1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * NET3 Protocol independent device support routines.
5 * Derived from the non IP parts of dev.c 1.0.19
7 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
8 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Florian la Roche <rzsfl@rz.uni-sb.de>
12 * Alan Cox <gw4pts@gw4pts.ampr.org>
13 * David Hinds <dahinds@users.sourceforge.net>
14 * Alexey Kuznetsov <kuznet@ms2.inr.ac.ru>
15 * Adam Sulmicki <adam@cfar.umd.edu>
16 * Pekka Riikonen <priikone@poesidon.pspt.fi>
19 * D.J. Barrow : Fixed bug where dev->refcnt gets set
20 * to 2 if register_netdev gets called
21 * before net_dev_init & also removed a
22 * few lines of code in the process.
23 * Alan Cox : device private ioctl copies fields back.
24 * Alan Cox : Transmit queue code does relevant
25 * stunts to keep the queue safe.
26 * Alan Cox : Fixed double lock.
27 * Alan Cox : Fixed promisc NULL pointer trap
28 * ???????? : Support the full private ioctl range
29 * Alan Cox : Moved ioctl permission check into
31 * Tim Kordas : SIOCADDMULTI/SIOCDELMULTI
32 * Alan Cox : 100 backlog just doesn't cut it when
33 * you start doing multicast video 8)
34 * Alan Cox : Rewrote net_bh and list manager.
35 * Alan Cox : Fix ETH_P_ALL echoback lengths.
36 * Alan Cox : Took out transmit every packet pass
37 * Saved a few bytes in the ioctl handler
38 * Alan Cox : Network driver sets packet type before
39 * calling netif_rx. Saves a function
41 * Alan Cox : Hashed net_bh()
42 * Richard Kooijman: Timestamp fixes.
43 * Alan Cox : Wrong field in SIOCGIFDSTADDR
44 * Alan Cox : Device lock protection.
45 * Alan Cox : Fixed nasty side effect of device close
47 * Rudi Cilibrasi : Pass the right thing to
49 * Dave Miller : 32bit quantity for the device lock to
50 * make it work out on a Sparc.
51 * Bjorn Ekwall : Added KERNELD hack.
52 * Alan Cox : Cleaned up the backlog initialise.
53 * Craig Metz : SIOCGIFCONF fix if space for under
55 * Thomas Bogendoerfer : Return ENODEV for dev_open, if there
56 * is no device open function.
57 * Andi Kleen : Fix error reporting for SIOCGIFCONF
58 * Michael Chastain : Fix signed/unsigned for SIOCGIFCONF
59 * Cyrus Durgin : Cleaned for KMOD
60 * Adam Sulmicki : Bug Fix : Network Device Unload
61 * A network device unload needs to purge
63 * Paul Rusty Russell : SIOCSIFNAME
64 * Pekka Riikonen : Netdev boot-time settings code
65 * Andrew Morton : Make unregister_netdevice wait
66 * indefinitely on dev->refcnt
67 * J Hadi Salim : - Backlog queue sampling
68 * - netif_rx() feedback
71 #include <linux/uaccess.h>
72 #include <linux/bitops.h>
73 #include <linux/capability.h>
74 #include <linux/cpu.h>
75 #include <linux/types.h>
76 #include <linux/kernel.h>
77 #include <linux/hash.h>
78 #include <linux/slab.h>
79 #include <linux/sched.h>
80 #include <linux/sched/mm.h>
81 #include <linux/mutex.h>
82 #include <linux/rwsem.h>
83 #include <linux/string.h>
85 #include <linux/socket.h>
86 #include <linux/sockios.h>
87 #include <linux/errno.h>
88 #include <linux/interrupt.h>
89 #include <linux/if_ether.h>
90 #include <linux/netdevice.h>
91 #include <linux/etherdevice.h>
92 #include <linux/ethtool.h>
93 #include <linux/skbuff.h>
94 #include <linux/kthread.h>
95 #include <linux/bpf.h>
96 #include <linux/bpf_trace.h>
97 #include <net/net_namespace.h>
99 #include <net/busy_poll.h>
100 #include <linux/rtnetlink.h>
101 #include <linux/stat.h>
104 #include <net/dst_metadata.h>
106 #include <net/pkt_sched.h>
107 #include <net/pkt_cls.h>
108 #include <net/checksum.h>
109 #include <net/xfrm.h>
110 #include <linux/highmem.h>
111 #include <linux/init.h>
112 #include <linux/module.h>
113 #include <linux/netpoll.h>
114 #include <linux/rcupdate.h>
115 #include <linux/delay.h>
116 #include <net/iw_handler.h>
117 #include <asm/current.h>
118 #include <linux/audit.h>
119 #include <linux/dmaengine.h>
120 #include <linux/err.h>
121 #include <linux/ctype.h>
122 #include <linux/if_arp.h>
123 #include <linux/if_vlan.h>
124 #include <linux/ip.h>
126 #include <net/mpls.h>
127 #include <linux/ipv6.h>
128 #include <linux/in.h>
129 #include <linux/jhash.h>
130 #include <linux/random.h>
131 #include <trace/events/napi.h>
132 #include <trace/events/net.h>
133 #include <trace/events/skb.h>
134 #include <linux/inetdevice.h>
135 #include <linux/cpu_rmap.h>
136 #include <linux/static_key.h>
137 #include <linux/hashtable.h>
138 #include <linux/vmalloc.h>
139 #include <linux/if_macvlan.h>
140 #include <linux/errqueue.h>
141 #include <linux/hrtimer.h>
142 #include <linux/netfilter_ingress.h>
143 #include <linux/crash_dump.h>
144 #include <linux/sctp.h>
145 #include <net/udp_tunnel.h>
146 #include <linux/net_namespace.h>
147 #include <linux/indirect_call_wrapper.h>
148 #include <net/devlink.h>
149 #include <linux/pm_runtime.h>
150 #include <linux/prandom.h>
151 #include <linux/once_lite.h>
153 #include "net-sysfs.h"
155 #define MAX_GRO_SKBS 8
157 /* This should be increased if a protocol with a bigger head is added. */
158 #define GRO_MAX_HEAD (MAX_HEADER + 128)
160 static DEFINE_SPINLOCK(ptype_lock);
161 static DEFINE_SPINLOCK(offload_lock);
162 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly;
163 struct list_head ptype_all __read_mostly; /* Taps */
164 static struct list_head offload_base __read_mostly;
166 static int netif_rx_internal(struct sk_buff *skb);
167 static int call_netdevice_notifiers_info(unsigned long val,
168 struct netdev_notifier_info *info);
169 static int call_netdevice_notifiers_extack(unsigned long val,
170 struct net_device *dev,
171 struct netlink_ext_ack *extack);
172 static struct napi_struct *napi_by_id(unsigned int napi_id);
175 * The @dev_base_head list is protected by @dev_base_lock and the rtnl
178 * Pure readers hold dev_base_lock for reading, or rcu_read_lock()
180 * Writers must hold the rtnl semaphore while they loop through the
181 * dev_base_head list, and hold dev_base_lock for writing when they do the
182 * actual updates. This allows pure readers to access the list even
183 * while a writer is preparing to update it.
185 * To put it another way, dev_base_lock is held for writing only to
186 * protect against pure readers; the rtnl semaphore provides the
187 * protection against other writers.
189 * See, for example usages, register_netdevice() and
190 * unregister_netdevice(), which must be called with the rtnl
193 DEFINE_RWLOCK(dev_base_lock);
194 EXPORT_SYMBOL(dev_base_lock);
196 static DEFINE_MUTEX(ifalias_mutex);
198 /* protects napi_hash addition/deletion and napi_gen_id */
199 static DEFINE_SPINLOCK(napi_hash_lock);
201 static unsigned int napi_gen_id = NR_CPUS;
202 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8);
204 static DECLARE_RWSEM(devnet_rename_sem);
206 static inline void dev_base_seq_inc(struct net *net)
208 while (++net->dev_base_seq == 0)
212 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name)
214 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ));
216 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)];
219 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex)
221 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)];
224 static inline void rps_lock(struct softnet_data *sd)
227 spin_lock(&sd->input_pkt_queue.lock);
231 static inline void rps_unlock(struct softnet_data *sd)
234 spin_unlock(&sd->input_pkt_queue.lock);
238 static struct netdev_name_node *netdev_name_node_alloc(struct net_device *dev,
241 struct netdev_name_node *name_node;
243 name_node = kmalloc(sizeof(*name_node), GFP_KERNEL);
246 INIT_HLIST_NODE(&name_node->hlist);
247 name_node->dev = dev;
248 name_node->name = name;
252 static struct netdev_name_node *
253 netdev_name_node_head_alloc(struct net_device *dev)
255 struct netdev_name_node *name_node;
257 name_node = netdev_name_node_alloc(dev, dev->name);
260 INIT_LIST_HEAD(&name_node->list);
264 static void netdev_name_node_free(struct netdev_name_node *name_node)
269 static void netdev_name_node_add(struct net *net,
270 struct netdev_name_node *name_node)
272 hlist_add_head_rcu(&name_node->hlist,
273 dev_name_hash(net, name_node->name));
276 static void netdev_name_node_del(struct netdev_name_node *name_node)
278 hlist_del_rcu(&name_node->hlist);
281 static struct netdev_name_node *netdev_name_node_lookup(struct net *net,
284 struct hlist_head *head = dev_name_hash(net, name);
285 struct netdev_name_node *name_node;
287 hlist_for_each_entry(name_node, head, hlist)
288 if (!strcmp(name_node->name, name))
293 static struct netdev_name_node *netdev_name_node_lookup_rcu(struct net *net,
296 struct hlist_head *head = dev_name_hash(net, name);
297 struct netdev_name_node *name_node;
299 hlist_for_each_entry_rcu(name_node, head, hlist)
300 if (!strcmp(name_node->name, name))
305 int netdev_name_node_alt_create(struct net_device *dev, const char *name)
307 struct netdev_name_node *name_node;
308 struct net *net = dev_net(dev);
310 name_node = netdev_name_node_lookup(net, name);
313 name_node = netdev_name_node_alloc(dev, name);
316 netdev_name_node_add(net, name_node);
317 /* The node that holds dev->name acts as a head of per-device list. */
318 list_add_tail(&name_node->list, &dev->name_node->list);
322 EXPORT_SYMBOL(netdev_name_node_alt_create);
324 static void __netdev_name_node_alt_destroy(struct netdev_name_node *name_node)
326 list_del(&name_node->list);
327 netdev_name_node_del(name_node);
328 kfree(name_node->name);
329 netdev_name_node_free(name_node);
332 int netdev_name_node_alt_destroy(struct net_device *dev, const char *name)
334 struct netdev_name_node *name_node;
335 struct net *net = dev_net(dev);
337 name_node = netdev_name_node_lookup(net, name);
340 /* lookup might have found our primary name or a name belonging
343 if (name_node == dev->name_node || name_node->dev != dev)
346 __netdev_name_node_alt_destroy(name_node);
350 EXPORT_SYMBOL(netdev_name_node_alt_destroy);
352 static void netdev_name_node_alt_flush(struct net_device *dev)
354 struct netdev_name_node *name_node, *tmp;
356 list_for_each_entry_safe(name_node, tmp, &dev->name_node->list, list)
357 __netdev_name_node_alt_destroy(name_node);
360 /* Device list insertion */
361 static void list_netdevice(struct net_device *dev)
363 struct net *net = dev_net(dev);
367 write_lock_bh(&dev_base_lock);
368 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head);
369 netdev_name_node_add(net, dev->name_node);
370 hlist_add_head_rcu(&dev->index_hlist,
371 dev_index_hash(net, dev->ifindex));
372 write_unlock_bh(&dev_base_lock);
374 dev_base_seq_inc(net);
377 /* Device list removal
378 * caller must respect a RCU grace period before freeing/reusing dev
380 static void unlist_netdevice(struct net_device *dev)
384 /* Unlink dev from the device chain */
385 write_lock_bh(&dev_base_lock);
386 list_del_rcu(&dev->dev_list);
387 netdev_name_node_del(dev->name_node);
388 hlist_del_rcu(&dev->index_hlist);
389 write_unlock_bh(&dev_base_lock);
391 dev_base_seq_inc(dev_net(dev));
398 static RAW_NOTIFIER_HEAD(netdev_chain);
401 * Device drivers call our routines to queue packets here. We empty the
402 * queue in the local softnet handler.
405 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data);
406 EXPORT_PER_CPU_SYMBOL(softnet_data);
408 #ifdef CONFIG_LOCKDEP
410 * register_netdevice() inits txq->_xmit_lock and sets lockdep class
411 * according to dev->type
413 static const unsigned short netdev_lock_type[] = {
414 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25,
415 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET,
416 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM,
417 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP,
418 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD,
419 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25,
420 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP,
421 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD,
422 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI,
423 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE,
424 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET,
425 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL,
426 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM,
427 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE,
428 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE};
430 static const char *const netdev_lock_name[] = {
431 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25",
432 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET",
433 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM",
434 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP",
435 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD",
436 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25",
437 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP",
438 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD",
439 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI",
440 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE",
441 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET",
442 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL",
443 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM",
444 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE",
445 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"};
447 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)];
448 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)];
450 static inline unsigned short netdev_lock_pos(unsigned short dev_type)
454 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++)
455 if (netdev_lock_type[i] == dev_type)
457 /* the last key is used by default */
458 return ARRAY_SIZE(netdev_lock_type) - 1;
461 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
462 unsigned short dev_type)
466 i = netdev_lock_pos(dev_type);
467 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i],
468 netdev_lock_name[i]);
471 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
475 i = netdev_lock_pos(dev->type);
476 lockdep_set_class_and_name(&dev->addr_list_lock,
477 &netdev_addr_lock_key[i],
478 netdev_lock_name[i]);
481 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock,
482 unsigned short dev_type)
486 static inline void netdev_set_addr_lockdep_class(struct net_device *dev)
491 /*******************************************************************************
493 * Protocol management and registration routines
495 *******************************************************************************/
499 * Add a protocol ID to the list. Now that the input handler is
500 * smarter we can dispense with all the messy stuff that used to be
503 * BEWARE!!! Protocol handlers, mangling input packets,
504 * MUST BE last in hash buckets and checking protocol handlers
505 * MUST start from promiscuous ptype_all chain in net_bh.
506 * It is true now, do not change it.
507 * Explanation follows: if protocol handler, mangling packet, will
508 * be the first on list, it is not able to sense, that packet
509 * is cloned and should be copied-on-write, so that it will
510 * change it and subsequent readers will get broken packet.
514 static inline struct list_head *ptype_head(const struct packet_type *pt)
516 if (pt->type == htons(ETH_P_ALL))
517 return pt->dev ? &pt->dev->ptype_all : &ptype_all;
519 return pt->dev ? &pt->dev->ptype_specific :
520 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK];
524 * dev_add_pack - add packet handler
525 * @pt: packet type declaration
527 * Add a protocol handler to the networking stack. The passed &packet_type
528 * is linked into kernel lists and may not be freed until it has been
529 * removed from the kernel lists.
531 * This call does not sleep therefore it can not
532 * guarantee all CPU's that are in middle of receiving packets
533 * will see the new packet type (until the next received packet).
536 void dev_add_pack(struct packet_type *pt)
538 struct list_head *head = ptype_head(pt);
540 spin_lock(&ptype_lock);
541 list_add_rcu(&pt->list, head);
542 spin_unlock(&ptype_lock);
544 EXPORT_SYMBOL(dev_add_pack);
547 * __dev_remove_pack - remove packet handler
548 * @pt: packet type declaration
550 * Remove a protocol handler that was previously added to the kernel
551 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
552 * from the kernel lists and can be freed or reused once this function
555 * The packet type might still be in use by receivers
556 * and must not be freed until after all the CPU's have gone
557 * through a quiescent state.
559 void __dev_remove_pack(struct packet_type *pt)
561 struct list_head *head = ptype_head(pt);
562 struct packet_type *pt1;
564 spin_lock(&ptype_lock);
566 list_for_each_entry(pt1, head, list) {
568 list_del_rcu(&pt->list);
573 pr_warn("dev_remove_pack: %p not found\n", pt);
575 spin_unlock(&ptype_lock);
577 EXPORT_SYMBOL(__dev_remove_pack);
580 * dev_remove_pack - remove packet handler
581 * @pt: packet type declaration
583 * Remove a protocol handler that was previously added to the kernel
584 * protocol handlers by dev_add_pack(). The passed &packet_type is removed
585 * from the kernel lists and can be freed or reused once this function
588 * This call sleeps to guarantee that no CPU is looking at the packet
591 void dev_remove_pack(struct packet_type *pt)
593 __dev_remove_pack(pt);
597 EXPORT_SYMBOL(dev_remove_pack);
601 * dev_add_offload - register offload handlers
602 * @po: protocol offload declaration
604 * Add protocol offload handlers to the networking stack. The passed
605 * &proto_offload is linked into kernel lists and may not be freed until
606 * it has been removed from the kernel lists.
608 * This call does not sleep therefore it can not
609 * guarantee all CPU's that are in middle of receiving packets
610 * will see the new offload handlers (until the next received packet).
612 void dev_add_offload(struct packet_offload *po)
614 struct packet_offload *elem;
616 spin_lock(&offload_lock);
617 list_for_each_entry(elem, &offload_base, list) {
618 if (po->priority < elem->priority)
621 list_add_rcu(&po->list, elem->list.prev);
622 spin_unlock(&offload_lock);
624 EXPORT_SYMBOL(dev_add_offload);
627 * __dev_remove_offload - remove offload handler
628 * @po: packet offload declaration
630 * Remove a protocol offload handler that was previously added to the
631 * kernel offload handlers by dev_add_offload(). The passed &offload_type
632 * is removed from the kernel lists and can be freed or reused once this
635 * The packet type might still be in use by receivers
636 * and must not be freed until after all the CPU's have gone
637 * through a quiescent state.
639 static void __dev_remove_offload(struct packet_offload *po)
641 struct list_head *head = &offload_base;
642 struct packet_offload *po1;
644 spin_lock(&offload_lock);
646 list_for_each_entry(po1, head, list) {
648 list_del_rcu(&po->list);
653 pr_warn("dev_remove_offload: %p not found\n", po);
655 spin_unlock(&offload_lock);
659 * dev_remove_offload - remove packet offload handler
660 * @po: packet offload declaration
662 * Remove a packet offload handler that was previously added to the kernel
663 * offload handlers by dev_add_offload(). The passed &offload_type is
664 * removed from the kernel lists and can be freed or reused once this
667 * This call sleeps to guarantee that no CPU is looking at the packet
670 void dev_remove_offload(struct packet_offload *po)
672 __dev_remove_offload(po);
676 EXPORT_SYMBOL(dev_remove_offload);
678 /******************************************************************************
680 * Device Boot-time Settings Routines
682 ******************************************************************************/
684 /* Boot time configuration table */
685 static struct netdev_boot_setup dev_boot_setup[NETDEV_BOOT_SETUP_MAX];
688 * netdev_boot_setup_add - add new setup entry
689 * @name: name of the device
690 * @map: configured settings for the device
692 * Adds new setup entry to the dev_boot_setup list. The function
693 * returns 0 on error and 1 on success. This is a generic routine to
696 static int netdev_boot_setup_add(char *name, struct ifmap *map)
698 struct netdev_boot_setup *s;
702 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
703 if (s[i].name[0] == '\0' || s[i].name[0] == ' ') {
704 memset(s[i].name, 0, sizeof(s[i].name));
705 strlcpy(s[i].name, name, IFNAMSIZ);
706 memcpy(&s[i].map, map, sizeof(s[i].map));
711 return i >= NETDEV_BOOT_SETUP_MAX ? 0 : 1;
715 * netdev_boot_setup_check - check boot time settings
716 * @dev: the netdevice
718 * Check boot time settings for the device.
719 * The found settings are set for the device to be used
720 * later in the device probing.
721 * Returns 0 if no settings found, 1 if they are.
723 int netdev_boot_setup_check(struct net_device *dev)
725 struct netdev_boot_setup *s = dev_boot_setup;
728 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++) {
729 if (s[i].name[0] != '\0' && s[i].name[0] != ' ' &&
730 !strcmp(dev->name, s[i].name)) {
731 dev->irq = s[i].map.irq;
732 dev->base_addr = s[i].map.base_addr;
733 dev->mem_start = s[i].map.mem_start;
734 dev->mem_end = s[i].map.mem_end;
740 EXPORT_SYMBOL(netdev_boot_setup_check);
744 * netdev_boot_base - get address from boot time settings
745 * @prefix: prefix for network device
746 * @unit: id for network device
748 * Check boot time settings for the base address of device.
749 * The found settings are set for the device to be used
750 * later in the device probing.
751 * Returns 0 if no settings found.
753 unsigned long netdev_boot_base(const char *prefix, int unit)
755 const struct netdev_boot_setup *s = dev_boot_setup;
759 sprintf(name, "%s%d", prefix, unit);
762 * If device already registered then return base of 1
763 * to indicate not to probe for this interface
765 if (__dev_get_by_name(&init_net, name))
768 for (i = 0; i < NETDEV_BOOT_SETUP_MAX; i++)
769 if (!strcmp(name, s[i].name))
770 return s[i].map.base_addr;
775 * Saves at boot time configured settings for any netdevice.
777 int __init netdev_boot_setup(char *str)
782 str = get_options(str, ARRAY_SIZE(ints), ints);
787 memset(&map, 0, sizeof(map));
791 map.base_addr = ints[2];
793 map.mem_start = ints[3];
795 map.mem_end = ints[4];
797 /* Add new entry to the list */
798 return netdev_boot_setup_add(str, &map);
801 __setup("netdev=", netdev_boot_setup);
803 /*******************************************************************************
805 * Device Interface Subroutines
807 *******************************************************************************/
810 * dev_get_iflink - get 'iflink' value of a interface
811 * @dev: targeted interface
813 * Indicates the ifindex the interface is linked to.
814 * Physical interfaces have the same 'ifindex' and 'iflink' values.
817 int dev_get_iflink(const struct net_device *dev)
819 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink)
820 return dev->netdev_ops->ndo_get_iflink(dev);
824 EXPORT_SYMBOL(dev_get_iflink);
827 * dev_fill_metadata_dst - Retrieve tunnel egress information.
828 * @dev: targeted interface
831 * For better visibility of tunnel traffic OVS needs to retrieve
832 * egress tunnel information for a packet. Following API allows
833 * user to get this info.
835 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb)
837 struct ip_tunnel_info *info;
839 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst)
842 info = skb_tunnel_info_unclone(skb);
845 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX)))
848 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb);
850 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst);
852 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack)
854 int k = stack->num_paths++;
856 if (WARN_ON_ONCE(k >= NET_DEVICE_PATH_STACK_MAX))
859 return &stack->path[k];
862 int dev_fill_forward_path(const struct net_device *dev, const u8 *daddr,
863 struct net_device_path_stack *stack)
865 const struct net_device *last_dev;
866 struct net_device_path_ctx ctx = {
870 struct net_device_path *path;
873 stack->num_paths = 0;
874 while (ctx.dev && ctx.dev->netdev_ops->ndo_fill_forward_path) {
876 path = dev_fwd_path(stack);
880 memset(path, 0, sizeof(struct net_device_path));
881 ret = ctx.dev->netdev_ops->ndo_fill_forward_path(&ctx, path);
885 if (WARN_ON_ONCE(last_dev == ctx.dev))
888 path = dev_fwd_path(stack);
891 path->type = DEV_PATH_ETHERNET;
896 EXPORT_SYMBOL_GPL(dev_fill_forward_path);
899 * __dev_get_by_name - find a device by its name
900 * @net: the applicable net namespace
901 * @name: name to find
903 * Find an interface by name. Must be called under RTNL semaphore
904 * or @dev_base_lock. If the name is found a pointer to the device
905 * is returned. If the name is not found then %NULL is returned. The
906 * reference counters are not incremented so the caller must be
907 * careful with locks.
910 struct net_device *__dev_get_by_name(struct net *net, const char *name)
912 struct netdev_name_node *node_name;
914 node_name = netdev_name_node_lookup(net, name);
915 return node_name ? node_name->dev : NULL;
917 EXPORT_SYMBOL(__dev_get_by_name);
920 * dev_get_by_name_rcu - find a device by its name
921 * @net: the applicable net namespace
922 * @name: name to find
924 * Find an interface by name.
925 * If the name is found a pointer to the device is returned.
926 * If the name is not found then %NULL is returned.
927 * The reference counters are not incremented so the caller must be
928 * careful with locks. The caller must hold RCU lock.
931 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name)
933 struct netdev_name_node *node_name;
935 node_name = netdev_name_node_lookup_rcu(net, name);
936 return node_name ? node_name->dev : NULL;
938 EXPORT_SYMBOL(dev_get_by_name_rcu);
941 * dev_get_by_name - find a device by its name
942 * @net: the applicable net namespace
943 * @name: name to find
945 * Find an interface by name. This can be called from any
946 * context and does its own locking. The returned handle has
947 * the usage count incremented and the caller must use dev_put() to
948 * release it when it is no longer needed. %NULL is returned if no
949 * matching device is found.
952 struct net_device *dev_get_by_name(struct net *net, const char *name)
954 struct net_device *dev;
957 dev = dev_get_by_name_rcu(net, name);
963 EXPORT_SYMBOL(dev_get_by_name);
966 * __dev_get_by_index - find a device by its ifindex
967 * @net: the applicable net namespace
968 * @ifindex: index of device
970 * Search for an interface by index. Returns %NULL if the device
971 * is not found or a pointer to the device. The device has not
972 * had its reference counter increased so the caller must be careful
973 * about locking. The caller must hold either the RTNL semaphore
977 struct net_device *__dev_get_by_index(struct net *net, int ifindex)
979 struct net_device *dev;
980 struct hlist_head *head = dev_index_hash(net, ifindex);
982 hlist_for_each_entry(dev, head, index_hlist)
983 if (dev->ifindex == ifindex)
988 EXPORT_SYMBOL(__dev_get_by_index);
991 * dev_get_by_index_rcu - find a device by its ifindex
992 * @net: the applicable net namespace
993 * @ifindex: index of device
995 * Search for an interface by index. Returns %NULL if the device
996 * is not found or a pointer to the device. The device has not
997 * had its reference counter increased so the caller must be careful
998 * about locking. The caller must hold RCU lock.
1001 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex)
1003 struct net_device *dev;
1004 struct hlist_head *head = dev_index_hash(net, ifindex);
1006 hlist_for_each_entry_rcu(dev, head, index_hlist)
1007 if (dev->ifindex == ifindex)
1012 EXPORT_SYMBOL(dev_get_by_index_rcu);
1016 * dev_get_by_index - find a device by its ifindex
1017 * @net: the applicable net namespace
1018 * @ifindex: index of device
1020 * Search for an interface by index. Returns NULL if the device
1021 * is not found or a pointer to the device. The device returned has
1022 * had a reference added and the pointer is safe until the user calls
1023 * dev_put to indicate they have finished with it.
1026 struct net_device *dev_get_by_index(struct net *net, int ifindex)
1028 struct net_device *dev;
1031 dev = dev_get_by_index_rcu(net, ifindex);
1037 EXPORT_SYMBOL(dev_get_by_index);
1040 * dev_get_by_napi_id - find a device by napi_id
1041 * @napi_id: ID of the NAPI struct
1043 * Search for an interface by NAPI ID. Returns %NULL if the device
1044 * is not found or a pointer to the device. The device has not had
1045 * its reference counter increased so the caller must be careful
1046 * about locking. The caller must hold RCU lock.
1049 struct net_device *dev_get_by_napi_id(unsigned int napi_id)
1051 struct napi_struct *napi;
1053 WARN_ON_ONCE(!rcu_read_lock_held());
1055 if (napi_id < MIN_NAPI_ID)
1058 napi = napi_by_id(napi_id);
1060 return napi ? napi->dev : NULL;
1062 EXPORT_SYMBOL(dev_get_by_napi_id);
1065 * netdev_get_name - get a netdevice name, knowing its ifindex.
1066 * @net: network namespace
1067 * @name: a pointer to the buffer where the name will be stored.
1068 * @ifindex: the ifindex of the interface to get the name from.
1070 int netdev_get_name(struct net *net, char *name, int ifindex)
1072 struct net_device *dev;
1075 down_read(&devnet_rename_sem);
1078 dev = dev_get_by_index_rcu(net, ifindex);
1084 strcpy(name, dev->name);
1089 up_read(&devnet_rename_sem);
1094 * dev_getbyhwaddr_rcu - find a device by its hardware address
1095 * @net: the applicable net namespace
1096 * @type: media type of device
1097 * @ha: hardware address
1099 * Search for an interface by MAC address. Returns NULL if the device
1100 * is not found or a pointer to the device.
1101 * The caller must hold RCU or RTNL.
1102 * The returned device has not had its ref count increased
1103 * and the caller must therefore be careful about locking
1107 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type,
1110 struct net_device *dev;
1112 for_each_netdev_rcu(net, dev)
1113 if (dev->type == type &&
1114 !memcmp(dev->dev_addr, ha, dev->addr_len))
1119 EXPORT_SYMBOL(dev_getbyhwaddr_rcu);
1121 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type)
1123 struct net_device *dev, *ret = NULL;
1126 for_each_netdev_rcu(net, dev)
1127 if (dev->type == type) {
1135 EXPORT_SYMBOL(dev_getfirstbyhwtype);
1138 * __dev_get_by_flags - find any device with given flags
1139 * @net: the applicable net namespace
1140 * @if_flags: IFF_* values
1141 * @mask: bitmask of bits in if_flags to check
1143 * Search for any interface with the given flags. Returns NULL if a device
1144 * is not found or a pointer to the device. Must be called inside
1145 * rtnl_lock(), and result refcount is unchanged.
1148 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags,
1149 unsigned short mask)
1151 struct net_device *dev, *ret;
1156 for_each_netdev(net, dev) {
1157 if (((dev->flags ^ if_flags) & mask) == 0) {
1164 EXPORT_SYMBOL(__dev_get_by_flags);
1167 * dev_valid_name - check if name is okay for network device
1168 * @name: name string
1170 * Network device names need to be valid file names to
1171 * allow sysfs to work. We also disallow any kind of
1174 bool dev_valid_name(const char *name)
1178 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ)
1180 if (!strcmp(name, ".") || !strcmp(name, ".."))
1184 if (*name == '/' || *name == ':' || isspace(*name))
1190 EXPORT_SYMBOL(dev_valid_name);
1193 * __dev_alloc_name - allocate a name for a device
1194 * @net: network namespace to allocate the device name in
1195 * @name: name format string
1196 * @buf: scratch buffer and result name string
1198 * Passed a format string - eg "lt%d" it will try and find a suitable
1199 * id. It scans list of devices to build up a free map, then chooses
1200 * the first empty slot. The caller must hold the dev_base or rtnl lock
1201 * while allocating the name and adding the device in order to avoid
1203 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1204 * Returns the number of the unit assigned or a negative errno code.
1207 static int __dev_alloc_name(struct net *net, const char *name, char *buf)
1211 const int max_netdevices = 8*PAGE_SIZE;
1212 unsigned long *inuse;
1213 struct net_device *d;
1215 if (!dev_valid_name(name))
1218 p = strchr(name, '%');
1221 * Verify the string as this thing may have come from
1222 * the user. There must be either one "%d" and no other "%"
1225 if (p[1] != 'd' || strchr(p + 2, '%'))
1228 /* Use one page as a bit array of possible slots */
1229 inuse = (unsigned long *) get_zeroed_page(GFP_ATOMIC);
1233 for_each_netdev(net, d) {
1234 struct netdev_name_node *name_node;
1235 list_for_each_entry(name_node, &d->name_node->list, list) {
1236 if (!sscanf(name_node->name, name, &i))
1238 if (i < 0 || i >= max_netdevices)
1241 /* avoid cases where sscanf is not exact inverse of printf */
1242 snprintf(buf, IFNAMSIZ, name, i);
1243 if (!strncmp(buf, name_node->name, IFNAMSIZ))
1246 if (!sscanf(d->name, name, &i))
1248 if (i < 0 || i >= max_netdevices)
1251 /* avoid cases where sscanf is not exact inverse of printf */
1252 snprintf(buf, IFNAMSIZ, name, i);
1253 if (!strncmp(buf, d->name, IFNAMSIZ))
1257 i = find_first_zero_bit(inuse, max_netdevices);
1258 free_page((unsigned long) inuse);
1261 snprintf(buf, IFNAMSIZ, name, i);
1262 if (!__dev_get_by_name(net, buf))
1265 /* It is possible to run out of possible slots
1266 * when the name is long and there isn't enough space left
1267 * for the digits, or if all bits are used.
1272 static int dev_alloc_name_ns(struct net *net,
1273 struct net_device *dev,
1280 ret = __dev_alloc_name(net, name, buf);
1282 strlcpy(dev->name, buf, IFNAMSIZ);
1287 * dev_alloc_name - allocate a name for a device
1289 * @name: name format string
1291 * Passed a format string - eg "lt%d" it will try and find a suitable
1292 * id. It scans list of devices to build up a free map, then chooses
1293 * the first empty slot. The caller must hold the dev_base or rtnl lock
1294 * while allocating the name and adding the device in order to avoid
1296 * Limited to bits_per_byte * page size devices (ie 32K on most platforms).
1297 * Returns the number of the unit assigned or a negative errno code.
1300 int dev_alloc_name(struct net_device *dev, const char *name)
1302 return dev_alloc_name_ns(dev_net(dev), dev, name);
1304 EXPORT_SYMBOL(dev_alloc_name);
1306 static int dev_get_valid_name(struct net *net, struct net_device *dev,
1311 if (!dev_valid_name(name))
1314 if (strchr(name, '%'))
1315 return dev_alloc_name_ns(net, dev, name);
1316 else if (__dev_get_by_name(net, name))
1318 else if (dev->name != name)
1319 strlcpy(dev->name, name, IFNAMSIZ);
1325 * dev_change_name - change name of a device
1327 * @newname: name (or format string) must be at least IFNAMSIZ
1329 * Change name of a device, can pass format strings "eth%d".
1332 int dev_change_name(struct net_device *dev, const char *newname)
1334 unsigned char old_assign_type;
1335 char oldname[IFNAMSIZ];
1341 BUG_ON(!dev_net(dev));
1345 /* Some auto-enslaved devices e.g. failover slaves are
1346 * special, as userspace might rename the device after
1347 * the interface had been brought up and running since
1348 * the point kernel initiated auto-enslavement. Allow
1349 * live name change even when these slave devices are
1352 * Typically, users of these auto-enslaving devices
1353 * don't actually care about slave name change, as
1354 * they are supposed to operate on master interface
1357 if (dev->flags & IFF_UP &&
1358 likely(!(dev->priv_flags & IFF_LIVE_RENAME_OK)))
1361 down_write(&devnet_rename_sem);
1363 if (strncmp(newname, dev->name, IFNAMSIZ) == 0) {
1364 up_write(&devnet_rename_sem);
1368 memcpy(oldname, dev->name, IFNAMSIZ);
1370 err = dev_get_valid_name(net, dev, newname);
1372 up_write(&devnet_rename_sem);
1376 if (oldname[0] && !strchr(oldname, '%'))
1377 netdev_info(dev, "renamed from %s\n", oldname);
1379 old_assign_type = dev->name_assign_type;
1380 dev->name_assign_type = NET_NAME_RENAMED;
1383 ret = device_rename(&dev->dev, dev->name);
1385 memcpy(dev->name, oldname, IFNAMSIZ);
1386 dev->name_assign_type = old_assign_type;
1387 up_write(&devnet_rename_sem);
1391 up_write(&devnet_rename_sem);
1393 netdev_adjacent_rename_links(dev, oldname);
1395 write_lock_bh(&dev_base_lock);
1396 netdev_name_node_del(dev->name_node);
1397 write_unlock_bh(&dev_base_lock);
1401 write_lock_bh(&dev_base_lock);
1402 netdev_name_node_add(net, dev->name_node);
1403 write_unlock_bh(&dev_base_lock);
1405 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev);
1406 ret = notifier_to_errno(ret);
1409 /* err >= 0 after dev_alloc_name() or stores the first errno */
1412 down_write(&devnet_rename_sem);
1413 memcpy(dev->name, oldname, IFNAMSIZ);
1414 memcpy(oldname, newname, IFNAMSIZ);
1415 dev->name_assign_type = old_assign_type;
1416 old_assign_type = NET_NAME_RENAMED;
1419 pr_err("%s: name change rollback failed: %d\n",
1428 * dev_set_alias - change ifalias of a device
1430 * @alias: name up to IFALIASZ
1431 * @len: limit of bytes to copy from info
1433 * Set ifalias for a device,
1435 int dev_set_alias(struct net_device *dev, const char *alias, size_t len)
1437 struct dev_ifalias *new_alias = NULL;
1439 if (len >= IFALIASZ)
1443 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL);
1447 memcpy(new_alias->ifalias, alias, len);
1448 new_alias->ifalias[len] = 0;
1451 mutex_lock(&ifalias_mutex);
1452 new_alias = rcu_replace_pointer(dev->ifalias, new_alias,
1453 mutex_is_locked(&ifalias_mutex));
1454 mutex_unlock(&ifalias_mutex);
1457 kfree_rcu(new_alias, rcuhead);
1461 EXPORT_SYMBOL(dev_set_alias);
1464 * dev_get_alias - get ifalias of a device
1466 * @name: buffer to store name of ifalias
1467 * @len: size of buffer
1469 * get ifalias for a device. Caller must make sure dev cannot go
1470 * away, e.g. rcu read lock or own a reference count to device.
