2 * NET3 Protocol independent device support routines.
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License
6 * as published by the Free Software Foundation; either version
7 * 2 of the License, or (at your option) any later version.
9 * Derived from the non IP parts of dev.c 1.0.19
11 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
12 * Mark Evans, <evansmp@uhura.aston.ac.uk>
15 * Florian la Roche <rzsfl@rz.uni-sb.de>
16 * Alan Cox <gw4pts@gw4pts.ampr.org>
17 * David Hinds <dahinds@users.sourceforge.net>
18 * Alexey Kuznetsov <kuznet@ms2.inr.ac.ru>
19 * Adam Sulmicki <adam@cfar.umd.edu>
20 * Pekka Riikonen <priikone@poesidon.pspt.fi>
23 * D.J. Barrow : Fixed bug where dev->refcnt gets set
24 * to 2 if register_netdev gets called
25 * before net_dev_init & also removed a
26 * few lines of code in the process.
27 * Alan Cox : device private ioctl copies fields back.
28 * Alan Cox : Transmit queue code does relevant
29 * stunts to keep the queue safe.
30 * Alan Cox : Fixed double lock.
31 * Alan Cox : Fixed promisc NULL pointer trap
32 * ???????? : Support the full private ioctl range
33 * Alan Cox : Moved ioctl permission check into
35 * Tim Kordas : SIOCADDMULTI/SIOCDELMULTI
36 * Alan Cox : 100 backlog just doesn't cut it when
37 * you start doing multicast video 8)
38 * Alan Cox : Rewrote net_bh and list manager.
39 * Alan Cox : Fix ETH_P_ALL echoback lengths.
40 * Alan Cox : Took out transmit every packet pass
41 * Saved a few bytes in the ioctl handler
42 * Alan Cox : Network driver sets packet type before
43 * calling netif_rx. Saves a function
45 * Alan Cox : Hashed net_bh()
46 * Richard Kooijman: Timestamp fixes.
47 * Alan Cox : Wrong field in SIOCGIFDSTADDR
48 * Alan Cox : Device lock protection.
49 * Alan Cox : Fixed nasty side effect of device close
51 * Rudi Cilibrasi : Pass the right thing to
53 * Dave Miller : 32bit quantity for the device lock to
54 * make it work out on a Sparc.
55 * Bjorn Ekwall : Added KERNELD hack.
56 * Alan Cox : Cleaned up the backlog initialise.
57 * Craig Metz : SIOCGIFCONF fix if space for under
59 * Thomas Bogendoerfer : Return ENODEV for dev_open, if there
60 * is no device open function.
61 * Andi Kleen : Fix error reporting for SIOCGIFCONF
62 * Michael Chastain : Fix signed/unsigned for SIOCGIFCONF
63 * Cyrus Durgin : Cleaned for KMOD
64 * Adam Sulmicki : Bug Fix : Network Device Unload
65 * A network device unload needs to purge
67 * Paul Rusty Russell : SIOCSIFNAME
68 * Pekka Riikonen : Netdev boot-time settings code
69 * Andrew Morton : Make unregister_netdevice wait
70 * indefinitely on dev->refcnt
71 * J Hadi Salim : - Backlog queue sampling
72 * - netif_rx() feedback
75 #include <linux/uaccess.h>
76 #include <linux/bitops.h>
77 #include <linux/capability.h>
78 #include <linux/cpu.h>
79 #include <linux/types.h>
80 #include <linux/kernel.h>
81 #include <linux/hash.h>
82 #include <linux/slab.h>
83 #include <linux/sched.h>
84 #include <linux/sched/mm.h>
85 #include <linux/mutex.h>
86 #include <linux/string.h>
88 #include <linux/socket.h>
89 #include <linux/sockios.h>
90 #include <linux/errno.h>
91 #include <linux/interrupt.h>
92 #include <linux/if_ether.h>
93 #include <linux/netdevice.h>
94 #include <linux/etherdevice.h>
95 #include <linux/ethtool.h>
96 #include <linux/notifier.h>
97 #include <linux/skbuff.h>
98 #include <linux/bpf.h>
99 #include <linux/bpf_trace.h>
100 #include <net/net_namespace.h>
101 #include <net/sock.h>
102 #include <net/busy_poll.h>
103 #include <linux/rtnetlink.h>
104 #include <linux/stat.h>
106 #include <net/dst_metadata.h>
107 #include <net/pkt_sched.h>
108 #include <net/pkt_cls.h>
109 #include <net/checksum.h>
110 #include <net/xfrm.h>
111 #include <linux/highmem.h>
112 #include <linux/init.h>
113 #include <linux/module.h>
114 #include <linux/netpoll.h>
115 #include <linux/rcupdate.h>
116 #include <linux/delay.h>
117 #include <net/iw_handler.h>
118 #include <asm/current.h>
119 #include <linux/audit.h>
120 #include <linux/dmaengine.h>
121 #include <linux/err.h>
122 #include <linux/ctype.h>
123 #include <linux/if_arp.h>
124 #include <linux/if_vlan.h>
125 #include <linux/ip.h>
127 #include <net/mpls.h>
128 #include <linux/ipv6.h>
129 #include <linux/in.h>
130 #include <linux/jhash.h>
131 #include <linux/random.h>
132 #include <trace/events/napi.h>
133 #include <trace/events/net.h>
134 #include <trace/events/skb.h>
135 #include <linux/pci.h>
136 #include <linux/inetdevice.h>
137 #include <linux/cpu_rmap.h>
138 #include <linux/static_key.h>
139 #include <linux/hashtable.h>
140 #include <linux/vmalloc.h>
141 #include <linux/if_macvlan.h>
142 #include <linux/errqueue.h>
143 #include <linux/hrtimer.h>
144 #include <linux/netfilter_ingress.h>
145 #include <linux/crash_dump.h>
146 #include <linux/sctp.h>
147 #include <net/udp_tunnel.h>
148 #include <linux/net_namespace.h>
150 #include "net-sysfs.h"
152 #define MAX_GRO_SKBS 8
154 /* This should be increased if a protocol with a bigger head is added. */
155 #define GRO_MAX_HEAD (MAX_HEADER + 128)
157 static DEFINE_SPINLOCK(ptype_lock);
158 static DEFINE_SPINLOCK(offload_lock);
159 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly;
160 struct list_head ptype_all __read_mostly; /* Taps */
161 static struct list_head offload_base __read_mostly;
163 static int netif_rx_internal(struct sk_buff *skb);
164 static int call_netdevice_notifiers_info(unsigned long val,
165 struct netdev_notifier_info *info);
166 static struct napi_struct *napi_by_id(unsigned int napi_id);
169 * The @dev_base_head list is protected by @dev_base_lock and the rtnl
172 * Pure readers hold dev_base_lock for reading, or rcu_read_lock()
174 * Writers must hold the rtnl semaphore while they loop through the
175 * dev_base_head list, and hold dev_base_lock for writing when they do the
176 * actual updates. This allows pure readers to access the list even
177 * while a writer is preparing to update it.
179 * To put it another way, dev_base_lock is held for writing only to
180 * protect against pure readers; the rtnl semaphore provides the
181 * protection against other writers.
183 * See, for example usages, register_netdevice() and
184 * unregister_netdevice(), which must be called with the rtnl
187 DEFINE_RWLOCK(dev_base_lock);
188 EXPORT_SYMBOL(dev_base_lock);
190 static DEFINE_MUTEX(ifalias_mutex);
192 /* protects napi_hash addition/deletion and napi_gen_id */
193 static DEFINE_SPINLOCK(napi_hash_lock);
195 static unsigned int napi_gen_id = NR_CPUS;
196 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8);
198 static seqcount_t devnet_rename_seq;
200 static inline void dev_base_seq_inc(struct net *net)
202 while (++net->dev_base_seq == 0)
206 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name)
208 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ));
210 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)];
213 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex)
215 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)];
218 static inline void rps_lock(struct softnet_data *sd)
221 spin_lock(&sd->input_pkt_queue.lock);
225 static inline void rps_unlock(struct softnet_data *sd)
228 spin_unlock(&sd->input_pkt_queue.lock);
232 /* Device list insertion */
233 static void list_netdevice(struct net_device *dev)
235 struct net *net = dev_net(dev);
239 write_lock_bh(&dev_base_lock);
240 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head);
241 hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
242 hlist_add_head_rcu(&dev->index_hlist,
243 dev_index_hash(net, dev->ifindex));
244 write_unlock_bh(&dev_base_lock);
246 dev_base_seq_inc(net);
249 /* Device list removal
250 * caller must respect a RCU grace period before freeing/reusing dev
252 static void unlist_netdevice(struct net_device *dev)
256 /* Unlink dev from the device chain */
257 write_lock_bh(&dev_base_lock);
258 list_del_rcu(&dev->dev_list);
259 hlist_del_rcu(&dev->name_hlist);
260 hlist_del_rcu(&dev->index_hlist);
261 write_unlock_bh(&dev_base_lock);
263 dev_base_seq_inc(dev_net(dev));
270 static RAW_NOTIFIER_HEAD(netdev_chain);
273 * Device drivers call our routines to queue packets here. We empty the
274 * queue in the local softnet handler.
277 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data);
278 EXPORT_PER_CPU_SYMBOL(softnet_data);
280 #ifdef CONFIG_LOCKDEP
282 * register_netdevice() inits txq->_xmit_lock and sets lockdep class
283 * according to dev->type
285 static const unsigned short netdev_lock_type[] = {
286 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25,
287 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET,
288 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM,
289 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP,
290 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD,
291 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25,
292 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP,
293 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD,
294 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI,
295 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE,
296 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET,
297 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL,
298 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM,
299 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE,
300 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE};
302 static const char *const netdev_lock_name[] = {
303 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25",
304 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET",
305 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM",
306 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP",
307 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD",
308 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25",
309 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP",
310 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD",
311 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI",
312 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE",
313 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET",
314 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL",
315 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM",
316 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE",
317 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"};
319 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)];
320 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)];
322 static inline unsigned short netdev_lock_pos(unsigned short dev_type)
326 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++)
327 if (netdev_lock_type[i] == dev_type)
329 /* the last key is used by default */
330 return ARRAY_SIZE(netdev_lock_type) - 1;
333 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
334 unsigned short dev_type)
338 i = netdev_lock_pos(dev_type);
339 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i],
340 netdev_lock_name[i]);
343 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
347 i = netdev_lock_pos(dev->type);
348 lockdep_set_class_and_name(&dev->addr_list_lock,
349 &netdev_addr_lock_key[i],
350 netdev_lock_name[i]);
353 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
354 unsigned short dev_type)
357 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
362 /*******************************************************************************
364 * Protocol management and registration routines
366 *******************************************************************************/
370 * Add a protocol ID to the list. Now that the input handler is
371 * smarter we can dispense with all the messy stuff that used to be
374 * BEWARE!!! Protocol handlers, mangling input packets,
375 * MUST BE last in hash buckets and checking protocol handlers
376 * MUST start from promiscuous ptype_all chain in net_bh.
377 * It is true now, do not change it.
378 * Explanation follows: if protocol handler, mangling packet, will
379 * be the first on list, it is not able to sense, that packet
380 * is cloned and should be copied-on-write, so that it will
381 * change it and subsequent readers will get broken packet.
385 static inline struct list_head *ptype_head(const struct packet_type *pt)
387 if (pt->type == htons(ETH_P_ALL))
388 return pt->dev ? &pt->dev->ptype_all : &ptype_all;
390 return pt->dev ? &pt->dev->ptype_specific :
391 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK];
395 * dev_add_pack - add packet handler
396 * @pt: packet type declaration
398 * Add a protocol handler to the networking stack. The passed &packet_type
399 * is linked into kernel lists and may not be freed until it has been
400 * removed from the kernel lists.
402 * This call does not sleep therefore it can not
403 * guarantee all CPU's that are in middle of receiving packets
404 * will see the new packet type (until the next received packet).
407 void dev_add_pack(struct packet_type *pt)
409 struct list_head *head = ptype_head(pt);
411 spin_lock(&ptype_lock);
412 list_add_rcu(&pt->list, head);
413 spin_unlock(&ptype_lock);
415 EXPORT_SYMBOL(dev_add_pack);
418 * __dev_remove_pack - remove packet handler
419 * @pt: packet type declaration
421 * Remove a protocol handler that was previously added to the kernel
422 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
423 * from the kernel lists and can be freed or reused once this function
426 * The packet type might still be in use by receivers
427 * and must not be freed until after all the CPU's have gone
428 * through a quiescent state.
430 void __dev_remove_pack(struct packet_type *pt)
432 struct list_head *head = ptype_head(pt);
433 struct packet_type *pt1;
435 spin_lock(&ptype_lock);
437 list_for_each_entry(pt1, head, list) {
439 list_del_rcu(&pt->list);
444 pr_warn("dev_remove_pack: %p not found\n", pt);
446 spin_unlock(&ptype_lock);
448 EXPORT_SYMBOL(__dev_remove_pack);
451 * dev_remove_pack - remove packet handler
452 * @pt: packet type declaration
454 * Remove a protocol handler that was previously added to the kernel
455 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
456 * from the kernel lists and can be freed or reused once this function
459 * This call sleeps to guarantee that no CPU is looking at the packet
462 void dev_remove_pack(struct packet_type *pt)
464 __dev_remove_pack(pt);
468 EXPORT_SYMBOL(dev_remove_pack);
472 * dev_add_offload - register offload handlers
473 * @po: protocol offload declaration
475 * Add protocol offload handlers to the networking stack. The passed
476 * &proto_offload is linked into kernel lists and may not be freed until
477 * it has been removed from the kernel lists.
479 * This call does not sleep therefore it can not
480 * guarantee all CPU's that are in middle of receiving packets
481 * will see the new offload handlers (until the next received packet).
483 void dev_add_offload(struct packet_offload *po)
485 struct packet_offload *elem;
487 spin_lock(&offload_lock);
488 list_for_each_entry(elem, &offload_base, list) {
489 if (po->priority < elem->priority)
492 list_add_rcu(&po->list, elem->list.prev);
493 spin_unlock(&offload_lock);
495 EXPORT_SYMBOL(dev_add_offload);
498 * __dev_remove_offload - remove offload handler
499 * @po: packet offload declaration
501 * Remove a protocol offload handler that was previously added to the
502 * kernel offload handlers by dev_add_offload(). The passed &offload_type
503 * is removed from the kernel lists and can be freed or reused once this
506 * The packet type might still be in use by receivers
507 * and must not be freed until after all the CPU's have gone
508 * through a quiescent state.
510 static void __dev_remove_offload(struct packet_offload *po)
512 struct list_head *head = &offload_base;
513 struct packet_offload *po1;
515 spin_lock(&offload_lock);
517 list_for_each_entry(po1, head, list) {
519 list_del_rcu(&po->list);
524 pr_warn("dev_remove_offload: %p not found\n", po);
526 spin_unlock(&offload_lock);
530 * dev_remove_offload - remove packet offload handler
531 * @po: packet offload declaration
533 * Remove a packet offload handler that was previously added to the kernel
534 * offload handlers by dev_add_offload(). The passed &offload_type is
535 * removed from the kernel lists and can be freed or reused once this
538 * This call sleeps to guarantee that no CPU is looking at the packet
541 void dev_remove_offload(struct packet_offload *po)
543 __dev_remove_offload(po);
547 EXPORT_SYMBOL(dev_remove_offload);
549 /******************************************************************************
551 * Device Boot-time Settings Routines
553 ******************************************************************************/
555 /* Boot time configuration table */
556 static struct netdev_boot_setup dev_boot_setup[NETDEV_BOOT_SETUP_MAX];
559 * netdev_boot_setup_add - add new setup entry
560 * @name: name of the device
561 * @map: configured settings for the device
563 * Adds new setup entry to the dev_boot_setup list. The function
564 * returns 0 on error and 1 on success. This is a generic routine to
567 static int netdev_boot_setup_add(char *name, struct ifmap *map)
569 struct netdev_boot_setup *s;
573 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
574 if (s[i].name[0] == '\0' || s[i].name[0] == ' ') {
575 memset(s[i].name, 0, sizeof(s[i].name));
576 strlcpy(s[i].name, name, IFNAMSIZ);
577 memcpy(&s[i].map, map, sizeof(s[i].map));
582 return i >= NETDEV_BOOT_SETUP_MAX ? 0 : 1;
586 * netdev_boot_setup_check - check boot time settings
587 * @dev: the netdevice
589 * Check boot time settings for the device.
590 * The found settings are set for the device to be used
591 * later in the device probing.
592 * Returns 0 if no settings found, 1 if they are.
594 int netdev_boot_setup_check(struct net_device *dev)
596 struct netdev_boot_setup *s = dev_boot_setup;
599 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
600 if (s[i].name[0] != '\0' && s[i].name[0] != ' ' &&
601 !strcmp(dev->name, s[i].name)) {
602 dev->irq = s[i].map.irq;
603 dev->base_addr = s[i].map.base_addr;
604 dev->mem_start = s[i].map.mem_start;
605 dev->mem_end = s[i].map.mem_end;
611 EXPORT_SYMBOL(netdev_boot_setup_check);
615 * netdev_boot_base - get address from boot time settings
616 * @prefix: prefix for network device
617 * @unit: id for network device
619 * Check boot time settings for the base address of device.
620 * The found settings are set for the device to be used
621 * later in the device probing.
622 * Returns 0 if no settings found.
624 unsigned long netdev_boot_base(const char *prefix, int unit)
626 const struct netdev_boot_setup *s = dev_boot_setup;
630 sprintf(name, "%s%d", prefix, unit);
633 * If device already registered then return base of 1
634 * to indicate not to probe for this interface
636 if (__dev_get_by_name(&init_net, name))
639 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++)
640 if (!strcmp(name, s[i].name))
641 return s[i].map.base_addr;
646 * Saves at boot time configured settings for any netdevice.
648 int __init netdev_boot_setup(char *str)
653 str = get_options(str, ARRAY_SIZE(ints), ints);
658 memset(&map, 0, sizeof(map));
662 map.base_addr = ints[2];
664 map.mem_start = ints[3];
666 map.mem_end = ints[4];
668 /* Add new entry to the list */
669 return netdev_boot_setup_add(str, &map);
672 __setup("netdev=", netdev_boot_setup);
674 /*******************************************************************************
676 * Device Interface Subroutines
678 *******************************************************************************/
681 * dev_get_iflink - get 'iflink' value of a interface
682 * @dev: targeted interface
684 * Indicates the ifindex the interface is linked to.
685 * Physical interfaces have the same 'ifindex' and 'iflink' values.
688 int dev_get_iflink(const struct net_device *dev)
690 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink)
691 return dev->netdev_ops->ndo_get_iflink(dev);
695 EXPORT_SYMBOL(dev_get_iflink);
698 * dev_fill_metadata_dst - Retrieve tunnel egress information.
699 * @dev: targeted interface
702 * For better visibility of tunnel traffic OVS needs to retrieve
703 * egress tunnel information for a packet. Following API allows
704 * user to get this info.
706 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb)
708 struct ip_tunnel_info *info;
710 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst)
713 info = skb_tunnel_info_unclone(skb);
716 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX)))
719 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb);
721 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst);
724 * __dev_get_by_name - find a device by its name
725 * @net: the applicable net namespace
726 * @name: name to find
728 * Find an interface by name. Must be called under RTNL semaphore
729 * or @dev_base_lock. If the name is found a pointer to the device
730 * is returned. If the name is not found then %NULL is returned. The
731 * reference counters are not incremented so the caller must be
732 * careful with locks.
735 struct net_device *__dev_get_by_name(struct net *net, const char *name)
737 struct net_device *dev;
738 struct hlist_head *head = dev_name_hash(net, name);
740 hlist_for_each_entry(dev, head, name_hlist)
741 if (!strncmp(dev->name, name, IFNAMSIZ))
746 EXPORT_SYMBOL(__dev_get_by_name);
749 * dev_get_by_name_rcu - find a device by its name
750 * @net: the applicable net namespace
751 * @name: name to find
753 * Find an interface by name.
754 * If the name is found a pointer to the device is returned.
755 * If the name is not found then %NULL is returned.
756 * The reference counters are not incremented so the caller must be
757 * careful with locks. The caller must hold RCU lock.
760 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name)
762 struct net_device *dev;
763 struct hlist_head *head = dev_name_hash(net, name);
765 hlist_for_each_entry_rcu(dev, head, name_hlist)
766 if (!strncmp(dev->name, name, IFNAMSIZ))
771 EXPORT_SYMBOL(dev_get_by_name_rcu);
774 * dev_get_by_name - find a device by its name
775 * @net: the applicable net namespace
776 * @name: name to find
778 * Find an interface by name. This can be called from any
779 * context and does its own locking. The returned handle has
780 * the usage count incremented and the caller must use dev_put() to
781 * release it when it is no longer needed. %NULL is returned if no
782 * matching device is found.
785 struct net_device *dev_get_by_name(struct net *net, const char *name)
787 struct net_device *dev;
790 dev = dev_get_by_name_rcu(net, name);
796 EXPORT_SYMBOL(dev_get_by_name);
799 * __dev_get_by_index - find a device by its ifindex
800 * @net: the applicable net namespace
801 * @ifindex: index of device
803 * Search for an interface by index. Returns %NULL if the device
804 * is not found or a pointer to the device. The device has not
805 * had its reference counter increased so the caller must be careful
806 * about locking. The caller must hold either the RTNL semaphore
810 struct net_device *__dev_get_by_index(struct net *net, int ifindex)
812 struct net_device *dev;
813 struct hlist_head *head = dev_index_hash(net, ifindex);
815 hlist_for_each_entry(dev, head, index_hlist)
816 if (dev->ifindex == ifindex)
821 EXPORT_SYMBOL(__dev_get_by_index);
824 * dev_get_by_index_rcu - find a device by its ifindex
825 * @net: the applicable net namespace
826 * @ifindex: index of device
828 * Search for an interface by index. Returns %NULL if the device
829 * is not found or a pointer to the device. The device has not
830 * had its reference counter increased so the caller must be careful
831 * about locking. The caller must hold RCU lock.
834 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex)
836 struct net_device *dev;
837 struct hlist_head *head = dev_index_hash(net, ifindex);
839 hlist_for_each_entry_rcu(dev, head, index_hlist)
840 if (dev->ifindex == ifindex)
845 EXPORT_SYMBOL(dev_get_by_index_rcu);
849 * dev_get_by_index - find a device by its ifindex
850 * @net: the applicable net namespace
851 * @ifindex: index of device
853 * Search for an interface by index. Returns NULL if the device
854 * is not found or a pointer to the device. The device returned has
855 * had a reference added and the pointer is safe until the user calls
856 * dev_put to indicate they have finished with it.
859 struct net_device *dev_get_by_index(struct net *net, int ifindex)
861 struct net_device *dev;
864 dev = dev_get_by_index_rcu(net, ifindex);
870 EXPORT_SYMBOL(dev_get_by_index);
873 * dev_get_by_napi_id - find a device by napi_id
874 * @napi_id: ID of the NAPI struct
876 * Search for an interface by NAPI ID. Returns %NULL if the device
877 * is not found or a pointer to the device. The device has not had
878 * its reference counter increased so the caller must be careful
879 * about locking. The caller must hold RCU lock.
882 struct net_device *dev_get_by_napi_id(unsigned int napi_id)
884 struct napi_struct *napi;
886 WARN_ON_ONCE(!rcu_read_lock_held());
888 if (napi_id < MIN_NAPI_ID)
891 napi = napi_by_id(napi_id);
893 return napi ? napi->dev : NULL;
895 EXPORT_SYMBOL(dev_get_by_napi_id);
898 * netdev_get_name - get a netdevice name, knowing its ifindex.
899 * @net: network namespace
900 * @name: a pointer to the buffer where the name will be stored.
901 * @ifindex: the ifindex of the interface to get the name from.
903 * The use of raw_seqcount_begin() and cond_resched() before
904 * retrying is required as we want to give the writers a chance
905 * to complete when CONFIG_PREEMPT is not set.
907 int netdev_get_name(struct net *net, char *name, int ifindex)
909 struct net_device *dev;
913 seq = raw_seqcount_begin(&devnet_rename_seq);
915 dev = dev_get_by_index_rcu(net, ifindex);
921 strcpy(name, dev->name);
923 if (read_seqcount_retry(&devnet_rename_seq, seq)) {
932 * dev_getbyhwaddr_rcu - find a device by its hardware address
933 * @net: the applicable net namespace
934 * @type: media type of device
935 * @ha: hardware address
937 * Search for an interface by MAC address. Returns NULL if the device
938 * is not found or a pointer to the device.
939 * The caller must hold RCU or RTNL.
940 * The returned device has not had its ref count increased
941 * and the caller must therefore be careful about locking
945 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type,
948 struct net_device *dev;
950 for_each_netdev_rcu(net, dev)
951 if (dev->type == type &&
952 !memcmp(dev->dev_addr, ha, dev->addr_len))
957 EXPORT_SYMBOL(dev_getbyhwaddr_rcu);
959 struct net_device *__dev_getfirstbyhwtype(struct net *net, unsigned short type)
961 struct net_device *dev;
964 for_each_netdev(net, dev)
965 if (dev->type == type)
970 EXPORT_SYMBOL(__dev_getfirstbyhwtype);
972 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type)
974 struct net_device *dev, *ret = NULL;
977 for_each_netdev_rcu(net, dev)
978 if (dev->type == type) {
986 EXPORT_SYMBOL(dev_getfirstbyhwtype);
989 * __dev_get_by_flags - find any device with given flags
990 * @net: the applicable net namespace
991 * @if_flags: IFF_* values
992 * @mask: bitmask of bits in if_flags to check
994 * Search for any interface with the given flags. Returns NULL if a device
995 * is not found or a pointer to the device. Must be called inside
996 * rtnl_lock(), and result refcount is unchanged.
999 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags,
1000 unsigned short mask)
1002 struct net_device *dev, *ret;
1007 for_each_netdev(net, dev) {
1008 if (((dev->flags ^ if_flags) & mask) == 0) {
1015 EXPORT_SYMBOL(__dev_get_by_flags);
1018 * dev_valid_name - check if name is okay for network device
1019 * @name: name string
1021 * Network device names need to be valid file names to
1022 * to allow sysfs to work. We also disallow any kind of
1025 bool dev_valid_name(const char *name)
1029 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ)
1031 if (!strcmp(name, ".") || !strcmp(name, ".."))
1035 if (*name == '/' || *name == ':' || isspace(*name))
1041 EXPORT_SYMBOL(dev_valid_name);
1044 * __dev_alloc_name - allocate a name for a device
1045 * @net: network namespace to allocate the device name in
1046 * @name: name format string
1047 * @buf: scratch buffer and result name string
1049 * Passed a format string - eg "lt%d" it will try and find a suitable
1050 * id. It scans list of devices to build up a free map, then chooses
1051 * the first empty slot. The caller must hold the dev_base or rtnl lock
1052 * while allocating the name and adding the device in order to avoid
1054 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1055 * Returns the number of the unit assigned or a negative errno code.
1058 static int __dev_alloc_name(struct net *net, const char *name, char *buf)
1062 const int max_netdevices = 8*PAGE_SIZE;
1063 unsigned long *inuse;
1064 struct net_device *d;
1066 if (!dev_valid_name(name))
1069 p = strchr(name, '%');
1072 * Verify the string as this thing may have come from
1073 * the user. There must be either one "%d" and no other "%"
1076 if (p[1] != 'd' || strchr(p + 2, '%'))
1079 /* Use one page as a bit array of possible slots */
1080 inuse = (unsigned long *) get_zeroed_page(GFP_ATOMIC);
1084 for_each_netdev(net, d) {
1085 if (!sscanf(d->name, name, &i))
1087 if (i < 0 || i >= max_netdevices)
1090 /* avoid cases where sscanf is not exact inverse of printf */
1091 snprintf(buf, IFNAMSIZ, name, i);
1092 if (!strncmp(buf, d->name, IFNAMSIZ))
1096 i = find_first_zero_bit(inuse, max_netdevices);
1097 free_page((unsigned long) inuse);
1100 snprintf(buf, IFNAMSIZ, name, i);
1101 if (!__dev_get_by_name(net, buf))
1104 /* It is possible to run out of possible slots
1105 * when the name is long and there isn't enough space left
1106 * for the digits, or if all bits are used.
1111 static int dev_alloc_name_ns(struct net *net,
1112 struct net_device *dev,
1119 ret = __dev_alloc_name(net, name, buf);
1121 strlcpy(dev->name, buf, IFNAMSIZ);
1126 * dev_alloc_name - allocate a name for a device
1128 * @name: name format string
1130 * Passed a format string - eg "lt%d" it will try and find a suitable
1131 * id. It scans list of devices to build up a free map, then chooses
1132 * the first empty slot. The caller must hold the dev_base or rtnl lock
1133 * while allocating the name and adding the device in order to avoid
1135 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1136 * Returns the number of the unit assigned or a negative errno code.
1139 int dev_alloc_name(struct net_device *dev, const char *name)
1141 return dev_alloc_name_ns(dev_net(dev), dev, name);
1143 EXPORT_SYMBOL(dev_alloc_name);
1145 int dev_get_valid_name(struct net *net, struct net_device *dev,
1150 if (!dev_valid_name(name))
1153 if (strchr(name, '%'))
1154 return dev_alloc_name_ns(net, dev, name);
1155 else if (__dev_get_by_name(net, name))
1157 else if (dev->name != name)
1158 strlcpy(dev->name, name, IFNAMSIZ);
1162 EXPORT_SYMBOL(dev_get_valid_name);
1165 * dev_change_name - change name of a device
1167 * @newname: name (or format string) must be at least IFNAMSIZ
1169 * Change name of a device, can pass format strings "eth%d".
