\n",
2061 netdev_set_prio_tc_map(dev, i, 0);
2066 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq)
2069 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2072 /* walk through the TCs and see if it falls into any of them */
2073 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) {
2074 if ((txq - tc->offset) < tc->count)
2078 /* didn't find it, just return -1 to indicate no match */
2084 EXPORT_SYMBOL(netdev_txq_to_tc);
2087 struct static_key xps_needed __read_mostly;
2088 EXPORT_SYMBOL(xps_needed);
2089 struct static_key xps_rxqs_needed __read_mostly;
2090 EXPORT_SYMBOL(xps_rxqs_needed);
2091 static DEFINE_MUTEX(xps_map_mutex);
2092 #define xmap_dereference(P) \
2093 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex))
2095 static bool remove_xps_queue(struct xps_dev_maps *dev_maps,
2098 struct xps_map *map = NULL;
2102 map = xmap_dereference(dev_maps->attr_map[tci]);
2106 for (pos = map->len; pos--;) {
2107 if (map->queues[pos] != index)
2111 map->queues[pos] = map->queues[--map->len];
2115 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL);
2116 kfree_rcu(map, rcu);
2123 static bool remove_xps_queue_cpu(struct net_device *dev,
2124 struct xps_dev_maps *dev_maps,
2125 int cpu, u16 offset, u16 count)
2127 int num_tc = dev->num_tc ? : 1;
2128 bool active = false;
2131 for (tci = cpu * num_tc; num_tc--; tci++) {
2134 for (i = count, j = offset; i--; j++) {
2135 if (!remove_xps_queue(dev_maps, tci, j))
2145 static void clean_xps_maps(struct net_device *dev, const unsigned long *mask,
2146 struct xps_dev_maps *dev_maps, unsigned int nr_ids,
2147 u16 offset, u16 count, bool is_rxqs_map)
2149 bool active = false;
2152 for (j = -1; j = netif_attrmask_next(j, mask, nr_ids),
2154 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset,
2158 RCU_INIT_POINTER(dev->xps_rxqs_map, NULL);
2160 RCU_INIT_POINTER(dev->xps_cpus_map, NULL);
2162 for (i = offset + (count - 1); count--; i--)
2163 netdev_queue_numa_node_write(
2164 netdev_get_tx_queue(dev, i),
2167 kfree_rcu(dev_maps, rcu);
2171 static void netif_reset_xps_queues(struct net_device *dev, u16 offset,
2174 const unsigned long *possible_mask = NULL;
2175 struct xps_dev_maps *dev_maps;
2176 unsigned int nr_ids;
2178 if (!static_key_false(&xps_needed))
2182 mutex_lock(&xps_map_mutex);
2184 if (static_key_false(&xps_rxqs_needed)) {
2185 dev_maps = xmap_dereference(dev->xps_rxqs_map);
2187 nr_ids = dev->num_rx_queues;
2188 clean_xps_maps(dev, possible_mask, dev_maps, nr_ids,
2189 offset, count, true);
2193 dev_maps = xmap_dereference(dev->xps_cpus_map);
2197 if (num_possible_cpus() > 1)
2198 possible_mask = cpumask_bits(cpu_possible_mask);
2199 nr_ids = nr_cpu_ids;
2200 clean_xps_maps(dev, possible_mask, dev_maps, nr_ids, offset, count,
2204 if (static_key_enabled(&xps_rxqs_needed))
2205 static_key_slow_dec_cpuslocked(&xps_rxqs_needed);
2207 static_key_slow_dec_cpuslocked(&xps_needed);
2208 mutex_unlock(&xps_map_mutex);
2212 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index)
2214 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index);
2217 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index,
2218 u16 index, bool is_rxqs_map)
2220 struct xps_map *new_map;
2221 int alloc_len = XPS_MIN_MAP_ALLOC;
2224 for (pos = 0; map && pos < map->len; pos++) {
2225 if (map->queues[pos] != index)
2230 /* Need to add tx-queue to this CPU's/rx-queue's existing map */
2232 if (pos < map->alloc_len)
2235 alloc_len = map->alloc_len * 2;
2238 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's
2242 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL);
2244 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL,
2245 cpu_to_node(attr_index));
2249 for (i = 0; i < pos; i++)
2250 new_map->queues[i] = map->queues[i];
2251 new_map->alloc_len = alloc_len;
2257 /* Must be called under cpus_read_lock */
2258 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask,
2259 u16 index, bool is_rxqs_map)
2261 const unsigned long *online_mask = NULL, *possible_mask = NULL;
2262 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL;
2263 int i, j, tci, numa_node_id = -2;
2264 int maps_sz, num_tc = 1, tc = 0;
2265 struct xps_map *map, *new_map;
2266 bool active = false;
2267 unsigned int nr_ids;
2270 /* Do not allow XPS on subordinate device directly */
2271 num_tc = dev->num_tc;
2275 /* If queue belongs to subordinate dev use its map */
2276 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev;
2278 tc = netdev_txq_to_tc(dev, index);
2283 mutex_lock(&xps_map_mutex);
2285 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues);
2286 dev_maps = xmap_dereference(dev->xps_rxqs_map);
2287 nr_ids = dev->num_rx_queues;
2289 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc);
2290 if (num_possible_cpus() > 1) {
2291 online_mask = cpumask_bits(cpu_online_mask);
2292 possible_mask = cpumask_bits(cpu_possible_mask);
2294 dev_maps = xmap_dereference(dev->xps_cpus_map);
2295 nr_ids = nr_cpu_ids;
2298 if (maps_sz < L1_CACHE_BYTES)
2299 maps_sz = L1_CACHE_BYTES;
2301 /* allocate memory for queue storage */
2302 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids),
2305 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL);
2306 if (!new_dev_maps) {
2307 mutex_unlock(&xps_map_mutex);
2311 tci = j * num_tc + tc;
2312 map = dev_maps ? xmap_dereference(dev_maps->attr_map[tci]) :
2315 map = expand_xps_map(map, j, index, is_rxqs_map);
2319 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map);
2323 goto out_no_new_maps;
2325 static_key_slow_inc_cpuslocked(&xps_needed);
2327 static_key_slow_inc_cpuslocked(&xps_rxqs_needed);
2329 for (j = -1; j = netif_attrmask_next(j, possible_mask, nr_ids),
2331 /* copy maps belonging to foreign traffic classes */
2332 for (i = tc, tci = j * num_tc; dev_maps && i--; tci++) {
2333 /* fill in the new device map from the old device map */
2334 map = xmap_dereference(dev_maps->attr_map[tci]);
2335 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map);
2338 /* We need to explicitly update tci as prevous loop
2339 * could break out early if dev_maps is NULL.
2341 tci = j * num_tc + tc;
2343 if (netif_attr_test_mask(j, mask, nr_ids) &&
2344 netif_attr_test_online(j, online_mask, nr_ids)) {
2345 /* add tx-queue to CPU/rx-queue maps */
2348 map = xmap_dereference(new_dev_maps->attr_map[tci]);
2349 while ((pos < map->len) && (map->queues[pos] != index))
2352 if (pos == map->len)
2353 map->queues[map->len++] = index;
2356 if (numa_node_id == -2)
2357 numa_node_id = cpu_to_node(j);
2358 else if (numa_node_id != cpu_to_node(j))
2362 } else if (dev_maps) {
2363 /* fill in the new device map from the old device map */
2364 map = xmap_dereference(dev_maps->attr_map[tci]);
2365 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map);
2368 /* copy maps belonging to foreign traffic classes */
2369 for (i = num_tc - tc, tci++; dev_maps && --i; tci++) {
2370 /* fill in the new device map from the old device map */
2371 map = xmap_dereference(dev_maps->attr_map[tci]);
2372 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map);
2377 rcu_assign_pointer(dev->xps_rxqs_map, new_dev_maps);
2379 rcu_assign_pointer(dev->xps_cpus_map, new_dev_maps);
2381 /* Cleanup old maps */
2383 goto out_no_old_maps;
2385 for (j = -1; j = netif_attrmask_next(j, possible_mask, nr_ids),
2387 for (i = num_tc, tci = j * num_tc; i--; tci++) {
2388 new_map = xmap_dereference(new_dev_maps->attr_map[tci]);
2389 map = xmap_dereference(dev_maps->attr_map[tci]);
2390 if (map && map != new_map)
2391 kfree_rcu(map, rcu);
2395 kfree_rcu(dev_maps, rcu);
2398 dev_maps = new_dev_maps;
2403 /* update Tx queue numa node */
2404 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index),
2405 (numa_node_id >= 0) ?
2406 numa_node_id : NUMA_NO_NODE);
2412 /* removes tx-queue from unused CPUs/rx-queues */
2413 for (j = -1; j = netif_attrmask_next(j, possible_mask, nr_ids),
2415 for (i = tc, tci = j * num_tc; i--; tci++)
2416 active |= remove_xps_queue(dev_maps, tci, index);
2417 if (!netif_attr_test_mask(j, mask, nr_ids) ||
2418 !netif_attr_test_online(j, online_mask, nr_ids))
2419 active |= remove_xps_queue(dev_maps, tci, index);
2420 for (i = num_tc - tc, tci++; --i; tci++)
2421 active |= remove_xps_queue(dev_maps, tci, index);
2424 /* free map if not active */
2427 RCU_INIT_POINTER(dev->xps_rxqs_map, NULL);
2429 RCU_INIT_POINTER(dev->xps_cpus_map, NULL);
2430 kfree_rcu(dev_maps, rcu);
2434 mutex_unlock(&xps_map_mutex);
2438 /* remove any maps that we added */
2439 for (j = -1; j = netif_attrmask_next(j, possible_mask, nr_ids),
2441 for (i = num_tc, tci = j * num_tc; i--; tci++) {
2442 new_map = xmap_dereference(new_dev_maps->attr_map[tci]);
2444 xmap_dereference(dev_maps->attr_map[tci]) :
2446 if (new_map && new_map != map)
2451 mutex_unlock(&xps_map_mutex);
2453 kfree(new_dev_maps);
2456 EXPORT_SYMBOL_GPL(__netif_set_xps_queue);
2458 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask,
2464 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, false);
2469 EXPORT_SYMBOL(netif_set_xps_queue);
2472 static void netdev_unbind_all_sb_channels(struct net_device *dev)
2474 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues];
2476 /* Unbind any subordinate channels */
2477 while (txq-- != &dev->_tx[0]) {
2479 netdev_unbind_sb_channel(dev, txq->sb_dev);
2483 void netdev_reset_tc(struct net_device *dev)
2486 netif_reset_xps_queues_gt(dev, 0);
2488 netdev_unbind_all_sb_channels(dev);
2490 /* Reset TC configuration of device */
2492 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq));
2493 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map));
2495 EXPORT_SYMBOL(netdev_reset_tc);
2497 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset)
2499 if (tc >= dev->num_tc)
2503 netif_reset_xps_queues(dev, offset, count);
2505 dev->tc_to_txq[tc].count = count;
2506 dev->tc_to_txq[tc].offset = offset;
2509 EXPORT_SYMBOL(netdev_set_tc_queue);
2511 int netdev_set_num_tc(struct net_device *dev, u8 num_tc)
2513 if (num_tc > TC_MAX_QUEUE)
2517 netif_reset_xps_queues_gt(dev, 0);
2519 netdev_unbind_all_sb_channels(dev);
2521 dev->num_tc = num_tc;
2524 EXPORT_SYMBOL(netdev_set_num_tc);
2526 void netdev_unbind_sb_channel(struct net_device *dev,
2527 struct net_device *sb_dev)
2529 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues];
2532 netif_reset_xps_queues_gt(sb_dev, 0);
2534 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq));
2535 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map));
2537 while (txq-- != &dev->_tx[0]) {
2538 if (txq->sb_dev == sb_dev)
2542 EXPORT_SYMBOL(netdev_unbind_sb_channel);
2544 int netdev_bind_sb_channel_queue(struct net_device *dev,
2545 struct net_device *sb_dev,
2546 u8 tc, u16 count, u16 offset)
2548 /* Make certain the sb_dev and dev are already configured */
2549 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc)
2552 /* We cannot hand out queues we don't have */
2553 if ((offset + count) > dev->real_num_tx_queues)
2556 /* Record the mapping */
2557 sb_dev->tc_to_txq[tc].count = count;
2558 sb_dev->tc_to_txq[tc].offset = offset;
2560 /* Provide a way for Tx queue to find the tc_to_txq map or
2561 * XPS map for itself.
2564 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev;
2568 EXPORT_SYMBOL(netdev_bind_sb_channel_queue);
2570 int netdev_set_sb_channel(struct net_device *dev, u16 channel)
2572 /* Do not use a multiqueue device to represent a subordinate channel */
2573 if (netif_is_multiqueue(dev))
2576 /* We allow channels 1 - 32767 to be used for subordinate channels.
2577 * Channel 0 is meant to be "native" mode and used only to represent
2578 * the main root device. We allow writing 0 to reset the device back
2579 * to normal mode after being used as a subordinate channel.
2581 if (channel > S16_MAX)
2584 dev->num_tc = -channel;
2588 EXPORT_SYMBOL(netdev_set_sb_channel);
2591 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues
2592 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed.
2594 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq)
2599 disabling = txq < dev->real_num_tx_queues;
2601 if (txq < 1 || txq > dev->num_tx_queues)
2604 if (dev->reg_state == NETREG_REGISTERED ||
2605 dev->reg_state == NETREG_UNREGISTERING) {
2608 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues,
2614 netif_setup_tc(dev, txq);
2616 dev->real_num_tx_queues = txq;
2620 qdisc_reset_all_tx_gt(dev, txq);
2622 netif_reset_xps_queues_gt(dev, txq);
2626 dev->real_num_tx_queues = txq;
2631 EXPORT_SYMBOL(netif_set_real_num_tx_queues);
2635 * netif_set_real_num_rx_queues - set actual number of RX queues used
2636 * @dev: Network device
2637 * @rxq: Actual number of RX queues
2639 * This must be called either with the rtnl_lock held or before
2640 * registration of the net device. Returns 0 on success, or a
2641 * negative error code. If called before registration, it always
2644 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq)
2648 if (rxq < 1 || rxq > dev->num_rx_queues)
2651 if (dev->reg_state == NETREG_REGISTERED) {
2654 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues,
2660 dev->real_num_rx_queues = rxq;
2663 EXPORT_SYMBOL(netif_set_real_num_rx_queues);
2667 * netif_get_num_default_rss_queues - default number of RSS queues
2669 * This routine should set an upper limit on the number of RSS queues
2670 * used by default by multiqueue devices.
2672 int netif_get_num_default_rss_queues(void)
2674 return is_kdump_kernel() ?
2675 1 : min_t(int, DEFAULT_MAX_NUM_RSS_QUEUES, num_online_cpus());
2677 EXPORT_SYMBOL(netif_get_num_default_rss_queues);
2679 static void __netif_reschedule(struct Qdisc *q)
2681 struct softnet_data *sd;
2682 unsigned long flags;
2684 local_irq_save(flags);
2685 sd = this_cpu_ptr(&softnet_data);
2686 q->next_sched = NULL;
2687 *sd->output_queue_tailp = q;
2688 sd->output_queue_tailp = &q->next_sched;
2689 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2690 local_irq_restore(flags);
2693 void __netif_schedule(struct Qdisc *q)
2695 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
2696 __netif_reschedule(q);
2698 EXPORT_SYMBOL(__netif_schedule);
2700 struct dev_kfree_skb_cb {
2701 enum skb_free_reason reason;
2704 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb)
2706 return (struct dev_kfree_skb_cb *)skb->cb;
2709 void netif_schedule_queue(struct netdev_queue *txq)
2712 if (!(txq->state & QUEUE_STATE_ANY_XOFF)) {
2713 struct Qdisc *q = rcu_dereference(txq->qdisc);
2715 __netif_schedule(q);
2719 EXPORT_SYMBOL(netif_schedule_queue);
2721 void netif_tx_wake_queue(struct netdev_queue *dev_queue)
2723 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) {
2727 q = rcu_dereference(dev_queue->qdisc);
2728 __netif_schedule(q);
2732 EXPORT_SYMBOL(netif_tx_wake_queue);
2734 void __dev_kfree_skb_irq(struct sk_buff *skb, enum skb_free_reason reason)
2736 unsigned long flags;
2741 if (likely(refcount_read(&skb->users) == 1)) {
2743 refcount_set(&skb->users, 0);
2744 } else if (likely(!refcount_dec_and_test(&skb->users))) {
2747 get_kfree_skb_cb(skb)->reason = reason;
2748 local_irq_save(flags);
2749 skb->next = __this_cpu_read(softnet_data.completion_queue);
2750 __this_cpu_write(softnet_data.completion_queue, skb);
2751 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2752 local_irq_restore(flags);
2754 EXPORT_SYMBOL(__dev_kfree_skb_irq);
2756 void __dev_kfree_skb_any(struct sk_buff *skb, enum skb_free_reason reason)
2758 if (in_irq() || irqs_disabled())
2759 __dev_kfree_skb_irq(skb, reason);
2763 EXPORT_SYMBOL(__dev_kfree_skb_any);
2767 * netif_device_detach - mark device as removed
2768 * @dev: network device
2770 * Mark device as removed from system and therefore no longer available.
2772 void netif_device_detach(struct net_device *dev)
2774 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) &&
2775 netif_running(dev)) {
2776 netif_tx_stop_all_queues(dev);
2779 EXPORT_SYMBOL(netif_device_detach);
2782 * netif_device_attach - mark device as attached
2783 * @dev: network device
2785 * Mark device as attached from system and restart if needed.
2787 void netif_device_attach(struct net_device *dev)
2789 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) &&
2790 netif_running(dev)) {
2791 netif_tx_wake_all_queues(dev);
2792 __netdev_watchdog_up(dev);
2795 EXPORT_SYMBOL(netif_device_attach);
2798 * Returns a Tx hash based on the given packet descriptor a Tx queues' number
2799 * to be used as a distribution range.
2801 static u16 skb_tx_hash(const struct net_device *dev,
2802 const struct net_device *sb_dev,
2803 struct sk_buff *skb)
2807 u16 qcount = dev->real_num_tx_queues;
2810 u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
2812 qoffset = sb_dev->tc_to_txq[tc].offset;
2813 qcount = sb_dev->tc_to_txq[tc].count;
2816 if (skb_rx_queue_recorded(skb)) {
2817 hash = skb_get_rx_queue(skb);
2818 while (unlikely(hash >= qcount))
2820 return hash + qoffset;
2823 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset;
2826 static void skb_warn_bad_offload(const struct sk_buff *skb)
2828 static const netdev_features_t null_features;
2829 struct net_device *dev = skb->dev;
2830 const char *name = "";
2832 if (!net_ratelimit())
2836 if (dev->dev.parent)
2837 name = dev_driver_string(dev->dev.parent);
2839 name = netdev_name(dev);
2841 WARN(1, "%s: caps=(%pNF, %pNF) len=%d data_len=%d gso_size=%d "
2842 "gso_type=%d ip_summed=%d\n",
2843 name, dev ? &dev->features : &null_features,
2844 skb->sk ? &skb->sk->sk_route_caps : &null_features,
2845 skb->len, skb->data_len, skb_shinfo(skb)->gso_size,
2846 skb_shinfo(skb)->gso_type, skb->ip_summed);
2850 * Invalidate hardware checksum when packet is to be mangled, and
2851 * complete checksum manually on outgoing path.
2853 int skb_checksum_help(struct sk_buff *skb)
2856 int ret = 0, offset;
2858 if (skb->ip_summed == CHECKSUM_COMPLETE)
2859 goto out_set_summed;
2861 if (unlikely(skb_shinfo(skb)->gso_size)) {
2862 skb_warn_bad_offload(skb);
2866 /* Before computing a checksum, we should make sure no frag could
2867 * be modified by an external entity : checksum could be wrong.
2869 if (skb_has_shared_frag(skb)) {
2870 ret = __skb_linearize(skb);
2875 offset = skb_checksum_start_offset(skb);
2876 BUG_ON(offset >= skb_headlen(skb));
2877 csum = skb_checksum(skb, offset, skb->len - offset, 0);
2879 offset += skb->csum_offset;
2880 BUG_ON(offset + sizeof(__sum16) > skb_headlen(skb));
2882 if (skb_cloned(skb) &&
2883 !skb_clone_writable(skb, offset + sizeof(__sum16))) {
2884 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2889 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0;
2891 skb->ip_summed = CHECKSUM_NONE;
2895 EXPORT_SYMBOL(skb_checksum_help);
2897 int skb_crc32c_csum_help(struct sk_buff *skb)
2900 int ret = 0, offset, start;
2902 if (skb->ip_summed != CHECKSUM_PARTIAL)
2905 if (unlikely(skb_is_gso(skb)))
2908 /* Before computing a checksum, we should make sure no frag could
2909 * be modified by an external entity : checksum could be wrong.
2911 if (unlikely(skb_has_shared_frag(skb))) {
2912 ret = __skb_linearize(skb);
2916 start = skb_checksum_start_offset(skb);
2917 offset = start + offsetof(struct sctphdr, checksum);
2918 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) {
2922 if (skb_cloned(skb) &&
2923 !skb_clone_writable(skb, offset + sizeof(__le32))) {
2924 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2928 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start,
2929 skb->len - start, ~(__u32)0,
2931 *(__le32 *)(skb->data + offset) = crc32c_csum;
2932 skb->ip_summed = CHECKSUM_NONE;
2933 skb->csum_not_inet = 0;
2938 __be16 skb_network_protocol(struct sk_buff *skb, int *depth)
2940 __be16 type = skb->protocol;
2942 /* Tunnel gso handlers can set protocol to ethernet. */
2943 if (type == htons(ETH_P_TEB)) {
2946 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr))))
2949 eth = (struct ethhdr *)skb->data;
2950 type = eth->h_proto;
2953 return __vlan_get_protocol(skb, type, depth);
2957 * skb_mac_gso_segment - mac layer segmentation handler.
2958 * @skb: buffer to segment
2959 * @features: features for the output path (see dev->features)
2961 struct sk_buff *skb_mac_gso_segment(struct sk_buff *skb,
2962 netdev_features_t features)
2964 struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT);
2965 struct packet_offload *ptype;
2966 int vlan_depth = skb->mac_len;
2967 __be16 type = skb_network_protocol(skb, &vlan_depth);
2969 if (unlikely(!type))
2970 return ERR_PTR(-EINVAL);
2972 __skb_pull(skb, vlan_depth);
2975 list_for_each_entry_rcu(ptype, &offload_base, list) {
2976 if (ptype->type == type && ptype->callbacks.gso_segment) {
2977 segs = ptype->callbacks.gso_segment(skb, features);
2983 __skb_push(skb, skb->data - skb_mac_header(skb));
2987 EXPORT_SYMBOL(skb_mac_gso_segment);
2990 /* openvswitch calls this on rx path, so we need a different check.
2992 static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
2995 return skb->ip_summed != CHECKSUM_PARTIAL &&
2996 skb->ip_summed != CHECKSUM_UNNECESSARY;
2998 return skb->ip_summed == CHECKSUM_NONE;
3002 * __skb_gso_segment - Perform segmentation on skb.
3003 * @skb: buffer to segment
3004 * @features: features for the output path (see dev->features)
3005 * @tx_path: whether it is called in TX path
3007 * This function segments the given skb and returns a list of segments.
3009 * It may return NULL if the skb requires no segmentation. This is
3010 * only possible when GSO is used for verifying header integrity.
3012 * Segmentation preserves SKB_SGO_CB_OFFSET bytes of previous skb cb.