1472 int dev_get_alias(const struct net_device *dev, char *name, size_t len)
1474 const struct dev_ifalias *alias;
1478 alias = rcu_dereference(dev->ifalias);
1480 ret = snprintf(name, len, "%s", alias->ifalias);
1487 * netdev_features_change - device changes features
1488 * @dev: device to cause notification
1490 * Called to indicate a device has changed features.
1492 void netdev_features_change(struct net_device *dev)
1494 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev);
1496 EXPORT_SYMBOL(netdev_features_change);
1499 * netdev_state_change - device changes state
1500 * @dev: device to cause notification
1502 * Called to indicate a device has changed state. This function calls
1503 * the notifier chains for netdev_chain and sends a NEWLINK message
1504 * to the routing socket.
1506 void netdev_state_change(struct net_device *dev)
1508 if (dev->flags & IFF_UP) {
1509 struct netdev_notifier_change_info change_info = {
1513 call_netdevice_notifiers_info(NETDEV_CHANGE,
1515 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL);
1518 EXPORT_SYMBOL(netdev_state_change);
1521 * __netdev_notify_peers - notify network peers about existence of @dev,
1522 * to be called when rtnl lock is already held.
1523 * @dev: network device
1525 * Generate traffic such that interested network peers are aware of
1526 * @dev, such as by generating a gratuitous ARP. This may be used when
1527 * a device wants to inform the rest of the network about some sort of
1528 * reconfiguration such as a failover event or virtual machine
1531 void __netdev_notify_peers(struct net_device *dev)
1534 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev);
1535 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev);
1537 EXPORT_SYMBOL(__netdev_notify_peers);
1540 * netdev_notify_peers - notify network peers about existence of @dev
1541 * @dev: network device
1543 * Generate traffic such that interested network peers are aware of
1544 * @dev, such as by generating a gratuitous ARP. This may be used when
1545 * a device wants to inform the rest of the network about some sort of
1546 * reconfiguration such as a failover event or virtual machine
1549 void netdev_notify_peers(struct net_device *dev)
1552 __netdev_notify_peers(dev);
1555 EXPORT_SYMBOL(netdev_notify_peers);
1557 static int napi_threaded_poll(void *data);
1559 static int napi_kthread_create(struct napi_struct *n)
1563 /* Create and wake up the kthread once to put it in
1564 * TASK_INTERRUPTIBLE mode to avoid the blocked task
1565 * warning and work with loadavg.
1567 n->thread = kthread_run(napi_threaded_poll, n, "napi/%s-%d",
1568 n->dev->name, n->napi_id);
1569 if (IS_ERR(n->thread)) {
1570 err = PTR_ERR(n->thread);
1571 pr_err("kthread_run failed with err %d\n", err);
1578 static int __dev_open(struct net_device *dev, struct netlink_ext_ack *extack)
1580 const struct net_device_ops *ops = dev->netdev_ops;
1585 if (!netif_device_present(dev)) {
1586 /* may be detached because parent is runtime-suspended */
1587 if (dev->dev.parent)
1588 pm_runtime_resume(dev->dev.parent);
1589 if (!netif_device_present(dev))
1593 /* Block netpoll from trying to do any rx path servicing.
1594 * If we don't do this there is a chance ndo_poll_controller
1595 * or ndo_poll may be running while we open the device
1597 netpoll_poll_disable(dev);
1599 ret = call_netdevice_notifiers_extack(NETDEV_PRE_UP, dev, extack);
1600 ret = notifier_to_errno(ret);
1604 set_bit(__LINK_STATE_START, &dev->state);
1606 if (ops->ndo_validate_addr)
1607 ret = ops->ndo_validate_addr(dev);
1609 if (!ret && ops->ndo_open)
1610 ret = ops->ndo_open(dev);
1612 netpoll_poll_enable(dev);
1615 clear_bit(__LINK_STATE_START, &dev->state);
1617 dev->flags |= IFF_UP;
1618 dev_set_rx_mode(dev);
1620 add_device_randomness(dev->dev_addr, dev->addr_len);
1627 * dev_open - prepare an interface for use.
1628 * @dev: device to open
1629 * @extack: netlink extended ack
1631 * Takes a device from down to up state. The device's private open
1632 * function is invoked and then the multicast lists are loaded. Finally
1633 * the device is moved into the up state and a %NETDEV_UP message is
1634 * sent to the netdev notifier chain.
1636 * Calling this function on an active interface is a nop. On a failure
1637 * a negative errno code is returned.
1639 int dev_open(struct net_device *dev, struct netlink_ext_ack *extack)
1643 if (dev->flags & IFF_UP)
1646 ret = __dev_open(dev, extack);
1650 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1651 call_netdevice_notifiers(NETDEV_UP, dev);
1655 EXPORT_SYMBOL(dev_open);
1657 static void __dev_close_many(struct list_head *head)
1659 struct net_device *dev;
1664 list_for_each_entry(dev, head, close_list) {
1665 /* Temporarily disable netpoll until the interface is down */
1666 netpoll_poll_disable(dev);
1668 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev);
1670 clear_bit(__LINK_STATE_START, &dev->state);
1672 /* Synchronize to scheduled poll. We cannot touch poll list, it
1673 * can be even on different cpu. So just clear netif_running().
1675 * dev->stop() will invoke napi_disable() on all of it's
1676 * napi_struct instances on this device.
1678 smp_mb__after_atomic(); /* Commit netif_running(). */
1681 dev_deactivate_many(head);
1683 list_for_each_entry(dev, head, close_list) {
1684 const struct net_device_ops *ops = dev->netdev_ops;
1687 * Call the device specific close. This cannot fail.
1688 * Only if device is UP
1690 * We allow it to be called even after a DETACH hot-plug
1696 dev->flags &= ~IFF_UP;
1697 netpoll_poll_enable(dev);
1701 static void __dev_close(struct net_device *dev)
1705 list_add(&dev->close_list, &single);
1706 __dev_close_many(&single);
1710 void dev_close_many(struct list_head *head, bool unlink)
1712 struct net_device *dev, *tmp;
1714 /* Remove the devices that don't need to be closed */
1715 list_for_each_entry_safe(dev, tmp, head, close_list)
1716 if (!(dev->flags & IFF_UP))
1717 list_del_init(&dev->close_list);
1719 __dev_close_many(head);
1721 list_for_each_entry_safe(dev, tmp, head, close_list) {
1722 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP|IFF_RUNNING, GFP_KERNEL);
1723 call_netdevice_notifiers(NETDEV_DOWN, dev);
1725 list_del_init(&dev->close_list);
1728 EXPORT_SYMBOL(dev_close_many);
1731 * dev_close - shutdown an interface.
1732 * @dev: device to shutdown
1734 * This function moves an active device into down state. A
1735 * %NETDEV_GOING_DOWN is sent to the netdev notifier chain. The device
1736 * is then deactivated and finally a %NETDEV_DOWN is sent to the notifier
1739 void dev_close(struct net_device *dev)
1741 if (dev->flags & IFF_UP) {
1744 list_add(&dev->close_list, &single);
1745 dev_close_many(&single, true);
1749 EXPORT_SYMBOL(dev_close);
1753 * dev_disable_lro - disable Large Receive Offload on a device
1756 * Disable Large Receive Offload (LRO) on a net device. Must be
1757 * called under RTNL. This is needed if received packets may be
1758 * forwarded to another interface.
1760 void dev_disable_lro(struct net_device *dev)
1762 struct net_device *lower_dev;
1763 struct list_head *iter;
1765 dev->wanted_features &= ~NETIF_F_LRO;
1766 netdev_update_features(dev);
1768 if (unlikely(dev->features & NETIF_F_LRO))
1769 netdev_WARN(dev, "failed to disable LRO!\n");
1771 netdev_for_each_lower_dev(dev, lower_dev, iter)
1772 dev_disable_lro(lower_dev);
1774 EXPORT_SYMBOL(dev_disable_lro);
1777 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device
1780 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be
1781 * called under RTNL. This is needed if Generic XDP is installed on
1784 static void dev_disable_gro_hw(struct net_device *dev)
1786 dev->wanted_features &= ~NETIF_F_GRO_HW;
1787 netdev_update_features(dev);
1789 if (unlikely(dev->features & NETIF_F_GRO_HW))
1790 netdev_WARN(dev, "failed to disable GRO_HW!\n");
1793 const char *netdev_cmd_to_name(enum netdev_cmd cmd)
1796 case NETDEV_##val: \
1797 return "NETDEV_" __stringify(val);
1799 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER)
1800 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE)
1801 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE)
1802 N(POST_INIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) N(CHANGEUPPER)
1803 N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) N(BONDING_INFO)
1804 N(PRECHANGEUPPER) N(CHANGELOWERSTATE) N(UDP_TUNNEL_PUSH_INFO)
1805 N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN)
1806 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO)
1807 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO)
1811 return "UNKNOWN_NETDEV_EVENT";
1813 EXPORT_SYMBOL_GPL(netdev_cmd_to_name);
1815 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val,
1816 struct net_device *dev)
1818 struct netdev_notifier_info info = {
1822 return nb->notifier_call(nb, val, &info);
1825 static int call_netdevice_register_notifiers(struct notifier_block *nb,
1826 struct net_device *dev)
1830 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev);
1831 err = notifier_to_errno(err);
1835 if (!(dev->flags & IFF_UP))
1838 call_netdevice_notifier(nb, NETDEV_UP, dev);
1842 static void call_netdevice_unregister_notifiers(struct notifier_block *nb,
1843 struct net_device *dev)
1845 if (dev->flags & IFF_UP) {
1846 call_netdevice_notifier(nb, NETDEV_GOING_DOWN,
1848 call_netdevice_notifier(nb, NETDEV_DOWN, dev);
1850 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev);
1853 static int call_netdevice_register_net_notifiers(struct notifier_block *nb,
1856 struct net_device *dev;
1859 for_each_netdev(net, dev) {
1860 err = call_netdevice_register_notifiers(nb, dev);
1867 for_each_netdev_continue_reverse(net, dev)
1868 call_netdevice_unregister_notifiers(nb, dev);
1872 static void call_netdevice_unregister_net_notifiers(struct notifier_block *nb,
1875 struct net_device *dev;
1877 for_each_netdev(net, dev)
1878 call_netdevice_unregister_notifiers(nb, dev);
1881 static int dev_boot_phase = 1;
1884 * register_netdevice_notifier - register a network notifier block
1887 * Register a notifier to be called when network device events occur.
1888 * The notifier passed is linked into the kernel structures and must
1889 * not be reused until it has been unregistered. A negative errno code
1890 * is returned on a failure.
1892 * When registered all registration and up events are replayed
1893 * to the new notifier to allow device to have a race free
1894 * view of the network device list.
1897 int register_netdevice_notifier(struct notifier_block *nb)
1902 /* Close race with setup_net() and cleanup_net() */
1903 down_write(&pernet_ops_rwsem);
1905 err = raw_notifier_chain_register(&netdev_chain, nb);
1911 err = call_netdevice_register_net_notifiers(nb, net);
1918 up_write(&pernet_ops_rwsem);
1922 for_each_net_continue_reverse(net)
1923 call_netdevice_unregister_net_notifiers(nb, net);
1925 raw_notifier_chain_unregister(&netdev_chain, nb);
1928 EXPORT_SYMBOL(register_netdevice_notifier);
1931 * unregister_netdevice_notifier - unregister a network notifier block
1934 * Unregister a notifier previously registered by
1935 * register_netdevice_notifier(). The notifier is unlinked into the
1936 * kernel structures and may then be reused. A negative errno code
1937 * is returned on a failure.
1939 * After unregistering unregister and down device events are synthesized
1940 * for all devices on the device list to the removed notifier to remove
1941 * the need for special case cleanup code.
1944 int unregister_netdevice_notifier(struct notifier_block *nb)
1949 /* Close race with setup_net() and cleanup_net() */
1950 down_write(&pernet_ops_rwsem);
1952 err = raw_notifier_chain_unregister(&netdev_chain, nb);
1957 call_netdevice_unregister_net_notifiers(nb, net);
1961 up_write(&pernet_ops_rwsem);
1964 EXPORT_SYMBOL(unregister_netdevice_notifier);
1966 static int __register_netdevice_notifier_net(struct net *net,
1967 struct notifier_block *nb,
1968 bool ignore_call_fail)
1972 err = raw_notifier_chain_register(&net->netdev_chain, nb);
1978 err = call_netdevice_register_net_notifiers(nb, net);
1979 if (err && !ignore_call_fail)
1980 goto chain_unregister;
1985 raw_notifier_chain_unregister(&net->netdev_chain, nb);
1989 static int __unregister_netdevice_notifier_net(struct net *net,
1990 struct notifier_block *nb)
1994 err = raw_notifier_chain_unregister(&net->netdev_chain, nb);
1998 call_netdevice_unregister_net_notifiers(nb, net);
2003 * register_netdevice_notifier_net - register a per-netns network notifier block
2004 * @net: network namespace
2007 * Register a notifier to be called when network device events occur.
2008 * The notifier passed is linked into the kernel structures and must
2009 * not be reused until it has been unregistered. A negative errno code
2010 * is returned on a failure.
2012 * When registered all registration and up events are replayed
2013 * to the new notifier to allow device to have a race free
2014 * view of the network device list.
2017 int register_netdevice_notifier_net(struct net *net, struct notifier_block *nb)
2022 err = __register_netdevice_notifier_net(net, nb, false);
2026 EXPORT_SYMBOL(register_netdevice_notifier_net);
2029 * unregister_netdevice_notifier_net - unregister a per-netns
2030 * network notifier block
2031 * @net: network namespace
2034 * Unregister a notifier previously registered by
2035 * register_netdevice_notifier(). The notifier is unlinked into the
2036 * kernel structures and may then be reused. A negative errno code
2037 * is returned on a failure.
2039 * After unregistering unregister and down device events are synthesized
2040 * for all devices on the device list to the removed notifier to remove
2041 * the need for special case cleanup code.
2044 int unregister_netdevice_notifier_net(struct net *net,
2045 struct notifier_block *nb)
2050 err = __unregister_netdevice_notifier_net(net, nb);
2054 EXPORT_SYMBOL(unregister_netdevice_notifier_net);
2056 int register_netdevice_notifier_dev_net(struct net_device *dev,
2057 struct notifier_block *nb,
2058 struct netdev_net_notifier *nn)
2063 err = __register_netdevice_notifier_net(dev_net(dev), nb, false);
2066 list_add(&nn->list, &dev->net_notifier_list);
2071 EXPORT_SYMBOL(register_netdevice_notifier_dev_net);
2073 int unregister_netdevice_notifier_dev_net(struct net_device *dev,
2074 struct notifier_block *nb,
2075 struct netdev_net_notifier *nn)
2080 list_del(&nn->list);
2081 err = __unregister_netdevice_notifier_net(dev_net(dev), nb);
2085 EXPORT_SYMBOL(unregister_netdevice_notifier_dev_net);
2087 static void move_netdevice_notifiers_dev_net(struct net_device *dev,
2090 struct netdev_net_notifier *nn;
2092 list_for_each_entry(nn, &dev->net_notifier_list, list) {
2093 __unregister_netdevice_notifier_net(dev_net(dev), nn->nb);
2094 __register_netdevice_notifier_net(net, nn->nb, true);
2099 * call_netdevice_notifiers_info - call all network notifier blocks
2100 * @val: value passed unmodified to notifier function
2101 * @info: notifier information data
2103 * Call all network notifier blocks. Parameters and return value
2104 * are as for raw_notifier_call_chain().
2107 static int call_netdevice_notifiers_info(unsigned long val,
2108 struct netdev_notifier_info *info)
2110 struct net *net = dev_net(info->dev);
2115 /* Run per-netns notifier block chain first, then run the global one.
2116 * Hopefully, one day, the global one is going to be removed after
2117 * all notifier block registrators get converted to be per-netns.
2119 ret = raw_notifier_call_chain(&net->netdev_chain, val, info);
2120 if (ret & NOTIFY_STOP_MASK)
2122 return raw_notifier_call_chain(&netdev_chain, val, info);
2125 static int call_netdevice_notifiers_extack(unsigned long val,
2126 struct net_device *dev,
2127 struct netlink_ext_ack *extack)
2129 struct netdev_notifier_info info = {
2134 return call_netdevice_notifiers_info(val, &info);
2138 * call_netdevice_notifiers - call all network notifier blocks
2139 * @val: value passed unmodified to notifier function
2140 * @dev: net_device pointer passed unmodified to notifier function
2142 * Call all network notifier blocks. Parameters and return value
2143 * are as for raw_notifier_call_chain().
2146 int call_netdevice_notifiers(unsigned long val, struct net_device *dev)
2148 return call_netdevice_notifiers_extack(val, dev, NULL);
2150 EXPORT_SYMBOL(call_netdevice_notifiers);
2153 * call_netdevice_notifiers_mtu - call all network notifier blocks
2154 * @val: value passed unmodified to notifier function
2155 * @dev: net_device pointer passed unmodified to notifier function
2156 * @arg: additional u32 argument passed to the notifier function
2158 * Call all network notifier blocks. Parameters and return value
2159 * are as for raw_notifier_call_chain().
2161 static int call_netdevice_notifiers_mtu(unsigned long val,
2162 struct net_device *dev, u32 arg)
2164 struct netdev_notifier_info_ext info = {
2169 BUILD_BUG_ON(offsetof(struct netdev_notifier_info_ext, info) != 0);
2171 return call_netdevice_notifiers_info(val, &info.info);
2174 #ifdef CONFIG_NET_INGRESS
2175 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key);
2177 void net_inc_ingress_queue(void)
2179 static_branch_inc(&ingress_needed_key);
2181 EXPORT_SYMBOL_GPL(net_inc_ingress_queue);
2183 void net_dec_ingress_queue(void)
2185 static_branch_dec(&ingress_needed_key);
2187 EXPORT_SYMBOL_GPL(net_dec_ingress_queue);
2190 #ifdef CONFIG_NET_EGRESS
2191 static DEFINE_STATIC_KEY_FALSE(egress_needed_key);
2193 void net_inc_egress_queue(void)
2195 static_branch_inc(&egress_needed_key);
2197 EXPORT_SYMBOL_GPL(net_inc_egress_queue);
2199 void net_dec_egress_queue(void)
2201 static_branch_dec(&egress_needed_key);
2203 EXPORT_SYMBOL_GPL(net_dec_egress_queue);
2206 static DEFINE_STATIC_KEY_FALSE(netstamp_needed_key);
2207 #ifdef CONFIG_JUMP_LABEL
2208 static atomic_t netstamp_needed_deferred;
2209 static atomic_t netstamp_wanted;
2210 static void netstamp_clear(struct work_struct *work)
2212 int deferred = atomic_xchg(&netstamp_needed_deferred, 0);
2215 wanted = atomic_add_return(deferred, &netstamp_wanted);
2217 static_branch_enable(&netstamp_needed_key);
2219 static_branch_disable(&netstamp_needed_key);
2221 static DECLARE_WORK(netstamp_work, netstamp_clear);
2224 void net_enable_timestamp(void)
2226 #ifdef CONFIG_JUMP_LABEL
2230 wanted = atomic_read(&netstamp_wanted);
2233 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted + 1) == wanted)
2236 atomic_inc(&netstamp_needed_deferred);
2237 schedule_work(&netstamp_work);
2239 static_branch_inc(&netstamp_needed_key);
2242 EXPORT_SYMBOL(net_enable_timestamp);
2244 void net_disable_timestamp(void)
2246 #ifdef CONFIG_JUMP_LABEL
2250 wanted = atomic_read(&netstamp_wanted);
2253 if (atomic_cmpxchg(&netstamp_wanted, wanted, wanted - 1) == wanted)
2256 atomic_dec(&netstamp_needed_deferred);
2257 schedule_work(&netstamp_work);
2259 static_branch_dec(&netstamp_needed_key);
2262 EXPORT_SYMBOL(net_disable_timestamp);
2264 static inline void net_timestamp_set(struct sk_buff *skb)
2267 if (static_branch_unlikely(&netstamp_needed_key))
2268 __net_timestamp(skb);
2271 #define net_timestamp_check(COND, SKB) \
2272 if (static_branch_unlikely(&netstamp_needed_key)) { \
2273 if ((COND) && !(SKB)->tstamp) \
2274 __net_timestamp(SKB); \
2277 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb)
2279 return __is_skb_forwardable(dev, skb, true);
2281 EXPORT_SYMBOL_GPL(is_skb_forwardable);
2283 static int __dev_forward_skb2(struct net_device *dev, struct sk_buff *skb,
2286 int ret = ____dev_forward_skb(dev, skb, check_mtu);
2289 skb->protocol = eth_type_trans(skb, dev);
2290 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN);
2296 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
2298 return __dev_forward_skb2(dev, skb, true);
2300 EXPORT_SYMBOL_GPL(__dev_forward_skb);
2303 * dev_forward_skb - loopback an skb to another netif
2305 * @dev: destination network device
2306 * @skb: buffer to forward
2309 * NET_RX_SUCCESS (no congestion)
2310 * NET_RX_DROP (packet was dropped, but freed)
2312 * dev_forward_skb can be used for injecting an skb from the
2313 * start_xmit function of one device into the receive queue
2314 * of another device.
2316 * The receiving device may be in another namespace, so
2317 * we have to clear all information in the skb that could
2318 * impact namespace isolation.
2320 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb)
2322 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb);
2324 EXPORT_SYMBOL_GPL(dev_forward_skb);
2326 int dev_forward_skb_nomtu(struct net_device *dev, struct sk_buff *skb)
2328 return __dev_forward_skb2(dev, skb, false) ?: netif_rx_internal(skb);
2331 static inline int deliver_skb(struct sk_buff *skb,
2332 struct packet_type *pt_prev,
2333 struct net_device *orig_dev)
2335 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
2337 refcount_inc(&skb->users);
2338 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
2341 static inline void deliver_ptype_list_skb(struct sk_buff *skb,
2342 struct packet_type **pt,
2343 struct net_device *orig_dev,
2345 struct list_head *ptype_list)
2347 struct packet_type *ptype, *pt_prev = *pt;
2349 list_for_each_entry_rcu(ptype, ptype_list, list) {
2350 if (ptype->type != type)
2353 deliver_skb(skb, pt_prev, orig_dev);
2359 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb)
2361 if (!ptype->af_packet_priv || !skb->sk)
2364 if (ptype->id_match)
2365 return ptype->id_match(ptype, skb->sk);
2366 else if ((struct sock *)ptype->af_packet_priv == skb->sk)
2373 * dev_nit_active - return true if any network interface taps are in use
2375 * @dev: network device to check for the presence of taps
2377 bool dev_nit_active(struct net_device *dev)
2379 return !list_empty(&ptype_all) || !list_empty(&dev->ptype_all);
2381 EXPORT_SYMBOL_GPL(dev_nit_active);
2384 * Support routine. Sends outgoing frames to any network
2385 * taps currently in use.
2388 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev)
2390 struct packet_type *ptype;
2391 struct sk_buff *skb2 = NULL;
2392 struct packet_type *pt_prev = NULL;
2393 struct list_head *ptype_list = &ptype_all;
2397 list_for_each_entry_rcu(ptype, ptype_list, list) {
2398 if (ptype->ignore_outgoing)
2401 /* Never send packets back to the socket
2402 * they originated from - MvS (miquels@drinkel.ow.org)
2404 if (skb_loop_sk(ptype, skb))
2408 deliver_skb(skb2, pt_prev, skb->dev);
2413 /* need to clone skb, done only once */
2414 skb2 = skb_clone(skb, GFP_ATOMIC);
2418 net_timestamp_set(skb2);
2420 /* skb->nh should be correctly
2421 * set by sender, so that the second statement is
2422 * just protection against buggy protocols.
2424 skb_reset_mac_header(skb2);
2426 if (skb_network_header(skb2) < skb2->data ||
2427 skb_network_header(skb2) > skb_tail_pointer(skb2)) {
2428 net_crit_ratelimited("protocol %04x is buggy, dev %s\n",
2429 ntohs(skb2->protocol),
2431 skb_reset_network_header(skb2);
2434 skb2->transport_header = skb2->network_header;
2435 skb2->pkt_type = PACKET_OUTGOING;
2439 if (ptype_list == &ptype_all) {
2440 ptype_list = &dev->ptype_all;
2445 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC))
2446 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev);
2452 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit);
2455 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change
2456 * @dev: Network device
2457 * @txq: number of queues available
2459 * If real_num_tx_queues is changed the tc mappings may no longer be
2460 * valid. To resolve this verify the tc mapping remains valid and if
2461 * not NULL the mapping. With no priorities mapping to this
2462 * offset/count pair it will no longer be used. In the worst case TC0
2463 * is invalid nothing can be done so disable priority mappings. If is
2464 * expected that drivers will fix this mapping if they can before
2465 * calling netif_set_real_num_tx_queues.
2467 static void netif_setup_tc(struct net_device *dev, unsigned int txq)
2470 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2472 /* If TC0 is invalidated disable TC mapping */
2473 if (tc->offset + tc->count > txq) {
2474 pr_warn("Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n");
2479 /* Invalidated prio to tc mappings set to TC0 */
2480 for (i = 1; i < TC_BITMASK + 1; i++) {
2481 int q = netdev_get_prio_tc_map(dev, i);
2483 tc = &dev->tc_to_txq[q];
2484 if (tc->offset + tc->count > txq) {
2485 pr_warn("Number of in use tx queues changed. Priority %i to tc mapping %i is no longer valid. Setting map to 0\n",
2487 netdev_set_prio_tc_map(dev, i, 0);
2492 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq)
2495 struct netdev_tc_txq *tc = &dev->tc_to_txq[0];
2498 /* walk through the TCs and see if it falls into any of them */
2499 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) {
2500 if ((txq - tc->offset) < tc->count)
2504 /* didn't find it, just return -1 to indicate no match */
2510 EXPORT_SYMBOL(netdev_txq_to_tc);
2513 static struct static_key xps_needed __read_mostly;
2514 static struct static_key xps_rxqs_needed __read_mostly;
2515 static DEFINE_MUTEX(xps_map_mutex);
2516 #define xmap_dereference(P) \
2517 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex))
2519 static bool remove_xps_queue(struct xps_dev_maps *dev_maps,
2520 struct xps_dev_maps *old_maps, int tci, u16 index)
2522 struct xps_map *map = NULL;
2526 map = xmap_dereference(dev_maps->attr_map[tci]);
2530 for (pos = map->len; pos--;) {
2531 if (map->queues[pos] != index)
2535 map->queues[pos] = map->queues[--map->len];
2540 RCU_INIT_POINTER(old_maps->attr_map[tci], NULL);
2541 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL);
2542 kfree_rcu(map, rcu);
2549 static bool remove_xps_queue_cpu(struct net_device *dev,
2550 struct xps_dev_maps *dev_maps,
2551 int cpu, u16 offset, u16 count)
2553 int num_tc = dev_maps->num_tc;
2554 bool active = false;
2557 for (tci = cpu * num_tc; num_tc--; tci++) {
2560 for (i = count, j = offset; i--; j++) {
2561 if (!remove_xps_queue(dev_maps, NULL, tci, j))
2571 static void reset_xps_maps(struct net_device *dev,
2572 struct xps_dev_maps *dev_maps,
2573 enum xps_map_type type)
2575 static_key_slow_dec_cpuslocked(&xps_needed);
2576 if (type == XPS_RXQS)
2577 static_key_slow_dec_cpuslocked(&xps_rxqs_needed);
2579 RCU_INIT_POINTER(dev->xps_maps[type], NULL);
2581 kfree_rcu(dev_maps, rcu);
2584 static void clean_xps_maps(struct net_device *dev, enum xps_map_type type,
2585 u16 offset, u16 count)
2587 struct xps_dev_maps *dev_maps;
2588 bool active = false;
2591 dev_maps = xmap_dereference(dev->xps_maps[type]);
2595 for (j = 0; j < dev_maps->nr_ids; j++)
2596 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, count);
2598 reset_xps_maps(dev, dev_maps, type);
2600 if (type == XPS_CPUS) {
2601 for (i = offset + (count - 1); count--; i--)
2602 netdev_queue_numa_node_write(
2603 netdev_get_tx_queue(dev, i), NUMA_NO_NODE);
2607 static void netif_reset_xps_queues(struct net_device *dev, u16 offset,
2610 if (!static_key_false(&xps_needed))
2614 mutex_lock(&xps_map_mutex);
2616 if (static_key_false(&xps_rxqs_needed))
2617 clean_xps_maps(dev, XPS_RXQS, offset, count);
2619 clean_xps_maps(dev, XPS_CPUS, offset, count);
2621 mutex_unlock(&xps_map_mutex);
2625 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index)
2627 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index);
2630 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index,
2631 u16 index, bool is_rxqs_map)
2633 struct xps_map *new_map;
2634 int alloc_len = XPS_MIN_MAP_ALLOC;
2637 for (pos = 0; map && pos < map->len; pos++) {
2638 if (map->queues[pos] != index)
2643 /* Need to add tx-queue to this CPU's/rx-queue's existing map */
2645 if (pos < map->alloc_len)
2648 alloc_len = map->alloc_len * 2;
2651 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's
2655 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL);
2657 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL,
2658 cpu_to_node(attr_index));
2662 for (i = 0; i < pos; i++)
2663 new_map->queues[i] = map->queues[i];
2664 new_map->alloc_len = alloc_len;
2670 /* Copy xps maps at a given index */
2671 static void xps_copy_dev_maps(struct xps_dev_maps *dev_maps,
2672 struct xps_dev_maps *new_dev_maps, int index,
2673 int tc, bool skip_tc)
2675 int i, tci = index * dev_maps->num_tc;
2676 struct xps_map *map;
2678 /* copy maps belonging to foreign traffic classes */
2679 for (i = 0; i < dev_maps->num_tc; i++, tci++) {
2680 if (i == tc && skip_tc)
2683 /* fill in the new device map from the old device map */
2684 map = xmap_dereference(dev_maps->attr_map[tci]);
2685 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map);
2689 /* Must be called under cpus_read_lock */
2690 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask,
2691 u16 index, enum xps_map_type type)
2693 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL, *old_dev_maps = NULL;
2694 const unsigned long *online_mask = NULL;
2695 bool active = false, copy = false;
2696 int i, j, tci, numa_node_id = -2;
2697 int maps_sz, num_tc = 1, tc = 0;
2698 struct xps_map *map, *new_map;
2699 unsigned int nr_ids;
2702 /* Do not allow XPS on subordinate device directly */
2703 num_tc = dev->num_tc;
2707 /* If queue belongs to subordinate dev use its map */
2708 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev;
2710 tc = netdev_txq_to_tc(dev, index);
2715 mutex_lock(&xps_map_mutex);
2717 dev_maps = xmap_dereference(dev->xps_maps[type]);
2718 if (type == XPS_RXQS) {
2719 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues);
2720 nr_ids = dev->num_rx_queues;
2722 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc);
2723 if (num_possible_cpus() > 1)
2724 online_mask = cpumask_bits(cpu_online_mask);
2725 nr_ids = nr_cpu_ids;
2728 if (maps_sz < L1_CACHE_BYTES)
2729 maps_sz = L1_CACHE_BYTES;
2731 /* The old dev_maps could be larger or smaller than the one we're
2732 * setting up now, as dev->num_tc or nr_ids could have been updated in
2733 * between. We could try to be smart, but let's be safe instead and only
2734 * copy foreign traffic classes if the two map sizes match.
2737 dev_maps->num_tc == num_tc && dev_maps->nr_ids == nr_ids)
2740 /* allocate memory for queue storage */
2741 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids),
2743 if (!new_dev_maps) {
2744 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL);
2745 if (!new_dev_maps) {
2746 mutex_unlock(&xps_map_mutex);
2750 new_dev_maps->nr_ids = nr_ids;
2751 new_dev_maps->num_tc = num_tc;
2754 tci = j * num_tc + tc;
2755 map = copy ? xmap_dereference(dev_maps->attr_map[tci]) : NULL;
2757 map = expand_xps_map(map, j, index, type == XPS_RXQS);
2761 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map);
2765 goto out_no_new_maps;
2768 /* Increment static keys at most once per type */
2769 static_key_slow_inc_cpuslocked(&xps_needed);
2770 if (type == XPS_RXQS)
2771 static_key_slow_inc_cpuslocked(&xps_rxqs_needed);
2774 for (j = 0; j < nr_ids; j++) {
2775 bool skip_tc = false;
2777 tci = j * num_tc + tc;
2778 if (netif_attr_test_mask(j, mask, nr_ids) &&
2779 netif_attr_test_online(j, online_mask, nr_ids)) {
2780 /* add tx-queue to CPU/rx-queue maps */
2785 map = xmap_dereference(new_dev_maps->attr_map[tci]);
2786 while ((pos < map->len) && (map->queues[pos] != index))
2789 if (pos == map->len)
2790 map->queues[map->len++] = index;
2792 if (type == XPS_CPUS) {
2793 if (numa_node_id == -2)
2794 numa_node_id = cpu_to_node(j);
2795 else if (numa_node_id != cpu_to_node(j))
2802 xps_copy_dev_maps(dev_maps, new_dev_maps, j, tc,
2806 rcu_assign_pointer(dev->xps_maps[type], new_dev_maps);
2808 /* Cleanup old maps */
2810 goto out_no_old_maps;
2812 for (j = 0; j < dev_maps->nr_ids; j++) {
2813 for (i = num_tc, tci = j * dev_maps->num_tc; i--; tci++) {
2814 map = xmap_dereference(dev_maps->attr_map[tci]);
2819 new_map = xmap_dereference(new_dev_maps->attr_map[tci]);
2824 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL);
2825 kfree_rcu(map, rcu);
2829 old_dev_maps = dev_maps;
2832 dev_maps = new_dev_maps;
2836 if (type == XPS_CPUS)
2837 /* update Tx queue numa node */
2838 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index),
2839 (numa_node_id >= 0) ?