1172 int dev_change_name(struct net_device *dev, const char *newname)
1174 unsigned char old_assign_type;
1175 char oldname[IFNAMSIZ];
1181 BUG_ON(!dev_net(dev));
1184 if (dev->flags & IFF_UP)
1187 write_seqcount_begin(&devnet_rename_seq);
1189 if (strncmp(newname, dev->name, IFNAMSIZ) == 0) {
1190 write_seqcount_end(&devnet_rename_seq);
1194 memcpy(oldname, dev->name, IFNAMSIZ);
1196 err = dev_get_valid_name(net, dev, newname);
1198 write_seqcount_end(&devnet_rename_seq);
1202 if (oldname[0] && !strchr(oldname, '%'))
1203 netdev_info(dev, "renamed from %s\n", oldname);
1205 old_assign_type = dev->name_assign_type;
1206 dev->name_assign_type = NET_NAME_RENAMED;
1209 ret = device_rename(&dev->dev, dev->name);
1211 memcpy(dev->name, oldname, IFNAMSIZ);
1212 dev->name_assign_type = old_assign_type;
1213 write_seqcount_end(&devnet_rename_seq);
1217 write_seqcount_end(&devnet_rename_seq);
1219 netdev_adjacent_rename_links(dev, oldname);
1221 write_lock_bh(&dev_base_lock);
1222 hlist_del_rcu(&dev->name_hlist);
1223 write_unlock_bh(&dev_base_lock);
1227 write_lock_bh(&dev_base_lock);
1228 hlist_add_head_rcu(&dev->name_hlist, dev_name_hash(net, dev->name));
1229 write_unlock_bh(&dev_base_lock);
1231 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev);
1232 ret = notifier_to_errno(ret);
1235 /* err >= 0 after dev_alloc_name() or stores the first errno */
1238 write_seqcount_begin(&devnet_rename_seq);
1239 memcpy(dev->name, oldname, IFNAMSIZ);
1240 memcpy(oldname, newname, IFNAMSIZ);
1241 dev->name_assign_type = old_assign_type;
1242 old_assign_type = NET_NAME_RENAMED;
1245 pr_err("%s: name change rollback failed: %d\n",
1254 * dev_set_alias - change ifalias of a device
1256 * @alias: name up to IFALIASZ
1257 * @len: limit of bytes to copy from info
1259 * Set ifalias for a device,
1261 int dev_set_alias(struct net_device *dev, const char *alias, size_t len)
1263 struct dev_ifalias *new_alias = NULL;
1265 if (len >= IFALIASZ)
1269 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL);
1273 memcpy(new_alias->ifalias, alias, len);
1274 new_alias->ifalias[len] = 0;
1277 mutex_lock(&ifalias_mutex);
1278 rcu_swap_protected(dev->ifalias, new_alias,
1279 mutex_is_locked(&ifalias_mutex));
1280 mutex_unlock(&ifalias_mutex);
1283 kfree_rcu(new_alias, rcuhead);
1287 EXPORT_SYMBOL(dev_set_alias);
1290 * dev_get_alias - get ifalias of a device
1292 * @name: buffer to store name of ifalias
1293 * @len: size of buffer
1295 * get ifalias for a device. Caller must make sure dev cannot go
1296 * away, e.g. rcu read lock or own a reference count to device.
1298 int dev_get_alias(const struct net_device *dev, char *name, size_t len)
1300 const struct dev_ifalias *alias;
1304 alias = rcu_dereference(dev->ifalias);
1306 ret = snprintf(name, len, "%s", alias->ifalias);
1313 * netdev_features_change - device changes features
1314 * @dev: device to cause notification
1316 * Called to indicate a device has changed features.
1318 void netdev_features_change(struct net_device *dev)
1320 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev);
1322 EXPORT_SYMBOL(netdev_features_change);
1325 * netdev_state_change - device changes state
1326 * @dev: device to cause notification
1328 * Called to indicate a device has changed state. This function calls
1329 * the notifier chains for netdev_chain and sends a NEWLINK message
1330 * to the routing socket.
1332 void netdev_state_change(struct net_device *dev)
1334 if (dev->flags & IFF_UP) {
1335 struct netdev_notifier_change_info change_info = {
1339 call_netdevice_notifiers_info(NETDEV_CHANGE,
1341 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL);
1344 EXPORT_SYMBOL(netdev_state_change);
1347 * netdev_notify_peers - notify network peers about existence of @dev
1348 * @dev: network device
1350 * Generate traffic such that interested network peers are aware of
1351 * @dev, such as by generating a gratuitous ARP. This may be used when
1352 * a device wants to inform the rest of the network about some sort of
1353 * reconfiguration such as a failover event or virtual machine
1356 void netdev_notify_peers(struct net_device *dev)
1359 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
1360 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev);
1363 EXPORT_SYMBOL(netdev_notify_peers);
1365 static int __dev_open(struct net_device *dev)
1367 const struct net_device_ops *ops = dev->netdev_ops;
1372 if (!netif_device_present(dev))
1375 /* Block netpoll from trying to do any rx path servicing.
1376 * If we don't do this there is a chance ndo_poll_controller
1377 * or ndo_poll may be running while we open the device
1379 netpoll_poll_disable(dev);
1381 ret = call_netdevice_notifiers(NETDEV_PRE_UP, dev);
1382 ret = notifier_to_errno(ret);
1386 set_bit(__LINK_STATE_START, &dev->state);
1388 if (ops->ndo_validate_addr)
1389 ret = ops->ndo_validate_addr(dev);
1391 if (!ret && ops->ndo_open)
1392 ret = ops->ndo_open(dev);
1394 netpoll_poll_enable(dev);
1397 clear_bit(__LINK_STATE_START, &dev->state);
1399 dev->flags |= IFF_UP;
1400 dev_set_rx_mode(dev);
1402 add_device_randomness(dev->dev_addr, dev->addr_len);
1409 * dev_open - prepare an interface for use.
1410 * @dev: device to open
1412 * Takes a device from down to up state. The device's private open
1413 * function is invoked and then the multicast lists are loaded. Finally
1414 * the device is moved into the up state and a %NETDEV_UP message is
1415 * sent to the netdev notifier chain.
1417 * Calling this function on an active interface is a nop. On a failure
1418 * a negative errno code is returned.
1420 int dev_open(struct net_device *dev)
1424 if (dev->flags & IFF_UP)
1427 ret = __dev_open(dev);
1431 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1432 call_netdevice_notifiers(NETDEV_UP, dev);
1436 EXPORT_SYMBOL(dev_open);
1438 static void __dev_close_many(struct list_head *head)
1440 struct net_device *dev;
1445 list_for_each_entry(dev, head, close_list) {
1446 /* Temporarily disable netpoll until the interface is down */
1447 netpoll_poll_disable(dev);
1449 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev);
1451 clear_bit(__LINK_STATE_START, &dev->state);
1453 /* Synchronize to scheduled poll. We cannot touch poll list, it
1454 * can be even on different cpu. So just clear netif_running().
1456 * dev->stop() will invoke napi_disable() on all of it's
1457 * napi_struct instances on this device.
1459 smp_mb__after_atomic(); /* Commit netif_running(). */
1462 dev_deactivate_many(head);
1464 list_for_each_entry(dev, head, close_list) {
1465 const struct net_device_ops *ops = dev->netdev_ops;
1468 * Call the device specific close. This cannot fail.
1469 * Only if device is UP
1471 * We allow it to be called even after a DETACH hot-plug
1477 dev->flags &= ~IFF_UP;
1478 netpoll_poll_enable(dev);
1482 static void __dev_close(struct net_device *dev)
1486 list_add(&dev->close_list, &single);
1487 __dev_close_many(&single);
1491 void dev_close_many(struct list_head *head, bool unlink)
1493 struct net_device *dev, *tmp;
1495 /* Remove the devices that don't need to be closed */
1496 list_for_each_entry_safe(dev, tmp, head, close_list)
1497 if (!(dev->flags & IFF_UP))
1498 list_del_init(&dev->close_list);
1500 __dev_close_many(head);
1502 list_for_each_entry_safe(dev, tmp, head, close_list) {
1503 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1504 call_netdevice_notifiers(NETDEV_DOWN, dev);
1506 list_del_init(&dev->close_list);
1509 EXPORT_SYMBOL(dev_close_many);
1512 * dev_close - shutdown an interface.
1513 * @dev: device to shutdown
1515 * This function moves an active device into down state. A
1516 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device
1517 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier
1520 void dev_close(struct net_device *dev)
1522 if (dev->flags & IFF_UP) {
1525 list_add(&dev->close_list, &single);
1526 dev_close_many(&single, true);
1530 EXPORT_SYMBOL(dev_close);
1534 * dev_disable_lro - disable Large Receive Offload on a device
1537 * Disable Large Receive Offload (LRO) on a net device. Must be
1538 * called under RTNL. This is needed if received packets may be
1539 * forwarded to another interface.
1541 void dev_disable_lro(struct net_device *dev)
1543 struct net_device *lower_dev;
1544 struct list_head *iter;
1546 dev->wanted_features &= ~NETIF_F_LRO;
1547 netdev_update_features(dev);
1549 if (unlikely(dev->features & NETIF_F_LRO))
1550 netdev_WARN(dev, "failed to disable LRO!\n");
1552 netdev_for_each_lower_dev(dev, lower_dev, iter)
1553 dev_disable_lro(lower_dev);
1555 EXPORT_SYMBOL(dev_disable_lro);
1558 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device
1561 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be
1562 * called under RTNL. This is needed if Generic XDP is installed on
1565 static void dev_disable_gro_hw(struct net_device *dev)
1567 dev->wanted_features &= ~NETIF_F_GRO_HW;
1568 netdev_update_features(dev);
1570 if (unlikely(dev->features & NETIF_F_GRO_HW))
1571 netdev_WARN(dev, "failed to disable GRO_HW!\n");
1574 const char *netdev_cmd_to_name(enum netdev_cmd cmd)
1577 case NETDEV_##val: \
1578 return "NETDEV_" __stringify(val);
1580 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER)
1581 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE)
1582 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE)
1583 N(POST_INIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) N(CHANGEUPPER)
1584 N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) N(BONDING_INFO)
1585 N(PRECHANGEUPPER) N(CHANGELOWERSTATE) N(UDP_TUNNEL_PUSH_INFO)
1586 N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN)
1587 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO)
1588 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO)
1591 return "UNKNOWN_NETDEV_EVENT";
1593 EXPORT_SYMBOL_GPL(netdev_cmd_to_name);
1595 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val,
1596 struct net_device *dev)
1598 struct netdev_notifier_info info = {
1602 return nb->notifier_call(nb, val, &info);
1605 static int dev_boot_phase = 1;
1608 * register_netdevice_notifier - register a network notifier block
1611 * Register a notifier to be called when network device events occur.
1612 * The notifier passed is linked into the kernel structures and must
1613 * not be reused until it has been unregistered. A negative errno code
1614 * is returned on a failure.
1616 * When registered all registration and up events are replayed
1617 * to the new notifier to allow device to have a race free
1618 * view of the network device list.
1621 int register_netdevice_notifier(struct notifier_block *nb)
1623 struct net_device *dev;
1624 struct net_device *last;
1628 /* Close race with setup_net() and cleanup_net() */
1629 down_write(&pernet_ops_rwsem);
1631 err = raw_notifier_chain_register(&netdev_chain, nb);
1637 for_each_netdev(net, dev) {
1638 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev);
1639 err = notifier_to_errno(err);
1643 if (!(dev->flags & IFF_UP))
1646 call_netdevice_notifier(nb, NETDEV_UP, dev);
1652 up_write(&pernet_ops_rwsem);
1658 for_each_netdev(net, dev) {
1662 if (dev->flags & IFF_UP) {
1663 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1665 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1667 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1672 raw_notifier_chain_unregister(&netdev_chain, nb);
1675 EXPORT_SYMBOL(register_netdevice_notifier);
1678 * unregister_netdevice_notifier - unregister a network notifier block
1681 * Unregister a notifier previously registered by
1682 * register_netdevice_notifier(). The notifier is unlinked into the
1683 * kernel structures and may then be reused. A negative errno code
1684 * is returned on a failure.
1686 * After unregistering unregister and down device events are synthesized
1687 * for all devices on the device list to the removed notifier to remove
1688 * the need for special case cleanup code.
1691 int unregister_netdevice_notifier(struct notifier_block *nb)
1693 struct net_device *dev;
1697 /* Close race with setup_net() and cleanup_net() */
1698 down_write(&pernet_ops_rwsem);
1700 err = raw_notifier_chain_unregister(&netdev_chain, nb);
1705 for_each_netdev(net, dev) {
1706 if (dev->flags & IFF_UP) {
1707 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1709 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1711 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1716 up_write(&pernet_ops_rwsem);
1719 EXPORT_SYMBOL(unregister_netdevice_notifier);
1722 * call_netdevice_notifiers_info - call all network notifier blocks
1723 * @val: value passed unmodified to notifier function
1724 * @info: notifier information data
1726 * Call all network notifier blocks. Parameters and return value
1727 * are as for raw_notifier_call_chain().
1730 static int call_netdevice_notifiers_info(unsigned long val,
1731 struct netdev_notifier_info *info)
1734 return raw_notifier_call_chain(&netdev_chain, val, info);
1738 * call_netdevice_notifiers - call all network notifier blocks
1739 * @val: value passed unmodified to notifier function
1740 * @dev: net_device pointer passed unmodified to notifier function
1742 * Call all network notifier blocks. Parameters and return value
1743 * are as for raw_notifier_call_chain().
1746 int call_netdevice_notifiers(unsigned long val, struct net_device *dev)
1748 struct netdev_notifier_info info = {
1752 return call_netdevice_notifiers_info(val, &info);
1754 EXPORT_SYMBOL(call_netdevice_notifiers);
1756 #ifdef CONFIG_NET_INGRESS
1757 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key);
1759 void net_inc_ingress_queue(void)
1761 static_branch_inc(&ingress_needed_key);
1763 EXPORT_SYMBOL_GPL(net_inc_ingress_queue);
1765 void net_dec_ingress_queue(void)
1767 static_branch_dec(&ingress_needed_key);
1769 EXPORT_SYMBOL_GPL(net_dec_ingress_queue);
1772 #ifdef CONFIG_NET_EGRESS
1773 static DEFINE_STATIC_KEY_FALSE(egress_needed_key);
1775 void net_inc_egress_queue(void)
1777 static_branch_inc(&egress_needed_key);
1779 EXPORT_SYMBOL_GPL(net_inc_egress_queue);
1781 void net_dec_egress_queue(void)
1783 static_branch_dec(&egress_needed_key);
1785 EXPORT_SYMBOL_GPL(net_dec_egress_queue);
1788 static DEFINE_STATIC_KEY_FALSE(netstamp_needed_key);
1789 #ifdef HAVE_JUMP_LABEL
1790 static atomic_t netstamp_needed_deferred;
1791 static atomic_t netstamp_wanted;
1792 static void netstamp_clear(struct work_struct *work)
1794 int deferred = atomic_xchg(&netstamp_needed_deferred, 0);
1797 wanted = atomic_add_return(deferred, &netstamp_wanted);
1799 static_branch_enable(&netstamp_needed_key);
1801 static_branch_disable(&netstamp_needed_key);
1803 static DECLARE_WORK(netstamp_work, netstamp_clear);
1806 void net_enable_timestamp(void)
1808 #ifdef HAVE_JUMP_LABEL
1812 wanted = atomic_read(&netstamp_wanted);
1815 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted + 1) == wanted)
1818 atomic_inc(&netstamp_needed_deferred);
1819 schedule_work(&netstamp_work);
1821 static_branch_inc(&netstamp_needed_key);
1824 EXPORT_SYMBOL(net_enable_timestamp);
1826 void net_disable_timestamp(void)
1828 #ifdef HAVE_JUMP_LABEL
1832 wanted = atomic_read(&netstamp_wanted);
1835 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted - 1) == wanted)
1838 atomic_dec(&netstamp_needed_deferred);
1839 schedule_work(&netstamp_work);
1841 static_branch_dec(&netstamp_needed_key);
1844 EXPORT_SYMBOL(net_disable_timestamp);
1846 static inline void net_timestamp_set(struct sk_buff *skb)
1849 if (static_branch_unlikely(&netstamp_needed_key))
1850 __net_timestamp(skb);
1853 #define net_timestamp_check(COND, SKB) \
1854 if (static_branch_unlikely(&netstamp_needed_key)) { \
1855 if ((COND) && !(SKB)->tstamp) \
1856 __net_timestamp(SKB); \
1859 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb)
1863 if (!(dev->flags & IFF_UP))
1866 len = dev->mtu + dev->hard_header_len + VLAN_HLEN;
1867 if (skb->len <= len)
1870 /* if TSO is enabled, we don't care about the length as the packet
1871 * could be forwarded without being segmented before
1873 if (skb_is_gso(skb))
1878 EXPORT_SYMBOL_GPL(is_skb_forwardable);
1880 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1882 int ret = ____dev_forward_skb(dev, skb);
1885 skb->protocol = eth_type_trans(skb, dev);
1886 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
1891 EXPORT_SYMBOL_GPL(__dev_forward_skb);
1894 * dev_forward_skb - loopback an skb to another netif
1896 * @dev: destination network device
1897 * @skb: buffer to forward
1900 * NET_RX_SUCCESS (no congestion)
1901 * NET_RX_DROP (packet was dropped, but freed)
1903 * dev_forward_skb can be used for injecting an skb from the
1904 * start_xmit function of one device into the receive queue
1905 * of another device.
1907 * The receiving device may be in another namespace, so
1908 * we have to clear all information in the skb that could
1909 * impact namespace isolation.
1911 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
1913 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb);
1915 EXPORT_SYMBOL_GPL(dev_forward_skb);
1917 static inline int deliver_skb(struct sk_buff *skb,
1918 struct packet_type *pt_prev,
1919 struct net_device *orig_dev)
1921 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
1923 refcount_inc(&skb->users);
1924 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
1927 static inline void deliver_ptype_list_skb(struct sk_buff *skb,
1928 struct packet_type **pt,
1929 struct net_device *orig_dev,
1931 struct list_head *ptype_list)
1933 struct packet_type *ptype, *pt_prev = *pt;
1935 list_for_each_entry_rcu(ptype, ptype_list, list) {
1936 if (ptype->type != type)
1939 deliver_skb(skb, pt_prev, orig_dev);
1945 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb)
1947 if (!ptype->af_packet_priv || !skb->sk)
1950 if (ptype->id_match)
1951 return ptype->id_match(ptype, skb->sk);
1952 else if ((struct sock *)ptype->af_packet_priv == skb->sk)
1959 * Support routine. Sends outgoing frames to any network
1960 * taps currently in use.
1963 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
1965 struct packet_type *ptype;
1966 struct sk_buff *skb2 = NULL;
1967 struct packet_type *pt_prev = NULL;
1968 struct list_head *ptype_list = &ptype_all;
1972 list_for_each_entry_rcu(ptype, ptype_list, list) {
1973 /* Never send packets back to the socket
1974 * they originated from - MvS (miquels@drinkel.ow.org)
1976 if (skb_loop_sk(ptype, skb))
1980 deliver_skb(skb2, pt_prev, skb->dev);
1985 /* need to clone skb, done only once */
1986 skb2 = skb_clone(skb, GFP_ATOMIC);
1990 net_timestamp_set(skb2);
1992 /* skb->nh should be correctly
1993 * set by sender, so that the second statement is
1994 * just protection against buggy protocols.
1996 skb_reset_mac_header(skb2);
1998 if (skb_network_header(skb2) < skb2->data ||
1999 skb_network_header(skb2) > skb_tail_pointer(skb2)) {
2000 net_crit_ratelimited("protocol %04x is buggy, dev %s\n",
2001 ntohs(skb2->protocol),
2003 skb_reset_network_header(skb2);
2006 skb2->transport_header = skb2->network_header;
2007 skb2->pkt_type = PACKET_OUTGOING;
2011 if (ptype_list == &ptype_all) {
2012 ptype_list = &dev->ptype_all;
2017 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC))
2018 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
2024 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit);
2027 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change
2028 * @dev: Network device
2029 * @txq: number of queues available
2031 * If real_num_tx_queues is changed the tc mappings may no longer be
2032 * valid. To resolve this verify the tc mapping remains valid and if
2033 * not NULL the mapping. With no priorities mapping to this
2034 * offset/count pair it will no longer be used. In the worst case TC0
2035 * is invalid nothing can be done so disable priority mappings. If is
2036 * expected that drivers will fix this mapping if they can before
2037 * calling netif_set_real_num_tx_queues.
2039 static void netif_setup_tc(struct net_device *dev, unsigned int txq)
2042 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2044 /* If TC0 is invalidated disable TC mapping */
2045 if (tc->offset + tc->count > txq) {
2046 pr_warn("Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n");
2051 /* Invalidated prio to tc mappings set to TC0 */
2052 for (i = 1; i < TC_BITMASK + 1; i++) {
2053 int q = netdev_get_prio_tc_map(dev, i);
2055 tc = &dev->tc_to_txq[q];
2056 if (tc->offset + tc->count > txq) {
2057 pr_warn("Number of in use tx queues changed. Priority %i to tc mapping %i is no longer valid. Setting map to 0\n",
2059 netdev_set_prio_tc_map(dev, i, 0);
2064 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq)
2067 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2070 /* walk through the TCs and see if it falls into any of them */
2071 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) {
2072 if ((txq - tc->offset) < tc->count)
2076 /* didn't find it, just return -1 to indicate no match */
2082 EXPORT_SYMBOL(netdev_txq_to_tc);
2085 struct static_key xps_needed __read_mostly;
2086 EXPORT_SYMBOL(xps_needed);
2087 struct static_key xps_rxqs_needed __read_mostly;
2088 EXPORT_SYMBOL(xps_rxqs_needed);
2089 static DEFINE_MUTEX(xps_map_mutex);
2090 #define xmap_dereference(P) \
2091 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex))
2093 static bool remove_xps_queue(struct xps_dev_maps *dev_maps,
2096 struct xps_map *map = NULL;
2100 map = xmap_dereference(dev_maps->attr_map[tci]);
2104 for (pos = map->len; pos--;) {
2105 if (map->queues[pos] != index)
2109 map->queues[pos] = map->queues[--map->len];
2113 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL);
2114 kfree_rcu(map, rcu);
2121 static bool remove_xps_queue_cpu(struct net_device *dev,
2122 struct xps_dev_maps *dev_maps,
2123 int cpu, u16 offset, u16 count)
2125 int num_tc = dev->num_tc ? : 1;
2126 bool active = false;
2129 for (tci = cpu * num_tc; num_tc--; tci++) {
2132 for (i = count, j = offset; i--; j++) {
2133 if (!remove_xps_queue(dev_maps, tci, j))
2143 static void clean_xps_maps(struct net_device *dev, const unsigned long *mask,
2144 struct xps_dev_maps *dev_maps, unsigned int nr_ids,
2145 u16 offset, u16 count, bool is_rxqs_map)
2147 bool active = false;
2150 for (j = -1; j = netif_attrmask_next(j, mask, nr_ids),
2152 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset,
2156 RCU_INIT_POINTER(dev->xps_rxqs_map, NULL);
2158 RCU_INIT_POINTER(dev->xps_cpus_map, NULL);
2160 for (i = offset + (count - 1); count--; i--)
2161 netdev_queue_numa_node_write(
2162 netdev_get_tx_queue(dev, i),
2165 kfree_rcu(dev_maps, rcu);
2169 static void netif_reset_xps_queues(struct net_device *dev, u16 offset,
2172 const unsigned long *possible_mask = NULL;
2173 struct xps_dev_maps *dev_maps;
2174 unsigned int nr_ids;
2176 if (!static_key_false(&xps_needed))
2179 mutex_lock(&xps_map_mutex);
2181 if (static_key_false(&xps_rxqs_needed)) {
2182 dev_maps = xmap_dereference(dev->xps_rxqs_map);
2184 nr_ids = dev->num_rx_queues;
2185 clean_xps_maps(dev, possible_mask, dev_maps, nr_ids,
2186 offset, count, true);
2190 dev_maps = xmap_dereference(dev->xps_cpus_map);
2194 if (num_possible_cpus() > 1)
2195 possible_mask = cpumask_bits(cpu_possible_mask);
2196 nr_ids = nr_cpu_ids;
2197 clean_xps_maps(dev, possible_mask, dev_maps, nr_ids, offset, count,
2201 if (static_key_enabled(&xps_rxqs_needed))
2202 static_key_slow_dec(&xps_rxqs_needed);
2204 static_key_slow_dec(&xps_needed);
2205 mutex_unlock(&xps_map_mutex);
2208 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index)
2210 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index);
2213 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index,
2214 u16 index, bool is_rxqs_map)
2216 struct xps_map *new_map;
2217 int alloc_len = XPS_MIN_MAP_ALLOC;
2220 for (pos = 0; map && pos < map->len; pos++) {
2221 if (map->queues[pos] != index)
2226 /* Need to add tx-queue to this CPU's/rx-queue's existing map */
2228 if (pos < map->alloc_len)
2231 alloc_len = map->alloc_len * 2;
2234 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's
2238 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL);
2240 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL,
2241 cpu_to_node(attr_index));
2245 for (i = 0; i < pos; i++)
2246 new_map->queues[i] = map->queues[i];
2247 new_map->alloc_len = alloc_len;
2253 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask,
2254 u16 index, bool is_rxqs_map)
2256 const unsigned long *online_mask = NULL, *possible_mask = NULL;
2257 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL;
2258 int i, j, tci, numa_node_id = -2;
2259 int maps_sz, num_tc = 1, tc = 0;
2260 struct xps_map *map, *new_map;
2261 bool active = false;
2262 unsigned int nr_ids;
2265 /* Do not allow XPS on subordinate device directly */
2266 num_tc = dev->num_tc;
2270 /* If queue belongs to subordinate dev use its map */
2271 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev;
2273 tc = netdev_txq_to_tc(dev, index);
2278 mutex_lock(&xps_map_mutex);
2280 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues);
2281 dev_maps = xmap_dereference(dev->xps_rxqs_map);
2282 nr_ids = dev->num_rx_queues;
2284 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc);
2285 if (num_possible_cpus() > 1) {
2286 online_mask = cpumask_bits(cpu_online_mask);
2287 possible_mask = cpumask_bits(cpu_possible_mask);
2289 dev_maps = xmap_dereference(dev->xps_cpus_map);
2290 nr_ids = nr_cpu_ids;
2293 if (maps_sz < L1_CACHE_BYTES)
2294 maps_sz = L1_CACHE_BYTES;
2296 /* allocate memory for queue storage */
2297 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids),
2300 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL);
2301 if (!new_dev_maps) {
2302 mutex_unlock(&xps_map_mutex);
2306 tci = j * num_tc + tc;
2307 map = dev_maps ? xmap_dereference(dev_maps->attr_map[tci]) :
2310 map = expand_xps_map(map, j, index, is_rxqs_map);
2314 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map);
2318 goto out_no_new_maps;
2320 static_key_slow_inc(&xps_needed);
2322 static_key_slow_inc(&xps_rxqs_needed);
2324 for (j = -1; j = netif_attrmask_next(j, possible_mask, nr_ids),
2326 /* copy maps belonging to foreign traffic classes */
2327 for (i = tc, tci = j * num_tc; dev_maps && i--; tci++) {
2328 /* fill in the new device map from the old device map */
2329 map = xmap_dereference(dev_maps->attr_map[tci]);
2330 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map);
2333 /* We need to explicitly update tci as prevous loop
2334 * could break out early if dev_maps is NULL.
2336 tci = j * num_tc + tc;
2338 if (netif_attr_test_mask(j, mask, nr_ids) &&
2339 netif_attr_test_online(j, online_mask, nr_ids)) {
2340 /* add tx-queue to CPU/rx-queue maps */
2343 map = xmap_dereference(new_dev_maps->attr_map[tci]);
2344 while ((pos < map->len) && (map->queues[pos] != index))
2347 if (pos == map->len)
2348 map->queues[map->len++] = index;
2351 if (numa_node_id == -2)
2352 numa_node_id = cpu_to_node(j);
2353 else if (numa_node_id != cpu_to_node(j))
2357 } else if (dev_maps) {
2358 /* fill in the new device map from the old device map */
2359 map = xmap_dereference(dev_maps->attr_map[tci]);
2360 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map);
2363 /* copy maps belonging to foreign traffic classes */
2364 for (i = num_tc - tc, tci++; dev_maps && --i; tci++) {
2365 /* fill in the new device map from the old device map */
2366 map = xmap_dereference(dev_maps->attr_map[tci]);
2367 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map);
2372 rcu_assign_pointer(dev->xps_rxqs_map, new_dev_maps);
2374 rcu_assign_pointer(dev->xps_cpus_map, new_dev_maps);
2376 /* Cleanup old maps */
2378 goto out_no_old_maps;
2380 for (j = -1; j = netif_attrmask_next(j, possible_mask, nr_ids),
2382 for (i = num_tc, tci = j * num_tc; i--; tci++) {
2383 new_map = xmap_dereference(new_dev_maps->attr_map[tci]);
2384 map = xmap_dereference(dev_maps->attr_map[tci]);
2385 if (map && map != new_map)
2386 kfree_rcu(map, rcu);
2390 kfree_rcu(dev_maps, rcu);
2393 dev_maps = new_dev_maps;
2398 /* update Tx queue numa node */
2399 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index),
2400 (numa_node_id >= 0) ?