3014 struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
3015 netdev_features_t features, bool tx_path)
3017 struct sk_buff *segs;
3019 if (unlikely(skb_needs_check(skb, tx_path))) {
3022 /* We're going to init ->check field in TCP or UDP header */
3023 err = skb_cow_head(skb, 0);
3025 return ERR_PTR(err);
3028 /* Only report GSO partial support if it will enable us to
3029 * support segmentation on this frame without needing additional
3032 if (features & NETIF_F_GSO_PARTIAL) {
3033 netdev_features_t partial_features = NETIF_F_GSO_ROBUST;
3034 struct net_device *dev = skb->dev;
3036 partial_features |= dev->features & dev->gso_partial_features;
3037 if (!skb_gso_ok(skb, features | partial_features))
3038 features &= ~NETIF_F_GSO_PARTIAL;
3041 BUILD_BUG_ON(SKB_SGO_CB_OFFSET +
3042 sizeof(*SKB_GSO_CB(skb)) > sizeof(skb->cb));
3044 SKB_GSO_CB(skb)->mac_offset = skb_headroom(skb);
3045 SKB_GSO_CB(skb)->encap_level = 0;
3047 skb_reset_mac_header(skb);
3048 skb_reset_mac_len(skb);
3050 segs = skb_mac_gso_segment(skb, features);
3052 if (unlikely(skb_needs_check(skb, tx_path) && !IS_ERR(segs)))
3053 skb_warn_bad_offload(skb);
3057 EXPORT_SYMBOL(__skb_gso_segment);
3059 /* Take action when hardware reception checksum errors are detected. */
3061 void netdev_rx_csum_fault(struct net_device *dev)
3063 if (net_ratelimit()) {
3064 pr_err("%s: hw csum failure\n", dev ? dev->name : "<unknown>");
3068 EXPORT_SYMBOL(netdev_rx_csum_fault);
3071 /* XXX: check that highmem exists at all on the given machine. */
3072 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
3074 #ifdef CONFIG_HIGHMEM
3077 if (!(dev->features & NETIF_F_HIGHDMA)) {
3078 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
3079 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
3081 if (PageHighMem(skb_frag_page(frag)))
3089 /* If MPLS offload request, verify we are testing hardware MPLS features
3090 * instead of standard features for the netdev.
3092 #if IS_ENABLED(CONFIG_NET_MPLS_GSO)
3093 static netdev_features_t net_mpls_features(struct sk_buff *skb,
3094 netdev_features_t features,
3097 if (eth_p_mpls(type))
3098 features &= skb->dev->mpls_features;
3103 static netdev_features_t net_mpls_features(struct sk_buff *skb,
3104 netdev_features_t features,
3111 static netdev_features_t harmonize_features(struct sk_buff *skb,
3112 netdev_features_t features)
3117 type = skb_network_protocol(skb, &tmp);
3118 features = net_mpls_features(skb, features, type);
3120 if (skb->ip_summed != CHECKSUM_NONE &&
3121 !can_checksum_protocol(features, type)) {
3122 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
3124 if (illegal_highdma(skb->dev, skb))
3125 features &= ~NETIF_F_SG;
3130 netdev_features_t passthru_features_check(struct sk_buff *skb,
3131 struct net_device *dev,
3132 netdev_features_t features)
3136 EXPORT_SYMBOL(passthru_features_check);
3138 static netdev_features_t dflt_features_check(struct sk_buff *skb,
3139 struct net_device *dev,
3140 netdev_features_t features)
3142 return vlan_features_check(skb, features);
3145 static netdev_features_t gso_features_check(const struct sk_buff *skb,
3146 struct net_device *dev,
3147 netdev_features_t features)
3149 u16 gso_segs = skb_shinfo(skb)->gso_segs;
3151 if (gso_segs > dev->gso_max_segs)
3152 return features & ~NETIF_F_GSO_MASK;
3154 /* Support for GSO partial features requires software
3155 * intervention before we can actually process the packets
3156 * so we need to strip support for any partial features now
3157 * and we can pull them back in after we have partially
3158 * segmented the frame.
3160 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL))
3161 features &= ~dev->gso_partial_features;
3163 /* Make sure to clear the IPv4 ID mangling feature if the
3164 * IPv4 header has the potential to be fragmented.
3166 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
3167 struct iphdr *iph = skb->encapsulation ?
3168 inner_ip_hdr(skb) : ip_hdr(skb);
3170 if (!(iph->frag_off & htons(IP_DF)))
3171 features &= ~NETIF_F_TSO_MANGLEID;
3177 netdev_features_t netif_skb_features(struct sk_buff *skb)
3179 struct net_device *dev = skb->dev;
3180 netdev_features_t features = dev->features;
3182 if (skb_is_gso(skb))
3183 features = gso_features_check(skb, dev, features);
3185 /* If encapsulation offload request, verify we are testing
3186 * hardware encapsulation features instead of standard
3187 * features for the netdev
3189 if (skb->encapsulation)
3190 features &= dev->hw_enc_features;
3192 if (skb_vlan_tagged(skb))
3193 features = netdev_intersect_features(features,
3194 dev->vlan_features |
3195 NETIF_F_HW_VLAN_CTAG_TX |
3196 NETIF_F_HW_VLAN_STAG_TX);
3198 if (dev->netdev_ops->ndo_features_check)
3199 features &= dev->netdev_ops->ndo_features_check(skb, dev,
3202 features &= dflt_features_check(skb, dev, features);
3204 return harmonize_features(skb, features);
3206 EXPORT_SYMBOL(netif_skb_features);
3208 static int xmit_one(struct sk_buff *skb, struct net_device *dev,
3209 struct netdev_queue *txq, bool more)
3214 if (!list_empty(&ptype_all) || !list_empty(&dev->ptype_all))
3215 dev_queue_xmit_nit(skb, dev);
3218 trace_net_dev_start_xmit(skb, dev);
3219 rc = netdev_start_xmit(skb, dev, txq, more);
3220 trace_net_dev_xmit(skb, rc, dev, len);
3225 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev,
3226 struct netdev_queue *txq, int *ret)
3228 struct sk_buff *skb = first;
3229 int rc = NETDEV_TX_OK;
3232 struct sk_buff *next = skb->next;
3234 skb_mark_not_on_list(skb);
3235 rc = xmit_one(skb, dev, txq, next != NULL);
3236 if (unlikely(!dev_xmit_complete(rc))) {
3242 if (netif_xmit_stopped(txq) && skb) {
3243 rc = NETDEV_TX_BUSY;
3253 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb,
3254 netdev_features_t features)
3256 if (skb_vlan_tag_present(skb) &&
3257 !vlan_hw_offload_capable(features, skb->vlan_proto))
3258 skb = __vlan_hwaccel_push_inside(skb);
3262 int skb_csum_hwoffload_help(struct sk_buff *skb,
3263 const netdev_features_t features)
3265 if (unlikely(skb->csum_not_inet))
3266 return !!(features & NETIF_F_SCTP_CRC) ? 0 :
3267 skb_crc32c_csum_help(skb);
3269 return !!(features & NETIF_F_CSUM_MASK) ? 0 : skb_checksum_help(skb);
3271 EXPORT_SYMBOL(skb_csum_hwoffload_help);
3273 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again)
3275 netdev_features_t features;
3277 features = netif_skb_features(skb);
3278 skb = validate_xmit_vlan(skb, features);
3282 skb = sk_validate_xmit_skb(skb, dev);
3286 if (netif_needs_gso(skb, features)) {
3287 struct sk_buff *segs;
3289 segs = skb_gso_segment(skb, features);
3297 if (skb_needs_linearize(skb, features) &&
3298 __skb_linearize(skb))
3301 /* If packet is not checksummed and device does not
3302 * support checksumming for this protocol, complete
3303 * checksumming here.
3305 if (skb->ip_summed == CHECKSUM_PARTIAL) {
3306 if (skb->encapsulation)
3307 skb_set_inner_transport_header(skb,
3308 skb_checksum_start_offset(skb));
3310 skb_set_transport_header(skb,
3311 skb_checksum_start_offset(skb));
3312 if (skb_csum_hwoffload_help(skb, features))
3317 skb = validate_xmit_xfrm(skb, features, again);
3324 atomic_long_inc(&dev->tx_dropped);
3328 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again)
3330 struct sk_buff *next, *head = NULL, *tail;
3332 for (; skb != NULL; skb = next) {
3334 skb_mark_not_on_list(skb);
3336 /* in case skb wont be segmented, point to itself */
3339 skb = validate_xmit_skb(skb, dev, again);
3347 /* If skb was segmented, skb->prev points to
3348 * the last segment. If not, it still contains skb.
3354 EXPORT_SYMBOL_GPL(validate_xmit_skb_list);
3356 static void qdisc_pkt_len_init(struct sk_buff *skb)
3358 const struct skb_shared_info *shinfo = skb_shinfo(skb);
3360 qdisc_skb_cb(skb)->pkt_len = skb->len;
3362 /* To get more precise estimation of bytes sent on wire,
3363 * we add to pkt_len the headers size of all segments
3365 if (shinfo->gso_size) {
3366 unsigned int hdr_len;
3367 u16 gso_segs = shinfo->gso_segs;
3369 /* mac layer + network layer */
3370 hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
3372 /* + transport layer */
3373 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) {
3374 const struct tcphdr *th;
3375 struct tcphdr _tcphdr;
3377 th = skb_header_pointer(skb, skb_transport_offset(skb),
3378 sizeof(_tcphdr), &_tcphdr);
3380 hdr_len += __tcp_hdrlen(th);
3382 struct udphdr _udphdr;
3384 if (skb_header_pointer(skb, skb_transport_offset(skb),
3385 sizeof(_udphdr), &_udphdr))
3386 hdr_len += sizeof(struct udphdr);
3389 if (shinfo->gso_type & SKB_GSO_DODGY)
3390 gso_segs = DIV_ROUND_UP(skb->len - hdr_len,
3393 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len;
3397 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
3398 struct net_device *dev,
3399 struct netdev_queue *txq)
3401 spinlock_t *root_lock = qdisc_lock(q);
3402 struct sk_buff *to_free = NULL;
3406 qdisc_calculate_pkt_len(skb, q);
3408 if (q->flags & TCQ_F_NOLOCK) {
3409 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3410 __qdisc_drop(skb, &to_free);
3413 rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
3417 if (unlikely(to_free))
3418 kfree_skb_list(to_free);
3423 * Heuristic to force contended enqueues to serialize on a
3424 * separate lock before trying to get qdisc main lock.
3425 * This permits qdisc->running owner to get the lock more
3426 * often and dequeue packets faster.
3428 contended = qdisc_is_running(q);
3429 if (unlikely(contended))
3430 spin_lock(&q->busylock);
3432 spin_lock(root_lock);
3433 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3434 __qdisc_drop(skb, &to_free);
3436 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
3437 qdisc_run_begin(q)) {
3439 * This is a work-conserving queue; there are no old skbs
3440 * waiting to be sent out; and the qdisc is not running -
3441 * xmit the skb directly.
3444 qdisc_bstats_update(q, skb);
3446 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) {
3447 if (unlikely(contended)) {
3448 spin_unlock(&q->busylock);
3455 rc = NET_XMIT_SUCCESS;
3457 rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
3458 if (qdisc_run_begin(q)) {
3459 if (unlikely(contended)) {
3460 spin_unlock(&q->busylock);
3467 spin_unlock(root_lock);
3468 if (unlikely(to_free))
3469 kfree_skb_list(to_free);
3470 if (unlikely(contended))
3471 spin_unlock(&q->busylock);
3475 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO)
3476 static void skb_update_prio(struct sk_buff *skb)
3478 const struct netprio_map *map;
3479 const struct sock *sk;
3480 unsigned int prioidx;
3484 map = rcu_dereference_bh(skb->dev->priomap);
3487 sk = skb_to_full_sk(skb);
3491 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data);
3493 if (prioidx < map->priomap_len)
3494 skb->priority = map->priomap[prioidx];
3497 #define skb_update_prio(skb)
3500 DEFINE_PER_CPU(int, xmit_recursion);
3501 EXPORT_SYMBOL(xmit_recursion);
3504 * dev_loopback_xmit - loop back @skb
3505 * @net: network namespace this loopback is happening in
3506 * @sk: sk needed to be a netfilter okfn
3507 * @skb: buffer to transmit
3509 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb)
3511 skb_reset_mac_header(skb);
3512 __skb_pull(skb, skb_network_offset(skb));
3513 skb->pkt_type = PACKET_LOOPBACK;
3514 skb->ip_summed = CHECKSUM_UNNECESSARY;
3515 WARN_ON(!skb_dst(skb));
3520 EXPORT_SYMBOL(dev_loopback_xmit);
3522 #ifdef CONFIG_NET_EGRESS
3523 static struct sk_buff *
3524 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
3526 struct mini_Qdisc *miniq = rcu_dereference_bh(dev->miniq_egress);
3527 struct tcf_result cl_res;
3532 /* qdisc_skb_cb(skb)->pkt_len was already set by the caller. */
3533 mini_qdisc_bstats_cpu_update(miniq, skb);
3535 switch (tcf_classify(skb, miniq->filter_list, &cl_res, false)) {
3537 case TC_ACT_RECLASSIFY:
3538 skb->tc_index = TC_H_MIN(cl_res.classid);
3541 mini_qdisc_qstats_cpu_drop(miniq);
3542 *ret = NET_XMIT_DROP;
3548 *ret = NET_XMIT_SUCCESS;
3551 case TC_ACT_REDIRECT:
3552 /* No need to push/pop skb's mac_header here on egress! */
3553 skb_do_redirect(skb);
3554 *ret = NET_XMIT_SUCCESS;
3562 #endif /* CONFIG_NET_EGRESS */
3565 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb,
3566 struct xps_dev_maps *dev_maps, unsigned int tci)
3568 struct xps_map *map;
3569 int queue_index = -1;
3573 tci += netdev_get_prio_tc_map(dev, skb->priority);
3576 map = rcu_dereference(dev_maps->attr_map[tci]);
3579 queue_index = map->queues[0];
3581 queue_index = map->queues[reciprocal_scale(
3582 skb_get_hash(skb), map->len)];
3583 if (unlikely(queue_index >= dev->real_num_tx_queues))
3590 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev,
3591 struct sk_buff *skb)
3594 struct xps_dev_maps *dev_maps;
3595 struct sock *sk = skb->sk;
3596 int queue_index = -1;
3598 if (!static_key_false(&xps_needed))
3602 if (!static_key_false(&xps_rxqs_needed))
3605 dev_maps = rcu_dereference(sb_dev->xps_rxqs_map);
3607 int tci = sk_rx_queue_get(sk);
3609 if (tci >= 0 && tci < dev->num_rx_queues)
3610 queue_index = __get_xps_queue_idx(dev, skb, dev_maps,
3615 if (queue_index < 0) {
3616 dev_maps = rcu_dereference(sb_dev->xps_cpus_map);
3618 unsigned int tci = skb->sender_cpu - 1;
3620 queue_index = __get_xps_queue_idx(dev, skb, dev_maps,
3632 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb,
3633 struct net_device *sb_dev,
3634 select_queue_fallback_t fallback)
3638 EXPORT_SYMBOL(dev_pick_tx_zero);
3640 u16 dev_pick_tx_cpu_id(struct net_device *dev, struct sk_buff *skb,
3641 struct net_device *sb_dev,
3642 select_queue_fallback_t fallback)
3644 return (u16)raw_smp_processor_id() % dev->real_num_tx_queues;
3646 EXPORT_SYMBOL(dev_pick_tx_cpu_id);
3648 static u16 __netdev_pick_tx(struct net_device *dev, struct sk_buff *skb,
3649 struct net_device *sb_dev)
3651 struct sock *sk = skb->sk;
3652 int queue_index = sk_tx_queue_get(sk);
3654 sb_dev = sb_dev ? : dev;
3656 if (queue_index < 0 || skb->ooo_okay ||
3657 queue_index >= dev->real_num_tx_queues) {
3658 int new_index = get_xps_queue(dev, sb_dev, skb);
3661 new_index = skb_tx_hash(dev, sb_dev, skb);
3663 if (queue_index != new_index && sk &&
3665 rcu_access_pointer(sk->sk_dst_cache))
3666 sk_tx_queue_set(sk, new_index);
3668 queue_index = new_index;
3674 struct netdev_queue *netdev_pick_tx(struct net_device *dev,
3675 struct sk_buff *skb,
3676 struct net_device *sb_dev)
3678 int queue_index = 0;
3681 u32 sender_cpu = skb->sender_cpu - 1;
3683 if (sender_cpu >= (u32)NR_CPUS)
3684 skb->sender_cpu = raw_smp_processor_id() + 1;
3687 if (dev->real_num_tx_queues != 1) {
3688 const struct net_device_ops *ops = dev->netdev_ops;
3690 if (ops->ndo_select_queue)
3691 queue_index = ops->ndo_select_queue(dev, skb, sb_dev,
3694 queue_index = __netdev_pick_tx(dev, skb, sb_dev);
3696 queue_index = netdev_cap_txqueue(dev, queue_index);
3699 skb_set_queue_mapping(skb, queue_index);
3700 return netdev_get_tx_queue(dev, queue_index);
3704 * __dev_queue_xmit - transmit a buffer
3705 * @skb: buffer to transmit
3706 * @sb_dev: suboordinate device used for L2 forwarding offload
3708 * Queue a buffer for transmission to a network device. The caller must
3709 * have set the device and priority and built the buffer before calling
3710 * this function. The function can be called from an interrupt.
3712 * A negative errno code is returned on a failure. A success does not
3713 * guarantee the frame will be transmitted as it may be dropped due
3714 * to congestion or traffic shaping.
3716 * -----------------------------------------------------------------------------------
3717 * I notice this method can also return errors from the queue disciplines,
3718 * including NET_XMIT_DROP, which is a positive value. So, errors can also
3721 * Regardless of the return value, the skb is consumed, so it is currently
3722 * difficult to retry a send to this method. (You can bump the ref count
3723 * before sending to hold a reference for retry if you are careful.)
3725 * When calling this method, interrupts MUST be enabled. This is because
3726 * the BH enable code must have IRQs enabled so that it will not deadlock.
3729 static int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev)
3731 struct net_device *dev = skb->dev;
3732 struct netdev_queue *txq;
3737 skb_reset_mac_header(skb);
3739 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP))
3740 __skb_tstamp_tx(skb, NULL, skb->sk, SCM_TSTAMP_SCHED);
3742 /* Disable soft irqs for various locks below. Also
3743 * stops preemption for RCU.
3747 skb_update_prio(skb);
3749 qdisc_pkt_len_init(skb);
3750 #ifdef CONFIG_NET_CLS_ACT
3751 skb->tc_at_ingress = 0;
3752 # ifdef CONFIG_NET_EGRESS
3753 if (static_branch_unlikely(&egress_needed_key)) {
3754 skb = sch_handle_egress(skb, &rc, dev);
3760 /* If device/qdisc don't need skb->dst, release it right now while
3761 * its hot in this cpu cache.
3763 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
3768 txq = netdev_pick_tx(dev, skb, sb_dev);
3769 q = rcu_dereference_bh(txq->qdisc);
3771 trace_net_dev_queue(skb);
3773 rc = __dev_xmit_skb(skb, q, dev, txq);
3777 /* The device has no queue. Common case for software devices:
3778 * loopback, all the sorts of tunnels...
3780 * Really, it is unlikely that netif_tx_lock protection is necessary
3781 * here. (f.e. loopback and IP tunnels are clean ignoring statistics
3783 * However, it is possible, that they rely on protection
3786 * Check this and shot the lock. It is not prone from deadlocks.
3787 *Either shot noqueue qdisc, it is even simpler 8)
3789 if (dev->flags & IFF_UP) {
3790 int cpu = smp_processor_id(); /* ok because BHs are off */
3792 if (txq->xmit_lock_owner != cpu) {
3793 if (unlikely(__this_cpu_read(xmit_recursion) >
3794 XMIT_RECURSION_LIMIT))
3795 goto recursion_alert;
3797 skb = validate_xmit_skb(skb, dev, &again);
3801 HARD_TX_LOCK(dev, txq, cpu);
3803 if (!netif_xmit_stopped(txq)) {
3804 __this_cpu_inc(xmit_recursion);
3805 skb = dev_hard_start_xmit(skb, dev, txq, &rc);
3806 __this_cpu_dec(xmit_recursion);
3807 if (dev_xmit_complete(rc)) {
3808 HARD_TX_UNLOCK(dev, txq);
3812 HARD_TX_UNLOCK(dev, txq);
3813 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
3816 /* Recursion is detected! It is possible,
3820 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
3826 rcu_read_unlock_bh();
3828 atomic_long_inc(&dev->tx_dropped);
3829 kfree_skb_list(skb);
3832 rcu_read_unlock_bh();
3836 int dev_queue_xmit(struct sk_buff *skb)
3838 return __dev_queue_xmit(skb, NULL);
3840 EXPORT_SYMBOL(dev_queue_xmit);
3842 int dev_queue_xmit_accel(struct sk_buff *skb, struct net_device *sb_dev)
3844 return __dev_queue_xmit(skb, sb_dev);
3846 EXPORT_SYMBOL(dev_queue_xmit_accel);
3848 int dev_direct_xmit(struct sk_buff *skb, u16 queue_id)
3850 struct net_device *dev = skb->dev;
3851 struct sk_buff *orig_skb = skb;
3852 struct netdev_queue *txq;
3853 int ret = NETDEV_TX_BUSY;
3856 if (unlikely(!netif_running(dev) ||
3857 !netif_carrier_ok(dev)))
3860 skb = validate_xmit_skb_list(skb, dev, &again);
3861 if (skb != orig_skb)
3864 skb_set_queue_mapping(skb, queue_id);
3865 txq = skb_get_tx_queue(dev, skb);
3869 HARD_TX_LOCK(dev, txq, smp_processor_id());
3870 if (!netif_xmit_frozen_or_drv_stopped(txq))
3871 ret = netdev_start_xmit(skb, dev, txq, false);
3872 HARD_TX_UNLOCK(dev, txq);
3876 if (!dev_xmit_complete(ret))
3881 atomic_long_inc(&dev->tx_dropped);
3882 kfree_skb_list(skb);
3883 return NET_XMIT_DROP;
3885 EXPORT_SYMBOL(dev_direct_xmit);
3887 /*************************************************************************
3889 *************************************************************************/
3891 int netdev_max_backlog __read_mostly = 1000;
3892 EXPORT_SYMBOL(netdev_max_backlog);
3894 int netdev_tstamp_prequeue __read_mostly = 1;
3895 int netdev_budget __read_mostly = 300;
3896 unsigned int __read_mostly netdev_budget_usecs = 2000;
3897 int weight_p __read_mostly = 64; /* old backlog weight */
3898 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */
3899 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */
3900 int dev_rx_weight __read_mostly = 64;
3901 int dev_tx_weight __read_mostly = 64;
3903 /* Called with irq disabled */
3904 static inline void ____napi_schedule(struct softnet_data *sd,
3905 struct napi_struct *napi)
3907 list_add_tail(&napi->poll_list, &sd->poll_list);
3908 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3913 /* One global table that all flow-based protocols share. */
3914 struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
3915 EXPORT_SYMBOL(rps_sock_flow_table);
3916 u32 rps_cpu_mask __read_mostly;
3917 EXPORT_SYMBOL(rps_cpu_mask);
3919 struct static_key rps_needed __read_mostly;
3920 EXPORT_SYMBOL(rps_needed);
3921 struct static_key rfs_needed __read_mostly;
3922 EXPORT_SYMBOL(rfs_needed);
3924 static struct rps_dev_flow *
3925 set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3926 struct rps_dev_flow *rflow, u16 next_cpu)
3928 if (next_cpu < nr_cpu_ids) {
3929 #ifdef CONFIG_RFS_ACCEL
3930 struct netdev_rx_queue *rxqueue;
3931 struct rps_dev_flow_table *flow_table;
3932 struct rps_dev_flow *old_rflow;
3937 /* Should we steer this flow to a different hardware queue? */
3938 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
3939 !(dev->features & NETIF_F_NTUPLE))
3941 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
3942 if (rxq_index == skb_get_rx_queue(skb))
3945 rxqueue = dev->_rx + rxq_index;
3946 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3949 flow_id = skb_get_hash(skb) & flow_table->mask;
3950 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
3951 rxq_index, flow_id);
3955 rflow = &flow_table->flows[flow_id];
3957 if (old_rflow->filter == rflow->filter)
3958 old_rflow->filter = RPS_NO_FILTER;
3962 per_cpu(softnet_data, next_cpu).input_queue_head;
3965 rflow->cpu = next_cpu;
3970 * get_rps_cpu is called from netif_receive_skb and returns the target
3971 * CPU from the RPS map of the receiving queue for a given skb.