2840 numa_node_id : NUMA_NO_NODE);
2845 /* removes tx-queue from unused CPUs/rx-queues */
2846 for (j = 0; j < dev_maps->nr_ids; j++) {
2847 tci = j * dev_maps->num_tc;
2849 for (i = 0; i < dev_maps->num_tc; i++, tci++) {
2851 netif_attr_test_mask(j, mask, dev_maps->nr_ids) &&
2852 netif_attr_test_online(j, online_mask, dev_maps->nr_ids))
2855 active |= remove_xps_queue(dev_maps,
2856 copy ? old_dev_maps : NULL,
2862 kfree_rcu(old_dev_maps, rcu);
2864 /* free map if not active */
2866 reset_xps_maps(dev, dev_maps, type);
2869 mutex_unlock(&xps_map_mutex);
2873 /* remove any maps that we added */
2874 for (j = 0; j < nr_ids; j++) {
2875 for (i = num_tc, tci = j * num_tc; i--; tci++) {
2876 new_map = xmap_dereference(new_dev_maps->attr_map[tci]);
2878 xmap_dereference(dev_maps->attr_map[tci]) :
2880 if (new_map && new_map != map)
2885 mutex_unlock(&xps_map_mutex);
2887 kfree(new_dev_maps);
2890 EXPORT_SYMBOL_GPL(__netif_set_xps_queue);
2892 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask,
2898 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, XPS_CPUS);
2903 EXPORT_SYMBOL(netif_set_xps_queue);
2906 static void netdev_unbind_all_sb_channels(struct net_device *dev)
2908 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues];
2910 /* Unbind any subordinate channels */
2911 while (txq-- != &dev->_tx[0]) {
2913 netdev_unbind_sb_channel(dev, txq->sb_dev);
2917 void netdev_reset_tc(struct net_device *dev)
2920 netif_reset_xps_queues_gt(dev, 0);
2922 netdev_unbind_all_sb_channels(dev);
2924 /* Reset TC configuration of device */
2926 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq));
2927 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map));
2929 EXPORT_SYMBOL(netdev_reset_tc);
2931 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset)
2933 if (tc >= dev->num_tc)
2937 netif_reset_xps_queues(dev, offset, count);
2939 dev->tc_to_txq[tc].count = count;
2940 dev->tc_to_txq[tc].offset = offset;
2943 EXPORT_SYMBOL(netdev_set_tc_queue);
2945 int netdev_set_num_tc(struct net_device *dev, u8 num_tc)
2947 if (num_tc > TC_MAX_QUEUE)
2951 netif_reset_xps_queues_gt(dev, 0);
2953 netdev_unbind_all_sb_channels(dev);
2955 dev->num_tc = num_tc;
2958 EXPORT_SYMBOL(netdev_set_num_tc);
2960 void netdev_unbind_sb_channel(struct net_device *dev,
2961 struct net_device *sb_dev)
2963 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues];
2966 netif_reset_xps_queues_gt(sb_dev, 0);
2968 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq));
2969 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map));
2971 while (txq-- != &dev->_tx[0]) {
2972 if (txq->sb_dev == sb_dev)
2976 EXPORT_SYMBOL(netdev_unbind_sb_channel);
2978 int netdev_bind_sb_channel_queue(struct net_device *dev,
2979 struct net_device *sb_dev,
2980 u8 tc, u16 count, u16 offset)
2982 /* Make certain the sb_dev and dev are already configured */
2983 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc)
2986 /* We cannot hand out queues we don't have */
2987 if ((offset + count) > dev->real_num_tx_queues)
2990 /* Record the mapping */
2991 sb_dev->tc_to_txq[tc].count = count;
2992 sb_dev->tc_to_txq[tc].offset = offset;
2994 /* Provide a way for Tx queue to find the tc_to_txq map or
2995 * XPS map for itself.
2998 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev;
3002 EXPORT_SYMBOL(netdev_bind_sb_channel_queue);
3004 int netdev_set_sb_channel(struct net_device *dev, u16 channel)
3006 /* Do not use a multiqueue device to represent a subordinate channel */
3007 if (netif_is_multiqueue(dev))
3010 /* We allow channels 1 - 32767 to be used for subordinate channels.
3011 * Channel 0 is meant to be "native" mode and used only to represent
3012 * the main root device. We allow writing 0 to reset the device back
3013 * to normal mode after being used as a subordinate channel.
3015 if (channel > S16_MAX)
3018 dev->num_tc = -channel;
3022 EXPORT_SYMBOL(netdev_set_sb_channel);
3025 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues
3026 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed.
3028 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq)
3033 disabling = txq < dev->real_num_tx_queues;
3035 if (txq < 1 || txq > dev->num_tx_queues)
3038 if (dev->reg_state == NETREG_REGISTERED ||
3039 dev->reg_state == NETREG_UNREGISTERING) {
3042 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues,
3048 netif_setup_tc(dev, txq);
3050 dev->real_num_tx_queues = txq;
3054 qdisc_reset_all_tx_gt(dev, txq);
3056 netif_reset_xps_queues_gt(dev, txq);
3060 dev->real_num_tx_queues = txq;
3065 EXPORT_SYMBOL(netif_set_real_num_tx_queues);
3069 * netif_set_real_num_rx_queues - set actual number of RX queues used
3070 * @dev: Network device
3071 * @rxq: Actual number of RX queues
3073 * This must be called either with the rtnl_lock held or before
3074 * registration of the net device. Returns 0 on success, or a
3075 * negative error code. If called before registration, it always
3078 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq)
3082 if (rxq < 1 || rxq > dev->num_rx_queues)
3085 if (dev->reg_state == NETREG_REGISTERED) {
3088 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues,
3094 dev->real_num_rx_queues = rxq;
3097 EXPORT_SYMBOL(netif_set_real_num_rx_queues);
3101 * netif_get_num_default_rss_queues - default number of RSS queues
3103 * This routine should set an upper limit on the number of RSS queues
3104 * used by default by multiqueue devices.
3106 int netif_get_num_default_rss_queues(void)
3108 return is_kdump_kernel() ?
3109 1 : min_t(int, DEFAULT_MAX_NUM_RSS_QUEUES, num_online_cpus());
3111 EXPORT_SYMBOL(netif_get_num_default_rss_queues);
3113 static void __netif_reschedule(struct Qdisc *q)
3115 struct softnet_data *sd;
3116 unsigned long flags;
3118 local_irq_save(flags);
3119 sd = this_cpu_ptr(&softnet_data);
3120 q->next_sched = NULL;
3121 *sd->output_queue_tailp = q;
3122 sd->output_queue_tailp = &q->next_sched;
3123 raise_softirq_irqoff(NET_TX_SOFTIRQ);
3124 local_irq_restore(flags);
3127 void __netif_schedule(struct Qdisc *q)
3129 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state))
3130 __netif_reschedule(q);
3132 EXPORT_SYMBOL(__netif_schedule);
3134 struct dev_kfree_skb_cb {
3135 enum skb_free_reason reason;
3138 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb)
3140 return (struct dev_kfree_skb_cb *)skb->cb;
3143 void netif_schedule_queue(struct netdev_queue *txq)
3146 if (!netif_xmit_stopped(txq)) {
3147 struct Qdisc *q = rcu_dereference(txq->qdisc);
3149 __netif_schedule(q);
3153 EXPORT_SYMBOL(netif_schedule_queue);
3155 void netif_tx_wake_queue(struct netdev_queue *dev_queue)
3157 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) {
3161 q = rcu_dereference(dev_queue->qdisc);
3162 __netif_schedule(q);
3166 EXPORT_SYMBOL(netif_tx_wake_queue);
3168 void __dev_kfree_skb_irq(struct sk_buff *skb, enum skb_free_reason reason)
3170 unsigned long flags;
3175 if (likely(refcount_read(&skb->users) == 1)) {
3177 refcount_set(&skb->users, 0);
3178 } else if (likely(!refcount_dec_and_test(&skb->users))) {
3181 get_kfree_skb_cb(skb)->reason = reason;
3182 local_irq_save(flags);
3183 skb->next = __this_cpu_read(softnet_data.completion_queue);
3184 __this_cpu_write(softnet_data.completion_queue, skb);
3185 raise_softirq_irqoff(NET_TX_SOFTIRQ);
3186 local_irq_restore(flags);
3188 EXPORT_SYMBOL(__dev_kfree_skb_irq);
3190 void __dev_kfree_skb_any(struct sk_buff *skb, enum skb_free_reason reason)
3192 if (in_irq() || irqs_disabled())
3193 __dev_kfree_skb_irq(skb, reason);
3197 EXPORT_SYMBOL(__dev_kfree_skb_any);
3201 * netif_device_detach - mark device as removed
3202 * @dev: network device
3204 * Mark device as removed from system and therefore no longer available.
3206 void netif_device_detach(struct net_device *dev)
3208 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) &&
3209 netif_running(dev)) {
3210 netif_tx_stop_all_queues(dev);
3213 EXPORT_SYMBOL(netif_device_detach);
3216 * netif_device_attach - mark device as attached
3217 * @dev: network device
3219 * Mark device as attached from system and restart if needed.
3221 void netif_device_attach(struct net_device *dev)
3223 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) &&
3224 netif_running(dev)) {
3225 netif_tx_wake_all_queues(dev);
3226 __netdev_watchdog_up(dev);
3229 EXPORT_SYMBOL(netif_device_attach);
3232 * Returns a Tx hash based on the given packet descriptor a Tx queues' number
3233 * to be used as a distribution range.
3235 static u16 skb_tx_hash(const struct net_device *dev,
3236 const struct net_device *sb_dev,
3237 struct sk_buff *skb)
3241 u16 qcount = dev->real_num_tx_queues;
3244 u8 tc = netdev_get_prio_tc_map(dev, skb->priority);
3246 qoffset = sb_dev->tc_to_txq[tc].offset;
3247 qcount = sb_dev->tc_to_txq[tc].count;
3250 if (skb_rx_queue_recorded(skb)) {
3251 hash = skb_get_rx_queue(skb);
3252 if (hash >= qoffset)
3254 while (unlikely(hash >= qcount))
3256 return hash + qoffset;
3259 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset;
3262 static void skb_warn_bad_offload(const struct sk_buff *skb)
3264 static const netdev_features_t null_features;
3265 struct net_device *dev = skb->dev;
3266 const char *name = "";
3268 if (!net_ratelimit())
3272 if (dev->dev.parent)
3273 name = dev_driver_string(dev->dev.parent);
3275 name = netdev_name(dev);
3277 skb_dump(KERN_WARNING, skb, false);
3278 WARN(1, "%s: caps=(%pNF, %pNF)\n",
3279 name, dev ? &dev->features : &null_features,
3280 skb->sk ? &skb->sk->sk_route_caps : &null_features);
3284 * Invalidate hardware checksum when packet is to be mangled, and
3285 * complete checksum manually on outgoing path.
3287 int skb_checksum_help(struct sk_buff *skb)
3290 int ret = 0, offset;
3292 if (skb->ip_summed == CHECKSUM_COMPLETE)
3293 goto out_set_summed;
3295 if (unlikely(skb_is_gso(skb))) {
3296 skb_warn_bad_offload(skb);
3300 /* Before computing a checksum, we should make sure no frag could
3301 * be modified by an external entity : checksum could be wrong.
3303 if (skb_has_shared_frag(skb)) {
3304 ret = __skb_linearize(skb);
3309 offset = skb_checksum_start_offset(skb);
3310 BUG_ON(offset >= skb_headlen(skb));
3311 csum = skb_checksum(skb, offset, skb->len - offset, 0);
3313 offset += skb->csum_offset;
3314 BUG_ON(offset + sizeof(__sum16) > skb_headlen(skb));
3316 ret = skb_ensure_writable(skb, offset + sizeof(__sum16));
3320 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0;
3322 skb->ip_summed = CHECKSUM_NONE;
3326 EXPORT_SYMBOL(skb_checksum_help);
3328 int skb_crc32c_csum_help(struct sk_buff *skb)
3331 int ret = 0, offset, start;
3333 if (skb->ip_summed != CHECKSUM_PARTIAL)
3336 if (unlikely(skb_is_gso(skb)))
3339 /* Before computing a checksum, we should make sure no frag could
3340 * be modified by an external entity : checksum could be wrong.
3342 if (unlikely(skb_has_shared_frag(skb))) {
3343 ret = __skb_linearize(skb);
3347 start = skb_checksum_start_offset(skb);
3348 offset = start + offsetof(struct sctphdr, checksum);
3349 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) {
3354 ret = skb_ensure_writable(skb, offset + sizeof(__le32));
3358 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start,
3359 skb->len - start, ~(__u32)0,
3361 *(__le32 *)(skb->data + offset) = crc32c_csum;
3362 skb->ip_summed = CHECKSUM_NONE;
3363 skb->csum_not_inet = 0;
3368 __be16 skb_network_protocol(struct sk_buff *skb, int *depth)
3370 __be16 type = skb->protocol;
3372 /* Tunnel gso handlers can set protocol to ethernet. */
3373 if (type == htons(ETH_P_TEB)) {
3376 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr))))
3379 eth = (struct ethhdr *)skb->data;
3380 type = eth->h_proto;
3383 return __vlan_get_protocol(skb, type, depth);
3387 * skb_mac_gso_segment - mac layer segmentation handler.
3388 * @skb: buffer to segment
3389 * @features: features for the output path (see dev->features)
3391 struct sk_buff *skb_mac_gso_segment(struct sk_buff *skb,
3392 netdev_features_t features)
3394 struct sk_buff *segs = ERR_PTR(-EPROTONOSUPPORT);
3395 struct packet_offload *ptype;
3396 int vlan_depth = skb->mac_len;
3397 __be16 type = skb_network_protocol(skb, &vlan_depth);
3399 if (unlikely(!type))
3400 return ERR_PTR(-EINVAL);
3402 __skb_pull(skb, vlan_depth);
3405 list_for_each_entry_rcu(ptype, &offload_base, list) {
3406 if (ptype->type == type && ptype->callbacks.gso_segment) {
3407 segs = ptype->callbacks.gso_segment(skb, features);
3413 __skb_push(skb, skb->data - skb_mac_header(skb));
3417 EXPORT_SYMBOL(skb_mac_gso_segment);
3420 /* openvswitch calls this on rx path, so we need a different check.
3422 static inline bool skb_needs_check(struct sk_buff *skb, bool tx_path)
3425 return skb->ip_summed != CHECKSUM_PARTIAL &&
3426 skb->ip_summed != CHECKSUM_UNNECESSARY;
3428 return skb->ip_summed == CHECKSUM_NONE;
3432 * __skb_gso_segment - Perform segmentation on skb.
3433 * @skb: buffer to segment
3434 * @features: features for the output path (see dev->features)
3435 * @tx_path: whether it is called in TX path
3437 * This function segments the given skb and returns a list of segments.
3439 * It may return NULL if the skb requires no segmentation. This is
3440 * only possible when GSO is used for verifying header integrity.
3442 * Segmentation preserves SKB_GSO_CB_OFFSET bytes of previous skb cb.
3444 struct sk_buff *__skb_gso_segment(struct sk_buff *skb,
3445 netdev_features_t features, bool tx_path)
3447 struct sk_buff *segs;
3449 if (unlikely(skb_needs_check(skb, tx_path))) {
3452 /* We're going to init ->check field in TCP or UDP header */
3453 err = skb_cow_head(skb, 0);
3455 return ERR_PTR(err);
3458 /* Only report GSO partial support if it will enable us to
3459 * support segmentation on this frame without needing additional
3462 if (features & NETIF_F_GSO_PARTIAL) {
3463 netdev_features_t partial_features = NETIF_F_GSO_ROBUST;
3464 struct net_device *dev = skb->dev;
3466 partial_features |= dev->features & dev->gso_partial_features;
3467 if (!skb_gso_ok(skb, features | partial_features))
3468 features &= ~NETIF_F_GSO_PARTIAL;
3471 BUILD_BUG_ON(SKB_GSO_CB_OFFSET +
3472 sizeof(*SKB_GSO_CB(skb)) > sizeof(skb->cb));
3474 SKB_GSO_CB(skb)->mac_offset = skb_headroom(skb);
3475 SKB_GSO_CB(skb)->encap_level = 0;
3477 skb_reset_mac_header(skb);
3478 skb_reset_mac_len(skb);
3480 segs = skb_mac_gso_segment(skb, features);
3482 if (segs != skb && unlikely(skb_needs_check(skb, tx_path) && !IS_ERR(segs)))
3483 skb_warn_bad_offload(skb);
3487 EXPORT_SYMBOL(__skb_gso_segment);
3489 /* Take action when hardware reception checksum errors are detected. */
3491 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb)
3493 pr_err("%s: hw csum failure\n", dev ? dev->name : "<unknown>");
3494 skb_dump(KERN_ERR, skb, true);
3498 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb)
3500 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb);
3502 EXPORT_SYMBOL(netdev_rx_csum_fault);
3505 /* XXX: check that highmem exists at all on the given machine. */
3506 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb)
3508 #ifdef CONFIG_HIGHMEM
3511 if (!(dev->features & NETIF_F_HIGHDMA)) {
3512 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
3513 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
3515 if (PageHighMem(skb_frag_page(frag)))
3523 /* If MPLS offload request, verify we are testing hardware MPLS features
3524 * instead of standard features for the netdev.
3526 #if IS_ENABLED(CONFIG_NET_MPLS_GSO)
3527 static netdev_features_t net_mpls_features(struct sk_buff *skb,
3528 netdev_features_t features,
3531 if (eth_p_mpls(type))
3532 features &= skb->dev->mpls_features;
3537 static netdev_features_t net_mpls_features(struct sk_buff *skb,
3538 netdev_features_t features,
3545 static netdev_features_t harmonize_features(struct sk_buff *skb,
3546 netdev_features_t features)
3550 type = skb_network_protocol(skb, NULL);
3551 features = net_mpls_features(skb, features, type);
3553 if (skb->ip_summed != CHECKSUM_NONE &&
3554 !can_checksum_protocol(features, type)) {
3555 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK);
3557 if (illegal_highdma(skb->dev, skb))
3558 features &= ~NETIF_F_SG;
3563 netdev_features_t passthru_features_check(struct sk_buff *skb,
3564 struct net_device *dev,
3565 netdev_features_t features)
3569 EXPORT_SYMBOL(passthru_features_check);
3571 static netdev_features_t dflt_features_check(struct sk_buff *skb,
3572 struct net_device *dev,
3573 netdev_features_t features)
3575 return vlan_features_check(skb, features);
3578 static netdev_features_t gso_features_check(const struct sk_buff *skb,
3579 struct net_device *dev,
3580 netdev_features_t features)
3582 u16 gso_segs = skb_shinfo(skb)->gso_segs;
3584 if (gso_segs > dev->gso_max_segs)
3585 return features & ~NETIF_F_GSO_MASK;
3587 if (!skb_shinfo(skb)->gso_type) {
3588 skb_warn_bad_offload(skb);
3589 return features & ~NETIF_F_GSO_MASK;
3592 /* Support for GSO partial features requires software
3593 * intervention before we can actually process the packets
3594 * so we need to strip support for any partial features now
3595 * and we can pull them back in after we have partially
3596 * segmented the frame.
3598 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL))
3599 features &= ~dev->gso_partial_features;
3601 /* Make sure to clear the IPv4 ID mangling feature if the
3602 * IPv4 header has the potential to be fragmented.
3604 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) {
3605 struct iphdr *iph = skb->encapsulation ?
3606 inner_ip_hdr(skb) : ip_hdr(skb);
3608 if (!(iph->frag_off & htons(IP_DF)))
3609 features &= ~NETIF_F_TSO_MANGLEID;
3615 netdev_features_t netif_skb_features(struct sk_buff *skb)
3617 struct net_device *dev = skb->dev;
3618 netdev_features_t features = dev->features;
3620 if (skb_is_gso(skb))
3621 features = gso_features_check(skb, dev, features);
3623 /* If encapsulation offload request, verify we are testing
3624 * hardware encapsulation features instead of standard
3625 * features for the netdev
3627 if (skb->encapsulation)
3628 features &= dev->hw_enc_features;
3630 if (skb_vlan_tagged(skb))
3631 features = netdev_intersect_features(features,
3632 dev->vlan_features |
3633 NETIF_F_HW_VLAN_CTAG_TX |
3634 NETIF_F_HW_VLAN_STAG_TX);
3636 if (dev->netdev_ops->ndo_features_check)
3637 features &= dev->netdev_ops->ndo_features_check(skb, dev,
3640 features &= dflt_features_check(skb, dev, features);
3642 return harmonize_features(skb, features);
3644 EXPORT_SYMBOL(netif_skb_features);
3646 static int xmit_one(struct sk_buff *skb, struct net_device *dev,
3647 struct netdev_queue *txq, bool more)
3652 if (dev_nit_active(dev))
3653 dev_queue_xmit_nit(skb, dev);
3656 PRANDOM_ADD_NOISE(skb, dev, txq, len + jiffies);
3657 trace_net_dev_start_xmit(skb, dev);
3658 rc = netdev_start_xmit(skb, dev, txq, more);
3659 trace_net_dev_xmit(skb, rc, dev, len);
3664 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev,
3665 struct netdev_queue *txq, int *ret)
3667 struct sk_buff *skb = first;
3668 int rc = NETDEV_TX_OK;
3671 struct sk_buff *next = skb->next;
3673 skb_mark_not_on_list(skb);
3674 rc = xmit_one(skb, dev, txq, next != NULL);
3675 if (unlikely(!dev_xmit_complete(rc))) {
3681 if (netif_tx_queue_stopped(txq) && skb) {
3682 rc = NETDEV_TX_BUSY;
3692 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb,
3693 netdev_features_t features)
3695 if (skb_vlan_tag_present(skb) &&
3696 !vlan_hw_offload_capable(features, skb->vlan_proto))
3697 skb = __vlan_hwaccel_push_inside(skb);
3701 int skb_csum_hwoffload_help(struct sk_buff *skb,
3702 const netdev_features_t features)
3704 if (unlikely(skb_csum_is_sctp(skb)))
3705 return !!(features & NETIF_F_SCTP_CRC) ? 0 :
3706 skb_crc32c_csum_help(skb);
3708 if (features & NETIF_F_HW_CSUM)
3711 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) {
3712 switch (skb->csum_offset) {
3713 case offsetof(struct tcphdr, check):
3714 case offsetof(struct udphdr, check):
3719 return skb_checksum_help(skb);
3721 EXPORT_SYMBOL(skb_csum_hwoffload_help);
3723 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again)
3725 netdev_features_t features;
3727 features = netif_skb_features(skb);
3728 skb = validate_xmit_vlan(skb, features);
3732 skb = sk_validate_xmit_skb(skb, dev);
3736 if (netif_needs_gso(skb, features)) {
3737 struct sk_buff *segs;
3739 segs = skb_gso_segment(skb, features);
3747 if (skb_needs_linearize(skb, features) &&
3748 __skb_linearize(skb))
3751 /* If packet is not checksummed and device does not
3752 * support checksumming for this protocol, complete
3753 * checksumming here.
3755 if (skb->ip_summed == CHECKSUM_PARTIAL) {
3756 if (skb->encapsulation)
3757 skb_set_inner_transport_header(skb,
3758 skb_checksum_start_offset(skb));
3760 skb_set_transport_header(skb,
3761 skb_checksum_start_offset(skb));
3762 if (skb_csum_hwoffload_help(skb, features))
3767 skb = validate_xmit_xfrm(skb, features, again);
3774 atomic_long_inc(&dev->tx_dropped);
3778 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again)
3780 struct sk_buff *next, *head = NULL, *tail;
3782 for (; skb != NULL; skb = next) {
3784 skb_mark_not_on_list(skb);
3786 /* in case skb wont be segmented, point to itself */
3789 skb = validate_xmit_skb(skb, dev, again);
3797 /* If skb was segmented, skb->prev points to
3798 * the last segment. If not, it still contains skb.
3804 EXPORT_SYMBOL_GPL(validate_xmit_skb_list);
3806 static void qdisc_pkt_len_init(struct sk_buff *skb)
3808 const struct skb_shared_info *shinfo = skb_shinfo(skb);
3810 qdisc_skb_cb(skb)->pkt_len = skb->len;
3812 /* To get more precise estimation of bytes sent on wire,
3813 * we add to pkt_len the headers size of all segments
3815 if (shinfo->gso_size && skb_transport_header_was_set(skb)) {
3816 unsigned int hdr_len;
3817 u16 gso_segs = shinfo->gso_segs;
3819 /* mac layer + network layer */
3820 hdr_len = skb_transport_header(skb) - skb_mac_header(skb);
3822 /* + transport layer */
3823 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) {
3824 const struct tcphdr *th;
3825 struct tcphdr _tcphdr;
3827 th = skb_header_pointer(skb, skb_transport_offset(skb),
3828 sizeof(_tcphdr), &_tcphdr);
3830 hdr_len += __tcp_hdrlen(th);
3832 struct udphdr _udphdr;
3834 if (skb_header_pointer(skb, skb_transport_offset(skb),
3835 sizeof(_udphdr), &_udphdr))
3836 hdr_len += sizeof(struct udphdr);
3839 if (shinfo->gso_type & SKB_GSO_DODGY)
3840 gso_segs = DIV_ROUND_UP(skb->len - hdr_len,
3843 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len;
3847 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q,
3848 struct net_device *dev,
3849 struct netdev_queue *txq)
3851 spinlock_t *root_lock = qdisc_lock(q);
3852 struct sk_buff *to_free = NULL;
3856 qdisc_calculate_pkt_len(skb, q);
3858 if (q->flags & TCQ_F_NOLOCK) {
3859 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) &&
3860 qdisc_run_begin(q)) {
3861 /* Retest nolock_qdisc_is_empty() within the protection
3862 * of q->seqlock to protect from racing with requeuing.
3864 if (unlikely(!nolock_qdisc_is_empty(q))) {
3865 rc = q->enqueue(skb, q, &to_free) &
3873 qdisc_bstats_cpu_update(q, skb);
3874 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) &&
3875 !nolock_qdisc_is_empty(q))
3879 return NET_XMIT_SUCCESS;
3882 rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
3886 if (unlikely(to_free))
3887 kfree_skb_list(to_free);
3892 * Heuristic to force contended enqueues to serialize on a
3893 * separate lock before trying to get qdisc main lock.
3894 * This permits qdisc->running owner to get the lock more
3895 * often and dequeue packets faster.
3897 contended = qdisc_is_running(q);
3898 if (unlikely(contended))
3899 spin_lock(&q->busylock);
3901 spin_lock(root_lock);
3902 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) {
3903 __qdisc_drop(skb, &to_free);
3905 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) &&
3906 qdisc_run_begin(q)) {
3908 * This is a work-conserving queue; there are no old skbs
3909 * waiting to be sent out; and the qdisc is not running -
3910 * xmit the skb directly.
3913 qdisc_bstats_update(q, skb);
3915 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) {
3916 if (unlikely(contended)) {
3917 spin_unlock(&q->busylock);
3924 rc = NET_XMIT_SUCCESS;
3926 rc = q->enqueue(skb, q, &to_free) & NET_XMIT_MASK;
3927 if (qdisc_run_begin(q)) {
3928 if (unlikely(contended)) {
3929 spin_unlock(&q->busylock);
3936 spin_unlock(root_lock);
3937 if (unlikely(to_free))
3938 kfree_skb_list(to_free);
3939 if (unlikely(contended))
3940 spin_unlock(&q->busylock);
3944 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO)
3945 static void skb_update_prio(struct sk_buff *skb)
3947 const struct netprio_map *map;
3948 const struct sock *sk;
3949 unsigned int prioidx;
3953 map = rcu_dereference_bh(skb->dev->priomap);
3956 sk = skb_to_full_sk(skb);
3960 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data);
3962 if (prioidx < map->priomap_len)
3963 skb->priority = map->priomap[prioidx];
3966 #define skb_update_prio(skb)
3970 * dev_loopback_xmit - loop back @skb
3971 * @net: network namespace this loopback is happening in
3972 * @sk: sk needed to be a netfilter okfn
3973 * @skb: buffer to transmit
3975 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb)
3977 skb_reset_mac_header(skb);
3978 __skb_pull(skb, skb_network_offset(skb));
3979 skb->pkt_type = PACKET_LOOPBACK;
3980 skb->ip_summed = CHECKSUM_UNNECESSARY;
3981 WARN_ON(!skb_dst(skb));
3986 EXPORT_SYMBOL(dev_loopback_xmit);
3988 #ifdef CONFIG_NET_EGRESS
3989 static struct sk_buff *
3990 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev)
3992 struct mini_Qdisc *miniq = rcu_dereference_bh(dev->miniq_egress);
3993 struct tcf_result cl_res;
3998 /* qdisc_skb_cb(skb)->pkt_len was already set by the caller. */
3999 qdisc_skb_cb(skb)->mru = 0;
4000 qdisc_skb_cb(skb)->post_ct = false;
4001 mini_qdisc_bstats_cpu_update(miniq, skb);
4003 switch (tcf_classify(skb, miniq->filter_list, &cl_res, false)) {
4005 case TC_ACT_RECLASSIFY:
4006 skb->tc_index = TC_H_MIN(cl_res.classid);
4009 mini_qdisc_qstats_cpu_drop(miniq);
4010 *ret = NET_XMIT_DROP;
4016 *ret = NET_XMIT_SUCCESS;
4019 case TC_ACT_REDIRECT:
4020 /* No need to push/pop skb's mac_header here on egress! */
4021 skb_do_redirect(skb);
4022 *ret = NET_XMIT_SUCCESS;
4030 #endif /* CONFIG_NET_EGRESS */
4033 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb,
4034 struct xps_dev_maps *dev_maps, unsigned int tci)
4036 int tc = netdev_get_prio_tc_map(dev, skb->priority);
4037 struct xps_map *map;
4038 int queue_index = -1;
4040 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids)
4043 tci *= dev_maps->num_tc;
4046 map = rcu_dereference(dev_maps->attr_map[tci]);
4049 queue_index = map->queues[0];
4051 queue_index = map->queues[reciprocal_scale(
4052 skb_get_hash(skb), map->len)];
4053 if (unlikely(queue_index >= dev->real_num_tx_queues))
4060 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev,
4061 struct sk_buff *skb)
4064 struct xps_dev_maps *dev_maps;
4065 struct sock *sk = skb->sk;
4066 int queue_index = -1;
4068 if (!static_key_false(&xps_needed))
4072 if (!static_key_false(&xps_rxqs_needed))
4075 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]);
4077 int tci = sk_rx_queue_get(sk);
4080 queue_index = __get_xps_queue_idx(dev, skb, dev_maps,
4085 if (queue_index < 0) {
4086 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]);
4088 unsigned int tci = skb->sender_cpu - 1;
4090 queue_index = __get_xps_queue_idx(dev, skb, dev_maps,
4102 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb,
4103 struct net_device *sb_dev)
4107 EXPORT_SYMBOL(dev_pick_tx_zero);
4109 u16 dev_pick_tx_cpu_id(struct net_device *dev, struct sk_buff *skb,
4110 struct net_device *sb_dev)
4112 return (u16)raw_smp_processor_id() % dev->real_num_tx_queues;
4114 EXPORT_SYMBOL(dev_pick_tx_cpu_id);
4116 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb,
4117 struct net_device *sb_dev)
4119 struct sock *sk = skb->sk;
4120 int queue_index = sk_tx_queue_get(sk);
4122 sb_dev = sb_dev ? : dev;
4124 if (queue_index < 0 || skb->ooo_okay ||
4125 queue_index >= dev->real_num_tx_queues) {
4126 int new_index = get_xps_queue(dev, sb_dev, skb);
4129 new_index = skb_tx_hash(dev, sb_dev, skb);
4131 if (queue_index != new_index && sk &&
4133 rcu_access_pointer(sk->sk_dst_cache))
4134 sk_tx_queue_set(sk, new_index);
4136 queue_index = new_index;
4141 EXPORT_SYMBOL(netdev_pick_tx);
4143 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev,
4144 struct sk_buff *skb,
4145 struct net_device *sb_dev)
4147 int queue_index = 0;
4150 u32 sender_cpu = skb->sender_cpu - 1;
4152 if (sender_cpu >= (u32)NR_CPUS)
4153 skb->sender_cpu = raw_smp_processor_id() + 1;
4156 if (dev->real_num_tx_queues != 1) {
4157 const struct net_device_ops *ops = dev->netdev_ops;
4159 if (ops->ndo_select_queue)
4160 queue_index = ops->ndo_select_queue(dev, skb, sb_dev);
4162 queue_index = netdev_pick_tx(dev, skb, sb_dev);
4164 queue_index = netdev_cap_txqueue(dev, queue_index);
4167 skb_set_queue_mapping(skb, queue_index);
4168 return netdev_get_tx_queue(dev, queue_index);
4172 * __dev_queue_xmit - transmit a buffer
4173 * @skb: buffer to transmit
4174 * @sb_dev: suboordinate device used for L2 forwarding offload
4176 * Queue a buffer for transmission to a network device. The caller must
4177 * have set the device and priority and built the buffer before calling
4178 * this function. The function can be called from an interrupt.
4180 * A negative errno code is returned on a failure. A success does not
4181 * guarantee the frame will be transmitted as it may be dropped due
4182 * to congestion or traffic shaping.
4184 * -----------------------------------------------------------------------------------
4185 * I notice this method can also return errors from the queue disciplines,
4186 * including NET_XMIT_DROP, which is a positive value. So, errors can also
4189 * Regardless of the return value, the skb is consumed, so it is currently
4190 * difficult to retry a send to this method. (You can bump the ref count
4191 * before sending to hold a reference for retry if you are careful.)
4193 * When calling this method, interrupts MUST be enabled. This is because
4194 * the BH enable code must have IRQs enabled so that it will not deadlock.
4197 static int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev)
4199 struct net_device *dev = skb->dev;
4200 struct netdev_queue *txq;
4205 skb_reset_mac_header(skb);
4207 if (unlikely(skb_shinfo(skb)->tx_flags & SKBTX_SCHED_TSTAMP))
4208 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED);
4210 /* Disable soft irqs for various locks below. Also
4211 * stops preemption for RCU.
4215 skb_update_prio(skb);
4217 qdisc_pkt_len_init(skb);
4218 #ifdef CONFIG_NET_CLS_ACT
4219 skb->tc_at_ingress = 0;
4220 # ifdef CONFIG_NET_EGRESS
4221 if (static_branch_unlikely(&egress_needed_key)) {
4222 skb = sch_handle_egress(skb, &rc, dev);
4228 /* If device/qdisc don't need skb->dst, release it right now while
4229 * its hot in this cpu cache.
4231 if (dev->priv_flags & IFF_XMIT_DST_RELEASE)
4236 txq = netdev_core_pick_tx(dev, skb, sb_dev);
4237 q = rcu_dereference_bh(txq->qdisc);
4239 trace_net_dev_queue(skb);
4241 rc = __dev_xmit_skb(skb, q, dev, txq);
4245 /* The device has no queue. Common case for software devices:
4246 * loopback, all the sorts of tunnels...
4248 * Really, it is unlikely that netif_tx_lock protection is necessary
4249 * here. (f.e. loopback and IP tunnels are clean ignoring statistics
4251 * However, it is possible, that they rely on protection
4254 * Check this and shot the lock. It is not prone from deadlocks.
4255 *Either shot noqueue qdisc, it is even simpler 8)
4257 if (dev->flags & IFF_UP) {
4258 int cpu = smp_processor_id(); /* ok because BHs are off */
4260 if (txq->xmit_lock_owner != cpu) {
4261 if (dev_xmit_recursion())
4262 goto recursion_alert;
4264 skb = validate_xmit_skb(skb, dev, &again);
4268 PRANDOM_ADD_NOISE(skb, dev, txq, jiffies);
4269 HARD_TX_LOCK(dev, txq, cpu);
4271 if (!netif_xmit_stopped(txq)) {
4272 dev_xmit_recursion_inc();
4273 skb = dev_hard_start_xmit(skb, dev, txq, &rc);
4274 dev_xmit_recursion_dec();
4275 if (dev_xmit_complete(rc)) {
4276 HARD_TX_UNLOCK(dev, txq);
4280 HARD_TX_UNLOCK(dev, txq);
4281 net_crit_ratelimited("Virtual device %s asks to queue packet!\n",
4284 /* Recursion is detected! It is possible,
4288 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n",
4294 rcu_read_unlock_bh();
4296 atomic_long_inc(&dev->tx_dropped);
4297 kfree_skb_list(skb);
4300 rcu_read_unlock_bh();
4304 int dev_queue_xmit(struct sk_buff *skb)
4306 return __dev_queue_xmit(skb, NULL);
4308 EXPORT_SYMBOL(dev_queue_xmit);
4310 int dev_queue_xmit_accel(struct sk_buff *skb, struct net_device *sb_dev)
4312 return __dev_queue_xmit(skb, sb_dev);
4314 EXPORT_SYMBOL(dev_queue_xmit_accel);
4316 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id)
4318 struct net_device *dev = skb->dev;
4319 struct sk_buff *orig_skb = skb;
4320 struct netdev_queue *txq;
4321 int ret = NETDEV_TX_BUSY;
4324 if (unlikely(!netif_running(dev) ||
4325 !netif_carrier_ok(dev)))
4328 skb = validate_xmit_skb_list(skb, dev, &again);
4329 if (skb != orig_skb)
4332 skb_set_queue_mapping(skb, queue_id);
4333 txq = skb_get_tx_queue(dev, skb);
4334 PRANDOM_ADD_NOISE(skb, dev, txq, jiffies);
4338 dev_xmit_recursion_inc();
4339 HARD_TX_LOCK(dev, txq, smp_processor_id());
4340 if (!netif_xmit_frozen_or_drv_stopped(txq))
4341 ret = netdev_start_xmit(skb, dev, txq, false);
4342 HARD_TX_UNLOCK(dev, txq);
4343 dev_xmit_recursion_dec();
4348 atomic_long_inc(&dev->tx_dropped);
4349 kfree_skb_list(skb);
4350 return NET_XMIT_DROP;
4352 EXPORT_SYMBOL(__dev_direct_xmit);
4354 /*************************************************************************
4356 *************************************************************************/
4358 int netdev_max_backlog __read_mostly = 1000;
4359 EXPORT_SYMBOL(netdev_max_backlog);
4361 int netdev_tstamp_prequeue __read_mostly = 1;
4362 int netdev_budget __read_mostly = 300;
4363 /* Must be at least 2 jiffes to guarantee 1 jiffy timeout */
4364 unsigned int __read_mostly netdev_budget_usecs = 2 * USEC_PER_SEC / HZ;
4365 int weight_p __read_mostly = 64; /* old backlog weight */
4366 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */
4367 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */
4368 int dev_rx_weight __read_mostly = 64;
4369 int dev_tx_weight __read_mostly = 64;
4370 /* Maximum number of GRO_NORMAL skbs to batch up for list-RX */
4371 int gro_normal_batch __read_mostly = 8;
4373 /* Called with irq disabled */
4374 static inline void ____napi_schedule(struct softnet_data *sd,
4375 struct napi_struct *napi)
4377 struct task_struct *thread;
4379 if (test_bit(NAPI_STATE_THREADED, &napi->state)) {
4380 /* Paired with smp_mb__before_atomic() in
4381 * napi_enable()/dev_set_threaded().