2401 numa_node_id : NUMA_NO_NODE);
2407 /* removes tx-queue from unused CPUs/rx-queues */
2408 for (j = -1; j = netif_attrmask_next(j, possible_mask, nr_ids),
2410 for (i = tc, tci = j * num_tc; i--; tci++)
2411 active |= remove_xps_queue(dev_maps, tci, index);
2412 if (!netif_attr_test_mask(j, mask, nr_ids) ||
2413 !netif_attr_test_online(j, online_mask, nr_ids))
2414 active |= remove_xps_queue(dev_maps, tci, index);
2415 for (i = num_tc - tc, tci++; --i; tci++)
2416 active |= remove_xps_queue(dev_maps, tci, index);
2419 /* free map if not active */
2422 RCU_INIT_POINTER(dev->xps_rxqs_map, NULL);
2424 RCU_INIT_POINTER(dev->xps_cpus_map, NULL);
2425 kfree_rcu(dev_maps, rcu);
2429 mutex_unlock(&xps_map_mutex);
2433 /* remove any maps that we added */
2434 for (j = -1; j = netif_attrmask_next(j, possible_mask, nr_ids),
2436 for (i = num_tc, tci = j * num_tc; i--; tci++) {
2437 new_map = xmap_dereference(new_dev_maps->attr_map[tci]);
2439 xmap_dereference(dev_maps->attr_map[tci]) :
2441 if (new_map && new_map != map)
2446 mutex_unlock(&xps_map_mutex);
2448 kfree(new_dev_maps);
2452 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask,
2455 return __netif_set_xps_queue(dev, cpumask_bits(mask), index, false);
2457 EXPORT_SYMBOL(netif_set_xps_queue);
2460 static void netdev_unbind_all_sb_channels(struct net_device *dev)
2462 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues];
2464 /* Unbind any subordinate channels */
2465 while (txq-- != &dev->_tx[0]) {
2467 netdev_unbind_sb_channel(dev, txq->sb_dev);
2471 void netdev_reset_tc(struct net_device *dev)
2474 netif_reset_xps_queues_gt(dev, 0);
2476 netdev_unbind_all_sb_channels(dev);
2478 /* Reset TC configuration of device */
2480 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq));
2481 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map));
2483 EXPORT_SYMBOL(netdev_reset_tc);
2485 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset)
2487 if (tc >= dev->num_tc)
2491 netif_reset_xps_queues(dev, offset, count);
2493 dev->tc_to_txq[tc].count = count;
2494 dev->tc_to_txq[tc].offset = offset;
2497 EXPORT_SYMBOL(netdev_set_tc_queue);
2499 int netdev_set_num_tc(struct net_device *dev, u8 num_tc)
2501 if (num_tc > TC_MAX_QUEUE)
2505 netif_reset_xps_queues_gt(dev, 0);
2507 netdev_unbind_all_sb_channels(dev);
2509 dev->num_tc = num_tc;
2512 EXPORT_SYMBOL(netdev_set_num_tc);
2514 void netdev_unbind_sb_channel(struct net_device *dev,
2515 struct net_device *sb_dev)
2517 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues];
2520 netif_reset_xps_queues_gt(sb_dev, 0);
2522 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq));
2523 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map));
2525 while (txq-- != &dev->_tx[0]) {
2526 if (txq->sb_dev == sb_dev)
2530 EXPORT_SYMBOL(netdev_unbind_sb_channel);
2532 int netdev_bind_sb_channel_queue(struct net_device *dev,
2533 struct net_device *sb_dev,
2534 u8 tc, u16 count, u16 offset)
2536 /* Make certain the sb_dev and dev are already configured */
2537 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc)
2540 /* We cannot hand out queues we don't have */
2541 if ((offset + count) > dev->real_num_tx_queues)
2544 /* Record the mapping */
2545 sb_dev->tc_to_txq[tc].count = count;
2546 sb_dev->tc_to_txq[tc].offset = offset;
2548 /* Provide a way for Tx queue to find the tc_to_txq map or
2549 * XPS map for itself.
2552 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev;
2556 EXPORT_SYMBOL(netdev_bind_sb_channel_queue);
2558 int netdev_set_sb_channel(struct net_device *dev, u16 channel)
2560 /* Do not use a multiqueue device to represent a subordinate channel */
2561 if (netif_is_multiqueue(dev))
2564 /* We allow channels 1 - 32767 to be used for subordinate channels.
2565 * Channel 0 is meant to be "native" mode and used only to represent
2566 * the main root device. We allow writing 0 to reset the device back
2567 * to normal mode after being used as a subordinate channel.
2569 if (channel > S16_MAX)
2572 dev->num_tc = -channel;
2576 EXPORT_SYMBOL(netdev_set_sb_channel);
2579 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues
2580 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed.
2582 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq)
2587 disabling = txq < dev->real_num_tx_queues;
2589 if (txq < 1 || txq > dev->num_tx_queues)
2592 if (dev->reg_state == NETREG_REGISTERED ||
2593 dev->reg_state == NETREG_UNREGISTERING) {
2596 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues,
2602 netif_setup_tc(dev, txq);
2604 dev->real_num_tx_queues = txq;
2608 qdisc_reset_all_tx_gt(dev, txq);
2610 netif_reset_xps_queues_gt(dev, txq);
2614 dev->real_num_tx_queues = txq;
2619 EXPORT_SYMBOL(netif_set_real_num_tx_queues);
2623 * netif_set_real_num_rx_queues - set actual number of RX queues used
2624 * @dev: Network device
2625 * @rxq: Actual number of RX queues
2627 * This must be called either with the rtnl_lock held or before
2628 * registration of the net device. Returns 0 on success, or a
2629 * negative error code. If called before registration, it always
2632 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq)
2636 if (rxq < 1 || rxq > dev->num_rx_queues)
2639 if (dev->reg_state == NETREG_REGISTERED) {
2642 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues,
2648 dev->real_num_rx_queues = rxq;
2651 EXPORT_SYMBOL(netif_set_real_num_rx_queues);
2655 * netif_get_num_default_rss_queues - default number of RSS queues
2657 * This routine should set an upper limit on the number of RSS queues
2658 * used by default by multiqueue devices.
2660 int netif_get_num_default_rss_queues(void)
2662 return is_kdump_kernel() ?
2663 1 : min_t(int, DEFAULT_MAX_NUM_RSS_QUEUES, num_online_cpus());
2665 EXPORT_SYMBOL(netif_get_num_default_rss_queues);
2667 static void __netif_reschedule(struct Qdisc *q)
2669 struct softnet_data *sd;
2670 unsigned long flags;
2672 local_irq_save(flags);
2673 sd = this_cpu_ptr(&softnet_data);
2674 q->next_sched = NULL;
2675 *sd->output_queue_tailp = q;
2676 sd->output_queue_tailp = &q->next_sched;
2677 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2678 local_irq_restore(flags);
2681 void __netif_schedule(struct Qdisc *q)
2683 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
2684 __netif_reschedule(q);
2686 EXPORT_SYMBOL(__netif_schedule);
2688 struct dev_kfree_skb_cb {
2689 enum skb_free_reason reason;
2692 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb)
2694 return (struct dev_kfree_skb_cb *)skb->cb;
2697 void netif_schedule_queue(struct netdev_queue *txq)
2700 if (!(txq->state & QUEUE_STATE_ANY_XOFF)) {
2701 struct Qdisc *q = rcu_dereference(txq->qdisc);
2703 __netif_schedule(q);
2707 EXPORT_SYMBOL(netif_schedule_queue);
2709 void netif_tx_wake_queue(struct netdev_queue *dev_queue)
2711 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) {
2715 q = rcu_dereference(dev_queue->qdisc);
2716 __netif_schedule(q);
2720 EXPORT_SYMBOL(netif_tx_wake_queue);
2722 void __dev_kfree_skb_irq(struct sk_buff *skb, enum skb_free_reason reason)
2724 unsigned long flags;
2729 if (likely(refcount_read(&skb->users) == 1)) {
2731 refcount_set(&skb->users, 0);
2732 } else if (likely(!refcount_dec_and_test(&skb->users))) {
2735 get_kfree_skb_cb(skb)->reason = reason;
2736 local_irq_save(flags);
2737 skb->next = __this_cpu_read(softnet_data.completion_queue);
2738 __this_cpu_write(softnet_data.completion_queue, skb);
2739 raise_softirq_irqoff(NET_TX_SOFTIRQ);
2740 local_irq_restore(flags);
2742 EXPORT_SYMBOL(__dev_kfree_skb_irq);
2744 void __dev_kfree_skb_any(struct sk_buff *skb, enum skb_free_reason reason)
2746 if (in_irq() || irqs_disabled())
2747 __dev_kfree_skb_irq(skb, reason);
2751 EXPORT_SYMBOL(__dev_kfree_skb_any);
2755 * netif_device_detach - mark device as removed
2756 * @dev: network device
2758 * Mark device as removed from system and therefore no longer available.
2760 void netif_device_detach(struct net_device *dev)
2762 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) &&
2763 netif_running(dev)) {
2764 netif_tx_stop_all_queues(dev);
2767 EXPORT_SYMBOL(netif_device_detach);
2770 * netif_device_attach - mark device as attached
2771 * @dev: network device
2773 * Mark device as attached from system and restart if needed.
2775 void netif_device_attach(struct net_device *dev)
2777 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) &&
2778 netif_running(dev)) {
2779 netif_tx_wake_all_queues(dev);
2780 __netdev_watchdog_up(dev);
2783 EXPORT_SYMBOL(netif_device_attach);
2786 * Returns a Tx hash based on the given packet descriptor a Tx queues' number
2787 * to be used as a distribution range.
2789 static u16 skb_tx_hash(const struct net_device *dev,
2790 const struct net_device *sb_dev,
2791 struct sk_buff *skb)
2795 u16 qcount = dev->real_num_tx_queues;
2798 u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
2800 qoffset = sb_dev->tc_to_txq[tc].offset;
2801 qcount = sb_dev->tc_to_txq[tc].count;
2804 if (skb_rx_queue_recorded(skb)) {
2805 hash = skb_get_rx_queue(skb);
2806 while (unlikely(hash >= qcount))
2808 return hash + qoffset;
2811 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset;
2814 static void skb_warn_bad_offload(const struct sk_buff *skb)
2816 static const netdev_features_t null_features;
2817 struct net_device *dev = skb->dev;
2818 const char *name = "";
2820 if (!net_ratelimit())
2824 if (dev->dev.parent)
2825 name = dev_driver_string(dev->dev.parent);
2827 name = netdev_name(dev);
2829 WARN(1, "%s: caps=(%pNF, %pNF) len=%d data_len=%d gso_size=%d "
2830 "gso_type=%d ip_summed=%d\n",
2831 name, dev ? &dev->features : &null_features,
2832 skb->sk ? &skb->sk->sk_route_caps : &null_features,
2833 skb->len, skb->data_len, skb_shinfo(skb)->gso_size,
2834 skb_shinfo(skb)->gso_type, skb->ip_summed);
2838 * Invalidate hardware checksum when packet is to be mangled, and
2839 * complete checksum manually on outgoing path.
2841 int skb_checksum_help(struct sk_buff *skb)
2844 int ret = 0, offset;
2846 if (skb->ip_summed == CHECKSUM_COMPLETE)
2847 goto out_set_summed;
2849 if (unlikely(skb_shinfo(skb)->gso_size)) {
2850 skb_warn_bad_offload(skb);
2854 /* Before computing a checksum, we should make sure no frag could
2855 * be modified by an external entity : checksum could be wrong.
2857 if (skb_has_shared_frag(skb)) {
2858 ret = __skb_linearize(skb);
2863 offset = skb_checksum_start_offset(skb);
2864 BUG_ON(offset >= skb_headlen(skb));
2865 csum = skb_checksum(skb, offset, skb->len - offset, 0);
2867 offset += skb->csum_offset;
2868 BUG_ON(offset + sizeof(__sum16) > skb_headlen(skb));
2870 if (skb_cloned(skb) &&
2871 !skb_clone_writable(skb, offset + sizeof(__sum16))) {
2872 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2877 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0;
2879 skb->ip_summed = CHECKSUM_NONE;
2883 EXPORT_SYMBOL(skb_checksum_help);
2885 int skb_crc32c_csum_help(struct sk_buff *skb)
2888 int ret = 0, offset, start;
2890 if (skb->ip_summed != CHECKSUM_PARTIAL)
2893 if (unlikely(skb_is_gso(skb)))
2896 /* Before computing a checksum, we should make sure no frag could
2897 * be modified by an external entity : checksum could be wrong.
2899 if (unlikely(skb_has_shared_frag(skb))) {
2900 ret = __skb_linearize(skb);
2904 start = skb_checksum_start_offset(skb);
2905 offset = start + offsetof(struct sctphdr, checksum);
2906 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) {
2910 if (skb_cloned(skb) &&
2911 !skb_clone_writable(skb, offset + sizeof(__le32))) {
2912 ret = pskb_expand_head(skb, 0, 0, GFP_ATOMIC);
2916 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start,
2917 skb->len - start, ~(__u32)0,
2919 *(__le32 *)(skb->data + offset) = crc32c_csum;
2920 skb->ip_summed = CHECKSUM_NONE;
2921 skb->csum_not_inet = 0;
2926 __be16 skb_network_protocol(struct sk_buff *skb, int *depth)
2928 __be16 type = skb->protocol;
2930 /* Tunnel gso handlers can set protocol to ethernet. */
2931 if (type == htons(ETH_P_TEB)) {
2934 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr))))
2937 eth = (struct ethhdr *)skb->data;
2938 type = eth->h_proto;
2941 return __vlan_get_protocol(skb, type, depth);
2945 * skb_mac_gso_segment - mac layer segmentation handler.
2946 * @skb: buffer to segment
2947 * @features: features for the output path (see dev->features)
2949 struct sk_buff *skb_mac_gso_segment(struct sk_buff *skb,
2950 netdev_features_t features)
2952 struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT);
2953 struct packet_offload *ptype;
2954 int vlan_depth = skb->mac_len;
2955 __be16 type = skb_network_protocol(skb, &vlan_depth);
2957 if (unlikely(!type))
2958 return ERR_PTR(-EINVAL);
2960 __skb_pull(skb, vlan_depth);
2963 list_for_each_entry_rcu(ptype, &offload_base, list) {
2964 if (ptype->type == type && ptype->callbacks.gso_segment) {
2965 segs = ptype->callbacks.gso_segment(skb, features);
2971 __skb_push(skb, skb->data - skb_mac_header(skb));
2975 EXPORT_SYMBOL(skb_mac_gso_segment);
2978 /* openvswitch calls this on rx path, so we need a different check.
2980 static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
2983 return skb->ip_summed != CHECKSUM_PARTIAL &&
2984 skb->ip_summed != CHECKSUM_UNNECESSARY;
2986 return skb->ip_summed == CHECKSUM_NONE;
2990 * __skb_gso_segment - Perform segmentation on skb.
2991 * @skb: buffer to segment
2992 * @features: features for the output path (see dev->features)
2993 * @tx_path: whether it is called in TX path
2995 * This function segments the given skb and returns a list of segments.
2997 * It may return NULL if the skb requires no segmentation. This is
2998 * only possible when GSO is used for verifying header integrity.
3000 * Segmentation preserves SKB_SGO_CB_OFFSET bytes of previous skb cb.
3002 struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
3003 netdev_features_t features, bool tx_path)
3005 struct sk_buff *segs;
3007 if (unlikely(skb_needs_check(skb, tx_path))) {
3010 /* We're going to init ->check field in TCP or UDP header */
3011 err = skb_cow_head(skb, 0);
3013 return ERR_PTR(err);
3016 /* Only report GSO partial support if it will enable us to
3017 * support segmentation on this frame without needing additional
3020 if (features & NETIF_F_GSO_PARTIAL) {
3021 netdev_features_t partial_features = NETIF_F_GSO_ROBUST;
3022 struct net_device *dev = skb->dev;
3024 partial_features |= dev->features & dev->gso_partial_features;
3025 if (!skb_gso_ok(skb, features | partial_features))
3026 features &= ~NETIF_F_GSO_PARTIAL;
3029 BUILD_BUG_ON(SKB_SGO_CB_OFFSET +
3030 sizeof(*SKB_GSO_CB(skb)) > sizeof(skb->cb));
3032 SKB_GSO_CB(skb)->mac_offset = skb_headroom(skb);
3033 SKB_GSO_CB(skb)->encap_level = 0;
3035 skb_reset_mac_header(skb);
3036 skb_reset_mac_len(skb);
3038 segs = skb_mac_gso_segment(skb, features);
3040 if (unlikely(skb_needs_check(skb, tx_path) && !IS_ERR(segs)))
3041 skb_warn_bad_offload(skb);
3045 EXPORT_SYMBOL(__skb_gso_segment);
3047 /* Take action when hardware reception checksum errors are detected. */
3049 void netdev_rx_csum_fault(struct net_device *dev)
3051 if (net_ratelimit()) {
3052 pr_err("%s: hw csum failure\n", dev ? dev->name : "<unknown>");
3056 EXPORT_SYMBOL(netdev_rx_csum_fault);
3059 /* XXX: check that highmem exists at all on the given machine. */
3060 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
3062 #ifdef CONFIG_HIGHMEM
3065 if (!(dev->features & NETIF_F_HIGHDMA)) {
3066 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
3067 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
3069 if (PageHighMem(skb_frag_page(frag)))
3077 /* If MPLS offload request, verify we are testing hardware MPLS features
3078 * instead of standard features for the netdev.
3080 #if IS_ENABLED(CONFIG_NET_MPLS_GSO)
3081 static netdev_features_t net_mpls_features(struct sk_buff *skb,
3082 netdev_features_t features,
3085 if (eth_p_mpls(type))
3086 features &= skb->dev->mpls_features;
3091 static netdev_features_t net_mpls_features(struct sk_buff *skb,
3092 netdev_features_t features,
3099 static netdev_features_t harmonize_features(struct sk_buff *skb,
3100 netdev_features_t features)
3105 type = skb_network_protocol(skb, &tmp);
3106 features = net_mpls_features(skb, features, type);
3108 if (skb->ip_summed != CHECKSUM_NONE &&
3109 !can_checksum_protocol(features, type)) {
3110 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
3112 if (illegal_highdma(skb->dev, skb))
3113 features &= ~NETIF_F_SG;
3118 netdev_features_t passthru_features_check(struct sk_buff *skb,
3119 struct net_device *dev,
3120 netdev_features_t features)
3124 EXPORT_SYMBOL(passthru_features_check);
3126 static netdev_features_t dflt_features_check(struct sk_buff *skb,
3127 struct net_device *dev,
3128 netdev_features_t features)
3130 return vlan_features_check(skb, features);
3133 static netdev_features_t gso_features_check(const struct sk_buff *skb,
3134 struct net_device *dev,
3135 netdev_features_t features)
3137 u16 gso_segs = skb_shinfo(skb)->gso_segs;
3139 if (gso_segs > dev->gso_max_segs)
3140 return features & ~NETIF_F_GSO_MASK;
3142 /* Support for GSO partial features requires software
3143 * intervention before we can actually process the packets
3144 * so we need to strip support for any partial features now
3145 * and we can pull them back in after we have partially
3146 * segmented the frame.
3148 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL))
3149 features &= ~dev->gso_partial_features;
3151 /* Make sure to clear the IPv4 ID mangling feature if the
3152 * IPv4 header has the potential to be fragmented.
3154 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
3155 struct iphdr *iph = skb->encapsulation ?
3156 inner_ip_hdr(skb) : ip_hdr(skb);
3158 if (!(iph->frag_off & htons(IP_DF)))
3159 features &= ~NETIF_F_TSO_MANGLEID;
3165 netdev_features_t netif_skb_features(struct sk_buff *skb)
3167 struct net_device *dev = skb->dev;
3168 netdev_features_t features = dev->features;
3170 if (skb_is_gso(skb))
3171 features = gso_features_check(skb, dev, features);
3173 /* If encapsulation offload request, verify we are testing
3174 * hardware encapsulation features instead of standard
3175 * features for the netdev
3177 if (skb->encapsulation)
3178 features &= dev->hw_enc_features;
3180 if (skb_vlan_tagged(skb))
3181 features = netdev_intersect_features(features,
3182 dev->vlan_features |
3183 NETIF_F_HW_VLAN_CTAG_TX |
3184 NETIF_F_HW_VLAN_STAG_TX);
3186 if (dev->netdev_ops->ndo_features_check)
3187 features &= dev->netdev_ops->ndo_features_check(skb, dev,
3190 features &= dflt_features_check(skb, dev, features);
3192 return harmonize_features(skb, features);
3194 EXPORT_SYMBOL(netif_skb_features);
3196 static int xmit_one(struct sk_buff *skb, struct net_device *dev,
3197 struct netdev_queue *txq, bool more)
3202 if (!list_empty(&ptype_all) || !list_empty(&dev->ptype_all))
3203 dev_queue_xmit_nit(skb, dev);
3206 trace_net_dev_start_xmit(skb, dev);
3207 rc = netdev_start_xmit(skb, dev, txq, more);
3208 trace_net_dev_xmit(skb, rc, dev, len);
3213 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev,
3214 struct netdev_queue *txq, int *ret)
3216 struct sk_buff *skb = first;
3217 int rc = NETDEV_TX_OK;
3220 struct sk_buff *next = skb->next;
3223 rc = xmit_one(skb, dev, txq, next != NULL);
3224 if (unlikely(!dev_xmit_complete(rc))) {
3230 if (netif_xmit_stopped(txq) && skb) {
3231 rc = NETDEV_TX_BUSY;
3241 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb,
3242 netdev_features_t features)
3244 if (skb_vlan_tag_present(skb) &&
3245 !vlan_hw_offload_capable(features, skb->vlan_proto))
3246 skb = __vlan_hwaccel_push_inside(skb);
3250 int skb_csum_hwoffload_help(struct sk_buff *skb,
3251 const netdev_features_t features)
3253 if (unlikely(skb->csum_not_inet))
3254 return !!(features & NETIF_F_SCTP_CRC) ? 0 :
3255 skb_crc32c_csum_help(skb);
3257 return !!(features & NETIF_F_CSUM_MASK) ? 0 : skb_checksum_help(skb);
3259 EXPORT_SYMBOL(skb_csum_hwoffload_help);
3261 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again)
3263 netdev_features_t features;
3265 features = netif_skb_features(skb);
3266 skb = validate_xmit_vlan(skb, features);
3270 skb = sk_validate_xmit_skb(skb, dev);
3274 if (netif_needs_gso(skb, features)) {
3275 struct sk_buff *segs;
3277 segs = skb_gso_segment(skb, features);
3285 if (skb_needs_linearize(skb, features) &&
3286 __skb_linearize(skb))
3289 /* If packet is not checksummed and device does not
3290 * support checksumming for this protocol, complete
3291 * checksumming here.
3293 if (skb->ip_summed == CHECKSUM_PARTIAL) {
3294 if (skb->encapsulation)
3295 skb_set_inner_transport_header(skb,
3296 skb_checksum_start_offset(skb));
3298 skb_set_transport_header(skb,
3299 skb_checksum_start_offset(skb));
3300 if (skb_csum_hwoffload_help(skb, features))
3305 skb = validate_xmit_xfrm(skb, features, again);
3312 atomic_long_inc(&dev->tx_dropped);
3316 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again)
3318 struct sk_buff *next, *head = NULL, *tail;
3320 for (; skb != NULL; skb = next) {
3324 /* in case skb wont be segmented, point to itself */
3327 skb = validate_xmit_skb(skb, dev, again);
3335 /* If skb was segmented, skb->prev points to
3336 * the last segment. If not, it still contains skb.
3342 EXPORT_SYMBOL_GPL(validate_xmit_skb_list);
3344 static void qdisc_pkt_len_init(struct sk_buff *skb)
3346 const struct skb_shared_info *shinfo = skb_shinfo(skb);
3348 qdisc_skb_cb(skb)->pkt_len = skb->len;
3350 /* To get more precise estimation of bytes sent on wire,
3351 * we add to pkt_len the headers size of all segments
3353 if (shinfo->gso_size) {
3354 unsigned int hdr_len;
3355 u16 gso_segs = shinfo->gso_segs;
3357 /* mac layer + network layer */
3358 hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
3360 /* + transport layer */
3361 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) {
3362 const struct tcphdr *th;
3363 struct tcphdr _tcphdr;
3365 th = skb_header_pointer(skb, skb_transport_offset(skb),
3366 sizeof(_tcphdr), &_tcphdr);
3368 hdr_len += __tcp_hdrlen(th);
3370 struct udphdr _udphdr;
3372 if (skb_header_pointer(skb, skb_transport_offset(skb),
3373 sizeof(_udphdr), &_udphdr))
3374 hdr_len += sizeof(struct udphdr);
3377 if (shinfo->gso_type & SKB_GSO_DODGY)
3378 gso_segs = DIV_ROUND_UP(skb->len - hdr_len,
3381 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len;
3385 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
3386 struct net_device *dev,
3387 struct netdev_queue *txq)
3389 spinlock_t *root_lock = qdisc_lock(q);
3390 struct sk_buff *to_free = NULL;
3394 qdisc_calculate_pkt_len(skb, q);
3396 if (q->flags & TCQ_F_NOLOCK) {
3397 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3398 __qdisc_drop(skb, &to_free);
3401 rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
3405 if (unlikely(to_free))
3406 kfree_skb_list(to_free);
3411 * Heuristic to force contended enqueues to serialize on a
3412 * separate lock before trying to get qdisc main lock.
3413 * This permits qdisc->running owner to get the lock more
3414 * often and dequeue packets faster.
3416 contended = qdisc_is_running(q);
3417 if (unlikely(contended))
3418 spin_lock(&q->busylock);
3420 spin_lock(root_lock);
3421 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3422 __qdisc_drop(skb, &to_free);
3424 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
3425 qdisc_run_begin(q)) {
3427 * This is a work-conserving queue; there are no old skbs
3428 * waiting to be sent out; and the qdisc is not running -
3429 * xmit the skb directly.
3432 qdisc_bstats_update(q, skb);
3434 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) {
3435 if (unlikely(contended)) {
3436 spin_unlock(&q->busylock);
3443 rc = NET_XMIT_SUCCESS;
3445 rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
3446 if (qdisc_run_begin(q)) {
3447 if (unlikely(contended)) {
3448 spin_unlock(&q->busylock);
3455 spin_unlock(root_lock);
3456 if (unlikely(to_free))
3457 kfree_skb_list(to_free);
3458 if (unlikely(contended))
3459 spin_unlock(&q->busylock);
3463 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO)
3464 static void skb_update_prio(struct sk_buff *skb)
3466 const struct netprio_map *map;
3467 const struct sock *sk;
3468 unsigned int prioidx;
3472 map = rcu_dereference_bh(skb->dev->priomap);
3475 sk = skb_to_full_sk(skb);
3479 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data);
3481 if (prioidx < map->priomap_len)
3482 skb->priority = map->priomap[prioidx];
3485 #define skb_update_prio(skb)
3488 DEFINE_PER_CPU(int, xmit_recursion);
3489 EXPORT_SYMBOL(xmit_recursion);
3492 * dev_loopback_xmit - loop back @skb
3493 * @net: network namespace this loopback is happening in
3494 * @sk: sk needed to be a netfilter okfn
3495 * @skb: buffer to transmit
3497 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb)
3499 skb_reset_mac_header(skb);
3500 __skb_pull(skb, skb_network_offset(skb));
3501 skb->pkt_type = PACKET_LOOPBACK;
3502 skb->ip_summed = CHECKSUM_UNNECESSARY;
3503 WARN_ON(!skb_dst(skb));
3508 EXPORT_SYMBOL(dev_loopback_xmit);
3510 #ifdef CONFIG_NET_EGRESS
3511 static struct sk_buff *
3512 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
3514 struct mini_Qdisc *miniq = rcu_dereference_bh(dev->miniq_egress);
3515 struct tcf_result cl_res;
3520 /* qdisc_skb_cb(skb)->pkt_len was already set by the caller. */
3521 mini_qdisc_bstats_cpu_update(miniq, skb);
3523 switch (tcf_classify(skb, miniq->filter_list, &cl_res, false)) {
3525 case TC_ACT_RECLASSIFY:
3526 skb->tc_index = TC_H_MIN(cl_res.classid);
3529 mini_qdisc_qstats_cpu_drop(miniq);
3530 *ret = NET_XMIT_DROP;
3536 *ret = NET_XMIT_SUCCESS;
3539 case TC_ACT_REDIRECT:
3540 /* No need to push/pop skb's mac_header here on egress! */
3541 skb_do_redirect(skb);
3542 *ret = NET_XMIT_SUCCESS;
3550 #endif /* CONFIG_NET_EGRESS */
3553 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb,
3554 struct xps_dev_maps *dev_maps, unsigned int tci)
3556 struct xps_map *map;
3557 int queue_index = -1;
3561 tci += netdev_get_prio_tc_map(dev, skb->priority);
3564 map = rcu_dereference(dev_maps->attr_map[tci]);
3567 queue_index = map->queues[0];
3569 queue_index = map->queues[reciprocal_scale(
3570 skb_get_hash(skb), map->len)];
3571 if (unlikely(queue_index >= dev->real_num_tx_queues))
3578 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev,
3579 struct sk_buff *skb)
3582 struct xps_dev_maps *dev_maps;
3583 struct sock *sk = skb->sk;
3584 int queue_index = -1;
3586 if (!static_key_false(&xps_needed))
3590 if (!static_key_false(&xps_rxqs_needed))
3593 dev_maps = rcu_dereference(sb_dev->xps_rxqs_map);
3595 int tci = sk_rx_queue_get(sk);
3597 if (tci >= 0 && tci < dev->num_rx_queues)
3598 queue_index = __get_xps_queue_idx(dev, skb, dev_maps,
3603 if (queue_index < 0) {
3604 dev_maps = rcu_dereference(sb_dev->xps_cpus_map);
3606 unsigned int tci = skb->sender_cpu - 1;
3608 queue_index = __get_xps_queue_idx(dev, skb, dev_maps,
3620 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb,
3621 struct net_device *sb_dev,
3622 select_queue_fallback_t fallback)
3626 EXPORT_SYMBOL(dev_pick_tx_zero);
3628 u16 dev_pick_tx_cpu_id(struct net_device *dev, struct sk_buff *skb,
3629 struct net_device *sb_dev,
3630 select_queue_fallback_t fallback)
3632 return (u16)raw_smp_processor_id() % dev->real_num_tx_queues;
3634 EXPORT_SYMBOL(dev_pick_tx_cpu_id);
3636 static u16 __netdev_pick_tx(struct net_device *dev, struct sk_buff *skb,
3637 struct net_device *sb_dev)
3639 struct sock *sk = skb->sk;
3640 int queue_index = sk_tx_queue_get(sk);
3642 sb_dev = sb_dev ? : dev;
3644 if (queue_index < 0 || skb->ooo_okay ||
3645 queue_index >= dev->real_num_tx_queues) {
3646 int new_index = get_xps_queue(dev, sb_dev, skb);
3649 new_index = skb_tx_hash(dev, sb_dev, skb);
3651 if (queue_index != new_index && sk &&
3653 rcu_access_pointer(sk->sk_dst_cache))
3654 sk_tx_queue_set(sk, new_index);
3656 queue_index = new_index;
3662 struct netdev_queue *netdev_pick_tx(struct net_device *dev,
3663 struct sk_buff *skb,
3664 struct net_device *sb_dev)
3666 int queue_index = 0;
3669 u32 sender_cpu = skb->sender_cpu - 1;
3671 if (sender_cpu >= (u32)NR_CPUS)
3672 skb->sender_cpu = raw_smp_processor_id() + 1;
3675 if (dev->real_num_tx_queues != 1) {
3676 const struct net_device_ops *ops = dev->netdev_ops;
3678 if (ops->ndo_select_queue)
3679 queue_index = ops->ndo_select_queue(dev, skb, sb_dev,
3682 queue_index = __netdev_pick_tx(dev, skb, sb_dev);
3684 queue_index = netdev_cap_txqueue(dev, queue_index);
3687 skb_set_queue_mapping(skb, queue_index);
3688 return netdev_get_tx_queue(dev, queue_index);
3692 * __dev_queue_xmit - transmit a buffer
3693 * @skb: buffer to transmit
3694 * @sb_dev: suboordinate device used for L2 forwarding offload
3696 * Queue a buffer for transmission to a network device. The caller must
3697 * have set the device and priority and built the buffer before calling
3698 * this function. The function can be called from an interrupt.