3972 * rcu_read_lock must be held on entry.
3974 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3975 struct rps_dev_flow **rflowp)
3977 const struct rps_sock_flow_table *sock_flow_table;
3978 struct netdev_rx_queue *rxqueue = dev->_rx;
3979 struct rps_dev_flow_table *flow_table;
3980 struct rps_map *map;
3985 if (skb_rx_queue_recorded(skb)) {
3986 u16 index = skb_get_rx_queue(skb);
3988 if (unlikely(index >= dev->real_num_rx_queues)) {
3989 WARN_ONCE(dev->real_num_rx_queues > 1,
3990 "%s received packet on queue %u, but number "
3991 "of RX queues is %u\n",
3992 dev->name, index, dev->real_num_rx_queues);
3998 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */
4000 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4001 map = rcu_dereference(rxqueue->rps_map);
4002 if (!flow_table && !map)
4005 skb_reset_network_header(skb);
4006 hash = skb_get_hash(skb);
4010 sock_flow_table = rcu_dereference(rps_sock_flow_table);
4011 if (flow_table && sock_flow_table) {
4012 struct rps_dev_flow *rflow;
4016 /* First check into global flow table if there is a match */
4017 ident = sock_flow_table->ents[hash & sock_flow_table->mask];
4018 if ((ident ^ hash) & ~rps_cpu_mask)
4021 next_cpu = ident & rps_cpu_mask;
4023 /* OK, now we know there is a match,
4024 * we can look at the local (per receive queue) flow table
4026 rflow = &flow_table->flows[hash & flow_table->mask];
4030 * If the desired CPU (where last recvmsg was done) is
4031 * different from current CPU (one in the rx-queue flow
4032 * table entry), switch if one of the following holds:
4033 * - Current CPU is unset (>= nr_cpu_ids).
4034 * - Current CPU is offline.
4035 * - The current CPU's queue tail has advanced beyond the
4036 * last packet that was enqueued using this table entry.
4037 * This guarantees that all previous packets for the flow
4038 * have been dequeued, thus preserving in order delivery.
4040 if (unlikely(tcpu != next_cpu) &&
4041 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) ||
4042 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
4043 rflow->last_qtail)) >= 0)) {
4045 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
4048 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) {
4058 tcpu = map->cpus[reciprocal_scale(hash, map->len)];
4059 if (cpu_online(tcpu)) {
4069 #ifdef CONFIG_RFS_ACCEL
4072 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
4073 * @dev: Device on which the filter was set
4074 * @rxq_index: RX queue index
4075 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
4076 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
4078 * Drivers that implement ndo_rx_flow_steer() should periodically call
4079 * this function for each installed filter and remove the filters for
4080 * which it returns %true.
4082 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
4083 u32 flow_id, u16 filter_id)
4085 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
4086 struct rps_dev_flow_table *flow_table;
4087 struct rps_dev_flow *rflow;
4092 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4093 if (flow_table && flow_id <= flow_table->mask) {
4094 rflow = &flow_table->flows[flow_id];
4095 cpu = READ_ONCE(rflow->cpu);
4096 if (rflow->filter == filter_id && cpu < nr_cpu_ids &&
4097 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
4098 rflow->last_qtail) <
4099 (int)(10 * flow_table->mask)))
4105 EXPORT_SYMBOL(rps_may_expire_flow);
4107 #endif /* CONFIG_RFS_ACCEL */
4109 /* Called from hardirq (IPI) context */
4110 static void rps_trigger_softirq(void *data)
4112 struct softnet_data *sd = data;
4114 ____napi_schedule(sd, &sd->backlog);
4118 #endif /* CONFIG_RPS */
4121 * Check if this softnet_data structure is another cpu one
4122 * If yes, queue it to our IPI list and return 1
4125 static int rps_ipi_queued(struct softnet_data *sd)
4128 struct softnet_data *mysd = this_cpu_ptr(&softnet_data);
4131 sd->rps_ipi_next = mysd->rps_ipi_list;
4132 mysd->rps_ipi_list = sd;
4134 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
4137 #endif /* CONFIG_RPS */
4141 #ifdef CONFIG_NET_FLOW_LIMIT
4142 int netdev_flow_limit_table_len __read_mostly = (1 << 12);
4145 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
4147 #ifdef CONFIG_NET_FLOW_LIMIT
4148 struct sd_flow_limit *fl;
4149 struct softnet_data *sd;
4150 unsigned int old_flow, new_flow;
4152 if (qlen < (netdev_max_backlog >> 1))
4155 sd = this_cpu_ptr(&softnet_data);
4158 fl = rcu_dereference(sd->flow_limit);
4160 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1);
4161 old_flow = fl->history[fl->history_head];
4162 fl->history[fl->history_head] = new_flow;
4165 fl->history_head &= FLOW_LIMIT_HISTORY - 1;
4167 if (likely(fl->buckets[old_flow]))
4168 fl->buckets[old_flow]--;
4170 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) {
4182 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
4183 * queue (may be a remote CPU queue).
4185 static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
4186 unsigned int *qtail)
4188 struct softnet_data *sd;
4189 unsigned long flags;
4192 sd = &per_cpu(softnet_data, cpu);
4194 local_irq_save(flags);
4197 if (!netif_running(skb->dev))
4199 qlen = skb_queue_len(&sd->input_pkt_queue);
4200 if (qlen <= netdev_max_backlog && !skb_flow_limit(skb, qlen)) {
4203 __skb_queue_tail(&sd->input_pkt_queue, skb);
4204 input_queue_tail_incr_save(sd, qtail);
4206 local_irq_restore(flags);
4207 return NET_RX_SUCCESS;
4210 /* Schedule NAPI for backlog device
4211 * We can use non atomic operation since we own the queue lock
4213 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
4214 if (!rps_ipi_queued(sd))
4215 ____napi_schedule(sd, &sd->backlog);
4224 local_irq_restore(flags);
4226 atomic_long_inc(&skb->dev->rx_dropped);
4231 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb)
4233 struct net_device *dev = skb->dev;
4234 struct netdev_rx_queue *rxqueue;
4238 if (skb_rx_queue_recorded(skb)) {
4239 u16 index = skb_get_rx_queue(skb);
4241 if (unlikely(index >= dev->real_num_rx_queues)) {
4242 WARN_ONCE(dev->real_num_rx_queues > 1,
4243 "%s received packet on queue %u, but number "
4244 "of RX queues is %u\n",
4245 dev->name, index, dev->real_num_rx_queues);
4247 return rxqueue; /* Return first rxqueue */
4254 static u32 netif_receive_generic_xdp(struct sk_buff *skb,
4255 struct xdp_buff *xdp,
4256 struct bpf_prog *xdp_prog)
4258 struct netdev_rx_queue *rxqueue;
4259 void *orig_data, *orig_data_end;
4260 u32 metalen, act = XDP_DROP;
4261 __be16 orig_eth_type;
4267 /* Reinjected packets coming from act_mirred or similar should
4268 * not get XDP generic processing.
4270 if (skb_cloned(skb) || skb_is_tc_redirected(skb))
4273 /* XDP packets must be linear and must have sufficient headroom
4274 * of XDP_PACKET_HEADROOM bytes. This is the guarantee that also
4275 * native XDP provides, thus we need to do it here as well.
4277 if (skb_is_nonlinear(skb) ||
4278 skb_headroom(skb) < XDP_PACKET_HEADROOM) {
4279 int hroom = XDP_PACKET_HEADROOM - skb_headroom(skb);
4280 int troom = skb->tail + skb->data_len - skb->end;
4282 /* In case we have to go down the path and also linearize,
4283 * then lets do the pskb_expand_head() work just once here.
4285 if (pskb_expand_head(skb,
4286 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0,
4287 troom > 0 ? troom + 128 : 0, GFP_ATOMIC))
4289 if (skb_linearize(skb))
4293 /* The XDP program wants to see the packet starting at the MAC
4296 mac_len = skb->data - skb_mac_header(skb);
4297 hlen = skb_headlen(skb) + mac_len;
4298 xdp->data = skb->data - mac_len;
4299 xdp->data_meta = xdp->data;
4300 xdp->data_end = xdp->data + hlen;
4301 xdp->data_hard_start = skb->data - skb_headroom(skb);
4302 orig_data_end = xdp->data_end;
4303 orig_data = xdp->data;
4304 eth = (struct ethhdr *)xdp->data;
4305 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest);
4306 orig_eth_type = eth->h_proto;
4308 rxqueue = netif_get_rxqueue(skb);
4309 xdp->rxq = &rxqueue->xdp_rxq;
4311 act = bpf_prog_run_xdp(xdp_prog, xdp);
4313 off = xdp->data - orig_data;
4315 __skb_pull(skb, off);
4317 __skb_push(skb, -off);
4318 skb->mac_header += off;
4320 /* check if bpf_xdp_adjust_tail was used. it can only "shrink"
4323 off = orig_data_end - xdp->data_end;
4325 skb_set_tail_pointer(skb, xdp->data_end - xdp->data);
4330 /* check if XDP changed eth hdr such SKB needs update */
4331 eth = (struct ethhdr *)xdp->data;
4332 if ((orig_eth_type != eth->h_proto) ||
4333 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) {
4334 __skb_push(skb, ETH_HLEN);
4335 skb->protocol = eth_type_trans(skb, skb->dev);
4341 __skb_push(skb, mac_len);
4344 metalen = xdp->data - xdp->data_meta;
4346 skb_metadata_set(skb, metalen);
4349 bpf_warn_invalid_xdp_action(act);
4352 trace_xdp_exception(skb->dev, xdp_prog, act);
4363 /* When doing generic XDP we have to bypass the qdisc layer and the
4364 * network taps in order to match in-driver-XDP behavior.
4366 void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog)
4368 struct net_device *dev = skb->dev;
4369 struct netdev_queue *txq;
4370 bool free_skb = true;
4373 txq = netdev_pick_tx(dev, skb, NULL);
4374 cpu = smp_processor_id();
4375 HARD_TX_LOCK(dev, txq, cpu);
4376 if (!netif_xmit_stopped(txq)) {
4377 rc = netdev_start_xmit(skb, dev, txq, 0);
4378 if (dev_xmit_complete(rc))
4381 HARD_TX_UNLOCK(dev, txq);
4383 trace_xdp_exception(dev, xdp_prog, XDP_TX);
4387 EXPORT_SYMBOL_GPL(generic_xdp_tx);
4389 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key);
4391 int do_xdp_generic(struct bpf_prog *xdp_prog, struct sk_buff *skb)
4394 struct xdp_buff xdp;
4398 act = netif_receive_generic_xdp(skb, &xdp, xdp_prog);
4399 if (act != XDP_PASS) {
4402 err = xdp_do_generic_redirect(skb->dev, skb,
4408 generic_xdp_tx(skb, xdp_prog);
4419 EXPORT_SYMBOL_GPL(do_xdp_generic);
4421 static int netif_rx_internal(struct sk_buff *skb)
4425 net_timestamp_check(netdev_tstamp_prequeue, skb);
4427 trace_netif_rx(skb);
4429 if (static_branch_unlikely(&generic_xdp_needed_key)) {
4434 ret = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
4438 /* Consider XDP consuming the packet a success from
4439 * the netdev point of view we do not want to count
4442 if (ret != XDP_PASS)
4443 return NET_RX_SUCCESS;
4447 if (static_key_false(&rps_needed)) {
4448 struct rps_dev_flow voidflow, *rflow = &voidflow;
4454 cpu = get_rps_cpu(skb->dev, skb, &rflow);
4456 cpu = smp_processor_id();
4458 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4467 ret = enqueue_to_backlog(skb, get_cpu(), &qtail);
4474 * netif_rx - post buffer to the network code
4475 * @skb: buffer to post
4477 * This function receives a packet from a device driver and queues it for
4478 * the upper (protocol) levels to process. It always succeeds. The buffer
4479 * may be dropped during processing for congestion control or by the
4483 * NET_RX_SUCCESS (no congestion)
4484 * NET_RX_DROP (packet was dropped)
4488 int netif_rx(struct sk_buff *skb)
4490 trace_netif_rx_entry(skb);
4492 return netif_rx_internal(skb);
4494 EXPORT_SYMBOL(netif_rx);
4496 int netif_rx_ni(struct sk_buff *skb)
4500 trace_netif_rx_ni_entry(skb);
4503 err = netif_rx_internal(skb);
4504 if (local_softirq_pending())
4510 EXPORT_SYMBOL(netif_rx_ni);
4512 static __latent_entropy void net_tx_action(struct softirq_action *h)
4514 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
4516 if (sd->completion_queue) {
4517 struct sk_buff *clist;
4519 local_irq_disable();
4520 clist = sd->completion_queue;
4521 sd->completion_queue = NULL;
4525 struct sk_buff *skb = clist;
4527 clist = clist->next;
4529 WARN_ON(refcount_read(&skb->users));
4530 if (likely(get_kfree_skb_cb(skb)->reason == SKB_REASON_CONSUMED))
4531 trace_consume_skb(skb);
4533 trace_kfree_skb(skb, net_tx_action);
4535 if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
4538 __kfree_skb_defer(skb);
4541 __kfree_skb_flush();
4544 if (sd->output_queue) {
4547 local_irq_disable();
4548 head = sd->output_queue;
4549 sd->output_queue = NULL;
4550 sd->output_queue_tailp = &sd->output_queue;
4554 struct Qdisc *q = head;
4555 spinlock_t *root_lock = NULL;
4557 head = head->next_sched;
4559 if (!(q->flags & TCQ_F_NOLOCK)) {
4560 root_lock = qdisc_lock(q);
4561 spin_lock(root_lock);
4563 /* We need to make sure head->next_sched is read
4564 * before clearing __QDISC_STATE_SCHED
4566 smp_mb__before_atomic();
4567 clear_bit(__QDISC_STATE_SCHED, &q->state);
4570 spin_unlock(root_lock);
4574 xfrm_dev_backlog(sd);
4577 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE)
4578 /* This hook is defined here for ATM LANE */
4579 int (*br_fdb_test_addr_hook)(struct net_device *dev,
4580 unsigned char *addr) __read_mostly;
4581 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook);
4584 static inline struct sk_buff *
4585 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
4586 struct net_device *orig_dev)
4588 #ifdef CONFIG_NET_CLS_ACT
4589 struct mini_Qdisc *miniq = rcu_dereference_bh(skb->dev->miniq_ingress);
4590 struct tcf_result cl_res;
4592 /* If there's at least one ingress present somewhere (so
4593 * we get here via enabled static key), remaining devices
4594 * that are not configured with an ingress qdisc will bail
4601 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4605 qdisc_skb_cb(skb)->pkt_len = skb->len;
4606 skb->tc_at_ingress = 1;
4607 mini_qdisc_bstats_cpu_update(miniq, skb);
4609 switch (tcf_classify(skb, miniq->filter_list, &cl_res, false)) {
4611 case TC_ACT_RECLASSIFY:
4612 skb->tc_index = TC_H_MIN(cl_res.classid);
4615 mini_qdisc_qstats_cpu_drop(miniq);
4623 case TC_ACT_REDIRECT:
4624 /* skb_mac_header check was done by cls/act_bpf, so
4625 * we can safely push the L2 header back before
4626 * redirecting to another netdev
4628 __skb_push(skb, skb->mac_len);
4629 skb_do_redirect(skb);
4631 case TC_ACT_REINSERT:
4632 /* this does not scrub the packet, and updates stats on error */
4633 skb_tc_reinsert(skb, &cl_res);
4638 #endif /* CONFIG_NET_CLS_ACT */
4643 * netdev_is_rx_handler_busy - check if receive handler is registered
4644 * @dev: device to check
4646 * Check if a receive handler is already registered for a given device.
4647 * Return true if there one.
4649 * The caller must hold the rtnl_mutex.
4651 bool netdev_is_rx_handler_busy(struct net_device *dev)
4654 return dev && rtnl_dereference(dev->rx_handler);
4656 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy);
4659 * netdev_rx_handler_register - register receive handler
4660 * @dev: device to register a handler for
4661 * @rx_handler: receive handler to register
4662 * @rx_handler_data: data pointer that is used by rx handler
4664 * Register a receive handler for a device. This handler will then be
4665 * called from __netif_receive_skb. A negative errno code is returned
4668 * The caller must hold the rtnl_mutex.
4670 * For a general description of rx_handler, see enum rx_handler_result.
4672 int netdev_rx_handler_register(struct net_device *dev,
4673 rx_handler_func_t *rx_handler,
4674 void *rx_handler_data)
4676 if (netdev_is_rx_handler_busy(dev))
4679 if (dev->priv_flags & IFF_NO_RX_HANDLER)
4682 /* Note: rx_handler_data must be set before rx_handler */
4683 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data);
4684 rcu_assign_pointer(dev->rx_handler, rx_handler);
4688 EXPORT_SYMBOL_GPL(netdev_rx_handler_register);
4691 * netdev_rx_handler_unregister - unregister receive handler
4692 * @dev: device to unregister a handler from
4694 * Unregister a receive handler from a device.
4696 * The caller must hold the rtnl_mutex.
4698 void netdev_rx_handler_unregister(struct net_device *dev)
4702 RCU_INIT_POINTER(dev->rx_handler, NULL);
4703 /* a reader seeing a non NULL rx_handler in a rcu_read_lock()
4704 * section has a guarantee to see a non NULL rx_handler_data
4708 RCU_INIT_POINTER(dev->rx_handler_data, NULL);
4710 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister);
4713 * Limit the use of PFMEMALLOC reserves to those protocols that implement
4714 * the special handling of PFMEMALLOC skbs.
4716 static bool skb_pfmemalloc_protocol(struct sk_buff *skb)
4718 switch (skb->protocol) {
4719 case htons(ETH_P_ARP):
4720 case htons(ETH_P_IP):
4721 case htons(ETH_P_IPV6):
4722 case htons(ETH_P_8021Q):
4723 case htons(ETH_P_8021AD):
4730 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev,
4731 int *ret, struct net_device *orig_dev)
4733 #ifdef CONFIG_NETFILTER_INGRESS
4734 if (nf_hook_ingress_active(skb)) {
4738 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4743 ingress_retval = nf_hook_ingress(skb);
4745 return ingress_retval;
4747 #endif /* CONFIG_NETFILTER_INGRESS */
4751 static int __netif_receive_skb_core(struct sk_buff *skb, bool pfmemalloc,
4752 struct packet_type **ppt_prev)
4754 struct packet_type *ptype, *pt_prev;
4755 rx_handler_func_t *rx_handler;
4756 struct net_device *orig_dev;
4757 bool deliver_exact = false;
4758 int ret = NET_RX_DROP;
4761 net_timestamp_check(!netdev_tstamp_prequeue, skb);
4763 trace_netif_receive_skb(skb);
4765 orig_dev = skb->dev;
4767 skb_reset_network_header(skb);
4768 if (!skb_transport_header_was_set(skb))
4769 skb_reset_transport_header(skb);
4770 skb_reset_mac_len(skb);
4775 skb->skb_iif = skb->dev->ifindex;
4777 __this_cpu_inc(softnet_data.processed);
4779 if (skb->protocol == cpu_to_be16(ETH_P_8021Q) ||
4780 skb->protocol == cpu_to_be16(ETH_P_8021AD)) {
4781 skb = skb_vlan_untag(skb);
4786 if (skb_skip_tc_classify(skb))
4792 list_for_each_entry_rcu(ptype, &ptype_all, list) {
4794 ret = deliver_skb(skb, pt_prev, orig_dev);
4798 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) {
4800 ret = deliver_skb(skb, pt_prev, orig_dev);
4805 #ifdef CONFIG_NET_INGRESS
4806 if (static_branch_unlikely(&ingress_needed_key)) {
4807 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev);
4811 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0)
4817 if (pfmemalloc && !skb_pfmemalloc_protocol(skb))
4820 if (skb_vlan_tag_present(skb)) {
4822 ret = deliver_skb(skb, pt_prev, orig_dev);
4825 if (vlan_do_receive(&skb))
4827 else if (unlikely(!skb))
4831 rx_handler = rcu_dereference(skb->dev->rx_handler);
4834 ret = deliver_skb(skb, pt_prev, orig_dev);
4837 switch (rx_handler(&skb)) {
4838 case RX_HANDLER_CONSUMED:
4839 ret = NET_RX_SUCCESS;
4841 case RX_HANDLER_ANOTHER:
4843 case RX_HANDLER_EXACT:
4844 deliver_exact = true;
4845 case RX_HANDLER_PASS:
4852 if (unlikely(skb_vlan_tag_present(skb))) {
4853 if (skb_vlan_tag_get_id(skb))
4854 skb->pkt_type = PACKET_OTHERHOST;
4855 /* Note: we might in the future use prio bits
4856 * and set skb->priority like in vlan_do_receive()
4857 * For the time being, just ignore Priority Code Point
4862 type = skb->protocol;
4864 /* deliver only exact match when indicated */
4865 if (likely(!deliver_exact)) {
4866 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4867 &ptype_base[ntohs(type) &
4871 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4872 &orig_dev->ptype_specific);
4874 if (unlikely(skb->dev != orig_dev)) {
4875 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4876 &skb->dev->ptype_specific);
4880 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
4882 *ppt_prev = pt_prev;
4886 atomic_long_inc(&skb->dev->rx_dropped);
4888 atomic_long_inc(&skb->dev->rx_nohandler);
4890 /* Jamal, now you will not able to escape explaining
4891 * me how you were going to use this. :-)
4900 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc)
4902 struct net_device *orig_dev = skb->dev;
4903 struct packet_type *pt_prev = NULL;
4906 ret = __netif_receive_skb_core(skb, pfmemalloc, &pt_prev);
4908 ret = pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
4913 * netif_receive_skb_core - special purpose version of netif_receive_skb
4914 * @skb: buffer to process
4916 * More direct receive version of netif_receive_skb(). It should
4917 * only be used by callers that have a need to skip RPS and Generic XDP.
4918 * Caller must also take care of handling if (page_is_)pfmemalloc.
4920 * This function may only be called from softirq context and interrupts
4921 * should be enabled.
4923 * Return values (usually ignored):
4924 * NET_RX_SUCCESS: no congestion
4925 * NET_RX_DROP: packet was dropped
4927 int netif_receive_skb_core(struct sk_buff *skb)
4932 ret = __netif_receive_skb_one_core(skb, false);
4937 EXPORT_SYMBOL(netif_receive_skb_core);
4939 static inline void __netif_receive_skb_list_ptype(struct list_head *head,
4940 struct packet_type *pt_prev,
4941 struct net_device *orig_dev)
4943 struct sk_buff *skb, *next;
4947 if (list_empty(head))
4949 if (pt_prev->list_func != NULL)
4950 pt_prev->list_func(head, pt_prev, orig_dev);
4952 list_for_each_entry_safe(skb, next, head, list)
4953 pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
4956 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc)
4958 /* Fast-path assumptions:
4959 * - There is no RX handler.