4382 * Use READ_ONCE() to guarantee a complete
4383 * read on napi->thread. Only call
4384 * wake_up_process() when it's not NULL.
4386 thread = READ_ONCE(napi->thread);
4388 /* Avoid doing set_bit() if the thread is in
4389 * INTERRUPTIBLE state, cause napi_thread_wait()
4390 * makes sure to proceed with napi polling
4391 * if the thread is explicitly woken from here.
4393 if (READ_ONCE(thread->__state) != TASK_INTERRUPTIBLE)
4394 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state);
4395 wake_up_process(thread);
4400 list_add_tail(&napi->poll_list, &sd->poll_list);
4401 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
4406 /* One global table that all flow-based protocols share. */
4407 struct rps_sock_flow_table __rcu *rps_sock_flow_table __read_mostly;
4408 EXPORT_SYMBOL(rps_sock_flow_table);
4409 u32 rps_cpu_mask __read_mostly;
4410 EXPORT_SYMBOL(rps_cpu_mask);
4412 struct static_key_false rps_needed __read_mostly;
4413 EXPORT_SYMBOL(rps_needed);
4414 struct static_key_false rfs_needed __read_mostly;
4415 EXPORT_SYMBOL(rfs_needed);
4417 static struct rps_dev_flow *
4418 set_rps_cpu(struct net_device *dev, struct sk_buff *skb,
4419 struct rps_dev_flow *rflow, u16 next_cpu)
4421 if (next_cpu < nr_cpu_ids) {
4422 #ifdef CONFIG_RFS_ACCEL
4423 struct netdev_rx_queue *rxqueue;
4424 struct rps_dev_flow_table *flow_table;
4425 struct rps_dev_flow *old_rflow;
4430 /* Should we steer this flow to a different hardware queue? */
4431 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap ||
4432 !(dev->features & NETIF_F_NTUPLE))
4434 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu);
4435 if (rxq_index == skb_get_rx_queue(skb))
4438 rxqueue = dev->_rx + rxq_index;
4439 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4442 flow_id = skb_get_hash(skb) & flow_table->mask;
4443 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb,
4444 rxq_index, flow_id);
4448 rflow = &flow_table->flows[flow_id];
4450 if (old_rflow->filter == rflow->filter)
4451 old_rflow->filter = RPS_NO_FILTER;
4455 per_cpu(softnet_data, next_cpu).input_queue_head;
4458 rflow->cpu = next_cpu;
4463 * get_rps_cpu is called from netif_receive_skb and returns the target
4464 * CPU from the RPS map of the receiving queue for a given skb.
4465 * rcu_read_lock must be held on entry.
4467 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb,
4468 struct rps_dev_flow **rflowp)
4470 const struct rps_sock_flow_table *sock_flow_table;
4471 struct netdev_rx_queue *rxqueue = dev->_rx;
4472 struct rps_dev_flow_table *flow_table;
4473 struct rps_map *map;
4478 if (skb_rx_queue_recorded(skb)) {
4479 u16 index = skb_get_rx_queue(skb);
4481 if (unlikely(index >= dev->real_num_rx_queues)) {
4482 WARN_ONCE(dev->real_num_rx_queues > 1,
4483 "%s received packet on queue %u, but number "
4484 "of RX queues is %u\n",
4485 dev->name, index, dev->real_num_rx_queues);
4491 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */
4493 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4494 map = rcu_dereference(rxqueue->rps_map);
4495 if (!flow_table && !map)
4498 skb_reset_network_header(skb);
4499 hash = skb_get_hash(skb);
4503 sock_flow_table = rcu_dereference(rps_sock_flow_table);
4504 if (flow_table && sock_flow_table) {
4505 struct rps_dev_flow *rflow;
4509 /* First check into global flow table if there is a match */
4510 ident = sock_flow_table->ents[hash & sock_flow_table->mask];
4511 if ((ident ^ hash) & ~rps_cpu_mask)
4514 next_cpu = ident & rps_cpu_mask;
4516 /* OK, now we know there is a match,
4517 * we can look at the local (per receive queue) flow table
4519 rflow = &flow_table->flows[hash & flow_table->mask];
4523 * If the desired CPU (where last recvmsg was done) is
4524 * different from current CPU (one in the rx-queue flow
4525 * table entry), switch if one of the following holds:
4526 * - Current CPU is unset (>= nr_cpu_ids).
4527 * - Current CPU is offline.
4528 * - The current CPU's queue tail has advanced beyond the
4529 * last packet that was enqueued using this table entry.
4530 * This guarantees that all previous packets for the flow
4531 * have been dequeued, thus preserving in order delivery.
4533 if (unlikely(tcpu != next_cpu) &&
4534 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) ||
4535 ((int)(per_cpu(softnet_data, tcpu).input_queue_head -
4536 rflow->last_qtail)) >= 0)) {
4538 rflow = set_rps_cpu(dev, skb, rflow, next_cpu);
4541 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) {
4551 tcpu = map->cpus[reciprocal_scale(hash, map->len)];
4552 if (cpu_online(tcpu)) {
4562 #ifdef CONFIG_RFS_ACCEL
4565 * rps_may_expire_flow - check whether an RFS hardware filter may be removed
4566 * @dev: Device on which the filter was set
4567 * @rxq_index: RX queue index
4568 * @flow_id: Flow ID passed to ndo_rx_flow_steer()
4569 * @filter_id: Filter ID returned by ndo_rx_flow_steer()
4571 * Drivers that implement ndo_rx_flow_steer() should periodically call
4572 * this function for each installed filter and remove the filters for
4573 * which it returns %true.
4575 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index,
4576 u32 flow_id, u16 filter_id)
4578 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index;
4579 struct rps_dev_flow_table *flow_table;
4580 struct rps_dev_flow *rflow;
4585 flow_table = rcu_dereference(rxqueue->rps_flow_table);
4586 if (flow_table && flow_id <= flow_table->mask) {
4587 rflow = &flow_table->flows[flow_id];
4588 cpu = READ_ONCE(rflow->cpu);
4589 if (rflow->filter == filter_id && cpu < nr_cpu_ids &&
4590 ((int)(per_cpu(softnet_data, cpu).input_queue_head -
4591 rflow->last_qtail) <
4592 (int)(10 * flow_table->mask)))
4598 EXPORT_SYMBOL(rps_may_expire_flow);
4600 #endif /* CONFIG_RFS_ACCEL */
4602 /* Called from hardirq (IPI) context */
4603 static void rps_trigger_softirq(void *data)
4605 struct softnet_data *sd = data;
4607 ____napi_schedule(sd, &sd->backlog);
4611 #endif /* CONFIG_RPS */
4614 * Check if this softnet_data structure is another cpu one
4615 * If yes, queue it to our IPI list and return 1
4618 static int rps_ipi_queued(struct softnet_data *sd)
4621 struct softnet_data *mysd = this_cpu_ptr(&softnet_data);
4624 sd->rps_ipi_next = mysd->rps_ipi_list;
4625 mysd->rps_ipi_list = sd;
4627 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
4630 #endif /* CONFIG_RPS */
4634 #ifdef CONFIG_NET_FLOW_LIMIT
4635 int netdev_flow_limit_table_len __read_mostly = (1 << 12);
4638 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen)
4640 #ifdef CONFIG_NET_FLOW_LIMIT
4641 struct sd_flow_limit *fl;
4642 struct softnet_data *sd;
4643 unsigned int old_flow, new_flow;
4645 if (qlen < (netdev_max_backlog >> 1))
4648 sd = this_cpu_ptr(&softnet_data);
4651 fl = rcu_dereference(sd->flow_limit);
4653 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1);
4654 old_flow = fl->history[fl->history_head];
4655 fl->history[fl->history_head] = new_flow;
4658 fl->history_head &= FLOW_LIMIT_HISTORY - 1;
4660 if (likely(fl->buckets[old_flow]))
4661 fl->buckets[old_flow]--;
4663 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) {
4675 * enqueue_to_backlog is called to queue an skb to a per CPU backlog
4676 * queue (may be a remote CPU queue).
4678 static int enqueue_to_backlog(struct sk_buff *skb, int cpu,
4679 unsigned int *qtail)
4681 struct softnet_data *sd;
4682 unsigned long flags;
4685 sd = &per_cpu(softnet_data, cpu);
4687 local_irq_save(flags);
4690 if (!netif_running(skb->dev))
4692 qlen = skb_queue_len(&sd->input_pkt_queue);
4693 if (qlen <= netdev_max_backlog && !skb_flow_limit(skb, qlen)) {
4696 __skb_queue_tail(&sd->input_pkt_queue, skb);
4697 input_queue_tail_incr_save(sd, qtail);
4699 local_irq_restore(flags);
4700 return NET_RX_SUCCESS;
4703 /* Schedule NAPI for backlog device
4704 * We can use non atomic operation since we own the queue lock
4706 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) {
4707 if (!rps_ipi_queued(sd))
4708 ____napi_schedule(sd, &sd->backlog);
4717 local_irq_restore(flags);
4719 atomic_long_inc(&skb->dev->rx_dropped);
4724 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb)
4726 struct net_device *dev = skb->dev;
4727 struct netdev_rx_queue *rxqueue;
4731 if (skb_rx_queue_recorded(skb)) {
4732 u16 index = skb_get_rx_queue(skb);
4734 if (unlikely(index >= dev->real_num_rx_queues)) {
4735 WARN_ONCE(dev->real_num_rx_queues > 1,
4736 "%s received packet on queue %u, but number "
4737 "of RX queues is %u\n",
4738 dev->name, index, dev->real_num_rx_queues);
4740 return rxqueue; /* Return first rxqueue */
4747 static u32 netif_receive_generic_xdp(struct sk_buff *skb,
4748 struct xdp_buff *xdp,
4749 struct bpf_prog *xdp_prog)
4751 void *orig_data, *orig_data_end, *hard_start;
4752 struct netdev_rx_queue *rxqueue;
4753 u32 metalen, act = XDP_DROP;
4754 bool orig_bcast, orig_host;
4755 u32 mac_len, frame_sz;
4756 __be16 orig_eth_type;
4760 /* Reinjected packets coming from act_mirred or similar should
4761 * not get XDP generic processing.
4763 if (skb_is_redirected(skb))
4766 /* XDP packets must be linear and must have sufficient headroom
4767 * of XDP_PACKET_HEADROOM bytes. This is the guarantee that also
4768 * native XDP provides, thus we need to do it here as well.
4770 if (skb_cloned(skb) || skb_is_nonlinear(skb) ||
4771 skb_headroom(skb) < XDP_PACKET_HEADROOM) {
4772 int hroom = XDP_PACKET_HEADROOM - skb_headroom(skb);
4773 int troom = skb->tail + skb->data_len - skb->end;
4775 /* In case we have to go down the path and also linearize,
4776 * then lets do the pskb_expand_head() work just once here.
4778 if (pskb_expand_head(skb,
4779 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0,
4780 troom > 0 ? troom + 128 : 0, GFP_ATOMIC))
4782 if (skb_linearize(skb))
4786 /* The XDP program wants to see the packet starting at the MAC
4789 mac_len = skb->data - skb_mac_header(skb);
4790 hard_start = skb->data - skb_headroom(skb);
4792 /* SKB "head" area always have tailroom for skb_shared_info */
4793 frame_sz = (void *)skb_end_pointer(skb) - hard_start;
4794 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
4796 rxqueue = netif_get_rxqueue(skb);
4797 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq);
4798 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len,
4799 skb_headlen(skb) + mac_len, true);
4801 orig_data_end = xdp->data_end;
4802 orig_data = xdp->data;
4803 eth = (struct ethhdr *)xdp->data;
4804 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr);
4805 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest);
4806 orig_eth_type = eth->h_proto;
4808 act = bpf_prog_run_xdp(xdp_prog, xdp);
4810 /* check if bpf_xdp_adjust_head was used */
4811 off = xdp->data - orig_data;
4814 __skb_pull(skb, off);
4816 __skb_push(skb, -off);
4818 skb->mac_header += off;
4819 skb_reset_network_header(skb);
4822 /* check if bpf_xdp_adjust_tail was used */
4823 off = xdp->data_end - orig_data_end;
4825 skb_set_tail_pointer(skb, xdp->data_end - xdp->data);
4826 skb->len += off; /* positive on grow, negative on shrink */
4829 /* check if XDP changed eth hdr such SKB needs update */
4830 eth = (struct ethhdr *)xdp->data;
4831 if ((orig_eth_type != eth->h_proto) ||
4832 (orig_host != ether_addr_equal_64bits(eth->h_dest,
4833 skb->dev->dev_addr)) ||
4834 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) {
4835 __skb_push(skb, ETH_HLEN);
4836 skb->pkt_type = PACKET_HOST;
4837 skb->protocol = eth_type_trans(skb, skb->dev);
4843 __skb_push(skb, mac_len);
4846 metalen = xdp->data - xdp->data_meta;
4848 skb_metadata_set(skb, metalen);
4851 bpf_warn_invalid_xdp_action(act);
4854 trace_xdp_exception(skb->dev, xdp_prog, act);
4865 /* When doing generic XDP we have to bypass the qdisc layer and the
4866 * network taps in order to match in-driver-XDP behavior.
4868 void generic_xdp_tx(struct sk_buff *skb, struct bpf_prog *xdp_prog)
4870 struct net_device *dev = skb->dev;
4871 struct netdev_queue *txq;
4872 bool free_skb = true;
4875 txq = netdev_core_pick_tx(dev, skb, NULL);
4876 cpu = smp_processor_id();
4877 HARD_TX_LOCK(dev, txq, cpu);
4878 if (!netif_xmit_stopped(txq)) {
4879 rc = netdev_start_xmit(skb, dev, txq, 0);
4880 if (dev_xmit_complete(rc))
4883 HARD_TX_UNLOCK(dev, txq);
4885 trace_xdp_exception(dev, xdp_prog, XDP_TX);
4890 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key);
4892 int do_xdp_generic(struct bpf_prog *xdp_prog, struct sk_buff *skb)
4895 struct xdp_buff xdp;
4899 act = netif_receive_generic_xdp(skb, &xdp, xdp_prog);
4900 if (act != XDP_PASS) {
4903 err = xdp_do_generic_redirect(skb->dev, skb,
4909 generic_xdp_tx(skb, xdp_prog);
4920 EXPORT_SYMBOL_GPL(do_xdp_generic);
4922 static int netif_rx_internal(struct sk_buff *skb)
4926 net_timestamp_check(netdev_tstamp_prequeue, skb);
4928 trace_netif_rx(skb);
4931 if (static_branch_unlikely(&rps_needed)) {
4932 struct rps_dev_flow voidflow, *rflow = &voidflow;
4938 cpu = get_rps_cpu(skb->dev, skb, &rflow);
4940 cpu = smp_processor_id();
4942 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
4951 ret = enqueue_to_backlog(skb, get_cpu(), &qtail);
4958 * netif_rx - post buffer to the network code
4959 * @skb: buffer to post
4961 * This function receives a packet from a device driver and queues it for
4962 * the upper (protocol) levels to process. It always succeeds. The buffer
4963 * may be dropped during processing for congestion control or by the
4967 * NET_RX_SUCCESS (no congestion)
4968 * NET_RX_DROP (packet was dropped)
4972 int netif_rx(struct sk_buff *skb)
4976 trace_netif_rx_entry(skb);
4978 ret = netif_rx_internal(skb);
4979 trace_netif_rx_exit(ret);
4983 EXPORT_SYMBOL(netif_rx);
4985 int netif_rx_ni(struct sk_buff *skb)
4989 trace_netif_rx_ni_entry(skb);
4992 err = netif_rx_internal(skb);
4993 if (local_softirq_pending())
4996 trace_netif_rx_ni_exit(err);
5000 EXPORT_SYMBOL(netif_rx_ni);
5002 int netif_rx_any_context(struct sk_buff *skb)
5005 * If invoked from contexts which do not invoke bottom half
5006 * processing either at return from interrupt or when softrqs are
5007 * reenabled, use netif_rx_ni() which invokes bottomhalf processing
5011 return netif_rx(skb);
5013 return netif_rx_ni(skb);
5015 EXPORT_SYMBOL(netif_rx_any_context);
5017 static __latent_entropy void net_tx_action(struct softirq_action *h)
5019 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
5021 if (sd->completion_queue) {
5022 struct sk_buff *clist;
5024 local_irq_disable();
5025 clist = sd->completion_queue;
5026 sd->completion_queue = NULL;
5030 struct sk_buff *skb = clist;
5032 clist = clist->next;
5034 WARN_ON(refcount_read(&skb->users));
5035 if (likely(get_kfree_skb_cb(skb)->reason == SKB_REASON_CONSUMED))
5036 trace_consume_skb(skb);
5038 trace_kfree_skb(skb, net_tx_action);
5040 if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
5043 __kfree_skb_defer(skb);
5047 if (sd->output_queue) {
5050 local_irq_disable();
5051 head = sd->output_queue;
5052 sd->output_queue = NULL;
5053 sd->output_queue_tailp = &sd->output_queue;
5059 struct Qdisc *q = head;
5060 spinlock_t *root_lock = NULL;
5062 head = head->next_sched;
5064 /* We need to make sure head->next_sched is read
5065 * before clearing __QDISC_STATE_SCHED
5067 smp_mb__before_atomic();
5069 if (!(q->flags & TCQ_F_NOLOCK)) {
5070 root_lock = qdisc_lock(q);
5071 spin_lock(root_lock);
5072 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED,
5074 /* There is a synchronize_net() between
5075 * STATE_DEACTIVATED flag being set and
5076 * qdisc_reset()/some_qdisc_is_busy() in
5077 * dev_deactivate(), so we can safely bail out
5078 * early here to avoid data race between
5079 * qdisc_deactivate() and some_qdisc_is_busy()
5080 * for lockless qdisc.
5082 clear_bit(__QDISC_STATE_SCHED, &q->state);
5086 clear_bit(__QDISC_STATE_SCHED, &q->state);
5089 spin_unlock(root_lock);
5095 xfrm_dev_backlog(sd);
5098 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE)
5099 /* This hook is defined here for ATM LANE */
5100 int (*br_fdb_test_addr_hook)(struct net_device *dev,
5101 unsigned char *addr) __read_mostly;
5102 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook);
5105 static inline struct sk_buff *
5106 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret,
5107 struct net_device *orig_dev, bool *another)
5109 #ifdef CONFIG_NET_CLS_ACT
5110 struct mini_Qdisc *miniq = rcu_dereference_bh(skb->dev->miniq_ingress);
5111 struct tcf_result cl_res;
5113 /* If there's at least one ingress present somewhere (so
5114 * we get here via enabled static key), remaining devices
5115 * that are not configured with an ingress qdisc will bail
5122 *ret = deliver_skb(skb, *pt_prev, orig_dev);
5126 qdisc_skb_cb(skb)->pkt_len = skb->len;
5127 qdisc_skb_cb(skb)->mru = 0;
5128 qdisc_skb_cb(skb)->post_ct = false;
5129 skb->tc_at_ingress = 1;
5130 mini_qdisc_bstats_cpu_update(miniq, skb);
5132 switch (tcf_classify_ingress(skb, miniq->block, miniq->filter_list,
5135 case TC_ACT_RECLASSIFY:
5136 skb->tc_index = TC_H_MIN(cl_res.classid);
5139 mini_qdisc_qstats_cpu_drop(miniq);
5147 case TC_ACT_REDIRECT:
5148 /* skb_mac_header check was done by cls/act_bpf, so
5149 * we can safely push the L2 header back before
5150 * redirecting to another netdev
5152 __skb_push(skb, skb->mac_len);
5153 if (skb_do_redirect(skb) == -EAGAIN) {
5154 __skb_pull(skb, skb->mac_len);
5159 case TC_ACT_CONSUMED:
5164 #endif /* CONFIG_NET_CLS_ACT */
5169 * netdev_is_rx_handler_busy - check if receive handler is registered
5170 * @dev: device to check
5172 * Check if a receive handler is already registered for a given device.
5173 * Return true if there one.
5175 * The caller must hold the rtnl_mutex.
5177 bool netdev_is_rx_handler_busy(struct net_device *dev)
5180 return dev && rtnl_dereference(dev->rx_handler);
5182 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy);
5185 * netdev_rx_handler_register - register receive handler
5186 * @dev: device to register a handler for
5187 * @rx_handler: receive handler to register
5188 * @rx_handler_data: data pointer that is used by rx handler
5190 * Register a receive handler for a device. This handler will then be
5191 * called from __netif_receive_skb. A negative errno code is returned
5194 * The caller must hold the rtnl_mutex.
5196 * For a general description of rx_handler, see enum rx_handler_result.
5198 int netdev_rx_handler_register(struct net_device *dev,
5199 rx_handler_func_t *rx_handler,
5200 void *rx_handler_data)
5202 if (netdev_is_rx_handler_busy(dev))
5205 if (dev->priv_flags & IFF_NO_RX_HANDLER)
5208 /* Note: rx_handler_data must be set before rx_handler */
5209 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data);
5210 rcu_assign_pointer(dev->rx_handler, rx_handler);
5214 EXPORT_SYMBOL_GPL(netdev_rx_handler_register);
5217 * netdev_rx_handler_unregister - unregister receive handler
5218 * @dev: device to unregister a handler from
5220 * Unregister a receive handler from a device.
5222 * The caller must hold the rtnl_mutex.
5224 void netdev_rx_handler_unregister(struct net_device *dev)
5228 RCU_INIT_POINTER(dev->rx_handler, NULL);
5229 /* a reader seeing a non NULL rx_handler in a rcu_read_lock()
5230 * section has a guarantee to see a non NULL rx_handler_data
5234 RCU_INIT_POINTER(dev->rx_handler_data, NULL);
5236 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister);
5239 * Limit the use of PFMEMALLOC reserves to those protocols that implement
5240 * the special handling of PFMEMALLOC skbs.
5242 static bool skb_pfmemalloc_protocol(struct sk_buff *skb)
5244 switch (skb->protocol) {
5245 case htons(ETH_P_ARP):
5246 case htons(ETH_P_IP):
5247 case htons(ETH_P_IPV6):
5248 case htons(ETH_P_8021Q):
5249 case htons(ETH_P_8021AD):
5256 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev,
5257 int *ret, struct net_device *orig_dev)
5259 if (nf_hook_ingress_active(skb)) {
5263 *ret = deliver_skb(skb, *pt_prev, orig_dev);
5268 ingress_retval = nf_hook_ingress(skb);
5270 return ingress_retval;
5275 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc,
5276 struct packet_type **ppt_prev)
5278 struct packet_type *ptype, *pt_prev;
5279 rx_handler_func_t *rx_handler;
5280 struct sk_buff *skb = *pskb;
5281 struct net_device *orig_dev;
5282 bool deliver_exact = false;
5283 int ret = NET_RX_DROP;
5286 net_timestamp_check(!netdev_tstamp_prequeue, skb);
5288 trace_netif_receive_skb(skb);
5290 orig_dev = skb->dev;
5292 skb_reset_network_header(skb);
5293 if (!skb_transport_header_was_set(skb))
5294 skb_reset_transport_header(skb);
5295 skb_reset_mac_len(skb);
5300 skb->skb_iif = skb->dev->ifindex;
5302 __this_cpu_inc(softnet_data.processed);
5304 if (static_branch_unlikely(&generic_xdp_needed_key)) {
5308 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), skb);
5311 if (ret2 != XDP_PASS) {
5315 skb_reset_mac_len(skb);
5318 if (eth_type_vlan(skb->protocol)) {
5319 skb = skb_vlan_untag(skb);
5324 if (skb_skip_tc_classify(skb))
5330 list_for_each_entry_rcu(ptype, &ptype_all, list) {
5332 ret = deliver_skb(skb, pt_prev, orig_dev);
5336 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) {
5338 ret = deliver_skb(skb, pt_prev, orig_dev);
5343 #ifdef CONFIG_NET_INGRESS
5344 if (static_branch_unlikely(&ingress_needed_key)) {
5345 bool another = false;
5347 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev,
5354 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0)
5358 skb_reset_redirect(skb);
5360 if (pfmemalloc && !skb_pfmemalloc_protocol(skb))
5363 if (skb_vlan_tag_present(skb)) {
5365 ret = deliver_skb(skb, pt_prev, orig_dev);
5368 if (vlan_do_receive(&skb))
5370 else if (unlikely(!skb))
5374 rx_handler = rcu_dereference(skb->dev->rx_handler);
5377 ret = deliver_skb(skb, pt_prev, orig_dev);
5380 switch (rx_handler(&skb)) {
5381 case RX_HANDLER_CONSUMED:
5382 ret = NET_RX_SUCCESS;
5384 case RX_HANDLER_ANOTHER:
5386 case RX_HANDLER_EXACT:
5387 deliver_exact = true;
5389 case RX_HANDLER_PASS:
5396 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) {
5398 if (skb_vlan_tag_get_id(skb)) {
5399 /* Vlan id is non 0 and vlan_do_receive() above couldn't
5402 skb->pkt_type = PACKET_OTHERHOST;
5403 } else if (eth_type_vlan(skb->protocol)) {
5404 /* Outer header is 802.1P with vlan 0, inner header is
5405 * 802.1Q or 802.1AD and vlan_do_receive() above could
5406 * not find vlan dev for vlan id 0.
5408 __vlan_hwaccel_clear_tag(skb);
5409 skb = skb_vlan_untag(skb);
5412 if (vlan_do_receive(&skb))
5413 /* After stripping off 802.1P header with vlan 0
5414 * vlan dev is found for inner header.
5417 else if (unlikely(!skb))
5420 /* We have stripped outer 802.1P vlan 0 header.
5421 * But could not find vlan dev.
5422 * check again for vlan id to set OTHERHOST.
5426 /* Note: we might in the future use prio bits
5427 * and set skb->priority like in vlan_do_receive()
5428 * For the time being, just ignore Priority Code Point
5430 __vlan_hwaccel_clear_tag(skb);
5433 type = skb->protocol;
5435 /* deliver only exact match when indicated */
5436 if (likely(!deliver_exact)) {
5437 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
5438 &ptype_base[ntohs(type) &
5442 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
5443 &orig_dev->ptype_specific);
5445 if (unlikely(skb->dev != orig_dev)) {
5446 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type,
5447 &skb->dev->ptype_specific);
5451 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC)))
5453 *ppt_prev = pt_prev;
5457 atomic_long_inc(&skb->dev->rx_dropped);
5459 atomic_long_inc(&skb->dev->rx_nohandler);
5461 /* Jamal, now you will not able to escape explaining
5462 * me how you were going to use this. :-)
5468 /* The invariant here is that if *ppt_prev is not NULL
5469 * then skb should also be non-NULL.
5471 * Apparently *ppt_prev assignment above holds this invariant due to
5472 * skb dereferencing near it.
5478 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc)
5480 struct net_device *orig_dev = skb->dev;
5481 struct packet_type *pt_prev = NULL;
5484 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev);
5486 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb,
5487 skb->dev, pt_prev, orig_dev);
5492 * netif_receive_skb_core - special purpose version of netif_receive_skb
5493 * @skb: buffer to process
5495 * More direct receive version of netif_receive_skb(). It should
5496 * only be used by callers that have a need to skip RPS and Generic XDP.
5497 * Caller must also take care of handling if ``(page_is_)pfmemalloc``.
5499 * This function may only be called from softirq context and interrupts
5500 * should be enabled.
5502 * Return values (usually ignored):
5503 * NET_RX_SUCCESS: no congestion
5504 * NET_RX_DROP: packet was dropped
5506 int netif_receive_skb_core(struct sk_buff *skb)
5511 ret = __netif_receive_skb_one_core(skb, false);
5516 EXPORT_SYMBOL(netif_receive_skb_core);
5518 static inline void __netif_receive_skb_list_ptype(struct list_head *head,
5519 struct packet_type *pt_prev,
5520 struct net_device *orig_dev)
5522 struct sk_buff *skb, *next;
5526 if (list_empty(head))
5528 if (pt_prev->list_func != NULL)
5529 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv,
5530 ip_list_rcv, head, pt_prev, orig_dev);
5532 list_for_each_entry_safe(skb, next, head, list) {
5533 skb_list_del_init(skb);
5534 pt_prev->func(skb, skb->dev, pt_prev, orig_dev);
5538 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc)
5540 /* Fast-path assumptions:
5541 * - There is no RX handler.
5542 * - Only one packet_type matches.
5543 * If either of these fails, we will end up doing some per-packet
5544 * processing in-line, then handling the 'last ptype' for the whole
5545 * sublist. This can't cause out-of-order delivery to any single ptype,
5546 * because the 'last ptype' must be constant across the sublist, and all
5547 * other ptypes are handled per-packet.
5549 /* Current (common) ptype of sublist */
5550 struct packet_type *pt_curr = NULL;
5551 /* Current (common) orig_dev of sublist */
5552 struct net_device *od_curr = NULL;
5553 struct list_head sublist;
5554 struct sk_buff *skb, *next;
5556 INIT_LIST_HEAD(&sublist);
5557 list_for_each_entry_safe(skb, next, head, list) {
5558 struct net_device *orig_dev = skb->dev;
5559 struct packet_type *pt_prev = NULL;
5561 skb_list_del_init(skb);
5562 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev);
5565 if (pt_curr != pt_prev || od_curr != orig_dev) {
5566 /* dispatch old sublist */
5567 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr);
5568 /* start new sublist */
5569 INIT_LIST_HEAD(&sublist);
5573 list_add_tail(&skb->list, &sublist);
5576 /* dispatch final sublist */
5577 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr);
5580 static int __netif_receive_skb(struct sk_buff *skb)
5584 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) {
5585 unsigned int noreclaim_flag;
5588 * PFMEMALLOC skbs are special, they should
5589 * - be delivered to SOCK_MEMALLOC sockets only
5590 * - stay away from userspace
5591 * - have bounded memory usage
5593 * Use PF_MEMALLOC as this saves us from propagating the allocation
5594 * context down to all allocation sites.
5596 noreclaim_flag = memalloc_noreclaim_save();
5597 ret = __netif_receive_skb_one_core(skb, true);
5598 memalloc_noreclaim_restore(noreclaim_flag);
5600 ret = __netif_receive_skb_one_core(skb, false);
5605 static void __netif_receive_skb_list(struct list_head *head)
5607 unsigned long noreclaim_flag = 0;
5608 struct sk_buff *skb, *next;
5609 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */
5611 list_for_each_entry_safe(skb, next, head, list) {
5612 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) {
5613 struct list_head sublist;
5615 /* Handle the previous sublist */
5616 list_cut_before(&sublist, head, &skb->list);
5617 if (!list_empty(&sublist))
5618 __netif_receive_skb_list_core(&sublist, pfmemalloc);
5619 pfmemalloc = !pfmemalloc;
5620 /* See comments in __netif_receive_skb */
5622 noreclaim_flag = memalloc_noreclaim_save();
5624 memalloc_noreclaim_restore(noreclaim_flag);
5627 /* Handle the remaining sublist */
5628 if (!list_empty(head))
5629 __netif_receive_skb_list_core(head, pfmemalloc);
5630 /* Restore pflags */
5632 memalloc_noreclaim_restore(noreclaim_flag);
5635 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp)
5637 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog);
5638 struct bpf_prog *new = xdp->prog;
5644 mutex_lock(&new->aux->used_maps_mutex);
5646 /* generic XDP does not work with DEVMAPs that can
5647 * have a bpf_prog installed on an entry
5649 for (i = 0; i < new->aux->used_map_cnt; i++) {
5650 if (dev_map_can_have_prog(new->aux->used_maps[i]) ||
5651 cpu_map_prog_allowed(new->aux->used_maps[i])) {
5652 mutex_unlock(&new->aux->used_maps_mutex);
5657 mutex_unlock(&new->aux->used_maps_mutex);
5660 switch (xdp->command) {
5661 case XDP_SETUP_PROG:
5662 rcu_assign_pointer(dev->xdp_prog, new);
5667 static_branch_dec(&generic_xdp_needed_key);
5668 } else if (new && !old) {
5669 static_branch_inc(&generic_xdp_needed_key);
5670 dev_disable_lro(dev);
5671 dev_disable_gro_hw(dev);
5683 static int netif_receive_skb_internal(struct sk_buff *skb)
5687 net_timestamp_check(netdev_tstamp_prequeue, skb);
5689 if (skb_defer_rx_timestamp(skb))
5690 return NET_RX_SUCCESS;
5694 if (static_branch_unlikely(&rps_needed)) {
5695 struct rps_dev_flow voidflow, *rflow = &voidflow;
5696 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
5699 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
5705 ret = __netif_receive_skb(skb);
5710 static void netif_receive_skb_list_internal(struct list_head *head)
5712 struct sk_buff *skb, *next;
5713 struct list_head sublist;
5715 INIT_LIST_HEAD(&sublist);
5716 list_for_each_entry_safe(skb, next, head, list) {
5717 net_timestamp_check(netdev_tstamp_prequeue, skb);
5718 skb_list_del_init(skb);
5719 if (!skb_defer_rx_timestamp(skb))
5720 list_add_tail(&skb->list, &sublist);
5722 list_splice_init(&sublist, head);
5726 if (static_branch_unlikely(&rps_needed)) {
5727 list_for_each_entry_safe(skb, next, head, list) {
5728 struct rps_dev_flow voidflow, *rflow = &voidflow;
5729 int cpu = get_rps_cpu(skb->dev, skb, &rflow);
5732 /* Will be handled, remove from list */
5733 skb_list_del_init(skb);
5734 enqueue_to_backlog(skb, cpu, &rflow->last_qtail);
5739 __netif_receive_skb_list(head);
5744 * netif_receive_skb - process receive buffer from network
5745 * @skb: buffer to process
5747 * netif_receive_skb() is the main receive data processing function.
5748 * It always succeeds. The buffer may be dropped during processing
5749 * for congestion control or by the protocol layers.
5751 * This function may only be called from softirq context and interrupts
5752 * should be enabled.
5754 * Return values (usually ignored):
5755 * NET_RX_SUCCESS: no congestion
5756 * NET_RX_DROP: packet was dropped
5758 int netif_receive_skb(struct sk_buff *skb)
5762 trace_netif_receive_skb_entry(skb);
5764 ret = netif_receive_skb_internal(skb);
5765 trace_netif_receive_skb_exit(ret);
5769 EXPORT_SYMBOL(netif_receive_skb);
5772 * netif_receive_skb_list - process many receive buffers from network
5773 * @head: list of skbs to process.
5775 * Since return value of netif_receive_skb() is normally ignored, and
5776 * wouldn't be meaningful for a list, this function returns void.
5778 * This function may only be called from softirq context and interrupts
5779 * should be enabled.