3700 * A negative errno code is returned on a failure. A success does not
3701 * guarantee the frame will be transmitted as it may be dropped due
3702 * to congestion or traffic shaping.
3704 * -----------------------------------------------------------------------------------
3705 * I notice this method can also return errors from the queue disciplines,
3706 * including NET_XMIT_DROP, which is a positive value. So, errors can also
3709 * Regardless of the return value, the skb is consumed, so it is currently
3710 * difficult to retry a send to this method. (You can bump the ref count
3711 * before sending to hold a reference for retry if you are careful.)
3713 * When calling this method, interrupts MUST be enabled. This is because
3714 * the BH enable code must have IRQs enabled so that it will not deadlock.
3717 static int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev)
3719 struct net_device *dev = skb->dev;
3720 struct netdev_queue *txq;
3725 skb_reset_mac_header(skb);
3727 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP))
3728 __skb_tstamp_tx(skb, NULL, skb->sk, SCM_TSTAMP_SCHED);
3730 /* Disable soft irqs for various locks below. Also
3731 * stops preemption for RCU.
3735 skb_update_prio(skb);
3737 qdisc_pkt_len_init(skb);
3738 #ifdef CONFIG_NET_CLS_ACT
3739 skb->tc_at_ingress = 0;
3740 # ifdef CONFIG_NET_EGRESS
3741 if (static_branch_unlikely(&egress_needed_key)) {
3742 skb = sch_handle_egress(skb, &rc, dev);
3748 /* If device/qdisc don't need skb->dst, release it right now while
3749 * its hot in this cpu cache.
3751 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
3756 txq = netdev_pick_tx(dev, skb, sb_dev);
3757 q = rcu_dereference_bh(txq->qdisc);
3759 trace_net_dev_queue(skb);
3761 rc = __dev_xmit_skb(skb, q, dev, txq);
3765 /* The device has no queue. Common case for software devices:
3766 * loopback, all the sorts of tunnels...
3768 * Really, it is unlikely that netif_tx_lock protection is necessary
3769 * here. (f.e. loopback and IP tunnels are clean ignoring statistics
3771 * However, it is possible, that they rely on protection
3774 * Check this and shot the lock. It is not prone from deadlocks.
3775 *Either shot noqueue qdisc, it is even simpler 8)
3777 if (dev->flags & IFF_UP) {
3778 int cpu = smp_processor_id(); /* ok because BHs are off */
3780 if (txq->xmit_lock_owner != cpu) {
3781 if (unlikely(__this_cpu_read(xmit_recursion) >
3782 XMIT_RECURSION_LIMIT))
3783 goto recursion_alert;
3785 skb = validate_xmit_skb(skb, dev, &again);
3789 HARD_TX_LOCK(dev, txq, cpu);
3791 if (!netif_xmit_stopped(txq)) {
3792 __this_cpu_inc(xmit_recursion);
3793 skb = dev_hard_start_xmit(skb, dev, txq, &rc);
3794 __this_cpu_dec(xmit_recursion);
3795 if (dev_xmit_complete(rc)) {
3796 HARD_TX_UNLOCK(dev, txq);
3800 HARD_TX_UNLOCK(dev, txq);
3801 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
3804 /* Recursion is detected! It is possible,
3808 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
3814 rcu_read_unlock_bh();
3816 atomic_long_inc(&dev->tx_dropped);
3817 kfree_skb_list(skb);
3820 rcu_read_unlock_bh();
3824 int dev_queue_xmit(struct sk_buff *skb)
3826 return __dev_queue_xmit(skb, NULL);
3828 EXPORT_SYMBOL(dev_queue_xmit);
3830 int dev_queue_xmit_accel(struct sk_buff *skb, struct net_device *sb_dev)
3832 return __dev_queue_xmit(skb, sb_dev);
3834 EXPORT_SYMBOL(dev_queue_xmit_accel);
3836 int dev_direct_xmit(struct sk_buff *skb, u16 queue_id)
3838 struct net_device *dev = skb->dev;
3839 struct sk_buff *orig_skb = skb;
3840 struct netdev_queue *txq;
3841 int ret = NETDEV_TX_BUSY;
3844 if (unlikely(!netif_running(dev) ||
3845 !netif_carrier_ok(dev)))
3848 skb = validate_xmit_skb_list(skb, dev, &again);
3849 if (skb != orig_skb)
3852 skb_set_queue_mapping(skb, queue_id);
3853 txq = skb_get_tx_queue(dev, skb);
3857 HARD_TX_LOCK(dev, txq, smp_processor_id());
3858 if (!netif_xmit_frozen_or_drv_stopped(txq))
3859 ret = netdev_start_xmit(skb, dev, txq, false);
3860 HARD_TX_UNLOCK(dev, txq);
3864 if (!dev_xmit_complete(ret))
3869 atomic_long_inc(&dev->tx_dropped);
3870 kfree_skb_list(skb);
3871 return NET_XMIT_DROP;
3873 EXPORT_SYMBOL(dev_direct_xmit);
3875 /*************************************************************************
3877 *************************************************************************/
3879 int netdev_max_backlog __read_mostly = 1000;
3880 EXPORT_SYMBOL(netdev_max_backlog);
3882 int netdev_tstamp_prequeue __read_mostly = 1;
3883 int netdev_budget __read_mostly = 300;
3884 unsigned int __read_mostly netdev_budget_usecs = 2000;
3885 int weight_p __read_mostly = 64; /* old backlog weight */
3886 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */
3887 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */
3888 int dev_rx_weight __read_mostly = 64;
3889 int dev_tx_weight __read_mostly = 64;
3891 /* Called with irq disabled */
3892 static inline void ____napi_schedule(struct softnet_data *sd,
3893 struct napi_struct *napi)
3895 list_add_tail(&napi->poll_list, &sd->poll_list);
3896 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
3901 /* One global table that all flow-based protocols share. */
3902 struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
3903 EXPORT_SYMBOL(rps_sock_flow_table);
3904 u32 rps_cpu_mask __read_mostly;
3905 EXPORT_SYMBOL(rps_cpu_mask);
3907 struct static_key rps_needed __read_mostly;
3908 EXPORT_SYMBOL(rps_needed);
3909 struct static_key rfs_needed __read_mostly;
3910 EXPORT_SYMBOL(rfs_needed);
3912 static struct rps_dev_flow *
3913 set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3914 struct rps_dev_flow *rflow, u16 next_cpu)
3916 if (next_cpu < nr_cpu_ids) {
3917 #ifdef CONFIG_RFS_ACCEL
3918 struct netdev_rx_queue *rxqueue;
3919 struct rps_dev_flow_table *flow_table;
3920 struct rps_dev_flow *old_rflow;
3925 /* Should we steer this flow to a different hardware queue? */
3926 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
3927 !(dev->features & NETIF_F_NTUPLE))
3929 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
3930 if (rxq_index == skb_get_rx_queue(skb))
3933 rxqueue = dev->_rx + rxq_index;
3934 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3937 flow_id = skb_get_hash(skb) & flow_table->mask;
3938 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
3939 rxq_index, flow_id);
3943 rflow = &flow_table->flows[flow_id];
3945 if (old_rflow->filter == rflow->filter)
3946 old_rflow->filter = RPS_NO_FILTER;
3950 per_cpu(softnet_data, next_cpu).input_queue_head;
3953 rflow->cpu = next_cpu;
3958 * get_rps_cpu is called from netif_receive_skb and returns the target
3959 * CPU from the RPS map of the receiving queue for a given skb.
3960 * rcu_read_lock must be held on entry.
3962 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
3963 struct rps_dev_flow **rflowp)
3965 const struct rps_sock_flow_table *sock_flow_table;
3966 struct netdev_rx_queue *rxqueue = dev->_rx;
3967 struct rps_dev_flow_table *flow_table;
3968 struct rps_map *map;
3973 if (skb_rx_queue_recorded(skb)) {
3974 u16 index = skb_get_rx_queue(skb);
3976 if (unlikely(index >= dev->real_num_rx_queues)) {
3977 WARN_ONCE(dev->real_num_rx_queues > 1,
3978 "%s received packet on queue %u, but number "
3979 "of RX queues is %u\n",
3980 dev->name, index, dev->real_num_rx_queues);
3986 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */
3988 flow_table = rcu_dereference(rxqueue->rps_flow_table);
3989 map = rcu_dereference(rxqueue->rps_map);
3990 if (!flow_table && !map)
3993 skb_reset_network_header(skb);
3994 hash = skb_get_hash(skb);
3998 sock_flow_table = rcu_dereference(rps_sock_flow_table);
3999 if (flow_table && sock_flow_table) {
4000 struct rps_dev_flow *rflow;
4004 /* First check into global flow table if there is a match */
4005 ident = sock_flow_table->ents[hash & sock_flow_table->mask];
4006 if ((ident ^ hash) & ~rps_cpu_mask)
4009 next_cpu = ident & rps_cpu_mask;
4011 /* OK, now we know there is a match,
4012 * we can look at the local (per receive queue) flow table
4014 rflow = &flow_table->flows[hash & flow_table->mask];
4018 * If the desired CPU (where last recvmsg was done) is
4019 * different from current CPU (one in the rx-queue flow
4020 * table entry), switch if one of the following holds:
4021 * - Current CPU is unset (>= nr_cpu_ids).
4022 * - Current CPU is offline.
4023 * - The current CPU's queue tail has advanced beyond the
4024 * last packet that was enqueued using this table entry.
4025 * This guarantees that all previous packets for the flow
4026 * have been dequeued, thus preserving in order delivery.
4028 if (unlikely(tcpu != next_cpu) &&
4029 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) ||
4030 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
4031 rflow->last_qtail)) >= 0)) {
4033 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
4036 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) {
4046 tcpu = map->cpus[reciprocal_scale(hash, map->len)];
4047 if (cpu_online(tcpu)) {
4057 #ifdef CONFIG_RFS_ACCEL
4060 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
4061 * @dev: Device on which the filter was set
4062 * @rxq_index: RX queue index
4063 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
4064 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
4066 * Drivers that implement ndo_rx_flow_steer() should periodically call
4067 * this function for each installed filter and remove the filters for
4068 * which it returns %true.
4070 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
4071 u32 flow_id, u16 filter_id)
4073 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
4074 struct rps_dev_flow_table *flow_table;
4075 struct rps_dev_flow *rflow;
4080 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4081 if (flow_table && flow_id <= flow_table->mask) {
4082 rflow = &flow_table->flows[flow_id];
4083 cpu = READ_ONCE(rflow->cpu);
4084 if (rflow->filter == filter_id && cpu < nr_cpu_ids &&
4085 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
4086 rflow->last_qtail) <
4087 (int)(10 * flow_table->mask)))
4093 EXPORT_SYMBOL(rps_may_expire_flow);
4095 #endif /* CONFIG_RFS_ACCEL */
4097 /* Called from hardirq (IPI) context */
4098 static void rps_trigger_softirq(void *data)
4100 struct softnet_data *sd = data;
4102 ____napi_schedule(sd, &sd->backlog);
4106 #endif /* CONFIG_RPS */
4109 * Check if this softnet_data structure is another cpu one
4110 * If yes, queue it to our IPI list and return 1
4113 static int rps_ipi_queued(struct softnet_data *sd)
4116 struct softnet_data *mysd = this_cpu_ptr(&softnet_data);
4119 sd->rps_ipi_next = mysd->rps_ipi_list;
4120 mysd->rps_ipi_list = sd;
4122 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
4125 #endif /* CONFIG_RPS */
4129 #ifdef CONFIG_NET_FLOW_LIMIT
4130 int netdev_flow_limit_table_len __read_mostly = (1 << 12);
4133 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
4135 #ifdef CONFIG_NET_FLOW_LIMIT
4136 struct sd_flow_limit *fl;
4137 struct softnet_data *sd;
4138 unsigned int old_flow, new_flow;
4140 if (qlen < (netdev_max_backlog >> 1))
4143 sd = this_cpu_ptr(&softnet_data);
4146 fl = rcu_dereference(sd->flow_limit);
4148 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1);
4149 old_flow = fl->history[fl->history_head];
4150 fl->history[fl->history_head] = new_flow;
4153 fl->history_head &= FLOW_LIMIT_HISTORY - 1;
4155 if (likely(fl->buckets[old_flow]))
4156 fl->buckets[old_flow]--;
4158 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) {
4170 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
4171 * queue (may be a remote CPU queue).
4173 static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
4174 unsigned int *qtail)
4176 struct softnet_data *sd;
4177 unsigned long flags;
4180 sd = &per_cpu(softnet_data, cpu);
4182 local_irq_save(flags);
4185 if (!netif_running(skb->dev))
4187 qlen = skb_queue_len(&sd->input_pkt_queue);
4188 if (qlen <= netdev_max_backlog && !skb_flow_limit(skb, qlen)) {
4191 __skb_queue_tail(&sd->input_pkt_queue, skb);
4192 input_queue_tail_incr_save(sd, qtail);
4194 local_irq_restore(flags);
4195 return NET_RX_SUCCESS;
4198 /* Schedule NAPI for backlog device
4199 * We can use non atomic operation since we own the queue lock
4201 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
4202 if (!rps_ipi_queued(sd))
4203 ____napi_schedule(sd, &sd->backlog);
4212 local_irq_restore(flags);
4214 atomic_long_inc(&skb->dev->rx_dropped);
4219 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb)
4221 struct net_device *dev = skb->dev;
4222 struct netdev_rx_queue *rxqueue;
4226 if (skb_rx_queue_recorded(skb)) {
4227 u16 index = skb_get_rx_queue(skb);
4229 if (unlikely(index >= dev->real_num_rx_queues)) {
4230 WARN_ONCE(dev->real_num_rx_queues > 1,
4231 "%s received packet on queue %u, but number "
4232 "of RX queues is %u\n",
4233 dev->name, index, dev->real_num_rx_queues);
4235 return rxqueue; /* Return first rxqueue */
4242 static u32 netif_receive_generic_xdp(struct sk_buff *skb,
4243 struct xdp_buff *xdp,
4244 struct bpf_prog *xdp_prog)
4246 struct netdev_rx_queue *rxqueue;
4247 void *orig_data, *orig_data_end;
4248 u32 metalen, act = XDP_DROP;
4252 /* Reinjected packets coming from act_mirred or similar should
4253 * not get XDP generic processing.
4255 if (skb_cloned(skb))
4258 /* XDP packets must be linear and must have sufficient headroom
4259 * of XDP_PACKET_HEADROOM bytes. This is the guarantee that also
4260 * native XDP provides, thus we need to do it here as well.
4262 if (skb_is_nonlinear(skb) ||
4263 skb_headroom(skb) < XDP_PACKET_HEADROOM) {
4264 int hroom = XDP_PACKET_HEADROOM - skb_headroom(skb);
4265 int troom = skb->tail + skb->data_len - skb->end;
4267 /* In case we have to go down the path and also linearize,
4268 * then lets do the pskb_expand_head() work just once here.
4270 if (pskb_expand_head(skb,
4271 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0,
4272 troom > 0 ? troom + 128 : 0, GFP_ATOMIC))
4274 if (skb_linearize(skb))
4278 /* The XDP program wants to see the packet starting at the MAC
4281 mac_len = skb->data - skb_mac_header(skb);
4282 hlen = skb_headlen(skb) + mac_len;
4283 xdp->data = skb->data - mac_len;
4284 xdp->data_meta = xdp->data;
4285 xdp->data_end = xdp->data + hlen;
4286 xdp->data_hard_start = skb->data - skb_headroom(skb);
4287 orig_data_end = xdp->data_end;
4288 orig_data = xdp->data;
4290 rxqueue = netif_get_rxqueue(skb);
4291 xdp->rxq = &rxqueue->xdp_rxq;
4293 act = bpf_prog_run_xdp(xdp_prog, xdp);
4295 off = xdp->data - orig_data;
4297 __skb_pull(skb, off);
4299 __skb_push(skb, -off);
4300 skb->mac_header += off;
4302 /* check if bpf_xdp_adjust_tail was used. it can only "shrink"
4305 off = orig_data_end - xdp->data_end;
4307 skb_set_tail_pointer(skb, xdp->data_end - xdp->data);
4315 __skb_push(skb, mac_len);
4318 metalen = xdp->data - xdp->data_meta;
4320 skb_metadata_set(skb, metalen);
4323 bpf_warn_invalid_xdp_action(act);
4326 trace_xdp_exception(skb->dev, xdp_prog, act);
4337 /* When doing generic XDP we have to bypass the qdisc layer and the
4338 * network taps in order to match in-driver-XDP behavior.
4340 void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog)
4342 struct net_device *dev = skb->dev;
4343 struct netdev_queue *txq;
4344 bool free_skb = true;
4347 txq = netdev_pick_tx(dev, skb, NULL);
4348 cpu = smp_processor_id();
4349 HARD_TX_LOCK(dev, txq, cpu);
4350 if (!netif_xmit_stopped(txq)) {
4351 rc = netdev_start_xmit(skb, dev, txq, 0);
4352 if (dev_xmit_complete(rc))
4355 HARD_TX_UNLOCK(dev, txq);
4357 trace_xdp_exception(dev, xdp_prog, XDP_TX);
4361 EXPORT_SYMBOL_GPL(generic_xdp_tx);
4363 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key);
4365 int do_xdp_generic(struct bpf_prog *xdp_prog, struct sk_buff *skb)
4368 struct xdp_buff xdp;
4372 act = netif_receive_generic_xdp(skb, &xdp, xdp_prog);
4373 if (act != XDP_PASS) {
4376 err = xdp_do_generic_redirect(skb->dev, skb,
4382 generic_xdp_tx(skb, xdp_prog);
4393 EXPORT_SYMBOL_GPL(do_xdp_generic);
4395 static int netif_rx_internal(struct sk_buff *skb)
4399 net_timestamp_check(netdev_tstamp_prequeue, skb);
4401 trace_netif_rx(skb);
4403 if (static_branch_unlikely(&generic_xdp_needed_key)) {
4408 ret = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
4412 /* Consider XDP consuming the packet a success from
4413 * the netdev point of view we do not want to count
4416 if (ret != XDP_PASS)
4417 return NET_RX_SUCCESS;
4421 if (static_key_false(&rps_needed)) {
4422 struct rps_dev_flow voidflow, *rflow = &voidflow;
4428 cpu = get_rps_cpu(skb->dev, skb, &rflow);
4430 cpu = smp_processor_id();
4432 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4441 ret = enqueue_to_backlog(skb, get_cpu(), &qtail);
4448 * netif_rx - post buffer to the network code
4449 * @skb: buffer to post
4451 * This function receives a packet from a device driver and queues it for
4452 * the upper (protocol) levels to process. It always succeeds. The buffer
4453 * may be dropped during processing for congestion control or by the
4457 * NET_RX_SUCCESS (no congestion)
4458 * NET_RX_DROP (packet was dropped)
4462 int netif_rx(struct sk_buff *skb)
4464 trace_netif_rx_entry(skb);
4466 return netif_rx_internal(skb);
4468 EXPORT_SYMBOL(netif_rx);
4470 int netif_rx_ni(struct sk_buff *skb)
4474 trace_netif_rx_ni_entry(skb);
4477 err = netif_rx_internal(skb);
4478 if (local_softirq_pending())
4484 EXPORT_SYMBOL(netif_rx_ni);
4486 static __latent_entropy void net_tx_action(struct softirq_action *h)
4488 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
4490 if (sd->completion_queue) {
4491 struct sk_buff *clist;
4493 local_irq_disable();
4494 clist = sd->completion_queue;
4495 sd->completion_queue = NULL;
4499 struct sk_buff *skb = clist;
4501 clist = clist->next;
4503 WARN_ON(refcount_read(&skb->users));
4504 if (likely(get_kfree_skb_cb(skb)->reason == SKB_REASON_CONSUMED))
4505 trace_consume_skb(skb);
4507 trace_kfree_skb(skb, net_tx_action);
4509 if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
4512 __kfree_skb_defer(skb);
4515 __kfree_skb_flush();
4518 if (sd->output_queue) {
4521 local_irq_disable();
4522 head = sd->output_queue;
4523 sd->output_queue = NULL;
4524 sd->output_queue_tailp = &sd->output_queue;
4528 struct Qdisc *q = head;
4529 spinlock_t *root_lock = NULL;
4531 head = head->next_sched;
4533 if (!(q->flags & TCQ_F_NOLOCK)) {
4534 root_lock = qdisc_lock(q);
4535 spin_lock(root_lock);
4537 /* We need to make sure head->next_sched is read
4538 * before clearing __QDISC_STATE_SCHED
4540 smp_mb__before_atomic();
4541 clear_bit(__QDISC_STATE_SCHED, &q->state);
4544 spin_unlock(root_lock);
4548 xfrm_dev_backlog(sd);
4551 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE)
4552 /* This hook is defined here for ATM LANE */
4553 int (*br_fdb_test_addr_hook)(struct net_device *dev,
4554 unsigned char *addr) __read_mostly;
4555 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook);
4558 static inline struct sk_buff *
4559 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
4560 struct net_device *orig_dev)
4562 #ifdef CONFIG_NET_CLS_ACT
4563 struct mini_Qdisc *miniq = rcu_dereference_bh(skb->dev->miniq_ingress);
4564 struct tcf_result cl_res;
4566 /* If there's at least one ingress present somewhere (so
4567 * we get here via enabled static key), remaining devices
4568 * that are not configured with an ingress qdisc will bail
4575 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4579 qdisc_skb_cb(skb)->pkt_len = skb->len;
4580 skb->tc_at_ingress = 1;
4581 mini_qdisc_bstats_cpu_update(miniq, skb);
4583 switch (tcf_classify(skb, miniq->filter_list, &cl_res, false)) {
4585 case TC_ACT_RECLASSIFY:
4586 skb->tc_index = TC_H_MIN(cl_res.classid);
4589 mini_qdisc_qstats_cpu_drop(miniq);
4597 case TC_ACT_REDIRECT:
4598 /* skb_mac_header check was done by cls/act_bpf, so
4599 * we can safely push the L2 header back before
4600 * redirecting to another netdev
4602 __skb_push(skb, skb->mac_len);
4603 skb_do_redirect(skb);
4608 #endif /* CONFIG_NET_CLS_ACT */
4613 * netdev_is_rx_handler_busy - check if receive handler is registered
4614 * @dev: device to check
4616 * Check if a receive handler is already registered for a given device.
4617 * Return true if there one.
4619 * The caller must hold the rtnl_mutex.
4621 bool netdev_is_rx_handler_busy(struct net_device *dev)
4624 return dev && rtnl_dereference(dev->rx_handler);
4626 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy);
4629 * netdev_rx_handler_register - register receive handler
4630 * @dev: device to register a handler for
4631 * @rx_handler: receive handler to register
4632 * @rx_handler_data: data pointer that is used by rx handler
4634 * Register a receive handler for a device. This handler will then be
4635 * called from __netif_receive_skb. A negative errno code is returned
4638 * The caller must hold the rtnl_mutex.
4640 * For a general description of rx_handler, see enum rx_handler_result.
4642 int netdev_rx_handler_register(struct net_device *dev,
4643 rx_handler_func_t *rx_handler,
4644 void *rx_handler_data)
4646 if (netdev_is_rx_handler_busy(dev))
4649 if (dev->priv_flags & IFF_NO_RX_HANDLER)
4652 /* Note: rx_handler_data must be set before rx_handler */
4653 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data);
4654 rcu_assign_pointer(dev->rx_handler, rx_handler);
4658 EXPORT_SYMBOL_GPL(netdev_rx_handler_register);
4661 * netdev_rx_handler_unregister - unregister receive handler
4662 * @dev: device to unregister a handler from
4664 * Unregister a receive handler from a device.
4666 * The caller must hold the rtnl_mutex.
4668 void netdev_rx_handler_unregister(struct net_device *dev)
4672 RCU_INIT_POINTER(dev->rx_handler, NULL);
4673 /* a reader seeing a non NULL rx_handler in a rcu_read_lock()
4674 * section has a guarantee to see a non NULL rx_handler_data
4678 RCU_INIT_POINTER(dev->rx_handler_data, NULL);
4680 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister);
4683 * Limit the use of PFMEMALLOC reserves to those protocols that implement
4684 * the special handling of PFMEMALLOC skbs.
4686 static bool skb_pfmemalloc_protocol(struct sk_buff *skb)
4688 switch (skb->protocol) {
4689 case htons(ETH_P_ARP):
4690 case htons(ETH_P_IP):
4691 case htons(ETH_P_IPV6):
4692 case htons(ETH_P_8021Q):
4693 case htons(ETH_P_8021AD):
4700 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev,
4701 int *ret, struct net_device *orig_dev)
4703 #ifdef CONFIG_NETFILTER_INGRESS
4704 if (nf_hook_ingress_active(skb)) {
4708 *ret = deliver_skb(skb, *pt_prev, orig_dev);
4713 ingress_retval = nf_hook_ingress(skb);
4715 return ingress_retval;
4717 #endif /* CONFIG_NETFILTER_INGRESS */
4721 static int __netif_receive_skb_core(struct sk_buff *skb, bool pfmemalloc,
4722 struct packet_type **ppt_prev)
4724 struct packet_type *ptype, *pt_prev;
4725 rx_handler_func_t *rx_handler;
4726 struct net_device *orig_dev;
4727 bool deliver_exact = false;
4728 int ret = NET_RX_DROP;
4731 net_timestamp_check(!netdev_tstamp_prequeue, skb);
4733 trace_netif_receive_skb(skb);
4735 orig_dev = skb->dev;
4737 skb_reset_network_header(skb);
4738 if (!skb_transport_header_was_set(skb))
4739 skb_reset_transport_header(skb);
4740 skb_reset_mac_len(skb);
4745 skb->skb_iif = skb->dev->ifindex;
4747 __this_cpu_inc(softnet_data.processed);
4749 if (skb->protocol == cpu_to_be16(ETH_P_8021Q) ||
4750 skb->protocol == cpu_to_be16(ETH_P_8021AD)) {
4751 skb = skb_vlan_untag(skb);
4756 if (skb_skip_tc_classify(skb))
4762 list_for_each_entry_rcu(ptype, &ptype_all, list) {
4764 ret = deliver_skb(skb, pt_prev, orig_dev);
4768 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) {
4770 ret = deliver_skb(skb, pt_prev, orig_dev);
4775 #ifdef CONFIG_NET_INGRESS
4776 if (static_branch_unlikely(&ingress_needed_key)) {
4777 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev);
4781 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0)
4787 if (pfmemalloc && !skb_pfmemalloc_protocol(skb))
4790 if (skb_vlan_tag_present(skb)) {
4792 ret = deliver_skb(skb, pt_prev, orig_dev);
4795 if (vlan_do_receive(&skb))
4797 else if (unlikely(!skb))
4801 rx_handler = rcu_dereference(skb->dev->rx_handler);
4804 ret = deliver_skb(skb, pt_prev, orig_dev);
4807 switch (rx_handler(&skb)) {
4808 case RX_HANDLER_CONSUMED:
4809 ret = NET_RX_SUCCESS;
4811 case RX_HANDLER_ANOTHER:
4813 case RX_HANDLER_EXACT:
4814 deliver_exact = true;
4815 case RX_HANDLER_PASS:
4822 if (unlikely(skb_vlan_tag_present(skb))) {
4823 if (skb_vlan_tag_get_id(skb))
4824 skb->pkt_type = PACKET_OTHERHOST;
4825 /* Note: we might in the future use prio bits
4826 * and set skb->priority like in vlan_do_receive()
4827 * For the time being, just ignore Priority Code Point
4832 type = skb->protocol;
4834 /* deliver only exact match when indicated */
4835 if (likely(!deliver_exact)) {
4836 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4837 &ptype_base[ntohs(type) &
4841 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4842 &orig_dev->ptype_specific);
4844 if (unlikely(skb->dev != orig_dev)) {
4845 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
4846 &skb->dev->ptype_specific);
4850 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
4852 *ppt_prev = pt_prev;
4856 atomic_long_inc(&skb->dev->rx_dropped);
4858 atomic_long_inc(&skb->dev->rx_nohandler);
4860 /* Jamal, now you will not able to escape explaining
4861 * me how you were going to use this. :-)
4870 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc)
4872 struct net_device *orig_dev = skb->dev;
4873 struct packet_type *pt_prev = NULL;
4876 ret = __netif_receive_skb_core(skb, pfmemalloc, &pt_prev);
4878 ret = pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
4883 * netif_receive_skb_core - special purpose version of netif_receive_skb
4884 * @skb: buffer to process
4886 * More direct receive version of netif_receive_skb(). It should
4887 * only be used by callers that have a need to skip RPS and Generic XDP.