4960 * - Only one packet_type matches.
4961 * If either of these fails, we will end up doing some per-packet
4962 * processing in-line, then handling the 'last ptype' for the whole
4963 * sublist. This can't cause out-of-order delivery to any single ptype,
4964 * because the 'last ptype' must be constant across the sublist, and all
4965 * other ptypes are handled per-packet.
4967 /* Current (common) ptype of sublist */
4968 struct packet_type *pt_curr = NULL;
4969 /* Current (common) orig_dev of sublist */
4970 struct net_device *od_curr = NULL;
4971 struct list_head sublist;
4972 struct sk_buff *skb, *next;
4974 INIT_LIST_HEAD(&sublist);
4975 list_for_each_entry_safe(skb, next, head, list) {
4976 struct net_device *orig_dev = skb->dev;
4977 struct packet_type *pt_prev = NULL;
4979 list_del(&skb->list);
4980 __netif_receive_skb_core(skb, pfmemalloc, &pt_prev);
4983 if (pt_curr != pt_prev || od_curr != orig_dev) {
4984 /* dispatch old sublist */
4985 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr);
4986 /* start new sublist */
4987 INIT_LIST_HEAD(&sublist);
4991 list_add_tail(&skb->list, &sublist);
4994 /* dispatch final sublist */
4995 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr);
4998 static int __netif_receive_skb(struct sk_buff *skb)
5002 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) {
5003 unsigned int noreclaim_flag;
5006 * PFMEMALLOC skbs are special, they should
5007 * - be delivered to SOCK_MEMALLOC sockets only
5008 * - stay away from userspace
5009 * - have bounded memory usage
5011 * Use PF_MEMALLOC as this saves us from propagating the allocation
5012 * context down to all allocation sites.
5014 noreclaim_flag = memalloc_noreclaim_save();
5015 ret = __netif_receive_skb_one_core(skb, true);
5016 memalloc_noreclaim_restore(noreclaim_flag);
5018 ret = __netif_receive_skb_one_core(skb, false);
5023 static void __netif_receive_skb_list(struct list_head *head)
5025 unsigned long noreclaim_flag = 0;
5026 struct sk_buff *skb, *next;
5027 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */
5029 list_for_each_entry_safe(skb, next, head, list) {
5030 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) {
5031 struct list_head sublist;
5033 /* Handle the previous sublist */
5034 list_cut_before(&sublist, head, &skb->list);
5035 if (!list_empty(&sublist))
5036 __netif_receive_skb_list_core(&sublist, pfmemalloc);
5037 pfmemalloc = !pfmemalloc;
5038 /* See comments in __netif_receive_skb */
5040 noreclaim_flag = memalloc_noreclaim_save();
5042 memalloc_noreclaim_restore(noreclaim_flag);
5045 /* Handle the remaining sublist */
5046 if (!list_empty(head))
5047 __netif_receive_skb_list_core(head, pfmemalloc);
5048 /* Restore pflags */
5050 memalloc_noreclaim_restore(noreclaim_flag);
5053 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp)
5055 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog);
5056 struct bpf_prog *new = xdp->prog;
5059 switch (xdp->command) {
5060 case XDP_SETUP_PROG:
5061 rcu_assign_pointer(dev->xdp_prog, new);
5066 static_branch_dec(&generic_xdp_needed_key);
5067 } else if (new && !old) {
5068 static_branch_inc(&generic_xdp_needed_key);
5069 dev_disable_lro(dev);
5070 dev_disable_gro_hw(dev);
5074 case XDP_QUERY_PROG:
5075 xdp->prog_id = old ? old->aux->id : 0;
5086 static int netif_receive_skb_internal(struct sk_buff *skb)
5090 net_timestamp_check(netdev_tstamp_prequeue, skb);
5092 if (skb_defer_rx_timestamp(skb))
5093 return NET_RX_SUCCESS;
5095 if (static_branch_unlikely(&generic_xdp_needed_key)) {
5100 ret = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
5104 if (ret != XDP_PASS)
5110 if (static_key_false(&rps_needed)) {
5111 struct rps_dev_flow voidflow, *rflow = &voidflow;
5112 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
5115 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
5121 ret = __netif_receive_skb(skb);
5126 static void netif_receive_skb_list_internal(struct list_head *head)
5128 struct bpf_prog *xdp_prog = NULL;
5129 struct sk_buff *skb, *next;
5130 struct list_head sublist;
5132 INIT_LIST_HEAD(&sublist);
5133 list_for_each_entry_safe(skb, next, head, list) {
5134 net_timestamp_check(netdev_tstamp_prequeue, skb);
5135 list_del(&skb->list);
5136 if (!skb_defer_rx_timestamp(skb))
5137 list_add_tail(&skb->list, &sublist);
5139 list_splice_init(&sublist, head);
5141 if (static_branch_unlikely(&generic_xdp_needed_key)) {
5144 list_for_each_entry_safe(skb, next, head, list) {
5145 xdp_prog = rcu_dereference(skb->dev->xdp_prog);
5146 list_del(&skb->list);
5147 if (do_xdp_generic(xdp_prog, skb) == XDP_PASS)
5148 list_add_tail(&skb->list, &sublist);
5152 /* Put passed packets back on main list */
5153 list_splice_init(&sublist, head);
5158 if (static_key_false(&rps_needed)) {
5159 list_for_each_entry_safe(skb, next, head, list) {
5160 struct rps_dev_flow voidflow, *rflow = &voidflow;
5161 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
5164 /* Will be handled, remove from list */
5165 list_del(&skb->list);
5166 enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
5171 __netif_receive_skb_list(head);
5176 * netif_receive_skb - process receive buffer from network
5177 * @skb: buffer to process
5179 * netif_receive_skb() is the main receive data processing function.
5180 * It always succeeds. The buffer may be dropped during processing
5181 * for congestion control or by the protocol layers.
5183 * This function may only be called from softirq context and interrupts
5184 * should be enabled.
5186 * Return values (usually ignored):
5187 * NET_RX_SUCCESS: no congestion
5188 * NET_RX_DROP: packet was dropped
5190 int netif_receive_skb(struct sk_buff *skb)
5192 trace_netif_receive_skb_entry(skb);
5194 return netif_receive_skb_internal(skb);
5196 EXPORT_SYMBOL(netif_receive_skb);
5199 * netif_receive_skb_list - process many receive buffers from network
5200 * @head: list of skbs to process.
5202 * Since return value of netif_receive_skb() is normally ignored, and
5203 * wouldn't be meaningful for a list, this function returns void.
5205 * This function may only be called from softirq context and interrupts
5206 * should be enabled.
5208 void netif_receive_skb_list(struct list_head *head)
5210 struct sk_buff *skb;
5212 if (list_empty(head))
5214 list_for_each_entry(skb, head, list)
5215 trace_netif_receive_skb_list_entry(skb);
5216 netif_receive_skb_list_internal(head);
5218 EXPORT_SYMBOL(netif_receive_skb_list);
5220 DEFINE_PER_CPU(struct work_struct, flush_works);
5222 /* Network device is going away, flush any packets still pending */
5223 static void flush_backlog(struct work_struct *work)
5225 struct sk_buff *skb, *tmp;
5226 struct softnet_data *sd;
5229 sd = this_cpu_ptr(&softnet_data);
5231 local_irq_disable();
5233 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) {
5234 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
5235 __skb_unlink(skb, &sd->input_pkt_queue);
5237 input_queue_head_incr(sd);
5243 skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
5244 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
5245 __skb_unlink(skb, &sd->process_queue);
5247 input_queue_head_incr(sd);
5253 static void flush_all_backlogs(void)
5259 for_each_online_cpu(cpu)
5260 queue_work_on(cpu, system_highpri_wq,
5261 per_cpu_ptr(&flush_works, cpu));
5263 for_each_online_cpu(cpu)
5264 flush_work(per_cpu_ptr(&flush_works, cpu));
5269 static int napi_gro_complete(struct sk_buff *skb)
5271 struct packet_offload *ptype;
5272 __be16 type = skb->protocol;
5273 struct list_head *head = &offload_base;
5276 BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
5278 if (NAPI_GRO_CB(skb)->count == 1) {
5279 skb_shinfo(skb)->gso_size = 0;
5284 list_for_each_entry_rcu(ptype, head, list) {
5285 if (ptype->type != type || !ptype->callbacks.gro_complete)
5288 err = ptype->callbacks.gro_complete(skb, 0);
5294 WARN_ON(&ptype->list == head);
5296 return NET_RX_SUCCESS;
5300 return netif_receive_skb_internal(skb);
5303 static void __napi_gro_flush_chain(struct napi_struct *napi, u32 index,
5306 struct list_head *head = &napi->gro_hash[index].list;
5307 struct sk_buff *skb, *p;
5309 list_for_each_entry_safe_reverse(skb, p, head, list) {
5310 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
5312 skb_list_del_init(skb);
5313 napi_gro_complete(skb);
5314 napi->gro_hash[index].count--;
5317 if (!napi->gro_hash[index].count)
5318 __clear_bit(index, &napi->gro_bitmask);
5321 /* napi->gro_hash[].list contains packets ordered by age.
5322 * youngest packets at the head of it.
5323 * Complete skbs in reverse order to reduce latencies.
5325 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
5329 for (i = 0; i < GRO_HASH_BUCKETS; i++) {
5330 if (test_bit(i, &napi->gro_bitmask))
5331 __napi_gro_flush_chain(napi, i, flush_old);
5334 EXPORT_SYMBOL(napi_gro_flush);
5336 static struct list_head *gro_list_prepare(struct napi_struct *napi,
5337 struct sk_buff *skb)
5339 unsigned int maclen = skb->dev->hard_header_len;
5340 u32 hash = skb_get_hash_raw(skb);
5341 struct list_head *head;
5344 head = &napi->gro_hash[hash & (GRO_HASH_BUCKETS - 1)].list;
5345 list_for_each_entry(p, head, list) {
5346 unsigned long diffs;
5348 NAPI_GRO_CB(p)->flush = 0;
5350 if (hash != skb_get_hash_raw(p)) {
5351 NAPI_GRO_CB(p)->same_flow = 0;
5355 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
5356 diffs |= p->vlan_tci ^ skb->vlan_tci;
5357 diffs |= skb_metadata_dst_cmp(p, skb);
5358 diffs |= skb_metadata_differs(p, skb);
5359 if (maclen == ETH_HLEN)
5360 diffs |= compare_ether_header(skb_mac_header(p),
5361 skb_mac_header(skb));
5363 diffs = memcmp(skb_mac_header(p),
5364 skb_mac_header(skb),
5366 NAPI_GRO_CB(p)->same_flow = !diffs;
5372 static void skb_gro_reset_offset(struct sk_buff *skb)
5374 const struct skb_shared_info *pinfo = skb_shinfo(skb);
5375 const skb_frag_t *frag0 = &pinfo->frags[0];
5377 NAPI_GRO_CB(skb)->data_offset = 0;
5378 NAPI_GRO_CB(skb)->frag0 = NULL;
5379 NAPI_GRO_CB(skb)->frag0_len = 0;
5381 if (skb_mac_header(skb) == skb_tail_pointer(skb) &&
5383 !PageHighMem(skb_frag_page(frag0))) {
5384 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
5385 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
5386 skb_frag_size(frag0),
5387 skb->end - skb->tail);
5391 static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
5393 struct skb_shared_info *pinfo = skb_shinfo(skb);
5395 BUG_ON(skb->end - skb->tail < grow);
5397 memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
5399 skb->data_len -= grow;
5402 pinfo->frags[0].page_offset += grow;
5403 skb_frag_size_sub(&pinfo->frags[0], grow);
5405 if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
5406 skb_frag_unref(skb, 0);
5407 memmove(pinfo->frags, pinfo->frags + 1,
5408 --pinfo->nr_frags * sizeof(pinfo->frags[0]));
5412 static void gro_flush_oldest(struct list_head *head)
5414 struct sk_buff *oldest;
5416 oldest = list_last_entry(head, struct sk_buff, list);
5418 /* We are called with head length >= MAX_GRO_SKBS, so this is
5421 if (WARN_ON_ONCE(!oldest))
5424 /* Do not adjust napi->gro_hash[].count, caller is adding a new
5427 list_del(&oldest->list);
5428 napi_gro_complete(oldest);
5431 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
5433 u32 hash = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1);
5434 struct list_head *head = &offload_base;
5435 struct packet_offload *ptype;
5436 __be16 type = skb->protocol;
5437 struct list_head *gro_head;
5438 struct sk_buff *pp = NULL;
5439 enum gro_result ret;
5443 if (netif_elide_gro(skb->dev))
5446 gro_head = gro_list_prepare(napi, skb);
5449 list_for_each_entry_rcu(ptype, head, list) {
5450 if (ptype->type != type || !ptype->callbacks.gro_receive)
5453 skb_set_network_header(skb, skb_gro_offset(skb));
5454 skb_reset_mac_len(skb);
5455 NAPI_GRO_CB(skb)->same_flow = 0;
5456 NAPI_GRO_CB(skb)->flush = skb_is_gso(skb) || skb_has_frag_list(skb);
5457 NAPI_GRO_CB(skb)->free = 0;
5458 NAPI_GRO_CB(skb)->encap_mark = 0;
5459 NAPI_GRO_CB(skb)->recursion_counter = 0;
5460 NAPI_GRO_CB(skb)->is_fou = 0;
5461 NAPI_GRO_CB(skb)->is_atomic = 1;
5462 NAPI_GRO_CB(skb)->gro_remcsum_start = 0;
5464 /* Setup for GRO checksum validation */
5465 switch (skb->ip_summed) {
5466 case CHECKSUM_COMPLETE:
5467 NAPI_GRO_CB(skb)->csum = skb->csum;
5468 NAPI_GRO_CB(skb)->csum_valid = 1;
5469 NAPI_GRO_CB(skb)->csum_cnt = 0;
5471 case CHECKSUM_UNNECESSARY:
5472 NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
5473 NAPI_GRO_CB(skb)->csum_valid = 0;
5476 NAPI_GRO_CB(skb)->csum_cnt = 0;
5477 NAPI_GRO_CB(skb)->csum_valid = 0;
5480 pp = ptype->callbacks.gro_receive(gro_head, skb);
5485 if (&ptype->list == head)
5488 if (IS_ERR(pp) && PTR_ERR(pp) == -EINPROGRESS) {
5493 same_flow = NAPI_GRO_CB(skb)->same_flow;
5494 ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
5497 skb_list_del_init(pp);
5498 napi_gro_complete(pp);
5499 napi->gro_hash[hash].count--;
5505 if (NAPI_GRO_CB(skb)->flush)
5508 if (unlikely(napi->gro_hash[hash].count >= MAX_GRO_SKBS)) {
5509 gro_flush_oldest(gro_head);
5511 napi->gro_hash[hash].count++;
5513 NAPI_GRO_CB(skb)->count = 1;
5514 NAPI_GRO_CB(skb)->age = jiffies;
5515 NAPI_GRO_CB(skb)->last = skb;
5516 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
5517 list_add(&skb->list, gro_head);
5521 grow = skb_gro_offset(skb) - skb_headlen(skb);
5523 gro_pull_from_frag0(skb, grow);
5525 if (napi->gro_hash[hash].count) {
5526 if (!test_bit(hash, &napi->gro_bitmask))
5527 __set_bit(hash, &napi->gro_bitmask);
5528 } else if (test_bit(hash, &napi->gro_bitmask)) {
5529 __clear_bit(hash, &napi->gro_bitmask);
5539 struct packet_offload *gro_find_receive_by_type(__be16 type)
5541 struct list_head *offload_head = &offload_base;
5542 struct packet_offload *ptype;
5544 list_for_each_entry_rcu(ptype, offload_head, list) {
5545 if (ptype->type != type || !ptype->callbacks.gro_receive)
5551 EXPORT_SYMBOL(gro_find_receive_by_type);
5553 struct packet_offload *gro_find_complete_by_type(__be16 type)
5555 struct list_head *offload_head = &offload_base;
5556 struct packet_offload *ptype;
5558 list_for_each_entry_rcu(ptype, offload_head, list) {
5559 if (ptype->type != type || !ptype->callbacks.gro_complete)
5565 EXPORT_SYMBOL(gro_find_complete_by_type);
5567 static void napi_skb_free_stolen_head(struct sk_buff *skb)
5571 kmem_cache_free(skbuff_head_cache, skb);
5574 static gro_result_t napi_skb_finish(gro_result_t ret, struct sk_buff *skb)
5578 if (netif_receive_skb_internal(skb))
5586 case GRO_MERGED_FREE:
5587 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
5588 napi_skb_free_stolen_head(skb);
5602 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
5604 skb_mark_napi_id(skb, napi);
5605 trace_napi_gro_receive_entry(skb);
5607 skb_gro_reset_offset(skb);
5609 return napi_skb_finish(dev_gro_receive(napi, skb), skb);
5611 EXPORT_SYMBOL(napi_gro_receive);
5613 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
5615 if (unlikely(skb->pfmemalloc)) {
5619 __skb_pull(skb, skb_headlen(skb));
5620 /* restore the reserve we had after netdev_alloc_skb_ip_align() */
5621 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
5623 skb->dev = napi->dev;
5625 skb->encapsulation = 0;
5626 skb_shinfo(skb)->gso_type = 0;
5627 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
5633 struct sk_buff *napi_get_frags(struct napi_struct *napi)
5635 struct sk_buff *skb = napi->skb;
5638 skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
5641 skb_mark_napi_id(skb, napi);
5646 EXPORT_SYMBOL(napi_get_frags);
5648 static gro_result_t napi_frags_finish(struct napi_struct *napi,
5649 struct sk_buff *skb,
5655 __skb_push(skb, ETH_HLEN);
5656 skb->protocol = eth_type_trans(skb, skb->dev);
5657 if (ret == GRO_NORMAL && netif_receive_skb_internal(skb))
5662 napi_reuse_skb(napi, skb);
5665 case GRO_MERGED_FREE:
5666 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
5667 napi_skb_free_stolen_head(skb);
5669 napi_reuse_skb(napi, skb);
5680 /* Upper GRO stack assumes network header starts at gro_offset=0
5681 * Drivers could call both napi_gro_frags() and napi_gro_receive()
5682 * We copy ethernet header into skb->data to have a common layout.
5684 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
5686 struct sk_buff *skb = napi->skb;
5687 const struct ethhdr *eth;
5688 unsigned int hlen = sizeof(*eth);
5692 skb_reset_mac_header(skb);
5693 skb_gro_reset_offset(skb);
5695 eth = skb_gro_header_fast(skb, 0);
5696 if (unlikely(skb_gro_header_hard(skb, hlen))) {
5697 eth = skb_gro_header_slow(skb, hlen, 0);
5698 if (unlikely(!eth)) {
5699 net_warn_ratelimited("%s: dropping impossible skb from %s\n",
5700 __func__, napi->dev->name);
5701 napi_reuse_skb(napi, skb);
5705 gro_pull_from_frag0(skb, hlen);
5706 NAPI_GRO_CB(skb)->frag0 += hlen;
5707 NAPI_GRO_CB(skb)->frag0_len -= hlen;
5709 __skb_pull(skb, hlen);
5712 * This works because the only protocols we care about don't require
5714 * We'll fix it up properly in napi_frags_finish()
5716 skb->protocol = eth->h_proto;
5721 gro_result_t napi_gro_frags(struct napi_struct *napi)
5723 struct sk_buff *skb = napi_frags_skb(napi);
5728 trace_napi_gro_frags_entry(skb);
5730 return napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
5732 EXPORT_SYMBOL(napi_gro_frags);
5734 /* Compute the checksum from gro_offset and return the folded value
5735 * after adding in any pseudo checksum.
5737 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
5742 wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
5744 /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
5745 sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
5747 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
5748 !skb->csum_complete_sw)
5749 netdev_rx_csum_fault(skb->dev);
5752 NAPI_GRO_CB(skb)->csum = wsum;
5753 NAPI_GRO_CB(skb)->csum_valid = 1;
5757 EXPORT_SYMBOL(__skb_gro_checksum_complete);
5759 static void net_rps_send_ipi(struct softnet_data *remsd)
5763 struct softnet_data *next = remsd->rps_ipi_next;
5765 if (cpu_online(remsd->cpu))
5766 smp_call_function_single_async(remsd->cpu, &remsd->csd);
5773 * net_rps_action_and_irq_enable sends any pending IPI's for rps.
5774 * Note: called with local irq disabled, but exits with local irq enabled.
5776 static void net_rps_action_and_irq_enable(struct softnet_data *sd)
5779 struct softnet_data *remsd = sd->rps_ipi_list;
5782 sd->rps_ipi_list = NULL;
5786 /* Send pending IPI's to kick RPS processing on remote cpus. */
5787 net_rps_send_ipi(remsd);
5793 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd)
5796 return sd->rps_ipi_list != NULL;
5802 static int process_backlog(struct napi_struct *napi, int quota)
5804 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
5808 /* Check if we have pending ipi, its better to send them now,
5809 * not waiting net_rx_action() end.
5811 if (sd_has_rps_ipi_waiting(sd)) {
5812 local_irq_disable();
5813 net_rps_action_and_irq_enable(sd);
5816 napi->weight = dev_rx_weight;
5818 struct sk_buff *skb;
5820 while ((skb = __skb_dequeue(&sd->process_queue))) {
5822 __netif_receive_skb(skb);
5824 input_queue_head_incr(sd);
5825 if (++work >= quota)
5830 local_irq_disable();
5832 if (skb_queue_empty(&sd->input_pkt_queue)) {
5834 * Inline a custom version of __napi_complete().
5835 * only current cpu owns and manipulates this napi,
5836 * and NAPI_STATE_SCHED is the only possible flag set
5838 * We can use a plain write instead of clear_bit(),
5839 * and we dont need an smp_mb() memory barrier.
5844 skb_queue_splice_tail_init(&sd->input_pkt_queue,
5845 &sd->process_queue);
5855 * __napi_schedule - schedule for receive
5856 * @n: entry to schedule
5858 * The entry's receive function will be scheduled to run.
5859 * Consider using __napi_schedule_irqoff() if hard irqs are masked.
5861 void __napi_schedule(struct napi_struct *n)
5863 unsigned long flags;
5865 local_irq_save(flags);
5866 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
5867 local_irq_restore(flags);
5869 EXPORT_SYMBOL(__napi_schedule);
5872 * napi_schedule_prep - check if napi can be scheduled
5875 * Test if NAPI routine is already running, and if not mark
5876 * it as running. This is used as a condition variable
5877 * insure only one NAPI poll instance runs. We also make
5878 * sure there is no pending NAPI disable.
5880 bool napi_schedule_prep(struct napi_struct *n)
5882 unsigned long val, new;
5885 val = READ_ONCE(n->state);
5886 if (unlikely(val & NAPIF_STATE_DISABLE))
5888 new = val | NAPIF_STATE_SCHED;
5890 /* Sets STATE_MISSED bit if STATE_SCHED was already set
5891 * This was suggested by Alexander Duyck, as compiler
5892 * emits better code than :
5893 * if (val & NAPIF_STATE_SCHED)
5894 * new |= NAPIF_STATE_MISSED;
5896 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED *
5898 } while (cmpxchg(&n->state, val, new) != val);
5900 return !(val & NAPIF_STATE_SCHED);
5902 EXPORT_SYMBOL(napi_schedule_prep);
5905 * __napi_schedule_irqoff - schedule for receive
5906 * @n: entry to schedule
5908 * Variant of __napi_schedule() assuming hard irqs are masked
5910 void __napi_schedule_irqoff(struct napi_struct *n)
5912 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
5914 EXPORT_SYMBOL(__napi_schedule_irqoff);
5916 bool napi_complete_done(struct napi_struct *n, int work_done)
5918 unsigned long flags, val, new;
5921 * 1) Don't let napi dequeue from the cpu poll list
5922 * just in case its running on a different cpu.