5781 void netif_receive_skb_list(struct list_head *head)
5783 struct sk_buff *skb;
5785 if (list_empty(head))
5787 if (trace_netif_receive_skb_list_entry_enabled()) {
5788 list_for_each_entry(skb, head, list)
5789 trace_netif_receive_skb_list_entry(skb);
5791 netif_receive_skb_list_internal(head);
5792 trace_netif_receive_skb_list_exit(0);
5794 EXPORT_SYMBOL(netif_receive_skb_list);
5796 static DEFINE_PER_CPU(struct work_struct, flush_works);
5798 /* Network device is going away, flush any packets still pending */
5799 static void flush_backlog(struct work_struct *work)
5801 struct sk_buff *skb, *tmp;
5802 struct softnet_data *sd;
5805 sd = this_cpu_ptr(&softnet_data);
5807 local_irq_disable();
5809 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) {
5810 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
5811 __skb_unlink(skb, &sd->input_pkt_queue);
5812 dev_kfree_skb_irq(skb);
5813 input_queue_head_incr(sd);
5819 skb_queue_walk_safe(&sd->process_queue, skb, tmp) {
5820 if (skb->dev->reg_state == NETREG_UNREGISTERING) {
5821 __skb_unlink(skb, &sd->process_queue);
5823 input_queue_head_incr(sd);
5829 static bool flush_required(int cpu)
5831 #if IS_ENABLED(CONFIG_RPS)
5832 struct softnet_data *sd = &per_cpu(softnet_data, cpu);
5835 local_irq_disable();
5838 /* as insertion into process_queue happens with the rps lock held,
5839 * process_queue access may race only with dequeue
5841 do_flush = !skb_queue_empty(&sd->input_pkt_queue) ||
5842 !skb_queue_empty_lockless(&sd->process_queue);
5848 /* without RPS we can't safely check input_pkt_queue: during a
5849 * concurrent remote skb_queue_splice() we can detect as empty both
5850 * input_pkt_queue and process_queue even if the latter could end-up
5851 * containing a lot of packets.
5856 static void flush_all_backlogs(void)
5858 static cpumask_t flush_cpus;
5861 /* since we are under rtnl lock protection we can use static data
5862 * for the cpumask and avoid allocating on stack the possibly
5869 cpumask_clear(&flush_cpus);
5870 for_each_online_cpu(cpu) {
5871 if (flush_required(cpu)) {
5872 queue_work_on(cpu, system_highpri_wq,
5873 per_cpu_ptr(&flush_works, cpu));
5874 cpumask_set_cpu(cpu, &flush_cpus);
5878 /* we can have in flight packet[s] on the cpus we are not flushing,
5879 * synchronize_net() in unregister_netdevice_many() will take care of
5882 for_each_cpu(cpu, &flush_cpus)
5883 flush_work(per_cpu_ptr(&flush_works, cpu));
5888 /* Pass the currently batched GRO_NORMAL SKBs up to the stack. */
5889 static void gro_normal_list(struct napi_struct *napi)
5891 if (!napi->rx_count)
5893 netif_receive_skb_list_internal(&napi->rx_list);
5894 INIT_LIST_HEAD(&napi->rx_list);
5898 /* Queue one GRO_NORMAL SKB up for list processing. If batch size exceeded,
5899 * pass the whole batch up to the stack.
5901 static void gro_normal_one(struct napi_struct *napi, struct sk_buff *skb, int segs)
5903 list_add_tail(&skb->list, &napi->rx_list);
5904 napi->rx_count += segs;
5905 if (napi->rx_count >= gro_normal_batch)
5906 gro_normal_list(napi);
5909 static int napi_gro_complete(struct napi_struct *napi, struct sk_buff *skb)
5911 struct packet_offload *ptype;
5912 __be16 type = skb->protocol;
5913 struct list_head *head = &offload_base;
5916 BUILD_BUG_ON(sizeof(struct napi_gro_cb) > sizeof(skb->cb));
5918 if (NAPI_GRO_CB(skb)->count == 1) {
5919 skb_shinfo(skb)->gso_size = 0;
5924 list_for_each_entry_rcu(ptype, head, list) {
5925 if (ptype->type != type || !ptype->callbacks.gro_complete)
5928 err = INDIRECT_CALL_INET(ptype->callbacks.gro_complete,
5929 ipv6_gro_complete, inet_gro_complete,
5936 WARN_ON(&ptype->list == head);
5938 return NET_RX_SUCCESS;
5942 gro_normal_one(napi, skb, NAPI_GRO_CB(skb)->count);
5943 return NET_RX_SUCCESS;
5946 static void __napi_gro_flush_chain(struct napi_struct *napi, u32 index,
5949 struct list_head *head = &napi->gro_hash[index].list;
5950 struct sk_buff *skb, *p;
5952 list_for_each_entry_safe_reverse(skb, p, head, list) {
5953 if (flush_old && NAPI_GRO_CB(skb)->age == jiffies)
5955 skb_list_del_init(skb);
5956 napi_gro_complete(napi, skb);
5957 napi->gro_hash[index].count--;
5960 if (!napi->gro_hash[index].count)
5961 __clear_bit(index, &napi->gro_bitmask);
5964 /* napi->gro_hash[].list contains packets ordered by age.
5965 * youngest packets at the head of it.
5966 * Complete skbs in reverse order to reduce latencies.
5968 void napi_gro_flush(struct napi_struct *napi, bool flush_old)
5970 unsigned long bitmask = napi->gro_bitmask;
5971 unsigned int i, base = ~0U;
5973 while ((i = ffs(bitmask)) != 0) {
5976 __napi_gro_flush_chain(napi, base, flush_old);
5979 EXPORT_SYMBOL(napi_gro_flush);
5981 static void gro_list_prepare(const struct list_head *head,
5982 const struct sk_buff *skb)
5984 unsigned int maclen = skb->dev->hard_header_len;
5985 u32 hash = skb_get_hash_raw(skb);
5988 list_for_each_entry(p, head, list) {
5989 unsigned long diffs;
5991 NAPI_GRO_CB(p)->flush = 0;
5993 if (hash != skb_get_hash_raw(p)) {
5994 NAPI_GRO_CB(p)->same_flow = 0;
5998 diffs = (unsigned long)p->dev ^ (unsigned long)skb->dev;
5999 diffs |= skb_vlan_tag_present(p) ^ skb_vlan_tag_present(skb);
6000 if (skb_vlan_tag_present(p))
6001 diffs |= skb_vlan_tag_get(p) ^ skb_vlan_tag_get(skb);
6002 diffs |= skb_metadata_dst_cmp(p, skb);
6003 diffs |= skb_metadata_differs(p, skb);
6004 if (maclen == ETH_HLEN)
6005 diffs |= compare_ether_header(skb_mac_header(p),
6006 skb_mac_header(skb));
6008 diffs = memcmp(skb_mac_header(p),
6009 skb_mac_header(skb),
6012 diffs |= skb_get_nfct(p) ^ skb_get_nfct(skb);
6013 #if IS_ENABLED(CONFIG_SKB_EXTENSIONS) && IS_ENABLED(CONFIG_NET_TC_SKB_EXT)
6015 struct tc_skb_ext *skb_ext = skb_ext_find(skb, TC_SKB_EXT);
6016 struct tc_skb_ext *p_ext = skb_ext_find(p, TC_SKB_EXT);
6018 diffs |= (!!p_ext) ^ (!!skb_ext);
6019 if (!diffs && unlikely(skb_ext))
6020 diffs |= p_ext->chain ^ skb_ext->chain;
6024 NAPI_GRO_CB(p)->same_flow = !diffs;
6028 static inline void skb_gro_reset_offset(struct sk_buff *skb, u32 nhoff)
6030 const struct skb_shared_info *pinfo = skb_shinfo(skb);
6031 const skb_frag_t *frag0 = &pinfo->frags[0];
6033 NAPI_GRO_CB(skb)->data_offset = 0;
6034 NAPI_GRO_CB(skb)->frag0 = NULL;
6035 NAPI_GRO_CB(skb)->frag0_len = 0;
6037 if (!skb_headlen(skb) && pinfo->nr_frags &&
6038 !PageHighMem(skb_frag_page(frag0)) &&
6039 (!NET_IP_ALIGN || !((skb_frag_off(frag0) + nhoff) & 3))) {
6040 NAPI_GRO_CB(skb)->frag0 = skb_frag_address(frag0);
6041 NAPI_GRO_CB(skb)->frag0_len = min_t(unsigned int,
6042 skb_frag_size(frag0),
6043 skb->end - skb->tail);
6047 static void gro_pull_from_frag0(struct sk_buff *skb, int grow)
6049 struct skb_shared_info *pinfo = skb_shinfo(skb);
6051 BUG_ON(skb->end - skb->tail < grow);
6053 memcpy(skb_tail_pointer(skb), NAPI_GRO_CB(skb)->frag0, grow);
6055 skb->data_len -= grow;
6058 skb_frag_off_add(&pinfo->frags[0], grow);
6059 skb_frag_size_sub(&pinfo->frags[0], grow);
6061 if (unlikely(!skb_frag_size(&pinfo->frags[0]))) {
6062 skb_frag_unref(skb, 0);
6063 memmove(pinfo->frags, pinfo->frags + 1,
6064 --pinfo->nr_frags * sizeof(pinfo->frags[0]));
6068 static void gro_flush_oldest(struct napi_struct *napi, struct list_head *head)
6070 struct sk_buff *oldest;
6072 oldest = list_last_entry(head, struct sk_buff, list);
6074 /* We are called with head length >= MAX_GRO_SKBS, so this is
6077 if (WARN_ON_ONCE(!oldest))
6080 /* Do not adjust napi->gro_hash[].count, caller is adding a new
6083 skb_list_del_init(oldest);
6084 napi_gro_complete(napi, oldest);
6087 static enum gro_result dev_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
6089 u32 bucket = skb_get_hash_raw(skb) & (GRO_HASH_BUCKETS - 1);
6090 struct gro_list *gro_list = &napi->gro_hash[bucket];
6091 struct list_head *head = &offload_base;
6092 struct packet_offload *ptype;
6093 __be16 type = skb->protocol;
6094 struct sk_buff *pp = NULL;
6095 enum gro_result ret;
6099 if (netif_elide_gro(skb->dev))
6102 gro_list_prepare(&gro_list->list, skb);
6105 list_for_each_entry_rcu(ptype, head, list) {
6106 if (ptype->type != type || !ptype->callbacks.gro_receive)
6109 skb_set_network_header(skb, skb_gro_offset(skb));
6110 skb_reset_mac_len(skb);
6111 NAPI_GRO_CB(skb)->same_flow = 0;
6112 NAPI_GRO_CB(skb)->flush = skb_is_gso(skb) || skb_has_frag_list(skb);
6113 NAPI_GRO_CB(skb)->free = 0;
6114 NAPI_GRO_CB(skb)->encap_mark = 0;
6115 NAPI_GRO_CB(skb)->recursion_counter = 0;
6116 NAPI_GRO_CB(skb)->is_fou = 0;
6117 NAPI_GRO_CB(skb)->is_atomic = 1;
6118 NAPI_GRO_CB(skb)->gro_remcsum_start = 0;
6120 /* Setup for GRO checksum validation */
6121 switch (skb->ip_summed) {
6122 case CHECKSUM_COMPLETE:
6123 NAPI_GRO_CB(skb)->csum = skb->csum;
6124 NAPI_GRO_CB(skb)->csum_valid = 1;
6125 NAPI_GRO_CB(skb)->csum_cnt = 0;
6127 case CHECKSUM_UNNECESSARY:
6128 NAPI_GRO_CB(skb)->csum_cnt = skb->csum_level + 1;
6129 NAPI_GRO_CB(skb)->csum_valid = 0;
6132 NAPI_GRO_CB(skb)->csum_cnt = 0;
6133 NAPI_GRO_CB(skb)->csum_valid = 0;
6136 pp = INDIRECT_CALL_INET(ptype->callbacks.gro_receive,
6137 ipv6_gro_receive, inet_gro_receive,
6138 &gro_list->list, skb);
6143 if (&ptype->list == head)
6146 if (PTR_ERR(pp) == -EINPROGRESS) {
6151 same_flow = NAPI_GRO_CB(skb)->same_flow;
6152 ret = NAPI_GRO_CB(skb)->free ? GRO_MERGED_FREE : GRO_MERGED;
6155 skb_list_del_init(pp);
6156 napi_gro_complete(napi, pp);
6163 if (NAPI_GRO_CB(skb)->flush)
6166 if (unlikely(gro_list->count >= MAX_GRO_SKBS))
6167 gro_flush_oldest(napi, &gro_list->list);
6171 NAPI_GRO_CB(skb)->count = 1;
6172 NAPI_GRO_CB(skb)->age = jiffies;
6173 NAPI_GRO_CB(skb)->last = skb;
6174 skb_shinfo(skb)->gso_size = skb_gro_len(skb);
6175 list_add(&skb->list, &gro_list->list);
6179 grow = skb_gro_offset(skb) - skb_headlen(skb);
6181 gro_pull_from_frag0(skb, grow);
6183 if (gro_list->count) {
6184 if (!test_bit(bucket, &napi->gro_bitmask))
6185 __set_bit(bucket, &napi->gro_bitmask);
6186 } else if (test_bit(bucket, &napi->gro_bitmask)) {
6187 __clear_bit(bucket, &napi->gro_bitmask);
6197 struct packet_offload *gro_find_receive_by_type(__be16 type)
6199 struct list_head *offload_head = &offload_base;
6200 struct packet_offload *ptype;
6202 list_for_each_entry_rcu(ptype, offload_head, list) {
6203 if (ptype->type != type || !ptype->callbacks.gro_receive)
6209 EXPORT_SYMBOL(gro_find_receive_by_type);
6211 struct packet_offload *gro_find_complete_by_type(__be16 type)
6213 struct list_head *offload_head = &offload_base;
6214 struct packet_offload *ptype;
6216 list_for_each_entry_rcu(ptype, offload_head, list) {
6217 if (ptype->type != type || !ptype->callbacks.gro_complete)
6223 EXPORT_SYMBOL(gro_find_complete_by_type);
6225 static gro_result_t napi_skb_finish(struct napi_struct *napi,
6226 struct sk_buff *skb,
6231 gro_normal_one(napi, skb, 1);
6234 case GRO_MERGED_FREE:
6235 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
6236 napi_skb_free_stolen_head(skb);
6237 else if (skb->fclone != SKB_FCLONE_UNAVAILABLE)
6240 __kfree_skb_defer(skb);
6252 gro_result_t napi_gro_receive(struct napi_struct *napi, struct sk_buff *skb)
6256 skb_mark_napi_id(skb, napi);
6257 trace_napi_gro_receive_entry(skb);
6259 skb_gro_reset_offset(skb, 0);
6261 ret = napi_skb_finish(napi, skb, dev_gro_receive(napi, skb));
6262 trace_napi_gro_receive_exit(ret);
6266 EXPORT_SYMBOL(napi_gro_receive);
6268 static void napi_reuse_skb(struct napi_struct *napi, struct sk_buff *skb)
6270 if (unlikely(skb->pfmemalloc)) {
6274 __skb_pull(skb, skb_headlen(skb));
6275 /* restore the reserve we had after netdev_alloc_skb_ip_align() */
6276 skb_reserve(skb, NET_SKB_PAD + NET_IP_ALIGN - skb_headroom(skb));
6277 __vlan_hwaccel_clear_tag(skb);
6278 skb->dev = napi->dev;
6281 /* eth_type_trans() assumes pkt_type is PACKET_HOST */
6282 skb->pkt_type = PACKET_HOST;
6284 skb->encapsulation = 0;
6285 skb_shinfo(skb)->gso_type = 0;
6286 skb->truesize = SKB_TRUESIZE(skb_end_offset(skb));
6293 struct sk_buff *napi_get_frags(struct napi_struct *napi)
6295 struct sk_buff *skb = napi->skb;
6298 skb = napi_alloc_skb(napi, GRO_MAX_HEAD);
6301 skb_mark_napi_id(skb, napi);
6306 EXPORT_SYMBOL(napi_get_frags);
6308 static gro_result_t napi_frags_finish(struct napi_struct *napi,
6309 struct sk_buff *skb,
6315 __skb_push(skb, ETH_HLEN);
6316 skb->protocol = eth_type_trans(skb, skb->dev);
6317 if (ret == GRO_NORMAL)
6318 gro_normal_one(napi, skb, 1);
6321 case GRO_MERGED_FREE:
6322 if (NAPI_GRO_CB(skb)->free == NAPI_GRO_FREE_STOLEN_HEAD)
6323 napi_skb_free_stolen_head(skb);
6325 napi_reuse_skb(napi, skb);
6336 /* Upper GRO stack assumes network header starts at gro_offset=0
6337 * Drivers could call both napi_gro_frags() and napi_gro_receive()
6338 * We copy ethernet header into skb->data to have a common layout.
6340 static struct sk_buff *napi_frags_skb(struct napi_struct *napi)
6342 struct sk_buff *skb = napi->skb;
6343 const struct ethhdr *eth;
6344 unsigned int hlen = sizeof(*eth);
6348 skb_reset_mac_header(skb);
6349 skb_gro_reset_offset(skb, hlen);
6351 if (unlikely(skb_gro_header_hard(skb, hlen))) {
6352 eth = skb_gro_header_slow(skb, hlen, 0);
6353 if (unlikely(!eth)) {
6354 net_warn_ratelimited("%s: dropping impossible skb from %s\n",
6355 __func__, napi->dev->name);
6356 napi_reuse_skb(napi, skb);
6360 eth = (const struct ethhdr *)skb->data;
6361 gro_pull_from_frag0(skb, hlen);
6362 NAPI_GRO_CB(skb)->frag0 += hlen;
6363 NAPI_GRO_CB(skb)->frag0_len -= hlen;
6365 __skb_pull(skb, hlen);
6368 * This works because the only protocols we care about don't require
6370 * We'll fix it up properly in napi_frags_finish()
6372 skb->protocol = eth->h_proto;
6377 gro_result_t napi_gro_frags(struct napi_struct *napi)
6380 struct sk_buff *skb = napi_frags_skb(napi);
6382 trace_napi_gro_frags_entry(skb);
6384 ret = napi_frags_finish(napi, skb, dev_gro_receive(napi, skb));
6385 trace_napi_gro_frags_exit(ret);
6389 EXPORT_SYMBOL(napi_gro_frags);
6391 /* Compute the checksum from gro_offset and return the folded value
6392 * after adding in any pseudo checksum.
6394 __sum16 __skb_gro_checksum_complete(struct sk_buff *skb)
6399 wsum = skb_checksum(skb, skb_gro_offset(skb), skb_gro_len(skb), 0);
6401 /* NAPI_GRO_CB(skb)->csum holds pseudo checksum */
6402 sum = csum_fold(csum_add(NAPI_GRO_CB(skb)->csum, wsum));
6403 /* See comments in __skb_checksum_complete(). */
6405 if (unlikely(skb->ip_summed == CHECKSUM_COMPLETE) &&
6406 !skb->csum_complete_sw)
6407 netdev_rx_csum_fault(skb->dev, skb);
6410 NAPI_GRO_CB(skb)->csum = wsum;
6411 NAPI_GRO_CB(skb)->csum_valid = 1;
6415 EXPORT_SYMBOL(__skb_gro_checksum_complete);
6417 static void net_rps_send_ipi(struct softnet_data *remsd)
6421 struct softnet_data *next = remsd->rps_ipi_next;
6423 if (cpu_online(remsd->cpu))
6424 smp_call_function_single_async(remsd->cpu, &remsd->csd);
6431 * net_rps_action_and_irq_enable sends any pending IPI's for rps.
6432 * Note: called with local irq disabled, but exits with local irq enabled.
6434 static void net_rps_action_and_irq_enable(struct softnet_data *sd)
6437 struct softnet_data *remsd = sd->rps_ipi_list;
6440 sd->rps_ipi_list = NULL;
6444 /* Send pending IPI's to kick RPS processing on remote cpus. */
6445 net_rps_send_ipi(remsd);
6451 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd)
6454 return sd->rps_ipi_list != NULL;
6460 static int process_backlog(struct napi_struct *napi, int quota)
6462 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog);
6466 /* Check if we have pending ipi, its better to send them now,
6467 * not waiting net_rx_action() end.
6469 if (sd_has_rps_ipi_waiting(sd)) {
6470 local_irq_disable();
6471 net_rps_action_and_irq_enable(sd);
6474 napi->weight = dev_rx_weight;
6476 struct sk_buff *skb;
6478 while ((skb = __skb_dequeue(&sd->process_queue))) {
6480 __netif_receive_skb(skb);
6482 input_queue_head_incr(sd);
6483 if (++work >= quota)
6488 local_irq_disable();
6490 if (skb_queue_empty(&sd->input_pkt_queue)) {
6492 * Inline a custom version of __napi_complete().
6493 * only current cpu owns and manipulates this napi,
6494 * and NAPI_STATE_SCHED is the only possible flag set
6496 * We can use a plain write instead of clear_bit(),
6497 * and we dont need an smp_mb() memory barrier.
6502 skb_queue_splice_tail_init(&sd->input_pkt_queue,
6503 &sd->process_queue);
6513 * __napi_schedule - schedule for receive
6514 * @n: entry to schedule
6516 * The entry's receive function will be scheduled to run.
6517 * Consider using __napi_schedule_irqoff() if hard irqs are masked.
6519 void __napi_schedule(struct napi_struct *n)
6521 unsigned long flags;
6523 local_irq_save(flags);
6524 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
6525 local_irq_restore(flags);
6527 EXPORT_SYMBOL(__napi_schedule);
6530 * napi_schedule_prep - check if napi can be scheduled
6533 * Test if NAPI routine is already running, and if not mark
6534 * it as running. This is used as a condition variable to
6535 * insure only one NAPI poll instance runs. We also make
6536 * sure there is no pending NAPI disable.
6538 bool napi_schedule_prep(struct napi_struct *n)
6540 unsigned long val, new;
6543 val = READ_ONCE(n->state);
6544 if (unlikely(val & NAPIF_STATE_DISABLE))
6546 new = val | NAPIF_STATE_SCHED;
6548 /* Sets STATE_MISSED bit if STATE_SCHED was already set
6549 * This was suggested by Alexander Duyck, as compiler
6550 * emits better code than :
6551 * if (val & NAPIF_STATE_SCHED)
6552 * new |= NAPIF_STATE_MISSED;
6554 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED *
6556 } while (cmpxchg(&n->state, val, new) != val);
6558 return !(val & NAPIF_STATE_SCHED);
6560 EXPORT_SYMBOL(napi_schedule_prep);
6563 * __napi_schedule_irqoff - schedule for receive
6564 * @n: entry to schedule
6566 * Variant of __napi_schedule() assuming hard irqs are masked.
6568 * On PREEMPT_RT enabled kernels this maps to __napi_schedule()
6569 * because the interrupt disabled assumption might not be true
6570 * due to force-threaded interrupts and spinlock substitution.
6572 void __napi_schedule_irqoff(struct napi_struct *n)
6574 if (!IS_ENABLED(CONFIG_PREEMPT_RT))
6575 ____napi_schedule(this_cpu_ptr(&softnet_data), n);
6579 EXPORT_SYMBOL(__napi_schedule_irqoff);
6581 bool napi_complete_done(struct napi_struct *n, int work_done)
6583 unsigned long flags, val, new, timeout = 0;
6587 * 1) Don't let napi dequeue from the cpu poll list
6588 * just in case its running on a different cpu.
6589 * 2) If we are busy polling, do nothing here, we have
6590 * the guarantee we will be called later.
6592 if (unlikely(n->state & (NAPIF_STATE_NPSVC |
6593 NAPIF_STATE_IN_BUSY_POLL)))
6598 timeout = READ_ONCE(n->dev->gro_flush_timeout);
6599 n->defer_hard_irqs_count = READ_ONCE(n->dev->napi_defer_hard_irqs);
6601 if (n->defer_hard_irqs_count > 0) {
6602 n->defer_hard_irqs_count--;
6603 timeout = READ_ONCE(n->dev->gro_flush_timeout);
6607 if (n->gro_bitmask) {
6608 /* When the NAPI instance uses a timeout and keeps postponing
6609 * it, we need to bound somehow the time packets are kept in
6612 napi_gro_flush(n, !!timeout);
6617 if (unlikely(!list_empty(&n->poll_list))) {
6618 /* If n->poll_list is not empty, we need to mask irqs */
6619 local_irq_save(flags);
6620 list_del_init(&n->poll_list);
6621 local_irq_restore(flags);
6625 val = READ_ONCE(n->state);
6627 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED));
6629 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED |
6630 NAPIF_STATE_SCHED_THREADED |
6631 NAPIF_STATE_PREFER_BUSY_POLL);
6633 /* If STATE_MISSED was set, leave STATE_SCHED set,
6634 * because we will call napi->poll() one more time.
6635 * This C code was suggested by Alexander Duyck to help gcc.
6637 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED *
6639 } while (cmpxchg(&n->state, val, new) != val);
6641 if (unlikely(val & NAPIF_STATE_MISSED)) {
6647 hrtimer_start(&n->timer, ns_to_ktime(timeout),
6648 HRTIMER_MODE_REL_PINNED);
6651 EXPORT_SYMBOL(napi_complete_done);
6653 /* must be called under rcu_read_lock(), as we dont take a reference */
6654 static struct napi_struct *napi_by_id(unsigned int napi_id)
6656 unsigned int hash = napi_id % HASH_SIZE(napi_hash);
6657 struct napi_struct *napi;
6659 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node)
6660 if (napi->napi_id == napi_id)
6666 #if defined(CONFIG_NET_RX_BUSY_POLL)
6668 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule)
6670 if (!skip_schedule) {
6671 gro_normal_list(napi);
6672 __napi_schedule(napi);
6676 if (napi->gro_bitmask) {
6677 /* flush too old packets
6678 * If HZ < 1000, flush all packets.
6680 napi_gro_flush(napi, HZ >= 1000);
6683 gro_normal_list(napi);
6684 clear_bit(NAPI_STATE_SCHED, &napi->state);
6687 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, bool prefer_busy_poll,
6690 bool skip_schedule = false;
6691 unsigned long timeout;
6694 /* Busy polling means there is a high chance device driver hard irq
6695 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was
6696 * set in napi_schedule_prep().
6697 * Since we are about to call napi->poll() once more, we can safely
6698 * clear NAPI_STATE_MISSED.
6700 * Note: x86 could use a single "lock and ..." instruction
6701 * to perform these two clear_bit()
6703 clear_bit(NAPI_STATE_MISSED, &napi->state);
6704 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state);
6708 if (prefer_busy_poll) {
6709 napi->defer_hard_irqs_count = READ_ONCE(napi->dev->napi_defer_hard_irqs);
6710 timeout = READ_ONCE(napi->dev->gro_flush_timeout);
6711 if (napi->defer_hard_irqs_count && timeout) {
6712 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED);
6713 skip_schedule = true;
6717 /* All we really want here is to re-enable device interrupts.
6718 * Ideally, a new ndo_busy_poll_stop() could avoid another round.
6720 rc = napi->poll(napi, budget);
6721 /* We can't gro_normal_list() here, because napi->poll() might have
6722 * rearmed the napi (napi_complete_done()) in which case it could
6723 * already be running on another CPU.
6725 trace_napi_poll(napi, rc, budget);
6726 netpoll_poll_unlock(have_poll_lock);
6728 __busy_poll_stop(napi, skip_schedule);
6732 void napi_busy_loop(unsigned int napi_id,
6733 bool (*loop_end)(void *, unsigned long),
6734 void *loop_end_arg, bool prefer_busy_poll, u16 budget)
6736 unsigned long start_time = loop_end ? busy_loop_current_time() : 0;
6737 int (*napi_poll)(struct napi_struct *napi, int budget);
6738 void *have_poll_lock = NULL;
6739 struct napi_struct *napi;
6746 napi = napi_by_id(napi_id);
6756 unsigned long val = READ_ONCE(napi->state);
6758 /* If multiple threads are competing for this napi,
6759 * we avoid dirtying napi->state as much as we can.
6761 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED |
6762 NAPIF_STATE_IN_BUSY_POLL)) {
6763 if (prefer_busy_poll)
6764 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state);
6767 if (cmpxchg(&napi->state, val,
6768 val | NAPIF_STATE_IN_BUSY_POLL |
6769 NAPIF_STATE_SCHED) != val) {
6770 if (prefer_busy_poll)
6771 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state);
6774 have_poll_lock = netpoll_poll_lock(napi);
6775 napi_poll = napi->poll;
6777 work = napi_poll(napi, budget);
6778 trace_napi_poll(napi, work, budget);
6779 gro_normal_list(napi);
6782 __NET_ADD_STATS(dev_net(napi->dev),
6783 LINUX_MIB_BUSYPOLLRXPACKETS, work);
6786 if (!loop_end || loop_end(loop_end_arg, start_time))
6789 if (unlikely(need_resched())) {
6791 busy_poll_stop(napi, have_poll_lock, prefer_busy_poll, budget);
6795 if (loop_end(loop_end_arg, start_time))
6802 busy_poll_stop(napi, have_poll_lock, prefer_busy_poll, budget);
6807 EXPORT_SYMBOL(napi_busy_loop);
6809 #endif /* CONFIG_NET_RX_BUSY_POLL */
6811 static void napi_hash_add(struct napi_struct *napi)
6813 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state))
6816 spin_lock(&napi_hash_lock);
6818 /* 0..NR_CPUS range is reserved for sender_cpu use */
6820 if (unlikely(++napi_gen_id < MIN_NAPI_ID))
6821 napi_gen_id = MIN_NAPI_ID;
6822 } while (napi_by_id(napi_gen_id));
6823 napi->napi_id = napi_gen_id;
6825 hlist_add_head_rcu(&napi->napi_hash_node,
6826 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]);
6828 spin_unlock(&napi_hash_lock);
6831 /* Warning : caller is responsible to make sure rcu grace period
6832 * is respected before freeing memory containing @napi
6834 static void napi_hash_del(struct napi_struct *napi)
6836 spin_lock(&napi_hash_lock);
6838 hlist_del_init_rcu(&napi->napi_hash_node);
6840 spin_unlock(&napi_hash_lock);
6843 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer)
6845 struct napi_struct *napi;
6847 napi = container_of(timer, struct napi_struct, timer);
6849 /* Note : we use a relaxed variant of napi_schedule_prep() not setting
6850 * NAPI_STATE_MISSED, since we do not react to a device IRQ.
6852 if (!napi_disable_pending(napi) &&
6853 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) {
6854 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state);
6855 __napi_schedule_irqoff(napi);
6858 return HRTIMER_NORESTART;
6861 static void init_gro_hash(struct napi_struct *napi)
6865 for (i = 0; i < GRO_HASH_BUCKETS; i++) {
6866 INIT_LIST_HEAD(&napi->gro_hash[i].list);
6867 napi->gro_hash[i].count = 0;
6869 napi->gro_bitmask = 0;
6872 int dev_set_threaded(struct net_device *dev, bool threaded)
6874 struct napi_struct *napi;
6877 if (dev->threaded == threaded)
6881 list_for_each_entry(napi, &dev->napi_list, dev_list) {
6882 if (!napi->thread) {
6883 err = napi_kthread_create(napi);
6892 dev->threaded = threaded;
6894 /* Make sure kthread is created before THREADED bit
6897 smp_mb__before_atomic();
6899 /* Setting/unsetting threaded mode on a napi might not immediately
6900 * take effect, if the current napi instance is actively being
6901 * polled. In this case, the switch between threaded mode and
6902 * softirq mode will happen in the next round of napi_schedule().
6903 * This should not cause hiccups/stalls to the live traffic.
6905 list_for_each_entry(napi, &dev->napi_list, dev_list) {
6907 set_bit(NAPI_STATE_THREADED, &napi->state);
6909 clear_bit(NAPI_STATE_THREADED, &napi->state);
6914 EXPORT_SYMBOL(dev_set_threaded);
6916 void netif_napi_add(struct net_device *dev, struct napi_struct *napi,
6917 int (*poll)(struct napi_struct *, int), int weight)
6919 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state)))
6922 INIT_LIST_HEAD(&napi->poll_list);
6923 INIT_HLIST_NODE(&napi->napi_hash_node);
6924 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
6925 napi->timer.function = napi_watchdog;
6926 init_gro_hash(napi);
6928 INIT_LIST_HEAD(&napi->rx_list);
6931 if (weight > NAPI_POLL_WEIGHT)
6932 netdev_err_once(dev, "%s() called with weight %d\n", __func__,
6934 napi->weight = weight;
6936 #ifdef CONFIG_NETPOLL
6937 napi->poll_owner = -1;
6939 set_bit(NAPI_STATE_SCHED, &napi->state);
6940 set_bit(NAPI_STATE_NPSVC, &napi->state);
6941 list_add_rcu(&napi->dev_list, &dev->napi_list);
6942 napi_hash_add(napi);
6943 /* Create kthread for this napi if dev->threaded is set.
6944 * Clear dev->threaded if kthread creation failed so that
6945 * threaded mode will not be enabled in napi_enable().
6947 if (dev->threaded && napi_kthread_create(napi))
6950 EXPORT_SYMBOL(netif_napi_add);
6952 void napi_disable(struct napi_struct *n)
6955 set_bit(NAPI_STATE_DISABLE, &n->state);
6957 while (test_and_set_bit(NAPI_STATE_SCHED, &n->state))
6959 while (test_and_set_bit(NAPI_STATE_NPSVC, &n->state))
6962 hrtimer_cancel(&n->timer);
6964 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &n->state);
6965 clear_bit(NAPI_STATE_DISABLE, &n->state);
6966 clear_bit(NAPI_STATE_THREADED, &n->state);
6968 EXPORT_SYMBOL(napi_disable);
6971 * napi_enable - enable NAPI scheduling
6974 * Resume NAPI from being scheduled on this context.
6975 * Must be paired with napi_disable.
6977 void napi_enable(struct napi_struct *n)
6979 BUG_ON(!test_bit(NAPI_STATE_SCHED, &n->state));
6980 smp_mb__before_atomic();
6981 clear_bit(NAPI_STATE_SCHED, &n->state);
6982 clear_bit(NAPI_STATE_NPSVC, &n->state);
6983 if (n->dev->threaded && n->thread)
6984 set_bit(NAPI_STATE_THREADED, &n->state);
6986 EXPORT_SYMBOL(napi_enable);
6988 static void flush_gro_hash(struct napi_struct *napi)
6992 for (i = 0; i < GRO_HASH_BUCKETS; i++) {
6993 struct sk_buff *skb, *n;
6995 list_for_each_entry_safe(skb, n, &napi->gro_hash[i].list, list)
6997 napi->gro_hash[i].count = 0;
7001 /* Must be called in process context */
7002 void __netif_napi_del(struct napi_struct *napi)
7004 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state))
7007 napi_hash_del(napi);
7008 list_del_rcu(&napi->dev_list);
7009 napi_free_frags(napi);
7011 flush_gro_hash(napi);
7012 napi->gro_bitmask = 0;
7015 kthread_stop(napi->thread);
7016 napi->thread = NULL;
7019 EXPORT_SYMBOL(__netif_napi_del);
7021 static int __napi_poll(struct napi_struct *n, bool *repoll)
7027 /* This NAPI_STATE_SCHED test is for avoiding a race
7028 * with netpoll's poll_napi(). Only the entity which
7029 * obtains the lock and sees NAPI_STATE_SCHED set will
7030 * actually make the ->poll() call. Therefore we avoid
7031 * accidentally calling ->poll() when NAPI is not scheduled.
7034 if (test_bit(NAPI_STATE_SCHED, &n->state)) {
7035 work = n->poll(n, weight);
7036 trace_napi_poll(n, work, weight);
7039 if (unlikely(work > weight))
7040 pr_err_once("NAPI poll function %pS returned %d, exceeding its budget of %d.\n",
7041 n->poll, work, weight);
7043 if (likely(work < weight))
7046 /* Drivers must not modify the NAPI state if they
7047 * consume the entire weight. In such cases this code
7048 * still "owns" the NAPI instance and therefore can
7049 * move the instance around on the list at-will.
7051 if (unlikely(napi_disable_pending(n))) {
7056 /* The NAPI context has more processing work, but busy-polling
7057 * is preferred. Exit early.
7059 if (napi_prefer_busy_poll(n)) {
7060 if (napi_complete_done(n, work)) {
7061 /* If timeout is not set, we need to make sure
7062 * that the NAPI is re-scheduled.
7069 if (n->gro_bitmask) {
7070 /* flush too old packets
7071 * If HZ < 1000, flush all packets.
7073 napi_gro_flush(n, HZ >= 1000);
7078 /* Some drivers may have called napi_schedule
7079 * prior to exhausting their budget.