4888 * Caller must also take care of handling if (page_is_)pfmemalloc.
4890 * This function may only be called from softirq context and interrupts
4891 * should be enabled.
4893 * Return values (usually ignored):
4894 * NET_RX_SUCCESS: no congestion
4895 * NET_RX_DROP: packet was dropped
4897 int netif_receive_skb_core(struct sk_buff *skb)
4902 ret = __netif_receive_skb_one_core(skb, false);
4907 EXPORT_SYMBOL(netif_receive_skb_core);
4909 static inline void __netif_receive_skb_list_ptype(struct list_head *head,
4910 struct packet_type *pt_prev,
4911 struct net_device *orig_dev)
4913 struct sk_buff *skb, *next;
4917 if (list_empty(head))
4919 if (pt_prev->list_func != NULL)
4920 pt_prev->list_func(head, pt_prev, orig_dev);
4922 list_for_each_entry_safe(skb, next, head, list)
4923 pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
4926 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc)
4928 /* Fast-path assumptions:
4929 * - There is no RX handler.
4930 * - Only one packet_type matches.
4931 * If either of these fails, we will end up doing some per-packet
4932 * processing in-line, then handling the 'last ptype' for the whole
4933 * sublist. This can't cause out-of-order delivery to any single ptype,
4934 * because the 'last ptype' must be constant across the sublist, and all
4935 * other ptypes are handled per-packet.
4937 /* Current (common) ptype of sublist */
4938 struct packet_type *pt_curr = NULL;
4939 /* Current (common) orig_dev of sublist */
4940 struct net_device *od_curr = NULL;
4941 struct list_head sublist;
4942 struct sk_buff *skb, *next;
4944 INIT_LIST_HEAD(&sublist);
4945 list_for_each_entry_safe(skb, next, head, list) {
4946 struct net_device *orig_dev = skb->dev;
4947 struct packet_type *pt_prev = NULL;
4949 list_del(&skb->list);
4950 __netif_receive_skb_core(skb, pfmemalloc, &pt_prev);
4953 if (pt_curr != pt_prev || od_curr != orig_dev) {
4954 /* dispatch old sublist */
4955 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr);
4956 /* start new sublist */
4957 INIT_LIST_HEAD(&sublist);
4961 list_add_tail(&skb->list, &sublist);
4964 /* dispatch final sublist */
4965 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr);
4968 static int __netif_receive_skb(struct sk_buff *skb)
4972 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) {
4973 unsigned int noreclaim_flag;
4976 * PFMEMALLOC skbs are special, they should
4977 * - be delivered to SOCK_MEMALLOC sockets only
4978 * - stay away from userspace
4979 * - have bounded memory usage
4981 * Use PF_MEMALLOC as this saves us from propagating the allocation
4982 * context down to all allocation sites.
4984 noreclaim_flag = memalloc_noreclaim_save();
4985 ret = __netif_receive_skb_one_core(skb, true);
4986 memalloc_noreclaim_restore(noreclaim_flag);
4988 ret = __netif_receive_skb_one_core(skb, false);
4993 static void __netif_receive_skb_list(struct list_head *head)
4995 unsigned long noreclaim_flag = 0;
4996 struct sk_buff *skb, *next;
4997 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */
4999 list_for_each_entry_safe(skb, next, head, list) {
5000 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) {
5001 struct list_head sublist;
5003 /* Handle the previous sublist */
5004 list_cut_before(&sublist, head, &skb->list);
5005 if (!list_empty(&sublist))
5006 __netif_receive_skb_list_core(&sublist, pfmemalloc);
5007 pfmemalloc = !pfmemalloc;
5008 /* See comments in __netif_receive_skb */
5010 noreclaim_flag = memalloc_noreclaim_save();
5012 memalloc_noreclaim_restore(noreclaim_flag);
5015 /* Handle the remaining sublist */
5016 if (!list_empty(head))
5017 __netif_receive_skb_list_core(head, pfmemalloc);
5018 /* Restore pflags */
5020 memalloc_noreclaim_restore(noreclaim_flag);
5023 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp)
5025 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog);
5026 struct bpf_prog *new = xdp->prog;
5029 switch (xdp->command) {
5030 case XDP_SETUP_PROG:
5031 rcu_assign_pointer(dev->xdp_prog, new);
5036 static_branch_dec(&generic_xdp_needed_key);
5037 } else if (new && !old) {
5038 static_branch_inc(&generic_xdp_needed_key);
5039 dev_disable_lro(dev);
5040 dev_disable_gro_hw(dev);
5044 case XDP_QUERY_PROG:
5045 xdp->prog_id = old ? old->aux->id : 0;
5056 static int netif_receive_skb_internal(struct sk_buff *skb)
5060 net_timestamp_check(netdev_tstamp_prequeue, skb);
5062 if (skb_defer_rx_timestamp(skb))
5063 return NET_RX_SUCCESS;
5065 if (static_branch_unlikely(&generic_xdp_needed_key)) {
5070 ret = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
5074 if (ret != XDP_PASS)
5080 if (static_key_false(&rps_needed)) {
5081 struct rps_dev_flow voidflow, *rflow = &voidflow;
5082 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
5085 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
5091 ret = __netif_receive_skb(skb);
5096 static void netif_receive_skb_list_internal(struct list_head *head)
5098 struct bpf_prog *xdp_prog = NULL;
5099 struct sk_buff *skb, *next;
5100 struct list_head sublist;
5102 INIT_LIST_HEAD(&sublist);
5103 list_for_each_entry_safe(skb, next, head, list) {
5104 net_timestamp_check(netdev_tstamp_prequeue, skb);
5105 list_del(&skb->list);
5106 if (!skb_defer_rx_timestamp(skb))
5107 list_add_tail(&skb->list, &sublist);
5109 list_splice_init(&sublist, head);
5111 if (static_branch_unlikely(&generic_xdp_needed_key)) {
5114 list_for_each_entry_safe(skb, next, head, list) {
5115 xdp_prog = rcu_dereference(skb->dev->xdp_prog);
5116 list_del(&skb->list);
5117 if (do_xdp_generic(xdp_prog, skb) == XDP_PASS)
5118 list_add_tail(&skb->list, &sublist);
5122 /* Put passed packets back on main list */
5123 list_splice_init(&sublist, head);
5128 if (static_key_false(&rps_needed)) {
5129 list_for_each_entry_safe(skb, next, head, list) {
5130 struct rps_dev_flow voidflow, *rflow = &voidflow;
5131 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
5134 /* Will be handled, remove from list */
5135 list_del(&skb->list);
5136 enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
5141 __netif_receive_skb_list(head);
5146 * netif_receive_skb - process receive buffer from network
5147 * @skb: buffer to process
5149 * netif_receive_skb() is the main receive data processing function.
5150 * It always succeeds. The buffer may be dropped during processing
5151 * for congestion control or by the protocol layers.
5153 * This function may only be called from softirq context and interrupts
5154 * should be enabled.
5156 * Return values (usually ignored):
5157 * NET_RX_SUCCESS: no congestion
5158 * NET_RX_DROP: packet was dropped
5160 int netif_receive_skb(struct sk_buff *skb)
5162 trace_netif_receive_skb_entry(skb);
5164 return netif_receive_skb_internal(skb);
5166 EXPORT_SYMBOL(netif_receive_skb);
5169 * netif_receive_skb_list - process many receive buffers from network
5170 * @head: list of skbs to process.
5172 * Since return value of netif_receive_skb() is normally ignored, and
5173 * wouldn't be meaningful for a list, this function returns void.
5175 * This function may only be called from softirq context and interrupts
5176 * should be enabled.
5178 void netif_receive_skb_list(struct list_head *head)
5180 struct sk_buff *skb;
5182 if (list_empty(head))
5184 list_for_each_entry(skb, head, list)
5185 trace_netif_receive_skb_list_entry(skb);
5186 netif_receive_skb_list_internal(head);
5188 EXPORT_SYMBOL(netif_receive_skb_list);
5190 DEFINE_PER_CPU(struct work_struct, flush_works);
5192 /* Network device is going away, flush any packets still pending */
5193 static void flush_backlog(struct work_struct *work)
5195 struct sk_buff *skb, *tmp;
5196 struct softnet_data *sd;
5199 sd = this_cpu_ptr(&softnet_data);
5201 local_irq_disable();
5203 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) {
5204 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
5205 __skb_unlink(skb, &sd->input_pkt_queue);
5207 input_queue_head_incr(sd);
5213 skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
5214 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
5215 __skb_unlink(skb, &sd->process_queue);
5217 input_queue_head_incr(sd);
5223 static void flush_all_backlogs(void)
5229 for_each_online_cpu(cpu)
5230 queue_work_on(cpu, system_highpri_wq,
5231 per_cpu_ptr(&flush_works, cpu));
5233 for_each_online_cpu(cpu)
5234 flush_work(per_cpu_ptr(&flush_works, cpu));
5239 static int napi_gro_complete(struct sk_buff *skb)
5241 struct packet_offload *ptype;
5242 __be16 type = skb->protocol;
5243 struct list_head *head = &offload_base;
5246 BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
5248 if (NAPI_GRO_CB(skb)->count == 1) {
5249 skb_shinfo(skb)->gso_size = 0;
5254 list_for_each_entry_rcu(ptype, head, list) {
5255 if (ptype->type != type || !ptype->callbacks.gro_complete)
5258 err = ptype->callbacks.gro_complete(skb, 0);
5264 WARN_ON(&ptype->list == head);
5266 return NET_RX_SUCCESS;
5270 return netif_receive_skb_internal(skb);
5273 static void __napi_gro_flush_chain(struct napi_struct *napi, u32 index,
5276 struct list_head *head = &napi->gro_hash[index].list;
5277 struct sk_buff *skb, *p;
5279 list_for_each_entry_safe_reverse(skb, p, head, list) {
5280 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
5282 list_del(&skb->list);
5284 napi_gro_complete(skb);
5285 napi->gro_hash[index].count--;
5288 if (!napi->gro_hash[index].count)
5289 __clear_bit(index, &napi->gro_bitmask);
5292 /* napi->gro_hash[].list contains packets ordered by age.
5293 * youngest packets at the head of it.
5294 * Complete skbs in reverse order to reduce latencies.
5296 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
5300 for (i = 0; i < GRO_HASH_BUCKETS; i++) {
5301 if (test_bit(i, &napi->gro_bitmask))
5302 __napi_gro_flush_chain(napi, i, flush_old);
5305 EXPORT_SYMBOL(napi_gro_flush);
5307 static struct list_head *gro_list_prepare(struct napi_struct *napi,
5308 struct sk_buff *skb)
5310 unsigned int maclen = skb->dev->hard_header_len;
5311 u32 hash = skb_get_hash_raw(skb);
5312 struct list_head *head;
5315 head = &napi->gro_hash[hash & (GRO_HASH_BUCKETS - 1)].list;
5316 list_for_each_entry(p, head, list) {
5317 unsigned long diffs;
5319 NAPI_GRO_CB(p)->flush = 0;
5321 if (hash != skb_get_hash_raw(p)) {
5322 NAPI_GRO_CB(p)->same_flow = 0;
5326 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
5327 diffs |= p->vlan_tci ^ skb->vlan_tci;
5328 diffs |= skb_metadata_dst_cmp(p, skb);
5329 diffs |= skb_metadata_differs(p, skb);
5330 if (maclen == ETH_HLEN)
5331 diffs |= compare_ether_header(skb_mac_header(p),
5332 skb_mac_header(skb));
5334 diffs = memcmp(skb_mac_header(p),
5335 skb_mac_header(skb),
5337 NAPI_GRO_CB(p)->same_flow = !diffs;
5343 static void skb_gro_reset_offset(struct sk_buff *skb)
5345 const struct skb_shared_info *pinfo = skb_shinfo(skb);
5346 const skb_frag_t *frag0 = &pinfo->frags[0];
5348 NAPI_GRO_CB(skb)->data_offset = 0;
5349 NAPI_GRO_CB(skb)->frag0 = NULL;
5350 NAPI_GRO_CB(skb)->frag0_len = 0;
5352 if (skb_mac_header(skb) == skb_tail_pointer(skb) &&
5354 !PageHighMem(skb_frag_page(frag0))) {
5355 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
5356 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
5357 skb_frag_size(frag0),
5358 skb->end - skb->tail);
5362 static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
5364 struct skb_shared_info *pinfo = skb_shinfo(skb);
5366 BUG_ON(skb->end - skb->tail < grow);
5368 memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
5370 skb->data_len -= grow;
5373 pinfo->frags[0].page_offset += grow;
5374 skb_frag_size_sub(&pinfo->frags[0], grow);
5376 if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
5377 skb_frag_unref(skb, 0);
5378 memmove(pinfo->frags, pinfo->frags + 1,
5379 --pinfo->nr_frags * sizeof(pinfo->frags[0]));
5383 static void gro_flush_oldest(struct list_head *head)
5385 struct sk_buff *oldest;
5387 oldest = list_last_entry(head, struct sk_buff, list);
5389 /* We are called with head length >= MAX_GRO_SKBS, so this is
5392 if (WARN_ON_ONCE(!oldest))
5395 /* Do not adjust napi->gro_hash[].count, caller is adding a new
5398 list_del(&oldest->list);
5399 napi_gro_complete(oldest);
5402 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
5404 u32 hash = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1);
5405 struct list_head *head = &offload_base;
5406 struct packet_offload *ptype;
5407 __be16 type = skb->protocol;
5408 struct list_head *gro_head;
5409 struct sk_buff *pp = NULL;
5410 enum gro_result ret;
5414 if (netif_elide_gro(skb->dev))
5417 gro_head = gro_list_prepare(napi, skb);
5420 list_for_each_entry_rcu(ptype, head, list) {
5421 if (ptype->type != type || !ptype->callbacks.gro_receive)
5424 skb_set_network_header(skb, skb_gro_offset(skb));
5425 skb_reset_mac_len(skb);
5426 NAPI_GRO_CB(skb)->same_flow = 0;
5427 NAPI_GRO_CB(skb)->flush = skb_is_gso(skb) || skb_has_frag_list(skb);
5428 NAPI_GRO_CB(skb)->free = 0;
5429 NAPI_GRO_CB(skb)->encap_mark = 0;
5430 NAPI_GRO_CB(skb)->recursion_counter = 0;
5431 NAPI_GRO_CB(skb)->is_fou = 0;
5432 NAPI_GRO_CB(skb)->is_atomic = 1;
5433 NAPI_GRO_CB(skb)->gro_remcsum_start = 0;
5435 /* Setup for GRO checksum validation */
5436 switch (skb->ip_summed) {
5437 case CHECKSUM_COMPLETE:
5438 NAPI_GRO_CB(skb)->csum = skb->csum;
5439 NAPI_GRO_CB(skb)->csum_valid = 1;
5440 NAPI_GRO_CB(skb)->csum_cnt = 0;
5442 case CHECKSUM_UNNECESSARY:
5443 NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
5444 NAPI_GRO_CB(skb)->csum_valid = 0;
5447 NAPI_GRO_CB(skb)->csum_cnt = 0;
5448 NAPI_GRO_CB(skb)->csum_valid = 0;
5451 pp = ptype->callbacks.gro_receive(gro_head, skb);
5456 if (&ptype->list == head)
5459 if (IS_ERR(pp) && PTR_ERR(pp) == -EINPROGRESS) {
5464 same_flow = NAPI_GRO_CB(skb)->same_flow;
5465 ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
5468 list_del(&pp->list);
5470 napi_gro_complete(pp);
5471 napi->gro_hash[hash].count--;
5477 if (NAPI_GRO_CB(skb)->flush)
5480 if (unlikely(napi->gro_hash[hash].count >= MAX_GRO_SKBS)) {
5481 gro_flush_oldest(gro_head);
5483 napi->gro_hash[hash].count++;
5485 NAPI_GRO_CB(skb)->count = 1;
5486 NAPI_GRO_CB(skb)->age = jiffies;
5487 NAPI_GRO_CB(skb)->last = skb;
5488 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
5489 list_add(&skb->list, gro_head);
5493 grow = skb_gro_offset(skb) - skb_headlen(skb);
5495 gro_pull_from_frag0(skb, grow);
5497 if (napi->gro_hash[hash].count) {
5498 if (!test_bit(hash, &napi->gro_bitmask))
5499 __set_bit(hash, &napi->gro_bitmask);
5500 } else if (test_bit(hash, &napi->gro_bitmask)) {
5501 __clear_bit(hash, &napi->gro_bitmask);
5511 struct packet_offload *gro_find_receive_by_type(__be16 type)
5513 struct list_head *offload_head = &offload_base;
5514 struct packet_offload *ptype;
5516 list_for_each_entry_rcu(ptype, offload_head, list) {
5517 if (ptype->type != type || !ptype->callbacks.gro_receive)
5523 EXPORT_SYMBOL(gro_find_receive_by_type);
5525 struct packet_offload *gro_find_complete_by_type(__be16 type)
5527 struct list_head *offload_head = &offload_base;
5528 struct packet_offload *ptype;
5530 list_for_each_entry_rcu(ptype, offload_head, list) {
5531 if (ptype->type != type || !ptype->callbacks.gro_complete)
5537 EXPORT_SYMBOL(gro_find_complete_by_type);
5539 static void napi_skb_free_stolen_head(struct sk_buff *skb)
5543 kmem_cache_free(skbuff_head_cache, skb);
5546 static gro_result_t napi_skb_finish(gro_result_t ret, struct sk_buff *skb)
5550 if (netif_receive_skb_internal(skb))
5558 case GRO_MERGED_FREE:
5559 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
5560 napi_skb_free_stolen_head(skb);
5574 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
5576 skb_mark_napi_id(skb, napi);
5577 trace_napi_gro_receive_entry(skb);
5579 skb_gro_reset_offset(skb);
5581 return napi_skb_finish(dev_gro_receive(napi, skb), skb);
5583 EXPORT_SYMBOL(napi_gro_receive);
5585 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
5587 if (unlikely(skb->pfmemalloc)) {
5591 __skb_pull(skb, skb_headlen(skb));
5592 /* restore the reserve we had after netdev_alloc_skb_ip_align() */
5593 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
5595 skb->dev = napi->dev;
5597 skb->encapsulation = 0;
5598 skb_shinfo(skb)->gso_type = 0;
5599 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
5605 struct sk_buff *napi_get_frags(struct napi_struct *napi)
5607 struct sk_buff *skb = napi->skb;
5610 skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
5613 skb_mark_napi_id(skb, napi);
5618 EXPORT_SYMBOL(napi_get_frags);
5620 static gro_result_t napi_frags_finish(struct napi_struct *napi,
5621 struct sk_buff *skb,
5627 __skb_push(skb, ETH_HLEN);
5628 skb->protocol = eth_type_trans(skb, skb->dev);
5629 if (ret == GRO_NORMAL && netif_receive_skb_internal(skb))
5634 napi_reuse_skb(napi, skb);
5637 case GRO_MERGED_FREE:
5638 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
5639 napi_skb_free_stolen_head(skb);
5641 napi_reuse_skb(napi, skb);
5652 /* Upper GRO stack assumes network header starts at gro_offset=0
5653 * Drivers could call both napi_gro_frags() and napi_gro_receive()
5654 * We copy ethernet header into skb->data to have a common layout.
5656 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
5658 struct sk_buff *skb = napi->skb;
5659 const struct ethhdr *eth;
5660 unsigned int hlen = sizeof(*eth);
5664 skb_reset_mac_header(skb);
5665 skb_gro_reset_offset(skb);
5667 eth = skb_gro_header_fast(skb, 0);
5668 if (unlikely(skb_gro_header_hard(skb, hlen))) {
5669 eth = skb_gro_header_slow(skb, hlen, 0);
5670 if (unlikely(!eth)) {
5671 net_warn_ratelimited("%s: dropping impossible skb from %s\n",
5672 __func__, napi->dev->name);
5673 napi_reuse_skb(napi, skb);
5677 gro_pull_from_frag0(skb, hlen);
5678 NAPI_GRO_CB(skb)->frag0 += hlen;
5679 NAPI_GRO_CB(skb)->frag0_len -= hlen;
5681 __skb_pull(skb, hlen);
5684 * This works because the only protocols we care about don't require
5686 * We'll fix it up properly in napi_frags_finish()
5688 skb->protocol = eth->h_proto;
5693 gro_result_t napi_gro_frags(struct napi_struct *napi)
5695 struct sk_buff *skb = napi_frags_skb(napi);
5700 trace_napi_gro_frags_entry(skb);
5702 return napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
5704 EXPORT_SYMBOL(napi_gro_frags);
5706 /* Compute the checksum from gro_offset and return the folded value
5707 * after adding in any pseudo checksum.
5709 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
5714 wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
5716 /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
5717 sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
5719 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
5720 !skb->csum_complete_sw)
5721 netdev_rx_csum_fault(skb->dev);
5724 NAPI_GRO_CB(skb)->csum = wsum;
5725 NAPI_GRO_CB(skb)->csum_valid = 1;
5729 EXPORT_SYMBOL(__skb_gro_checksum_complete);
5731 static void net_rps_send_ipi(struct softnet_data *remsd)
5735 struct softnet_data *next = remsd->rps_ipi_next;
5737 if (cpu_online(remsd->cpu))
5738 smp_call_function_single_async(remsd->cpu, &remsd->csd);
5745 * net_rps_action_and_irq_enable sends any pending IPI's for rps.
5746 * Note: called with local irq disabled, but exits with local irq enabled.
5748 static void net_rps_action_and_irq_enable(struct softnet_data *sd)
5751 struct softnet_data *remsd = sd->rps_ipi_list;
5754 sd->rps_ipi_list = NULL;
5758 /* Send pending IPI's to kick RPS processing on remote cpus. */
5759 net_rps_send_ipi(remsd);
5765 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd)
5768 return sd->rps_ipi_list != NULL;
5774 static int process_backlog(struct napi_struct *napi, int quota)
5776 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
5780 /* Check if we have pending ipi, its better to send them now,
5781 * not waiting net_rx_action() end.
5783 if (sd_has_rps_ipi_waiting(sd)) {
5784 local_irq_disable();
5785 net_rps_action_and_irq_enable(sd);
5788 napi->weight = dev_rx_weight;
5790 struct sk_buff *skb;
5792 while ((skb = __skb_dequeue(&sd->process_queue))) {
5794 __netif_receive_skb(skb);
5796 input_queue_head_incr(sd);
5797 if (++work >= quota)
5802 local_irq_disable();
5804 if (skb_queue_empty(&sd->input_pkt_queue)) {
5806 * Inline a custom version of __napi_complete().
5807 * only current cpu owns and manipulates this napi,
5808 * and NAPI_STATE_SCHED is the only possible flag set
5810 * We can use a plain write instead of clear_bit(),
5811 * and we dont need an smp_mb() memory barrier.
5816 skb_queue_splice_tail_init(&sd->input_pkt_queue,
5817 &sd->process_queue);
5827 * __napi_schedule - schedule for receive
5828 * @n: entry to schedule
5830 * The entry's receive function will be scheduled to run.
5831 * Consider using __napi_schedule_irqoff() if hard irqs are masked.
5833 void __napi_schedule(struct napi_struct *n)
5835 unsigned long flags;
5837 local_irq_save(flags);
5838 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
5839 local_irq_restore(flags);
5841 EXPORT_SYMBOL(__napi_schedule);
5844 * napi_schedule_prep - check if napi can be scheduled
5847 * Test if NAPI routine is already running, and if not mark
5848 * it as running. This is used as a condition variable
5849 * insure only one NAPI poll instance runs. We also make
5850 * sure there is no pending NAPI disable.
5852 bool napi_schedule_prep(struct napi_struct *n)
5854 unsigned long val, new;
5857 val = READ_ONCE(n->state);
5858 if (unlikely(val & NAPIF_STATE_DISABLE))
5860 new = val | NAPIF_STATE_SCHED;
5862 /* Sets STATE_MISSED bit if STATE_SCHED was already set
5863 * This was suggested by Alexander Duyck, as compiler
5864 * emits better code than :
5865 * if (val & NAPIF_STATE_SCHED)
5866 * new |= NAPIF_STATE_MISSED;
5868 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED *
5870 } while (cmpxchg(&n->state, val, new) != val);
5872 return !(val & NAPIF_STATE_SCHED);
5874 EXPORT_SYMBOL(napi_schedule_prep);
5877 * __napi_schedule_irqoff - schedule for receive
5878 * @n: entry to schedule
5880 * Variant of __napi_schedule() assuming hard irqs are masked
5882 void __napi_schedule_irqoff(struct napi_struct *n)
5884 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
5886 EXPORT_SYMBOL(__napi_schedule_irqoff);
5888 bool napi_complete_done(struct napi_struct *n, int work_done)
5890 unsigned long flags, val, new;
5893 * 1) Don't let napi dequeue from the cpu poll list
5894 * just in case its running on a different cpu.
5895 * 2) If we are busy polling, do nothing here, we have
5896 * the guarantee we will be called later.
5898 if (unlikely(n->state & (NAPIF_STATE_NPSVC |
5899 NAPIF_STATE_IN_BUSY_POLL)))
5902 if (n->gro_bitmask) {
5903 unsigned long timeout = 0;
5906 timeout = n->dev->gro_flush_timeout;
5909 hrtimer_start(&n->timer, ns_to_ktime(timeout),
5910 HRTIMER_MODE_REL_PINNED);
5912 napi_gro_flush(n, false);
5914 if (unlikely(!list_empty(&n->poll_list))) {
5915 /* If n->poll_list is not empty, we need to mask irqs */
5916 local_irq_save(flags);
5917 list_del_init(&n->poll_list);
5918 local_irq_restore(flags);
5922 val = READ_ONCE(n->state);
5924 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED));
5926 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED);
5928 /* If STATE_MISSED was set, leave STATE_SCHED set,
5929 * because we will call napi->poll() one more time.
5930 * This C code was suggested by Alexander Duyck to help gcc.
5932 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED *
5934 } while (cmpxchg(&n->state, val, new) != val);
5936 if (unlikely(val & NAPIF_STATE_MISSED)) {
5943 EXPORT_SYMBOL(napi_complete_done);
5945 /* must be called under rcu_read_lock(), as we dont take a reference */
5946 static struct napi_struct *napi_by_id(unsigned int napi_id)
5948 unsigned int hash = napi_id % HASH_SIZE(napi_hash);
5949 struct napi_struct *napi;
5951 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node)
5952 if (napi->napi_id == napi_id)
5958 #if defined(CONFIG_NET_RX_BUSY_POLL)
5960 #define BUSY_POLL_BUDGET 8
5962 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock)
5966 /* Busy polling means there is a high chance device driver hard irq
5967 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was
5968 * set in napi_schedule_prep().
5969 * Since we are about to call napi->poll() once more, we can safely
5970 * clear NAPI_STATE_MISSED.
5972 * Note: x86 could use a single "lock and ..." instruction
5973 * to perform these two clear_bit()
5975 clear_bit(NAPI_STATE_MISSED, &napi->state);
5976 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state);
5980 /* All we really want here is to re-enable device interrupts.
5981 * Ideally, a new ndo_busy_poll_stop() could avoid another round.
5983 rc = napi->poll(napi, BUSY_POLL_BUDGET);
5984 trace_napi_poll(napi, rc, BUSY_POLL_BUDGET);
5985 netpoll_poll_unlock(have_poll_lock);
5986 if (rc == BUSY_POLL_BUDGET)
5987 __napi_schedule(napi);
5991 void napi_busy_loop(unsigned int napi_id,
5992 bool (*loop_end)(void *, unsigned long),
5995 unsigned long start_time = loop_end ? busy_loop_current_time() : 0;
5996 int (*napi_poll)(struct napi_struct *napi, int budget);
5997 void *have_poll_lock = NULL;
5998 struct napi_struct *napi;
6005 napi = napi_by_id(napi_id);
6015 unsigned long val = READ_ONCE(napi->state);
6017 /* If multiple threads are competing for this napi,
6018 * we avoid dirtying napi->state as much as we can.