5923 * 2) If we are busy polling, do nothing here, we have
5924 * the guarantee we will be called later.
5926 if (unlikely(n->state & (NAPIF_STATE_NPSVC |
5927 NAPIF_STATE_IN_BUSY_POLL)))
5930 if (n->gro_bitmask) {
5931 unsigned long timeout = 0;
5934 timeout = n->dev->gro_flush_timeout;
5937 hrtimer_start(&n->timer, ns_to_ktime(timeout),
5938 HRTIMER_MODE_REL_PINNED);
5940 napi_gro_flush(n, false);
5942 if (unlikely(!list_empty(&n->poll_list))) {
5943 /* If n->poll_list is not empty, we need to mask irqs */
5944 local_irq_save(flags);
5945 list_del_init(&n->poll_list);
5946 local_irq_restore(flags);
5950 val = READ_ONCE(n->state);
5952 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED));
5954 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED);
5956 /* If STATE_MISSED was set, leave STATE_SCHED set,
5957 * because we will call napi->poll() one more time.
5958 * This C code was suggested by Alexander Duyck to help gcc.
5960 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED *
5962 } while (cmpxchg(&n->state, val, new) != val);
5964 if (unlikely(val & NAPIF_STATE_MISSED)) {
5971 EXPORT_SYMBOL(napi_complete_done);
5973 /* must be called under rcu_read_lock(), as we dont take a reference */
5974 static struct napi_struct *napi_by_id(unsigned int napi_id)
5976 unsigned int hash = napi_id % HASH_SIZE(napi_hash);
5977 struct napi_struct *napi;
5979 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node)
5980 if (napi->napi_id == napi_id)
5986 #if defined(CONFIG_NET_RX_BUSY_POLL)
5988 #define BUSY_POLL_BUDGET 8
5990 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock)
5994 /* Busy polling means there is a high chance device driver hard irq
5995 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was
5996 * set in napi_schedule_prep().
5997 * Since we are about to call napi->poll() once more, we can safely
5998 * clear NAPI_STATE_MISSED.
6000 * Note: x86 could use a single "lock and ..." instruction
6001 * to perform these two clear_bit()
6003 clear_bit(NAPI_STATE_MISSED, &napi->state);
6004 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state);
6008 /* All we really want here is to re-enable device interrupts.
6009 * Ideally, a new ndo_busy_poll_stop() could avoid another round.
6011 rc = napi->poll(napi, BUSY_POLL_BUDGET);
6012 trace_napi_poll(napi, rc, BUSY_POLL_BUDGET);
6013 netpoll_poll_unlock(have_poll_lock);
6014 if (rc == BUSY_POLL_BUDGET)
6015 __napi_schedule(napi);
6019 void napi_busy_loop(unsigned int napi_id,
6020 bool (*loop_end)(void *, unsigned long),
6023 unsigned long start_time = loop_end ? busy_loop_current_time() : 0;
6024 int (*napi_poll)(struct napi_struct *napi, int budget);
6025 void *have_poll_lock = NULL;
6026 struct napi_struct *napi;
6033 napi = napi_by_id(napi_id);
6043 unsigned long val = READ_ONCE(napi->state);
6045 /* If multiple threads are competing for this napi,
6046 * we avoid dirtying napi->state as much as we can.
6048 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED |
6049 NAPIF_STATE_IN_BUSY_POLL))
6051 if (cmpxchg(&napi->state, val,
6052 val | NAPIF_STATE_IN_BUSY_POLL |
6053 NAPIF_STATE_SCHED) != val)
6055 have_poll_lock = netpoll_poll_lock(napi);
6056 napi_poll = napi->poll;
6058 work = napi_poll(napi, BUSY_POLL_BUDGET);
6059 trace_napi_poll(napi, work, BUSY_POLL_BUDGET);
6062 __NET_ADD_STATS(dev_net(napi->dev),
6063 LINUX_MIB_BUSYPOLLRXPACKETS, work);
6066 if (!loop_end || loop_end(loop_end_arg, start_time))
6069 if (unlikely(need_resched())) {
6071 busy_poll_stop(napi, have_poll_lock);
6075 if (loop_end(loop_end_arg, start_time))
6082 busy_poll_stop(napi, have_poll_lock);
6087 EXPORT_SYMBOL(napi_busy_loop);
6089 #endif /* CONFIG_NET_RX_BUSY_POLL */
6091 static void napi_hash_add(struct napi_struct *napi)
6093 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state) ||
6094 test_and_set_bit(NAPI_STATE_HASHED, &napi->state))
6097 spin_lock(&napi_hash_lock);
6099 /* 0..NR_CPUS range is reserved for sender_cpu use */
6101 if (unlikely(++napi_gen_id < MIN_NAPI_ID))
6102 napi_gen_id = MIN_NAPI_ID;
6103 } while (napi_by_id(napi_gen_id));
6104 napi->napi_id = napi_gen_id;
6106 hlist_add_head_rcu(&napi->napi_hash_node,
6107 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]);
6109 spin_unlock(&napi_hash_lock);
6112 /* Warning : caller is responsible to make sure rcu grace period
6113 * is respected before freeing memory containing @napi
6115 bool napi_hash_del(struct napi_struct *napi)
6117 bool rcu_sync_needed = false;
6119 spin_lock(&napi_hash_lock);
6121 if (test_and_clear_bit(NAPI_STATE_HASHED, &napi->state)) {
6122 rcu_sync_needed = true;
6123 hlist_del_rcu(&napi->napi_hash_node);
6125 spin_unlock(&napi_hash_lock);
6126 return rcu_sync_needed;
6128 EXPORT_SYMBOL_GPL(napi_hash_del);
6130 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer)
6132 struct napi_struct *napi;
6134 napi = container_of(timer, struct napi_struct, timer);
6136 /* Note : we use a relaxed variant of napi_schedule_prep() not setting
6137 * NAPI_STATE_MISSED, since we do not react to a device IRQ.
6139 if (napi->gro_bitmask && !napi_disable_pending(napi) &&
6140 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state))
6141 __napi_schedule_irqoff(napi);
6143 return HRTIMER_NORESTART;
6146 static void init_gro_hash(struct napi_struct *napi)
6150 for (i = 0; i < GRO_HASH_BUCKETS; i++) {
6151 INIT_LIST_HEAD(&napi->gro_hash[i].list);
6152 napi->gro_hash[i].count = 0;
6154 napi->gro_bitmask = 0;
6157 void netif_napi_add(struct net_device *dev, struct napi_struct *napi,
6158 int (*poll)(struct napi_struct *, int), int weight)
6160 INIT_LIST_HEAD(&napi->poll_list);
6161 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
6162 napi->timer.function = napi_watchdog;
6163 init_gro_hash(napi);
6166 if (weight > NAPI_POLL_WEIGHT)
6167 pr_err_once("netif_napi_add() called with weight %d on device %s\n",
6169 napi->weight = weight;
6170 list_add(&napi->dev_list, &dev->napi_list);
6172 #ifdef CONFIG_NETPOLL
6173 napi->poll_owner = -1;
6175 set_bit(NAPI_STATE_SCHED, &napi->state);
6176 napi_hash_add(napi);
6178 EXPORT_SYMBOL(netif_napi_add);
6180 void napi_disable(struct napi_struct *n)
6183 set_bit(NAPI_STATE_DISABLE, &n->state);
6185 while (test_and_set_bit(NAPI_STATE_SCHED, &n->state))
6187 while (test_and_set_bit(NAPI_STATE_NPSVC, &n->state))
6190 hrtimer_cancel(&n->timer);
6192 clear_bit(NAPI_STATE_DISABLE, &n->state);
6194 EXPORT_SYMBOL(napi_disable);
6196 static void flush_gro_hash(struct napi_struct *napi)
6200 for (i = 0; i < GRO_HASH_BUCKETS; i++) {
6201 struct sk_buff *skb, *n;
6203 list_for_each_entry_safe(skb, n, &napi->gro_hash[i].list, list)
6205 napi->gro_hash[i].count = 0;
6209 /* Must be called in process context */
6210 void netif_napi_del(struct napi_struct *napi)
6213 if (napi_hash_del(napi))
6215 list_del_init(&napi->dev_list);
6216 napi_free_frags(napi);
6218 flush_gro_hash(napi);
6219 napi->gro_bitmask = 0;
6221 EXPORT_SYMBOL(netif_napi_del);
6223 static int napi_poll(struct napi_struct *n, struct list_head *repoll)
6228 list_del_init(&n->poll_list);
6230 have = netpoll_poll_lock(n);
6234 /* This NAPI_STATE_SCHED test is for avoiding a race
6235 * with netpoll's poll_napi(). Only the entity which
6236 * obtains the lock and sees NAPI_STATE_SCHED set will
6237 * actually make the ->poll() call. Therefore we avoid
6238 * accidentally calling ->poll() when NAPI is not scheduled.
6241 if (test_bit(NAPI_STATE_SCHED, &n->state)) {
6242 work = n->poll(n, weight);
6243 trace_napi_poll(n, work, weight);
6246 WARN_ON_ONCE(work > weight);
6248 if (likely(work < weight))
6251 /* Drivers must not modify the NAPI state if they
6252 * consume the entire weight. In such cases this code
6253 * still "owns" the NAPI instance and therefore can
6254 * move the instance around on the list at-will.
6256 if (unlikely(napi_disable_pending(n))) {
6261 if (n->gro_bitmask) {
6262 /* flush too old packets
6263 * If HZ < 1000, flush all packets.
6265 napi_gro_flush(n, HZ >= 1000);
6268 /* Some drivers may have called napi_schedule
6269 * prior to exhausting their budget.
6271 if (unlikely(!list_empty(&n->poll_list))) {
6272 pr_warn_once("%s: Budget exhausted after napi rescheduled\n",
6273 n->dev ? n->dev->name : "backlog");
6277 list_add_tail(&n->poll_list, repoll);
6280 netpoll_poll_unlock(have);
6285 static __latent_entropy void net_rx_action(struct softirq_action *h)
6287 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
6288 unsigned long time_limit = jiffies +
6289 usecs_to_jiffies(netdev_budget_usecs);
6290 int budget = netdev_budget;
6294 local_irq_disable();
6295 list_splice_init(&sd->poll_list, &list);
6299 struct napi_struct *n;
6301 if (list_empty(&list)) {
6302 if (!sd_has_rps_ipi_waiting(sd) && list_empty(&repoll))
6307 n = list_first_entry(&list, struct napi_struct, poll_list);
6308 budget -= napi_poll(n, &repoll);
6310 /* If softirq window is exhausted then punt.
6311 * Allow this to run for 2 jiffies since which will allow
6312 * an average latency of 1.5/HZ.
6314 if (unlikely(budget <= 0 ||
6315 time_after_eq(jiffies, time_limit))) {
6321 local_irq_disable();
6323 list_splice_tail_init(&sd->poll_list, &list);
6324 list_splice_tail(&repoll, &list);
6325 list_splice(&list, &sd->poll_list);
6326 if (!list_empty(&sd->poll_list))
6327 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
6329 net_rps_action_and_irq_enable(sd);
6331 __kfree_skb_flush();
6334 struct netdev_adjacent {
6335 struct net_device *dev;
6337 /* upper master flag, there can only be one master device per list */
6340 /* counter for the number of times this device was added to us */
6343 /* private field for the users */
6346 struct list_head list;
6347 struct rcu_head rcu;
6350 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev,
6351 struct list_head *adj_list)
6353 struct netdev_adjacent *adj;
6355 list_for_each_entry(adj, adj_list, list) {
6356 if (adj->dev == adj_dev)
6362 static int __netdev_has_upper_dev(struct net_device *upper_dev, void *data)
6364 struct net_device *dev = data;
6366 return upper_dev == dev;
6370 * netdev_has_upper_dev - Check if device is linked to an upper device
6372 * @upper_dev: upper device to check
6374 * Find out if a device is linked to specified upper device and return true
6375 * in case it is. Note that this checks only immediate upper device,
6376 * not through a complete stack of devices. The caller must hold the RTNL lock.
6378 bool netdev_has_upper_dev(struct net_device *dev,
6379 struct net_device *upper_dev)
6383 return netdev_walk_all_upper_dev_rcu(dev, __netdev_has_upper_dev,
6386 EXPORT_SYMBOL(netdev_has_upper_dev);
6389 * netdev_has_upper_dev_all - Check if device is linked to an upper device
6391 * @upper_dev: upper device to check
6393 * Find out if a device is linked to specified upper device and return true
6394 * in case it is. Note that this checks the entire upper device chain.
6395 * The caller must hold rcu lock.
6398 bool netdev_has_upper_dev_all_rcu(struct net_device *dev,
6399 struct net_device *upper_dev)
6401 return !!netdev_walk_all_upper_dev_rcu(dev, __netdev_has_upper_dev,
6404 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu);
6407 * netdev_has_any_upper_dev - Check if device is linked to some device
6410 * Find out if a device is linked to an upper device and return true in case
6411 * it is. The caller must hold the RTNL lock.
6413 bool netdev_has_any_upper_dev(struct net_device *dev)
6417 return !list_empty(&dev->adj_list.upper);
6419 EXPORT_SYMBOL(netdev_has_any_upper_dev);
6422 * netdev_master_upper_dev_get - Get master upper device
6425 * Find a master upper device and return pointer to it or NULL in case
6426 * it's not there. The caller must hold the RTNL lock.
6428 struct net_device *netdev_master_upper_dev_get(struct net_device *dev)
6430 struct netdev_adjacent *upper;
6434 if (list_empty(&dev->adj_list.upper))
6437 upper = list_first_entry(&dev->adj_list.upper,
6438 struct netdev_adjacent, list);
6439 if (likely(upper->master))
6443 EXPORT_SYMBOL(netdev_master_upper_dev_get);
6446 * netdev_has_any_lower_dev - Check if device is linked to some device
6449 * Find out if a device is linked to a lower device and return true in case
6450 * it is. The caller must hold the RTNL lock.
6452 static bool netdev_has_any_lower_dev(struct net_device *dev)
6456 return !list_empty(&dev->adj_list.lower);
6459 void *netdev_adjacent_get_private(struct list_head *adj_list)
6461 struct netdev_adjacent *adj;
6463 adj = list_entry(adj_list, struct netdev_adjacent, list);
6465 return adj->private;
6467 EXPORT_SYMBOL(netdev_adjacent_get_private);
6470 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list
6472 * @iter: list_head ** of the current position
6474 * Gets the next device from the dev's upper list, starting from iter
6475 * position. The caller must hold RCU read lock.
6477 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev,
6478 struct list_head **iter)
6480 struct netdev_adjacent *upper;
6482 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
6484 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6486 if (&upper->list == &dev->adj_list.upper)
6489 *iter = &upper->list;
6493 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu);
6495 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev,
6496 struct list_head **iter)
6498 struct netdev_adjacent *upper;
6500 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
6502 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6504 if (&upper->list == &dev->adj_list.upper)
6507 *iter = &upper->list;
6512 int netdev_walk_all_upper_dev_rcu(struct net_device *dev,
6513 int (*fn)(struct net_device *dev,
6517 struct net_device *udev;
6518 struct list_head *iter;
6521 for (iter = &dev->adj_list.upper,
6522 udev = netdev_next_upper_dev_rcu(dev, &iter);
6524 udev = netdev_next_upper_dev_rcu(dev, &iter)) {
6525 /* first is the upper device itself */
6526 ret = fn(udev, data);
6530 /* then look at all of its upper devices */
6531 ret = netdev_walk_all_upper_dev_rcu(udev, fn, data);
6538 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu);
6541 * netdev_lower_get_next_private - Get the next ->private from the
6542 * lower neighbour list
6544 * @iter: list_head ** of the current position
6546 * Gets the next netdev_adjacent->private from the dev's lower neighbour
6547 * list, starting from iter position. The caller must hold either hold the
6548 * RTNL lock or its own locking that guarantees that the neighbour lower
6549 * list will remain unchanged.
6551 void *netdev_lower_get_next_private(struct net_device *dev,
6552 struct list_head **iter)
6554 struct netdev_adjacent *lower;
6556 lower = list_entry(*iter, struct netdev_adjacent, list);
6558 if (&lower->list == &dev->adj_list.lower)
6561 *iter = lower->list.next;
6563 return lower->private;
6565 EXPORT_SYMBOL(netdev_lower_get_next_private);
6568 * netdev_lower_get_next_private_rcu - Get the next ->private from the
6569 * lower neighbour list, RCU
6572 * @iter: list_head ** of the current position
6574 * Gets the next netdev_adjacent->private from the dev's lower neighbour
6575 * list, starting from iter position. The caller must hold RCU read lock.
6577 void *netdev_lower_get_next_private_rcu(struct net_device *dev,
6578 struct list_head **iter)
6580 struct netdev_adjacent *lower;
6582 WARN_ON_ONCE(!rcu_read_lock_held());
6584 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6586 if (&lower->list == &dev->adj_list.lower)
6589 *iter = &lower->list;
6591 return lower->private;
6593 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu);
6596 * netdev_lower_get_next - Get the next device from the lower neighbour
6599 * @iter: list_head ** of the current position
6601 * Gets the next netdev_adjacent from the dev's lower neighbour
6602 * list, starting from iter position. The caller must hold RTNL lock or
6603 * its own locking that guarantees that the neighbour lower
6604 * list will remain unchanged.
6606 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter)
6608 struct netdev_adjacent *lower;
6610 lower = list_entry(*iter, struct netdev_adjacent, list);
6612 if (&lower->list == &dev->adj_list.lower)
6615 *iter = lower->list.next;
6619 EXPORT_SYMBOL(netdev_lower_get_next);
6621 static struct net_device *netdev_next_lower_dev(struct net_device *dev,
6622 struct list_head **iter)
6624 struct netdev_adjacent *lower;
6626 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
6628 if (&lower->list == &dev->adj_list.lower)
6631 *iter = &lower->list;
6636 int netdev_walk_all_lower_dev(struct net_device *dev,
6637 int (*fn)(struct net_device *dev,
6641 struct net_device *ldev;
6642 struct list_head *iter;
6645 for (iter = &dev->adj_list.lower,
6646 ldev = netdev_next_lower_dev(dev, &iter);
6648 ldev = netdev_next_lower_dev(dev, &iter)) {
6649 /* first is the lower device itself */
6650 ret = fn(ldev, data);
6654 /* then look at all of its lower devices */
6655 ret = netdev_walk_all_lower_dev(ldev, fn, data);
6662 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev);
6664 static struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev,
6665 struct list_head **iter)
6667 struct netdev_adjacent *lower;
6669 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6670 if (&lower->list == &dev->adj_list.lower)
6673 *iter = &lower->list;
6678 int netdev_walk_all_lower_dev_rcu(struct net_device *dev,
6679 int (*fn)(struct net_device *dev,
6683 struct net_device *ldev;
6684 struct list_head *iter;
6687 for (iter = &dev->adj_list.lower,
6688 ldev = netdev_next_lower_dev_rcu(dev, &iter);
6690 ldev = netdev_next_lower_dev_rcu(dev, &iter)) {
6691 /* first is the lower device itself */
6692 ret = fn(ldev, data);
6696 /* then look at all of its lower devices */
6697 ret = netdev_walk_all_lower_dev_rcu(ldev, fn, data);
6704 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu);
6707 * netdev_lower_get_first_private_rcu - Get the first ->private from the
6708 * lower neighbour list, RCU
6712 * Gets the first netdev_adjacent->private from the dev's lower neighbour
6713 * list. The caller must hold RCU read lock.
6715 void *netdev_lower_get_first_private_rcu(struct net_device *dev)
6717 struct netdev_adjacent *lower;
6719 lower = list_first_or_null_rcu(&dev->adj_list.lower,
6720 struct netdev_adjacent, list);
6722 return lower->private;
6725 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu);
6728 * netdev_master_upper_dev_get_rcu - Get master upper device
6731 * Find a master upper device and return pointer to it or NULL in case
6732 * it's not there. The caller must hold the RCU read lock.
6734 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev)
6736 struct netdev_adjacent *upper;
6738 upper = list_first_or_null_rcu(&dev->adj_list.upper,
6739 struct netdev_adjacent, list);
6740 if (upper && likely(upper->master))
6744 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu);
6746 static int netdev_adjacent_sysfs_add(struct net_device *dev,
6747 struct net_device *adj_dev,
6748 struct list_head *dev_list)
6750 char linkname[IFNAMSIZ+7];
6752 sprintf(linkname, dev_list == &dev->adj_list.upper ?
6753 "upper_%s" : "lower_%s", adj_dev->name);
6754 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj),
6757 static void netdev_adjacent_sysfs_del(struct net_device *dev,
6759 struct list_head *dev_list)
6761 char linkname[IFNAMSIZ+7];
6763 sprintf(linkname, dev_list == &dev->adj_list.upper ?
6764 "upper_%s" : "lower_%s", name);
6765 sysfs_remove_link(&(dev->dev.kobj), linkname);
6768 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev,
6769 struct net_device *adj_dev,
6770 struct list_head *dev_list)
6772 return (dev_list == &dev->adj_list.upper ||
6773 dev_list == &dev->adj_list.lower) &&
6774 net_eq(dev_net(dev), dev_net(adj_dev));
6777 static int __netdev_adjacent_dev_insert(struct net_device *dev,
6778 struct net_device *adj_dev,
6779 struct list_head *dev_list,
6780 void *private, bool master)
6782 struct netdev_adjacent *adj;
6785 adj = __netdev_find_adj(adj_dev, dev_list);
6789 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n",
6790 dev->name, adj_dev->name, adj->ref_nr);
6795 adj = kmalloc(sizeof(*adj), GFP_KERNEL);
6800 adj->master = master;
6802 adj->private = private;
6805 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n",
6806 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name);
6808 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) {
6809 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list);
6814 /* Ensure that master link is always the first item in list. */
6816 ret = sysfs_create_link(&(dev->dev.kobj),
6817 &(adj_dev->dev.kobj), "master");
6819 goto remove_symlinks;
6821 list_add_rcu(&adj->list, dev_list);
6823 list_add_tail_rcu(&adj->list, dev_list);
6829 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
6830 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
6838 static void __netdev_adjacent_dev_remove(struct net_device *dev,
6839 struct net_device *adj_dev,
6841 struct list_head *dev_list)
6843 struct netdev_adjacent *adj;
6845 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n",
6846 dev->name, adj_dev->name, ref_nr);
6848 adj = __netdev_find_adj(adj_dev, dev_list);
6851 pr_err("Adjacency does not exist for device %s from %s\n",
6852 dev->name, adj_dev->name);
6857 if (adj->ref_nr > ref_nr) {
6858 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n",
6859 dev->name, adj_dev->name, ref_nr,
6860 adj->ref_nr - ref_nr);
6861 adj->ref_nr -= ref_nr;
6866 sysfs_remove_link(&(dev->dev.kobj), "master");
6868 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
6869 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
6871 list_del_rcu(&adj->list);
6872 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n",
6873 adj_dev->name, dev->name, adj_dev->name);
6875 kfree_rcu(adj, rcu);
6878 static int __netdev_adjacent_dev_link_lists(struct net_device *dev,
6879 struct net_device *upper_dev,
6880 struct list_head *up_list,
6881 struct list_head *down_list,
6882 void *private, bool master)
6886 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list,
6891 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list,
6894 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list);
6901 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev,
6902 struct net_device *upper_dev,
6904 struct list_head *up_list,
6905 struct list_head *down_list)
6907 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list);
6908 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list);
6911 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev,
6912 struct net_device *upper_dev,
6913 void *private, bool master)
6915 return __netdev_adjacent_dev_link_lists(dev, upper_dev,
6916 &dev->adj_list.upper,
6917 &upper_dev->adj_list.lower,
6921 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev,
6922 struct net_device *upper_dev)
6924 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1,
6925 &dev->adj_list.upper,
6926 &upper_dev->adj_list.lower);
6929 static int __netdev_upper_dev_link(struct net_device *dev,
6930 struct net_device *upper_dev, bool master,
6931 void *upper_priv, void *upper_info,
6932 struct netlink_ext_ack *extack)
6934 struct netdev_notifier_changeupper_info changeupper_info = {
6939 .upper_dev = upper_dev,
6942 .upper_info = upper_info,
6944 struct net_device *master_dev;
6949 if (dev == upper_dev)
6952 /* To prevent loops, check if dev is not upper device to upper_dev. */
6953 if (netdev_has_upper_dev(upper_dev, dev))
6957 if (netdev_has_upper_dev(dev, upper_dev))
6960 master_dev = netdev_master_upper_dev_get(dev);
6962 return master_dev == upper_dev ? -EEXIST : -EBUSY;
6965 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
6966 &changeupper_info.info);
6967 ret = notifier_to_errno(ret);
6971 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv,
6976 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
6977 &changeupper_info.info);
6978 ret = notifier_to_errno(ret);
6985 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
6991 * netdev_upper_dev_link - Add a link to the upper device
6993 * @upper_dev: new upper device
6994 * @extack: netlink extended ack
6996 * Adds a link to device which is upper to this one. The caller must hold
6997 * the RTNL lock. On a failure a negative errno code is returned.