7081 if (unlikely(!list_empty(&n->poll_list))) {
7082 pr_warn_once("%s: Budget exhausted after napi rescheduled\n",
7083 n->dev ? n->dev->name : "backlog");
7092 static int napi_poll(struct napi_struct *n, struct list_head *repoll)
7094 bool do_repoll = false;
7098 list_del_init(&n->poll_list);
7100 have = netpoll_poll_lock(n);
7102 work = __napi_poll(n, &do_repoll);
7105 list_add_tail(&n->poll_list, repoll);
7107 netpoll_poll_unlock(have);
7112 static int napi_thread_wait(struct napi_struct *napi)
7116 set_current_state(TASK_INTERRUPTIBLE);
7118 while (!kthread_should_stop()) {
7119 /* Testing SCHED_THREADED bit here to make sure the current
7120 * kthread owns this napi and could poll on this napi.
7121 * Testing SCHED bit is not enough because SCHED bit might be
7122 * set by some other busy poll thread or by napi_disable().
7124 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state) || woken) {
7125 WARN_ON(!list_empty(&napi->poll_list));
7126 __set_current_state(TASK_RUNNING);
7131 /* woken being true indicates this thread owns this napi. */
7133 set_current_state(TASK_INTERRUPTIBLE);
7135 __set_current_state(TASK_RUNNING);
7140 static int napi_threaded_poll(void *data)
7142 struct napi_struct *napi = data;
7145 while (!napi_thread_wait(napi)) {
7147 bool repoll = false;
7151 have = netpoll_poll_lock(napi);
7152 __napi_poll(napi, &repoll);
7153 netpoll_poll_unlock(have);
7166 static __latent_entropy void net_rx_action(struct softirq_action *h)
7168 struct softnet_data *sd = this_cpu_ptr(&softnet_data);
7169 unsigned long time_limit = jiffies +
7170 usecs_to_jiffies(netdev_budget_usecs);
7171 int budget = netdev_budget;
7175 local_irq_disable();
7176 list_splice_init(&sd->poll_list, &list);
7180 struct napi_struct *n;
7182 if (list_empty(&list)) {
7183 if (!sd_has_rps_ipi_waiting(sd) && list_empty(&repoll))
7188 n = list_first_entry(&list, struct napi_struct, poll_list);
7189 budget -= napi_poll(n, &repoll);
7191 /* If softirq window is exhausted then punt.
7192 * Allow this to run for 2 jiffies since which will allow
7193 * an average latency of 1.5/HZ.
7195 if (unlikely(budget <= 0 ||
7196 time_after_eq(jiffies, time_limit))) {
7202 local_irq_disable();
7204 list_splice_tail_init(&sd->poll_list, &list);
7205 list_splice_tail(&repoll, &list);
7206 list_splice(&list, &sd->poll_list);
7207 if (!list_empty(&sd->poll_list))
7208 __raise_softirq_irqoff(NET_RX_SOFTIRQ);
7210 net_rps_action_and_irq_enable(sd);
7213 struct netdev_adjacent {
7214 struct net_device *dev;
7216 /* upper master flag, there can only be one master device per list */
7219 /* lookup ignore flag */
7222 /* counter for the number of times this device was added to us */
7225 /* private field for the users */
7228 struct list_head list;
7229 struct rcu_head rcu;
7232 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev,
7233 struct list_head *adj_list)
7235 struct netdev_adjacent *adj;
7237 list_for_each_entry(adj, adj_list, list) {
7238 if (adj->dev == adj_dev)
7244 static int ____netdev_has_upper_dev(struct net_device *upper_dev,
7245 struct netdev_nested_priv *priv)
7247 struct net_device *dev = (struct net_device *)priv->data;
7249 return upper_dev == dev;
7253 * netdev_has_upper_dev - Check if device is linked to an upper device
7255 * @upper_dev: upper device to check
7257 * Find out if a device is linked to specified upper device and return true
7258 * in case it is. Note that this checks only immediate upper device,
7259 * not through a complete stack of devices. The caller must hold the RTNL lock.
7261 bool netdev_has_upper_dev(struct net_device *dev,
7262 struct net_device *upper_dev)
7264 struct netdev_nested_priv priv = {
7265 .data = (void *)upper_dev,
7270 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev,
7273 EXPORT_SYMBOL(netdev_has_upper_dev);
7276 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device
7278 * @upper_dev: upper device to check
7280 * Find out if a device is linked to specified upper device and return true
7281 * in case it is. Note that this checks the entire upper device chain.
7282 * The caller must hold rcu lock.
7285 bool netdev_has_upper_dev_all_rcu(struct net_device *dev,
7286 struct net_device *upper_dev)
7288 struct netdev_nested_priv priv = {
7289 .data = (void *)upper_dev,
7292 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev,
7295 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu);
7298 * netdev_has_any_upper_dev - Check if device is linked to some device
7301 * Find out if a device is linked to an upper device and return true in case
7302 * it is. The caller must hold the RTNL lock.
7304 bool netdev_has_any_upper_dev(struct net_device *dev)
7308 return !list_empty(&dev->adj_list.upper);
7310 EXPORT_SYMBOL(netdev_has_any_upper_dev);
7313 * netdev_master_upper_dev_get - Get master upper device
7316 * Find a master upper device and return pointer to it or NULL in case
7317 * it's not there. The caller must hold the RTNL lock.
7319 struct net_device *netdev_master_upper_dev_get(struct net_device *dev)
7321 struct netdev_adjacent *upper;
7325 if (list_empty(&dev->adj_list.upper))
7328 upper = list_first_entry(&dev->adj_list.upper,
7329 struct netdev_adjacent, list);
7330 if (likely(upper->master))
7334 EXPORT_SYMBOL(netdev_master_upper_dev_get);
7336 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev)
7338 struct netdev_adjacent *upper;
7342 if (list_empty(&dev->adj_list.upper))
7345 upper = list_first_entry(&dev->adj_list.upper,
7346 struct netdev_adjacent, list);
7347 if (likely(upper->master) && !upper->ignore)
7353 * netdev_has_any_lower_dev - Check if device is linked to some device
7356 * Find out if a device is linked to a lower device and return true in case
7357 * it is. The caller must hold the RTNL lock.
7359 static bool netdev_has_any_lower_dev(struct net_device *dev)
7363 return !list_empty(&dev->adj_list.lower);
7366 void *netdev_adjacent_get_private(struct list_head *adj_list)
7368 struct netdev_adjacent *adj;
7370 adj = list_entry(adj_list, struct netdev_adjacent, list);
7372 return adj->private;
7374 EXPORT_SYMBOL(netdev_adjacent_get_private);
7377 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list
7379 * @iter: list_head ** of the current position
7381 * Gets the next device from the dev's upper list, starting from iter
7382 * position. The caller must hold RCU read lock.
7384 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev,
7385 struct list_head **iter)
7387 struct netdev_adjacent *upper;
7389 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
7391 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7393 if (&upper->list == &dev->adj_list.upper)
7396 *iter = &upper->list;
7400 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu);
7402 static struct net_device *__netdev_next_upper_dev(struct net_device *dev,
7403 struct list_head **iter,
7406 struct netdev_adjacent *upper;
7408 upper = list_entry((*iter)->next, struct netdev_adjacent, list);
7410 if (&upper->list == &dev->adj_list.upper)
7413 *iter = &upper->list;
7414 *ignore = upper->ignore;
7419 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev,
7420 struct list_head **iter)
7422 struct netdev_adjacent *upper;
7424 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held());
7426 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7428 if (&upper->list == &dev->adj_list.upper)
7431 *iter = &upper->list;
7436 static int __netdev_walk_all_upper_dev(struct net_device *dev,
7437 int (*fn)(struct net_device *dev,
7438 struct netdev_nested_priv *priv),
7439 struct netdev_nested_priv *priv)
7441 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7442 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7447 iter = &dev->adj_list.upper;
7451 ret = fn(now, priv);
7458 udev = __netdev_next_upper_dev(now, &iter, &ignore);
7465 niter = &udev->adj_list.upper;
7466 dev_stack[cur] = now;
7467 iter_stack[cur++] = iter;
7474 next = dev_stack[--cur];
7475 niter = iter_stack[cur];
7485 int netdev_walk_all_upper_dev_rcu(struct net_device *dev,
7486 int (*fn)(struct net_device *dev,
7487 struct netdev_nested_priv *priv),
7488 struct netdev_nested_priv *priv)
7490 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7491 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7495 iter = &dev->adj_list.upper;
7499 ret = fn(now, priv);
7506 udev = netdev_next_upper_dev_rcu(now, &iter);
7511 niter = &udev->adj_list.upper;
7512 dev_stack[cur] = now;
7513 iter_stack[cur++] = iter;
7520 next = dev_stack[--cur];
7521 niter = iter_stack[cur];
7530 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu);
7532 static bool __netdev_has_upper_dev(struct net_device *dev,
7533 struct net_device *upper_dev)
7535 struct netdev_nested_priv priv = {
7537 .data = (void *)upper_dev,
7542 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev,
7547 * netdev_lower_get_next_private - Get the next ->private from the
7548 * lower neighbour list
7550 * @iter: list_head ** of the current position
7552 * Gets the next netdev_adjacent->private from the dev's lower neighbour
7553 * list, starting from iter position. The caller must hold either hold the
7554 * RTNL lock or its own locking that guarantees that the neighbour lower
7555 * list will remain unchanged.
7557 void *netdev_lower_get_next_private(struct net_device *dev,
7558 struct list_head **iter)
7560 struct netdev_adjacent *lower;
7562 lower = list_entry(*iter, struct netdev_adjacent, list);
7564 if (&lower->list == &dev->adj_list.lower)
7567 *iter = lower->list.next;
7569 return lower->private;
7571 EXPORT_SYMBOL(netdev_lower_get_next_private);
7574 * netdev_lower_get_next_private_rcu - Get the next ->private from the
7575 * lower neighbour list, RCU
7578 * @iter: list_head ** of the current position
7580 * Gets the next netdev_adjacent->private from the dev's lower neighbour
7581 * list, starting from iter position. The caller must hold RCU read lock.
7583 void *netdev_lower_get_next_private_rcu(struct net_device *dev,
7584 struct list_head **iter)
7586 struct netdev_adjacent *lower;
7588 WARN_ON_ONCE(!rcu_read_lock_held());
7590 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7592 if (&lower->list == &dev->adj_list.lower)
7595 *iter = &lower->list;
7597 return lower->private;
7599 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu);
7602 * netdev_lower_get_next - Get the next device from the lower neighbour
7605 * @iter: list_head ** of the current position
7607 * Gets the next netdev_adjacent from the dev's lower neighbour
7608 * list, starting from iter position. The caller must hold RTNL lock or
7609 * its own locking that guarantees that the neighbour lower
7610 * list will remain unchanged.
7612 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter)
7614 struct netdev_adjacent *lower;
7616 lower = list_entry(*iter, struct netdev_adjacent, list);
7618 if (&lower->list == &dev->adj_list.lower)
7621 *iter = lower->list.next;
7625 EXPORT_SYMBOL(netdev_lower_get_next);
7627 static struct net_device *netdev_next_lower_dev(struct net_device *dev,
7628 struct list_head **iter)
7630 struct netdev_adjacent *lower;
7632 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
7634 if (&lower->list == &dev->adj_list.lower)
7637 *iter = &lower->list;
7642 static struct net_device *__netdev_next_lower_dev(struct net_device *dev,
7643 struct list_head **iter,
7646 struct netdev_adjacent *lower;
7648 lower = list_entry((*iter)->next, struct netdev_adjacent, list);
7650 if (&lower->list == &dev->adj_list.lower)
7653 *iter = &lower->list;
7654 *ignore = lower->ignore;
7659 int netdev_walk_all_lower_dev(struct net_device *dev,
7660 int (*fn)(struct net_device *dev,
7661 struct netdev_nested_priv *priv),
7662 struct netdev_nested_priv *priv)
7664 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7665 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7669 iter = &dev->adj_list.lower;
7673 ret = fn(now, priv);
7680 ldev = netdev_next_lower_dev(now, &iter);
7685 niter = &ldev->adj_list.lower;
7686 dev_stack[cur] = now;
7687 iter_stack[cur++] = iter;
7694 next = dev_stack[--cur];
7695 niter = iter_stack[cur];
7704 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev);
7706 static int __netdev_walk_all_lower_dev(struct net_device *dev,
7707 int (*fn)(struct net_device *dev,
7708 struct netdev_nested_priv *priv),
7709 struct netdev_nested_priv *priv)
7711 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7712 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7717 iter = &dev->adj_list.lower;
7721 ret = fn(now, priv);
7728 ldev = __netdev_next_lower_dev(now, &iter, &ignore);
7735 niter = &ldev->adj_list.lower;
7736 dev_stack[cur] = now;
7737 iter_stack[cur++] = iter;
7744 next = dev_stack[--cur];
7745 niter = iter_stack[cur];
7755 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev,
7756 struct list_head **iter)
7758 struct netdev_adjacent *lower;
7760 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list);
7761 if (&lower->list == &dev->adj_list.lower)
7764 *iter = &lower->list;
7768 EXPORT_SYMBOL(netdev_next_lower_dev_rcu);
7770 static u8 __netdev_upper_depth(struct net_device *dev)
7772 struct net_device *udev;
7773 struct list_head *iter;
7777 for (iter = &dev->adj_list.upper,
7778 udev = __netdev_next_upper_dev(dev, &iter, &ignore);
7780 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) {
7783 if (max_depth < udev->upper_level)
7784 max_depth = udev->upper_level;
7790 static u8 __netdev_lower_depth(struct net_device *dev)
7792 struct net_device *ldev;
7793 struct list_head *iter;
7797 for (iter = &dev->adj_list.lower,
7798 ldev = __netdev_next_lower_dev(dev, &iter, &ignore);
7800 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) {
7803 if (max_depth < ldev->lower_level)
7804 max_depth = ldev->lower_level;
7810 static int __netdev_update_upper_level(struct net_device *dev,
7811 struct netdev_nested_priv *__unused)
7813 dev->upper_level = __netdev_upper_depth(dev) + 1;
7817 static int __netdev_update_lower_level(struct net_device *dev,
7818 struct netdev_nested_priv *priv)
7820 dev->lower_level = __netdev_lower_depth(dev) + 1;
7822 #ifdef CONFIG_LOCKDEP
7826 if (priv->flags & NESTED_SYNC_IMM)
7827 dev->nested_level = dev->lower_level - 1;
7828 if (priv->flags & NESTED_SYNC_TODO)
7829 net_unlink_todo(dev);
7834 int netdev_walk_all_lower_dev_rcu(struct net_device *dev,
7835 int (*fn)(struct net_device *dev,
7836 struct netdev_nested_priv *priv),
7837 struct netdev_nested_priv *priv)
7839 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1];
7840 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1];
7844 iter = &dev->adj_list.lower;
7848 ret = fn(now, priv);
7855 ldev = netdev_next_lower_dev_rcu(now, &iter);
7860 niter = &ldev->adj_list.lower;
7861 dev_stack[cur] = now;
7862 iter_stack[cur++] = iter;
7869 next = dev_stack[--cur];
7870 niter = iter_stack[cur];
7879 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu);
7882 * netdev_lower_get_first_private_rcu - Get the first ->private from the
7883 * lower neighbour list, RCU
7887 * Gets the first netdev_adjacent->private from the dev's lower neighbour
7888 * list. The caller must hold RCU read lock.
7890 void *netdev_lower_get_first_private_rcu(struct net_device *dev)
7892 struct netdev_adjacent *lower;
7894 lower = list_first_or_null_rcu(&dev->adj_list.lower,
7895 struct netdev_adjacent, list);
7897 return lower->private;
7900 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu);
7903 * netdev_master_upper_dev_get_rcu - Get master upper device
7906 * Find a master upper device and return pointer to it or NULL in case
7907 * it's not there. The caller must hold the RCU read lock.
7909 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev)
7911 struct netdev_adjacent *upper;
7913 upper = list_first_or_null_rcu(&dev->adj_list.upper,
7914 struct netdev_adjacent, list);
7915 if (upper && likely(upper->master))
7919 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu);
7921 static int netdev_adjacent_sysfs_add(struct net_device *dev,
7922 struct net_device *adj_dev,
7923 struct list_head *dev_list)
7925 char linkname[IFNAMSIZ+7];
7927 sprintf(linkname, dev_list == &dev->adj_list.upper ?
7928 "upper_%s" : "lower_%s", adj_dev->name);
7929 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj),
7932 static void netdev_adjacent_sysfs_del(struct net_device *dev,
7934 struct list_head *dev_list)
7936 char linkname[IFNAMSIZ+7];
7938 sprintf(linkname, dev_list == &dev->adj_list.upper ?
7939 "upper_%s" : "lower_%s", name);
7940 sysfs_remove_link(&(dev->dev.kobj), linkname);
7943 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev,
7944 struct net_device *adj_dev,
7945 struct list_head *dev_list)
7947 return (dev_list == &dev->adj_list.upper ||
7948 dev_list == &dev->adj_list.lower) &&
7949 net_eq(dev_net(dev), dev_net(adj_dev));
7952 static int __netdev_adjacent_dev_insert(struct net_device *dev,
7953 struct net_device *adj_dev,
7954 struct list_head *dev_list,
7955 void *private, bool master)
7957 struct netdev_adjacent *adj;
7960 adj = __netdev_find_adj(adj_dev, dev_list);
7964 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n",
7965 dev->name, adj_dev->name, adj->ref_nr);
7970 adj = kmalloc(sizeof(*adj), GFP_KERNEL);
7975 adj->master = master;
7977 adj->private = private;
7978 adj->ignore = false;
7981 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n",
7982 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name);
7984 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) {
7985 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list);
7990 /* Ensure that master link is always the first item in list. */
7992 ret = sysfs_create_link(&(dev->dev.kobj),
7993 &(adj_dev->dev.kobj), "master");
7995 goto remove_symlinks;
7997 list_add_rcu(&adj->list, dev_list);
7999 list_add_tail_rcu(&adj->list, dev_list);
8005 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
8006 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
8014 static void __netdev_adjacent_dev_remove(struct net_device *dev,
8015 struct net_device *adj_dev,
8017 struct list_head *dev_list)
8019 struct netdev_adjacent *adj;
8021 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n",
8022 dev->name, adj_dev->name, ref_nr);
8024 adj = __netdev_find_adj(adj_dev, dev_list);
8027 pr_err("Adjacency does not exist for device %s from %s\n",
8028 dev->name, adj_dev->name);
8033 if (adj->ref_nr > ref_nr) {
8034 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n",
8035 dev->name, adj_dev->name, ref_nr,
8036 adj->ref_nr - ref_nr);
8037 adj->ref_nr -= ref_nr;
8042 sysfs_remove_link(&(dev->dev.kobj), "master");
8044 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list))
8045 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list);
8047 list_del_rcu(&adj->list);
8048 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n",
8049 adj_dev->name, dev->name, adj_dev->name);
8051 kfree_rcu(adj, rcu);
8054 static int __netdev_adjacent_dev_link_lists(struct net_device *dev,
8055 struct net_device *upper_dev,
8056 struct list_head *up_list,
8057 struct list_head *down_list,
8058 void *private, bool master)
8062 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list,
8067 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list,
8070 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list);
8077 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev,
8078 struct net_device *upper_dev,
8080 struct list_head *up_list,
8081 struct list_head *down_list)
8083 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list);
8084 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list);
8087 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev,
8088 struct net_device *upper_dev,
8089 void *private, bool master)
8091 return __netdev_adjacent_dev_link_lists(dev, upper_dev,
8092 &dev->adj_list.upper,
8093 &upper_dev->adj_list.lower,
8097 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev,
8098 struct net_device *upper_dev)
8100 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1,
8101 &dev->adj_list.upper,
8102 &upper_dev->adj_list.lower);
8105 static int __netdev_upper_dev_link(struct net_device *dev,
8106 struct net_device *upper_dev, bool master,
8107 void *upper_priv, void *upper_info,
8108 struct netdev_nested_priv *priv,
8109 struct netlink_ext_ack *extack)
8111 struct netdev_notifier_changeupper_info changeupper_info = {
8116 .upper_dev = upper_dev,
8119 .upper_info = upper_info,
8121 struct net_device *master_dev;
8126 if (dev == upper_dev)
8129 /* To prevent loops, check if dev is not upper device to upper_dev. */
8130 if (__netdev_has_upper_dev(upper_dev, dev))
8133 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV)
8137 if (__netdev_has_upper_dev(dev, upper_dev))
8140 master_dev = __netdev_master_upper_dev_get(dev);
8142 return master_dev == upper_dev ? -EEXIST : -EBUSY;
8145 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
8146 &changeupper_info.info);
8147 ret = notifier_to_errno(ret);
8151 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv,
8156 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
8157 &changeupper_info.info);
8158 ret = notifier_to_errno(ret);
8162 __netdev_update_upper_level(dev, NULL);
8163 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL);
8165 __netdev_update_lower_level(upper_dev, priv);
8166 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level,
8172 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
8178 * netdev_upper_dev_link - Add a link to the upper device
8180 * @upper_dev: new upper device
8181 * @extack: netlink extended ack
8183 * Adds a link to device which is upper to this one. The caller must hold
8184 * the RTNL lock. On a failure a negative errno code is returned.
8185 * On success the reference counts are adjusted and the function
8188 int netdev_upper_dev_link(struct net_device *dev,
8189 struct net_device *upper_dev,
8190 struct netlink_ext_ack *extack)
8192 struct netdev_nested_priv priv = {
8193 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO,
8197 return __netdev_upper_dev_link(dev, upper_dev, false,
8198 NULL, NULL, &priv, extack);
8200 EXPORT_SYMBOL(netdev_upper_dev_link);
8203 * netdev_master_upper_dev_link - Add a master link to the upper device
8205 * @upper_dev: new upper device
8206 * @upper_priv: upper device private
8207 * @upper_info: upper info to be passed down via notifier
8208 * @extack: netlink extended ack
8210 * Adds a link to device which is upper to this one. In this case, only
8211 * one master upper device can be linked, although other non-master devices
8212 * might be linked as well. The caller must hold the RTNL lock.
8213 * On a failure a negative errno code is returned. On success the reference
8214 * counts are adjusted and the function returns zero.
8216 int netdev_master_upper_dev_link(struct net_device *dev,
8217 struct net_device *upper_dev,
8218 void *upper_priv, void *upper_info,
8219 struct netlink_ext_ack *extack)
8221 struct netdev_nested_priv priv = {
8222 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO,
8226 return __netdev_upper_dev_link(dev, upper_dev, true,
8227 upper_priv, upper_info, &priv, extack);
8229 EXPORT_SYMBOL(netdev_master_upper_dev_link);
8231 static void __netdev_upper_dev_unlink(struct net_device *dev,
8232 struct net_device *upper_dev,
8233 struct netdev_nested_priv *priv)
8235 struct netdev_notifier_changeupper_info changeupper_info = {
8239 .upper_dev = upper_dev,
8245 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev;
8247 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER,
8248 &changeupper_info.info);
8250 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev);
8252 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER,
8253 &changeupper_info.info);
8255 __netdev_update_upper_level(dev, NULL);
8256 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL);
8258 __netdev_update_lower_level(upper_dev, priv);
8259 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level,
8264 * netdev_upper_dev_unlink - Removes a link to upper device
8266 * @upper_dev: new upper device
8268 * Removes a link to device which is upper to this one. The caller must hold
8271 void netdev_upper_dev_unlink(struct net_device *dev,
8272 struct net_device *upper_dev)
8274 struct netdev_nested_priv priv = {
8275 .flags = NESTED_SYNC_TODO,
8279 __netdev_upper_dev_unlink(dev, upper_dev, &priv);
8281 EXPORT_SYMBOL(netdev_upper_dev_unlink);
8283 static void __netdev_adjacent_dev_set(struct net_device *upper_dev,
8284 struct net_device *lower_dev,
8287 struct netdev_adjacent *adj;
8289 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower);
8293 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper);
8298 static void netdev_adjacent_dev_disable(struct net_device *upper_dev,
8299 struct net_device *lower_dev)
8301 __netdev_adjacent_dev_set(upper_dev, lower_dev, true);
8304 static void netdev_adjacent_dev_enable(struct net_device *upper_dev,
8305 struct net_device *lower_dev)
8307 __netdev_adjacent_dev_set(upper_dev, lower_dev, false);
8310 int netdev_adjacent_change_prepare(struct net_device *old_dev,
8311 struct net_device *new_dev,
8312 struct net_device *dev,
8313 struct netlink_ext_ack *extack)
8315 struct netdev_nested_priv priv = {
8324 if (old_dev && new_dev != old_dev)
8325 netdev_adjacent_dev_disable(dev, old_dev);
8326 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv,
8329 if (old_dev && new_dev != old_dev)
8330 netdev_adjacent_dev_enable(dev, old_dev);
8336 EXPORT_SYMBOL(netdev_adjacent_change_prepare);
8338 void netdev_adjacent_change_commit(struct net_device *old_dev,
8339 struct net_device *new_dev,
8340 struct net_device *dev)
8342 struct netdev_nested_priv priv = {
8343 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO,
8347 if (!new_dev || !old_dev)
8350 if (new_dev == old_dev)
8353 netdev_adjacent_dev_enable(dev, old_dev);
8354 __netdev_upper_dev_unlink(old_dev, dev, &priv);
8356 EXPORT_SYMBOL(netdev_adjacent_change_commit);
8358 void netdev_adjacent_change_abort(struct net_device *old_dev,
8359 struct net_device *new_dev,
8360 struct net_device *dev)
8362 struct netdev_nested_priv priv = {
8370 if (old_dev && new_dev != old_dev)
8371 netdev_adjacent_dev_enable(dev, old_dev);
8373 __netdev_upper_dev_unlink(new_dev, dev, &priv);
8375 EXPORT_SYMBOL(netdev_adjacent_change_abort);
8378 * netdev_bonding_info_change - Dispatch event about slave change
8380 * @bonding_info: info to dispatch
8382 * Send NETDEV_BONDING_INFO to netdev notifiers with info.
8383 * The caller must hold the RTNL lock.
8385 void netdev_bonding_info_change(struct net_device *dev,
8386 struct netdev_bonding_info *bonding_info)
8388 struct netdev_notifier_bonding_info info = {
8392 memcpy(&info.bonding_info, bonding_info,
8393 sizeof(struct netdev_bonding_info));
8394 call_netdevice_notifiers_info(NETDEV_BONDING_INFO,
8397 EXPORT_SYMBOL(netdev_bonding_info_change);
8400 * netdev_get_xmit_slave - Get the xmit slave of master device
8403 * @all_slaves: assume all the slaves are active
8405 * The reference counters are not incremented so the caller must be
8406 * careful with locks. The caller must hold RCU lock.
8407 * %NULL is returned if no slave is found.
8410 struct net_device *netdev_get_xmit_slave(struct net_device *dev,
8411 struct sk_buff *skb,
8414 const struct net_device_ops *ops = dev->netdev_ops;
8416 if (!ops->ndo_get_xmit_slave)
8418 return ops->ndo_get_xmit_slave(dev, skb, all_slaves);
8420 EXPORT_SYMBOL(netdev_get_xmit_slave);
8422 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev,
8425 const struct net_device_ops *ops = dev->netdev_ops;
8427 if (!ops->ndo_sk_get_lower_dev)
8429 return ops->ndo_sk_get_lower_dev(dev, sk);
8433 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket
8437 * %NULL is returned if no lower device is found.
8440 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev,
8443 struct net_device *lower;
8445 lower = netdev_sk_get_lower_dev(dev, sk);
8448 lower = netdev_sk_get_lower_dev(dev, sk);
8453 EXPORT_SYMBOL(netdev_sk_get_lowest_dev);
8455 static void netdev_adjacent_add_links(struct net_device *dev)
8457 struct netdev_adjacent *iter;
8459 struct net *net = dev_net(dev);
8461 list_for_each_entry(iter, &dev->adj_list.upper, list) {
8462 if (!net_eq(net, dev_net(iter->dev)))
8464 netdev_adjacent_sysfs_add(iter->dev, dev,
8465 &iter->dev->adj_list.lower);
8466 netdev_adjacent_sysfs_add(dev, iter->dev,
8467 &dev->adj_list.upper);
8470 list_for_each_entry(iter, &dev->adj_list.lower, list) {
8471 if (!net_eq(net, dev_net(iter->dev)))
8473 netdev_adjacent_sysfs_add(iter->dev, dev,
8474 &iter->dev->adj_list.upper);
8475 netdev_adjacent_sysfs_add(dev, iter->dev,
8476 &dev->adj_list.lower);
8480 static void netdev_adjacent_del_links(struct net_device *dev)
8482 struct netdev_adjacent *iter;
8484 struct net *net = dev_net(dev);
8486 list_for_each_entry(iter, &dev->adj_list.upper, list) {
8487 if (!net_eq(net, dev_net(iter->dev)))
8489 netdev_adjacent_sysfs_del(iter->dev, dev->name,
8490 &iter->dev->adj_list.lower);
8491 netdev_adjacent_sysfs_del(dev, iter->dev->name,
8492 &dev->adj_list.upper);
8495 list_for_each_entry(iter, &dev->adj_list.lower, list) {
8496 if (!net_eq(net, dev_net(iter->dev)))
8498 netdev_adjacent_sysfs_del(iter->dev, dev->name,
8499 &iter->dev->adj_list.upper);
8500 netdev_adjacent_sysfs_del(dev, iter->dev->name,
8501 &dev->adj_list.lower);
8505 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname)
8507 struct netdev_adjacent *iter;
8509 struct net *net = dev_net(dev);
8511 list_for_each_entry(iter, &dev->adj_list.upper, list) {
8512 if (!net_eq(net, dev_net(iter->dev)))
8514 netdev_adjacent_sysfs_del(iter->dev, oldname,
8515 &iter->dev->adj_list.lower);
8516 netdev_adjacent_sysfs_add(iter->dev, dev,
8517 &iter->dev->adj_list.lower);
8520 list_for_each_entry(iter, &dev->adj_list.lower, list) {
8521 if (!net_eq(net, dev_net(iter->dev)))
8523 netdev_adjacent_sysfs_del(iter->dev, oldname,
8524 &iter->dev->adj_list.upper);
8525 netdev_adjacent_sysfs_add(iter->dev, dev,
8526 &iter->dev->adj_list.upper);
8530 void *netdev_lower_dev_get_private(struct net_device *dev,
8531 struct net_device *lower_dev)
8533 struct netdev_adjacent *lower;
8537 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower);
8541 return lower->private;
8543 EXPORT_SYMBOL(netdev_lower_dev_get_private);
8547 * netdev_lower_state_changed - Dispatch event about lower device state change
8548 * @lower_dev: device
8549 * @lower_state_info: state to dispatch
8551 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info.
8552 * The caller must hold the RTNL lock.
8554 void netdev_lower_state_changed(struct net_device *lower_dev,
8555 void *lower_state_info)
8557 struct netdev_notifier_changelowerstate_info changelowerstate_info = {
8558 .info.dev = lower_dev,
8562 changelowerstate_info.lower_state_info = lower_state_info;
8563 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE,
8564 &changelowerstate_info.info);
8566 EXPORT_SYMBOL(netdev_lower_state_changed);
8568 static void dev_change_rx_flags(struct net_device *dev, int flags)
8570 const struct net_device_ops *ops = dev->netdev_ops;
8572 if (ops->ndo_change_rx_flags)
8573 ops->ndo_change_rx_flags(dev, flags);
8576 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify)
8578 unsigned int old_flags = dev->flags;
8584 dev->flags |= IFF_PROMISC;
8585 dev->promiscuity += inc;
8586 if (dev->promiscuity == 0) {
8589 * If inc causes overflow, untouch promisc and return error.
8592 dev->flags &= ~IFF_PROMISC;
8594 dev->promiscuity -= inc;
8595 pr_warn("%s: promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n",
8600 if (dev->flags != old_flags) {
8601 pr_info("device %s %s promiscuous mode\n",
8603 dev->flags & IFF_PROMISC ? "entered" : "left");
8604 if (audit_enabled) {
8605 current_uid_gid(&uid, &gid);
8606 audit_log(audit_context(), GFP_ATOMIC,
8607 AUDIT_ANOM_PROMISCUOUS,
8608 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u",
8609 dev->name, (dev->flags & IFF_PROMISC),
8610 (old_flags & IFF_PROMISC),
8611 from_kuid(&init_user_ns, audit_get_loginuid(current)),
8612 from_kuid(&init_user_ns, uid),
8613 from_kgid(&init_user_ns, gid),
8614 audit_get_sessionid(current));
8617 dev_change_rx_flags(dev, IFF_PROMISC);
8620 __dev_notify_flags(dev, old_flags, IFF_PROMISC);
8625 * dev_set_promiscuity - update promiscuity count on a device
8629 * Add or remove promiscuity from a device. While the count in the device
8630 * remains above zero the interface remains promiscuous. Once it hits zero
8631 * the device reverts back to normal filtering operation. A negative inc
8632 * value is used to drop promiscuity on the device.
8633 * Return 0 if successful or a negative errno code on error.
8635 int dev_set_promiscuity(struct net_device *dev, int inc)
8637 unsigned int old_flags = dev->flags;
8640 err = __dev_set_promiscuity(dev, inc, true);
8643 if (dev->flags != old_flags)
8644 dev_set_rx_mode(dev);
8647 EXPORT_SYMBOL(dev_set_promiscuity);
8649 static int __dev_set_allmulti(struct net_device *dev, int inc, bool notify)
8651 unsigned int old_flags = dev->flags, old_gflags = dev->gflags;
8655 dev->flags |= IFF_ALLMULTI;
8656 dev->allmulti += inc;
8657 if (dev->allmulti == 0) {
8660 * If inc causes overflow, untouch allmulti and return error.
8663 dev->flags &= ~IFF_ALLMULTI;
8665 dev->allmulti -= inc;
8666 pr_warn("%s: allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n",
8671 if (dev->flags ^ old_flags) {
8672 dev_change_rx_flags(dev, IFF_ALLMULTI);
8673 dev_set_rx_mode(dev);
8675 __dev_notify_flags(dev, old_flags,
8676 dev->gflags ^ old_gflags);
8682 * dev_set_allmulti - update allmulti count on a device
8686 * Add or remove reception of all multicast frames to a device. While the
8687 * count in the device remains above zero the interface remains listening
8688 * to all interfaces. Once it hits zero the device reverts back to normal
8689 * filtering operation. A negative @inc value is used to drop the counter
8690 * when releasing a resource needing all multicasts.
8691 * Return 0 if successful or a negative errno code on error.
8694 int dev_set_allmulti(struct net_device *dev, int inc)
8696 return __dev_set_allmulti(dev, inc, true);
8698 EXPORT_SYMBOL(dev_set_allmulti);
8701 * Upload unicast and multicast address lists to device and
8702 * configure RX filtering. When the device doesn't support unicast
8703 * filtering it is put in promiscuous mode while unicast addresses
8706 void __dev_set_rx_mode(struct net_device *dev)
8708 const struct net_device_ops *ops = dev->netdev_ops;
8710 /* dev_open will call this function so the list will stay sane. */
8711 if (!(dev->flags&IFF_UP))
8714 if (!netif_device_present(dev))
8717 if (!(dev->priv_flags & IFF_UNICAST_FLT)) {
8718 /* Unicast addresses changes may only happen under the rtnl,
8719 * therefore calling __dev_set_promiscuity here is safe.
8721 if (!netdev_uc_empty(dev) && !dev->uc_promisc) {
8722 __dev_set_promiscuity(dev, 1, false);
8723 dev->uc_promisc = true;
8724 } else if (netdev_uc_empty(dev) && dev->uc_promisc) {
8725 __dev_set_promiscuity(dev, -1, false);
8726 dev->uc_promisc = false;
8730 if (ops->ndo_set_rx_mode)
8731 ops->ndo_set_rx_mode(dev);
8734 void dev_set_rx_mode(struct net_device *dev)
8736 netif_addr_lock_bh(dev);
8737 __dev_set_rx_mode(dev);
8738 netif_addr_unlock_bh(dev);
8742 * dev_get_flags - get flags reported to userspace
8745 * Get the combination of flag bits exported through APIs to userspace.