6020 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED |
6021 NAPIF_STATE_IN_BUSY_POLL))
6023 if (cmpxchg(&napi->state, val,
6024 val | NAPIF_STATE_IN_BUSY_POLL |
6025 NAPIF_STATE_SCHED) != val)
6027 have_poll_lock = netpoll_poll_lock(napi);
6028 napi_poll = napi->poll;
6030 work = napi_poll(napi, BUSY_POLL_BUDGET);
6031 trace_napi_poll(napi, work, BUSY_POLL_BUDGET);
6034 __NET_ADD_STATS(dev_net(napi->dev),
6035 LINUX_MIB_BUSYPOLLRXPACKETS, work);
6038 if (!loop_end || loop_end(loop_end_arg, start_time))
6041 if (unlikely(need_resched())) {
6043 busy_poll_stop(napi, have_poll_lock);
6047 if (loop_end(loop_end_arg, start_time))
6054 busy_poll_stop(napi, have_poll_lock);
6059 EXPORT_SYMBOL(napi_busy_loop);
6061 #endif /* CONFIG_NET_RX_BUSY_POLL */
6063 static void napi_hash_add(struct napi_struct *napi)
6065 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state) ||
6066 test_and_set_bit(NAPI_STATE_HASHED, &napi->state))
6069 spin_lock(&napi_hash_lock);
6071 /* 0..NR_CPUS range is reserved for sender_cpu use */
6073 if (unlikely(++napi_gen_id < MIN_NAPI_ID))
6074 napi_gen_id = MIN_NAPI_ID;
6075 } while (napi_by_id(napi_gen_id));
6076 napi->napi_id = napi_gen_id;
6078 hlist_add_head_rcu(&napi->napi_hash_node,
6079 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]);
6081 spin_unlock(&napi_hash_lock);
6084 /* Warning : caller is responsible to make sure rcu grace period
6085 * is respected before freeing memory containing @napi
6087 bool napi_hash_del(struct napi_struct *napi)
6089 bool rcu_sync_needed = false;
6091 spin_lock(&napi_hash_lock);
6093 if (test_and_clear_bit(NAPI_STATE_HASHED, &napi->state)) {
6094 rcu_sync_needed = true;
6095 hlist_del_rcu(&napi->napi_hash_node);
6097 spin_unlock(&napi_hash_lock);
6098 return rcu_sync_needed;
6100 EXPORT_SYMBOL_GPL(napi_hash_del);
6102 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer)
6104 struct napi_struct *napi;
6106 napi = container_of(timer, struct napi_struct, timer);
6108 /* Note : we use a relaxed variant of napi_schedule_prep() not setting
6109 * NAPI_STATE_MISSED, since we do not react to a device IRQ.
6111 if (napi->gro_bitmask && !napi_disable_pending(napi) &&
6112 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state))
6113 __napi_schedule_irqoff(napi);
6115 return HRTIMER_NORESTART;
6118 static void init_gro_hash(struct napi_struct *napi)
6122 for (i = 0; i < GRO_HASH_BUCKETS; i++) {
6123 INIT_LIST_HEAD(&napi->gro_hash[i].list);
6124 napi->gro_hash[i].count = 0;
6126 napi->gro_bitmask = 0;
6129 void netif_napi_add(struct net_device *dev, struct napi_struct *napi,
6130 int (*poll)(struct napi_struct *, int), int weight)
6132 INIT_LIST_HEAD(&napi->poll_list);
6133 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
6134 napi->timer.function = napi_watchdog;
6135 init_gro_hash(napi);
6138 if (weight > NAPI_POLL_WEIGHT)
6139 pr_err_once("netif_napi_add() called with weight %d on device %s\n",
6141 napi->weight = weight;
6142 list_add(&napi->dev_list, &dev->napi_list);
6144 #ifdef CONFIG_NETPOLL
6145 napi->poll_owner = -1;
6147 set_bit(NAPI_STATE_SCHED, &napi->state);
6148 napi_hash_add(napi);
6150 EXPORT_SYMBOL(netif_napi_add);
6152 void napi_disable(struct napi_struct *n)
6155 set_bit(NAPI_STATE_DISABLE, &n->state);
6157 while (test_and_set_bit(NAPI_STATE_SCHED, &n->state))
6159 while (test_and_set_bit(NAPI_STATE_NPSVC, &n->state))
6162 hrtimer_cancel(&n->timer);
6164 clear_bit(NAPI_STATE_DISABLE, &n->state);
6166 EXPORT_SYMBOL(napi_disable);
6168 static void flush_gro_hash(struct napi_struct *napi)
6172 for (i = 0; i < GRO_HASH_BUCKETS; i++) {
6173 struct sk_buff *skb, *n;
6175 list_for_each_entry_safe(skb, n, &napi->gro_hash[i].list, list)
6177 napi->gro_hash[i].count = 0;
6181 /* Must be called in process context */
6182 void netif_napi_del(struct napi_struct *napi)
6185 if (napi_hash_del(napi))
6187 list_del_init(&napi->dev_list);
6188 napi_free_frags(napi);
6190 flush_gro_hash(napi);
6191 napi->gro_bitmask = 0;
6193 EXPORT_SYMBOL(netif_napi_del);
6195 static int napi_poll(struct napi_struct *n, struct list_head *repoll)
6200 list_del_init(&n->poll_list);
6202 have = netpoll_poll_lock(n);
6206 /* This NAPI_STATE_SCHED test is for avoiding a race
6207 * with netpoll's poll_napi(). Only the entity which
6208 * obtains the lock and sees NAPI_STATE_SCHED set will
6209 * actually make the ->poll() call. Therefore we avoid
6210 * accidentally calling ->poll() when NAPI is not scheduled.
6213 if (test_bit(NAPI_STATE_SCHED, &n->state)) {
6214 work = n->poll(n, weight);
6215 trace_napi_poll(n, work, weight);
6218 WARN_ON_ONCE(work > weight);
6220 if (likely(work < weight))
6223 /* Drivers must not modify the NAPI state if they
6224 * consume the entire weight. In such cases this code
6225 * still "owns" the NAPI instance and therefore can
6226 * move the instance around on the list at-will.
6228 if (unlikely(napi_disable_pending(n))) {
6233 if (n->gro_bitmask) {
6234 /* flush too old packets
6235 * If HZ < 1000, flush all packets.
6237 napi_gro_flush(n, HZ >= 1000);
6240 /* Some drivers may have called napi_schedule
6241 * prior to exhausting their budget.
6243 if (unlikely(!list_empty(&n->poll_list))) {
6244 pr_warn_once("%s: Budget exhausted after napi rescheduled\n",
6245 n->dev ? n->dev->name : "backlog");
6249 list_add_tail(&n->poll_list, repoll);
6252 netpoll_poll_unlock(have);
6257 static __latent_entropy void net_rx_action(struct softirq_action *h)
6259 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
6260 unsigned long time_limit = jiffies +
6261 usecs_to_jiffies(netdev_budget_usecs);
6262 int budget = netdev_budget;
6266 local_irq_disable();
6267 list_splice_init(&sd->poll_list, &list);
6271 struct napi_struct *n;
6273 if (list_empty(&list)) {
6274 if (!sd_has_rps_ipi_waiting(sd) && list_empty(&repoll))
6279 n = list_first_entry(&list, struct napi_struct, poll_list);
6280 budget -= napi_poll(n, &repoll);
6282 /* If softirq window is exhausted then punt.
6283 * Allow this to run for 2 jiffies since which will allow
6284 * an average latency of 1.5/HZ.
6286 if (unlikely(budget <= 0 ||
6287 time_after_eq(jiffies, time_limit))) {
6293 local_irq_disable();
6295 list_splice_tail_init(&sd->poll_list, &list);
6296 list_splice_tail(&repoll, &list);
6297 list_splice(&list, &sd->poll_list);
6298 if (!list_empty(&sd->poll_list))
6299 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
6301 net_rps_action_and_irq_enable(sd);
6303 __kfree_skb_flush();
6306 struct netdev_adjacent {
6307 struct net_device *dev;
6309 /* upper master flag, there can only be one master device per list */
6312 /* counter for the number of times this device was added to us */
6315 /* private field for the users */
6318 struct list_head list;
6319 struct rcu_head rcu;
6322 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev,
6323 struct list_head *adj_list)
6325 struct netdev_adjacent *adj;
6327 list_for_each_entry(adj, adj_list, list) {
6328 if (adj->dev == adj_dev)
6334 static int __netdev_has_upper_dev(struct net_device *upper_dev, void *data)
6336 struct net_device *dev = data;
6338 return upper_dev == dev;
6342 * netdev_has_upper_dev - Check if device is linked to an upper device
6344 * @upper_dev: upper device to check
6346 * Find out if a device is linked to specified upper device and return true
6347 * in case it is. Note that this checks only immediate upper device,
6348 * not through a complete stack of devices. The caller must hold the RTNL lock.
6350 bool netdev_has_upper_dev(struct net_device *dev,
6351 struct net_device *upper_dev)
6355 return netdev_walk_all_upper_dev_rcu(dev, __netdev_has_upper_dev,
6358 EXPORT_SYMBOL(netdev_has_upper_dev);
6361 * netdev_has_upper_dev_all - Check if device is linked to an upper device
6363 * @upper_dev: upper device to check
6365 * Find out if a device is linked to specified upper device and return true
6366 * in case it is. Note that this checks the entire upper device chain.
6367 * The caller must hold rcu lock.
6370 bool netdev_has_upper_dev_all_rcu(struct net_device *dev,
6371 struct net_device *upper_dev)
6373 return !!netdev_walk_all_upper_dev_rcu(dev, __netdev_has_upper_dev,
6376 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu);
6379 * netdev_has_any_upper_dev - Check if device is linked to some device
6382 * Find out if a device is linked to an upper device and return true in case
6383 * it is. The caller must hold the RTNL lock.
6385 bool netdev_has_any_upper_dev(struct net_device *dev)
6389 return !list_empty(&dev->adj_list.upper);
6391 EXPORT_SYMBOL(netdev_has_any_upper_dev);
6394 * netdev_master_upper_dev_get - Get master upper device
6397 * Find a master upper device and return pointer to it or NULL in case
6398 * it's not there. The caller must hold the RTNL lock.
6400 struct net_device *netdev_master_upper_dev_get(struct net_device *dev)
6402 struct netdev_adjacent *upper;
6406 if (list_empty(&dev->adj_list.upper))
6409 upper = list_first_entry(&dev->adj_list.upper,
6410 struct netdev_adjacent, list);
6411 if (likely(upper->master))
6415 EXPORT_SYMBOL(netdev_master_upper_dev_get);
6418 * netdev_has_any_lower_dev - Check if device is linked to some device
6421 * Find out if a device is linked to a lower device and return true in case
6422 * it is. The caller must hold the RTNL lock.
6424 static bool netdev_has_any_lower_dev(struct net_device *dev)
6428 return !list_empty(&dev->adj_list.lower);
6431 void *netdev_adjacent_get_private(struct list_head *adj_list)
6433 struct netdev_adjacent *adj;
6435 adj = list_entry(adj_list, struct netdev_adjacent, list);
6437 return adj->private;
6439 EXPORT_SYMBOL(netdev_adjacent_get_private);
6442 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list
6444 * @iter: list_head ** of the current position
6446 * Gets the next device from the dev's upper list, starting from iter
6447 * position. The caller must hold RCU read lock.
6449 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev,
6450 struct list_head **iter)
6452 struct netdev_adjacent *upper;
6454 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
6456 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6458 if (&upper->list == &dev->adj_list.upper)
6461 *iter = &upper->list;
6465 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu);
6467 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev,
6468 struct list_head **iter)
6470 struct netdev_adjacent *upper;
6472 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
6474 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6476 if (&upper->list == &dev->adj_list.upper)
6479 *iter = &upper->list;
6484 int netdev_walk_all_upper_dev_rcu(struct net_device *dev,
6485 int (*fn)(struct net_device *dev,
6489 struct net_device *udev;
6490 struct list_head *iter;
6493 for (iter = &dev->adj_list.upper,
6494 udev = netdev_next_upper_dev_rcu(dev, &iter);
6496 udev = netdev_next_upper_dev_rcu(dev, &iter)) {
6497 /* first is the upper device itself */
6498 ret = fn(udev, data);
6502 /* then look at all of its upper devices */
6503 ret = netdev_walk_all_upper_dev_rcu(udev, fn, data);
6510 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu);
6513 * netdev_lower_get_next_private - Get the next ->private from the
6514 * lower neighbour list
6516 * @iter: list_head ** of the current position
6518 * Gets the next netdev_adjacent->private from the dev's lower neighbour
6519 * list, starting from iter position. The caller must hold either hold the
6520 * RTNL lock or its own locking that guarantees that the neighbour lower
6521 * list will remain unchanged.
6523 void *netdev_lower_get_next_private(struct net_device *dev,
6524 struct list_head **iter)
6526 struct netdev_adjacent *lower;
6528 lower = list_entry(*iter, struct netdev_adjacent, list);
6530 if (&lower->list == &dev->adj_list.lower)
6533 *iter = lower->list.next;
6535 return lower->private;
6537 EXPORT_SYMBOL(netdev_lower_get_next_private);
6540 * netdev_lower_get_next_private_rcu - Get the next ->private from the
6541 * lower neighbour list, RCU
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 RCU read lock.
6549 void *netdev_lower_get_next_private_rcu(struct net_device *dev,
6550 struct list_head **iter)
6552 struct netdev_adjacent *lower;
6554 WARN_ON_ONCE(!rcu_read_lock_held());
6556 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6558 if (&lower->list == &dev->adj_list.lower)
6561 *iter = &lower->list;
6563 return lower->private;
6565 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu);
6568 * netdev_lower_get_next - Get the next device from the lower neighbour
6571 * @iter: list_head ** of the current position
6573 * Gets the next netdev_adjacent from the dev's lower neighbour
6574 * list, starting from iter position. The caller must hold RTNL lock or
6575 * its own locking that guarantees that the neighbour lower
6576 * list will remain unchanged.
6578 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter)
6580 struct netdev_adjacent *lower;
6582 lower = list_entry(*iter, struct netdev_adjacent, list);
6584 if (&lower->list == &dev->adj_list.lower)
6587 *iter = lower->list.next;
6591 EXPORT_SYMBOL(netdev_lower_get_next);
6593 static struct net_device *netdev_next_lower_dev(struct net_device *dev,
6594 struct list_head **iter)
6596 struct netdev_adjacent *lower;
6598 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
6600 if (&lower->list == &dev->adj_list.lower)
6603 *iter = &lower->list;
6608 int netdev_walk_all_lower_dev(struct net_device *dev,
6609 int (*fn)(struct net_device *dev,
6613 struct net_device *ldev;
6614 struct list_head *iter;
6617 for (iter = &dev->adj_list.lower,
6618 ldev = netdev_next_lower_dev(dev, &iter);
6620 ldev = netdev_next_lower_dev(dev, &iter)) {
6621 /* first is the lower device itself */
6622 ret = fn(ldev, data);
6626 /* then look at all of its lower devices */
6627 ret = netdev_walk_all_lower_dev(ldev, fn, data);
6634 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev);
6636 static struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev,
6637 struct list_head **iter)
6639 struct netdev_adjacent *lower;
6641 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
6642 if (&lower->list == &dev->adj_list.lower)
6645 *iter = &lower->list;
6650 int netdev_walk_all_lower_dev_rcu(struct net_device *dev,
6651 int (*fn)(struct net_device *dev,
6655 struct net_device *ldev;
6656 struct list_head *iter;
6659 for (iter = &dev->adj_list.lower,
6660 ldev = netdev_next_lower_dev_rcu(dev, &iter);
6662 ldev = netdev_next_lower_dev_rcu(dev, &iter)) {
6663 /* first is the lower device itself */
6664 ret = fn(ldev, data);
6668 /* then look at all of its lower devices */
6669 ret = netdev_walk_all_lower_dev_rcu(ldev, fn, data);
6676 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu);
6679 * netdev_lower_get_first_private_rcu - Get the first ->private from the
6680 * lower neighbour list, RCU
6684 * Gets the first netdev_adjacent->private from the dev's lower neighbour
6685 * list. The caller must hold RCU read lock.
6687 void *netdev_lower_get_first_private_rcu(struct net_device *dev)
6689 struct netdev_adjacent *lower;
6691 lower = list_first_or_null_rcu(&dev->adj_list.lower,
6692 struct netdev_adjacent, list);
6694 return lower->private;
6697 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu);
6700 * netdev_master_upper_dev_get_rcu - Get master upper device
6703 * Find a master upper device and return pointer to it or NULL in case
6704 * it's not there. The caller must hold the RCU read lock.
6706 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev)
6708 struct netdev_adjacent *upper;
6710 upper = list_first_or_null_rcu(&dev->adj_list.upper,
6711 struct netdev_adjacent, list);
6712 if (upper && likely(upper->master))
6716 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu);
6718 static int netdev_adjacent_sysfs_add(struct net_device *dev,
6719 struct net_device *adj_dev,
6720 struct list_head *dev_list)
6722 char linkname[IFNAMSIZ+7];
6724 sprintf(linkname, dev_list == &dev->adj_list.upper ?
6725 "upper_%s" : "lower_%s", adj_dev->name);
6726 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj),
6729 static void netdev_adjacent_sysfs_del(struct net_device *dev,
6731 struct list_head *dev_list)
6733 char linkname[IFNAMSIZ+7];
6735 sprintf(linkname, dev_list == &dev->adj_list.upper ?
6736 "upper_%s" : "lower_%s", name);
6737 sysfs_remove_link(&(dev->dev.kobj), linkname);
6740 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev,
6741 struct net_device *adj_dev,
6742 struct list_head *dev_list)
6744 return (dev_list == &dev->adj_list.upper ||
6745 dev_list == &dev->adj_list.lower) &&
6746 net_eq(dev_net(dev), dev_net(adj_dev));
6749 static int __netdev_adjacent_dev_insert(struct net_device *dev,
6750 struct net_device *adj_dev,
6751 struct list_head *dev_list,
6752 void *private, bool master)
6754 struct netdev_adjacent *adj;
6757 adj = __netdev_find_adj(adj_dev, dev_list);
6761 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n",
6762 dev->name, adj_dev->name, adj->ref_nr);
6767 adj = kmalloc(sizeof(*adj), GFP_KERNEL);
6772 adj->master = master;
6774 adj->private = private;
6777 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n",
6778 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name);
6780 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) {
6781 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list);
6786 /* Ensure that master link is always the first item in list. */
6788 ret = sysfs_create_link(&(dev->dev.kobj),
6789 &(adj_dev->dev.kobj), "master");
6791 goto remove_symlinks;
6793 list_add_rcu(&adj->list, dev_list);
6795 list_add_tail_rcu(&adj->list, dev_list);
6801 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
6802 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
6810 static void __netdev_adjacent_dev_remove(struct net_device *dev,
6811 struct net_device *adj_dev,
6813 struct list_head *dev_list)
6815 struct netdev_adjacent *adj;
6817 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n",
6818 dev->name, adj_dev->name, ref_nr);
6820 adj = __netdev_find_adj(adj_dev, dev_list);
6823 pr_err("Adjacency does not exist for device %s from %s\n",
6824 dev->name, adj_dev->name);
6829 if (adj->ref_nr > ref_nr) {
6830 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n",
6831 dev->name, adj_dev->name, ref_nr,
6832 adj->ref_nr - ref_nr);
6833 adj->ref_nr -= ref_nr;
6838 sysfs_remove_link(&(dev->dev.kobj), "master");
6840 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
6841 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
6843 list_del_rcu(&adj->list);
6844 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n",
6845 adj_dev->name, dev->name, adj_dev->name);
6847 kfree_rcu(adj, rcu);
6850 static int __netdev_adjacent_dev_link_lists(struct net_device *dev,
6851 struct net_device *upper_dev,
6852 struct list_head *up_list,
6853 struct list_head *down_list,
6854 void *private, bool master)
6858 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list,
6863 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list,
6866 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list);
6873 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev,
6874 struct net_device *upper_dev,
6876 struct list_head *up_list,
6877 struct list_head *down_list)
6879 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list);
6880 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list);
6883 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev,
6884 struct net_device *upper_dev,
6885 void *private, bool master)
6887 return __netdev_adjacent_dev_link_lists(dev, upper_dev,
6888 &dev->adj_list.upper,
6889 &upper_dev->adj_list.lower,
6893 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev,
6894 struct net_device *upper_dev)
6896 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1,
6897 &dev->adj_list.upper,
6898 &upper_dev->adj_list.lower);
6901 static int __netdev_upper_dev_link(struct net_device *dev,
6902 struct net_device *upper_dev, bool master,
6903 void *upper_priv, void *upper_info,
6904 struct netlink_ext_ack *extack)
6906 struct netdev_notifier_changeupper_info changeupper_info = {
6911 .upper_dev = upper_dev,
6914 .upper_info = upper_info,
6916 struct net_device *master_dev;
6921 if (dev == upper_dev)
6924 /* To prevent loops, check if dev is not upper device to upper_dev. */
6925 if (netdev_has_upper_dev(upper_dev, dev))
6929 if (netdev_has_upper_dev(dev, upper_dev))
6932 master_dev = netdev_master_upper_dev_get(dev);
6934 return master_dev == upper_dev ? -EEXIST : -EBUSY;
6937 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
6938 &changeupper_info.info);
6939 ret = notifier_to_errno(ret);
6943 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv,
6948 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
6949 &changeupper_info.info);
6950 ret = notifier_to_errno(ret);
6957 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
6963 * netdev_upper_dev_link - Add a link to the upper device
6965 * @upper_dev: new upper device
6966 * @extack: netlink extended ack
6968 * Adds a link to device which is upper to this one. The caller must hold
6969 * the RTNL lock. On a failure a negative errno code is returned.
6970 * On success the reference counts are adjusted and the function
6973 int netdev_upper_dev_link(struct net_device *dev,
6974 struct net_device *upper_dev,
6975 struct netlink_ext_ack *extack)
6977 return __netdev_upper_dev_link(dev, upper_dev, false,
6978 NULL, NULL, extack);
6980 EXPORT_SYMBOL(netdev_upper_dev_link);
6983 * netdev_master_upper_dev_link - Add a master link to the upper device
6985 * @upper_dev: new upper device
6986 * @upper_priv: upper device private
6987 * @upper_info: upper info to be passed down via notifier
6988 * @extack: netlink extended ack
6990 * Adds a link to device which is upper to this one. In this case, only
6991 * one master upper device can be linked, although other non-master devices
6992 * might be linked as well. The caller must hold the RTNL lock.
6993 * On a failure a negative errno code is returned. On success the reference
6994 * counts are adjusted and the function returns zero.
6996 int netdev_master_upper_dev_link(struct net_device *dev,
6997 struct net_device *upper_dev,
6998 void *upper_priv, void *upper_info,
6999 struct netlink_ext_ack *extack)
7001 return __netdev_upper_dev_link(dev, upper_dev, true,
7002 upper_priv, upper_info, extack);
7004 EXPORT_SYMBOL(netdev_master_upper_dev_link);
7007 * netdev_upper_dev_unlink - Removes a link to upper device
7009 * @upper_dev: new upper device
7011 * Removes a link to device which is upper to this one. The caller must hold
7014 void netdev_upper_dev_unlink(struct net_device *dev,
7015 struct net_device *upper_dev)
7017 struct netdev_notifier_changeupper_info changeupper_info = {
7021 .upper_dev = upper_dev,
7027 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev;
7029 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
7030 &changeupper_info.info);
7032 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
7034 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
7035 &changeupper_info.info);
7037 EXPORT_SYMBOL(netdev_upper_dev_unlink);
7040 * netdev_bonding_info_change - Dispatch event about slave change
7042 * @bonding_info: info to dispatch
7044 * Send NETDEV_BONDING_INFO to netdev notifiers with info.
7045 * The caller must hold the RTNL lock.
7047 void netdev_bonding_info_change(struct net_device *dev,
7048 struct netdev_bonding_info *bonding_info)
7050 struct netdev_notifier_bonding_info info = {
7054 memcpy(&info.bonding_info, bonding_info,
7055 sizeof(struct netdev_bonding_info));
7056 call_netdevice_notifiers_info(NETDEV_BONDING_INFO,
7059 EXPORT_SYMBOL(netdev_bonding_info_change);
7061 static void netdev_adjacent_add_links(struct net_device *dev)
7063 struct netdev_adjacent *iter;
7065 struct net *net = dev_net(dev);
7067 list_for_each_entry(iter, &dev->adj_list.upper, list) {
7068 if (!net_eq(net, dev_net(iter->dev)))
7070 netdev_adjacent_sysfs_add(iter->dev, dev,
7071 &iter->dev->adj_list.lower);
7072 netdev_adjacent_sysfs_add(dev, iter->dev,
7073 &dev->adj_list.upper);
7076 list_for_each_entry(iter, &dev->adj_list.lower, list) {
7077 if (!net_eq(net, dev_net(iter->dev)))
7079 netdev_adjacent_sysfs_add(iter->dev, dev,
7080 &iter->dev->adj_list.upper);
7081 netdev_adjacent_sysfs_add(dev, iter->dev,
7082 &dev->adj_list.lower);
7086 static void netdev_adjacent_del_links(struct net_device *dev)
7088 struct netdev_adjacent *iter;
7090 struct net *net = dev_net(dev);
7092 list_for_each_entry(iter, &dev->adj_list.upper, list) {
7093 if (!net_eq(net, dev_net(iter->dev)))
7095 netdev_adjacent_sysfs_del(iter->dev, dev->name,
7096 &iter->dev->adj_list.lower);
7097 netdev_adjacent_sysfs_del(dev, iter->dev->name,
7098 &dev->adj_list.upper);
7101 list_for_each_entry(iter, &dev->adj_list.lower, list) {
7102 if (!net_eq(net, dev_net(iter->dev)))
7104 netdev_adjacent_sysfs_del(iter->dev, dev->name,
7105 &iter->dev->adj_list.upper);
7106 netdev_adjacent_sysfs_del(dev, iter->dev->name,
7107 &dev->adj_list.lower);
7111 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname)
7113 struct netdev_adjacent *iter;
7115 struct net *net = dev_net(dev);
7117 list_for_each_entry(iter, &dev->adj_list.upper, list) {
7118 if (!net_eq(net, dev_net(iter->dev)))
7120 netdev_adjacent_sysfs_del(iter->dev, oldname,
7121 &iter->dev->adj_list.lower);
7122 netdev_adjacent_sysfs_add(iter->dev, dev,
7123 &iter->dev->adj_list.lower);
7126 list_for_each_entry(iter, &dev->adj_list.lower, list) {
7127 if (!net_eq(net, dev_net(iter->dev)))
7129 netdev_adjacent_sysfs_del(iter->dev, oldname,
7130 &iter->dev->adj_list.upper);
7131 netdev_adjacent_sysfs_add(iter->dev, dev,
7132 &iter->dev->adj_list.upper);
7136 void *netdev_lower_dev_get_private(struct net_device *dev,
7137 struct net_device *lower_dev)
7139 struct netdev_adjacent *lower;
7143 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower);
7147 return lower->private;
7149 EXPORT_SYMBOL(netdev_lower_dev_get_private);
7152 int dev_get_nest_level(struct net_device *dev)
7154 struct net_device *lower = NULL;
7155 struct list_head *iter;
7161 netdev_for_each_lower_dev(dev, lower, iter) {
7162 nest = dev_get_nest_level(lower);
7163 if (max_nest < nest)
7167 return max_nest + 1;
7169 EXPORT_SYMBOL(dev_get_nest_level);
7172 * netdev_lower_change - Dispatch event about lower device state change
7173 * @lower_dev: device
7174 * @lower_state_info: state to dispatch
7176 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info.
7177 * The caller must hold the RTNL lock.
7179 void netdev_lower_state_changed(struct net_device *lower_dev,
7180 void *lower_state_info)
7182 struct netdev_notifier_changelowerstate_info changelowerstate_info = {
7183 .info.dev = lower_dev,
7187 changelowerstate_info.lower_state_info = lower_state_info;
7188 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE,
7189 &changelowerstate_info.info);
7191 EXPORT_SYMBOL(netdev_lower_state_changed);
7193 static void dev_change_rx_flags(struct net_device *dev, int flags)
7195 const struct net_device_ops *ops = dev->netdev_ops;
7197 if (ops->ndo_change_rx_flags)
7198 ops->ndo_change_rx_flags(dev, flags);
7201 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify)
7203 unsigned int old_flags = dev->flags;
7209 dev->flags |= IFF_PROMISC;
7210 dev->promiscuity += inc;
7211 if (dev->promiscuity == 0) {
7214 * If inc causes overflow, untouch promisc and return error.
7217 dev->flags &= ~IFF_PROMISC;
7219 dev->promiscuity -= inc;
7220 pr_warn("%s: promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n",
7225 if (dev->flags != old_flags) {
7226 pr_info("device %s %s promiscuous mode\n",
7228 dev->flags & IFF_PROMISC ? "entered" : "left");
7229 if (audit_enabled) {
7230 current_uid_gid(&uid, &gid);
7231 audit_log(audit_context(), GFP_ATOMIC,
7232 AUDIT_ANOM_PROMISCUOUS,
7233 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u",
7234 dev->name, (dev->flags & IFF_PROMISC),
7235 (old_flags & IFF_PROMISC),
7236 from_kuid(&init_user_ns, audit_get_loginuid(current)),
7237 from_kuid(&init_user_ns, uid),
7238 from_kgid(&init_user_ns, gid),
7239 audit_get_sessionid(current));
7242 dev_change_rx_flags(dev, IFF_PROMISC);
7245 __dev_notify_flags(dev, old_flags, IFF_PROMISC);
7250 * dev_set_promiscuity - update promiscuity count on a device
7254 * Add or remove promiscuity from a device. While the count in the device
7255 * remains above zero the interface remains promiscuous. Once it hits zero
7256 * the device reverts back to normal filtering operation. A negative inc
7257 * value is used to drop promiscuity on the device.
7258 * Return 0 if successful or a negative errno code on error.
7260 int dev_set_promiscuity(struct net_device *dev, int inc)
7262 unsigned int old_flags = dev->flags;
7265 err = __dev_set_promiscuity(dev, inc, true);
7268 if (dev->flags != old_flags)
7269 dev_set_rx_mode(dev);
7272 EXPORT_SYMBOL(dev_set_promiscuity);
7274 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify)
7276 unsigned int old_flags = dev->flags, old_gflags = dev->gflags;
7280 dev->flags |= IFF_ALLMULTI;
7281 dev->allmulti += inc;
7282 if (dev->allmulti == 0) {
7285 * If inc causes overflow, untouch allmulti and return error.