6998 * On success the reference counts are adjusted and the function
7001 int netdev_upper_dev_link(struct net_device *dev,
7002 struct net_device *upper_dev,
7003 struct netlink_ext_ack *extack)
7005 return __netdev_upper_dev_link(dev, upper_dev, false,
7006 NULL, NULL, extack);
7008 EXPORT_SYMBOL(netdev_upper_dev_link);
7011 * netdev_master_upper_dev_link - Add a master link to the upper device
7013 * @upper_dev: new upper device
7014 * @upper_priv: upper device private
7015 * @upper_info: upper info to be passed down via notifier
7016 * @extack: netlink extended ack
7018 * Adds a link to device which is upper to this one. In this case, only
7019 * one master upper device can be linked, although other non-master devices
7020 * might be linked as well. The caller must hold the RTNL lock.
7021 * On a failure a negative errno code is returned. On success the reference
7022 * counts are adjusted and the function returns zero.
7024 int netdev_master_upper_dev_link(struct net_device *dev,
7025 struct net_device *upper_dev,
7026 void *upper_priv, void *upper_info,
7027 struct netlink_ext_ack *extack)
7029 return __netdev_upper_dev_link(dev, upper_dev, true,
7030 upper_priv, upper_info, extack);
7032 EXPORT_SYMBOL(netdev_master_upper_dev_link);
7035 * netdev_upper_dev_unlink - Removes a link to upper device
7037 * @upper_dev: new upper device
7039 * Removes a link to device which is upper to this one. The caller must hold
7042 void netdev_upper_dev_unlink(struct net_device *dev,
7043 struct net_device *upper_dev)
7045 struct netdev_notifier_changeupper_info changeupper_info = {
7049 .upper_dev = upper_dev,
7055 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev;
7057 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
7058 &changeupper_info.info);
7060 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
7062 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
7063 &changeupper_info.info);
7065 EXPORT_SYMBOL(netdev_upper_dev_unlink);
7068 * netdev_bonding_info_change - Dispatch event about slave change
7070 * @bonding_info: info to dispatch
7072 * Send NETDEV_BONDING_INFO to netdev notifiers with info.
7073 * The caller must hold the RTNL lock.
7075 void netdev_bonding_info_change(struct net_device *dev,
7076 struct netdev_bonding_info *bonding_info)
7078 struct netdev_notifier_bonding_info info = {
7082 memcpy(&info.bonding_info, bonding_info,
7083 sizeof(struct netdev_bonding_info));
7084 call_netdevice_notifiers_info(NETDEV_BONDING_INFO,
7087 EXPORT_SYMBOL(netdev_bonding_info_change);
7089 static void netdev_adjacent_add_links(struct net_device *dev)
7091 struct netdev_adjacent *iter;
7093 struct net *net = dev_net(dev);
7095 list_for_each_entry(iter, &dev->adj_list.upper, list) {
7096 if (!net_eq(net, dev_net(iter->dev)))
7098 netdev_adjacent_sysfs_add(iter->dev, dev,
7099 &iter->dev->adj_list.lower);
7100 netdev_adjacent_sysfs_add(dev, iter->dev,
7101 &dev->adj_list.upper);
7104 list_for_each_entry(iter, &dev->adj_list.lower, list) {
7105 if (!net_eq(net, dev_net(iter->dev)))
7107 netdev_adjacent_sysfs_add(iter->dev, dev,
7108 &iter->dev->adj_list.upper);
7109 netdev_adjacent_sysfs_add(dev, iter->dev,
7110 &dev->adj_list.lower);
7114 static void netdev_adjacent_del_links(struct net_device *dev)
7116 struct netdev_adjacent *iter;
7118 struct net *net = dev_net(dev);
7120 list_for_each_entry(iter, &dev->adj_list.upper, list) {
7121 if (!net_eq(net, dev_net(iter->dev)))
7123 netdev_adjacent_sysfs_del(iter->dev, dev->name,
7124 &iter->dev->adj_list.lower);
7125 netdev_adjacent_sysfs_del(dev, iter->dev->name,
7126 &dev->adj_list.upper);
7129 list_for_each_entry(iter, &dev->adj_list.lower, list) {
7130 if (!net_eq(net, dev_net(iter->dev)))
7132 netdev_adjacent_sysfs_del(iter->dev, dev->name,
7133 &iter->dev->adj_list.upper);
7134 netdev_adjacent_sysfs_del(dev, iter->dev->name,
7135 &dev->adj_list.lower);
7139 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname)
7141 struct netdev_adjacent *iter;
7143 struct net *net = dev_net(dev);
7145 list_for_each_entry(iter, &dev->adj_list.upper, list) {
7146 if (!net_eq(net, dev_net(iter->dev)))
7148 netdev_adjacent_sysfs_del(iter->dev, oldname,
7149 &iter->dev->adj_list.lower);
7150 netdev_adjacent_sysfs_add(iter->dev, dev,
7151 &iter->dev->adj_list.lower);
7154 list_for_each_entry(iter, &dev->adj_list.lower, list) {
7155 if (!net_eq(net, dev_net(iter->dev)))
7157 netdev_adjacent_sysfs_del(iter->dev, oldname,
7158 &iter->dev->adj_list.upper);
7159 netdev_adjacent_sysfs_add(iter->dev, dev,
7160 &iter->dev->adj_list.upper);
7164 void *netdev_lower_dev_get_private(struct net_device *dev,
7165 struct net_device *lower_dev)
7167 struct netdev_adjacent *lower;
7171 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower);
7175 return lower->private;
7177 EXPORT_SYMBOL(netdev_lower_dev_get_private);
7180 int dev_get_nest_level(struct net_device *dev)
7182 struct net_device *lower = NULL;
7183 struct list_head *iter;
7189 netdev_for_each_lower_dev(dev, lower, iter) {
7190 nest = dev_get_nest_level(lower);
7191 if (max_nest < nest)
7195 return max_nest + 1;
7197 EXPORT_SYMBOL(dev_get_nest_level);
7200 * netdev_lower_change - Dispatch event about lower device state change
7201 * @lower_dev: device
7202 * @lower_state_info: state to dispatch
7204 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info.
7205 * The caller must hold the RTNL lock.
7207 void netdev_lower_state_changed(struct net_device *lower_dev,
7208 void *lower_state_info)
7210 struct netdev_notifier_changelowerstate_info changelowerstate_info = {
7211 .info.dev = lower_dev,
7215 changelowerstate_info.lower_state_info = lower_state_info;
7216 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE,
7217 &changelowerstate_info.info);
7219 EXPORT_SYMBOL(netdev_lower_state_changed);
7221 static void dev_change_rx_flags(struct net_device *dev, int flags)
7223 const struct net_device_ops *ops = dev->netdev_ops;
7225 if (ops->ndo_change_rx_flags)
7226 ops->ndo_change_rx_flags(dev, flags);
7229 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify)
7231 unsigned int old_flags = dev->flags;
7237 dev->flags |= IFF_PROMISC;
7238 dev->promiscuity += inc;
7239 if (dev->promiscuity == 0) {
7242 * If inc causes overflow, untouch promisc and return error.
7245 dev->flags &= ~IFF_PROMISC;
7247 dev->promiscuity -= inc;
7248 pr_warn("%s: promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n",
7253 if (dev->flags != old_flags) {
7254 pr_info("device %s %s promiscuous mode\n",
7256 dev->flags & IFF_PROMISC ? "entered" : "left");
7257 if (audit_enabled) {
7258 current_uid_gid(&uid, &gid);
7259 audit_log(audit_context(), GFP_ATOMIC,
7260 AUDIT_ANOM_PROMISCUOUS,
7261 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u",
7262 dev->name, (dev->flags & IFF_PROMISC),
7263 (old_flags & IFF_PROMISC),
7264 from_kuid(&init_user_ns, audit_get_loginuid(current)),
7265 from_kuid(&init_user_ns, uid),
7266 from_kgid(&init_user_ns, gid),
7267 audit_get_sessionid(current));
7270 dev_change_rx_flags(dev, IFF_PROMISC);
7273 __dev_notify_flags(dev, old_flags, IFF_PROMISC);
7278 * dev_set_promiscuity - update promiscuity count on a device
7282 * Add or remove promiscuity from a device. While the count in the device
7283 * remains above zero the interface remains promiscuous. Once it hits zero
7284 * the device reverts back to normal filtering operation. A negative inc
7285 * value is used to drop promiscuity on the device.
7286 * Return 0 if successful or a negative errno code on error.
7288 int dev_set_promiscuity(struct net_device *dev, int inc)
7290 unsigned int old_flags = dev->flags;
7293 err = __dev_set_promiscuity(dev, inc, true);
7296 if (dev->flags != old_flags)
7297 dev_set_rx_mode(dev);
7300 EXPORT_SYMBOL(dev_set_promiscuity);
7302 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify)
7304 unsigned int old_flags = dev->flags, old_gflags = dev->gflags;
7308 dev->flags |= IFF_ALLMULTI;
7309 dev->allmulti += inc;
7310 if (dev->allmulti == 0) {
7313 * If inc causes overflow, untouch allmulti and return error.
7316 dev->flags &= ~IFF_ALLMULTI;
7318 dev->allmulti -= inc;
7319 pr_warn("%s: allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n",
7324 if (dev->flags ^ old_flags) {
7325 dev_change_rx_flags(dev, IFF_ALLMULTI);
7326 dev_set_rx_mode(dev);
7328 __dev_notify_flags(dev, old_flags,
7329 dev->gflags ^ old_gflags);
7335 * dev_set_allmulti - update allmulti count on a device
7339 * Add or remove reception of all multicast frames to a device. While the
7340 * count in the device remains above zero the interface remains listening
7341 * to all interfaces. Once it hits zero the device reverts back to normal
7342 * filtering operation. A negative @inc value is used to drop the counter
7343 * when releasing a resource needing all multicasts.
7344 * Return 0 if successful or a negative errno code on error.
7347 int dev_set_allmulti(struct net_device *dev, int inc)
7349 return __dev_set_allmulti(dev, inc, true);
7351 EXPORT_SYMBOL(dev_set_allmulti);
7354 * Upload unicast and multicast address lists to device and
7355 * configure RX filtering. When the device doesn't support unicast
7356 * filtering it is put in promiscuous mode while unicast addresses
7359 void __dev_set_rx_mode(struct net_device *dev)
7361 const struct net_device_ops *ops = dev->netdev_ops;
7363 /* dev_open will call this function so the list will stay sane. */
7364 if (!(dev->flags&IFF_UP))
7367 if (!netif_device_present(dev))
7370 if (!(dev->priv_flags & IFF_UNICAST_FLT)) {
7371 /* Unicast addresses changes may only happen under the rtnl,
7372 * therefore calling __dev_set_promiscuity here is safe.
7374 if (!netdev_uc_empty(dev) && !dev->uc_promisc) {
7375 __dev_set_promiscuity(dev, 1, false);
7376 dev->uc_promisc = true;
7377 } else if (netdev_uc_empty(dev) && dev->uc_promisc) {
7378 __dev_set_promiscuity(dev, -1, false);
7379 dev->uc_promisc = false;
7383 if (ops->ndo_set_rx_mode)
7384 ops->ndo_set_rx_mode(dev);
7387 void dev_set_rx_mode(struct net_device *dev)
7389 netif_addr_lock_bh(dev);
7390 __dev_set_rx_mode(dev);
7391 netif_addr_unlock_bh(dev);
7395 * dev_get_flags - get flags reported to userspace
7398 * Get the combination of flag bits exported through APIs to userspace.
7400 unsigned int dev_get_flags(const struct net_device *dev)
7404 flags = (dev->flags & ~(IFF_PROMISC |
7409 (dev->gflags & (IFF_PROMISC |
7412 if (netif_running(dev)) {
7413 if (netif_oper_up(dev))
7414 flags |= IFF_RUNNING;
7415 if (netif_carrier_ok(dev))
7416 flags |= IFF_LOWER_UP;
7417 if (netif_dormant(dev))
7418 flags |= IFF_DORMANT;
7423 EXPORT_SYMBOL(dev_get_flags);
7425 int __dev_change_flags(struct net_device *dev, unsigned int flags)
7427 unsigned int old_flags = dev->flags;
7433 * Set the flags on our device.
7436 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP |
7437 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL |
7439 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC |
7443 * Load in the correct multicast list now the flags have changed.
7446 if ((old_flags ^ flags) & IFF_MULTICAST)
7447 dev_change_rx_flags(dev, IFF_MULTICAST);
7449 dev_set_rx_mode(dev);
7452 * Have we downed the interface. We handle IFF_UP ourselves
7453 * according to user attempts to set it, rather than blindly
7458 if ((old_flags ^ flags) & IFF_UP) {
7459 if (old_flags & IFF_UP)
7462 ret = __dev_open(dev);
7465 if ((flags ^ dev->gflags) & IFF_PROMISC) {
7466 int inc = (flags & IFF_PROMISC) ? 1 : -1;
7467 unsigned int old_flags = dev->flags;
7469 dev->gflags ^= IFF_PROMISC;
7471 if (__dev_set_promiscuity(dev, inc, false) >= 0)
7472 if (dev->flags != old_flags)
7473 dev_set_rx_mode(dev);
7476 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI
7477 * is important. Some (broken) drivers set IFF_PROMISC, when
7478 * IFF_ALLMULTI is requested not asking us and not reporting.
7480 if ((flags ^ dev->gflags) & IFF_ALLMULTI) {
7481 int inc = (flags & IFF_ALLMULTI) ? 1 : -1;
7483 dev->gflags ^= IFF_ALLMULTI;
7484 __dev_set_allmulti(dev, inc, false);
7490 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags,
7491 unsigned int gchanges)
7493 unsigned int changes = dev->flags ^ old_flags;
7496 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC);
7498 if (changes & IFF_UP) {
7499 if (dev->flags & IFF_UP)
7500 call_netdevice_notifiers(NETDEV_UP, dev);
7502 call_netdevice_notifiers(NETDEV_DOWN, dev);
7505 if (dev->flags & IFF_UP &&
7506 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) {
7507 struct netdev_notifier_change_info change_info = {
7511 .flags_changed = changes,
7514 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info);
7519 * dev_change_flags - change device settings
7521 * @flags: device state flags
7523 * Change settings on device based state flags. The flags are
7524 * in the userspace exported format.
7526 int dev_change_flags(struct net_device *dev, unsigned int flags)
7529 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags;
7531 ret = __dev_change_flags(dev, flags);
7535 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags);
7536 __dev_notify_flags(dev, old_flags, changes);
7539 EXPORT_SYMBOL(dev_change_flags);
7541 int __dev_set_mtu(struct net_device *dev, int new_mtu)
7543 const struct net_device_ops *ops = dev->netdev_ops;
7545 if (ops->ndo_change_mtu)
7546 return ops->ndo_change_mtu(dev, new_mtu);
7551 EXPORT_SYMBOL(__dev_set_mtu);
7554 * dev_set_mtu_ext - Change maximum transfer unit
7556 * @new_mtu: new transfer unit
7557 * @extack: netlink extended ack
7559 * Change the maximum transfer size of the network device.
7561 int dev_set_mtu_ext(struct net_device *dev, int new_mtu,
7562 struct netlink_ext_ack *extack)
7566 if (new_mtu == dev->mtu)
7569 /* MTU must be positive, and in range */
7570 if (new_mtu < 0 || new_mtu < dev->min_mtu) {
7571 NL_SET_ERR_MSG(extack, "mtu less than device minimum");
7575 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) {
7576 NL_SET_ERR_MSG(extack, "mtu greater than device maximum");
7580 if (!netif_device_present(dev))
7583 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev);
7584 err = notifier_to_errno(err);
7588 orig_mtu = dev->mtu;
7589 err = __dev_set_mtu(dev, new_mtu);
7592 err = call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
7593 err = notifier_to_errno(err);
7595 /* setting mtu back and notifying everyone again,
7596 * so that they have a chance to revert changes.
7598 __dev_set_mtu(dev, orig_mtu);
7599 call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
7605 int dev_set_mtu(struct net_device *dev, int new_mtu)
7607 struct netlink_ext_ack extack;
7610 memset(&extack, 0, sizeof(extack));
7611 err = dev_set_mtu_ext(dev, new_mtu, &extack);
7612 if (err && extack._msg)
7613 net_err_ratelimited("%s: %s\n", dev->name, extack._msg);
7616 EXPORT_SYMBOL(dev_set_mtu);
7619 * dev_change_tx_queue_len - Change TX queue length of a netdevice
7621 * @new_len: new tx queue length
7623 int dev_change_tx_queue_len(struct net_device *dev, unsigned long new_len)
7625 unsigned int orig_len = dev->tx_queue_len;
7628 if (new_len != (unsigned int)new_len)
7631 if (new_len != orig_len) {
7632 dev->tx_queue_len = new_len;
7633 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev);
7634 res = notifier_to_errno(res);
7637 res = dev_qdisc_change_tx_queue_len(dev);
7645 netdev_err(dev, "refused to change device tx_queue_len\n");
7646 dev->tx_queue_len = orig_len;
7651 * dev_set_group - Change group this device belongs to
7653 * @new_group: group this device should belong to
7655 void dev_set_group(struct net_device *dev, int new_group)
7657 dev->group = new_group;
7659 EXPORT_SYMBOL(dev_set_group);
7662 * dev_set_mac_address - Change Media Access Control Address
7666 * Change the hardware (MAC) address of the device
7668 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa)
7670 const struct net_device_ops *ops = dev->netdev_ops;
7673 if (!ops->ndo_set_mac_address)
7675 if (sa->sa_family != dev->type)
7677 if (!netif_device_present(dev))
7679 err = ops->ndo_set_mac_address(dev, sa);
7682 dev->addr_assign_type = NET_ADDR_SET;
7683 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
7684 add_device_randomness(dev->dev_addr, dev->addr_len);
7687 EXPORT_SYMBOL(dev_set_mac_address);
7690 * dev_change_carrier - Change device carrier
7692 * @new_carrier: new value
7694 * Change device carrier
7696 int dev_change_carrier(struct net_device *dev, bool new_carrier)
7698 const struct net_device_ops *ops = dev->netdev_ops;
7700 if (!ops->ndo_change_carrier)
7702 if (!netif_device_present(dev))
7704 return ops->ndo_change_carrier(dev, new_carrier);
7706 EXPORT_SYMBOL(dev_change_carrier);
7709 * dev_get_phys_port_id - Get device physical port ID
7713 * Get device physical port ID
7715 int dev_get_phys_port_id(struct net_device *dev,
7716 struct netdev_phys_item_id *ppid)
7718 const struct net_device_ops *ops = dev->netdev_ops;
7720 if (!ops->ndo_get_phys_port_id)
7722 return ops->ndo_get_phys_port_id(dev, ppid);
7724 EXPORT_SYMBOL(dev_get_phys_port_id);
7727 * dev_get_phys_port_name - Get device physical port name
7730 * @len: limit of bytes to copy to name
7732 * Get device physical port name
7734 int dev_get_phys_port_name(struct net_device *dev,
7735 char *name, size_t len)
7737 const struct net_device_ops *ops = dev->netdev_ops;
7739 if (!ops->ndo_get_phys_port_name)
7741 return ops->ndo_get_phys_port_name(dev, name, len);
7743 EXPORT_SYMBOL(dev_get_phys_port_name);
7746 * dev_change_proto_down - update protocol port state information
7748 * @proto_down: new value
7750 * This info can be used by switch drivers to set the phys state of the
7753 int dev_change_proto_down(struct net_device *dev, bool proto_down)
7755 const struct net_device_ops *ops = dev->netdev_ops;
7757 if (!ops->ndo_change_proto_down)
7759 if (!netif_device_present(dev))
7761 return ops->ndo_change_proto_down(dev, proto_down);
7763 EXPORT_SYMBOL(dev_change_proto_down);
7765 u32 __dev_xdp_query(struct net_device *dev, bpf_op_t bpf_op,
7766 enum bpf_netdev_command cmd)
7768 struct netdev_bpf xdp;
7773 memset(&xdp, 0, sizeof(xdp));
7776 /* Query must always succeed. */
7777 WARN_ON(bpf_op(dev, &xdp) < 0 && cmd == XDP_QUERY_PROG);
7782 static int dev_xdp_install(struct net_device *dev, bpf_op_t bpf_op,
7783 struct netlink_ext_ack *extack, u32 flags,
7784 struct bpf_prog *prog)
7786 struct netdev_bpf xdp;
7788 memset(&xdp, 0, sizeof(xdp));
7789 if (flags & XDP_FLAGS_HW_MODE)
7790 xdp.command = XDP_SETUP_PROG_HW;
7792 xdp.command = XDP_SETUP_PROG;
7793 xdp.extack = extack;
7797 return bpf_op(dev, &xdp);
7800 static void dev_xdp_uninstall(struct net_device *dev)
7802 struct netdev_bpf xdp;
7805 /* Remove generic XDP */
7806 WARN_ON(dev_xdp_install(dev, generic_xdp_install, NULL, 0, NULL));
7808 /* Remove from the driver */
7809 ndo_bpf = dev->netdev_ops->ndo_bpf;
7813 memset(&xdp, 0, sizeof(xdp));
7814 xdp.command = XDP_QUERY_PROG;
7815 WARN_ON(ndo_bpf(dev, &xdp));
7817 WARN_ON(dev_xdp_install(dev, ndo_bpf, NULL, xdp.prog_flags,
7820 /* Remove HW offload */
7821 memset(&xdp, 0, sizeof(xdp));
7822 xdp.command = XDP_QUERY_PROG_HW;
7823 if (!ndo_bpf(dev, &xdp) && xdp.prog_id)
7824 WARN_ON(dev_xdp_install(dev, ndo_bpf, NULL, xdp.prog_flags,
7829 * dev_change_xdp_fd - set or clear a bpf program for a device rx path
7831 * @extack: netlink extended ack
7832 * @fd: new program fd or negative value to clear
7833 * @flags: xdp-related flags
7835 * Set or clear a bpf program for a device
7837 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack,
7840 const struct net_device_ops *ops = dev->netdev_ops;
7841 enum bpf_netdev_command query;
7842 struct bpf_prog *prog = NULL;
7843 bpf_op_t bpf_op, bpf_chk;
7848 query = flags & XDP_FLAGS_HW_MODE ? XDP_QUERY_PROG_HW : XDP_QUERY_PROG;
7850 bpf_op = bpf_chk = ops->ndo_bpf;
7851 if (!bpf_op && (flags & (XDP_FLAGS_DRV_MODE | XDP_FLAGS_HW_MODE)))
7853 if (!bpf_op || (flags & XDP_FLAGS_SKB_MODE))
7854 bpf_op = generic_xdp_install;
7855 if (bpf_op == bpf_chk)
7856 bpf_chk = generic_xdp_install;
7859 if (__dev_xdp_query(dev, bpf_chk, XDP_QUERY_PROG) ||
7860 __dev_xdp_query(dev, bpf_chk, XDP_QUERY_PROG_HW))
7862 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) &&
7863 __dev_xdp_query(dev, bpf_op, query))
7866 prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP,
7867 bpf_op == ops->ndo_bpf);
7869 return PTR_ERR(prog);
7871 if (!(flags & XDP_FLAGS_HW_MODE) &&
7872 bpf_prog_is_dev_bound(prog->aux)) {
7873 NL_SET_ERR_MSG(extack, "using device-bound program without HW_MODE flag is not supported");
7879 err = dev_xdp_install(dev, bpf_op, extack, flags, prog);
7880 if (err < 0 && prog)
7887 * dev_new_index - allocate an ifindex
7888 * @net: the applicable net namespace
7890 * Returns a suitable unique value for a new device interface
7891 * number. The caller must hold the rtnl semaphore or the
7892 * dev_base_lock to be sure it remains unique.