8747 unsigned int dev_get_flags(const struct net_device *dev)
8751 flags = (dev->flags & ~(IFF_PROMISC |
8756 (dev->gflags & (IFF_PROMISC |
8759 if (netif_running(dev)) {
8760 if (netif_oper_up(dev))
8761 flags |= IFF_RUNNING;
8762 if (netif_carrier_ok(dev))
8763 flags |= IFF_LOWER_UP;
8764 if (netif_dormant(dev))
8765 flags |= IFF_DORMANT;
8770 EXPORT_SYMBOL(dev_get_flags);
8772 int __dev_change_flags(struct net_device *dev, unsigned int flags,
8773 struct netlink_ext_ack *extack)
8775 unsigned int old_flags = dev->flags;
8781 * Set the flags on our device.
8784 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP |
8785 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL |
8787 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC |
8791 * Load in the correct multicast list now the flags have changed.
8794 if ((old_flags ^ flags) & IFF_MULTICAST)
8795 dev_change_rx_flags(dev, IFF_MULTICAST);
8797 dev_set_rx_mode(dev);
8800 * Have we downed the interface. We handle IFF_UP ourselves
8801 * according to user attempts to set it, rather than blindly
8806 if ((old_flags ^ flags) & IFF_UP) {
8807 if (old_flags & IFF_UP)
8810 ret = __dev_open(dev, extack);
8813 if ((flags ^ dev->gflags) & IFF_PROMISC) {
8814 int inc = (flags & IFF_PROMISC) ? 1 : -1;
8815 unsigned int old_flags = dev->flags;
8817 dev->gflags ^= IFF_PROMISC;
8819 if (__dev_set_promiscuity(dev, inc, false) >= 0)
8820 if (dev->flags != old_flags)
8821 dev_set_rx_mode(dev);
8824 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI
8825 * is important. Some (broken) drivers set IFF_PROMISC, when
8826 * IFF_ALLMULTI is requested not asking us and not reporting.
8828 if ((flags ^ dev->gflags) & IFF_ALLMULTI) {
8829 int inc = (flags & IFF_ALLMULTI) ? 1 : -1;
8831 dev->gflags ^= IFF_ALLMULTI;
8832 __dev_set_allmulti(dev, inc, false);
8838 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags,
8839 unsigned int gchanges)
8841 unsigned int changes = dev->flags ^ old_flags;
8844 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC);
8846 if (changes & IFF_UP) {
8847 if (dev->flags & IFF_UP)
8848 call_netdevice_notifiers(NETDEV_UP, dev);
8850 call_netdevice_notifiers(NETDEV_DOWN, dev);
8853 if (dev->flags & IFF_UP &&
8854 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) {
8855 struct netdev_notifier_change_info change_info = {
8859 .flags_changed = changes,
8862 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info);
8867 * dev_change_flags - change device settings
8869 * @flags: device state flags
8870 * @extack: netlink extended ack
8872 * Change settings on device based state flags. The flags are
8873 * in the userspace exported format.
8875 int dev_change_flags(struct net_device *dev, unsigned int flags,
8876 struct netlink_ext_ack *extack)
8879 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags;
8881 ret = __dev_change_flags(dev, flags, extack);
8885 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags);
8886 __dev_notify_flags(dev, old_flags, changes);
8889 EXPORT_SYMBOL(dev_change_flags);
8891 int __dev_set_mtu(struct net_device *dev, int new_mtu)
8893 const struct net_device_ops *ops = dev->netdev_ops;
8895 if (ops->ndo_change_mtu)
8896 return ops->ndo_change_mtu(dev, new_mtu);
8898 /* Pairs with all the lockless reads of dev->mtu in the stack */
8899 WRITE_ONCE(dev->mtu, new_mtu);
8902 EXPORT_SYMBOL(__dev_set_mtu);
8904 int dev_validate_mtu(struct net_device *dev, int new_mtu,
8905 struct netlink_ext_ack *extack)
8907 /* MTU must be positive, and in range */
8908 if (new_mtu < 0 || new_mtu < dev->min_mtu) {
8909 NL_SET_ERR_MSG(extack, "mtu less than device minimum");
8913 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) {
8914 NL_SET_ERR_MSG(extack, "mtu greater than device maximum");
8921 * dev_set_mtu_ext - Change maximum transfer unit
8923 * @new_mtu: new transfer unit
8924 * @extack: netlink extended ack
8926 * Change the maximum transfer size of the network device.
8928 int dev_set_mtu_ext(struct net_device *dev, int new_mtu,
8929 struct netlink_ext_ack *extack)
8933 if (new_mtu == dev->mtu)
8936 err = dev_validate_mtu(dev, new_mtu, extack);
8940 if (!netif_device_present(dev))
8943 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev);
8944 err = notifier_to_errno(err);
8948 orig_mtu = dev->mtu;
8949 err = __dev_set_mtu(dev, new_mtu);
8952 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev,
8954 err = notifier_to_errno(err);
8956 /* setting mtu back and notifying everyone again,
8957 * so that they have a chance to revert changes.
8959 __dev_set_mtu(dev, orig_mtu);
8960 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev,
8967 int dev_set_mtu(struct net_device *dev, int new_mtu)
8969 struct netlink_ext_ack extack;
8972 memset(&extack, 0, sizeof(extack));
8973 err = dev_set_mtu_ext(dev, new_mtu, &extack);
8974 if (err && extack._msg)
8975 net_err_ratelimited("%s: %s\n", dev->name, extack._msg);
8978 EXPORT_SYMBOL(dev_set_mtu);
8981 * dev_change_tx_queue_len - Change TX queue length of a netdevice
8983 * @new_len: new tx queue length
8985 int dev_change_tx_queue_len(struct net_device *dev, unsigned long new_len)
8987 unsigned int orig_len = dev->tx_queue_len;
8990 if (new_len != (unsigned int)new_len)
8993 if (new_len != orig_len) {
8994 dev->tx_queue_len = new_len;
8995 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev);
8996 res = notifier_to_errno(res);
8999 res = dev_qdisc_change_tx_queue_len(dev);
9007 netdev_err(dev, "refused to change device tx_queue_len\n");
9008 dev->tx_queue_len = orig_len;
9013 * dev_set_group - Change group this device belongs to
9015 * @new_group: group this device should belong to
9017 void dev_set_group(struct net_device *dev, int new_group)
9019 dev->group = new_group;
9021 EXPORT_SYMBOL(dev_set_group);
9024 * dev_pre_changeaddr_notify - Call NETDEV_PRE_CHANGEADDR.
9026 * @addr: new address
9027 * @extack: netlink extended ack
9029 int dev_pre_changeaddr_notify(struct net_device *dev, const char *addr,
9030 struct netlink_ext_ack *extack)
9032 struct netdev_notifier_pre_changeaddr_info info = {
9034 .info.extack = extack,
9039 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info);
9040 return notifier_to_errno(rc);
9042 EXPORT_SYMBOL(dev_pre_changeaddr_notify);
9045 * dev_set_mac_address - Change Media Access Control Address
9048 * @extack: netlink extended ack
9050 * Change the hardware (MAC) address of the device
9052 int dev_set_mac_address(struct net_device *dev, struct sockaddr *sa,
9053 struct netlink_ext_ack *extack)
9055 const struct net_device_ops *ops = dev->netdev_ops;
9058 if (!ops->ndo_set_mac_address)
9060 if (sa->sa_family != dev->type)
9062 if (!netif_device_present(dev))
9064 err = dev_pre_changeaddr_notify(dev, sa->sa_data, extack);
9067 err = ops->ndo_set_mac_address(dev, sa);
9070 dev->addr_assign_type = NET_ADDR_SET;
9071 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev);
9072 add_device_randomness(dev->dev_addr, dev->addr_len);
9075 EXPORT_SYMBOL(dev_set_mac_address);
9077 static DECLARE_RWSEM(dev_addr_sem);
9079 int dev_set_mac_address_user(struct net_device *dev, struct sockaddr *sa,
9080 struct netlink_ext_ack *extack)
9084 down_write(&dev_addr_sem);
9085 ret = dev_set_mac_address(dev, sa, extack);
9086 up_write(&dev_addr_sem);
9089 EXPORT_SYMBOL(dev_set_mac_address_user);
9091 int dev_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name)
9093 size_t size = sizeof(sa->sa_data);
9094 struct net_device *dev;
9097 down_read(&dev_addr_sem);
9100 dev = dev_get_by_name_rcu(net, dev_name);
9106 memset(sa->sa_data, 0, size);
9108 memcpy(sa->sa_data, dev->dev_addr,
9109 min_t(size_t, size, dev->addr_len));
9110 sa->sa_family = dev->type;
9114 up_read(&dev_addr_sem);
9117 EXPORT_SYMBOL(dev_get_mac_address);
9120 * dev_change_carrier - Change device carrier
9122 * @new_carrier: new value
9124 * Change device carrier
9126 int dev_change_carrier(struct net_device *dev, bool new_carrier)
9128 const struct net_device_ops *ops = dev->netdev_ops;
9130 if (!ops->ndo_change_carrier)
9132 if (!netif_device_present(dev))
9134 return ops->ndo_change_carrier(dev, new_carrier);
9136 EXPORT_SYMBOL(dev_change_carrier);
9139 * dev_get_phys_port_id - Get device physical port ID
9143 * Get device physical port ID
9145 int dev_get_phys_port_id(struct net_device *dev,
9146 struct netdev_phys_item_id *ppid)
9148 const struct net_device_ops *ops = dev->netdev_ops;
9150 if (!ops->ndo_get_phys_port_id)
9152 return ops->ndo_get_phys_port_id(dev, ppid);
9154 EXPORT_SYMBOL(dev_get_phys_port_id);
9157 * dev_get_phys_port_name - Get device physical port name
9160 * @len: limit of bytes to copy to name
9162 * Get device physical port name
9164 int dev_get_phys_port_name(struct net_device *dev,
9165 char *name, size_t len)
9167 const struct net_device_ops *ops = dev->netdev_ops;
9170 if (ops->ndo_get_phys_port_name) {
9171 err = ops->ndo_get_phys_port_name(dev, name, len);
9172 if (err != -EOPNOTSUPP)
9175 return devlink_compat_phys_port_name_get(dev, name, len);
9177 EXPORT_SYMBOL(dev_get_phys_port_name);
9180 * dev_get_port_parent_id - Get the device's port parent identifier
9181 * @dev: network device
9182 * @ppid: pointer to a storage for the port's parent identifier
9183 * @recurse: allow/disallow recursion to lower devices
9185 * Get the devices's port parent identifier
9187 int dev_get_port_parent_id(struct net_device *dev,
9188 struct netdev_phys_item_id *ppid,
9191 const struct net_device_ops *ops = dev->netdev_ops;
9192 struct netdev_phys_item_id first = { };
9193 struct net_device *lower_dev;
9194 struct list_head *iter;
9197 if (ops->ndo_get_port_parent_id) {
9198 err = ops->ndo_get_port_parent_id(dev, ppid);
9199 if (err != -EOPNOTSUPP)
9203 err = devlink_compat_switch_id_get(dev, ppid);
9204 if (!err || err != -EOPNOTSUPP)
9210 netdev_for_each_lower_dev(dev, lower_dev, iter) {
9211 err = dev_get_port_parent_id(lower_dev, ppid, recurse);
9216 else if (memcmp(&first, ppid, sizeof(*ppid)))
9222 EXPORT_SYMBOL(dev_get_port_parent_id);
9225 * netdev_port_same_parent_id - Indicate if two network devices have
9226 * the same port parent identifier
9227 * @a: first network device
9228 * @b: second network device
9230 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b)
9232 struct netdev_phys_item_id a_id = { };
9233 struct netdev_phys_item_id b_id = { };
9235 if (dev_get_port_parent_id(a, &a_id, true) ||
9236 dev_get_port_parent_id(b, &b_id, true))
9239 return netdev_phys_item_id_same(&a_id, &b_id);
9241 EXPORT_SYMBOL(netdev_port_same_parent_id);
9244 * dev_change_proto_down - update protocol port state information
9246 * @proto_down: new value
9248 * This info can be used by switch drivers to set the phys state of the
9251 int dev_change_proto_down(struct net_device *dev, bool proto_down)
9253 const struct net_device_ops *ops = dev->netdev_ops;
9255 if (!ops->ndo_change_proto_down)
9257 if (!netif_device_present(dev))
9259 return ops->ndo_change_proto_down(dev, proto_down);
9261 EXPORT_SYMBOL(dev_change_proto_down);
9264 * dev_change_proto_down_generic - generic implementation for
9265 * ndo_change_proto_down that sets carrier according to
9269 * @proto_down: new value
9271 int dev_change_proto_down_generic(struct net_device *dev, bool proto_down)
9274 netif_carrier_off(dev);
9276 netif_carrier_on(dev);
9277 dev->proto_down = proto_down;
9280 EXPORT_SYMBOL(dev_change_proto_down_generic);
9283 * dev_change_proto_down_reason - proto down reason
9286 * @mask: proto down mask
9287 * @value: proto down value
9289 void dev_change_proto_down_reason(struct net_device *dev, unsigned long mask,
9295 dev->proto_down_reason = value;
9297 for_each_set_bit(b, &mask, 32) {
9298 if (value & (1 << b))
9299 dev->proto_down_reason |= BIT(b);
9301 dev->proto_down_reason &= ~BIT(b);
9305 EXPORT_SYMBOL(dev_change_proto_down_reason);
9307 struct bpf_xdp_link {
9308 struct bpf_link link;
9309 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */
9313 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags)
9315 if (flags & XDP_FLAGS_HW_MODE)
9317 if (flags & XDP_FLAGS_DRV_MODE)
9318 return XDP_MODE_DRV;
9319 if (flags & XDP_FLAGS_SKB_MODE)
9320 return XDP_MODE_SKB;
9321 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB;
9324 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode)
9328 return generic_xdp_install;
9331 return dev->netdev_ops->ndo_bpf;
9337 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev,
9338 enum bpf_xdp_mode mode)
9340 return dev->xdp_state[mode].link;
9343 static struct bpf_prog *dev_xdp_prog(struct net_device *dev,
9344 enum bpf_xdp_mode mode)
9346 struct bpf_xdp_link *link = dev_xdp_link(dev, mode);
9349 return link->link.prog;
9350 return dev->xdp_state[mode].prog;
9353 static u8 dev_xdp_prog_count(struct net_device *dev)
9358 for (i = 0; i < __MAX_XDP_MODE; i++)
9359 if (dev->xdp_state[i].prog || dev->xdp_state[i].link)
9364 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode)
9366 struct bpf_prog *prog = dev_xdp_prog(dev, mode);
9368 return prog ? prog->aux->id : 0;
9371 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode,
9372 struct bpf_xdp_link *link)
9374 dev->xdp_state[mode].link = link;
9375 dev->xdp_state[mode].prog = NULL;
9378 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode,
9379 struct bpf_prog *prog)
9381 dev->xdp_state[mode].link = NULL;
9382 dev->xdp_state[mode].prog = prog;
9385 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode,
9386 bpf_op_t bpf_op, struct netlink_ext_ack *extack,
9387 u32 flags, struct bpf_prog *prog)
9389 struct netdev_bpf xdp;
9392 memset(&xdp, 0, sizeof(xdp));
9393 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG;
9394 xdp.extack = extack;
9398 /* Drivers assume refcnt is already incremented (i.e, prog pointer is
9399 * "moved" into driver), so they don't increment it on their own, but
9400 * they do decrement refcnt when program is detached or replaced.
9401 * Given net_device also owns link/prog, we need to bump refcnt here
9402 * to prevent drivers from underflowing it.
9406 err = bpf_op(dev, &xdp);
9413 if (mode != XDP_MODE_HW)
9414 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog);
9419 static void dev_xdp_uninstall(struct net_device *dev)
9421 struct bpf_xdp_link *link;
9422 struct bpf_prog *prog;
9423 enum bpf_xdp_mode mode;
9428 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) {
9429 prog = dev_xdp_prog(dev, mode);
9433 bpf_op = dev_xdp_bpf_op(dev, mode);
9437 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL));
9439 /* auto-detach link from net device */
9440 link = dev_xdp_link(dev, mode);
9446 dev_xdp_set_link(dev, mode, NULL);
9450 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack,
9451 struct bpf_xdp_link *link, struct bpf_prog *new_prog,
9452 struct bpf_prog *old_prog, u32 flags)
9454 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES);
9455 struct bpf_prog *cur_prog;
9456 enum bpf_xdp_mode mode;
9462 /* either link or prog attachment, never both */
9463 if (link && (new_prog || old_prog))
9465 /* link supports only XDP mode flags */
9466 if (link && (flags & ~XDP_FLAGS_MODES)) {
9467 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment");
9470 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */
9471 if (num_modes > 1) {
9472 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set");
9475 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */
9476 if (!num_modes && dev_xdp_prog_count(dev) > 1) {
9477 NL_SET_ERR_MSG(extack,
9478 "More than one program loaded, unset mode is ambiguous");
9481 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */
9482 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) {
9483 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified");
9487 mode = dev_xdp_mode(dev, flags);
9488 /* can't replace attached link */
9489 if (dev_xdp_link(dev, mode)) {
9490 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link");
9494 cur_prog = dev_xdp_prog(dev, mode);
9495 /* can't replace attached prog with link */
9496 if (link && cur_prog) {
9497 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link");
9500 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) {
9501 NL_SET_ERR_MSG(extack, "Active program does not match expected");
9505 /* put effective new program into new_prog */
9507 new_prog = link->link.prog;
9510 bool offload = mode == XDP_MODE_HW;
9511 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB
9512 ? XDP_MODE_DRV : XDP_MODE_SKB;
9514 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) {
9515 NL_SET_ERR_MSG(extack, "XDP program already attached");
9518 if (!offload && dev_xdp_prog(dev, other_mode)) {
9519 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time");
9522 if (!offload && bpf_prog_is_dev_bound(new_prog->aux)) {
9523 NL_SET_ERR_MSG(extack, "Using device-bound program without HW_MODE flag is not supported");
9526 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) {
9527 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device");
9530 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) {
9531 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device");
9536 /* don't call drivers if the effective program didn't change */
9537 if (new_prog != cur_prog) {
9538 bpf_op = dev_xdp_bpf_op(dev, mode);
9540 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode");
9544 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog);
9550 dev_xdp_set_link(dev, mode, link);
9552 dev_xdp_set_prog(dev, mode, new_prog);
9554 bpf_prog_put(cur_prog);
9559 static int dev_xdp_attach_link(struct net_device *dev,
9560 struct netlink_ext_ack *extack,
9561 struct bpf_xdp_link *link)
9563 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags);
9566 static int dev_xdp_detach_link(struct net_device *dev,
9567 struct netlink_ext_ack *extack,
9568 struct bpf_xdp_link *link)
9570 enum bpf_xdp_mode mode;
9575 mode = dev_xdp_mode(dev, link->flags);
9576 if (dev_xdp_link(dev, mode) != link)
9579 bpf_op = dev_xdp_bpf_op(dev, mode);
9580 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL));
9581 dev_xdp_set_link(dev, mode, NULL);
9585 static void bpf_xdp_link_release(struct bpf_link *link)
9587 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9591 /* if racing with net_device's tear down, xdp_link->dev might be
9592 * already NULL, in which case link was already auto-detached
9594 if (xdp_link->dev) {
9595 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link));
9596 xdp_link->dev = NULL;
9602 static int bpf_xdp_link_detach(struct bpf_link *link)
9604 bpf_xdp_link_release(link);
9608 static void bpf_xdp_link_dealloc(struct bpf_link *link)
9610 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9615 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link,
9616 struct seq_file *seq)
9618 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9623 ifindex = xdp_link->dev->ifindex;
9626 seq_printf(seq, "ifindex:\t%u\n", ifindex);
9629 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link,
9630 struct bpf_link_info *info)
9632 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9637 ifindex = xdp_link->dev->ifindex;
9640 info->xdp.ifindex = ifindex;
9644 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog,
9645 struct bpf_prog *old_prog)
9647 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link);
9648 enum bpf_xdp_mode mode;
9654 /* link might have been auto-released already, so fail */
9655 if (!xdp_link->dev) {
9660 if (old_prog && link->prog != old_prog) {
9664 old_prog = link->prog;
9665 if (old_prog == new_prog) {
9666 /* no-op, don't disturb drivers */
9667 bpf_prog_put(new_prog);
9671 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags);
9672 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode);
9673 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL,
9674 xdp_link->flags, new_prog);
9678 old_prog = xchg(&link->prog, new_prog);
9679 bpf_prog_put(old_prog);
9686 static const struct bpf_link_ops bpf_xdp_link_lops = {
9687 .release = bpf_xdp_link_release,
9688 .dealloc = bpf_xdp_link_dealloc,
9689 .detach = bpf_xdp_link_detach,
9690 .show_fdinfo = bpf_xdp_link_show_fdinfo,
9691 .fill_link_info = bpf_xdp_link_fill_link_info,
9692 .update_prog = bpf_xdp_link_update,
9695 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog)
9697 struct net *net = current->nsproxy->net_ns;
9698 struct bpf_link_primer link_primer;
9699 struct bpf_xdp_link *link;
9700 struct net_device *dev;
9703 dev = dev_get_by_index(net, attr->link_create.target_ifindex);
9707 link = kzalloc(sizeof(*link), GFP_USER);
9713 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog);
9715 link->flags = attr->link_create.flags;
9717 err = bpf_link_prime(&link->link, &link_primer);
9724 err = dev_xdp_attach_link(dev, NULL, link);
9728 bpf_link_cleanup(&link_primer);
9732 fd = bpf_link_settle(&link_primer);
9733 /* link itself doesn't hold dev's refcnt to not complicate shutdown */
9743 * dev_change_xdp_fd - set or clear a bpf program for a device rx path
9745 * @extack: netlink extended ack
9746 * @fd: new program fd or negative value to clear
9747 * @expected_fd: old program fd that userspace expects to replace or clear
9748 * @flags: xdp-related flags
9750 * Set or clear a bpf program for a device
9752 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack,
9753 int fd, int expected_fd, u32 flags)
9755 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags);
9756 struct bpf_prog *new_prog = NULL, *old_prog = NULL;
9762 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP,
9763 mode != XDP_MODE_SKB);
9764 if (IS_ERR(new_prog))
9765 return PTR_ERR(new_prog);
9768 if (expected_fd >= 0) {
9769 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP,
9770 mode != XDP_MODE_SKB);
9771 if (IS_ERR(old_prog)) {
9772 err = PTR_ERR(old_prog);
9778 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags);
9781 if (err && new_prog)
9782 bpf_prog_put(new_prog);
9784 bpf_prog_put(old_prog);
9789 * dev_new_index - allocate an ifindex
9790 * @net: the applicable net namespace
9792 * Returns a suitable unique value for a new device interface
9793 * number. The caller must hold the rtnl semaphore or the
9794 * dev_base_lock to be sure it remains unique.
9796 static int dev_new_index(struct net *net)
9798 int ifindex = net->ifindex;
9803 if (!__dev_get_by_index(net, ifindex))
9804 return net->ifindex = ifindex;
9808 /* Delayed registration/unregisteration */
9809 static LIST_HEAD(net_todo_list);
9810 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq);
9812 static void net_set_todo(struct net_device *dev)
9814 list_add_tail(&dev->todo_list, &net_todo_list);
9815 dev_net(dev)->dev_unreg_count++;
9818 static netdev_features_t netdev_sync_upper_features(struct net_device *lower,
9819 struct net_device *upper, netdev_features_t features)
9821 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
9822 netdev_features_t feature;
9825 for_each_netdev_feature(upper_disables, feature_bit) {
9826 feature = __NETIF_F_BIT(feature_bit);
9827 if (!(upper->wanted_features & feature)
9828 && (features & feature)) {
9829 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n",
9830 &feature, upper->name);
9831 features &= ~feature;
9838 static void netdev_sync_lower_features(struct net_device *upper,
9839 struct net_device *lower, netdev_features_t features)
9841 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES;
9842 netdev_features_t feature;
9845 for_each_netdev_feature(upper_disables, feature_bit) {
9846 feature = __NETIF_F_BIT(feature_bit);
9847 if (!(features & feature) && (lower->features & feature)) {
9848 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n",
9849 &feature, lower->name);
9850 lower->wanted_features &= ~feature;
9851 __netdev_update_features(lower);
9853 if (unlikely(lower->features & feature))
9854 netdev_WARN(upper, "failed to disable %pNF on %s!\n",
9855 &feature, lower->name);
9857 netdev_features_change(lower);
9862 static netdev_features_t netdev_fix_features(struct net_device *dev,
9863 netdev_features_t features)
9865 /* Fix illegal checksum combinations */
9866 if ((features & NETIF_F_HW_CSUM) &&
9867 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) {
9868 netdev_warn(dev, "mixed HW and IP checksum settings.\n");
9869 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM);
9872 /* TSO requires that SG is present as well. */
9873 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) {
9874 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n");
9875 features &= ~NETIF_F_ALL_TSO;
9878 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) &&
9879 !(features & NETIF_F_IP_CSUM)) {
9880 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n");
9881 features &= ~NETIF_F_TSO;
9882 features &= ~NETIF_F_TSO_ECN;
9885 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) &&
9886 !(features & NETIF_F_IPV6_CSUM)) {
9887 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n");
9888 features &= ~NETIF_F_TSO6;
9891 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */
9892 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO))
9893 features &= ~NETIF_F_TSO_MANGLEID;
9895 /* TSO ECN requires that TSO is present as well. */
9896 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN)
9897 features &= ~NETIF_F_TSO_ECN;
9899 /* Software GSO depends on SG. */
9900 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) {
9901 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n");
9902 features &= ~NETIF_F_GSO;
9905 /* GSO partial features require GSO partial be set */
9906 if ((features & dev->gso_partial_features) &&
9907 !(features & NETIF_F_GSO_PARTIAL)) {
9909 "Dropping partially supported GSO features since no GSO partial.\n");
9910 features &= ~dev->gso_partial_features;
9913 if (!(features & NETIF_F_RXCSUM)) {
9914 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet
9915 * successfully merged by hardware must also have the
9916 * checksum verified by hardware. If the user does not
9917 * want to enable RXCSUM, logically, we should disable GRO_HW.
9919 if (features & NETIF_F_GRO_HW) {
9920 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n");
9921 features &= ~NETIF_F_GRO_HW;
9925 /* LRO/HW-GRO features cannot be combined with RX-FCS */
9926 if (features & NETIF_F_RXFCS) {
9927 if (features & NETIF_F_LRO) {
9928 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n");
9929 features &= ~NETIF_F_LRO;
9932 if (features & NETIF_F_GRO_HW) {
9933 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n");
9934 features &= ~NETIF_F_GRO_HW;
9938 if (features & NETIF_F_HW_TLS_TX) {
9939 bool ip_csum = (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) ==
9940 (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM);
9941 bool hw_csum = features & NETIF_F_HW_CSUM;
9943 if (!ip_csum && !hw_csum) {
9944 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n");
9945 features &= ~NETIF_F_HW_TLS_TX;
9949 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) {
9950 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n");
9951 features &= ~NETIF_F_HW_TLS_RX;
9957 int __netdev_update_features(struct net_device *dev)
9959 struct net_device *upper, *lower;
9960 netdev_features_t features;
9961 struct list_head *iter;
9966 features = netdev_get_wanted_features(dev);
9968 if (dev->netdev_ops->ndo_fix_features)
9969 features = dev->netdev_ops->ndo_fix_features(dev, features);
9971 /* driver might be less strict about feature dependencies */
9972 features = netdev_fix_features(dev, features);
9974 /* some features can't be enabled if they're off on an upper device */
9975 netdev_for_each_upper_dev_rcu(dev, upper, iter)
9976 features = netdev_sync_upper_features(dev, upper, features);
9978 if (dev->features == features)
9981 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n",
9982 &dev->features, &features);
9984 if (dev->netdev_ops->ndo_set_features)
9985 err = dev->netdev_ops->ndo_set_features(dev, features);
9989 if (unlikely(err < 0)) {
9991 "set_features() failed (%d); wanted %pNF, left %pNF\n",
9992 err, &features, &dev->features);
9993 /* return non-0 since some features might have changed and
9994 * it's better to fire a spurious notification than miss it
10000 /* some features must be disabled on lower devices when disabled
10001 * on an upper device (think: bonding master or bridge)
10003 netdev_for_each_lower_dev(dev, lower, iter)
10004 netdev_sync_lower_features(dev, lower, features);
10007 netdev_features_t diff = features ^ dev->features;
10009 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) {
10010 /* udp_tunnel_{get,drop}_rx_info both need
10011 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the
10012 * device, or they won't do anything.
10013 * Thus we need to update dev->features
10014 * *before* calling udp_tunnel_get_rx_info,
10015 * but *after* calling udp_tunnel_drop_rx_info.
10017 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) {
10018 dev->features = features;
10019 udp_tunnel_get_rx_info(dev);
10021 udp_tunnel_drop_rx_info(dev);
10025 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) {
10026 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) {
10027 dev->features = features;
10028 err |= vlan_get_rx_ctag_filter_info(dev);
10030 vlan_drop_rx_ctag_filter_info(dev);
10034 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) {
10035 if (features & NETIF_F_HW_VLAN_STAG_FILTER) {
10036 dev->features = features;
10037 err |= vlan_get_rx_stag_filter_info(dev);
10039 vlan_drop_rx_stag_filter_info(dev);
10043 dev->features = features;
10046 return err < 0 ? 0 : 1;
10050 * netdev_update_features - recalculate device features
10051 * @dev: the device to check
10053 * Recalculate dev->features set and send notifications if it
10054 * has changed. Should be called after driver or hardware dependent
10055 * conditions might have changed that influence the features.
10057 void netdev_update_features(struct net_device *dev)
10059 if (__netdev_update_features(dev))
10060 netdev_features_change(dev);
10062 EXPORT_SYMBOL(netdev_update_features);
10065 * netdev_change_features - recalculate device features
10066 * @dev: the device to check
10068 * Recalculate dev->features set and send notifications even
10069 * if they have not changed. Should be called instead of
10070 * netdev_update_features() if also dev->vlan_features might
10071 * have changed to allow the changes to be propagated to stacked
10074 void netdev_change_features(struct net_device *dev)
10076 __netdev_update_features(dev);
10077 netdev_features_change(dev);
10079 EXPORT_SYMBOL(netdev_change_features);
10082 * netif_stacked_transfer_operstate - transfer operstate
10083 * @rootdev: the root or lower level device to transfer state from
10084 * @dev: the device to transfer operstate to
10086 * Transfer operational state from root to device. This is normally
10087 * called when a stacking relationship exists between the root
10088 * device and the device(a leaf device).
10090 void netif_stacked_transfer_operstate(const struct net_device *rootdev,
10091 struct net_device *dev)
10093 if (rootdev->operstate == IF_OPER_DORMANT)
10094 netif_dormant_on(dev);
10096 netif_dormant_off(dev);
10098 if (rootdev->operstate == IF_OPER_TESTING)
10099 netif_testing_on(dev);
10101 netif_testing_off(dev);
10103 if (netif_carrier_ok(rootdev))
10104 netif_carrier_on(dev);
10106 netif_carrier_off(dev);
10108 EXPORT_SYMBOL(netif_stacked_transfer_operstate);
10110 static int netif_alloc_rx_queues(struct net_device *dev)
10112 unsigned int i, count = dev->num_rx_queues;
10113 struct netdev_rx_queue *rx;
10114 size_t sz = count * sizeof(*rx);
10119 rx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
10125 for (i = 0; i < count; i++) {
10128 /* XDP RX-queue setup */
10129 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0);
10136 /* Rollback successful reg's and free other resources */
10138 xdp_rxq_info_unreg(&rx[i].xdp_rxq);
10144 static void netif_free_rx_queues(struct net_device *dev)
10146 unsigned int i, count = dev->num_rx_queues;
10148 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */
10152 for (i = 0; i < count; i++)
10153 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq);
10158 static void netdev_init_one_queue(struct net_device *dev,
10159 struct netdev_queue *queue, void *_unused)
10161 /* Initialize queue lock */
10162 spin_lock_init(&queue->_xmit_lock);
10163 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type);
10164 queue->xmit_lock_owner = -1;
10165 netdev_queue_numa_node_write(queue, NUMA_NO_NODE);
10168 dql_init(&queue->dql, HZ);
10172 static void netif_free_tx_queues(struct net_device *dev)
10177 static int netif_alloc_netdev_queues(struct net_device *dev)
10179 unsigned int count = dev->num_tx_queues;
10180 struct netdev_queue *tx;
10181 size_t sz = count * sizeof(*tx);
10183 if (count < 1 || count > 0xffff)
10186 tx = kvzalloc(sz, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
10192 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL);
10193 spin_lock_init(&dev->tx_global_lock);
10198 void netif_tx_stop_all_queues(struct net_device *dev)
10202 for (i = 0; i < dev->num_tx_queues; i++) {
10203 struct netdev_queue *txq = netdev_get_tx_queue(dev, i);
10205 netif_tx_stop_queue(txq);
10208 EXPORT_SYMBOL(netif_tx_stop_all_queues);
10211 * register_netdevice - register a network device
10212 * @dev: device to register
10214 * Take a completed network device structure and add it to the kernel
10215 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
10216 * chain. 0 is returned on success. A negative errno code is returned
10217 * on a failure to set up the device, or if the name is a duplicate.
10219 * Callers must hold the rtnl semaphore. You may want
10220 * register_netdev() instead of this.
10223 * The locking appears insufficient to guarantee two parallel registers
10224 * will not get the same name.
10227 int register_netdevice(struct net_device *dev)
10230 struct net *net = dev_net(dev);
10232 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE <
10233 NETDEV_FEATURE_COUNT);
10234 BUG_ON(dev_boot_phase);
10239 /* When net_device's are persistent, this will be fatal. */
10240 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED);
10243 ret = ethtool_check_ops(dev->ethtool_ops);
10247 spin_lock_init(&dev->addr_list_lock);
10248 netdev_set_addr_lockdep_class(dev);
10250 ret = dev_get_valid_name(net, dev, dev->name);
10255 dev->name_node = netdev_name_node_head_alloc(dev);
10256 if (!dev->name_node)
10259 /* Init, if this function is available */
10260 if (dev->netdev_ops->ndo_init) {
10261 ret = dev->netdev_ops->ndo_init(dev);
10265 goto err_free_name;
10269 if (((dev->hw_features | dev->features) &
10270 NETIF_F_HW_VLAN_CTAG_FILTER) &&
10271 (!dev->netdev_ops->ndo_vlan_rx_add_vid ||
10272 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) {
10273 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n");
10280 dev->ifindex = dev_new_index(net);
10281 else if (__dev_get_by_index(net, dev->ifindex))
10284 /* Transfer changeable features to wanted_features and enable
10285 * software offloads (GSO and GRO).
10287 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF);
10288 dev->features |= NETIF_F_SOFT_FEATURES;
10290 if (dev->udp_tunnel_nic_info) {
10291 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT;
10292 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT;
10295 dev->wanted_features = dev->features & dev->hw_features;
10297 if (!(dev->flags & IFF_LOOPBACK))
10298 dev->hw_features |= NETIF_F_NOCACHE_COPY;
10300 /* If IPv4 TCP segmentation offload is supported we should also
10301 * allow the device to enable segmenting the frame with the option
10302 * of ignoring a static IP ID value. This doesn't enable the
10303 * feature itself but allows the user to enable it later.
10305 if (dev->hw_features & NETIF_F_TSO)
10306 dev->hw_features |= NETIF_F_TSO_MANGLEID;
10307 if (dev->vlan_features & NETIF_F_TSO)
10308 dev->vlan_features |= NETIF_F_TSO_MANGLEID;
10309 if (dev->mpls_features & NETIF_F_TSO)
10310 dev->mpls_features |= NETIF_F_TSO_MANGLEID;
10311 if (dev->hw_enc_features & NETIF_F_TSO)
10312 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID;
10314 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices.
10316 dev->vlan_features |= NETIF_F_HIGHDMA;
10318 /* Make NETIF_F_SG inheritable to tunnel devices.
10320 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL;
10322 /* Make NETIF_F_SG inheritable to MPLS.
10324 dev->mpls_features |= NETIF_F_SG;
10326 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev);
10327 ret = notifier_to_errno(ret);
10331 ret = netdev_register_kobject(dev);
10333 dev->reg_state = NETREG_UNREGISTERED;
10336 dev->reg_state = NETREG_REGISTERED;
10338 __netdev_update_features(dev);
10341 * Default initial state at registry is that the
10342 * device is present.