7288 dev->flags &= ~IFF_ALLMULTI;
7290 dev->allmulti -= inc;
7291 pr_warn("%s: allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n",
7296 if (dev->flags ^ old_flags) {
7297 dev_change_rx_flags(dev, IFF_ALLMULTI);
7298 dev_set_rx_mode(dev);
7300 __dev_notify_flags(dev, old_flags,
7301 dev->gflags ^ old_gflags);
7307 * dev_set_allmulti - update allmulti count on a device
7311 * Add or remove reception of all multicast frames to a device. While the
7312 * count in the device remains above zero the interface remains listening
7313 * to all interfaces. Once it hits zero the device reverts back to normal
7314 * filtering operation. A negative @inc value is used to drop the counter
7315 * when releasing a resource needing all multicasts.
7316 * Return 0 if successful or a negative errno code on error.
7319 int dev_set_allmulti(struct net_device *dev, int inc)
7321 return __dev_set_allmulti(dev, inc, true);
7323 EXPORT_SYMBOL(dev_set_allmulti);
7326 * Upload unicast and multicast address lists to device and
7327 * configure RX filtering. When the device doesn't support unicast
7328 * filtering it is put in promiscuous mode while unicast addresses
7331 void __dev_set_rx_mode(struct net_device *dev)
7333 const struct net_device_ops *ops = dev->netdev_ops;
7335 /* dev_open will call this function so the list will stay sane. */
7336 if (!(dev->flags&IFF_UP))
7339 if (!netif_device_present(dev))
7342 if (!(dev->priv_flags & IFF_UNICAST_FLT)) {
7343 /* Unicast addresses changes may only happen under the rtnl,
7344 * therefore calling __dev_set_promiscuity here is safe.
7346 if (!netdev_uc_empty(dev) && !dev->uc_promisc) {
7347 __dev_set_promiscuity(dev, 1, false);
7348 dev->uc_promisc = true;
7349 } else if (netdev_uc_empty(dev) && dev->uc_promisc) {
7350 __dev_set_promiscuity(dev, -1, false);
7351 dev->uc_promisc = false;
7355 if (ops->ndo_set_rx_mode)
7356 ops->ndo_set_rx_mode(dev);
7359 void dev_set_rx_mode(struct net_device *dev)
7361 netif_addr_lock_bh(dev);
7362 __dev_set_rx_mode(dev);
7363 netif_addr_unlock_bh(dev);
7367 * dev_get_flags - get flags reported to userspace
7370 * Get the combination of flag bits exported through APIs to userspace.
7372 unsigned int dev_get_flags(const struct net_device *dev)
7376 flags = (dev->flags & ~(IFF_PROMISC |
7381 (dev->gflags & (IFF_PROMISC |
7384 if (netif_running(dev)) {
7385 if (netif_oper_up(dev))
7386 flags |= IFF_RUNNING;
7387 if (netif_carrier_ok(dev))
7388 flags |= IFF_LOWER_UP;
7389 if (netif_dormant(dev))
7390 flags |= IFF_DORMANT;
7395 EXPORT_SYMBOL(dev_get_flags);
7397 int __dev_change_flags(struct net_device *dev, unsigned int flags)
7399 unsigned int old_flags = dev->flags;
7405 * Set the flags on our device.
7408 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP |
7409 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL |
7411 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC |
7415 * Load in the correct multicast list now the flags have changed.
7418 if ((old_flags ^ flags) & IFF_MULTICAST)
7419 dev_change_rx_flags(dev, IFF_MULTICAST);
7421 dev_set_rx_mode(dev);
7424 * Have we downed the interface. We handle IFF_UP ourselves
7425 * according to user attempts to set it, rather than blindly
7430 if ((old_flags ^ flags) & IFF_UP) {
7431 if (old_flags & IFF_UP)
7434 ret = __dev_open(dev);
7437 if ((flags ^ dev->gflags) & IFF_PROMISC) {
7438 int inc = (flags & IFF_PROMISC) ? 1 : -1;
7439 unsigned int old_flags = dev->flags;
7441 dev->gflags ^= IFF_PROMISC;
7443 if (__dev_set_promiscuity(dev, inc, false) >= 0)
7444 if (dev->flags != old_flags)
7445 dev_set_rx_mode(dev);
7448 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI
7449 * is important. Some (broken) drivers set IFF_PROMISC, when
7450 * IFF_ALLMULTI is requested not asking us and not reporting.
7452 if ((flags ^ dev->gflags) & IFF_ALLMULTI) {
7453 int inc = (flags & IFF_ALLMULTI) ? 1 : -1;
7455 dev->gflags ^= IFF_ALLMULTI;
7456 __dev_set_allmulti(dev, inc, false);
7462 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags,
7463 unsigned int gchanges)
7465 unsigned int changes = dev->flags ^ old_flags;
7468 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC);
7470 if (changes & IFF_UP) {
7471 if (dev->flags & IFF_UP)
7472 call_netdevice_notifiers(NETDEV_UP, dev);
7474 call_netdevice_notifiers(NETDEV_DOWN, dev);
7477 if (dev->flags & IFF_UP &&
7478 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) {
7479 struct netdev_notifier_change_info change_info = {
7483 .flags_changed = changes,
7486 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info);
7491 * dev_change_flags - change device settings
7493 * @flags: device state flags
7495 * Change settings on device based state flags. The flags are
7496 * in the userspace exported format.
7498 int dev_change_flags(struct net_device *dev, unsigned int flags)
7501 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags;
7503 ret = __dev_change_flags(dev, flags);
7507 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags);
7508 __dev_notify_flags(dev, old_flags, changes);
7511 EXPORT_SYMBOL(dev_change_flags);
7513 int __dev_set_mtu(struct net_device *dev, int new_mtu)
7515 const struct net_device_ops *ops = dev->netdev_ops;
7517 if (ops->ndo_change_mtu)
7518 return ops->ndo_change_mtu(dev, new_mtu);
7523 EXPORT_SYMBOL(__dev_set_mtu);
7526 * dev_set_mtu - Change maximum transfer unit
7528 * @new_mtu: new transfer unit
7530 * Change the maximum transfer size of the network device.
7532 int dev_set_mtu(struct net_device *dev, int new_mtu)
7536 if (new_mtu == dev->mtu)
7539 /* MTU must be positive, and in range */
7540 if (new_mtu < 0 || new_mtu < dev->min_mtu) {
7541 net_err_ratelimited("%s: Invalid MTU %d requested, hw min %d\n",
7542 dev->name, new_mtu, dev->min_mtu);
7546 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) {
7547 net_err_ratelimited("%s: Invalid MTU %d requested, hw max %d\n",
7548 dev->name, new_mtu, dev->max_mtu);
7552 if (!netif_device_present(dev))
7555 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev);
7556 err = notifier_to_errno(err);
7560 orig_mtu = dev->mtu;
7561 err = __dev_set_mtu(dev, new_mtu);
7564 err = call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
7565 err = notifier_to_errno(err);
7567 /* setting mtu back and notifying everyone again,
7568 * so that they have a chance to revert changes.
7570 __dev_set_mtu(dev, orig_mtu);
7571 call_netdevice_notifiers(NETDEV_CHANGEMTU, dev);
7576 EXPORT_SYMBOL(dev_set_mtu);
7579 * dev_change_tx_queue_len - Change TX queue length of a netdevice
7581 * @new_len: new tx queue length
7583 int dev_change_tx_queue_len(struct net_device *dev, unsigned long new_len)
7585 unsigned int orig_len = dev->tx_queue_len;
7588 if (new_len != (unsigned int)new_len)
7591 if (new_len != orig_len) {
7592 dev->tx_queue_len = new_len;
7593 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev);
7594 res = notifier_to_errno(res);
7597 "refused to change device tx_queue_len\n");
7598 dev->tx_queue_len = orig_len;
7601 return dev_qdisc_change_tx_queue_len(dev);
7608 * dev_set_group - Change group this device belongs to
7610 * @new_group: group this device should belong to
7612 void dev_set_group(struct net_device *dev, int new_group)
7614 dev->group = new_group;
7616 EXPORT_SYMBOL(dev_set_group);
7619 * dev_set_mac_address - Change Media Access Control Address
7623 * Change the hardware (MAC) address of the device
7625 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa)
7627 const struct net_device_ops *ops = dev->netdev_ops;
7630 if (!ops->ndo_set_mac_address)
7632 if (sa->sa_family != dev->type)
7634 if (!netif_device_present(dev))
7636 err = ops->ndo_set_mac_address(dev, sa);
7639 dev->addr_assign_type = NET_ADDR_SET;
7640 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
7641 add_device_randomness(dev->dev_addr, dev->addr_len);
7644 EXPORT_SYMBOL(dev_set_mac_address);
7647 * dev_change_carrier - Change device carrier
7649 * @new_carrier: new value
7651 * Change device carrier
7653 int dev_change_carrier(struct net_device *dev, bool new_carrier)
7655 const struct net_device_ops *ops = dev->netdev_ops;
7657 if (!ops->ndo_change_carrier)
7659 if (!netif_device_present(dev))
7661 return ops->ndo_change_carrier(dev, new_carrier);
7663 EXPORT_SYMBOL(dev_change_carrier);
7666 * dev_get_phys_port_id - Get device physical port ID
7670 * Get device physical port ID
7672 int dev_get_phys_port_id(struct net_device *dev,
7673 struct netdev_phys_item_id *ppid)
7675 const struct net_device_ops *ops = dev->netdev_ops;
7677 if (!ops->ndo_get_phys_port_id)
7679 return ops->ndo_get_phys_port_id(dev, ppid);
7681 EXPORT_SYMBOL(dev_get_phys_port_id);
7684 * dev_get_phys_port_name - Get device physical port name
7687 * @len: limit of bytes to copy to name
7689 * Get device physical port name
7691 int dev_get_phys_port_name(struct net_device *dev,
7692 char *name, size_t len)
7694 const struct net_device_ops *ops = dev->netdev_ops;
7696 if (!ops->ndo_get_phys_port_name)
7698 return ops->ndo_get_phys_port_name(dev, name, len);
7700 EXPORT_SYMBOL(dev_get_phys_port_name);
7703 * dev_change_proto_down - update protocol port state information
7705 * @proto_down: new value
7707 * This info can be used by switch drivers to set the phys state of the
7710 int dev_change_proto_down(struct net_device *dev, bool proto_down)
7712 const struct net_device_ops *ops = dev->netdev_ops;
7714 if (!ops->ndo_change_proto_down)
7716 if (!netif_device_present(dev))
7718 return ops->ndo_change_proto_down(dev, proto_down);
7720 EXPORT_SYMBOL(dev_change_proto_down);
7722 u32 __dev_xdp_query(struct net_device *dev, bpf_op_t bpf_op,
7723 enum bpf_netdev_command cmd)
7725 struct netdev_bpf xdp;
7730 memset(&xdp, 0, sizeof(xdp));
7733 /* Query must always succeed. */
7734 WARN_ON(bpf_op(dev, &xdp) < 0 && cmd == XDP_QUERY_PROG);
7739 static int dev_xdp_install(struct net_device *dev, bpf_op_t bpf_op,
7740 struct netlink_ext_ack *extack, u32 flags,
7741 struct bpf_prog *prog)
7743 struct netdev_bpf xdp;
7745 memset(&xdp, 0, sizeof(xdp));
7746 if (flags & XDP_FLAGS_HW_MODE)
7747 xdp.command = XDP_SETUP_PROG_HW;
7749 xdp.command = XDP_SETUP_PROG;
7750 xdp.extack = extack;
7754 return bpf_op(dev, &xdp);
7757 static void dev_xdp_uninstall(struct net_device *dev)
7759 struct netdev_bpf xdp;
7762 /* Remove generic XDP */
7763 WARN_ON(dev_xdp_install(dev, generic_xdp_install, NULL, 0, NULL));
7765 /* Remove from the driver */
7766 ndo_bpf = dev->netdev_ops->ndo_bpf;
7770 memset(&xdp, 0, sizeof(xdp));
7771 xdp.command = XDP_QUERY_PROG;
7772 WARN_ON(ndo_bpf(dev, &xdp));
7774 WARN_ON(dev_xdp_install(dev, ndo_bpf, NULL, xdp.prog_flags,
7777 /* Remove HW offload */
7778 memset(&xdp, 0, sizeof(xdp));
7779 xdp.command = XDP_QUERY_PROG_HW;
7780 if (!ndo_bpf(dev, &xdp) && xdp.prog_id)
7781 WARN_ON(dev_xdp_install(dev, ndo_bpf, NULL, xdp.prog_flags,
7786 * dev_change_xdp_fd - set or clear a bpf program for a device rx path
7788 * @extack: netlink extended ack
7789 * @fd: new program fd or negative value to clear
7790 * @flags: xdp-related flags
7792 * Set or clear a bpf program for a device
7794 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack,
7797 const struct net_device_ops *ops = dev->netdev_ops;
7798 enum bpf_netdev_command query;
7799 struct bpf_prog *prog = NULL;
7800 bpf_op_t bpf_op, bpf_chk;
7805 query = flags & XDP_FLAGS_HW_MODE ? XDP_QUERY_PROG_HW : XDP_QUERY_PROG;
7807 bpf_op = bpf_chk = ops->ndo_bpf;
7808 if (!bpf_op && (flags & (XDP_FLAGS_DRV_MODE | XDP_FLAGS_HW_MODE)))
7810 if (!bpf_op || (flags & XDP_FLAGS_SKB_MODE))
7811 bpf_op = generic_xdp_install;
7812 if (bpf_op == bpf_chk)
7813 bpf_chk = generic_xdp_install;
7816 if (__dev_xdp_query(dev, bpf_chk, XDP_QUERY_PROG) ||
7817 __dev_xdp_query(dev, bpf_chk, XDP_QUERY_PROG_HW))
7819 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) &&
7820 __dev_xdp_query(dev, bpf_op, query))
7823 prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP,
7824 bpf_op == ops->ndo_bpf);
7826 return PTR_ERR(prog);
7828 if (!(flags & XDP_FLAGS_HW_MODE) &&
7829 bpf_prog_is_dev_bound(prog->aux)) {
7830 NL_SET_ERR_MSG(extack, "using device-bound program without HW_MODE flag is not supported");
7836 err = dev_xdp_install(dev, bpf_op, extack, flags, prog);
7837 if (err < 0 && prog)
7844 * dev_new_index - allocate an ifindex
7845 * @net: the applicable net namespace
7847 * Returns a suitable unique value for a new device interface
7848 * number. The caller must hold the rtnl semaphore or the
7849 * dev_base_lock to be sure it remains unique.
7851 static int dev_new_index(struct net *net)
7853 int ifindex = net->ifindex;
7858 if (!__dev_get_by_index(net, ifindex))
7859 return net->ifindex = ifindex;
7863 /* Delayed registration/unregisteration */
7864 static LIST_HEAD(net_todo_list);
7865 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq);
7867 static void net_set_todo(struct net_device *dev)
7869 list_add_tail(&dev->todo_list, &net_todo_list);
7870 dev_net(dev)->dev_unreg_count++;
7873 static void rollback_registered_many(struct list_head *head)
7875 struct net_device *dev, *tmp;
7876 LIST_HEAD(close_head);
7878 BUG_ON(dev_boot_phase);
7881 list_for_each_entry_safe(dev, tmp, head, unreg_list) {
7882 /* Some devices call without registering
7883 * for initialization unwind. Remove those
7884 * devices and proceed with the remaining.
7886 if (dev->reg_state == NETREG_UNINITIALIZED) {
7887 pr_debug("unregister_netdevice: device %s/%p never was registered\n",
7891 list_del(&dev->unreg_list);
7894 dev->dismantle = true;
7895 BUG_ON(dev->reg_state != NETREG_REGISTERED);
7898 /* If device is running, close it first. */
7899 list_for_each_entry(dev, head, unreg_list)
7900 list_add_tail(&dev->close_list, &close_head);
7901 dev_close_many(&close_head, true);
7903 list_for_each_entry(dev, head, unreg_list) {
7904 /* And unlink it from device chain. */
7905 unlist_netdevice(dev);
7907 dev->reg_state = NETREG_UNREGISTERING;
7909 flush_all_backlogs();
7913 list_for_each_entry(dev, head, unreg_list) {
7914 struct sk_buff *skb = NULL;
7916 /* Shutdown queueing discipline. */
7919 dev_xdp_uninstall(dev);
7921 /* Notify protocols, that we are about to destroy
7922 * this device. They should clean all the things.
7924 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
7926 if (!dev->rtnl_link_ops ||
7927 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
7928 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0,
7929 GFP_KERNEL, NULL, 0);
7932 * Flush the unicast and multicast chains
7937 if (dev->netdev_ops->ndo_uninit)
7938 dev->netdev_ops->ndo_uninit(dev);
7941 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL);
7943 /* Notifier chain MUST detach us all upper devices. */
7944 WARN_ON(netdev_has_any_upper_dev(dev));
7945 WARN_ON(netdev_has_any_lower_dev(dev));
7947 /* Remove entries from kobject tree */
7948 netdev_unregister_kobject(dev);
7950 /* Remove XPS queueing entries */
7951 netif_reset_xps_queues_gt(dev, 0);
7957 list_for_each_entry(dev, head, unreg_list)
7961 static void rollback_registered(struct net_device *dev)
7965 list_add(&dev->unreg_list, &single);
7966 rollback_registered_many(&single);
7970 static netdev_features_t netdev_sync_upper_features(struct net_device *lower,
7971 struct net_device *upper, netdev_features_t features)
7973 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
7974 netdev_features_t feature;
7977 for_each_netdev_feature(&upper_disables, feature_bit) {
7978 feature = __NETIF_F_BIT(feature_bit);
7979 if (!(upper->wanted_features & feature)
7980 && (features & feature)) {
7981 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n",
7982 &feature, upper->name);
7983 features &= ~feature;
7990 static void netdev_sync_lower_features(struct net_device *upper,
7991 struct net_device *lower, netdev_features_t features)
7993 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
7994 netdev_features_t feature;
7997 for_each_netdev_feature(&upper_disables, feature_bit) {
7998 feature = __NETIF_F_BIT(feature_bit);
7999 if (!(features & feature) && (lower->features & feature)) {
8000 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n",
8001 &feature, lower->name);
8002 lower->wanted_features &= ~feature;
8003 netdev_update_features(lower);
8005 if (unlikely(lower->features & feature))
8006 netdev_WARN(upper, "failed to disable %pNF on %s!\n",
8007 &feature, lower->name);
8012 static netdev_features_t netdev_fix_features(struct net_device *dev,
8013 netdev_features_t features)
8015 /* Fix illegal checksum combinations */
8016 if ((features & NETIF_F_HW_CSUM) &&
8017 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
8018 netdev_warn(dev, "mixed HW and IP checksum settings.\n");
8019 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
8022 /* TSO requires that SG is present as well. */
8023 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) {
8024 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n");
8025 features &= ~NETIF_F_ALL_TSO;
8028 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) &&
8029 !(features & NETIF_F_IP_CSUM)) {
8030 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n");
8031 features &= ~NETIF_F_TSO;
8032 features &= ~NETIF_F_TSO_ECN;
8035 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) &&
8036 !(features & NETIF_F_IPV6_CSUM)) {
8037 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n");
8038 features &= ~NETIF_F_TSO6;
8041 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */
8042 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO))
8043 features &= ~NETIF_F_TSO_MANGLEID;
8045 /* TSO ECN requires that TSO is present as well. */
8046 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN)
8047 features &= ~NETIF_F_TSO_ECN;
8049 /* Software GSO depends on SG. */
8050 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) {
8051 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n");
8052 features &= ~NETIF_F_GSO;
8055 /* GSO partial features require GSO partial be set */
8056 if ((features & dev->gso_partial_features) &&
8057 !(features & NETIF_F_GSO_PARTIAL)) {
8059 "Dropping partially supported GSO features since no GSO partial.\n");
8060 features &= ~dev->gso_partial_features;
8063 if (!(features & NETIF_F_RXCSUM)) {
8064 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet
8065 * successfully merged by hardware must also have the
8066 * checksum verified by hardware. If the user does not
8067 * want to enable RXCSUM, logically, we should disable GRO_HW.
8069 if (features & NETIF_F_GRO_HW) {
8070 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n");
8071 features &= ~NETIF_F_GRO_HW;
8075 /* LRO/HW-GRO features cannot be combined with RX-FCS */
8076 if (features & NETIF_F_RXFCS) {
8077 if (features & NETIF_F_LRO) {
8078 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n");
8079 features &= ~NETIF_F_LRO;
8082 if (features & NETIF_F_GRO_HW) {
8083 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n");
8084 features &= ~NETIF_F_GRO_HW;
8091 int __netdev_update_features(struct net_device *dev)
8093 struct net_device *upper, *lower;
8094 netdev_features_t features;
8095 struct list_head *iter;
8100 features = netdev_get_wanted_features(dev);
8102 if (dev->netdev_ops->ndo_fix_features)
8103 features = dev->netdev_ops->ndo_fix_features(dev, features);
8105 /* driver might be less strict about feature dependencies */
8106 features = netdev_fix_features(dev, features);
8108 /* some features can't be enabled if they're off an an upper device */
8109 netdev_for_each_upper_dev_rcu(dev, upper, iter)
8110 features = netdev_sync_upper_features(dev, upper, features);
8112 if (dev->features == features)
8115 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n",
8116 &dev->features, &features);
8118 if (dev->netdev_ops->ndo_set_features)
8119 err = dev->netdev_ops->ndo_set_features(dev, features);
8123 if (unlikely(err < 0)) {
8125 "set_features() failed (%d); wanted %pNF, left %pNF\n",
8126 err, &features, &dev->features);
8127 /* return non-0 since some features might have changed and
8128 * it's better to fire a spurious notification than miss it
8134 /* some features must be disabled on lower devices when disabled
8135 * on an upper device (think: bonding master or bridge)
8137 netdev_for_each_lower_dev(dev, lower, iter)
8138 netdev_sync_lower_features(dev, lower, features);
8141 netdev_features_t diff = features ^ dev->features;
8143 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) {
8144 /* udp_tunnel_{get,drop}_rx_info both need
8145 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the
8146 * device, or they won't do anything.
8147 * Thus we need to update dev->features
8148 * *before* calling udp_tunnel_get_rx_info,
8149 * but *after* calling udp_tunnel_drop_rx_info.
8151 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) {
8152 dev->features = features;
8153 udp_tunnel_get_rx_info(dev);
8155 udp_tunnel_drop_rx_info(dev);
8159 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) {
8160 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) {
8161 dev->features = features;
8162 err |= vlan_get_rx_ctag_filter_info(dev);
8164 vlan_drop_rx_ctag_filter_info(dev);
8168 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) {
8169 if (features & NETIF_F_HW_VLAN_STAG_FILTER) {
8170 dev->features = features;
8171 err |= vlan_get_rx_stag_filter_info(dev);
8173 vlan_drop_rx_stag_filter_info(dev);
8177 dev->features = features;
8180 return err < 0 ? 0 : 1;
8184 * netdev_update_features - recalculate device features
8185 * @dev: the device to check
8187 * Recalculate dev->features set and send notifications if it
8188 * has changed. Should be called after driver or hardware dependent
8189 * conditions might have changed that influence the features.
8191 void netdev_update_features(struct net_device *dev)
8193 if (__netdev_update_features(dev))
8194 netdev_features_change(dev);
8196 EXPORT_SYMBOL(netdev_update_features);
8199 * netdev_change_features - recalculate device features
8200 * @dev: the device to check
8202 * Recalculate dev->features set and send notifications even
8203 * if they have not changed. Should be called instead of
8204 * netdev_update_features() if also dev->vlan_features might
8205 * have changed to allow the changes to be propagated to stacked
8208 void netdev_change_features(struct net_device *dev)
8210 __netdev_update_features(dev);
8211 netdev_features_change(dev);
8213 EXPORT_SYMBOL(netdev_change_features);
8216 * netif_stacked_transfer_operstate - transfer operstate
8217 * @rootdev: the root or lower level device to transfer state from
8218 * @dev: the device to transfer operstate to
8220 * Transfer operational state from root to device. This is normally
8221 * called when a stacking relationship exists between the root
8222 * device and the device(a leaf device).
8224 void netif_stacked_transfer_operstate(const struct net_device *rootdev,
8225 struct net_device *dev)
8227 if (rootdev->operstate == IF_OPER_DORMANT)
8228 netif_dormant_on(dev);
8230 netif_dormant_off(dev);
8232 if (netif_carrier_ok(rootdev))
8233 netif_carrier_on(dev);
8235 netif_carrier_off(dev);
8237 EXPORT_SYMBOL(netif_stacked_transfer_operstate);
8239 static int netif_alloc_rx_queues(struct net_device *dev)
8241 unsigned int i, count = dev->num_rx_queues;
8242 struct netdev_rx_queue *rx;
8243 size_t sz = count * sizeof(*rx);
8248 rx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
8254 for (i = 0; i < count; i++) {
8257 /* XDP RX-queue setup */
8258 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i);
8265 /* Rollback successful reg's and free other resources */
8267 xdp_rxq_info_unreg(&rx[i].xdp_rxq);
8273 static void netif_free_rx_queues(struct net_device *dev)
8275 unsigned int i, count = dev->num_rx_queues;
8277 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */
8281 for (i = 0; i < count; i++)
8282 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq);
8287 static void netdev_init_one_queue(struct net_device *dev,
8288 struct netdev_queue *queue, void *_unused)
8290 /* Initialize queue lock */
8291 spin_lock_init(&queue->_xmit_lock);
8292 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type);
8293 queue->xmit_lock_owner = -1;
8294 netdev_queue_numa_node_write(queue, NUMA_NO_NODE);
8297 dql_init(&queue->dql, HZ);
8301 static void netif_free_tx_queues(struct net_device *dev)
8306 static int netif_alloc_netdev_queues(struct net_device *dev)
8308 unsigned int count = dev->num_tx_queues;
8309 struct netdev_queue *tx;
8310 size_t sz = count * sizeof(*tx);
8312 if (count < 1 || count > 0xffff)
8315 tx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
8321 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL);
8322 spin_lock_init(&dev->tx_global_lock);
8327 void netif_tx_stop_all_queues(struct net_device *dev)
8331 for (i = 0; i < dev->num_tx_queues; i++) {
8332 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
8334 netif_tx_stop_queue(txq);
8337 EXPORT_SYMBOL(netif_tx_stop_all_queues);
8340 * register_netdevice - register a network device
8341 * @dev: device to register
8343 * Take a completed network device structure and add it to the kernel
8344 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
8345 * chain. 0 is returned on success. A negative errno code is returned
8346 * on a failure to set up the device, or if the name is a duplicate.
8348 * Callers must hold the rtnl semaphore. You may want
8349 * register_netdev() instead of this.
8352 * The locking appears insufficient to guarantee two parallel registers
8353 * will not get the same name.
8356 int register_netdevice(struct net_device *dev)
8359 struct net *net = dev_net(dev);
8361 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE <
8362 NETDEV_FEATURE_COUNT);
8363 BUG_ON(dev_boot_phase);
8368 /* When net_device's are persistent, this will be fatal. */
8369 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
8372 spin_lock_init(&dev->addr_list_lock);
8373 netdev_set_addr_lockdep_class(dev);
8375 ret = dev_get_valid_name(net, dev, dev->name);
8379 /* Init, if this function is available */
8380 if (dev->netdev_ops->ndo_init) {
8381 ret = dev->netdev_ops->ndo_init(dev);
8389 if (((dev->hw_features | dev->features) &
8390 NETIF_F_HW_VLAN_CTAG_FILTER) &&
8391 (!dev->netdev_ops->ndo_vlan_rx_add_vid ||
8392 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) {
8393 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n");
8400 dev->ifindex = dev_new_index(net);
8401 else if (__dev_get_by_index(net, dev->ifindex))
8404 /* Transfer changeable features to wanted_features and enable
8405 * software offloads (GSO and GRO).
8407 dev->hw_features |= NETIF_F_SOFT_FEATURES;
8408 dev->features |= NETIF_F_SOFT_FEATURES;
8410 if (dev->netdev_ops->ndo_udp_tunnel_add) {
8411 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT;
8412 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT;
8415 dev->wanted_features = dev->features & dev->hw_features;
8417 if (!(dev->flags & IFF_LOOPBACK))
8418 dev->hw_features |= NETIF_F_NOCACHE_COPY;
8420 /* If IPv4 TCP segmentation offload is supported we should also
8421 * allow the device to enable segmenting the frame with the option
8422 * of ignoring a static IP ID value. This doesn't enable the
8423 * feature itself but allows the user to enable it later.
8425 if (dev->hw_features & NETIF_F_TSO)
8426 dev->hw_features |= NETIF_F_TSO_MANGLEID;
8427 if (dev->vlan_features & NETIF_F_TSO)
8428 dev->vlan_features |= NETIF_F_TSO_MANGLEID;
8429 if (dev->mpls_features & NETIF_F_TSO)
8430 dev->mpls_features |= NETIF_F_TSO_MANGLEID;
8431 if (dev->hw_enc_features & NETIF_F_TSO)
8432 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID;
8434 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices.
8436 dev->vlan_features |= NETIF_F_HIGHDMA;
8438 /* Make NETIF_F_SG inheritable to tunnel devices.
8440 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL;
8442 /* Make NETIF_F_SG inheritable to MPLS.
8444 dev->mpls_features |= NETIF_F_SG;
8446 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev);
8447 ret = notifier_to_errno(ret);
8451 ret = netdev_register_kobject(dev);
8454 dev->reg_state = NETREG_REGISTERED;
8456 __netdev_update_features(dev);
8459 * Default initial state at registry is that the
8460 * device is present.