7894 static int dev_new_index(struct net *net)
7896 int ifindex = net->ifindex;
7901 if (!__dev_get_by_index(net, ifindex))
7902 return net->ifindex = ifindex;
7906 /* Delayed registration/unregisteration */
7907 static LIST_HEAD(net_todo_list);
7908 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq);
7910 static void net_set_todo(struct net_device *dev)
7912 list_add_tail(&dev->todo_list, &net_todo_list);
7913 dev_net(dev)->dev_unreg_count++;
7916 static void rollback_registered_many(struct list_head *head)
7918 struct net_device *dev, *tmp;
7919 LIST_HEAD(close_head);
7921 BUG_ON(dev_boot_phase);
7924 list_for_each_entry_safe(dev, tmp, head, unreg_list) {
7925 /* Some devices call without registering
7926 * for initialization unwind. Remove those
7927 * devices and proceed with the remaining.
7929 if (dev->reg_state == NETREG_UNINITIALIZED) {
7930 pr_debug("unregister_netdevice: device %s/%p never was registered\n",
7934 list_del(&dev->unreg_list);
7937 dev->dismantle = true;
7938 BUG_ON(dev->reg_state != NETREG_REGISTERED);
7941 /* If device is running, close it first. */
7942 list_for_each_entry(dev, head, unreg_list)
7943 list_add_tail(&dev->close_list, &close_head);
7944 dev_close_many(&close_head, true);
7946 list_for_each_entry(dev, head, unreg_list) {
7947 /* And unlink it from device chain. */
7948 unlist_netdevice(dev);
7950 dev->reg_state = NETREG_UNREGISTERING;
7952 flush_all_backlogs();
7956 list_for_each_entry(dev, head, unreg_list) {
7957 struct sk_buff *skb = NULL;
7959 /* Shutdown queueing discipline. */
7962 dev_xdp_uninstall(dev);
7964 /* Notify protocols, that we are about to destroy
7965 * this device. They should clean all the things.
7967 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
7969 if (!dev->rtnl_link_ops ||
7970 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
7971 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0,
7972 GFP_KERNEL, NULL, 0);
7975 * Flush the unicast and multicast chains
7980 if (dev->netdev_ops->ndo_uninit)
7981 dev->netdev_ops->ndo_uninit(dev);
7984 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL);
7986 /* Notifier chain MUST detach us all upper devices. */
7987 WARN_ON(netdev_has_any_upper_dev(dev));
7988 WARN_ON(netdev_has_any_lower_dev(dev));
7990 /* Remove entries from kobject tree */
7991 netdev_unregister_kobject(dev);
7993 /* Remove XPS queueing entries */
7994 netif_reset_xps_queues_gt(dev, 0);
8000 list_for_each_entry(dev, head, unreg_list)
8004 static void rollback_registered(struct net_device *dev)
8008 list_add(&dev->unreg_list, &single);
8009 rollback_registered_many(&single);
8013 static netdev_features_t netdev_sync_upper_features(struct net_device *lower,
8014 struct net_device *upper, netdev_features_t features)
8016 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
8017 netdev_features_t feature;
8020 for_each_netdev_feature(&upper_disables, feature_bit) {
8021 feature = __NETIF_F_BIT(feature_bit);
8022 if (!(upper->wanted_features & feature)
8023 && (features & feature)) {
8024 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n",
8025 &feature, upper->name);
8026 features &= ~feature;
8033 static void netdev_sync_lower_features(struct net_device *upper,
8034 struct net_device *lower, netdev_features_t features)
8036 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
8037 netdev_features_t feature;
8040 for_each_netdev_feature(&upper_disables, feature_bit) {
8041 feature = __NETIF_F_BIT(feature_bit);
8042 if (!(features & feature) && (lower->features & feature)) {
8043 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n",
8044 &feature, lower->name);
8045 lower->wanted_features &= ~feature;
8046 netdev_update_features(lower);
8048 if (unlikely(lower->features & feature))
8049 netdev_WARN(upper, "failed to disable %pNF on %s!\n",
8050 &feature, lower->name);
8055 static netdev_features_t netdev_fix_features(struct net_device *dev,
8056 netdev_features_t features)
8058 /* Fix illegal checksum combinations */
8059 if ((features & NETIF_F_HW_CSUM) &&
8060 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
8061 netdev_warn(dev, "mixed HW and IP checksum settings.\n");
8062 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
8065 /* TSO requires that SG is present as well. */
8066 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) {
8067 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n");
8068 features &= ~NETIF_F_ALL_TSO;
8071 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) &&
8072 !(features & NETIF_F_IP_CSUM)) {
8073 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n");
8074 features &= ~NETIF_F_TSO;
8075 features &= ~NETIF_F_TSO_ECN;
8078 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) &&
8079 !(features & NETIF_F_IPV6_CSUM)) {
8080 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n");
8081 features &= ~NETIF_F_TSO6;
8084 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */
8085 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO))
8086 features &= ~NETIF_F_TSO_MANGLEID;
8088 /* TSO ECN requires that TSO is present as well. */
8089 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN)
8090 features &= ~NETIF_F_TSO_ECN;
8092 /* Software GSO depends on SG. */
8093 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) {
8094 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n");
8095 features &= ~NETIF_F_GSO;
8098 /* GSO partial features require GSO partial be set */
8099 if ((features & dev->gso_partial_features) &&
8100 !(features & NETIF_F_GSO_PARTIAL)) {
8102 "Dropping partially supported GSO features since no GSO partial.\n");
8103 features &= ~dev->gso_partial_features;
8106 if (!(features & NETIF_F_RXCSUM)) {
8107 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet
8108 * successfully merged by hardware must also have the
8109 * checksum verified by hardware. If the user does not
8110 * want to enable RXCSUM, logically, we should disable GRO_HW.
8112 if (features & NETIF_F_GRO_HW) {
8113 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n");
8114 features &= ~NETIF_F_GRO_HW;
8118 /* LRO/HW-GRO features cannot be combined with RX-FCS */
8119 if (features & NETIF_F_RXFCS) {
8120 if (features & NETIF_F_LRO) {
8121 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n");
8122 features &= ~NETIF_F_LRO;
8125 if (features & NETIF_F_GRO_HW) {
8126 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n");
8127 features &= ~NETIF_F_GRO_HW;
8134 int __netdev_update_features(struct net_device *dev)
8136 struct net_device *upper, *lower;
8137 netdev_features_t features;
8138 struct list_head *iter;
8143 features = netdev_get_wanted_features(dev);
8145 if (dev->netdev_ops->ndo_fix_features)
8146 features = dev->netdev_ops->ndo_fix_features(dev, features);
8148 /* driver might be less strict about feature dependencies */
8149 features = netdev_fix_features(dev, features);
8151 /* some features can't be enabled if they're off an an upper device */
8152 netdev_for_each_upper_dev_rcu(dev, upper, iter)
8153 features = netdev_sync_upper_features(dev, upper, features);
8155 if (dev->features == features)
8158 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n",
8159 &dev->features, &features);
8161 if (dev->netdev_ops->ndo_set_features)
8162 err = dev->netdev_ops->ndo_set_features(dev, features);
8166 if (unlikely(err < 0)) {
8168 "set_features() failed (%d); wanted %pNF, left %pNF\n",
8169 err, &features, &dev->features);
8170 /* return non-0 since some features might have changed and
8171 * it's better to fire a spurious notification than miss it
8177 /* some features must be disabled on lower devices when disabled
8178 * on an upper device (think: bonding master or bridge)
8180 netdev_for_each_lower_dev(dev, lower, iter)
8181 netdev_sync_lower_features(dev, lower, features);
8184 netdev_features_t diff = features ^ dev->features;
8186 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) {
8187 /* udp_tunnel_{get,drop}_rx_info both need
8188 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the
8189 * device, or they won't do anything.
8190 * Thus we need to update dev->features
8191 * *before* calling udp_tunnel_get_rx_info,
8192 * but *after* calling udp_tunnel_drop_rx_info.
8194 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) {
8195 dev->features = features;
8196 udp_tunnel_get_rx_info(dev);
8198 udp_tunnel_drop_rx_info(dev);
8202 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) {
8203 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) {
8204 dev->features = features;
8205 err |= vlan_get_rx_ctag_filter_info(dev);
8207 vlan_drop_rx_ctag_filter_info(dev);
8211 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) {
8212 if (features & NETIF_F_HW_VLAN_STAG_FILTER) {
8213 dev->features = features;
8214 err |= vlan_get_rx_stag_filter_info(dev);
8216 vlan_drop_rx_stag_filter_info(dev);
8220 dev->features = features;
8223 return err < 0 ? 0 : 1;
8227 * netdev_update_features - recalculate device features
8228 * @dev: the device to check
8230 * Recalculate dev->features set and send notifications if it
8231 * has changed. Should be called after driver or hardware dependent
8232 * conditions might have changed that influence the features.
8234 void netdev_update_features(struct net_device *dev)
8236 if (__netdev_update_features(dev))
8237 netdev_features_change(dev);
8239 EXPORT_SYMBOL(netdev_update_features);
8242 * netdev_change_features - recalculate device features
8243 * @dev: the device to check
8245 * Recalculate dev->features set and send notifications even
8246 * if they have not changed. Should be called instead of
8247 * netdev_update_features() if also dev->vlan_features might
8248 * have changed to allow the changes to be propagated to stacked
8251 void netdev_change_features(struct net_device *dev)
8253 __netdev_update_features(dev);
8254 netdev_features_change(dev);
8256 EXPORT_SYMBOL(netdev_change_features);
8259 * netif_stacked_transfer_operstate - transfer operstate
8260 * @rootdev: the root or lower level device to transfer state from
8261 * @dev: the device to transfer operstate to
8263 * Transfer operational state from root to device. This is normally
8264 * called when a stacking relationship exists between the root
8265 * device and the device(a leaf device).
8267 void netif_stacked_transfer_operstate(const struct net_device *rootdev,
8268 struct net_device *dev)
8270 if (rootdev->operstate == IF_OPER_DORMANT)
8271 netif_dormant_on(dev);
8273 netif_dormant_off(dev);
8275 if (netif_carrier_ok(rootdev))
8276 netif_carrier_on(dev);
8278 netif_carrier_off(dev);
8280 EXPORT_SYMBOL(netif_stacked_transfer_operstate);
8282 static int netif_alloc_rx_queues(struct net_device *dev)
8284 unsigned int i, count = dev->num_rx_queues;
8285 struct netdev_rx_queue *rx;
8286 size_t sz = count * sizeof(*rx);
8291 rx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
8297 for (i = 0; i < count; i++) {
8300 /* XDP RX-queue setup */
8301 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i);
8308 /* Rollback successful reg's and free other resources */
8310 xdp_rxq_info_unreg(&rx[i].xdp_rxq);
8316 static void netif_free_rx_queues(struct net_device *dev)
8318 unsigned int i, count = dev->num_rx_queues;
8320 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */
8324 for (i = 0; i < count; i++)
8325 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq);
8330 static void netdev_init_one_queue(struct net_device *dev,
8331 struct netdev_queue *queue, void *_unused)
8333 /* Initialize queue lock */
8334 spin_lock_init(&queue->_xmit_lock);
8335 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type);
8336 queue->xmit_lock_owner = -1;
8337 netdev_queue_numa_node_write(queue, NUMA_NO_NODE);
8340 dql_init(&queue->dql, HZ);
8344 static void netif_free_tx_queues(struct net_device *dev)
8349 static int netif_alloc_netdev_queues(struct net_device *dev)
8351 unsigned int count = dev->num_tx_queues;
8352 struct netdev_queue *tx;
8353 size_t sz = count * sizeof(*tx);
8355 if (count < 1 || count > 0xffff)
8358 tx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
8364 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL);
8365 spin_lock_init(&dev->tx_global_lock);
8370 void netif_tx_stop_all_queues(struct net_device *dev)
8374 for (i = 0; i < dev->num_tx_queues; i++) {
8375 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
8377 netif_tx_stop_queue(txq);
8380 EXPORT_SYMBOL(netif_tx_stop_all_queues);
8383 * register_netdevice - register a network device
8384 * @dev: device to register
8386 * Take a completed network device structure and add it to the kernel
8387 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
8388 * chain. 0 is returned on success. A negative errno code is returned
8389 * on a failure to set up the device, or if the name is a duplicate.
8391 * Callers must hold the rtnl semaphore. You may want
8392 * register_netdev() instead of this.
8395 * The locking appears insufficient to guarantee two parallel registers
8396 * will not get the same name.
8399 int register_netdevice(struct net_device *dev)
8402 struct net *net = dev_net(dev);
8404 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE <
8405 NETDEV_FEATURE_COUNT);
8406 BUG_ON(dev_boot_phase);
8411 /* When net_device's are persistent, this will be fatal. */
8412 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
8415 spin_lock_init(&dev->addr_list_lock);
8416 netdev_set_addr_lockdep_class(dev);
8418 ret = dev_get_valid_name(net, dev, dev->name);
8422 /* Init, if this function is available */
8423 if (dev->netdev_ops->ndo_init) {
8424 ret = dev->netdev_ops->ndo_init(dev);
8432 if (((dev->hw_features | dev->features) &
8433 NETIF_F_HW_VLAN_CTAG_FILTER) &&
8434 (!dev->netdev_ops->ndo_vlan_rx_add_vid ||
8435 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) {
8436 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n");
8443 dev->ifindex = dev_new_index(net);
8444 else if (__dev_get_by_index(net, dev->ifindex))
8447 /* Transfer changeable features to wanted_features and enable
8448 * software offloads (GSO and GRO).
8450 dev->hw_features |= NETIF_F_SOFT_FEATURES;
8451 dev->features |= NETIF_F_SOFT_FEATURES;
8453 if (dev->netdev_ops->ndo_udp_tunnel_add) {
8454 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT;
8455 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT;
8458 dev->wanted_features = dev->features & dev->hw_features;
8460 if (!(dev->flags & IFF_LOOPBACK))
8461 dev->hw_features |= NETIF_F_NOCACHE_COPY;
8463 /* If IPv4 TCP segmentation offload is supported we should also
8464 * allow the device to enable segmenting the frame with the option
8465 * of ignoring a static IP ID value. This doesn't enable the
8466 * feature itself but allows the user to enable it later.
8468 if (dev->hw_features & NETIF_F_TSO)
8469 dev->hw_features |= NETIF_F_TSO_MANGLEID;
8470 if (dev->vlan_features & NETIF_F_TSO)
8471 dev->vlan_features |= NETIF_F_TSO_MANGLEID;
8472 if (dev->mpls_features & NETIF_F_TSO)
8473 dev->mpls_features |= NETIF_F_TSO_MANGLEID;
8474 if (dev->hw_enc_features & NETIF_F_TSO)
8475 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID;
8477 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices.
8479 dev->vlan_features |= NETIF_F_HIGHDMA;
8481 /* Make NETIF_F_SG inheritable to tunnel devices.
8483 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL;
8485 /* Make NETIF_F_SG inheritable to MPLS.
8487 dev->mpls_features |= NETIF_F_SG;
8489 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev);
8490 ret = notifier_to_errno(ret);
8494 ret = netdev_register_kobject(dev);
8497 dev->reg_state = NETREG_REGISTERED;
8499 __netdev_update_features(dev);
8502 * Default initial state at registry is that the
8503 * device is present.
8506 set_bit(__LINK_STATE_PRESENT, &dev->state);
8508 linkwatch_init_dev(dev);
8510 dev_init_scheduler(dev);
8512 list_netdevice(dev);
8513 add_device_randomness(dev->dev_addr, dev->addr_len);
8515 /* If the device has permanent device address, driver should
8516 * set dev_addr and also addr_assign_type should be set to
8517 * NET_ADDR_PERM (default value).
8519 if (dev->addr_assign_type == NET_ADDR_PERM)
8520 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len);
8522 /* Notify protocols, that a new device appeared. */
8523 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
8524 ret = notifier_to_errno(ret);
8526 rollback_registered(dev);
8527 dev->reg_state = NETREG_UNREGISTERED;
8530 * Prevent userspace races by waiting until the network
8531 * device is fully setup before sending notifications.
8533 if (!dev->rtnl_link_ops ||
8534 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
8535 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
8541 if (dev->netdev_ops->ndo_uninit)
8542 dev->netdev_ops->ndo_uninit(dev);
8543 if (dev->priv_destructor)
8544 dev->priv_destructor(dev);
8547 EXPORT_SYMBOL(register_netdevice);
8550 * init_dummy_netdev - init a dummy network device for NAPI
8551 * @dev: device to init
8553 * This takes a network device structure and initialize the minimum
8554 * amount of fields so it can be used to schedule NAPI polls without
8555 * registering a full blown interface. This is to be used by drivers
8556 * that need to tie several hardware interfaces to a single NAPI
8557 * poll scheduler due to HW limitations.
8559 int init_dummy_netdev(struct net_device *dev)
8561 /* Clear everything. Note we don't initialize spinlocks
8562 * are they aren't supposed to be taken by any of the
8563 * NAPI code and this dummy netdev is supposed to be
8564 * only ever used for NAPI polls
8566 memset(dev, 0, sizeof(struct net_device));
8568 /* make sure we BUG if trying to hit standard
8569 * register/unregister code path
8571 dev->reg_state = NETREG_DUMMY;
8573 /* NAPI wants this */
8574 INIT_LIST_HEAD(&dev->napi_list);
8576 /* a dummy interface is started by default */
8577 set_bit(__LINK_STATE_PRESENT, &dev->state);
8578 set_bit(__LINK_STATE_START, &dev->state);
8580 /* Note : We dont allocate pcpu_refcnt for dummy devices,
8581 * because users of this 'device' dont need to change
8587 EXPORT_SYMBOL_GPL(init_dummy_netdev);
8591 * register_netdev - register a network device
8592 * @dev: device to register
8594 * Take a completed network device structure and add it to the kernel
8595 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
8596 * chain. 0 is returned on success. A negative errno code is returned
8597 * on a failure to set up the device, or if the name is a duplicate.
8599 * This is a wrapper around register_netdevice that takes the rtnl semaphore
8600 * and expands the device name if you passed a format string to
8603 int register_netdev(struct net_device *dev)
8607 if (rtnl_lock_killable())
8609 err = register_netdevice(dev);
8613 EXPORT_SYMBOL(register_netdev);
8615 int netdev_refcnt_read(const struct net_device *dev)
8619 for_each_possible_cpu(i)
8620 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
8623 EXPORT_SYMBOL(netdev_refcnt_read);
8626 * netdev_wait_allrefs - wait until all references are gone.
8627 * @dev: target net_device
8629 * This is called when unregistering network devices.
8631 * Any protocol or device that holds a reference should register
8632 * for netdevice notification, and cleanup and put back the
8633 * reference if they receive an UNREGISTER event.
8634 * We can get stuck here if buggy protocols don't correctly
8637 static void netdev_wait_allrefs(struct net_device *dev)
8639 unsigned long rebroadcast_time, warning_time;
8642 linkwatch_forget_dev(dev);
8644 rebroadcast_time = warning_time = jiffies;
8645 refcnt = netdev_refcnt_read(dev);
8647 while (refcnt != 0) {
8648 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
8651 /* Rebroadcast unregister notification */
8652 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
8658 if (test_bit(__LINK_STATE_LINKWATCH_PENDING,
8660 /* We must not have linkwatch events
8661 * pending on unregister. If this
8662 * happens, we simply run the queue
8663 * unscheduled, resulting in a noop
8666 linkwatch_run_queue();
8671 rebroadcast_time = jiffies;
8676 refcnt = netdev_refcnt_read(dev);
8678 if (time_after(jiffies, warning_time + 10 * HZ)) {
8679 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n",
8681 warning_time = jiffies;
8690 * register_netdevice(x1);
8691 * register_netdevice(x2);
8693 * unregister_netdevice(y1);
8694 * unregister_netdevice(y2);
8700 * We are invoked by rtnl_unlock().
8701 * This allows us to deal with problems:
8702 * 1) We can delete sysfs objects which invoke hotplug
8703 * without deadlocking with linkwatch via keventd.
8704 * 2) Since we run with the RTNL semaphore not held, we can sleep
8705 * safely in order to wait for the netdev refcnt to drop to zero.
8707 * We must not return until all unregister events added during
8708 * the interval the lock was held have been completed.
8710 void netdev_run_todo(void)
8712 struct list_head list;
8714 /* Snapshot list, allow later requests */
8715 list_replace_init(&net_todo_list, &list);
8720 /* Wait for rcu callbacks to finish before next phase */
8721 if (!list_empty(&list))
8724 while (!list_empty(&list)) {
8725 struct net_device *dev
8726 = list_first_entry(&list, struct net_device, todo_list);
8727 list_del(&dev->todo_list);
8729 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) {
8730 pr_err("network todo '%s' but state %d\n",
8731 dev->name, dev->reg_state);
8736 dev->reg_state = NETREG_UNREGISTERED;
8738 netdev_wait_allrefs(dev);
8741 BUG_ON(netdev_refcnt_read(dev));
8742 BUG_ON(!list_empty(&dev->ptype_all));
8743 BUG_ON(!list_empty(&dev->ptype_specific));
8744 WARN_ON(rcu_access_pointer(dev->ip_ptr));
8745 WARN_ON(rcu_access_pointer(dev->ip6_ptr));
8746 #if IS_ENABLED(CONFIG_DECNET)
8747 WARN_ON(dev->dn_ptr);
8749 if (dev->priv_destructor)
8750 dev->priv_destructor(dev);
8751 if (dev->needs_free_netdev)
8754 /* Report a network device has been unregistered */
8756 dev_net(dev)->dev_unreg_count--;
8758 wake_up(&netdev_unregistering_wq);
8760 /* Free network device */
8761 kobject_put(&dev->dev.kobj);
8765 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has
8766 * all the same fields in the same order as net_device_stats, with only
8767 * the type differing, but rtnl_link_stats64 may have additional fields
8768 * at the end for newer counters.