10345 set_bit(__LINK_STATE_PRESENT, &dev->state);
10347 linkwatch_init_dev(dev);
10349 dev_init_scheduler(dev);
10351 list_netdevice(dev);
10352 add_device_randomness(dev->dev_addr, dev->addr_len);
10354 /* If the device has permanent device address, driver should
10355 * set dev_addr and also addr_assign_type should be set to
10356 * NET_ADDR_PERM (default value).
10358 if (dev->addr_assign_type == NET_ADDR_PERM)
10359 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len);
10361 /* Notify protocols, that a new device appeared. */
10362 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev);
10363 ret = notifier_to_errno(ret);
10365 /* Expect explicit free_netdev() on failure */
10366 dev->needs_free_netdev = false;
10367 unregister_netdevice_queue(dev, NULL);
10371 * Prevent userspace races by waiting until the network
10372 * device is fully setup before sending notifications.
10374 if (!dev->rtnl_link_ops ||
10375 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
10376 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
10382 if (dev->netdev_ops->ndo_uninit)
10383 dev->netdev_ops->ndo_uninit(dev);
10384 if (dev->priv_destructor)
10385 dev->priv_destructor(dev);
10387 netdev_name_node_free(dev->name_node);
10390 EXPORT_SYMBOL(register_netdevice);
10393 * init_dummy_netdev - init a dummy network device for NAPI
10394 * @dev: device to init
10396 * This takes a network device structure and initialize the minimum
10397 * amount of fields so it can be used to schedule NAPI polls without
10398 * registering a full blown interface. This is to be used by drivers
10399 * that need to tie several hardware interfaces to a single NAPI
10400 * poll scheduler due to HW limitations.
10402 int init_dummy_netdev(struct net_device *dev)
10404 /* Clear everything. Note we don't initialize spinlocks
10405 * are they aren't supposed to be taken by any of the
10406 * NAPI code and this dummy netdev is supposed to be
10407 * only ever used for NAPI polls
10409 memset(dev, 0, sizeof(struct net_device));
10411 /* make sure we BUG if trying to hit standard
10412 * register/unregister code path
10414 dev->reg_state = NETREG_DUMMY;
10416 /* NAPI wants this */
10417 INIT_LIST_HEAD(&dev->napi_list);
10419 /* a dummy interface is started by default */
10420 set_bit(__LINK_STATE_PRESENT, &dev->state);
10421 set_bit(__LINK_STATE_START, &dev->state);
10423 /* napi_busy_loop stats accounting wants this */
10424 dev_net_set(dev, &init_net);
10426 /* Note : We dont allocate pcpu_refcnt for dummy devices,
10427 * because users of this 'device' dont need to change
10433 EXPORT_SYMBOL_GPL(init_dummy_netdev);
10437 * register_netdev - register a network device
10438 * @dev: device to register
10440 * Take a completed network device structure and add it to the kernel
10441 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier
10442 * chain. 0 is returned on success. A negative errno code is returned
10443 * on a failure to set up the device, or if the name is a duplicate.
10445 * This is a wrapper around register_netdevice that takes the rtnl semaphore
10446 * and expands the device name if you passed a format string to
10449 int register_netdev(struct net_device *dev)
10453 if (rtnl_lock_killable())
10455 err = register_netdevice(dev);
10459 EXPORT_SYMBOL(register_netdev);
10461 int netdev_refcnt_read(const struct net_device *dev)
10463 #ifdef CONFIG_PCPU_DEV_REFCNT
10466 for_each_possible_cpu(i)
10467 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i);
10470 return refcount_read(&dev->dev_refcnt);
10473 EXPORT_SYMBOL(netdev_refcnt_read);
10475 int netdev_unregister_timeout_secs __read_mostly = 10;
10477 #define WAIT_REFS_MIN_MSECS 1
10478 #define WAIT_REFS_MAX_MSECS 250
10480 * netdev_wait_allrefs - wait until all references are gone.
10481 * @dev: target net_device
10483 * This is called when unregistering network devices.
10485 * Any protocol or device that holds a reference should register
10486 * for netdevice notification, and cleanup and put back the
10487 * reference if they receive an UNREGISTER event.
10488 * We can get stuck here if buggy protocols don't correctly
10491 static void netdev_wait_allrefs(struct net_device *dev)
10493 unsigned long rebroadcast_time, warning_time;
10494 int wait = 0, refcnt;
10496 linkwatch_forget_dev(dev);
10498 rebroadcast_time = warning_time = jiffies;
10499 refcnt = netdev_refcnt_read(dev);
10501 while (refcnt != 1) {
10502 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) {
10505 /* Rebroadcast unregister notification */
10506 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
10512 if (test_bit(__LINK_STATE_LINKWATCH_PENDING,
10514 /* We must not have linkwatch events
10515 * pending on unregister. If this
10516 * happens, we simply run the queue
10517 * unscheduled, resulting in a noop
10520 linkwatch_run_queue();
10525 rebroadcast_time = jiffies;
10530 wait = WAIT_REFS_MIN_MSECS;
10533 wait = min(wait << 1, WAIT_REFS_MAX_MSECS);
10536 refcnt = netdev_refcnt_read(dev);
10539 time_after(jiffies, warning_time +
10540 netdev_unregister_timeout_secs * HZ)) {
10541 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n",
10542 dev->name, refcnt);
10543 warning_time = jiffies;
10548 /* The sequence is:
10552 * register_netdevice(x1);
10553 * register_netdevice(x2);
10555 * unregister_netdevice(y1);
10556 * unregister_netdevice(y2);
10562 * We are invoked by rtnl_unlock().
10563 * This allows us to deal with problems:
10564 * 1) We can delete sysfs objects which invoke hotplug
10565 * without deadlocking with linkwatch via keventd.
10566 * 2) Since we run with the RTNL semaphore not held, we can sleep
10567 * safely in order to wait for the netdev refcnt to drop to zero.
10569 * We must not return until all unregister events added during
10570 * the interval the lock was held have been completed.
10572 void netdev_run_todo(void)
10574 struct list_head list;
10575 #ifdef CONFIG_LOCKDEP
10576 struct list_head unlink_list;
10578 list_replace_init(&net_unlink_list, &unlink_list);
10580 while (!list_empty(&unlink_list)) {
10581 struct net_device *dev = list_first_entry(&unlink_list,
10584 list_del_init(&dev->unlink_list);
10585 dev->nested_level = dev->lower_level - 1;
10589 /* Snapshot list, allow later requests */
10590 list_replace_init(&net_todo_list, &list);
10595 /* Wait for rcu callbacks to finish before next phase */
10596 if (!list_empty(&list))
10599 while (!list_empty(&list)) {
10600 struct net_device *dev
10601 = list_first_entry(&list, struct net_device, todo_list);
10602 list_del(&dev->todo_list);
10604 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) {
10605 pr_err("network todo '%s' but state %d\n",
10606 dev->name, dev->reg_state);
10611 dev->reg_state = NETREG_UNREGISTERED;
10613 netdev_wait_allrefs(dev);
10616 BUG_ON(netdev_refcnt_read(dev) != 1);
10617 BUG_ON(!list_empty(&dev->ptype_all));
10618 BUG_ON(!list_empty(&dev->ptype_specific));
10619 WARN_ON(rcu_access_pointer(dev->ip_ptr));
10620 WARN_ON(rcu_access_pointer(dev->ip6_ptr));
10621 #if IS_ENABLED(CONFIG_DECNET)
10622 WARN_ON(dev->dn_ptr);
10624 if (dev->priv_destructor)
10625 dev->priv_destructor(dev);
10626 if (dev->needs_free_netdev)
10629 /* Report a network device has been unregistered */
10631 dev_net(dev)->dev_unreg_count--;
10633 wake_up(&netdev_unregistering_wq);
10635 /* Free network device */
10636 kobject_put(&dev->dev.kobj);
10640 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has
10641 * all the same fields in the same order as net_device_stats, with only
10642 * the type differing, but rtnl_link_stats64 may have additional fields
10643 * at the end for newer counters.
10645 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64,
10646 const struct net_device_stats *netdev_stats)
10648 #if BITS_PER_LONG == 64
10649 BUILD_BUG_ON(sizeof(*stats64) < sizeof(*netdev_stats));
10650 memcpy(stats64, netdev_stats, sizeof(*netdev_stats));
10651 /* zero out counters that only exist in rtnl_link_stats64 */
10652 memset((char *)stats64 + sizeof(*netdev_stats), 0,
10653 sizeof(*stats64) - sizeof(*netdev_stats));
10655 size_t i, n = sizeof(*netdev_stats) / sizeof(unsigned long);
10656 const unsigned long *src = (const unsigned long *)netdev_stats;
10657 u64 *dst = (u64 *)stats64;
10659 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64));
10660 for (i = 0; i < n; i++)
10662 /* zero out counters that only exist in rtnl_link_stats64 */
10663 memset((char *)stats64 + n * sizeof(u64), 0,
10664 sizeof(*stats64) - n * sizeof(u64));
10667 EXPORT_SYMBOL(netdev_stats_to_stats64);
10670 * dev_get_stats - get network device statistics
10671 * @dev: device to get statistics from
10672 * @storage: place to store stats
10674 * Get network statistics from device. Return @storage.
10675 * The device driver may provide its own method by setting
10676 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats;
10677 * otherwise the internal statistics structure is used.
10679 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev,
10680 struct rtnl_link_stats64 *storage)
10682 const struct net_device_ops *ops = dev->netdev_ops;
10684 if (ops->ndo_get_stats64) {
10685 memset(storage, 0, sizeof(*storage));
10686 ops->ndo_get_stats64(dev, storage);
10687 } else if (ops->ndo_get_stats) {
10688 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev));
10690 netdev_stats_to_stats64(storage, &dev->stats);
10692 storage->rx_dropped += (unsigned long)atomic_long_read(&dev->rx_dropped);
10693 storage->tx_dropped += (unsigned long)atomic_long_read(&dev->tx_dropped);
10694 storage->rx_nohandler += (unsigned long)atomic_long_read(&dev->rx_nohandler);
10697 EXPORT_SYMBOL(dev_get_stats);
10700 * dev_fetch_sw_netstats - get per-cpu network device statistics
10701 * @s: place to store stats
10702 * @netstats: per-cpu network stats to read from
10704 * Read per-cpu network statistics and populate the related fields in @s.
10706 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s,
10707 const struct pcpu_sw_netstats __percpu *netstats)
10711 for_each_possible_cpu(cpu) {
10712 const struct pcpu_sw_netstats *stats;
10713 struct pcpu_sw_netstats tmp;
10714 unsigned int start;
10716 stats = per_cpu_ptr(netstats, cpu);
10718 start = u64_stats_fetch_begin_irq(&stats->syncp);
10719 tmp.rx_packets = stats->rx_packets;
10720 tmp.rx_bytes = stats->rx_bytes;
10721 tmp.tx_packets = stats->tx_packets;
10722 tmp.tx_bytes = stats->tx_bytes;
10723 } while (u64_stats_fetch_retry_irq(&stats->syncp, start));
10725 s->rx_packets += tmp.rx_packets;
10726 s->rx_bytes += tmp.rx_bytes;
10727 s->tx_packets += tmp.tx_packets;
10728 s->tx_bytes += tmp.tx_bytes;
10731 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats);
10734 * dev_get_tstats64 - ndo_get_stats64 implementation
10735 * @dev: device to get statistics from
10736 * @s: place to store stats
10738 * Populate @s from dev->stats and dev->tstats. Can be used as
10739 * ndo_get_stats64() callback.
10741 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s)
10743 netdev_stats_to_stats64(s, &dev->stats);
10744 dev_fetch_sw_netstats(s, dev->tstats);
10746 EXPORT_SYMBOL_GPL(dev_get_tstats64);
10748 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev)
10750 struct netdev_queue *queue = dev_ingress_queue(dev);
10752 #ifdef CONFIG_NET_CLS_ACT
10755 queue = kzalloc(sizeof(*queue), GFP_KERNEL);
10758 netdev_init_one_queue(dev, queue, NULL);
10759 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc);
10760 queue->qdisc_sleeping = &noop_qdisc;
10761 rcu_assign_pointer(dev->ingress_queue, queue);
10766 static const struct ethtool_ops default_ethtool_ops;
10768 void netdev_set_default_ethtool_ops(struct net_device *dev,
10769 const struct ethtool_ops *ops)
10771 if (dev->ethtool_ops == &default_ethtool_ops)
10772 dev->ethtool_ops = ops;
10774 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops);
10776 void netdev_freemem(struct net_device *dev)
10778 char *addr = (char *)dev - dev->padded;
10784 * alloc_netdev_mqs - allocate network device
10785 * @sizeof_priv: size of private data to allocate space for
10786 * @name: device name format string
10787 * @name_assign_type: origin of device name
10788 * @setup: callback to initialize device
10789 * @txqs: the number of TX subqueues to allocate
10790 * @rxqs: the number of RX subqueues to allocate
10792 * Allocates a struct net_device with private data area for driver use
10793 * and performs basic initialization. Also allocates subqueue structs
10794 * for each queue on the device.
10796 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name,
10797 unsigned char name_assign_type,
10798 void (*setup)(struct net_device *),
10799 unsigned int txqs, unsigned int rxqs)
10801 struct net_device *dev;
10802 unsigned int alloc_size;
10803 struct net_device *p;
10805 BUG_ON(strlen(name) >= sizeof(dev->name));
10808 pr_err("alloc_netdev: Unable to allocate device with zero queues\n");
10813 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n");
10817 alloc_size = sizeof(struct net_device);
10819 /* ensure 32-byte alignment of private area */
10820 alloc_size = ALIGN(alloc_size, NETDEV_ALIGN);
10821 alloc_size += sizeof_priv;
10823 /* ensure 32-byte alignment of whole construct */
10824 alloc_size += NETDEV_ALIGN - 1;
10826 p = kvzalloc(alloc_size, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
10830 dev = PTR_ALIGN(p, NETDEV_ALIGN);
10831 dev->padded = (char *)dev - (char *)p;
10833 #ifdef CONFIG_PCPU_DEV_REFCNT
10834 dev->pcpu_refcnt = alloc_percpu(int);
10835 if (!dev->pcpu_refcnt)
10839 refcount_set(&dev->dev_refcnt, 1);
10842 if (dev_addr_init(dev))
10848 dev_net_set(dev, &init_net);
10850 dev->gso_max_size = GSO_MAX_SIZE;
10851 dev->gso_max_segs = GSO_MAX_SEGS;
10852 dev->upper_level = 1;
10853 dev->lower_level = 1;
10854 #ifdef CONFIG_LOCKDEP
10855 dev->nested_level = 0;
10856 INIT_LIST_HEAD(&dev->unlink_list);
10859 INIT_LIST_HEAD(&dev->napi_list);
10860 INIT_LIST_HEAD(&dev->unreg_list);
10861 INIT_LIST_HEAD(&dev->close_list);
10862 INIT_LIST_HEAD(&dev->link_watch_list);
10863 INIT_LIST_HEAD(&dev->adj_list.upper);
10864 INIT_LIST_HEAD(&dev->adj_list.lower);
10865 INIT_LIST_HEAD(&dev->ptype_all);
10866 INIT_LIST_HEAD(&dev->ptype_specific);
10867 INIT_LIST_HEAD(&dev->net_notifier_list);
10868 #ifdef CONFIG_NET_SCHED
10869 hash_init(dev->qdisc_hash);
10871 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM;
10874 if (!dev->tx_queue_len) {
10875 dev->priv_flags |= IFF_NO_QUEUE;
10876 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN;
10879 dev->num_tx_queues = txqs;
10880 dev->real_num_tx_queues = txqs;
10881 if (netif_alloc_netdev_queues(dev))
10884 dev->num_rx_queues = rxqs;
10885 dev->real_num_rx_queues = rxqs;
10886 if (netif_alloc_rx_queues(dev))
10889 strcpy(dev->name, name);
10890 dev->name_assign_type = name_assign_type;
10891 dev->group = INIT_NETDEV_GROUP;
10892 if (!dev->ethtool_ops)
10893 dev->ethtool_ops = &default_ethtool_ops;
10895 nf_hook_ingress_init(dev);
10904 #ifdef CONFIG_PCPU_DEV_REFCNT
10905 free_percpu(dev->pcpu_refcnt);
10908 netdev_freemem(dev);
10911 EXPORT_SYMBOL(alloc_netdev_mqs);
10914 * free_netdev - free network device
10917 * This function does the last stage of destroying an allocated device
10918 * interface. The reference to the device object is released. If this
10919 * is the last reference then it will be freed.Must be called in process
10922 void free_netdev(struct net_device *dev)
10924 struct napi_struct *p, *n;
10928 /* When called immediately after register_netdevice() failed the unwind
10929 * handling may still be dismantling the device. Handle that case by
10930 * deferring the free.
10932 if (dev->reg_state == NETREG_UNREGISTERING) {
10934 dev->needs_free_netdev = true;
10938 netif_free_tx_queues(dev);
10939 netif_free_rx_queues(dev);
10941 kfree(rcu_dereference_protected(dev->ingress_queue, 1));
10943 /* Flush device addresses */
10944 dev_addr_flush(dev);
10946 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list)
10949 #ifdef CONFIG_PCPU_DEV_REFCNT
10950 free_percpu(dev->pcpu_refcnt);
10951 dev->pcpu_refcnt = NULL;
10953 free_percpu(dev->xdp_bulkq);
10954 dev->xdp_bulkq = NULL;
10956 /* Compatibility with error handling in drivers */
10957 if (dev->reg_state == NETREG_UNINITIALIZED) {
10958 netdev_freemem(dev);
10962 BUG_ON(dev->reg_state != NETREG_UNREGISTERED);
10963 dev->reg_state = NETREG_RELEASED;
10965 /* will free via device release */
10966 put_device(&dev->dev);
10968 EXPORT_SYMBOL(free_netdev);
10971 * synchronize_net - Synchronize with packet receive processing
10973 * Wait for packets currently being received to be done.
10974 * Does not block later packets from starting.
10976 void synchronize_net(void)
10979 if (rtnl_is_locked())
10980 synchronize_rcu_expedited();
10984 EXPORT_SYMBOL(synchronize_net);
10987 * unregister_netdevice_queue - remove device from the kernel
10991 * This function shuts down a device interface and removes it
10992 * from the kernel tables.
10993 * If head not NULL, device is queued to be unregistered later.
10995 * Callers must hold the rtnl semaphore. You may want
10996 * unregister_netdev() instead of this.
10999 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head)
11004 list_move_tail(&dev->unreg_list, head);
11008 list_add(&dev->unreg_list, &single);
11009 unregister_netdevice_many(&single);
11012 EXPORT_SYMBOL(unregister_netdevice_queue);
11015 * unregister_netdevice_many - unregister many devices
11016 * @head: list of devices
11018 * Note: As most callers use a stack allocated list_head,
11019 * we force a list_del() to make sure stack wont be corrupted later.
11021 void unregister_netdevice_many(struct list_head *head)
11023 struct net_device *dev, *tmp;
11024 LIST_HEAD(close_head);
11026 BUG_ON(dev_boot_phase);
11029 if (list_empty(head))
11032 list_for_each_entry_safe(dev, tmp, head, unreg_list) {
11033 /* Some devices call without registering
11034 * for initialization unwind. Remove those
11035 * devices and proceed with the remaining.
11037 if (dev->reg_state == NETREG_UNINITIALIZED) {
11038 pr_debug("unregister_netdevice: device %s/%p never was registered\n",
11042 list_del(&dev->unreg_list);
11045 dev->dismantle = true;
11046 BUG_ON(dev->reg_state != NETREG_REGISTERED);
11049 /* If device is running, close it first. */
11050 list_for_each_entry(dev, head, unreg_list)
11051 list_add_tail(&dev->close_list, &close_head);
11052 dev_close_many(&close_head, true);
11054 list_for_each_entry(dev, head, unreg_list) {
11055 /* And unlink it from device chain. */
11056 unlist_netdevice(dev);
11058 dev->reg_state = NETREG_UNREGISTERING;
11060 flush_all_backlogs();
11064 list_for_each_entry(dev, head, unreg_list) {
11065 struct sk_buff *skb = NULL;
11067 /* Shutdown queueing discipline. */
11070 dev_xdp_uninstall(dev);
11072 /* Notify protocols, that we are about to destroy
11073 * this device. They should clean all the things.
11075 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
11077 if (!dev->rtnl_link_ops ||
11078 dev->rtnl_link_state == RTNL_LINK_INITIALIZED)
11079 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0,
11080 GFP_KERNEL, NULL, 0);
11083 * Flush the unicast and multicast chains
11088 netdev_name_node_alt_flush(dev);
11089 netdev_name_node_free(dev->name_node);
11091 if (dev->netdev_ops->ndo_uninit)
11092 dev->netdev_ops->ndo_uninit(dev);
11095 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL);
11097 /* Notifier chain MUST detach us all upper devices. */
11098 WARN_ON(netdev_has_any_upper_dev(dev));
11099 WARN_ON(netdev_has_any_lower_dev(dev));
11101 /* Remove entries from kobject tree */
11102 netdev_unregister_kobject(dev);
11104 /* Remove XPS queueing entries */
11105 netif_reset_xps_queues_gt(dev, 0);
11111 list_for_each_entry(dev, head, unreg_list) {
11118 EXPORT_SYMBOL(unregister_netdevice_many);
11121 * unregister_netdev - remove device from the kernel
11124 * This function shuts down a device interface and removes it
11125 * from the kernel tables.
11127 * This is just a wrapper for unregister_netdevice that takes
11128 * the rtnl semaphore. In general you want to use this and not
11129 * unregister_netdevice.
11131 void unregister_netdev(struct net_device *dev)
11134 unregister_netdevice(dev);
11137 EXPORT_SYMBOL(unregister_netdev);
11140 * __dev_change_net_namespace - move device to different nethost namespace
11142 * @net: network namespace
11143 * @pat: If not NULL name pattern to try if the current device name
11144 * is already taken in the destination network namespace.
11145 * @new_ifindex: If not zero, specifies device index in the target
11148 * This function shuts down a device interface and moves it
11149 * to a new network namespace. On success 0 is returned, on
11150 * a failure a netagive errno code is returned.
11152 * Callers must hold the rtnl semaphore.
11155 int __dev_change_net_namespace(struct net_device *dev, struct net *net,
11156 const char *pat, int new_ifindex)
11158 struct net *net_old = dev_net(dev);
11163 /* Don't allow namespace local devices to be moved. */
11165 if (dev->features & NETIF_F_NETNS_LOCAL)
11168 /* Ensure the device has been registrered */
11169 if (dev->reg_state != NETREG_REGISTERED)
11172 /* Get out if there is nothing todo */
11174 if (net_eq(net_old, net))
11177 /* Pick the destination device name, and ensure
11178 * we can use it in the destination network namespace.
11181 if (__dev_get_by_name(net, dev->name)) {
11182 /* We get here if we can't use the current device name */
11185 err = dev_get_valid_name(net, dev, pat);
11190 /* Check that new_ifindex isn't used yet. */
11192 if (new_ifindex && __dev_get_by_index(net, new_ifindex))
11196 * And now a mini version of register_netdevice unregister_netdevice.
11199 /* If device is running close it first. */
11202 /* And unlink it from device chain */
11203 unlist_netdevice(dev);
11207 /* Shutdown queueing discipline. */
11210 /* Notify protocols, that we are about to destroy
11211 * this device. They should clean all the things.
11213 * Note that dev->reg_state stays at NETREG_REGISTERED.
11214 * This is wanted because this way 8021q and macvlan know
11215 * the device is just moving and can keep their slaves up.
11217 call_netdevice_notifiers(NETDEV_UNREGISTER, dev);
11220 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL);
11221 /* If there is an ifindex conflict assign a new one */
11222 if (!new_ifindex) {
11223 if (__dev_get_by_index(net, dev->ifindex))
11224 new_ifindex = dev_new_index(net);
11226 new_ifindex = dev->ifindex;
11229 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid,
11233 * Flush the unicast and multicast chains
11238 /* Send a netdev-removed uevent to the old namespace */
11239 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE);
11240 netdev_adjacent_del_links(dev);
11242 /* Move per-net netdevice notifiers that are following the netdevice */
11243 move_netdevice_notifiers_dev_net(dev, net);
11245 /* Actually switch the network namespace */
11246 dev_net_set(dev, net);
11247 dev->ifindex = new_ifindex;
11249 /* Send a netdev-add uevent to the new namespace */
11250 kobject_uevent(&dev->dev.kobj, KOBJ_ADD);
11251 netdev_adjacent_add_links(dev);
11253 /* Fixup kobjects */
11254 err = device_rename(&dev->dev, dev->name);
11257 /* Adapt owner in case owning user namespace of target network
11258 * namespace is different from the original one.
11260 err = netdev_change_owner(dev, net_old, net);
11263 /* Add the device back in the hashes */
11264 list_netdevice(dev);
11266 /* Notify protocols, that a new device appeared. */
11267 call_netdevice_notifiers(NETDEV_REGISTER, dev);
11270 * Prevent userspace races by waiting until the network
11271 * device is fully setup before sending notifications.
11273 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL);
11280 EXPORT_SYMBOL_GPL(__dev_change_net_namespace);
11282 static int dev_cpu_dead(unsigned int oldcpu)
11284 struct sk_buff **list_skb;
11285 struct sk_buff *skb;
11287 struct softnet_data *sd, *oldsd, *remsd = NULL;
11289 local_irq_disable();
11290 cpu = smp_processor_id();
11291 sd = &per_cpu(softnet_data, cpu);
11292 oldsd = &per_cpu(softnet_data, oldcpu);
11294 /* Find end of our completion_queue. */
11295 list_skb = &sd->completion_queue;
11297 list_skb = &(*list_skb)->next;
11298 /* Append completion queue from offline CPU. */
11299 *list_skb = oldsd->completion_queue;
11300 oldsd->completion_queue = NULL;
11302 /* Append output queue from offline CPU. */
11303 if (oldsd->output_queue) {
11304 *sd->output_queue_tailp = oldsd->output_queue;
11305 sd->output_queue_tailp = oldsd->output_queue_tailp;
11306 oldsd->output_queue = NULL;
11307 oldsd->output_queue_tailp = &oldsd->output_queue;
11309 /* Append NAPI poll list from offline CPU, with one exception :
11310 * process_backlog() must be called by cpu owning percpu backlog.
11311 * We properly handle process_queue & input_pkt_queue later.
11313 while (!list_empty(&oldsd->poll_list)) {
11314 struct napi_struct *napi = list_first_entry(&oldsd->poll_list,
11315 struct napi_struct,
11318 list_del_init(&napi->poll_list);
11319 if (napi->poll == process_backlog)
11322 ____napi_schedule(sd, napi);
11325 raise_softirq_irqoff(NET_TX_SOFTIRQ);
11326 local_irq_enable();
11329 remsd = oldsd->rps_ipi_list;
11330 oldsd->rps_ipi_list = NULL;
11332 /* send out pending IPI's on offline CPU */
11333 net_rps_send_ipi(remsd);
11335 /* Process offline CPU's input_pkt_queue */
11336 while ((skb = __skb_dequeue(&oldsd->process_queue))) {
11338 input_queue_head_incr(oldsd);
11340 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) {
11342 input_queue_head_incr(oldsd);
11349 * netdev_increment_features - increment feature set by one
11350 * @all: current feature set
11351 * @one: new feature set
11352 * @mask: mask feature set
11354 * Computes a new feature set after adding a device with feature set
11355 * @one to the master device with current feature set @all. Will not
11356 * enable anything that is off in @mask. Returns the new feature set.
11358 netdev_features_t netdev_increment_features(netdev_features_t all,
11359 netdev_features_t one, netdev_features_t mask)
11361 if (mask & NETIF_F_HW_CSUM)
11362 mask |= NETIF_F_CSUM_MASK;
11363 mask |= NETIF_F_VLAN_CHALLENGED;
11365 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask;
11366 all &= one | ~NETIF_F_ALL_FOR_ALL;
11368 /* If one device supports hw checksumming, set for all. */
11369 if (all & NETIF_F_HW_CSUM)
11370 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM);
11374 EXPORT_SYMBOL(netdev_increment_features);
11376 static struct hlist_head * __net_init netdev_create_hash(void)
11379 struct hlist_head *hash;
11381 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL);
11383 for (i = 0; i < NETDEV_HASHENTRIES; i++)
11384 INIT_HLIST_HEAD(&hash[i]);
11389 /* Initialize per network namespace state */
11390 static int __net_init netdev_init(struct net *net)
11392 BUILD_BUG_ON(GRO_HASH_BUCKETS >
11393 8 * sizeof_field(struct napi_struct, gro_bitmask));
11395 if (net != &init_net)
11396 INIT_LIST_HEAD(&net->dev_base_head);
11398 net->dev_name_head = netdev_create_hash();
11399 if (net->dev_name_head == NULL)
11402 net->dev_index_head = netdev_create_hash();
11403 if (net->dev_index_head == NULL)
11406 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain);
11411 kfree(net->dev_name_head);
11417 * netdev_drivername - network driver for the device
11418 * @dev: network device
11420 * Determine network driver for device.
11422 const char *netdev_drivername(const struct net_device *dev)
11424 const struct device_driver *driver;
11425 const struct device *parent;
11426 const char *empty = "";
11428 parent = dev->dev.parent;
11432 driver = parent->driver;
11433 if (driver && driver->name)
11434 return driver->name;
11438 static void __netdev_printk(const char *level, const struct net_device *dev,
11439 struct va_format *vaf)
11441 if (dev && dev->dev.parent) {
11442 dev_printk_emit(level[1] - '0',
11445 dev_driver_string(dev->dev.parent),
11446 dev_name(dev->dev.parent),
11447 netdev_name(dev), netdev_reg_state(dev),
11450 printk("%s%s%s: %pV",
11451 level, netdev_name(dev), netdev_reg_state(dev), vaf);
11453 printk("%s(NULL net_device): %pV", level, vaf);
11457 void netdev_printk(const char *level, const struct net_device *dev,
11458 const char *format, ...)
11460 struct va_format vaf;
11463 va_start(args, format);
11468 __netdev_printk(level, dev, &vaf);
11472 EXPORT_SYMBOL(netdev_printk);
11474 #define define_netdev_printk_level(func, level) \
11475 void func(const struct net_device *dev, const char *fmt, ...) \
11477 struct va_format vaf; \
11480 va_start(args, fmt); \
11485 __netdev_printk(level, dev, &vaf); \
11489 EXPORT_SYMBOL(func);
11491 define_netdev_printk_level(netdev_emerg, KERN_EMERG);
11492 define_netdev_printk_level(netdev_alert, KERN_ALERT);
11493 define_netdev_printk_level(netdev_crit, KERN_CRIT);
11494 define_netdev_printk_level(netdev_err, KERN_ERR);
11495 define_netdev_printk_level(netdev_warn, KERN_WARNING);
11496 define_netdev_printk_level(netdev_notice, KERN_NOTICE);
11497 define_netdev_printk_level(netdev_info, KERN_INFO);
11499 static void __net_exit netdev_exit(struct net *net)
11501 kfree(net->dev_name_head);
11502 kfree(net->dev_index_head);
11503 if (net != &init_net)
11504 WARN_ON_ONCE(!list_empty(&net->dev_base_head));
11507 static struct pernet_operations __net_initdata netdev_net_ops = {
11508 .init = netdev_init,
11509 .exit = netdev_exit,
11512 static void __net_exit default_device_exit(struct net *net)
11514 struct net_device *dev, *aux;
11516 * Push all migratable network devices back to the
11517 * initial network namespace
11520 for_each_netdev_safe(net, dev, aux) {
11522 char fb_name[IFNAMSIZ];
11524 /* Ignore unmoveable devices (i.e. loopback) */
11525 if (dev->features & NETIF_F_NETNS_LOCAL)
11528 /* Leave virtual devices for the generic cleanup */
11529 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund)
11532 /* Push remaining network devices to init_net */
11533 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex);
11534 if (__dev_get_by_name(&init_net, fb_name))
11535 snprintf(fb_name, IFNAMSIZ, "dev%%d");
11536 err = dev_change_net_namespace(dev, &init_net, fb_name);
11538 pr_emerg("%s: failed to move %s to init_net: %d\n",
11539 __func__, dev->name, err);
11546 static void __net_exit rtnl_lock_unregistering(struct list_head *net_list)
11548 /* Return with the rtnl_lock held when there are no network
11549 * devices unregistering in any network namespace in net_list.
11552 bool unregistering;
11553 DEFINE_WAIT_FUNC(wait, woken_wake_function);
11555 add_wait_queue(&netdev_unregistering_wq, &wait);
11557 unregistering = false;
11559 list_for_each_entry(net, net_list, exit_list) {
11560 if (net->dev_unreg_count > 0) {
11561 unregistering = true;
11565 if (!unregistering)
11569 wait_woken(&wait, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
11571 remove_wait_queue(&netdev_unregistering_wq, &wait);
11574 static void __net_exit default_device_exit_batch(struct list_head *net_list)
11576 /* At exit all network devices most be removed from a network
11577 * namespace. Do this in the reverse order of registration.
11578 * Do this across as many network namespaces as possible to
11579 * improve batching efficiency.
11581 struct net_device *dev;
11583 LIST_HEAD(dev_kill_list);
11585 /* To prevent network device cleanup code from dereferencing
11586 * loopback devices or network devices that have been freed
11587 * wait here for all pending unregistrations to complete,
11588 * before unregistring the loopback device and allowing the
11589 * network namespace be freed.
11591 * The netdev todo list containing all network devices
11592 * unregistrations that happen in default_device_exit_batch
11593 * will run in the rtnl_unlock() at the end of
11594 * default_device_exit_batch.
11596 rtnl_lock_unregistering(net_list);
11597 list_for_each_entry(net, net_list, exit_list) {
11598 for_each_netdev_reverse(net, dev) {
11599 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink)
11600 dev->rtnl_link_ops->dellink(dev, &dev_kill_list);
11602 unregister_netdevice_queue(dev, &dev_kill_list);
11605 unregister_netdevice_many(&dev_kill_list);
11609 static struct pernet_operations __net_initdata default_device_ops = {
11610 .exit = default_device_exit,
11611 .exit_batch = default_device_exit_batch,
11615 * Initialize the DEV module. At boot time this walks the device list and
11616 * unhooks any devices that fail to initialise (normally hardware not
11617 * present) and leaves us with a valid list of present and active devices.
11622 * This is called single threaded during boot, so no need
11623 * to take the rtnl semaphore.
11625 static int __init net_dev_init(void)
11627 int i, rc = -ENOMEM;
11629 BUG_ON(!dev_boot_phase);
11631 if (dev_proc_init())
11634 if (netdev_kobject_init())
11637 INIT_LIST_HEAD(&ptype_all);
11638 for (i = 0; i < PTYPE_HASH_SIZE; i++)
11639 INIT_LIST_HEAD(&ptype_base[i]);
11641 INIT_LIST_HEAD(&offload_base);
11643 if (register_pernet_subsys(&netdev_net_ops))
11647 * Initialise the packet receive queues.
11650 for_each_possible_cpu(i) {
11651 struct work_struct *flush = per_cpu_ptr(&flush_works, i);
11652 struct softnet_data *sd = &per_cpu(softnet_data, i);
11654 INIT_WORK(flush, flush_backlog);
11656 skb_queue_head_init(&sd->input_pkt_queue);
11657 skb_queue_head_init(&sd->process_queue);
11658 #ifdef CONFIG_XFRM_OFFLOAD
11659 skb_queue_head_init(&sd->xfrm_backlog);
11661 INIT_LIST_HEAD(&sd->poll_list);
11662 sd->output_queue_tailp = &sd->output_queue;
11664 INIT_CSD(&sd->csd, rps_trigger_softirq, sd);
11668 init_gro_hash(&sd->backlog);
11669 sd->backlog.poll = process_backlog;
11670 sd->backlog.weight = weight_p;
11673 dev_boot_phase = 0;
11675 /* The loopback device is special if any other network devices
11676 * is present in a network namespace the loopback device must
11677 * be present. Since we now dynamically allocate and free the
11678 * loopback device ensure this invariant is maintained by
11679 * keeping the loopback device as the first device on the
11680 * list of network devices. Ensuring the loopback devices
11681 * is the first device that appears and the last network device
11684 if (register_pernet_device(&loopback_net_ops))
11687 if (register_pernet_device(&default_device_ops))
11690 open_softirq(NET_TX_SOFTIRQ, net_tx_action);
11691 open_softirq(NET_RX_SOFTIRQ, net_rx_action);
11693 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead",
11694 NULL, dev_cpu_dead);
11701 subsys_initcall(net_dev_init);