8463 set_bit(__LINK_STATE_PRESENT, &dev->state);
8465 linkwatch_init_dev(dev);
8467 dev_init_scheduler(dev);
8469 list_netdevice(dev);
8470 add_device_randomness(dev->dev_addr, dev->addr_len);
8472 /* If the device has permanent device address, driver should
8473 * set dev_addr and also addr_assign_type should be set to
8474 * NET_ADDR_PERM (default value).
8476 if (dev->addr_assign_type == NET_ADDR_PERM)
8477 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len);
8479 /* Notify protocols, that a new device appeared. */
8480 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
8481 ret = notifier_to_errno(ret);
8483 rollback_registered(dev);
8484 dev->reg_state = NETREG_UNREGISTERED;
8487 * Prevent userspace races by waiting until the network
8488 * device is fully setup before sending notifications.
8490 if (!dev->rtnl_link_ops ||
8491 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
8492 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
8498 if (dev->netdev_ops->ndo_uninit)
8499 dev->netdev_ops->ndo_uninit(dev);
8500 if (dev->priv_destructor)
8501 dev->priv_destructor(dev);
8504 EXPORT_SYMBOL(register_netdevice);
8507 * init_dummy_netdev - init a dummy network device for NAPI
8508 * @dev: device to init
8510 * This takes a network device structure and initialize the minimum
8511 * amount of fields so it can be used to schedule NAPI polls without
8512 * registering a full blown interface. This is to be used by drivers
8513 * that need to tie several hardware interfaces to a single NAPI
8514 * poll scheduler due to HW limitations.
8516 int init_dummy_netdev(struct net_device *dev)
8518 /* Clear everything. Note we don't initialize spinlocks
8519 * are they aren't supposed to be taken by any of the
8520 * NAPI code and this dummy netdev is supposed to be
8521 * only ever used for NAPI polls
8523 memset(dev, 0, sizeof(struct net_device));
8525 /* make sure we BUG if trying to hit standard
8526 * register/unregister code path
8528 dev->reg_state = NETREG_DUMMY;
8530 /* NAPI wants this */
8531 INIT_LIST_HEAD(&dev->napi_list);
8533 /* a dummy interface is started by default */
8534 set_bit(__LINK_STATE_PRESENT, &dev->state);
8535 set_bit(__LINK_STATE_START, &dev->state);
8537 /* Note : We dont allocate pcpu_refcnt for dummy devices,
8538 * because users of this 'device' dont need to change
8544 EXPORT_SYMBOL_GPL(init_dummy_netdev);
8548 * register_netdev - register a network device
8549 * @dev: device to register
8551 * Take a completed network device structure and add it to the kernel
8552 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
8553 * chain. 0 is returned on success. A negative errno code is returned
8554 * on a failure to set up the device, or if the name is a duplicate.
8556 * This is a wrapper around register_netdevice that takes the rtnl semaphore
8557 * and expands the device name if you passed a format string to
8560 int register_netdev(struct net_device *dev)
8564 if (rtnl_lock_killable())
8566 err = register_netdevice(dev);
8570 EXPORT_SYMBOL(register_netdev);
8572 int netdev_refcnt_read(const struct net_device *dev)
8576 for_each_possible_cpu(i)
8577 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
8580 EXPORT_SYMBOL(netdev_refcnt_read);
8583 * netdev_wait_allrefs - wait until all references are gone.
8584 * @dev: target net_device
8586 * This is called when unregistering network devices.
8588 * Any protocol or device that holds a reference should register
8589 * for netdevice notification, and cleanup and put back the
8590 * reference if they receive an UNREGISTER event.
8591 * We can get stuck here if buggy protocols don't correctly
8594 static void netdev_wait_allrefs(struct net_device *dev)
8596 unsigned long rebroadcast_time, warning_time;
8599 linkwatch_forget_dev(dev);
8601 rebroadcast_time = warning_time = jiffies;
8602 refcnt = netdev_refcnt_read(dev);
8604 while (refcnt != 0) {
8605 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
8608 /* Rebroadcast unregister notification */
8609 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
8615 if (test_bit(__LINK_STATE_LINKWATCH_PENDING,
8617 /* We must not have linkwatch events
8618 * pending on unregister. If this
8619 * happens, we simply run the queue
8620 * unscheduled, resulting in a noop
8623 linkwatch_run_queue();
8628 rebroadcast_time = jiffies;
8633 refcnt = netdev_refcnt_read(dev);
8635 if (time_after(jiffies, warning_time + 10 * HZ)) {
8636 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n",
8638 warning_time = jiffies;
8647 * register_netdevice(x1);
8648 * register_netdevice(x2);
8650 * unregister_netdevice(y1);
8651 * unregister_netdevice(y2);
8657 * We are invoked by rtnl_unlock().
8658 * This allows us to deal with problems:
8659 * 1) We can delete sysfs objects which invoke hotplug
8660 * without deadlocking with linkwatch via keventd.
8661 * 2) Since we run with the RTNL semaphore not held, we can sleep
8662 * safely in order to wait for the netdev refcnt to drop to zero.
8664 * We must not return until all unregister events added during
8665 * the interval the lock was held have been completed.
8667 void netdev_run_todo(void)
8669 struct list_head list;
8671 /* Snapshot list, allow later requests */
8672 list_replace_init(&net_todo_list, &list);
8677 /* Wait for rcu callbacks to finish before next phase */
8678 if (!list_empty(&list))
8681 while (!list_empty(&list)) {
8682 struct net_device *dev
8683 = list_first_entry(&list, struct net_device, todo_list);
8684 list_del(&dev->todo_list);
8686 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) {
8687 pr_err("network todo '%s' but state %d\n",
8688 dev->name, dev->reg_state);
8693 dev->reg_state = NETREG_UNREGISTERED;
8695 netdev_wait_allrefs(dev);
8698 BUG_ON(netdev_refcnt_read(dev));
8699 BUG_ON(!list_empty(&dev->ptype_all));
8700 BUG_ON(!list_empty(&dev->ptype_specific));
8701 WARN_ON(rcu_access_pointer(dev->ip_ptr));
8702 WARN_ON(rcu_access_pointer(dev->ip6_ptr));
8703 #if IS_ENABLED(CONFIG_DECNET)
8704 WARN_ON(dev->dn_ptr);
8706 if (dev->priv_destructor)
8707 dev->priv_destructor(dev);
8708 if (dev->needs_free_netdev)
8711 /* Report a network device has been unregistered */
8713 dev_net(dev)->dev_unreg_count--;
8715 wake_up(&netdev_unregistering_wq);
8717 /* Free network device */
8718 kobject_put(&dev->dev.kobj);
8722 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has
8723 * all the same fields in the same order as net_device_stats, with only
8724 * the type differing, but rtnl_link_stats64 may have additional fields
8725 * at the end for newer counters.
8727 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
8728 const struct net_device_stats *netdev_stats)
8730 #if BITS_PER_LONG == 64
8731 BUILD_BUG_ON(sizeof(*stats64) < sizeof(*netdev_stats));
8732 memcpy(stats64, netdev_stats, sizeof(*netdev_stats));
8733 /* zero out counters that only exist in rtnl_link_stats64 */
8734 memset((char *)stats64 + sizeof(*netdev_stats), 0,
8735 sizeof(*stats64) - sizeof(*netdev_stats));
8737 size_t i, n = sizeof(*netdev_stats) / sizeof(unsigned long);
8738 const unsigned long *src = (const unsigned long *)netdev_stats;
8739 u64 *dst = (u64 *)stats64;
8741 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64));
8742 for (i = 0; i < n; i++)
8744 /* zero out counters that only exist in rtnl_link_stats64 */
8745 memset((char *)stats64 + n * sizeof(u64), 0,
8746 sizeof(*stats64) - n * sizeof(u64));
8749 EXPORT_SYMBOL(netdev_stats_to_stats64);
8752 * dev_get_stats - get network device statistics
8753 * @dev: device to get statistics from
8754 * @storage: place to store stats
8756 * Get network statistics from device. Return @storage.
8757 * The device driver may provide its own method by setting
8758 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats;
8759 * otherwise the internal statistics structure is used.
8761 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev,
8762 struct rtnl_link_stats64 *storage)
8764 const struct net_device_ops *ops = dev->netdev_ops;
8766 if (ops->ndo_get_stats64) {
8767 memset(storage, 0, sizeof(*storage));
8768 ops->ndo_get_stats64(dev, storage);
8769 } else if (ops->ndo_get_stats) {
8770 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev));
8772 netdev_stats_to_stats64(storage, &dev->stats);
8774 storage->rx_dropped += (unsigned long)atomic_long_read(&dev->rx_dropped);
8775 storage->tx_dropped += (unsigned long)atomic_long_read(&dev->tx_dropped);
8776 storage->rx_nohandler += (unsigned long)atomic_long_read(&dev->rx_nohandler);
8779 EXPORT_SYMBOL(dev_get_stats);
8781 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev)
8783 struct netdev_queue *queue = dev_ingress_queue(dev);
8785 #ifdef CONFIG_NET_CLS_ACT
8788 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
8791 netdev_init_one_queue(dev, queue, NULL);
8792 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc);
8793 queue->qdisc_sleeping = &noop_qdisc;
8794 rcu_assign_pointer(dev->ingress_queue, queue);
8799 static const struct ethtool_ops default_ethtool_ops;
8801 void netdev_set_default_ethtool_ops(struct net_device *dev,
8802 const struct ethtool_ops *ops)
8804 if (dev->ethtool_ops == &default_ethtool_ops)
8805 dev->ethtool_ops = ops;
8807 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops);
8809 void netdev_freemem(struct net_device *dev)
8811 char *addr = (char *)dev - dev->padded;
8817 * alloc_netdev_mqs - allocate network device
8818 * @sizeof_priv: size of private data to allocate space for
8819 * @name: device name format string
8820 * @name_assign_type: origin of device name
8821 * @setup: callback to initialize device
8822 * @txqs: the number of TX subqueues to allocate
8823 * @rxqs: the number of RX subqueues to allocate
8825 * Allocates a struct net_device with private data area for driver use
8826 * and performs basic initialization. Also allocates subqueue structs
8827 * for each queue on the device.
8829 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
8830 unsigned char name_assign_type,
8831 void (*setup)(struct net_device *),
8832 unsigned int txqs, unsigned int rxqs)
8834 struct net_device *dev;
8835 unsigned int alloc_size;
8836 struct net_device *p;
8838 BUG_ON(strlen(name) >= sizeof(dev->name));
8841 pr_err("alloc_netdev: Unable to allocate device with zero queues\n");
8846 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n");
8850 alloc_size = sizeof(struct net_device);
8852 /* ensure 32-byte alignment of private area */
8853 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
8854 alloc_size += sizeof_priv;
8856 /* ensure 32-byte alignment of whole construct */
8857 alloc_size += NETDEV_ALIGN - 1;
8859 p = kvzalloc(alloc_size, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
8863 dev = PTR_ALIGN(p, NETDEV_ALIGN);
8864 dev->padded = (char *)dev - (char *)p;
8866 dev->pcpu_refcnt = alloc_percpu(int);
8867 if (!dev->pcpu_refcnt)
8870 if (dev_addr_init(dev))
8876 dev_net_set(dev, &init_net);
8878 dev->gso_max_size = GSO_MAX_SIZE;
8879 dev->gso_max_segs = GSO_MAX_SEGS;
8881 INIT_LIST_HEAD(&dev->napi_list);
8882 INIT_LIST_HEAD(&dev->unreg_list);
8883 INIT_LIST_HEAD(&dev->close_list);
8884 INIT_LIST_HEAD(&dev->link_watch_list);
8885 INIT_LIST_HEAD(&dev->adj_list.upper);
8886 INIT_LIST_HEAD(&dev->adj_list.lower);
8887 INIT_LIST_HEAD(&dev->ptype_all);
8888 INIT_LIST_HEAD(&dev->ptype_specific);
8889 #ifdef CONFIG_NET_SCHED
8890 hash_init(dev->qdisc_hash);
8892 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM;
8895 if (!dev->tx_queue_len) {
8896 dev->priv_flags |= IFF_NO_QUEUE;
8897 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN;
8900 dev->num_tx_queues = txqs;
8901 dev->real_num_tx_queues = txqs;
8902 if (netif_alloc_netdev_queues(dev))
8905 dev->num_rx_queues = rxqs;
8906 dev->real_num_rx_queues = rxqs;
8907 if (netif_alloc_rx_queues(dev))
8910 strcpy(dev->name, name);
8911 dev->name_assign_type = name_assign_type;
8912 dev->group = INIT_NETDEV_GROUP;
8913 if (!dev->ethtool_ops)
8914 dev->ethtool_ops = &default_ethtool_ops;
8916 nf_hook_ingress_init(dev);
8925 free_percpu(dev->pcpu_refcnt);
8927 netdev_freemem(dev);
8930 EXPORT_SYMBOL(alloc_netdev_mqs);
8933 * free_netdev - free network device
8936 * This function does the last stage of destroying an allocated device
8937 * interface. The reference to the device object is released. If this
8938 * is the last reference then it will be freed.Must be called in process
8941 void free_netdev(struct net_device *dev)
8943 struct napi_struct *p, *n;
8946 netif_free_tx_queues(dev);
8947 netif_free_rx_queues(dev);
8949 kfree(rcu_dereference_protected(dev->ingress_queue, 1));
8951 /* Flush device addresses */
8952 dev_addr_flush(dev);
8954 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
8957 free_percpu(dev->pcpu_refcnt);
8958 dev->pcpu_refcnt = NULL;
8960 /* Compatibility with error handling in drivers */
8961 if (dev->reg_state == NETREG_UNINITIALIZED) {
8962 netdev_freemem(dev);
8966 BUG_ON(dev->reg_state != NETREG_UNREGISTERED);
8967 dev->reg_state = NETREG_RELEASED;
8969 /* will free via device release */
8970 put_device(&dev->dev);
8972 EXPORT_SYMBOL(free_netdev);
8975 * synchronize_net - Synchronize with packet receive processing
8977 * Wait for packets currently being received to be done.
8978 * Does not block later packets from starting.
8980 void synchronize_net(void)
8983 if (rtnl_is_locked())
8984 synchronize_rcu_expedited();
8988 EXPORT_SYMBOL(synchronize_net);
8991 * unregister_netdevice_queue - remove device from the kernel
8995 * This function shuts down a device interface and removes it
8996 * from the kernel tables.
8997 * If head not NULL, device is queued to be unregistered later.
8999 * Callers must hold the rtnl semaphore. You may want
9000 * unregister_netdev() instead of this.
9003 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head)
9008 list_move_tail(&dev->unreg_list, head);
9010 rollback_registered(dev);
9011 /* Finish processing unregister after unlock */
9015 EXPORT_SYMBOL(unregister_netdevice_queue);
9018 * unregister_netdevice_many - unregister many devices
9019 * @head: list of devices
9021 * Note: As most callers use a stack allocated list_head,
9022 * we force a list_del() to make sure stack wont be corrupted later.
9024 void unregister_netdevice_many(struct list_head *head)
9026 struct net_device *dev;
9028 if (!list_empty(head)) {
9029 rollback_registered_many(head);
9030 list_for_each_entry(dev, head, unreg_list)
9035 EXPORT_SYMBOL(unregister_netdevice_many);
9038 * unregister_netdev - remove device from the kernel
9041 * This function shuts down a device interface and removes it
9042 * from the kernel tables.
9044 * This is just a wrapper for unregister_netdevice that takes
9045 * the rtnl semaphore. In general you want to use this and not
9046 * unregister_netdevice.
9048 void unregister_netdev(struct net_device *dev)
9051 unregister_netdevice(dev);
9054 EXPORT_SYMBOL(unregister_netdev);
9057 * dev_change_net_namespace - move device to different nethost namespace
9059 * @net: network namespace
9060 * @pat: If not NULL name pattern to try if the current device name
9061 * is already taken in the destination network namespace.
9063 * This function shuts down a device interface and moves it
9064 * to a new network namespace. On success 0 is returned, on
9065 * a failure a netagive errno code is returned.
9067 * Callers must hold the rtnl semaphore.
9070 int dev_change_net_namespace(struct net_device *dev, struct net *net, const char *pat)
9072 int err, new_nsid, new_ifindex;
9076 /* Don't allow namespace local devices to be moved. */
9078 if (dev->features & NETIF_F_NETNS_LOCAL)
9081 /* Ensure the device has been registrered */
9082 if (dev->reg_state != NETREG_REGISTERED)
9085 /* Get out if there is nothing todo */
9087 if (net_eq(dev_net(dev), net))
9090 /* Pick the destination device name, and ensure
9091 * we can use it in the destination network namespace.
9094 if (__dev_get_by_name(net, dev->name)) {
9095 /* We get here if we can't use the current device name */
9098 err = dev_get_valid_name(net, dev, pat);
9104 * And now a mini version of register_netdevice unregister_netdevice.
9107 /* If device is running close it first. */
9110 /* And unlink it from device chain */
9111 unlist_netdevice(dev);
9115 /* Shutdown queueing discipline. */
9118 /* Notify protocols, that we are about to destroy
9119 * this device. They should clean all the things.
9121 * Note that dev->reg_state stays at NETREG_REGISTERED.
9122 * This is wanted because this way 8021q and macvlan know
9123 * the device is just moving and can keep their slaves up.
9125 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
9128 new_nsid = peernet2id_alloc(dev_net(dev), net);
9129 /* If there is an ifindex conflict assign a new one */
9130 if (__dev_get_by_index(net, dev->ifindex))
9131 new_ifindex = dev_new_index(net);
9133 new_ifindex = dev->ifindex;
9135 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid,
9139 * Flush the unicast and multicast chains
9144 /* Send a netdev-removed uevent to the old namespace */
9145 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE);
9146 netdev_adjacent_del_links(dev);
9148 /* Actually switch the network namespace */
9149 dev_net_set(dev, net);
9150 dev->ifindex = new_ifindex;
9152 /* Send a netdev-add uevent to the new namespace */
9153 kobject_uevent(&dev->dev.kobj, KOBJ_ADD);
9154 netdev_adjacent_add_links(dev);
9156 /* Fixup kobjects */
9157 err = device_rename(&dev->dev, dev->name);
9160 /* Add the device back in the hashes */
9161 list_netdevice(dev);
9163 /* Notify protocols, that a new device appeared. */
9164 call_netdevice_notifiers(NETDEV_REGISTER, dev);
9167 * Prevent userspace races by waiting until the network
9168 * device is fully setup before sending notifications.
9170 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
9177 EXPORT_SYMBOL_GPL(dev_change_net_namespace);
9179 static int dev_cpu_dead(unsigned int oldcpu)
9181 struct sk_buff **list_skb;
9182 struct sk_buff *skb;
9184 struct softnet_data *sd, *oldsd, *remsd = NULL;
9186 local_irq_disable();
9187 cpu = smp_processor_id();
9188 sd = &per_cpu(softnet_data, cpu);
9189 oldsd = &per_cpu(softnet_data, oldcpu);
9191 /* Find end of our completion_queue. */
9192 list_skb = &sd->completion_queue;
9194 list_skb = &(*list_skb)->next;
9195 /* Append completion queue from offline CPU. */
9196 *list_skb = oldsd->completion_queue;
9197 oldsd->completion_queue = NULL;
9199 /* Append output queue from offline CPU. */
9200 if (oldsd->output_queue) {
9201 *sd->output_queue_tailp = oldsd->output_queue;
9202 sd->output_queue_tailp = oldsd->output_queue_tailp;
9203 oldsd->output_queue = NULL;
9204 oldsd->output_queue_tailp = &oldsd->output_queue;
9206 /* Append NAPI poll list from offline CPU, with one exception :
9207 * process_backlog() must be called by cpu owning percpu backlog.
9208 * We properly handle process_queue & input_pkt_queue later.
9210 while (!list_empty(&oldsd->poll_list)) {
9211 struct napi_struct *napi = list_first_entry(&oldsd->poll_list,
9215 list_del_init(&napi->poll_list);
9216 if (napi->poll == process_backlog)
9219 ____napi_schedule(sd, napi);
9222 raise_softirq_irqoff(NET_TX_SOFTIRQ);
9226 remsd = oldsd->rps_ipi_list;
9227 oldsd->rps_ipi_list = NULL;
9229 /* send out pending IPI's on offline CPU */
9230 net_rps_send_ipi(remsd);
9232 /* Process offline CPU's input_pkt_queue */
9233 while ((skb = __skb_dequeue(&oldsd->process_queue))) {
9235 input_queue_head_incr(oldsd);
9237 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) {
9239 input_queue_head_incr(oldsd);
9246 * netdev_increment_features - increment feature set by one
9247 * @all: current feature set
9248 * @one: new feature set
9249 * @mask: mask feature set
9251 * Computes a new feature set after adding a device with feature set
9252 * @one to the master device with current feature set @all. Will not
9253 * enable anything that is off in @mask. Returns the new feature set.
9255 netdev_features_t netdev_increment_features(netdev_features_t all,
9256 netdev_features_t one, netdev_features_t mask)
9258 if (mask & NETIF_F_HW_CSUM)
9259 mask |= NETIF_F_CSUM_MASK;
9260 mask |= NETIF_F_VLAN_CHALLENGED;
9262 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask;
9263 all &= one | ~NETIF_F_ALL_FOR_ALL;
9265 /* If one device supports hw checksumming, set for all. */
9266 if (all & NETIF_F_HW_CSUM)
9267 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM);
9271 EXPORT_SYMBOL(netdev_increment_features);
9273 static struct hlist_head * __net_init netdev_create_hash(void)
9276 struct hlist_head *hash;
9278 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL);
9280 for (i = 0; i < NETDEV_HASHENTRIES; i++)
9281 INIT_HLIST_HEAD(&hash[i]);
9286 /* Initialize per network namespace state */
9287 static int __net_init netdev_init(struct net *net)
9289 BUILD_BUG_ON(GRO_HASH_BUCKETS >
9290 8 * FIELD_SIZEOF(struct napi_struct, gro_bitmask));
9292 if (net != &init_net)
9293 INIT_LIST_HEAD(&net->dev_base_head);
9295 net->dev_name_head = netdev_create_hash();
9296 if (net->dev_name_head == NULL)
9299 net->dev_index_head = netdev_create_hash();
9300 if (net->dev_index_head == NULL)
9306 kfree(net->dev_name_head);
9312 * netdev_drivername - network driver for the device
9313 * @dev: network device
9315 * Determine network driver for device.
9317 const char *netdev_drivername(const struct net_device *dev)
9319 const struct device_driver *driver;
9320 const struct device *parent;
9321 const char *empty = "";
9323 parent = dev->dev.parent;
9327 driver = parent->driver;
9328 if (driver && driver->name)
9329 return driver->name;
9333 static void __netdev_printk(const char *level, const struct net_device *dev,
9334 struct va_format *vaf)
9336 if (dev && dev->dev.parent) {
9337 dev_printk_emit(level[1] - '0',
9340 dev_driver_string(dev->dev.parent),
9341 dev_name(dev->dev.parent),
9342 netdev_name(dev), netdev_reg_state(dev),
9345 printk("%s%s%s: %pV",
9346 level, netdev_name(dev), netdev_reg_state(dev), vaf);
9348 printk("%s(NULL net_device): %pV", level, vaf);
9352 void netdev_printk(const char *level, const struct net_device *dev,
9353 const char *format, ...)
9355 struct va_format vaf;
9358 va_start(args, format);
9363 __netdev_printk(level, dev, &vaf);
9367 EXPORT_SYMBOL(netdev_printk);
9369 #define define_netdev_printk_level(func, level) \
9370 void func(const struct net_device *dev, const char *fmt, ...) \
9372 struct va_format vaf; \
9375 va_start(args, fmt); \
9380 __netdev_printk(level, dev, &vaf); \
9384 EXPORT_SYMBOL(func);
9386 define_netdev_printk_level(netdev_emerg, KERN_EMERG);
9387 define_netdev_printk_level(netdev_alert, KERN_ALERT);
9388 define_netdev_printk_level(netdev_crit, KERN_CRIT);
9389 define_netdev_printk_level(netdev_err, KERN_ERR);
9390 define_netdev_printk_level(netdev_warn, KERN_WARNING);
9391 define_netdev_printk_level(netdev_notice, KERN_NOTICE);
9392 define_netdev_printk_level(netdev_info, KERN_INFO);
9394 static void __net_exit netdev_exit(struct net *net)
9396 kfree(net->dev_name_head);
9397 kfree(net->dev_index_head);
9398 if (net != &init_net)
9399 WARN_ON_ONCE(!list_empty(&net->dev_base_head));
9402 static struct pernet_operations __net_initdata netdev_net_ops = {
9403 .init = netdev_init,
9404 .exit = netdev_exit,
9407 static void __net_exit default_device_exit(struct net *net)
9409 struct net_device *dev, *aux;
9411 * Push all migratable network devices back to the
9412 * initial network namespace
9415 for_each_netdev_safe(net, dev, aux) {
9417 char fb_name[IFNAMSIZ];
9419 /* Ignore unmoveable devices (i.e. loopback) */
9420 if (dev->features & NETIF_F_NETNS_LOCAL)
9423 /* Leave virtual devices for the generic cleanup */
9424 if (dev->rtnl_link_ops)
9427 /* Push remaining network devices to init_net */
9428 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex);
9429 err = dev_change_net_namespace(dev, &init_net, fb_name);
9431 pr_emerg("%s: failed to move %s to init_net: %d\n",
9432 __func__, dev->name, err);
9439 static void __net_exit rtnl_lock_unregistering(struct list_head *net_list)
9441 /* Return with the rtnl_lock held when there are no network
9442 * devices unregistering in any network namespace in net_list.
9446 DEFINE_WAIT_FUNC(wait, woken_wake_function);
9448 add_wait_queue(&netdev_unregistering_wq, &wait);
9450 unregistering = false;
9452 list_for_each_entry(net, net_list, exit_list) {
9453 if (net->dev_unreg_count > 0) {
9454 unregistering = true;
9462 wait_woken(&wait, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
9464 remove_wait_queue(&netdev_unregistering_wq, &wait);
9467 static void __net_exit default_device_exit_batch(struct list_head *net_list)
9469 /* At exit all network devices most be removed from a network
9470 * namespace. Do this in the reverse order of registration.
9471 * Do this across as many network namespaces as possible to
9472 * improve batching efficiency.
9474 struct net_device *dev;
9476 LIST_HEAD(dev_kill_list);
9478 /* To prevent network device cleanup code from dereferencing
9479 * loopback devices or network devices that have been freed
9480 * wait here for all pending unregistrations to complete,
9481 * before unregistring the loopback device and allowing the
9482 * network namespace be freed.
9484 * The netdev todo list containing all network devices
9485 * unregistrations that happen in default_device_exit_batch
9486 * will run in the rtnl_unlock() at the end of
9487 * default_device_exit_batch.
9489 rtnl_lock_unregistering(net_list);
9490 list_for_each_entry(net, net_list, exit_list) {
9491 for_each_netdev_reverse(net, dev) {
9492 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink)
9493 dev->rtnl_link_ops->dellink(dev, &dev_kill_list);
9495 unregister_netdevice_queue(dev, &dev_kill_list);
9498 unregister_netdevice_many(&dev_kill_list);
9502 static struct pernet_operations __net_initdata default_device_ops = {
9503 .exit = default_device_exit,
9504 .exit_batch = default_device_exit_batch,
9508 * Initialize the DEV module. At boot time this walks the device list and
9509 * unhooks any devices that fail to initialise (normally hardware not
9510 * present) and leaves us with a valid list of present and active devices.
9515 * This is called single threaded during boot, so no need
9516 * to take the rtnl semaphore.
9518 static int __init net_dev_init(void)
9520 int i, rc = -ENOMEM;
9522 BUG_ON(!dev_boot_phase);
9524 if (dev_proc_init())
9527 if (netdev_kobject_init())
9530 INIT_LIST_HEAD(&ptype_all);
9531 for (i = 0; i < PTYPE_HASH_SIZE; i++)
9532 INIT_LIST_HEAD(&ptype_base[i]);
9534 INIT_LIST_HEAD(&offload_base);
9536 if (register_pernet_subsys(&netdev_net_ops))
9540 * Initialise the packet receive queues.
9543 for_each_possible_cpu(i) {
9544 struct work_struct *flush = per_cpu_ptr(&flush_works, i);
9545 struct softnet_data *sd = &per_cpu(softnet_data, i);
9547 INIT_WORK(flush, flush_backlog);
9549 skb_queue_head_init(&sd->input_pkt_queue);
9550 skb_queue_head_init(&sd->process_queue);
9551 #ifdef CONFIG_XFRM_OFFLOAD
9552 skb_queue_head_init(&sd->xfrm_backlog);
9554 INIT_LIST_HEAD(&sd->poll_list);
9555 sd->output_queue_tailp = &sd->output_queue;
9557 sd->csd.func = rps_trigger_softirq;
9562 init_gro_hash(&sd->backlog);
9563 sd->backlog.poll = process_backlog;
9564 sd->backlog.weight = weight_p;
9569 /* The loopback device is special if any other network devices
9570 * is present in a network namespace the loopback device must
9571 * be present. Since we now dynamically allocate and free the
9572 * loopback device ensure this invariant is maintained by
9573 * keeping the loopback device as the first device on the
9574 * list of network devices. Ensuring the loopback devices
9575 * is the first device that appears and the last network device
9578 if (register_pernet_device(&loopback_net_ops))
9581 if (register_pernet_device(&default_device_ops))
9584 open_softirq(NET_TX_SOFTIRQ, net_tx_action);
9585 open_softirq(NET_RX_SOFTIRQ, net_rx_action);
9587 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead",
9588 NULL, dev_cpu_dead);
9595 subsys_initcall(net_dev_init);