8770 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
8771 const struct net_device_stats *netdev_stats)
8773 #if BITS_PER_LONG == 64
8774 BUILD_BUG_ON(sizeof(*stats64) < sizeof(*netdev_stats));
8775 memcpy(stats64, netdev_stats, sizeof(*netdev_stats));
8776 /* zero out counters that only exist in rtnl_link_stats64 */
8777 memset((char *)stats64 + sizeof(*netdev_stats), 0,
8778 sizeof(*stats64) - sizeof(*netdev_stats));
8780 size_t i, n = sizeof(*netdev_stats) / sizeof(unsigned long);
8781 const unsigned long *src = (const unsigned long *)netdev_stats;
8782 u64 *dst = (u64 *)stats64;
8784 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64));
8785 for (i = 0; i < n; i++)
8787 /* zero out counters that only exist in rtnl_link_stats64 */
8788 memset((char *)stats64 + n * sizeof(u64), 0,
8789 sizeof(*stats64) - n * sizeof(u64));
8792 EXPORT_SYMBOL(netdev_stats_to_stats64);
8795 * dev_get_stats - get network device statistics
8796 * @dev: device to get statistics from
8797 * @storage: place to store stats
8799 * Get network statistics from device. Return @storage.
8800 * The device driver may provide its own method by setting
8801 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats;
8802 * otherwise the internal statistics structure is used.
8804 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev,
8805 struct rtnl_link_stats64 *storage)
8807 const struct net_device_ops *ops = dev->netdev_ops;
8809 if (ops->ndo_get_stats64) {
8810 memset(storage, 0, sizeof(*storage));
8811 ops->ndo_get_stats64(dev, storage);
8812 } else if (ops->ndo_get_stats) {
8813 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev));
8815 netdev_stats_to_stats64(storage, &dev->stats);
8817 storage->rx_dropped += (unsigned long)atomic_long_read(&dev->rx_dropped);
8818 storage->tx_dropped += (unsigned long)atomic_long_read(&dev->tx_dropped);
8819 storage->rx_nohandler += (unsigned long)atomic_long_read(&dev->rx_nohandler);
8822 EXPORT_SYMBOL(dev_get_stats);
8824 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev)
8826 struct netdev_queue *queue = dev_ingress_queue(dev);
8828 #ifdef CONFIG_NET_CLS_ACT
8831 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
8834 netdev_init_one_queue(dev, queue, NULL);
8835 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc);
8836 queue->qdisc_sleeping = &noop_qdisc;
8837 rcu_assign_pointer(dev->ingress_queue, queue);
8842 static const struct ethtool_ops default_ethtool_ops;
8844 void netdev_set_default_ethtool_ops(struct net_device *dev,
8845 const struct ethtool_ops *ops)
8847 if (dev->ethtool_ops == &default_ethtool_ops)
8848 dev->ethtool_ops = ops;
8850 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops);
8852 void netdev_freemem(struct net_device *dev)
8854 char *addr = (char *)dev - dev->padded;
8860 * alloc_netdev_mqs - allocate network device
8861 * @sizeof_priv: size of private data to allocate space for
8862 * @name: device name format string
8863 * @name_assign_type: origin of device name
8864 * @setup: callback to initialize device
8865 * @txqs: the number of TX subqueues to allocate
8866 * @rxqs: the number of RX subqueues to allocate
8868 * Allocates a struct net_device with private data area for driver use
8869 * and performs basic initialization. Also allocates subqueue structs
8870 * for each queue on the device.
8872 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
8873 unsigned char name_assign_type,
8874 void (*setup)(struct net_device *),
8875 unsigned int txqs, unsigned int rxqs)
8877 struct net_device *dev;
8878 unsigned int alloc_size;
8879 struct net_device *p;
8881 BUG_ON(strlen(name) >= sizeof(dev->name));
8884 pr_err("alloc_netdev: Unable to allocate device with zero queues\n");
8889 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n");
8893 alloc_size = sizeof(struct net_device);
8895 /* ensure 32-byte alignment of private area */
8896 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
8897 alloc_size += sizeof_priv;
8899 /* ensure 32-byte alignment of whole construct */
8900 alloc_size += NETDEV_ALIGN - 1;
8902 p = kvzalloc(alloc_size, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
8906 dev = PTR_ALIGN(p, NETDEV_ALIGN);
8907 dev->padded = (char *)dev - (char *)p;
8909 dev->pcpu_refcnt = alloc_percpu(int);
8910 if (!dev->pcpu_refcnt)
8913 if (dev_addr_init(dev))
8919 dev_net_set(dev, &init_net);
8921 dev->gso_max_size = GSO_MAX_SIZE;
8922 dev->gso_max_segs = GSO_MAX_SEGS;
8924 INIT_LIST_HEAD(&dev->napi_list);
8925 INIT_LIST_HEAD(&dev->unreg_list);
8926 INIT_LIST_HEAD(&dev->close_list);
8927 INIT_LIST_HEAD(&dev->link_watch_list);
8928 INIT_LIST_HEAD(&dev->adj_list.upper);
8929 INIT_LIST_HEAD(&dev->adj_list.lower);
8930 INIT_LIST_HEAD(&dev->ptype_all);
8931 INIT_LIST_HEAD(&dev->ptype_specific);
8932 #ifdef CONFIG_NET_SCHED
8933 hash_init(dev->qdisc_hash);
8935 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM;
8938 if (!dev->tx_queue_len) {
8939 dev->priv_flags |= IFF_NO_QUEUE;
8940 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN;
8943 dev->num_tx_queues = txqs;
8944 dev->real_num_tx_queues = txqs;
8945 if (netif_alloc_netdev_queues(dev))
8948 dev->num_rx_queues = rxqs;
8949 dev->real_num_rx_queues = rxqs;
8950 if (netif_alloc_rx_queues(dev))
8953 strcpy(dev->name, name);
8954 dev->name_assign_type = name_assign_type;
8955 dev->group = INIT_NETDEV_GROUP;
8956 if (!dev->ethtool_ops)
8957 dev->ethtool_ops = &default_ethtool_ops;
8959 nf_hook_ingress_init(dev);
8968 free_percpu(dev->pcpu_refcnt);
8970 netdev_freemem(dev);
8973 EXPORT_SYMBOL(alloc_netdev_mqs);
8976 * free_netdev - free network device
8979 * This function does the last stage of destroying an allocated device
8980 * interface. The reference to the device object is released. If this
8981 * is the last reference then it will be freed.Must be called in process
8984 void free_netdev(struct net_device *dev)
8986 struct napi_struct *p, *n;
8989 netif_free_tx_queues(dev);
8990 netif_free_rx_queues(dev);
8992 kfree(rcu_dereference_protected(dev->ingress_queue, 1));
8994 /* Flush device addresses */
8995 dev_addr_flush(dev);
8997 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
9000 free_percpu(dev->pcpu_refcnt);
9001 dev->pcpu_refcnt = NULL;
9003 /* Compatibility with error handling in drivers */
9004 if (dev->reg_state == NETREG_UNINITIALIZED) {
9005 netdev_freemem(dev);
9009 BUG_ON(dev->reg_state != NETREG_UNREGISTERED);
9010 dev->reg_state = NETREG_RELEASED;
9012 /* will free via device release */
9013 put_device(&dev->dev);
9015 EXPORT_SYMBOL(free_netdev);
9018 * synchronize_net - Synchronize with packet receive processing
9020 * Wait for packets currently being received to be done.
9021 * Does not block later packets from starting.
9023 void synchronize_net(void)
9026 if (rtnl_is_locked())
9027 synchronize_rcu_expedited();
9031 EXPORT_SYMBOL(synchronize_net);
9034 * unregister_netdevice_queue - remove device from the kernel
9038 * This function shuts down a device interface and removes it
9039 * from the kernel tables.
9040 * If head not NULL, device is queued to be unregistered later.
9042 * Callers must hold the rtnl semaphore. You may want
9043 * unregister_netdev() instead of this.
9046 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head)
9051 list_move_tail(&dev->unreg_list, head);
9053 rollback_registered(dev);
9054 /* Finish processing unregister after unlock */
9058 EXPORT_SYMBOL(unregister_netdevice_queue);
9061 * unregister_netdevice_many - unregister many devices
9062 * @head: list of devices
9064 * Note: As most callers use a stack allocated list_head,
9065 * we force a list_del() to make sure stack wont be corrupted later.
9067 void unregister_netdevice_many(struct list_head *head)
9069 struct net_device *dev;
9071 if (!list_empty(head)) {
9072 rollback_registered_many(head);
9073 list_for_each_entry(dev, head, unreg_list)
9078 EXPORT_SYMBOL(unregister_netdevice_many);
9081 * unregister_netdev - remove device from the kernel
9084 * This function shuts down a device interface and removes it
9085 * from the kernel tables.
9087 * This is just a wrapper for unregister_netdevice that takes
9088 * the rtnl semaphore. In general you want to use this and not
9089 * unregister_netdevice.
9091 void unregister_netdev(struct net_device *dev)
9094 unregister_netdevice(dev);
9097 EXPORT_SYMBOL(unregister_netdev);
9100 * dev_change_net_namespace - move device to different nethost namespace
9102 * @net: network namespace
9103 * @pat: If not NULL name pattern to try if the current device name
9104 * is already taken in the destination network namespace.
9106 * This function shuts down a device interface and moves it
9107 * to a new network namespace. On success 0 is returned, on
9108 * a failure a netagive errno code is returned.
9110 * Callers must hold the rtnl semaphore.
9113 int dev_change_net_namespace(struct net_device *dev, struct net *net, const char *pat)
9115 int err, new_nsid, new_ifindex;
9119 /* Don't allow namespace local devices to be moved. */
9121 if (dev->features & NETIF_F_NETNS_LOCAL)
9124 /* Ensure the device has been registrered */
9125 if (dev->reg_state != NETREG_REGISTERED)
9128 /* Get out if there is nothing todo */
9130 if (net_eq(dev_net(dev), net))
9133 /* Pick the destination device name, and ensure
9134 * we can use it in the destination network namespace.
9137 if (__dev_get_by_name(net, dev->name)) {
9138 /* We get here if we can't use the current device name */
9141 err = dev_get_valid_name(net, dev, pat);
9147 * And now a mini version of register_netdevice unregister_netdevice.
9150 /* If device is running close it first. */
9153 /* And unlink it from device chain */
9154 unlist_netdevice(dev);
9158 /* Shutdown queueing discipline. */
9161 /* Notify protocols, that we are about to destroy
9162 * this device. They should clean all the things.
9164 * Note that dev->reg_state stays at NETREG_REGISTERED.
9165 * This is wanted because this way 8021q and macvlan know
9166 * the device is just moving and can keep their slaves up.
9168 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
9171 new_nsid = peernet2id_alloc(dev_net(dev), net);
9172 /* If there is an ifindex conflict assign a new one */
9173 if (__dev_get_by_index(net, dev->ifindex))
9174 new_ifindex = dev_new_index(net);
9176 new_ifindex = dev->ifindex;
9178 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid,
9182 * Flush the unicast and multicast chains
9187 /* Send a netdev-removed uevent to the old namespace */
9188 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE);
9189 netdev_adjacent_del_links(dev);
9191 /* Actually switch the network namespace */
9192 dev_net_set(dev, net);
9193 dev->ifindex = new_ifindex;
9195 /* Send a netdev-add uevent to the new namespace */
9196 kobject_uevent(&dev->dev.kobj, KOBJ_ADD);
9197 netdev_adjacent_add_links(dev);
9199 /* Fixup kobjects */
9200 err = device_rename(&dev->dev, dev->name);
9203 /* Add the device back in the hashes */
9204 list_netdevice(dev);
9206 /* Notify protocols, that a new device appeared. */
9207 call_netdevice_notifiers(NETDEV_REGISTER, dev);
9210 * Prevent userspace races by waiting until the network
9211 * device is fully setup before sending notifications.
9213 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
9220 EXPORT_SYMBOL_GPL(dev_change_net_namespace);
9222 static int dev_cpu_dead(unsigned int oldcpu)
9224 struct sk_buff **list_skb;
9225 struct sk_buff *skb;
9227 struct softnet_data *sd, *oldsd, *remsd = NULL;
9229 local_irq_disable();
9230 cpu = smp_processor_id();
9231 sd = &per_cpu(softnet_data, cpu);
9232 oldsd = &per_cpu(softnet_data, oldcpu);
9234 /* Find end of our completion_queue. */
9235 list_skb = &sd->completion_queue;
9237 list_skb = &(*list_skb)->next;
9238 /* Append completion queue from offline CPU. */
9239 *list_skb = oldsd->completion_queue;
9240 oldsd->completion_queue = NULL;
9242 /* Append output queue from offline CPU. */
9243 if (oldsd->output_queue) {
9244 *sd->output_queue_tailp = oldsd->output_queue;
9245 sd->output_queue_tailp = oldsd->output_queue_tailp;
9246 oldsd->output_queue = NULL;
9247 oldsd->output_queue_tailp = &oldsd->output_queue;
9249 /* Append NAPI poll list from offline CPU, with one exception :
9250 * process_backlog() must be called by cpu owning percpu backlog.
9251 * We properly handle process_queue & input_pkt_queue later.
9253 while (!list_empty(&oldsd->poll_list)) {
9254 struct napi_struct *napi = list_first_entry(&oldsd->poll_list,
9258 list_del_init(&napi->poll_list);
9259 if (napi->poll == process_backlog)
9262 ____napi_schedule(sd, napi);
9265 raise_softirq_irqoff(NET_TX_SOFTIRQ);
9269 remsd = oldsd->rps_ipi_list;
9270 oldsd->rps_ipi_list = NULL;
9272 /* send out pending IPI's on offline CPU */
9273 net_rps_send_ipi(remsd);
9275 /* Process offline CPU's input_pkt_queue */
9276 while ((skb = __skb_dequeue(&oldsd->process_queue))) {
9278 input_queue_head_incr(oldsd);
9280 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) {
9282 input_queue_head_incr(oldsd);
9289 * netdev_increment_features - increment feature set by one
9290 * @all: current feature set
9291 * @one: new feature set
9292 * @mask: mask feature set
9294 * Computes a new feature set after adding a device with feature set
9295 * @one to the master device with current feature set @all. Will not
9296 * enable anything that is off in @mask. Returns the new feature set.
9298 netdev_features_t netdev_increment_features(netdev_features_t all,
9299 netdev_features_t one, netdev_features_t mask)
9301 if (mask & NETIF_F_HW_CSUM)
9302 mask |= NETIF_F_CSUM_MASK;
9303 mask |= NETIF_F_VLAN_CHALLENGED;
9305 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask;
9306 all &= one | ~NETIF_F_ALL_FOR_ALL;
9308 /* If one device supports hw checksumming, set for all. */
9309 if (all & NETIF_F_HW_CSUM)
9310 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM);
9314 EXPORT_SYMBOL(netdev_increment_features);
9316 static struct hlist_head * __net_init netdev_create_hash(void)
9319 struct hlist_head *hash;
9321 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL);
9323 for (i = 0; i < NETDEV_HASHENTRIES; i++)
9324 INIT_HLIST_HEAD(&hash[i]);
9329 /* Initialize per network namespace state */
9330 static int __net_init netdev_init(struct net *net)
9332 BUILD_BUG_ON(GRO_HASH_BUCKETS >
9333 8 * FIELD_SIZEOF(struct napi_struct, gro_bitmask));
9335 if (net != &init_net)
9336 INIT_LIST_HEAD(&net->dev_base_head);
9338 net->dev_name_head = netdev_create_hash();
9339 if (net->dev_name_head == NULL)
9342 net->dev_index_head = netdev_create_hash();
9343 if (net->dev_index_head == NULL)
9349 kfree(net->dev_name_head);
9355 * netdev_drivername - network driver for the device
9356 * @dev: network device
9358 * Determine network driver for device.
9360 const char *netdev_drivername(const struct net_device *dev)
9362 const struct device_driver *driver;
9363 const struct device *parent;
9364 const char *empty = "";
9366 parent = dev->dev.parent;
9370 driver = parent->driver;
9371 if (driver && driver->name)
9372 return driver->name;
9376 static void __netdev_printk(const char *level, const struct net_device *dev,
9377 struct va_format *vaf)
9379 if (dev && dev->dev.parent) {
9380 dev_printk_emit(level[1] - '0',
9383 dev_driver_string(dev->dev.parent),
9384 dev_name(dev->dev.parent),
9385 netdev_name(dev), netdev_reg_state(dev),
9388 printk("%s%s%s: %pV",
9389 level, netdev_name(dev), netdev_reg_state(dev), vaf);
9391 printk("%s(NULL net_device): %pV", level, vaf);
9395 void netdev_printk(const char *level, const struct net_device *dev,
9396 const char *format, ...)
9398 struct va_format vaf;
9401 va_start(args, format);
9406 __netdev_printk(level, dev, &vaf);
9410 EXPORT_SYMBOL(netdev_printk);
9412 #define define_netdev_printk_level(func, level) \
9413 void func(const struct net_device *dev, const char *fmt, ...) \
9415 struct va_format vaf; \
9418 va_start(args, fmt); \
9423 __netdev_printk(level, dev, &vaf); \
9427 EXPORT_SYMBOL(func);
9429 define_netdev_printk_level(netdev_emerg, KERN_EMERG);
9430 define_netdev_printk_level(netdev_alert, KERN_ALERT);
9431 define_netdev_printk_level(netdev_crit, KERN_CRIT);
9432 define_netdev_printk_level(netdev_err, KERN_ERR);
9433 define_netdev_printk_level(netdev_warn, KERN_WARNING);
9434 define_netdev_printk_level(netdev_notice, KERN_NOTICE);
9435 define_netdev_printk_level(netdev_info, KERN_INFO);
9437 static void __net_exit netdev_exit(struct net *net)
9439 kfree(net->dev_name_head);
9440 kfree(net->dev_index_head);
9441 if (net != &init_net)
9442 WARN_ON_ONCE(!list_empty(&net->dev_base_head));
9445 static struct pernet_operations __net_initdata netdev_net_ops = {
9446 .init = netdev_init,
9447 .exit = netdev_exit,
9450 static void __net_exit default_device_exit(struct net *net)
9452 struct net_device *dev, *aux;
9454 * Push all migratable network devices back to the
9455 * initial network namespace
9458 for_each_netdev_safe(net, dev, aux) {
9460 char fb_name[IFNAMSIZ];
9462 /* Ignore unmoveable devices (i.e. loopback) */
9463 if (dev->features & NETIF_F_NETNS_LOCAL)
9466 /* Leave virtual devices for the generic cleanup */
9467 if (dev->rtnl_link_ops)
9470 /* Push remaining network devices to init_net */
9471 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex);
9472 err = dev_change_net_namespace(dev, &init_net, fb_name);
9474 pr_emerg("%s: failed to move %s to init_net: %d\n",
9475 __func__, dev->name, err);
9482 static void __net_exit rtnl_lock_unregistering(struct list_head *net_list)
9484 /* Return with the rtnl_lock held when there are no network
9485 * devices unregistering in any network namespace in net_list.
9489 DEFINE_WAIT_FUNC(wait, woken_wake_function);
9491 add_wait_queue(&netdev_unregistering_wq, &wait);
9493 unregistering = false;
9495 list_for_each_entry(net, net_list, exit_list) {
9496 if (net->dev_unreg_count > 0) {
9497 unregistering = true;
9505 wait_woken(&wait, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
9507 remove_wait_queue(&netdev_unregistering_wq, &wait);
9510 static void __net_exit default_device_exit_batch(struct list_head *net_list)
9512 /* At exit all network devices most be removed from a network
9513 * namespace. Do this in the reverse order of registration.
9514 * Do this across as many network namespaces as possible to
9515 * improve batching efficiency.
9517 struct net_device *dev;
9519 LIST_HEAD(dev_kill_list);
9521 /* To prevent network device cleanup code from dereferencing
9522 * loopback devices or network devices that have been freed
9523 * wait here for all pending unregistrations to complete,
9524 * before unregistring the loopback device and allowing the
9525 * network namespace be freed.
9527 * The netdev todo list containing all network devices
9528 * unregistrations that happen in default_device_exit_batch
9529 * will run in the rtnl_unlock() at the end of
9530 * default_device_exit_batch.
9532 rtnl_lock_unregistering(net_list);
9533 list_for_each_entry(net, net_list, exit_list) {
9534 for_each_netdev_reverse(net, dev) {
9535 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink)
9536 dev->rtnl_link_ops->dellink(dev, &dev_kill_list);
9538 unregister_netdevice_queue(dev, &dev_kill_list);
9541 unregister_netdevice_many(&dev_kill_list);
9545 static struct pernet_operations __net_initdata default_device_ops = {
9546 .exit = default_device_exit,
9547 .exit_batch = default_device_exit_batch,
9551 * Initialize the DEV module. At boot time this walks the device list and
9552 * unhooks any devices that fail to initialise (normally hardware not
9553 * present) and leaves us with a valid list of present and active devices.
9558 * This is called single threaded during boot, so no need
9559 * to take the rtnl semaphore.
9561 static int __init net_dev_init(void)
9563 int i, rc = -ENOMEM;
9565 BUG_ON(!dev_boot_phase);
9567 if (dev_proc_init())
9570 if (netdev_kobject_init())
9573 INIT_LIST_HEAD(&ptype_all);
9574 for (i = 0; i < PTYPE_HASH_SIZE; i++)
9575 INIT_LIST_HEAD(&ptype_base[i]);
9577 INIT_LIST_HEAD(&offload_base);
9579 if (register_pernet_subsys(&netdev_net_ops))
9583 * Initialise the packet receive queues.
9586 for_each_possible_cpu(i) {
9587 struct work_struct *flush = per_cpu_ptr(&flush_works, i);
9588 struct softnet_data *sd = &per_cpu(softnet_data, i);
9590 INIT_WORK(flush, flush_backlog);
9592 skb_queue_head_init(&sd->input_pkt_queue);
9593 skb_queue_head_init(&sd->process_queue);
9594 #ifdef CONFIG_XFRM_OFFLOAD
9595 skb_queue_head_init(&sd->xfrm_backlog);
9597 INIT_LIST_HEAD(&sd->poll_list);
9598 sd->output_queue_tailp = &sd->output_queue;
9600 sd->csd.func = rps_trigger_softirq;
9605 init_gro_hash(&sd->backlog);
9606 sd->backlog.poll = process_backlog;
9607 sd->backlog.weight = weight_p;
9612 /* The loopback device is special if any other network devices
9613 * is present in a network namespace the loopback device must
9614 * be present. Since we now dynamically allocate and free the
9615 * loopback device ensure this invariant is maintained by
9616 * keeping the loopback device as the first device on the
9617 * list of network devices. Ensuring the loopback devices
9618 * is the first device that appears and the last network device
9621 if (register_pernet_device(&loopback_net_ops))
9624 if (register_pernet_device(&default_device_ops))
9627 open_softirq(NET_TX_SOFTIRQ, net_tx_action);
9628 open_softirq(NET_RX_SOFTIRQ, net_rx_action);
9630 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead",
9631 NULL, dev_cpu_dead);
9638 subsys_initcall(net_dev_init);